vp9_rdopt.c revision c927526be9a7b72fb5edb3f29c4e8ceabe0ec98a 
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    ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | \
46                                 (1 << INTRA_FRAME))
47#define GOLDEN_FRAME_MODE_MASK  ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | \
48                                 (1 << INTRA_FRAME))
49#define ALT_REF_MODE_MASK       ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | \
50                                 (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
57typedef struct {
58  PREDICTION_MODE mode;
59  MV_REFERENCE_FRAME ref_frame[2];
60} MODE_DEFINITION;
61
62typedef struct {
63  MV_REFERENCE_FRAME ref_frame[2];
64} REF_DEFINITION;
65
66struct rdcost_block_args {
67  MACROBLOCK *x;
68  ENTROPY_CONTEXT t_above[16];
69  ENTROPY_CONTEXT t_left[16];
70  int this_rate;
71  int64_t this_dist;
72  int64_t this_sse;
73  int64_t this_rd;
74  int64_t best_rd;
75  int exit_early;
76  int use_fast_coef_costing;
77  const scan_order *so;
78  uint8_t skippable;
79};
80
81#define LAST_NEW_MV_INDEX 6
82static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
83  {NEARESTMV, {LAST_FRAME,   NONE}},
84  {NEARESTMV, {ALTREF_FRAME, NONE}},
85  {NEARESTMV, {GOLDEN_FRAME, NONE}},
86
87  {DC_PRED,   {INTRA_FRAME,  NONE}},
88
89  {NEWMV,     {LAST_FRAME,   NONE}},
90  {NEWMV,     {ALTREF_FRAME, NONE}},
91  {NEWMV,     {GOLDEN_FRAME, NONE}},
92
93  {NEARMV,    {LAST_FRAME,   NONE}},
94  {NEARMV,    {ALTREF_FRAME, NONE}},
95  {NEARMV,    {GOLDEN_FRAME, NONE}},
96
97  {ZEROMV,    {LAST_FRAME,   NONE}},
98  {ZEROMV,    {GOLDEN_FRAME, NONE}},
99  {ZEROMV,    {ALTREF_FRAME, NONE}},
100
101  {NEARESTMV, {LAST_FRAME,   ALTREF_FRAME}},
102  {NEARESTMV, {GOLDEN_FRAME, ALTREF_FRAME}},
103
104  {TM_PRED,   {INTRA_FRAME,  NONE}},
105
106  {NEARMV,    {LAST_FRAME,   ALTREF_FRAME}},
107  {NEWMV,     {LAST_FRAME,   ALTREF_FRAME}},
108  {NEARMV,    {GOLDEN_FRAME, ALTREF_FRAME}},
109  {NEWMV,     {GOLDEN_FRAME, ALTREF_FRAME}},
110
111  {ZEROMV,    {LAST_FRAME,   ALTREF_FRAME}},
112  {ZEROMV,    {GOLDEN_FRAME, ALTREF_FRAME}},
113
114  {H_PRED,    {INTRA_FRAME,  NONE}},
115  {V_PRED,    {INTRA_FRAME,  NONE}},
116  {D135_PRED, {INTRA_FRAME,  NONE}},
117  {D207_PRED, {INTRA_FRAME,  NONE}},
118  {D153_PRED, {INTRA_FRAME,  NONE}},
119  {D63_PRED,  {INTRA_FRAME,  NONE}},
120  {D117_PRED, {INTRA_FRAME,  NONE}},
121  {D45_PRED,  {INTRA_FRAME,  NONE}},
122};
123
124static const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
125  {{LAST_FRAME,   NONE}},
126  {{GOLDEN_FRAME, NONE}},
127  {{ALTREF_FRAME, NONE}},
128  {{LAST_FRAME,   ALTREF_FRAME}},
129  {{GOLDEN_FRAME, ALTREF_FRAME}},
130  {{INTRA_FRAME,  NONE}},
131};
132
133static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
134                           int m, int n, int min_plane, int max_plane) {
135  int i;
136
137  for (i = min_plane; i < max_plane; ++i) {
138    struct macroblock_plane *const p = &x->plane[i];
139    struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
140
141    p->coeff    = ctx->coeff_pbuf[i][m];
142    p->qcoeff   = ctx->qcoeff_pbuf[i][m];
143    pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
144    p->eobs     = ctx->eobs_pbuf[i][m];
145
146    ctx->coeff_pbuf[i][m]   = ctx->coeff_pbuf[i][n];
147    ctx->qcoeff_pbuf[i][m]  = ctx->qcoeff_pbuf[i][n];
148    ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
149    ctx->eobs_pbuf[i][m]    = ctx->eobs_pbuf[i][n];
150
151    ctx->coeff_pbuf[i][n]   = p->coeff;
152    ctx->qcoeff_pbuf[i][n]  = p->qcoeff;
153    ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
154    ctx->eobs_pbuf[i][n]    = p->eobs;
155  }
156}
157
158static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize,
159                            MACROBLOCK *x, MACROBLOCKD *xd,
160                            int *out_rate_sum, int64_t *out_dist_sum,
161                            int *skip_txfm_sb, int64_t *skip_sse_sb) {
162  // Note our transform coeffs are 8 times an orthogonal transform.
163  // Hence quantizer step is also 8 times. To get effective quantizer
164  // we need to divide by 8 before sending to modeling function.
165  int i;
166  int64_t rate_sum = 0;
167  int64_t dist_sum = 0;
168  const int ref = xd->mi[0]->mbmi.ref_frame[0];
169  unsigned int sse;
170  unsigned int var = 0;
171  unsigned int sum_sse = 0;
172  int64_t total_sse = 0;
173  int skip_flag = 1;
174  const int shift = 6;
175  int rate;
176  int64_t dist;
177  const int dequant_shift =
178#if CONFIG_VP9_HIGHBITDEPTH
179      (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ?
180          xd->bd - 5 :
181#endif  // CONFIG_VP9_HIGHBITDEPTH
182          3;
183
184  x->pred_sse[ref] = 0;
185
186  for (i = 0; i < MAX_MB_PLANE; ++i) {
187    struct macroblock_plane *const p = &x->plane[i];
188    struct macroblockd_plane *const pd = &xd->plane[i];
189    const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
190    const TX_SIZE max_tx_size = max_txsize_lookup[bs];
191    const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
192    const int64_t dc_thr = p->quant_thred[0] >> shift;
193    const int64_t ac_thr = p->quant_thred[1] >> shift;
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    sum_sse = 0;
205
206    for (idy = 0; idy < bh; ++idy) {
207      for (idx = 0; idx < bw; ++idx) {
208        uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
209        uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
210        int block_idx = (idy << 1) + idx;
211        int low_err_skip = 0;
212
213        var = cpi->fn_ptr[unit_size].vf(src, p->src.stride,
214                                        dst, pd->dst.stride, &sse);
215        x->bsse[(i << 2) + block_idx] = sse;
216        sum_sse += sse;
217
218        x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_NONE;
219        if (!x->select_tx_size) {
220          // Check if all ac coefficients can be quantized to zero.
221          if (var < ac_thr || var == 0) {
222            x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_ONLY;
223
224            // Check if dc coefficient can be quantized to zero.
225            if (sse - var < dc_thr || sse == var) {
226              x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_DC;
227
228              if (!sse || (var < low_ac_thr && sse - var < low_dc_thr))
229                low_err_skip = 1;
230            }
231          }
232        }
233
234        if (skip_flag && !low_err_skip)
235          skip_flag = 0;
236
237        if (i == 0)
238          x->pred_sse[ref] += sse;
239      }
240    }
241
242    total_sse += sum_sse;
243
244    // Fast approximate the modelling function.
245    if (cpi->sf.simple_model_rd_from_var) {
246      int64_t rate;
247      const int64_t square_error = sum_sse;
248      int quantizer = (pd->dequant[1] >> dequant_shift);
249
250      if (quantizer < 120)
251        rate = (square_error * (280 - quantizer)) >> 8;
252      else
253        rate = 0;
254      dist = (square_error * quantizer) >> 8;
255      rate_sum += rate;
256      dist_sum += dist;
257    } else {
258      vp9_model_rd_from_var_lapndz(sum_sse, num_pels_log2_lookup[bs],
259                                   pd->dequant[1] >> dequant_shift,
260                                   &rate, &dist);
261      rate_sum += rate;
262      dist_sum += dist;
263    }
264  }
265
266  *skip_txfm_sb = skip_flag;
267  *skip_sse_sb = total_sse << 4;
268  *out_rate_sum = (int)rate_sum;
269  *out_dist_sum = dist_sum << 4;
270}
271
272#if CONFIG_VP9_HIGHBITDEPTH
273int64_t vp9_highbd_block_error_c(const tran_low_t *coeff,
274                                 const tran_low_t *dqcoeff,
275                                 intptr_t block_size,
276                                 int64_t *ssz, int bd) {
277  int i;
278  int64_t error = 0, sqcoeff = 0;
279  int shift = 2 * (bd - 8);
280  int rounding = shift > 0 ? 1 << (shift - 1) : 0;
281
282  for (i = 0; i < block_size; i++) {
283    const int64_t diff = coeff[i] - dqcoeff[i];
284    error +=  diff * diff;
285    sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
286  }
287  assert(error >= 0 && sqcoeff >= 0);
288  error = (error + rounding) >> shift;
289  sqcoeff = (sqcoeff + rounding) >> shift;
290
291  *ssz = sqcoeff;
292  return error;
293}
294
295int64_t vp9_highbd_block_error_8bit_c(const tran_low_t *coeff,
296                                      const tran_low_t *dqcoeff,
297                                      intptr_t block_size,
298                                      int64_t *ssz) {
299  // Note that the C versions of these 2 functions (vp9_block_error and
300  // vp9_highbd_block_error_8bit are the same, but the optimized assembly
301  // routines are not compatible in the non high bitdepth configuration, so
302  // they still cannot share the same name.
303  return vp9_block_error_c(coeff, dqcoeff, block_size, ssz);
304}
305
306static int64_t vp9_highbd_block_error_dispatch(const tran_low_t *coeff,
307                                               const tran_low_t *dqcoeff,
308                                               intptr_t block_size,
309                                               int64_t *ssz, int bd) {
310  if (bd == 8) {
311    return vp9_highbd_block_error_8bit(coeff, dqcoeff, block_size, ssz);
312  } else {
313    return vp9_highbd_block_error(coeff, dqcoeff, block_size, ssz, bd);
314  }
315}
316#endif  // CONFIG_VP9_HIGHBITDEPTH
317
318int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
319                          intptr_t block_size, int64_t *ssz) {
320  int i;
321  int64_t error = 0, sqcoeff = 0;
322
323  for (i = 0; i < block_size; i++) {
324    const int diff = coeff[i] - dqcoeff[i];
325    error +=  diff * diff;
326    sqcoeff += coeff[i] * coeff[i];
327  }
328
329  *ssz = sqcoeff;
330  return error;
331}
332
333int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff,
334                             int block_size) {
335  int i;
336  int64_t error = 0;
337
338  for (i = 0; i < block_size; i++) {
339    const int diff = coeff[i] - dqcoeff[i];
340    error +=  diff * diff;
341  }
342
343  return error;
344}
345
346/* The trailing '0' is a terminator which is used inside cost_coeffs() to
347 * decide whether to include cost of a trailing EOB node or not (i.e. we
348 * can skip this if the last coefficient in this transform block, e.g. the
349 * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
350 * were non-zero). */
351static const int16_t band_counts[TX_SIZES][8] = {
352  { 1, 2, 3, 4,  3,   16 - 13, 0 },
353  { 1, 2, 3, 4, 11,   64 - 21, 0 },
354  { 1, 2, 3, 4, 11,  256 - 21, 0 },
355  { 1, 2, 3, 4, 11, 1024 - 21, 0 },
356};
357static int cost_coeffs(MACROBLOCK *x,
358                       int plane, int block,
359                       ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L,
360                       TX_SIZE tx_size,
361                       const int16_t *scan, const int16_t *nb,
362                       int use_fast_coef_costing) {
363  MACROBLOCKD *const xd = &x->e_mbd;
364  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
365  const struct macroblock_plane *p = &x->plane[plane];
366  const PLANE_TYPE type = get_plane_type(plane);
367  const int16_t *band_count = &band_counts[tx_size][1];
368  const int eob = p->eobs[block];
369  const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
370  unsigned int (*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
371                   x->token_costs[tx_size][type][is_inter_block(mbmi)];
372  uint8_t token_cache[32 * 32];
373  int pt = combine_entropy_contexts(*A, *L);
374  int c, cost;
375#if CONFIG_VP9_HIGHBITDEPTH
376  const int16_t *cat6_high_cost = vp9_get_high_cost_table(xd->bd);
377#else
378  const int16_t *cat6_high_cost = vp9_get_high_cost_table(8);
379#endif
380
381  // Check for consistency of tx_size with mode info
382  assert(type == PLANE_TYPE_Y ? mbmi->tx_size == tx_size :
383         get_uv_tx_size(mbmi, &xd->plane[plane]) == tx_size);
384
385  if (eob == 0) {
386    // single eob token
387    cost = token_costs[0][0][pt][EOB_TOKEN];
388    c = 0;
389  } else {
390    int band_left = *band_count++;
391
392    // dc token
393    int v = qcoeff[0];
394    int16_t prev_t;
395    EXTRABIT e;
396    vp9_get_token_extra(v, &prev_t, &e);
397    cost = (*token_costs)[0][pt][prev_t] +
398        vp9_get_cost(prev_t, e, cat6_high_cost);
399
400    token_cache[0] = vp9_pt_energy_class[prev_t];
401    ++token_costs;
402
403    // ac tokens
404    for (c = 1; c < eob; c++) {
405      const int rc = scan[c];
406      int16_t t;
407
408      v = qcoeff[rc];
409      vp9_get_token_extra(v, &t, &e);
410      if (use_fast_coef_costing) {
411        cost += (*token_costs)[!prev_t][!prev_t][t] +
412            vp9_get_cost(t, e, cat6_high_cost);
413      } else {
414        pt = get_coef_context(nb, token_cache, c);
415        cost += (*token_costs)[!prev_t][pt][t] +
416            vp9_get_cost(t, e, cat6_high_cost);
417        token_cache[rc] = vp9_pt_energy_class[t];
418      }
419      prev_t = t;
420      if (!--band_left) {
421        band_left = *band_count++;
422        ++token_costs;
423      }
424    }
425
426    // eob token
427    if (band_left) {
428      if (use_fast_coef_costing) {
429        cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
430      } else {
431        pt = get_coef_context(nb, token_cache, c);
432        cost += (*token_costs)[0][pt][EOB_TOKEN];
433      }
434    }
435  }
436
437  // is eob first coefficient;
438  *A = *L = (c > 0);
439
440  return cost;
441}
442
443static void dist_block(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
444                       int64_t *out_dist, int64_t *out_sse) {
445  const int ss_txfrm_size = tx_size << 1;
446  MACROBLOCKD* const xd = &x->e_mbd;
447  const struct macroblock_plane *const p = &x->plane[plane];
448  const struct macroblockd_plane *const pd = &xd->plane[plane];
449  int64_t this_sse;
450  int shift = tx_size == TX_32X32 ? 0 : 2;
451  tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
452  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
453#if CONFIG_VP9_HIGHBITDEPTH
454  const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
455  *out_dist = vp9_highbd_block_error_dispatch(coeff, dqcoeff,
456                                              16 << ss_txfrm_size,
457                                              &this_sse, bd) >> shift;
458#else
459  *out_dist = vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
460                              &this_sse) >> shift;
461#endif  // CONFIG_VP9_HIGHBITDEPTH
462  *out_sse = this_sse >> shift;
463
464  if (x->skip_encode && !is_inter_block(&xd->mi[0]->mbmi)) {
465    // TODO(jingning): tune the model to better capture the distortion.
466    int64_t p = (pd->dequant[1] * pd->dequant[1] *
467                    (1 << ss_txfrm_size)) >>
468#if CONFIG_VP9_HIGHBITDEPTH
469                        (shift + 2 + (bd - 8) * 2);
470#else
471                        (shift + 2);
472#endif  // CONFIG_VP9_HIGHBITDEPTH
473    *out_dist += (p >> 4);
474    *out_sse  += p;
475  }
476}
477
478static int rate_block(int plane, int block, BLOCK_SIZE plane_bsize,
479                      TX_SIZE tx_size, struct rdcost_block_args* args) {
480  int x_idx, y_idx;
481  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x_idx, &y_idx);
482
483  return cost_coeffs(args->x, plane, block, args->t_above + x_idx,
484                     args->t_left + y_idx, tx_size,
485                     args->so->scan, args->so->neighbors,
486                     args->use_fast_coef_costing);
487}
488
489static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
490                          TX_SIZE tx_size, void *arg) {
491  struct rdcost_block_args *args = arg;
492  MACROBLOCK *const x = args->x;
493  MACROBLOCKD *const xd = &x->e_mbd;
494  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
495  int64_t rd1, rd2, rd;
496  int rate;
497  int64_t dist;
498  int64_t sse;
499
500  if (args->exit_early)
501    return;
502
503  if (!is_inter_block(mbmi)) {
504    struct encode_b_args arg = {x, NULL, &mbmi->skip};
505    vp9_encode_block_intra(plane, block, plane_bsize, tx_size, &arg);
506    dist_block(x, plane, block, tx_size, &dist, &sse);
507  } else if (max_txsize_lookup[plane_bsize] == tx_size) {
508    if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
509        SKIP_TXFM_NONE) {
510      // full forward transform and quantization
511      vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
512      dist_block(x, plane, block, tx_size, &dist, &sse);
513    } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
514               SKIP_TXFM_AC_ONLY) {
515      // compute DC coefficient
516      tran_low_t *const coeff   = BLOCK_OFFSET(x->plane[plane].coeff, block);
517      tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
518      vp9_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
519      sse  = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
520      dist = sse;
521      if (x->plane[plane].eobs[block]) {
522        const int64_t orig_sse = (int64_t)coeff[0] * coeff[0];
523        const int64_t resd_sse = coeff[0] - dqcoeff[0];
524        int64_t dc_correct = orig_sse - resd_sse * resd_sse;
525#if CONFIG_VP9_HIGHBITDEPTH
526        dc_correct >>= ((xd->bd - 8) * 2);
527#endif
528        if (tx_size != TX_32X32)
529          dc_correct >>= 2;
530
531        dist = VPXMAX(0, sse - dc_correct);
532      }
533    } else {
534      // SKIP_TXFM_AC_DC
535      // skip forward transform
536      x->plane[plane].eobs[block] = 0;
537      sse  = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
538      dist = sse;
539    }
540  } else {
541    // full forward transform and quantization
542    vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
543    dist_block(x, plane, block, tx_size, &dist, &sse);
544  }
545
546  rd = RDCOST(x->rdmult, x->rddiv, 0, dist);
547  if (args->this_rd + rd > args->best_rd) {
548    args->exit_early = 1;
549    return;
550  }
551
552  rate = rate_block(plane, block, plane_bsize, tx_size, args);
553  rd1 = RDCOST(x->rdmult, x->rddiv, rate, dist);
554  rd2 = RDCOST(x->rdmult, x->rddiv, 0, sse);
555
556  // TODO(jingning): temporarily enabled only for luma component
557  rd = VPXMIN(rd1, rd2);
558  if (plane == 0)
559    x->zcoeff_blk[tx_size][block] = !x->plane[plane].eobs[block] ||
560                                    (rd1 > rd2 && !xd->lossless);
561
562  args->this_rate += rate;
563  args->this_dist += dist;
564  args->this_sse += sse;
565  args->this_rd += rd;
566
567  if (args->this_rd > args->best_rd) {
568    args->exit_early = 1;
569    return;
570  }
571
572  args->skippable &= !x->plane[plane].eobs[block];
573}
574
575static void txfm_rd_in_plane(MACROBLOCK *x,
576                             int *rate, int64_t *distortion,
577                             int *skippable, int64_t *sse,
578                             int64_t ref_best_rd, int plane,
579                             BLOCK_SIZE bsize, TX_SIZE tx_size,
580                             int use_fast_coef_casting) {
581  MACROBLOCKD *const xd = &x->e_mbd;
582  const struct macroblockd_plane *const pd = &xd->plane[plane];
583  struct rdcost_block_args args;
584  vp9_zero(args);
585  args.x = x;
586  args.best_rd = ref_best_rd;
587  args.use_fast_coef_costing = use_fast_coef_casting;
588  args.skippable = 1;
589
590  if (plane == 0)
591    xd->mi[0]->mbmi.tx_size = tx_size;
592
593  vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
594
595  args.so = get_scan(xd, tx_size, get_plane_type(plane), 0);
596
597  vp9_foreach_transformed_block_in_plane(xd, bsize, plane,
598                                         block_rd_txfm, &args);
599  if (args.exit_early) {
600    *rate       = INT_MAX;
601    *distortion = INT64_MAX;
602    *sse        = INT64_MAX;
603    *skippable  = 0;
604  } else {
605    *distortion = args.this_dist;
606    *rate       = args.this_rate;
607    *sse        = args.this_sse;
608    *skippable  = args.skippable;
609  }
610}
611
612static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x,
613                                   int *rate, int64_t *distortion,
614                                   int *skip, int64_t *sse,
615                                   int64_t ref_best_rd,
616                                   BLOCK_SIZE bs) {
617  const TX_SIZE max_tx_size = max_txsize_lookup[bs];
618  VP9_COMMON *const cm = &cpi->common;
619  const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
620  MACROBLOCKD *const xd = &x->e_mbd;
621  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
622
623  mbmi->tx_size = VPXMIN(max_tx_size, largest_tx_size);
624
625  txfm_rd_in_plane(x, rate, distortion, skip,
626                   sse, ref_best_rd, 0, bs,
627                   mbmi->tx_size, cpi->sf.use_fast_coef_costing);
628}
629
630static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
631                                   int *rate,
632                                   int64_t *distortion,
633                                   int *skip,
634                                   int64_t *psse,
635                                   int64_t ref_best_rd,
636                                   BLOCK_SIZE bs) {
637  const TX_SIZE max_tx_size = max_txsize_lookup[bs];
638  VP9_COMMON *const cm = &cpi->common;
639  MACROBLOCKD *const xd = &x->e_mbd;
640  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
641  vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
642  int r[TX_SIZES][2], s[TX_SIZES];
643  int64_t d[TX_SIZES], sse[TX_SIZES];
644  int64_t rd[TX_SIZES][2] = {{INT64_MAX, INT64_MAX},
645                             {INT64_MAX, INT64_MAX},
646                             {INT64_MAX, INT64_MAX},
647                             {INT64_MAX, INT64_MAX}};
648  int n, m;
649  int s0, s1;
650  int64_t best_rd = INT64_MAX;
651  TX_SIZE best_tx = max_tx_size;
652  int start_tx, end_tx;
653
654  const vpx_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc->tx_probs);
655  assert(skip_prob > 0);
656  s0 = vp9_cost_bit(skip_prob, 0);
657  s1 = vp9_cost_bit(skip_prob, 1);
658
659  if (cm->tx_mode == TX_MODE_SELECT) {
660    start_tx = max_tx_size;
661    end_tx = 0;
662  } else {
663    TX_SIZE chosen_tx_size = VPXMIN(max_tx_size,
664                                    tx_mode_to_biggest_tx_size[cm->tx_mode]);
665    start_tx = chosen_tx_size;
666    end_tx = chosen_tx_size;
667  }
668
669  for (n = start_tx; n >= end_tx; n--) {
670    int r_tx_size = 0;
671    for (m = 0; m <= n - (n == (int) max_tx_size); m++) {
672      if (m == n)
673        r_tx_size += vp9_cost_zero(tx_probs[m]);
674      else
675        r_tx_size += vp9_cost_one(tx_probs[m]);
676    }
677    txfm_rd_in_plane(x, &r[n][0], &d[n], &s[n],
678                     &sse[n], ref_best_rd, 0, bs, n,
679                     cpi->sf.use_fast_coef_costing);
680    r[n][1] = r[n][0];
681    if (r[n][0] < INT_MAX) {
682      r[n][1] += r_tx_size;
683    }
684    if (d[n] == INT64_MAX || r[n][0] == INT_MAX) {
685      rd[n][0] = rd[n][1] = INT64_MAX;
686    } else if (s[n]) {
687      if (is_inter_block(mbmi)) {
688        rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
689        r[n][1] -= r_tx_size;
690      } else {
691        rd[n][0] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
692        rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1 + r_tx_size, sse[n]);
693      }
694    } else {
695      rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
696      rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
697    }
698
699    if (is_inter_block(mbmi) && !xd->lossless && !s[n] && sse[n] != INT64_MAX) {
700      rd[n][0] = VPXMIN(rd[n][0], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
701      rd[n][1] = VPXMIN(rd[n][1], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
702    }
703
704    // Early termination in transform size search.
705    if (cpi->sf.tx_size_search_breakout &&
706        (rd[n][1] == INT64_MAX ||
707        (n < (int) max_tx_size && rd[n][1] > rd[n + 1][1]) ||
708        s[n] == 1))
709      break;
710
711    if (rd[n][1] < best_rd) {
712      best_tx = n;
713      best_rd = rd[n][1];
714    }
715  }
716  mbmi->tx_size = best_tx;
717
718  *distortion = d[mbmi->tx_size];
719  *rate       = r[mbmi->tx_size][cm->tx_mode == TX_MODE_SELECT];
720  *skip       = s[mbmi->tx_size];
721  *psse       = sse[mbmi->tx_size];
722}
723
724static void super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
725                            int64_t *distortion, int *skip,
726                            int64_t *psse, BLOCK_SIZE bs,
727                            int64_t ref_best_rd) {
728  MACROBLOCKD *xd = &x->e_mbd;
729  int64_t sse;
730  int64_t *ret_sse = psse ? psse : &sse;
731
732  assert(bs == xd->mi[0]->mbmi.sb_type);
733
734  if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
735    choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
736                           bs);
737  } else {
738    choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse,
739                           ref_best_rd, bs);
740  }
741}
742
743static int conditional_skipintra(PREDICTION_MODE mode,
744                                 PREDICTION_MODE best_intra_mode) {
745  if (mode == D117_PRED &&
746      best_intra_mode != V_PRED &&
747      best_intra_mode != D135_PRED)
748    return 1;
749  if (mode == D63_PRED &&
750      best_intra_mode != V_PRED &&
751      best_intra_mode != D45_PRED)
752    return 1;
753  if (mode == D207_PRED &&
754      best_intra_mode != H_PRED &&
755      best_intra_mode != D45_PRED)
756    return 1;
757  if (mode == D153_PRED &&
758      best_intra_mode != H_PRED &&
759      best_intra_mode != D135_PRED)
760    return 1;
761  return 0;
762}
763
764static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x,
765                                     int row, int col,
766                                     PREDICTION_MODE *best_mode,
767                                     const int *bmode_costs,
768                                     ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
769                                     int *bestrate, int *bestratey,
770                                     int64_t *bestdistortion,
771                                     BLOCK_SIZE bsize, int64_t rd_thresh) {
772  PREDICTION_MODE mode;
773  MACROBLOCKD *const xd = &x->e_mbd;
774  int64_t best_rd = rd_thresh;
775  struct macroblock_plane *p = &x->plane[0];
776  struct macroblockd_plane *pd = &xd->plane[0];
777  const int src_stride = p->src.stride;
778  const int dst_stride = pd->dst.stride;
779  const uint8_t *src_init = &p->src.buf[row * 4 * src_stride + col * 4];
780  uint8_t *dst_init = &pd->dst.buf[row * 4 * src_stride + col * 4];
781  ENTROPY_CONTEXT ta[2], tempa[2];
782  ENTROPY_CONTEXT tl[2], templ[2];
783  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
784  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
785  int idx, idy;
786  uint8_t best_dst[8 * 8];
787#if CONFIG_VP9_HIGHBITDEPTH
788  uint16_t best_dst16[8 * 8];
789#endif
790
791  memcpy(ta, a, sizeof(ta));
792  memcpy(tl, l, sizeof(tl));
793  xd->mi[0]->mbmi.tx_size = TX_4X4;
794
795#if CONFIG_VP9_HIGHBITDEPTH
796  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
797    for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
798      int64_t this_rd;
799      int ratey = 0;
800      int64_t distortion = 0;
801      int rate = bmode_costs[mode];
802
803      if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
804        continue;
805
806      // Only do the oblique modes if the best so far is
807      // one of the neighboring directional modes
808      if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
809        if (conditional_skipintra(mode, *best_mode))
810            continue;
811      }
812
813      memcpy(tempa, ta, sizeof(ta));
814      memcpy(templ, tl, sizeof(tl));
815
816      for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
817        for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
818          const int block = (row + idy) * 2 + (col + idx);
819          const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
820          uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
821          int16_t *const src_diff = vp9_raster_block_offset_int16(BLOCK_8X8,
822                                                                  block,
823                                                                  p->src_diff);
824          tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
825          xd->mi[0]->bmi[block].as_mode = mode;
826          vp9_predict_intra_block(xd, 1, TX_4X4, mode,
827                                  x->skip_encode ? src : dst,
828                                  x->skip_encode ? src_stride : dst_stride,
829                                  dst, dst_stride,
830                                  col + idx, row + idy, 0);
831          vpx_highbd_subtract_block(4, 4, src_diff, 8, src, src_stride,
832                                    dst, dst_stride, xd->bd);
833          if (xd->lossless) {
834            const scan_order *so = &vp9_default_scan_orders[TX_4X4];
835            vp9_highbd_fwht4x4(src_diff, coeff, 8);
836            vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
837            ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
838                                 so->scan, so->neighbors,
839                                 cpi->sf.use_fast_coef_costing);
840            if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
841              goto next_highbd;
842            vp9_highbd_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block),
843                                   dst, dst_stride,
844                                   p->eobs[block], xd->bd);
845          } else {
846            int64_t unused;
847            const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
848            const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
849            if (tx_type == DCT_DCT)
850              vpx_highbd_fdct4x4(src_diff, coeff, 8);
851            else
852              vp9_highbd_fht4x4(src_diff, coeff, 8, tx_type);
853            vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
854            ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
855                                 so->scan, so->neighbors,
856                                 cpi->sf.use_fast_coef_costing);
857            distortion += vp9_highbd_block_error_dispatch(
858                coeff, BLOCK_OFFSET(pd->dqcoeff, block),
859                16, &unused, xd->bd) >> 2;
860            if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
861              goto next_highbd;
862            vp9_highbd_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
863                                  dst, dst_stride, p->eobs[block], xd->bd);
864          }
865        }
866      }
867
868      rate += ratey;
869      this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
870
871      if (this_rd < best_rd) {
872        *bestrate = rate;
873        *bestratey = ratey;
874        *bestdistortion = distortion;
875        best_rd = this_rd;
876        *best_mode = mode;
877        memcpy(a, tempa, sizeof(tempa));
878        memcpy(l, templ, sizeof(templ));
879        for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
880          memcpy(best_dst16 + idy * 8,
881                 CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
882                 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
883        }
884      }
885    next_highbd:
886      {}
887    }
888    if (best_rd >= rd_thresh || x->skip_encode)
889      return best_rd;
890
891    for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
892      memcpy(CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
893             best_dst16 + idy * 8,
894             num_4x4_blocks_wide * 4 * sizeof(uint16_t));
895    }
896
897    return best_rd;
898  }
899#endif  // CONFIG_VP9_HIGHBITDEPTH
900
901  for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
902    int64_t this_rd;
903    int ratey = 0;
904    int64_t distortion = 0;
905    int rate = bmode_costs[mode];
906
907    if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
908      continue;
909
910    // Only do the oblique modes if the best so far is
911    // one of the neighboring directional modes
912    if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
913      if (conditional_skipintra(mode, *best_mode))
914          continue;
915    }
916
917    memcpy(tempa, ta, sizeof(ta));
918    memcpy(templ, tl, sizeof(tl));
919
920    for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
921      for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
922        const int block = (row + idy) * 2 + (col + idx);
923        const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
924        uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
925        int16_t *const src_diff =
926            vp9_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
927        tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
928        xd->mi[0]->bmi[block].as_mode = mode;
929        vp9_predict_intra_block(xd, 1, TX_4X4, mode,
930                                x->skip_encode ? src : dst,
931                                x->skip_encode ? src_stride : dst_stride,
932                                dst, dst_stride, col + idx, row + idy, 0);
933        vpx_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
934
935        if (xd->lossless) {
936          const scan_order *so = &vp9_default_scan_orders[TX_4X4];
937          vp9_fwht4x4(src_diff, coeff, 8);
938          vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
939          ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
940                               so->scan, so->neighbors,
941                               cpi->sf.use_fast_coef_costing);
942          if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
943            goto next;
944          vp9_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride,
945                          p->eobs[block]);
946        } else {
947          int64_t unused;
948          const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
949          const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
950          vp9_fht4x4(src_diff, coeff, 8, tx_type);
951          vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
952          ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
953                             so->scan, so->neighbors,
954                             cpi->sf.use_fast_coef_costing);
955#if CONFIG_VP9_HIGHBITDEPTH
956          distortion += vp9_highbd_block_error_8bit(
957              coeff, BLOCK_OFFSET(pd->dqcoeff, block), 16, &unused) >> 2;
958#else
959          distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
960                                        16, &unused) >> 2;
961#endif
962          if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
963            goto next;
964          vp9_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
965                         dst, dst_stride, p->eobs[block]);
966        }
967      }
968    }
969
970    rate += ratey;
971    this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
972
973    if (this_rd < best_rd) {
974      *bestrate = rate;
975      *bestratey = ratey;
976      *bestdistortion = distortion;
977      best_rd = this_rd;
978      *best_mode = mode;
979      memcpy(a, tempa, sizeof(tempa));
980      memcpy(l, templ, sizeof(templ));
981      for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
982        memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
983               num_4x4_blocks_wide * 4);
984    }
985  next:
986    {}
987  }
988
989  if (best_rd >= rd_thresh || x->skip_encode)
990    return best_rd;
991
992  for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
993    memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
994           num_4x4_blocks_wide * 4);
995
996  return best_rd;
997}
998
999static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
1000                                            int *rate, int *rate_y,
1001                                            int64_t *distortion,
1002                                            int64_t best_rd) {
1003  int i, j;
1004  const MACROBLOCKD *const xd = &mb->e_mbd;
1005  MODE_INFO *const mic = xd->mi[0];
1006  const MODE_INFO *above_mi = xd->above_mi;
1007  const MODE_INFO *left_mi = xd->left_mi;
1008  const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
1009  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1010  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1011  int idx, idy;
1012  int cost = 0;
1013  int64_t total_distortion = 0;
1014  int tot_rate_y = 0;
1015  int64_t total_rd = 0;
1016  ENTROPY_CONTEXT t_above[4], t_left[4];
1017  const int *bmode_costs = cpi->mbmode_cost;
1018
1019  memcpy(t_above, xd->plane[0].above_context, sizeof(t_above));
1020  memcpy(t_left, xd->plane[0].left_context, sizeof(t_left));
1021
1022  // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
1023  for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1024    for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1025      PREDICTION_MODE best_mode = DC_PRED;
1026      int r = INT_MAX, ry = INT_MAX;
1027      int64_t d = INT64_MAX, this_rd = INT64_MAX;
1028      i = idy * 2 + idx;
1029      if (cpi->common.frame_type == KEY_FRAME) {
1030        const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i);
1031        const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i);
1032
1033        bmode_costs  = cpi->y_mode_costs[A][L];
1034      }
1035
1036      this_rd = rd_pick_intra4x4block(cpi, mb, idy, idx, &best_mode,
1037                                      bmode_costs, t_above + idx, t_left + idy,
1038                                      &r, &ry, &d, bsize, best_rd - total_rd);
1039      if (this_rd >= best_rd - total_rd)
1040        return INT64_MAX;
1041
1042      total_rd += this_rd;
1043      cost += r;
1044      total_distortion += d;
1045      tot_rate_y += ry;
1046
1047      mic->bmi[i].as_mode = best_mode;
1048      for (j = 1; j < num_4x4_blocks_high; ++j)
1049        mic->bmi[i + j * 2].as_mode = best_mode;
1050      for (j = 1; j < num_4x4_blocks_wide; ++j)
1051        mic->bmi[i + j].as_mode = best_mode;
1052
1053      if (total_rd >= best_rd)
1054        return INT64_MAX;
1055    }
1056  }
1057
1058  *rate = cost;
1059  *rate_y = tot_rate_y;
1060  *distortion = total_distortion;
1061  mic->mbmi.mode = mic->bmi[3].as_mode;
1062
1063  return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
1064}
1065
1066// This function is used only for intra_only frames
1067static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x,
1068                                      int *rate, int *rate_tokenonly,
1069                                      int64_t *distortion, int *skippable,
1070                                      BLOCK_SIZE bsize,
1071                                      int64_t best_rd) {
1072  PREDICTION_MODE mode;
1073  PREDICTION_MODE mode_selected = DC_PRED;
1074  MACROBLOCKD *const xd = &x->e_mbd;
1075  MODE_INFO *const mic = xd->mi[0];
1076  int this_rate, this_rate_tokenonly, s;
1077  int64_t this_distortion, this_rd;
1078  TX_SIZE best_tx = TX_4X4;
1079  int *bmode_costs;
1080  const MODE_INFO *above_mi = xd->above_mi;
1081  const MODE_INFO *left_mi = xd->left_mi;
1082  const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1083  const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1084  bmode_costs = cpi->y_mode_costs[A][L];
1085
1086  memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1087  /* Y Search for intra prediction mode */
1088  for (mode = DC_PRED; mode <= TM_PRED; mode++) {
1089    if (cpi->sf.use_nonrd_pick_mode) {
1090      // These speed features are turned on in hybrid non-RD and RD mode
1091      // for key frame coding in the context of real-time setting.
1092      if (conditional_skipintra(mode, mode_selected))
1093          continue;
1094      if (*skippable)
1095        break;
1096    }
1097
1098    mic->mbmi.mode = mode;
1099
1100    super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion,
1101        &s, NULL, bsize, best_rd);
1102
1103    if (this_rate_tokenonly == INT_MAX)
1104      continue;
1105
1106    this_rate = this_rate_tokenonly + bmode_costs[mode];
1107    this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1108
1109    if (this_rd < best_rd) {
1110      mode_selected   = mode;
1111      best_rd         = this_rd;
1112      best_tx         = mic->mbmi.tx_size;
1113      *rate           = this_rate;
1114      *rate_tokenonly = this_rate_tokenonly;
1115      *distortion     = this_distortion;
1116      *skippable      = s;
1117    }
1118  }
1119
1120  mic->mbmi.mode = mode_selected;
1121  mic->mbmi.tx_size = best_tx;
1122
1123  return best_rd;
1124}
1125
1126// Return value 0: early termination triggered, no valid rd cost available;
1127//              1: rd cost values are valid.
1128static int super_block_uvrd(const VP9_COMP *cpi, MACROBLOCK *x,
1129                            int *rate, int64_t *distortion, int *skippable,
1130                            int64_t *sse, BLOCK_SIZE bsize,
1131                            int64_t ref_best_rd) {
1132  MACROBLOCKD *const xd = &x->e_mbd;
1133  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1134  const TX_SIZE uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
1135  int plane;
1136  int pnrate = 0, pnskip = 1;
1137  int64_t pndist = 0, pnsse = 0;
1138  int is_cost_valid = 1;
1139
1140  if (ref_best_rd < 0)
1141    is_cost_valid = 0;
1142
1143  if (is_inter_block(mbmi) && is_cost_valid) {
1144    int plane;
1145    for (plane = 1; plane < MAX_MB_PLANE; ++plane)
1146      vp9_subtract_plane(x, bsize, plane);
1147  }
1148
1149  *rate = 0;
1150  *distortion = 0;
1151  *sse = 0;
1152  *skippable = 1;
1153
1154  for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
1155    txfm_rd_in_plane(x, &pnrate, &pndist, &pnskip, &pnsse,
1156                     ref_best_rd, plane, bsize, uv_tx_size,
1157                     cpi->sf.use_fast_coef_costing);
1158    if (pnrate == INT_MAX) {
1159      is_cost_valid = 0;
1160      break;
1161    }
1162    *rate += pnrate;
1163    *distortion += pndist;
1164    *sse += pnsse;
1165    *skippable &= pnskip;
1166  }
1167
1168  if (!is_cost_valid) {
1169    // reset cost value
1170    *rate = INT_MAX;
1171    *distortion = INT64_MAX;
1172    *sse = INT64_MAX;
1173    *skippable = 0;
1174  }
1175
1176  return is_cost_valid;
1177}
1178
1179static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
1180                                       PICK_MODE_CONTEXT *ctx,
1181                                       int *rate, int *rate_tokenonly,
1182                                       int64_t *distortion, int *skippable,
1183                                       BLOCK_SIZE bsize, TX_SIZE max_tx_size) {
1184  MACROBLOCKD *xd = &x->e_mbd;
1185  PREDICTION_MODE mode;
1186  PREDICTION_MODE mode_selected = DC_PRED;
1187  int64_t best_rd = INT64_MAX, this_rd;
1188  int this_rate_tokenonly, this_rate, s;
1189  int64_t this_distortion, this_sse;
1190
1191  memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1192  for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1193    if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode)))
1194      continue;
1195
1196    xd->mi[0]->mbmi.uv_mode = mode;
1197
1198    if (!super_block_uvrd(cpi, x, &this_rate_tokenonly,
1199                          &this_distortion, &s, &this_sse, bsize, best_rd))
1200      continue;
1201    this_rate = this_rate_tokenonly +
1202                cpi->intra_uv_mode_cost[cpi->common.frame_type][mode];
1203    this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1204
1205    if (this_rd < best_rd) {
1206      mode_selected   = mode;
1207      best_rd         = this_rd;
1208      *rate           = this_rate;
1209      *rate_tokenonly = this_rate_tokenonly;
1210      *distortion     = this_distortion;
1211      *skippable      = s;
1212      if (!x->select_tx_size)
1213        swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
1214    }
1215  }
1216
1217  xd->mi[0]->mbmi.uv_mode = mode_selected;
1218  return best_rd;
1219}
1220
1221static int64_t rd_sbuv_dcpred(const VP9_COMP *cpi, MACROBLOCK *x,
1222                              int *rate, int *rate_tokenonly,
1223                              int64_t *distortion, int *skippable,
1224                              BLOCK_SIZE bsize) {
1225  const VP9_COMMON *cm = &cpi->common;
1226  int64_t unused;
1227
1228  x->e_mbd.mi[0]->mbmi.uv_mode = DC_PRED;
1229  memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1230  super_block_uvrd(cpi, x, rate_tokenonly, distortion,
1231                   skippable, &unused, bsize, INT64_MAX);
1232  *rate = *rate_tokenonly + cpi->intra_uv_mode_cost[cm->frame_type][DC_PRED];
1233  return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
1234}
1235
1236static void choose_intra_uv_mode(VP9_COMP *cpi, MACROBLOCK *const x,
1237                                 PICK_MODE_CONTEXT *ctx,
1238                                 BLOCK_SIZE bsize, TX_SIZE max_tx_size,
1239                                 int *rate_uv, int *rate_uv_tokenonly,
1240                                 int64_t *dist_uv, int *skip_uv,
1241                                 PREDICTION_MODE *mode_uv) {
1242  // Use an estimated rd for uv_intra based on DC_PRED if the
1243  // appropriate speed flag is set.
1244  if (cpi->sf.use_uv_intra_rd_estimate) {
1245    rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv,
1246                   skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
1247  // Else do a proper rd search for each possible transform size that may
1248  // be considered in the main rd loop.
1249  } else {
1250    rd_pick_intra_sbuv_mode(cpi, x, ctx,
1251                            rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
1252                            bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize, max_tx_size);
1253  }
1254  *mode_uv = x->e_mbd.mi[0]->mbmi.uv_mode;
1255}
1256
1257static int cost_mv_ref(const VP9_COMP *cpi, PREDICTION_MODE mode,
1258                       int mode_context) {
1259  assert(is_inter_mode(mode));
1260  return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
1261}
1262
1263static int set_and_cost_bmi_mvs(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
1264                                int i,
1265                                PREDICTION_MODE mode, int_mv this_mv[2],
1266                                int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1267                                int_mv seg_mvs[MAX_REF_FRAMES],
1268                                int_mv *best_ref_mv[2], const int *mvjcost,
1269                                int *mvcost[2]) {
1270  MODE_INFO *const mic = xd->mi[0];
1271  const MB_MODE_INFO *const mbmi = &mic->mbmi;
1272  const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1273  int thismvcost = 0;
1274  int idx, idy;
1275  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type];
1276  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
1277  const int is_compound = has_second_ref(mbmi);
1278
1279  switch (mode) {
1280    case NEWMV:
1281      this_mv[0].as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
1282      thismvcost += vp9_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
1283                                    mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1284      if (is_compound) {
1285        this_mv[1].as_int = seg_mvs[mbmi->ref_frame[1]].as_int;
1286        thismvcost += vp9_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
1287                                      mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1288      }
1289      break;
1290    case NEARMV:
1291    case NEARESTMV:
1292      this_mv[0].as_int = frame_mv[mode][mbmi->ref_frame[0]].as_int;
1293      if (is_compound)
1294        this_mv[1].as_int = frame_mv[mode][mbmi->ref_frame[1]].as_int;
1295      break;
1296    case ZEROMV:
1297      this_mv[0].as_int = 0;
1298      if (is_compound)
1299        this_mv[1].as_int = 0;
1300      break;
1301    default:
1302      break;
1303  }
1304
1305  mic->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
1306  if (is_compound)
1307    mic->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
1308
1309  mic->bmi[i].as_mode = mode;
1310
1311  for (idy = 0; idy < num_4x4_blocks_high; ++idy)
1312    for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
1313      memmove(&mic->bmi[i + idy * 2 + idx], &mic->bmi[i], sizeof(mic->bmi[i]));
1314
1315  return cost_mv_ref(cpi, mode, mbmi_ext->mode_context[mbmi->ref_frame[0]]) +
1316            thismvcost;
1317}
1318
1319static int64_t encode_inter_mb_segment(VP9_COMP *cpi,
1320                                       MACROBLOCK *x,
1321                                       int64_t best_yrd,
1322                                       int i,
1323                                       int *labelyrate,
1324                                       int64_t *distortion, int64_t *sse,
1325                                       ENTROPY_CONTEXT *ta,
1326                                       ENTROPY_CONTEXT *tl,
1327                                       int mi_row, int mi_col) {
1328  int k;
1329  MACROBLOCKD *xd = &x->e_mbd;
1330  struct macroblockd_plane *const pd = &xd->plane[0];
1331  struct macroblock_plane *const p = &x->plane[0];
1332  MODE_INFO *const mi = xd->mi[0];
1333  const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->mbmi.sb_type, pd);
1334  const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
1335  const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
1336  int idx, idy;
1337
1338  const uint8_t *const src =
1339      &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1340  uint8_t *const dst = &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i,
1341                                                            pd->dst.stride)];
1342  int64_t thisdistortion = 0, thissse = 0;
1343  int thisrate = 0, ref;
1344  const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1345  const int is_compound = has_second_ref(&mi->mbmi);
1346  const InterpKernel *kernel = vp9_filter_kernels[mi->mbmi.interp_filter];
1347
1348  for (ref = 0; ref < 1 + is_compound; ++ref) {
1349    const uint8_t *pre = &pd->pre[ref].buf[vp9_raster_block_offset(BLOCK_8X8, i,
1350                                               pd->pre[ref].stride)];
1351#if CONFIG_VP9_HIGHBITDEPTH
1352  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1353    vp9_highbd_build_inter_predictor(pre, pd->pre[ref].stride,
1354                                     dst, pd->dst.stride,
1355                                     &mi->bmi[i].as_mv[ref].as_mv,
1356                                     &xd->block_refs[ref]->sf, width, height,
1357                                     ref, kernel, MV_PRECISION_Q3,
1358                                     mi_col * MI_SIZE + 4 * (i % 2),
1359                                     mi_row * MI_SIZE + 4 * (i / 2), xd->bd);
1360  } else {
1361    vp9_build_inter_predictor(pre, pd->pre[ref].stride,
1362                              dst, pd->dst.stride,
1363                              &mi->bmi[i].as_mv[ref].as_mv,
1364                              &xd->block_refs[ref]->sf, width, height, ref,
1365                              kernel, MV_PRECISION_Q3,
1366                              mi_col * MI_SIZE + 4 * (i % 2),
1367                              mi_row * MI_SIZE + 4 * (i / 2));
1368  }
1369#else
1370    vp9_build_inter_predictor(pre, pd->pre[ref].stride,
1371                              dst, pd->dst.stride,
1372                              &mi->bmi[i].as_mv[ref].as_mv,
1373                              &xd->block_refs[ref]->sf, width, height, ref,
1374                              kernel, MV_PRECISION_Q3,
1375                              mi_col * MI_SIZE + 4 * (i % 2),
1376                              mi_row * MI_SIZE + 4 * (i / 2));
1377#endif  // CONFIG_VP9_HIGHBITDEPTH
1378  }
1379
1380#if CONFIG_VP9_HIGHBITDEPTH
1381  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1382    vpx_highbd_subtract_block(
1383        height, width, vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1384        8, src, p->src.stride, dst, pd->dst.stride, xd->bd);
1385  } else {
1386    vpx_subtract_block(
1387        height, width, vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1388        8, src, p->src.stride, dst, pd->dst.stride);
1389  }
1390#else
1391  vpx_subtract_block(height, width,
1392                     vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1393                     8, src, p->src.stride, dst, pd->dst.stride);
1394#endif  // CONFIG_VP9_HIGHBITDEPTH
1395
1396  k = i;
1397  for (idy = 0; idy < height / 4; ++idy) {
1398    for (idx = 0; idx < width / 4; ++idx) {
1399#if CONFIG_VP9_HIGHBITDEPTH
1400      const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
1401#endif
1402      int64_t ssz, rd, rd1, rd2;
1403      tran_low_t* coeff;
1404
1405      k += (idy * 2 + idx);
1406      coeff = BLOCK_OFFSET(p->coeff, k);
1407      x->fwd_txm4x4(vp9_raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
1408                    coeff, 8);
1409      vp9_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan);
1410#if CONFIG_VP9_HIGHBITDEPTH
1411      thisdistortion += vp9_highbd_block_error_dispatch(
1412          coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz, bd);
1413#else
1414      thisdistortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
1415                                        16, &ssz);
1416#endif  // CONFIG_VP9_HIGHBITDEPTH
1417      thissse += ssz;
1418      thisrate += cost_coeffs(x, 0, k, ta + (k & 1), tl + (k >> 1), TX_4X4,
1419                              so->scan, so->neighbors,
1420                              cpi->sf.use_fast_coef_costing);
1421      rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
1422      rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
1423      rd = VPXMIN(rd1, rd2);
1424      if (rd >= best_yrd)
1425        return INT64_MAX;
1426    }
1427  }
1428
1429  *distortion = thisdistortion >> 2;
1430  *labelyrate = thisrate;
1431  *sse = thissse >> 2;
1432
1433  return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
1434}
1435
1436typedef struct {
1437  int eobs;
1438  int brate;
1439  int byrate;
1440  int64_t bdist;
1441  int64_t bsse;
1442  int64_t brdcost;
1443  int_mv mvs[2];
1444  ENTROPY_CONTEXT ta[2];
1445  ENTROPY_CONTEXT tl[2];
1446} SEG_RDSTAT;
1447
1448typedef struct {
1449  int_mv *ref_mv[2];
1450  int_mv mvp;
1451
1452  int64_t segment_rd;
1453  int r;
1454  int64_t d;
1455  int64_t sse;
1456  int segment_yrate;
1457  PREDICTION_MODE modes[4];
1458  SEG_RDSTAT rdstat[4][INTER_MODES];
1459  int mvthresh;
1460} BEST_SEG_INFO;
1461
1462static INLINE int mv_check_bounds(const MACROBLOCK *x, const MV *mv) {
1463  return (mv->row >> 3) < x->mv_row_min ||
1464         (mv->row >> 3) > x->mv_row_max ||
1465         (mv->col >> 3) < x->mv_col_min ||
1466         (mv->col >> 3) > x->mv_col_max;
1467}
1468
1469static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
1470  MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0]->mbmi;
1471  struct macroblock_plane *const p = &x->plane[0];
1472  struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
1473
1474  p->src.buf = &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i,
1475                                                   p->src.stride)];
1476  assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
1477  pd->pre[0].buf = &pd->pre[0].buf[vp9_raster_block_offset(BLOCK_8X8, i,
1478                                                           pd->pre[0].stride)];
1479  if (has_second_ref(mbmi))
1480    pd->pre[1].buf = &pd->pre[1].buf[vp9_raster_block_offset(BLOCK_8X8, i,
1481                                                           pd->pre[1].stride)];
1482}
1483
1484static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
1485                                  struct buf_2d orig_pre[2]) {
1486  MB_MODE_INFO *mbmi = &x->e_mbd.mi[0]->mbmi;
1487  x->plane[0].src = orig_src;
1488  x->e_mbd.plane[0].pre[0] = orig_pre[0];
1489  if (has_second_ref(mbmi))
1490    x->e_mbd.plane[0].pre[1] = orig_pre[1];
1491}
1492
1493static INLINE int mv_has_subpel(const MV *mv) {
1494  return (mv->row & 0x0F) || (mv->col & 0x0F);
1495}
1496
1497// Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
1498// TODO(aconverse): Find out if this is still productive then clean up or remove
1499static int check_best_zero_mv(
1500    const VP9_COMP *cpi, const uint8_t mode_context[MAX_REF_FRAMES],
1501    int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int this_mode,
1502    const MV_REFERENCE_FRAME ref_frames[2]) {
1503  if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
1504      frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
1505      (ref_frames[1] == NONE ||
1506       frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
1507    int rfc = mode_context[ref_frames[0]];
1508    int c1 = cost_mv_ref(cpi, NEARMV, rfc);
1509    int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
1510    int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
1511
1512    if (this_mode == NEARMV) {
1513      if (c1 > c3) return 0;
1514    } else if (this_mode == NEARESTMV) {
1515      if (c2 > c3) return 0;
1516    } else {
1517      assert(this_mode == ZEROMV);
1518      if (ref_frames[1] == NONE) {
1519        if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
1520            (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
1521          return 0;
1522      } else {
1523        if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
1524             frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
1525            (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
1526             frame_mv[NEARMV][ref_frames[1]].as_int == 0))
1527          return 0;
1528      }
1529    }
1530  }
1531  return 1;
1532}
1533
1534static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
1535                                BLOCK_SIZE bsize,
1536                                int_mv *frame_mv,
1537                                int mi_row, int mi_col,
1538                                int_mv single_newmv[MAX_REF_FRAMES],
1539                                int *rate_mv) {
1540  const VP9_COMMON *const cm = &cpi->common;
1541  const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
1542  const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
1543  MACROBLOCKD *xd = &x->e_mbd;
1544  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
1545  const int refs[2] = {mbmi->ref_frame[0],
1546                       mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]};
1547  int_mv ref_mv[2];
1548  int ite, ref;
1549  const InterpKernel *kernel = vp9_filter_kernels[mbmi->interp_filter];
1550  struct scale_factors sf;
1551
1552  // Do joint motion search in compound mode to get more accurate mv.
1553  struct buf_2d backup_yv12[2][MAX_MB_PLANE];
1554  int last_besterr[2] = {INT_MAX, INT_MAX};
1555  const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
1556    vp9_get_scaled_ref_frame(cpi, mbmi->ref_frame[0]),
1557    vp9_get_scaled_ref_frame(cpi, mbmi->ref_frame[1])
1558  };
1559
1560  // Prediction buffer from second frame.
1561#if CONFIG_VP9_HIGHBITDEPTH
1562  DECLARE_ALIGNED(16, uint16_t, second_pred_alloc_16[64 * 64]);
1563  uint8_t *second_pred;
1564#else
1565  DECLARE_ALIGNED(16, uint8_t, second_pred[64 * 64]);
1566#endif  // CONFIG_VP9_HIGHBITDEPTH
1567
1568  for (ref = 0; ref < 2; ++ref) {
1569    ref_mv[ref] = x->mbmi_ext->ref_mvs[refs[ref]][0];
1570
1571    if (scaled_ref_frame[ref]) {
1572      int i;
1573      // Swap out the reference frame for a version that's been scaled to
1574      // match the resolution of the current frame, allowing the existing
1575      // motion search code to be used without additional modifications.
1576      for (i = 0; i < MAX_MB_PLANE; i++)
1577        backup_yv12[ref][i] = xd->plane[i].pre[ref];
1578      vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
1579                           NULL);
1580    }
1581
1582    frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
1583  }
1584
1585  // Since we have scaled the reference frames to match the size of the current
1586  // frame we must use a unit scaling factor during mode selection.
1587#if CONFIG_VP9_HIGHBITDEPTH
1588  vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height,
1589                                    cm->width, cm->height,
1590                                    cm->use_highbitdepth);
1591#else
1592  vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height,
1593                                    cm->width, cm->height);
1594#endif  // CONFIG_VP9_HIGHBITDEPTH
1595
1596  // Allow joint search multiple times iteratively for each reference frame
1597  // and break out of the search loop if it couldn't find a better mv.
1598  for (ite = 0; ite < 4; ite++) {
1599    struct buf_2d ref_yv12[2];
1600    int bestsme = INT_MAX;
1601    int sadpb = x->sadperbit16;
1602    MV tmp_mv;
1603    int search_range = 3;
1604
1605    int tmp_col_min = x->mv_col_min;
1606    int tmp_col_max = x->mv_col_max;
1607    int tmp_row_min = x->mv_row_min;
1608    int tmp_row_max = x->mv_row_max;
1609    int id = ite % 2;  // Even iterations search in the first reference frame,
1610                       // odd iterations search in the second. The predictor
1611                       // found for the 'other' reference frame is factored in.
1612
1613    // Initialized here because of compiler problem in Visual Studio.
1614    ref_yv12[0] = xd->plane[0].pre[0];
1615    ref_yv12[1] = xd->plane[0].pre[1];
1616
1617    // Get the prediction block from the 'other' reference frame.
1618#if CONFIG_VP9_HIGHBITDEPTH
1619    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1620      second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc_16);
1621      vp9_highbd_build_inter_predictor(ref_yv12[!id].buf,
1622                                       ref_yv12[!id].stride,
1623                                       second_pred, pw,
1624                                       &frame_mv[refs[!id]].as_mv,
1625                                       &sf, pw, ph, 0,
1626                                       kernel, MV_PRECISION_Q3,
1627                                       mi_col * MI_SIZE, mi_row * MI_SIZE,
1628                                       xd->bd);
1629    } else {
1630      second_pred = (uint8_t *)second_pred_alloc_16;
1631      vp9_build_inter_predictor(ref_yv12[!id].buf,
1632                                ref_yv12[!id].stride,
1633                                second_pred, pw,
1634                                &frame_mv[refs[!id]].as_mv,
1635                                &sf, pw, ph, 0,
1636                                kernel, MV_PRECISION_Q3,
1637                                mi_col * MI_SIZE, mi_row * MI_SIZE);
1638    }
1639#else
1640    vp9_build_inter_predictor(ref_yv12[!id].buf,
1641                              ref_yv12[!id].stride,
1642                              second_pred, pw,
1643                              &frame_mv[refs[!id]].as_mv,
1644                              &sf, pw, ph, 0,
1645                              kernel, MV_PRECISION_Q3,
1646                              mi_col * MI_SIZE, mi_row * MI_SIZE);
1647#endif  // CONFIG_VP9_HIGHBITDEPTH
1648
1649    // Do compound motion search on the current reference frame.
1650    if (id)
1651      xd->plane[0].pre[0] = ref_yv12[id];
1652    vp9_set_mv_search_range(x, &ref_mv[id].as_mv);
1653
1654    // Use the mv result from the single mode as mv predictor.
1655    tmp_mv = frame_mv[refs[id]].as_mv;
1656
1657    tmp_mv.col >>= 3;
1658    tmp_mv.row >>= 3;
1659
1660    // Small-range full-pixel motion search.
1661    bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb,
1662                                       search_range,
1663                                       &cpi->fn_ptr[bsize],
1664                                       &ref_mv[id].as_mv, second_pred);
1665    if (bestsme < INT_MAX)
1666      bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
1667                                      second_pred, &cpi->fn_ptr[bsize], 1);
1668
1669    x->mv_col_min = tmp_col_min;
1670    x->mv_col_max = tmp_col_max;
1671    x->mv_row_min = tmp_row_min;
1672    x->mv_row_max = tmp_row_max;
1673
1674    if (bestsme < INT_MAX) {
1675      int dis; /* TODO: use dis in distortion calculation later. */
1676      unsigned int sse;
1677      bestsme = cpi->find_fractional_mv_step(
1678          x, &tmp_mv,
1679          &ref_mv[id].as_mv,
1680          cpi->common.allow_high_precision_mv,
1681          x->errorperbit,
1682          &cpi->fn_ptr[bsize],
1683          0, cpi->sf.mv.subpel_iters_per_step,
1684          NULL,
1685          x->nmvjointcost, x->mvcost,
1686          &dis, &sse, second_pred,
1687          pw, ph);
1688    }
1689
1690    // Restore the pointer to the first (possibly scaled) prediction buffer.
1691    if (id)
1692      xd->plane[0].pre[0] = ref_yv12[0];
1693
1694    if (bestsme < last_besterr[id]) {
1695      frame_mv[refs[id]].as_mv = tmp_mv;
1696      last_besterr[id] = bestsme;
1697    } else {
1698      break;
1699    }
1700  }
1701
1702  *rate_mv = 0;
1703
1704  for (ref = 0; ref < 2; ++ref) {
1705    if (scaled_ref_frame[ref]) {
1706      // Restore the prediction frame pointers to their unscaled versions.
1707      int i;
1708      for (i = 0; i < MAX_MB_PLANE; i++)
1709        xd->plane[i].pre[ref] = backup_yv12[ref][i];
1710    }
1711
1712    *rate_mv += vp9_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
1713                                &x->mbmi_ext->ref_mvs[refs[ref]][0].as_mv,
1714                                x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1715  }
1716}
1717
1718static int64_t rd_pick_best_sub8x8_mode(VP9_COMP *cpi, MACROBLOCK *x,
1719                                        int_mv *best_ref_mv,
1720                                        int_mv *second_best_ref_mv,
1721                                        int64_t best_rd, int *returntotrate,
1722                                        int *returnyrate,
1723                                        int64_t *returndistortion,
1724                                        int *skippable, int64_t *psse,
1725                                        int mvthresh,
1726                                        int_mv seg_mvs[4][MAX_REF_FRAMES],
1727                                        BEST_SEG_INFO *bsi_buf, int filter_idx,
1728                                        int mi_row, int mi_col) {
1729  int i;
1730  BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
1731  MACROBLOCKD *xd = &x->e_mbd;
1732  MODE_INFO *mi = xd->mi[0];
1733  MB_MODE_INFO *mbmi = &mi->mbmi;
1734  int mode_idx;
1735  int k, br = 0, idx, idy;
1736  int64_t bd = 0, block_sse = 0;
1737  PREDICTION_MODE this_mode;
1738  VP9_COMMON *cm = &cpi->common;
1739  struct macroblock_plane *const p = &x->plane[0];
1740  struct macroblockd_plane *const pd = &xd->plane[0];
1741  const int label_count = 4;
1742  int64_t this_segment_rd = 0;
1743  int label_mv_thresh;
1744  int segmentyrate = 0;
1745  const BLOCK_SIZE bsize = mbmi->sb_type;
1746  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1747  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1748  ENTROPY_CONTEXT t_above[2], t_left[2];
1749  int subpelmv = 1, have_ref = 0;
1750  const int has_second_rf = has_second_ref(mbmi);
1751  const int inter_mode_mask = cpi->sf.inter_mode_mask[bsize];
1752  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1753
1754  vp9_zero(*bsi);
1755
1756  bsi->segment_rd = best_rd;
1757  bsi->ref_mv[0] = best_ref_mv;
1758  bsi->ref_mv[1] = second_best_ref_mv;
1759  bsi->mvp.as_int = best_ref_mv->as_int;
1760  bsi->mvthresh = mvthresh;
1761
1762  for (i = 0; i < 4; i++)
1763    bsi->modes[i] = ZEROMV;
1764
1765  memcpy(t_above, pd->above_context, sizeof(t_above));
1766  memcpy(t_left, pd->left_context, sizeof(t_left));
1767
1768  // 64 makes this threshold really big effectively
1769  // making it so that we very rarely check mvs on
1770  // segments.   setting this to 1 would make mv thresh
1771  // roughly equal to what it is for macroblocks
1772  label_mv_thresh = 1 * bsi->mvthresh / label_count;
1773
1774  // Segmentation method overheads
1775  for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1776    for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1777      // TODO(jingning,rbultje): rewrite the rate-distortion optimization
1778      // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
1779      int_mv mode_mv[MB_MODE_COUNT][2];
1780      int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1781      PREDICTION_MODE mode_selected = ZEROMV;
1782      int64_t best_rd = INT64_MAX;
1783      const int i = idy * 2 + idx;
1784      int ref;
1785
1786      for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1787        const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
1788        frame_mv[ZEROMV][frame].as_int = 0;
1789        vp9_append_sub8x8_mvs_for_idx(cm, xd, i, ref, mi_row, mi_col,
1790                                      &frame_mv[NEARESTMV][frame],
1791                                      &frame_mv[NEARMV][frame],
1792                                      mbmi_ext->mode_context);
1793      }
1794
1795      // search for the best motion vector on this segment
1796      for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1797        const struct buf_2d orig_src = x->plane[0].src;
1798        struct buf_2d orig_pre[2];
1799
1800        mode_idx = INTER_OFFSET(this_mode);
1801        bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
1802        if (!(inter_mode_mask & (1 << this_mode)))
1803          continue;
1804
1805        if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
1806                                this_mode, mbmi->ref_frame))
1807          continue;
1808
1809        memcpy(orig_pre, pd->pre, sizeof(orig_pre));
1810        memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
1811               sizeof(bsi->rdstat[i][mode_idx].ta));
1812        memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
1813               sizeof(bsi->rdstat[i][mode_idx].tl));
1814
1815        // motion search for newmv (single predictor case only)
1816        if (!has_second_rf && this_mode == NEWMV &&
1817            seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV) {
1818          MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
1819          int step_param = 0;
1820          int thissme, bestsme = INT_MAX;
1821          int sadpb = x->sadperbit4;
1822          MV mvp_full;
1823          int max_mv;
1824          int cost_list[5];
1825
1826          /* Is the best so far sufficiently good that we cant justify doing
1827           * and new motion search. */
1828          if (best_rd < label_mv_thresh)
1829            break;
1830
1831          if (cpi->oxcf.mode != BEST) {
1832            // use previous block's result as next block's MV predictor.
1833            if (i > 0) {
1834              bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
1835              if (i == 2)
1836                bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
1837            }
1838          }
1839          if (i == 0)
1840            max_mv = x->max_mv_context[mbmi->ref_frame[0]];
1841          else
1842            max_mv =
1843                VPXMAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
1844
1845          if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
1846            // Take wtd average of the step_params based on the last frame's
1847            // max mv magnitude and the best ref mvs of the current block for
1848            // the given reference.
1849            step_param = (vp9_init_search_range(max_mv) +
1850                              cpi->mv_step_param) / 2;
1851          } else {
1852            step_param = cpi->mv_step_param;
1853          }
1854
1855          mvp_full.row = bsi->mvp.as_mv.row >> 3;
1856          mvp_full.col = bsi->mvp.as_mv.col >> 3;
1857
1858          if (cpi->sf.adaptive_motion_search) {
1859            mvp_full.row = x->pred_mv[mbmi->ref_frame[0]].row >> 3;
1860            mvp_full.col = x->pred_mv[mbmi->ref_frame[0]].col >> 3;
1861            step_param = VPXMAX(step_param, 8);
1862          }
1863
1864          // adjust src pointer for this block
1865          mi_buf_shift(x, i);
1866
1867          vp9_set_mv_search_range(x, &bsi->ref_mv[0]->as_mv);
1868
1869          bestsme = vp9_full_pixel_search(
1870              cpi, x, bsize, &mvp_full, step_param, sadpb,
1871              cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
1872              &bsi->ref_mv[0]->as_mv, new_mv,
1873              INT_MAX, 1);
1874
1875          // Should we do a full search (best quality only)
1876          if (cpi->oxcf.mode == BEST) {
1877            int_mv *const best_mv = &mi->bmi[i].as_mv[0];
1878            /* Check if mvp_full is within the range. */
1879            clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max,
1880                     x->mv_row_min, x->mv_row_max);
1881            thissme = cpi->full_search_sad(x, &mvp_full,
1882                                           sadpb, 16, &cpi->fn_ptr[bsize],
1883                                           &bsi->ref_mv[0]->as_mv,
1884                                           &best_mv->as_mv);
1885            cost_list[1] = cost_list[2] = cost_list[3] = cost_list[4] = INT_MAX;
1886            if (thissme < bestsme) {
1887              bestsme = thissme;
1888              *new_mv = best_mv->as_mv;
1889            } else {
1890              // The full search result is actually worse so re-instate the
1891              // previous best vector
1892              best_mv->as_mv = *new_mv;
1893            }
1894          }
1895
1896          if (bestsme < INT_MAX) {
1897            int distortion;
1898            cpi->find_fractional_mv_step(
1899                x,
1900                new_mv,
1901                &bsi->ref_mv[0]->as_mv,
1902                cm->allow_high_precision_mv,
1903                x->errorperbit, &cpi->fn_ptr[bsize],
1904                cpi->sf.mv.subpel_force_stop,
1905                cpi->sf.mv.subpel_iters_per_step,
1906                cond_cost_list(cpi, cost_list),
1907                x->nmvjointcost, x->mvcost,
1908                &distortion,
1909                &x->pred_sse[mbmi->ref_frame[0]],
1910                NULL, 0, 0);
1911
1912            // save motion search result for use in compound prediction
1913            seg_mvs[i][mbmi->ref_frame[0]].as_mv = *new_mv;
1914          }
1915
1916          if (cpi->sf.adaptive_motion_search)
1917            x->pred_mv[mbmi->ref_frame[0]] = *new_mv;
1918
1919          // restore src pointers
1920          mi_buf_restore(x, orig_src, orig_pre);
1921        }
1922
1923        if (has_second_rf) {
1924          if (seg_mvs[i][mbmi->ref_frame[1]].as_int == INVALID_MV ||
1925              seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV)
1926            continue;
1927        }
1928
1929        if (has_second_rf && this_mode == NEWMV &&
1930            mbmi->interp_filter == EIGHTTAP) {
1931          // adjust src pointers
1932          mi_buf_shift(x, i);
1933          if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
1934            int rate_mv;
1935            joint_motion_search(cpi, x, bsize, frame_mv[this_mode],
1936                                mi_row, mi_col, seg_mvs[i],
1937                                &rate_mv);
1938            seg_mvs[i][mbmi->ref_frame[0]].as_int =
1939                frame_mv[this_mode][mbmi->ref_frame[0]].as_int;
1940            seg_mvs[i][mbmi->ref_frame[1]].as_int =
1941                frame_mv[this_mode][mbmi->ref_frame[1]].as_int;
1942          }
1943          // restore src pointers
1944          mi_buf_restore(x, orig_src, orig_pre);
1945        }
1946
1947        bsi->rdstat[i][mode_idx].brate =
1948            set_and_cost_bmi_mvs(cpi, x, xd, i, this_mode, mode_mv[this_mode],
1949                                 frame_mv, seg_mvs[i], bsi->ref_mv,
1950                                 x->nmvjointcost, x->mvcost);
1951
1952        for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1953          bsi->rdstat[i][mode_idx].mvs[ref].as_int =
1954              mode_mv[this_mode][ref].as_int;
1955          if (num_4x4_blocks_wide > 1)
1956            bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int =
1957                mode_mv[this_mode][ref].as_int;
1958          if (num_4x4_blocks_high > 1)
1959            bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int =
1960                mode_mv[this_mode][ref].as_int;
1961        }
1962
1963        // Trap vectors that reach beyond the UMV borders
1964        if (mv_check_bounds(x, &mode_mv[this_mode][0].as_mv) ||
1965            (has_second_rf &&
1966             mv_check_bounds(x, &mode_mv[this_mode][1].as_mv)))
1967          continue;
1968
1969        if (filter_idx > 0) {
1970          BEST_SEG_INFO *ref_bsi = bsi_buf;
1971          subpelmv = 0;
1972          have_ref = 1;
1973
1974          for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1975            subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
1976            have_ref &= mode_mv[this_mode][ref].as_int ==
1977                ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1978          }
1979
1980          if (filter_idx > 1 && !subpelmv && !have_ref) {
1981            ref_bsi = bsi_buf + 1;
1982            have_ref = 1;
1983            for (ref = 0; ref < 1 + has_second_rf; ++ref)
1984              have_ref &= mode_mv[this_mode][ref].as_int ==
1985                  ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1986          }
1987
1988          if (!subpelmv && have_ref &&
1989              ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1990            memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
1991                   sizeof(SEG_RDSTAT));
1992            if (num_4x4_blocks_wide > 1)
1993              bsi->rdstat[i + 1][mode_idx].eobs =
1994                  ref_bsi->rdstat[i + 1][mode_idx].eobs;
1995            if (num_4x4_blocks_high > 1)
1996              bsi->rdstat[i + 2][mode_idx].eobs =
1997                  ref_bsi->rdstat[i + 2][mode_idx].eobs;
1998
1999            if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
2000              mode_selected = this_mode;
2001              best_rd = bsi->rdstat[i][mode_idx].brdcost;
2002            }
2003            continue;
2004          }
2005        }
2006
2007        bsi->rdstat[i][mode_idx].brdcost =
2008            encode_inter_mb_segment(cpi, x,
2009                                    bsi->segment_rd - this_segment_rd, i,
2010                                    &bsi->rdstat[i][mode_idx].byrate,
2011                                    &bsi->rdstat[i][mode_idx].bdist,
2012                                    &bsi->rdstat[i][mode_idx].bsse,
2013                                    bsi->rdstat[i][mode_idx].ta,
2014                                    bsi->rdstat[i][mode_idx].tl,
2015                                    mi_row, mi_col);
2016        if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
2017          bsi->rdstat[i][mode_idx].brdcost += RDCOST(x->rdmult, x->rddiv,
2018                                            bsi->rdstat[i][mode_idx].brate, 0);
2019          bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
2020          bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
2021          if (num_4x4_blocks_wide > 1)
2022            bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
2023          if (num_4x4_blocks_high > 1)
2024            bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
2025        }
2026
2027        if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
2028          mode_selected = this_mode;
2029          best_rd = bsi->rdstat[i][mode_idx].brdcost;
2030        }
2031      } /*for each 4x4 mode*/
2032
2033      if (best_rd == INT64_MAX) {
2034        int iy, midx;
2035        for (iy = i + 1; iy < 4; ++iy)
2036          for (midx = 0; midx < INTER_MODES; ++midx)
2037            bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2038        bsi->segment_rd = INT64_MAX;
2039        return INT64_MAX;
2040      }
2041
2042      mode_idx = INTER_OFFSET(mode_selected);
2043      memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
2044      memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
2045
2046      set_and_cost_bmi_mvs(cpi, x, xd, i, mode_selected, mode_mv[mode_selected],
2047                           frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
2048                           x->mvcost);
2049
2050      br += bsi->rdstat[i][mode_idx].brate;
2051      bd += bsi->rdstat[i][mode_idx].bdist;
2052      block_sse += bsi->rdstat[i][mode_idx].bsse;
2053      segmentyrate += bsi->rdstat[i][mode_idx].byrate;
2054      this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
2055
2056      if (this_segment_rd > bsi->segment_rd) {
2057        int iy, midx;
2058        for (iy = i + 1; iy < 4; ++iy)
2059          for (midx = 0; midx < INTER_MODES; ++midx)
2060            bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2061        bsi->segment_rd = INT64_MAX;
2062        return INT64_MAX;
2063      }
2064    }
2065  } /* for each label */
2066
2067  bsi->r = br;
2068  bsi->d = bd;
2069  bsi->segment_yrate = segmentyrate;
2070  bsi->segment_rd = this_segment_rd;
2071  bsi->sse = block_sse;
2072
2073  // update the coding decisions
2074  for (k = 0; k < 4; ++k)
2075    bsi->modes[k] = mi->bmi[k].as_mode;
2076
2077  if (bsi->segment_rd > best_rd)
2078    return INT64_MAX;
2079  /* set it to the best */
2080  for (i = 0; i < 4; i++) {
2081    mode_idx = INTER_OFFSET(bsi->modes[i]);
2082    mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
2083    if (has_second_ref(mbmi))
2084      mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
2085    x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
2086    mi->bmi[i].as_mode = bsi->modes[i];
2087  }
2088
2089  /*
2090   * used to set mbmi->mv.as_int
2091   */
2092  *returntotrate = bsi->r;
2093  *returndistortion = bsi->d;
2094  *returnyrate = bsi->segment_yrate;
2095  *skippable = vp9_is_skippable_in_plane(x, BLOCK_8X8, 0);
2096  *psse = bsi->sse;
2097  mbmi->mode = bsi->modes[3];
2098
2099  return bsi->segment_rd;
2100}
2101
2102static void estimate_ref_frame_costs(const VP9_COMMON *cm,
2103                                     const MACROBLOCKD *xd,
2104                                     int segment_id,
2105                                     unsigned int *ref_costs_single,
2106                                     unsigned int *ref_costs_comp,
2107                                     vpx_prob *comp_mode_p) {
2108  int seg_ref_active = segfeature_active(&cm->seg, segment_id,
2109                                         SEG_LVL_REF_FRAME);
2110  if (seg_ref_active) {
2111    memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
2112    memset(ref_costs_comp,   0, MAX_REF_FRAMES * sizeof(*ref_costs_comp));
2113    *comp_mode_p = 128;
2114  } else {
2115    vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
2116    vpx_prob comp_inter_p = 128;
2117
2118    if (cm->reference_mode == REFERENCE_MODE_SELECT) {
2119      comp_inter_p = vp9_get_reference_mode_prob(cm, xd);
2120      *comp_mode_p = comp_inter_p;
2121    } else {
2122      *comp_mode_p = 128;
2123    }
2124
2125    ref_costs_single[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
2126
2127    if (cm->reference_mode != COMPOUND_REFERENCE) {
2128      vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
2129      vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
2130      unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
2131
2132      if (cm->reference_mode == REFERENCE_MODE_SELECT)
2133        base_cost += vp9_cost_bit(comp_inter_p, 0);
2134
2135      ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] =
2136          ref_costs_single[ALTREF_FRAME] = base_cost;
2137      ref_costs_single[LAST_FRAME]   += vp9_cost_bit(ref_single_p1, 0);
2138      ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
2139      ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
2140      ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
2141      ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
2142    } else {
2143      ref_costs_single[LAST_FRAME]   = 512;
2144      ref_costs_single[GOLDEN_FRAME] = 512;
2145      ref_costs_single[ALTREF_FRAME] = 512;
2146    }
2147    if (cm->reference_mode != SINGLE_REFERENCE) {
2148      vpx_prob ref_comp_p = vp9_get_pred_prob_comp_ref_p(cm, xd);
2149      unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
2150
2151      if (cm->reference_mode == REFERENCE_MODE_SELECT)
2152        base_cost += vp9_cost_bit(comp_inter_p, 1);
2153
2154      ref_costs_comp[LAST_FRAME]   = base_cost + vp9_cost_bit(ref_comp_p, 0);
2155      ref_costs_comp[GOLDEN_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 1);
2156    } else {
2157      ref_costs_comp[LAST_FRAME]   = 512;
2158      ref_costs_comp[GOLDEN_FRAME] = 512;
2159    }
2160  }
2161}
2162
2163static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
2164                         int mode_index,
2165                         int64_t comp_pred_diff[REFERENCE_MODES],
2166                         int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS],
2167                         int skippable) {
2168  MACROBLOCKD *const xd = &x->e_mbd;
2169
2170  // Take a snapshot of the coding context so it can be
2171  // restored if we decide to encode this way
2172  ctx->skip = x->skip;
2173  ctx->skippable = skippable;
2174  ctx->best_mode_index = mode_index;
2175  ctx->mic = *xd->mi[0];
2176  ctx->mbmi_ext = *x->mbmi_ext;
2177  ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
2178  ctx->comp_pred_diff   = (int)comp_pred_diff[COMPOUND_REFERENCE];
2179  ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
2180
2181  memcpy(ctx->best_filter_diff, best_filter_diff,
2182         sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
2183}
2184
2185static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
2186                               MV_REFERENCE_FRAME ref_frame,
2187                               BLOCK_SIZE block_size,
2188                               int mi_row, int mi_col,
2189                               int_mv frame_nearest_mv[MAX_REF_FRAMES],
2190                               int_mv frame_near_mv[MAX_REF_FRAMES],
2191                               struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
2192  const VP9_COMMON *cm = &cpi->common;
2193  const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2194  MACROBLOCKD *const xd = &x->e_mbd;
2195  MODE_INFO *const mi = xd->mi[0];
2196  int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
2197  const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2198  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2199
2200  assert(yv12 != NULL);
2201
2202  // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
2203  // use the UV scaling factors.
2204  vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
2205
2206  // Gets an initial list of candidate vectors from neighbours and orders them
2207  vp9_find_mv_refs(cm, xd, mi, ref_frame, candidates, mi_row, mi_col,
2208                   NULL, NULL, mbmi_ext->mode_context);
2209
2210  // Candidate refinement carried out at encoder and decoder
2211  vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2212                        &frame_nearest_mv[ref_frame],
2213                        &frame_near_mv[ref_frame]);
2214
2215  // Further refinement that is encode side only to test the top few candidates
2216  // in full and choose the best as the centre point for subsequent searches.
2217  // The current implementation doesn't support scaling.
2218  if (!vp9_is_scaled(sf) && block_size >= BLOCK_8X8)
2219    vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
2220                ref_frame, block_size);
2221}
2222
2223static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
2224                                 BLOCK_SIZE bsize,
2225                                 int mi_row, int mi_col,
2226                                 int_mv *tmp_mv, int *rate_mv) {
2227  MACROBLOCKD *xd = &x->e_mbd;
2228  const VP9_COMMON *cm = &cpi->common;
2229  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
2230  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
2231  int bestsme = INT_MAX;
2232  int step_param;
2233  int sadpb = x->sadperbit16;
2234  MV mvp_full;
2235  int ref = mbmi->ref_frame[0];
2236  MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2237
2238  int tmp_col_min = x->mv_col_min;
2239  int tmp_col_max = x->mv_col_max;
2240  int tmp_row_min = x->mv_row_min;
2241  int tmp_row_max = x->mv_row_max;
2242  int cost_list[5];
2243
2244  const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi,
2245                                                                        ref);
2246
2247  MV pred_mv[3];
2248  pred_mv[0] = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2249  pred_mv[1] = x->mbmi_ext->ref_mvs[ref][1].as_mv;
2250  pred_mv[2] = x->pred_mv[ref];
2251
2252  if (scaled_ref_frame) {
2253    int i;
2254    // Swap out the reference frame for a version that's been scaled to
2255    // match the resolution of the current frame, allowing the existing
2256    // motion search code to be used without additional modifications.
2257    for (i = 0; i < MAX_MB_PLANE; i++)
2258      backup_yv12[i] = xd->plane[i].pre[0];
2259
2260    vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
2261  }
2262
2263  vp9_set_mv_search_range(x, &ref_mv);
2264
2265  // Work out the size of the first step in the mv step search.
2266  // 0 here is maximum length first step. 1 is VPXMAX >> 1 etc.
2267  if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
2268    // Take wtd average of the step_params based on the last frame's
2269    // max mv magnitude and that based on the best ref mvs of the current
2270    // block for the given reference.
2271    step_param = (vp9_init_search_range(x->max_mv_context[ref]) +
2272                    cpi->mv_step_param) / 2;
2273  } else {
2274    step_param = cpi->mv_step_param;
2275  }
2276
2277  if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64) {
2278    int boffset =
2279        2 * (b_width_log2_lookup[BLOCK_64X64] -
2280             VPXMIN(b_height_log2_lookup[bsize], b_width_log2_lookup[bsize]));
2281    step_param = VPXMAX(step_param, boffset);
2282  }
2283
2284  if (cpi->sf.adaptive_motion_search) {
2285    int bwl = b_width_log2_lookup[bsize];
2286    int bhl = b_height_log2_lookup[bsize];
2287    int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
2288
2289    if (tlevel < 5)
2290      step_param += 2;
2291
2292    // prev_mv_sad is not setup for dynamically scaled frames.
2293    if (cpi->oxcf.resize_mode != RESIZE_DYNAMIC) {
2294      int i;
2295      for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
2296        if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
2297          x->pred_mv[ref].row = 0;
2298          x->pred_mv[ref].col = 0;
2299          tmp_mv->as_int = INVALID_MV;
2300
2301          if (scaled_ref_frame) {
2302            int i;
2303            for (i = 0; i < MAX_MB_PLANE; ++i)
2304              xd->plane[i].pre[0] = backup_yv12[i];
2305          }
2306          return;
2307        }
2308      }
2309    }
2310  }
2311
2312  mvp_full = pred_mv[x->mv_best_ref_index[ref]];
2313
2314  mvp_full.col >>= 3;
2315  mvp_full.row >>= 3;
2316
2317  bestsme = vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
2318                                  cond_cost_list(cpi, cost_list),
2319                                  &ref_mv, &tmp_mv->as_mv, INT_MAX, 1);
2320
2321  x->mv_col_min = tmp_col_min;
2322  x->mv_col_max = tmp_col_max;
2323  x->mv_row_min = tmp_row_min;
2324  x->mv_row_max = tmp_row_max;
2325
2326  if (bestsme < INT_MAX) {
2327    int dis;  /* TODO: use dis in distortion calculation later. */
2328    cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
2329                                 cm->allow_high_precision_mv,
2330                                 x->errorperbit,
2331                                 &cpi->fn_ptr[bsize],
2332                                 cpi->sf.mv.subpel_force_stop,
2333                                 cpi->sf.mv.subpel_iters_per_step,
2334                                 cond_cost_list(cpi, cost_list),
2335                                 x->nmvjointcost, x->mvcost,
2336                                 &dis, &x->pred_sse[ref], NULL, 0, 0);
2337  }
2338  *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
2339                             x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2340
2341  if (cpi->sf.adaptive_motion_search)
2342    x->pred_mv[ref] = tmp_mv->as_mv;
2343
2344  if (scaled_ref_frame) {
2345    int i;
2346    for (i = 0; i < MAX_MB_PLANE; i++)
2347      xd->plane[i].pre[0] = backup_yv12[i];
2348  }
2349}
2350
2351
2352
2353static INLINE void restore_dst_buf(MACROBLOCKD *xd,
2354                                   uint8_t *orig_dst[MAX_MB_PLANE],
2355                                   int orig_dst_stride[MAX_MB_PLANE]) {
2356  int i;
2357  for (i = 0; i < MAX_MB_PLANE; i++) {
2358    xd->plane[i].dst.buf = orig_dst[i];
2359    xd->plane[i].dst.stride = orig_dst_stride[i];
2360  }
2361}
2362
2363// In some situations we want to discount tha pparent cost of a new motion
2364// vector. Where there is a subtle motion field and especially where there is
2365// low spatial complexity then it can be hard to cover the cost of a new motion
2366// vector in a single block, even if that motion vector reduces distortion.
2367// However, once established that vector may be usable through the nearest and
2368// near mv modes to reduce distortion in subsequent blocks and also improve
2369// visual quality.
2370static int discount_newmv_test(const VP9_COMP *cpi,
2371                               int this_mode,
2372                               int_mv this_mv,
2373                               int_mv (*mode_mv)[MAX_REF_FRAMES],
2374                               int ref_frame) {
2375  return (!cpi->rc.is_src_frame_alt_ref &&
2376          (this_mode == NEWMV) &&
2377          (this_mv.as_int != 0) &&
2378          ((mode_mv[NEARESTMV][ref_frame].as_int == 0) ||
2379           (mode_mv[NEARESTMV][ref_frame].as_int == INVALID_MV)) &&
2380          ((mode_mv[NEARMV][ref_frame].as_int == 0) ||
2381           (mode_mv[NEARMV][ref_frame].as_int == INVALID_MV)));
2382}
2383
2384static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
2385                                 BLOCK_SIZE bsize,
2386                                 int *rate2, int64_t *distortion,
2387                                 int *skippable,
2388                                 int *rate_y, int *rate_uv,
2389                                 int *disable_skip,
2390                                 int_mv (*mode_mv)[MAX_REF_FRAMES],
2391                                 int mi_row, int mi_col,
2392                                 int_mv single_newmv[MAX_REF_FRAMES],
2393                                 INTERP_FILTER (*single_filter)[MAX_REF_FRAMES],
2394                                 int (*single_skippable)[MAX_REF_FRAMES],
2395                                 int64_t *psse,
2396                                 const int64_t ref_best_rd,
2397                                 int64_t *mask_filter,
2398                                 int64_t filter_cache[]) {
2399  VP9_COMMON *cm = &cpi->common;
2400  MACROBLOCKD *xd = &x->e_mbd;
2401  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
2402  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2403  const int is_comp_pred = has_second_ref(mbmi);
2404  const int this_mode = mbmi->mode;
2405  int_mv *frame_mv = mode_mv[this_mode];
2406  int i;
2407  int refs[2] = { mbmi->ref_frame[0],
2408    (mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]) };
2409  int_mv cur_mv[2];
2410#if CONFIG_VP9_HIGHBITDEPTH
2411  DECLARE_ALIGNED(16, uint16_t, tmp_buf16[MAX_MB_PLANE * 64 * 64]);
2412  uint8_t *tmp_buf;
2413#else
2414  DECLARE_ALIGNED(16, uint8_t, tmp_buf[MAX_MB_PLANE * 64 * 64]);
2415#endif  // CONFIG_VP9_HIGHBITDEPTH
2416  int pred_exists = 0;
2417  int intpel_mv;
2418  int64_t rd, tmp_rd, best_rd = INT64_MAX;
2419  int best_needs_copy = 0;
2420  uint8_t *orig_dst[MAX_MB_PLANE];
2421  int orig_dst_stride[MAX_MB_PLANE];
2422  int rs = 0;
2423  INTERP_FILTER best_filter = SWITCHABLE;
2424  uint8_t skip_txfm[MAX_MB_PLANE << 2] = {0};
2425  int64_t bsse[MAX_MB_PLANE << 2] = {0};
2426
2427  int bsl = mi_width_log2_lookup[bsize];
2428  int pred_filter_search = cpi->sf.cb_pred_filter_search ?
2429      (((mi_row + mi_col) >> bsl) +
2430       get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
2431
2432  int skip_txfm_sb = 0;
2433  int64_t skip_sse_sb = INT64_MAX;
2434  int64_t distortion_y = 0, distortion_uv = 0;
2435
2436#if CONFIG_VP9_HIGHBITDEPTH
2437  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2438    tmp_buf = CONVERT_TO_BYTEPTR(tmp_buf16);
2439  } else {
2440    tmp_buf = (uint8_t *)tmp_buf16;
2441  }
2442#endif  // CONFIG_VP9_HIGHBITDEPTH
2443
2444  if (pred_filter_search) {
2445    INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE;
2446    if (xd->up_available)
2447      af = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
2448    if (xd->left_available)
2449      lf = xd->mi[-1]->mbmi.interp_filter;
2450
2451    if ((this_mode != NEWMV) || (af == lf))
2452      best_filter = af;
2453  }
2454
2455  if (is_comp_pred) {
2456    if (frame_mv[refs[0]].as_int == INVALID_MV ||
2457        frame_mv[refs[1]].as_int == INVALID_MV)
2458      return INT64_MAX;
2459
2460    if (cpi->sf.adaptive_mode_search) {
2461      if (single_filter[this_mode][refs[0]] ==
2462          single_filter[this_mode][refs[1]])
2463        best_filter = single_filter[this_mode][refs[0]];
2464    }
2465  }
2466
2467  if (this_mode == NEWMV) {
2468    int rate_mv;
2469    if (is_comp_pred) {
2470      // Initialize mv using single prediction mode result.
2471      frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
2472      frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
2473
2474      if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
2475        joint_motion_search(cpi, x, bsize, frame_mv,
2476                            mi_row, mi_col, single_newmv, &rate_mv);
2477      } else {
2478        rate_mv  = vp9_mv_bit_cost(&frame_mv[refs[0]].as_mv,
2479                                   &x->mbmi_ext->ref_mvs[refs[0]][0].as_mv,
2480                                   x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2481        rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]].as_mv,
2482                                   &x->mbmi_ext->ref_mvs[refs[1]][0].as_mv,
2483                                   x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2484      }
2485      *rate2 += rate_mv;
2486    } else {
2487      int_mv tmp_mv;
2488      single_motion_search(cpi, x, bsize, mi_row, mi_col,
2489                           &tmp_mv, &rate_mv);
2490      if (tmp_mv.as_int == INVALID_MV)
2491        return INT64_MAX;
2492
2493      frame_mv[refs[0]].as_int =
2494          xd->mi[0]->bmi[0].as_mv[0].as_int = tmp_mv.as_int;
2495      single_newmv[refs[0]].as_int = tmp_mv.as_int;
2496
2497      // Estimate the rate implications of a new mv but discount this
2498      // under certain circumstances where we want to help initiate a weak
2499      // motion field, where the distortion gain for a single block may not
2500      // be enough to overcome the cost of a new mv.
2501      if (discount_newmv_test(cpi, this_mode, tmp_mv, mode_mv, refs[0])) {
2502        *rate2 += VPXMAX((rate_mv / NEW_MV_DISCOUNT_FACTOR), 1);
2503      } else {
2504        *rate2 += rate_mv;
2505      }
2506    }
2507  }
2508
2509  for (i = 0; i < is_comp_pred + 1; ++i) {
2510    cur_mv[i] = frame_mv[refs[i]];
2511    // Clip "next_nearest" so that it does not extend to far out of image
2512    if (this_mode != NEWMV)
2513      clamp_mv2(&cur_mv[i].as_mv, xd);
2514
2515    if (mv_check_bounds(x, &cur_mv[i].as_mv))
2516      return INT64_MAX;
2517    mbmi->mv[i].as_int = cur_mv[i].as_int;
2518  }
2519
2520  // do first prediction into the destination buffer. Do the next
2521  // prediction into a temporary buffer. Then keep track of which one
2522  // of these currently holds the best predictor, and use the other
2523  // one for future predictions. In the end, copy from tmp_buf to
2524  // dst if necessary.
2525  for (i = 0; i < MAX_MB_PLANE; i++) {
2526    orig_dst[i] = xd->plane[i].dst.buf;
2527    orig_dst_stride[i] = xd->plane[i].dst.stride;
2528  }
2529
2530  // We don't include the cost of the second reference here, because there
2531  // are only two options: Last/ARF or Golden/ARF; The second one is always
2532  // known, which is ARF.
2533  //
2534  // Under some circumstances we discount the cost of new mv mode to encourage
2535  // initiation of a motion field.
2536  if (discount_newmv_test(cpi, this_mode, frame_mv[refs[0]],
2537                          mode_mv, refs[0])) {
2538    *rate2 += VPXMIN(cost_mv_ref(cpi, this_mode,
2539                                 mbmi_ext->mode_context[refs[0]]),
2540                     cost_mv_ref(cpi, NEARESTMV,
2541                                 mbmi_ext->mode_context[refs[0]]));
2542  } else {
2543    *rate2 += cost_mv_ref(cpi, this_mode, mbmi_ext->mode_context[refs[0]]);
2544  }
2545
2546  if (RDCOST(x->rdmult, x->rddiv, *rate2, 0) > ref_best_rd &&
2547      mbmi->mode != NEARESTMV)
2548    return INT64_MAX;
2549
2550  pred_exists = 0;
2551  // Are all MVs integer pel for Y and UV
2552  intpel_mv = !mv_has_subpel(&mbmi->mv[0].as_mv);
2553  if (is_comp_pred)
2554    intpel_mv &= !mv_has_subpel(&mbmi->mv[1].as_mv);
2555
2556  // Search for best switchable filter by checking the variance of
2557  // pred error irrespective of whether the filter will be used
2558  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
2559    filter_cache[i] = INT64_MAX;
2560
2561  if (cm->interp_filter != BILINEAR) {
2562    if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
2563      best_filter = EIGHTTAP;
2564    } else if (best_filter == SWITCHABLE) {
2565      int newbest;
2566      int tmp_rate_sum = 0;
2567      int64_t tmp_dist_sum = 0;
2568
2569      for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
2570        int j;
2571        int64_t rs_rd;
2572        int tmp_skip_sb = 0;
2573        int64_t tmp_skip_sse = INT64_MAX;
2574
2575        mbmi->interp_filter = i;
2576        rs = vp9_get_switchable_rate(cpi, xd);
2577        rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
2578
2579        if (i > 0 && intpel_mv) {
2580          rd = RDCOST(x->rdmult, x->rddiv, tmp_rate_sum, tmp_dist_sum);
2581          filter_cache[i] = rd;
2582          filter_cache[SWITCHABLE_FILTERS] =
2583              VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2584          if (cm->interp_filter == SWITCHABLE)
2585            rd += rs_rd;
2586          *mask_filter = VPXMAX(*mask_filter, rd);
2587        } else {
2588          int rate_sum = 0;
2589          int64_t dist_sum = 0;
2590          if (i > 0 && cpi->sf.adaptive_interp_filter_search &&
2591              (cpi->sf.interp_filter_search_mask & (1 << i))) {
2592            rate_sum = INT_MAX;
2593            dist_sum = INT64_MAX;
2594            continue;
2595          }
2596
2597          if ((cm->interp_filter == SWITCHABLE &&
2598               (!i || best_needs_copy)) ||
2599              (cm->interp_filter != SWITCHABLE &&
2600               (cm->interp_filter == mbmi->interp_filter ||
2601                (i == 0 && intpel_mv)))) {
2602            restore_dst_buf(xd, orig_dst, orig_dst_stride);
2603          } else {
2604            for (j = 0; j < MAX_MB_PLANE; j++) {
2605              xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
2606              xd->plane[j].dst.stride = 64;
2607            }
2608          }
2609          vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2610          model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum,
2611                          &tmp_skip_sb, &tmp_skip_sse);
2612
2613          rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum);
2614          filter_cache[i] = rd;
2615          filter_cache[SWITCHABLE_FILTERS] =
2616              VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2617          if (cm->interp_filter == SWITCHABLE)
2618            rd += rs_rd;
2619          *mask_filter = VPXMAX(*mask_filter, rd);
2620
2621          if (i == 0 && intpel_mv) {
2622            tmp_rate_sum = rate_sum;
2623            tmp_dist_sum = dist_sum;
2624          }
2625        }
2626
2627        if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2628          if (rd / 2 > ref_best_rd) {
2629            restore_dst_buf(xd, orig_dst, orig_dst_stride);
2630            return INT64_MAX;
2631          }
2632        }
2633        newbest = i == 0 || rd < best_rd;
2634
2635        if (newbest) {
2636          best_rd = rd;
2637          best_filter = mbmi->interp_filter;
2638          if (cm->interp_filter == SWITCHABLE && i && !intpel_mv)
2639            best_needs_copy = !best_needs_copy;
2640        }
2641
2642        if ((cm->interp_filter == SWITCHABLE && newbest) ||
2643            (cm->interp_filter != SWITCHABLE &&
2644             cm->interp_filter == mbmi->interp_filter)) {
2645          pred_exists = 1;
2646          tmp_rd = best_rd;
2647
2648          skip_txfm_sb = tmp_skip_sb;
2649          skip_sse_sb = tmp_skip_sse;
2650          memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2651          memcpy(bsse, x->bsse, sizeof(bsse));
2652        }
2653      }
2654      restore_dst_buf(xd, orig_dst, orig_dst_stride);
2655    }
2656  }
2657  // Set the appropriate filter
2658  mbmi->interp_filter = cm->interp_filter != SWITCHABLE ?
2659      cm->interp_filter : best_filter;
2660  rs = cm->interp_filter == SWITCHABLE ? vp9_get_switchable_rate(cpi, xd) : 0;
2661
2662  if (pred_exists) {
2663    if (best_needs_copy) {
2664      // again temporarily set the buffers to local memory to prevent a memcpy
2665      for (i = 0; i < MAX_MB_PLANE; i++) {
2666        xd->plane[i].dst.buf = tmp_buf + i * 64 * 64;
2667        xd->plane[i].dst.stride = 64;
2668      }
2669    }
2670    rd = tmp_rd + RDCOST(x->rdmult, x->rddiv, rs, 0);
2671  } else {
2672    int tmp_rate;
2673    int64_t tmp_dist;
2674    // Handles the special case when a filter that is not in the
2675    // switchable list (ex. bilinear) is indicated at the frame level, or
2676    // skip condition holds.
2677    vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2678    model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist,
2679                    &skip_txfm_sb, &skip_sse_sb);
2680    rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
2681    memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2682    memcpy(bsse, x->bsse, sizeof(bsse));
2683  }
2684
2685  if (!is_comp_pred)
2686    single_filter[this_mode][refs[0]] = mbmi->interp_filter;
2687
2688  if (cpi->sf.adaptive_mode_search)
2689    if (is_comp_pred)
2690      if (single_skippable[this_mode][refs[0]] &&
2691          single_skippable[this_mode][refs[1]])
2692        memset(skip_txfm, SKIP_TXFM_AC_DC, sizeof(skip_txfm));
2693
2694  if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2695    // if current pred_error modeled rd is substantially more than the best
2696    // so far, do not bother doing full rd
2697    if (rd / 2 > ref_best_rd) {
2698      restore_dst_buf(xd, orig_dst, orig_dst_stride);
2699      return INT64_MAX;
2700    }
2701  }
2702
2703  if (cm->interp_filter == SWITCHABLE)
2704    *rate2 += rs;
2705
2706  memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm));
2707  memcpy(x->bsse, bsse, sizeof(bsse));
2708
2709  if (!skip_txfm_sb) {
2710    int skippable_y, skippable_uv;
2711    int64_t sseuv = INT64_MAX;
2712    int64_t rdcosty = INT64_MAX;
2713
2714    // Y cost and distortion
2715    vp9_subtract_plane(x, bsize, 0);
2716    super_block_yrd(cpi, x, rate_y, &distortion_y, &skippable_y, psse,
2717                    bsize, ref_best_rd);
2718
2719    if (*rate_y == INT_MAX) {
2720      *rate2 = INT_MAX;
2721      *distortion = INT64_MAX;
2722      restore_dst_buf(xd, orig_dst, orig_dst_stride);
2723      return INT64_MAX;
2724    }
2725
2726    *rate2 += *rate_y;
2727    *distortion += distortion_y;
2728
2729    rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
2730    rdcosty = VPXMIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
2731
2732    if (!super_block_uvrd(cpi, x, rate_uv, &distortion_uv, &skippable_uv,
2733                          &sseuv, bsize, ref_best_rd - rdcosty)) {
2734      *rate2 = INT_MAX;
2735      *distortion = INT64_MAX;
2736      restore_dst_buf(xd, orig_dst, orig_dst_stride);
2737      return INT64_MAX;
2738    }
2739
2740    *psse += sseuv;
2741    *rate2 += *rate_uv;
2742    *distortion += distortion_uv;
2743    *skippable = skippable_y && skippable_uv;
2744  } else {
2745    x->skip = 1;
2746    *disable_skip = 1;
2747
2748    // The cost of skip bit needs to be added.
2749    *rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2750
2751    *distortion = skip_sse_sb;
2752  }
2753
2754  if (!is_comp_pred)
2755    single_skippable[this_mode][refs[0]] = *skippable;
2756
2757  restore_dst_buf(xd, orig_dst, orig_dst_stride);
2758  return 0;  // The rate-distortion cost will be re-calculated by caller.
2759}
2760
2761void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
2762                               RD_COST *rd_cost, BLOCK_SIZE bsize,
2763                               PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
2764  VP9_COMMON *const cm = &cpi->common;
2765  MACROBLOCKD *const xd = &x->e_mbd;
2766  struct macroblockd_plane *const pd = xd->plane;
2767  int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
2768  int y_skip = 0, uv_skip = 0;
2769  int64_t dist_y = 0, dist_uv = 0;
2770  TX_SIZE max_uv_tx_size;
2771  x->skip_encode = 0;
2772  ctx->skip = 0;
2773  xd->mi[0]->mbmi.ref_frame[0] = INTRA_FRAME;
2774  xd->mi[0]->mbmi.ref_frame[1] = NONE;
2775
2776  if (bsize >= BLOCK_8X8) {
2777    if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2778                               &dist_y, &y_skip, bsize,
2779                               best_rd) >= best_rd) {
2780      rd_cost->rate = INT_MAX;
2781      return;
2782    }
2783  } else {
2784    y_skip = 0;
2785    if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2786                                     &dist_y, best_rd) >= best_rd) {
2787      rd_cost->rate = INT_MAX;
2788      return;
2789    }
2790  }
2791  max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0]->mbmi.tx_size, bsize,
2792                                       pd[1].subsampling_x,
2793                                       pd[1].subsampling_y);
2794  rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
2795                          &dist_uv, &uv_skip, VPXMAX(BLOCK_8X8, bsize),
2796                          max_uv_tx_size);
2797
2798  if (y_skip && uv_skip) {
2799    rd_cost->rate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
2800                    vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2801    rd_cost->dist = dist_y + dist_uv;
2802  } else {
2803    rd_cost->rate = rate_y + rate_uv +
2804                      vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
2805    rd_cost->dist = dist_y + dist_uv;
2806  }
2807
2808  ctx->mic = *xd->mi[0];
2809  ctx->mbmi_ext = *x->mbmi_ext;
2810  rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
2811}
2812
2813// This function is designed to apply a bias or adjustment to an rd value based
2814// on the relative variance of the source and reconstruction.
2815#define LOW_VAR_THRESH 16
2816#define VLOW_ADJ_MAX 25
2817#define VHIGH_ADJ_MAX 8
2818static void rd_variance_adjustment(VP9_COMP *cpi,
2819                                   MACROBLOCK *x,
2820                                   BLOCK_SIZE bsize,
2821                                   int64_t *this_rd,
2822                                   MV_REFERENCE_FRAME ref_frame,
2823                                   unsigned int source_variance) {
2824  MACROBLOCKD *const xd = &x->e_mbd;
2825  unsigned int recon_variance;
2826  unsigned int absvar_diff = 0;
2827  int64_t var_error = 0;
2828  int64_t var_factor = 0;
2829
2830  if (*this_rd == INT64_MAX)
2831    return;
2832
2833#if CONFIG_VP9_HIGHBITDEPTH
2834  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2835    recon_variance =
2836      vp9_high_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize, xd->bd);
2837  } else {
2838    recon_variance =
2839      vp9_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
2840  }
2841#else
2842  recon_variance =
2843    vp9_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
2844#endif  // CONFIG_VP9_HIGHBITDEPTH
2845
2846  if ((source_variance + recon_variance) > LOW_VAR_THRESH) {
2847    absvar_diff = (source_variance > recon_variance)
2848      ? (source_variance - recon_variance)
2849      : (recon_variance - source_variance);
2850
2851    var_error = (200 * source_variance * recon_variance) /
2852      ((source_variance * source_variance) +
2853       (recon_variance * recon_variance));
2854    var_error = 100 - var_error;
2855  }
2856
2857  // Source variance above a threshold and ref frame is intra.
2858  // This case is targeted mainly at discouraging intra modes that give rise
2859  // to a predictor with a low spatial complexity compared to the source.
2860  if ((source_variance > LOW_VAR_THRESH) && (ref_frame == INTRA_FRAME) &&
2861      (source_variance > recon_variance)) {
2862    var_factor = VPXMIN(absvar_diff, VPXMIN(VLOW_ADJ_MAX, var_error));
2863  // A second possible case of interest is where the source variance
2864  // is very low and we wish to discourage false texture or motion trails.
2865  } else if ((source_variance < (LOW_VAR_THRESH >> 1)) &&
2866             (recon_variance > source_variance)) {
2867    var_factor = VPXMIN(absvar_diff, VPXMIN(VHIGH_ADJ_MAX, var_error));
2868  }
2869  *this_rd += (*this_rd * var_factor) / 100;
2870}
2871
2872
2873// Do we have an internal image edge (e.g. formatting bars).
2874int vp9_internal_image_edge(VP9_COMP *cpi) {
2875  return (cpi->oxcf.pass == 2) &&
2876    ((cpi->twopass.this_frame_stats.inactive_zone_rows > 0) ||
2877    (cpi->twopass.this_frame_stats.inactive_zone_cols > 0));
2878}
2879
2880// Checks to see if a super block is on a horizontal image edge.
2881// In most cases this is the "real" edge unless there are formatting
2882// bars embedded in the stream.
2883int vp9_active_h_edge(VP9_COMP *cpi, int mi_row, int mi_step) {
2884  int top_edge = 0;
2885  int bottom_edge = cpi->common.mi_rows;
2886  int is_active_h_edge = 0;
2887
2888  // For two pass account for any formatting bars detected.
2889  if (cpi->oxcf.pass == 2) {
2890    TWO_PASS *twopass = &cpi->twopass;
2891
2892    // The inactive region is specified in MBs not mi units.
2893    // The image edge is in the following MB row.
2894    top_edge += (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
2895
2896    bottom_edge -= (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
2897    bottom_edge = VPXMAX(top_edge, bottom_edge);
2898  }
2899
2900  if (((top_edge >= mi_row) && (top_edge < (mi_row + mi_step))) ||
2901      ((bottom_edge >= mi_row) && (bottom_edge < (mi_row + mi_step)))) {
2902    is_active_h_edge = 1;
2903  }
2904  return is_active_h_edge;
2905}
2906
2907// Checks to see if a super block is on a vertical image edge.
2908// In most cases this is the "real" edge unless there are formatting
2909// bars embedded in the stream.
2910int vp9_active_v_edge(VP9_COMP *cpi, int mi_col, int mi_step) {
2911  int left_edge = 0;
2912  int right_edge = cpi->common.mi_cols;
2913  int is_active_v_edge = 0;
2914
2915  // For two pass account for any formatting bars detected.
2916  if (cpi->oxcf.pass == 2) {
2917    TWO_PASS *twopass = &cpi->twopass;
2918
2919    // The inactive region is specified in MBs not mi units.
2920    // The image edge is in the following MB row.
2921    left_edge += (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
2922
2923    right_edge -= (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
2924    right_edge = VPXMAX(left_edge, right_edge);
2925  }
2926
2927  if (((left_edge >= mi_col) && (left_edge < (mi_col + mi_step))) ||
2928      ((right_edge >= mi_col) && (right_edge < (mi_col + mi_step)))) {
2929    is_active_v_edge = 1;
2930  }
2931  return is_active_v_edge;
2932}
2933
2934// Checks to see if a super block is at the edge of the active image.
2935// In most cases this is the "real" edge unless there are formatting
2936// bars embedded in the stream.
2937int vp9_active_edge_sb(VP9_COMP *cpi,
2938                       int mi_row, int mi_col) {
2939  return vp9_active_h_edge(cpi, mi_row, MI_BLOCK_SIZE) ||
2940         vp9_active_v_edge(cpi, mi_col, MI_BLOCK_SIZE);
2941}
2942
2943void vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi,
2944                               TileDataEnc *tile_data,
2945                               MACROBLOCK *x,
2946                               int mi_row, int mi_col,
2947                               RD_COST *rd_cost, BLOCK_SIZE bsize,
2948                               PICK_MODE_CONTEXT *ctx,
2949                               int64_t best_rd_so_far) {
2950  VP9_COMMON *const cm = &cpi->common;
2951  TileInfo *const tile_info = &tile_data->tile_info;
2952  RD_OPT *const rd_opt = &cpi->rd;
2953  SPEED_FEATURES *const sf = &cpi->sf;
2954  MACROBLOCKD *const xd = &x->e_mbd;
2955  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
2956  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2957  const struct segmentation *const seg = &cm->seg;
2958  PREDICTION_MODE this_mode;
2959  MV_REFERENCE_FRAME ref_frame, second_ref_frame;
2960  unsigned char segment_id = mbmi->segment_id;
2961  int comp_pred, i, k;
2962  int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
2963  struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2964  int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
2965  INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES];
2966  int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES];
2967  static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2968                                    VP9_ALT_FLAG };
2969  int64_t best_rd = best_rd_so_far;
2970  int64_t best_pred_diff[REFERENCE_MODES];
2971  int64_t best_pred_rd[REFERENCE_MODES];
2972  int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
2973  int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
2974  MB_MODE_INFO best_mbmode;
2975  int best_mode_skippable = 0;
2976  int midx, best_mode_index = -1;
2977  unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
2978  vpx_prob comp_mode_p;
2979  int64_t best_intra_rd = INT64_MAX;
2980  unsigned int best_pred_sse = UINT_MAX;
2981  PREDICTION_MODE best_intra_mode = DC_PRED;
2982  int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
2983  int64_t dist_uv[TX_SIZES];
2984  int skip_uv[TX_SIZES];
2985  PREDICTION_MODE mode_uv[TX_SIZES];
2986  const int intra_cost_penalty = vp9_get_intra_cost_penalty(
2987      cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
2988  int best_skip2 = 0;
2989  uint8_t ref_frame_skip_mask[2] = { 0 };
2990  uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 };
2991  int mode_skip_start = sf->mode_skip_start + 1;
2992  const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
2993  const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
2994  int64_t mode_threshold[MAX_MODES];
2995  int *mode_map = tile_data->mode_map[bsize];
2996  const int mode_search_skip_flags = sf->mode_search_skip_flags;
2997  int64_t mask_filter = 0;
2998  int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
2999
3000  vp9_zero(best_mbmode);
3001
3002  x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3003
3004  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3005    filter_cache[i] = INT64_MAX;
3006
3007  estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3008                           &comp_mode_p);
3009
3010  for (i = 0; i < REFERENCE_MODES; ++i)
3011    best_pred_rd[i] = INT64_MAX;
3012  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3013    best_filter_rd[i] = INT64_MAX;
3014  for (i = 0; i < TX_SIZES; i++)
3015    rate_uv_intra[i] = INT_MAX;
3016  for (i = 0; i < MAX_REF_FRAMES; ++i)
3017    x->pred_sse[i] = INT_MAX;
3018  for (i = 0; i < MB_MODE_COUNT; ++i) {
3019    for (k = 0; k < MAX_REF_FRAMES; ++k) {
3020      single_inter_filter[i][k] = SWITCHABLE;
3021      single_skippable[i][k] = 0;
3022    }
3023  }
3024
3025  rd_cost->rate = INT_MAX;
3026
3027  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3028    x->pred_mv_sad[ref_frame] = INT_MAX;
3029    if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3030      assert(get_ref_frame_buffer(cpi, ref_frame) != NULL);
3031      setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
3032                         frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
3033    }
3034    frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3035    frame_mv[ZEROMV][ref_frame].as_int = 0;
3036  }
3037
3038  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3039    if (!(cpi->ref_frame_flags & flag_list[ref_frame])) {
3040      // Skip checking missing references in both single and compound reference
3041      // modes. Note that a mode will be skipped iff both reference frames
3042      // are masked out.
3043      ref_frame_skip_mask[0] |= (1 << ref_frame);
3044      ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3045    } else if (sf->reference_masking) {
3046      for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
3047        // Skip fixed mv modes for poor references
3048        if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
3049          mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
3050          break;
3051        }
3052      }
3053    }
3054    // If the segment reference frame feature is enabled....
3055    // then do nothing if the current ref frame is not allowed..
3056    if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3057        get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3058      ref_frame_skip_mask[0] |= (1 << ref_frame);
3059      ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3060    }
3061  }
3062
3063  // Disable this drop out case if the ref frame
3064  // segment level feature is enabled for this segment. This is to
3065  // prevent the possibility that we end up unable to pick any mode.
3066  if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3067    // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3068    // unless ARNR filtering is enabled in which case we want
3069    // an unfiltered alternative. We allow near/nearest as well
3070    // because they may result in zero-zero MVs but be cheaper.
3071    if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
3072      ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME);
3073      ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3074      mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
3075      if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
3076        mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
3077      if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
3078        mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
3079    }
3080  }
3081
3082  if (cpi->rc.is_src_frame_alt_ref) {
3083    if (sf->alt_ref_search_fp) {
3084      mode_skip_mask[ALTREF_FRAME] = 0;
3085      ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME);
3086      ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3087    }
3088  }
3089
3090  if (sf->alt_ref_search_fp)
3091    if (!cm->show_frame && x->pred_mv_sad[GOLDEN_FRAME] < INT_MAX)
3092      if (x->pred_mv_sad[ALTREF_FRAME] > (x->pred_mv_sad[GOLDEN_FRAME] << 1))
3093        mode_skip_mask[ALTREF_FRAME] |= INTER_ALL;
3094
3095  if (sf->adaptive_mode_search) {
3096    if (cm->show_frame && !cpi->rc.is_src_frame_alt_ref &&
3097        cpi->rc.frames_since_golden >= 3)
3098      if (x->pred_mv_sad[GOLDEN_FRAME] > (x->pred_mv_sad[LAST_FRAME] << 1))
3099        mode_skip_mask[GOLDEN_FRAME] |= INTER_ALL;
3100  }
3101
3102  if (bsize > sf->max_intra_bsize) {
3103    ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
3104    ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
3105  }
3106
3107  mode_skip_mask[INTRA_FRAME] |=
3108      ~(sf->intra_y_mode_mask[max_txsize_lookup[bsize]]);
3109
3110  for (i = 0; i <= LAST_NEW_MV_INDEX; ++i)
3111    mode_threshold[i] = 0;
3112  for (i = LAST_NEW_MV_INDEX + 1; i < MAX_MODES; ++i)
3113    mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
3114
3115  midx =  sf->schedule_mode_search ? mode_skip_start : 0;
3116  while (midx > 4) {
3117    uint8_t end_pos = 0;
3118    for (i = 5; i < midx; ++i) {
3119      if (mode_threshold[mode_map[i - 1]] > mode_threshold[mode_map[i]]) {
3120        uint8_t tmp = mode_map[i];
3121        mode_map[i] = mode_map[i - 1];
3122        mode_map[i - 1] = tmp;
3123        end_pos = i;
3124      }
3125    }
3126    midx = end_pos;
3127  }
3128
3129  for (midx = 0; midx < MAX_MODES; ++midx) {
3130    int mode_index = mode_map[midx];
3131    int mode_excluded = 0;
3132    int64_t this_rd = INT64_MAX;
3133    int disable_skip = 0;
3134    int compmode_cost = 0;
3135    int rate2 = 0, rate_y = 0, rate_uv = 0;
3136    int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3137    int skippable = 0;
3138    int this_skip2 = 0;
3139    int64_t total_sse = INT64_MAX;
3140    int early_term = 0;
3141
3142    this_mode = vp9_mode_order[mode_index].mode;
3143    ref_frame = vp9_mode_order[mode_index].ref_frame[0];
3144    second_ref_frame = vp9_mode_order[mode_index].ref_frame[1];
3145
3146    // Look at the reference frame of the best mode so far and set the
3147    // skip mask to look at a subset of the remaining modes.
3148    if (midx == mode_skip_start && best_mode_index >= 0) {
3149      switch (best_mbmode.ref_frame[0]) {
3150        case INTRA_FRAME:
3151          break;
3152        case LAST_FRAME:
3153          ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK;
3154          ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3155          break;
3156        case GOLDEN_FRAME:
3157          ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
3158          ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3159          break;
3160        case ALTREF_FRAME:
3161          ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK;
3162          break;
3163        case NONE:
3164        case MAX_REF_FRAMES:
3165          assert(0 && "Invalid Reference frame");
3166          break;
3167      }
3168    }
3169
3170    if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3171        (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
3172      continue;
3173
3174    if (mode_skip_mask[ref_frame] & (1 << this_mode))
3175      continue;
3176
3177    // Test best rd so far against threshold for trying this mode.
3178    if (best_mode_skippable && sf->schedule_mode_search)
3179      mode_threshold[mode_index] <<= 1;
3180
3181    if (best_rd < mode_threshold[mode_index])
3182      continue;
3183
3184    if (sf->motion_field_mode_search) {
3185      const int mi_width  = VPXMIN(num_8x8_blocks_wide_lookup[bsize],
3186                                   tile_info->mi_col_end - mi_col);
3187      const int mi_height = VPXMIN(num_8x8_blocks_high_lookup[bsize],
3188                                   tile_info->mi_row_end - mi_row);
3189      const int bsl = mi_width_log2_lookup[bsize];
3190      int cb_partition_search_ctrl = (((mi_row + mi_col) >> bsl)
3191          + get_chessboard_index(cm->current_video_frame)) & 0x1;
3192      MB_MODE_INFO *ref_mbmi;
3193      int const_motion = 1;
3194      int skip_ref_frame = !cb_partition_search_ctrl;
3195      MV_REFERENCE_FRAME rf = NONE;
3196      int_mv ref_mv;
3197      ref_mv.as_int = INVALID_MV;
3198
3199      if ((mi_row - 1) >= tile_info->mi_row_start) {
3200        ref_mv = xd->mi[-xd->mi_stride]->mbmi.mv[0];
3201        rf = xd->mi[-xd->mi_stride]->mbmi.ref_frame[0];
3202        for (i = 0; i < mi_width; ++i) {
3203          ref_mbmi = &xd->mi[-xd->mi_stride + i]->mbmi;
3204          const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
3205                          (ref_frame == ref_mbmi->ref_frame[0]);
3206          skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
3207        }
3208      }
3209
3210      if ((mi_col - 1) >= tile_info->mi_col_start) {
3211        if (ref_mv.as_int == INVALID_MV)
3212          ref_mv = xd->mi[-1]->mbmi.mv[0];
3213        if (rf == NONE)
3214          rf = xd->mi[-1]->mbmi.ref_frame[0];
3215        for (i = 0; i < mi_height; ++i) {
3216          ref_mbmi = &xd->mi[i * xd->mi_stride - 1]->mbmi;
3217          const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
3218                          (ref_frame == ref_mbmi->ref_frame[0]);
3219          skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
3220        }
3221      }
3222
3223      if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV)
3224        if (rf > INTRA_FRAME)
3225          if (ref_frame != rf)
3226            continue;
3227
3228      if (const_motion)
3229        if (this_mode == NEARMV || this_mode == ZEROMV)
3230          continue;
3231    }
3232
3233    comp_pred = second_ref_frame > INTRA_FRAME;
3234    if (comp_pred) {
3235      if (!cpi->allow_comp_inter_inter)
3236        continue;
3237
3238      // Skip compound inter modes if ARF is not available.
3239      if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
3240        continue;
3241
3242      // Do not allow compound prediction if the segment level reference frame
3243      // feature is in use as in this case there can only be one reference.
3244      if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
3245        continue;
3246
3247      if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3248          best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
3249        continue;
3250
3251      mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3252    } else {
3253      if (ref_frame != INTRA_FRAME)
3254        mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3255    }
3256
3257    if (ref_frame == INTRA_FRAME) {
3258      if (sf->adaptive_mode_search)
3259        if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
3260          continue;
3261
3262      if (this_mode != DC_PRED) {
3263        // Disable intra modes other than DC_PRED for blocks with low variance
3264        // Threshold for intra skipping based on source variance
3265        // TODO(debargha): Specialize the threshold for super block sizes
3266        const unsigned int skip_intra_var_thresh = 64;
3267        if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
3268            x->source_variance < skip_intra_var_thresh)
3269          continue;
3270        // Only search the oblique modes if the best so far is
3271        // one of the neighboring directional modes
3272        if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
3273            (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
3274          if (best_mode_index >= 0 &&
3275              best_mbmode.ref_frame[0] > INTRA_FRAME)
3276            continue;
3277        }
3278        if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
3279          if (conditional_skipintra(this_mode, best_intra_mode))
3280              continue;
3281        }
3282      }
3283    } else {
3284      const MV_REFERENCE_FRAME ref_frames[2] = {ref_frame, second_ref_frame};
3285      if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
3286                              this_mode, ref_frames))
3287        continue;
3288    }
3289
3290    mbmi->mode = this_mode;
3291    mbmi->uv_mode = DC_PRED;
3292    mbmi->ref_frame[0] = ref_frame;
3293    mbmi->ref_frame[1] = second_ref_frame;
3294    // Evaluate all sub-pel filters irrespective of whether we can use
3295    // them for this frame.
3296    mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
3297                                                          : cm->interp_filter;
3298    mbmi->mv[0].as_int = mbmi->mv[1].as_int = 0;
3299
3300    x->skip = 0;
3301    set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3302
3303    // Select prediction reference frames.
3304    for (i = 0; i < MAX_MB_PLANE; i++) {
3305      xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3306      if (comp_pred)
3307        xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3308    }
3309
3310    if (ref_frame == INTRA_FRAME) {
3311      TX_SIZE uv_tx;
3312      struct macroblockd_plane *const pd = &xd->plane[1];
3313      memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
3314      super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable,
3315                      NULL, bsize, best_rd);
3316      if (rate_y == INT_MAX)
3317        continue;
3318
3319      uv_tx = get_uv_tx_size_impl(mbmi->tx_size, bsize, pd->subsampling_x,
3320                                  pd->subsampling_y);
3321      if (rate_uv_intra[uv_tx] == INT_MAX) {
3322        choose_intra_uv_mode(cpi, x, ctx, bsize, uv_tx,
3323                             &rate_uv_intra[uv_tx], &rate_uv_tokenonly[uv_tx],
3324                             &dist_uv[uv_tx], &skip_uv[uv_tx], &mode_uv[uv_tx]);
3325      }
3326
3327      rate_uv = rate_uv_tokenonly[uv_tx];
3328      distortion_uv = dist_uv[uv_tx];
3329      skippable = skippable && skip_uv[uv_tx];
3330      mbmi->uv_mode = mode_uv[uv_tx];
3331
3332      rate2 = rate_y + cpi->mbmode_cost[mbmi->mode] + rate_uv_intra[uv_tx];
3333      if (this_mode != DC_PRED && this_mode != TM_PRED)
3334        rate2 += intra_cost_penalty;
3335      distortion2 = distortion_y + distortion_uv;
3336    } else {
3337      this_rd = handle_inter_mode(cpi, x, bsize,
3338                                  &rate2, &distortion2, &skippable,
3339                                  &rate_y, &rate_uv,
3340                                  &disable_skip, frame_mv,
3341                                  mi_row, mi_col,
3342                                  single_newmv, single_inter_filter,
3343                                  single_skippable, &total_sse, best_rd,
3344                                  &mask_filter, filter_cache);
3345      if (this_rd == INT64_MAX)
3346        continue;
3347
3348      compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
3349
3350      if (cm->reference_mode == REFERENCE_MODE_SELECT)
3351        rate2 += compmode_cost;
3352    }
3353
3354    // Estimate the reference frame signaling cost and add it
3355    // to the rolling cost variable.
3356    if (comp_pred) {
3357      rate2 += ref_costs_comp[ref_frame];
3358    } else {
3359      rate2 += ref_costs_single[ref_frame];
3360    }
3361
3362    if (!disable_skip) {
3363      if (skippable) {
3364        // Back out the coefficient coding costs
3365        rate2 -= (rate_y + rate_uv);
3366
3367        // Cost the skip mb case
3368        rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
3369      } else if (ref_frame != INTRA_FRAME && !xd->lossless) {
3370        if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
3371            RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
3372          // Add in the cost of the no skip flag.
3373          rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
3374        } else {
3375          // FIXME(rbultje) make this work for splitmv also
3376          rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
3377          distortion2 = total_sse;
3378          assert(total_sse >= 0);
3379          rate2 -= (rate_y + rate_uv);
3380          this_skip2 = 1;
3381        }
3382      } else {
3383        // Add in the cost of the no skip flag.
3384        rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
3385      }
3386
3387      // Calculate the final RD estimate for this mode.
3388      this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3389    }
3390
3391    // Apply an adjustment to the rd value based on the similarity of the
3392    // source variance and reconstructed variance.
3393    rd_variance_adjustment(cpi, x, bsize, &this_rd,
3394                           ref_frame, x->source_variance);
3395
3396    if (ref_frame == INTRA_FRAME) {
3397    // Keep record of best intra rd
3398      if (this_rd < best_intra_rd) {
3399        best_intra_rd = this_rd;
3400        best_intra_mode = mbmi->mode;
3401      }
3402    }
3403
3404    if (!disable_skip && ref_frame == INTRA_FRAME) {
3405      for (i = 0; i < REFERENCE_MODES; ++i)
3406        best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
3407      for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3408        best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
3409    }
3410
3411    // Did this mode help.. i.e. is it the new best mode
3412    if (this_rd < best_rd || x->skip) {
3413      int max_plane = MAX_MB_PLANE;
3414      if (!mode_excluded) {
3415        // Note index of best mode so far
3416        best_mode_index = mode_index;
3417
3418        if (ref_frame == INTRA_FRAME) {
3419          /* required for left and above block mv */
3420          mbmi->mv[0].as_int = 0;
3421          max_plane = 1;
3422        } else {
3423          best_pred_sse = x->pred_sse[ref_frame];
3424        }
3425
3426        rd_cost->rate = rate2;
3427        rd_cost->dist = distortion2;
3428        rd_cost->rdcost = this_rd;
3429        best_rd = this_rd;
3430        best_mbmode = *mbmi;
3431        best_skip2 = this_skip2;
3432        best_mode_skippable = skippable;
3433
3434        if (!x->select_tx_size)
3435          swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
3436        memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size],
3437               sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
3438
3439        // TODO(debargha): enhance this test with a better distortion prediction
3440        // based on qp, activity mask and history
3441        if ((mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
3442            (mode_index > MIN_EARLY_TERM_INDEX)) {
3443          int qstep = xd->plane[0].dequant[1];
3444          // TODO(debargha): Enhance this by specializing for each mode_index
3445          int scale = 4;
3446#if CONFIG_VP9_HIGHBITDEPTH
3447          if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
3448            qstep >>= (xd->bd - 8);
3449          }
3450#endif  // CONFIG_VP9_HIGHBITDEPTH
3451          if (x->source_variance < UINT_MAX) {
3452            const int var_adjust = (x->source_variance < 16);
3453            scale -= var_adjust;
3454          }
3455          if (ref_frame > INTRA_FRAME &&
3456              distortion2 * scale < qstep * qstep) {
3457            early_term = 1;
3458          }
3459        }
3460      }
3461    }
3462
3463    /* keep record of best compound/single-only prediction */
3464    if (!disable_skip && ref_frame != INTRA_FRAME) {
3465      int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
3466
3467      if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3468        single_rate = rate2 - compmode_cost;
3469        hybrid_rate = rate2;
3470      } else {
3471        single_rate = rate2;
3472        hybrid_rate = rate2 + compmode_cost;
3473      }
3474
3475      single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
3476      hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
3477
3478      if (!comp_pred) {
3479        if (single_rd < best_pred_rd[SINGLE_REFERENCE])
3480          best_pred_rd[SINGLE_REFERENCE] = single_rd;
3481      } else {
3482        if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
3483          best_pred_rd[COMPOUND_REFERENCE] = single_rd;
3484      }
3485      if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
3486        best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
3487
3488      /* keep record of best filter type */
3489      if (!mode_excluded && cm->interp_filter != BILINEAR) {
3490        int64_t ref = filter_cache[cm->interp_filter == SWITCHABLE ?
3491                              SWITCHABLE_FILTERS : cm->interp_filter];
3492
3493        for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3494          int64_t adj_rd;
3495          if (ref == INT64_MAX)
3496            adj_rd = 0;
3497          else if (filter_cache[i] == INT64_MAX)
3498            // when early termination is triggered, the encoder does not have
3499            // access to the rate-distortion cost. it only knows that the cost
3500            // should be above the maximum valid value. hence it takes the known
3501            // maximum plus an arbitrary constant as the rate-distortion cost.
3502            adj_rd = mask_filter - ref + 10;
3503          else
3504            adj_rd = filter_cache[i] - ref;
3505
3506          adj_rd += this_rd;
3507          best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
3508        }
3509      }
3510    }
3511
3512    if (early_term)
3513      break;
3514
3515    if (x->skip && !comp_pred)
3516      break;
3517  }
3518
3519  // The inter modes' rate costs are not calculated precisely in some cases.
3520  // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and
3521  // ZEROMV. Here, checks are added for those cases, and the mode decisions
3522  // are corrected.
3523  if (best_mbmode.mode == NEWMV) {
3524    const MV_REFERENCE_FRAME refs[2] = {best_mbmode.ref_frame[0],
3525        best_mbmode.ref_frame[1]};
3526    int comp_pred_mode = refs[1] > INTRA_FRAME;
3527
3528    if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3529        ((comp_pred_mode && frame_mv[NEARESTMV][refs[1]].as_int ==
3530            best_mbmode.mv[1].as_int) || !comp_pred_mode))
3531      best_mbmode.mode = NEARESTMV;
3532    else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3533        ((comp_pred_mode && frame_mv[NEARMV][refs[1]].as_int ==
3534            best_mbmode.mv[1].as_int) || !comp_pred_mode))
3535      best_mbmode.mode = NEARMV;
3536    else if (best_mbmode.mv[0].as_int == 0 &&
3537        ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) || !comp_pred_mode))
3538      best_mbmode.mode = ZEROMV;
3539  }
3540
3541  if (best_mode_index < 0 || best_rd >= best_rd_so_far) {
3542    rd_cost->rate = INT_MAX;
3543    rd_cost->rdcost = INT64_MAX;
3544    return;
3545  }
3546
3547  // If we used an estimate for the uv intra rd in the loop above...
3548  if (sf->use_uv_intra_rd_estimate) {
3549    // Do Intra UV best rd mode selection if best mode choice above was intra.
3550    if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
3551      TX_SIZE uv_tx_size;
3552      *mbmi = best_mbmode;
3553      uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
3554      rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
3555                              &rate_uv_tokenonly[uv_tx_size],
3556                              &dist_uv[uv_tx_size],
3557                              &skip_uv[uv_tx_size],
3558                              bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
3559                              uv_tx_size);
3560    }
3561  }
3562
3563  assert((cm->interp_filter == SWITCHABLE) ||
3564         (cm->interp_filter == best_mbmode.interp_filter) ||
3565         !is_inter_block(&best_mbmode));
3566
3567  if (!cpi->rc.is_src_frame_alt_ref)
3568    vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3569                              sf->adaptive_rd_thresh, bsize, best_mode_index);
3570
3571  // macroblock modes
3572  *mbmi = best_mbmode;
3573  x->skip |= best_skip2;
3574
3575  for (i = 0; i < REFERENCE_MODES; ++i) {
3576    if (best_pred_rd[i] == INT64_MAX)
3577      best_pred_diff[i] = INT_MIN;
3578    else
3579      best_pred_diff[i] = best_rd - best_pred_rd[i];
3580  }
3581
3582  if (!x->skip) {
3583    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3584      if (best_filter_rd[i] == INT64_MAX)
3585        best_filter_diff[i] = 0;
3586      else
3587        best_filter_diff[i] = best_rd - best_filter_rd[i];
3588    }
3589    if (cm->interp_filter == SWITCHABLE)
3590      assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
3591  } else {
3592    vp9_zero(best_filter_diff);
3593  }
3594
3595  // TODO(yunqingwang): Moving this line in front of the above best_filter_diff
3596  // updating code causes PSNR loss. Need to figure out the confliction.
3597  x->skip |= best_mode_skippable;
3598
3599  if (!x->skip && !x->select_tx_size) {
3600    int has_high_freq_coeff = 0;
3601    int plane;
3602    int max_plane = is_inter_block(&xd->mi[0]->mbmi)
3603                        ? MAX_MB_PLANE : 1;
3604    for (plane = 0; plane < max_plane; ++plane) {
3605      x->plane[plane].eobs = ctx->eobs_pbuf[plane][1];
3606      has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
3607    }
3608
3609    for (plane = max_plane; plane < MAX_MB_PLANE; ++plane) {
3610      x->plane[plane].eobs = ctx->eobs_pbuf[plane][2];
3611      has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
3612    }
3613
3614    best_mode_skippable |= !has_high_freq_coeff;
3615  }
3616
3617  assert(best_mode_index >= 0);
3618
3619  store_coding_context(x, ctx, best_mode_index, best_pred_diff,
3620                       best_filter_diff, best_mode_skippable);
3621}
3622
3623void vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi,
3624                                        TileDataEnc *tile_data,
3625                                        MACROBLOCK *x,
3626                                        RD_COST *rd_cost,
3627                                        BLOCK_SIZE bsize,
3628                                        PICK_MODE_CONTEXT *ctx,
3629                                        int64_t best_rd_so_far) {
3630  VP9_COMMON *const cm = &cpi->common;
3631  MACROBLOCKD *const xd = &x->e_mbd;
3632  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
3633  unsigned char segment_id = mbmi->segment_id;
3634  const int comp_pred = 0;
3635  int i;
3636  int64_t best_pred_diff[REFERENCE_MODES];
3637  int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3638  unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3639  vpx_prob comp_mode_p;
3640  INTERP_FILTER best_filter = SWITCHABLE;
3641  int64_t this_rd = INT64_MAX;
3642  int rate2 = 0;
3643  const int64_t distortion2 = 0;
3644
3645  x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3646
3647  estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3648                           &comp_mode_p);
3649
3650  for (i = 0; i < MAX_REF_FRAMES; ++i)
3651    x->pred_sse[i] = INT_MAX;
3652  for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i)
3653    x->pred_mv_sad[i] = INT_MAX;
3654
3655  rd_cost->rate = INT_MAX;
3656
3657  assert(segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
3658
3659  mbmi->mode = ZEROMV;
3660  mbmi->uv_mode = DC_PRED;
3661  mbmi->ref_frame[0] = LAST_FRAME;
3662  mbmi->ref_frame[1] = NONE;
3663  mbmi->mv[0].as_int = 0;
3664  x->skip = 1;
3665
3666  if (cm->interp_filter != BILINEAR) {
3667    best_filter = EIGHTTAP;
3668    if (cm->interp_filter == SWITCHABLE &&
3669        x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
3670      int rs;
3671      int best_rs = INT_MAX;
3672      for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
3673        mbmi->interp_filter = i;
3674        rs = vp9_get_switchable_rate(cpi, xd);
3675        if (rs < best_rs) {
3676          best_rs = rs;
3677          best_filter = mbmi->interp_filter;
3678        }
3679      }
3680    }
3681  }
3682  // Set the appropriate filter
3683  if (cm->interp_filter == SWITCHABLE) {
3684    mbmi->interp_filter = best_filter;
3685    rate2 += vp9_get_switchable_rate(cpi, xd);
3686  } else {
3687    mbmi->interp_filter = cm->interp_filter;
3688  }
3689
3690  if (cm->reference_mode == REFERENCE_MODE_SELECT)
3691    rate2 += vp9_cost_bit(comp_mode_p, comp_pred);
3692
3693  // Estimate the reference frame signaling cost and add it
3694  // to the rolling cost variable.
3695  rate2 += ref_costs_single[LAST_FRAME];
3696  this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3697
3698  rd_cost->rate = rate2;
3699  rd_cost->dist = distortion2;
3700  rd_cost->rdcost = this_rd;
3701
3702  if (this_rd >= best_rd_so_far) {
3703    rd_cost->rate = INT_MAX;
3704    rd_cost->rdcost = INT64_MAX;
3705    return;
3706  }
3707
3708  assert((cm->interp_filter == SWITCHABLE) ||
3709         (cm->interp_filter == mbmi->interp_filter));
3710
3711  vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3712                            cpi->sf.adaptive_rd_thresh, bsize, THR_ZEROMV);
3713
3714  vp9_zero(best_pred_diff);
3715  vp9_zero(best_filter_diff);
3716
3717  if (!x->select_tx_size)
3718    swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE);
3719  store_coding_context(x, ctx, THR_ZEROMV,
3720                       best_pred_diff, best_filter_diff, 0);
3721}
3722
3723void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi,
3724                                   TileDataEnc *tile_data,
3725                                   MACROBLOCK *x,
3726                                   int mi_row, int mi_col,
3727                                   RD_COST *rd_cost,
3728                                   BLOCK_SIZE bsize,
3729                                   PICK_MODE_CONTEXT *ctx,
3730                                   int64_t best_rd_so_far) {
3731  VP9_COMMON *const cm = &cpi->common;
3732  RD_OPT *const rd_opt = &cpi->rd;
3733  SPEED_FEATURES *const sf = &cpi->sf;
3734  MACROBLOCKD *const xd = &x->e_mbd;
3735  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
3736  const struct segmentation *const seg = &cm->seg;
3737  MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3738  unsigned char segment_id = mbmi->segment_id;
3739  int comp_pred, i;
3740  int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3741  struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3742  static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
3743                                    VP9_ALT_FLAG };
3744  int64_t best_rd = best_rd_so_far;
3745  int64_t best_yrd = best_rd_so_far;  // FIXME(rbultje) more precise
3746  int64_t best_pred_diff[REFERENCE_MODES];
3747  int64_t best_pred_rd[REFERENCE_MODES];
3748  int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3749  int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3750  MB_MODE_INFO best_mbmode;
3751  int ref_index, best_ref_index = 0;
3752  unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3753  vpx_prob comp_mode_p;
3754  INTERP_FILTER tmp_best_filter = SWITCHABLE;
3755  int rate_uv_intra, rate_uv_tokenonly;
3756  int64_t dist_uv;
3757  int skip_uv;
3758  PREDICTION_MODE mode_uv = DC_PRED;
3759  const int intra_cost_penalty = vp9_get_intra_cost_penalty(
3760    cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
3761  int_mv seg_mvs[4][MAX_REF_FRAMES];
3762  b_mode_info best_bmodes[4];
3763  int best_skip2 = 0;
3764  int ref_frame_skip_mask[2] = { 0 };
3765  int64_t mask_filter = 0;
3766  int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
3767  int internal_active_edge =
3768    vp9_active_edge_sb(cpi, mi_row, mi_col) && vp9_internal_image_edge(cpi);
3769
3770  x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3771  memset(x->zcoeff_blk[TX_4X4], 0, 4);
3772  vp9_zero(best_mbmode);
3773
3774  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3775    filter_cache[i] = INT64_MAX;
3776
3777  for (i = 0; i < 4; i++) {
3778    int j;
3779    for (j = 0; j < MAX_REF_FRAMES; j++)
3780      seg_mvs[i][j].as_int = INVALID_MV;
3781  }
3782
3783  estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3784                           &comp_mode_p);
3785
3786  for (i = 0; i < REFERENCE_MODES; ++i)
3787    best_pred_rd[i] = INT64_MAX;
3788  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3789    best_filter_rd[i] = INT64_MAX;
3790  rate_uv_intra = INT_MAX;
3791
3792  rd_cost->rate = INT_MAX;
3793
3794  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
3795    if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3796      setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
3797                         frame_mv[NEARESTMV], frame_mv[NEARMV],
3798                         yv12_mb);
3799    } else {
3800      ref_frame_skip_mask[0] |= (1 << ref_frame);
3801      ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3802    }
3803    frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3804    frame_mv[ZEROMV][ref_frame].as_int = 0;
3805  }
3806
3807  for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
3808    int mode_excluded = 0;
3809    int64_t this_rd = INT64_MAX;
3810    int disable_skip = 0;
3811    int compmode_cost = 0;
3812    int rate2 = 0, rate_y = 0, rate_uv = 0;
3813    int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3814    int skippable = 0;
3815    int i;
3816    int this_skip2 = 0;
3817    int64_t total_sse = INT_MAX;
3818    int early_term = 0;
3819    struct buf_2d backup_yv12[2][MAX_MB_PLANE];
3820
3821    ref_frame = vp9_ref_order[ref_index].ref_frame[0];
3822    second_ref_frame = vp9_ref_order[ref_index].ref_frame[1];
3823
3824    // Look at the reference frame of the best mode so far and set the
3825    // skip mask to look at a subset of the remaining modes.
3826    if (ref_index > 2 && sf->mode_skip_start < MAX_MODES) {
3827      if (ref_index == 3) {
3828        switch (best_mbmode.ref_frame[0]) {
3829          case INTRA_FRAME:
3830            break;
3831          case LAST_FRAME:
3832            ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME);
3833            ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3834            break;
3835          case GOLDEN_FRAME:
3836            ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME);
3837            ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3838            break;
3839          case ALTREF_FRAME:
3840            ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME);
3841            break;
3842          case NONE:
3843          case MAX_REF_FRAMES:
3844            assert(0 && "Invalid Reference frame");
3845            break;
3846        }
3847      }
3848    }
3849
3850    if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3851        (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
3852      continue;
3853
3854    // Test best rd so far against threshold for trying this mode.
3855    if (!internal_active_edge &&
3856        rd_less_than_thresh(best_rd,
3857                            rd_opt->threshes[segment_id][bsize][ref_index],
3858                            tile_data->thresh_freq_fact[bsize][ref_index]))
3859      continue;
3860
3861    comp_pred = second_ref_frame > INTRA_FRAME;
3862    if (comp_pred) {
3863      if (!cpi->allow_comp_inter_inter)
3864        continue;
3865      if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
3866        continue;
3867      // Do not allow compound prediction if the segment level reference frame
3868      // feature is in use as in this case there can only be one reference.
3869      if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
3870        continue;
3871
3872      if ((sf->mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3873          best_mbmode.ref_frame[0] == INTRA_FRAME)
3874        continue;
3875    }
3876
3877    if (comp_pred)
3878      mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3879    else if (ref_frame != INTRA_FRAME)
3880      mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3881
3882    // If the segment reference frame feature is enabled....
3883    // then do nothing if the current ref frame is not allowed..
3884    if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3885        get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3886      continue;
3887    // Disable this drop out case if the ref frame
3888    // segment level feature is enabled for this segment. This is to
3889    // prevent the possibility that we end up unable to pick any mode.
3890    } else if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3891      // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3892      // unless ARNR filtering is enabled in which case we want
3893      // an unfiltered alternative. We allow near/nearest as well
3894      // because they may result in zero-zero MVs but be cheaper.
3895      if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
3896        continue;
3897    }
3898
3899    mbmi->tx_size = TX_4X4;
3900    mbmi->uv_mode = DC_PRED;
3901    mbmi->ref_frame[0] = ref_frame;
3902    mbmi->ref_frame[1] = second_ref_frame;
3903    // Evaluate all sub-pel filters irrespective of whether we can use
3904    // them for this frame.
3905    mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
3906                                                          : cm->interp_filter;
3907    x->skip = 0;
3908    set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3909
3910    // Select prediction reference frames.
3911    for (i = 0; i < MAX_MB_PLANE; i++) {
3912      xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3913      if (comp_pred)
3914        xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3915    }
3916
3917    if (ref_frame == INTRA_FRAME) {
3918      int rate;
3919      if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y,
3920                                       &distortion_y, best_rd) >= best_rd)
3921        continue;
3922      rate2 += rate;
3923      rate2 += intra_cost_penalty;
3924      distortion2 += distortion_y;
3925
3926      if (rate_uv_intra == INT_MAX) {
3927        choose_intra_uv_mode(cpi, x, ctx, bsize, TX_4X4,
3928                             &rate_uv_intra,
3929                             &rate_uv_tokenonly,
3930                             &dist_uv, &skip_uv,
3931                             &mode_uv);
3932      }
3933      rate2 += rate_uv_intra;
3934      rate_uv = rate_uv_tokenonly;
3935      distortion2 += dist_uv;
3936      distortion_uv = dist_uv;
3937      mbmi->uv_mode = mode_uv;
3938    } else {
3939      int rate;
3940      int64_t distortion;
3941      int64_t this_rd_thresh;
3942      int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
3943      int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
3944      int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
3945      int tmp_best_skippable = 0;
3946      int switchable_filter_index;
3947      int_mv *second_ref = comp_pred ?
3948                             &x->mbmi_ext->ref_mvs[second_ref_frame][0] : NULL;
3949      b_mode_info tmp_best_bmodes[16];
3950      MB_MODE_INFO tmp_best_mbmode;
3951      BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
3952      int pred_exists = 0;
3953      int uv_skippable;
3954
3955      YV12_BUFFER_CONFIG *scaled_ref_frame[2] = {NULL, NULL};
3956      int ref;
3957
3958      for (ref = 0; ref < 2; ++ref) {
3959        scaled_ref_frame[ref] = mbmi->ref_frame[ref] > INTRA_FRAME ?
3960            vp9_get_scaled_ref_frame(cpi, mbmi->ref_frame[ref]) : NULL;
3961
3962        if (scaled_ref_frame[ref]) {
3963          int i;
3964          // Swap out the reference frame for a version that's been scaled to
3965          // match the resolution of the current frame, allowing the existing
3966          // motion search code to be used without additional modifications.
3967          for (i = 0; i < MAX_MB_PLANE; i++)
3968            backup_yv12[ref][i] = xd->plane[i].pre[ref];
3969          vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
3970                               NULL);
3971        }
3972      }
3973
3974      this_rd_thresh = (ref_frame == LAST_FRAME) ?
3975          rd_opt->threshes[segment_id][bsize][THR_LAST] :
3976          rd_opt->threshes[segment_id][bsize][THR_ALTR];
3977      this_rd_thresh = (ref_frame == GOLDEN_FRAME) ?
3978      rd_opt->threshes[segment_id][bsize][THR_GOLD] : this_rd_thresh;
3979      for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3980        filter_cache[i] = INT64_MAX;
3981
3982      if (cm->interp_filter != BILINEAR) {
3983        tmp_best_filter = EIGHTTAP;
3984        if (x->source_variance < sf->disable_filter_search_var_thresh) {
3985          tmp_best_filter = EIGHTTAP;
3986        } else if (sf->adaptive_pred_interp_filter == 1 &&
3987                   ctx->pred_interp_filter < SWITCHABLE) {
3988          tmp_best_filter = ctx->pred_interp_filter;
3989        } else if (sf->adaptive_pred_interp_filter == 2) {
3990          tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE ?
3991                              ctx->pred_interp_filter : 0;
3992        } else {
3993          for (switchable_filter_index = 0;
3994               switchable_filter_index < SWITCHABLE_FILTERS;
3995               ++switchable_filter_index) {
3996            int newbest, rs;
3997            int64_t rs_rd;
3998            MB_MODE_INFO_EXT *mbmi_ext = x->mbmi_ext;
3999            mbmi->interp_filter = switchable_filter_index;
4000            tmp_rd = rd_pick_best_sub8x8_mode(cpi, x,
4001                                              &mbmi_ext->ref_mvs[ref_frame][0],
4002                                              second_ref, best_yrd, &rate,
4003                                              &rate_y, &distortion,
4004                                              &skippable, &total_sse,
4005                                              (int) this_rd_thresh, seg_mvs,
4006                                              bsi, switchable_filter_index,
4007                                              mi_row, mi_col);
4008
4009            if (tmp_rd == INT64_MAX)
4010              continue;
4011            rs = vp9_get_switchable_rate(cpi, xd);
4012            rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
4013            filter_cache[switchable_filter_index] = tmp_rd;
4014            filter_cache[SWITCHABLE_FILTERS] =
4015                VPXMIN(filter_cache[SWITCHABLE_FILTERS], tmp_rd + rs_rd);
4016            if (cm->interp_filter == SWITCHABLE)
4017              tmp_rd += rs_rd;
4018
4019            mask_filter = VPXMAX(mask_filter, tmp_rd);
4020
4021            newbest = (tmp_rd < tmp_best_rd);
4022            if (newbest) {
4023              tmp_best_filter = mbmi->interp_filter;
4024              tmp_best_rd = tmp_rd;
4025            }
4026            if ((newbest && cm->interp_filter == SWITCHABLE) ||
4027                (mbmi->interp_filter == cm->interp_filter &&
4028                 cm->interp_filter != SWITCHABLE)) {
4029              tmp_best_rdu = tmp_rd;
4030              tmp_best_rate = rate;
4031              tmp_best_ratey = rate_y;
4032              tmp_best_distortion = distortion;
4033              tmp_best_sse = total_sse;
4034              tmp_best_skippable = skippable;
4035              tmp_best_mbmode = *mbmi;
4036              for (i = 0; i < 4; i++) {
4037                tmp_best_bmodes[i] = xd->mi[0]->bmi[i];
4038                x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
4039              }
4040              pred_exists = 1;
4041              if (switchable_filter_index == 0 &&
4042                  sf->use_rd_breakout &&
4043                  best_rd < INT64_MAX) {
4044                if (tmp_best_rdu / 2 > best_rd) {
4045                  // skip searching the other filters if the first is
4046                  // already substantially larger than the best so far
4047                  tmp_best_filter = mbmi->interp_filter;
4048                  tmp_best_rdu = INT64_MAX;
4049                  break;
4050                }
4051              }
4052            }
4053          }  // switchable_filter_index loop
4054        }
4055      }
4056
4057      if (tmp_best_rdu == INT64_MAX && pred_exists)
4058        continue;
4059
4060      mbmi->interp_filter = (cm->interp_filter == SWITCHABLE ?
4061                             tmp_best_filter : cm->interp_filter);
4062      if (!pred_exists) {
4063        // Handles the special case when a filter that is not in the
4064        // switchable list (bilinear, 6-tap) is indicated at the frame level
4065        tmp_rd = rd_pick_best_sub8x8_mode(cpi, x,
4066                                          &x->mbmi_ext->ref_mvs[ref_frame][0],
4067                                          second_ref, best_yrd, &rate, &rate_y,
4068                                          &distortion, &skippable, &total_sse,
4069                                          (int) this_rd_thresh, seg_mvs, bsi, 0,
4070                                          mi_row, mi_col);
4071        if (tmp_rd == INT64_MAX)
4072          continue;
4073      } else {
4074        total_sse = tmp_best_sse;
4075        rate = tmp_best_rate;
4076        rate_y = tmp_best_ratey;
4077        distortion = tmp_best_distortion;
4078        skippable = tmp_best_skippable;
4079        *mbmi = tmp_best_mbmode;
4080        for (i = 0; i < 4; i++)
4081          xd->mi[0]->bmi[i] = tmp_best_bmodes[i];
4082      }
4083
4084      rate2 += rate;
4085      distortion2 += distortion;
4086
4087      if (cm->interp_filter == SWITCHABLE)
4088        rate2 += vp9_get_switchable_rate(cpi, xd);
4089
4090      if (!mode_excluded)
4091        mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
4092                                  : cm->reference_mode == COMPOUND_REFERENCE;
4093
4094      compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
4095
4096      tmp_best_rdu =
4097          best_rd - VPXMIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
4098                           RDCOST(x->rdmult, x->rddiv, 0, total_sse));
4099
4100      if (tmp_best_rdu > 0) {
4101        // If even the 'Y' rd value of split is higher than best so far
4102        // then dont bother looking at UV
4103        vp9_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col,
4104                                        BLOCK_8X8);
4105        memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
4106        if (!super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
4107                              &uv_sse, BLOCK_8X8, tmp_best_rdu)) {
4108          for (ref = 0; ref < 2; ++ref) {
4109            if (scaled_ref_frame[ref]) {
4110              int i;
4111              for (i = 0; i < MAX_MB_PLANE; ++i)
4112                xd->plane[i].pre[ref] = backup_yv12[ref][i];
4113            }
4114          }
4115          continue;
4116        }
4117
4118        rate2 += rate_uv;
4119        distortion2 += distortion_uv;
4120        skippable = skippable && uv_skippable;
4121        total_sse += uv_sse;
4122      }
4123
4124      for (ref = 0; ref < 2; ++ref) {
4125        if (scaled_ref_frame[ref]) {
4126          // Restore the prediction frame pointers to their unscaled versions.
4127          int i;
4128          for (i = 0; i < MAX_MB_PLANE; ++i)
4129            xd->plane[i].pre[ref] = backup_yv12[ref][i];
4130        }
4131      }
4132    }
4133
4134    if (cm->reference_mode == REFERENCE_MODE_SELECT)
4135      rate2 += compmode_cost;
4136
4137    // Estimate the reference frame signaling cost and add it
4138    // to the rolling cost variable.
4139    if (second_ref_frame > INTRA_FRAME) {
4140      rate2 += ref_costs_comp[ref_frame];
4141    } else {
4142      rate2 += ref_costs_single[ref_frame];
4143    }
4144
4145    if (!disable_skip) {
4146      // Skip is never coded at the segment level for sub8x8 blocks and instead
4147      // always coded in the bitstream at the mode info level.
4148
4149      if (ref_frame != INTRA_FRAME && !xd->lossless) {
4150        if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
4151            RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
4152          // Add in the cost of the no skip flag.
4153          rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
4154        } else {
4155          // FIXME(rbultje) make this work for splitmv also
4156          rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
4157          distortion2 = total_sse;
4158          assert(total_sse >= 0);
4159          rate2 -= (rate_y + rate_uv);
4160          rate_y = 0;
4161          rate_uv = 0;
4162          this_skip2 = 1;
4163        }
4164      } else {
4165        // Add in the cost of the no skip flag.
4166        rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
4167      }
4168
4169      // Calculate the final RD estimate for this mode.
4170      this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
4171    }
4172
4173    if (!disable_skip && ref_frame == INTRA_FRAME) {
4174      for (i = 0; i < REFERENCE_MODES; ++i)
4175        best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
4176      for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
4177        best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
4178    }
4179
4180    // Did this mode help.. i.e. is it the new best mode
4181    if (this_rd < best_rd || x->skip) {
4182      if (!mode_excluded) {
4183        int max_plane = MAX_MB_PLANE;
4184        // Note index of best mode so far
4185        best_ref_index = ref_index;
4186
4187        if (ref_frame == INTRA_FRAME) {
4188          /* required for left and above block mv */
4189          mbmi->mv[0].as_int = 0;
4190          max_plane = 1;
4191        }
4192
4193        rd_cost->rate = rate2;
4194        rd_cost->dist = distortion2;
4195        rd_cost->rdcost = this_rd;
4196        best_rd = this_rd;
4197        best_yrd = best_rd -
4198                   RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
4199        best_mbmode = *mbmi;
4200        best_skip2 = this_skip2;
4201        if (!x->select_tx_size)
4202          swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
4203        memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
4204               sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
4205
4206        for (i = 0; i < 4; i++)
4207          best_bmodes[i] = xd->mi[0]->bmi[i];
4208
4209        // TODO(debargha): enhance this test with a better distortion prediction
4210        // based on qp, activity mask and history
4211        if ((sf->mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
4212            (ref_index > MIN_EARLY_TERM_INDEX)) {
4213          int qstep = xd->plane[0].dequant[1];
4214          // TODO(debargha): Enhance this by specializing for each mode_index
4215          int scale = 4;
4216#if CONFIG_VP9_HIGHBITDEPTH
4217          if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
4218            qstep >>= (xd->bd - 8);
4219          }
4220#endif  // CONFIG_VP9_HIGHBITDEPTH
4221          if (x->source_variance < UINT_MAX) {
4222            const int var_adjust = (x->source_variance < 16);
4223            scale -= var_adjust;
4224          }
4225          if (ref_frame > INTRA_FRAME &&
4226              distortion2 * scale < qstep * qstep) {
4227            early_term = 1;
4228          }
4229        }
4230      }
4231    }
4232
4233    /* keep record of best compound/single-only prediction */
4234    if (!disable_skip && ref_frame != INTRA_FRAME) {
4235      int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
4236
4237      if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4238        single_rate = rate2 - compmode_cost;
4239        hybrid_rate = rate2;
4240      } else {
4241        single_rate = rate2;
4242        hybrid_rate = rate2 + compmode_cost;
4243      }
4244
4245      single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
4246      hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
4247
4248      if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
4249        best_pred_rd[SINGLE_REFERENCE] = single_rd;
4250      else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
4251        best_pred_rd[COMPOUND_REFERENCE] = single_rd;
4252
4253      if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
4254        best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
4255    }
4256
4257    /* keep record of best filter type */
4258    if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
4259        cm->interp_filter != BILINEAR) {
4260      int64_t ref = filter_cache[cm->interp_filter == SWITCHABLE ?
4261                              SWITCHABLE_FILTERS : cm->interp_filter];
4262      int64_t adj_rd;
4263      for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4264        if (ref == INT64_MAX)
4265          adj_rd = 0;
4266        else if (filter_cache[i] == INT64_MAX)
4267          // when early termination is triggered, the encoder does not have
4268          // access to the rate-distortion cost. it only knows that the cost
4269          // should be above the maximum valid value. hence it takes the known
4270          // maximum plus an arbitrary constant as the rate-distortion cost.
4271          adj_rd = mask_filter - ref + 10;
4272        else
4273          adj_rd = filter_cache[i] - ref;
4274
4275        adj_rd += this_rd;
4276        best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
4277      }
4278    }
4279
4280    if (early_term)
4281      break;
4282
4283    if (x->skip && !comp_pred)
4284      break;
4285  }
4286
4287  if (best_rd >= best_rd_so_far) {
4288    rd_cost->rate = INT_MAX;
4289    rd_cost->rdcost = INT64_MAX;
4290    return;
4291  }
4292
4293  // If we used an estimate for the uv intra rd in the loop above...
4294  if (sf->use_uv_intra_rd_estimate) {
4295    // Do Intra UV best rd mode selection if best mode choice above was intra.
4296    if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
4297      *mbmi = best_mbmode;
4298      rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra,
4299                              &rate_uv_tokenonly,
4300                              &dist_uv,
4301                              &skip_uv,
4302                              BLOCK_8X8, TX_4X4);
4303    }
4304  }
4305
4306  if (best_rd == INT64_MAX) {
4307    rd_cost->rate = INT_MAX;
4308    rd_cost->dist = INT64_MAX;
4309    rd_cost->rdcost = INT64_MAX;
4310    return;
4311  }
4312
4313  assert((cm->interp_filter == SWITCHABLE) ||
4314         (cm->interp_filter == best_mbmode.interp_filter) ||
4315         !is_inter_block(&best_mbmode));
4316
4317  vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
4318                            sf->adaptive_rd_thresh, bsize, best_ref_index);
4319
4320  // macroblock modes
4321  *mbmi = best_mbmode;
4322  x->skip |= best_skip2;
4323  if (!is_inter_block(&best_mbmode)) {
4324    for (i = 0; i < 4; i++)
4325      xd->mi[0]->bmi[i].as_mode = best_bmodes[i].as_mode;
4326  } else {
4327    for (i = 0; i < 4; ++i)
4328      memcpy(&xd->mi[0]->bmi[i], &best_bmodes[i], sizeof(b_mode_info));
4329
4330    mbmi->mv[0].as_int = xd->mi[0]->bmi[3].as_mv[0].as_int;
4331    mbmi->mv[1].as_int = xd->mi[0]->bmi[3].as_mv[1].as_int;
4332  }
4333
4334  for (i = 0; i < REFERENCE_MODES; ++i) {
4335    if (best_pred_rd[i] == INT64_MAX)
4336      best_pred_diff[i] = INT_MIN;
4337    else
4338      best_pred_diff[i] = best_rd - best_pred_rd[i];
4339  }
4340
4341  if (!x->skip) {
4342    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4343      if (best_filter_rd[i] == INT64_MAX)
4344        best_filter_diff[i] = 0;
4345      else
4346        best_filter_diff[i] = best_rd - best_filter_rd[i];
4347    }
4348    if (cm->interp_filter == SWITCHABLE)
4349      assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
4350  } else {
4351    vp9_zero(best_filter_diff);
4352  }
4353
4354  store_coding_context(x, ctx, best_ref_index,
4355                       best_pred_diff, best_filter_diff, 0);
4356}
4357