?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Add non420 code in multi-threaded loopfilter
[aom.git] / vp9 / encoder / vp9_rdopt.c
bloba183fdc6917309cace2b323d6967d96bd050d685
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
16 #include "vpx_mem/vpx_mem.h"
17
18 #include "vp9/common/vp9_common.h"
19 #include "vp9/common/vp9_entropy.h"
20 #include "vp9/common/vp9_entropymode.h"
21 #include "vp9/common/vp9_idct.h"
22 #include "vp9/common/vp9_mvref_common.h"
23 #include "vp9/common/vp9_pred_common.h"
24 #include "vp9/common/vp9_quant_common.h"
25 #include "vp9/common/vp9_reconinter.h"
26 #include "vp9/common/vp9_reconintra.h"
27 #include "vp9/common/vp9_seg_common.h"
28 #include "vp9/common/vp9_systemdependent.h"
29
30 #include "vp9/encoder/vp9_cost.h"
31 #include "vp9/encoder/vp9_encodemb.h"
32 #include "vp9/encoder/vp9_encodemv.h"
33 #include "vp9/encoder/vp9_encoder.h"
34 #include "vp9/encoder/vp9_mcomp.h"
35 #include "vp9/encoder/vp9_quantize.h"
36 #include "vp9/encoder/vp9_ratectrl.h"
37 #include "vp9/encoder/vp9_rd.h"
38 #include "vp9/encoder/vp9_rdopt.h"
39 #include "vp9/encoder/vp9_variance.h"
40 #include "vp9/encoder/vp9_aq_variance.h"
41
42 #define LAST_FRAME_MODE_MASK ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | \
43 (1 << INTRA_FRAME))
44 #define GOLDEN_FRAME_MODE_MASK ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | \
45 (1 << INTRA_FRAME))
46 #define ALT_REF_MODE_MASK ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | \
47 (1 << INTRA_FRAME))
48
49 #define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
50
51 #define MIN_EARLY_TERM_INDEX 3
52 #define NEW_MV_DISCOUNT_FACTOR 8
53
54 typedef struct {
55 PREDICTION_MODE mode;
56 MV_REFERENCE_FRAME ref_frame[2];
57 } MODE_DEFINITION;
58
59 typedef struct {
60 MV_REFERENCE_FRAME ref_frame[2];
61 } REF_DEFINITION;
62
63 struct rdcost_block_args {
64 MACROBLOCK *x;
65 ENTROPY_CONTEXT t_above[16];
66 ENTROPY_CONTEXT t_left[16];
67 int rate;
68 int64_t dist;
69 int64_t sse;
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 skip;
76 int use_fast_coef_costing;
77 const scan_order *so;
78 };
79
80 #define LAST_NEW_MV_INDEX 6
81 static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
82 {NEARESTMV, {LAST_FRAME, NONE}},
83 {NEARESTMV, {ALTREF_FRAME, NONE}},
84 {NEARESTMV, {GOLDEN_FRAME, NONE}},
85
86 {DC_PRED, {INTRA_FRAME, NONE}},
87
88 {NEWMV, {LAST_FRAME, NONE}},
89 {NEWMV, {ALTREF_FRAME, NONE}},
90 {NEWMV, {GOLDEN_FRAME, NONE}},
91
92 {NEARMV, {LAST_FRAME, NONE}},
93 {NEARMV, {ALTREF_FRAME, NONE}},
94 {NEARMV, {GOLDEN_FRAME, NONE}},
95
96 {ZEROMV, {LAST_FRAME, NONE}},
97 {ZEROMV, {GOLDEN_FRAME, NONE}},
98 {ZEROMV, {ALTREF_FRAME, NONE}},
99
100 {NEARESTMV, {LAST_FRAME, ALTREF_FRAME}},
101 {NEARESTMV, {GOLDEN_FRAME, ALTREF_FRAME}},
102
103 {TM_PRED, {INTRA_FRAME, NONE}},
104
105 {NEARMV, {LAST_FRAME, ALTREF_FRAME}},
106 {NEWMV, {LAST_FRAME, ALTREF_FRAME}},
107 {NEARMV, {GOLDEN_FRAME, ALTREF_FRAME}},
108 {NEWMV, {GOLDEN_FRAME, ALTREF_FRAME}},
109
110 {ZEROMV, {LAST_FRAME, ALTREF_FRAME}},
111 {ZEROMV, {GOLDEN_FRAME, ALTREF_FRAME}},
112
113 {H_PRED, {INTRA_FRAME, NONE}},
114 {V_PRED, {INTRA_FRAME, NONE}},
115 {D135_PRED, {INTRA_FRAME, NONE}},
116 {D207_PRED, {INTRA_FRAME, NONE}},
117 {D153_PRED, {INTRA_FRAME, NONE}},
118 {D63_PRED, {INTRA_FRAME, NONE}},
119 {D117_PRED, {INTRA_FRAME, NONE}},
120 {D45_PRED, {INTRA_FRAME, NONE}},
121 };
122
123 static const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
124 {{LAST_FRAME, NONE}},
125 {{GOLDEN_FRAME, NONE}},
126 {{ALTREF_FRAME, NONE}},
127 {{LAST_FRAME, ALTREF_FRAME}},
128 {{GOLDEN_FRAME, ALTREF_FRAME}},
129 {{INTRA_FRAME, NONE}},
130 };
131
132 static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
133 int m, int n, int min_plane, int max_plane) {
134 int i;
135
136 for (i = min_plane; i < max_plane; ++i) {
137 struct macroblock_plane *const p = &x->plane[i];
138 struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
139
140 p->coeff = ctx->coeff_pbuf[i][m];
141 p->qcoeff = ctx->qcoeff_pbuf[i][m];
142 pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
143 p->eobs = ctx->eobs_pbuf[i][m];
144
145 ctx->coeff_pbuf[i][m] = ctx->coeff_pbuf[i][n];
146 ctx->qcoeff_pbuf[i][m] = ctx->qcoeff_pbuf[i][n];
147 ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
148 ctx->eobs_pbuf[i][m] = ctx->eobs_pbuf[i][n];
149
150 ctx->coeff_pbuf[i][n] = p->coeff;
151 ctx->qcoeff_pbuf[i][n] = p->qcoeff;
152 ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
153 ctx->eobs_pbuf[i][n] = p->eobs;
154 }
155 }
156
157 static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize,
158 MACROBLOCK *x, MACROBLOCKD *xd,
159 int *out_rate_sum, int64_t *out_dist_sum,
160 int *skip_txfm_sb, int64_t *skip_sse_sb) {
161 // Note our transform coeffs are 8 times an orthogonal transform.
162 // Hence quantizer step is also 8 times. To get effective quantizer
163 // we need to divide by 8 before sending to modeling function.
164 int i;
165 int64_t rate_sum = 0;
166 int64_t dist_sum = 0;
167 const int ref = xd->mi[0].src_mi->mbmi.ref_frame[0];
168 unsigned int sse;
169 unsigned int var = 0;
170 unsigned int sum_sse = 0;
171 int64_t total_sse = 0;
172 int skip_flag = 1;
173 const int shift = 6;
174 int rate;
175 int64_t dist;
176
177 x->pred_sse[ref] = 0;
178
179 for (i = 0; i < MAX_MB_PLANE; ++i) {
180 struct macroblock_plane *const p = &x->plane[i];
181 struct macroblockd_plane *const pd = &xd->plane[i];
182 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
183 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
184 const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
185 const int64_t dc_thr = p->quant_thred[0] >> shift;
186 const int64_t ac_thr = p->quant_thred[1] >> shift;
187 // The low thresholds are used to measure if the prediction errors are
188 // low enough so that we can skip the mode search.
189 const int64_t low_dc_thr = MIN(50, dc_thr >> 2);
190 const int64_t low_ac_thr = MIN(80, ac_thr >> 2);
191 int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
192 int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
193 int idx, idy;
194 int lw = b_width_log2_lookup[unit_size] + 2;
195 int lh = b_height_log2_lookup[unit_size] + 2;
196
197 sum_sse = 0;
198
199 for (idy = 0; idy < bh; ++idy) {
200 for (idx = 0; idx < bw; ++idx) {
201 uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
202 uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
203 int block_idx = (idy << 1) + idx;
204 int low_err_skip = 0;
205
206 var = cpi->fn_ptr[unit_size].vf(src, p->src.stride,
207 dst, pd->dst.stride, &sse);
208 x->bsse[(i << 2) + block_idx] = sse;
209 sum_sse += sse;
210
211 x->skip_txfm[(i << 2) + block_idx] = 0;
212 if (!x->select_tx_size) {
213 // Check if all ac coefficients can be quantized to zero.
214 if (var < ac_thr || var == 0) {
215 x->skip_txfm[(i << 2) + block_idx] = 2;
216
217 // Check if dc coefficient can be quantized to zero.
218 if (sse - var < dc_thr || sse == var) {
219 x->skip_txfm[(i << 2) + block_idx] = 1;
220
221 if (!sse || (var < low_ac_thr && sse - var < low_dc_thr))
222 low_err_skip = 1;
223 }
224 }
225 }
226
227 if (skip_flag && !low_err_skip)
228 skip_flag = 0;
229
230 if (i == 0)
231 x->pred_sse[ref] += sse;
232 }
233 }
234
235 total_sse += sum_sse;
236
237 // Fast approximate the modelling function.
238 if (cpi->oxcf.speed > 4) {
239 int64_t rate;
240 const int64_t square_error = sum_sse;
241 int quantizer = (pd->dequant[1] >> 3);
242 #if CONFIG_VP9_HIGHBITDEPTH
243 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
244 quantizer >>= (xd->bd - 8);
245 }
246 #endif // CONFIG_VP9_HIGHBITDEPTH
247
248 if (quantizer < 120)
249 rate = (square_error * (280 - quantizer)) >> 8;
250 else
251 rate = 0;
252 dist = (square_error * quantizer) >> 8;
253 rate_sum += rate;
254 dist_sum += dist;
255 } else {
256 #if CONFIG_VP9_HIGHBITDEPTH
257 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
258 vp9_model_rd_from_var_lapndz(sum_sse, num_pels_log2_lookup[bs],
259 pd->dequant[1] >> (xd->bd - 5),
260 &rate, &dist);
261 } else {
262 vp9_model_rd_from_var_lapndz(sum_sse, num_pels_log2_lookup[bs],
263 pd->dequant[1] >> 3, &rate, &dist);
264 }
265 #else
266 vp9_model_rd_from_var_lapndz(sum_sse, num_pels_log2_lookup[bs],
267 pd->dequant[1] >> 3, &rate, &dist);
268 #endif // CONFIG_VP9_HIGHBITDEPTH
269 rate_sum += rate;
270 dist_sum += dist;
271 }
272 }
273
274 *skip_txfm_sb = skip_flag;
275 *skip_sse_sb = total_sse << 4;
276 *out_rate_sum = (int)rate_sum;
277 *out_dist_sum = dist_sum << 4;
278 }
279
280 int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
281 intptr_t block_size, int64_t *ssz) {
282 int i;
283 int64_t error = 0, sqcoeff = 0;
284
285 for (i = 0; i < block_size; i++) {
286 const int diff = coeff[i] - dqcoeff[i];
287 error += diff * diff;
288 sqcoeff += coeff[i] * coeff[i];
289 }
290
291 *ssz = sqcoeff;
292 return error;
293 }
294
295
296 #if CONFIG_VP9_HIGHBITDEPTH
297 int64_t vp9_highbd_block_error_c(const tran_low_t *coeff,
298 const tran_low_t *dqcoeff,
299 intptr_t block_size,
300 int64_t *ssz, int bd) {
301 int i;
302 int64_t error = 0, sqcoeff = 0;
303 int shift = 2 * (bd - 8);
304 int rounding = shift > 0 ? 1 << (shift - 1) : 0;
305
306 for (i = 0; i < block_size; i++) {
307 const int64_t diff = coeff[i] - dqcoeff[i];
308 error += diff * diff;
309 sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
310 }
311 assert(error >= 0 && sqcoeff >= 0);
312 error = (error + rounding) >> shift;
313 sqcoeff = (sqcoeff + rounding) >> shift;
314
315 *ssz = sqcoeff;
316 return error;
317 }
318 #endif // CONFIG_VP9_HIGHBITDEPTH
319
320 /* The trailing '0' is a terminator which is used inside cost_coeffs() to
321 * decide whether to include cost of a trailing EOB node or not (i.e. we
322 * can skip this if the last coefficient in this transform block, e.g. the
323 * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
324 * were non-zero). */
325 static const int16_t band_counts[TX_SIZES][8] = {
326 { 1, 2, 3, 4, 3, 16 - 13, 0 },
327 { 1, 2, 3, 4, 11, 64 - 21, 0 },
328 { 1, 2, 3, 4, 11, 256 - 21, 0 },
329 { 1, 2, 3, 4, 11, 1024 - 21, 0 },
330 };
331 static INLINE int cost_coeffs(MACROBLOCK *x,
332 int plane, int block,
333 ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L,
334 TX_SIZE tx_size,
335 const int16_t *scan, const int16_t *nb,
336 int use_fast_coef_costing) {
337 MACROBLOCKD *const xd = &x->e_mbd;
338 MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
339 const struct macroblock_plane *p = &x->plane[plane];
340 const struct macroblockd_plane *pd = &xd->plane[plane];
341 const PLANE_TYPE type = pd->plane_type;
342 const int16_t *band_count = &band_counts[tx_size][1];
343 const int eob = p->eobs[block];
344 const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
345 unsigned int (*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
346 x->token_costs[tx_size][type][is_inter_block(mbmi)];
347 uint8_t token_cache[32 * 32];
348 int pt = combine_entropy_contexts(*A, *L);
349 int c, cost;
350 #if CONFIG_VP9_HIGHBITDEPTH
351 const int16_t *cat6_high_cost = vp9_get_high_cost_table(xd->bd);
352 #else
353 const int16_t *cat6_high_cost = vp9_get_high_cost_table(8);
354 #endif
355
356 // Check for consistency of tx_size with mode info
357 assert(type == PLANE_TYPE_Y ? mbmi->tx_size == tx_size
358 : get_uv_tx_size(mbmi, pd) == tx_size);
359
360 if (eob == 0) {
361 // single eob token
362 cost = token_costs[0][0][pt][EOB_TOKEN];
363 c = 0;
364 } else {
365 int band_left = *band_count++;
366
367 // dc token
368 int v = qcoeff[0];
369 int16_t prev_t;
370 EXTRABIT e;
371 vp9_get_token_extra(v, &prev_t, &e);
372 cost = (*token_costs)[0][pt][prev_t] +
373 vp9_get_cost(prev_t, e, cat6_high_cost);
374
375 token_cache[0] = vp9_pt_energy_class[prev_t];
376 ++token_costs;
377
378 // ac tokens
379 for (c = 1; c < eob; c++) {
380 const int rc = scan[c];
381 int16_t t;
382
383 v = qcoeff[rc];
384 vp9_get_token_extra(v, &t, &e);
385 if (use_fast_coef_costing) {
386 cost += (*token_costs)[!prev_t][!prev_t][t] +
387 vp9_get_cost(t, e, cat6_high_cost);
388 } else {
389 pt = get_coef_context(nb, token_cache, c);
390 cost += (*token_costs)[!prev_t][pt][t] +
391 vp9_get_cost(t, e, cat6_high_cost);
392 token_cache[rc] = vp9_pt_energy_class[t];
393 }
394 prev_t = t;
395 if (!--band_left) {
396 band_left = *band_count++;
397 ++token_costs;
398 }
399 }
400
401 // eob token
402 if (band_left) {
403 if (use_fast_coef_costing) {
404 cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
405 } else {
406 pt = get_coef_context(nb, token_cache, c);
407 cost += (*token_costs)[0][pt][EOB_TOKEN];
408 }
409 }
410 }
411
412 // is eob first coefficient;
413 *A = *L = (c > 0);
414
415 return cost;
416 }
417
418 #if CONFIG_VP9_HIGHBITDEPTH
419 static void dist_block(int plane, int block, TX_SIZE tx_size,
420 struct rdcost_block_args* args, int bd) {
421 #else
422 static void dist_block(int plane, int block, TX_SIZE tx_size,
423 struct rdcost_block_args* args) {
424 #endif // CONFIG_VP9_HIGHBITDEPTH
425 const int ss_txfrm_size = tx_size << 1;
426 MACROBLOCK* const x = args->x;
427 MACROBLOCKD* const xd = &x->e_mbd;
428 const struct macroblock_plane *const p = &x->plane[plane];
429 const struct macroblockd_plane *const pd = &xd->plane[plane];
430 int64_t this_sse;
431 int shift = tx_size == TX_32X32 ? 0 : 2;
432 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
433 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
434 #if CONFIG_VP9_HIGHBITDEPTH
435 args->dist = vp9_highbd_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
436 &this_sse, bd) >> shift;
437 #else
438 args->dist = vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
439 &this_sse) >> shift;
440 #endif // CONFIG_VP9_HIGHBITDEPTH
441 args->sse = this_sse >> shift;
442
443 if (x->skip_encode && !is_inter_block(&xd->mi[0].src_mi->mbmi)) {
444 // TODO(jingning): tune the model to better capture the distortion.
445 int64_t p = (pd->dequant[1] * pd->dequant[1] *
446 (1 << ss_txfrm_size)) >> (shift + 2);
447 #if CONFIG_VP9_HIGHBITDEPTH
448 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
449 p >>= ((xd->bd - 8) * 2);
450 }
451 #endif // CONFIG_VP9_HIGHBITDEPTH
452 args->dist += (p >> 4);
453 args->sse += p;
454 }
455 }
456
457 static void rate_block(int plane, int block, BLOCK_SIZE plane_bsize,
458 TX_SIZE tx_size, struct rdcost_block_args* args) {
459 int x_idx, y_idx;
460 txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x_idx, &y_idx);
461
462 args->rate = cost_coeffs(args->x, plane, block, args->t_above + x_idx,
463 args->t_left + y_idx, tx_size,
464 args->so->scan, args->so->neighbors,
465 args->use_fast_coef_costing);
466 }
467
468 static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
469 TX_SIZE tx_size, void *arg) {
470 struct rdcost_block_args *args = arg;
471 MACROBLOCK *const x = args->x;
472 MACROBLOCKD *const xd = &x->e_mbd;
473 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
474 int64_t rd1, rd2, rd;
475
476 if (args->skip)
477 return;
478
479 if (!is_inter_block(mbmi)) {
480 struct encode_b_args arg = {x, NULL, &mbmi->skip};
481 vp9_encode_block_intra(plane, block, plane_bsize, tx_size, &arg);
482 #if CONFIG_VP9_HIGHBITDEPTH
483 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
484 dist_block(plane, block, tx_size, args, xd->bd);
485 } else {
486 dist_block(plane, block, tx_size, args, 8);
487 }
488 #else
489 dist_block(plane, block, tx_size, args);
490 #endif // CONFIG_VP9_HIGHBITDEPTH
491 } else if (max_txsize_lookup[plane_bsize] == tx_size) {
492 if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] == 0) {
493 // full forward transform and quantization
494 vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
495 #if CONFIG_VP9_HIGHBITDEPTH
496 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
497 dist_block(plane, block, tx_size, args, xd->bd);
498 } else {
499 dist_block(plane, block, tx_size, args, 8);
500 }
501 #else
502 dist_block(plane, block, tx_size, args);
503 #endif // CONFIG_VP9_HIGHBITDEPTH
504 } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] == 2) {
505 // compute DC coefficient
506 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[plane].coeff, block);
507 tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
508 vp9_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
509 args->sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
510 args->dist = args->sse;
511 if (x->plane[plane].eobs[block]) {
512 int64_t dc_correct = coeff[0] * coeff[0] -
513 (coeff[0] - dqcoeff[0]) * (coeff[0] - dqcoeff[0]);
514 #if CONFIG_VP9_HIGHBITDEPTH
515 dc_correct >>= ((xd->bd - 8) * 2);
516 #endif
517 if (tx_size != TX_32X32)
518 dc_correct >>= 2;
519
520 args->dist = MAX(0, args->sse - dc_correct);
521 }
522 } else {
523 // skip forward transform
524 x->plane[plane].eobs[block] = 0;
525 args->sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
526 args->dist = args->sse;
527 }
528 } else {
529 // full forward transform and quantization
530 vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
531 #if CONFIG_VP9_HIGHBITDEPTH
532 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
533 dist_block(plane, block, tx_size, args, xd->bd);
534 } else {
535 dist_block(plane, block, tx_size, args, 8);
536 }
537 #else
538 dist_block(plane, block, tx_size, args);
539 #endif // CONFIG_VP9_HIGHBITDEPTH
540 }
541
542 rate_block(plane, block, plane_bsize, tx_size, args);
543 rd1 = RDCOST(x->rdmult, x->rddiv, args->rate, args->dist);
544 rd2 = RDCOST(x->rdmult, x->rddiv, 0, args->sse);
545
546 // TODO(jingning): temporarily enabled only for luma component
547 rd = MIN(rd1, rd2);
548 if (plane == 0)
549 x->zcoeff_blk[tx_size][block] = !x->plane[plane].eobs[block] ||
550 (rd1 > rd2 && !xd->lossless);
551
552 args->this_rate += args->rate;
553 args->this_dist += args->dist;
554 args->this_sse += args->sse;
555 args->this_rd += rd;
556
557 if (args->this_rd > args->best_rd) {
558 args->skip = 1;
559 return;
560 }
561 }
562
563 static void txfm_rd_in_plane(MACROBLOCK *x,
564 int *rate, int64_t *distortion,
565 int *skippable, int64_t *sse,
566 int64_t ref_best_rd, int plane,
567 BLOCK_SIZE bsize, TX_SIZE tx_size,
568 int use_fast_coef_casting) {
569 MACROBLOCKD *const xd = &x->e_mbd;
570 const struct macroblockd_plane *const pd = &xd->plane[plane];
571 struct rdcost_block_args args;
572 vp9_zero(args);
573 args.x = x;
574 args.best_rd = ref_best_rd;
575 args.use_fast_coef_costing = use_fast_coef_casting;
576
577 if (plane == 0)
578 xd->mi[0].src_mi->mbmi.tx_size = tx_size;
579
580 vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
581
582 args.so = get_scan(xd, tx_size, pd->plane_type, 0);
583
584 vp9_foreach_transformed_block_in_plane(xd, bsize, plane,
585 block_rd_txfm, &args);
586 if (args.skip) {
587 *rate = INT_MAX;
588 *distortion = INT64_MAX;
589 *sse = INT64_MAX;
590 *skippable = 0;
591 } else {
592 *distortion = args.this_dist;
593 *rate = args.this_rate;
594 *sse = args.this_sse;
595 *skippable = vp9_is_skippable_in_plane(x, bsize, plane);
596 }
597 }
598
599 static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x,
600 int *rate, int64_t *distortion,
601 int *skip, int64_t *sse,
602 int64_t ref_best_rd,
603 BLOCK_SIZE bs) {
604 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
605 VP9_COMMON *const cm = &cpi->common;
606 const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
607 MACROBLOCKD *const xd = &x->e_mbd;
608 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
609
610 mbmi->tx_size = MIN(max_tx_size, largest_tx_size);
611
612 txfm_rd_in_plane(x, rate, distortion, skip,
613 sse, ref_best_rd, 0, bs,
614 mbmi->tx_size, cpi->sf.use_fast_coef_costing);
615 }
616
617 static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
618 int *rate,
619 int64_t *distortion,
620 int *skip,
621 int64_t *psse,
622 int64_t tx_cache[TX_MODES],
623 int64_t ref_best_rd,
624 BLOCK_SIZE bs) {
625 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
626 VP9_COMMON *const cm = &cpi->common;
627 MACROBLOCKD *const xd = &x->e_mbd;
628 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
629 vp9_prob skip_prob = vp9_get_skip_prob(cm, xd);
630 int r[TX_SIZES][2], s[TX_SIZES];
631 int64_t d[TX_SIZES], sse[TX_SIZES];
632 int64_t rd[TX_SIZES][2] = {{INT64_MAX, INT64_MAX},
633 {INT64_MAX, INT64_MAX},
634 {INT64_MAX, INT64_MAX},
635 {INT64_MAX, INT64_MAX}};
636 int n, m;
637 int s0, s1;
638 const TX_SIZE max_mode_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
639 int64_t best_rd = INT64_MAX;
640 TX_SIZE best_tx = max_tx_size;
641
642 const vp9_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc->tx_probs);
643 assert(skip_prob > 0);
644 s0 = vp9_cost_bit(skip_prob, 0);
645 s1 = vp9_cost_bit(skip_prob, 1);
646
647 for (n = max_tx_size; n >= 0; n--) {
648 txfm_rd_in_plane(x, &r[n][0], &d[n], &s[n],
649 &sse[n], ref_best_rd, 0, bs, n,
650 cpi->sf.use_fast_coef_costing);
651 r[n][1] = r[n][0];
652 if (r[n][0] < INT_MAX) {
653 for (m = 0; m <= n - (n == (int) max_tx_size); m++) {
654 if (m == n)
655 r[n][1] += vp9_cost_zero(tx_probs[m]);
656 else
657 r[n][1] += vp9_cost_one(tx_probs[m]);
658 }
659 }
660 if (d[n] == INT64_MAX) {
661 rd[n][0] = rd[n][1] = INT64_MAX;
662 } else if (s[n]) {
663 rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, d[n]);
664 } else {
665 rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
666 rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
667 }
668
669 // Early termination in transform size search.
670 if (cpi->sf.tx_size_search_breakout &&
671 (rd[n][1] == INT64_MAX ||
672 (n < (int) max_tx_size && rd[n][1] > rd[n + 1][1]) ||
673 s[n] == 1))
674 break;
675
676 if (rd[n][1] < best_rd) {
677 best_tx = n;
678 best_rd = rd[n][1];
679 }
680 }
681 mbmi->tx_size = cm->tx_mode == TX_MODE_SELECT ?
682 best_tx : MIN(max_tx_size, max_mode_tx_size);
683
684
685 *distortion = d[mbmi->tx_size];
686 *rate = r[mbmi->tx_size][cm->tx_mode == TX_MODE_SELECT];
687 *skip = s[mbmi->tx_size];
688 *psse = sse[mbmi->tx_size];
689
690 tx_cache[ONLY_4X4] = rd[TX_4X4][0];
691 tx_cache[ALLOW_8X8] = rd[TX_8X8][0];
692 tx_cache[ALLOW_16X16] = rd[MIN(max_tx_size, TX_16X16)][0];
693 tx_cache[ALLOW_32X32] = rd[MIN(max_tx_size, TX_32X32)][0];
694
695 if (max_tx_size == TX_32X32 && best_tx == TX_32X32) {
696 tx_cache[TX_MODE_SELECT] = rd[TX_32X32][1];
697 } else if (max_tx_size >= TX_16X16 && best_tx == TX_16X16) {
698 tx_cache[TX_MODE_SELECT] = rd[TX_16X16][1];
699 } else if (rd[TX_8X8][1] < rd[TX_4X4][1]) {
700 tx_cache[TX_MODE_SELECT] = rd[TX_8X8][1];
701 } else {
702 tx_cache[TX_MODE_SELECT] = rd[TX_4X4][1];
703 }
704 }
705
706 static void super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
707 int64_t *distortion, int *skip,
708 int64_t *psse, BLOCK_SIZE bs,
709 int64_t txfm_cache[TX_MODES],
710 int64_t ref_best_rd) {
711 MACROBLOCKD *xd = &x->e_mbd;
712 int64_t sse;
713 int64_t *ret_sse = psse ? psse : &sse;
714
715 assert(bs == xd->mi[0].src_mi->mbmi.sb_type);
716
717 if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
718 vpx_memset(txfm_cache, 0, TX_MODES * sizeof(int64_t));
719 choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
720 bs);
721 } else {
722 choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse,
723 txfm_cache, ref_best_rd, bs);
724 }
725 }
726
727 static int conditional_skipintra(PREDICTION_MODE mode,
728 PREDICTION_MODE best_intra_mode) {
729 if (mode == D117_PRED &&
730 best_intra_mode != V_PRED &&
731 best_intra_mode != D135_PRED)
732 return 1;
733 if (mode == D63_PRED &&
734 best_intra_mode != V_PRED &&
735 best_intra_mode != D45_PRED)
736 return 1;
737 if (mode == D207_PRED &&
738 best_intra_mode != H_PRED &&
739 best_intra_mode != D45_PRED)
740 return 1;
741 if (mode == D153_PRED &&
742 best_intra_mode != H_PRED &&
743 best_intra_mode != D135_PRED)
744 return 1;
745 return 0;
746 }
747
748 static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int ib,
749 PREDICTION_MODE *best_mode,
750 const int *bmode_costs,
751 ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
752 int *bestrate, int *bestratey,
753 int64_t *bestdistortion,
754 BLOCK_SIZE bsize, int64_t rd_thresh) {
755 PREDICTION_MODE mode;
756 MACROBLOCKD *const xd = &x->e_mbd;
757 int64_t best_rd = rd_thresh;
758
759 struct macroblock_plane *p = &x->plane[0];
760 struct macroblockd_plane *pd = &xd->plane[0];
761 const int src_stride = p->src.stride;
762 const int dst_stride = pd->dst.stride;
763 const uint8_t *src_init = &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, ib,
764 src_stride)];
765 uint8_t *dst_init = &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, ib,
766 dst_stride)];
767 ENTROPY_CONTEXT ta[2], tempa[2];
768 ENTROPY_CONTEXT tl[2], templ[2];
769
770 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
771 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
772 int idx, idy;
773 uint8_t best_dst[8 * 8];
774 #if CONFIG_VP9_HIGHBITDEPTH
775 uint16_t best_dst16[8 * 8];
776 #endif
777
778 assert(ib < 4);
779
780 vpx_memcpy(ta, a, sizeof(ta));
781 vpx_memcpy(tl, l, sizeof(tl));
782 xd->mi[0].src_mi->mbmi.tx_size = TX_4X4;
783
784 #if CONFIG_VP9_HIGHBITDEPTH
785 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
786 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
787 int64_t this_rd;
788 int ratey = 0;
789 int64_t distortion = 0;
790 int rate = bmode_costs[mode];
791
792 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
793 continue;
794
795 // Only do the oblique modes if the best so far is
796 // one of the neighboring directional modes
797 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
798 if (conditional_skipintra(mode, *best_mode))
799 continue;
800 }
801
802 vpx_memcpy(tempa, ta, sizeof(ta));
803 vpx_memcpy(templ, tl, sizeof(tl));
804
805 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
806 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
807 const int block = ib + idy * 2 + idx;
808 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
809 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
810 int16_t *const src_diff = vp9_raster_block_offset_int16(BLOCK_8X8,
811 block,
812 p->src_diff);
813 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
814 xd->mi[0].src_mi->bmi[block].as_mode = mode;
815 vp9_predict_intra_block(xd, block, 1,
816 TX_4X4, mode,
817 x->skip_encode ? src : dst,
818 x->skip_encode ? src_stride : dst_stride,
819 dst, dst_stride, idx, idy, 0);
820 vp9_highbd_subtract_block(4, 4, src_diff, 8, src, src_stride,
821 dst, dst_stride, xd->bd);
822 if (xd->lossless) {
823 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
824 vp9_highbd_fwht4x4(src_diff, coeff, 8);
825 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
826 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
827 so->scan, so->neighbors,
828 cpi->sf.use_fast_coef_costing);
829 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
830 goto next_highbd;
831 vp9_highbd_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block),
832 dst, dst_stride,
833 p->eobs[block], xd->bd);
834 } else {
835 int64_t unused;
836 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
837 const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
838 vp9_highbd_fht4x4(src_diff, coeff, 8, tx_type);
839 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
840 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
841 so->scan, so->neighbors,
842 cpi->sf.use_fast_coef_costing);
843 distortion += vp9_highbd_block_error(
844 coeff, BLOCK_OFFSET(pd->dqcoeff, block),
845 16, &unused, xd->bd) >> 2;
846 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
847 goto next_highbd;
848 vp9_highbd_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
849 dst, dst_stride, p->eobs[block], xd->bd);
850 }
851 }
852 }
853
854 rate += ratey;
855 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
856
857 if (this_rd < best_rd) {
858 *bestrate = rate;
859 *bestratey = ratey;
860 *bestdistortion = distortion;
861 best_rd = this_rd;
862 *best_mode = mode;
863 vpx_memcpy(a, tempa, sizeof(tempa));
864 vpx_memcpy(l, templ, sizeof(templ));
865 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
866 vpx_memcpy(best_dst16 + idy * 8,
867 CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
868 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
869 }
870 }
871 next_highbd:
872 {}
873 }
874 if (best_rd >= rd_thresh || x->skip_encode)
875 return best_rd;
876
877 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
878 vpx_memcpy(CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
879 best_dst16 + idy * 8,
880 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
881 }
882
883 return best_rd;
884 }
885 #endif // CONFIG_VP9_HIGHBITDEPTH
886
887 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
888 int64_t this_rd;
889 int ratey = 0;
890 int64_t distortion = 0;
891 int rate = bmode_costs[mode];
892
893 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
894 continue;
895
896 // Only do the oblique modes if the best so far is
897 // one of the neighboring directional modes
898 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
899 if (conditional_skipintra(mode, *best_mode))
900 continue;
901 }
902
903 vpx_memcpy(tempa, ta, sizeof(ta));
904 vpx_memcpy(templ, tl, sizeof(tl));
905
906 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
907 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
908 const int block = ib + idy * 2 + idx;
909 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
910 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
911 int16_t *const src_diff =
912 vp9_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
913 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
914 xd->mi[0].src_mi->bmi[block].as_mode = mode;
915 vp9_predict_intra_block(xd, block, 1,
916 TX_4X4, mode,
917 x->skip_encode ? src : dst,
918 x->skip_encode ? src_stride : dst_stride,
919 dst, dst_stride, idx, idy, 0);
920 vp9_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
921
922 if (xd->lossless) {
923 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
924 vp9_fwht4x4(src_diff, coeff, 8);
925 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
926 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
927 so->scan, so->neighbors,
928 cpi->sf.use_fast_coef_costing);
929 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
930 goto next;
931 vp9_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride,
932 p->eobs[block]);
933 } else {
934 int64_t unused;
935 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
936 const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
937 vp9_fht4x4(src_diff, coeff, 8, tx_type);
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 distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
943 16, &unused) >> 2;
944 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
945 goto next;
946 vp9_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
947 dst, dst_stride, p->eobs[block]);
948 }
949 }
950 }
951
952 rate += ratey;
953 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
954
955 if (this_rd < best_rd) {
956 *bestrate = rate;
957 *bestratey = ratey;
958 *bestdistortion = distortion;
959 best_rd = this_rd;
960 *best_mode = mode;
961 vpx_memcpy(a, tempa, sizeof(tempa));
962 vpx_memcpy(l, templ, sizeof(templ));
963 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
964 vpx_memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
965 num_4x4_blocks_wide * 4);
966 }
967 next:
968 {}
969 }
970
971 if (best_rd >= rd_thresh || x->skip_encode)
972 return best_rd;
973
974 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
975 vpx_memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
976 num_4x4_blocks_wide * 4);
977
978 return best_rd;
979 }
980
981 static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
982 int *rate, int *rate_y,
983 int64_t *distortion,
984 int64_t best_rd) {
985 int i, j;
986 const MACROBLOCKD *const xd = &mb->e_mbd;
987 MODE_INFO *const mic = xd->mi[0].src_mi;
988 const MODE_INFO *above_mi = xd->above_mi;
989 const MODE_INFO *left_mi = xd->left_mi;
990 const BLOCK_SIZE bsize = xd->mi[0].src_mi->mbmi.sb_type;
991 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
992 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
993 int idx, idy;
994 int cost = 0;
995 int64_t total_distortion = 0;
996 int tot_rate_y = 0;
997 int64_t total_rd = 0;
998 ENTROPY_CONTEXT t_above[4], t_left[4];
999 const int *bmode_costs = cpi->mbmode_cost;
1000
1001 vpx_memcpy(t_above, xd->plane[0].above_context, sizeof(t_above));
1002 vpx_memcpy(t_left, xd->plane[0].left_context, sizeof(t_left));
1003
1004 // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
1005 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1006 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1007 PREDICTION_MODE best_mode = DC_PRED;
1008 int r = INT_MAX, ry = INT_MAX;
1009 int64_t d = INT64_MAX, this_rd = INT64_MAX;
1010 i = idy * 2 + idx;
1011 if (cpi->common.frame_type == KEY_FRAME) {
1012 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i);
1013 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i);
1014
1015 bmode_costs = cpi->y_mode_costs[A][L];
1016 }
1017
1018 this_rd = rd_pick_intra4x4block(cpi, mb, i, &best_mode, bmode_costs,
1019 t_above + idx, t_left + idy, &r, &ry, &d,
1020 bsize, best_rd - total_rd);
1021 if (this_rd >= best_rd - total_rd)
1022 return INT64_MAX;
1023
1024 total_rd += this_rd;
1025 cost += r;
1026 total_distortion += d;
1027 tot_rate_y += ry;
1028
1029 mic->bmi[i].as_mode = best_mode;
1030 for (j = 1; j < num_4x4_blocks_high; ++j)
1031 mic->bmi[i + j * 2].as_mode = best_mode;
1032 for (j = 1; j < num_4x4_blocks_wide; ++j)
1033 mic->bmi[i + j].as_mode = best_mode;
1034
1035 if (total_rd >= best_rd)
1036 return INT64_MAX;
1037 }
1038 }
1039
1040 *rate = cost;
1041 *rate_y = tot_rate_y;
1042 *distortion = total_distortion;
1043 mic->mbmi.mode = mic->bmi[3].as_mode;
1044
1045 return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
1046 }
1047
1048 // This function is used only for intra_only frames
1049 static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x,
1050 int *rate, int *rate_tokenonly,
1051 int64_t *distortion, int *skippable,
1052 BLOCK_SIZE bsize,
1053 int64_t tx_cache[TX_MODES],
1054 int64_t best_rd) {
1055 PREDICTION_MODE mode;
1056 PREDICTION_MODE mode_selected = DC_PRED;
1057 MACROBLOCKD *const xd = &x->e_mbd;
1058 MODE_INFO *const mic = xd->mi[0].src_mi;
1059 int this_rate, this_rate_tokenonly, s;
1060 int64_t this_distortion, this_rd;
1061 TX_SIZE best_tx = TX_4X4;
1062 int i;
1063 int *bmode_costs;
1064 const MODE_INFO *above_mi = xd->above_mi;
1065 const MODE_INFO *left_mi = xd->left_mi;
1066 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1067 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1068 bmode_costs = cpi->y_mode_costs[A][L];
1069
1070 if (cpi->sf.tx_size_search_method == USE_FULL_RD)
1071 for (i = 0; i < TX_MODES; i++)
1072 tx_cache[i] = INT64_MAX;
1073
1074 vpx_memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
1075 /* Y Search for intra prediction mode */
1076 for (mode = DC_PRED; mode <= TM_PRED; mode++) {
1077 int64_t local_tx_cache[TX_MODES];
1078
1079 if (cpi->sf.use_nonrd_pick_mode) {
1080 // These speed features are turned on in hybrid non-RD and RD mode
1081 // for key frame coding in the context of real-time setting.
1082 if (conditional_skipintra(mode, mode_selected))
1083 continue;
1084 if (*skippable)
1085 break;
1086 }
1087
1088 mic->mbmi.mode = mode;
1089
1090 super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion,
1091 &s, NULL, bsize, local_tx_cache, best_rd);
1092
1093 if (this_rate_tokenonly == INT_MAX)
1094 continue;
1095
1096 this_rate = this_rate_tokenonly + bmode_costs[mode];
1097 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1098
1099 if (this_rd < best_rd) {
1100 mode_selected = mode;
1101 best_rd = this_rd;
1102 best_tx = mic->mbmi.tx_size;
1103 *rate = this_rate;
1104 *rate_tokenonly = this_rate_tokenonly;
1105 *distortion = this_distortion;
1106 *skippable = s;
1107 }
1108
1109 if (cpi->sf.tx_size_search_method == USE_FULL_RD && this_rd < INT64_MAX) {
1110 for (i = 0; i < TX_MODES && local_tx_cache[i] < INT64_MAX; i++) {
1111 const int64_t adj_rd = this_rd + local_tx_cache[i] -
1112 local_tx_cache[cpi->common.tx_mode];
1113 if (adj_rd < tx_cache[i]) {
1114 tx_cache[i] = adj_rd;
1115 }
1116 }
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.
1128 static 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].src_mi->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
1179 static 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 vpx_memset(x->skip_txfm, 0, 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].src_mi->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].src_mi->mbmi.uv_mode = mode_selected;
1218 return best_rd;
1219 }
1220
1221 static 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].src_mi->mbmi.uv_mode = DC_PRED;
1229 vpx_memset(x->skip_txfm, 0, 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
1236 static 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].src_mi->mbmi.uv_mode;
1255 }
1256
1257 static 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
1263 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
1264 BLOCK_SIZE bsize,
1265 int_mv *frame_mv,
1266 int mi_row, int mi_col,
1267 int_mv single_newmv[MAX_REF_FRAMES],
1268 int *rate_mv);
1269
1270 static int set_and_cost_bmi_mvs(VP9_COMP *cpi, MACROBLOCKD *xd, int i,
1271 PREDICTION_MODE mode, int_mv this_mv[2],
1272 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1273 int_mv seg_mvs[MAX_REF_FRAMES],
1274 int_mv *best_ref_mv[2], const int *mvjcost,
1275 int *mvcost[2]) {
1276 MODE_INFO *const mic = xd->mi[0].src_mi;
1277 const MB_MODE_INFO *const mbmi = &mic->mbmi;
1278 int thismvcost = 0;
1279 int idx, idy;
1280 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type];
1281 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
1282 const int is_compound = has_second_ref(mbmi);
1283
1284 switch (mode) {
1285 case NEWMV:
1286 this_mv[0].as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
1287 thismvcost += vp9_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
1288 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1289 if (is_compound) {
1290 this_mv[1].as_int = seg_mvs[mbmi->ref_frame[1]].as_int;
1291 thismvcost += vp9_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
1292 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1293 }
1294 break;
1295 case NEARMV:
1296 case NEARESTMV:
1297 this_mv[0].as_int = frame_mv[mode][mbmi->ref_frame[0]].as_int;
1298 if (is_compound)
1299 this_mv[1].as_int = frame_mv[mode][mbmi->ref_frame[1]].as_int;
1300 break;
1301 case ZEROMV:
1302 this_mv[0].as_int = 0;
1303 if (is_compound)
1304 this_mv[1].as_int = 0;
1305 break;
1306 default:
1307 break;
1308 }
1309
1310 mic->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
1311 if (is_compound)
1312 mic->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
1313
1314 mic->bmi[i].as_mode = mode;
1315
1316 for (idy = 0; idy < num_4x4_blocks_high; ++idy)
1317 for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
1318 vpx_memcpy(&mic->bmi[i + idy * 2 + idx],
1319 &mic->bmi[i], sizeof(mic->bmi[i]));
1320
1321 return cost_mv_ref(cpi, mode, mbmi->mode_context[mbmi->ref_frame[0]]) +
1322 thismvcost;
1323 }
1324
1325 static int64_t encode_inter_mb_segment(VP9_COMP *cpi,
1326 MACROBLOCK *x,
1327 int64_t best_yrd,
1328 int i,
1329 int *labelyrate,
1330 int64_t *distortion, int64_t *sse,
1331 ENTROPY_CONTEXT *ta,
1332 ENTROPY_CONTEXT *tl,
1333 int mi_row, int mi_col) {
1334 int k;
1335 MACROBLOCKD *xd = &x->e_mbd;
1336 struct macroblockd_plane *const pd = &xd->plane[0];
1337 struct macroblock_plane *const p = &x->plane[0];
1338 MODE_INFO *const mi = xd->mi[0].src_mi;
1339 const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->mbmi.sb_type, pd);
1340 const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
1341 const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
1342 int idx, idy;
1343
1344 const uint8_t *const src =
1345 &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1346 uint8_t *const dst = &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i,
1347 pd->dst.stride)];
1348 int64_t thisdistortion = 0, thissse = 0;
1349 int thisrate = 0, ref;
1350 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1351 const int is_compound = has_second_ref(&mi->mbmi);
1352 const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter);
1353
1354 for (ref = 0; ref < 1 + is_compound; ++ref) {
1355 const uint8_t *pre = &pd->pre[ref].buf[vp9_raster_block_offset(BLOCK_8X8, i,
1356 pd->pre[ref].stride)];
1357 #if CONFIG_VP9_HIGHBITDEPTH
1358 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1359 vp9_highbd_build_inter_predictor(pre, pd->pre[ref].stride,
1360 dst, pd->dst.stride,
1361 &mi->bmi[i].as_mv[ref].as_mv,
1362 &xd->block_refs[ref]->sf, width, height,
1363 ref, kernel, MV_PRECISION_Q3,
1364 mi_col * MI_SIZE + 4 * (i % 2),
1365 mi_row * MI_SIZE + 4 * (i / 2), xd->bd);
1366 } else {
1367 vp9_build_inter_predictor(pre, pd->pre[ref].stride,
1368 dst, pd->dst.stride,
1369 &mi->bmi[i].as_mv[ref].as_mv,
1370 &xd->block_refs[ref]->sf, width, height, ref,
1371 kernel, MV_PRECISION_Q3,
1372 mi_col * MI_SIZE + 4 * (i % 2),
1373 mi_row * MI_SIZE + 4 * (i / 2));
1374 }
1375 #else
1376 vp9_build_inter_predictor(pre, pd->pre[ref].stride,
1377 dst, pd->dst.stride,
1378 &mi->bmi[i].as_mv[ref].as_mv,
1379 &xd->block_refs[ref]->sf, width, height, ref,
1380 kernel, MV_PRECISION_Q3,
1381 mi_col * MI_SIZE + 4 * (i % 2),
1382 mi_row * MI_SIZE + 4 * (i / 2));
1383 #endif // CONFIG_VP9_HIGHBITDEPTH
1384 }
1385
1386 #if CONFIG_VP9_HIGHBITDEPTH
1387 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1388 vp9_highbd_subtract_block(
1389 height, width, vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1390 8, src, p->src.stride, dst, pd->dst.stride, xd->bd);
1391 } else {
1392 vp9_subtract_block(
1393 height, width, vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1394 8, src, p->src.stride, dst, pd->dst.stride);
1395 }
1396 #else
1397 vp9_subtract_block(height, width,
1398 vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1399 8, src, p->src.stride, dst, pd->dst.stride);
1400 #endif // CONFIG_VP9_HIGHBITDEPTH
1401
1402 k = i;
1403 for (idy = 0; idy < height / 4; ++idy) {
1404 for (idx = 0; idx < width / 4; ++idx) {
1405 int64_t ssz, rd, rd1, rd2;
1406 tran_low_t* coeff;
1407
1408 k += (idy * 2 + idx);
1409 coeff = BLOCK_OFFSET(p->coeff, k);
1410 x->fwd_txm4x4(vp9_raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
1411 coeff, 8);
1412 vp9_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan);
1413 #if CONFIG_VP9_HIGHBITDEPTH
1414 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1415 thisdistortion += vp9_highbd_block_error(coeff,
1416 BLOCK_OFFSET(pd->dqcoeff, k),
1417 16, &ssz, xd->bd);
1418 } else {
1419 thisdistortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
1420 16, &ssz);
1421 }
1422 #else
1423 thisdistortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
1424 16, &ssz);
1425 #endif // CONFIG_VP9_HIGHBITDEPTH
1426 thissse += ssz;
1427 thisrate += cost_coeffs(x, 0, k, ta + (k & 1), tl + (k >> 1), TX_4X4,
1428 so->scan, so->neighbors,
1429 cpi->sf.use_fast_coef_costing);
1430 rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
1431 rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
1432 rd = MIN(rd1, rd2);
1433 if (rd >= best_yrd)
1434 return INT64_MAX;
1435 }
1436 }
1437
1438 *distortion = thisdistortion >> 2;
1439 *labelyrate = thisrate;
1440 *sse = thissse >> 2;
1441
1442 return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
1443 }
1444
1445 typedef struct {
1446 int eobs;
1447 int brate;
1448 int byrate;
1449 int64_t bdist;
1450 int64_t bsse;
1451 int64_t brdcost;
1452 int_mv mvs[2];
1453 ENTROPY_CONTEXT ta[2];
1454 ENTROPY_CONTEXT tl[2];
1455 } SEG_RDSTAT;
1456
1457 typedef struct {
1458 int_mv *ref_mv[2];
1459 int_mv mvp;
1460
1461 int64_t segment_rd;
1462 int r;
1463 int64_t d;
1464 int64_t sse;
1465 int segment_yrate;
1466 PREDICTION_MODE modes[4];
1467 SEG_RDSTAT rdstat[4][INTER_MODES];
1468 int mvthresh;
1469 } BEST_SEG_INFO;
1470
1471 static INLINE int mv_check_bounds(const MACROBLOCK *x, const MV *mv) {
1472 return (mv->row >> 3) < x->mv_row_min ||
1473 (mv->row >> 3) > x->mv_row_max ||
1474 (mv->col >> 3) < x->mv_col_min ||
1475 (mv->col >> 3) > x->mv_col_max;
1476 }
1477
1478 static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
1479 MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0].src_mi->mbmi;
1480 struct macroblock_plane *const p = &x->plane[0];
1481 struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
1482
1483 p->src.buf = &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i,
1484 p->src.stride)];
1485 assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
1486 pd->pre[0].buf = &pd->pre[0].buf[vp9_raster_block_offset(BLOCK_8X8, i,
1487 pd->pre[0].stride)];
1488 if (has_second_ref(mbmi))
1489 pd->pre[1].buf = &pd->pre[1].buf[vp9_raster_block_offset(BLOCK_8X8, i,
1490 pd->pre[1].stride)];
1491 }
1492
1493 static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
1494 struct buf_2d orig_pre[2]) {
1495 MB_MODE_INFO *mbmi = &x->e_mbd.mi[0].src_mi->mbmi;
1496 x->plane[0].src = orig_src;
1497 x->e_mbd.plane[0].pre[0] = orig_pre[0];
1498 if (has_second_ref(mbmi))
1499 x->e_mbd.plane[0].pre[1] = orig_pre[1];
1500 }
1501
1502 static INLINE int mv_has_subpel(const MV *mv) {
1503 return (mv->row & 0x0F) || (mv->col & 0x0F);
1504 }
1505
1506 // Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
1507 // TODO(aconverse): Find out if this is still productive then clean up or remove
1508 static int check_best_zero_mv(
1509 const VP9_COMP *cpi, const uint8_t mode_context[MAX_REF_FRAMES],
1510 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int this_mode,
1511 const MV_REFERENCE_FRAME ref_frames[2]) {
1512 if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
1513 frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
1514 (ref_frames[1] == NONE ||
1515 frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
1516 int rfc = mode_context[ref_frames[0]];
1517 int c1 = cost_mv_ref(cpi, NEARMV, rfc);
1518 int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
1519 int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
1520
1521 if (this_mode == NEARMV) {
1522 if (c1 > c3) return 0;
1523 } else if (this_mode == NEARESTMV) {
1524 if (c2 > c3) return 0;
1525 } else {
1526 assert(this_mode == ZEROMV);
1527 if (ref_frames[1] == NONE) {
1528 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
1529 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
1530 return 0;
1531 } else {
1532 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
1533 frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
1534 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
1535 frame_mv[NEARMV][ref_frames[1]].as_int == 0))
1536 return 0;
1537 }
1538 }
1539 }
1540 return 1;
1541 }
1542
1543 static int64_t rd_pick_best_sub8x8_mode(VP9_COMP *cpi, MACROBLOCK *x,
1544 const TileInfo * const tile,
1545 int_mv *best_ref_mv,
1546 int_mv *second_best_ref_mv,
1547 int64_t best_rd, int *returntotrate,
1548 int *returnyrate,
1549 int64_t *returndistortion,
1550 int *skippable, int64_t *psse,
1551 int mvthresh,
1552 int_mv seg_mvs[4][MAX_REF_FRAMES],
1553 BEST_SEG_INFO *bsi_buf, int filter_idx,
1554 int mi_row, int mi_col) {
1555 int i;
1556 BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
1557 MACROBLOCKD *xd = &x->e_mbd;
1558 MODE_INFO *mi = xd->mi[0].src_mi;
1559 MB_MODE_INFO *mbmi = &mi->mbmi;
1560 int mode_idx;
1561 int k, br = 0, idx, idy;
1562 int64_t bd = 0, block_sse = 0;
1563 PREDICTION_MODE this_mode;
1564 VP9_COMMON *cm = &cpi->common;
1565 struct macroblock_plane *const p = &x->plane[0];
1566 struct macroblockd_plane *const pd = &xd->plane[0];
1567 const int label_count = 4;
1568 int64_t this_segment_rd = 0;
1569 int label_mv_thresh;
1570 int segmentyrate = 0;
1571 const BLOCK_SIZE bsize = mbmi->sb_type;
1572 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1573 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1574 ENTROPY_CONTEXT t_above[2], t_left[2];
1575 int subpelmv = 1, have_ref = 0;
1576 const int has_second_rf = has_second_ref(mbmi);
1577 const int inter_mode_mask = cpi->sf.inter_mode_mask[bsize];
1578
1579 vp9_zero(*bsi);
1580
1581 bsi->segment_rd = best_rd;
1582 bsi->ref_mv[0] = best_ref_mv;
1583 bsi->ref_mv[1] = second_best_ref_mv;
1584 bsi->mvp.as_int = best_ref_mv->as_int;
1585 bsi->mvthresh = mvthresh;
1586
1587 for (i = 0; i < 4; i++)
1588 bsi->modes[i] = ZEROMV;
1589
1590 vpx_memcpy(t_above, pd->above_context, sizeof(t_above));
1591 vpx_memcpy(t_left, pd->left_context, sizeof(t_left));
1592
1593 // 64 makes this threshold really big effectively
1594 // making it so that we very rarely check mvs on
1595 // segments. setting this to 1 would make mv thresh
1596 // roughly equal to what it is for macroblocks
1597 label_mv_thresh = 1 * bsi->mvthresh / label_count;
1598
1599 // Segmentation method overheads
1600 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1601 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1602 // TODO(jingning,rbultje): rewrite the rate-distortion optimization
1603 // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
1604 int_mv mode_mv[MB_MODE_COUNT][2];
1605 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1606 PREDICTION_MODE mode_selected = ZEROMV;
1607 int64_t best_rd = INT64_MAX;
1608 const int i = idy * 2 + idx;
1609 int ref;
1610
1611 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1612 const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
1613 frame_mv[ZEROMV][frame].as_int = 0;
1614 vp9_append_sub8x8_mvs_for_idx(cm, xd, tile, i, ref, mi_row, mi_col,
1615 &frame_mv[NEARESTMV][frame],
1616 &frame_mv[NEARMV][frame]);
1617 }
1618
1619 // search for the best motion vector on this segment
1620 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1621 const struct buf_2d orig_src = x->plane[0].src;
1622 struct buf_2d orig_pre[2];
1623
1624 mode_idx = INTER_OFFSET(this_mode);
1625 bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
1626 if (!(inter_mode_mask & (1 << this_mode)))
1627 continue;
1628
1629 if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv,
1630 this_mode, mbmi->ref_frame))
1631 continue;
1632
1633 vpx_memcpy(orig_pre, pd->pre, sizeof(orig_pre));
1634 vpx_memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
1635 sizeof(bsi->rdstat[i][mode_idx].ta));
1636 vpx_memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
1637 sizeof(bsi->rdstat[i][mode_idx].tl));
1638
1639 // motion search for newmv (single predictor case only)
1640 if (!has_second_rf && this_mode == NEWMV &&
1641 seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV) {
1642 MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
1643 int step_param = 0;
1644 int thissme, bestsme = INT_MAX;
1645 int sadpb = x->sadperbit4;
1646 MV mvp_full;
1647 int max_mv;
1648 int cost_list[5];
1649
1650 /* Is the best so far sufficiently good that we cant justify doing
1651 * and new motion search. */
1652 if (best_rd < label_mv_thresh)
1653 break;
1654
1655 if (cpi->oxcf.mode != BEST) {
1656 // use previous block's result as next block's MV predictor.
1657 if (i > 0) {
1658 bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
1659 if (i == 2)
1660 bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
1661 }
1662 }
1663 if (i == 0)
1664 max_mv = x->max_mv_context[mbmi->ref_frame[0]];
1665 else
1666 max_mv = MAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
1667
1668 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
1669 // Take wtd average of the step_params based on the last frame's
1670 // max mv magnitude and the best ref mvs of the current block for
1671 // the given reference.
1672 step_param = (vp9_init_search_range(max_mv) +
1673 cpi->mv_step_param) / 2;
1674 } else {
1675 step_param = cpi->mv_step_param;
1676 }
1677
1678 mvp_full.row = bsi->mvp.as_mv.row >> 3;
1679 mvp_full.col = bsi->mvp.as_mv.col >> 3;
1680
1681 if (cpi->sf.adaptive_motion_search) {
1682 mvp_full.row = x->pred_mv[mbmi->ref_frame[0]].row >> 3;
1683 mvp_full.col = x->pred_mv[mbmi->ref_frame[0]].col >> 3;
1684 step_param = MAX(step_param, 8);
1685 }
1686
1687 // adjust src pointer for this block
1688 mi_buf_shift(x, i);
1689
1690 vp9_set_mv_search_range(x, &bsi->ref_mv[0]->as_mv);
1691
1692 bestsme = vp9_full_pixel_search(
1693 cpi, x, bsize, &mvp_full, step_param, sadpb,
1694 cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
1695 &bsi->ref_mv[0]->as_mv, new_mv,
1696 INT_MAX, 1);
1697
1698 // Should we do a full search (best quality only)
1699 if (cpi->oxcf.mode == BEST) {
1700 int_mv *const best_mv = &mi->bmi[i].as_mv[0];
1701 /* Check if mvp_full is within the range. */
1702 clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max,
1703 x->mv_row_min, x->mv_row_max);
1704 thissme = cpi->full_search_sad(x, &mvp_full,
1705 sadpb, 16, &cpi->fn_ptr[bsize],
1706 &bsi->ref_mv[0]->as_mv,
1707 &best_mv->as_mv);
1708 cost_list[1] = cost_list[2] = cost_list[3] = cost_list[4] = INT_MAX;
1709 if (thissme < bestsme) {
1710 bestsme = thissme;
1711 *new_mv = best_mv->as_mv;
1712 } else {
1713 // The full search result is actually worse so re-instate the
1714 // previous best vector
1715 best_mv->as_mv = *new_mv;
1716 }
1717 }
1718
1719 if (bestsme < INT_MAX) {
1720 int distortion;
1721 cpi->find_fractional_mv_step(
1722 x,
1723 new_mv,
1724 &bsi->ref_mv[0]->as_mv,
1725 cm->allow_high_precision_mv,
1726 x->errorperbit, &cpi->fn_ptr[bsize],
1727 cpi->sf.mv.subpel_force_stop,
1728 cpi->sf.mv.subpel_iters_per_step,
1729 cond_cost_list(cpi, cost_list),
1730 x->nmvjointcost, x->mvcost,
1731 &distortion,
1732 &x->pred_sse[mbmi->ref_frame[0]],
1733 NULL, 0, 0);
1734
1735 // save motion search result for use in compound prediction
1736 seg_mvs[i][mbmi->ref_frame[0]].as_mv = *new_mv;
1737 }
1738
1739 if (cpi->sf.adaptive_motion_search)
1740 x->pred_mv[mbmi->ref_frame[0]] = *new_mv;
1741
1742 // restore src pointers
1743 mi_buf_restore(x, orig_src, orig_pre);
1744 }
1745
1746 if (has_second_rf) {
1747 if (seg_mvs[i][mbmi->ref_frame[1]].as_int == INVALID_MV ||
1748 seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV)
1749 continue;
1750 }
1751
1752 if (has_second_rf && this_mode == NEWMV &&
1753 mbmi->interp_filter == EIGHTTAP) {
1754 // adjust src pointers
1755 mi_buf_shift(x, i);
1756 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
1757 int rate_mv;
1758 joint_motion_search(cpi, x, bsize, frame_mv[this_mode],
1759 mi_row, mi_col, seg_mvs[i],
1760 &rate_mv);
1761 seg_mvs[i][mbmi->ref_frame[0]].as_int =
1762 frame_mv[this_mode][mbmi->ref_frame[0]].as_int;
1763 seg_mvs[i][mbmi->ref_frame[1]].as_int =
1764 frame_mv[this_mode][mbmi->ref_frame[1]].as_int;
1765 }
1766 // restore src pointers
1767 mi_buf_restore(x, orig_src, orig_pre);
1768 }
1769
1770 bsi->rdstat[i][mode_idx].brate =
1771 set_and_cost_bmi_mvs(cpi, xd, i, this_mode, mode_mv[this_mode],
1772 frame_mv, seg_mvs[i], bsi->ref_mv,
1773 x->nmvjointcost, x->mvcost);
1774
1775 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1776 bsi->rdstat[i][mode_idx].mvs[ref].as_int =
1777 mode_mv[this_mode][ref].as_int;
1778 if (num_4x4_blocks_wide > 1)
1779 bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int =
1780 mode_mv[this_mode][ref].as_int;
1781 if (num_4x4_blocks_high > 1)
1782 bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int =
1783 mode_mv[this_mode][ref].as_int;
1784 }
1785
1786 // Trap vectors that reach beyond the UMV borders
1787 if (mv_check_bounds(x, &mode_mv[this_mode][0].as_mv) ||
1788 (has_second_rf &&
1789 mv_check_bounds(x, &mode_mv[this_mode][1].as_mv)))
1790 continue;
1791
1792 if (filter_idx > 0) {
1793 BEST_SEG_INFO *ref_bsi = bsi_buf;
1794 subpelmv = 0;
1795 have_ref = 1;
1796
1797 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1798 subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
1799 have_ref &= mode_mv[this_mode][ref].as_int ==
1800 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1801 }
1802
1803 if (filter_idx > 1 && !subpelmv && !have_ref) {
1804 ref_bsi = bsi_buf + 1;
1805 have_ref = 1;
1806 for (ref = 0; ref < 1 + has_second_rf; ++ref)
1807 have_ref &= mode_mv[this_mode][ref].as_int ==
1808 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1809 }
1810
1811 if (!subpelmv && have_ref &&
1812 ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1813 vpx_memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
1814 sizeof(SEG_RDSTAT));
1815 if (num_4x4_blocks_wide > 1)
1816 bsi->rdstat[i + 1][mode_idx].eobs =
1817 ref_bsi->rdstat[i + 1][mode_idx].eobs;
1818 if (num_4x4_blocks_high > 1)
1819 bsi->rdstat[i + 2][mode_idx].eobs =
1820 ref_bsi->rdstat[i + 2][mode_idx].eobs;
1821
1822 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
1823 mode_selected = this_mode;
1824 best_rd = bsi->rdstat[i][mode_idx].brdcost;
1825 }
1826 continue;
1827 }
1828 }
1829
1830 bsi->rdstat[i][mode_idx].brdcost =
1831 encode_inter_mb_segment(cpi, x,
1832 bsi->segment_rd - this_segment_rd, i,
1833 &bsi->rdstat[i][mode_idx].byrate,
1834 &bsi->rdstat[i][mode_idx].bdist,
1835 &bsi->rdstat[i][mode_idx].bsse,
1836 bsi->rdstat[i][mode_idx].ta,
1837 bsi->rdstat[i][mode_idx].tl,
1838 mi_row, mi_col);
1839 if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1840 bsi->rdstat[i][mode_idx].brdcost += RDCOST(x->rdmult, x->rddiv,
1841 bsi->rdstat[i][mode_idx].brate, 0);
1842 bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
1843 bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
1844 if (num_4x4_blocks_wide > 1)
1845 bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
1846 if (num_4x4_blocks_high > 1)
1847 bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
1848 }
1849
1850 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
1851 mode_selected = this_mode;
1852 best_rd = bsi->rdstat[i][mode_idx].brdcost;
1853 }
1854 } /*for each 4x4 mode*/
1855
1856 if (best_rd == INT64_MAX) {
1857 int iy, midx;
1858 for (iy = i + 1; iy < 4; ++iy)
1859 for (midx = 0; midx < INTER_MODES; ++midx)
1860 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
1861 bsi->segment_rd = INT64_MAX;
1862 return INT64_MAX;;
1863 }
1864
1865 mode_idx = INTER_OFFSET(mode_selected);
1866 vpx_memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
1867 vpx_memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
1868
1869 set_and_cost_bmi_mvs(cpi, xd, i, mode_selected, mode_mv[mode_selected],
1870 frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
1871 x->mvcost);
1872
1873 br += bsi->rdstat[i][mode_idx].brate;
1874 bd += bsi->rdstat[i][mode_idx].bdist;
1875 block_sse += bsi->rdstat[i][mode_idx].bsse;
1876 segmentyrate += bsi->rdstat[i][mode_idx].byrate;
1877 this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
1878
1879 if (this_segment_rd > bsi->segment_rd) {
1880 int iy, midx;
1881 for (iy = i + 1; iy < 4; ++iy)
1882 for (midx = 0; midx < INTER_MODES; ++midx)
1883 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
1884 bsi->segment_rd = INT64_MAX;
1885 return INT64_MAX;;
1886 }
1887 }
1888 } /* for each label */
1889
1890 bsi->r = br;
1891 bsi->d = bd;
1892 bsi->segment_yrate = segmentyrate;
1893 bsi->segment_rd = this_segment_rd;
1894 bsi->sse = block_sse;
1895
1896 // update the coding decisions
1897 for (k = 0; k < 4; ++k)
1898 bsi->modes[k] = mi->bmi[k].as_mode;
1899
1900 if (bsi->segment_rd > best_rd)
1901 return INT64_MAX;
1902 /* set it to the best */
1903 for (i = 0; i < 4; i++) {
1904 mode_idx = INTER_OFFSET(bsi->modes[i]);
1905 mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
1906 if (has_second_ref(mbmi))
1907 mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
1908 x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
1909 mi->bmi[i].as_mode = bsi->modes[i];
1910 }
1911
1912 /*
1913 * used to set mbmi->mv.as_int
1914 */
1915 *returntotrate = bsi->r;
1916 *returndistortion = bsi->d;
1917 *returnyrate = bsi->segment_yrate;
1918 *skippable = vp9_is_skippable_in_plane(x, BLOCK_8X8, 0);
1919 *psse = bsi->sse;
1920 mbmi->mode = bsi->modes[3];
1921
1922 return bsi->segment_rd;
1923 }
1924
1925 static void estimate_ref_frame_costs(const VP9_COMMON *cm,
1926 const MACROBLOCKD *xd,
1927 int segment_id,
1928 unsigned int *ref_costs_single,
1929 unsigned int *ref_costs_comp,
1930 vp9_prob *comp_mode_p) {
1931 int seg_ref_active = vp9_segfeature_active(&cm->seg, segment_id,
1932 SEG_LVL_REF_FRAME);
1933 if (seg_ref_active) {
1934 vpx_memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
1935 vpx_memset(ref_costs_comp, 0, MAX_REF_FRAMES * sizeof(*ref_costs_comp));
1936 *comp_mode_p = 128;
1937 } else {
1938 vp9_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
1939 vp9_prob comp_inter_p = 128;
1940
1941 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
1942 comp_inter_p = vp9_get_reference_mode_prob(cm, xd);
1943 *comp_mode_p = comp_inter_p;
1944 } else {
1945 *comp_mode_p = 128;
1946 }
1947
1948 ref_costs_single[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1949
1950 if (cm->reference_mode != COMPOUND_REFERENCE) {
1951 vp9_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1952 vp9_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1953 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
1954
1955 if (cm->reference_mode == REFERENCE_MODE_SELECT)
1956 base_cost += vp9_cost_bit(comp_inter_p, 0);
1957
1958 ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] =
1959 ref_costs_single[ALTREF_FRAME] = base_cost;
1960 ref_costs_single[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1961 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1962 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1963 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1964 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1965 } else {
1966 ref_costs_single[LAST_FRAME] = 512;
1967 ref_costs_single[GOLDEN_FRAME] = 512;
1968 ref_costs_single[ALTREF_FRAME] = 512;
1969 }
1970 if (cm->reference_mode != SINGLE_REFERENCE) {
1971 vp9_prob ref_comp_p = vp9_get_pred_prob_comp_ref_p(cm, xd);
1972 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
1973
1974 if (cm->reference_mode == REFERENCE_MODE_SELECT)
1975 base_cost += vp9_cost_bit(comp_inter_p, 1);
1976
1977 ref_costs_comp[LAST_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 0);
1978 ref_costs_comp[GOLDEN_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 1);
1979 } else {
1980 ref_costs_comp[LAST_FRAME] = 512;
1981 ref_costs_comp[GOLDEN_FRAME] = 512;
1982 }
1983 }
1984 }
1985
1986 static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
1987 int mode_index,
1988 int64_t comp_pred_diff[REFERENCE_MODES],
1989 const int64_t tx_size_diff[TX_MODES],
1990 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS],
1991 int skippable) {
1992 MACROBLOCKD *const xd = &x->e_mbd;
1993
1994 // Take a snapshot of the coding context so it can be
1995 // restored if we decide to encode this way
1996 ctx->skip = x->skip;
1997 ctx->skippable = skippable;
1998 ctx->best_mode_index = mode_index;
1999 ctx->mic = *xd->mi[0].src_mi;
2000 ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
2001 ctx->comp_pred_diff = (int)comp_pred_diff[COMPOUND_REFERENCE];
2002 ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
2003
2004 vpx_memcpy(ctx->tx_rd_diff, tx_size_diff, sizeof(ctx->tx_rd_diff));
2005 vpx_memcpy(ctx->best_filter_diff, best_filter_diff,
2006 sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
2007 }
2008
2009 static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
2010 const TileInfo *const tile,
2011 MV_REFERENCE_FRAME ref_frame,
2012 BLOCK_SIZE block_size,
2013 int mi_row, int mi_col,
2014 int_mv frame_nearest_mv[MAX_REF_FRAMES],
2015 int_mv frame_near_mv[MAX_REF_FRAMES],
2016 struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
2017 const VP9_COMMON *cm = &cpi->common;
2018 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2019 MACROBLOCKD *const xd = &x->e_mbd;
2020 MODE_INFO *const mi = xd->mi[0].src_mi;
2021 int_mv *const candidates = mi->mbmi.ref_mvs[ref_frame];
2022 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2023
2024 // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
2025 // use the UV scaling factors.
2026 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
2027
2028 // Gets an initial list of candidate vectors from neighbours and orders them
2029 vp9_find_mv_refs(cm, xd, tile, mi, ref_frame, candidates, mi_row, mi_col);
2030
2031 // Candidate refinement carried out at encoder and decoder
2032 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2033 &frame_nearest_mv[ref_frame],
2034 &frame_near_mv[ref_frame]);
2035
2036 // Further refinement that is encode side only to test the top few candidates
2037 // in full and choose the best as the centre point for subsequent searches.
2038 // The current implementation doesn't support scaling.
2039 if (!vp9_is_scaled(sf) && block_size >= BLOCK_8X8)
2040 vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
2041 ref_frame, block_size);
2042 }
2043
2044 static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
2045 BLOCK_SIZE bsize,
2046 int mi_row, int mi_col,
2047 int_mv *tmp_mv, int *rate_mv) {
2048 MACROBLOCKD *xd = &x->e_mbd;
2049 const VP9_COMMON *cm = &cpi->common;
2050 MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
2051 struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
2052 int bestsme = INT_MAX;
2053 int step_param;
2054 int sadpb = x->sadperbit16;
2055 MV mvp_full;
2056 int ref = mbmi->ref_frame[0];
2057 MV ref_mv = mbmi->ref_mvs[ref][0].as_mv;
2058
2059 int tmp_col_min = x->mv_col_min;
2060 int tmp_col_max = x->mv_col_max;
2061 int tmp_row_min = x->mv_row_min;
2062 int tmp_row_max = x->mv_row_max;
2063 int cost_list[5];
2064
2065 const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi,
2066 ref);
2067
2068 MV pred_mv[3];
2069 pred_mv[0] = mbmi->ref_mvs[ref][0].as_mv;
2070 pred_mv[1] = mbmi->ref_mvs[ref][1].as_mv;
2071 pred_mv[2] = x->pred_mv[ref];
2072
2073 if (scaled_ref_frame) {
2074 int i;
2075 // Swap out the reference frame for a version that's been scaled to
2076 // match the resolution of the current frame, allowing the existing
2077 // motion search code to be used without additional modifications.
2078 for (i = 0; i < MAX_MB_PLANE; i++)
2079 backup_yv12[i] = xd->plane[i].pre[0];
2080
2081 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
2082 }
2083
2084 vp9_set_mv_search_range(x, &ref_mv);
2085
2086 // Work out the size of the first step in the mv step search.
2087 // 0 here is maximum length first step. 1 is MAX >> 1 etc.
2088 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
2089 // Take wtd average of the step_params based on the last frame's
2090 // max mv magnitude and that based on the best ref mvs of the current
2091 // block for the given reference.
2092 step_param = (vp9_init_search_range(x->max_mv_context[ref]) +
2093 cpi->mv_step_param) / 2;
2094 } else {
2095 step_param = cpi->mv_step_param;
2096 }
2097
2098 if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64) {
2099 int boffset = 2 * (b_width_log2_lookup[BLOCK_64X64] -
2100 MIN(b_height_log2_lookup[bsize], b_width_log2_lookup[bsize]));
2101 step_param = MAX(step_param, boffset);
2102 }
2103
2104 if (cpi->sf.adaptive_motion_search) {
2105 int bwl = b_width_log2_lookup[bsize];
2106 int bhl = b_height_log2_lookup[bsize];
2107 int i;
2108 int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
2109
2110 if (tlevel < 5)
2111 step_param += 2;
2112
2113 for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
2114 if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
2115 x->pred_mv[ref].row = 0;
2116 x->pred_mv[ref].col = 0;
2117 tmp_mv->as_int = INVALID_MV;
2118
2119 if (scaled_ref_frame) {
2120 int i;
2121 for (i = 0; i < MAX_MB_PLANE; i++)
2122 xd->plane[i].pre[0] = backup_yv12[i];
2123 }
2124 return;
2125 }
2126 }
2127 }
2128
2129 mvp_full = pred_mv[x->mv_best_ref_index[ref]];
2130
2131 mvp_full.col >>= 3;
2132 mvp_full.row >>= 3;
2133
2134 bestsme = vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
2135 cond_cost_list(cpi, cost_list),
2136 &ref_mv, &tmp_mv->as_mv, INT_MAX, 1);
2137
2138 x->mv_col_min = tmp_col_min;
2139 x->mv_col_max = tmp_col_max;
2140 x->mv_row_min = tmp_row_min;
2141 x->mv_row_max = tmp_row_max;
2142
2143 if (bestsme < INT_MAX) {
2144 int dis; /* TODO: use dis in distortion calculation later. */
2145 cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
2146 cm->allow_high_precision_mv,
2147 x->errorperbit,
2148 &cpi->fn_ptr[bsize],
2149 cpi->sf.mv.subpel_force_stop,
2150 cpi->sf.mv.subpel_iters_per_step,
2151 cond_cost_list(cpi, cost_list),
2152 x->nmvjointcost, x->mvcost,
2153 &dis, &x->pred_sse[ref], NULL, 0, 0);
2154 }
2155 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
2156 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2157
2158 if (cpi->sf.adaptive_motion_search)
2159 x->pred_mv[ref] = tmp_mv->as_mv;
2160
2161 if (scaled_ref_frame) {
2162 int i;
2163 for (i = 0; i < MAX_MB_PLANE; i++)
2164 xd->plane[i].pre[0] = backup_yv12[i];
2165 }
2166 }
2167
2168 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
2169 BLOCK_SIZE bsize,
2170 int_mv *frame_mv,
2171 int mi_row, int mi_col,
2172 int_mv single_newmv[MAX_REF_FRAMES],
2173 int *rate_mv) {
2174 const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
2175 const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
2176 MACROBLOCKD *xd = &x->e_mbd;
2177 MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
2178 const int refs[2] = { mbmi->ref_frame[0],
2179 mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1] };
2180 int_mv ref_mv[2];
2181 int ite, ref;
2182 // Prediction buffer from second frame.
2183 #if CONFIG_VP9_HIGHBITDEPTH
2184 uint8_t *second_pred;
2185 uint8_t *second_pred_alloc;
2186 #else
2187 uint8_t *second_pred = vpx_memalign(16, pw * ph * sizeof(uint8_t));
2188 #endif // CONFIG_VP9_HIGHBITDEPTH
2189 const InterpKernel *kernel = vp9_get_interp_kernel(mbmi->interp_filter);
2190
2191 // Do joint motion search in compound mode to get more accurate mv.
2192 struct buf_2d backup_yv12[2][MAX_MB_PLANE];
2193 struct buf_2d scaled_first_yv12 = xd->plane[0].pre[0];
2194 int last_besterr[2] = {INT_MAX, INT_MAX};
2195 const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
2196 vp9_get_scaled_ref_frame(cpi, mbmi->ref_frame[0]),
2197 vp9_get_scaled_ref_frame(cpi, mbmi->ref_frame[1])
2198 };
2199 #if CONFIG_VP9_HIGHBITDEPTH
2200 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2201 second_pred_alloc = vpx_memalign(16, pw * ph * sizeof(uint16_t));
2202 second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc);
2203 } else {
2204 second_pred_alloc = vpx_memalign(16, pw * ph * sizeof(uint8_t));
2205 second_pred = second_pred_alloc;
2206 }
2207 #endif // CONFIG_VP9_HIGHBITDEPTH
2208
2209 for (ref = 0; ref < 2; ++ref) {
2210 ref_mv[ref] = mbmi->ref_mvs[refs[ref]][0];
2211
2212 if (scaled_ref_frame[ref]) {
2213 int i;
2214 // Swap out the reference frame for a version that's been scaled to
2215 // match the resolution of the current frame, allowing the existing
2216 // motion search code to be used without additional modifications.
2217 for (i = 0; i < MAX_MB_PLANE; i++)
2218 backup_yv12[ref][i] = xd->plane[i].pre[ref];
2219 vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
2220 NULL);
2221 }
2222
2223 frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
2224 }
2225
2226 // Allow joint search multiple times iteratively for each ref frame
2227 // and break out the search loop if it couldn't find better mv.
2228 for (ite = 0; ite < 4; ite++) {
2229 struct buf_2d ref_yv12[2];
2230 int bestsme = INT_MAX;
2231 int sadpb = x->sadperbit16;
2232 MV tmp_mv;
2233 int search_range = 3;
2234
2235 int tmp_col_min = x->mv_col_min;
2236 int tmp_col_max = x->mv_col_max;
2237 int tmp_row_min = x->mv_row_min;
2238 int tmp_row_max = x->mv_row_max;
2239 int id = ite % 2;
2240
2241 // Initialized here because of compiler problem in Visual Studio.
2242 ref_yv12[0] = xd->plane[0].pre[0];
2243 ref_yv12[1] = xd->plane[0].pre[1];
2244
2245 // Get pred block from second frame.
2246 #if CONFIG_VP9_HIGHBITDEPTH
2247 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2248 vp9_highbd_build_inter_predictor(ref_yv12[!id].buf,
2249 ref_yv12[!id].stride,
2250 second_pred, pw,
2251 &frame_mv[refs[!id]].as_mv,
2252 &xd->block_refs[!id]->sf,
2253 pw, ph, 0,
2254 kernel, MV_PRECISION_Q3,
2255 mi_col * MI_SIZE, mi_row * MI_SIZE,
2256 xd->bd);
2257 } else {
2258 vp9_build_inter_predictor(ref_yv12[!id].buf,
2259 ref_yv12[!id].stride,
2260 second_pred, pw,
2261 &frame_mv[refs[!id]].as_mv,
2262 &xd->block_refs[!id]->sf,
2263 pw, ph, 0,
2264 kernel, MV_PRECISION_Q3,
2265 mi_col * MI_SIZE, mi_row * MI_SIZE);
2266 }
2267 #else
2268 vp9_build_inter_predictor(ref_yv12[!id].buf,
2269 ref_yv12[!id].stride,
2270 second_pred, pw,
2271 &frame_mv[refs[!id]].as_mv,
2272 &xd->block_refs[!id]->sf,
2273 pw, ph, 0,
2274 kernel, MV_PRECISION_Q3,
2275 mi_col * MI_SIZE, mi_row * MI_SIZE);
2276 #endif // CONFIG_VP9_HIGHBITDEPTH
2277
2278 // Compound motion search on first ref frame.
2279 if (id)
2280 xd->plane[0].pre[0] = ref_yv12[id];
2281 vp9_set_mv_search_range(x, &ref_mv[id].as_mv);
2282
2283 // Use mv result from single mode as mvp.
2284 tmp_mv = frame_mv[refs[id]].as_mv;
2285
2286 tmp_mv.col >>= 3;
2287 tmp_mv.row >>= 3;
2288
2289 // Small-range full-pixel motion search
2290 bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb,
2291 search_range,
2292 &cpi->fn_ptr[bsize],
2293 &ref_mv[id].as_mv, second_pred);
2294 if (bestsme < INT_MAX)
2295 bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
2296 second_pred, &cpi->fn_ptr[bsize], 1);
2297
2298 x->mv_col_min = tmp_col_min;
2299 x->mv_col_max = tmp_col_max;
2300 x->mv_row_min = tmp_row_min;
2301 x->mv_row_max = tmp_row_max;
2302
2303 if (bestsme < INT_MAX) {
2304 int dis; /* TODO: use dis in distortion calculation later. */
2305 unsigned int sse;
2306 bestsme = cpi->find_fractional_mv_step(
2307 x, &tmp_mv,
2308 &ref_mv[id].as_mv,
2309 cpi->common.allow_high_precision_mv,
2310 x->errorperbit,
2311 &cpi->fn_ptr[bsize],
2312 0, cpi->sf.mv.subpel_iters_per_step,
2313 NULL,
2314 x->nmvjointcost, x->mvcost,
2315 &dis, &sse, second_pred,
2316 pw, ph);
2317 }
2318
2319 if (id)
2320 xd->plane[0].pre[0] = scaled_first_yv12;
2321
2322 if (bestsme < last_besterr[id]) {
2323 frame_mv[refs[id]].as_mv = tmp_mv;
2324 last_besterr[id] = bestsme;
2325 } else {
2326 break;
2327 }
2328 }
2329
2330 *rate_mv = 0;
2331
2332 for (ref = 0; ref < 2; ++ref) {
2333 if (scaled_ref_frame[ref]) {
2334 // restore the predictor
2335 int i;
2336 for (i = 0; i < MAX_MB_PLANE; i++)
2337 xd->plane[i].pre[ref] = backup_yv12[ref][i];
2338 }
2339
2340 *rate_mv += vp9_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
2341 &mbmi->ref_mvs[refs[ref]][0].as_mv,
2342 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2343 }
2344
2345 #if CONFIG_VP9_HIGHBITDEPTH
2346 vpx_free(second_pred_alloc);
2347 #else
2348 vpx_free(second_pred);
2349 #endif // CONFIG_VP9_HIGHBITDEPTH
2350 }
2351
2352 static INLINE void restore_dst_buf(MACROBLOCKD *xd,
2353 uint8_t *orig_dst[MAX_MB_PLANE],
2354 int orig_dst_stride[MAX_MB_PLANE]) {
2355 int i;
2356 for (i = 0; i < MAX_MB_PLANE; i++) {
2357 xd->plane[i].dst.buf = orig_dst[i];
2358 xd->plane[i].dst.stride = orig_dst_stride[i];
2359 }
2360 }
2361
2362 // In some situations we want to discount tha pparent cost of a new motion
2363 // vector. Where there is a subtle motion field and especially where there is
2364 // low spatial complexity then it can be hard to cover the cost of a new motion
2365 // vector in a single block, even if that motion vector reduces distortion.
2366 // However, once established that vector may be usable through the nearest and
2367 // near mv modes to reduce distortion in subsequent blocks and also improve
2368 // visual quality.
2369 static int discount_newmv_test(const VP9_COMP *cpi,
2370 int this_mode,
2371 int_mv this_mv,
2372 int_mv (*mode_mv)[MAX_REF_FRAMES],
2373 int ref_frame) {
2374 return (!cpi->rc.is_src_frame_alt_ref &&
2375 (this_mode == NEWMV) &&
2376 (this_mv.as_int != 0) &&
2377 ((mode_mv[NEARESTMV][ref_frame].as_int == 0) ||
2378 (mode_mv[NEARESTMV][ref_frame].as_int == INVALID_MV)) &&
2379 ((mode_mv[NEARMV][ref_frame].as_int == 0) ||
2380 (mode_mv[NEARMV][ref_frame].as_int == INVALID_MV)));
2381 }
2382
2383 static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
2384 BLOCK_SIZE bsize,
2385 int64_t txfm_cache[],
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].src_mi->mbmi;
2402 const int is_comp_pred = has_second_ref(mbmi);
2403 const int this_mode = mbmi->mode;
2404 int_mv *frame_mv = mode_mv[this_mode];
2405 int i;
2406 int refs[2] = { mbmi->ref_frame[0],
2407 (mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]) };
2408 int_mv cur_mv[2];
2409 #if CONFIG_VP9_HIGHBITDEPTH
2410 DECLARE_ALIGNED_ARRAY(16, uint16_t, tmp_buf16, MAX_MB_PLANE * 64 * 64);
2411 DECLARE_ALIGNED_ARRAY(16, uint8_t, tmp_buf8, MAX_MB_PLANE * 64 * 64);
2412 uint8_t *tmp_buf;
2413 #else
2414 DECLARE_ALIGNED_ARRAY(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 = tmp_buf8;
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].src_mi->mbmi.interp_filter;
2448 if (xd->left_available)
2449 lf = xd->mi[-1].src_mi->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 &mbmi->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 &mbmi->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].src_mi->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 += MAX((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 three options: Last/Golden, ARF/Last or Golden/ARF, or in other
2532 // words if you present them in that order, the second one is always known
2533 // if the first is known.
2534 //
2535 // Under some circumstances we discount the cost of new mv mode to encourage
2536 // initiation of a motion field.
2537 if (discount_newmv_test(cpi, this_mode, frame_mv[refs[0]],
2538 mode_mv, refs[0])) {
2539 *rate2 += MIN(cost_mv_ref(cpi, this_mode, mbmi->mode_context[refs[0]]),
2540 cost_mv_ref(cpi, NEARESTMV, mbmi->mode_context[refs[0]]));
2541 } else {
2542 *rate2 += cost_mv_ref(cpi, this_mode, mbmi->mode_context[refs[0]]);
2543 }
2544
2545 if (RDCOST(x->rdmult, x->rddiv, *rate2, 0) > ref_best_rd &&
2546 mbmi->mode != NEARESTMV)
2547 return INT64_MAX;
2548
2549 pred_exists = 0;
2550 // Are all MVs integer pel for Y and UV
2551 intpel_mv = !mv_has_subpel(&mbmi->mv[0].as_mv);
2552 if (is_comp_pred)
2553 intpel_mv &= !mv_has_subpel(&mbmi->mv[1].as_mv);
2554
2555 // Search for best switchable filter by checking the variance of
2556 // pred error irrespective of whether the filter will be used
2557 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
2558 filter_cache[i] = INT64_MAX;
2559
2560 if (cm->interp_filter != BILINEAR) {
2561 if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
2562 best_filter = EIGHTTAP;
2563 } else if (best_filter == SWITCHABLE) {
2564 int newbest;
2565 int tmp_rate_sum = 0;
2566 int64_t tmp_dist_sum = 0;
2567
2568 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
2569 int j;
2570 int64_t rs_rd;
2571 int tmp_skip_sb = 0;
2572 int64_t tmp_skip_sse = INT64_MAX;
2573
2574 mbmi->interp_filter = i;
2575 rs = vp9_get_switchable_rate(cpi, xd);
2576 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
2577
2578 if (i > 0 && intpel_mv) {
2579 rd = RDCOST(x->rdmult, x->rddiv, tmp_rate_sum, tmp_dist_sum);
2580 filter_cache[i] = rd;
2581 filter_cache[SWITCHABLE_FILTERS] =
2582 MIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2583 if (cm->interp_filter == SWITCHABLE)
2584 rd += rs_rd;
2585 *mask_filter = MAX(*mask_filter, rd);
2586 } else {
2587 int rate_sum = 0;
2588 int64_t dist_sum = 0;
2589 if (i > 0 && cpi->sf.adaptive_interp_filter_search &&
2590 (cpi->sf.interp_filter_search_mask & (1 << i))) {
2591 rate_sum = INT_MAX;
2592 dist_sum = INT64_MAX;
2593 continue;
2594 }
2595
2596 if ((cm->interp_filter == SWITCHABLE &&
2597 (!i || best_needs_copy)) ||
2598 (cm->interp_filter != SWITCHABLE &&
2599 (cm->interp_filter == mbmi->interp_filter ||
2600 (i == 0 && intpel_mv)))) {
2601 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2602 } else {
2603 for (j = 0; j < MAX_MB_PLANE; j++) {
2604 xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
2605 xd->plane[j].dst.stride = 64;
2606 }
2607 }
2608 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2609 model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum,
2610 &tmp_skip_sb, &tmp_skip_sse);
2611
2612 rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum);
2613 filter_cache[i] = rd;
2614 filter_cache[SWITCHABLE_FILTERS] =
2615 MIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2616 if (cm->interp_filter == SWITCHABLE)
2617 rd += rs_rd;
2618 *mask_filter = MAX(*mask_filter, rd);
2619
2620 if (i == 0 && intpel_mv) {
2621 tmp_rate_sum = rate_sum;
2622 tmp_dist_sum = dist_sum;
2623 }
2624 }
2625
2626 if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2627 if (rd / 2 > ref_best_rd) {
2628 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2629 return INT64_MAX;
2630 }
2631 }
2632 newbest = i == 0 || rd < best_rd;
2633
2634 if (newbest) {
2635 best_rd = rd;
2636 best_filter = mbmi->interp_filter;
2637 if (cm->interp_filter == SWITCHABLE && i && !intpel_mv)
2638 best_needs_copy = !best_needs_copy;
2639 }
2640
2641 if ((cm->interp_filter == SWITCHABLE && newbest) ||
2642 (cm->interp_filter != SWITCHABLE &&
2643 cm->interp_filter == mbmi->interp_filter)) {
2644 pred_exists = 1;
2645 tmp_rd = best_rd;
2646
2647 skip_txfm_sb = tmp_skip_sb;
2648 skip_sse_sb = tmp_skip_sse;
2649 vpx_memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2650 vpx_memcpy(bsse, x->bsse, sizeof(bsse));
2651 }
2652 }
2653 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2654 }
2655 }
2656 // Set the appropriate filter
2657 mbmi->interp_filter = cm->interp_filter != SWITCHABLE ?
2658 cm->interp_filter : best_filter;
2659 rs = cm->interp_filter == SWITCHABLE ? vp9_get_switchable_rate(cpi, xd) : 0;
2660
2661 if (pred_exists) {
2662 if (best_needs_copy) {
2663 // again temporarily set the buffers to local memory to prevent a memcpy
2664 for (i = 0; i < MAX_MB_PLANE; i++) {
2665 xd->plane[i].dst.buf = tmp_buf + i * 64 * 64;
2666 xd->plane[i].dst.stride = 64;
2667 }
2668 }
2669 rd = tmp_rd + RDCOST(x->rdmult, x->rddiv, rs, 0);
2670 } else {
2671 int tmp_rate;
2672 int64_t tmp_dist;
2673 // Handles the special case when a filter that is not in the
2674 // switchable list (ex. bilinear) is indicated at the frame level, or
2675 // skip condition holds.
2676 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2677 model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist,
2678 &skip_txfm_sb, &skip_sse_sb);
2679 rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
2680 vpx_memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2681 vpx_memcpy(bsse, x->bsse, sizeof(bsse));
2682 }
2683
2684 if (!is_comp_pred)
2685 single_filter[this_mode][refs[0]] = mbmi->interp_filter;
2686
2687 if (cpi->sf.adaptive_mode_search)
2688 if (is_comp_pred)
2689 if (single_skippable[this_mode][refs[0]] &&
2690 single_skippable[this_mode][refs[1]])
2691 vpx_memset(skip_txfm, 1, sizeof(skip_txfm));
2692
2693 if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2694 // if current pred_error modeled rd is substantially more than the best
2695 // so far, do not bother doing full rd
2696 if (rd / 2 > ref_best_rd) {
2697 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2698 return INT64_MAX;
2699 }
2700 }
2701
2702 if (cm->interp_filter == SWITCHABLE)
2703 *rate2 += rs;
2704
2705 vpx_memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm));
2706 vpx_memcpy(x->bsse, bsse, sizeof(bsse));
2707
2708 if (!skip_txfm_sb) {
2709 int skippable_y, skippable_uv;
2710 int64_t sseuv = INT64_MAX;
2711 int64_t rdcosty = INT64_MAX;
2712
2713 // Y cost and distortion
2714 vp9_subtract_plane(x, bsize, 0);
2715 super_block_yrd(cpi, x, rate_y, &distortion_y, &skippable_y, psse,
2716 bsize, txfm_cache, ref_best_rd);
2717
2718 if (*rate_y == INT_MAX) {
2719 *rate2 = INT_MAX;
2720 *distortion = INT64_MAX;
2721 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2722 return INT64_MAX;
2723 }
2724
2725 *rate2 += *rate_y;
2726 *distortion += distortion_y;
2727
2728 rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
2729 rdcosty = MIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
2730
2731 if (!super_block_uvrd(cpi, x, rate_uv, &distortion_uv, &skippable_uv,
2732 &sseuv, bsize, ref_best_rd - rdcosty)) {
2733 *rate2 = INT_MAX;
2734 *distortion = INT64_MAX;
2735 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2736 return INT64_MAX;
2737 }
2738
2739 *psse += sseuv;
2740 *rate2 += *rate_uv;
2741 *distortion += distortion_uv;
2742 *skippable = skippable_y && skippable_uv;
2743 } else {
2744 x->skip = 1;
2745 *disable_skip = 1;
2746
2747 // The cost of skip bit needs to be added.
2748 *rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2749
2750 *distortion = skip_sse_sb;
2751 }
2752
2753 if (!is_comp_pred)
2754 single_skippable[this_mode][refs[0]] = *skippable;
2755
2756 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2757 return 0; // The rate-distortion cost will be re-calculated by caller.
2758 }
2759
2760 void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
2761 RD_COST *rd_cost, BLOCK_SIZE bsize,
2762 PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
2763 VP9_COMMON *const cm = &cpi->common;
2764 MACROBLOCKD *const xd = &x->e_mbd;
2765 struct macroblockd_plane *const pd = xd->plane;
2766 int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
2767 int y_skip = 0, uv_skip = 0;
2768 int64_t dist_y = 0, dist_uv = 0, tx_cache[TX_MODES] = { 0 };
2769 TX_SIZE max_uv_tx_size;
2770 x->skip_encode = 0;
2771 ctx->skip = 0;
2772 xd->mi[0].src_mi->mbmi.ref_frame[0] = INTRA_FRAME;
2773 xd->mi[0].src_mi->mbmi.ref_frame[1] = NONE;
2774
2775 if (bsize >= BLOCK_8X8) {
2776 if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2777 &dist_y, &y_skip, bsize, tx_cache,
2778 best_rd) >= best_rd) {
2779 rd_cost->rate = INT_MAX;
2780 return;
2781 }
2782 } else {
2783 y_skip = 0;
2784 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2785 &dist_y, best_rd) >= best_rd) {
2786 rd_cost->rate = INT_MAX;
2787 return;
2788 }
2789 }
2790 max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0].src_mi->mbmi.tx_size, bsize,
2791 pd[1].subsampling_x,
2792 pd[1].subsampling_y);
2793 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
2794 &dist_uv, &uv_skip, MAX(BLOCK_8X8, bsize),
2795 max_uv_tx_size);
2796
2797 if (y_skip && uv_skip) {
2798 rd_cost->rate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
2799 vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2800 rd_cost->dist = dist_y + dist_uv;
2801 vp9_zero(ctx->tx_rd_diff);
2802 } else {
2803 int i;
2804 rd_cost->rate = rate_y + rate_uv +
2805 vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
2806 rd_cost->dist = dist_y + dist_uv;
2807 if (cpi->sf.tx_size_search_method == USE_FULL_RD)
2808 for (i = 0; i < TX_MODES; i++) {
2809 if (tx_cache[i] < INT64_MAX && tx_cache[cm->tx_mode] < INT64_MAX)
2810 ctx->tx_rd_diff[i] = tx_cache[i] - tx_cache[cm->tx_mode];
2811 else
2812 ctx->tx_rd_diff[i] = 0;
2813 }
2814 }
2815
2816 ctx->mic = *xd->mi[0].src_mi;
2817 rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
2818 }
2819
2820 void vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi,
2821 TileDataEnc *tile_data,
2822 MACROBLOCK *x,
2823 int mi_row, int mi_col,
2824 RD_COST *rd_cost, BLOCK_SIZE bsize,
2825 PICK_MODE_CONTEXT *ctx,
2826 int64_t best_rd_so_far) {
2827 VP9_COMMON *const cm = &cpi->common;
2828 TileInfo *const tile_info = &tile_data->tile_info;
2829 RD_OPT *const rd_opt = &cpi->rd;
2830 SPEED_FEATURES *const sf = &cpi->sf;
2831 MACROBLOCKD *const xd = &x->e_mbd;
2832 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
2833 const struct segmentation *const seg = &cm->seg;
2834 PREDICTION_MODE this_mode;
2835 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
2836 unsigned char segment_id = mbmi->segment_id;
2837 int comp_pred, i, k;
2838 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
2839 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2840 int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
2841 INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES];
2842 int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES];
2843 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2844 VP9_ALT_FLAG };
2845 int64_t best_rd = best_rd_so_far;
2846 int64_t best_tx_rd[TX_MODES];
2847 int64_t best_tx_diff[TX_MODES];
2848 int64_t best_pred_diff[REFERENCE_MODES];
2849 int64_t best_pred_rd[REFERENCE_MODES];
2850 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
2851 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
2852 MB_MODE_INFO best_mbmode;
2853 int best_mode_skippable = 0;
2854 int midx, best_mode_index = -1;
2855 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
2856 vp9_prob comp_mode_p;
2857 int64_t best_intra_rd = INT64_MAX;
2858 unsigned int best_pred_sse = UINT_MAX;
2859 PREDICTION_MODE best_intra_mode = DC_PRED;
2860 int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
2861 int64_t dist_uv[TX_SIZES];
2862 int skip_uv[TX_SIZES];
2863 PREDICTION_MODE mode_uv[TX_SIZES];
2864 const int intra_cost_penalty = vp9_get_intra_cost_penalty(
2865 cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
2866 int best_skip2 = 0;
2867 uint8_t ref_frame_skip_mask[2] = { 0 };
2868 uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 };
2869 int mode_skip_start = sf->mode_skip_start + 1;
2870 const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
2871 const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
2872 int64_t mode_threshold[MAX_MODES];
2873 int *mode_map = tile_data->mode_map[bsize];
2874 const int mode_search_skip_flags = sf->mode_search_skip_flags;
2875 int64_t mask_filter = 0;
2876 int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
2877
2878 vp9_zero(best_mbmode);
2879
2880 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
2881
2882 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
2883 filter_cache[i] = INT64_MAX;
2884
2885 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
2886 &comp_mode_p);
2887
2888 for (i = 0; i < REFERENCE_MODES; ++i)
2889 best_pred_rd[i] = INT64_MAX;
2890 for (i = 0; i < TX_MODES; i++)
2891 best_tx_rd[i] = INT64_MAX;
2892 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
2893 best_filter_rd[i] = INT64_MAX;
2894 for (i = 0; i < TX_SIZES; i++)
2895 rate_uv_intra[i] = INT_MAX;
2896 for (i = 0; i < MAX_REF_FRAMES; ++i)
2897 x->pred_sse[i] = INT_MAX;
2898 for (i = 0; i < MB_MODE_COUNT; ++i) {
2899 for (k = 0; k < MAX_REF_FRAMES; ++k) {
2900 single_inter_filter[i][k] = SWITCHABLE;
2901 single_skippable[i][k] = 0;
2902 }
2903 }
2904
2905 rd_cost->rate = INT_MAX;
2906
2907 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
2908 x->pred_mv_sad[ref_frame] = INT_MAX;
2909 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
2910 setup_buffer_inter(cpi, x, tile_info, ref_frame, bsize, mi_row, mi_col,
2911 frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
2912 }
2913 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
2914 frame_mv[ZEROMV][ref_frame].as_int = 0;
2915 }
2916
2917 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
2918 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) {
2919 // Skip checking missing references in both single and compound reference
2920 // modes. Note that a mode will be skipped iff both reference frames
2921 // are masked out.
2922 ref_frame_skip_mask[0] |= (1 << ref_frame);
2923 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
2924 } else if (sf->reference_masking) {
2925 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
2926 // Skip fixed mv modes for poor references
2927 if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
2928 mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
2929 break;
2930 }
2931 }
2932 }
2933 // If the segment reference frame feature is enabled....
2934 // then do nothing if the current ref frame is not allowed..
2935 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
2936 vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
2937 ref_frame_skip_mask[0] |= (1 << ref_frame);
2938 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
2939 }
2940 }
2941
2942 // Disable this drop out case if the ref frame
2943 // segment level feature is enabled for this segment. This is to
2944 // prevent the possibility that we end up unable to pick any mode.
2945 if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
2946 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
2947 // unless ARNR filtering is enabled in which case we want
2948 // an unfiltered alternative. We allow near/nearest as well
2949 // because they may result in zero-zero MVs but be cheaper.
2950 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
2951 ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME);
2952 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
2953 mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
2954 if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
2955 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
2956 if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
2957 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
2958 }
2959 }
2960
2961 if (cpi->rc.is_src_frame_alt_ref) {
2962 if (sf->alt_ref_search_fp) {
2963 mode_skip_mask[ALTREF_FRAME] = 0;
2964 ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME);
2965 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
2966 }
2967 }
2968
2969 if (sf->alt_ref_search_fp)
2970 if (!cm->show_frame && x->pred_mv_sad[GOLDEN_FRAME] < INT_MAX)
2971 if (x->pred_mv_sad[ALTREF_FRAME] > (x->pred_mv_sad[GOLDEN_FRAME] << 1))
2972 mode_skip_mask[ALTREF_FRAME] |= INTER_ALL;
2973
2974 if (sf->adaptive_mode_search) {
2975 if (cm->show_frame && !cpi->rc.is_src_frame_alt_ref &&
2976 cpi->rc.frames_since_golden >= 3)
2977 if (x->pred_mv_sad[GOLDEN_FRAME] > (x->pred_mv_sad[LAST_FRAME] << 1))
2978 mode_skip_mask[GOLDEN_FRAME] |= INTER_ALL;
2979 }
2980
2981 if (bsize > sf->max_intra_bsize) {
2982 ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
2983 ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
2984 }
2985
2986 mode_skip_mask[INTRA_FRAME] |=
2987 ~(sf->intra_y_mode_mask[max_txsize_lookup[bsize]]);
2988
2989 for (i = 0; i <= LAST_NEW_MV_INDEX; ++i)
2990 mode_threshold[i] = 0;
2991 for (i = LAST_NEW_MV_INDEX + 1; i < MAX_MODES; ++i)
2992 mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
2993
2994 midx = sf->schedule_mode_search ? mode_skip_start : 0;
2995 while (midx > 4) {
2996 uint8_t end_pos = 0;
2997 for (i = 5; i < midx; ++i) {
2998 if (mode_threshold[mode_map[i - 1]] > mode_threshold[mode_map[i]]) {
2999 uint8_t tmp = mode_map[i];
3000 mode_map[i] = mode_map[i - 1];
3001 mode_map[i - 1] = tmp;
3002 end_pos = i;
3003 }
3004 }
3005 midx = end_pos;
3006 }
3007
3008 for (midx = 0; midx < MAX_MODES; ++midx) {
3009 int mode_index = mode_map[midx];
3010 int mode_excluded = 0;
3011 int64_t this_rd = INT64_MAX;
3012 int disable_skip = 0;
3013 int compmode_cost = 0;
3014 int rate2 = 0, rate_y = 0, rate_uv = 0;
3015 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3016 int skippable = 0;
3017 int64_t tx_cache[TX_MODES];
3018 int this_skip2 = 0;
3019 int64_t total_sse = INT64_MAX;
3020 int early_term = 0;
3021
3022 this_mode = vp9_mode_order[mode_index].mode;
3023 ref_frame = vp9_mode_order[mode_index].ref_frame[0];
3024 second_ref_frame = vp9_mode_order[mode_index].ref_frame[1];
3025
3026 // Look at the reference frame of the best mode so far and set the
3027 // skip mask to look at a subset of the remaining modes.
3028 if (midx == mode_skip_start && best_mode_index >= 0) {
3029 switch (best_mbmode.ref_frame[0]) {
3030 case INTRA_FRAME:
3031 break;
3032 case LAST_FRAME:
3033 ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK;
3034 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3035 break;
3036 case GOLDEN_FRAME:
3037 ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
3038 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3039 break;
3040 case ALTREF_FRAME:
3041 ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK;
3042 break;
3043 case NONE:
3044 case MAX_REF_FRAMES:
3045 assert(0 && "Invalid Reference frame");
3046 break;
3047 }
3048 }
3049
3050 if (ref_frame_skip_mask[0] & (1 << ref_frame) &&
3051 ref_frame_skip_mask[1] & (1 << MAX(0, second_ref_frame)))
3052 continue;
3053
3054 if (mode_skip_mask[ref_frame] & (1 << this_mode))
3055 continue;
3056
3057 // Test best rd so far against threshold for trying this mode.
3058 if (best_mode_skippable && sf->schedule_mode_search)
3059 mode_threshold[mode_index] <<= 1;
3060
3061 if (best_rd < mode_threshold[mode_index])
3062 continue;
3063
3064 if (sf->motion_field_mode_search) {
3065 const int mi_width = MIN(num_8x8_blocks_wide_lookup[bsize],
3066 tile_info->mi_col_end - mi_col);
3067 const int mi_height = MIN(num_8x8_blocks_high_lookup[bsize],
3068 tile_info->mi_row_end - mi_row);
3069 const int bsl = mi_width_log2_lookup[bsize];
3070 int cb_partition_search_ctrl = (((mi_row + mi_col) >> bsl)
3071 + get_chessboard_index(cm->current_video_frame)) & 0x1;
3072 MB_MODE_INFO *ref_mbmi;
3073 int const_motion = 1;
3074 int skip_ref_frame = !cb_partition_search_ctrl;
3075 MV_REFERENCE_FRAME rf = NONE;
3076 int_mv ref_mv;
3077 ref_mv.as_int = INVALID_MV;
3078
3079 if ((mi_row - 1) >= tile_info->mi_row_start) {
3080 ref_mv = xd->mi[-xd->mi_stride].src_mi->mbmi.mv[0];
3081 rf = xd->mi[-xd->mi_stride].src_mi->mbmi.ref_frame[0];
3082 for (i = 0; i < mi_width; ++i) {
3083 ref_mbmi = &xd->mi[-xd->mi_stride + i].src_mi->mbmi;
3084 const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
3085 (ref_frame == ref_mbmi->ref_frame[0]);
3086 skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
3087 }
3088 }
3089
3090 if ((mi_col - 1) >= tile_info->mi_col_start) {
3091 if (ref_mv.as_int == INVALID_MV)
3092 ref_mv = xd->mi[-1].src_mi->mbmi.mv[0];
3093 if (rf == NONE)
3094 rf = xd->mi[-1].src_mi->mbmi.ref_frame[0];
3095 for (i = 0; i < mi_height; ++i) {
3096 ref_mbmi = &xd->mi[i * xd->mi_stride - 1].src_mi->mbmi;
3097 const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
3098 (ref_frame == ref_mbmi->ref_frame[0]);
3099 skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
3100 }
3101 }
3102
3103 if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV)
3104 if (rf > INTRA_FRAME)
3105 if (ref_frame != rf)
3106 continue;
3107
3108 if (const_motion)
3109 if (this_mode == NEARMV || this_mode == ZEROMV)
3110 continue;
3111 }
3112
3113 comp_pred = second_ref_frame > INTRA_FRAME;
3114 if (comp_pred) {
3115 if (!cpi->allow_comp_inter_inter)
3116 continue;
3117
3118 // Skip compound inter modes if ARF is not available.
3119 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
3120 continue;
3121
3122 // Do not allow compound prediction if the segment level reference frame
3123 // feature is in use as in this case there can only be one reference.
3124 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
3125 continue;
3126
3127 if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3128 best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
3129 continue;
3130
3131 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3132 } else {
3133 if (ref_frame != INTRA_FRAME)
3134 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3135 }
3136
3137 if (ref_frame == INTRA_FRAME) {
3138 if (sf->adaptive_mode_search)
3139 if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
3140 continue;
3141
3142 if (this_mode != DC_PRED) {
3143 // Disable intra modes other than DC_PRED for blocks with low variance
3144 // Threshold for intra skipping based on source variance
3145 // TODO(debargha): Specialize the threshold for super block sizes
3146 const unsigned int skip_intra_var_thresh = 64;
3147 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
3148 x->source_variance < skip_intra_var_thresh)
3149 continue;
3150 // Only search the oblique modes if the best so far is
3151 // one of the neighboring directional modes
3152 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
3153 (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
3154 if (best_mode_index >= 0 &&
3155 best_mbmode.ref_frame[0] > INTRA_FRAME)
3156 continue;
3157 }
3158 if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
3159 if (conditional_skipintra(this_mode, best_intra_mode))
3160 continue;
3161 }
3162 }
3163 } else {
3164 const MV_REFERENCE_FRAME ref_frames[2] = {ref_frame, second_ref_frame};
3165 if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv,
3166 this_mode, ref_frames))
3167 continue;
3168 }
3169
3170 mbmi->mode = this_mode;
3171 mbmi->uv_mode = DC_PRED;
3172 mbmi->ref_frame[0] = ref_frame;
3173 mbmi->ref_frame[1] = second_ref_frame;
3174 // Evaluate all sub-pel filters irrespective of whether we can use
3175 // them for this frame.
3176 mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
3177 : cm->interp_filter;
3178 mbmi->mv[0].as_int = mbmi->mv[1].as_int = 0;
3179
3180 x->skip = 0;
3181 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3182
3183 // Select prediction reference frames.
3184 for (i = 0; i < MAX_MB_PLANE; i++) {
3185 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3186 if (comp_pred)
3187 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3188 }
3189
3190 for (i = 0; i < TX_MODES; ++i)
3191 tx_cache[i] = INT64_MAX;
3192
3193 if (ref_frame == INTRA_FRAME) {
3194 TX_SIZE uv_tx;
3195 struct macroblockd_plane *const pd = &xd->plane[1];
3196 vpx_memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
3197 super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable,
3198 NULL, bsize, tx_cache, best_rd);
3199 if (rate_y == INT_MAX)
3200 continue;
3201
3202 uv_tx = get_uv_tx_size_impl(mbmi->tx_size, bsize, pd->subsampling_x,
3203 pd->subsampling_y);
3204 if (rate_uv_intra[uv_tx] == INT_MAX) {
3205 choose_intra_uv_mode(cpi, x, ctx, bsize, uv_tx,
3206 &rate_uv_intra[uv_tx], &rate_uv_tokenonly[uv_tx],
3207 &dist_uv[uv_tx], &skip_uv[uv_tx], &mode_uv[uv_tx]);
3208 }
3209
3210 rate_uv = rate_uv_tokenonly[uv_tx];
3211 distortion_uv = dist_uv[uv_tx];
3212 skippable = skippable && skip_uv[uv_tx];
3213 mbmi->uv_mode = mode_uv[uv_tx];
3214
3215 rate2 = rate_y + cpi->mbmode_cost[mbmi->mode] + rate_uv_intra[uv_tx];
3216 if (this_mode != DC_PRED && this_mode != TM_PRED)
3217 rate2 += intra_cost_penalty;
3218 distortion2 = distortion_y + distortion_uv;
3219 } else {
3220 this_rd = handle_inter_mode(cpi, x, bsize,
3221 tx_cache,
3222 &rate2, &distortion2, &skippable,
3223 &rate_y, &rate_uv,
3224 &disable_skip, frame_mv,
3225 mi_row, mi_col,
3226 single_newmv, single_inter_filter,
3227 single_skippable, &total_sse, best_rd,
3228 &mask_filter, filter_cache);
3229 if (this_rd == INT64_MAX)
3230 continue;
3231
3232 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
3233
3234 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3235 rate2 += compmode_cost;
3236 }
3237
3238 // Estimate the reference frame signaling cost and add it
3239 // to the rolling cost variable.
3240 if (comp_pred) {
3241 rate2 += ref_costs_comp[ref_frame];
3242 } else {
3243 rate2 += ref_costs_single[ref_frame];
3244 }
3245
3246 if (!disable_skip) {
3247 if (skippable) {
3248 // Back out the coefficient coding costs
3249 rate2 -= (rate_y + rate_uv);
3250
3251 // Cost the skip mb case
3252 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
3253 } else if (ref_frame != INTRA_FRAME && !xd->lossless) {
3254 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
3255 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
3256 // Add in the cost of the no skip flag.
3257 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
3258 } else {
3259 // FIXME(rbultje) make this work for splitmv also
3260 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
3261 distortion2 = total_sse;
3262 assert(total_sse >= 0);
3263 rate2 -= (rate_y + rate_uv);
3264 this_skip2 = 1;
3265 }
3266 } else {
3267 // Add in the cost of the no skip flag.
3268 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
3269 }
3270
3271 // Calculate the final RD estimate for this mode.
3272 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3273 }
3274
3275 if (ref_frame == INTRA_FRAME) {
3276 // Keep record of best intra rd
3277 if (this_rd < best_intra_rd) {
3278 best_intra_rd = this_rd;
3279 best_intra_mode = mbmi->mode;
3280 }
3281 }
3282
3283 if (!disable_skip && ref_frame == INTRA_FRAME) {
3284 for (i = 0; i < REFERENCE_MODES; ++i)
3285 best_pred_rd[i] = MIN(best_pred_rd[i], this_rd);
3286 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3287 best_filter_rd[i] = MIN(best_filter_rd[i], this_rd);
3288 }
3289
3290 // Did this mode help.. i.e. is it the new best mode
3291 if (this_rd < best_rd || x->skip) {
3292 int max_plane = MAX_MB_PLANE;
3293 if (!mode_excluded) {
3294 // Note index of best mode so far
3295 best_mode_index = mode_index;
3296
3297 if (ref_frame == INTRA_FRAME) {
3298 /* required for left and above block mv */
3299 mbmi->mv[0].as_int = 0;
3300 max_plane = 1;
3301 } else {
3302 best_pred_sse = x->pred_sse[ref_frame];
3303 }
3304
3305 rd_cost->rate = rate2;
3306 rd_cost->dist = distortion2;
3307 rd_cost->rdcost = this_rd;
3308 best_rd = this_rd;
3309 best_mbmode = *mbmi;
3310 best_skip2 = this_skip2;
3311 best_mode_skippable = skippable;
3312
3313 if (!x->select_tx_size)
3314 swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
3315 vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size],
3316 sizeof(uint8_t) * ctx->num_4x4_blk);
3317
3318 // TODO(debargha): enhance this test with a better distortion prediction
3319 // based on qp, activity mask and history
3320 if ((mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
3321 (mode_index > MIN_EARLY_TERM_INDEX)) {
3322 int qstep = xd->plane[0].dequant[1];
3323 // TODO(debargha): Enhance this by specializing for each mode_index
3324 int scale = 4;
3325 #if CONFIG_VP9_HIGHBITDEPTH
3326 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
3327 qstep >>= (xd->bd - 8);
3328 }
3329 #endif // CONFIG_VP9_HIGHBITDEPTH
3330 if (x->source_variance < UINT_MAX) {
3331 const int var_adjust = (x->source_variance < 16);
3332 scale -= var_adjust;
3333 }
3334 if (ref_frame > INTRA_FRAME &&
3335 distortion2 * scale < qstep * qstep) {
3336 early_term = 1;
3337 }
3338 }
3339 }
3340 }
3341
3342 /* keep record of best compound/single-only prediction */
3343 if (!disable_skip && ref_frame != INTRA_FRAME) {
3344 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
3345
3346 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3347 single_rate = rate2 - compmode_cost;
3348 hybrid_rate = rate2;
3349 } else {
3350 single_rate = rate2;
3351 hybrid_rate = rate2 + compmode_cost;
3352 }
3353
3354 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
3355 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
3356
3357 if (!comp_pred) {
3358 if (single_rd < best_pred_rd[SINGLE_REFERENCE])
3359 best_pred_rd[SINGLE_REFERENCE] = single_rd;
3360 } else {
3361 if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
3362 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
3363 }
3364 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
3365 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
3366
3367 /* keep record of best filter type */
3368 if (!mode_excluded && cm->interp_filter != BILINEAR) {
3369 int64_t ref = filter_cache[cm->interp_filter == SWITCHABLE ?
3370 SWITCHABLE_FILTERS : cm->interp_filter];
3371
3372 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3373 int64_t adj_rd;
3374 if (ref == INT64_MAX)
3375 adj_rd = 0;
3376 else if (filter_cache[i] == INT64_MAX)
3377 // when early termination is triggered, the encoder does not have
3378 // access to the rate-distortion cost. it only knows that the cost
3379 // should be above the maximum valid value. hence it takes the known
3380 // maximum plus an arbitrary constant as the rate-distortion cost.
3381 adj_rd = mask_filter - ref + 10;
3382 else
3383 adj_rd = filter_cache[i] - ref;
3384
3385 adj_rd += this_rd;
3386 best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd);
3387 }
3388 }
3389 }
3390
3391 /* keep record of best txfm size */
3392 if (bsize < BLOCK_32X32) {
3393 if (bsize < BLOCK_16X16)
3394 tx_cache[ALLOW_16X16] = tx_cache[ALLOW_8X8];
3395
3396 tx_cache[ALLOW_32X32] = tx_cache[ALLOW_16X16];
3397 }
3398 if (!mode_excluded && this_rd != INT64_MAX) {
3399 for (i = 0; i < TX_MODES && tx_cache[i] < INT64_MAX; i++) {
3400 int64_t adj_rd = INT64_MAX;
3401 adj_rd = this_rd + tx_cache[i] - tx_cache[cm->tx_mode];
3402
3403 if (adj_rd < best_tx_rd[i])
3404 best_tx_rd[i] = adj_rd;
3405 }
3406 }
3407
3408 if (early_term)
3409 break;
3410
3411 if (x->skip && !comp_pred)
3412 break;
3413 }
3414
3415 // The inter modes' rate costs are not calculated precisely in some cases.
3416 // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and
3417 // ZEROMV. Here, checks are added for those cases, and the mode decisions
3418 // are corrected.
3419 if (best_mbmode.mode == NEWMV) {
3420 const MV_REFERENCE_FRAME refs[2] = {best_mbmode.ref_frame[0],
3421 best_mbmode.ref_frame[1]};
3422 int comp_pred_mode = refs[1] > INTRA_FRAME;
3423
3424 if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3425 ((comp_pred_mode && frame_mv[NEARESTMV][refs[1]].as_int ==
3426 best_mbmode.mv[1].as_int) || !comp_pred_mode))
3427 best_mbmode.mode = NEARESTMV;
3428 else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3429 ((comp_pred_mode && frame_mv[NEARMV][refs[1]].as_int ==
3430 best_mbmode.mv[1].as_int) || !comp_pred_mode))
3431 best_mbmode.mode = NEARMV;
3432 else if (best_mbmode.mv[0].as_int == 0 &&
3433 ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) || !comp_pred_mode))
3434 best_mbmode.mode = ZEROMV;
3435 }
3436
3437 if (best_mode_index < 0 || best_rd >= best_rd_so_far) {
3438 rd_cost->rate = INT_MAX;
3439 rd_cost->rdcost = INT64_MAX;
3440 return;
3441 }
3442
3443 // If we used an estimate for the uv intra rd in the loop above...
3444 if (sf->use_uv_intra_rd_estimate) {
3445 // Do Intra UV best rd mode selection if best mode choice above was intra.
3446 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
3447 TX_SIZE uv_tx_size;
3448 *mbmi = best_mbmode;
3449 uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
3450 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
3451 &rate_uv_tokenonly[uv_tx_size],
3452 &dist_uv[uv_tx_size],
3453 &skip_uv[uv_tx_size],
3454 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
3455 uv_tx_size);
3456 }
3457 }
3458
3459 assert((cm->interp_filter == SWITCHABLE) ||
3460 (cm->interp_filter == best_mbmode.interp_filter) ||
3461 !is_inter_block(&best_mbmode));
3462
3463 if (!cpi->rc.is_src_frame_alt_ref)
3464 vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3465 sf->adaptive_rd_thresh, bsize, best_mode_index);
3466
3467 // macroblock modes
3468 *mbmi = best_mbmode;
3469 x->skip |= best_skip2;
3470
3471 for (i = 0; i < REFERENCE_MODES; ++i) {
3472 if (best_pred_rd[i] == INT64_MAX)
3473 best_pred_diff[i] = INT_MIN;
3474 else
3475 best_pred_diff[i] = best_rd - best_pred_rd[i];
3476 }
3477
3478 if (!x->skip) {
3479 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3480 if (best_filter_rd[i] == INT64_MAX)
3481 best_filter_diff[i] = 0;
3482 else
3483 best_filter_diff[i] = best_rd - best_filter_rd[i];
3484 }
3485 if (cm->interp_filter == SWITCHABLE)
3486 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
3487 for (i = 0; i < TX_MODES; i++) {
3488 if (best_tx_rd[i] == INT64_MAX)
3489 best_tx_diff[i] = 0;
3490 else
3491 best_tx_diff[i] = best_rd - best_tx_rd[i];
3492 }
3493 } else {
3494 vp9_zero(best_filter_diff);
3495 vp9_zero(best_tx_diff);
3496 }
3497
3498 // TODO(yunqingwang): Moving this line in front of the above best_filter_diff
3499 // updating code causes PSNR loss. Need to figure out the confliction.
3500 x->skip |= best_mode_skippable;
3501
3502 if (!x->skip && !x->select_tx_size) {
3503 int has_high_freq_coeff = 0;
3504 int plane;
3505 int max_plane = is_inter_block(&xd->mi[0].src_mi->mbmi)
3506 ? MAX_MB_PLANE : 1;
3507 for (plane = 0; plane < max_plane; ++plane) {
3508 x->plane[plane].eobs = ctx->eobs_pbuf[plane][1];
3509 has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
3510 }
3511
3512 for (plane = max_plane; plane < MAX_MB_PLANE; ++plane) {
3513 x->plane[plane].eobs = ctx->eobs_pbuf[plane][2];
3514 has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
3515 }
3516
3517 best_mode_skippable |= !has_high_freq_coeff;
3518 }
3519
3520 store_coding_context(x, ctx, best_mode_index, best_pred_diff,
3521 best_tx_diff, best_filter_diff, best_mode_skippable);
3522 }
3523
3524 void vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi,
3525 TileDataEnc *tile_data,
3526 MACROBLOCK *x,
3527 RD_COST *rd_cost,
3528 BLOCK_SIZE bsize,
3529 PICK_MODE_CONTEXT *ctx,
3530 int64_t best_rd_so_far) {
3531 VP9_COMMON *const cm = &cpi->common;
3532 MACROBLOCKD *const xd = &x->e_mbd;
3533 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
3534 unsigned char segment_id = mbmi->segment_id;
3535 const int comp_pred = 0;
3536 int i;
3537 int64_t best_tx_diff[TX_MODES];
3538 int64_t best_pred_diff[REFERENCE_MODES];
3539 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3540 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3541 vp9_prob comp_mode_p;
3542 INTERP_FILTER best_filter = SWITCHABLE;
3543 int64_t this_rd = INT64_MAX;
3544 int rate2 = 0;
3545 const int64_t distortion2 = 0;
3546
3547 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3548
3549 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3550 &comp_mode_p);
3551
3552 for (i = 0; i < MAX_REF_FRAMES; ++i)
3553 x->pred_sse[i] = INT_MAX;
3554 for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i)
3555 x->pred_mv_sad[i] = INT_MAX;
3556
3557 rd_cost->rate = INT_MAX;
3558
3559 assert(vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
3560
3561 mbmi->mode = ZEROMV;
3562 mbmi->uv_mode = DC_PRED;
3563 mbmi->ref_frame[0] = LAST_FRAME;
3564 mbmi->ref_frame[1] = NONE;
3565 mbmi->mv[0].as_int = 0;
3566 x->skip = 1;
3567
3568 if (cm->interp_filter != BILINEAR) {
3569 best_filter = EIGHTTAP;
3570 if (cm->interp_filter == SWITCHABLE &&
3571 x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
3572 int rs;
3573 int best_rs = INT_MAX;
3574 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
3575 mbmi->interp_filter = i;
3576 rs = vp9_get_switchable_rate(cpi, xd);
3577 if (rs < best_rs) {
3578 best_rs = rs;
3579 best_filter = mbmi->interp_filter;
3580 }
3581 }
3582 }
3583 }
3584 // Set the appropriate filter
3585 if (cm->interp_filter == SWITCHABLE) {
3586 mbmi->interp_filter = best_filter;
3587 rate2 += vp9_get_switchable_rate(cpi, xd);
3588 } else {
3589 mbmi->interp_filter = cm->interp_filter;
3590 }
3591
3592 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3593 rate2 += vp9_cost_bit(comp_mode_p, comp_pred);
3594
3595 // Estimate the reference frame signaling cost and add it
3596 // to the rolling cost variable.
3597 rate2 += ref_costs_single[LAST_FRAME];
3598 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3599
3600 rd_cost->rate = rate2;
3601 rd_cost->dist = distortion2;
3602 rd_cost->rdcost = this_rd;
3603
3604 if (this_rd >= best_rd_so_far) {
3605 rd_cost->rate = INT_MAX;
3606 rd_cost->rdcost = INT64_MAX;
3607 return;
3608 }
3609
3610 assert((cm->interp_filter == SWITCHABLE) ||
3611 (cm->interp_filter == mbmi->interp_filter));
3612
3613 vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3614 cpi->sf.adaptive_rd_thresh, bsize, THR_ZEROMV);
3615
3616 vp9_zero(best_pred_diff);
3617 vp9_zero(best_filter_diff);
3618 vp9_zero(best_tx_diff);
3619
3620 if (!x->select_tx_size)
3621 swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE);
3622 store_coding_context(x, ctx, THR_ZEROMV,
3623 best_pred_diff, best_tx_diff, best_filter_diff, 0);
3624 }
3625
3626 void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi,
3627 TileDataEnc *tile_data,
3628 MACROBLOCK *x,
3629 int mi_row, int mi_col,
3630 RD_COST *rd_cost,
3631 BLOCK_SIZE bsize,
3632 PICK_MODE_CONTEXT *ctx,
3633 int64_t best_rd_so_far) {
3634 VP9_COMMON *const cm = &cpi->common;
3635 TileInfo *const tile_info = &tile_data->tile_info;
3636 RD_OPT *const rd_opt = &cpi->rd;
3637 SPEED_FEATURES *const sf = &cpi->sf;
3638 MACROBLOCKD *const xd = &x->e_mbd;
3639 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
3640 const struct segmentation *const seg = &cm->seg;
3641 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3642 unsigned char segment_id = mbmi->segment_id;
3643 int comp_pred, i;
3644 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3645 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3646 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
3647 VP9_ALT_FLAG };
3648 int64_t best_rd = best_rd_so_far;
3649 int64_t best_yrd = best_rd_so_far; // FIXME(rbultje) more precise
3650 static const int64_t best_tx_diff[TX_MODES] = { 0 };
3651 int64_t best_pred_diff[REFERENCE_MODES];
3652 int64_t best_pred_rd[REFERENCE_MODES];
3653 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3654 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3655 MB_MODE_INFO best_mbmode;
3656 int ref_index, best_ref_index = 0;
3657 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3658 vp9_prob comp_mode_p;
3659 INTERP_FILTER tmp_best_filter = SWITCHABLE;
3660 int rate_uv_intra, rate_uv_tokenonly;
3661 int64_t dist_uv;
3662 int skip_uv;
3663 PREDICTION_MODE mode_uv = DC_PRED;
3664 const int intra_cost_penalty = vp9_get_intra_cost_penalty(
3665 cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
3666 int_mv seg_mvs[4][MAX_REF_FRAMES];
3667 b_mode_info best_bmodes[4];
3668 int best_skip2 = 0;
3669 int ref_frame_skip_mask[2] = { 0 };
3670 int64_t mask_filter = 0;
3671 int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
3672
3673 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3674 vpx_memset(x->zcoeff_blk[TX_4X4], 0, 4);
3675 vp9_zero(best_mbmode);
3676
3677 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3678 filter_cache[i] = INT64_MAX;
3679
3680 for (i = 0; i < 4; i++) {
3681 int j;
3682 for (j = 0; j < MAX_REF_FRAMES; j++)
3683 seg_mvs[i][j].as_int = INVALID_MV;
3684 }
3685
3686 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3687 &comp_mode_p);
3688
3689 for (i = 0; i < REFERENCE_MODES; ++i)
3690 best_pred_rd[i] = INT64_MAX;
3691 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3692 best_filter_rd[i] = INT64_MAX;
3693 rate_uv_intra = INT_MAX;
3694
3695 rd_cost->rate = INT_MAX;
3696
3697 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
3698 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3699 setup_buffer_inter(cpi, x, tile_info,
3700 ref_frame, bsize, mi_row, mi_col,
3701 frame_mv[NEARESTMV], frame_mv[NEARMV],
3702 yv12_mb);
3703 } else {
3704 ref_frame_skip_mask[0] |= (1 << ref_frame);
3705 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3706 }
3707 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3708 frame_mv[ZEROMV][ref_frame].as_int = 0;
3709 }
3710
3711 for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
3712 int mode_excluded = 0;
3713 int64_t this_rd = INT64_MAX;
3714 int disable_skip = 0;
3715 int compmode_cost = 0;
3716 int rate2 = 0, rate_y = 0, rate_uv = 0;
3717 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3718 int skippable = 0;
3719 int i;
3720 int this_skip2 = 0;
3721 int64_t total_sse = INT_MAX;
3722 int early_term = 0;
3723
3724 ref_frame = vp9_ref_order[ref_index].ref_frame[0];
3725 second_ref_frame = vp9_ref_order[ref_index].ref_frame[1];
3726
3727 // Look at the reference frame of the best mode so far and set the
3728 // skip mask to look at a subset of the remaining modes.
3729 if (ref_index > 2 && sf->mode_skip_start < MAX_MODES) {
3730 if (ref_index == 3) {
3731 switch (best_mbmode.ref_frame[0]) {
3732 case INTRA_FRAME:
3733 break;
3734 case LAST_FRAME:
3735 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME);
3736 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3737 break;
3738 case GOLDEN_FRAME:
3739 ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME);
3740 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3741 break;
3742 case ALTREF_FRAME:
3743 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME);
3744 break;
3745 case NONE:
3746 case MAX_REF_FRAMES:
3747 assert(0 && "Invalid Reference frame");
3748 break;
3749 }
3750 }
3751 }
3752
3753 if (ref_frame_skip_mask[0] & (1 << ref_frame) &&
3754 ref_frame_skip_mask[1] & (1 << MAX(0, second_ref_frame)))
3755 continue;
3756
3757 // Test best rd so far against threshold for trying this mode.
3758 if (rd_less_than_thresh(best_rd,
3759 rd_opt->threshes[segment_id][bsize][ref_index],
3760 tile_data->thresh_freq_fact[bsize][ref_index]))
3761 continue;
3762
3763 comp_pred = second_ref_frame > INTRA_FRAME;
3764 if (comp_pred) {
3765 if (!cpi->allow_comp_inter_inter)
3766 continue;
3767 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
3768 continue;
3769 // Do not allow compound prediction if the segment level reference frame
3770 // feature is in use as in this case there can only be one reference.
3771 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
3772 continue;
3773
3774 if ((sf->mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3775 best_mbmode.ref_frame[0] == INTRA_FRAME)
3776 continue;
3777 }
3778
3779 // TODO(jingning, jkoleszar): scaling reference frame not supported for
3780 // sub8x8 blocks.
3781 if (ref_frame > INTRA_FRAME &&
3782 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
3783 continue;
3784
3785 if (second_ref_frame > INTRA_FRAME &&
3786 vp9_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf))
3787 continue;
3788
3789 if (comp_pred)
3790 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3791 else if (ref_frame != INTRA_FRAME)
3792 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3793
3794 // If the segment reference frame feature is enabled....
3795 // then do nothing if the current ref frame is not allowed..
3796 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3797 vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3798 continue;
3799 // Disable this drop out case if the ref frame
3800 // segment level feature is enabled for this segment. This is to
3801 // prevent the possibility that we end up unable to pick any mode.
3802 } else if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3803 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3804 // unless ARNR filtering is enabled in which case we want
3805 // an unfiltered alternative. We allow near/nearest as well
3806 // because they may result in zero-zero MVs but be cheaper.
3807 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
3808 continue;
3809 }
3810
3811 mbmi->tx_size = TX_4X4;
3812 mbmi->uv_mode = DC_PRED;
3813 mbmi->ref_frame[0] = ref_frame;
3814 mbmi->ref_frame[1] = second_ref_frame;
3815 // Evaluate all sub-pel filters irrespective of whether we can use
3816 // them for this frame.
3817 mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
3818 : cm->interp_filter;
3819 x->skip = 0;
3820 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3821
3822 // Select prediction reference frames.
3823 for (i = 0; i < MAX_MB_PLANE; i++) {
3824 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3825 if (comp_pred)
3826 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3827 }
3828
3829 if (ref_frame == INTRA_FRAME) {
3830 int rate;
3831 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y,
3832 &distortion_y, best_rd) >= best_rd)
3833 continue;
3834 rate2 += rate;
3835 rate2 += intra_cost_penalty;
3836 distortion2 += distortion_y;
3837
3838 if (rate_uv_intra == INT_MAX) {
3839 choose_intra_uv_mode(cpi, x, ctx, bsize, TX_4X4,
3840 &rate_uv_intra,
3841 &rate_uv_tokenonly,
3842 &dist_uv, &skip_uv,
3843 &mode_uv);
3844 }
3845 rate2 += rate_uv_intra;
3846 rate_uv = rate_uv_tokenonly;
3847 distortion2 += dist_uv;
3848 distortion_uv = dist_uv;
3849 mbmi->uv_mode = mode_uv;
3850 } else {
3851 int rate;
3852 int64_t distortion;
3853 int64_t this_rd_thresh;
3854 int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
3855 int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
3856 int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
3857 int tmp_best_skippable = 0;
3858 int switchable_filter_index;
3859 int_mv *second_ref = comp_pred ?
3860 &mbmi->ref_mvs[second_ref_frame][0] : NULL;
3861 b_mode_info tmp_best_bmodes[16];
3862 MB_MODE_INFO tmp_best_mbmode;
3863 BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
3864 int pred_exists = 0;
3865 int uv_skippable;
3866
3867 this_rd_thresh = (ref_frame == LAST_FRAME) ?
3868 rd_opt->threshes[segment_id][bsize][THR_LAST] :
3869 rd_opt->threshes[segment_id][bsize][THR_ALTR];
3870 this_rd_thresh = (ref_frame == GOLDEN_FRAME) ?
3871 rd_opt->threshes[segment_id][bsize][THR_GOLD] : this_rd_thresh;
3872 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3873 filter_cache[i] = INT64_MAX;
3874
3875 if (cm->interp_filter != BILINEAR) {
3876 tmp_best_filter = EIGHTTAP;
3877 if (x->source_variance < sf->disable_filter_search_var_thresh) {
3878 tmp_best_filter = EIGHTTAP;
3879 } else if (sf->adaptive_pred_interp_filter == 1 &&
3880 ctx->pred_interp_filter < SWITCHABLE) {
3881 tmp_best_filter = ctx->pred_interp_filter;
3882 } else if (sf->adaptive_pred_interp_filter == 2) {
3883 tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE ?
3884 ctx->pred_interp_filter : 0;
3885 } else {
3886 for (switchable_filter_index = 0;
3887 switchable_filter_index < SWITCHABLE_FILTERS;
3888 ++switchable_filter_index) {
3889 int newbest, rs;
3890 int64_t rs_rd;
3891 mbmi->interp_filter = switchable_filter_index;
3892 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x, tile_info,
3893 &mbmi->ref_mvs[ref_frame][0],
3894 second_ref, best_yrd, &rate,
3895 &rate_y, &distortion,
3896 &skippable, &total_sse,
3897 (int) this_rd_thresh, seg_mvs,
3898 bsi, switchable_filter_index,
3899 mi_row, mi_col);
3900
3901 if (tmp_rd == INT64_MAX)
3902 continue;
3903 rs = vp9_get_switchable_rate(cpi, xd);
3904 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
3905 filter_cache[switchable_filter_index] = tmp_rd;
3906 filter_cache[SWITCHABLE_FILTERS] =
3907 MIN(filter_cache[SWITCHABLE_FILTERS],
3908 tmp_rd + rs_rd);
3909 if (cm->interp_filter == SWITCHABLE)
3910 tmp_rd += rs_rd;
3911
3912 mask_filter = MAX(mask_filter, tmp_rd);
3913
3914 newbest = (tmp_rd < tmp_best_rd);
3915 if (newbest) {
3916 tmp_best_filter = mbmi->interp_filter;
3917 tmp_best_rd = tmp_rd;
3918 }
3919 if ((newbest && cm->interp_filter == SWITCHABLE) ||
3920 (mbmi->interp_filter == cm->interp_filter &&
3921 cm->interp_filter != SWITCHABLE)) {
3922 tmp_best_rdu = tmp_rd;
3923 tmp_best_rate = rate;
3924 tmp_best_ratey = rate_y;
3925 tmp_best_distortion = distortion;
3926 tmp_best_sse = total_sse;
3927 tmp_best_skippable = skippable;
3928 tmp_best_mbmode = *mbmi;
3929 for (i = 0; i < 4; i++) {
3930 tmp_best_bmodes[i] = xd->mi[0].src_mi->bmi[i];
3931 x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
3932 }
3933 pred_exists = 1;
3934 if (switchable_filter_index == 0 &&
3935 sf->use_rd_breakout &&
3936 best_rd < INT64_MAX) {
3937 if (tmp_best_rdu / 2 > best_rd) {
3938 // skip searching the other filters if the first is
3939 // already substantially larger than the best so far
3940 tmp_best_filter = mbmi->interp_filter;
3941 tmp_best_rdu = INT64_MAX;
3942 break;
3943 }
3944 }
3945 }
3946 } // switchable_filter_index loop
3947 }
3948 }
3949
3950 if (tmp_best_rdu == INT64_MAX && pred_exists)
3951 continue;
3952
3953 mbmi->interp_filter = (cm->interp_filter == SWITCHABLE ?
3954 tmp_best_filter : cm->interp_filter);
3955 if (!pred_exists) {
3956 // Handles the special case when a filter that is not in the
3957 // switchable list (bilinear, 6-tap) is indicated at the frame level
3958 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x, tile_info,
3959 &mbmi->ref_mvs[ref_frame][0],
3960 second_ref, best_yrd, &rate, &rate_y,
3961 &distortion, &skippable, &total_sse,
3962 (int) this_rd_thresh, seg_mvs, bsi, 0,
3963 mi_row, mi_col);
3964 if (tmp_rd == INT64_MAX)
3965 continue;
3966 } else {
3967 total_sse = tmp_best_sse;
3968 rate = tmp_best_rate;
3969 rate_y = tmp_best_ratey;
3970 distortion = tmp_best_distortion;
3971 skippable = tmp_best_skippable;
3972 *mbmi = tmp_best_mbmode;
3973 for (i = 0; i < 4; i++)
3974 xd->mi[0].src_mi->bmi[i] = tmp_best_bmodes[i];
3975 }
3976
3977 rate2 += rate;
3978 distortion2 += distortion;
3979
3980 if (cm->interp_filter == SWITCHABLE)
3981 rate2 += vp9_get_switchable_rate(cpi, xd);
3982
3983 if (!mode_excluded)
3984 mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
3985 : cm->reference_mode == COMPOUND_REFERENCE;
3986
3987 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
3988
3989 tmp_best_rdu = best_rd -
3990 MIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
3991 RDCOST(x->rdmult, x->rddiv, 0, total_sse));
3992
3993 if (tmp_best_rdu > 0) {
3994 // If even the 'Y' rd value of split is higher than best so far
3995 // then dont bother looking at UV
3996 vp9_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col,
3997 BLOCK_8X8);
3998 vpx_memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
3999 if (!super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
4000 &uv_sse, BLOCK_8X8, tmp_best_rdu))
4001 continue;
4002
4003 rate2 += rate_uv;
4004 distortion2 += distortion_uv;
4005 skippable = skippable && uv_skippable;
4006 total_sse += uv_sse;
4007 }
4008 }
4009
4010 if (cm->reference_mode == REFERENCE_MODE_SELECT)
4011 rate2 += compmode_cost;
4012
4013 // Estimate the reference frame signaling cost and add it
4014 // to the rolling cost variable.
4015 if (second_ref_frame > INTRA_FRAME) {
4016 rate2 += ref_costs_comp[ref_frame];
4017 } else {
4018 rate2 += ref_costs_single[ref_frame];
4019 }
4020
4021 if (!disable_skip) {
4022 // Skip is never coded at the segment level for sub8x8 blocks and instead
4023 // always coded in the bitstream at the mode info level.
4024
4025 if (ref_frame != INTRA_FRAME && !xd->lossless) {
4026 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
4027 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
4028 // Add in the cost of the no skip flag.
4029 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
4030 } else {
4031 // FIXME(rbultje) make this work for splitmv also
4032 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
4033 distortion2 = total_sse;
4034 assert(total_sse >= 0);
4035 rate2 -= (rate_y + rate_uv);
4036 rate_y = 0;
4037 rate_uv = 0;
4038 this_skip2 = 1;
4039 }
4040 } else {
4041 // Add in the cost of the no skip flag.
4042 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
4043 }
4044
4045 // Calculate the final RD estimate for this mode.
4046 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
4047 }
4048
4049 if (!disable_skip && ref_frame == INTRA_FRAME) {
4050 for (i = 0; i < REFERENCE_MODES; ++i)
4051 best_pred_rd[i] = MIN(best_pred_rd[i], this_rd);
4052 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
4053 best_filter_rd[i] = MIN(best_filter_rd[i], this_rd);
4054 }
4055
4056 // Did this mode help.. i.e. is it the new best mode
4057 if (this_rd < best_rd || x->skip) {
4058 if (!mode_excluded) {
4059 int max_plane = MAX_MB_PLANE;
4060 // Note index of best mode so far
4061 best_ref_index = ref_index;
4062
4063 if (ref_frame == INTRA_FRAME) {
4064 /* required for left and above block mv */
4065 mbmi->mv[0].as_int = 0;
4066 max_plane = 1;
4067 }
4068
4069 rd_cost->rate = rate2;
4070 rd_cost->dist = distortion2;
4071 rd_cost->rdcost = this_rd;
4072 best_rd = this_rd;
4073 best_yrd = best_rd -
4074 RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
4075 best_mbmode = *mbmi;
4076 best_skip2 = this_skip2;
4077 if (!x->select_tx_size)
4078 swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
4079 vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
4080 sizeof(uint8_t) * ctx->num_4x4_blk);
4081
4082 for (i = 0; i < 4; i++)
4083 best_bmodes[i] = xd->mi[0].src_mi->bmi[i];
4084
4085 // TODO(debargha): enhance this test with a better distortion prediction
4086 // based on qp, activity mask and history
4087 if ((sf->mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
4088 (ref_index > MIN_EARLY_TERM_INDEX)) {
4089 int qstep = xd->plane[0].dequant[1];
4090 // TODO(debargha): Enhance this by specializing for each mode_index
4091 int scale = 4;
4092 #if CONFIG_VP9_HIGHBITDEPTH
4093 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
4094 qstep >>= (xd->bd - 8);
4095 }
4096 #endif // CONFIG_VP9_HIGHBITDEPTH
4097 if (x->source_variance < UINT_MAX) {
4098 const int var_adjust = (x->source_variance < 16);
4099 scale -= var_adjust;
4100 }
4101 if (ref_frame > INTRA_FRAME &&
4102 distortion2 * scale < qstep * qstep) {
4103 early_term = 1;
4104 }
4105 }
4106 }
4107 }
4108
4109 /* keep record of best compound/single-only prediction */
4110 if (!disable_skip && ref_frame != INTRA_FRAME) {
4111 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
4112
4113 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4114 single_rate = rate2 - compmode_cost;
4115 hybrid_rate = rate2;
4116 } else {
4117 single_rate = rate2;
4118 hybrid_rate = rate2 + compmode_cost;
4119 }
4120
4121 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
4122 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
4123
4124 if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
4125 best_pred_rd[SINGLE_REFERENCE] = single_rd;
4126 else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
4127 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
4128
4129 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
4130 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
4131 }
4132
4133 /* keep record of best filter type */
4134 if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
4135 cm->interp_filter != BILINEAR) {
4136 int64_t ref = filter_cache[cm->interp_filter == SWITCHABLE ?
4137 SWITCHABLE_FILTERS : cm->interp_filter];
4138 int64_t adj_rd;
4139 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4140 if (ref == INT64_MAX)
4141 adj_rd = 0;
4142 else if (filter_cache[i] == INT64_MAX)
4143 // when early termination is triggered, the encoder does not have
4144 // access to the rate-distortion cost. it only knows that the cost
4145 // should be above the maximum valid value. hence it takes the known
4146 // maximum plus an arbitrary constant as the rate-distortion cost.
4147 adj_rd = mask_filter - ref + 10;
4148 else
4149 adj_rd = filter_cache[i] - ref;
4150
4151 adj_rd += this_rd;
4152 best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd);
4153 }
4154 }
4155
4156 if (early_term)
4157 break;
4158
4159 if (x->skip && !comp_pred)
4160 break;
4161 }
4162
4163 if (best_rd >= best_rd_so_far) {
4164 rd_cost->rate = INT_MAX;
4165 rd_cost->rdcost = INT64_MAX;
4166 return;
4167 }
4168
4169 // If we used an estimate for the uv intra rd in the loop above...
4170 if (sf->use_uv_intra_rd_estimate) {
4171 // Do Intra UV best rd mode selection if best mode choice above was intra.
4172 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
4173 *mbmi = best_mbmode;
4174 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra,
4175 &rate_uv_tokenonly,
4176 &dist_uv,
4177 &skip_uv,
4178 BLOCK_8X8, TX_4X4);
4179 }
4180 }
4181
4182 if (best_rd == INT64_MAX) {
4183 rd_cost->rate = INT_MAX;
4184 rd_cost->dist = INT64_MAX;
4185 rd_cost->rdcost = INT64_MAX;
4186 return;
4187 }
4188
4189 assert((cm->interp_filter == SWITCHABLE) ||
4190 (cm->interp_filter == best_mbmode.interp_filter) ||
4191 !is_inter_block(&best_mbmode));
4192
4193 vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
4194 sf->adaptive_rd_thresh, bsize, best_ref_index);
4195
4196 // macroblock modes
4197 *mbmi = best_mbmode;
4198 x->skip |= best_skip2;
4199 if (!is_inter_block(&best_mbmode)) {
4200 for (i = 0; i < 4; i++)
4201 xd->mi[0].src_mi->bmi[i].as_mode = best_bmodes[i].as_mode;
4202 } else {
4203 for (i = 0; i < 4; ++i)
4204 vpx_memcpy(&xd->mi[0].src_mi->bmi[i], &best_bmodes[i],
4205 sizeof(b_mode_info));
4206
4207 mbmi->mv[0].as_int = xd->mi[0].src_mi->bmi[3].as_mv[0].as_int;
4208 mbmi->mv[1].as_int = xd->mi[0].src_mi->bmi[3].as_mv[1].as_int;
4209 }
4210
4211 for (i = 0; i < REFERENCE_MODES; ++i) {
4212 if (best_pred_rd[i] == INT64_MAX)
4213 best_pred_diff[i] = INT_MIN;
4214 else
4215 best_pred_diff[i] = best_rd - best_pred_rd[i];
4216 }
4217
4218 if (!x->skip) {
4219 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4220 if (best_filter_rd[i] == INT64_MAX)
4221 best_filter_diff[i] = 0;
4222 else
4223 best_filter_diff[i] = best_rd - best_filter_rd[i];
4224 }
4225 if (cm->interp_filter == SWITCHABLE)
4226 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
4227 } else {
4228 vp9_zero(best_filter_diff);
4229 }
4230
4231 store_coding_context(x, ctx, best_ref_index,
4232 best_pred_diff, best_tx_diff, best_filter_diff, 0);
4233 }
4234
AV1 Codec Library