vp9_rd.c 23.8 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
/*
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include <assert.h>
#include <math.h>
#include <stdio.h>

#include "./vp9_rtcd.h"

#include "vpx_mem/vpx_mem.h"
18
#include "vpx_ports/bitops.h"
19
#include "vpx_ports/mem.h"
20
#include "vpx_ports/system_state.h"
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

#include "vp9/common/vp9_common.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_entropymode.h"
#include "vp9/common/vp9_mvref_common.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_quant_common.h"
#include "vp9/common/vp9_reconinter.h"
#include "vp9/common/vp9_reconintra.h"
#include "vp9/common/vp9_seg_common.h"

#include "vp9/encoder/vp9_cost.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/encoder/vp9_encoder.h"
#include "vp9/encoder/vp9_mcomp.h"
#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_ratectrl.h"
#include "vp9/encoder/vp9_rd.h"
#include "vp9/encoder/vp9_tokenize.h"

#define RD_THRESH_POW      1.25
#define RD_MULT_EPB_RATIO  64

Alex Converse's avatar
Alex Converse committed
45
// Factor to weigh the rate for switchable interp filters.
46
47
#define SWITCHABLE_INTERP_RATE_FACTOR 1

48
49
50
51
52
53
54
55
56
57
58
59
void vp9_rd_cost_reset(RD_COST *rd_cost) {
  rd_cost->rate = INT_MAX;
  rd_cost->dist = INT64_MAX;
  rd_cost->rdcost = INT64_MAX;
}

void vp9_rd_cost_init(RD_COST *rd_cost) {
  rd_cost->rate = 0;
  rd_cost->dist = 0;
  rd_cost->rdcost = 0;
}

60
61
// The baseline rd thresholds for breaking out of the rd loop for
// certain modes are assumed to be based on 8x8 blocks.
Alex Converse's avatar
Alex Converse committed
62
// This table is used to correct for block size.
63
64
65
66
67
68
// The factors here are << 2 (2 = x0.5, 32 = x8 etc).
static const uint8_t rd_thresh_block_size_factor[BLOCK_SIZES] = {
  2, 3, 3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32
};

static void fill_mode_costs(VP9_COMP *cpi) {
69
  const FRAME_CONTEXT *const fc = cpi->common.fc;
70
71
  int i, j;

Alex Converse's avatar
Alex Converse committed
72
73
  for (i = 0; i < INTRA_MODES; ++i)
    for (j = 0; j < INTRA_MODES; ++j)
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
      vp9_cost_tokens(cpi->y_mode_costs[i][j], vp9_kf_y_mode_prob[i][j],
                      vp9_intra_mode_tree);

  vp9_cost_tokens(cpi->mbmode_cost, fc->y_mode_prob[1], vp9_intra_mode_tree);
  vp9_cost_tokens(cpi->intra_uv_mode_cost[KEY_FRAME],
                  vp9_kf_uv_mode_prob[TM_PRED], vp9_intra_mode_tree);
  vp9_cost_tokens(cpi->intra_uv_mode_cost[INTER_FRAME],
                  fc->uv_mode_prob[TM_PRED], vp9_intra_mode_tree);

  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
    vp9_cost_tokens(cpi->switchable_interp_costs[i],
                    fc->switchable_interp_prob[i], vp9_switchable_interp_tree);
}

static void fill_token_costs(vp9_coeff_cost *c,
                             vp9_coeff_probs_model (*p)[PLANE_TYPES]) {
  int i, j, k, l;
  TX_SIZE t;
  for (t = TX_4X4; t <= TX_32X32; ++t)
    for (i = 0; i < PLANE_TYPES; ++i)
      for (j = 0; j < REF_TYPES; ++j)
        for (k = 0; k < COEF_BANDS; ++k)
          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
Yaowu Xu's avatar
Yaowu Xu committed
97
            vpx_prob probs[ENTROPY_NODES];
98
99
100
101
102
103
104
105
106
107
            vp9_model_to_full_probs(p[t][i][j][k][l], probs);
            vp9_cost_tokens((int *)c[t][i][j][k][0][l], probs,
                            vp9_coef_tree);
            vp9_cost_tokens_skip((int *)c[t][i][j][k][1][l], probs,
                                 vp9_coef_tree);
            assert(c[t][i][j][k][0][l][EOB_TOKEN] ==
                   c[t][i][j][k][1][l][EOB_TOKEN]);
          }
}

Alex Converse's avatar
Alex Converse committed
108
// Values are now correlated to quantizer.
109
110
111
112
113
114
115
116
117
118
119
120
static int sad_per_bit16lut_8[QINDEX_RANGE];
static int sad_per_bit4lut_8[QINDEX_RANGE];

#if CONFIG_VP9_HIGHBITDEPTH
static int sad_per_bit16lut_10[QINDEX_RANGE];
static int sad_per_bit4lut_10[QINDEX_RANGE];
static int sad_per_bit16lut_12[QINDEX_RANGE];
static int sad_per_bit4lut_12[QINDEX_RANGE];
#endif

static void init_me_luts_bd(int *bit16lut, int *bit4lut, int range,
                            vpx_bit_depth_t bit_depth) {
121
  int i;
Alex Converse's avatar
Alex Converse committed
122
  // Initialize the sad lut tables using a formulaic calculation for now.
123
124
  // This is to make it easier to resolve the impact of experimental changes
  // to the quantizer tables.
125
126
127
128
  for (i = 0; i < range; i++) {
    const double q = vp9_convert_qindex_to_q(i, bit_depth);
    bit16lut[i] = (int)(0.0418 * q + 2.4107);
    bit4lut[i] = (int)(0.063 * q + 2.742);
129
130
131
  }
}

132
void vp9_init_me_luts(void) {
133
134
135
136
137
138
139
140
141
142
  init_me_luts_bd(sad_per_bit16lut_8, sad_per_bit4lut_8, QINDEX_RANGE,
                  VPX_BITS_8);
#if CONFIG_VP9_HIGHBITDEPTH
  init_me_luts_bd(sad_per_bit16lut_10, sad_per_bit4lut_10, QINDEX_RANGE,
                  VPX_BITS_10);
  init_me_luts_bd(sad_per_bit16lut_12, sad_per_bit4lut_12, QINDEX_RANGE,
                  VPX_BITS_12);
#endif
}

143
144
145
146
147
static const int rd_boost_factor[16] = {
  64, 32, 32, 32, 24, 16, 12, 12,
  8, 8, 4, 4, 2, 2, 1, 0
};
static const int rd_frame_type_factor[FRAME_UPDATE_TYPES] = {
148
  128, 144, 128, 128, 144
149
150
};

151
int vp9_compute_rd_mult(const VP9_COMP *cpi, int qindex) {
152
  const int64_t q = vp9_dc_quant(qindex, 0, cpi->common.bit_depth);
153
#if CONFIG_VP9_HIGHBITDEPTH
Deb Mukherjee's avatar
Deb Mukherjee committed
154
  int64_t rdmult = 0;
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
  switch (cpi->common.bit_depth) {
    case VPX_BITS_8:
      rdmult = 88 * q * q / 24;
      break;
    case VPX_BITS_10:
      rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 4);
      break;
    case VPX_BITS_12:
      rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 8);
      break;
    default:
      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
      return -1;
  }
#else
Deb Mukherjee's avatar
Deb Mukherjee committed
170
  int64_t rdmult = 88 * q * q / 24;
171
#endif  // CONFIG_VP9_HIGHBITDEPTH
172
  if (cpi->oxcf.pass == 2 && (cpi->common.frame_type != KEY_FRAME)) {
173
174
175
176
177
178
    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
    const FRAME_UPDATE_TYPE frame_type = gf_group->update_type[gf_group->index];
    const int boost_index = MIN(15, (cpi->rc.gfu_boost / 100));

    rdmult = (rdmult * rd_frame_type_factor[frame_type]) >> 7;
    rdmult += ((rdmult * rd_boost_factor[boost_index]) >> 7);
179
  }
hui su's avatar
hui su committed
180
181
  if (rdmult < 1)
    rdmult = 1;
Deb Mukherjee's avatar
Deb Mukherjee committed
182
  return (int)rdmult;
183
184
}

185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
static int compute_rd_thresh_factor(int qindex, vpx_bit_depth_t bit_depth) {
  double q;
#if CONFIG_VP9_HIGHBITDEPTH
  switch (bit_depth) {
    case VPX_BITS_8:
      q = vp9_dc_quant(qindex, 0, VPX_BITS_8) / 4.0;
      break;
    case VPX_BITS_10:
      q = vp9_dc_quant(qindex, 0, VPX_BITS_10) / 16.0;
      break;
    case VPX_BITS_12:
      q = vp9_dc_quant(qindex, 0, VPX_BITS_12) / 64.0;
      break;
    default:
      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
      return -1;
  }
#else
  (void) bit_depth;
  q = vp9_dc_quant(qindex, 0, VPX_BITS_8) / 4.0;
205
#endif  // CONFIG_VP9_HIGHBITDEPTH
Alex Converse's avatar
Alex Converse committed
206
  // TODO(debargha): Adjust the function below.
207
  return MAX((int)(pow(q, RD_THRESH_POW) * 5.12), 8);
208
209
}

210
void vp9_initialize_me_consts(VP9_COMP *cpi, MACROBLOCK *x, int qindex) {
211
212
213
#if CONFIG_VP9_HIGHBITDEPTH
  switch (cpi->common.bit_depth) {
    case VPX_BITS_8:
214
215
      x->sadperbit16 = sad_per_bit16lut_8[qindex];
      x->sadperbit4 = sad_per_bit4lut_8[qindex];
216
217
      break;
    case VPX_BITS_10:
218
219
      x->sadperbit16 = sad_per_bit16lut_10[qindex];
      x->sadperbit4 = sad_per_bit4lut_10[qindex];
220
221
      break;
    case VPX_BITS_12:
222
223
      x->sadperbit16 = sad_per_bit16lut_12[qindex];
      x->sadperbit4 = sad_per_bit4lut_12[qindex];
224
225
226
227
228
      break;
    default:
      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
  }
#else
229
230
231
  (void)cpi;
  x->sadperbit16 = sad_per_bit16lut_8[qindex];
  x->sadperbit4 = sad_per_bit4lut_8[qindex];
232
#endif  // CONFIG_VP9_HIGHBITDEPTH
233
234
235
236
237
238
}

static void set_block_thresholds(const VP9_COMMON *cm, RD_OPT *rd) {
  int i, bsize, segment_id;

  for (segment_id = 0; segment_id < MAX_SEGMENTS; ++segment_id) {
Alex Converse's avatar
Alex Converse committed
239
240
    const int qindex =
        clamp(vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex) +
241
242
              cm->y_dc_delta_q, 0, MAXQ);
    const int q = compute_rd_thresh_factor(qindex, cm->bit_depth);
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268

    for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) {
      // Threshold here seems unnecessarily harsh but fine given actual
      // range of values used for cpi->sf.thresh_mult[].
      const int t = q * rd_thresh_block_size_factor[bsize];
      const int thresh_max = INT_MAX / t;

      if (bsize >= BLOCK_8X8) {
        for (i = 0; i < MAX_MODES; ++i)
          rd->threshes[segment_id][bsize][i] =
              rd->thresh_mult[i] < thresh_max
                  ? rd->thresh_mult[i] * t / 4
                  : INT_MAX;
      } else {
        for (i = 0; i < MAX_REFS; ++i)
          rd->threshes[segment_id][bsize][i] =
              rd->thresh_mult_sub8x8[i] < thresh_max
                  ? rd->thresh_mult_sub8x8[i] * t / 4
                  : INT_MAX;
      }
    }
  }
}

void vp9_initialize_rd_consts(VP9_COMP *cpi) {
  VP9_COMMON *const cm = &cpi->common;
269
  MACROBLOCK *const x = &cpi->td.mb;
270
  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
271
272
273
  RD_OPT *const rd = &cpi->rd;
  int i;

274
  vpx_clear_system_state();
275

Alex Converse's avatar
Alex Converse committed
276
  rd->RDDIV = RDDIV_BITS;  // In bits (to multiply D by 128).
277
278
279
280
281
282
283
284
285
  rd->RDMULT = vp9_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q);

  x->errorperbit = rd->RDMULT / RD_MULT_EPB_RATIO;
  x->errorperbit += (x->errorperbit == 0);

  x->select_tx_size = (cpi->sf.tx_size_search_method == USE_LARGESTALL &&
                       cm->frame_type != KEY_FRAME) ? 0 : 1;

  set_block_thresholds(cm, rd);
286
  set_partition_probs(cm, xd);
287

288
  if (!cpi->sf.use_nonrd_pick_mode || cm->frame_type == KEY_FRAME)
289
    fill_token_costs(x->token_costs, cm->fc->coef_probs);
290

291
292
  if (cpi->sf.partition_search_type != VAR_BASED_PARTITION ||
      cm->frame_type == KEY_FRAME) {
Alex Converse's avatar
Alex Converse committed
293
    for (i = 0; i < PARTITION_CONTEXTS; ++i)
294
      vp9_cost_tokens(cpi->partition_cost[i], get_partition_probs(xd, i),
295
296
297
298
299
300
301
302
303
304
305
                      vp9_partition_tree);
  }

  if (!cpi->sf.use_nonrd_pick_mode || (cm->current_video_frame & 0x07) == 1 ||
      cm->frame_type == KEY_FRAME) {
    fill_mode_costs(cpi);

    if (!frame_is_intra_only(cm)) {
      vp9_build_nmv_cost_table(x->nmvjointcost,
                               cm->allow_high_precision_mv ? x->nmvcost_hp
                                                           : x->nmvcost,
306
                               &cm->fc->nmvc, cm->allow_high_precision_mv);
307
308
309

      for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
        vp9_cost_tokens((int *)cpi->inter_mode_cost[i],
310
                        cm->fc->inter_mode_probs[i], vp9_inter_mode_tree);
311
312
313
314
315
    }
  }
}

static void model_rd_norm(int xsq_q10, int *r_q10, int *d_q10) {
Alex Converse's avatar
Alex Converse committed
316
  // NOTE: The tables below must be of the same size.
317
318

  // The functions described below are sampled at the four most significant
Alex Converse's avatar
Alex Converse committed
319
  // bits of x^2 + 8 / 256.
320

Alex Converse's avatar
Alex Converse committed
321
322
323
324
  // Normalized rate:
  // This table models the rate for a Laplacian source with given variance
  // when quantized with a uniform quantizer with given stepsize. The
  // closed form expression is:
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
  // Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)],
  // where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance),
  // and H(x) is the binary entropy function.
  static const int rate_tab_q10[] = {
    65536,  6086,  5574,  5275,  5063,  4899,  4764,  4651,
     4553,  4389,  4255,  4142,  4044,  3958,  3881,  3811,
     3748,  3635,  3538,  3453,  3376,  3307,  3244,  3186,
     3133,  3037,  2952,  2877,  2809,  2747,  2690,  2638,
     2589,  2501,  2423,  2353,  2290,  2232,  2179,  2130,
     2084,  2001,  1928,  1862,  1802,  1748,  1698,  1651,
     1608,  1530,  1460,  1398,  1342,  1290,  1243,  1199,
     1159,  1086,  1021,   963,   911,   864,   821,   781,
      745,   680,   623,   574,   530,   490,   455,   424,
      395,   345,   304,   269,   239,   213,   190,   171,
      154,   126,   104,    87,    73,    61,    52,    44,
       38,    28,    21,    16,    12,    10,     8,     6,
        5,     3,     2,     1,     1,     1,     0,     0,
  };
Alex Converse's avatar
Alex Converse committed
343
  // Normalized distortion:
344
  // This table models the normalized distortion for a Laplacian source
Alex Converse's avatar
Alex Converse committed
345
  // with given variance when quantized with a uniform quantizer
346
347
  // with given stepsize. The closed form expression is:
  // Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2))
Alex Converse's avatar
Alex Converse committed
348
  // where x = qpstep / sqrt(variance).
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
  // Note the actual distortion is Dn * variance.
  static const int dist_tab_q10[] = {
       0,     0,     1,     1,     1,     2,     2,     2,
       3,     3,     4,     5,     5,     6,     7,     7,
       8,     9,    11,    12,    13,    15,    16,    17,
      18,    21,    24,    26,    29,    31,    34,    36,
      39,    44,    49,    54,    59,    64,    69,    73,
      78,    88,    97,   106,   115,   124,   133,   142,
     151,   167,   184,   200,   215,   231,   245,   260,
     274,   301,   327,   351,   375,   397,   418,   439,
     458,   495,   528,   559,   587,   613,   637,   659,
     680,   717,   749,   777,   801,   823,   842,   859,
     874,   899,   919,   936,   949,   960,   969,   977,
     983,   994,  1001,  1006,  1010,  1013,  1015,  1017,
    1018,  1020,  1022,  1022,  1023,  1023,  1023,  1024,
  };
  static const int xsq_iq_q10[] = {
         0,      4,      8,     12,     16,     20,     24,     28,
        32,     40,     48,     56,     64,     72,     80,     88,
        96,    112,    128,    144,    160,    176,    192,    208,
       224,    256,    288,    320,    352,    384,    416,    448,
       480,    544,    608,    672,    736,    800,    864,    928,
       992,   1120,   1248,   1376,   1504,   1632,   1760,   1888,
      2016,   2272,   2528,   2784,   3040,   3296,   3552,   3808,
      4064,   4576,   5088,   5600,   6112,   6624,   7136,   7648,
      8160,   9184,  10208,  11232,  12256,  13280,  14304,  15328,
     16352,  18400,  20448,  22496,  24544,  26592,  28640,  30688,
     32736,  36832,  40928,  45024,  49120,  53216,  57312,  61408,
     65504,  73696,  81888,  90080,  98272, 106464, 114656, 122848,
    131040, 147424, 163808, 180192, 196576, 212960, 229344, 245728,
  };
Alex Converse's avatar
Alex Converse committed
380
381
382
  const int tmp = (xsq_q10 >> 2) + 8;
  const int k = get_msb(tmp) - 3;
  const int xq = (k << 3) + ((tmp >> k) & 0x7);
383
384
385
386
387
388
389
  const int one_q10 = 1 << 10;
  const int a_q10 = ((xsq_q10 - xsq_iq_q10[xq]) << 10) >> (2 + k);
  const int b_q10 = one_q10 - a_q10;
  *r_q10 = (rate_tab_q10[xq] * b_q10 + rate_tab_q10[xq + 1] * a_q10) >> 10;
  *d_q10 = (dist_tab_q10[xq] * b_q10 + dist_tab_q10[xq + 1] * a_q10) >> 10;
}

390
void vp9_model_rd_from_var_lapndz(unsigned int var, unsigned int n_log2,
391
392
393
394
395
396
397
398
399
400
401
402
403
                                  unsigned int qstep, int *rate,
                                  int64_t *dist) {
  // This function models the rate and distortion for a Laplacian
  // source with given variance when quantized with a uniform quantizer
  // with given stepsize. The closed form expressions are in:
  // Hang and Chen, "Source Model for transform video coder and its
  // application - Part I: Fundamental Theory", IEEE Trans. Circ.
  // Sys. for Video Tech., April 1997.
  if (var == 0) {
    *rate = 0;
    *dist = 0;
  } else {
    int d_q10, r_q10;
404
    static const uint32_t MAX_XSQ_Q10 = 245727;
405
    const uint64_t xsq_q10_64 =
406
        (((uint64_t)qstep * qstep << (n_log2 + 10)) + (var >> 1)) / var;
407
    const int xsq_q10 = (int)MIN(xsq_q10_64, MAX_XSQ_Q10);
408
    model_rd_norm(xsq_q10, &r_q10, &d_q10);
409
    *rate = ((r_q10 << n_log2) + 2) >> 2;
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
    *dist = (var * (int64_t)d_q10 + 512) >> 10;
  }
}

void vp9_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
                              const struct macroblockd_plane *pd,
                              ENTROPY_CONTEXT t_above[16],
                              ENTROPY_CONTEXT t_left[16]) {
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
  const ENTROPY_CONTEXT *const above = pd->above_context;
  const ENTROPY_CONTEXT *const left = pd->left_context;

  int i;
  switch (tx_size) {
    case TX_4X4:
James Zern's avatar
James Zern committed
427
428
      memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
      memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
      break;
    case TX_8X8:
      for (i = 0; i < num_4x4_w; i += 2)
        t_above[i] = !!*(const uint16_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 2)
        t_left[i] = !!*(const uint16_t *)&left[i];
      break;
    case TX_16X16:
      for (i = 0; i < num_4x4_w; i += 4)
        t_above[i] = !!*(const uint32_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 4)
        t_left[i] = !!*(const uint32_t *)&left[i];
      break;
    case TX_32X32:
      for (i = 0; i < num_4x4_w; i += 8)
        t_above[i] = !!*(const uint64_t *)&above[i];
      for (i = 0; i < num_4x4_h; i += 8)
        t_left[i] = !!*(const uint64_t *)&left[i];
      break;
    default:
      assert(0 && "Invalid transform size.");
450
      break;
451
452
453
454
455
456
457
458
459
460
461
462
  }
}

void vp9_mv_pred(VP9_COMP *cpi, MACROBLOCK *x,
                 uint8_t *ref_y_buffer, int ref_y_stride,
                 int ref_frame, BLOCK_SIZE block_size) {
  int i;
  int zero_seen = 0;
  int best_index = 0;
  int best_sad = INT_MAX;
  int this_sad = INT_MAX;
  int max_mv = 0;
463
  int near_same_nearest;
464
465
  uint8_t *src_y_ptr = x->plane[0].src.buf;
  uint8_t *ref_y_ptr;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
466
  const int num_mv_refs = MAX_MV_REF_CANDIDATES +
467
                    (cpi->sf.adaptive_motion_search &&
468
                     block_size < x->max_partition_size);
469
470

  MV pred_mv[3];
471
472
  pred_mv[0] = x->mbmi_ext->ref_mvs[ref_frame][0].as_mv;
  pred_mv[1] = x->mbmi_ext->ref_mvs[ref_frame][1].as_mv;
473
  pred_mv[2] = x->pred_mv[ref_frame];
474
  assert(num_mv_refs <= (int)(sizeof(pred_mv) / sizeof(pred_mv[0])));
475

476
  near_same_nearest =
477
478
      x->mbmi_ext->ref_mvs[ref_frame][0].as_int ==
          x->mbmi_ext->ref_mvs[ref_frame][1].as_int;
Alex Converse's avatar
Alex Converse committed
479
480
  // Get the sad for each candidate reference mv.
  for (i = 0; i < num_mv_refs; ++i) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
481
    const MV *this_mv = &pred_mv[i];
482
    int fp_row, fp_col;
483

484
    if (i == 1 && near_same_nearest)
485
      continue;
486
487
488
    fp_row = (this_mv->row + 3 + (this_mv->row >= 0)) >> 3;
    fp_col = (this_mv->col + 3 + (this_mv->col >= 0)) >> 3;
    max_mv = MAX(max_mv, MAX(abs(this_mv->row), abs(this_mv->col)) >> 3);
489

490
491
492
    if (fp_row ==0 && fp_col == 0 && zero_seen)
      continue;
    zero_seen |= (fp_row ==0 && fp_col == 0);
493

494
    ref_y_ptr =&ref_y_buffer[ref_y_stride * fp_row + fp_col];
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
    // Find sad for current vector.
    this_sad = cpi->fn_ptr[block_size].sdf(src_y_ptr, x->plane[0].src.stride,
                                           ref_y_ptr, ref_y_stride);
    // Note if it is the best so far.
    if (this_sad < best_sad) {
      best_sad = this_sad;
      best_index = i;
    }
  }

  // Note the index of the mv that worked best in the reference list.
  x->mv_best_ref_index[ref_frame] = best_index;
  x->max_mv_context[ref_frame] = max_mv;
  x->pred_mv_sad[ref_frame] = best_sad;
}

void vp9_setup_pred_block(const MACROBLOCKD *xd,
                          struct buf_2d dst[MAX_MB_PLANE],
                          const YV12_BUFFER_CONFIG *src,
                          int mi_row, int mi_col,
                          const struct scale_factors *scale,
                          const struct scale_factors *scale_uv) {
  int i;

  dst[0].buf = src->y_buffer;
  dst[0].stride = src->y_stride;
  dst[1].buf = src->u_buffer;
  dst[2].buf = src->v_buffer;
  dst[1].stride = dst[2].stride = src->uv_stride;

Alex Converse's avatar
Alex Converse committed
525
  for (i = 0; i < MAX_MB_PLANE; ++i) {
526
527
528
529
530
531
    setup_pred_plane(dst + i, dst[i].buf, dst[i].stride, mi_row, mi_col,
                     i ? scale_uv : scale,
                     xd->plane[i].subsampling_x, xd->plane[i].subsampling_y);
  }
}

532
533
534
535
536
537
538
539
540
541
542
543
544
545
int vp9_raster_block_offset(BLOCK_SIZE plane_bsize,
                            int raster_block, int stride) {
  const int bw = b_width_log2_lookup[plane_bsize];
  const int y = 4 * (raster_block >> bw);
  const int x = 4 * (raster_block & ((1 << bw) - 1));
  return y * stride + x;
}

int16_t* vp9_raster_block_offset_int16(BLOCK_SIZE plane_bsize,
                                       int raster_block, int16_t *base) {
  const int stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
  return base + vp9_raster_block_offset(plane_bsize, raster_block, stride);
}

546
547
YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const VP9_COMP *cpi,
                                             int ref_frame) {
548
549
  const VP9_COMMON *const cm = &cpi->common;
  const int scaled_idx = cpi->scaled_ref_idx[ref_frame - 1];
550
551
552
553
  const int ref_idx = get_ref_frame_buf_idx(cpi, ref_frame);
  return
      (scaled_idx != ref_idx && scaled_idx != INVALID_IDX) ?
          &cm->buffer_pool->frame_bufs[scaled_idx].buf : NULL;
554
555
}

556
int vp9_get_switchable_rate(const VP9_COMP *cpi, const MACROBLOCKD *const xd) {
557
  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
558
559
560
561
562
563
564
565
566
567
  const int ctx = vp9_get_pred_context_switchable_interp(xd);
  return SWITCHABLE_INTERP_RATE_FACTOR *
             cpi->switchable_interp_costs[ctx][mbmi->interp_filter];
}

void vp9_set_rd_speed_thresholds(VP9_COMP *cpi) {
  int i;
  RD_OPT *const rd = &cpi->rd;
  SPEED_FEATURES *const sf = &cpi->sf;

Alex Converse's avatar
Alex Converse committed
568
  // Set baseline threshold values.
569
  for (i = 0; i < MAX_MODES; ++i)
570
    rd->thresh_mult[i] = cpi->oxcf.mode == BEST ? -500 : 0;
571

572
573
574
575
576
577
578
579
580
  if (sf->adaptive_rd_thresh) {
    rd->thresh_mult[THR_NEARESTMV] = 300;
    rd->thresh_mult[THR_NEARESTG] = 300;
    rd->thresh_mult[THR_NEARESTA] = 300;
  } else {
    rd->thresh_mult[THR_NEARESTMV] = 0;
    rd->thresh_mult[THR_NEARESTG] = 0;
    rd->thresh_mult[THR_NEARESTA] = 0;
  }
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617

  rd->thresh_mult[THR_DC] += 1000;

  rd->thresh_mult[THR_NEWMV] += 1000;
  rd->thresh_mult[THR_NEWA] += 1000;
  rd->thresh_mult[THR_NEWG] += 1000;

  rd->thresh_mult[THR_NEARMV] += 1000;
  rd->thresh_mult[THR_NEARA] += 1000;
  rd->thresh_mult[THR_COMP_NEARESTLA] += 1000;
  rd->thresh_mult[THR_COMP_NEARESTGA] += 1000;

  rd->thresh_mult[THR_TM] += 1000;

  rd->thresh_mult[THR_COMP_NEARLA] += 1500;
  rd->thresh_mult[THR_COMP_NEWLA] += 2000;
  rd->thresh_mult[THR_NEARG] += 1000;
  rd->thresh_mult[THR_COMP_NEARGA] += 1500;
  rd->thresh_mult[THR_COMP_NEWGA] += 2000;

  rd->thresh_mult[THR_ZEROMV] += 2000;
  rd->thresh_mult[THR_ZEROG] += 2000;
  rd->thresh_mult[THR_ZEROA] += 2000;
  rd->thresh_mult[THR_COMP_ZEROLA] += 2500;
  rd->thresh_mult[THR_COMP_ZEROGA] += 2500;

  rd->thresh_mult[THR_H_PRED] += 2000;
  rd->thresh_mult[THR_V_PRED] += 2000;
  rd->thresh_mult[THR_D45_PRED ] += 2500;
  rd->thresh_mult[THR_D135_PRED] += 2500;
  rd->thresh_mult[THR_D117_PRED] += 2500;
  rd->thresh_mult[THR_D153_PRED] += 2500;
  rd->thresh_mult[THR_D207_PRED] += 2500;
  rd->thresh_mult[THR_D63_PRED] += 2500;
}

void vp9_set_rd_speed_thresholds_sub8x8(VP9_COMP *cpi) {
618
619
620
  static const int thresh_mult[2][MAX_REFS] =
      {{2500, 2500, 2500, 4500, 4500, 2500},
       {2000, 2000, 2000, 4000, 4000, 2000}};
621
622
  RD_OPT *const rd = &cpi->rd;
  const int idx = cpi->oxcf.mode == BEST;
James Zern's avatar
James Zern committed
623
  memcpy(rd->thresh_mult_sub8x8, thresh_mult[idx], sizeof(thresh_mult[idx]));
624
}
625

626
void vp9_update_rd_thresh_fact(int (*factor_buf)[MAX_MODES], int rd_thresh,
627
                               int bsize, int best_mode_index) {
628
  if (rd_thresh > 0) {
629
630
631
632
633
634
635
    const int top_mode = bsize < BLOCK_8X8 ? MAX_REFS : MAX_MODES;
    int mode;
    for (mode = 0; mode < top_mode; ++mode) {
      const BLOCK_SIZE min_size = MAX(bsize - 1, BLOCK_4X4);
      const BLOCK_SIZE max_size = MIN(bsize + 2, BLOCK_64X64);
      BLOCK_SIZE bs;
      for (bs = min_size; bs <= max_size; ++bs) {
636
        int *const fact = &factor_buf[bs][mode];
637
638
639
640
        if (mode == best_mode_index) {
          *fact -= (*fact >> 4);
        } else {
          *fact = MIN(*fact + RD_THRESH_INC,
641
                      rd_thresh * RD_THRESH_MAX_FACT);
642
643
644
645
646
647
        }
      }
    }
  }
}

648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
int vp9_get_intra_cost_penalty(int qindex, int qdelta,
                               vpx_bit_depth_t bit_depth) {
  const int q = vp9_dc_quant(qindex, qdelta, bit_depth);
#if CONFIG_VP9_HIGHBITDEPTH
  switch (bit_depth) {
    case VPX_BITS_8:
      return 20 * q;
    case VPX_BITS_10:
      return 5 * q;
    case VPX_BITS_12:
      return ROUND_POWER_OF_TWO(5 * q, 2);
    default:
      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
      return -1;
  }
#else
  return 20 * q;
#endif  // CONFIG_VP9_HIGHBITDEPTH
}