• 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 #include "vp9/encoder/vp9_cost.h"
35 #include "vp9/encoder/vp9_encodemb.h"
36 #include "vp9/encoder/vp9_encodemv.h"
37 #include "vp9/encoder/vp9_encoder.h"
38 #include "vp9/encoder/vp9_mcomp.h"
39 #include "vp9/encoder/vp9_quantize.h"
40 #include "vp9/encoder/vp9_ratectrl.h"
41 #include "vp9/encoder/vp9_rd.h"
42 #include "vp9/encoder/vp9_rdopt.h"
43 #include "vp9/encoder/vp9_aq_variance.h"
44 
45 #define LAST_FRAME_MODE_MASK \
46   ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
47 #define GOLDEN_FRAME_MODE_MASK \
48   ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
49 #define ALT_REF_MODE_MASK \
50   ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | (1 << INTRA_FRAME))
51 
52 #define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
53 
54 #define MIN_EARLY_TERM_INDEX 3
55 #define NEW_MV_DISCOUNT_FACTOR 8
56 
57 typedef struct {
58   PREDICTION_MODE mode;
59   MV_REFERENCE_FRAME ref_frame[2];
60 } MODE_DEFINITION;
61 
62 typedef struct {
63   MV_REFERENCE_FRAME ref_frame[2];
64 } REF_DEFINITION;
65 
66 struct rdcost_block_args {
67   const VP9_COMP *cpi;
68   MACROBLOCK *x;
69   ENTROPY_CONTEXT t_above[16];
70   ENTROPY_CONTEXT t_left[16];
71   int this_rate;
72   int64_t this_dist;
73   int64_t this_sse;
74   int64_t this_rd;
75   int64_t best_rd;
76   int exit_early;
77   int use_fast_coef_costing;
78   const scan_order *so;
79   uint8_t skippable;
80 };
81 
82 #define LAST_NEW_MV_INDEX 6
83 static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
84   { NEARESTMV, { LAST_FRAME, NONE } },
85   { NEARESTMV, { ALTREF_FRAME, NONE } },
86   { NEARESTMV, { GOLDEN_FRAME, NONE } },
87 
88   { DC_PRED, { INTRA_FRAME, NONE } },
89 
90   { NEWMV, { LAST_FRAME, NONE } },
91   { NEWMV, { ALTREF_FRAME, NONE } },
92   { NEWMV, { GOLDEN_FRAME, NONE } },
93 
94   { NEARMV, { LAST_FRAME, NONE } },
95   { NEARMV, { ALTREF_FRAME, NONE } },
96   { NEARMV, { GOLDEN_FRAME, NONE } },
97 
98   { ZEROMV, { LAST_FRAME, NONE } },
99   { ZEROMV, { GOLDEN_FRAME, NONE } },
100   { ZEROMV, { ALTREF_FRAME, NONE } },
101 
102   { NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
103   { NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
104 
105   { TM_PRED, { INTRA_FRAME, NONE } },
106 
107   { NEARMV, { LAST_FRAME, ALTREF_FRAME } },
108   { NEWMV, { LAST_FRAME, ALTREF_FRAME } },
109   { NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
110   { NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
111 
112   { ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
113   { ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },
114 
115   { H_PRED, { INTRA_FRAME, NONE } },
116   { V_PRED, { INTRA_FRAME, NONE } },
117   { D135_PRED, { INTRA_FRAME, NONE } },
118   { D207_PRED, { INTRA_FRAME, NONE } },
119   { D153_PRED, { INTRA_FRAME, NONE } },
120   { D63_PRED, { INTRA_FRAME, NONE } },
121   { D117_PRED, { INTRA_FRAME, NONE } },
122   { D45_PRED, { INTRA_FRAME, NONE } },
123 };
124 
125 static const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
126   { { LAST_FRAME, NONE } },           { { GOLDEN_FRAME, NONE } },
127   { { ALTREF_FRAME, NONE } },         { { LAST_FRAME, ALTREF_FRAME } },
128   { { GOLDEN_FRAME, ALTREF_FRAME } }, { { INTRA_FRAME, NONE } },
129 };
130 
swap_block_ptr(MACROBLOCK * x,PICK_MODE_CONTEXT * ctx,int m,int n,int min_plane,int max_plane)131 static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx, int m, int n,
132                            int min_plane, int max_plane) {
133   int i;
134 
135   for (i = min_plane; i < max_plane; ++i) {
136     struct macroblock_plane *const p = &x->plane[i];
137     struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
138 
139     p->coeff = ctx->coeff_pbuf[i][m];
140     p->qcoeff = ctx->qcoeff_pbuf[i][m];
141     pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
142     p->eobs = ctx->eobs_pbuf[i][m];
143 
144     ctx->coeff_pbuf[i][m] = ctx->coeff_pbuf[i][n];
145     ctx->qcoeff_pbuf[i][m] = ctx->qcoeff_pbuf[i][n];
146     ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
147     ctx->eobs_pbuf[i][m] = ctx->eobs_pbuf[i][n];
148 
149     ctx->coeff_pbuf[i][n] = p->coeff;
150     ctx->qcoeff_pbuf[i][n] = p->qcoeff;
151     ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
152     ctx->eobs_pbuf[i][n] = p->eobs;
153   }
154 }
155 
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)156 static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
157                             MACROBLOCKD *xd, int *out_rate_sum,
158                             int64_t *out_dist_sum, int *skip_txfm_sb,
159                             int64_t *skip_sse_sb) {
160   // Note our transform coeffs are 8 times an orthogonal transform.
161   // Hence quantizer step is also 8 times. To get effective quantizer
162   // we need to divide by 8 before sending to modeling function.
163   int i;
164   int64_t rate_sum = 0;
165   int64_t dist_sum = 0;
166   const int ref = xd->mi[0]->ref_frame[0];
167   unsigned int sse;
168   unsigned int var = 0;
169   int64_t total_sse = 0;
170   int skip_flag = 1;
171   const int shift = 6;
172   int64_t dist;
173   const int dequant_shift =
174 #if CONFIG_VP9_HIGHBITDEPTH
175       (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 :
176 #endif  // CONFIG_VP9_HIGHBITDEPTH
177                                                     3;
178   unsigned int qstep_vec[MAX_MB_PLANE];
179   unsigned int nlog2_vec[MAX_MB_PLANE];
180   unsigned int sum_sse_vec[MAX_MB_PLANE];
181   int any_zero_sum_sse = 0;
182 
183   x->pred_sse[ref] = 0;
184 
185   for (i = 0; i < MAX_MB_PLANE; ++i) {
186     struct macroblock_plane *const p = &x->plane[i];
187     struct macroblockd_plane *const pd = &xd->plane[i];
188     const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
189     const TX_SIZE max_tx_size = max_txsize_lookup[bs];
190     const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
191     const int64_t dc_thr = p->quant_thred[0] >> shift;
192     const int64_t ac_thr = p->quant_thred[1] >> shift;
193     unsigned int sum_sse = 0;
194     // The low thresholds are used to measure if the prediction errors are
195     // low enough so that we can skip the mode search.
196     const int64_t low_dc_thr = VPXMIN(50, dc_thr >> 2);
197     const int64_t low_ac_thr = VPXMIN(80, ac_thr >> 2);
198     int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
199     int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
200     int idx, idy;
201     int lw = b_width_log2_lookup[unit_size] + 2;
202     int lh = b_height_log2_lookup[unit_size] + 2;
203 
204     for (idy = 0; idy < bh; ++idy) {
205       for (idx = 0; idx < bw; ++idx) {
206         uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
207         uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
208         int block_idx = (idy << 1) + idx;
209         int low_err_skip = 0;
210 
211         var = cpi->fn_ptr[unit_size].vf(src, p->src.stride, dst, pd->dst.stride,
212                                         &sse);
213         x->bsse[(i << 2) + block_idx] = sse;
214         sum_sse += sse;
215 
216         x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_NONE;
217         if (!x->select_tx_size) {
218           // Check if all ac coefficients can be quantized to zero.
219           if (var < ac_thr || var == 0) {
220             x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_ONLY;
221 
222             // Check if dc coefficient can be quantized to zero.
223             if (sse - var < dc_thr || sse == var) {
224               x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_DC;
225 
226               if (!sse || (var < low_ac_thr && sse - var < low_dc_thr))
227                 low_err_skip = 1;
228             }
229           }
230         }
231 
232         if (skip_flag && !low_err_skip) skip_flag = 0;
233 
234         if (i == 0) x->pred_sse[ref] += sse;
235       }
236     }
237 
238     total_sse += sum_sse;
239     sum_sse_vec[i] = sum_sse;
240     any_zero_sum_sse = any_zero_sum_sse || (sum_sse == 0);
241     qstep_vec[i] = pd->dequant[1] >> dequant_shift;
242     nlog2_vec[i] = num_pels_log2_lookup[bs];
243   }
244 
245   // Fast approximate the modelling function.
246   if (cpi->sf.simple_model_rd_from_var) {
247     for (i = 0; i < MAX_MB_PLANE; ++i) {
248       int64_t rate;
249       const int64_t square_error = sum_sse_vec[i];
250       int quantizer = qstep_vec[i];
251 
252       if (quantizer < 120)
253         rate = (square_error * (280 - quantizer)) >> (16 - VP9_PROB_COST_SHIFT);
254       else
255         rate = 0;
256       dist = (square_error * quantizer) >> 8;
257       rate_sum += rate;
258       dist_sum += dist;
259     }
260   } else {
261     if (any_zero_sum_sse) {
262       for (i = 0; i < MAX_MB_PLANE; ++i) {
263         int rate;
264         vp9_model_rd_from_var_lapndz(sum_sse_vec[i], nlog2_vec[i], qstep_vec[i],
265                                      &rate, &dist);
266         rate_sum += rate;
267         dist_sum += dist;
268       }
269     } else {
270       vp9_model_rd_from_var_lapndz_vec(sum_sse_vec, nlog2_vec, qstep_vec,
271                                        &rate_sum, &dist_sum);
272     }
273   }
274 
275   *skip_txfm_sb = skip_flag;
276   *skip_sse_sb = total_sse << 4;
277   *out_rate_sum = (int)rate_sum;
278   *out_dist_sum = dist_sum << 4;
279 }
280 
281 #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)282 int64_t vp9_highbd_block_error_c(const tran_low_t *coeff,
283                                  const tran_low_t *dqcoeff, intptr_t block_size,
284                                  int64_t *ssz, int bd) {
285   int i;
286   int64_t error = 0, sqcoeff = 0;
287   int shift = 2 * (bd - 8);
288   int rounding = shift > 0 ? 1 << (shift - 1) : 0;
289 
290   for (i = 0; i < block_size; i++) {
291     const int64_t diff = coeff[i] - dqcoeff[i];
292     error += diff * diff;
293     sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
294   }
295   assert(error >= 0 && sqcoeff >= 0);
296   error = (error + rounding) >> shift;
297   sqcoeff = (sqcoeff + rounding) >> shift;
298 
299   *ssz = sqcoeff;
300   return error;
301 }
302 
vp9_highbd_block_error_dispatch(const tran_low_t * coeff,const tran_low_t * dqcoeff,intptr_t block_size,int64_t * ssz,int bd)303 static int64_t vp9_highbd_block_error_dispatch(const tran_low_t *coeff,
304                                                const tran_low_t *dqcoeff,
305                                                intptr_t block_size,
306                                                int64_t *ssz, int bd) {
307   if (bd == 8) {
308     return vp9_block_error(coeff, dqcoeff, block_size, ssz);
309   } else {
310     return vp9_highbd_block_error(coeff, dqcoeff, block_size, ssz, bd);
311   }
312 }
313 #endif  // CONFIG_VP9_HIGHBITDEPTH
314 
vp9_block_error_c(const tran_low_t * coeff,const tran_low_t * dqcoeff,intptr_t block_size,int64_t * ssz)315 int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
316                           intptr_t block_size, int64_t *ssz) {
317   int i;
318   int64_t error = 0, sqcoeff = 0;
319 
320   for (i = 0; i < block_size; i++) {
321     const int diff = coeff[i] - dqcoeff[i];
322     error += diff * diff;
323     sqcoeff += coeff[i] * coeff[i];
324   }
325 
326   *ssz = sqcoeff;
327   return error;
328 }
329 
vp9_block_error_fp_c(const tran_low_t * coeff,const tran_low_t * dqcoeff,int block_size)330 int64_t vp9_block_error_fp_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
331                              int block_size) {
332   int i;
333   int64_t error = 0;
334 
335   for (i = 0; i < block_size; i++) {
336     const int diff = coeff[i] - dqcoeff[i];
337     error += diff * diff;
338   }
339 
340   return error;
341 }
342 
343 /* The trailing '0' is a terminator which is used inside cost_coeffs() to
344  * decide whether to include cost of a trailing EOB node or not (i.e. we
345  * can skip this if the last coefficient in this transform block, e.g. the
346  * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
347  * were non-zero). */
348 static const int16_t band_counts[TX_SIZES][8] = {
349   { 1, 2, 3, 4, 3, 16 - 13, 0 },
350   { 1, 2, 3, 4, 11, 64 - 21, 0 },
351   { 1, 2, 3, 4, 11, 256 - 21, 0 },
352   { 1, 2, 3, 4, 11, 1024 - 21, 0 },
353 };
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)354 static int cost_coeffs(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
355                        int pt, const int16_t *scan, const int16_t *nb,
356                        int use_fast_coef_costing) {
357   MACROBLOCKD *const xd = &x->e_mbd;
358   MODE_INFO *mi = xd->mi[0];
359   const struct macroblock_plane *p = &x->plane[plane];
360   const PLANE_TYPE type = get_plane_type(plane);
361   const int16_t *band_count = &band_counts[tx_size][1];
362   const int eob = p->eobs[block];
363   const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
364   unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
365       x->token_costs[tx_size][type][is_inter_block(mi)];
366   uint8_t token_cache[32 * 32];
367   int cost;
368 #if CONFIG_VP9_HIGHBITDEPTH
369   const uint16_t *cat6_high_cost = vp9_get_high_cost_table(xd->bd);
370 #else
371   const uint16_t *cat6_high_cost = vp9_get_high_cost_table(8);
372 #endif
373 
374   // Check for consistency of tx_size with mode info
375   assert(type == PLANE_TYPE_Y
376              ? mi->tx_size == tx_size
377              : get_uv_tx_size(mi, &xd->plane[plane]) == tx_size);
378 
379   if (eob == 0) {
380     // single eob token
381     cost = token_costs[0][0][pt][EOB_TOKEN];
382   } else {
383     if (use_fast_coef_costing) {
384       int band_left = *band_count++;
385       int c;
386 
387       // dc token
388       int v = qcoeff[0];
389       int16_t prev_t;
390       cost = vp9_get_token_cost(v, &prev_t, cat6_high_cost);
391       cost += (*token_costs)[0][pt][prev_t];
392 
393       token_cache[0] = vp9_pt_energy_class[prev_t];
394       ++token_costs;
395 
396       // ac tokens
397       for (c = 1; c < eob; c++) {
398         const int rc = scan[c];
399         int16_t t;
400 
401         v = qcoeff[rc];
402         cost += vp9_get_token_cost(v, &t, cat6_high_cost);
403         cost += (*token_costs)[!prev_t][!prev_t][t];
404         prev_t = t;
405         if (!--band_left) {
406           band_left = *band_count++;
407           ++token_costs;
408         }
409       }
410 
411       // eob token
412       if (band_left) cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
413 
414     } else {  // !use_fast_coef_costing
415       int band_left = *band_count++;
416       int c;
417 
418       // dc token
419       int v = qcoeff[0];
420       int16_t tok;
421       unsigned int(*tok_cost_ptr)[COEFF_CONTEXTS][ENTROPY_TOKENS];
422       cost = vp9_get_token_cost(v, &tok, cat6_high_cost);
423       cost += (*token_costs)[0][pt][tok];
424 
425       token_cache[0] = vp9_pt_energy_class[tok];
426       ++token_costs;
427 
428       tok_cost_ptr = &((*token_costs)[!tok]);
429 
430       // ac tokens
431       for (c = 1; c < eob; c++) {
432         const int rc = scan[c];
433 
434         v = qcoeff[rc];
435         cost += vp9_get_token_cost(v, &tok, cat6_high_cost);
436         pt = get_coef_context(nb, token_cache, c);
437         cost += (*tok_cost_ptr)[pt][tok];
438         token_cache[rc] = vp9_pt_energy_class[tok];
439         if (!--band_left) {
440           band_left = *band_count++;
441           ++token_costs;
442         }
443         tok_cost_ptr = &((*token_costs)[!tok]);
444       }
445 
446       // eob token
447       if (band_left) {
448         pt = get_coef_context(nb, token_cache, c);
449         cost += (*token_costs)[0][pt][EOB_TOKEN];
450       }
451     }
452   }
453 
454   return cost;
455 }
456 
num_4x4_to_edge(int plane_4x4_dim,int mb_to_edge_dim,int subsampling_dim,int blk_dim)457 static INLINE int num_4x4_to_edge(int plane_4x4_dim, int mb_to_edge_dim,
458                                   int subsampling_dim, int blk_dim) {
459   return plane_4x4_dim + (mb_to_edge_dim >> (5 + subsampling_dim)) - blk_dim;
460 }
461 
462 // Compute the pixel domain sum square error on all visible 4x4s in the
463 // 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)464 static unsigned pixel_sse(const VP9_COMP *const cpi, const MACROBLOCKD *xd,
465                           const struct macroblockd_plane *const pd,
466                           const uint8_t *src, const int src_stride,
467                           const uint8_t *dst, const int dst_stride, int blk_row,
468                           int blk_col, const BLOCK_SIZE plane_bsize,
469                           const BLOCK_SIZE tx_bsize) {
470   unsigned int sse = 0;
471   const int plane_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
472   const int plane_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
473   const int tx_4x4_w = num_4x4_blocks_wide_lookup[tx_bsize];
474   const int tx_4x4_h = num_4x4_blocks_high_lookup[tx_bsize];
475   int b4x4s_to_right_edge = num_4x4_to_edge(plane_4x4_w, xd->mb_to_right_edge,
476                                             pd->subsampling_x, blk_col);
477   int b4x4s_to_bottom_edge = num_4x4_to_edge(plane_4x4_h, xd->mb_to_bottom_edge,
478                                              pd->subsampling_y, blk_row);
479   if (tx_bsize == BLOCK_4X4 ||
480       (b4x4s_to_right_edge >= tx_4x4_w && b4x4s_to_bottom_edge >= tx_4x4_h)) {
481     cpi->fn_ptr[tx_bsize].vf(src, src_stride, dst, dst_stride, &sse);
482   } else {
483     const vpx_variance_fn_t vf_4x4 = cpi->fn_ptr[BLOCK_4X4].vf;
484     int r, c;
485     unsigned this_sse = 0;
486     int max_r = VPXMIN(b4x4s_to_bottom_edge, tx_4x4_h);
487     int max_c = VPXMIN(b4x4s_to_right_edge, tx_4x4_w);
488     sse = 0;
489     // if we are in the unrestricted motion border.
490     for (r = 0; r < max_r; ++r) {
491       // Skip visiting the sub blocks that are wholly within the UMV.
492       for (c = 0; c < max_c; ++c) {
493         vf_4x4(src + r * src_stride * 4 + c * 4, src_stride,
494                dst + r * dst_stride * 4 + c * 4, dst_stride, &this_sse);
495         sse += this_sse;
496       }
497     }
498   }
499   return sse;
500 }
501 
502 // 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)503 static int64_t sum_squares_visible(const MACROBLOCKD *xd,
504                                    const struct macroblockd_plane *const pd,
505                                    const int16_t *diff, const int diff_stride,
506                                    int blk_row, int blk_col,
507                                    const BLOCK_SIZE plane_bsize,
508                                    const BLOCK_SIZE tx_bsize) {
509   int64_t sse;
510   const int plane_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
511   const int plane_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
512   const int tx_4x4_w = num_4x4_blocks_wide_lookup[tx_bsize];
513   const int tx_4x4_h = num_4x4_blocks_high_lookup[tx_bsize];
514   int b4x4s_to_right_edge = num_4x4_to_edge(plane_4x4_w, xd->mb_to_right_edge,
515                                             pd->subsampling_x, blk_col);
516   int b4x4s_to_bottom_edge = num_4x4_to_edge(plane_4x4_h, xd->mb_to_bottom_edge,
517                                              pd->subsampling_y, blk_row);
518   if (tx_bsize == BLOCK_4X4 ||
519       (b4x4s_to_right_edge >= tx_4x4_w && b4x4s_to_bottom_edge >= tx_4x4_h)) {
520     assert(tx_4x4_w == tx_4x4_h);
521     sse = (int64_t)vpx_sum_squares_2d_i16(diff, diff_stride, tx_4x4_w << 2);
522   } else {
523     int r, c;
524     int max_r = VPXMIN(b4x4s_to_bottom_edge, tx_4x4_h);
525     int max_c = VPXMIN(b4x4s_to_right_edge, tx_4x4_w);
526     sse = 0;
527     // if we are in the unrestricted motion border.
528     for (r = 0; r < max_r; ++r) {
529       // Skip visiting the sub blocks that are wholly within the UMV.
530       for (c = 0; c < max_c; ++c) {
531         sse += (int64_t)vpx_sum_squares_2d_i16(
532             diff + r * diff_stride * 4 + c * 4, diff_stride, 4);
533       }
534     }
535   }
536   return sse;
537 }
538 
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)539 static void dist_block(const VP9_COMP *cpi, MACROBLOCK *x, int plane,
540                        BLOCK_SIZE plane_bsize, int block, int blk_row,
541                        int blk_col, TX_SIZE tx_size, int64_t *out_dist,
542                        int64_t *out_sse) {
543   MACROBLOCKD *const xd = &x->e_mbd;
544   const struct macroblock_plane *const p = &x->plane[plane];
545   const struct macroblockd_plane *const pd = &xd->plane[plane];
546   const int eob = p->eobs[block];
547 
548   if (x->block_tx_domain && eob) {
549     const int ss_txfrm_size = tx_size << 1;
550     int64_t this_sse;
551     const int shift = tx_size == TX_32X32 ? 0 : 2;
552     const tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
553     const tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
554 #if CONFIG_VP9_HIGHBITDEPTH
555     const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
556     *out_dist = vp9_highbd_block_error_dispatch(
557                     coeff, dqcoeff, 16 << ss_txfrm_size, &this_sse, bd) >>
558                 shift;
559 #else
560     *out_dist =
561         vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size, &this_sse) >>
562         shift;
563 #endif  // CONFIG_VP9_HIGHBITDEPTH
564     *out_sse = this_sse >> shift;
565 
566     if (x->skip_encode && !is_inter_block(xd->mi[0])) {
567       // TODO(jingning): tune the model to better capture the distortion.
568       const int64_t p =
569           (pd->dequant[1] * pd->dequant[1] * (1 << ss_txfrm_size)) >>
570 #if CONFIG_VP9_HIGHBITDEPTH
571           (shift + 2 + (bd - 8) * 2);
572 #else
573           (shift + 2);
574 #endif  // CONFIG_VP9_HIGHBITDEPTH
575       *out_dist += (p >> 4);
576       *out_sse += p;
577     }
578   } else {
579     const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
580     const int bs = 4 * num_4x4_blocks_wide_lookup[tx_bsize];
581     const int src_stride = p->src.stride;
582     const int dst_stride = pd->dst.stride;
583     const int src_idx = 4 * (blk_row * src_stride + blk_col);
584     const int dst_idx = 4 * (blk_row * dst_stride + blk_col);
585     const uint8_t *src = &p->src.buf[src_idx];
586     const uint8_t *dst = &pd->dst.buf[dst_idx];
587     const tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
588     unsigned int tmp;
589 
590     tmp = pixel_sse(cpi, xd, pd, src, src_stride, dst, dst_stride, blk_row,
591                     blk_col, plane_bsize, tx_bsize);
592     *out_sse = (int64_t)tmp * 16;
593 
594     if (eob) {
595 #if CONFIG_VP9_HIGHBITDEPTH
596       DECLARE_ALIGNED(16, uint16_t, recon16[1024]);
597       uint8_t *recon = (uint8_t *)recon16;
598 #else
599       DECLARE_ALIGNED(16, uint8_t, recon[1024]);
600 #endif  // CONFIG_VP9_HIGHBITDEPTH
601 
602 #if CONFIG_VP9_HIGHBITDEPTH
603       if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
604         vpx_highbd_convolve_copy(CONVERT_TO_SHORTPTR(dst), dst_stride, recon16,
605                                  32, NULL, 0, 0, 0, 0, bs, bs, xd->bd);
606         if (xd->lossless) {
607           vp9_highbd_iwht4x4_add(dqcoeff, recon16, 32, eob, xd->bd);
608         } else {
609           switch (tx_size) {
610             case TX_4X4:
611               vp9_highbd_idct4x4_add(dqcoeff, recon16, 32, eob, xd->bd);
612               break;
613             case TX_8X8:
614               vp9_highbd_idct8x8_add(dqcoeff, recon16, 32, eob, xd->bd);
615               break;
616             case TX_16X16:
617               vp9_highbd_idct16x16_add(dqcoeff, recon16, 32, eob, xd->bd);
618               break;
619             default:
620               assert(tx_size == TX_32X32);
621               vp9_highbd_idct32x32_add(dqcoeff, recon16, 32, eob, xd->bd);
622               break;
623           }
624         }
625         recon = CONVERT_TO_BYTEPTR(recon16);
626       } else {
627 #endif  // CONFIG_VP9_HIGHBITDEPTH
628         vpx_convolve_copy(dst, dst_stride, recon, 32, NULL, 0, 0, 0, 0, bs, bs);
629         switch (tx_size) {
630           case TX_32X32: vp9_idct32x32_add(dqcoeff, recon, 32, eob); break;
631           case TX_16X16: vp9_idct16x16_add(dqcoeff, recon, 32, eob); break;
632           case TX_8X8: vp9_idct8x8_add(dqcoeff, recon, 32, eob); break;
633           default:
634             assert(tx_size == TX_4X4);
635             // this is like vp9_short_idct4x4 but has a special case around
636             // eob<=1, which is significant (not just an optimization) for
637             // the lossless case.
638             x->inv_txfm_add(dqcoeff, recon, 32, eob);
639             break;
640         }
641 #if CONFIG_VP9_HIGHBITDEPTH
642       }
643 #endif  // CONFIG_VP9_HIGHBITDEPTH
644 
645       tmp = pixel_sse(cpi, xd, pd, src, src_stride, recon, 32, blk_row, blk_col,
646                       plane_bsize, tx_bsize);
647     }
648 
649     *out_dist = (int64_t)tmp * 16;
650   }
651 }
652 
rate_block(int plane,int block,TX_SIZE tx_size,int coeff_ctx,struct rdcost_block_args * args)653 static int rate_block(int plane, int block, TX_SIZE tx_size, int coeff_ctx,
654                       struct rdcost_block_args *args) {
655   return cost_coeffs(args->x, plane, block, tx_size, coeff_ctx, args->so->scan,
656                      args->so->neighbors, args->use_fast_coef_costing);
657 }
658 
block_rd_txfm(int plane,int block,int blk_row,int blk_col,BLOCK_SIZE plane_bsize,TX_SIZE tx_size,void * arg)659 static void block_rd_txfm(int plane, int block, int blk_row, int blk_col,
660                           BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
661   struct rdcost_block_args *args = arg;
662   MACROBLOCK *const x = args->x;
663   MACROBLOCKD *const xd = &x->e_mbd;
664   MODE_INFO *const mi = xd->mi[0];
665   int64_t rd1, rd2, rd;
666   int rate;
667   int64_t dist;
668   int64_t sse;
669   const int coeff_ctx =
670       combine_entropy_contexts(args->t_left[blk_row], args->t_above[blk_col]);
671 
672   if (args->exit_early) return;
673 
674   if (!is_inter_block(mi)) {
675     struct encode_b_args intra_arg = { x, x->block_qcoeff_opt, args->t_above,
676                                        args->t_left, &mi->skip };
677     vp9_encode_block_intra(plane, block, blk_row, blk_col, plane_bsize, tx_size,
678                            &intra_arg);
679     if (x->block_tx_domain) {
680       dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
681                  tx_size, &dist, &sse);
682     } else {
683       const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
684       const struct macroblock_plane *const p = &x->plane[plane];
685       const struct macroblockd_plane *const pd = &xd->plane[plane];
686       const int src_stride = p->src.stride;
687       const int dst_stride = pd->dst.stride;
688       const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
689       const uint8_t *src = &p->src.buf[4 * (blk_row * src_stride + blk_col)];
690       const uint8_t *dst = &pd->dst.buf[4 * (blk_row * dst_stride + blk_col)];
691       const int16_t *diff = &p->src_diff[4 * (blk_row * diff_stride + blk_col)];
692       unsigned int tmp;
693       sse = sum_squares_visible(xd, pd, diff, diff_stride, blk_row, blk_col,
694                                 plane_bsize, tx_bsize);
695 #if CONFIG_VP9_HIGHBITDEPTH
696       if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && (xd->bd > 8))
697         sse = ROUND64_POWER_OF_TWO(sse, (xd->bd - 8) * 2);
698 #endif  // CONFIG_VP9_HIGHBITDEPTH
699       sse = sse * 16;
700       tmp = pixel_sse(args->cpi, xd, pd, src, src_stride, dst, dst_stride,
701                       blk_row, blk_col, plane_bsize, tx_bsize);
702       dist = (int64_t)tmp * 16;
703     }
704   } else {
705     int skip_txfm_flag = SKIP_TXFM_NONE;
706     if (max_txsize_lookup[plane_bsize] == tx_size)
707       skip_txfm_flag = x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))];
708     if (skip_txfm_flag == SKIP_TXFM_NONE) {
709       // full forward transform and quantization
710       vp9_xform_quant(x, plane, block, blk_row, blk_col, plane_bsize, tx_size);
711       if (x->block_qcoeff_opt)
712         vp9_optimize_b(x, plane, block, tx_size, coeff_ctx);
713       dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
714                  tx_size, &dist, &sse);
715     } else if (skip_txfm_flag == SKIP_TXFM_AC_ONLY) {
716       // compute DC coefficient
717       tran_low_t *const coeff = BLOCK_OFFSET(x->plane[plane].coeff, block);
718       tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
719       vp9_xform_quant_dc(x, plane, block, blk_row, blk_col, plane_bsize,
720                          tx_size);
721       sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
722       dist = sse;
723       if (x->plane[plane].eobs[block]) {
724         const int64_t orig_sse = (int64_t)coeff[0] * coeff[0];
725         const int64_t resd_sse = coeff[0] - dqcoeff[0];
726         int64_t dc_correct = orig_sse - resd_sse * resd_sse;
727 #if CONFIG_VP9_HIGHBITDEPTH
728         dc_correct >>= ((xd->bd - 8) * 2);
729 #endif
730         if (tx_size != TX_32X32) dc_correct >>= 2;
731 
732         dist = VPXMAX(0, sse - dc_correct);
733       }
734     } else {
735       // SKIP_TXFM_AC_DC
736       // skip forward transform. Because this is handled here, the quantization
737       // does not need to do it.
738       x->plane[plane].eobs[block] = 0;
739       sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
740       dist = sse;
741     }
742   }
743 
744   rd = RDCOST(x->rdmult, x->rddiv, 0, dist);
745   if (args->this_rd + rd > args->best_rd) {
746     args->exit_early = 1;
747     return;
748   }
749 
750   rate = rate_block(plane, block, tx_size, coeff_ctx, args);
751   args->t_above[blk_col] = (x->plane[plane].eobs[block] > 0) ? 1 : 0;
752   args->t_left[blk_row] = (x->plane[plane].eobs[block] > 0) ? 1 : 0;
753   rd1 = RDCOST(x->rdmult, x->rddiv, rate, dist);
754   rd2 = RDCOST(x->rdmult, x->rddiv, 0, sse);
755 
756   // TODO(jingning): temporarily enabled only for luma component
757   rd = VPXMIN(rd1, rd2);
758   if (plane == 0) {
759     x->zcoeff_blk[tx_size][block] =
760         !x->plane[plane].eobs[block] ||
761         (x->sharpness == 0 && rd1 > rd2 && !xd->lossless);
762     x->sum_y_eobs[tx_size] += x->plane[plane].eobs[block];
763   }
764 
765   args->this_rate += rate;
766   args->this_dist += dist;
767   args->this_sse += sse;
768   args->this_rd += rd;
769 
770   if (args->this_rd > args->best_rd) {
771     args->exit_early = 1;
772     return;
773   }
774 
775   args->skippable &= !x->plane[plane].eobs[block];
776 }
777 
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)778 static void txfm_rd_in_plane(const VP9_COMP *cpi, MACROBLOCK *x, int *rate,
779                              int64_t *distortion, int *skippable, int64_t *sse,
780                              int64_t ref_best_rd, int plane, BLOCK_SIZE bsize,
781                              TX_SIZE tx_size, int use_fast_coef_costing) {
782   MACROBLOCKD *const xd = &x->e_mbd;
783   const struct macroblockd_plane *const pd = &xd->plane[plane];
784   struct rdcost_block_args args;
785   vp9_zero(args);
786   args.cpi = cpi;
787   args.x = x;
788   args.best_rd = ref_best_rd;
789   args.use_fast_coef_costing = use_fast_coef_costing;
790   args.skippable = 1;
791 
792   if (plane == 0) xd->mi[0]->tx_size = tx_size;
793 
794   vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
795 
796   args.so = get_scan(xd, tx_size, get_plane_type(plane), 0);
797 
798   vp9_foreach_transformed_block_in_plane(xd, bsize, plane, block_rd_txfm,
799                                          &args);
800   if (args.exit_early) {
801     *rate = INT_MAX;
802     *distortion = INT64_MAX;
803     *sse = INT64_MAX;
804     *skippable = 0;
805   } else {
806     *distortion = args.this_dist;
807     *rate = args.this_rate;
808     *sse = args.this_sse;
809     *skippable = args.skippable;
810   }
811 }
812 
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)813 static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
814                                    int64_t *distortion, int *skip, int64_t *sse,
815                                    int64_t ref_best_rd, BLOCK_SIZE bs) {
816   const TX_SIZE max_tx_size = max_txsize_lookup[bs];
817   VP9_COMMON *const cm = &cpi->common;
818   const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
819   MACROBLOCKD *const xd = &x->e_mbd;
820   MODE_INFO *const mi = xd->mi[0];
821 
822   mi->tx_size = VPXMIN(max_tx_size, largest_tx_size);
823 
824   txfm_rd_in_plane(cpi, x, rate, distortion, skip, sse, ref_best_rd, 0, bs,
825                    mi->tx_size, cpi->sf.use_fast_coef_costing);
826 }
827 
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)828 static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
829                                    int64_t *distortion, int *skip,
830                                    int64_t *psse, int64_t ref_best_rd,
831                                    BLOCK_SIZE bs) {
832   const TX_SIZE max_tx_size = max_txsize_lookup[bs];
833   VP9_COMMON *const cm = &cpi->common;
834   MACROBLOCKD *const xd = &x->e_mbd;
835   MODE_INFO *const mi = xd->mi[0];
836   vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
837   int r[TX_SIZES][2], s[TX_SIZES];
838   int64_t d[TX_SIZES], sse[TX_SIZES];
839   int64_t rd[TX_SIZES][2] = { { INT64_MAX, INT64_MAX },
840                               { INT64_MAX, INT64_MAX },
841                               { INT64_MAX, INT64_MAX },
842                               { INT64_MAX, INT64_MAX } };
843   int n;
844   int s0, s1;
845   int64_t best_rd = ref_best_rd;
846   TX_SIZE best_tx = max_tx_size;
847   int start_tx, end_tx;
848   const int tx_size_ctx = get_tx_size_context(xd);
849   assert(skip_prob > 0);
850   s0 = vp9_cost_bit(skip_prob, 0);
851   s1 = vp9_cost_bit(skip_prob, 1);
852 
853   if (cm->tx_mode == TX_MODE_SELECT) {
854     start_tx = max_tx_size;
855     end_tx = VPXMAX(start_tx - cpi->sf.tx_size_search_depth, 0);
856     if (bs > BLOCK_32X32) end_tx = VPXMIN(end_tx + 1, start_tx);
857   } else {
858     TX_SIZE chosen_tx_size =
859         VPXMIN(max_tx_size, tx_mode_to_biggest_tx_size[cm->tx_mode]);
860     start_tx = chosen_tx_size;
861     end_tx = chosen_tx_size;
862   }
863 
864   for (n = start_tx; n >= end_tx; n--) {
865     const int r_tx_size = cpi->tx_size_cost[max_tx_size - 1][tx_size_ctx][n];
866     txfm_rd_in_plane(cpi, x, &r[n][0], &d[n], &s[n], &sse[n], best_rd, 0, bs, n,
867                      cpi->sf.use_fast_coef_costing);
868     r[n][1] = r[n][0];
869     if (r[n][0] < INT_MAX) {
870       r[n][1] += r_tx_size;
871     }
872     if (d[n] == INT64_MAX || r[n][0] == INT_MAX) {
873       rd[n][0] = rd[n][1] = INT64_MAX;
874     } else if (s[n]) {
875       if (is_inter_block(mi)) {
876         rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
877         r[n][1] -= r_tx_size;
878       } else {
879         rd[n][0] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
880         rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1 + r_tx_size, sse[n]);
881       }
882     } else {
883       rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
884       rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
885     }
886 
887     if (is_inter_block(mi) && !xd->lossless && !s[n] && sse[n] != INT64_MAX) {
888       rd[n][0] = VPXMIN(rd[n][0], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
889       rd[n][1] = VPXMIN(rd[n][1], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
890     }
891 
892     // Early termination in transform size search.
893     if (cpi->sf.tx_size_search_breakout &&
894         (rd[n][1] == INT64_MAX ||
895          (n < (int)max_tx_size && rd[n][1] > rd[n + 1][1]) || s[n] == 1))
896       break;
897 
898     if (rd[n][1] < best_rd) {
899       best_tx = n;
900       best_rd = rd[n][1];
901     }
902   }
903   mi->tx_size = best_tx;
904 
905   *distortion = d[mi->tx_size];
906   *rate = r[mi->tx_size][cm->tx_mode == TX_MODE_SELECT];
907   *skip = s[mi->tx_size];
908   *psse = sse[mi->tx_size];
909 }
910 
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)911 static void super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
912                             int64_t *distortion, int *skip, int64_t *psse,
913                             BLOCK_SIZE bs, int64_t ref_best_rd) {
914   MACROBLOCKD *xd = &x->e_mbd;
915   int64_t sse;
916   int64_t *ret_sse = psse ? psse : &sse;
917 
918   assert(bs == xd->mi[0]->sb_type);
919 
920   if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
921     choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
922                            bs);
923   } else {
924     choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
925                            bs);
926   }
927 }
928 
conditional_skipintra(PREDICTION_MODE mode,PREDICTION_MODE best_intra_mode)929 static int conditional_skipintra(PREDICTION_MODE mode,
930                                  PREDICTION_MODE best_intra_mode) {
931   if (mode == D117_PRED && best_intra_mode != V_PRED &&
932       best_intra_mode != D135_PRED)
933     return 1;
934   if (mode == D63_PRED && best_intra_mode != V_PRED &&
935       best_intra_mode != D45_PRED)
936     return 1;
937   if (mode == D207_PRED && best_intra_mode != H_PRED &&
938       best_intra_mode != D45_PRED)
939     return 1;
940   if (mode == D153_PRED && best_intra_mode != H_PRED &&
941       best_intra_mode != D135_PRED)
942     return 1;
943   return 0;
944 }
945 
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)946 static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int row,
947                                      int col, PREDICTION_MODE *best_mode,
948                                      const int *bmode_costs, ENTROPY_CONTEXT *a,
949                                      ENTROPY_CONTEXT *l, int *bestrate,
950                                      int *bestratey, int64_t *bestdistortion,
951                                      BLOCK_SIZE bsize, int64_t rd_thresh) {
952   PREDICTION_MODE mode;
953   MACROBLOCKD *const xd = &x->e_mbd;
954   int64_t best_rd = rd_thresh;
955   struct macroblock_plane *p = &x->plane[0];
956   struct macroblockd_plane *pd = &xd->plane[0];
957   const int src_stride = p->src.stride;
958   const int dst_stride = pd->dst.stride;
959   const uint8_t *src_init = &p->src.buf[row * 4 * src_stride + col * 4];
960   uint8_t *dst_init = &pd->dst.buf[row * 4 * src_stride + col * 4];
961   ENTROPY_CONTEXT ta[2], tempa[2];
962   ENTROPY_CONTEXT tl[2], templ[2];
963   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
964   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
965   int idx, idy;
966   uint8_t best_dst[8 * 8];
967 #if CONFIG_VP9_HIGHBITDEPTH
968   uint16_t best_dst16[8 * 8];
969 #endif
970   memcpy(ta, a, num_4x4_blocks_wide * sizeof(a[0]));
971   memcpy(tl, l, num_4x4_blocks_high * sizeof(l[0]));
972 
973   xd->mi[0]->tx_size = TX_4X4;
974 
975 #if CONFIG_VP9_HIGHBITDEPTH
976   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
977     for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
978       int64_t this_rd;
979       int ratey = 0;
980       int64_t distortion = 0;
981       int rate = bmode_costs[mode];
982 
983       if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode))) continue;
984 
985       // Only do the oblique modes if the best so far is
986       // one of the neighboring directional modes
987       if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
988         if (conditional_skipintra(mode, *best_mode)) continue;
989       }
990 
991       memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0]));
992       memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0]));
993 
994       for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
995         for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
996           const int block = (row + idy) * 2 + (col + idx);
997           const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
998           uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
999           uint16_t *const dst16 = CONVERT_TO_SHORTPTR(dst);
1000           int16_t *const src_diff =
1001               vp9_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
1002           tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
1003           xd->mi[0]->bmi[block].as_mode = mode;
1004           vp9_predict_intra_block(xd, 1, TX_4X4, mode,
1005                                   x->skip_encode ? src : dst,
1006                                   x->skip_encode ? src_stride : dst_stride, dst,
1007                                   dst_stride, col + idx, row + idy, 0);
1008           vpx_highbd_subtract_block(4, 4, src_diff, 8, src, src_stride, dst,
1009                                     dst_stride, xd->bd);
1010           if (xd->lossless) {
1011             const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1012             const int coeff_ctx =
1013                 combine_entropy_contexts(tempa[idx], templ[idy]);
1014             vp9_highbd_fwht4x4(src_diff, coeff, 8);
1015             vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
1016             ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1017                                  so->neighbors, cpi->sf.use_fast_coef_costing);
1018             tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0 ? 1 : 0);
1019             if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1020               goto next_highbd;
1021             vp9_highbd_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst16,
1022                                    dst_stride, p->eobs[block], xd->bd);
1023           } else {
1024             int64_t unused;
1025             const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
1026             const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
1027             const int coeff_ctx =
1028                 combine_entropy_contexts(tempa[idx], templ[idy]);
1029             if (tx_type == DCT_DCT)
1030               vpx_highbd_fdct4x4(src_diff, coeff, 8);
1031             else
1032               vp9_highbd_fht4x4(src_diff, coeff, 8, tx_type);
1033             vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
1034             ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1035                                  so->neighbors, cpi->sf.use_fast_coef_costing);
1036             distortion += vp9_highbd_block_error_dispatch(
1037                               coeff, BLOCK_OFFSET(pd->dqcoeff, block), 16,
1038                               &unused, xd->bd) >>
1039                           2;
1040             tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0 ? 1 : 0);
1041             if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1042               goto next_highbd;
1043             vp9_highbd_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
1044                                   dst16, dst_stride, p->eobs[block], xd->bd);
1045           }
1046         }
1047       }
1048 
1049       rate += ratey;
1050       this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
1051 
1052       if (this_rd < best_rd) {
1053         *bestrate = rate;
1054         *bestratey = ratey;
1055         *bestdistortion = distortion;
1056         best_rd = this_rd;
1057         *best_mode = mode;
1058         memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0]));
1059         memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0]));
1060         for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
1061           memcpy(best_dst16 + idy * 8,
1062                  CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
1063                  num_4x4_blocks_wide * 4 * sizeof(uint16_t));
1064         }
1065       }
1066     next_highbd : {}
1067     }
1068     if (best_rd >= rd_thresh || x->skip_encode) return best_rd;
1069 
1070     for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
1071       memcpy(CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
1072              best_dst16 + idy * 8, num_4x4_blocks_wide * 4 * sizeof(uint16_t));
1073     }
1074 
1075     return best_rd;
1076   }
1077 #endif  // CONFIG_VP9_HIGHBITDEPTH
1078 
1079   for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1080     int64_t this_rd;
1081     int ratey = 0;
1082     int64_t distortion = 0;
1083     int rate = bmode_costs[mode];
1084 
1085     if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode))) continue;
1086 
1087     // Only do the oblique modes if the best so far is
1088     // one of the neighboring directional modes
1089     if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
1090       if (conditional_skipintra(mode, *best_mode)) continue;
1091     }
1092 
1093     memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0]));
1094     memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0]));
1095 
1096     for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
1097       for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
1098         const int block = (row + idy) * 2 + (col + idx);
1099         const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
1100         uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
1101         int16_t *const src_diff =
1102             vp9_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
1103         tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
1104         xd->mi[0]->bmi[block].as_mode = mode;
1105         vp9_predict_intra_block(xd, 1, TX_4X4, mode, x->skip_encode ? src : dst,
1106                                 x->skip_encode ? src_stride : dst_stride, dst,
1107                                 dst_stride, col + idx, row + idy, 0);
1108         vpx_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
1109 
1110         if (xd->lossless) {
1111           const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1112           const int coeff_ctx =
1113               combine_entropy_contexts(tempa[idx], templ[idy]);
1114           vp9_fwht4x4(src_diff, coeff, 8);
1115           vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
1116           ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1117                                so->neighbors, cpi->sf.use_fast_coef_costing);
1118           tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0) ? 1 : 0;
1119           if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1120             goto next;
1121           vp9_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride,
1122                           p->eobs[block]);
1123         } else {
1124           int64_t unused;
1125           const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
1126           const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
1127           const int coeff_ctx =
1128               combine_entropy_contexts(tempa[idx], templ[idy]);
1129           vp9_fht4x4(src_diff, coeff, 8, tx_type);
1130           vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
1131           ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1132                                so->neighbors, cpi->sf.use_fast_coef_costing);
1133           tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0) ? 1 : 0;
1134           distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
1135                                         16, &unused) >>
1136                         2;
1137           if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1138             goto next;
1139           vp9_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block), dst,
1140                          dst_stride, p->eobs[block]);
1141         }
1142       }
1143     }
1144 
1145     rate += ratey;
1146     this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
1147 
1148     if (this_rd < best_rd) {
1149       *bestrate = rate;
1150       *bestratey = ratey;
1151       *bestdistortion = distortion;
1152       best_rd = this_rd;
1153       *best_mode = mode;
1154       memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0]));
1155       memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0]));
1156       for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
1157         memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
1158                num_4x4_blocks_wide * 4);
1159     }
1160   next : {}
1161   }
1162 
1163   if (best_rd >= rd_thresh || x->skip_encode) return best_rd;
1164 
1165   for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
1166     memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
1167            num_4x4_blocks_wide * 4);
1168 
1169   return best_rd;
1170 }
1171 
rd_pick_intra_sub_8x8_y_mode(VP9_COMP * cpi,MACROBLOCK * mb,int * rate,int * rate_y,int64_t * distortion,int64_t best_rd)1172 static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
1173                                             int *rate, int *rate_y,
1174                                             int64_t *distortion,
1175                                             int64_t best_rd) {
1176   int i, j;
1177   const MACROBLOCKD *const xd = &mb->e_mbd;
1178   MODE_INFO *const mic = xd->mi[0];
1179   const MODE_INFO *above_mi = xd->above_mi;
1180   const MODE_INFO *left_mi = xd->left_mi;
1181   const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
1182   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1183   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1184   int idx, idy;
1185   int cost = 0;
1186   int64_t total_distortion = 0;
1187   int tot_rate_y = 0;
1188   int64_t total_rd = 0;
1189   const int *bmode_costs = cpi->mbmode_cost;
1190 
1191   // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
1192   for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1193     for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1194       PREDICTION_MODE best_mode = DC_PRED;
1195       int r = INT_MAX, ry = INT_MAX;
1196       int64_t d = INT64_MAX, this_rd = INT64_MAX;
1197       i = idy * 2 + idx;
1198       if (cpi->common.frame_type == KEY_FRAME) {
1199         const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i);
1200         const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i);
1201 
1202         bmode_costs = cpi->y_mode_costs[A][L];
1203       }
1204 
1205       this_rd = rd_pick_intra4x4block(
1206           cpi, mb, idy, idx, &best_mode, bmode_costs,
1207           xd->plane[0].above_context + idx, xd->plane[0].left_context + idy, &r,
1208           &ry, &d, bsize, best_rd - total_rd);
1209 
1210       if (this_rd >= best_rd - total_rd) return INT64_MAX;
1211 
1212       total_rd += this_rd;
1213       cost += r;
1214       total_distortion += d;
1215       tot_rate_y += ry;
1216 
1217       mic->bmi[i].as_mode = best_mode;
1218       for (j = 1; j < num_4x4_blocks_high; ++j)
1219         mic->bmi[i + j * 2].as_mode = best_mode;
1220       for (j = 1; j < num_4x4_blocks_wide; ++j)
1221         mic->bmi[i + j].as_mode = best_mode;
1222 
1223       if (total_rd >= best_rd) return INT64_MAX;
1224     }
1225   }
1226 
1227   *rate = cost;
1228   *rate_y = tot_rate_y;
1229   *distortion = total_distortion;
1230   mic->mode = mic->bmi[3].as_mode;
1231 
1232   return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
1233 }
1234 
1235 // 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)1236 static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
1237                                       int *rate_tokenonly, int64_t *distortion,
1238                                       int *skippable, BLOCK_SIZE bsize,
1239                                       int64_t best_rd) {
1240   PREDICTION_MODE mode;
1241   PREDICTION_MODE mode_selected = DC_PRED;
1242   MACROBLOCKD *const xd = &x->e_mbd;
1243   MODE_INFO *const mic = xd->mi[0];
1244   int this_rate, this_rate_tokenonly, s;
1245   int64_t this_distortion, this_rd;
1246   TX_SIZE best_tx = TX_4X4;
1247   int *bmode_costs;
1248   const MODE_INFO *above_mi = xd->above_mi;
1249   const MODE_INFO *left_mi = xd->left_mi;
1250   const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1251   const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1252   bmode_costs = cpi->y_mode_costs[A][L];
1253 
1254   memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1255   /* Y Search for intra prediction mode */
1256   for (mode = DC_PRED; mode <= TM_PRED; mode++) {
1257     if (cpi->sf.use_nonrd_pick_mode) {
1258       // These speed features are turned on in hybrid non-RD and RD mode
1259       // for key frame coding in the context of real-time setting.
1260       if (conditional_skipintra(mode, mode_selected)) continue;
1261       if (*skippable) break;
1262     }
1263 
1264     mic->mode = mode;
1265 
1266     super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s, NULL,
1267                     bsize, best_rd);
1268 
1269     if (this_rate_tokenonly == INT_MAX) continue;
1270 
1271     this_rate = this_rate_tokenonly + bmode_costs[mode];
1272     this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1273 
1274     if (this_rd < best_rd) {
1275       mode_selected = mode;
1276       best_rd = this_rd;
1277       best_tx = mic->tx_size;
1278       *rate = this_rate;
1279       *rate_tokenonly = this_rate_tokenonly;
1280       *distortion = this_distortion;
1281       *skippable = s;
1282     }
1283   }
1284 
1285   mic->mode = mode_selected;
1286   mic->tx_size = best_tx;
1287 
1288   return best_rd;
1289 }
1290 
1291 // Return value 0: early termination triggered, no valid rd cost available;
1292 //              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)1293 static int super_block_uvrd(const VP9_COMP *cpi, MACROBLOCK *x, int *rate,
1294                             int64_t *distortion, int *skippable, int64_t *sse,
1295                             BLOCK_SIZE bsize, int64_t ref_best_rd) {
1296   MACROBLOCKD *const xd = &x->e_mbd;
1297   MODE_INFO *const mi = xd->mi[0];
1298   const TX_SIZE uv_tx_size = get_uv_tx_size(mi, &xd->plane[1]);
1299   int plane;
1300   int pnrate = 0, pnskip = 1;
1301   int64_t pndist = 0, pnsse = 0;
1302   int is_cost_valid = 1;
1303 
1304   if (ref_best_rd < 0) is_cost_valid = 0;
1305 
1306   if (is_inter_block(mi) && is_cost_valid) {
1307     int plane;
1308     for (plane = 1; plane < MAX_MB_PLANE; ++plane)
1309       vp9_subtract_plane(x, bsize, plane);
1310   }
1311 
1312   *rate = 0;
1313   *distortion = 0;
1314   *sse = 0;
1315   *skippable = 1;
1316 
1317   for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
1318     txfm_rd_in_plane(cpi, x, &pnrate, &pndist, &pnskip, &pnsse, ref_best_rd,
1319                      plane, bsize, uv_tx_size, cpi->sf.use_fast_coef_costing);
1320     if (pnrate == INT_MAX) {
1321       is_cost_valid = 0;
1322       break;
1323     }
1324     *rate += pnrate;
1325     *distortion += pndist;
1326     *sse += pnsse;
1327     *skippable &= pnskip;
1328   }
1329 
1330   if (!is_cost_valid) {
1331     // reset cost value
1332     *rate = INT_MAX;
1333     *distortion = INT64_MAX;
1334     *sse = INT64_MAX;
1335     *skippable = 0;
1336   }
1337 
1338   return is_cost_valid;
1339 }
1340 
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)1341 static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
1342                                        PICK_MODE_CONTEXT *ctx, int *rate,
1343                                        int *rate_tokenonly, int64_t *distortion,
1344                                        int *skippable, BLOCK_SIZE bsize,
1345                                        TX_SIZE max_tx_size) {
1346   MACROBLOCKD *xd = &x->e_mbd;
1347   PREDICTION_MODE mode;
1348   PREDICTION_MODE mode_selected = DC_PRED;
1349   int64_t best_rd = INT64_MAX, this_rd;
1350   int this_rate_tokenonly, this_rate, s;
1351   int64_t this_distortion, this_sse;
1352 
1353   memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1354   for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1355     if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode))) continue;
1356 #if CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
1357     if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) &&
1358         (xd->above_mi == NULL || xd->left_mi == NULL) && need_top_left[mode])
1359       continue;
1360 #endif  // CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
1361 
1362     xd->mi[0]->uv_mode = mode;
1363 
1364     if (!super_block_uvrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s,
1365                           &this_sse, bsize, best_rd))
1366       continue;
1367     this_rate =
1368         this_rate_tokenonly +
1369         cpi->intra_uv_mode_cost[cpi->common.frame_type][xd->mi[0]->mode][mode];
1370     this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1371 
1372     if (this_rd < best_rd) {
1373       mode_selected = mode;
1374       best_rd = this_rd;
1375       *rate = this_rate;
1376       *rate_tokenonly = this_rate_tokenonly;
1377       *distortion = this_distortion;
1378       *skippable = s;
1379       if (!x->select_tx_size) swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
1380     }
1381   }
1382 
1383   xd->mi[0]->uv_mode = mode_selected;
1384   return best_rd;
1385 }
1386 
rd_sbuv_dcpred(const VP9_COMP * cpi,MACROBLOCK * x,int * rate,int * rate_tokenonly,int64_t * distortion,int * skippable,BLOCK_SIZE bsize)1387 static int64_t rd_sbuv_dcpred(const VP9_COMP *cpi, MACROBLOCK *x, int *rate,
1388                               int *rate_tokenonly, int64_t *distortion,
1389                               int *skippable, BLOCK_SIZE bsize) {
1390   const VP9_COMMON *cm = &cpi->common;
1391   int64_t unused;
1392 
1393   x->e_mbd.mi[0]->uv_mode = DC_PRED;
1394   memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1395   super_block_uvrd(cpi, x, rate_tokenonly, distortion, skippable, &unused,
1396                    bsize, INT64_MAX);
1397   *rate =
1398       *rate_tokenonly +
1399       cpi->intra_uv_mode_cost[cm->frame_type][x->e_mbd.mi[0]->mode][DC_PRED];
1400   return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
1401 }
1402 
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)1403 static void choose_intra_uv_mode(VP9_COMP *cpi, MACROBLOCK *const x,
1404                                  PICK_MODE_CONTEXT *ctx, BLOCK_SIZE bsize,
1405                                  TX_SIZE max_tx_size, int *rate_uv,
1406                                  int *rate_uv_tokenonly, int64_t *dist_uv,
1407                                  int *skip_uv, PREDICTION_MODE *mode_uv) {
1408   // Use an estimated rd for uv_intra based on DC_PRED if the
1409   // appropriate speed flag is set.
1410   if (cpi->sf.use_uv_intra_rd_estimate) {
1411     rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
1412                    bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
1413     // Else do a proper rd search for each possible transform size that may
1414     // be considered in the main rd loop.
1415   } else {
1416     rd_pick_intra_sbuv_mode(cpi, x, ctx, rate_uv, rate_uv_tokenonly, dist_uv,
1417                             skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
1418                             max_tx_size);
1419   }
1420   *mode_uv = x->e_mbd.mi[0]->uv_mode;
1421 }
1422 
cost_mv_ref(const VP9_COMP * cpi,PREDICTION_MODE mode,int mode_context)1423 static int cost_mv_ref(const VP9_COMP *cpi, PREDICTION_MODE mode,
1424                        int mode_context) {
1425   assert(is_inter_mode(mode));
1426   return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
1427 }
1428 
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])1429 static int set_and_cost_bmi_mvs(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
1430                                 int i, PREDICTION_MODE mode, int_mv this_mv[2],
1431                                 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1432                                 int_mv seg_mvs[MAX_REF_FRAMES],
1433                                 int_mv *best_ref_mv[2], const int *mvjcost,
1434                                 int *mvcost[2]) {
1435   MODE_INFO *const mi = xd->mi[0];
1436   const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1437   int thismvcost = 0;
1438   int idx, idy;
1439   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mi->sb_type];
1440   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mi->sb_type];
1441   const int is_compound = has_second_ref(mi);
1442 
1443   switch (mode) {
1444     case NEWMV:
1445       this_mv[0].as_int = seg_mvs[mi->ref_frame[0]].as_int;
1446       thismvcost += vp9_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
1447                                     mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1448       if (is_compound) {
1449         this_mv[1].as_int = seg_mvs[mi->ref_frame[1]].as_int;
1450         thismvcost += vp9_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
1451                                       mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1452       }
1453       break;
1454     case NEARMV:
1455     case NEARESTMV:
1456       this_mv[0].as_int = frame_mv[mode][mi->ref_frame[0]].as_int;
1457       if (is_compound)
1458         this_mv[1].as_int = frame_mv[mode][mi->ref_frame[1]].as_int;
1459       break;
1460     default:
1461       assert(mode == ZEROMV);
1462       this_mv[0].as_int = 0;
1463       if (is_compound) this_mv[1].as_int = 0;
1464       break;
1465   }
1466 
1467   mi->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
1468   if (is_compound) mi->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
1469 
1470   mi->bmi[i].as_mode = mode;
1471 
1472   for (idy = 0; idy < num_4x4_blocks_high; ++idy)
1473     for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
1474       memmove(&mi->bmi[i + idy * 2 + idx], &mi->bmi[i], sizeof(mi->bmi[i]));
1475 
1476   return cost_mv_ref(cpi, mode, mbmi_ext->mode_context[mi->ref_frame[0]]) +
1477          thismvcost;
1478 }
1479 
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)1480 static int64_t encode_inter_mb_segment(VP9_COMP *cpi, MACROBLOCK *x,
1481                                        int64_t best_yrd, int i, int *labelyrate,
1482                                        int64_t *distortion, int64_t *sse,
1483                                        ENTROPY_CONTEXT *ta, ENTROPY_CONTEXT *tl,
1484                                        int mi_row, int mi_col) {
1485   int k;
1486   MACROBLOCKD *xd = &x->e_mbd;
1487   struct macroblockd_plane *const pd = &xd->plane[0];
1488   struct macroblock_plane *const p = &x->plane[0];
1489   MODE_INFO *const mi = xd->mi[0];
1490   const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->sb_type, pd);
1491   const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
1492   const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
1493   int idx, idy;
1494 
1495   const uint8_t *const src =
1496       &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1497   uint8_t *const dst =
1498       &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
1499   int64_t thisdistortion = 0, thissse = 0;
1500   int thisrate = 0, ref;
1501   const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1502   const int is_compound = has_second_ref(mi);
1503   const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
1504 
1505   for (ref = 0; ref < 1 + is_compound; ++ref) {
1506     const int bw = b_width_log2_lookup[BLOCK_8X8];
1507     const int h = 4 * (i >> bw);
1508     const int w = 4 * (i & ((1 << bw) - 1));
1509     const struct scale_factors *sf = &xd->block_refs[ref]->sf;
1510     int y_stride = pd->pre[ref].stride;
1511     uint8_t *pre = pd->pre[ref].buf + (h * pd->pre[ref].stride + w);
1512 
1513     if (vp9_is_scaled(sf)) {
1514       const int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
1515       const int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
1516 
1517       y_stride = xd->block_refs[ref]->buf->y_stride;
1518       pre = xd->block_refs[ref]->buf->y_buffer;
1519       pre += scaled_buffer_offset(x_start + w, y_start + h, y_stride, sf);
1520     }
1521 #if CONFIG_VP9_HIGHBITDEPTH
1522     if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1523       vp9_highbd_build_inter_predictor(
1524           CONVERT_TO_SHORTPTR(pre), y_stride, CONVERT_TO_SHORTPTR(dst),
1525           pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
1526           &xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
1527           mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2),
1528           xd->bd);
1529     } else {
1530       vp9_build_inter_predictor(
1531           pre, y_stride, dst, pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
1532           &xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
1533           mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2));
1534     }
1535 #else
1536     vp9_build_inter_predictor(
1537         pre, y_stride, dst, pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
1538         &xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
1539         mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2));
1540 #endif  // CONFIG_VP9_HIGHBITDEPTH
1541   }
1542 
1543 #if CONFIG_VP9_HIGHBITDEPTH
1544   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1545     vpx_highbd_subtract_block(
1546         height, width, vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1547         8, src, p->src.stride, dst, pd->dst.stride, xd->bd);
1548   } else {
1549     vpx_subtract_block(height, width,
1550                        vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1551                        8, src, p->src.stride, dst, pd->dst.stride);
1552   }
1553 #else
1554   vpx_subtract_block(height, width,
1555                      vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1556                      8, src, p->src.stride, dst, pd->dst.stride);
1557 #endif  // CONFIG_VP9_HIGHBITDEPTH
1558 
1559   k = i;
1560   for (idy = 0; idy < height / 4; ++idy) {
1561     for (idx = 0; idx < width / 4; ++idx) {
1562 #if CONFIG_VP9_HIGHBITDEPTH
1563       const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
1564 #endif
1565       int64_t ssz, rd, rd1, rd2;
1566       tran_low_t *coeff;
1567       int coeff_ctx;
1568       k += (idy * 2 + idx);
1569       coeff_ctx = combine_entropy_contexts(ta[k & 1], tl[k >> 1]);
1570       coeff = BLOCK_OFFSET(p->coeff, k);
1571       x->fwd_txfm4x4(vp9_raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
1572                      coeff, 8);
1573       vp9_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan);
1574 #if CONFIG_VP9_HIGHBITDEPTH
1575       thisdistortion += vp9_highbd_block_error_dispatch(
1576           coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz, bd);
1577 #else
1578       thisdistortion +=
1579           vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz);
1580 #endif  // CONFIG_VP9_HIGHBITDEPTH
1581       thissse += ssz;
1582       thisrate += cost_coeffs(x, 0, k, TX_4X4, coeff_ctx, so->scan,
1583                               so->neighbors, cpi->sf.use_fast_coef_costing);
1584       ta[k & 1] = tl[k >> 1] = (x->plane[0].eobs[k] > 0) ? 1 : 0;
1585       rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
1586       rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
1587       rd = VPXMIN(rd1, rd2);
1588       if (rd >= best_yrd) return INT64_MAX;
1589     }
1590   }
1591 
1592   *distortion = thisdistortion >> 2;
1593   *labelyrate = thisrate;
1594   *sse = thissse >> 2;
1595 
1596   return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
1597 }
1598 
1599 typedef struct {
1600   int eobs;
1601   int brate;
1602   int byrate;
1603   int64_t bdist;
1604   int64_t bsse;
1605   int64_t brdcost;
1606   int_mv mvs[2];
1607   ENTROPY_CONTEXT ta[2];
1608   ENTROPY_CONTEXT tl[2];
1609 } SEG_RDSTAT;
1610 
1611 typedef struct {
1612   int_mv *ref_mv[2];
1613   int_mv mvp;
1614 
1615   int64_t segment_rd;
1616   int r;
1617   int64_t d;
1618   int64_t sse;
1619   int segment_yrate;
1620   PREDICTION_MODE modes[4];
1621   SEG_RDSTAT rdstat[4][INTER_MODES];
1622   int mvthresh;
1623 } BEST_SEG_INFO;
1624 
mv_check_bounds(const MvLimits * mv_limits,const MV * mv)1625 static INLINE int mv_check_bounds(const MvLimits *mv_limits, const MV *mv) {
1626   return (mv->row >> 3) < mv_limits->row_min ||
1627          (mv->row >> 3) > mv_limits->row_max ||
1628          (mv->col >> 3) < mv_limits->col_min ||
1629          (mv->col >> 3) > mv_limits->col_max;
1630 }
1631 
mi_buf_shift(MACROBLOCK * x,int i)1632 static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
1633   MODE_INFO *const mi = x->e_mbd.mi[0];
1634   struct macroblock_plane *const p = &x->plane[0];
1635   struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
1636 
1637   p->src.buf =
1638       &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1639   assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
1640   pd->pre[0].buf =
1641       &pd->pre[0].buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
1642   if (has_second_ref(mi))
1643     pd->pre[1].buf =
1644         &pd->pre[1]
1645              .buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[1].stride)];
1646 }
1647 
mi_buf_restore(MACROBLOCK * x,struct buf_2d orig_src,struct buf_2d orig_pre[2])1648 static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
1649                                   struct buf_2d orig_pre[2]) {
1650   MODE_INFO *mi = x->e_mbd.mi[0];
1651   x->plane[0].src = orig_src;
1652   x->e_mbd.plane[0].pre[0] = orig_pre[0];
1653   if (has_second_ref(mi)) x->e_mbd.plane[0].pre[1] = orig_pre[1];
1654 }
1655 
mv_has_subpel(const MV * mv)1656 static INLINE int mv_has_subpel(const MV *mv) {
1657   return (mv->row & 0x0F) || (mv->col & 0x0F);
1658 }
1659 
1660 // Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
1661 // 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])1662 static int check_best_zero_mv(const VP9_COMP *cpi,
1663                               const uint8_t mode_context[MAX_REF_FRAMES],
1664                               int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1665                               int this_mode,
1666                               const MV_REFERENCE_FRAME ref_frames[2]) {
1667   if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
1668       frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
1669       (ref_frames[1] == NONE ||
1670        frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
1671     int rfc = mode_context[ref_frames[0]];
1672     int c1 = cost_mv_ref(cpi, NEARMV, rfc);
1673     int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
1674     int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
1675 
1676     if (this_mode == NEARMV) {
1677       if (c1 > c3) return 0;
1678     } else if (this_mode == NEARESTMV) {
1679       if (c2 > c3) return 0;
1680     } else {
1681       assert(this_mode == ZEROMV);
1682       if (ref_frames[1] == NONE) {
1683         if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
1684             (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
1685           return 0;
1686       } else {
1687         if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
1688              frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
1689             (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
1690              frame_mv[NEARMV][ref_frames[1]].as_int == 0))
1691           return 0;
1692       }
1693     }
1694   }
1695   return 1;
1696 }
1697 
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)1698 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
1699                                 int_mv *frame_mv, int mi_row, int mi_col,
1700                                 int_mv single_newmv[MAX_REF_FRAMES],
1701                                 int *rate_mv) {
1702   const VP9_COMMON *const cm = &cpi->common;
1703   const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
1704   const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
1705   MACROBLOCKD *xd = &x->e_mbd;
1706   MODE_INFO *mi = xd->mi[0];
1707   const int refs[2] = { mi->ref_frame[0],
1708                         mi->ref_frame[1] < 0 ? 0 : mi->ref_frame[1] };
1709   int_mv ref_mv[2];
1710   int ite, ref;
1711   const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
1712   struct scale_factors sf;
1713 
1714   // Do joint motion search in compound mode to get more accurate mv.
1715   struct buf_2d backup_yv12[2][MAX_MB_PLANE];
1716   uint32_t last_besterr[2] = { UINT_MAX, UINT_MAX };
1717   const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
1718     vp9_get_scaled_ref_frame(cpi, mi->ref_frame[0]),
1719     vp9_get_scaled_ref_frame(cpi, mi->ref_frame[1])
1720   };
1721 
1722 // Prediction buffer from second frame.
1723 #if CONFIG_VP9_HIGHBITDEPTH
1724   DECLARE_ALIGNED(16, uint16_t, second_pred_alloc_16[64 * 64]);
1725   uint8_t *second_pred;
1726 #else
1727   DECLARE_ALIGNED(16, uint8_t, second_pred[64 * 64]);
1728 #endif  // CONFIG_VP9_HIGHBITDEPTH
1729 
1730   for (ref = 0; ref < 2; ++ref) {
1731     ref_mv[ref] = x->mbmi_ext->ref_mvs[refs[ref]][0];
1732 
1733     if (scaled_ref_frame[ref]) {
1734       int i;
1735       // Swap out the reference frame for a version that's been scaled to
1736       // match the resolution of the current frame, allowing the existing
1737       // motion search code to be used without additional modifications.
1738       for (i = 0; i < MAX_MB_PLANE; i++)
1739         backup_yv12[ref][i] = xd->plane[i].pre[ref];
1740       vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
1741                            NULL);
1742     }
1743 
1744     frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
1745   }
1746 
1747 // Since we have scaled the reference frames to match the size of the current
1748 // frame we must use a unit scaling factor during mode selection.
1749 #if CONFIG_VP9_HIGHBITDEPTH
1750   vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
1751                                     cm->height, cm->use_highbitdepth);
1752 #else
1753   vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
1754                                     cm->height);
1755 #endif  // CONFIG_VP9_HIGHBITDEPTH
1756 
1757   // Allow joint search multiple times iteratively for each reference frame
1758   // and break out of the search loop if it couldn't find a better mv.
1759   for (ite = 0; ite < 4; ite++) {
1760     struct buf_2d ref_yv12[2];
1761     uint32_t bestsme = UINT_MAX;
1762     int sadpb = x->sadperbit16;
1763     MV tmp_mv;
1764     int search_range = 3;
1765 
1766     const MvLimits tmp_mv_limits = x->mv_limits;
1767     int id = ite % 2;  // Even iterations search in the first reference frame,
1768                        // odd iterations search in the second. The predictor
1769                        // found for the 'other' reference frame is factored in.
1770 
1771     // Initialized here because of compiler problem in Visual Studio.
1772     ref_yv12[0] = xd->plane[0].pre[0];
1773     ref_yv12[1] = xd->plane[0].pre[1];
1774 
1775 // Get the prediction block from the 'other' reference frame.
1776 #if CONFIG_VP9_HIGHBITDEPTH
1777     if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1778       second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc_16);
1779       vp9_highbd_build_inter_predictor(
1780           CONVERT_TO_SHORTPTR(ref_yv12[!id].buf), ref_yv12[!id].stride,
1781           second_pred_alloc_16, pw, &frame_mv[refs[!id]].as_mv, &sf, pw, ph, 0,
1782           kernel, MV_PRECISION_Q3, mi_col * MI_SIZE, mi_row * MI_SIZE, xd->bd);
1783     } else {
1784       second_pred = (uint8_t *)second_pred_alloc_16;
1785       vp9_build_inter_predictor(ref_yv12[!id].buf, ref_yv12[!id].stride,
1786                                 second_pred, pw, &frame_mv[refs[!id]].as_mv,
1787                                 &sf, pw, ph, 0, kernel, MV_PRECISION_Q3,
1788                                 mi_col * MI_SIZE, mi_row * MI_SIZE);
1789     }
1790 #else
1791     vp9_build_inter_predictor(ref_yv12[!id].buf, ref_yv12[!id].stride,
1792                               second_pred, pw, &frame_mv[refs[!id]].as_mv, &sf,
1793                               pw, ph, 0, kernel, MV_PRECISION_Q3,
1794                               mi_col * MI_SIZE, mi_row * MI_SIZE);
1795 #endif  // CONFIG_VP9_HIGHBITDEPTH
1796 
1797     // Do compound motion search on the current reference frame.
1798     if (id) xd->plane[0].pre[0] = ref_yv12[id];
1799     vp9_set_mv_search_range(&x->mv_limits, &ref_mv[id].as_mv);
1800 
1801     // Use the mv result from the single mode as mv predictor.
1802     tmp_mv = frame_mv[refs[id]].as_mv;
1803 
1804     tmp_mv.col >>= 3;
1805     tmp_mv.row >>= 3;
1806 
1807     // Small-range full-pixel motion search.
1808     bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb, search_range,
1809                                        &cpi->fn_ptr[bsize], &ref_mv[id].as_mv,
1810                                        second_pred);
1811     if (bestsme < UINT_MAX)
1812       bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
1813                                       second_pred, &cpi->fn_ptr[bsize], 1);
1814 
1815     x->mv_limits = tmp_mv_limits;
1816 
1817     if (bestsme < UINT_MAX) {
1818       uint32_t dis; /* TODO: use dis in distortion calculation later. */
1819       uint32_t sse;
1820       bestsme = cpi->find_fractional_mv_step(
1821           x, &tmp_mv, &ref_mv[id].as_mv, cpi->common.allow_high_precision_mv,
1822           x->errorperbit, &cpi->fn_ptr[bsize], 0,
1823           cpi->sf.mv.subpel_search_level, NULL, x->nmvjointcost, x->mvcost,
1824           &dis, &sse, second_pred, pw, ph, cpi->sf.use_accurate_subpel_search);
1825     }
1826 
1827     // Restore the pointer to the first (possibly scaled) prediction buffer.
1828     if (id) xd->plane[0].pre[0] = ref_yv12[0];
1829 
1830     if (bestsme < last_besterr[id]) {
1831       frame_mv[refs[id]].as_mv = tmp_mv;
1832       last_besterr[id] = bestsme;
1833     } else {
1834       break;
1835     }
1836   }
1837 
1838   *rate_mv = 0;
1839 
1840   for (ref = 0; ref < 2; ++ref) {
1841     if (scaled_ref_frame[ref]) {
1842       // Restore the prediction frame pointers to their unscaled versions.
1843       int i;
1844       for (i = 0; i < MAX_MB_PLANE; i++)
1845         xd->plane[i].pre[ref] = backup_yv12[ref][i];
1846     }
1847 
1848     *rate_mv += vp9_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
1849                                 &x->mbmi_ext->ref_mvs[refs[ref]][0].as_mv,
1850                                 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1851   }
1852 }
1853 
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)1854 static int64_t rd_pick_best_sub8x8_mode(
1855     VP9_COMP *cpi, MACROBLOCK *x, int_mv *best_ref_mv,
1856     int_mv *second_best_ref_mv, int64_t best_rd, int *returntotrate,
1857     int *returnyrate, int64_t *returndistortion, int *skippable, int64_t *psse,
1858     int mvthresh, int_mv seg_mvs[4][MAX_REF_FRAMES], BEST_SEG_INFO *bsi_buf,
1859     int filter_idx, int mi_row, int mi_col) {
1860   int i;
1861   BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
1862   MACROBLOCKD *xd = &x->e_mbd;
1863   MODE_INFO *mi = xd->mi[0];
1864   int mode_idx;
1865   int k, br = 0, idx, idy;
1866   int64_t bd = 0, block_sse = 0;
1867   PREDICTION_MODE this_mode;
1868   VP9_COMMON *cm = &cpi->common;
1869   struct macroblock_plane *const p = &x->plane[0];
1870   struct macroblockd_plane *const pd = &xd->plane[0];
1871   const int label_count = 4;
1872   int64_t this_segment_rd = 0;
1873   int label_mv_thresh;
1874   int segmentyrate = 0;
1875   const BLOCK_SIZE bsize = mi->sb_type;
1876   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1877   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1878   const int pw = num_4x4_blocks_wide << 2;
1879   const int ph = num_4x4_blocks_high << 2;
1880   ENTROPY_CONTEXT t_above[2], t_left[2];
1881   int subpelmv = 1, have_ref = 0;
1882   SPEED_FEATURES *const sf = &cpi->sf;
1883   const int has_second_rf = has_second_ref(mi);
1884   const int inter_mode_mask = sf->inter_mode_mask[bsize];
1885   MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1886 
1887   vp9_zero(*bsi);
1888 
1889   bsi->segment_rd = best_rd;
1890   bsi->ref_mv[0] = best_ref_mv;
1891   bsi->ref_mv[1] = second_best_ref_mv;
1892   bsi->mvp.as_int = best_ref_mv->as_int;
1893   bsi->mvthresh = mvthresh;
1894 
1895   for (i = 0; i < 4; i++) bsi->modes[i] = ZEROMV;
1896 
1897   memcpy(t_above, pd->above_context, sizeof(t_above));
1898   memcpy(t_left, pd->left_context, sizeof(t_left));
1899 
1900   // 64 makes this threshold really big effectively
1901   // making it so that we very rarely check mvs on
1902   // segments.   setting this to 1 would make mv thresh
1903   // roughly equal to what it is for macroblocks
1904   label_mv_thresh = 1 * bsi->mvthresh / label_count;
1905 
1906   // Segmentation method overheads
1907   for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1908     for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1909       // TODO(jingning,rbultje): rewrite the rate-distortion optimization
1910       // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
1911       int_mv mode_mv[MB_MODE_COUNT][2];
1912       int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1913       PREDICTION_MODE mode_selected = ZEROMV;
1914       int64_t best_rd = INT64_MAX;
1915       const int i = idy * 2 + idx;
1916       int ref;
1917 
1918       for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1919         const MV_REFERENCE_FRAME frame = mi->ref_frame[ref];
1920         frame_mv[ZEROMV][frame].as_int = 0;
1921         vp9_append_sub8x8_mvs_for_idx(
1922             cm, xd, i, ref, mi_row, mi_col, &frame_mv[NEARESTMV][frame],
1923             &frame_mv[NEARMV][frame], mbmi_ext->mode_context);
1924       }
1925 
1926       // search for the best motion vector on this segment
1927       for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1928         const struct buf_2d orig_src = x->plane[0].src;
1929         struct buf_2d orig_pre[2];
1930 
1931         mode_idx = INTER_OFFSET(this_mode);
1932         bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
1933         if (!(inter_mode_mask & (1 << this_mode))) continue;
1934 
1935         if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
1936                                 this_mode, mi->ref_frame))
1937           continue;
1938 
1939         memcpy(orig_pre, pd->pre, sizeof(orig_pre));
1940         memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
1941                sizeof(bsi->rdstat[i][mode_idx].ta));
1942         memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
1943                sizeof(bsi->rdstat[i][mode_idx].tl));
1944 
1945         // motion search for newmv (single predictor case only)
1946         if (!has_second_rf && this_mode == NEWMV &&
1947             seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV) {
1948           MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
1949           int step_param = 0;
1950           uint32_t bestsme = UINT_MAX;
1951           int sadpb = x->sadperbit4;
1952           MV mvp_full;
1953           int max_mv;
1954           int cost_list[5];
1955           const MvLimits tmp_mv_limits = x->mv_limits;
1956 
1957           /* Is the best so far sufficiently good that we cant justify doing
1958            * and new motion search. */
1959           if (best_rd < label_mv_thresh) break;
1960 
1961           if (cpi->oxcf.mode != BEST) {
1962             // use previous block's result as next block's MV predictor.
1963             if (i > 0) {
1964               bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
1965               if (i == 2) bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
1966             }
1967           }
1968           if (i == 0)
1969             max_mv = x->max_mv_context[mi->ref_frame[0]];
1970           else
1971             max_mv =
1972                 VPXMAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
1973 
1974           if (sf->mv.auto_mv_step_size && cm->show_frame) {
1975             // Take wtd average of the step_params based on the last frame's
1976             // max mv magnitude and the best ref mvs of the current block for
1977             // the given reference.
1978             step_param =
1979                 (vp9_init_search_range(max_mv) + cpi->mv_step_param) / 2;
1980           } else {
1981             step_param = cpi->mv_step_param;
1982           }
1983 
1984           mvp_full.row = bsi->mvp.as_mv.row >> 3;
1985           mvp_full.col = bsi->mvp.as_mv.col >> 3;
1986 
1987           if (sf->adaptive_motion_search) {
1988             if (x->pred_mv[mi->ref_frame[0]].row != INT16_MAX &&
1989                 x->pred_mv[mi->ref_frame[0]].col != INT16_MAX) {
1990               mvp_full.row = x->pred_mv[mi->ref_frame[0]].row >> 3;
1991               mvp_full.col = x->pred_mv[mi->ref_frame[0]].col >> 3;
1992             }
1993             step_param = VPXMAX(step_param, 8);
1994           }
1995 
1996           // adjust src pointer for this block
1997           mi_buf_shift(x, i);
1998 
1999           vp9_set_mv_search_range(&x->mv_limits, &bsi->ref_mv[0]->as_mv);
2000 
2001           bestsme = vp9_full_pixel_search(
2002               cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method,
2003               sadpb,
2004               sf->mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
2005               &bsi->ref_mv[0]->as_mv, new_mv, INT_MAX, 1);
2006 
2007           x->mv_limits = tmp_mv_limits;
2008 
2009           if (bestsme < UINT_MAX) {
2010             uint32_t distortion;
2011             cpi->find_fractional_mv_step(
2012                 x, new_mv, &bsi->ref_mv[0]->as_mv, cm->allow_high_precision_mv,
2013                 x->errorperbit, &cpi->fn_ptr[bsize], sf->mv.subpel_force_stop,
2014                 sf->mv.subpel_search_level, cond_cost_list(cpi, cost_list),
2015                 x->nmvjointcost, x->mvcost, &distortion,
2016                 &x->pred_sse[mi->ref_frame[0]], NULL, pw, ph,
2017                 cpi->sf.use_accurate_subpel_search);
2018 
2019             // save motion search result for use in compound prediction
2020             seg_mvs[i][mi->ref_frame[0]].as_mv = *new_mv;
2021           }
2022 
2023           x->pred_mv[mi->ref_frame[0]] = *new_mv;
2024 
2025           // restore src pointers
2026           mi_buf_restore(x, orig_src, orig_pre);
2027         }
2028 
2029         if (has_second_rf) {
2030           if (seg_mvs[i][mi->ref_frame[1]].as_int == INVALID_MV ||
2031               seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV)
2032             continue;
2033         }
2034 
2035         if (has_second_rf && this_mode == NEWMV &&
2036             mi->interp_filter == EIGHTTAP) {
2037           // adjust src pointers
2038           mi_buf_shift(x, i);
2039           if (sf->comp_inter_joint_search_thresh <= bsize) {
2040             int rate_mv;
2041             joint_motion_search(cpi, x, bsize, frame_mv[this_mode], mi_row,
2042                                 mi_col, seg_mvs[i], &rate_mv);
2043             seg_mvs[i][mi->ref_frame[0]].as_int =
2044                 frame_mv[this_mode][mi->ref_frame[0]].as_int;
2045             seg_mvs[i][mi->ref_frame[1]].as_int =
2046                 frame_mv[this_mode][mi->ref_frame[1]].as_int;
2047           }
2048           // restore src pointers
2049           mi_buf_restore(x, orig_src, orig_pre);
2050         }
2051 
2052         bsi->rdstat[i][mode_idx].brate = set_and_cost_bmi_mvs(
2053             cpi, x, xd, i, this_mode, mode_mv[this_mode], frame_mv, seg_mvs[i],
2054             bsi->ref_mv, x->nmvjointcost, x->mvcost);
2055 
2056         for (ref = 0; ref < 1 + has_second_rf; ++ref) {
2057           bsi->rdstat[i][mode_idx].mvs[ref].as_int =
2058               mode_mv[this_mode][ref].as_int;
2059           if (num_4x4_blocks_wide > 1)
2060             bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int =
2061                 mode_mv[this_mode][ref].as_int;
2062           if (num_4x4_blocks_high > 1)
2063             bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int =
2064                 mode_mv[this_mode][ref].as_int;
2065         }
2066 
2067         // Trap vectors that reach beyond the UMV borders
2068         if (mv_check_bounds(&x->mv_limits, &mode_mv[this_mode][0].as_mv) ||
2069             (has_second_rf &&
2070              mv_check_bounds(&x->mv_limits, &mode_mv[this_mode][1].as_mv)))
2071           continue;
2072 
2073         if (filter_idx > 0) {
2074           BEST_SEG_INFO *ref_bsi = bsi_buf;
2075           subpelmv = 0;
2076           have_ref = 1;
2077 
2078           for (ref = 0; ref < 1 + has_second_rf; ++ref) {
2079             subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
2080             have_ref &= mode_mv[this_mode][ref].as_int ==
2081                         ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
2082           }
2083 
2084           if (filter_idx > 1 && !subpelmv && !have_ref) {
2085             ref_bsi = bsi_buf + 1;
2086             have_ref = 1;
2087             for (ref = 0; ref < 1 + has_second_rf; ++ref)
2088               have_ref &= mode_mv[this_mode][ref].as_int ==
2089                           ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
2090           }
2091 
2092           if (!subpelmv && have_ref &&
2093               ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
2094             memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
2095                    sizeof(SEG_RDSTAT));
2096             if (num_4x4_blocks_wide > 1)
2097               bsi->rdstat[i + 1][mode_idx].eobs =
2098                   ref_bsi->rdstat[i + 1][mode_idx].eobs;
2099             if (num_4x4_blocks_high > 1)
2100               bsi->rdstat[i + 2][mode_idx].eobs =
2101                   ref_bsi->rdstat[i + 2][mode_idx].eobs;
2102 
2103             if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
2104               mode_selected = this_mode;
2105               best_rd = bsi->rdstat[i][mode_idx].brdcost;
2106             }
2107             continue;
2108           }
2109         }
2110 
2111         bsi->rdstat[i][mode_idx].brdcost = encode_inter_mb_segment(
2112             cpi, x, bsi->segment_rd - this_segment_rd, i,
2113             &bsi->rdstat[i][mode_idx].byrate, &bsi->rdstat[i][mode_idx].bdist,
2114             &bsi->rdstat[i][mode_idx].bsse, bsi->rdstat[i][mode_idx].ta,
2115             bsi->rdstat[i][mode_idx].tl, mi_row, mi_col);
2116         if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
2117           bsi->rdstat[i][mode_idx].brdcost +=
2118               RDCOST(x->rdmult, x->rddiv, bsi->rdstat[i][mode_idx].brate, 0);
2119           bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
2120           bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
2121           if (num_4x4_blocks_wide > 1)
2122             bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
2123           if (num_4x4_blocks_high > 1)
2124             bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
2125         }
2126 
2127         if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
2128           mode_selected = this_mode;
2129           best_rd = bsi->rdstat[i][mode_idx].brdcost;
2130         }
2131       } /*for each 4x4 mode*/
2132 
2133       if (best_rd == INT64_MAX) {
2134         int iy, midx;
2135         for (iy = i + 1; iy < 4; ++iy)
2136           for (midx = 0; midx < INTER_MODES; ++midx)
2137             bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2138         bsi->segment_rd = INT64_MAX;
2139         return INT64_MAX;
2140       }
2141 
2142       mode_idx = INTER_OFFSET(mode_selected);
2143       memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
2144       memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
2145 
2146       set_and_cost_bmi_mvs(cpi, x, xd, i, mode_selected, mode_mv[mode_selected],
2147                            frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
2148                            x->mvcost);
2149 
2150       br += bsi->rdstat[i][mode_idx].brate;
2151       bd += bsi->rdstat[i][mode_idx].bdist;
2152       block_sse += bsi->rdstat[i][mode_idx].bsse;
2153       segmentyrate += bsi->rdstat[i][mode_idx].byrate;
2154       this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
2155 
2156       if (this_segment_rd > bsi->segment_rd) {
2157         int iy, midx;
2158         for (iy = i + 1; iy < 4; ++iy)
2159           for (midx = 0; midx < INTER_MODES; ++midx)
2160             bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2161         bsi->segment_rd = INT64_MAX;
2162         return INT64_MAX;
2163       }
2164     }
2165   } /* for each label */
2166 
2167   bsi->r = br;
2168   bsi->d = bd;
2169   bsi->segment_yrate = segmentyrate;
2170   bsi->segment_rd = this_segment_rd;
2171   bsi->sse = block_sse;
2172 
2173   // update the coding decisions
2174   for (k = 0; k < 4; ++k) bsi->modes[k] = mi->bmi[k].as_mode;
2175 
2176   if (bsi->segment_rd > best_rd) return INT64_MAX;
2177   /* set it to the best */
2178   for (i = 0; i < 4; i++) {
2179     mode_idx = INTER_OFFSET(bsi->modes[i]);
2180     mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
2181     if (has_second_ref(mi))
2182       mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
2183     x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
2184     mi->bmi[i].as_mode = bsi->modes[i];
2185   }
2186 
2187   /*
2188    * used to set mbmi->mv.as_int
2189    */
2190   *returntotrate = bsi->r;
2191   *returndistortion = bsi->d;
2192   *returnyrate = bsi->segment_yrate;
2193   *skippable = vp9_is_skippable_in_plane(x, BLOCK_8X8, 0);
2194   *psse = bsi->sse;
2195   mi->mode = bsi->modes[3];
2196 
2197   return bsi->segment_rd;
2198 }
2199 
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)2200 static void estimate_ref_frame_costs(const VP9_COMMON *cm,
2201                                      const MACROBLOCKD *xd, int segment_id,
2202                                      unsigned int *ref_costs_single,
2203                                      unsigned int *ref_costs_comp,
2204                                      vpx_prob *comp_mode_p) {
2205   int seg_ref_active =
2206       segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME);
2207   if (seg_ref_active) {
2208     memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
2209     memset(ref_costs_comp, 0, MAX_REF_FRAMES * sizeof(*ref_costs_comp));
2210     *comp_mode_p = 128;
2211   } else {
2212     vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
2213     vpx_prob comp_inter_p = 128;
2214 
2215     if (cm->reference_mode == REFERENCE_MODE_SELECT) {
2216       comp_inter_p = vp9_get_reference_mode_prob(cm, xd);
2217       *comp_mode_p = comp_inter_p;
2218     } else {
2219       *comp_mode_p = 128;
2220     }
2221 
2222     ref_costs_single[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
2223 
2224     if (cm->reference_mode != COMPOUND_REFERENCE) {
2225       vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
2226       vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
2227       unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
2228 
2229       if (cm->reference_mode == REFERENCE_MODE_SELECT)
2230         base_cost += vp9_cost_bit(comp_inter_p, 0);
2231 
2232       ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] =
2233           ref_costs_single[ALTREF_FRAME] = base_cost;
2234       ref_costs_single[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
2235       ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
2236       ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
2237       ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
2238       ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
2239     } else {
2240       ref_costs_single[LAST_FRAME] = 512;
2241       ref_costs_single[GOLDEN_FRAME] = 512;
2242       ref_costs_single[ALTREF_FRAME] = 512;
2243     }
2244     if (cm->reference_mode != SINGLE_REFERENCE) {
2245       vpx_prob ref_comp_p = vp9_get_pred_prob_comp_ref_p(cm, xd);
2246       unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
2247 
2248       if (cm->reference_mode == REFERENCE_MODE_SELECT)
2249         base_cost += vp9_cost_bit(comp_inter_p, 1);
2250 
2251       ref_costs_comp[LAST_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 0);
2252       ref_costs_comp[GOLDEN_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 1);
2253     } else {
2254       ref_costs_comp[LAST_FRAME] = 512;
2255       ref_costs_comp[GOLDEN_FRAME] = 512;
2256     }
2257   }
2258 }
2259 
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)2260 static void store_coding_context(
2261     MACROBLOCK *x, PICK_MODE_CONTEXT *ctx, int mode_index,
2262     int64_t comp_pred_diff[REFERENCE_MODES],
2263     int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS], int skippable) {
2264   MACROBLOCKD *const xd = &x->e_mbd;
2265 
2266   // Take a snapshot of the coding context so it can be
2267   // restored if we decide to encode this way
2268   ctx->skip = x->skip;
2269   ctx->skippable = skippable;
2270   ctx->best_mode_index = mode_index;
2271   ctx->mic = *xd->mi[0];
2272   ctx->mbmi_ext = *x->mbmi_ext;
2273   ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
2274   ctx->comp_pred_diff = (int)comp_pred_diff[COMPOUND_REFERENCE];
2275   ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
2276 
2277   memcpy(ctx->best_filter_diff, best_filter_diff,
2278          sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
2279 }
2280 
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])2281 static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
2282                                MV_REFERENCE_FRAME ref_frame,
2283                                BLOCK_SIZE block_size, int mi_row, int mi_col,
2284                                int_mv frame_nearest_mv[MAX_REF_FRAMES],
2285                                int_mv frame_near_mv[MAX_REF_FRAMES],
2286                                struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
2287   const VP9_COMMON *cm = &cpi->common;
2288   const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2289   MACROBLOCKD *const xd = &x->e_mbd;
2290   MODE_INFO *const mi = xd->mi[0];
2291   int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
2292   const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2293   MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2294 
2295   assert(yv12 != NULL);
2296 
2297   // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
2298   // use the UV scaling factors.
2299   vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
2300 
2301   // Gets an initial list of candidate vectors from neighbours and orders them
2302   vp9_find_mv_refs(cm, xd, mi, ref_frame, candidates, mi_row, mi_col,
2303                    mbmi_ext->mode_context);
2304 
2305   // Candidate refinement carried out at encoder and decoder
2306   vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2307                         &frame_nearest_mv[ref_frame],
2308                         &frame_near_mv[ref_frame]);
2309 
2310   // Further refinement that is encode side only to test the top few candidates
2311   // in full and choose the best as the centre point for subsequent searches.
2312   // The current implementation doesn't support scaling.
2313   if (!vp9_is_scaled(sf) && block_size >= BLOCK_8X8)
2314     vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
2315                 block_size);
2316 }
2317 
2318 #if CONFIG_NON_GREEDY_MV
2319 #define MAX_PREV_NB_FULL_MV_NUM 8
find_prev_nb_full_mvs(const VP9_COMMON * cm,const MACROBLOCKD * xd,int ref_frame,BLOCK_SIZE bsize,int mi_row,int mi_col,int_mv * nb_full_mvs)2320 static int find_prev_nb_full_mvs(const VP9_COMMON *cm, const MACROBLOCKD *xd,
2321                                  int ref_frame, BLOCK_SIZE bsize, int mi_row,
2322                                  int mi_col, int_mv *nb_full_mvs) {
2323   int i;
2324   const TileInfo *tile = &xd->tile;
2325   int full_mv_num = 0;
2326   assert(bsize >= BLOCK_8X8);
2327   for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
2328     const POSITION *mv_ref = &mv_ref_blocks[bsize][i];
2329     if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
2330       const MODE_INFO *nb_mi =
2331           xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
2332       if (nb_mi->sb_type >= BLOCK_8X8) {
2333         if (nb_mi->ref_frame[0] == ref_frame) {
2334           nb_full_mvs[full_mv_num].as_mv = get_full_mv(&nb_mi->mv[0].as_mv);
2335           ++full_mv_num;
2336           if (full_mv_num >= MAX_PREV_NB_FULL_MV_NUM) {
2337             return full_mv_num;
2338           }
2339         } else if (nb_mi->ref_frame[1] == ref_frame) {
2340           nb_full_mvs[full_mv_num].as_mv = get_full_mv(&nb_mi->mv[1].as_mv);
2341           ++full_mv_num;
2342           if (full_mv_num >= MAX_PREV_NB_FULL_MV_NUM) {
2343             return full_mv_num;
2344           }
2345         }
2346       } else {
2347         int j;
2348         for (j = 0; j < 4; ++j) {
2349           // TODO(angiebird): avoid using duplicated mvs
2350           if (nb_mi->ref_frame[0] == ref_frame) {
2351             nb_full_mvs[full_mv_num].as_mv =
2352                 get_full_mv(&nb_mi->bmi[j].as_mv[0].as_mv);
2353             ++full_mv_num;
2354             if (full_mv_num >= MAX_PREV_NB_FULL_MV_NUM) {
2355               return full_mv_num;
2356             }
2357           } else if (nb_mi->ref_frame[1] == ref_frame) {
2358             nb_full_mvs[full_mv_num].as_mv =
2359                 get_full_mv(&nb_mi->bmi[j].as_mv[1].as_mv);
2360             ++full_mv_num;
2361             if (full_mv_num >= MAX_PREV_NB_FULL_MV_NUM) {
2362               return full_mv_num;
2363             }
2364           }
2365         }
2366       }
2367     }
2368   }
2369   return full_mv_num;
2370 }
2371 #endif  // CONFIG_NON_GREEDY_MV
2372 
single_motion_search(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int mi_row,int mi_col,int_mv * tmp_mv,int * rate_mv)2373 static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
2374                                  int mi_row, int mi_col, int_mv *tmp_mv,
2375                                  int *rate_mv) {
2376   MACROBLOCKD *xd = &x->e_mbd;
2377   const VP9_COMMON *cm = &cpi->common;
2378   MODE_INFO *mi = xd->mi[0];
2379   struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } };
2380   int step_param;
2381   MV mvp_full;
2382   int ref = mi->ref_frame[0];
2383   MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2384   const MvLimits tmp_mv_limits = x->mv_limits;
2385   int cost_list[5];
2386   const int best_predmv_idx = x->mv_best_ref_index[ref];
2387   const YV12_BUFFER_CONFIG *scaled_ref_frame =
2388       vp9_get_scaled_ref_frame(cpi, ref);
2389   const int pw = num_4x4_blocks_wide_lookup[bsize] << 2;
2390   const int ph = num_4x4_blocks_high_lookup[bsize] << 2;
2391   MV pred_mv[3];
2392 
2393 #if CONFIG_NON_GREEDY_MV
2394   double mv_dist = 0;
2395   double mv_cost = 0;
2396   double lambda = (pw * ph) / 4;
2397   double bestsme;
2398   int_mv nb_full_mvs[MAX_PREV_NB_FULL_MV_NUM];
2399 
2400   const int nb_full_mv_num =
2401       find_prev_nb_full_mvs(cm, xd, ref, bsize, mi_row, mi_col, nb_full_mvs);
2402 #else   // CONFIG_NON_GREEDY_MV
2403   int bestsme = INT_MAX;
2404   int sadpb = x->sadperbit16;
2405 #endif  // CONFIG_NON_GREEDY_MV
2406 
2407   pred_mv[0] = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2408   pred_mv[1] = x->mbmi_ext->ref_mvs[ref][1].as_mv;
2409   pred_mv[2] = x->pred_mv[ref];
2410 
2411   if (scaled_ref_frame) {
2412     int i;
2413     // Swap out the reference frame for a version that's been scaled to
2414     // match the resolution of the current frame, allowing the existing
2415     // motion search code to be used without additional modifications.
2416     for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
2417 
2418     vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
2419   }
2420 
2421   // Work out the size of the first step in the mv step search.
2422   // 0 here is maximum length first step. 1 is VPXMAX >> 1 etc.
2423   if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
2424     // Take wtd average of the step_params based on the last frame's
2425     // max mv magnitude and that based on the best ref mvs of the current
2426     // block for the given reference.
2427     step_param =
2428         (vp9_init_search_range(x->max_mv_context[ref]) + cpi->mv_step_param) /
2429         2;
2430   } else {
2431     step_param = cpi->mv_step_param;
2432   }
2433 
2434   if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64) {
2435     const int boffset =
2436         2 * (b_width_log2_lookup[BLOCK_64X64] -
2437              VPXMIN(b_height_log2_lookup[bsize], b_width_log2_lookup[bsize]));
2438     step_param = VPXMAX(step_param, boffset);
2439   }
2440 
2441   if (cpi->sf.adaptive_motion_search) {
2442     int bwl = b_width_log2_lookup[bsize];
2443     int bhl = b_height_log2_lookup[bsize];
2444     int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
2445 
2446     if (tlevel < 5) step_param += 2;
2447 
2448     // prev_mv_sad is not setup for dynamically scaled frames.
2449     if (cpi->oxcf.resize_mode != RESIZE_DYNAMIC) {
2450       int i;
2451       for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
2452         if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
2453           x->pred_mv[ref].row = INT16_MAX;
2454           x->pred_mv[ref].col = INT16_MAX;
2455           tmp_mv->as_int = INVALID_MV;
2456 
2457           if (scaled_ref_frame) {
2458             int i;
2459             for (i = 0; i < MAX_MB_PLANE; ++i)
2460               xd->plane[i].pre[0] = backup_yv12[i];
2461           }
2462           return;
2463         }
2464       }
2465     }
2466   }
2467 
2468   // Note: MV limits are modified here. Always restore the original values
2469   // after full-pixel motion search.
2470   vp9_set_mv_search_range(&x->mv_limits, &ref_mv);
2471 
2472   mvp_full = pred_mv[best_predmv_idx];
2473   mvp_full.col >>= 3;
2474   mvp_full.row >>= 3;
2475 
2476 #if CONFIG_NON_GREEDY_MV
2477   bestsme = vp9_full_pixel_diamond_new(
2478       cpi, x, &mvp_full, step_param, lambda, 1, &cpi->fn_ptr[bsize],
2479       nb_full_mvs, nb_full_mv_num, &tmp_mv->as_mv, &mv_dist, &mv_cost);
2480 #else   // CONFIG_NON_GREEDY_MV
2481   bestsme = vp9_full_pixel_search(
2482       cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method, sadpb,
2483       cond_cost_list(cpi, cost_list), &ref_mv, &tmp_mv->as_mv, INT_MAX, 1);
2484 #endif  // CONFIG_NON_GREEDY_MV
2485 
2486   if (cpi->sf.enhanced_full_pixel_motion_search) {
2487     int i;
2488     for (i = 0; i < 3; ++i) {
2489 #if CONFIG_NON_GREEDY_MV
2490       double this_me;
2491 #else   // CONFIG_NON_GREEDY_MV
2492       int this_me;
2493 #endif  // CONFIG_NON_GREEDY_MV
2494       MV this_mv;
2495       int diff_row;
2496       int diff_col;
2497       int step;
2498 
2499       if (pred_mv[i].row == INT16_MAX || pred_mv[i].col == INT16_MAX) continue;
2500       if (i == best_predmv_idx) continue;
2501 
2502       diff_row = ((int)pred_mv[i].row -
2503                   pred_mv[i > 0 ? (i - 1) : best_predmv_idx].row) >>
2504                  3;
2505       diff_col = ((int)pred_mv[i].col -
2506                   pred_mv[i > 0 ? (i - 1) : best_predmv_idx].col) >>
2507                  3;
2508       if (diff_row == 0 && diff_col == 0) continue;
2509       if (diff_row < 0) diff_row = -diff_row;
2510       if (diff_col < 0) diff_col = -diff_col;
2511       step = get_msb((diff_row + diff_col + 1) >> 1);
2512       if (step <= 0) continue;
2513 
2514       mvp_full = pred_mv[i];
2515       mvp_full.col >>= 3;
2516       mvp_full.row >>= 3;
2517 #if CONFIG_NON_GREEDY_MV
2518       this_me = vp9_full_pixel_diamond_new(
2519           cpi, x, &mvp_full, VPXMAX(step_param, MAX_MVSEARCH_STEPS - step),
2520           lambda, 1, &cpi->fn_ptr[bsize], nb_full_mvs, nb_full_mv_num, &this_mv,
2521           &mv_dist, &mv_cost);
2522 #else   // CONFIG_NON_GREEDY_MV
2523       this_me = vp9_full_pixel_search(
2524           cpi, x, bsize, &mvp_full,
2525           VPXMAX(step_param, MAX_MVSEARCH_STEPS - step),
2526           cpi->sf.mv.search_method, sadpb, cond_cost_list(cpi, cost_list),
2527           &ref_mv, &this_mv, INT_MAX, 1);
2528 #endif  // CONFIG_NON_GREEDY_MV
2529       if (this_me < bestsme) {
2530         tmp_mv->as_mv = this_mv;
2531         bestsme = this_me;
2532       }
2533     }
2534   }
2535 
2536   x->mv_limits = tmp_mv_limits;
2537 
2538   if (bestsme < INT_MAX) {
2539     uint32_t dis; /* TODO: use dis in distortion calculation later. */
2540     cpi->find_fractional_mv_step(
2541         x, &tmp_mv->as_mv, &ref_mv, cm->allow_high_precision_mv, x->errorperbit,
2542         &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
2543         cpi->sf.mv.subpel_search_level, cond_cost_list(cpi, cost_list),
2544         x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, pw, ph,
2545         cpi->sf.use_accurate_subpel_search);
2546   }
2547   *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost,
2548                              x->mvcost, MV_COST_WEIGHT);
2549 
2550   x->pred_mv[ref] = tmp_mv->as_mv;
2551 
2552   if (scaled_ref_frame) {
2553     int i;
2554     for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
2555   }
2556 }
2557 
restore_dst_buf(MACROBLOCKD * xd,uint8_t * orig_dst[MAX_MB_PLANE],int orig_dst_stride[MAX_MB_PLANE])2558 static INLINE void restore_dst_buf(MACROBLOCKD *xd,
2559                                    uint8_t *orig_dst[MAX_MB_PLANE],
2560                                    int orig_dst_stride[MAX_MB_PLANE]) {
2561   int i;
2562   for (i = 0; i < MAX_MB_PLANE; i++) {
2563     xd->plane[i].dst.buf = orig_dst[i];
2564     xd->plane[i].dst.stride = orig_dst_stride[i];
2565   }
2566 }
2567 
2568 // In some situations we want to discount tha pparent cost of a new motion
2569 // vector. Where there is a subtle motion field and especially where there is
2570 // low spatial complexity then it can be hard to cover the cost of a new motion
2571 // vector in a single block, even if that motion vector reduces distortion.
2572 // However, once established that vector may be usable through the nearest and
2573 // near mv modes to reduce distortion in subsequent blocks and also improve
2574 // visual quality.
discount_newmv_test(const VP9_COMP * cpi,int this_mode,int_mv this_mv,int_mv (* mode_mv)[MAX_REF_FRAMES],int ref_frame)2575 static int discount_newmv_test(const VP9_COMP *cpi, int this_mode,
2576                                int_mv this_mv,
2577                                int_mv (*mode_mv)[MAX_REF_FRAMES],
2578                                int ref_frame) {
2579   return (!cpi->rc.is_src_frame_alt_ref && (this_mode == NEWMV) &&
2580           (this_mv.as_int != 0) &&
2581           ((mode_mv[NEARESTMV][ref_frame].as_int == 0) ||
2582            (mode_mv[NEARESTMV][ref_frame].as_int == INVALID_MV)) &&
2583           ((mode_mv[NEARMV][ref_frame].as_int == 0) ||
2584            (mode_mv[NEARMV][ref_frame].as_int == INVALID_MV)));
2585 }
2586 
handle_inter_mode(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int * rate2,int64_t * distortion,int * skippable,int * rate_y,int * rate_uv,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[])2587 static int64_t handle_inter_mode(
2588     VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, int *rate2,
2589     int64_t *distortion, int *skippable, int *rate_y, int *rate_uv,
2590     int *disable_skip, int_mv (*mode_mv)[MAX_REF_FRAMES], int mi_row,
2591     int mi_col, int_mv single_newmv[MAX_REF_FRAMES],
2592     INTERP_FILTER (*single_filter)[MAX_REF_FRAMES],
2593     int (*single_skippable)[MAX_REF_FRAMES], int64_t *psse,
2594     const int64_t ref_best_rd, int64_t *mask_filter, int64_t filter_cache[]) {
2595   VP9_COMMON *cm = &cpi->common;
2596   MACROBLOCKD *xd = &x->e_mbd;
2597   MODE_INFO *mi = xd->mi[0];
2598   MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2599   const int is_comp_pred = has_second_ref(mi);
2600   const int this_mode = mi->mode;
2601   int_mv *frame_mv = mode_mv[this_mode];
2602   int i;
2603   int refs[2] = { mi->ref_frame[0],
2604                   (mi->ref_frame[1] < 0 ? 0 : mi->ref_frame[1]) };
2605   int_mv cur_mv[2];
2606 #if CONFIG_VP9_HIGHBITDEPTH
2607   DECLARE_ALIGNED(16, uint16_t, tmp_buf16[MAX_MB_PLANE * 64 * 64]);
2608   uint8_t *tmp_buf;
2609 #else
2610   DECLARE_ALIGNED(16, uint8_t, tmp_buf[MAX_MB_PLANE * 64 * 64]);
2611 #endif  // CONFIG_VP9_HIGHBITDEPTH
2612   int pred_exists = 0;
2613   int intpel_mv;
2614   int64_t rd, tmp_rd, best_rd = INT64_MAX;
2615   int best_needs_copy = 0;
2616   uint8_t *orig_dst[MAX_MB_PLANE];
2617   int orig_dst_stride[MAX_MB_PLANE];
2618   int rs = 0;
2619   INTERP_FILTER best_filter = SWITCHABLE;
2620   uint8_t skip_txfm[MAX_MB_PLANE << 2] = { 0 };
2621   int64_t bsse[MAX_MB_PLANE << 2] = { 0 };
2622 
2623   int bsl = mi_width_log2_lookup[bsize];
2624   int pred_filter_search =
2625       cpi->sf.cb_pred_filter_search
2626           ? (((mi_row + mi_col) >> bsl) +
2627              get_chessboard_index(cm->current_video_frame)) &
2628                 0x1
2629           : 0;
2630 
2631   int skip_txfm_sb = 0;
2632   int64_t skip_sse_sb = INT64_MAX;
2633   int64_t distortion_y = 0, distortion_uv = 0;
2634 
2635 #if CONFIG_VP9_HIGHBITDEPTH
2636   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2637     tmp_buf = CONVERT_TO_BYTEPTR(tmp_buf16);
2638   } else {
2639     tmp_buf = (uint8_t *)tmp_buf16;
2640   }
2641 #endif  // CONFIG_VP9_HIGHBITDEPTH
2642 
2643   if (pred_filter_search) {
2644     INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE;
2645     if (xd->above_mi && is_inter_block(xd->above_mi))
2646       af = xd->above_mi->interp_filter;
2647     if (xd->left_mi && is_inter_block(xd->left_mi))
2648       lf = xd->left_mi->interp_filter;
2649 
2650     if ((this_mode != NEWMV) || (af == lf)) best_filter = af;
2651   }
2652 
2653   if (is_comp_pred) {
2654     if (frame_mv[refs[0]].as_int == INVALID_MV ||
2655         frame_mv[refs[1]].as_int == INVALID_MV)
2656       return INT64_MAX;
2657 
2658     if (cpi->sf.adaptive_mode_search) {
2659       if (single_filter[this_mode][refs[0]] ==
2660           single_filter[this_mode][refs[1]])
2661         best_filter = single_filter[this_mode][refs[0]];
2662     }
2663   }
2664 
2665   if (this_mode == NEWMV) {
2666     int rate_mv;
2667     if (is_comp_pred) {
2668       // Initialize mv using single prediction mode result.
2669       frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
2670       frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
2671 
2672       if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
2673         joint_motion_search(cpi, x, bsize, frame_mv, mi_row, mi_col,
2674                             single_newmv, &rate_mv);
2675       } else {
2676         rate_mv = vp9_mv_bit_cost(&frame_mv[refs[0]].as_mv,
2677                                   &x->mbmi_ext->ref_mvs[refs[0]][0].as_mv,
2678                                   x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2679         rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]].as_mv,
2680                                    &x->mbmi_ext->ref_mvs[refs[1]][0].as_mv,
2681                                    x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2682       }
2683       *rate2 += rate_mv;
2684     } else {
2685       int_mv tmp_mv;
2686       single_motion_search(cpi, x, bsize, mi_row, mi_col, &tmp_mv, &rate_mv);
2687       if (tmp_mv.as_int == INVALID_MV) return INT64_MAX;
2688 
2689       frame_mv[refs[0]].as_int = xd->mi[0]->bmi[0].as_mv[0].as_int =
2690           tmp_mv.as_int;
2691       single_newmv[refs[0]].as_int = tmp_mv.as_int;
2692 
2693       // Estimate the rate implications of a new mv but discount this
2694       // under certain circumstances where we want to help initiate a weak
2695       // motion field, where the distortion gain for a single block may not
2696       // be enough to overcome the cost of a new mv.
2697       if (discount_newmv_test(cpi, this_mode, tmp_mv, mode_mv, refs[0])) {
2698         *rate2 += VPXMAX((rate_mv / NEW_MV_DISCOUNT_FACTOR), 1);
2699       } else {
2700         *rate2 += rate_mv;
2701       }
2702     }
2703   }
2704 
2705   for (i = 0; i < is_comp_pred + 1; ++i) {
2706     cur_mv[i] = frame_mv[refs[i]];
2707     // Clip "next_nearest" so that it does not extend to far out of image
2708     if (this_mode != NEWMV) clamp_mv2(&cur_mv[i].as_mv, xd);
2709 
2710     if (mv_check_bounds(&x->mv_limits, &cur_mv[i].as_mv)) return INT64_MAX;
2711     mi->mv[i].as_int = cur_mv[i].as_int;
2712   }
2713 
2714   // do first prediction into the destination buffer. Do the next
2715   // prediction into a temporary buffer. Then keep track of which one
2716   // of these currently holds the best predictor, and use the other
2717   // one for future predictions. In the end, copy from tmp_buf to
2718   // dst if necessary.
2719   for (i = 0; i < MAX_MB_PLANE; i++) {
2720     orig_dst[i] = xd->plane[i].dst.buf;
2721     orig_dst_stride[i] = xd->plane[i].dst.stride;
2722   }
2723 
2724   // We don't include the cost of the second reference here, because there
2725   // are only two options: Last/ARF or Golden/ARF; The second one is always
2726   // known, which is ARF.
2727   //
2728   // Under some circumstances we discount the cost of new mv mode to encourage
2729   // initiation of a motion field.
2730   if (discount_newmv_test(cpi, this_mode, frame_mv[refs[0]], mode_mv,
2731                           refs[0])) {
2732     *rate2 +=
2733         VPXMIN(cost_mv_ref(cpi, this_mode, mbmi_ext->mode_context[refs[0]]),
2734                cost_mv_ref(cpi, NEARESTMV, mbmi_ext->mode_context[refs[0]]));
2735   } else {
2736     *rate2 += cost_mv_ref(cpi, this_mode, mbmi_ext->mode_context[refs[0]]);
2737   }
2738 
2739   if (RDCOST(x->rdmult, x->rddiv, *rate2, 0) > ref_best_rd &&
2740       mi->mode != NEARESTMV)
2741     return INT64_MAX;
2742 
2743   pred_exists = 0;
2744   // Are all MVs integer pel for Y and UV
2745   intpel_mv = !mv_has_subpel(&mi->mv[0].as_mv);
2746   if (is_comp_pred) intpel_mv &= !mv_has_subpel(&mi->mv[1].as_mv);
2747 
2748   // Search for best switchable filter by checking the variance of
2749   // pred error irrespective of whether the filter will be used
2750   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) filter_cache[i] = INT64_MAX;
2751 
2752   if (cm->interp_filter != BILINEAR) {
2753     if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
2754       best_filter = EIGHTTAP;
2755     } else if (best_filter == SWITCHABLE) {
2756       int newbest;
2757       int tmp_rate_sum = 0;
2758       int64_t tmp_dist_sum = 0;
2759 
2760       for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
2761         int j;
2762         int64_t rs_rd;
2763         int tmp_skip_sb = 0;
2764         int64_t tmp_skip_sse = INT64_MAX;
2765 
2766         mi->interp_filter = i;
2767         rs = vp9_get_switchable_rate(cpi, xd);
2768         rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
2769 
2770         if (i > 0 && intpel_mv) {
2771           rd = RDCOST(x->rdmult, x->rddiv, tmp_rate_sum, tmp_dist_sum);
2772           filter_cache[i] = rd;
2773           filter_cache[SWITCHABLE_FILTERS] =
2774               VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2775           if (cm->interp_filter == SWITCHABLE) rd += rs_rd;
2776           *mask_filter = VPXMAX(*mask_filter, rd);
2777         } else {
2778           int rate_sum = 0;
2779           int64_t dist_sum = 0;
2780           if (i > 0 && cpi->sf.adaptive_interp_filter_search &&
2781               (cpi->sf.interp_filter_search_mask & (1 << i))) {
2782             rate_sum = INT_MAX;
2783             dist_sum = INT64_MAX;
2784             continue;
2785           }
2786 
2787           if ((cm->interp_filter == SWITCHABLE && (!i || best_needs_copy)) ||
2788               (cm->interp_filter != SWITCHABLE &&
2789                (cm->interp_filter == mi->interp_filter ||
2790                 (i == 0 && intpel_mv)))) {
2791             restore_dst_buf(xd, orig_dst, orig_dst_stride);
2792           } else {
2793             for (j = 0; j < MAX_MB_PLANE; j++) {
2794               xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
2795               xd->plane[j].dst.stride = 64;
2796             }
2797           }
2798           vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2799           model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum, &tmp_skip_sb,
2800                           &tmp_skip_sse);
2801 
2802           rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum);
2803           filter_cache[i] = rd;
2804           filter_cache[SWITCHABLE_FILTERS] =
2805               VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2806           if (cm->interp_filter == SWITCHABLE) rd += rs_rd;
2807           *mask_filter = VPXMAX(*mask_filter, rd);
2808 
2809           if (i == 0 && intpel_mv) {
2810             tmp_rate_sum = rate_sum;
2811             tmp_dist_sum = dist_sum;
2812           }
2813         }
2814 
2815         if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2816           if (rd / 2 > ref_best_rd) {
2817             restore_dst_buf(xd, orig_dst, orig_dst_stride);
2818             return INT64_MAX;
2819           }
2820         }
2821         newbest = i == 0 || rd < best_rd;
2822 
2823         if (newbest) {
2824           best_rd = rd;
2825           best_filter = mi->interp_filter;
2826           if (cm->interp_filter == SWITCHABLE && i && !intpel_mv)
2827             best_needs_copy = !best_needs_copy;
2828         }
2829 
2830         if ((cm->interp_filter == SWITCHABLE && newbest) ||
2831             (cm->interp_filter != SWITCHABLE &&
2832              cm->interp_filter == mi->interp_filter)) {
2833           pred_exists = 1;
2834           tmp_rd = best_rd;
2835 
2836           skip_txfm_sb = tmp_skip_sb;
2837           skip_sse_sb = tmp_skip_sse;
2838           memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2839           memcpy(bsse, x->bsse, sizeof(bsse));
2840         }
2841       }
2842       restore_dst_buf(xd, orig_dst, orig_dst_stride);
2843     }
2844   }
2845   // Set the appropriate filter
2846   mi->interp_filter =
2847       cm->interp_filter != SWITCHABLE ? cm->interp_filter : best_filter;
2848   rs = cm->interp_filter == SWITCHABLE ? vp9_get_switchable_rate(cpi, xd) : 0;
2849 
2850   if (pred_exists) {
2851     if (best_needs_copy) {
2852       // again temporarily set the buffers to local memory to prevent a memcpy
2853       for (i = 0; i < MAX_MB_PLANE; i++) {
2854         xd->plane[i].dst.buf = tmp_buf + i * 64 * 64;
2855         xd->plane[i].dst.stride = 64;
2856       }
2857     }
2858     rd = tmp_rd + RDCOST(x->rdmult, x->rddiv, rs, 0);
2859   } else {
2860     int tmp_rate;
2861     int64_t tmp_dist;
2862     // Handles the special case when a filter that is not in the
2863     // switchable list (ex. bilinear) is indicated at the frame level, or
2864     // skip condition holds.
2865     vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2866     model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist, &skip_txfm_sb,
2867                     &skip_sse_sb);
2868     rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
2869     memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2870     memcpy(bsse, x->bsse, sizeof(bsse));
2871   }
2872 
2873   if (!is_comp_pred) single_filter[this_mode][refs[0]] = mi->interp_filter;
2874 
2875   if (cpi->sf.adaptive_mode_search)
2876     if (is_comp_pred)
2877       if (single_skippable[this_mode][refs[0]] &&
2878           single_skippable[this_mode][refs[1]])
2879         memset(skip_txfm, SKIP_TXFM_AC_DC, sizeof(skip_txfm));
2880 
2881   if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2882     // if current pred_error modeled rd is substantially more than the best
2883     // so far, do not bother doing full rd
2884     if (rd / 2 > ref_best_rd) {
2885       restore_dst_buf(xd, orig_dst, orig_dst_stride);
2886       return INT64_MAX;
2887     }
2888   }
2889 
2890   if (cm->interp_filter == SWITCHABLE) *rate2 += rs;
2891 
2892   memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm));
2893   memcpy(x->bsse, bsse, sizeof(bsse));
2894 
2895   if (!skip_txfm_sb || xd->lossless) {
2896     int skippable_y, skippable_uv;
2897     int64_t sseuv = INT64_MAX;
2898     int64_t rdcosty = INT64_MAX;
2899 
2900     // Y cost and distortion
2901     vp9_subtract_plane(x, bsize, 0);
2902     super_block_yrd(cpi, x, rate_y, &distortion_y, &skippable_y, psse, bsize,
2903                     ref_best_rd);
2904 
2905     if (*rate_y == INT_MAX) {
2906       *rate2 = INT_MAX;
2907       *distortion = INT64_MAX;
2908       restore_dst_buf(xd, orig_dst, orig_dst_stride);
2909       return INT64_MAX;
2910     }
2911 
2912     *rate2 += *rate_y;
2913     *distortion += distortion_y;
2914 
2915     rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
2916     rdcosty = VPXMIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
2917 
2918     if (!super_block_uvrd(cpi, x, rate_uv, &distortion_uv, &skippable_uv,
2919                           &sseuv, bsize, ref_best_rd - rdcosty)) {
2920       *rate2 = INT_MAX;
2921       *distortion = INT64_MAX;
2922       restore_dst_buf(xd, orig_dst, orig_dst_stride);
2923       return INT64_MAX;
2924     }
2925 
2926     *psse += sseuv;
2927     *rate2 += *rate_uv;
2928     *distortion += distortion_uv;
2929     *skippable = skippable_y && skippable_uv;
2930   } else {
2931     x->skip = 1;
2932     *disable_skip = 1;
2933 
2934     // The cost of skip bit needs to be added.
2935     *rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2936 
2937     *distortion = skip_sse_sb;
2938   }
2939 
2940   if (!is_comp_pred) single_skippable[this_mode][refs[0]] = *skippable;
2941 
2942   restore_dst_buf(xd, orig_dst, orig_dst_stride);
2943   return 0;  // The rate-distortion cost will be re-calculated by caller.
2944 }
2945 
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)2946 void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
2947                                BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
2948                                int64_t best_rd) {
2949   VP9_COMMON *const cm = &cpi->common;
2950   MACROBLOCKD *const xd = &x->e_mbd;
2951   struct macroblockd_plane *const pd = xd->plane;
2952   int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
2953   int y_skip = 0, uv_skip = 0;
2954   int64_t dist_y = 0, dist_uv = 0;
2955   TX_SIZE max_uv_tx_size;
2956   x->skip_encode = 0;
2957   ctx->skip = 0;
2958   xd->mi[0]->ref_frame[0] = INTRA_FRAME;
2959   xd->mi[0]->ref_frame[1] = NONE;
2960   // Initialize interp_filter here so we do not have to check for inter block
2961   // modes in get_pred_context_switchable_interp()
2962   xd->mi[0]->interp_filter = SWITCHABLE_FILTERS;
2963 
2964   if (bsize >= BLOCK_8X8) {
2965     if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly, &dist_y,
2966                                &y_skip, bsize, best_rd) >= best_rd) {
2967       rd_cost->rate = INT_MAX;
2968       return;
2969     }
2970   } else {
2971     y_skip = 0;
2972     if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2973                                      &dist_y, best_rd) >= best_rd) {
2974       rd_cost->rate = INT_MAX;
2975       return;
2976     }
2977   }
2978   max_uv_tx_size = uv_txsize_lookup[bsize][xd->mi[0]->tx_size]
2979                                    [pd[1].subsampling_x][pd[1].subsampling_y];
2980   rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly, &dist_uv,
2981                           &uv_skip, VPXMAX(BLOCK_8X8, bsize), max_uv_tx_size);
2982 
2983   if (y_skip && uv_skip) {
2984     rd_cost->rate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
2985                     vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2986     rd_cost->dist = dist_y + dist_uv;
2987   } else {
2988     rd_cost->rate =
2989         rate_y + rate_uv + vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
2990     rd_cost->dist = dist_y + dist_uv;
2991   }
2992 
2993   ctx->mic = *xd->mi[0];
2994   ctx->mbmi_ext = *x->mbmi_ext;
2995   rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
2996 }
2997 
2998 // This function is designed to apply a bias or adjustment to an rd value based
2999 // on the relative variance of the source and reconstruction.
3000 #define VERY_LOW_VAR_THRESH 2
3001 #define LOW_VAR_THRESH 5
3002 #define VAR_MULT 100
3003 static unsigned int max_var_adjust[VP9E_CONTENT_INVALID] = { 16, 16, 100 };
3004 
rd_variance_adjustment(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int64_t * this_rd,MV_REFERENCE_FRAME ref_frame,unsigned int source_variance)3005 static void rd_variance_adjustment(VP9_COMP *cpi, MACROBLOCK *x,
3006                                    BLOCK_SIZE bsize, int64_t *this_rd,
3007                                    MV_REFERENCE_FRAME ref_frame,
3008                                    unsigned int source_variance) {
3009   MACROBLOCKD *const xd = &x->e_mbd;
3010   unsigned int rec_variance;
3011   unsigned int src_variance;
3012   unsigned int src_rec_min;
3013   unsigned int absvar_diff = 0;
3014   unsigned int var_factor = 0;
3015   unsigned int adj_max;
3016   vp9e_tune_content content_type = cpi->oxcf.content;
3017 
3018   if (*this_rd == INT64_MAX) return;
3019 
3020 #if CONFIG_VP9_HIGHBITDEPTH
3021   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
3022     if (source_variance > 100) {
3023       rec_variance = vp9_high_get_sby_perpixel_variance(cpi, &xd->plane[0].dst,
3024                                                         bsize, xd->bd);
3025       src_variance = source_variance;
3026     } else {
3027       rec_variance =
3028           vp9_high_get_sby_variance(cpi, &xd->plane[0].dst, bsize, xd->bd);
3029       src_variance =
3030           vp9_high_get_sby_variance(cpi, &x->plane[0].src, bsize, xd->bd);
3031     }
3032   } else {
3033     if (source_variance > 100) {
3034       rec_variance =
3035           vp9_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
3036       src_variance = source_variance;
3037     } else {
3038       rec_variance = vp9_get_sby_variance(cpi, &xd->plane[0].dst, bsize);
3039       src_variance = vp9_get_sby_variance(cpi, &x->plane[0].src, bsize);
3040     }
3041   }
3042 #else
3043   if (source_variance > 100) {
3044     rec_variance = vp9_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
3045     src_variance = source_variance;
3046   } else {
3047     rec_variance = vp9_get_sby_variance(cpi, &xd->plane[0].dst, bsize);
3048     src_variance = vp9_get_sby_variance(cpi, &x->plane[0].src, bsize);
3049   }
3050 #endif  // CONFIG_VP9_HIGHBITDEPTH
3051 
3052   // Lower of source (raw per pixel value) and recon variance. Note that
3053   // if the source per pixel is 0 then the recon value here will not be per
3054   // pixel (see above) so will likely be much larger.
3055   src_rec_min = VPXMIN(source_variance, rec_variance);
3056 
3057   if (src_rec_min > LOW_VAR_THRESH) return;
3058 
3059   absvar_diff = (src_variance > rec_variance) ? (src_variance - rec_variance)
3060                                               : (rec_variance - src_variance);
3061 
3062   adj_max = max_var_adjust[content_type];
3063 
3064   var_factor =
3065       (unsigned int)((int64_t)VAR_MULT * absvar_diff) / VPXMAX(1, src_variance);
3066   var_factor = VPXMIN(adj_max, var_factor);
3067 
3068   *this_rd += (*this_rd * var_factor) / 100;
3069 
3070   if (content_type == VP9E_CONTENT_FILM) {
3071     if (src_rec_min <= VERY_LOW_VAR_THRESH) {
3072       if (ref_frame == INTRA_FRAME) *this_rd *= 2;
3073       if (bsize > 6) *this_rd *= 2;
3074     }
3075   }
3076 }
3077 
3078 // Do we have an internal image edge (e.g. formatting bars).
vp9_internal_image_edge(VP9_COMP * cpi)3079 int vp9_internal_image_edge(VP9_COMP *cpi) {
3080   return (cpi->oxcf.pass == 2) &&
3081          ((cpi->twopass.this_frame_stats.inactive_zone_rows > 0) ||
3082           (cpi->twopass.this_frame_stats.inactive_zone_cols > 0));
3083 }
3084 
3085 // Checks to see if a super block is on a horizontal image edge.
3086 // In most cases this is the "real" edge unless there are formatting
3087 // bars embedded in the stream.
vp9_active_h_edge(VP9_COMP * cpi,int mi_row,int mi_step)3088 int vp9_active_h_edge(VP9_COMP *cpi, int mi_row, int mi_step) {
3089   int top_edge = 0;
3090   int bottom_edge = cpi->common.mi_rows;
3091   int is_active_h_edge = 0;
3092 
3093   // For two pass account for any formatting bars detected.
3094   if (cpi->oxcf.pass == 2) {
3095     TWO_PASS *twopass = &cpi->twopass;
3096 
3097     // The inactive region is specified in MBs not mi units.
3098     // The image edge is in the following MB row.
3099     top_edge += (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
3100 
3101     bottom_edge -= (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
3102     bottom_edge = VPXMAX(top_edge, bottom_edge);
3103   }
3104 
3105   if (((top_edge >= mi_row) && (top_edge < (mi_row + mi_step))) ||
3106       ((bottom_edge >= mi_row) && (bottom_edge < (mi_row + mi_step)))) {
3107     is_active_h_edge = 1;
3108   }
3109   return is_active_h_edge;
3110 }
3111 
3112 // Checks to see if a super block is on a vertical image edge.
3113 // In most cases this is the "real" edge unless there are formatting
3114 // bars embedded in the stream.
vp9_active_v_edge(VP9_COMP * cpi,int mi_col,int mi_step)3115 int vp9_active_v_edge(VP9_COMP *cpi, int mi_col, int mi_step) {
3116   int left_edge = 0;
3117   int right_edge = cpi->common.mi_cols;
3118   int is_active_v_edge = 0;
3119 
3120   // For two pass account for any formatting bars detected.
3121   if (cpi->oxcf.pass == 2) {
3122     TWO_PASS *twopass = &cpi->twopass;
3123 
3124     // The inactive region is specified in MBs not mi units.
3125     // The image edge is in the following MB row.
3126     left_edge += (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
3127 
3128     right_edge -= (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
3129     right_edge = VPXMAX(left_edge, right_edge);
3130   }
3131 
3132   if (((left_edge >= mi_col) && (left_edge < (mi_col + mi_step))) ||
3133       ((right_edge >= mi_col) && (right_edge < (mi_col + mi_step)))) {
3134     is_active_v_edge = 1;
3135   }
3136   return is_active_v_edge;
3137 }
3138 
3139 // Checks to see if a super block is at the edge of the active image.
3140 // In most cases this is the "real" edge unless there are formatting
3141 // bars embedded in the stream.
vp9_active_edge_sb(VP9_COMP * cpi,int mi_row,int mi_col)3142 int vp9_active_edge_sb(VP9_COMP *cpi, int mi_row, int mi_col) {
3143   return vp9_active_h_edge(cpi, mi_row, MI_BLOCK_SIZE) ||
3144          vp9_active_v_edge(cpi, mi_col, MI_BLOCK_SIZE);
3145 }
3146 
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)3147 void vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, TileDataEnc *tile_data,
3148                                MACROBLOCK *x, int mi_row, int mi_col,
3149                                RD_COST *rd_cost, BLOCK_SIZE bsize,
3150                                PICK_MODE_CONTEXT *ctx, int64_t best_rd_so_far) {
3151   VP9_COMMON *const cm = &cpi->common;
3152   TileInfo *const tile_info = &tile_data->tile_info;
3153   RD_OPT *const rd_opt = &cpi->rd;
3154   SPEED_FEATURES *const sf = &cpi->sf;
3155   MACROBLOCKD *const xd = &x->e_mbd;
3156   MODE_INFO *const mi = xd->mi[0];
3157   MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
3158   const struct segmentation *const seg = &cm->seg;
3159   PREDICTION_MODE this_mode;
3160   MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3161   unsigned char segment_id = mi->segment_id;
3162   int comp_pred, i, k;
3163   int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3164   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3165   int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
3166   INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES];
3167   int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES];
3168   static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
3169                                     VP9_ALT_FLAG };
3170   int64_t best_rd = best_rd_so_far;
3171   int64_t best_pred_diff[REFERENCE_MODES];
3172   int64_t best_pred_rd[REFERENCE_MODES];
3173   int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3174   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3175   MODE_INFO best_mbmode;
3176   int best_mode_skippable = 0;
3177   int midx, best_mode_index = -1;
3178   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3179   vpx_prob comp_mode_p;
3180   int64_t best_intra_rd = INT64_MAX;
3181   unsigned int best_pred_sse = UINT_MAX;
3182   PREDICTION_MODE best_intra_mode = DC_PRED;
3183   int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
3184   int64_t dist_uv[TX_SIZES];
3185   int skip_uv[TX_SIZES];
3186   PREDICTION_MODE mode_uv[TX_SIZES];
3187   const int intra_cost_penalty =
3188       vp9_get_intra_cost_penalty(cpi, bsize, cm->base_qindex, cm->y_dc_delta_q);
3189   int best_skip2 = 0;
3190   uint8_t ref_frame_skip_mask[2] = { 0, 1 };
3191   uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 };
3192   int mode_skip_start = sf->mode_skip_start + 1;
3193   const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
3194   const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
3195   int64_t mode_threshold[MAX_MODES];
3196   int8_t *tile_mode_map = tile_data->mode_map[bsize];
3197   int8_t mode_map[MAX_MODES];  // Maintain mode_map information locally to avoid
3198                                // lock mechanism involved with reads from
3199                                // tile_mode_map
3200   const int mode_search_skip_flags = sf->mode_search_skip_flags;
3201   const int is_rect_partition =
3202       num_4x4_blocks_wide_lookup[bsize] != num_4x4_blocks_high_lookup[bsize];
3203   int64_t mask_filter = 0;
3204   int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
3205 
3206   vp9_zero(best_mbmode);
3207 
3208   x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3209 
3210   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) filter_cache[i] = INT64_MAX;
3211 
3212   estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3213                            &comp_mode_p);
3214 
3215   for (i = 0; i < REFERENCE_MODES; ++i) best_pred_rd[i] = INT64_MAX;
3216   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3217     best_filter_rd[i] = INT64_MAX;
3218   for (i = 0; i < TX_SIZES; i++) rate_uv_intra[i] = INT_MAX;
3219   for (i = 0; i < MAX_REF_FRAMES; ++i) x->pred_sse[i] = INT_MAX;
3220   for (i = 0; i < MB_MODE_COUNT; ++i) {
3221     for (k = 0; k < MAX_REF_FRAMES; ++k) {
3222       single_inter_filter[i][k] = SWITCHABLE;
3223       single_skippable[i][k] = 0;
3224     }
3225   }
3226 
3227   rd_cost->rate = INT_MAX;
3228 
3229   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3230     x->pred_mv_sad[ref_frame] = INT_MAX;
3231     if ((cpi->ref_frame_flags & flag_list[ref_frame]) &&
3232         !(is_rect_partition && (ctx->skip_ref_frame_mask & (1 << ref_frame)))) {
3233       assert(get_ref_frame_buffer(cpi, ref_frame) != NULL);
3234       setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
3235                          frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
3236     }
3237     frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3238     frame_mv[ZEROMV][ref_frame].as_int = 0;
3239   }
3240 
3241   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3242     if (!(cpi->ref_frame_flags & flag_list[ref_frame])) {
3243       // Skip checking missing references in both single and compound reference
3244       // modes. Note that a mode will be skipped if both reference frames
3245       // are masked out.
3246       ref_frame_skip_mask[0] |= (1 << ref_frame);
3247       ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3248     } else if (sf->reference_masking) {
3249       for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
3250         // Skip fixed mv modes for poor references
3251         if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
3252           mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
3253           break;
3254         }
3255       }
3256     }
3257     // If the segment reference frame feature is enabled....
3258     // then do nothing if the current ref frame is not allowed..
3259     if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3260         get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3261       ref_frame_skip_mask[0] |= (1 << ref_frame);
3262       ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3263     }
3264   }
3265 
3266   // Disable this drop out case if the ref frame
3267   // segment level feature is enabled for this segment. This is to
3268   // prevent the possibility that we end up unable to pick any mode.
3269   if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3270     // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3271     // unless ARNR filtering is enabled in which case we want
3272     // an unfiltered alternative. We allow near/nearest as well
3273     // because they may result in zero-zero MVs but be cheaper.
3274     if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
3275       ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME);
3276       ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3277       mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
3278       if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
3279         mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
3280       if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
3281         mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
3282     }
3283   }
3284 
3285   if (cpi->rc.is_src_frame_alt_ref) {
3286     if (sf->alt_ref_search_fp) {
3287       mode_skip_mask[ALTREF_FRAME] = 0;
3288       ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME);
3289       ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3290     }
3291   }
3292 
3293   if (sf->alt_ref_search_fp)
3294     if (!cm->show_frame && x->pred_mv_sad[GOLDEN_FRAME] < INT_MAX)
3295       if (x->pred_mv_sad[ALTREF_FRAME] > (x->pred_mv_sad[GOLDEN_FRAME] << 1))
3296         mode_skip_mask[ALTREF_FRAME] |= INTER_ALL;
3297 
3298   if (sf->adaptive_mode_search) {
3299     if (cm->show_frame && !cpi->rc.is_src_frame_alt_ref &&
3300         cpi->rc.frames_since_golden >= 3)
3301       if (x->pred_mv_sad[GOLDEN_FRAME] > (x->pred_mv_sad[LAST_FRAME] << 1))
3302         mode_skip_mask[GOLDEN_FRAME] |= INTER_ALL;
3303   }
3304 
3305   if (bsize > sf->max_intra_bsize) {
3306     ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
3307     ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
3308   }
3309 
3310   mode_skip_mask[INTRA_FRAME] |=
3311       ~(sf->intra_y_mode_mask[max_txsize_lookup[bsize]]);
3312 
3313   for (i = 0; i <= LAST_NEW_MV_INDEX; ++i) mode_threshold[i] = 0;
3314 
3315   for (i = LAST_NEW_MV_INDEX + 1; i < MAX_MODES; ++i)
3316     mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
3317 
3318   midx = sf->schedule_mode_search ? mode_skip_start : 0;
3319 
3320   while (midx > 4) {
3321     uint8_t end_pos = 0;
3322     for (i = 5; i < midx; ++i) {
3323       if (mode_threshold[tile_mode_map[i - 1]] >
3324           mode_threshold[tile_mode_map[i]]) {
3325         uint8_t tmp = tile_mode_map[i];
3326         tile_mode_map[i] = tile_mode_map[i - 1];
3327         tile_mode_map[i - 1] = tmp;
3328         end_pos = i;
3329       }
3330     }
3331     midx = end_pos;
3332   }
3333 
3334   memcpy(mode_map, tile_mode_map, sizeof(mode_map));
3335 
3336   for (midx = 0; midx < MAX_MODES; ++midx) {
3337     int mode_index = mode_map[midx];
3338     int mode_excluded = 0;
3339     int64_t this_rd = INT64_MAX;
3340     int disable_skip = 0;
3341     int compmode_cost = 0;
3342     int rate2 = 0, rate_y = 0, rate_uv = 0;
3343     int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3344     int skippable = 0;
3345     int this_skip2 = 0;
3346     int64_t total_sse = INT64_MAX;
3347     int early_term = 0;
3348 
3349     this_mode = vp9_mode_order[mode_index].mode;
3350     ref_frame = vp9_mode_order[mode_index].ref_frame[0];
3351     second_ref_frame = vp9_mode_order[mode_index].ref_frame[1];
3352 
3353     vp9_zero(x->sum_y_eobs);
3354 
3355     if (is_rect_partition) {
3356       if (ctx->skip_ref_frame_mask & (1 << ref_frame)) continue;
3357       if (second_ref_frame > 0 &&
3358           (ctx->skip_ref_frame_mask & (1 << second_ref_frame)))
3359         continue;
3360     }
3361 
3362     // Look at the reference frame of the best mode so far and set the
3363     // skip mask to look at a subset of the remaining modes.
3364     if (midx == mode_skip_start && best_mode_index >= 0) {
3365       switch (best_mbmode.ref_frame[0]) {
3366         case INTRA_FRAME: break;
3367         case LAST_FRAME: ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK; break;
3368         case GOLDEN_FRAME:
3369           ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
3370           break;
3371         case ALTREF_FRAME: ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK; break;
3372         case NONE:
3373         case MAX_REF_FRAMES: assert(0 && "Invalid Reference frame"); break;
3374       }
3375     }
3376 
3377     if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3378         (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
3379       continue;
3380 
3381     if (mode_skip_mask[ref_frame] & (1 << this_mode)) continue;
3382 
3383     // Test best rd so far against threshold for trying this mode.
3384     if (best_mode_skippable && sf->schedule_mode_search)
3385       mode_threshold[mode_index] <<= 1;
3386 
3387     if (best_rd < mode_threshold[mode_index]) continue;
3388 
3389     // This is only used in motion vector unit test.
3390     if (cpi->oxcf.motion_vector_unit_test && ref_frame == INTRA_FRAME) continue;
3391 
3392     if (sf->motion_field_mode_search) {
3393       const int mi_width = VPXMIN(num_8x8_blocks_wide_lookup[bsize],
3394                                   tile_info->mi_col_end - mi_col);
3395       const int mi_height = VPXMIN(num_8x8_blocks_high_lookup[bsize],
3396                                    tile_info->mi_row_end - mi_row);
3397       const int bsl = mi_width_log2_lookup[bsize];
3398       int cb_partition_search_ctrl =
3399           (((mi_row + mi_col) >> bsl) +
3400            get_chessboard_index(cm->current_video_frame)) &
3401           0x1;
3402       MODE_INFO *ref_mi;
3403       int const_motion = 1;
3404       int skip_ref_frame = !cb_partition_search_ctrl;
3405       MV_REFERENCE_FRAME rf = NONE;
3406       int_mv ref_mv;
3407       ref_mv.as_int = INVALID_MV;
3408 
3409       if ((mi_row - 1) >= tile_info->mi_row_start) {
3410         ref_mv = xd->mi[-xd->mi_stride]->mv[0];
3411         rf = xd->mi[-xd->mi_stride]->ref_frame[0];
3412         for (i = 0; i < mi_width; ++i) {
3413           ref_mi = xd->mi[-xd->mi_stride + i];
3414           const_motion &= (ref_mv.as_int == ref_mi->mv[0].as_int) &&
3415                           (ref_frame == ref_mi->ref_frame[0]);
3416           skip_ref_frame &= (rf == ref_mi->ref_frame[0]);
3417         }
3418       }
3419 
3420       if ((mi_col - 1) >= tile_info->mi_col_start) {
3421         if (ref_mv.as_int == INVALID_MV) ref_mv = xd->mi[-1]->mv[0];
3422         if (rf == NONE) rf = xd->mi[-1]->ref_frame[0];
3423         for (i = 0; i < mi_height; ++i) {
3424           ref_mi = xd->mi[i * xd->mi_stride - 1];
3425           const_motion &= (ref_mv.as_int == ref_mi->mv[0].as_int) &&
3426                           (ref_frame == ref_mi->ref_frame[0]);
3427           skip_ref_frame &= (rf == ref_mi->ref_frame[0]);
3428         }
3429       }
3430 
3431       if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV)
3432         if (rf > INTRA_FRAME)
3433           if (ref_frame != rf) continue;
3434 
3435       if (const_motion)
3436         if (this_mode == NEARMV || this_mode == ZEROMV) continue;
3437     }
3438 
3439     comp_pred = second_ref_frame > INTRA_FRAME;
3440     if (comp_pred) {
3441       if (!cpi->allow_comp_inter_inter) continue;
3442 
3443       if (cm->ref_frame_sign_bias[ref_frame] ==
3444           cm->ref_frame_sign_bias[second_ref_frame])
3445         continue;
3446 
3447       // Skip compound inter modes if ARF is not available.
3448       if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) continue;
3449 
3450       // Do not allow compound prediction if the segment level reference frame
3451       // feature is in use as in this case there can only be one reference.
3452       if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) continue;
3453 
3454       if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3455           best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
3456         continue;
3457 
3458       mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3459     } else {
3460       if (ref_frame != INTRA_FRAME)
3461         mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3462     }
3463 
3464     if (ref_frame == INTRA_FRAME) {
3465       if (sf->adaptive_mode_search)
3466         if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
3467           continue;
3468 
3469       if (this_mode != DC_PRED) {
3470         // Disable intra modes other than DC_PRED for blocks with low variance
3471         // Threshold for intra skipping based on source variance
3472         // TODO(debargha): Specialize the threshold for super block sizes
3473         const unsigned int skip_intra_var_thresh = 64;
3474         if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
3475             x->source_variance < skip_intra_var_thresh)
3476           continue;
3477         // Only search the oblique modes if the best so far is
3478         // one of the neighboring directional modes
3479         if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
3480             (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
3481           if (best_mode_index >= 0 && best_mbmode.ref_frame[0] > INTRA_FRAME)
3482             continue;
3483         }
3484         if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
3485           if (conditional_skipintra(this_mode, best_intra_mode)) continue;
3486         }
3487       }
3488     } else {
3489       const MV_REFERENCE_FRAME ref_frames[2] = { ref_frame, second_ref_frame };
3490       if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv, this_mode,
3491                               ref_frames))
3492         continue;
3493     }
3494 
3495     mi->mode = this_mode;
3496     mi->uv_mode = DC_PRED;
3497     mi->ref_frame[0] = ref_frame;
3498     mi->ref_frame[1] = second_ref_frame;
3499     // Evaluate all sub-pel filters irrespective of whether we can use
3500     // them for this frame.
3501     mi->interp_filter =
3502         cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
3503     mi->mv[0].as_int = mi->mv[1].as_int = 0;
3504 
3505     x->skip = 0;
3506     set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3507 
3508     // Select prediction reference frames.
3509     for (i = 0; i < MAX_MB_PLANE; i++) {
3510       xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3511       if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3512     }
3513 
3514     if (ref_frame == INTRA_FRAME) {
3515       TX_SIZE uv_tx;
3516       struct macroblockd_plane *const pd = &xd->plane[1];
3517       memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
3518       super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable, NULL, bsize,
3519                       best_rd);
3520       if (rate_y == INT_MAX) continue;
3521 
3522       uv_tx = uv_txsize_lookup[bsize][mi->tx_size][pd->subsampling_x]
3523                               [pd->subsampling_y];
3524       if (rate_uv_intra[uv_tx] == INT_MAX) {
3525         choose_intra_uv_mode(cpi, x, ctx, bsize, uv_tx, &rate_uv_intra[uv_tx],
3526                              &rate_uv_tokenonly[uv_tx], &dist_uv[uv_tx],
3527                              &skip_uv[uv_tx], &mode_uv[uv_tx]);
3528       }
3529 
3530       rate_uv = rate_uv_tokenonly[uv_tx];
3531       distortion_uv = dist_uv[uv_tx];
3532       skippable = skippable && skip_uv[uv_tx];
3533       mi->uv_mode = mode_uv[uv_tx];
3534 
3535       rate2 = rate_y + cpi->mbmode_cost[mi->mode] + rate_uv_intra[uv_tx];
3536       if (this_mode != DC_PRED && this_mode != TM_PRED)
3537         rate2 += intra_cost_penalty;
3538       distortion2 = distortion_y + distortion_uv;
3539     } else {
3540       this_rd = handle_inter_mode(
3541           cpi, x, bsize, &rate2, &distortion2, &skippable, &rate_y, &rate_uv,
3542           &disable_skip, frame_mv, mi_row, mi_col, single_newmv,
3543           single_inter_filter, single_skippable, &total_sse, best_rd,
3544           &mask_filter, filter_cache);
3545       if (this_rd == INT64_MAX) continue;
3546 
3547       compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
3548 
3549       if (cm->reference_mode == REFERENCE_MODE_SELECT) rate2 += compmode_cost;
3550     }
3551 
3552     // Estimate the reference frame signaling cost and add it
3553     // to the rolling cost variable.
3554     if (comp_pred) {
3555       rate2 += ref_costs_comp[ref_frame];
3556     } else {
3557       rate2 += ref_costs_single[ref_frame];
3558     }
3559 
3560     if (!disable_skip) {
3561       const vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
3562       const int skip_cost0 = vp9_cost_bit(skip_prob, 0);
3563       const int skip_cost1 = vp9_cost_bit(skip_prob, 1);
3564 
3565       if (skippable) {
3566         // Back out the coefficient coding costs
3567         rate2 -= (rate_y + rate_uv);
3568 
3569         // Cost the skip mb case
3570         rate2 += skip_cost1;
3571       } else if (ref_frame != INTRA_FRAME && !xd->lossless) {
3572         if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv + skip_cost0,
3573                    distortion2) <
3574             RDCOST(x->rdmult, x->rddiv, skip_cost1, total_sse)) {
3575           // Add in the cost of the no skip flag.
3576           rate2 += skip_cost0;
3577         } else {
3578           // FIXME(rbultje) make this work for splitmv also
3579           assert(total_sse >= 0);
3580 
3581           rate2 += skip_cost1;
3582           distortion2 = total_sse;
3583           rate2 -= (rate_y + rate_uv);
3584           this_skip2 = 1;
3585         }
3586       } else {
3587         // Add in the cost of the no skip flag.
3588         rate2 += skip_cost0;
3589       }
3590 
3591       // Calculate the final RD estimate for this mode.
3592       this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3593     }
3594 
3595     // Apply an adjustment to the rd value based on the similarity of the
3596     // source variance and reconstructed variance.
3597     rd_variance_adjustment(cpi, x, bsize, &this_rd, ref_frame,
3598                            x->source_variance);
3599 
3600     if (ref_frame == INTRA_FRAME) {
3601       // Keep record of best intra rd
3602       if (this_rd < best_intra_rd) {
3603         best_intra_rd = this_rd;
3604         best_intra_mode = mi->mode;
3605       }
3606     }
3607 
3608     if (!disable_skip && ref_frame == INTRA_FRAME) {
3609       for (i = 0; i < REFERENCE_MODES; ++i)
3610         best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
3611       for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3612         best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
3613     }
3614 
3615     // Did this mode help.. i.e. is it the new best mode
3616     if (this_rd < best_rd || x->skip) {
3617       int max_plane = MAX_MB_PLANE;
3618       if (!mode_excluded) {
3619         // Note index of best mode so far
3620         best_mode_index = mode_index;
3621 
3622         if (ref_frame == INTRA_FRAME) {
3623           /* required for left and above block mv */
3624           mi->mv[0].as_int = 0;
3625           max_plane = 1;
3626           // Initialize interp_filter here so we do not have to check for
3627           // inter block modes in get_pred_context_switchable_interp()
3628           mi->interp_filter = SWITCHABLE_FILTERS;
3629         } else {
3630           best_pred_sse = x->pred_sse[ref_frame];
3631         }
3632 
3633         rd_cost->rate = rate2;
3634         rd_cost->dist = distortion2;
3635         rd_cost->rdcost = this_rd;
3636         best_rd = this_rd;
3637         best_mbmode = *mi;
3638         best_skip2 = this_skip2;
3639         best_mode_skippable = skippable;
3640 
3641         if (!x->select_tx_size) swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
3642         memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mi->tx_size],
3643                sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
3644         ctx->sum_y_eobs = x->sum_y_eobs[mi->tx_size];
3645 
3646         // TODO(debargha): enhance this test with a better distortion prediction
3647         // based on qp, activity mask and history
3648         if ((mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
3649             (mode_index > MIN_EARLY_TERM_INDEX)) {
3650           int qstep = xd->plane[0].dequant[1];
3651           // TODO(debargha): Enhance this by specializing for each mode_index
3652           int scale = 4;
3653 #if CONFIG_VP9_HIGHBITDEPTH
3654           if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
3655             qstep >>= (xd->bd - 8);
3656           }
3657 #endif  // CONFIG_VP9_HIGHBITDEPTH
3658           if (x->source_variance < UINT_MAX) {
3659             const int var_adjust = (x->source_variance < 16);
3660             scale -= var_adjust;
3661           }
3662           if (ref_frame > INTRA_FRAME && distortion2 * scale < qstep * qstep) {
3663             early_term = 1;
3664           }
3665         }
3666       }
3667     }
3668 
3669     /* keep record of best compound/single-only prediction */
3670     if (!disable_skip && ref_frame != INTRA_FRAME) {
3671       int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
3672 
3673       if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3674         single_rate = rate2 - compmode_cost;
3675         hybrid_rate = rate2;
3676       } else {
3677         single_rate = rate2;
3678         hybrid_rate = rate2 + compmode_cost;
3679       }
3680 
3681       single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
3682       hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
3683 
3684       if (!comp_pred) {
3685         if (single_rd < best_pred_rd[SINGLE_REFERENCE])
3686           best_pred_rd[SINGLE_REFERENCE] = single_rd;
3687       } else {
3688         if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
3689           best_pred_rd[COMPOUND_REFERENCE] = single_rd;
3690       }
3691       if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
3692         best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
3693 
3694       /* keep record of best filter type */
3695       if (!mode_excluded && cm->interp_filter != BILINEAR) {
3696         int64_t ref =
3697             filter_cache[cm->interp_filter == SWITCHABLE ? SWITCHABLE_FILTERS
3698                                                          : cm->interp_filter];
3699 
3700         for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3701           int64_t adj_rd;
3702           if (ref == INT64_MAX)
3703             adj_rd = 0;
3704           else if (filter_cache[i] == INT64_MAX)
3705             // when early termination is triggered, the encoder does not have
3706             // access to the rate-distortion cost. it only knows that the cost
3707             // should be above the maximum valid value. hence it takes the known
3708             // maximum plus an arbitrary constant as the rate-distortion cost.
3709             adj_rd = mask_filter - ref + 10;
3710           else
3711             adj_rd = filter_cache[i] - ref;
3712 
3713           adj_rd += this_rd;
3714           best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
3715         }
3716       }
3717     }
3718 
3719     if (early_term) break;
3720 
3721     if (x->skip && !comp_pred) break;
3722   }
3723 
3724   // The inter modes' rate costs are not calculated precisely in some cases.
3725   // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and
3726   // ZEROMV. Here, checks are added for those cases, and the mode decisions
3727   // are corrected.
3728   if (best_mbmode.mode == NEWMV) {
3729     const MV_REFERENCE_FRAME refs[2] = { best_mbmode.ref_frame[0],
3730                                          best_mbmode.ref_frame[1] };
3731     int comp_pred_mode = refs[1] > INTRA_FRAME;
3732 
3733     if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3734         ((comp_pred_mode &&
3735           frame_mv[NEARESTMV][refs[1]].as_int == best_mbmode.mv[1].as_int) ||
3736          !comp_pred_mode))
3737       best_mbmode.mode = NEARESTMV;
3738     else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3739              ((comp_pred_mode &&
3740                frame_mv[NEARMV][refs[1]].as_int == best_mbmode.mv[1].as_int) ||
3741               !comp_pred_mode))
3742       best_mbmode.mode = NEARMV;
3743     else if (best_mbmode.mv[0].as_int == 0 &&
3744              ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) ||
3745               !comp_pred_mode))
3746       best_mbmode.mode = ZEROMV;
3747   }
3748 
3749   if (best_mode_index < 0 || best_rd >= best_rd_so_far) {
3750 // If adaptive interp filter is enabled, then the current leaf node of 8x8
3751 // data is needed for sub8x8. Hence preserve the context.
3752 #if CONFIG_CONSISTENT_RECODE
3753     if (bsize == BLOCK_8X8) ctx->mic = *xd->mi[0];
3754 #else
3755     if (cpi->row_mt && bsize == BLOCK_8X8) ctx->mic = *xd->mi[0];
3756 #endif
3757     rd_cost->rate = INT_MAX;
3758     rd_cost->rdcost = INT64_MAX;
3759     return;
3760   }
3761 
3762   // If we used an estimate for the uv intra rd in the loop above...
3763   if (sf->use_uv_intra_rd_estimate) {
3764     // Do Intra UV best rd mode selection if best mode choice above was intra.
3765     if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
3766       TX_SIZE uv_tx_size;
3767       *mi = best_mbmode;
3768       uv_tx_size = get_uv_tx_size(mi, &xd->plane[1]);
3769       rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
3770                               &rate_uv_tokenonly[uv_tx_size],
3771                               &dist_uv[uv_tx_size], &skip_uv[uv_tx_size],
3772                               bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
3773                               uv_tx_size);
3774     }
3775   }
3776 
3777   assert((cm->interp_filter == SWITCHABLE) ||
3778          (cm->interp_filter == best_mbmode.interp_filter) ||
3779          !is_inter_block(&best_mbmode));
3780 
3781   if (!cpi->rc.is_src_frame_alt_ref)
3782     vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3783                               sf->adaptive_rd_thresh, bsize, best_mode_index);
3784 
3785   // macroblock modes
3786   *mi = best_mbmode;
3787   x->skip |= best_skip2;
3788 
3789   for (i = 0; i < REFERENCE_MODES; ++i) {
3790     if (best_pred_rd[i] == INT64_MAX)
3791       best_pred_diff[i] = INT_MIN;
3792     else
3793       best_pred_diff[i] = best_rd - best_pred_rd[i];
3794   }
3795 
3796   if (!x->skip) {
3797     for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3798       if (best_filter_rd[i] == INT64_MAX)
3799         best_filter_diff[i] = 0;
3800       else
3801         best_filter_diff[i] = best_rd - best_filter_rd[i];
3802     }
3803     if (cm->interp_filter == SWITCHABLE)
3804       assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
3805   } else {
3806     vp9_zero(best_filter_diff);
3807   }
3808 
3809   // TODO(yunqingwang): Moving this line in front of the above best_filter_diff
3810   // updating code causes PSNR loss. Need to figure out the confliction.
3811   x->skip |= best_mode_skippable;
3812 
3813   if (!x->skip && !x->select_tx_size) {
3814     int has_high_freq_coeff = 0;
3815     int plane;
3816     int max_plane = is_inter_block(xd->mi[0]) ? MAX_MB_PLANE : 1;
3817     for (plane = 0; plane < max_plane; ++plane) {
3818       x->plane[plane].eobs = ctx->eobs_pbuf[plane][1];
3819       has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
3820     }
3821 
3822     for (plane = max_plane; plane < MAX_MB_PLANE; ++plane) {
3823       x->plane[plane].eobs = ctx->eobs_pbuf[plane][2];
3824       has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
3825     }
3826 
3827     best_mode_skippable |= !has_high_freq_coeff;
3828   }
3829 
3830   assert(best_mode_index >= 0);
3831 
3832   store_coding_context(x, ctx, best_mode_index, best_pred_diff,
3833                        best_filter_diff, best_mode_skippable);
3834 }
3835 
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)3836 void vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi, TileDataEnc *tile_data,
3837                                         MACROBLOCK *x, RD_COST *rd_cost,
3838                                         BLOCK_SIZE bsize,
3839                                         PICK_MODE_CONTEXT *ctx,
3840                                         int64_t best_rd_so_far) {
3841   VP9_COMMON *const cm = &cpi->common;
3842   MACROBLOCKD *const xd = &x->e_mbd;
3843   MODE_INFO *const mi = xd->mi[0];
3844   unsigned char segment_id = mi->segment_id;
3845   const int comp_pred = 0;
3846   int i;
3847   int64_t best_pred_diff[REFERENCE_MODES];
3848   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3849   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3850   vpx_prob comp_mode_p;
3851   INTERP_FILTER best_filter = SWITCHABLE;
3852   int64_t this_rd = INT64_MAX;
3853   int rate2 = 0;
3854   const int64_t distortion2 = 0;
3855 
3856   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3857 
3858   estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3859                            &comp_mode_p);
3860 
3861   for (i = 0; i < MAX_REF_FRAMES; ++i) x->pred_sse[i] = INT_MAX;
3862   for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i) x->pred_mv_sad[i] = INT_MAX;
3863 
3864   rd_cost->rate = INT_MAX;
3865 
3866   assert(segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
3867 
3868   mi->mode = ZEROMV;
3869   mi->uv_mode = DC_PRED;
3870   mi->ref_frame[0] = LAST_FRAME;
3871   mi->ref_frame[1] = NONE;
3872   mi->mv[0].as_int = 0;
3873   x->skip = 1;
3874 
3875   ctx->sum_y_eobs = 0;
3876 
3877   if (cm->interp_filter != BILINEAR) {
3878     best_filter = EIGHTTAP;
3879     if (cm->interp_filter == SWITCHABLE &&
3880         x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
3881       int rs;
3882       int best_rs = INT_MAX;
3883       for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
3884         mi->interp_filter = i;
3885         rs = vp9_get_switchable_rate(cpi, xd);
3886         if (rs < best_rs) {
3887           best_rs = rs;
3888           best_filter = mi->interp_filter;
3889         }
3890       }
3891     }
3892   }
3893   // Set the appropriate filter
3894   if (cm->interp_filter == SWITCHABLE) {
3895     mi->interp_filter = best_filter;
3896     rate2 += vp9_get_switchable_rate(cpi, xd);
3897   } else {
3898     mi->interp_filter = cm->interp_filter;
3899   }
3900 
3901   if (cm->reference_mode == REFERENCE_MODE_SELECT)
3902     rate2 += vp9_cost_bit(comp_mode_p, comp_pred);
3903 
3904   // Estimate the reference frame signaling cost and add it
3905   // to the rolling cost variable.
3906   rate2 += ref_costs_single[LAST_FRAME];
3907   this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3908 
3909   rd_cost->rate = rate2;
3910   rd_cost->dist = distortion2;
3911   rd_cost->rdcost = this_rd;
3912 
3913   if (this_rd >= best_rd_so_far) {
3914     rd_cost->rate = INT_MAX;
3915     rd_cost->rdcost = INT64_MAX;
3916     return;
3917   }
3918 
3919   assert((cm->interp_filter == SWITCHABLE) ||
3920          (cm->interp_filter == mi->interp_filter));
3921 
3922   vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3923                             cpi->sf.adaptive_rd_thresh, bsize, THR_ZEROMV);
3924 
3925   vp9_zero(best_pred_diff);
3926   vp9_zero(best_filter_diff);
3927 
3928   if (!x->select_tx_size) swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE);
3929   store_coding_context(x, ctx, THR_ZEROMV, best_pred_diff, best_filter_diff, 0);
3930 }
3931 
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)3932 void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, TileDataEnc *tile_data,
3933                                    MACROBLOCK *x, int mi_row, int mi_col,
3934                                    RD_COST *rd_cost, BLOCK_SIZE bsize,
3935                                    PICK_MODE_CONTEXT *ctx,
3936                                    int64_t best_rd_so_far) {
3937   VP9_COMMON *const cm = &cpi->common;
3938   RD_OPT *const rd_opt = &cpi->rd;
3939   SPEED_FEATURES *const sf = &cpi->sf;
3940   MACROBLOCKD *const xd = &x->e_mbd;
3941   MODE_INFO *const mi = xd->mi[0];
3942   const struct segmentation *const seg = &cm->seg;
3943   MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3944   unsigned char segment_id = mi->segment_id;
3945   int comp_pred, i;
3946   int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3947   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3948   static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
3949                                     VP9_ALT_FLAG };
3950   int64_t best_rd = best_rd_so_far;
3951   int64_t best_yrd = best_rd_so_far;  // FIXME(rbultje) more precise
3952   int64_t best_pred_diff[REFERENCE_MODES];
3953   int64_t best_pred_rd[REFERENCE_MODES];
3954   int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3955   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3956   MODE_INFO best_mbmode;
3957   int ref_index, best_ref_index = 0;
3958   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3959   vpx_prob comp_mode_p;
3960   INTERP_FILTER tmp_best_filter = SWITCHABLE;
3961   int rate_uv_intra, rate_uv_tokenonly;
3962   int64_t dist_uv;
3963   int skip_uv;
3964   PREDICTION_MODE mode_uv = DC_PRED;
3965   const int intra_cost_penalty =
3966       vp9_get_intra_cost_penalty(cpi, bsize, cm->base_qindex, cm->y_dc_delta_q);
3967   int_mv seg_mvs[4][MAX_REF_FRAMES];
3968   b_mode_info best_bmodes[4];
3969   int best_skip2 = 0;
3970   int ref_frame_skip_mask[2] = { 0 };
3971   int64_t mask_filter = 0;
3972   int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
3973   int internal_active_edge =
3974       vp9_active_edge_sb(cpi, mi_row, mi_col) && vp9_internal_image_edge(cpi);
3975   const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
3976 
3977   x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3978   memset(x->zcoeff_blk[TX_4X4], 0, 4);
3979   vp9_zero(best_mbmode);
3980 
3981   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) filter_cache[i] = INT64_MAX;
3982 
3983   for (i = 0; i < 4; i++) {
3984     int j;
3985     for (j = 0; j < MAX_REF_FRAMES; j++) seg_mvs[i][j].as_int = INVALID_MV;
3986   }
3987 
3988   estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3989                            &comp_mode_p);
3990 
3991   for (i = 0; i < REFERENCE_MODES; ++i) best_pred_rd[i] = INT64_MAX;
3992   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3993     best_filter_rd[i] = INT64_MAX;
3994   rate_uv_intra = INT_MAX;
3995 
3996   rd_cost->rate = INT_MAX;
3997 
3998   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
3999     if (cpi->ref_frame_flags & flag_list[ref_frame]) {
4000       setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
4001                          frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
4002     } else {
4003       ref_frame_skip_mask[0] |= (1 << ref_frame);
4004       ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
4005     }
4006     frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
4007     frame_mv[ZEROMV][ref_frame].as_int = 0;
4008   }
4009 
4010   for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
4011     int mode_excluded = 0;
4012     int64_t this_rd = INT64_MAX;
4013     int disable_skip = 0;
4014     int compmode_cost = 0;
4015     int rate2 = 0, rate_y = 0, rate_uv = 0;
4016     int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
4017     int skippable = 0;
4018     int i;
4019     int this_skip2 = 0;
4020     int64_t total_sse = INT_MAX;
4021     int early_term = 0;
4022     struct buf_2d backup_yv12[2][MAX_MB_PLANE];
4023 
4024     ref_frame = vp9_ref_order[ref_index].ref_frame[0];
4025     second_ref_frame = vp9_ref_order[ref_index].ref_frame[1];
4026 
4027     vp9_zero(x->sum_y_eobs);
4028 
4029 #if CONFIG_BETTER_HW_COMPATIBILITY
4030     // forbid 8X4 and 4X8 partitions if any reference frame is scaled.
4031     if (bsize == BLOCK_8X4 || bsize == BLOCK_4X8) {
4032       int ref_scaled = ref_frame > INTRA_FRAME &&
4033                        vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf);
4034       if (second_ref_frame > INTRA_FRAME)
4035         ref_scaled += vp9_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf);
4036       if (ref_scaled) continue;
4037     }
4038 #endif
4039     // Look at the reference frame of the best mode so far and set the
4040     // skip mask to look at a subset of the remaining modes.
4041     if (ref_index > 2 && sf->mode_skip_start < MAX_MODES) {
4042       if (ref_index == 3) {
4043         switch (best_mbmode.ref_frame[0]) {
4044           case INTRA_FRAME: break;
4045           case LAST_FRAME:
4046             ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME);
4047             ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
4048             break;
4049           case GOLDEN_FRAME:
4050             ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME);
4051             ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
4052             break;
4053           case ALTREF_FRAME:
4054             ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME);
4055             break;
4056           case NONE:
4057           case MAX_REF_FRAMES: assert(0 && "Invalid Reference frame"); break;
4058         }
4059       }
4060     }
4061 
4062     if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
4063         (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
4064       continue;
4065 
4066     // Test best rd so far against threshold for trying this mode.
4067     if (!internal_active_edge &&
4068         rd_less_than_thresh(best_rd,
4069                             rd_opt->threshes[segment_id][bsize][ref_index],
4070                             &rd_thresh_freq_fact[ref_index]))
4071       continue;
4072 
4073     // This is only used in motion vector unit test.
4074     if (cpi->oxcf.motion_vector_unit_test && ref_frame == INTRA_FRAME) continue;
4075 
4076     comp_pred = second_ref_frame > INTRA_FRAME;
4077     if (comp_pred) {
4078       if (!cpi->allow_comp_inter_inter) continue;
4079 
4080       if (cm->ref_frame_sign_bias[ref_frame] ==
4081           cm->ref_frame_sign_bias[second_ref_frame])
4082         continue;
4083 
4084       if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) continue;
4085       // Do not allow compound prediction if the segment level reference frame
4086       // feature is in use as in this case there can only be one reference.
4087       if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) continue;
4088 
4089       if ((sf->mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
4090           best_mbmode.ref_frame[0] == INTRA_FRAME)
4091         continue;
4092     }
4093 
4094     if (comp_pred)
4095       mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
4096     else if (ref_frame != INTRA_FRAME)
4097       mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
4098 
4099     // If the segment reference frame feature is enabled....
4100     // then do nothing if the current ref frame is not allowed..
4101     if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
4102         get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
4103       continue;
4104       // Disable this drop out case if the ref frame
4105       // segment level feature is enabled for this segment. This is to
4106       // prevent the possibility that we end up unable to pick any mode.
4107     } else if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
4108       // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
4109       // unless ARNR filtering is enabled in which case we want
4110       // an unfiltered alternative. We allow near/nearest as well
4111       // because they may result in zero-zero MVs but be cheaper.
4112       if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
4113         continue;
4114     }
4115 
4116     mi->tx_size = TX_4X4;
4117     mi->uv_mode = DC_PRED;
4118     mi->ref_frame[0] = ref_frame;
4119     mi->ref_frame[1] = second_ref_frame;
4120     // Evaluate all sub-pel filters irrespective of whether we can use
4121     // them for this frame.
4122     mi->interp_filter =
4123         cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
4124     x->skip = 0;
4125     set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
4126 
4127     // Select prediction reference frames.
4128     for (i = 0; i < MAX_MB_PLANE; i++) {
4129       xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
4130       if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
4131     }
4132 
4133     if (ref_frame == INTRA_FRAME) {
4134       int rate;
4135       if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y, &distortion_y,
4136                                        best_rd) >= best_rd)
4137         continue;
4138       rate2 += rate;
4139       rate2 += intra_cost_penalty;
4140       distortion2 += distortion_y;
4141 
4142       if (rate_uv_intra == INT_MAX) {
4143         choose_intra_uv_mode(cpi, x, ctx, bsize, TX_4X4, &rate_uv_intra,
4144                              &rate_uv_tokenonly, &dist_uv, &skip_uv, &mode_uv);
4145       }
4146       rate2 += rate_uv_intra;
4147       rate_uv = rate_uv_tokenonly;
4148       distortion2 += dist_uv;
4149       distortion_uv = dist_uv;
4150       mi->uv_mode = mode_uv;
4151     } else {
4152       int rate;
4153       int64_t distortion;
4154       int64_t this_rd_thresh;
4155       int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
4156       int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
4157       int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
4158       int tmp_best_skippable = 0;
4159       int switchable_filter_index;
4160       int_mv *second_ref =
4161           comp_pred ? &x->mbmi_ext->ref_mvs[second_ref_frame][0] : NULL;
4162       b_mode_info tmp_best_bmodes[16];
4163       MODE_INFO tmp_best_mbmode;
4164       BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
4165       int pred_exists = 0;
4166       int uv_skippable;
4167 
4168       YV12_BUFFER_CONFIG *scaled_ref_frame[2] = { NULL, NULL };
4169       int ref;
4170 
4171       for (ref = 0; ref < 2; ++ref) {
4172         scaled_ref_frame[ref] =
4173             mi->ref_frame[ref] > INTRA_FRAME
4174                 ? vp9_get_scaled_ref_frame(cpi, mi->ref_frame[ref])
4175                 : NULL;
4176 
4177         if (scaled_ref_frame[ref]) {
4178           int i;
4179           // Swap out the reference frame for a version that's been scaled to
4180           // match the resolution of the current frame, allowing the existing
4181           // motion search code to be used without additional modifications.
4182           for (i = 0; i < MAX_MB_PLANE; i++)
4183             backup_yv12[ref][i] = xd->plane[i].pre[ref];
4184           vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
4185                                NULL);
4186         }
4187       }
4188 
4189       this_rd_thresh = (ref_frame == LAST_FRAME)
4190                            ? rd_opt->threshes[segment_id][bsize][THR_LAST]
4191                            : rd_opt->threshes[segment_id][bsize][THR_ALTR];
4192       this_rd_thresh = (ref_frame == GOLDEN_FRAME)
4193                            ? rd_opt->threshes[segment_id][bsize][THR_GOLD]
4194                            : this_rd_thresh;
4195       for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
4196         filter_cache[i] = INT64_MAX;
4197 
4198       if (cm->interp_filter != BILINEAR) {
4199         tmp_best_filter = EIGHTTAP;
4200         if (x->source_variance < sf->disable_filter_search_var_thresh) {
4201           tmp_best_filter = EIGHTTAP;
4202         } else if (sf->adaptive_pred_interp_filter == 1 &&
4203                    ctx->pred_interp_filter < SWITCHABLE) {
4204           tmp_best_filter = ctx->pred_interp_filter;
4205         } else if (sf->adaptive_pred_interp_filter == 2) {
4206           tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE
4207                                 ? ctx->pred_interp_filter
4208                                 : 0;
4209         } else {
4210           for (switchable_filter_index = 0;
4211                switchable_filter_index < SWITCHABLE_FILTERS;
4212                ++switchable_filter_index) {
4213             int newbest, rs;
4214             int64_t rs_rd;
4215             MB_MODE_INFO_EXT *mbmi_ext = x->mbmi_ext;
4216             mi->interp_filter = switchable_filter_index;
4217             tmp_rd = rd_pick_best_sub8x8_mode(
4218                 cpi, x, &mbmi_ext->ref_mvs[ref_frame][0], second_ref, best_yrd,
4219                 &rate, &rate_y, &distortion, &skippable, &total_sse,
4220                 (int)this_rd_thresh, seg_mvs, bsi, switchable_filter_index,
4221                 mi_row, mi_col);
4222 
4223             if (tmp_rd == INT64_MAX) continue;
4224             rs = vp9_get_switchable_rate(cpi, xd);
4225             rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
4226             filter_cache[switchable_filter_index] = tmp_rd;
4227             filter_cache[SWITCHABLE_FILTERS] =
4228                 VPXMIN(filter_cache[SWITCHABLE_FILTERS], tmp_rd + rs_rd);
4229             if (cm->interp_filter == SWITCHABLE) tmp_rd += rs_rd;
4230 
4231             mask_filter = VPXMAX(mask_filter, tmp_rd);
4232 
4233             newbest = (tmp_rd < tmp_best_rd);
4234             if (newbest) {
4235               tmp_best_filter = mi->interp_filter;
4236               tmp_best_rd = tmp_rd;
4237             }
4238             if ((newbest && cm->interp_filter == SWITCHABLE) ||
4239                 (mi->interp_filter == cm->interp_filter &&
4240                  cm->interp_filter != SWITCHABLE)) {
4241               tmp_best_rdu = tmp_rd;
4242               tmp_best_rate = rate;
4243               tmp_best_ratey = rate_y;
4244               tmp_best_distortion = distortion;
4245               tmp_best_sse = total_sse;
4246               tmp_best_skippable = skippable;
4247               tmp_best_mbmode = *mi;
4248               for (i = 0; i < 4; i++) {
4249                 tmp_best_bmodes[i] = xd->mi[0]->bmi[i];
4250                 x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
4251                 x->sum_y_eobs[TX_4X4] += x->plane[0].eobs[i];
4252               }
4253               pred_exists = 1;
4254               if (switchable_filter_index == 0 && sf->use_rd_breakout &&
4255                   best_rd < INT64_MAX) {
4256                 if (tmp_best_rdu / 2 > best_rd) {
4257                   // skip searching the other filters if the first is
4258                   // already substantially larger than the best so far
4259                   tmp_best_filter = mi->interp_filter;
4260                   tmp_best_rdu = INT64_MAX;
4261                   break;
4262                 }
4263               }
4264             }
4265           }  // switchable_filter_index loop
4266         }
4267       }
4268 
4269       if (tmp_best_rdu == INT64_MAX && pred_exists) continue;
4270 
4271       mi->interp_filter = (cm->interp_filter == SWITCHABLE ? tmp_best_filter
4272                                                            : cm->interp_filter);
4273       if (!pred_exists) {
4274         // Handles the special case when a filter that is not in the
4275         // switchable list (bilinear, 6-tap) is indicated at the frame level
4276         tmp_rd = rd_pick_best_sub8x8_mode(
4277             cpi, x, &x->mbmi_ext->ref_mvs[ref_frame][0], second_ref, best_yrd,
4278             &rate, &rate_y, &distortion, &skippable, &total_sse,
4279             (int)this_rd_thresh, seg_mvs, bsi, 0, mi_row, mi_col);
4280         if (tmp_rd == INT64_MAX) continue;
4281       } else {
4282         total_sse = tmp_best_sse;
4283         rate = tmp_best_rate;
4284         rate_y = tmp_best_ratey;
4285         distortion = tmp_best_distortion;
4286         skippable = tmp_best_skippable;
4287         *mi = tmp_best_mbmode;
4288         for (i = 0; i < 4; i++) xd->mi[0]->bmi[i] = tmp_best_bmodes[i];
4289       }
4290 
4291       rate2 += rate;
4292       distortion2 += distortion;
4293 
4294       if (cm->interp_filter == SWITCHABLE)
4295         rate2 += vp9_get_switchable_rate(cpi, xd);
4296 
4297       if (!mode_excluded)
4298         mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
4299                                   : cm->reference_mode == COMPOUND_REFERENCE;
4300 
4301       compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
4302 
4303       tmp_best_rdu =
4304           best_rd - VPXMIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
4305                            RDCOST(x->rdmult, x->rddiv, 0, total_sse));
4306 
4307       if (tmp_best_rdu > 0) {
4308         // If even the 'Y' rd value of split is higher than best so far
4309         // then dont bother looking at UV
4310         vp9_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col, BLOCK_8X8);
4311         memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
4312         if (!super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
4313                               &uv_sse, BLOCK_8X8, tmp_best_rdu)) {
4314           for (ref = 0; ref < 2; ++ref) {
4315             if (scaled_ref_frame[ref]) {
4316               int i;
4317               for (i = 0; i < MAX_MB_PLANE; ++i)
4318                 xd->plane[i].pre[ref] = backup_yv12[ref][i];
4319             }
4320           }
4321           continue;
4322         }
4323 
4324         rate2 += rate_uv;
4325         distortion2 += distortion_uv;
4326         skippable = skippable && uv_skippable;
4327         total_sse += uv_sse;
4328       }
4329 
4330       for (ref = 0; ref < 2; ++ref) {
4331         if (scaled_ref_frame[ref]) {
4332           // Restore the prediction frame pointers to their unscaled versions.
4333           int i;
4334           for (i = 0; i < MAX_MB_PLANE; ++i)
4335             xd->plane[i].pre[ref] = backup_yv12[ref][i];
4336         }
4337       }
4338     }
4339 
4340     if (cm->reference_mode == REFERENCE_MODE_SELECT) rate2 += compmode_cost;
4341 
4342     // Estimate the reference frame signaling cost and add it
4343     // to the rolling cost variable.
4344     if (second_ref_frame > INTRA_FRAME) {
4345       rate2 += ref_costs_comp[ref_frame];
4346     } else {
4347       rate2 += ref_costs_single[ref_frame];
4348     }
4349 
4350     if (!disable_skip) {
4351       const vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
4352       const int skip_cost0 = vp9_cost_bit(skip_prob, 0);
4353       const int skip_cost1 = vp9_cost_bit(skip_prob, 1);
4354 
4355       // Skip is never coded at the segment level for sub8x8 blocks and instead
4356       // always coded in the bitstream at the mode info level.
4357       if (ref_frame != INTRA_FRAME && !xd->lossless) {
4358         if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv + skip_cost0,
4359                    distortion2) <
4360             RDCOST(x->rdmult, x->rddiv, skip_cost1, total_sse)) {
4361           // Add in the cost of the no skip flag.
4362           rate2 += skip_cost0;
4363         } else {
4364           // FIXME(rbultje) make this work for splitmv also
4365           rate2 += skip_cost1;
4366           distortion2 = total_sse;
4367           assert(total_sse >= 0);
4368           rate2 -= (rate_y + rate_uv);
4369           rate_y = 0;
4370           rate_uv = 0;
4371           this_skip2 = 1;
4372         }
4373       } else {
4374         // Add in the cost of the no skip flag.
4375         rate2 += skip_cost0;
4376       }
4377 
4378       // Calculate the final RD estimate for this mode.
4379       this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
4380     }
4381 
4382     if (!disable_skip && ref_frame == INTRA_FRAME) {
4383       for (i = 0; i < REFERENCE_MODES; ++i)
4384         best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
4385       for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
4386         best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
4387     }
4388 
4389     // Did this mode help.. i.e. is it the new best mode
4390     if (this_rd < best_rd || x->skip) {
4391       if (!mode_excluded) {
4392         int max_plane = MAX_MB_PLANE;
4393         // Note index of best mode so far
4394         best_ref_index = ref_index;
4395 
4396         if (ref_frame == INTRA_FRAME) {
4397           /* required for left and above block mv */
4398           mi->mv[0].as_int = 0;
4399           max_plane = 1;
4400           // Initialize interp_filter here so we do not have to check for
4401           // inter block modes in get_pred_context_switchable_interp()
4402           mi->interp_filter = SWITCHABLE_FILTERS;
4403         }
4404 
4405         rd_cost->rate = rate2;
4406         rd_cost->dist = distortion2;
4407         rd_cost->rdcost = this_rd;
4408         best_rd = this_rd;
4409         best_yrd =
4410             best_rd - RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
4411         best_mbmode = *mi;
4412         best_skip2 = this_skip2;
4413         if (!x->select_tx_size) swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
4414         memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
4415                sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
4416         ctx->sum_y_eobs = x->sum_y_eobs[TX_4X4];
4417 
4418         for (i = 0; i < 4; i++) best_bmodes[i] = xd->mi[0]->bmi[i];
4419 
4420         // TODO(debargha): enhance this test with a better distortion prediction
4421         // based on qp, activity mask and history
4422         if ((sf->mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
4423             (ref_index > MIN_EARLY_TERM_INDEX)) {
4424           int qstep = xd->plane[0].dequant[1];
4425           // TODO(debargha): Enhance this by specializing for each mode_index
4426           int scale = 4;
4427 #if CONFIG_VP9_HIGHBITDEPTH
4428           if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
4429             qstep >>= (xd->bd - 8);
4430           }
4431 #endif  // CONFIG_VP9_HIGHBITDEPTH
4432           if (x->source_variance < UINT_MAX) {
4433             const int var_adjust = (x->source_variance < 16);
4434             scale -= var_adjust;
4435           }
4436           if (ref_frame > INTRA_FRAME && distortion2 * scale < qstep * qstep) {
4437             early_term = 1;
4438           }
4439         }
4440       }
4441     }
4442 
4443     /* keep record of best compound/single-only prediction */
4444     if (!disable_skip && ref_frame != INTRA_FRAME) {
4445       int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
4446 
4447       if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4448         single_rate = rate2 - compmode_cost;
4449         hybrid_rate = rate2;
4450       } else {
4451         single_rate = rate2;
4452         hybrid_rate = rate2 + compmode_cost;
4453       }
4454 
4455       single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
4456       hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
4457 
4458       if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
4459         best_pred_rd[SINGLE_REFERENCE] = single_rd;
4460       else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
4461         best_pred_rd[COMPOUND_REFERENCE] = single_rd;
4462 
4463       if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
4464         best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
4465     }
4466 
4467     /* keep record of best filter type */
4468     if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
4469         cm->interp_filter != BILINEAR) {
4470       int64_t ref =
4471           filter_cache[cm->interp_filter == SWITCHABLE ? SWITCHABLE_FILTERS
4472                                                        : cm->interp_filter];
4473       int64_t adj_rd;
4474       for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4475         if (ref == INT64_MAX)
4476           adj_rd = 0;
4477         else if (filter_cache[i] == INT64_MAX)
4478           // when early termination is triggered, the encoder does not have
4479           // access to the rate-distortion cost. it only knows that the cost
4480           // should be above the maximum valid value. hence it takes the known
4481           // maximum plus an arbitrary constant as the rate-distortion cost.
4482           adj_rd = mask_filter - ref + 10;
4483         else
4484           adj_rd = filter_cache[i] - ref;
4485 
4486         adj_rd += this_rd;
4487         best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
4488       }
4489     }
4490 
4491     if (early_term) break;
4492 
4493     if (x->skip && !comp_pred) break;
4494   }
4495 
4496   if (best_rd >= best_rd_so_far) {
4497     rd_cost->rate = INT_MAX;
4498     rd_cost->rdcost = INT64_MAX;
4499     return;
4500   }
4501 
4502   // If we used an estimate for the uv intra rd in the loop above...
4503   if (sf->use_uv_intra_rd_estimate) {
4504     // Do Intra UV best rd mode selection if best mode choice above was intra.
4505     if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
4506       *mi = best_mbmode;
4507       rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra, &rate_uv_tokenonly,
4508                               &dist_uv, &skip_uv, BLOCK_8X8, TX_4X4);
4509     }
4510   }
4511 
4512   if (best_rd == INT64_MAX) {
4513     rd_cost->rate = INT_MAX;
4514     rd_cost->dist = INT64_MAX;
4515     rd_cost->rdcost = INT64_MAX;
4516     return;
4517   }
4518 
4519   assert((cm->interp_filter == SWITCHABLE) ||
4520          (cm->interp_filter == best_mbmode.interp_filter) ||
4521          !is_inter_block(&best_mbmode));
4522 
4523   vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact, sf->adaptive_rd_thresh,
4524                             bsize, best_ref_index);
4525 
4526   // macroblock modes
4527   *mi = best_mbmode;
4528   x->skip |= best_skip2;
4529   if (!is_inter_block(&best_mbmode)) {
4530     for (i = 0; i < 4; i++) xd->mi[0]->bmi[i].as_mode = best_bmodes[i].as_mode;
4531   } else {
4532     for (i = 0; i < 4; ++i)
4533       memcpy(&xd->mi[0]->bmi[i], &best_bmodes[i], sizeof(b_mode_info));
4534 
4535     mi->mv[0].as_int = xd->mi[0]->bmi[3].as_mv[0].as_int;
4536     mi->mv[1].as_int = xd->mi[0]->bmi[3].as_mv[1].as_int;
4537   }
4538 
4539   for (i = 0; i < REFERENCE_MODES; ++i) {
4540     if (best_pred_rd[i] == INT64_MAX)
4541       best_pred_diff[i] = INT_MIN;
4542     else
4543       best_pred_diff[i] = best_rd - best_pred_rd[i];
4544   }
4545 
4546   if (!x->skip) {
4547     for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4548       if (best_filter_rd[i] == INT64_MAX)
4549         best_filter_diff[i] = 0;
4550       else
4551         best_filter_diff[i] = best_rd - best_filter_rd[i];
4552     }
4553     if (cm->interp_filter == SWITCHABLE)
4554       assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
4555   } else {
4556     vp9_zero(best_filter_diff);
4557   }
4558 
4559   store_coding_context(x, ctx, best_ref_index, best_pred_diff, best_filter_diff,
4560                        0);
4561 }
4562