2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
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.
14 #include "./vp10_rtcd.h"
15 #include "./vpx_dsp_rtcd.h"
17 #include "vpx_mem/vpx_mem.h"
18 #include "vpx_ports/mem.h"
19 #include "vpx_ports/system_state.h"
21 #include "vp10/common/common.h"
22 #include "vp10/common/entropy.h"
23 #include "vp10/common/entropymode.h"
24 #include "vp10/common/idct.h"
25 #include "vp10/common/mvref_common.h"
26 #include "vp10/common/pred_common.h"
27 #include "vp10/common/quant_common.h"
28 #include "vp10/common/reconinter.h"
29 #include "vp10/common/reconintra.h"
30 #include "vp10/common/scan.h"
31 #include "vp10/common/seg_common.h"
33 #include "vp10/encoder/cost.h"
34 #include "vp10/encoder/encodemb.h"
35 #include "vp10/encoder/encodemv.h"
36 #include "vp10/encoder/encoder.h"
37 #include "vp10/encoder/mcomp.h"
38 #include "vp10/encoder/quantize.h"
39 #include "vp10/encoder/ratectrl.h"
40 #include "vp10/encoder/rd.h"
41 #include "vp10/encoder/rdopt.h"
42 #include "vp10/encoder/aq_variance.h"
44 #define LAST_FRAME_MODE_MASK ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | \
46 #define GOLDEN_FRAME_MODE_MASK ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | \
48 #define ALT_REF_MODE_MASK ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | \
51 #define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
53 #define MIN_EARLY_TERM_INDEX 3
54 #define NEW_MV_DISCOUNT_FACTOR 8
58 MV_REFERENCE_FRAME ref_frame[2];
62 MV_REFERENCE_FRAME ref_frame[2];
65 struct rdcost_block_args {
67 ENTROPY_CONTEXT t_above[16];
68 ENTROPY_CONTEXT t_left[16];
75 int use_fast_coef_costing;
80 #define LAST_NEW_MV_INDEX 6
81 static const MODE_DEFINITION vp10_mode_order[MAX_MODES] = {
82 {NEARESTMV, {LAST_FRAME, NONE}},
83 {NEARESTMV, {ALTREF_FRAME, NONE}},
84 {NEARESTMV, {GOLDEN_FRAME, NONE}},
86 {DC_PRED, {INTRA_FRAME, NONE}},
88 {NEWMV, {LAST_FRAME, NONE}},
89 {NEWMV, {ALTREF_FRAME, NONE}},
90 {NEWMV, {GOLDEN_FRAME, NONE}},
92 {NEARMV, {LAST_FRAME, NONE}},
93 {NEARMV, {ALTREF_FRAME, NONE}},
94 {NEARMV, {GOLDEN_FRAME, NONE}},
96 {ZEROMV, {LAST_FRAME, NONE}},
97 {ZEROMV, {GOLDEN_FRAME, NONE}},
98 {ZEROMV, {ALTREF_FRAME, NONE}},
100 {NEARESTMV, {LAST_FRAME, ALTREF_FRAME}},
101 {NEARESTMV, {GOLDEN_FRAME, ALTREF_FRAME}},
103 {TM_PRED, {INTRA_FRAME, NONE}},
105 {NEARMV, {LAST_FRAME, ALTREF_FRAME}},
106 {NEWMV, {LAST_FRAME, ALTREF_FRAME}},
107 {NEARMV, {GOLDEN_FRAME, ALTREF_FRAME}},
108 {NEWMV, {GOLDEN_FRAME, ALTREF_FRAME}},
110 {ZEROMV, {LAST_FRAME, ALTREF_FRAME}},
111 {ZEROMV, {GOLDEN_FRAME, ALTREF_FRAME}},
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}},
123 static const REF_DEFINITION vp10_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}},
132 static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
133 int m, int n, int min_plane, int max_plane) {
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];
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];
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];
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;
157 static void model_rd_for_sb(VP10_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.
165 int64_t rate_sum = 0;
166 int64_t dist_sum = 0;
167 const int ref = xd->mi[0]->mbmi.ref_frame[0];
169 unsigned int var = 0;
170 unsigned int sum_sse = 0;
171 int64_t total_sse = 0;
176 const int dequant_shift =
177 #if CONFIG_VP9_HIGHBITDEPTH
178 (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ?
180 #endif // CONFIG_VP9_HIGHBITDEPTH
183 x->pred_sse[ref] = 0;
185 for (i = 0; i < MAX_MB_PLANE; ++i) {
186 struct macroblock_plane *const p = &x->plane[i];
187 struct macroblockd_plane *const pd = &xd->plane[i];
188 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
189 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
190 const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
191 const int64_t dc_thr = p->quant_thred[0] >> shift;
192 const int64_t ac_thr = p->quant_thred[1] >> shift;
193 // The low thresholds are used to measure if the prediction errors are
194 // low enough so that we can skip the mode search.
195 const int64_t low_dc_thr = MIN(50, dc_thr >> 2);
196 const int64_t low_ac_thr = MIN(80, ac_thr >> 2);
197 int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
198 int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
200 int lw = b_width_log2_lookup[unit_size] + 2;
201 int lh = b_height_log2_lookup[unit_size] + 2;
205 for (idy = 0; idy < bh; ++idy) {
206 for (idx = 0; idx < bw; ++idx) {
207 uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
208 uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
209 int block_idx = (idy << 1) + idx;
210 int low_err_skip = 0;
212 var = cpi->fn_ptr[unit_size].vf(src, p->src.stride,
213 dst, pd->dst.stride, &sse);
214 x->bsse[(i << 2) + block_idx] = sse;
217 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_NONE;
218 if (!x->select_tx_size) {
219 // Check if all ac coefficients can be quantized to zero.
220 if (var < ac_thr || var == 0) {
221 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_ONLY;
223 // Check if dc coefficient can be quantized to zero.
224 if (sse - var < dc_thr || sse == var) {
225 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_DC;
227 if (!sse || (var < low_ac_thr && sse - var < low_dc_thr))
233 if (skip_flag && !low_err_skip)
237 x->pred_sse[ref] += sse;
241 total_sse += sum_sse;
243 // Fast approximate the modelling function.
244 if (cpi->sf.simple_model_rd_from_var) {
246 const int64_t square_error = sum_sse;
247 int quantizer = (pd->dequant[1] >> dequant_shift);
250 rate = (square_error * (280 - quantizer)) >> 8;
253 dist = (square_error * quantizer) >> 8;
257 vp10_model_rd_from_var_lapndz(sum_sse, num_pels_log2_lookup[bs],
258 pd->dequant[1] >> dequant_shift,
265 *skip_txfm_sb = skip_flag;
266 *skip_sse_sb = total_sse << 4;
267 *out_rate_sum = (int)rate_sum;
268 *out_dist_sum = dist_sum << 4;
271 int64_t vp10_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
272 intptr_t block_size, int64_t *ssz) {
274 int64_t error = 0, sqcoeff = 0;
276 for (i = 0; i < block_size; i++) {
277 const int diff = coeff[i] - dqcoeff[i];
278 error += diff * diff;
279 sqcoeff += coeff[i] * coeff[i];
286 int64_t vp10_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff,
291 for (i = 0; i < block_size; i++) {
292 const int diff = coeff[i] - dqcoeff[i];
293 error += diff * diff;
299 #if CONFIG_VP9_HIGHBITDEPTH
300 int64_t vp10_highbd_block_error_c(const tran_low_t *coeff,
301 const tran_low_t *dqcoeff,
303 int64_t *ssz, int bd) {
305 int64_t error = 0, sqcoeff = 0;
306 int shift = 2 * (bd - 8);
307 int rounding = shift > 0 ? 1 << (shift - 1) : 0;
309 for (i = 0; i < block_size; i++) {
310 const int64_t diff = coeff[i] - dqcoeff[i];
311 error += diff * diff;
312 sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
314 assert(error >= 0 && sqcoeff >= 0);
315 error = (error + rounding) >> shift;
316 sqcoeff = (sqcoeff + rounding) >> shift;
321 #endif // CONFIG_VP9_HIGHBITDEPTH
323 /* The trailing '0' is a terminator which is used inside cost_coeffs() to
324 * decide whether to include cost of a trailing EOB node or not (i.e. we
325 * can skip this if the last coefficient in this transform block, e.g. the
326 * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
328 static const int16_t band_counts[TX_SIZES][8] = {
329 { 1, 2, 3, 4, 3, 16 - 13, 0 },
330 { 1, 2, 3, 4, 11, 64 - 21, 0 },
331 { 1, 2, 3, 4, 11, 256 - 21, 0 },
332 { 1, 2, 3, 4, 11, 1024 - 21, 0 },
334 static int cost_coeffs(MACROBLOCK *x,
335 int plane, int block,
336 ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L,
338 const int16_t *scan, const int16_t *nb,
339 int use_fast_coef_costing) {
340 MACROBLOCKD *const xd = &x->e_mbd;
341 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
342 const struct macroblock_plane *p = &x->plane[plane];
343 const struct macroblockd_plane *pd = &xd->plane[plane];
344 const PLANE_TYPE type = pd->plane_type;
345 const int16_t *band_count = &band_counts[tx_size][1];
346 const int eob = p->eobs[block];
347 const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
348 unsigned int (*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
349 x->token_costs[tx_size][type][is_inter_block(mbmi)];
350 uint8_t token_cache[32 * 32];
351 int pt = combine_entropy_contexts(*A, *L);
353 #if CONFIG_VP9_HIGHBITDEPTH
354 const int16_t *cat6_high_cost = vp10_get_high_cost_table(xd->bd);
356 const int16_t *cat6_high_cost = vp10_get_high_cost_table(8);
359 // Check for consistency of tx_size with mode info
360 assert(type == PLANE_TYPE_Y ? mbmi->tx_size == tx_size
361 : get_uv_tx_size(mbmi, pd) == tx_size);
365 cost = token_costs[0][0][pt][EOB_TOKEN];
368 int band_left = *band_count++;
374 vp10_get_token_extra(v, &prev_t, &e);
375 cost = (*token_costs)[0][pt][prev_t] +
376 vp10_get_cost(prev_t, e, cat6_high_cost);
378 token_cache[0] = vp10_pt_energy_class[prev_t];
382 for (c = 1; c < eob; c++) {
383 const int rc = scan[c];
387 vp10_get_token_extra(v, &t, &e);
388 if (use_fast_coef_costing) {
389 cost += (*token_costs)[!prev_t][!prev_t][t] +
390 vp10_get_cost(t, e, cat6_high_cost);
392 pt = get_coef_context(nb, token_cache, c);
393 cost += (*token_costs)[!prev_t][pt][t] +
394 vp10_get_cost(t, e, cat6_high_cost);
395 token_cache[rc] = vp10_pt_energy_class[t];
399 band_left = *band_count++;
406 if (use_fast_coef_costing) {
407 cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
409 pt = get_coef_context(nb, token_cache, c);
410 cost += (*token_costs)[0][pt][EOB_TOKEN];
415 // is eob first coefficient;
421 static void dist_block(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
422 int64_t *out_dist, int64_t *out_sse) {
423 const int ss_txfrm_size = tx_size << 1;
424 MACROBLOCKD* const xd = &x->e_mbd;
425 const struct macroblock_plane *const p = &x->plane[plane];
426 const struct macroblockd_plane *const pd = &xd->plane[plane];
428 int shift = tx_size == TX_32X32 ? 0 : 2;
429 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
430 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
431 #if CONFIG_VP9_HIGHBITDEPTH
432 const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
433 *out_dist = vp10_highbd_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
434 &this_sse, bd) >> shift;
436 *out_dist = vp10_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
438 #endif // CONFIG_VP9_HIGHBITDEPTH
439 *out_sse = this_sse >> shift;
441 if (x->skip_encode && !is_inter_block(&xd->mi[0]->mbmi)) {
442 // TODO(jingning): tune the model to better capture the distortion.
443 int64_t p = (pd->dequant[1] * pd->dequant[1] *
444 (1 << ss_txfrm_size)) >>
445 #if CONFIG_VP9_HIGHBITDEPTH
446 (shift + 2 + (bd - 8) * 2);
449 #endif // CONFIG_VP9_HIGHBITDEPTH
450 *out_dist += (p >> 4);
455 static int rate_block(int plane, int block, BLOCK_SIZE plane_bsize,
456 TX_SIZE tx_size, struct rdcost_block_args* args) {
458 txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x_idx, &y_idx);
460 return cost_coeffs(args->x, plane, block, args->t_above + x_idx,
461 args->t_left + y_idx, tx_size,
462 args->so->scan, args->so->neighbors,
463 args->use_fast_coef_costing);
466 static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
467 TX_SIZE tx_size, void *arg) {
468 struct rdcost_block_args *args = arg;
469 MACROBLOCK *const x = args->x;
470 MACROBLOCKD *const xd = &x->e_mbd;
471 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
472 int64_t rd1, rd2, rd;
477 if (args->exit_early)
480 if (!is_inter_block(mbmi)) {
481 struct encode_b_args arg = {x, NULL, &mbmi->skip};
482 vp10_encode_block_intra(plane, block, plane_bsize, tx_size, &arg);
483 dist_block(x, plane, block, tx_size, &dist, &sse);
484 } else if (max_txsize_lookup[plane_bsize] == tx_size) {
485 if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
487 // full forward transform and quantization
488 vp10_xform_quant(x, plane, block, plane_bsize, tx_size);
489 dist_block(x, plane, block, tx_size, &dist, &sse);
490 } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
492 // compute DC coefficient
493 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[plane].coeff, block);
494 tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
495 vp10_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
496 sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
498 if (x->plane[plane].eobs[block]) {
499 const int64_t orig_sse = (int64_t)coeff[0] * coeff[0];
500 const int64_t resd_sse = coeff[0] - dqcoeff[0];
501 int64_t dc_correct = orig_sse - resd_sse * resd_sse;
502 #if CONFIG_VP9_HIGHBITDEPTH
503 dc_correct >>= ((xd->bd - 8) * 2);
505 if (tx_size != TX_32X32)
508 dist = MAX(0, sse - dc_correct);
512 // skip forward transform
513 x->plane[plane].eobs[block] = 0;
514 sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
518 // full forward transform and quantization
519 vp10_xform_quant(x, plane, block, plane_bsize, tx_size);
520 dist_block(x, plane, block, tx_size, &dist, &sse);
523 rd = RDCOST(x->rdmult, x->rddiv, 0, dist);
524 if (args->this_rd + rd > args->best_rd) {
525 args->exit_early = 1;
529 rate = rate_block(plane, block, plane_bsize, tx_size, args);
530 rd1 = RDCOST(x->rdmult, x->rddiv, rate, dist);
531 rd2 = RDCOST(x->rdmult, x->rddiv, 0, sse);
533 // TODO(jingning): temporarily enabled only for luma component
536 x->zcoeff_blk[tx_size][block] = !x->plane[plane].eobs[block] ||
537 (rd1 > rd2 && !xd->lossless);
539 args->this_rate += rate;
540 args->this_dist += dist;
541 args->this_sse += sse;
544 if (args->this_rd > args->best_rd) {
545 args->exit_early = 1;
549 args->skippable &= !x->plane[plane].eobs[block];
552 static void txfm_rd_in_plane(MACROBLOCK *x,
553 int *rate, int64_t *distortion,
554 int *skippable, int64_t *sse,
555 int64_t ref_best_rd, int plane,
556 BLOCK_SIZE bsize, TX_SIZE tx_size,
557 int use_fast_coef_casting) {
558 MACROBLOCKD *const xd = &x->e_mbd;
559 const struct macroblockd_plane *const pd = &xd->plane[plane];
560 struct rdcost_block_args args;
563 args.best_rd = ref_best_rd;
564 args.use_fast_coef_costing = use_fast_coef_casting;
568 xd->mi[0]->mbmi.tx_size = tx_size;
570 vp10_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
572 args.so = get_scan(xd, tx_size, pd->plane_type, 0);
574 vp10_foreach_transformed_block_in_plane(xd, bsize, plane,
575 block_rd_txfm, &args);
576 if (args.exit_early) {
578 *distortion = INT64_MAX;
582 *distortion = args.this_dist;
583 *rate = args.this_rate;
584 *sse = args.this_sse;
585 *skippable = args.skippable;
589 static void choose_largest_tx_size(VP10_COMP *cpi, MACROBLOCK *x,
590 int *rate, int64_t *distortion,
591 int *skip, int64_t *sse,
594 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
595 VP10_COMMON *const cm = &cpi->common;
596 const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
597 MACROBLOCKD *const xd = &x->e_mbd;
598 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
600 mbmi->tx_size = MIN(max_tx_size, largest_tx_size);
602 txfm_rd_in_plane(x, rate, distortion, skip,
603 sse, ref_best_rd, 0, bs,
604 mbmi->tx_size, cpi->sf.use_fast_coef_costing);
607 static void choose_tx_size_from_rd(VP10_COMP *cpi, MACROBLOCK *x,
614 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
615 VP10_COMMON *const cm = &cpi->common;
616 MACROBLOCKD *const xd = &x->e_mbd;
617 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
618 vpx_prob skip_prob = vp10_get_skip_prob(cm, xd);
619 int r[TX_SIZES][2], s[TX_SIZES];
620 int64_t d[TX_SIZES], sse[TX_SIZES];
621 int64_t rd[TX_SIZES][2] = {{INT64_MAX, INT64_MAX},
622 {INT64_MAX, INT64_MAX},
623 {INT64_MAX, INT64_MAX},
624 {INT64_MAX, INT64_MAX}};
627 int64_t best_rd = INT64_MAX;
628 TX_SIZE best_tx = max_tx_size;
629 int start_tx, end_tx;
631 const vpx_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc->tx_probs);
632 assert(skip_prob > 0);
633 s0 = vp10_cost_bit(skip_prob, 0);
634 s1 = vp10_cost_bit(skip_prob, 1);
636 if (cm->tx_mode == TX_MODE_SELECT) {
637 start_tx = max_tx_size;
640 TX_SIZE chosen_tx_size = MIN(max_tx_size,
641 tx_mode_to_biggest_tx_size[cm->tx_mode]);
642 start_tx = chosen_tx_size;
643 end_tx = chosen_tx_size;
646 for (n = start_tx; n >= end_tx; n--) {
647 txfm_rd_in_plane(x, &r[n][0], &d[n], &s[n],
648 &sse[n], ref_best_rd, 0, bs, n,
649 cpi->sf.use_fast_coef_costing);
651 if (r[n][0] < INT_MAX) {
652 for (m = 0; m <= n - (n == (int) max_tx_size); m++) {
654 r[n][1] += vp10_cost_zero(tx_probs[m]);
656 r[n][1] += vp10_cost_one(tx_probs[m]);
659 if (d[n] == INT64_MAX) {
660 rd[n][0] = rd[n][1] = INT64_MAX;
662 rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, d[n]);
664 rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
665 rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
668 // Early termination in transform size search.
669 if (cpi->sf.tx_size_search_breakout &&
670 (rd[n][1] == INT64_MAX ||
671 (n < (int) max_tx_size && rd[n][1] > rd[n + 1][1]) ||
675 if (rd[n][1] < best_rd) {
680 mbmi->tx_size = best_tx;
682 *distortion = d[mbmi->tx_size];
683 *rate = r[mbmi->tx_size][cm->tx_mode == TX_MODE_SELECT];
684 *skip = s[mbmi->tx_size];
685 *psse = sse[mbmi->tx_size];
688 static void super_block_yrd(VP10_COMP *cpi, MACROBLOCK *x, int *rate,
689 int64_t *distortion, int *skip,
690 int64_t *psse, BLOCK_SIZE bs,
691 int64_t ref_best_rd) {
692 MACROBLOCKD *xd = &x->e_mbd;
694 int64_t *ret_sse = psse ? psse : &sse;
696 assert(bs == xd->mi[0]->mbmi.sb_type);
698 if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
699 choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
702 choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse,
707 static int conditional_skipintra(PREDICTION_MODE mode,
708 PREDICTION_MODE best_intra_mode) {
709 if (mode == D117_PRED &&
710 best_intra_mode != V_PRED &&
711 best_intra_mode != D135_PRED)
713 if (mode == D63_PRED &&
714 best_intra_mode != V_PRED &&
715 best_intra_mode != D45_PRED)
717 if (mode == D207_PRED &&
718 best_intra_mode != H_PRED &&
719 best_intra_mode != D45_PRED)
721 if (mode == D153_PRED &&
722 best_intra_mode != H_PRED &&
723 best_intra_mode != D135_PRED)
728 static int64_t rd_pick_intra4x4block(VP10_COMP *cpi, MACROBLOCK *x,
730 PREDICTION_MODE *best_mode,
731 const int *bmode_costs,
732 ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
733 int *bestrate, int *bestratey,
734 int64_t *bestdistortion,
735 BLOCK_SIZE bsize, int64_t rd_thresh) {
736 PREDICTION_MODE mode;
737 MACROBLOCKD *const xd = &x->e_mbd;
738 int64_t best_rd = rd_thresh;
739 struct macroblock_plane *p = &x->plane[0];
740 struct macroblockd_plane *pd = &xd->plane[0];
741 const int src_stride = p->src.stride;
742 const int dst_stride = pd->dst.stride;
743 const uint8_t *src_init = &p->src.buf[row * 4 * src_stride + col * 4];
744 uint8_t *dst_init = &pd->dst.buf[row * 4 * src_stride + col * 4];
745 ENTROPY_CONTEXT ta[2], tempa[2];
746 ENTROPY_CONTEXT tl[2], templ[2];
747 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
748 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
750 uint8_t best_dst[8 * 8];
751 #if CONFIG_VP9_HIGHBITDEPTH
752 uint16_t best_dst16[8 * 8];
755 memcpy(ta, a, sizeof(ta));
756 memcpy(tl, l, sizeof(tl));
757 xd->mi[0]->mbmi.tx_size = TX_4X4;
759 #if CONFIG_VP9_HIGHBITDEPTH
760 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
761 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
764 int64_t distortion = 0;
765 int rate = bmode_costs[mode];
767 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
770 // Only do the oblique modes if the best so far is
771 // one of the neighboring directional modes
772 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
773 if (conditional_skipintra(mode, *best_mode))
777 memcpy(tempa, ta, sizeof(ta));
778 memcpy(templ, tl, sizeof(tl));
780 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
781 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
782 const int block = (row + idy) * 2 + (col + idx);
783 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
784 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
785 int16_t *const src_diff = vp10_raster_block_offset_int16(BLOCK_8X8,
788 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
789 xd->mi[0]->bmi[block].as_mode = mode;
790 vp10_predict_intra_block(xd, 1, TX_4X4, mode,
791 x->skip_encode ? src : dst,
792 x->skip_encode ? src_stride : dst_stride,
794 col + idx, row + idy, 0);
795 vpx_highbd_subtract_block(4, 4, src_diff, 8, src, src_stride,
796 dst, dst_stride, xd->bd);
798 const scan_order *so = &vp10_default_scan_orders[TX_4X4];
799 vp10_highbd_fwht4x4(src_diff, coeff, 8);
800 vp10_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
801 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
802 so->scan, so->neighbors,
803 cpi->sf.use_fast_coef_costing);
804 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
806 vp10_highbd_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block),
808 p->eobs[block], xd->bd);
811 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
812 const scan_order *so = &vp10_scan_orders[TX_4X4][tx_type];
813 if (tx_type == DCT_DCT)
814 vpx_highbd_fdct4x4(src_diff, coeff, 8);
816 vp10_highbd_fht4x4(src_diff, coeff, 8, tx_type);
817 vp10_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
818 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
819 so->scan, so->neighbors,
820 cpi->sf.use_fast_coef_costing);
821 distortion += vp10_highbd_block_error(
822 coeff, BLOCK_OFFSET(pd->dqcoeff, block),
823 16, &unused, xd->bd) >> 2;
824 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
826 vp10_highbd_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
827 dst, dst_stride, p->eobs[block], xd->bd);
833 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
835 if (this_rd < best_rd) {
838 *bestdistortion = distortion;
841 memcpy(a, tempa, sizeof(tempa));
842 memcpy(l, templ, sizeof(templ));
843 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
844 memcpy(best_dst16 + idy * 8,
845 CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
846 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
852 if (best_rd >= rd_thresh || x->skip_encode)
855 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
856 memcpy(CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
857 best_dst16 + idy * 8,
858 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
863 #endif // CONFIG_VP9_HIGHBITDEPTH
865 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
868 int64_t distortion = 0;
869 int rate = bmode_costs[mode];
871 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
874 // Only do the oblique modes if the best so far is
875 // one of the neighboring directional modes
876 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
877 if (conditional_skipintra(mode, *best_mode))
881 memcpy(tempa, ta, sizeof(ta));
882 memcpy(templ, tl, sizeof(tl));
884 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
885 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
886 const int block = (row + idy) * 2 + (col + idx);
887 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
888 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
889 int16_t *const src_diff =
890 vp10_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
891 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
892 xd->mi[0]->bmi[block].as_mode = mode;
893 vp10_predict_intra_block(xd, 1, TX_4X4, mode,
894 x->skip_encode ? src : dst,
895 x->skip_encode ? src_stride : dst_stride,
896 dst, dst_stride, col + idx, row + idy, 0);
897 vpx_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
900 const scan_order *so = &vp10_default_scan_orders[TX_4X4];
901 vp10_fwht4x4(src_diff, coeff, 8);
902 vp10_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
903 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
904 so->scan, so->neighbors,
905 cpi->sf.use_fast_coef_costing);
906 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
908 vp10_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride,
912 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
913 const scan_order *so = &vp10_scan_orders[TX_4X4][tx_type];
914 vp10_fht4x4(src_diff, coeff, 8, tx_type);
915 vp10_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
916 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
917 so->scan, so->neighbors,
918 cpi->sf.use_fast_coef_costing);
919 distortion += vp10_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
921 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
923 vp10_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
924 dst, dst_stride, p->eobs[block]);
930 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
932 if (this_rd < best_rd) {
935 *bestdistortion = distortion;
938 memcpy(a, tempa, sizeof(tempa));
939 memcpy(l, templ, sizeof(templ));
940 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
941 memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
942 num_4x4_blocks_wide * 4);
948 if (best_rd >= rd_thresh || x->skip_encode)
951 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
952 memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
953 num_4x4_blocks_wide * 4);
958 static int64_t rd_pick_intra_sub_8x8_y_mode(VP10_COMP *cpi, MACROBLOCK *mb,
959 int *rate, int *rate_y,
963 const MACROBLOCKD *const xd = &mb->e_mbd;
964 MODE_INFO *const mic = xd->mi[0];
965 const MODE_INFO *above_mi = xd->above_mi;
966 const MODE_INFO *left_mi = xd->left_mi;
967 const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
968 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
969 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
972 int64_t total_distortion = 0;
974 int64_t total_rd = 0;
975 ENTROPY_CONTEXT t_above[4], t_left[4];
976 const int *bmode_costs = cpi->mbmode_cost;
978 memcpy(t_above, xd->plane[0].above_context, sizeof(t_above));
979 memcpy(t_left, xd->plane[0].left_context, sizeof(t_left));
981 // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
982 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
983 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
984 PREDICTION_MODE best_mode = DC_PRED;
985 int r = INT_MAX, ry = INT_MAX;
986 int64_t d = INT64_MAX, this_rd = INT64_MAX;
988 if (cpi->common.frame_type == KEY_FRAME) {
989 const PREDICTION_MODE A = vp10_above_block_mode(mic, above_mi, i);
990 const PREDICTION_MODE L = vp10_left_block_mode(mic, left_mi, i);
992 bmode_costs = cpi->y_mode_costs[A][L];
995 this_rd = rd_pick_intra4x4block(cpi, mb, idy, idx, &best_mode,
996 bmode_costs, t_above + idx, t_left + idy,
997 &r, &ry, &d, bsize, best_rd - total_rd);
998 if (this_rd >= best_rd - total_rd)
1001 total_rd += this_rd;
1003 total_distortion += d;
1006 mic->bmi[i].as_mode = best_mode;
1007 for (j = 1; j < num_4x4_blocks_high; ++j)
1008 mic->bmi[i + j * 2].as_mode = best_mode;
1009 for (j = 1; j < num_4x4_blocks_wide; ++j)
1010 mic->bmi[i + j].as_mode = best_mode;
1012 if (total_rd >= best_rd)
1018 *rate_y = tot_rate_y;
1019 *distortion = total_distortion;
1020 mic->mbmi.mode = mic->bmi[3].as_mode;
1022 return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
1025 // This function is used only for intra_only frames
1026 static int64_t rd_pick_intra_sby_mode(VP10_COMP *cpi, MACROBLOCK *x,
1027 int *rate, int *rate_tokenonly,
1028 int64_t *distortion, int *skippable,
1031 PREDICTION_MODE mode;
1032 PREDICTION_MODE mode_selected = DC_PRED;
1033 MACROBLOCKD *const xd = &x->e_mbd;
1034 MODE_INFO *const mic = xd->mi[0];
1035 int this_rate, this_rate_tokenonly, s;
1036 int64_t this_distortion, this_rd;
1037 TX_SIZE best_tx = TX_4X4;
1039 const MODE_INFO *above_mi = xd->above_mi;
1040 const MODE_INFO *left_mi = xd->left_mi;
1041 const PREDICTION_MODE A = vp10_above_block_mode(mic, above_mi, 0);
1042 const PREDICTION_MODE L = vp10_left_block_mode(mic, left_mi, 0);
1043 bmode_costs = cpi->y_mode_costs[A][L];
1045 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1046 /* Y Search for intra prediction mode */
1047 for (mode = DC_PRED; mode <= TM_PRED; mode++) {
1049 if (cpi->sf.use_nonrd_pick_mode) {
1050 // These speed features are turned on in hybrid non-RD and RD mode
1051 // for key frame coding in the context of real-time setting.
1052 if (conditional_skipintra(mode, mode_selected))
1058 mic->mbmi.mode = mode;
1060 super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion,
1061 &s, NULL, bsize, best_rd);
1063 if (this_rate_tokenonly == INT_MAX)
1066 this_rate = this_rate_tokenonly + bmode_costs[mode];
1067 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1069 if (this_rd < best_rd) {
1070 mode_selected = mode;
1072 best_tx = mic->mbmi.tx_size;
1074 *rate_tokenonly = this_rate_tokenonly;
1075 *distortion = this_distortion;
1080 mic->mbmi.mode = mode_selected;
1081 mic->mbmi.tx_size = best_tx;
1086 // Return value 0: early termination triggered, no valid rd cost available;
1087 // 1: rd cost values are valid.
1088 static int super_block_uvrd(const VP10_COMP *cpi, MACROBLOCK *x,
1089 int *rate, int64_t *distortion, int *skippable,
1090 int64_t *sse, BLOCK_SIZE bsize,
1091 int64_t ref_best_rd) {
1092 MACROBLOCKD *const xd = &x->e_mbd;
1093 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1094 const TX_SIZE uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
1096 int pnrate = 0, pnskip = 1;
1097 int64_t pndist = 0, pnsse = 0;
1098 int is_cost_valid = 1;
1100 if (ref_best_rd < 0)
1103 if (is_inter_block(mbmi) && is_cost_valid) {
1105 for (plane = 1; plane < MAX_MB_PLANE; ++plane)
1106 vp10_subtract_plane(x, bsize, plane);
1114 for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
1115 txfm_rd_in_plane(x, &pnrate, &pndist, &pnskip, &pnsse,
1116 ref_best_rd, plane, bsize, uv_tx_size,
1117 cpi->sf.use_fast_coef_costing);
1118 if (pnrate == INT_MAX) {
1123 *distortion += pndist;
1125 *skippable &= pnskip;
1128 if (!is_cost_valid) {
1131 *distortion = INT64_MAX;
1136 return is_cost_valid;
1139 static int64_t rd_pick_intra_sbuv_mode(VP10_COMP *cpi, MACROBLOCK *x,
1140 PICK_MODE_CONTEXT *ctx,
1141 int *rate, int *rate_tokenonly,
1142 int64_t *distortion, int *skippable,
1143 BLOCK_SIZE bsize, TX_SIZE max_tx_size) {
1144 MACROBLOCKD *xd = &x->e_mbd;
1145 PREDICTION_MODE mode;
1146 PREDICTION_MODE mode_selected = DC_PRED;
1147 int64_t best_rd = INT64_MAX, this_rd;
1148 int this_rate_tokenonly, this_rate, s;
1149 int64_t this_distortion, this_sse;
1151 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1152 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1153 if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode)))
1156 xd->mi[0]->mbmi.uv_mode = mode;
1158 if (!super_block_uvrd(cpi, x, &this_rate_tokenonly,
1159 &this_distortion, &s, &this_sse, bsize, best_rd))
1161 this_rate = this_rate_tokenonly +
1162 cpi->intra_uv_mode_cost[cpi->common.frame_type][mode];
1163 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1165 if (this_rd < best_rd) {
1166 mode_selected = mode;
1169 *rate_tokenonly = this_rate_tokenonly;
1170 *distortion = this_distortion;
1172 if (!x->select_tx_size)
1173 swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
1177 xd->mi[0]->mbmi.uv_mode = mode_selected;
1181 static int64_t rd_sbuv_dcpred(const VP10_COMP *cpi, MACROBLOCK *x,
1182 int *rate, int *rate_tokenonly,
1183 int64_t *distortion, int *skippable,
1185 const VP10_COMMON *cm = &cpi->common;
1188 x->e_mbd.mi[0]->mbmi.uv_mode = DC_PRED;
1189 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1190 super_block_uvrd(cpi, x, rate_tokenonly, distortion,
1191 skippable, &unused, bsize, INT64_MAX);
1192 *rate = *rate_tokenonly + cpi->intra_uv_mode_cost[cm->frame_type][DC_PRED];
1193 return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
1196 static void choose_intra_uv_mode(VP10_COMP *cpi, MACROBLOCK *const x,
1197 PICK_MODE_CONTEXT *ctx,
1198 BLOCK_SIZE bsize, TX_SIZE max_tx_size,
1199 int *rate_uv, int *rate_uv_tokenonly,
1200 int64_t *dist_uv, int *skip_uv,
1201 PREDICTION_MODE *mode_uv) {
1202 // Use an estimated rd for uv_intra based on DC_PRED if the
1203 // appropriate speed flag is set.
1204 if (cpi->sf.use_uv_intra_rd_estimate) {
1205 rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv,
1206 skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
1207 // Else do a proper rd search for each possible transform size that may
1208 // be considered in the main rd loop.
1210 rd_pick_intra_sbuv_mode(cpi, x, ctx,
1211 rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
1212 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize, max_tx_size);
1214 *mode_uv = x->e_mbd.mi[0]->mbmi.uv_mode;
1217 static int cost_mv_ref(const VP10_COMP *cpi, PREDICTION_MODE mode,
1219 assert(is_inter_mode(mode));
1220 return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
1223 static int set_and_cost_bmi_mvs(VP10_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
1225 PREDICTION_MODE mode, int_mv this_mv[2],
1226 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1227 int_mv seg_mvs[MAX_REF_FRAMES],
1228 int_mv *best_ref_mv[2], const int *mvjcost,
1230 MODE_INFO *const mic = xd->mi[0];
1231 const MB_MODE_INFO *const mbmi = &mic->mbmi;
1232 const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1235 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type];
1236 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
1237 const int is_compound = has_second_ref(mbmi);
1241 this_mv[0].as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
1242 thismvcost += vp10_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
1243 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1245 this_mv[1].as_int = seg_mvs[mbmi->ref_frame[1]].as_int;
1246 thismvcost += vp10_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
1247 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1252 this_mv[0].as_int = frame_mv[mode][mbmi->ref_frame[0]].as_int;
1254 this_mv[1].as_int = frame_mv[mode][mbmi->ref_frame[1]].as_int;
1257 this_mv[0].as_int = 0;
1259 this_mv[1].as_int = 0;
1265 mic->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
1267 mic->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
1269 mic->bmi[i].as_mode = mode;
1271 for (idy = 0; idy < num_4x4_blocks_high; ++idy)
1272 for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
1273 memmove(&mic->bmi[i + idy * 2 + idx], &mic->bmi[i], sizeof(mic->bmi[i]));
1275 return cost_mv_ref(cpi, mode, mbmi_ext->mode_context[mbmi->ref_frame[0]]) +
1279 static int64_t encode_inter_mb_segment(VP10_COMP *cpi,
1284 int64_t *distortion, int64_t *sse,
1285 ENTROPY_CONTEXT *ta,
1286 ENTROPY_CONTEXT *tl,
1287 int mi_row, int mi_col) {
1289 MACROBLOCKD *xd = &x->e_mbd;
1290 struct macroblockd_plane *const pd = &xd->plane[0];
1291 struct macroblock_plane *const p = &x->plane[0];
1292 MODE_INFO *const mi = xd->mi[0];
1293 const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->mbmi.sb_type, pd);
1294 const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
1295 const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
1298 const uint8_t *const src =
1299 &p->src.buf[vp10_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1300 uint8_t *const dst = &pd->dst.buf[vp10_raster_block_offset(BLOCK_8X8, i,
1302 int64_t thisdistortion = 0, thissse = 0;
1303 int thisrate = 0, ref;
1304 const scan_order *so = &vp10_default_scan_orders[TX_4X4];
1305 const int is_compound = has_second_ref(&mi->mbmi);
1306 const InterpKernel *kernel = vp10_filter_kernels[mi->mbmi.interp_filter];
1308 for (ref = 0; ref < 1 + is_compound; ++ref) {
1309 const uint8_t *pre = &pd->pre[ref].buf[vp10_raster_block_offset(BLOCK_8X8, i,
1310 pd->pre[ref].stride)];
1311 #if CONFIG_VP9_HIGHBITDEPTH
1312 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1313 vp10_highbd_build_inter_predictor(pre, pd->pre[ref].stride,
1314 dst, pd->dst.stride,
1315 &mi->bmi[i].as_mv[ref].as_mv,
1316 &xd->block_refs[ref]->sf, width, height,
1317 ref, kernel, MV_PRECISION_Q3,
1318 mi_col * MI_SIZE + 4 * (i % 2),
1319 mi_row * MI_SIZE + 4 * (i / 2), xd->bd);
1321 vp10_build_inter_predictor(pre, pd->pre[ref].stride,
1322 dst, pd->dst.stride,
1323 &mi->bmi[i].as_mv[ref].as_mv,
1324 &xd->block_refs[ref]->sf, width, height, ref,
1325 kernel, MV_PRECISION_Q3,
1326 mi_col * MI_SIZE + 4 * (i % 2),
1327 mi_row * MI_SIZE + 4 * (i / 2));
1330 vp10_build_inter_predictor(pre, pd->pre[ref].stride,
1331 dst, pd->dst.stride,
1332 &mi->bmi[i].as_mv[ref].as_mv,
1333 &xd->block_refs[ref]->sf, width, height, ref,
1334 kernel, MV_PRECISION_Q3,
1335 mi_col * MI_SIZE + 4 * (i % 2),
1336 mi_row * MI_SIZE + 4 * (i / 2));
1337 #endif // CONFIG_VP9_HIGHBITDEPTH
1340 #if CONFIG_VP9_HIGHBITDEPTH
1341 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1342 vpx_highbd_subtract_block(
1343 height, width, vp10_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1344 8, src, p->src.stride, dst, pd->dst.stride, xd->bd);
1347 height, width, vp10_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1348 8, src, p->src.stride, dst, pd->dst.stride);
1351 vpx_subtract_block(height, width,
1352 vp10_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1353 8, src, p->src.stride, dst, pd->dst.stride);
1354 #endif // CONFIG_VP9_HIGHBITDEPTH
1357 for (idy = 0; idy < height / 4; ++idy) {
1358 for (idx = 0; idx < width / 4; ++idx) {
1359 int64_t ssz, rd, rd1, rd2;
1362 k += (idy * 2 + idx);
1363 coeff = BLOCK_OFFSET(p->coeff, k);
1364 x->fwd_txm4x4(vp10_raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
1366 vp10_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan);
1367 #if CONFIG_VP9_HIGHBITDEPTH
1368 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1369 thisdistortion += vp10_highbd_block_error(coeff,
1370 BLOCK_OFFSET(pd->dqcoeff, k),
1373 thisdistortion += vp10_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
1377 thisdistortion += vp10_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
1379 #endif // CONFIG_VP9_HIGHBITDEPTH
1381 thisrate += cost_coeffs(x, 0, k, ta + (k & 1), tl + (k >> 1), TX_4X4,
1382 so->scan, so->neighbors,
1383 cpi->sf.use_fast_coef_costing);
1384 rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
1385 rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
1392 *distortion = thisdistortion >> 2;
1393 *labelyrate = thisrate;
1394 *sse = thissse >> 2;
1396 return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
1407 ENTROPY_CONTEXT ta[2];
1408 ENTROPY_CONTEXT tl[2];
1420 PREDICTION_MODE modes[4];
1421 SEG_RDSTAT rdstat[4][INTER_MODES];
1425 static INLINE int mv_check_bounds(const MACROBLOCK *x, const MV *mv) {
1426 return (mv->row >> 3) < x->mv_row_min ||
1427 (mv->row >> 3) > x->mv_row_max ||
1428 (mv->col >> 3) < x->mv_col_min ||
1429 (mv->col >> 3) > x->mv_col_max;
1432 static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
1433 MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0]->mbmi;
1434 struct macroblock_plane *const p = &x->plane[0];
1435 struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
1437 p->src.buf = &p->src.buf[vp10_raster_block_offset(BLOCK_8X8, i,
1439 assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
1440 pd->pre[0].buf = &pd->pre[0].buf[vp10_raster_block_offset(BLOCK_8X8, i,
1441 pd->pre[0].stride)];
1442 if (has_second_ref(mbmi))
1443 pd->pre[1].buf = &pd->pre[1].buf[vp10_raster_block_offset(BLOCK_8X8, i,
1444 pd->pre[1].stride)];
1447 static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
1448 struct buf_2d orig_pre[2]) {
1449 MB_MODE_INFO *mbmi = &x->e_mbd.mi[0]->mbmi;
1450 x->plane[0].src = orig_src;
1451 x->e_mbd.plane[0].pre[0] = orig_pre[0];
1452 if (has_second_ref(mbmi))
1453 x->e_mbd.plane[0].pre[1] = orig_pre[1];
1456 static INLINE int mv_has_subpel(const MV *mv) {
1457 return (mv->row & 0x0F) || (mv->col & 0x0F);
1460 // Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
1461 // TODO(aconverse): Find out if this is still productive then clean up or remove
1462 static int check_best_zero_mv(
1463 const VP10_COMP *cpi, const uint8_t mode_context[MAX_REF_FRAMES],
1464 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int this_mode,
1465 const MV_REFERENCE_FRAME ref_frames[2]) {
1466 if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
1467 frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
1468 (ref_frames[1] == NONE ||
1469 frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
1470 int rfc = mode_context[ref_frames[0]];
1471 int c1 = cost_mv_ref(cpi, NEARMV, rfc);
1472 int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
1473 int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
1475 if (this_mode == NEARMV) {
1476 if (c1 > c3) return 0;
1477 } else if (this_mode == NEARESTMV) {
1478 if (c2 > c3) return 0;
1480 assert(this_mode == ZEROMV);
1481 if (ref_frames[1] == NONE) {
1482 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
1483 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
1486 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
1487 frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
1488 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
1489 frame_mv[NEARMV][ref_frames[1]].as_int == 0))
1497 static void joint_motion_search(VP10_COMP *cpi, MACROBLOCK *x,
1500 int mi_row, int mi_col,
1501 int_mv single_newmv[MAX_REF_FRAMES],
1503 const VP10_COMMON *const cm = &cpi->common;
1504 const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
1505 const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
1506 MACROBLOCKD *xd = &x->e_mbd;
1507 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
1508 const int refs[2] = {mbmi->ref_frame[0],
1509 mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]};
1512 const InterpKernel *kernel = vp10_filter_kernels[mbmi->interp_filter];
1513 struct scale_factors sf;
1515 // Do joint motion search in compound mode to get more accurate mv.
1516 struct buf_2d backup_yv12[2][MAX_MB_PLANE];
1517 int last_besterr[2] = {INT_MAX, INT_MAX};
1518 const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
1519 vp10_get_scaled_ref_frame(cpi, mbmi->ref_frame[0]),
1520 vp10_get_scaled_ref_frame(cpi, mbmi->ref_frame[1])
1523 // Prediction buffer from second frame.
1524 #if CONFIG_VP9_HIGHBITDEPTH
1525 DECLARE_ALIGNED(16, uint16_t, second_pred_alloc_16[64 * 64]);
1526 uint8_t *second_pred;
1528 DECLARE_ALIGNED(16, uint8_t, second_pred[64 * 64]);
1529 #endif // CONFIG_VP9_HIGHBITDEPTH
1531 for (ref = 0; ref < 2; ++ref) {
1532 ref_mv[ref] = x->mbmi_ext->ref_mvs[refs[ref]][0];
1534 if (scaled_ref_frame[ref]) {
1536 // Swap out the reference frame for a version that's been scaled to
1537 // match the resolution of the current frame, allowing the existing
1538 // motion search code to be used without additional modifications.
1539 for (i = 0; i < MAX_MB_PLANE; i++)
1540 backup_yv12[ref][i] = xd->plane[i].pre[ref];
1541 vp10_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
1545 frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
1548 // Since we have scaled the reference frames to match the size of the current
1549 // frame we must use a unit scaling factor during mode selection.
1550 #if CONFIG_VP9_HIGHBITDEPTH
1551 vp10_setup_scale_factors_for_frame(&sf, cm->width, cm->height,
1552 cm->width, cm->height,
1553 cm->use_highbitdepth);
1555 vp10_setup_scale_factors_for_frame(&sf, cm->width, cm->height,
1556 cm->width, cm->height);
1557 #endif // CONFIG_VP9_HIGHBITDEPTH
1559 // Allow joint search multiple times iteratively for each reference frame
1560 // and break out of the search loop if it couldn't find a better mv.
1561 for (ite = 0; ite < 4; ite++) {
1562 struct buf_2d ref_yv12[2];
1563 int bestsme = INT_MAX;
1564 int sadpb = x->sadperbit16;
1566 int search_range = 3;
1568 int tmp_col_min = x->mv_col_min;
1569 int tmp_col_max = x->mv_col_max;
1570 int tmp_row_min = x->mv_row_min;
1571 int tmp_row_max = x->mv_row_max;
1572 int id = ite % 2; // Even iterations search in the first reference frame,
1573 // odd iterations search in the second. The predictor
1574 // found for the 'other' reference frame is factored in.
1576 // Initialized here because of compiler problem in Visual Studio.
1577 ref_yv12[0] = xd->plane[0].pre[0];
1578 ref_yv12[1] = xd->plane[0].pre[1];
1580 // Get the prediction block from the 'other' reference frame.
1581 #if CONFIG_VP9_HIGHBITDEPTH
1582 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1583 second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc_16);
1584 vp10_highbd_build_inter_predictor(ref_yv12[!id].buf,
1585 ref_yv12[!id].stride,
1587 &frame_mv[refs[!id]].as_mv,
1589 kernel, MV_PRECISION_Q3,
1590 mi_col * MI_SIZE, mi_row * MI_SIZE,
1593 second_pred = (uint8_t *)second_pred_alloc_16;
1594 vp10_build_inter_predictor(ref_yv12[!id].buf,
1595 ref_yv12[!id].stride,
1597 &frame_mv[refs[!id]].as_mv,
1599 kernel, MV_PRECISION_Q3,
1600 mi_col * MI_SIZE, mi_row * MI_SIZE);
1603 vp10_build_inter_predictor(ref_yv12[!id].buf,
1604 ref_yv12[!id].stride,
1606 &frame_mv[refs[!id]].as_mv,
1608 kernel, MV_PRECISION_Q3,
1609 mi_col * MI_SIZE, mi_row * MI_SIZE);
1610 #endif // CONFIG_VP9_HIGHBITDEPTH
1612 // Do compound motion search on the current reference frame.
1614 xd->plane[0].pre[0] = ref_yv12[id];
1615 vp10_set_mv_search_range(x, &ref_mv[id].as_mv);
1617 // Use the mv result from the single mode as mv predictor.
1618 tmp_mv = frame_mv[refs[id]].as_mv;
1623 // Small-range full-pixel motion search.
1624 bestsme = vp10_refining_search_8p_c(x, &tmp_mv, sadpb,
1626 &cpi->fn_ptr[bsize],
1627 &ref_mv[id].as_mv, second_pred);
1628 if (bestsme < INT_MAX)
1629 bestsme = vp10_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
1630 second_pred, &cpi->fn_ptr[bsize], 1);
1632 x->mv_col_min = tmp_col_min;
1633 x->mv_col_max = tmp_col_max;
1634 x->mv_row_min = tmp_row_min;
1635 x->mv_row_max = tmp_row_max;
1637 if (bestsme < INT_MAX) {
1638 int dis; /* TODO: use dis in distortion calculation later. */
1640 bestsme = cpi->find_fractional_mv_step(
1643 cpi->common.allow_high_precision_mv,
1645 &cpi->fn_ptr[bsize],
1646 0, cpi->sf.mv.subpel_iters_per_step,
1648 x->nmvjointcost, x->mvcost,
1649 &dis, &sse, second_pred,
1653 // Restore the pointer to the first (possibly scaled) prediction buffer.
1655 xd->plane[0].pre[0] = ref_yv12[0];
1657 if (bestsme < last_besterr[id]) {
1658 frame_mv[refs[id]].as_mv = tmp_mv;
1659 last_besterr[id] = bestsme;
1667 for (ref = 0; ref < 2; ++ref) {
1668 if (scaled_ref_frame[ref]) {
1669 // Restore the prediction frame pointers to their unscaled versions.
1671 for (i = 0; i < MAX_MB_PLANE; i++)
1672 xd->plane[i].pre[ref] = backup_yv12[ref][i];
1675 *rate_mv += vp10_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
1676 &x->mbmi_ext->ref_mvs[refs[ref]][0].as_mv,
1677 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1681 static int64_t rd_pick_best_sub8x8_mode(VP10_COMP *cpi, MACROBLOCK *x,
1682 int_mv *best_ref_mv,
1683 int_mv *second_best_ref_mv,
1684 int64_t best_rd, int *returntotrate,
1686 int64_t *returndistortion,
1687 int *skippable, int64_t *psse,
1689 int_mv seg_mvs[4][MAX_REF_FRAMES],
1690 BEST_SEG_INFO *bsi_buf, int filter_idx,
1691 int mi_row, int mi_col) {
1693 BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
1694 MACROBLOCKD *xd = &x->e_mbd;
1695 MODE_INFO *mi = xd->mi[0];
1696 MB_MODE_INFO *mbmi = &mi->mbmi;
1698 int k, br = 0, idx, idy;
1699 int64_t bd = 0, block_sse = 0;
1700 PREDICTION_MODE this_mode;
1701 VP10_COMMON *cm = &cpi->common;
1702 struct macroblock_plane *const p = &x->plane[0];
1703 struct macroblockd_plane *const pd = &xd->plane[0];
1704 const int label_count = 4;
1705 int64_t this_segment_rd = 0;
1706 int label_mv_thresh;
1707 int segmentyrate = 0;
1708 const BLOCK_SIZE bsize = mbmi->sb_type;
1709 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1710 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1711 ENTROPY_CONTEXT t_above[2], t_left[2];
1712 int subpelmv = 1, have_ref = 0;
1713 const int has_second_rf = has_second_ref(mbmi);
1714 const int inter_mode_mask = cpi->sf.inter_mode_mask[bsize];
1715 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1719 bsi->segment_rd = best_rd;
1720 bsi->ref_mv[0] = best_ref_mv;
1721 bsi->ref_mv[1] = second_best_ref_mv;
1722 bsi->mvp.as_int = best_ref_mv->as_int;
1723 bsi->mvthresh = mvthresh;
1725 for (i = 0; i < 4; i++)
1726 bsi->modes[i] = ZEROMV;
1728 memcpy(t_above, pd->above_context, sizeof(t_above));
1729 memcpy(t_left, pd->left_context, sizeof(t_left));
1731 // 64 makes this threshold really big effectively
1732 // making it so that we very rarely check mvs on
1733 // segments. setting this to 1 would make mv thresh
1734 // roughly equal to what it is for macroblocks
1735 label_mv_thresh = 1 * bsi->mvthresh / label_count;
1737 // Segmentation method overheads
1738 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1739 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1740 // TODO(jingning,rbultje): rewrite the rate-distortion optimization
1741 // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
1742 int_mv mode_mv[MB_MODE_COUNT][2];
1743 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1744 PREDICTION_MODE mode_selected = ZEROMV;
1745 int64_t best_rd = INT64_MAX;
1746 const int i = idy * 2 + idx;
1749 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1750 const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
1751 frame_mv[ZEROMV][frame].as_int = 0;
1752 vp10_append_sub8x8_mvs_for_idx(cm, xd, i, ref, mi_row, mi_col,
1753 &frame_mv[NEARESTMV][frame],
1754 &frame_mv[NEARMV][frame],
1755 mbmi_ext->mode_context);
1758 // search for the best motion vector on this segment
1759 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1760 const struct buf_2d orig_src = x->plane[0].src;
1761 struct buf_2d orig_pre[2];
1763 mode_idx = INTER_OFFSET(this_mode);
1764 bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
1765 if (!(inter_mode_mask & (1 << this_mode)))
1768 if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
1769 this_mode, mbmi->ref_frame))
1772 memcpy(orig_pre, pd->pre, sizeof(orig_pre));
1773 memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
1774 sizeof(bsi->rdstat[i][mode_idx].ta));
1775 memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
1776 sizeof(bsi->rdstat[i][mode_idx].tl));
1778 // motion search for newmv (single predictor case only)
1779 if (!has_second_rf && this_mode == NEWMV &&
1780 seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV) {
1781 MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
1783 int thissme, bestsme = INT_MAX;
1784 int sadpb = x->sadperbit4;
1789 /* Is the best so far sufficiently good that we cant justify doing
1790 * and new motion search. */
1791 if (best_rd < label_mv_thresh)
1794 if (cpi->oxcf.mode != BEST) {
1795 // use previous block's result as next block's MV predictor.
1797 bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
1799 bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
1803 max_mv = x->max_mv_context[mbmi->ref_frame[0]];
1805 max_mv = MAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
1807 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
1808 // Take wtd average of the step_params based on the last frame's
1809 // max mv magnitude and the best ref mvs of the current block for
1810 // the given reference.
1811 step_param = (vp10_init_search_range(max_mv) +
1812 cpi->mv_step_param) / 2;
1814 step_param = cpi->mv_step_param;
1817 mvp_full.row = bsi->mvp.as_mv.row >> 3;
1818 mvp_full.col = bsi->mvp.as_mv.col >> 3;
1820 if (cpi->sf.adaptive_motion_search) {
1821 mvp_full.row = x->pred_mv[mbmi->ref_frame[0]].row >> 3;
1822 mvp_full.col = x->pred_mv[mbmi->ref_frame[0]].col >> 3;
1823 step_param = MAX(step_param, 8);
1826 // adjust src pointer for this block
1829 vp10_set_mv_search_range(x, &bsi->ref_mv[0]->as_mv);
1831 bestsme = vp10_full_pixel_search(
1832 cpi, x, bsize, &mvp_full, step_param, sadpb,
1833 cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
1834 &bsi->ref_mv[0]->as_mv, new_mv,
1837 // Should we do a full search (best quality only)
1838 if (cpi->oxcf.mode == BEST) {
1839 int_mv *const best_mv = &mi->bmi[i].as_mv[0];
1840 /* Check if mvp_full is within the range. */
1841 clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max,
1842 x->mv_row_min, x->mv_row_max);
1843 thissme = cpi->full_search_sad(x, &mvp_full,
1844 sadpb, 16, &cpi->fn_ptr[bsize],
1845 &bsi->ref_mv[0]->as_mv,
1847 cost_list[1] = cost_list[2] = cost_list[3] = cost_list[4] = INT_MAX;
1848 if (thissme < bestsme) {
1850 *new_mv = best_mv->as_mv;
1852 // The full search result is actually worse so re-instate the
1853 // previous best vector
1854 best_mv->as_mv = *new_mv;
1858 if (bestsme < INT_MAX) {
1860 cpi->find_fractional_mv_step(
1863 &bsi->ref_mv[0]->as_mv,
1864 cm->allow_high_precision_mv,
1865 x->errorperbit, &cpi->fn_ptr[bsize],
1866 cpi->sf.mv.subpel_force_stop,
1867 cpi->sf.mv.subpel_iters_per_step,
1868 cond_cost_list(cpi, cost_list),
1869 x->nmvjointcost, x->mvcost,
1871 &x->pred_sse[mbmi->ref_frame[0]],
1874 // save motion search result for use in compound prediction
1875 seg_mvs[i][mbmi->ref_frame[0]].as_mv = *new_mv;
1878 if (cpi->sf.adaptive_motion_search)
1879 x->pred_mv[mbmi->ref_frame[0]] = *new_mv;
1881 // restore src pointers
1882 mi_buf_restore(x, orig_src, orig_pre);
1885 if (has_second_rf) {
1886 if (seg_mvs[i][mbmi->ref_frame[1]].as_int == INVALID_MV ||
1887 seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV)
1891 if (has_second_rf && this_mode == NEWMV &&
1892 mbmi->interp_filter == EIGHTTAP) {
1893 // adjust src pointers
1895 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
1897 joint_motion_search(cpi, x, bsize, frame_mv[this_mode],
1898 mi_row, mi_col, seg_mvs[i],
1900 seg_mvs[i][mbmi->ref_frame[0]].as_int =
1901 frame_mv[this_mode][mbmi->ref_frame[0]].as_int;
1902 seg_mvs[i][mbmi->ref_frame[1]].as_int =
1903 frame_mv[this_mode][mbmi->ref_frame[1]].as_int;
1905 // restore src pointers
1906 mi_buf_restore(x, orig_src, orig_pre);
1909 bsi->rdstat[i][mode_idx].brate =
1910 set_and_cost_bmi_mvs(cpi, x, xd, i, this_mode, mode_mv[this_mode],
1911 frame_mv, seg_mvs[i], bsi->ref_mv,
1912 x->nmvjointcost, x->mvcost);
1914 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1915 bsi->rdstat[i][mode_idx].mvs[ref].as_int =
1916 mode_mv[this_mode][ref].as_int;
1917 if (num_4x4_blocks_wide > 1)
1918 bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int =
1919 mode_mv[this_mode][ref].as_int;
1920 if (num_4x4_blocks_high > 1)
1921 bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int =
1922 mode_mv[this_mode][ref].as_int;
1925 // Trap vectors that reach beyond the UMV borders
1926 if (mv_check_bounds(x, &mode_mv[this_mode][0].as_mv) ||
1928 mv_check_bounds(x, &mode_mv[this_mode][1].as_mv)))
1931 if (filter_idx > 0) {
1932 BEST_SEG_INFO *ref_bsi = bsi_buf;
1936 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1937 subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
1938 have_ref &= mode_mv[this_mode][ref].as_int ==
1939 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1942 if (filter_idx > 1 && !subpelmv && !have_ref) {
1943 ref_bsi = bsi_buf + 1;
1945 for (ref = 0; ref < 1 + has_second_rf; ++ref)
1946 have_ref &= mode_mv[this_mode][ref].as_int ==
1947 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1950 if (!subpelmv && have_ref &&
1951 ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1952 memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
1953 sizeof(SEG_RDSTAT));
1954 if (num_4x4_blocks_wide > 1)
1955 bsi->rdstat[i + 1][mode_idx].eobs =
1956 ref_bsi->rdstat[i + 1][mode_idx].eobs;
1957 if (num_4x4_blocks_high > 1)
1958 bsi->rdstat[i + 2][mode_idx].eobs =
1959 ref_bsi->rdstat[i + 2][mode_idx].eobs;
1961 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
1962 mode_selected = this_mode;
1963 best_rd = bsi->rdstat[i][mode_idx].brdcost;
1969 bsi->rdstat[i][mode_idx].brdcost =
1970 encode_inter_mb_segment(cpi, x,
1971 bsi->segment_rd - this_segment_rd, i,
1972 &bsi->rdstat[i][mode_idx].byrate,
1973 &bsi->rdstat[i][mode_idx].bdist,
1974 &bsi->rdstat[i][mode_idx].bsse,
1975 bsi->rdstat[i][mode_idx].ta,
1976 bsi->rdstat[i][mode_idx].tl,
1978 if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1979 bsi->rdstat[i][mode_idx].brdcost += RDCOST(x->rdmult, x->rddiv,
1980 bsi->rdstat[i][mode_idx].brate, 0);
1981 bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
1982 bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
1983 if (num_4x4_blocks_wide > 1)
1984 bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
1985 if (num_4x4_blocks_high > 1)
1986 bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
1989 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
1990 mode_selected = this_mode;
1991 best_rd = bsi->rdstat[i][mode_idx].brdcost;
1993 } /*for each 4x4 mode*/
1995 if (best_rd == INT64_MAX) {
1997 for (iy = i + 1; iy < 4; ++iy)
1998 for (midx = 0; midx < INTER_MODES; ++midx)
1999 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2000 bsi->segment_rd = INT64_MAX;
2004 mode_idx = INTER_OFFSET(mode_selected);
2005 memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
2006 memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
2008 set_and_cost_bmi_mvs(cpi, x, xd, i, mode_selected, mode_mv[mode_selected],
2009 frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
2012 br += bsi->rdstat[i][mode_idx].brate;
2013 bd += bsi->rdstat[i][mode_idx].bdist;
2014 block_sse += bsi->rdstat[i][mode_idx].bsse;
2015 segmentyrate += bsi->rdstat[i][mode_idx].byrate;
2016 this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
2018 if (this_segment_rd > bsi->segment_rd) {
2020 for (iy = i + 1; iy < 4; ++iy)
2021 for (midx = 0; midx < INTER_MODES; ++midx)
2022 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2023 bsi->segment_rd = INT64_MAX;
2027 } /* for each label */
2031 bsi->segment_yrate = segmentyrate;
2032 bsi->segment_rd = this_segment_rd;
2033 bsi->sse = block_sse;
2035 // update the coding decisions
2036 for (k = 0; k < 4; ++k)
2037 bsi->modes[k] = mi->bmi[k].as_mode;
2039 if (bsi->segment_rd > best_rd)
2041 /* set it to the best */
2042 for (i = 0; i < 4; i++) {
2043 mode_idx = INTER_OFFSET(bsi->modes[i]);
2044 mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
2045 if (has_second_ref(mbmi))
2046 mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
2047 x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
2048 mi->bmi[i].as_mode = bsi->modes[i];
2052 * used to set mbmi->mv.as_int
2054 *returntotrate = bsi->r;
2055 *returndistortion = bsi->d;
2056 *returnyrate = bsi->segment_yrate;
2057 *skippable = vp10_is_skippable_in_plane(x, BLOCK_8X8, 0);
2059 mbmi->mode = bsi->modes[3];
2061 return bsi->segment_rd;
2064 static void estimate_ref_frame_costs(const VP10_COMMON *cm,
2065 const MACROBLOCKD *xd,
2067 unsigned int *ref_costs_single,
2068 unsigned int *ref_costs_comp,
2069 vpx_prob *comp_mode_p) {
2070 int seg_ref_active = segfeature_active(&cm->seg, segment_id,
2072 if (seg_ref_active) {
2073 memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
2074 memset(ref_costs_comp, 0, MAX_REF_FRAMES * sizeof(*ref_costs_comp));
2077 vpx_prob intra_inter_p = vp10_get_intra_inter_prob(cm, xd);
2078 vpx_prob comp_inter_p = 128;
2080 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
2081 comp_inter_p = vp10_get_reference_mode_prob(cm, xd);
2082 *comp_mode_p = comp_inter_p;
2087 ref_costs_single[INTRA_FRAME] = vp10_cost_bit(intra_inter_p, 0);
2089 if (cm->reference_mode != COMPOUND_REFERENCE) {
2090 vpx_prob ref_single_p1 = vp10_get_pred_prob_single_ref_p1(cm, xd);
2091 vpx_prob ref_single_p2 = vp10_get_pred_prob_single_ref_p2(cm, xd);
2092 unsigned int base_cost = vp10_cost_bit(intra_inter_p, 1);
2094 if (cm->reference_mode == REFERENCE_MODE_SELECT)
2095 base_cost += vp10_cost_bit(comp_inter_p, 0);
2097 ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] =
2098 ref_costs_single[ALTREF_FRAME] = base_cost;
2099 ref_costs_single[LAST_FRAME] += vp10_cost_bit(ref_single_p1, 0);
2100 ref_costs_single[GOLDEN_FRAME] += vp10_cost_bit(ref_single_p1, 1);
2101 ref_costs_single[ALTREF_FRAME] += vp10_cost_bit(ref_single_p1, 1);
2102 ref_costs_single[GOLDEN_FRAME] += vp10_cost_bit(ref_single_p2, 0);
2103 ref_costs_single[ALTREF_FRAME] += vp10_cost_bit(ref_single_p2, 1);
2105 ref_costs_single[LAST_FRAME] = 512;
2106 ref_costs_single[GOLDEN_FRAME] = 512;
2107 ref_costs_single[ALTREF_FRAME] = 512;
2109 if (cm->reference_mode != SINGLE_REFERENCE) {
2110 vpx_prob ref_comp_p = vp10_get_pred_prob_comp_ref_p(cm, xd);
2111 unsigned int base_cost = vp10_cost_bit(intra_inter_p, 1);
2113 if (cm->reference_mode == REFERENCE_MODE_SELECT)
2114 base_cost += vp10_cost_bit(comp_inter_p, 1);
2116 ref_costs_comp[LAST_FRAME] = base_cost + vp10_cost_bit(ref_comp_p, 0);
2117 ref_costs_comp[GOLDEN_FRAME] = base_cost + vp10_cost_bit(ref_comp_p, 1);
2119 ref_costs_comp[LAST_FRAME] = 512;
2120 ref_costs_comp[GOLDEN_FRAME] = 512;
2125 static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
2127 int64_t comp_pred_diff[REFERENCE_MODES],
2128 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS],
2130 MACROBLOCKD *const xd = &x->e_mbd;
2132 // Take a snapshot of the coding context so it can be
2133 // restored if we decide to encode this way
2134 ctx->skip = x->skip;
2135 ctx->skippable = skippable;
2136 ctx->best_mode_index = mode_index;
2137 ctx->mic = *xd->mi[0];
2138 ctx->mbmi_ext = *x->mbmi_ext;
2139 ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
2140 ctx->comp_pred_diff = (int)comp_pred_diff[COMPOUND_REFERENCE];
2141 ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
2143 memcpy(ctx->best_filter_diff, best_filter_diff,
2144 sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
2147 static void setup_buffer_inter(VP10_COMP *cpi, MACROBLOCK *x,
2148 MV_REFERENCE_FRAME ref_frame,
2149 BLOCK_SIZE block_size,
2150 int mi_row, int mi_col,
2151 int_mv frame_nearest_mv[MAX_REF_FRAMES],
2152 int_mv frame_near_mv[MAX_REF_FRAMES],
2153 struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
2154 const VP10_COMMON *cm = &cpi->common;
2155 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2156 MACROBLOCKD *const xd = &x->e_mbd;
2157 MODE_INFO *const mi = xd->mi[0];
2158 int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
2159 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2160 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2162 assert(yv12 != NULL);
2164 // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
2165 // use the UV scaling factors.
2166 vp10_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
2168 // Gets an initial list of candidate vectors from neighbours and orders them
2169 vp10_find_mv_refs(cm, xd, mi, ref_frame, candidates, mi_row, mi_col,
2170 NULL, NULL, mbmi_ext->mode_context);
2172 // Candidate refinement carried out at encoder and decoder
2173 vp10_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2174 &frame_nearest_mv[ref_frame],
2175 &frame_near_mv[ref_frame]);
2177 // Further refinement that is encode side only to test the top few candidates
2178 // in full and choose the best as the centre point for subsequent searches.
2179 // The current implementation doesn't support scaling.
2180 if (!vp10_is_scaled(sf) && block_size >= BLOCK_8X8)
2181 vp10_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
2182 ref_frame, block_size);
2185 static void single_motion_search(VP10_COMP *cpi, MACROBLOCK *x,
2187 int mi_row, int mi_col,
2188 int_mv *tmp_mv, int *rate_mv) {
2189 MACROBLOCKD *xd = &x->e_mbd;
2190 const VP10_COMMON *cm = &cpi->common;
2191 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
2192 struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
2193 int bestsme = INT_MAX;
2195 int sadpb = x->sadperbit16;
2197 int ref = mbmi->ref_frame[0];
2198 MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2200 int tmp_col_min = x->mv_col_min;
2201 int tmp_col_max = x->mv_col_max;
2202 int tmp_row_min = x->mv_row_min;
2203 int tmp_row_max = x->mv_row_max;
2206 const YV12_BUFFER_CONFIG *scaled_ref_frame = vp10_get_scaled_ref_frame(cpi,
2210 pred_mv[0] = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2211 pred_mv[1] = x->mbmi_ext->ref_mvs[ref][1].as_mv;
2212 pred_mv[2] = x->pred_mv[ref];
2214 if (scaled_ref_frame) {
2216 // Swap out the reference frame for a version that's been scaled to
2217 // match the resolution of the current frame, allowing the existing
2218 // motion search code to be used without additional modifications.
2219 for (i = 0; i < MAX_MB_PLANE; i++)
2220 backup_yv12[i] = xd->plane[i].pre[0];
2222 vp10_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
2225 vp10_set_mv_search_range(x, &ref_mv);
2227 // Work out the size of the first step in the mv step search.
2228 // 0 here is maximum length first step. 1 is MAX >> 1 etc.
2229 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
2230 // Take wtd average of the step_params based on the last frame's
2231 // max mv magnitude and that based on the best ref mvs of the current
2232 // block for the given reference.
2233 step_param = (vp10_init_search_range(x->max_mv_context[ref]) +
2234 cpi->mv_step_param) / 2;
2236 step_param = cpi->mv_step_param;
2239 if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64) {
2240 int boffset = 2 * (b_width_log2_lookup[BLOCK_64X64] -
2241 MIN(b_height_log2_lookup[bsize], b_width_log2_lookup[bsize]));
2242 step_param = MAX(step_param, boffset);
2245 if (cpi->sf.adaptive_motion_search) {
2246 int bwl = b_width_log2_lookup[bsize];
2247 int bhl = b_height_log2_lookup[bsize];
2248 int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
2253 // prev_mv_sad is not setup for dynamically scaled frames.
2254 if (cpi->oxcf.resize_mode != RESIZE_DYNAMIC) {
2256 for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
2257 if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
2258 x->pred_mv[ref].row = 0;
2259 x->pred_mv[ref].col = 0;
2260 tmp_mv->as_int = INVALID_MV;
2262 if (scaled_ref_frame) {
2264 for (i = 0; i < MAX_MB_PLANE; ++i)
2265 xd->plane[i].pre[0] = backup_yv12[i];
2273 mvp_full = pred_mv[x->mv_best_ref_index[ref]];
2278 bestsme = vp10_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
2279 cond_cost_list(cpi, cost_list),
2280 &ref_mv, &tmp_mv->as_mv, INT_MAX, 1);
2282 x->mv_col_min = tmp_col_min;
2283 x->mv_col_max = tmp_col_max;
2284 x->mv_row_min = tmp_row_min;
2285 x->mv_row_max = tmp_row_max;
2287 if (bestsme < INT_MAX) {
2288 int dis; /* TODO: use dis in distortion calculation later. */
2289 cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
2290 cm->allow_high_precision_mv,
2292 &cpi->fn_ptr[bsize],
2293 cpi->sf.mv.subpel_force_stop,
2294 cpi->sf.mv.subpel_iters_per_step,
2295 cond_cost_list(cpi, cost_list),
2296 x->nmvjointcost, x->mvcost,
2297 &dis, &x->pred_sse[ref], NULL, 0, 0);
2299 *rate_mv = vp10_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
2300 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2302 if (cpi->sf.adaptive_motion_search)
2303 x->pred_mv[ref] = tmp_mv->as_mv;
2305 if (scaled_ref_frame) {
2307 for (i = 0; i < MAX_MB_PLANE; i++)
2308 xd->plane[i].pre[0] = backup_yv12[i];
2314 static INLINE void restore_dst_buf(MACROBLOCKD *xd,
2315 uint8_t *orig_dst[MAX_MB_PLANE],
2316 int orig_dst_stride[MAX_MB_PLANE]) {
2318 for (i = 0; i < MAX_MB_PLANE; i++) {
2319 xd->plane[i].dst.buf = orig_dst[i];
2320 xd->plane[i].dst.stride = orig_dst_stride[i];
2324 // In some situations we want to discount tha pparent cost of a new motion
2325 // vector. Where there is a subtle motion field and especially where there is
2326 // low spatial complexity then it can be hard to cover the cost of a new motion
2327 // vector in a single block, even if that motion vector reduces distortion.
2328 // However, once established that vector may be usable through the nearest and
2329 // near mv modes to reduce distortion in subsequent blocks and also improve
2331 static int discount_newmv_test(const VP10_COMP *cpi,
2334 int_mv (*mode_mv)[MAX_REF_FRAMES],
2336 return (!cpi->rc.is_src_frame_alt_ref &&
2337 (this_mode == NEWMV) &&
2338 (this_mv.as_int != 0) &&
2339 ((mode_mv[NEARESTMV][ref_frame].as_int == 0) ||
2340 (mode_mv[NEARESTMV][ref_frame].as_int == INVALID_MV)) &&
2341 ((mode_mv[NEARMV][ref_frame].as_int == 0) ||
2342 (mode_mv[NEARMV][ref_frame].as_int == INVALID_MV)));
2345 static int64_t handle_inter_mode(VP10_COMP *cpi, MACROBLOCK *x,
2347 int *rate2, int64_t *distortion,
2349 int *rate_y, int *rate_uv,
2351 int_mv (*mode_mv)[MAX_REF_FRAMES],
2352 int mi_row, int mi_col,
2353 int_mv single_newmv[MAX_REF_FRAMES],
2354 INTERP_FILTER (*single_filter)[MAX_REF_FRAMES],
2355 int (*single_skippable)[MAX_REF_FRAMES],
2357 const int64_t ref_best_rd,
2358 int64_t *mask_filter,
2359 int64_t filter_cache[]) {
2360 VP10_COMMON *cm = &cpi->common;
2361 MACROBLOCKD *xd = &x->e_mbd;
2362 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
2363 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2364 const int is_comp_pred = has_second_ref(mbmi);
2365 const int this_mode = mbmi->mode;
2366 int_mv *frame_mv = mode_mv[this_mode];
2368 int refs[2] = { mbmi->ref_frame[0],
2369 (mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]) };
2371 #if CONFIG_VP9_HIGHBITDEPTH
2372 DECLARE_ALIGNED(16, uint16_t, tmp_buf16[MAX_MB_PLANE * 64 * 64]);
2375 DECLARE_ALIGNED(16, uint8_t, tmp_buf[MAX_MB_PLANE * 64 * 64]);
2376 #endif // CONFIG_VP9_HIGHBITDEPTH
2377 int pred_exists = 0;
2379 int64_t rd, tmp_rd, best_rd = INT64_MAX;
2380 int best_needs_copy = 0;
2381 uint8_t *orig_dst[MAX_MB_PLANE];
2382 int orig_dst_stride[MAX_MB_PLANE];
2384 INTERP_FILTER best_filter = SWITCHABLE;
2385 uint8_t skip_txfm[MAX_MB_PLANE << 2] = {0};
2386 int64_t bsse[MAX_MB_PLANE << 2] = {0};
2388 int bsl = mi_width_log2_lookup[bsize];
2389 int pred_filter_search = cpi->sf.cb_pred_filter_search ?
2390 (((mi_row + mi_col) >> bsl) +
2391 get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
2393 int skip_txfm_sb = 0;
2394 int64_t skip_sse_sb = INT64_MAX;
2395 int64_t distortion_y = 0, distortion_uv = 0;
2397 #if CONFIG_VP9_HIGHBITDEPTH
2398 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2399 tmp_buf = CONVERT_TO_BYTEPTR(tmp_buf16);
2401 tmp_buf = (uint8_t *)tmp_buf16;
2403 #endif // CONFIG_VP9_HIGHBITDEPTH
2405 if (pred_filter_search) {
2406 INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE;
2407 if (xd->up_available)
2408 af = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
2409 if (xd->left_available)
2410 lf = xd->mi[-1]->mbmi.interp_filter;
2412 if ((this_mode != NEWMV) || (af == lf))
2417 if (frame_mv[refs[0]].as_int == INVALID_MV ||
2418 frame_mv[refs[1]].as_int == INVALID_MV)
2421 if (cpi->sf.adaptive_mode_search) {
2422 if (single_filter[this_mode][refs[0]] ==
2423 single_filter[this_mode][refs[1]])
2424 best_filter = single_filter[this_mode][refs[0]];
2428 if (this_mode == NEWMV) {
2431 // Initialize mv using single prediction mode result.
2432 frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
2433 frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
2435 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
2436 joint_motion_search(cpi, x, bsize, frame_mv,
2437 mi_row, mi_col, single_newmv, &rate_mv);
2439 rate_mv = vp10_mv_bit_cost(&frame_mv[refs[0]].as_mv,
2440 &x->mbmi_ext->ref_mvs[refs[0]][0].as_mv,
2441 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2442 rate_mv += vp10_mv_bit_cost(&frame_mv[refs[1]].as_mv,
2443 &x->mbmi_ext->ref_mvs[refs[1]][0].as_mv,
2444 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2449 single_motion_search(cpi, x, bsize, mi_row, mi_col,
2451 if (tmp_mv.as_int == INVALID_MV)
2454 frame_mv[refs[0]].as_int =
2455 xd->mi[0]->bmi[0].as_mv[0].as_int = tmp_mv.as_int;
2456 single_newmv[refs[0]].as_int = tmp_mv.as_int;
2458 // Estimate the rate implications of a new mv but discount this
2459 // under certain circumstances where we want to help initiate a weak
2460 // motion field, where the distortion gain for a single block may not
2461 // be enough to overcome the cost of a new mv.
2462 if (discount_newmv_test(cpi, this_mode, tmp_mv, mode_mv, refs[0])) {
2463 *rate2 += MAX((rate_mv / NEW_MV_DISCOUNT_FACTOR), 1);
2470 for (i = 0; i < is_comp_pred + 1; ++i) {
2471 cur_mv[i] = frame_mv[refs[i]];
2472 // Clip "next_nearest" so that it does not extend to far out of image
2473 if (this_mode != NEWMV)
2474 clamp_mv2(&cur_mv[i].as_mv, xd);
2476 if (mv_check_bounds(x, &cur_mv[i].as_mv))
2478 mbmi->mv[i].as_int = cur_mv[i].as_int;
2481 // do first prediction into the destination buffer. Do the next
2482 // prediction into a temporary buffer. Then keep track of which one
2483 // of these currently holds the best predictor, and use the other
2484 // one for future predictions. In the end, copy from tmp_buf to
2485 // dst if necessary.
2486 for (i = 0; i < MAX_MB_PLANE; i++) {
2487 orig_dst[i] = xd->plane[i].dst.buf;
2488 orig_dst_stride[i] = xd->plane[i].dst.stride;
2491 // We don't include the cost of the second reference here, because there
2492 // are only three options: Last/Golden, ARF/Last or Golden/ARF, or in other
2493 // words if you present them in that order, the second one is always known
2494 // if the first is known.
2496 // Under some circumstances we discount the cost of new mv mode to encourage
2497 // initiation of a motion field.
2498 if (discount_newmv_test(cpi, this_mode, frame_mv[refs[0]],
2499 mode_mv, refs[0])) {
2500 *rate2 += MIN(cost_mv_ref(cpi, this_mode,
2501 mbmi_ext->mode_context[refs[0]]),
2502 cost_mv_ref(cpi, NEARESTMV,
2503 mbmi_ext->mode_context[refs[0]]));
2505 *rate2 += cost_mv_ref(cpi, this_mode, mbmi_ext->mode_context[refs[0]]);
2508 if (RDCOST(x->rdmult, x->rddiv, *rate2, 0) > ref_best_rd &&
2509 mbmi->mode != NEARESTMV)
2513 // Are all MVs integer pel for Y and UV
2514 intpel_mv = !mv_has_subpel(&mbmi->mv[0].as_mv);
2516 intpel_mv &= !mv_has_subpel(&mbmi->mv[1].as_mv);
2518 // Search for best switchable filter by checking the variance of
2519 // pred error irrespective of whether the filter will be used
2520 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
2521 filter_cache[i] = INT64_MAX;
2523 if (cm->interp_filter != BILINEAR) {
2524 if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
2525 best_filter = EIGHTTAP;
2526 } else if (best_filter == SWITCHABLE) {
2528 int tmp_rate_sum = 0;
2529 int64_t tmp_dist_sum = 0;
2531 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
2534 int tmp_skip_sb = 0;
2535 int64_t tmp_skip_sse = INT64_MAX;
2537 mbmi->interp_filter = i;
2538 rs = vp10_get_switchable_rate(cpi, xd);
2539 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
2541 if (i > 0 && intpel_mv) {
2542 rd = RDCOST(x->rdmult, x->rddiv, tmp_rate_sum, tmp_dist_sum);
2543 filter_cache[i] = rd;
2544 filter_cache[SWITCHABLE_FILTERS] =
2545 MIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2546 if (cm->interp_filter == SWITCHABLE)
2548 *mask_filter = MAX(*mask_filter, rd);
2551 int64_t dist_sum = 0;
2552 if (i > 0 && cpi->sf.adaptive_interp_filter_search &&
2553 (cpi->sf.interp_filter_search_mask & (1 << i))) {
2555 dist_sum = INT64_MAX;
2559 if ((cm->interp_filter == SWITCHABLE &&
2560 (!i || best_needs_copy)) ||
2561 (cm->interp_filter != SWITCHABLE &&
2562 (cm->interp_filter == mbmi->interp_filter ||
2563 (i == 0 && intpel_mv)))) {
2564 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2566 for (j = 0; j < MAX_MB_PLANE; j++) {
2567 xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
2568 xd->plane[j].dst.stride = 64;
2571 vp10_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2572 model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum,
2573 &tmp_skip_sb, &tmp_skip_sse);
2575 rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum);
2576 filter_cache[i] = rd;
2577 filter_cache[SWITCHABLE_FILTERS] =
2578 MIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2579 if (cm->interp_filter == SWITCHABLE)
2581 *mask_filter = MAX(*mask_filter, rd);
2583 if (i == 0 && intpel_mv) {
2584 tmp_rate_sum = rate_sum;
2585 tmp_dist_sum = dist_sum;
2589 if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2590 if (rd / 2 > ref_best_rd) {
2591 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2595 newbest = i == 0 || rd < best_rd;
2599 best_filter = mbmi->interp_filter;
2600 if (cm->interp_filter == SWITCHABLE && i && !intpel_mv)
2601 best_needs_copy = !best_needs_copy;
2604 if ((cm->interp_filter == SWITCHABLE && newbest) ||
2605 (cm->interp_filter != SWITCHABLE &&
2606 cm->interp_filter == mbmi->interp_filter)) {
2610 skip_txfm_sb = tmp_skip_sb;
2611 skip_sse_sb = tmp_skip_sse;
2612 memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2613 memcpy(bsse, x->bsse, sizeof(bsse));
2616 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2619 // Set the appropriate filter
2620 mbmi->interp_filter = cm->interp_filter != SWITCHABLE ?
2621 cm->interp_filter : best_filter;
2622 rs = cm->interp_filter == SWITCHABLE ? vp10_get_switchable_rate(cpi, xd) : 0;
2625 if (best_needs_copy) {
2626 // again temporarily set the buffers to local memory to prevent a memcpy
2627 for (i = 0; i < MAX_MB_PLANE; i++) {
2628 xd->plane[i].dst.buf = tmp_buf + i * 64 * 64;
2629 xd->plane[i].dst.stride = 64;
2632 rd = tmp_rd + RDCOST(x->rdmult, x->rddiv, rs, 0);
2636 // Handles the special case when a filter that is not in the
2637 // switchable list (ex. bilinear) is indicated at the frame level, or
2638 // skip condition holds.
2639 vp10_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2640 model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist,
2641 &skip_txfm_sb, &skip_sse_sb);
2642 rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
2643 memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2644 memcpy(bsse, x->bsse, sizeof(bsse));
2648 single_filter[this_mode][refs[0]] = mbmi->interp_filter;
2650 if (cpi->sf.adaptive_mode_search)
2652 if (single_skippable[this_mode][refs[0]] &&
2653 single_skippable[this_mode][refs[1]])
2654 memset(skip_txfm, SKIP_TXFM_AC_DC, sizeof(skip_txfm));
2656 if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2657 // if current pred_error modeled rd is substantially more than the best
2658 // so far, do not bother doing full rd
2659 if (rd / 2 > ref_best_rd) {
2660 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2665 if (cm->interp_filter == SWITCHABLE)
2668 memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm));
2669 memcpy(x->bsse, bsse, sizeof(bsse));
2671 if (!skip_txfm_sb) {
2672 int skippable_y, skippable_uv;
2673 int64_t sseuv = INT64_MAX;
2674 int64_t rdcosty = INT64_MAX;
2676 // Y cost and distortion
2677 vp10_subtract_plane(x, bsize, 0);
2678 super_block_yrd(cpi, x, rate_y, &distortion_y, &skippable_y, psse,
2679 bsize, ref_best_rd);
2681 if (*rate_y == INT_MAX) {
2683 *distortion = INT64_MAX;
2684 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2689 *distortion += distortion_y;
2691 rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
2692 rdcosty = MIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
2694 if (!super_block_uvrd(cpi, x, rate_uv, &distortion_uv, &skippable_uv,
2695 &sseuv, bsize, ref_best_rd - rdcosty)) {
2697 *distortion = INT64_MAX;
2698 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2704 *distortion += distortion_uv;
2705 *skippable = skippable_y && skippable_uv;
2710 // The cost of skip bit needs to be added.
2711 *rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 1);
2713 *distortion = skip_sse_sb;
2717 single_skippable[this_mode][refs[0]] = *skippable;
2719 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2720 return 0; // The rate-distortion cost will be re-calculated by caller.
2723 void vp10_rd_pick_intra_mode_sb(VP10_COMP *cpi, MACROBLOCK *x,
2724 RD_COST *rd_cost, BLOCK_SIZE bsize,
2725 PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
2726 VP10_COMMON *const cm = &cpi->common;
2727 MACROBLOCKD *const xd = &x->e_mbd;
2728 struct macroblockd_plane *const pd = xd->plane;
2729 int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
2730 int y_skip = 0, uv_skip = 0;
2731 int64_t dist_y = 0, dist_uv = 0;
2732 TX_SIZE max_uv_tx_size;
2735 xd->mi[0]->mbmi.ref_frame[0] = INTRA_FRAME;
2736 xd->mi[0]->mbmi.ref_frame[1] = NONE;
2738 if (bsize >= BLOCK_8X8) {
2739 if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2740 &dist_y, &y_skip, bsize,
2741 best_rd) >= best_rd) {
2742 rd_cost->rate = INT_MAX;
2747 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2748 &dist_y, best_rd) >= best_rd) {
2749 rd_cost->rate = INT_MAX;
2753 max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0]->mbmi.tx_size, bsize,
2754 pd[1].subsampling_x,
2755 pd[1].subsampling_y);
2756 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
2757 &dist_uv, &uv_skip, MAX(BLOCK_8X8, bsize),
2760 if (y_skip && uv_skip) {
2761 rd_cost->rate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
2762 vp10_cost_bit(vp10_get_skip_prob(cm, xd), 1);
2763 rd_cost->dist = dist_y + dist_uv;
2765 rd_cost->rate = rate_y + rate_uv +
2766 vp10_cost_bit(vp10_get_skip_prob(cm, xd), 0);
2767 rd_cost->dist = dist_y + dist_uv;
2770 ctx->mic = *xd->mi[0];
2771 ctx->mbmi_ext = *x->mbmi_ext;
2772 rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
2775 // This function is designed to apply a bias or adjustment to an rd value based
2776 // on the relative variance of the source and reconstruction.
2777 #define LOW_VAR_THRESH 16
2778 #define VLOW_ADJ_MAX 25
2779 #define VHIGH_ADJ_MAX 8
2780 static void rd_variance_adjustment(VP10_COMP *cpi,
2784 MV_REFERENCE_FRAME ref_frame,
2785 unsigned int source_variance) {
2786 MACROBLOCKD *const xd = &x->e_mbd;
2787 unsigned int recon_variance;
2788 unsigned int absvar_diff = 0;
2789 int64_t var_error = 0;
2790 int64_t var_factor = 0;
2792 if (*this_rd == INT64_MAX)
2795 #if CONFIG_VP9_HIGHBITDEPTH
2796 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2798 vp10_high_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize, xd->bd);
2801 vp10_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
2805 vp10_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
2806 #endif // CONFIG_VP9_HIGHBITDEPTH
2808 if ((source_variance + recon_variance) > LOW_VAR_THRESH) {
2809 absvar_diff = (source_variance > recon_variance)
2810 ? (source_variance - recon_variance)
2811 : (recon_variance - source_variance);
2813 var_error = (200 * source_variance * recon_variance) /
2814 ((source_variance * source_variance) +
2815 (recon_variance * recon_variance));
2816 var_error = 100 - var_error;
2819 // Source variance above a threshold and ref frame is intra.
2820 // This case is targeted mainly at discouraging intra modes that give rise
2821 // to a predictor with a low spatial complexity compared to the source.
2822 if ((source_variance > LOW_VAR_THRESH) && (ref_frame == INTRA_FRAME) &&
2823 (source_variance > recon_variance)) {
2824 var_factor = MIN(absvar_diff, MIN(VLOW_ADJ_MAX, var_error));
2825 // A second possible case of interest is where the source variance
2826 // is very low and we wish to discourage false texture or motion trails.
2827 } else if ((source_variance < (LOW_VAR_THRESH >> 1)) &&
2828 (recon_variance > source_variance)) {
2829 var_factor = MIN(absvar_diff, MIN(VHIGH_ADJ_MAX, var_error));
2831 *this_rd += (*this_rd * var_factor) / 100;
2835 // Do we have an internal image edge (e.g. formatting bars).
2836 int vp10_internal_image_edge(VP10_COMP *cpi) {
2837 return (cpi->oxcf.pass == 2) &&
2838 ((cpi->twopass.this_frame_stats.inactive_zone_rows > 0) ||
2839 (cpi->twopass.this_frame_stats.inactive_zone_cols > 0));
2842 // Checks to see if a super block is on a horizontal image edge.
2843 // In most cases this is the "real" edge unless there are formatting
2844 // bars embedded in the stream.
2845 int vp10_active_h_edge(VP10_COMP *cpi, int mi_row, int mi_step) {
2847 int bottom_edge = cpi->common.mi_rows;
2848 int is_active_h_edge = 0;
2850 // For two pass account for any formatting bars detected.
2851 if (cpi->oxcf.pass == 2) {
2852 TWO_PASS *twopass = &cpi->twopass;
2854 // The inactive region is specified in MBs not mi units.
2855 // The image edge is in the following MB row.
2856 top_edge += (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
2858 bottom_edge -= (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
2859 bottom_edge = MAX(top_edge, bottom_edge);
2862 if (((top_edge >= mi_row) && (top_edge < (mi_row + mi_step))) ||
2863 ((bottom_edge >= mi_row) && (bottom_edge < (mi_row + mi_step)))) {
2864 is_active_h_edge = 1;
2866 return is_active_h_edge;
2869 // Checks to see if a super block is on a vertical image edge.
2870 // In most cases this is the "real" edge unless there are formatting
2871 // bars embedded in the stream.
2872 int vp10_active_v_edge(VP10_COMP *cpi, int mi_col, int mi_step) {
2874 int right_edge = cpi->common.mi_cols;
2875 int is_active_v_edge = 0;
2877 // For two pass account for any formatting bars detected.
2878 if (cpi->oxcf.pass == 2) {
2879 TWO_PASS *twopass = &cpi->twopass;
2881 // The inactive region is specified in MBs not mi units.
2882 // The image edge is in the following MB row.
2883 left_edge += (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
2885 right_edge -= (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
2886 right_edge = MAX(left_edge, right_edge);
2889 if (((left_edge >= mi_col) && (left_edge < (mi_col + mi_step))) ||
2890 ((right_edge >= mi_col) && (right_edge < (mi_col + mi_step)))) {
2891 is_active_v_edge = 1;
2893 return is_active_v_edge;
2896 // Checks to see if a super block is at the edge of the active image.
2897 // In most cases this is the "real" edge unless there are formatting
2898 // bars embedded in the stream.
2899 int vp10_active_edge_sb(VP10_COMP *cpi,
2900 int mi_row, int mi_col) {
2901 return vp10_active_h_edge(cpi, mi_row, MI_BLOCK_SIZE) ||
2902 vp10_active_v_edge(cpi, mi_col, MI_BLOCK_SIZE);
2905 void vp10_rd_pick_inter_mode_sb(VP10_COMP *cpi,
2906 TileDataEnc *tile_data,
2908 int mi_row, int mi_col,
2909 RD_COST *rd_cost, BLOCK_SIZE bsize,
2910 PICK_MODE_CONTEXT *ctx,
2911 int64_t best_rd_so_far) {
2912 VP10_COMMON *const cm = &cpi->common;
2913 TileInfo *const tile_info = &tile_data->tile_info;
2914 RD_OPT *const rd_opt = &cpi->rd;
2915 SPEED_FEATURES *const sf = &cpi->sf;
2916 MACROBLOCKD *const xd = &x->e_mbd;
2917 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
2918 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2919 const struct segmentation *const seg = &cm->seg;
2920 PREDICTION_MODE this_mode;
2921 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
2922 unsigned char segment_id = mbmi->segment_id;
2923 int comp_pred, i, k;
2924 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
2925 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2926 int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
2927 INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES];
2928 int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES];
2929 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2931 int64_t best_rd = best_rd_so_far;
2932 int64_t best_pred_diff[REFERENCE_MODES];
2933 int64_t best_pred_rd[REFERENCE_MODES];
2934 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
2935 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
2936 MB_MODE_INFO best_mbmode;
2937 int best_mode_skippable = 0;
2938 int midx, best_mode_index = -1;
2939 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
2940 vpx_prob comp_mode_p;
2941 int64_t best_intra_rd = INT64_MAX;
2942 unsigned int best_pred_sse = UINT_MAX;
2943 PREDICTION_MODE best_intra_mode = DC_PRED;
2944 int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
2945 int64_t dist_uv[TX_SIZES];
2946 int skip_uv[TX_SIZES];
2947 PREDICTION_MODE mode_uv[TX_SIZES];
2948 const int intra_cost_penalty = vp10_get_intra_cost_penalty(
2949 cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
2951 uint8_t ref_frame_skip_mask[2] = { 0 };
2952 uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 };
2953 int mode_skip_start = sf->mode_skip_start + 1;
2954 const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
2955 const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
2956 int64_t mode_threshold[MAX_MODES];
2957 int *mode_map = tile_data->mode_map[bsize];
2958 const int mode_search_skip_flags = sf->mode_search_skip_flags;
2959 int64_t mask_filter = 0;
2960 int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
2962 vp10_zero(best_mbmode);
2964 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
2966 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
2967 filter_cache[i] = INT64_MAX;
2969 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
2972 for (i = 0; i < REFERENCE_MODES; ++i)
2973 best_pred_rd[i] = INT64_MAX;
2974 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
2975 best_filter_rd[i] = INT64_MAX;
2976 for (i = 0; i < TX_SIZES; i++)
2977 rate_uv_intra[i] = INT_MAX;
2978 for (i = 0; i < MAX_REF_FRAMES; ++i)
2979 x->pred_sse[i] = INT_MAX;
2980 for (i = 0; i < MB_MODE_COUNT; ++i) {
2981 for (k = 0; k < MAX_REF_FRAMES; ++k) {
2982 single_inter_filter[i][k] = SWITCHABLE;
2983 single_skippable[i][k] = 0;
2987 rd_cost->rate = INT_MAX;
2989 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
2990 x->pred_mv_sad[ref_frame] = INT_MAX;
2991 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
2992 assert(get_ref_frame_buffer(cpi, ref_frame) != NULL);
2993 setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
2994 frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
2996 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
2997 frame_mv[ZEROMV][ref_frame].as_int = 0;
3000 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3001 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) {
3002 // Skip checking missing references in both single and compound reference
3003 // modes. Note that a mode will be skipped iff both reference frames
3005 ref_frame_skip_mask[0] |= (1 << ref_frame);
3006 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3007 } else if (sf->reference_masking) {
3008 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
3009 // Skip fixed mv modes for poor references
3010 if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
3011 mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
3016 // If the segment reference frame feature is enabled....
3017 // then do nothing if the current ref frame is not allowed..
3018 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3019 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3020 ref_frame_skip_mask[0] |= (1 << ref_frame);
3021 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3025 // Disable this drop out case if the ref frame
3026 // segment level feature is enabled for this segment. This is to
3027 // prevent the possibility that we end up unable to pick any mode.
3028 if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3029 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3030 // unless ARNR filtering is enabled in which case we want
3031 // an unfiltered alternative. We allow near/nearest as well
3032 // because they may result in zero-zero MVs but be cheaper.
3033 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
3034 ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME);
3035 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3036 mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
3037 if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
3038 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
3039 if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
3040 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
3044 if (cpi->rc.is_src_frame_alt_ref) {
3045 if (sf->alt_ref_search_fp) {
3046 mode_skip_mask[ALTREF_FRAME] = 0;
3047 ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME);
3048 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3052 if (sf->alt_ref_search_fp)
3053 if (!cm->show_frame && x->pred_mv_sad[GOLDEN_FRAME] < INT_MAX)
3054 if (x->pred_mv_sad[ALTREF_FRAME] > (x->pred_mv_sad[GOLDEN_FRAME] << 1))
3055 mode_skip_mask[ALTREF_FRAME] |= INTER_ALL;
3057 if (sf->adaptive_mode_search) {
3058 if (cm->show_frame && !cpi->rc.is_src_frame_alt_ref &&
3059 cpi->rc.frames_since_golden >= 3)
3060 if (x->pred_mv_sad[GOLDEN_FRAME] > (x->pred_mv_sad[LAST_FRAME] << 1))
3061 mode_skip_mask[GOLDEN_FRAME] |= INTER_ALL;
3064 if (bsize > sf->max_intra_bsize) {
3065 ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
3066 ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
3069 mode_skip_mask[INTRA_FRAME] |=
3070 ~(sf->intra_y_mode_mask[max_txsize_lookup[bsize]]);
3072 for (i = 0; i <= LAST_NEW_MV_INDEX; ++i)
3073 mode_threshold[i] = 0;
3074 for (i = LAST_NEW_MV_INDEX + 1; i < MAX_MODES; ++i)
3075 mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
3077 midx = sf->schedule_mode_search ? mode_skip_start : 0;
3079 uint8_t end_pos = 0;
3080 for (i = 5; i < midx; ++i) {
3081 if (mode_threshold[mode_map[i - 1]] > mode_threshold[mode_map[i]]) {
3082 uint8_t tmp = mode_map[i];
3083 mode_map[i] = mode_map[i - 1];
3084 mode_map[i - 1] = tmp;
3091 for (midx = 0; midx < MAX_MODES; ++midx) {
3092 int mode_index = mode_map[midx];
3093 int mode_excluded = 0;
3094 int64_t this_rd = INT64_MAX;
3095 int disable_skip = 0;
3096 int compmode_cost = 0;
3097 int rate2 = 0, rate_y = 0, rate_uv = 0;
3098 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3101 int64_t total_sse = INT64_MAX;
3104 this_mode = vp10_mode_order[mode_index].mode;
3105 ref_frame = vp10_mode_order[mode_index].ref_frame[0];
3106 second_ref_frame = vp10_mode_order[mode_index].ref_frame[1];
3108 // Look at the reference frame of the best mode so far and set the
3109 // skip mask to look at a subset of the remaining modes.
3110 if (midx == mode_skip_start && best_mode_index >= 0) {
3111 switch (best_mbmode.ref_frame[0]) {
3115 ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK;
3116 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3119 ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
3120 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3123 ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK;
3126 case MAX_REF_FRAMES:
3127 assert(0 && "Invalid Reference frame");
3132 if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3133 (ref_frame_skip_mask[1] & (1 << MAX(0, second_ref_frame))))
3136 if (mode_skip_mask[ref_frame] & (1 << this_mode))
3139 // Test best rd so far against threshold for trying this mode.
3140 if (best_mode_skippable && sf->schedule_mode_search)
3141 mode_threshold[mode_index] <<= 1;
3143 if (best_rd < mode_threshold[mode_index])
3146 if (sf->motion_field_mode_search) {
3147 const int mi_width = MIN(num_8x8_blocks_wide_lookup[bsize],
3148 tile_info->mi_col_end - mi_col);
3149 const int mi_height = MIN(num_8x8_blocks_high_lookup[bsize],
3150 tile_info->mi_row_end - mi_row);
3151 const int bsl = mi_width_log2_lookup[bsize];
3152 int cb_partition_search_ctrl = (((mi_row + mi_col) >> bsl)
3153 + get_chessboard_index(cm->current_video_frame)) & 0x1;
3154 MB_MODE_INFO *ref_mbmi;
3155 int const_motion = 1;
3156 int skip_ref_frame = !cb_partition_search_ctrl;
3157 MV_REFERENCE_FRAME rf = NONE;
3159 ref_mv.as_int = INVALID_MV;
3161 if ((mi_row - 1) >= tile_info->mi_row_start) {
3162 ref_mv = xd->mi[-xd->mi_stride]->mbmi.mv[0];
3163 rf = xd->mi[-xd->mi_stride]->mbmi.ref_frame[0];
3164 for (i = 0; i < mi_width; ++i) {
3165 ref_mbmi = &xd->mi[-xd->mi_stride + i]->mbmi;
3166 const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
3167 (ref_frame == ref_mbmi->ref_frame[0]);
3168 skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
3172 if ((mi_col - 1) >= tile_info->mi_col_start) {
3173 if (ref_mv.as_int == INVALID_MV)
3174 ref_mv = xd->mi[-1]->mbmi.mv[0];
3176 rf = xd->mi[-1]->mbmi.ref_frame[0];
3177 for (i = 0; i < mi_height; ++i) {
3178 ref_mbmi = &xd->mi[i * xd->mi_stride - 1]->mbmi;
3179 const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
3180 (ref_frame == ref_mbmi->ref_frame[0]);
3181 skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
3185 if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV)
3186 if (rf > INTRA_FRAME)
3187 if (ref_frame != rf)
3191 if (this_mode == NEARMV || this_mode == ZEROMV)
3195 comp_pred = second_ref_frame > INTRA_FRAME;
3197 if (!cpi->allow_comp_inter_inter)
3200 // Skip compound inter modes if ARF is not available.
3201 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
3204 // Do not allow compound prediction if the segment level reference frame
3205 // feature is in use as in this case there can only be one reference.
3206 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
3209 if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3210 best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
3213 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3215 if (ref_frame != INTRA_FRAME)
3216 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3219 if (ref_frame == INTRA_FRAME) {
3220 if (sf->adaptive_mode_search)
3221 if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
3224 if (this_mode != DC_PRED) {
3225 // Disable intra modes other than DC_PRED for blocks with low variance
3226 // Threshold for intra skipping based on source variance
3227 // TODO(debargha): Specialize the threshold for super block sizes
3228 const unsigned int skip_intra_var_thresh = 64;
3229 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
3230 x->source_variance < skip_intra_var_thresh)
3232 // Only search the oblique modes if the best so far is
3233 // one of the neighboring directional modes
3234 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
3235 (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
3236 if (best_mode_index >= 0 &&
3237 best_mbmode.ref_frame[0] > INTRA_FRAME)
3240 if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
3241 if (conditional_skipintra(this_mode, best_intra_mode))
3246 const MV_REFERENCE_FRAME ref_frames[2] = {ref_frame, second_ref_frame};
3247 if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
3248 this_mode, ref_frames))
3252 mbmi->mode = this_mode;
3253 mbmi->uv_mode = DC_PRED;
3254 mbmi->ref_frame[0] = ref_frame;
3255 mbmi->ref_frame[1] = second_ref_frame;
3256 // Evaluate all sub-pel filters irrespective of whether we can use
3257 // them for this frame.
3258 mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
3259 : cm->interp_filter;
3260 mbmi->mv[0].as_int = mbmi->mv[1].as_int = 0;
3263 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3265 // Select prediction reference frames.
3266 for (i = 0; i < MAX_MB_PLANE; i++) {
3267 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3269 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3272 if (ref_frame == INTRA_FRAME) {
3274 struct macroblockd_plane *const pd = &xd->plane[1];
3275 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
3276 super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable,
3277 NULL, bsize, best_rd);
3278 if (rate_y == INT_MAX)
3281 uv_tx = get_uv_tx_size_impl(mbmi->tx_size, bsize, pd->subsampling_x,
3283 if (rate_uv_intra[uv_tx] == INT_MAX) {
3284 choose_intra_uv_mode(cpi, x, ctx, bsize, uv_tx,
3285 &rate_uv_intra[uv_tx], &rate_uv_tokenonly[uv_tx],
3286 &dist_uv[uv_tx], &skip_uv[uv_tx], &mode_uv[uv_tx]);
3289 rate_uv = rate_uv_tokenonly[uv_tx];
3290 distortion_uv = dist_uv[uv_tx];
3291 skippable = skippable && skip_uv[uv_tx];
3292 mbmi->uv_mode = mode_uv[uv_tx];
3294 rate2 = rate_y + cpi->mbmode_cost[mbmi->mode] + rate_uv_intra[uv_tx];
3295 if (this_mode != DC_PRED && this_mode != TM_PRED)
3296 rate2 += intra_cost_penalty;
3297 distortion2 = distortion_y + distortion_uv;
3299 this_rd = handle_inter_mode(cpi, x, bsize,
3300 &rate2, &distortion2, &skippable,
3302 &disable_skip, frame_mv,
3304 single_newmv, single_inter_filter,
3305 single_skippable, &total_sse, best_rd,
3306 &mask_filter, filter_cache);
3307 if (this_rd == INT64_MAX)
3310 compmode_cost = vp10_cost_bit(comp_mode_p, comp_pred);
3312 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3313 rate2 += compmode_cost;
3316 // Estimate the reference frame signaling cost and add it
3317 // to the rolling cost variable.
3319 rate2 += ref_costs_comp[ref_frame];
3321 rate2 += ref_costs_single[ref_frame];
3324 if (!disable_skip) {
3326 // Back out the coefficient coding costs
3327 rate2 -= (rate_y + rate_uv);
3329 // Cost the skip mb case
3330 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 1);
3331 } else if (ref_frame != INTRA_FRAME && !xd->lossless) {
3332 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
3333 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
3334 // Add in the cost of the no skip flag.
3335 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 0);
3337 // FIXME(rbultje) make this work for splitmv also
3338 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 1);
3339 distortion2 = total_sse;
3340 assert(total_sse >= 0);
3341 rate2 -= (rate_y + rate_uv);
3345 // Add in the cost of the no skip flag.
3346 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 0);
3349 // Calculate the final RD estimate for this mode.
3350 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3353 // Apply an adjustment to the rd value based on the similarity of the
3354 // source variance and reconstructed variance.
3355 rd_variance_adjustment(cpi, x, bsize, &this_rd,
3356 ref_frame, x->source_variance);
3358 if (ref_frame == INTRA_FRAME) {
3359 // Keep record of best intra rd
3360 if (this_rd < best_intra_rd) {
3361 best_intra_rd = this_rd;
3362 best_intra_mode = mbmi->mode;
3366 if (!disable_skip && ref_frame == INTRA_FRAME) {
3367 for (i = 0; i < REFERENCE_MODES; ++i)
3368 best_pred_rd[i] = MIN(best_pred_rd[i], this_rd);
3369 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3370 best_filter_rd[i] = MIN(best_filter_rd[i], this_rd);
3373 // Did this mode help.. i.e. is it the new best mode
3374 if (this_rd < best_rd || x->skip) {
3375 int max_plane = MAX_MB_PLANE;
3376 if (!mode_excluded) {
3377 // Note index of best mode so far
3378 best_mode_index = mode_index;
3380 if (ref_frame == INTRA_FRAME) {
3381 /* required for left and above block mv */
3382 mbmi->mv[0].as_int = 0;
3385 best_pred_sse = x->pred_sse[ref_frame];
3388 rd_cost->rate = rate2;
3389 rd_cost->dist = distortion2;
3390 rd_cost->rdcost = this_rd;
3392 best_mbmode = *mbmi;
3393 best_skip2 = this_skip2;
3394 best_mode_skippable = skippable;
3396 if (!x->select_tx_size)
3397 swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
3398 memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size],
3399 sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
3401 // TODO(debargha): enhance this test with a better distortion prediction
3402 // based on qp, activity mask and history
3403 if ((mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
3404 (mode_index > MIN_EARLY_TERM_INDEX)) {
3405 int qstep = xd->plane[0].dequant[1];
3406 // TODO(debargha): Enhance this by specializing for each mode_index
3408 #if CONFIG_VP9_HIGHBITDEPTH
3409 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
3410 qstep >>= (xd->bd - 8);
3412 #endif // CONFIG_VP9_HIGHBITDEPTH
3413 if (x->source_variance < UINT_MAX) {
3414 const int var_adjust = (x->source_variance < 16);
3415 scale -= var_adjust;
3417 if (ref_frame > INTRA_FRAME &&
3418 distortion2 * scale < qstep * qstep) {
3425 /* keep record of best compound/single-only prediction */
3426 if (!disable_skip && ref_frame != INTRA_FRAME) {
3427 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
3429 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3430 single_rate = rate2 - compmode_cost;
3431 hybrid_rate = rate2;
3433 single_rate = rate2;
3434 hybrid_rate = rate2 + compmode_cost;
3437 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
3438 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
3441 if (single_rd < best_pred_rd[SINGLE_REFERENCE])
3442 best_pred_rd[SINGLE_REFERENCE] = single_rd;
3444 if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
3445 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
3447 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
3448 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
3450 /* keep record of best filter type */
3451 if (!mode_excluded && cm->interp_filter != BILINEAR) {
3452 int64_t ref = filter_cache[cm->interp_filter == SWITCHABLE ?
3453 SWITCHABLE_FILTERS : cm->interp_filter];
3455 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3457 if (ref == INT64_MAX)
3459 else if (filter_cache[i] == INT64_MAX)
3460 // when early termination is triggered, the encoder does not have
3461 // access to the rate-distortion cost. it only knows that the cost
3462 // should be above the maximum valid value. hence it takes the known
3463 // maximum plus an arbitrary constant as the rate-distortion cost.
3464 adj_rd = mask_filter - ref + 10;
3466 adj_rd = filter_cache[i] - ref;
3469 best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd);
3477 if (x->skip && !comp_pred)
3481 // The inter modes' rate costs are not calculated precisely in some cases.
3482 // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and
3483 // ZEROMV. Here, checks are added for those cases, and the mode decisions
3485 if (best_mbmode.mode == NEWMV) {
3486 const MV_REFERENCE_FRAME refs[2] = {best_mbmode.ref_frame[0],
3487 best_mbmode.ref_frame[1]};
3488 int comp_pred_mode = refs[1] > INTRA_FRAME;
3490 if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3491 ((comp_pred_mode && frame_mv[NEARESTMV][refs[1]].as_int ==
3492 best_mbmode.mv[1].as_int) || !comp_pred_mode))
3493 best_mbmode.mode = NEARESTMV;
3494 else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3495 ((comp_pred_mode && frame_mv[NEARMV][refs[1]].as_int ==
3496 best_mbmode.mv[1].as_int) || !comp_pred_mode))
3497 best_mbmode.mode = NEARMV;
3498 else if (best_mbmode.mv[0].as_int == 0 &&
3499 ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) || !comp_pred_mode))
3500 best_mbmode.mode = ZEROMV;
3503 if (best_mode_index < 0 || best_rd >= best_rd_so_far) {
3504 rd_cost->rate = INT_MAX;
3505 rd_cost->rdcost = INT64_MAX;
3509 // If we used an estimate for the uv intra rd in the loop above...
3510 if (sf->use_uv_intra_rd_estimate) {
3511 // Do Intra UV best rd mode selection if best mode choice above was intra.
3512 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
3514 *mbmi = best_mbmode;
3515 uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
3516 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
3517 &rate_uv_tokenonly[uv_tx_size],
3518 &dist_uv[uv_tx_size],
3519 &skip_uv[uv_tx_size],
3520 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
3525 assert((cm->interp_filter == SWITCHABLE) ||
3526 (cm->interp_filter == best_mbmode.interp_filter) ||
3527 !is_inter_block(&best_mbmode));
3529 if (!cpi->rc.is_src_frame_alt_ref)
3530 vp10_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3531 sf->adaptive_rd_thresh, bsize, best_mode_index);
3534 *mbmi = best_mbmode;
3535 x->skip |= best_skip2;
3537 for (i = 0; i < REFERENCE_MODES; ++i) {
3538 if (best_pred_rd[i] == INT64_MAX)
3539 best_pred_diff[i] = INT_MIN;
3541 best_pred_diff[i] = best_rd - best_pred_rd[i];
3545 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3546 if (best_filter_rd[i] == INT64_MAX)
3547 best_filter_diff[i] = 0;
3549 best_filter_diff[i] = best_rd - best_filter_rd[i];
3551 if (cm->interp_filter == SWITCHABLE)
3552 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
3554 vp10_zero(best_filter_diff);
3557 // TODO(yunqingwang): Moving this line in front of the above best_filter_diff
3558 // updating code causes PSNR loss. Need to figure out the confliction.
3559 x->skip |= best_mode_skippable;
3561 if (!x->skip && !x->select_tx_size) {
3562 int has_high_freq_coeff = 0;
3564 int max_plane = is_inter_block(&xd->mi[0]->mbmi)
3566 for (plane = 0; plane < max_plane; ++plane) {
3567 x->plane[plane].eobs = ctx->eobs_pbuf[plane][1];
3568 has_high_freq_coeff |= vp10_has_high_freq_in_plane(x, bsize, plane);
3571 for (plane = max_plane; plane < MAX_MB_PLANE; ++plane) {
3572 x->plane[plane].eobs = ctx->eobs_pbuf[plane][2];
3573 has_high_freq_coeff |= vp10_has_high_freq_in_plane(x, bsize, plane);
3576 best_mode_skippable |= !has_high_freq_coeff;
3579 assert(best_mode_index >= 0);
3581 store_coding_context(x, ctx, best_mode_index, best_pred_diff,
3582 best_filter_diff, best_mode_skippable);
3585 void vp10_rd_pick_inter_mode_sb_seg_skip(VP10_COMP *cpi,
3586 TileDataEnc *tile_data,
3590 PICK_MODE_CONTEXT *ctx,
3591 int64_t best_rd_so_far) {
3592 VP10_COMMON *const cm = &cpi->common;
3593 MACROBLOCKD *const xd = &x->e_mbd;
3594 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
3595 unsigned char segment_id = mbmi->segment_id;
3596 const int comp_pred = 0;
3598 int64_t best_pred_diff[REFERENCE_MODES];
3599 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3600 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3601 vpx_prob comp_mode_p;
3602 INTERP_FILTER best_filter = SWITCHABLE;
3603 int64_t this_rd = INT64_MAX;
3605 const int64_t distortion2 = 0;
3607 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3609 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3612 for (i = 0; i < MAX_REF_FRAMES; ++i)
3613 x->pred_sse[i] = INT_MAX;
3614 for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i)
3615 x->pred_mv_sad[i] = INT_MAX;
3617 rd_cost->rate = INT_MAX;
3619 assert(segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
3621 mbmi->mode = ZEROMV;
3622 mbmi->uv_mode = DC_PRED;
3623 mbmi->ref_frame[0] = LAST_FRAME;
3624 mbmi->ref_frame[1] = NONE;
3625 mbmi->mv[0].as_int = 0;
3628 if (cm->interp_filter != BILINEAR) {
3629 best_filter = EIGHTTAP;
3630 if (cm->interp_filter == SWITCHABLE &&
3631 x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
3633 int best_rs = INT_MAX;
3634 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
3635 mbmi->interp_filter = i;
3636 rs = vp10_get_switchable_rate(cpi, xd);
3639 best_filter = mbmi->interp_filter;
3644 // Set the appropriate filter
3645 if (cm->interp_filter == SWITCHABLE) {
3646 mbmi->interp_filter = best_filter;
3647 rate2 += vp10_get_switchable_rate(cpi, xd);
3649 mbmi->interp_filter = cm->interp_filter;
3652 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3653 rate2 += vp10_cost_bit(comp_mode_p, comp_pred);
3655 // Estimate the reference frame signaling cost and add it
3656 // to the rolling cost variable.
3657 rate2 += ref_costs_single[LAST_FRAME];
3658 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3660 rd_cost->rate = rate2;
3661 rd_cost->dist = distortion2;
3662 rd_cost->rdcost = this_rd;
3664 if (this_rd >= best_rd_so_far) {
3665 rd_cost->rate = INT_MAX;
3666 rd_cost->rdcost = INT64_MAX;
3670 assert((cm->interp_filter == SWITCHABLE) ||
3671 (cm->interp_filter == mbmi->interp_filter));
3673 vp10_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3674 cpi->sf.adaptive_rd_thresh, bsize, THR_ZEROMV);
3676 vp10_zero(best_pred_diff);
3677 vp10_zero(best_filter_diff);
3679 if (!x->select_tx_size)
3680 swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE);
3681 store_coding_context(x, ctx, THR_ZEROMV,
3682 best_pred_diff, best_filter_diff, 0);
3685 void vp10_rd_pick_inter_mode_sub8x8(VP10_COMP *cpi,
3686 TileDataEnc *tile_data,
3688 int mi_row, int mi_col,
3691 PICK_MODE_CONTEXT *ctx,
3692 int64_t best_rd_so_far) {
3693 VP10_COMMON *const cm = &cpi->common;
3694 RD_OPT *const rd_opt = &cpi->rd;
3695 SPEED_FEATURES *const sf = &cpi->sf;
3696 MACROBLOCKD *const xd = &x->e_mbd;
3697 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
3698 const struct segmentation *const seg = &cm->seg;
3699 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3700 unsigned char segment_id = mbmi->segment_id;
3702 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3703 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3704 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
3706 int64_t best_rd = best_rd_so_far;
3707 int64_t best_yrd = best_rd_so_far; // FIXME(rbultje) more precise
3708 int64_t best_pred_diff[REFERENCE_MODES];
3709 int64_t best_pred_rd[REFERENCE_MODES];
3710 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3711 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3712 MB_MODE_INFO best_mbmode;
3713 int ref_index, best_ref_index = 0;
3714 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3715 vpx_prob comp_mode_p;
3716 INTERP_FILTER tmp_best_filter = SWITCHABLE;
3717 int rate_uv_intra, rate_uv_tokenonly;
3720 PREDICTION_MODE mode_uv = DC_PRED;
3721 const int intra_cost_penalty = vp10_get_intra_cost_penalty(
3722 cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
3723 int_mv seg_mvs[4][MAX_REF_FRAMES];
3724 b_mode_info best_bmodes[4];
3726 int ref_frame_skip_mask[2] = { 0 };
3727 int64_t mask_filter = 0;
3728 int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
3729 int internal_active_edge =
3730 vp10_active_edge_sb(cpi, mi_row, mi_col) && vp10_internal_image_edge(cpi);
3732 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3733 memset(x->zcoeff_blk[TX_4X4], 0, 4);
3734 vp10_zero(best_mbmode);
3736 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3737 filter_cache[i] = INT64_MAX;
3739 for (i = 0; i < 4; i++) {
3741 for (j = 0; j < MAX_REF_FRAMES; j++)
3742 seg_mvs[i][j].as_int = INVALID_MV;
3745 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3748 for (i = 0; i < REFERENCE_MODES; ++i)
3749 best_pred_rd[i] = INT64_MAX;
3750 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3751 best_filter_rd[i] = INT64_MAX;
3752 rate_uv_intra = INT_MAX;
3754 rd_cost->rate = INT_MAX;
3756 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
3757 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3758 setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
3759 frame_mv[NEARESTMV], frame_mv[NEARMV],
3762 ref_frame_skip_mask[0] |= (1 << ref_frame);
3763 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3765 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3766 frame_mv[ZEROMV][ref_frame].as_int = 0;
3769 for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
3770 int mode_excluded = 0;
3771 int64_t this_rd = INT64_MAX;
3772 int disable_skip = 0;
3773 int compmode_cost = 0;
3774 int rate2 = 0, rate_y = 0, rate_uv = 0;
3775 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3779 int64_t total_sse = INT_MAX;
3782 ref_frame = vp10_ref_order[ref_index].ref_frame[0];
3783 second_ref_frame = vp10_ref_order[ref_index].ref_frame[1];
3785 // Look at the reference frame of the best mode so far and set the
3786 // skip mask to look at a subset of the remaining modes.
3787 if (ref_index > 2 && sf->mode_skip_start < MAX_MODES) {
3788 if (ref_index == 3) {
3789 switch (best_mbmode.ref_frame[0]) {
3793 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME);
3794 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3797 ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME);
3798 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3801 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME);
3804 case MAX_REF_FRAMES:
3805 assert(0 && "Invalid Reference frame");
3811 if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3812 (ref_frame_skip_mask[1] & (1 << MAX(0, second_ref_frame))))
3815 // Test best rd so far against threshold for trying this mode.
3816 if (!internal_active_edge &&
3817 rd_less_than_thresh(best_rd,
3818 rd_opt->threshes[segment_id][bsize][ref_index],
3819 tile_data->thresh_freq_fact[bsize][ref_index]))
3822 comp_pred = second_ref_frame > INTRA_FRAME;
3824 if (!cpi->allow_comp_inter_inter)
3826 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
3828 // Do not allow compound prediction if the segment level reference frame
3829 // feature is in use as in this case there can only be one reference.
3830 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
3833 if ((sf->mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3834 best_mbmode.ref_frame[0] == INTRA_FRAME)
3838 // TODO(jingning, jkoleszar): scaling reference frame not supported for
3840 if (ref_frame > INTRA_FRAME &&
3841 vp10_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
3844 if (second_ref_frame > INTRA_FRAME &&
3845 vp10_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf))
3849 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3850 else if (ref_frame != INTRA_FRAME)
3851 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3853 // If the segment reference frame feature is enabled....
3854 // then do nothing if the current ref frame is not allowed..
3855 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3856 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3858 // Disable this drop out case if the ref frame
3859 // segment level feature is enabled for this segment. This is to
3860 // prevent the possibility that we end up unable to pick any mode.
3861 } else if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3862 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3863 // unless ARNR filtering is enabled in which case we want
3864 // an unfiltered alternative. We allow near/nearest as well
3865 // because they may result in zero-zero MVs but be cheaper.
3866 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
3870 mbmi->tx_size = TX_4X4;
3871 mbmi->uv_mode = DC_PRED;
3872 mbmi->ref_frame[0] = ref_frame;
3873 mbmi->ref_frame[1] = second_ref_frame;
3874 // Evaluate all sub-pel filters irrespective of whether we can use
3875 // them for this frame.
3876 mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
3877 : cm->interp_filter;
3879 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3881 // Select prediction reference frames.
3882 for (i = 0; i < MAX_MB_PLANE; i++) {
3883 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3885 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3888 if (ref_frame == INTRA_FRAME) {
3890 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y,
3891 &distortion_y, best_rd) >= best_rd)
3894 rate2 += intra_cost_penalty;
3895 distortion2 += distortion_y;
3897 if (rate_uv_intra == INT_MAX) {
3898 choose_intra_uv_mode(cpi, x, ctx, bsize, TX_4X4,
3904 rate2 += rate_uv_intra;
3905 rate_uv = rate_uv_tokenonly;
3906 distortion2 += dist_uv;
3907 distortion_uv = dist_uv;
3908 mbmi->uv_mode = mode_uv;
3912 int64_t this_rd_thresh;
3913 int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
3914 int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
3915 int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
3916 int tmp_best_skippable = 0;
3917 int switchable_filter_index;
3918 int_mv *second_ref = comp_pred ?
3919 &x->mbmi_ext->ref_mvs[second_ref_frame][0] : NULL;
3920 b_mode_info tmp_best_bmodes[16];
3921 MB_MODE_INFO tmp_best_mbmode;
3922 BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
3923 int pred_exists = 0;
3926 this_rd_thresh = (ref_frame == LAST_FRAME) ?
3927 rd_opt->threshes[segment_id][bsize][THR_LAST] :
3928 rd_opt->threshes[segment_id][bsize][THR_ALTR];
3929 this_rd_thresh = (ref_frame == GOLDEN_FRAME) ?
3930 rd_opt->threshes[segment_id][bsize][THR_GOLD] : this_rd_thresh;
3931 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3932 filter_cache[i] = INT64_MAX;
3934 if (cm->interp_filter != BILINEAR) {
3935 tmp_best_filter = EIGHTTAP;
3936 if (x->source_variance < sf->disable_filter_search_var_thresh) {
3937 tmp_best_filter = EIGHTTAP;
3938 } else if (sf->adaptive_pred_interp_filter == 1 &&
3939 ctx->pred_interp_filter < SWITCHABLE) {
3940 tmp_best_filter = ctx->pred_interp_filter;
3941 } else if (sf->adaptive_pred_interp_filter == 2) {
3942 tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE ?
3943 ctx->pred_interp_filter : 0;
3945 for (switchable_filter_index = 0;
3946 switchable_filter_index < SWITCHABLE_FILTERS;
3947 ++switchable_filter_index) {
3950 MB_MODE_INFO_EXT *mbmi_ext = x->mbmi_ext;
3951 mbmi->interp_filter = switchable_filter_index;
3952 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x,
3953 &mbmi_ext->ref_mvs[ref_frame][0],
3954 second_ref, best_yrd, &rate,
3955 &rate_y, &distortion,
3956 &skippable, &total_sse,
3957 (int) this_rd_thresh, seg_mvs,
3958 bsi, switchable_filter_index,
3961 if (tmp_rd == INT64_MAX)
3963 rs = vp10_get_switchable_rate(cpi, xd);
3964 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
3965 filter_cache[switchable_filter_index] = tmp_rd;
3966 filter_cache[SWITCHABLE_FILTERS] =
3967 MIN(filter_cache[SWITCHABLE_FILTERS],
3969 if (cm->interp_filter == SWITCHABLE)
3972 mask_filter = MAX(mask_filter, tmp_rd);
3974 newbest = (tmp_rd < tmp_best_rd);
3976 tmp_best_filter = mbmi->interp_filter;
3977 tmp_best_rd = tmp_rd;
3979 if ((newbest && cm->interp_filter == SWITCHABLE) ||
3980 (mbmi->interp_filter == cm->interp_filter &&
3981 cm->interp_filter != SWITCHABLE)) {
3982 tmp_best_rdu = tmp_rd;
3983 tmp_best_rate = rate;
3984 tmp_best_ratey = rate_y;
3985 tmp_best_distortion = distortion;
3986 tmp_best_sse = total_sse;
3987 tmp_best_skippable = skippable;
3988 tmp_best_mbmode = *mbmi;
3989 for (i = 0; i < 4; i++) {
3990 tmp_best_bmodes[i] = xd->mi[0]->bmi[i];
3991 x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
3994 if (switchable_filter_index == 0 &&
3995 sf->use_rd_breakout &&
3996 best_rd < INT64_MAX) {
3997 if (tmp_best_rdu / 2 > best_rd) {
3998 // skip searching the other filters if the first is
3999 // already substantially larger than the best so far
4000 tmp_best_filter = mbmi->interp_filter;
4001 tmp_best_rdu = INT64_MAX;
4006 } // switchable_filter_index loop
4010 if (tmp_best_rdu == INT64_MAX && pred_exists)
4013 mbmi->interp_filter = (cm->interp_filter == SWITCHABLE ?
4014 tmp_best_filter : cm->interp_filter);
4016 // Handles the special case when a filter that is not in the
4017 // switchable list (bilinear, 6-tap) is indicated at the frame level
4018 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x,
4019 &x->mbmi_ext->ref_mvs[ref_frame][0],
4020 second_ref, best_yrd, &rate, &rate_y,
4021 &distortion, &skippable, &total_sse,
4022 (int) this_rd_thresh, seg_mvs, bsi, 0,
4024 if (tmp_rd == INT64_MAX)
4027 total_sse = tmp_best_sse;
4028 rate = tmp_best_rate;
4029 rate_y = tmp_best_ratey;
4030 distortion = tmp_best_distortion;
4031 skippable = tmp_best_skippable;
4032 *mbmi = tmp_best_mbmode;
4033 for (i = 0; i < 4; i++)
4034 xd->mi[0]->bmi[i] = tmp_best_bmodes[i];
4038 distortion2 += distortion;
4040 if (cm->interp_filter == SWITCHABLE)
4041 rate2 += vp10_get_switchable_rate(cpi, xd);
4044 mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
4045 : cm->reference_mode == COMPOUND_REFERENCE;
4047 compmode_cost = vp10_cost_bit(comp_mode_p, comp_pred);
4049 tmp_best_rdu = best_rd -
4050 MIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
4051 RDCOST(x->rdmult, x->rddiv, 0, total_sse));
4053 if (tmp_best_rdu > 0) {
4054 // If even the 'Y' rd value of split is higher than best so far
4055 // then dont bother looking at UV
4056 vp10_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col,
4058 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
4059 if (!super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
4060 &uv_sse, BLOCK_8X8, tmp_best_rdu))
4064 distortion2 += distortion_uv;
4065 skippable = skippable && uv_skippable;
4066 total_sse += uv_sse;
4070 if (cm->reference_mode == REFERENCE_MODE_SELECT)
4071 rate2 += compmode_cost;
4073 // Estimate the reference frame signaling cost and add it
4074 // to the rolling cost variable.
4075 if (second_ref_frame > INTRA_FRAME) {
4076 rate2 += ref_costs_comp[ref_frame];
4078 rate2 += ref_costs_single[ref_frame];
4081 if (!disable_skip) {
4082 // Skip is never coded at the segment level for sub8x8 blocks and instead
4083 // always coded in the bitstream at the mode info level.
4085 if (ref_frame != INTRA_FRAME && !xd->lossless) {
4086 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
4087 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
4088 // Add in the cost of the no skip flag.
4089 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 0);
4091 // FIXME(rbultje) make this work for splitmv also
4092 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 1);
4093 distortion2 = total_sse;
4094 assert(total_sse >= 0);
4095 rate2 -= (rate_y + rate_uv);
4101 // Add in the cost of the no skip flag.
4102 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 0);
4105 // Calculate the final RD estimate for this mode.
4106 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
4109 if (!disable_skip && ref_frame == INTRA_FRAME) {
4110 for (i = 0; i < REFERENCE_MODES; ++i)
4111 best_pred_rd[i] = MIN(best_pred_rd[i], this_rd);
4112 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
4113 best_filter_rd[i] = MIN(best_filter_rd[i], this_rd);
4116 // Did this mode help.. i.e. is it the new best mode
4117 if (this_rd < best_rd || x->skip) {
4118 if (!mode_excluded) {
4119 int max_plane = MAX_MB_PLANE;
4120 // Note index of best mode so far
4121 best_ref_index = ref_index;
4123 if (ref_frame == INTRA_FRAME) {
4124 /* required for left and above block mv */
4125 mbmi->mv[0].as_int = 0;
4129 rd_cost->rate = rate2;
4130 rd_cost->dist = distortion2;
4131 rd_cost->rdcost = this_rd;
4133 best_yrd = best_rd -
4134 RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
4135 best_mbmode = *mbmi;
4136 best_skip2 = this_skip2;
4137 if (!x->select_tx_size)
4138 swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
4139 memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
4140 sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
4142 for (i = 0; i < 4; i++)
4143 best_bmodes[i] = xd->mi[0]->bmi[i];
4145 // TODO(debargha): enhance this test with a better distortion prediction
4146 // based on qp, activity mask and history
4147 if ((sf->mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
4148 (ref_index > MIN_EARLY_TERM_INDEX)) {
4149 int qstep = xd->plane[0].dequant[1];
4150 // TODO(debargha): Enhance this by specializing for each mode_index
4152 #if CONFIG_VP9_HIGHBITDEPTH
4153 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
4154 qstep >>= (xd->bd - 8);
4156 #endif // CONFIG_VP9_HIGHBITDEPTH
4157 if (x->source_variance < UINT_MAX) {
4158 const int var_adjust = (x->source_variance < 16);
4159 scale -= var_adjust;
4161 if (ref_frame > INTRA_FRAME &&
4162 distortion2 * scale < qstep * qstep) {
4169 /* keep record of best compound/single-only prediction */
4170 if (!disable_skip && ref_frame != INTRA_FRAME) {
4171 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
4173 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4174 single_rate = rate2 - compmode_cost;
4175 hybrid_rate = rate2;
4177 single_rate = rate2;
4178 hybrid_rate = rate2 + compmode_cost;
4181 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
4182 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
4184 if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
4185 best_pred_rd[SINGLE_REFERENCE] = single_rd;
4186 else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
4187 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
4189 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
4190 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
4193 /* keep record of best filter type */
4194 if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
4195 cm->interp_filter != BILINEAR) {
4196 int64_t ref = filter_cache[cm->interp_filter == SWITCHABLE ?
4197 SWITCHABLE_FILTERS : cm->interp_filter];
4199 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4200 if (ref == INT64_MAX)
4202 else if (filter_cache[i] == INT64_MAX)
4203 // when early termination is triggered, the encoder does not have
4204 // access to the rate-distortion cost. it only knows that the cost
4205 // should be above the maximum valid value. hence it takes the known
4206 // maximum plus an arbitrary constant as the rate-distortion cost.
4207 adj_rd = mask_filter - ref + 10;
4209 adj_rd = filter_cache[i] - ref;
4212 best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd);
4219 if (x->skip && !comp_pred)
4223 if (best_rd >= best_rd_so_far) {
4224 rd_cost->rate = INT_MAX;
4225 rd_cost->rdcost = INT64_MAX;
4229 // If we used an estimate for the uv intra rd in the loop above...
4230 if (sf->use_uv_intra_rd_estimate) {
4231 // Do Intra UV best rd mode selection if best mode choice above was intra.
4232 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
4233 *mbmi = best_mbmode;
4234 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra,
4242 if (best_rd == INT64_MAX) {
4243 rd_cost->rate = INT_MAX;
4244 rd_cost->dist = INT64_MAX;
4245 rd_cost->rdcost = INT64_MAX;
4249 assert((cm->interp_filter == SWITCHABLE) ||
4250 (cm->interp_filter == best_mbmode.interp_filter) ||
4251 !is_inter_block(&best_mbmode));
4253 vp10_update_rd_thresh_fact(tile_data->thresh_freq_fact,
4254 sf->adaptive_rd_thresh, bsize, best_ref_index);
4257 *mbmi = best_mbmode;
4258 x->skip |= best_skip2;
4259 if (!is_inter_block(&best_mbmode)) {
4260 for (i = 0; i < 4; i++)
4261 xd->mi[0]->bmi[i].as_mode = best_bmodes[i].as_mode;
4263 for (i = 0; i < 4; ++i)
4264 memcpy(&xd->mi[0]->bmi[i], &best_bmodes[i], sizeof(b_mode_info));
4266 mbmi->mv[0].as_int = xd->mi[0]->bmi[3].as_mv[0].as_int;
4267 mbmi->mv[1].as_int = xd->mi[0]->bmi[3].as_mv[1].as_int;
4270 for (i = 0; i < REFERENCE_MODES; ++i) {
4271 if (best_pred_rd[i] == INT64_MAX)
4272 best_pred_diff[i] = INT_MIN;
4274 best_pred_diff[i] = best_rd - best_pred_rd[i];
4278 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4279 if (best_filter_rd[i] == INT64_MAX)
4280 best_filter_diff[i] = 0;
4282 best_filter_diff[i] = best_rd - best_filter_rd[i];
4284 if (cm->interp_filter == SWITCHABLE)
4285 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
4287 vp10_zero(best_filter_diff);
4290 store_coding_context(x, ctx, best_ref_index,
4291 best_pred_diff, best_filter_diff, 0);