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_dsp/vpx_dsp_common.h"
18 #include "vpx_mem/vpx_mem.h"
19 #include "vpx_ports/mem.h"
20 #include "vpx_ports/system_state.h"
22 #include "vp10/common/common.h"
23 #include "vp10/common/entropy.h"
24 #include "vp10/common/entropymode.h"
25 #include "vp10/common/idct.h"
26 #include "vp10/common/mvref_common.h"
27 #include "vp10/common/pred_common.h"
28 #include "vp10/common/quant_common.h"
29 #include "vp10/common/reconinter.h"
30 #include "vp10/common/reconintra.h"
31 #include "vp10/common/scan.h"
32 #include "vp10/common/seg_common.h"
34 #include "vp10/encoder/cost.h"
35 #include "vp10/encoder/encodemb.h"
36 #include "vp10/encoder/encodemv.h"
37 #include "vp10/encoder/encoder.h"
38 #include "vp10/encoder/mcomp.h"
39 #include "vp10/encoder/quantize.h"
40 #include "vp10/encoder/ratectrl.h"
41 #include "vp10/encoder/rd.h"
42 #include "vp10/encoder/rdopt.h"
43 #include "vp10/encoder/aq_variance.h"
45 #define LAST_FRAME_MODE_MASK ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | \
47 #define GOLDEN_FRAME_MODE_MASK ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | \
49 #define ALT_REF_MODE_MASK ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | \
52 #define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
54 #define MIN_EARLY_TERM_INDEX 3
55 #define NEW_MV_DISCOUNT_FACTOR 8
59 MV_REFERENCE_FRAME ref_frame[2];
63 MV_REFERENCE_FRAME ref_frame[2];
66 struct rdcost_block_args {
68 ENTROPY_CONTEXT t_above[16];
69 ENTROPY_CONTEXT t_left[16];
76 int use_fast_coef_costing;
81 #define LAST_NEW_MV_INDEX 6
82 static const MODE_DEFINITION vp10_mode_order[MAX_MODES] = {
83 {NEARESTMV, {LAST_FRAME, NONE}},
84 {NEARESTMV, {ALTREF_FRAME, NONE}},
85 {NEARESTMV, {GOLDEN_FRAME, NONE}},
87 {DC_PRED, {INTRA_FRAME, NONE}},
89 {NEWMV, {LAST_FRAME, NONE}},
90 {NEWMV, {ALTREF_FRAME, NONE}},
91 {NEWMV, {GOLDEN_FRAME, NONE}},
93 {NEARMV, {LAST_FRAME, NONE}},
94 {NEARMV, {ALTREF_FRAME, NONE}},
95 {NEARMV, {GOLDEN_FRAME, NONE}},
97 {ZEROMV, {LAST_FRAME, NONE}},
98 {ZEROMV, {GOLDEN_FRAME, NONE}},
99 {ZEROMV, {ALTREF_FRAME, NONE}},
101 {NEARESTMV, {LAST_FRAME, ALTREF_FRAME}},
102 {NEARESTMV, {GOLDEN_FRAME, ALTREF_FRAME}},
104 {TM_PRED, {INTRA_FRAME, NONE}},
106 {NEARMV, {LAST_FRAME, ALTREF_FRAME}},
107 {NEWMV, {LAST_FRAME, ALTREF_FRAME}},
108 {NEARMV, {GOLDEN_FRAME, ALTREF_FRAME}},
109 {NEWMV, {GOLDEN_FRAME, ALTREF_FRAME}},
111 {ZEROMV, {LAST_FRAME, ALTREF_FRAME}},
112 {ZEROMV, {GOLDEN_FRAME, ALTREF_FRAME}},
114 {H_PRED, {INTRA_FRAME, NONE}},
115 {V_PRED, {INTRA_FRAME, NONE}},
116 {D135_PRED, {INTRA_FRAME, NONE}},
117 {D207_PRED, {INTRA_FRAME, NONE}},
118 {D153_PRED, {INTRA_FRAME, NONE}},
119 {D63_PRED, {INTRA_FRAME, NONE}},
120 {D117_PRED, {INTRA_FRAME, NONE}},
121 {D45_PRED, {INTRA_FRAME, NONE}},
124 static const REF_DEFINITION vp10_ref_order[MAX_REFS] = {
125 {{LAST_FRAME, NONE}},
126 {{GOLDEN_FRAME, NONE}},
127 {{ALTREF_FRAME, NONE}},
128 {{LAST_FRAME, ALTREF_FRAME}},
129 {{GOLDEN_FRAME, ALTREF_FRAME}},
130 {{INTRA_FRAME, NONE}},
133 static INLINE int write_uniform_cost(int n, int v) {
134 int l = get_unsigned_bits(n), m = (1 << l) - n;
138 return (l - 1) * vp10_cost_bit(128, 0);
140 return l * vp10_cost_bit(128, 0);
143 static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
144 int m, int n, int min_plane, int max_plane) {
147 for (i = min_plane; i < max_plane; ++i) {
148 struct macroblock_plane *const p = &x->plane[i];
149 struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
151 p->coeff = ctx->coeff_pbuf[i][m];
152 p->qcoeff = ctx->qcoeff_pbuf[i][m];
153 pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
154 p->eobs = ctx->eobs_pbuf[i][m];
156 ctx->coeff_pbuf[i][m] = ctx->coeff_pbuf[i][n];
157 ctx->qcoeff_pbuf[i][m] = ctx->qcoeff_pbuf[i][n];
158 ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
159 ctx->eobs_pbuf[i][m] = ctx->eobs_pbuf[i][n];
161 ctx->coeff_pbuf[i][n] = p->coeff;
162 ctx->qcoeff_pbuf[i][n] = p->qcoeff;
163 ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
164 ctx->eobs_pbuf[i][n] = p->eobs;
168 static void model_rd_for_sb(VP10_COMP *cpi, BLOCK_SIZE bsize,
169 MACROBLOCK *x, MACROBLOCKD *xd,
170 int *out_rate_sum, int64_t *out_dist_sum,
171 int *skip_txfm_sb, int64_t *skip_sse_sb) {
172 // Note our transform coeffs are 8 times an orthogonal transform.
173 // Hence quantizer step is also 8 times. To get effective quantizer
174 // we need to divide by 8 before sending to modeling function.
176 int64_t rate_sum = 0;
177 int64_t dist_sum = 0;
178 const int ref = xd->mi[0]->mbmi.ref_frame[0];
180 unsigned int var = 0;
181 unsigned int sum_sse = 0;
182 int64_t total_sse = 0;
187 const int dequant_shift =
188 #if CONFIG_VP9_HIGHBITDEPTH
189 (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ?
191 #endif // CONFIG_VP9_HIGHBITDEPTH
194 x->pred_sse[ref] = 0;
196 for (i = 0; i < MAX_MB_PLANE; ++i) {
197 struct macroblock_plane *const p = &x->plane[i];
198 struct macroblockd_plane *const pd = &xd->plane[i];
199 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
200 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
201 const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
202 const int64_t dc_thr = p->quant_thred[0] >> shift;
203 const int64_t ac_thr = p->quant_thred[1] >> shift;
204 // The low thresholds are used to measure if the prediction errors are
205 // low enough so that we can skip the mode search.
206 const int64_t low_dc_thr = VPXMIN(50, dc_thr >> 2);
207 const int64_t low_ac_thr = VPXMIN(80, ac_thr >> 2);
208 int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
209 int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
211 int lw = b_width_log2_lookup[unit_size] + 2;
212 int lh = b_height_log2_lookup[unit_size] + 2;
216 for (idy = 0; idy < bh; ++idy) {
217 for (idx = 0; idx < bw; ++idx) {
218 uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
219 uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
220 int block_idx = (idy << 1) + idx;
221 int low_err_skip = 0;
223 var = cpi->fn_ptr[unit_size].vf(src, p->src.stride,
224 dst, pd->dst.stride, &sse);
225 x->bsse[(i << 2) + block_idx] = sse;
228 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_NONE;
229 if (!x->select_tx_size) {
230 // Check if all ac coefficients can be quantized to zero.
231 if (var < ac_thr || var == 0) {
232 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_ONLY;
234 // Check if dc coefficient can be quantized to zero.
235 if (sse - var < dc_thr || sse == var) {
236 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_DC;
238 if (!sse || (var < low_ac_thr && sse - var < low_dc_thr))
244 if (skip_flag && !low_err_skip)
248 x->pred_sse[ref] += sse;
252 total_sse += sum_sse;
254 // Fast approximate the modelling function.
255 if (cpi->sf.simple_model_rd_from_var) {
257 const int64_t square_error = sum_sse;
258 int quantizer = (pd->dequant[1] >> dequant_shift);
261 rate = (square_error * (280 - quantizer)) >> 8;
264 dist = (square_error * quantizer) >> 8;
268 vp10_model_rd_from_var_lapndz(sum_sse, num_pels_log2_lookup[bs],
269 pd->dequant[1] >> dequant_shift,
276 *skip_txfm_sb = skip_flag;
277 *skip_sse_sb = total_sse << 4;
278 *out_rate_sum = (int)rate_sum;
279 *out_dist_sum = dist_sum << 4;
282 int64_t vp10_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
283 intptr_t block_size, int64_t *ssz) {
285 int64_t error = 0, sqcoeff = 0;
287 for (i = 0; i < block_size; i++) {
288 const int diff = coeff[i] - dqcoeff[i];
289 error += diff * diff;
290 sqcoeff += coeff[i] * coeff[i];
297 int64_t vp10_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff,
302 for (i = 0; i < block_size; i++) {
303 const int diff = coeff[i] - dqcoeff[i];
304 error += diff * diff;
310 #if CONFIG_VP9_HIGHBITDEPTH
311 int64_t vp10_highbd_block_error_c(const tran_low_t *coeff,
312 const tran_low_t *dqcoeff,
314 int64_t *ssz, int bd) {
316 int64_t error = 0, sqcoeff = 0;
317 int shift = 2 * (bd - 8);
318 int rounding = shift > 0 ? 1 << (shift - 1) : 0;
320 for (i = 0; i < block_size; i++) {
321 const int64_t diff = coeff[i] - dqcoeff[i];
322 error += diff * diff;
323 sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
325 assert(error >= 0 && sqcoeff >= 0);
326 error = (error + rounding) >> shift;
327 sqcoeff = (sqcoeff + rounding) >> shift;
332 #endif // CONFIG_VP9_HIGHBITDEPTH
334 /* The trailing '0' is a terminator which is used inside cost_coeffs() to
335 * decide whether to include cost of a trailing EOB node or not (i.e. we
336 * can skip this if the last coefficient in this transform block, e.g. the
337 * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
339 static const int16_t band_counts[TX_SIZES][8] = {
340 { 1, 2, 3, 4, 3, 16 - 13, 0 },
341 { 1, 2, 3, 4, 11, 64 - 21, 0 },
342 { 1, 2, 3, 4, 11, 256 - 21, 0 },
343 { 1, 2, 3, 4, 11, 1024 - 21, 0 },
345 static int cost_coeffs(MACROBLOCK *x,
346 int plane, int block,
347 ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L,
349 const int16_t *scan, const int16_t *nb,
350 int use_fast_coef_costing) {
351 MACROBLOCKD *const xd = &x->e_mbd;
352 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
353 const struct macroblock_plane *p = &x->plane[plane];
354 const struct macroblockd_plane *pd = &xd->plane[plane];
355 const PLANE_TYPE type = pd->plane_type;
356 const int16_t *band_count = &band_counts[tx_size][1];
357 const int eob = p->eobs[block];
358 const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
359 unsigned int (*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
360 x->token_costs[tx_size][type][is_inter_block(mbmi)];
361 uint8_t token_cache[32 * 32];
362 int pt = combine_entropy_contexts(*A, *L);
364 #if CONFIG_VP9_HIGHBITDEPTH
365 const int16_t *cat6_high_cost = vp10_get_high_cost_table(xd->bd);
367 const int16_t *cat6_high_cost = vp10_get_high_cost_table(8);
370 // Check for consistency of tx_size with mode info
371 assert(type == PLANE_TYPE_Y ? mbmi->tx_size == tx_size
372 : get_uv_tx_size(mbmi, pd) == tx_size);
376 cost = token_costs[0][0][pt][EOB_TOKEN];
379 int band_left = *band_count++;
385 vp10_get_token_extra(v, &prev_t, &e);
386 cost = (*token_costs)[0][pt][prev_t] +
387 vp10_get_cost(prev_t, e, cat6_high_cost);
389 token_cache[0] = vp10_pt_energy_class[prev_t];
393 for (c = 1; c < eob; c++) {
394 const int rc = scan[c];
398 vp10_get_token_extra(v, &t, &e);
399 if (use_fast_coef_costing) {
400 cost += (*token_costs)[!prev_t][!prev_t][t] +
401 vp10_get_cost(t, e, cat6_high_cost);
403 pt = get_coef_context(nb, token_cache, c);
404 cost += (*token_costs)[!prev_t][pt][t] +
405 vp10_get_cost(t, e, cat6_high_cost);
406 token_cache[rc] = vp10_pt_energy_class[t];
410 band_left = *band_count++;
417 if (use_fast_coef_costing) {
418 cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
420 pt = get_coef_context(nb, token_cache, c);
421 cost += (*token_costs)[0][pt][EOB_TOKEN];
426 // is eob first coefficient;
432 static void dist_block(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
433 int64_t *out_dist, int64_t *out_sse) {
434 const int ss_txfrm_size = tx_size << 1;
435 MACROBLOCKD* const xd = &x->e_mbd;
436 const struct macroblock_plane *const p = &x->plane[plane];
437 const struct macroblockd_plane *const pd = &xd->plane[plane];
439 int shift = tx_size == TX_32X32 ? 0 : 2;
440 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
441 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
442 #if CONFIG_VP9_HIGHBITDEPTH
443 const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
444 *out_dist = vp10_highbd_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
445 &this_sse, bd) >> shift;
447 *out_dist = vp10_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
449 #endif // CONFIG_VP9_HIGHBITDEPTH
450 *out_sse = this_sse >> shift;
453 static int rate_block(int plane, int block, int blk_row, int blk_col,
454 TX_SIZE tx_size, struct rdcost_block_args* args) {
455 return cost_coeffs(args->x, plane, block, args->t_above + blk_col,
456 args->t_left + blk_row, tx_size,
457 args->so->scan, args->so->neighbors,
458 args->use_fast_coef_costing);
461 static void block_rd_txfm(int plane, int block, int blk_row, int blk_col,
462 BLOCK_SIZE plane_bsize,
463 TX_SIZE tx_size, void *arg) {
464 struct rdcost_block_args *args = arg;
465 MACROBLOCK *const x = args->x;
466 MACROBLOCKD *const xd = &x->e_mbd;
467 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
468 int64_t rd1, rd2, rd;
473 if (args->exit_early)
476 if (!is_inter_block(mbmi)) {
477 struct encode_b_args arg = {x, NULL, &mbmi->skip};
478 vp10_encode_block_intra(plane, block, blk_row, blk_col,
479 plane_bsize, tx_size, &arg);
480 dist_block(x, plane, block, tx_size, &dist, &sse);
481 } else if (max_txsize_lookup[plane_bsize] == tx_size) {
482 if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
484 // full forward transform and quantization
485 vp10_xform_quant(x, plane, block, blk_row, blk_col,
486 plane_bsize, tx_size);
487 dist_block(x, plane, block, tx_size, &dist, &sse);
488 } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
490 // compute DC coefficient
491 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[plane].coeff, block);
492 tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
493 vp10_xform_quant_dc(x, plane, block, blk_row, blk_col,
494 plane_bsize, tx_size);
495 sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
497 if (x->plane[plane].eobs[block]) {
498 const int64_t orig_sse = (int64_t)coeff[0] * coeff[0];
499 const int64_t resd_sse = coeff[0] - dqcoeff[0];
500 int64_t dc_correct = orig_sse - resd_sse * resd_sse;
501 #if CONFIG_VP9_HIGHBITDEPTH
502 dc_correct >>= ((xd->bd - 8) * 2);
504 if (tx_size != TX_32X32)
507 dist = VPXMAX(0, sse - dc_correct);
511 // skip forward transform
512 x->plane[plane].eobs[block] = 0;
513 sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
517 // full forward transform and quantization
518 vp10_xform_quant(x, plane, block, blk_row, blk_col, plane_bsize, tx_size);
519 dist_block(x, plane, block, tx_size, &dist, &sse);
522 rd = RDCOST(x->rdmult, x->rddiv, 0, dist);
523 if (args->this_rd + rd > args->best_rd) {
524 args->exit_early = 1;
528 rate = rate_block(plane, block, blk_row, blk_col, tx_size, args);
529 rd1 = RDCOST(x->rdmult, x->rddiv, rate, dist);
530 rd2 = RDCOST(x->rdmult, x->rddiv, 0, sse);
532 // TODO(jingning): temporarily enabled only for luma component
533 rd = VPXMIN(rd1, rd2);
535 x->zcoeff_blk[tx_size][block] = !x->plane[plane].eobs[block] ||
536 (rd1 > rd2 && !xd->lossless[mbmi->segment_id]);
538 args->this_rate += rate;
539 args->this_dist += dist;
540 args->this_sse += sse;
543 if (args->this_rd > args->best_rd) {
544 args->exit_early = 1;
548 args->skippable &= !x->plane[plane].eobs[block];
551 static void txfm_rd_in_plane(MACROBLOCK *x,
552 int *rate, int64_t *distortion,
553 int *skippable, int64_t *sse,
554 int64_t ref_best_rd, int plane,
555 BLOCK_SIZE bsize, TX_SIZE tx_size,
556 int use_fast_coef_casting) {
557 MACROBLOCKD *const xd = &x->e_mbd;
558 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 tx_type = get_tx_type(pd->plane_type, xd, 0);
573 args.so = get_scan(tx_size, tx_type);
575 vp10_foreach_transformed_block_in_plane(xd, bsize, plane,
576 block_rd_txfm, &args);
577 if (args.exit_early) {
579 *distortion = INT64_MAX;
583 *distortion = args.this_dist;
584 *rate = args.this_rate;
585 *sse = args.this_sse;
586 *skippable = args.skippable;
590 static void choose_largest_tx_size(VP10_COMP *cpi, MACROBLOCK *x,
591 int *rate, int64_t *distortion,
592 int *skip, int64_t *sse,
595 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
596 VP10_COMMON *const cm = &cpi->common;
597 const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
598 MACROBLOCKD *const xd = &x->e_mbd;
599 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
601 mbmi->tx_size = VPXMIN(max_tx_size, largest_tx_size);
603 txfm_rd_in_plane(x, rate, distortion, skip,
604 sse, ref_best_rd, 0, bs,
605 mbmi->tx_size, cpi->sf.use_fast_coef_costing);
608 static void choose_smallest_tx_size(VP10_COMP *cpi, MACROBLOCK *x,
609 int *rate, int64_t *distortion,
610 int *skip, int64_t *sse,
613 MACROBLOCKD *const xd = &x->e_mbd;
614 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
616 mbmi->tx_size = TX_4X4;
618 txfm_rd_in_plane(x, rate, distortion, skip,
619 sse, ref_best_rd, 0, bs,
620 mbmi->tx_size, cpi->sf.use_fast_coef_costing);
623 static void choose_tx_size_from_rd(VP10_COMP *cpi, MACROBLOCK *x,
630 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
631 VP10_COMMON *const cm = &cpi->common;
632 MACROBLOCKD *const xd = &x->e_mbd;
633 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
634 vpx_prob skip_prob = vp10_get_skip_prob(cm, xd);
635 int r[TX_SIZES][2], s[TX_SIZES];
636 int64_t d[TX_SIZES], sse[TX_SIZES];
637 int64_t rd[TX_SIZES][2] = {{INT64_MAX, INT64_MAX},
638 {INT64_MAX, INT64_MAX},
639 {INT64_MAX, INT64_MAX},
640 {INT64_MAX, INT64_MAX}};
643 int64_t best_rd = INT64_MAX;
644 TX_SIZE best_tx = max_tx_size;
645 int start_tx, end_tx;
647 const vpx_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc->tx_probs);
648 assert(skip_prob > 0);
649 s0 = vp10_cost_bit(skip_prob, 0);
650 s1 = vp10_cost_bit(skip_prob, 1);
652 if (cm->tx_mode == TX_MODE_SELECT) {
653 start_tx = max_tx_size;
656 TX_SIZE chosen_tx_size = VPXMIN(max_tx_size,
657 tx_mode_to_biggest_tx_size[cm->tx_mode]);
658 start_tx = chosen_tx_size;
659 end_tx = chosen_tx_size;
662 for (n = start_tx; n >= end_tx; n--) {
664 for (m = 0; m <= n - (n == (int) max_tx_size); m++) {
666 r_tx_size += vp10_cost_zero(tx_probs[m]);
668 r_tx_size += vp10_cost_one(tx_probs[m]);
670 txfm_rd_in_plane(x, &r[n][0], &d[n], &s[n],
671 &sse[n], ref_best_rd, 0, bs, n,
672 cpi->sf.use_fast_coef_costing);
674 if (r[n][0] < INT_MAX) {
675 r[n][1] += r_tx_size;
677 if (d[n] == INT64_MAX || r[n][0] == INT_MAX) {
678 rd[n][0] = rd[n][1] = INT64_MAX;
680 if (is_inter_block(mbmi)) {
681 rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
682 r[n][1] -= r_tx_size;
684 rd[n][0] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
685 rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1 + r_tx_size, sse[n]);
688 rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
689 rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
692 if (is_inter_block(mbmi) && !xd->lossless[mbmi->segment_id] &&
693 !s[n] && sse[n] != INT64_MAX) {
694 rd[n][0] = VPXMIN(rd[n][0], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
695 rd[n][1] = VPXMIN(rd[n][1], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
698 // Early termination in transform size search.
699 if (cpi->sf.tx_size_search_breakout &&
700 (rd[n][1] == INT64_MAX ||
701 (n < (int) max_tx_size && rd[n][1] > rd[n + 1][1]) ||
705 if (rd[n][1] < best_rd) {
710 mbmi->tx_size = best_tx;
712 *distortion = d[mbmi->tx_size];
713 *rate = r[mbmi->tx_size][cm->tx_mode == TX_MODE_SELECT];
714 *skip = s[mbmi->tx_size];
715 *psse = sse[mbmi->tx_size];
718 static void super_block_yrd(VP10_COMP *cpi, MACROBLOCK *x, int *rate,
719 int64_t *distortion, int *skip,
720 int64_t *psse, BLOCK_SIZE bs,
721 int64_t ref_best_rd) {
722 MACROBLOCKD *xd = &x->e_mbd;
724 int64_t *ret_sse = psse ? psse : &sse;
726 assert(bs == xd->mi[0]->mbmi.sb_type);
728 if (CONFIG_MISC_FIXES && xd->lossless[xd->mi[0]->mbmi.segment_id]) {
729 choose_smallest_tx_size(cpi, x, rate, distortion, skip, ret_sse,
731 } else if (cpi->sf.tx_size_search_method == USE_LARGESTALL ||
732 xd->lossless[xd->mi[0]->mbmi.segment_id]) {
733 choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
736 choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse,
741 static int conditional_skipintra(PREDICTION_MODE mode,
742 PREDICTION_MODE best_intra_mode) {
743 if (mode == D117_PRED &&
744 best_intra_mode != V_PRED &&
745 best_intra_mode != D135_PRED)
747 if (mode == D63_PRED &&
748 best_intra_mode != V_PRED &&
749 best_intra_mode != D45_PRED)
751 if (mode == D207_PRED &&
752 best_intra_mode != H_PRED &&
753 best_intra_mode != D45_PRED)
755 if (mode == D153_PRED &&
756 best_intra_mode != H_PRED &&
757 best_intra_mode != D135_PRED)
762 static int64_t rd_pick_intra4x4block(VP10_COMP *cpi, MACROBLOCK *x,
764 PREDICTION_MODE *best_mode,
765 const int *bmode_costs,
766 ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
767 int *bestrate, int *bestratey,
768 int64_t *bestdistortion,
769 BLOCK_SIZE bsize, int64_t rd_thresh) {
770 PREDICTION_MODE mode;
771 MACROBLOCKD *const xd = &x->e_mbd;
772 int64_t best_rd = rd_thresh;
773 struct macroblock_plane *p = &x->plane[0];
774 struct macroblockd_plane *pd = &xd->plane[0];
775 const int src_stride = p->src.stride;
776 const int dst_stride = pd->dst.stride;
777 const uint8_t *src_init = &p->src.buf[row * 4 * src_stride + col * 4];
778 uint8_t *dst_init = &pd->dst.buf[row * 4 * src_stride + col * 4];
779 ENTROPY_CONTEXT ta[2], tempa[2];
780 ENTROPY_CONTEXT tl[2], templ[2];
781 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
782 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
784 uint8_t best_dst[8 * 8];
785 #if CONFIG_VP9_HIGHBITDEPTH
786 uint16_t best_dst16[8 * 8];
789 memcpy(ta, a, sizeof(ta));
790 memcpy(tl, l, sizeof(tl));
791 xd->mi[0]->mbmi.tx_size = TX_4X4;
793 #if CONFIG_VP9_HIGHBITDEPTH
794 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
795 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
798 int64_t distortion = 0;
799 int rate = bmode_costs[mode];
801 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
804 // Only do the oblique modes if the best so far is
805 // one of the neighboring directional modes
806 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
807 if (conditional_skipintra(mode, *best_mode))
811 memcpy(tempa, ta, sizeof(ta));
812 memcpy(templ, tl, sizeof(tl));
814 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
815 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
816 const int block = (row + idy) * 2 + (col + idx);
817 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
818 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
819 int16_t *const src_diff = vp10_raster_block_offset_int16(BLOCK_8X8,
822 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
823 xd->mi[0]->bmi[block].as_mode = mode;
824 vp10_predict_intra_block(xd, 1, 1, TX_4X4, mode, dst, dst_stride,
826 col + idx, row + idy, 0);
827 vpx_highbd_subtract_block(4, 4, src_diff, 8, src, src_stride,
828 dst, dst_stride, xd->bd);
829 if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
830 TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block);
831 const scan_order *so = get_scan(TX_4X4, tx_type);
832 vp10_highbd_fwd_txfm_4x4(src_diff, coeff, 8, DCT_DCT, 1);
833 vp10_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
834 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
835 so->scan, so->neighbors,
836 cpi->sf.use_fast_coef_costing);
837 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
839 vp10_highbd_inv_txfm_add_4x4(BLOCK_OFFSET(pd->dqcoeff, block),
840 dst, dst_stride, p->eobs[block],
844 TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block);
845 const scan_order *so = get_scan(TX_4X4, tx_type);
846 vp10_highbd_fwd_txfm_4x4(src_diff, coeff, 8, tx_type, 0);
847 vp10_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
848 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
849 so->scan, so->neighbors,
850 cpi->sf.use_fast_coef_costing);
851 distortion += vp10_highbd_block_error(
852 coeff, BLOCK_OFFSET(pd->dqcoeff, block),
853 16, &unused, xd->bd) >> 2;
854 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
856 vp10_highbd_inv_txfm_add_4x4(BLOCK_OFFSET(pd->dqcoeff, block),
857 dst, dst_stride, p->eobs[block],
864 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
866 if (this_rd < best_rd) {
869 *bestdistortion = distortion;
872 memcpy(a, tempa, sizeof(tempa));
873 memcpy(l, templ, sizeof(templ));
874 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
875 memcpy(best_dst16 + idy * 8,
876 CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
877 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
883 if (best_rd >= rd_thresh)
886 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
887 memcpy(CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
888 best_dst16 + idy * 8,
889 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
894 #endif // CONFIG_VP9_HIGHBITDEPTH
896 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
899 int64_t distortion = 0;
900 int rate = bmode_costs[mode];
902 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
905 // Only do the oblique modes if the best so far is
906 // one of the neighboring directional modes
907 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
908 if (conditional_skipintra(mode, *best_mode))
912 memcpy(tempa, ta, sizeof(ta));
913 memcpy(templ, tl, sizeof(tl));
915 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
916 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
917 const int block = (row + idy) * 2 + (col + idx);
918 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
919 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
920 int16_t *const src_diff =
921 vp10_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
922 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
923 xd->mi[0]->bmi[block].as_mode = mode;
924 vp10_predict_intra_block(xd, 1, 1, TX_4X4, mode, dst, dst_stride,
925 dst, dst_stride, col + idx, row + idy, 0);
926 vpx_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
928 if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
929 TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block);
930 const scan_order *so = get_scan(TX_4X4, tx_type);
931 vp10_fwd_txfm_4x4(src_diff, coeff, 8, DCT_DCT, 1);
932 vp10_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
933 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
934 so->scan, so->neighbors,
935 cpi->sf.use_fast_coef_costing);
936 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
938 vp10_inv_txfm_add_4x4(BLOCK_OFFSET(pd->dqcoeff, block),
939 dst, dst_stride, p->eobs[block], DCT_DCT, 1);
942 TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block);
943 const scan_order *so = get_scan(TX_4X4, tx_type);
944 vp10_fwd_txfm_4x4(src_diff, coeff, 8, tx_type, 0);
945 vp10_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
946 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
947 so->scan, so->neighbors,
948 cpi->sf.use_fast_coef_costing);
949 distortion += vp10_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
951 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
953 vp10_inv_txfm_add_4x4(BLOCK_OFFSET(pd->dqcoeff, block),
954 dst, dst_stride, p->eobs[block], tx_type, 0);
960 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
962 if (this_rd < best_rd) {
965 *bestdistortion = distortion;
968 memcpy(a, tempa, sizeof(tempa));
969 memcpy(l, templ, sizeof(templ));
970 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
971 memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
972 num_4x4_blocks_wide * 4);
978 if (best_rd >= rd_thresh)
981 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
982 memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
983 num_4x4_blocks_wide * 4);
988 static int64_t rd_pick_intra_sub_8x8_y_mode(VP10_COMP *cpi, MACROBLOCK *mb,
989 int *rate, int *rate_y,
993 const MACROBLOCKD *const xd = &mb->e_mbd;
994 MODE_INFO *const mic = xd->mi[0];
995 const MODE_INFO *above_mi = xd->above_mi;
996 const MODE_INFO *left_mi = xd->left_mi;
997 const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
998 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
999 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1002 int64_t total_distortion = 0;
1004 int64_t total_rd = 0;
1005 ENTROPY_CONTEXT t_above[4], t_left[4];
1006 const int *bmode_costs = cpi->mbmode_cost;
1008 memcpy(t_above, xd->plane[0].above_context, sizeof(t_above));
1009 memcpy(t_left, xd->plane[0].left_context, sizeof(t_left));
1011 // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
1012 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1013 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1014 PREDICTION_MODE best_mode = DC_PRED;
1015 int r = INT_MAX, ry = INT_MAX;
1016 int64_t d = INT64_MAX, this_rd = INT64_MAX;
1018 if (cpi->common.frame_type == KEY_FRAME) {
1019 const PREDICTION_MODE A = vp10_above_block_mode(mic, above_mi, i);
1020 const PREDICTION_MODE L = vp10_left_block_mode(mic, left_mi, i);
1022 bmode_costs = cpi->y_mode_costs[A][L];
1025 this_rd = rd_pick_intra4x4block(cpi, mb, idy, idx, &best_mode,
1026 bmode_costs, t_above + idx, t_left + idy,
1027 &r, &ry, &d, bsize, best_rd - total_rd);
1028 if (this_rd >= best_rd - total_rd)
1031 total_rd += this_rd;
1033 total_distortion += d;
1036 mic->bmi[i].as_mode = best_mode;
1037 for (j = 1; j < num_4x4_blocks_high; ++j)
1038 mic->bmi[i + j * 2].as_mode = best_mode;
1039 for (j = 1; j < num_4x4_blocks_wide; ++j)
1040 mic->bmi[i + j].as_mode = best_mode;
1042 if (total_rd >= best_rd)
1048 *rate_y = tot_rate_y;
1049 *distortion = total_distortion;
1050 mic->mbmi.mode = mic->bmi[3].as_mode;
1052 return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
1055 // This function is used only for intra_only frames
1056 static int64_t rd_pick_intra_sby_mode(VP10_COMP *cpi, MACROBLOCK *x,
1057 int *rate, int *rate_tokenonly,
1058 int64_t *distortion, int *skippable,
1061 PREDICTION_MODE mode;
1062 PREDICTION_MODE mode_selected = DC_PRED;
1063 MACROBLOCKD *const xd = &x->e_mbd;
1064 MODE_INFO *const mic = xd->mi[0];
1065 int this_rate, this_rate_tokenonly, s;
1066 int64_t this_distortion, this_rd;
1067 TX_SIZE best_tx = TX_4X4;
1069 const MODE_INFO *above_mi = xd->above_mi;
1070 const MODE_INFO *left_mi = xd->left_mi;
1071 const PREDICTION_MODE A = vp10_above_block_mode(mic, above_mi, 0);
1072 const PREDICTION_MODE L = vp10_left_block_mode(mic, left_mi, 0);
1073 bmode_costs = cpi->y_mode_costs[A][L];
1075 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1077 /* Y Search for intra prediction mode */
1078 for (mode = DC_PRED; mode <= TM_PRED; mode++) {
1079 mic->mbmi.mode = mode;
1081 super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion,
1082 &s, NULL, bsize, best_rd);
1084 if (this_rate_tokenonly == INT_MAX)
1087 this_rate = this_rate_tokenonly + bmode_costs[mode];
1088 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1090 if (this_rd < best_rd) {
1091 mode_selected = mode;
1093 best_tx = mic->mbmi.tx_size;
1095 *rate_tokenonly = this_rate_tokenonly;
1096 *distortion = this_distortion;
1101 mic->mbmi.mode = mode_selected;
1102 mic->mbmi.tx_size = best_tx;
1107 // Return value 0: early termination triggered, no valid rd cost available;
1108 // 1: rd cost values are valid.
1109 static int super_block_uvrd(const VP10_COMP *cpi, MACROBLOCK *x,
1110 int *rate, int64_t *distortion, int *skippable,
1111 int64_t *sse, BLOCK_SIZE bsize,
1112 int64_t ref_best_rd) {
1113 MACROBLOCKD *const xd = &x->e_mbd;
1114 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1115 const TX_SIZE uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
1117 int pnrate = 0, pnskip = 1;
1118 int64_t pndist = 0, pnsse = 0;
1119 int is_cost_valid = 1;
1121 if (ref_best_rd < 0)
1124 if (is_inter_block(mbmi) && is_cost_valid) {
1126 for (plane = 1; plane < MAX_MB_PLANE; ++plane)
1127 vp10_subtract_plane(x, bsize, plane);
1135 for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
1136 txfm_rd_in_plane(x, &pnrate, &pndist, &pnskip, &pnsse,
1137 ref_best_rd, plane, bsize, uv_tx_size,
1138 cpi->sf.use_fast_coef_costing);
1139 if (pnrate == INT_MAX) {
1144 *distortion += pndist;
1146 *skippable &= pnskip;
1149 if (!is_cost_valid) {
1152 *distortion = INT64_MAX;
1157 return is_cost_valid;
1160 static int64_t rd_pick_intra_sbuv_mode(VP10_COMP *cpi, MACROBLOCK *x,
1161 PICK_MODE_CONTEXT *ctx,
1162 int *rate, int *rate_tokenonly,
1163 int64_t *distortion, int *skippable,
1164 BLOCK_SIZE bsize, TX_SIZE max_tx_size) {
1165 MACROBLOCKD *xd = &x->e_mbd;
1166 PREDICTION_MODE mode;
1167 PREDICTION_MODE mode_selected = DC_PRED;
1168 int64_t best_rd = INT64_MAX, this_rd;
1169 int this_rate_tokenonly, this_rate, s;
1170 int64_t this_distortion, this_sse;
1172 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1173 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1174 if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode)))
1177 xd->mi[0]->mbmi.uv_mode = mode;
1179 if (!super_block_uvrd(cpi, x, &this_rate_tokenonly,
1180 &this_distortion, &s, &this_sse, bsize, best_rd))
1182 this_rate = this_rate_tokenonly +
1183 cpi->intra_uv_mode_cost[xd->mi[0]->mbmi.mode][mode];
1184 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1186 if (this_rd < best_rd) {
1187 mode_selected = mode;
1190 *rate_tokenonly = this_rate_tokenonly;
1191 *distortion = this_distortion;
1193 if (!x->select_tx_size)
1194 swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
1198 xd->mi[0]->mbmi.uv_mode = mode_selected;
1202 static int64_t rd_sbuv_dcpred(const VP10_COMP *cpi, MACROBLOCK *x,
1203 int *rate, int *rate_tokenonly,
1204 int64_t *distortion, int *skippable,
1208 x->e_mbd.mi[0]->mbmi.uv_mode = DC_PRED;
1209 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1210 super_block_uvrd(cpi, x, rate_tokenonly, distortion,
1211 skippable, &unused, bsize, INT64_MAX);
1212 *rate = *rate_tokenonly +
1213 cpi->intra_uv_mode_cost[x->e_mbd.mi[0]->mbmi.mode][DC_PRED];
1214 return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
1217 static void choose_intra_uv_mode(VP10_COMP *cpi, MACROBLOCK *const x,
1218 PICK_MODE_CONTEXT *ctx,
1219 BLOCK_SIZE bsize, TX_SIZE max_tx_size,
1220 int *rate_uv, int *rate_uv_tokenonly,
1221 int64_t *dist_uv, int *skip_uv,
1222 PREDICTION_MODE *mode_uv) {
1223 // Use an estimated rd for uv_intra based on DC_PRED if the
1224 // appropriate speed flag is set.
1225 if (cpi->sf.use_uv_intra_rd_estimate) {
1226 rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv,
1227 skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
1228 // Else do a proper rd search for each possible transform size that may
1229 // be considered in the main rd loop.
1231 rd_pick_intra_sbuv_mode(cpi, x, ctx,
1232 rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
1233 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize, max_tx_size);
1235 *mode_uv = x->e_mbd.mi[0]->mbmi.uv_mode;
1238 static int cost_mv_ref(const VP10_COMP *cpi, PREDICTION_MODE mode,
1240 assert(is_inter_mode(mode));
1241 return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
1244 static int set_and_cost_bmi_mvs(VP10_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
1246 PREDICTION_MODE mode, int_mv this_mv[2],
1247 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1248 int_mv seg_mvs[MAX_REF_FRAMES],
1249 int_mv *best_ref_mv[2], const int *mvjcost,
1251 MODE_INFO *const mic = xd->mi[0];
1252 const MB_MODE_INFO *const mbmi = &mic->mbmi;
1253 const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1256 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type];
1257 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
1258 const int is_compound = has_second_ref(mbmi);
1262 this_mv[0].as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
1263 thismvcost += vp10_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
1264 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1266 this_mv[1].as_int = seg_mvs[mbmi->ref_frame[1]].as_int;
1267 thismvcost += vp10_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
1268 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1273 this_mv[0].as_int = frame_mv[mode][mbmi->ref_frame[0]].as_int;
1275 this_mv[1].as_int = frame_mv[mode][mbmi->ref_frame[1]].as_int;
1278 this_mv[0].as_int = 0;
1280 this_mv[1].as_int = 0;
1286 mic->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
1288 mic->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
1290 mic->bmi[i].as_mode = mode;
1292 for (idy = 0; idy < num_4x4_blocks_high; ++idy)
1293 for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
1294 memmove(&mic->bmi[i + idy * 2 + idx], &mic->bmi[i], sizeof(mic->bmi[i]));
1296 return cost_mv_ref(cpi, mode, mbmi_ext->mode_context[mbmi->ref_frame[0]]) +
1300 static int64_t encode_inter_mb_segment(VP10_COMP *cpi,
1305 int64_t *distortion, int64_t *sse,
1306 ENTROPY_CONTEXT *ta,
1307 ENTROPY_CONTEXT *tl,
1309 int mi_row, int mi_col) {
1311 MACROBLOCKD *xd = &x->e_mbd;
1312 struct macroblockd_plane *const pd = &xd->plane[0];
1313 struct macroblock_plane *const p = &x->plane[0];
1314 MODE_INFO *const mi = xd->mi[0];
1315 const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->mbmi.sb_type, pd);
1316 const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
1317 const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
1319 void (*fwd_txm4x4)(const int16_t *input, tran_low_t *output, int stride);
1321 const uint8_t *const src =
1322 &p->src.buf[vp10_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1323 uint8_t *const dst = &pd->dst.buf[vp10_raster_block_offset(BLOCK_8X8, i,
1325 int64_t thisdistortion = 0, thissse = 0;
1327 TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, i);
1328 const scan_order *so = get_scan(TX_4X4, tx_type);
1330 vp10_build_inter_predictor_sub8x8(xd, 0, i, ir, ic, mi_row, mi_col);
1332 #if CONFIG_VP9_HIGHBITDEPTH
1333 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1334 fwd_txm4x4 = xd->lossless[mi->mbmi.segment_id] ? vp10_highbd_fwht4x4
1335 : vpx_highbd_fdct4x4;
1337 fwd_txm4x4 = xd->lossless[mi->mbmi.segment_id] ? vp10_fwht4x4 : vpx_fdct4x4;
1340 fwd_txm4x4 = xd->lossless[mi->mbmi.segment_id] ? vp10_fwht4x4 : vpx_fdct4x4;
1341 #endif // CONFIG_VP9_HIGHBITDEPTH
1343 #if CONFIG_VP9_HIGHBITDEPTH
1344 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1345 vpx_highbd_subtract_block(
1346 height, width, vp10_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1347 8, src, p->src.stride, dst, pd->dst.stride, xd->bd);
1350 height, width, vp10_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1351 8, src, p->src.stride, dst, pd->dst.stride);
1354 vpx_subtract_block(height, width,
1355 vp10_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1356 8, src, p->src.stride, dst, pd->dst.stride);
1357 #endif // CONFIG_VP9_HIGHBITDEPTH
1360 for (idy = 0; idy < height / 4; ++idy) {
1361 for (idx = 0; idx < width / 4; ++idx) {
1362 int64_t ssz, rd, rd1, rd2;
1365 k += (idy * 2 + idx);
1366 coeff = BLOCK_OFFSET(p->coeff, k);
1367 fwd_txm4x4(vp10_raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
1369 vp10_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan);
1370 #if CONFIG_VP9_HIGHBITDEPTH
1371 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1372 thisdistortion += vp10_highbd_block_error(coeff,
1373 BLOCK_OFFSET(pd->dqcoeff, k),
1376 thisdistortion += vp10_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
1380 thisdistortion += vp10_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
1382 #endif // CONFIG_VP9_HIGHBITDEPTH
1384 thisrate += cost_coeffs(x, 0, k, ta + (k & 1), tl + (k >> 1), TX_4X4,
1385 so->scan, so->neighbors,
1386 cpi->sf.use_fast_coef_costing);
1387 rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
1388 rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
1389 rd = VPXMIN(rd1, rd2);
1395 *distortion = thisdistortion >> 2;
1396 *labelyrate = thisrate;
1397 *sse = thissse >> 2;
1399 return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
1410 ENTROPY_CONTEXT ta[2];
1411 ENTROPY_CONTEXT tl[2];
1423 PREDICTION_MODE modes[4];
1424 SEG_RDSTAT rdstat[4][INTER_MODES];
1428 static INLINE int mv_check_bounds(const MACROBLOCK *x, const MV *mv) {
1429 return (mv->row >> 3) < x->mv_row_min ||
1430 (mv->row >> 3) > x->mv_row_max ||
1431 (mv->col >> 3) < x->mv_col_min ||
1432 (mv->col >> 3) > x->mv_col_max;
1435 static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
1436 MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0]->mbmi;
1437 struct macroblock_plane *const p = &x->plane[0];
1438 struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
1440 p->src.buf = &p->src.buf[vp10_raster_block_offset(BLOCK_8X8, i,
1442 assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
1443 pd->pre[0].buf = &pd->pre[0].buf[vp10_raster_block_offset(BLOCK_8X8, i,
1444 pd->pre[0].stride)];
1445 if (has_second_ref(mbmi))
1446 pd->pre[1].buf = &pd->pre[1].buf[vp10_raster_block_offset(BLOCK_8X8, i,
1447 pd->pre[1].stride)];
1450 static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
1451 struct buf_2d orig_pre[2]) {
1452 MB_MODE_INFO *mbmi = &x->e_mbd.mi[0]->mbmi;
1453 x->plane[0].src = orig_src;
1454 x->e_mbd.plane[0].pre[0] = orig_pre[0];
1455 if (has_second_ref(mbmi))
1456 x->e_mbd.plane[0].pre[1] = orig_pre[1];
1459 static INLINE int mv_has_subpel(const MV *mv) {
1460 return (mv->row & 0x0F) || (mv->col & 0x0F);
1463 // Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
1464 // TODO(aconverse): Find out if this is still productive then clean up or remove
1465 static int check_best_zero_mv(
1466 const VP10_COMP *cpi, const uint8_t mode_context[MAX_REF_FRAMES],
1467 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int this_mode,
1468 const MV_REFERENCE_FRAME ref_frames[2]) {
1469 if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
1470 frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
1471 (ref_frames[1] == NONE ||
1472 frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
1473 int rfc = mode_context[ref_frames[0]];
1474 int c1 = cost_mv_ref(cpi, NEARMV, rfc);
1475 int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
1476 int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
1478 if (this_mode == NEARMV) {
1479 if (c1 > c3) return 0;
1480 } else if (this_mode == NEARESTMV) {
1481 if (c2 > c3) return 0;
1483 assert(this_mode == ZEROMV);
1484 if (ref_frames[1] == NONE) {
1485 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
1486 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
1489 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
1490 frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
1491 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
1492 frame_mv[NEARMV][ref_frames[1]].as_int == 0))
1500 static void joint_motion_search(VP10_COMP *cpi, MACROBLOCK *x,
1503 int mi_row, int mi_col,
1504 int_mv single_newmv[MAX_REF_FRAMES],
1506 const VP10_COMMON *const cm = &cpi->common;
1507 const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
1508 const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
1509 MACROBLOCKD *xd = &x->e_mbd;
1510 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
1511 const int refs[2] = {mbmi->ref_frame[0],
1512 mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]};
1515 const InterpKernel *kernel = vp10_filter_kernels[mbmi->interp_filter];
1516 struct scale_factors sf;
1518 // Do joint motion search in compound mode to get more accurate mv.
1519 struct buf_2d backup_yv12[2][MAX_MB_PLANE];
1520 int last_besterr[2] = {INT_MAX, INT_MAX};
1521 const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
1522 vp10_get_scaled_ref_frame(cpi, mbmi->ref_frame[0]),
1523 vp10_get_scaled_ref_frame(cpi, mbmi->ref_frame[1])
1526 // Prediction buffer from second frame.
1527 #if CONFIG_VP9_HIGHBITDEPTH
1528 DECLARE_ALIGNED(16, uint16_t, second_pred_alloc_16[64 * 64]);
1529 uint8_t *second_pred;
1531 DECLARE_ALIGNED(16, uint8_t, second_pred[64 * 64]);
1532 #endif // CONFIG_VP9_HIGHBITDEPTH
1534 for (ref = 0; ref < 2; ++ref) {
1535 ref_mv[ref] = x->mbmi_ext->ref_mvs[refs[ref]][0];
1537 if (scaled_ref_frame[ref]) {
1539 // Swap out the reference frame for a version that's been scaled to
1540 // match the resolution of the current frame, allowing the existing
1541 // motion search code to be used without additional modifications.
1542 for (i = 0; i < MAX_MB_PLANE; i++)
1543 backup_yv12[ref][i] = xd->plane[i].pre[ref];
1544 vp10_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
1548 frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
1551 // Since we have scaled the reference frames to match the size of the current
1552 // frame we must use a unit scaling factor during mode selection.
1553 #if CONFIG_VP9_HIGHBITDEPTH
1554 vp10_setup_scale_factors_for_frame(&sf, cm->width, cm->height,
1555 cm->width, cm->height,
1556 cm->use_highbitdepth);
1558 vp10_setup_scale_factors_for_frame(&sf, cm->width, cm->height,
1559 cm->width, cm->height);
1560 #endif // CONFIG_VP9_HIGHBITDEPTH
1562 // Allow joint search multiple times iteratively for each reference frame
1563 // and break out of the search loop if it couldn't find a better mv.
1564 for (ite = 0; ite < 4; ite++) {
1565 struct buf_2d ref_yv12[2];
1566 int bestsme = INT_MAX;
1567 int sadpb = x->sadperbit16;
1569 int search_range = 3;
1571 int tmp_col_min = x->mv_col_min;
1572 int tmp_col_max = x->mv_col_max;
1573 int tmp_row_min = x->mv_row_min;
1574 int tmp_row_max = x->mv_row_max;
1575 int id = ite % 2; // Even iterations search in the first reference frame,
1576 // odd iterations search in the second. The predictor
1577 // found for the 'other' reference frame is factored in.
1579 // Initialized here because of compiler problem in Visual Studio.
1580 ref_yv12[0] = xd->plane[0].pre[0];
1581 ref_yv12[1] = xd->plane[0].pre[1];
1583 // Get the prediction block from the 'other' reference frame.
1584 #if CONFIG_VP9_HIGHBITDEPTH
1585 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1586 second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc_16);
1587 vp10_highbd_build_inter_predictor(ref_yv12[!id].buf,
1588 ref_yv12[!id].stride,
1590 &frame_mv[refs[!id]].as_mv,
1592 kernel, MV_PRECISION_Q3,
1593 mi_col * MI_SIZE, mi_row * MI_SIZE,
1596 second_pred = (uint8_t *)second_pred_alloc_16;
1597 vp10_build_inter_predictor(ref_yv12[!id].buf,
1598 ref_yv12[!id].stride,
1600 &frame_mv[refs[!id]].as_mv,
1602 kernel, MV_PRECISION_Q3,
1603 mi_col * MI_SIZE, mi_row * MI_SIZE);
1606 vp10_build_inter_predictor(ref_yv12[!id].buf,
1607 ref_yv12[!id].stride,
1609 &frame_mv[refs[!id]].as_mv,
1611 kernel, MV_PRECISION_Q3,
1612 mi_col * MI_SIZE, mi_row * MI_SIZE);
1613 #endif // CONFIG_VP9_HIGHBITDEPTH
1615 // Do compound motion search on the current reference frame.
1617 xd->plane[0].pre[0] = ref_yv12[id];
1618 vp10_set_mv_search_range(x, &ref_mv[id].as_mv);
1620 // Use the mv result from the single mode as mv predictor.
1621 tmp_mv = frame_mv[refs[id]].as_mv;
1626 // Small-range full-pixel motion search.
1627 bestsme = vp10_refining_search_8p_c(x, &tmp_mv, sadpb,
1629 &cpi->fn_ptr[bsize],
1630 &ref_mv[id].as_mv, second_pred);
1631 if (bestsme < INT_MAX)
1632 bestsme = vp10_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
1633 second_pred, &cpi->fn_ptr[bsize], 1);
1635 x->mv_col_min = tmp_col_min;
1636 x->mv_col_max = tmp_col_max;
1637 x->mv_row_min = tmp_row_min;
1638 x->mv_row_max = tmp_row_max;
1640 if (bestsme < INT_MAX) {
1641 int dis; /* TODO: use dis in distortion calculation later. */
1643 bestsme = cpi->find_fractional_mv_step(
1646 cpi->common.allow_high_precision_mv,
1648 &cpi->fn_ptr[bsize],
1649 0, cpi->sf.mv.subpel_iters_per_step,
1651 x->nmvjointcost, x->mvcost,
1652 &dis, &sse, second_pred,
1656 // Restore the pointer to the first (possibly scaled) prediction buffer.
1658 xd->plane[0].pre[0] = ref_yv12[0];
1660 if (bestsme < last_besterr[id]) {
1661 frame_mv[refs[id]].as_mv = tmp_mv;
1662 last_besterr[id] = bestsme;
1670 for (ref = 0; ref < 2; ++ref) {
1671 if (scaled_ref_frame[ref]) {
1672 // Restore the prediction frame pointers to their unscaled versions.
1674 for (i = 0; i < MAX_MB_PLANE; i++)
1675 xd->plane[i].pre[ref] = backup_yv12[ref][i];
1678 *rate_mv += vp10_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
1679 &x->mbmi_ext->ref_mvs[refs[ref]][0].as_mv,
1680 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1684 static int64_t rd_pick_best_sub8x8_mode(VP10_COMP *cpi, MACROBLOCK *x,
1685 int_mv *best_ref_mv,
1686 int_mv *second_best_ref_mv,
1687 int64_t best_rd, int *returntotrate,
1689 int64_t *returndistortion,
1690 int *skippable, int64_t *psse,
1692 int_mv seg_mvs[4][MAX_REF_FRAMES],
1693 BEST_SEG_INFO *bsi_buf, int filter_idx,
1694 int mi_row, int mi_col) {
1696 BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
1697 MACROBLOCKD *xd = &x->e_mbd;
1698 MODE_INFO *mi = xd->mi[0];
1699 MB_MODE_INFO *mbmi = &mi->mbmi;
1701 int k, br = 0, idx, idy;
1702 int64_t bd = 0, block_sse = 0;
1703 PREDICTION_MODE this_mode;
1704 VP10_COMMON *cm = &cpi->common;
1705 struct macroblock_plane *const p = &x->plane[0];
1706 struct macroblockd_plane *const pd = &xd->plane[0];
1707 const int label_count = 4;
1708 int64_t this_segment_rd = 0;
1709 int label_mv_thresh;
1710 int segmentyrate = 0;
1711 const BLOCK_SIZE bsize = mbmi->sb_type;
1712 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1713 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1714 ENTROPY_CONTEXT t_above[2], t_left[2];
1715 int subpelmv = 1, have_ref = 0;
1716 const int has_second_rf = has_second_ref(mbmi);
1717 const int inter_mode_mask = cpi->sf.inter_mode_mask[bsize];
1718 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1722 bsi->segment_rd = best_rd;
1723 bsi->ref_mv[0] = best_ref_mv;
1724 bsi->ref_mv[1] = second_best_ref_mv;
1725 bsi->mvp.as_int = best_ref_mv->as_int;
1726 bsi->mvthresh = mvthresh;
1728 for (i = 0; i < 4; i++)
1729 bsi->modes[i] = ZEROMV;
1731 memcpy(t_above, pd->above_context, sizeof(t_above));
1732 memcpy(t_left, pd->left_context, sizeof(t_left));
1734 // 64 makes this threshold really big effectively
1735 // making it so that we very rarely check mvs on
1736 // segments. setting this to 1 would make mv thresh
1737 // roughly equal to what it is for macroblocks
1738 label_mv_thresh = 1 * bsi->mvthresh / label_count;
1740 // Segmentation method overheads
1741 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1742 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1743 // TODO(jingning,rbultje): rewrite the rate-distortion optimization
1744 // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
1745 int_mv mode_mv[MB_MODE_COUNT][2];
1746 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1747 PREDICTION_MODE mode_selected = ZEROMV;
1748 int64_t best_rd = INT64_MAX;
1749 const int i = idy * 2 + idx;
1752 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1753 const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
1754 frame_mv[ZEROMV][frame].as_int = 0;
1755 vp10_append_sub8x8_mvs_for_idx(cm, xd, i, ref, mi_row, mi_col,
1756 &frame_mv[NEARESTMV][frame],
1757 &frame_mv[NEARMV][frame],
1758 mbmi_ext->mode_context);
1761 // search for the best motion vector on this segment
1762 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1763 const struct buf_2d orig_src = x->plane[0].src;
1764 struct buf_2d orig_pre[2];
1766 mode_idx = INTER_OFFSET(this_mode);
1767 bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
1768 if (!(inter_mode_mask & (1 << this_mode)))
1771 if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
1772 this_mode, mbmi->ref_frame))
1775 memcpy(orig_pre, pd->pre, sizeof(orig_pre));
1776 memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
1777 sizeof(bsi->rdstat[i][mode_idx].ta));
1778 memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
1779 sizeof(bsi->rdstat[i][mode_idx].tl));
1781 // motion search for newmv (single predictor case only)
1782 if (!has_second_rf && this_mode == NEWMV &&
1783 seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV) {
1784 MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
1786 int thissme, bestsme = INT_MAX;
1787 int sadpb = x->sadperbit4;
1792 /* Is the best so far sufficiently good that we cant justify doing
1793 * and new motion search. */
1794 if (best_rd < label_mv_thresh)
1797 if (cpi->oxcf.mode != BEST) {
1798 // use previous block's result as next block's MV predictor.
1800 bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
1802 bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
1806 max_mv = x->max_mv_context[mbmi->ref_frame[0]];
1809 VPXMAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
1811 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
1812 // Take wtd average of the step_params based on the last frame's
1813 // max mv magnitude and the best ref mvs of the current block for
1814 // the given reference.
1815 step_param = (vp10_init_search_range(max_mv) +
1816 cpi->mv_step_param) / 2;
1818 step_param = cpi->mv_step_param;
1821 mvp_full.row = bsi->mvp.as_mv.row >> 3;
1822 mvp_full.col = bsi->mvp.as_mv.col >> 3;
1824 if (cpi->sf.adaptive_motion_search) {
1825 mvp_full.row = x->pred_mv[mbmi->ref_frame[0]].row >> 3;
1826 mvp_full.col = x->pred_mv[mbmi->ref_frame[0]].col >> 3;
1827 step_param = VPXMAX(step_param, 8);
1830 // adjust src pointer for this block
1833 vp10_set_mv_search_range(x, &bsi->ref_mv[0]->as_mv);
1835 bestsme = vp10_full_pixel_search(
1836 cpi, x, bsize, &mvp_full, step_param, sadpb,
1837 cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
1838 &bsi->ref_mv[0]->as_mv, new_mv,
1841 // Should we do a full search (best quality only)
1842 if (cpi->oxcf.mode == BEST) {
1843 int_mv *const best_mv = &mi->bmi[i].as_mv[0];
1844 /* Check if mvp_full is within the range. */
1845 clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max,
1846 x->mv_row_min, x->mv_row_max);
1847 thissme = cpi->full_search_sad(x, &mvp_full,
1848 sadpb, 16, &cpi->fn_ptr[bsize],
1849 &bsi->ref_mv[0]->as_mv,
1851 cost_list[1] = cost_list[2] = cost_list[3] = cost_list[4] = INT_MAX;
1852 if (thissme < bestsme) {
1854 *new_mv = best_mv->as_mv;
1856 // The full search result is actually worse so re-instate the
1857 // previous best vector
1858 best_mv->as_mv = *new_mv;
1862 if (bestsme < INT_MAX) {
1864 cpi->find_fractional_mv_step(
1867 &bsi->ref_mv[0]->as_mv,
1868 cm->allow_high_precision_mv,
1869 x->errorperbit, &cpi->fn_ptr[bsize],
1870 cpi->sf.mv.subpel_force_stop,
1871 cpi->sf.mv.subpel_iters_per_step,
1872 cond_cost_list(cpi, cost_list),
1873 x->nmvjointcost, x->mvcost,
1875 &x->pred_sse[mbmi->ref_frame[0]],
1878 // save motion search result for use in compound prediction
1879 seg_mvs[i][mbmi->ref_frame[0]].as_mv = *new_mv;
1882 if (cpi->sf.adaptive_motion_search)
1883 x->pred_mv[mbmi->ref_frame[0]] = *new_mv;
1885 // restore src pointers
1886 mi_buf_restore(x, orig_src, orig_pre);
1889 if (has_second_rf) {
1890 if (seg_mvs[i][mbmi->ref_frame[1]].as_int == INVALID_MV ||
1891 seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV)
1895 if (has_second_rf && this_mode == NEWMV &&
1896 mbmi->interp_filter == EIGHTTAP) {
1897 // adjust src pointers
1899 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
1901 joint_motion_search(cpi, x, bsize, frame_mv[this_mode],
1902 mi_row, mi_col, seg_mvs[i],
1904 seg_mvs[i][mbmi->ref_frame[0]].as_int =
1905 frame_mv[this_mode][mbmi->ref_frame[0]].as_int;
1906 seg_mvs[i][mbmi->ref_frame[1]].as_int =
1907 frame_mv[this_mode][mbmi->ref_frame[1]].as_int;
1909 // restore src pointers
1910 mi_buf_restore(x, orig_src, orig_pre);
1913 bsi->rdstat[i][mode_idx].brate =
1914 set_and_cost_bmi_mvs(cpi, x, xd, i, this_mode, mode_mv[this_mode],
1915 frame_mv, seg_mvs[i], bsi->ref_mv,
1916 x->nmvjointcost, x->mvcost);
1918 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1919 bsi->rdstat[i][mode_idx].mvs[ref].as_int =
1920 mode_mv[this_mode][ref].as_int;
1921 if (num_4x4_blocks_wide > 1)
1922 bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int =
1923 mode_mv[this_mode][ref].as_int;
1924 if (num_4x4_blocks_high > 1)
1925 bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int =
1926 mode_mv[this_mode][ref].as_int;
1929 // Trap vectors that reach beyond the UMV borders
1930 if (mv_check_bounds(x, &mode_mv[this_mode][0].as_mv) ||
1932 mv_check_bounds(x, &mode_mv[this_mode][1].as_mv)))
1935 if (filter_idx > 0) {
1936 BEST_SEG_INFO *ref_bsi = bsi_buf;
1940 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1941 subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
1942 have_ref &= mode_mv[this_mode][ref].as_int ==
1943 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1946 if (filter_idx > 1 && !subpelmv && !have_ref) {
1947 ref_bsi = bsi_buf + 1;
1949 for (ref = 0; ref < 1 + has_second_rf; ++ref)
1950 have_ref &= mode_mv[this_mode][ref].as_int ==
1951 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1954 if (!subpelmv && have_ref &&
1955 ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1956 memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
1957 sizeof(SEG_RDSTAT));
1958 if (num_4x4_blocks_wide > 1)
1959 bsi->rdstat[i + 1][mode_idx].eobs =
1960 ref_bsi->rdstat[i + 1][mode_idx].eobs;
1961 if (num_4x4_blocks_high > 1)
1962 bsi->rdstat[i + 2][mode_idx].eobs =
1963 ref_bsi->rdstat[i + 2][mode_idx].eobs;
1965 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
1966 mode_selected = this_mode;
1967 best_rd = bsi->rdstat[i][mode_idx].brdcost;
1973 bsi->rdstat[i][mode_idx].brdcost =
1974 encode_inter_mb_segment(cpi, x,
1975 bsi->segment_rd - this_segment_rd, i,
1976 &bsi->rdstat[i][mode_idx].byrate,
1977 &bsi->rdstat[i][mode_idx].bdist,
1978 &bsi->rdstat[i][mode_idx].bsse,
1979 bsi->rdstat[i][mode_idx].ta,
1980 bsi->rdstat[i][mode_idx].tl,
1983 if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1984 bsi->rdstat[i][mode_idx].brdcost += RDCOST(x->rdmult, x->rddiv,
1985 bsi->rdstat[i][mode_idx].brate, 0);
1986 bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
1987 bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
1988 if (num_4x4_blocks_wide > 1)
1989 bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
1990 if (num_4x4_blocks_high > 1)
1991 bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
1994 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
1995 mode_selected = this_mode;
1996 best_rd = bsi->rdstat[i][mode_idx].brdcost;
1998 } /*for each 4x4 mode*/
2000 if (best_rd == INT64_MAX) {
2002 for (iy = i + 1; iy < 4; ++iy)
2003 for (midx = 0; midx < INTER_MODES; ++midx)
2004 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2005 bsi->segment_rd = INT64_MAX;
2009 mode_idx = INTER_OFFSET(mode_selected);
2010 memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
2011 memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
2013 set_and_cost_bmi_mvs(cpi, x, xd, i, mode_selected, mode_mv[mode_selected],
2014 frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
2017 br += bsi->rdstat[i][mode_idx].brate;
2018 bd += bsi->rdstat[i][mode_idx].bdist;
2019 block_sse += bsi->rdstat[i][mode_idx].bsse;
2020 segmentyrate += bsi->rdstat[i][mode_idx].byrate;
2021 this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
2023 if (this_segment_rd > bsi->segment_rd) {
2025 for (iy = i + 1; iy < 4; ++iy)
2026 for (midx = 0; midx < INTER_MODES; ++midx)
2027 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2028 bsi->segment_rd = INT64_MAX;
2032 } /* for each label */
2036 bsi->segment_yrate = segmentyrate;
2037 bsi->segment_rd = this_segment_rd;
2038 bsi->sse = block_sse;
2040 // update the coding decisions
2041 for (k = 0; k < 4; ++k)
2042 bsi->modes[k] = mi->bmi[k].as_mode;
2044 if (bsi->segment_rd > best_rd)
2046 /* set it to the best */
2047 for (i = 0; i < 4; i++) {
2048 mode_idx = INTER_OFFSET(bsi->modes[i]);
2049 mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
2050 if (has_second_ref(mbmi))
2051 mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
2052 x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
2053 mi->bmi[i].as_mode = bsi->modes[i];
2057 * used to set mbmi->mv.as_int
2059 *returntotrate = bsi->r;
2060 *returndistortion = bsi->d;
2061 *returnyrate = bsi->segment_yrate;
2062 *skippable = vp10_is_skippable_in_plane(x, BLOCK_8X8, 0);
2064 mbmi->mode = bsi->modes[3];
2066 return bsi->segment_rd;
2069 static void estimate_ref_frame_costs(const VP10_COMMON *cm,
2070 const MACROBLOCKD *xd,
2072 unsigned int *ref_costs_single,
2073 unsigned int *ref_costs_comp,
2074 vpx_prob *comp_mode_p) {
2075 int seg_ref_active = segfeature_active(&cm->seg, segment_id,
2077 if (seg_ref_active) {
2078 memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
2079 memset(ref_costs_comp, 0, MAX_REF_FRAMES * sizeof(*ref_costs_comp));
2082 vpx_prob intra_inter_p = vp10_get_intra_inter_prob(cm, xd);
2083 vpx_prob comp_inter_p = 128;
2085 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
2086 comp_inter_p = vp10_get_reference_mode_prob(cm, xd);
2087 *comp_mode_p = comp_inter_p;
2092 ref_costs_single[INTRA_FRAME] = vp10_cost_bit(intra_inter_p, 0);
2094 if (cm->reference_mode != COMPOUND_REFERENCE) {
2095 vpx_prob ref_single_p1 = vp10_get_pred_prob_single_ref_p1(cm, xd);
2096 vpx_prob ref_single_p2 = vp10_get_pred_prob_single_ref_p2(cm, xd);
2097 unsigned int base_cost = vp10_cost_bit(intra_inter_p, 1);
2099 if (cm->reference_mode == REFERENCE_MODE_SELECT)
2100 base_cost += vp10_cost_bit(comp_inter_p, 0);
2102 ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] =
2103 ref_costs_single[ALTREF_FRAME] = base_cost;
2104 ref_costs_single[LAST_FRAME] += vp10_cost_bit(ref_single_p1, 0);
2105 ref_costs_single[GOLDEN_FRAME] += vp10_cost_bit(ref_single_p1, 1);
2106 ref_costs_single[ALTREF_FRAME] += vp10_cost_bit(ref_single_p1, 1);
2107 ref_costs_single[GOLDEN_FRAME] += vp10_cost_bit(ref_single_p2, 0);
2108 ref_costs_single[ALTREF_FRAME] += vp10_cost_bit(ref_single_p2, 1);
2110 ref_costs_single[LAST_FRAME] = 512;
2111 ref_costs_single[GOLDEN_FRAME] = 512;
2112 ref_costs_single[ALTREF_FRAME] = 512;
2114 if (cm->reference_mode != SINGLE_REFERENCE) {
2115 vpx_prob ref_comp_p = vp10_get_pred_prob_comp_ref_p(cm, xd);
2116 unsigned int base_cost = vp10_cost_bit(intra_inter_p, 1);
2118 if (cm->reference_mode == REFERENCE_MODE_SELECT)
2119 base_cost += vp10_cost_bit(comp_inter_p, 1);
2121 ref_costs_comp[LAST_FRAME] = base_cost + vp10_cost_bit(ref_comp_p, 0);
2122 ref_costs_comp[GOLDEN_FRAME] = base_cost + vp10_cost_bit(ref_comp_p, 1);
2124 ref_costs_comp[LAST_FRAME] = 512;
2125 ref_costs_comp[GOLDEN_FRAME] = 512;
2130 static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
2132 int64_t comp_pred_diff[REFERENCE_MODES],
2133 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS],
2135 MACROBLOCKD *const xd = &x->e_mbd;
2137 // Take a snapshot of the coding context so it can be
2138 // restored if we decide to encode this way
2139 ctx->skip = x->skip;
2140 ctx->skippable = skippable;
2141 ctx->best_mode_index = mode_index;
2142 ctx->mic = *xd->mi[0];
2143 ctx->mbmi_ext = *x->mbmi_ext;
2144 ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
2145 ctx->comp_pred_diff = (int)comp_pred_diff[COMPOUND_REFERENCE];
2146 ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
2148 memcpy(ctx->best_filter_diff, best_filter_diff,
2149 sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
2152 static void setup_buffer_inter(VP10_COMP *cpi, MACROBLOCK *x,
2153 MV_REFERENCE_FRAME ref_frame,
2154 BLOCK_SIZE block_size,
2155 int mi_row, int mi_col,
2156 int_mv frame_nearest_mv[MAX_REF_FRAMES],
2157 int_mv frame_near_mv[MAX_REF_FRAMES],
2158 struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
2159 const VP10_COMMON *cm = &cpi->common;
2160 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2161 MACROBLOCKD *const xd = &x->e_mbd;
2162 MODE_INFO *const mi = xd->mi[0];
2163 int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
2164 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2165 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2167 assert(yv12 != NULL);
2169 // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
2170 // use the UV scaling factors.
2171 vp10_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
2173 // Gets an initial list of candidate vectors from neighbours and orders them
2174 vp10_find_mv_refs(cm, xd, mi, ref_frame, candidates, mi_row, mi_col,
2175 NULL, NULL, mbmi_ext->mode_context);
2177 // Candidate refinement carried out at encoder and decoder
2178 vp10_find_best_ref_mvs(cm->allow_high_precision_mv, candidates,
2179 &frame_nearest_mv[ref_frame],
2180 &frame_near_mv[ref_frame]);
2182 // Further refinement that is encode side only to test the top few candidates
2183 // in full and choose the best as the centre point for subsequent searches.
2184 // The current implementation doesn't support scaling.
2185 if (!vp10_is_scaled(sf) && block_size >= BLOCK_8X8)
2186 vp10_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
2187 ref_frame, block_size);
2190 static void single_motion_search(VP10_COMP *cpi, MACROBLOCK *x,
2192 int mi_row, int mi_col,
2193 int_mv *tmp_mv, int *rate_mv) {
2194 MACROBLOCKD *xd = &x->e_mbd;
2195 const VP10_COMMON *cm = &cpi->common;
2196 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
2197 struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
2198 int bestsme = INT_MAX;
2200 int sadpb = x->sadperbit16;
2202 int ref = mbmi->ref_frame[0];
2203 MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2205 int tmp_col_min = x->mv_col_min;
2206 int tmp_col_max = x->mv_col_max;
2207 int tmp_row_min = x->mv_row_min;
2208 int tmp_row_max = x->mv_row_max;
2211 const YV12_BUFFER_CONFIG *scaled_ref_frame = vp10_get_scaled_ref_frame(cpi,
2215 pred_mv[0] = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2216 pred_mv[1] = x->mbmi_ext->ref_mvs[ref][1].as_mv;
2217 pred_mv[2] = x->pred_mv[ref];
2219 if (scaled_ref_frame) {
2221 // Swap out the reference frame for a version that's been scaled to
2222 // match the resolution of the current frame, allowing the existing
2223 // motion search code to be used without additional modifications.
2224 for (i = 0; i < MAX_MB_PLANE; i++)
2225 backup_yv12[i] = xd->plane[i].pre[0];
2227 vp10_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
2230 vp10_set_mv_search_range(x, &ref_mv);
2232 // Work out the size of the first step in the mv step search.
2233 // 0 here is maximum length first step. 1 is VPXMAX >> 1 etc.
2234 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
2235 // Take wtd average of the step_params based on the last frame's
2236 // max mv magnitude and that based on the best ref mvs of the current
2237 // block for the given reference.
2238 step_param = (vp10_init_search_range(x->max_mv_context[ref]) +
2239 cpi->mv_step_param) / 2;
2241 step_param = cpi->mv_step_param;
2244 if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64) {
2246 2 * (b_width_log2_lookup[BLOCK_64X64] -
2247 VPXMIN(b_height_log2_lookup[bsize], b_width_log2_lookup[bsize]));
2248 step_param = VPXMAX(step_param, boffset);
2251 if (cpi->sf.adaptive_motion_search) {
2252 int bwl = b_width_log2_lookup[bsize];
2253 int bhl = b_height_log2_lookup[bsize];
2254 int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
2259 // prev_mv_sad is not setup for dynamically scaled frames.
2260 if (cpi->oxcf.resize_mode != RESIZE_DYNAMIC) {
2262 for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
2263 if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
2264 x->pred_mv[ref].row = 0;
2265 x->pred_mv[ref].col = 0;
2266 tmp_mv->as_int = INVALID_MV;
2268 if (scaled_ref_frame) {
2270 for (i = 0; i < MAX_MB_PLANE; ++i)
2271 xd->plane[i].pre[0] = backup_yv12[i];
2279 mvp_full = pred_mv[x->mv_best_ref_index[ref]];
2284 bestsme = vp10_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
2285 cond_cost_list(cpi, cost_list),
2286 &ref_mv, &tmp_mv->as_mv, INT_MAX, 1);
2288 x->mv_col_min = tmp_col_min;
2289 x->mv_col_max = tmp_col_max;
2290 x->mv_row_min = tmp_row_min;
2291 x->mv_row_max = tmp_row_max;
2293 if (bestsme < INT_MAX) {
2294 int dis; /* TODO: use dis in distortion calculation later. */
2295 cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
2296 cm->allow_high_precision_mv,
2298 &cpi->fn_ptr[bsize],
2299 cpi->sf.mv.subpel_force_stop,
2300 cpi->sf.mv.subpel_iters_per_step,
2301 cond_cost_list(cpi, cost_list),
2302 x->nmvjointcost, x->mvcost,
2303 &dis, &x->pred_sse[ref], NULL, 0, 0);
2305 *rate_mv = vp10_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
2306 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2308 if (cpi->sf.adaptive_motion_search)
2309 x->pred_mv[ref] = tmp_mv->as_mv;
2311 if (scaled_ref_frame) {
2313 for (i = 0; i < MAX_MB_PLANE; i++)
2314 xd->plane[i].pre[0] = backup_yv12[i];
2320 static INLINE void restore_dst_buf(MACROBLOCKD *xd,
2321 uint8_t *orig_dst[MAX_MB_PLANE],
2322 int orig_dst_stride[MAX_MB_PLANE]) {
2324 for (i = 0; i < MAX_MB_PLANE; i++) {
2325 xd->plane[i].dst.buf = orig_dst[i];
2326 xd->plane[i].dst.stride = orig_dst_stride[i];
2330 // In some situations we want to discount tha pparent cost of a new motion
2331 // vector. Where there is a subtle motion field and especially where there is
2332 // low spatial complexity then it can be hard to cover the cost of a new motion
2333 // vector in a single block, even if that motion vector reduces distortion.
2334 // However, once established that vector may be usable through the nearest and
2335 // near mv modes to reduce distortion in subsequent blocks and also improve
2337 static int discount_newmv_test(const VP10_COMP *cpi,
2340 int_mv (*mode_mv)[MAX_REF_FRAMES],
2342 return (!cpi->rc.is_src_frame_alt_ref &&
2343 (this_mode == NEWMV) &&
2344 (this_mv.as_int != 0) &&
2345 ((mode_mv[NEARESTMV][ref_frame].as_int == 0) ||
2346 (mode_mv[NEARESTMV][ref_frame].as_int == INVALID_MV)) &&
2347 ((mode_mv[NEARMV][ref_frame].as_int == 0) ||
2348 (mode_mv[NEARMV][ref_frame].as_int == INVALID_MV)));
2351 #define LEFT_TOP_MARGIN ((VP9_ENC_BORDER_IN_PIXELS - VP9_INTERP_EXTEND) << 3)
2352 #define RIGHT_BOTTOM_MARGIN ((VP9_ENC_BORDER_IN_PIXELS -\
2353 VP9_INTERP_EXTEND) << 3)
2355 // TODO(jingning): this mv clamping function should be block size dependent.
2356 static INLINE void clamp_mv2(MV *mv, const MACROBLOCKD *xd) {
2357 clamp_mv(mv, xd->mb_to_left_edge - LEFT_TOP_MARGIN,
2358 xd->mb_to_right_edge + RIGHT_BOTTOM_MARGIN,
2359 xd->mb_to_top_edge - LEFT_TOP_MARGIN,
2360 xd->mb_to_bottom_edge + RIGHT_BOTTOM_MARGIN);
2363 static int64_t handle_inter_mode(VP10_COMP *cpi, MACROBLOCK *x,
2365 int *rate2, int64_t *distortion,
2367 int *rate_y, int *rate_uv,
2369 int_mv (*mode_mv)[MAX_REF_FRAMES],
2370 int mi_row, int mi_col,
2371 int_mv single_newmv[MAX_REF_FRAMES],
2372 INTERP_FILTER (*single_filter)[MAX_REF_FRAMES],
2373 int (*single_skippable)[MAX_REF_FRAMES],
2375 const int64_t ref_best_rd,
2376 int64_t *mask_filter,
2377 int64_t filter_cache[]) {
2378 VP10_COMMON *cm = &cpi->common;
2379 MACROBLOCKD *xd = &x->e_mbd;
2380 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
2381 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2382 const int is_comp_pred = has_second_ref(mbmi);
2383 const int this_mode = mbmi->mode;
2384 int_mv *frame_mv = mode_mv[this_mode];
2386 int refs[2] = { mbmi->ref_frame[0],
2387 (mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]) };
2389 #if CONFIG_VP9_HIGHBITDEPTH
2390 DECLARE_ALIGNED(16, uint16_t, tmp_buf16[MAX_MB_PLANE * 64 * 64]);
2393 DECLARE_ALIGNED(16, uint8_t, tmp_buf[MAX_MB_PLANE * 64 * 64]);
2394 #endif // CONFIG_VP9_HIGHBITDEPTH
2395 int pred_exists = 0;
2397 int64_t rd, tmp_rd, best_rd = INT64_MAX;
2398 int best_needs_copy = 0;
2399 uint8_t *orig_dst[MAX_MB_PLANE];
2400 int orig_dst_stride[MAX_MB_PLANE];
2402 INTERP_FILTER best_filter = SWITCHABLE;
2403 uint8_t skip_txfm[MAX_MB_PLANE << 2] = {0};
2404 int64_t bsse[MAX_MB_PLANE << 2] = {0};
2406 int bsl = mi_width_log2_lookup[bsize];
2407 int pred_filter_search = cpi->sf.cb_pred_filter_search ?
2408 (((mi_row + mi_col) >> bsl) +
2409 get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
2411 int skip_txfm_sb = 0;
2412 int64_t skip_sse_sb = INT64_MAX;
2413 int64_t distortion_y = 0, distortion_uv = 0;
2415 #if CONFIG_VP9_HIGHBITDEPTH
2416 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2417 tmp_buf = CONVERT_TO_BYTEPTR(tmp_buf16);
2419 tmp_buf = (uint8_t *)tmp_buf16;
2421 #endif // CONFIG_VP9_HIGHBITDEPTH
2423 if (pred_filter_search) {
2424 INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE;
2425 if (xd->up_available)
2426 af = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
2427 if (xd->left_available)
2428 lf = xd->mi[-1]->mbmi.interp_filter;
2430 if ((this_mode != NEWMV) || (af == lf))
2435 if (frame_mv[refs[0]].as_int == INVALID_MV ||
2436 frame_mv[refs[1]].as_int == INVALID_MV)
2439 if (cpi->sf.adaptive_mode_search) {
2440 if (single_filter[this_mode][refs[0]] ==
2441 single_filter[this_mode][refs[1]])
2442 best_filter = single_filter[this_mode][refs[0]];
2446 if (this_mode == NEWMV) {
2449 // Initialize mv using single prediction mode result.
2450 frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
2451 frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
2453 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
2454 joint_motion_search(cpi, x, bsize, frame_mv,
2455 mi_row, mi_col, single_newmv, &rate_mv);
2457 rate_mv = vp10_mv_bit_cost(&frame_mv[refs[0]].as_mv,
2458 &x->mbmi_ext->ref_mvs[refs[0]][0].as_mv,
2459 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2460 rate_mv += vp10_mv_bit_cost(&frame_mv[refs[1]].as_mv,
2461 &x->mbmi_ext->ref_mvs[refs[1]][0].as_mv,
2462 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2467 single_motion_search(cpi, x, bsize, mi_row, mi_col,
2469 if (tmp_mv.as_int == INVALID_MV)
2472 frame_mv[refs[0]].as_int =
2473 xd->mi[0]->bmi[0].as_mv[0].as_int = tmp_mv.as_int;
2474 single_newmv[refs[0]].as_int = tmp_mv.as_int;
2476 // Estimate the rate implications of a new mv but discount this
2477 // under certain circumstances where we want to help initiate a weak
2478 // motion field, where the distortion gain for a single block may not
2479 // be enough to overcome the cost of a new mv.
2480 if (discount_newmv_test(cpi, this_mode, tmp_mv, mode_mv, refs[0])) {
2481 *rate2 += VPXMAX((rate_mv / NEW_MV_DISCOUNT_FACTOR), 1);
2488 for (i = 0; i < is_comp_pred + 1; ++i) {
2489 cur_mv[i] = frame_mv[refs[i]];
2490 // Clip "next_nearest" so that it does not extend to far out of image
2491 if (this_mode != NEWMV)
2492 clamp_mv2(&cur_mv[i].as_mv, xd);
2494 if (mv_check_bounds(x, &cur_mv[i].as_mv))
2496 mbmi->mv[i].as_int = cur_mv[i].as_int;
2499 // do first prediction into the destination buffer. Do the next
2500 // prediction into a temporary buffer. Then keep track of which one
2501 // of these currently holds the best predictor, and use the other
2502 // one for future predictions. In the end, copy from tmp_buf to
2503 // dst if necessary.
2504 for (i = 0; i < MAX_MB_PLANE; i++) {
2505 orig_dst[i] = xd->plane[i].dst.buf;
2506 orig_dst_stride[i] = xd->plane[i].dst.stride;
2509 // We don't include the cost of the second reference here, because there
2510 // are only three options: Last/Golden, ARF/Last or Golden/ARF, or in other
2511 // words if you present them in that order, the second one is always known
2512 // if the first is known.
2514 // Under some circumstances we discount the cost of new mv mode to encourage
2515 // initiation of a motion field.
2516 if (discount_newmv_test(cpi, this_mode, frame_mv[refs[0]],
2517 mode_mv, refs[0])) {
2518 *rate2 += VPXMIN(cost_mv_ref(cpi, this_mode,
2519 mbmi_ext->mode_context[refs[0]]),
2520 cost_mv_ref(cpi, NEARESTMV,
2521 mbmi_ext->mode_context[refs[0]]));
2523 *rate2 += cost_mv_ref(cpi, this_mode, mbmi_ext->mode_context[refs[0]]);
2526 if (RDCOST(x->rdmult, x->rddiv, *rate2, 0) > ref_best_rd &&
2527 mbmi->mode != NEARESTMV)
2531 // Are all MVs integer pel for Y and UV
2532 intpel_mv = !mv_has_subpel(&mbmi->mv[0].as_mv);
2534 intpel_mv &= !mv_has_subpel(&mbmi->mv[1].as_mv);
2536 // Search for best switchable filter by checking the variance of
2537 // pred error irrespective of whether the filter will be used
2538 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
2539 filter_cache[i] = INT64_MAX;
2541 if (cm->interp_filter != BILINEAR) {
2542 if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
2543 best_filter = EIGHTTAP;
2544 } else if (best_filter == SWITCHABLE) {
2546 int tmp_rate_sum = 0;
2547 int64_t tmp_dist_sum = 0;
2549 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
2552 int tmp_skip_sb = 0;
2553 int64_t tmp_skip_sse = INT64_MAX;
2555 mbmi->interp_filter = i;
2556 rs = vp10_get_switchable_rate(cpi, xd);
2557 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
2559 if (i > 0 && intpel_mv) {
2560 rd = RDCOST(x->rdmult, x->rddiv, tmp_rate_sum, tmp_dist_sum);
2561 filter_cache[i] = rd;
2562 filter_cache[SWITCHABLE_FILTERS] =
2563 VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2564 if (cm->interp_filter == SWITCHABLE)
2566 *mask_filter = VPXMAX(*mask_filter, rd);
2569 int64_t dist_sum = 0;
2570 if (i > 0 && cpi->sf.adaptive_interp_filter_search &&
2571 (cpi->sf.interp_filter_search_mask & (1 << i))) {
2573 dist_sum = INT64_MAX;
2577 if ((cm->interp_filter == SWITCHABLE &&
2578 (!i || best_needs_copy)) ||
2579 (cm->interp_filter != SWITCHABLE &&
2580 (cm->interp_filter == mbmi->interp_filter ||
2581 (i == 0 && intpel_mv)))) {
2582 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2584 for (j = 0; j < MAX_MB_PLANE; j++) {
2585 xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
2586 xd->plane[j].dst.stride = 64;
2589 vp10_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2590 model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum,
2591 &tmp_skip_sb, &tmp_skip_sse);
2593 rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum);
2594 filter_cache[i] = rd;
2595 filter_cache[SWITCHABLE_FILTERS] =
2596 VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2597 if (cm->interp_filter == SWITCHABLE)
2599 *mask_filter = VPXMAX(*mask_filter, rd);
2601 if (i == 0 && intpel_mv) {
2602 tmp_rate_sum = rate_sum;
2603 tmp_dist_sum = dist_sum;
2607 if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2608 if (rd / 2 > ref_best_rd) {
2609 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2613 newbest = i == 0 || rd < best_rd;
2617 best_filter = mbmi->interp_filter;
2618 if (cm->interp_filter == SWITCHABLE && i && !intpel_mv)
2619 best_needs_copy = !best_needs_copy;
2622 if ((cm->interp_filter == SWITCHABLE && newbest) ||
2623 (cm->interp_filter != SWITCHABLE &&
2624 cm->interp_filter == mbmi->interp_filter)) {
2628 skip_txfm_sb = tmp_skip_sb;
2629 skip_sse_sb = tmp_skip_sse;
2630 memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2631 memcpy(bsse, x->bsse, sizeof(bsse));
2634 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2637 // Set the appropriate filter
2638 mbmi->interp_filter = cm->interp_filter != SWITCHABLE ?
2639 cm->interp_filter : best_filter;
2640 rs = cm->interp_filter == SWITCHABLE ? vp10_get_switchable_rate(cpi, xd) : 0;
2643 if (best_needs_copy) {
2644 // again temporarily set the buffers to local memory to prevent a memcpy
2645 for (i = 0; i < MAX_MB_PLANE; i++) {
2646 xd->plane[i].dst.buf = tmp_buf + i * 64 * 64;
2647 xd->plane[i].dst.stride = 64;
2650 rd = tmp_rd + RDCOST(x->rdmult, x->rddiv, rs, 0);
2654 // Handles the special case when a filter that is not in the
2655 // switchable list (ex. bilinear) is indicated at the frame level, or
2656 // skip condition holds.
2657 vp10_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2658 model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist,
2659 &skip_txfm_sb, &skip_sse_sb);
2660 rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
2661 memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2662 memcpy(bsse, x->bsse, sizeof(bsse));
2666 single_filter[this_mode][refs[0]] = mbmi->interp_filter;
2668 if (cpi->sf.adaptive_mode_search)
2670 if (single_skippable[this_mode][refs[0]] &&
2671 single_skippable[this_mode][refs[1]])
2672 memset(skip_txfm, SKIP_TXFM_AC_DC, sizeof(skip_txfm));
2674 if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2675 // if current pred_error modeled rd is substantially more than the best
2676 // so far, do not bother doing full rd
2677 if (rd / 2 > ref_best_rd) {
2678 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2683 if (cm->interp_filter == SWITCHABLE)
2686 memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm));
2687 memcpy(x->bsse, bsse, sizeof(bsse));
2689 if (!skip_txfm_sb) {
2690 int skippable_y, skippable_uv;
2691 int64_t sseuv = INT64_MAX;
2692 int64_t rdcosty = INT64_MAX;
2694 // Y cost and distortion
2695 vp10_subtract_plane(x, bsize, 0);
2696 super_block_yrd(cpi, x, rate_y, &distortion_y, &skippable_y, psse,
2697 bsize, ref_best_rd);
2699 if (*rate_y == INT_MAX) {
2701 *distortion = INT64_MAX;
2702 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2707 *distortion += distortion_y;
2709 rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
2710 rdcosty = VPXMIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
2712 if (!super_block_uvrd(cpi, x, rate_uv, &distortion_uv, &skippable_uv,
2713 &sseuv, bsize, ref_best_rd - rdcosty)) {
2715 *distortion = INT64_MAX;
2716 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2722 *distortion += distortion_uv;
2723 *skippable = skippable_y && skippable_uv;
2728 // The cost of skip bit needs to be added.
2729 *rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 1);
2731 *distortion = skip_sse_sb;
2735 single_skippable[this_mode][refs[0]] = *skippable;
2737 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2738 return 0; // The rate-distortion cost will be re-calculated by caller.
2741 void vp10_rd_pick_intra_mode_sb(VP10_COMP *cpi, MACROBLOCK *x,
2742 RD_COST *rd_cost, BLOCK_SIZE bsize,
2743 PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
2744 VP10_COMMON *const cm = &cpi->common;
2745 MACROBLOCKD *const xd = &x->e_mbd;
2746 struct macroblockd_plane *const pd = xd->plane;
2747 int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
2748 int y_skip = 0, uv_skip = 0;
2749 int64_t dist_y = 0, dist_uv = 0;
2750 TX_SIZE max_uv_tx_size;
2752 xd->mi[0]->mbmi.ref_frame[0] = INTRA_FRAME;
2753 xd->mi[0]->mbmi.ref_frame[1] = NONE;
2755 if (bsize >= BLOCK_8X8) {
2756 if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2757 &dist_y, &y_skip, bsize,
2758 best_rd) >= best_rd) {
2759 rd_cost->rate = INT_MAX;
2764 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2765 &dist_y, best_rd) >= best_rd) {
2766 rd_cost->rate = INT_MAX;
2770 max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0]->mbmi.tx_size, bsize,
2771 pd[1].subsampling_x,
2772 pd[1].subsampling_y);
2773 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
2774 &dist_uv, &uv_skip, VPXMAX(BLOCK_8X8, bsize),
2777 if (y_skip && uv_skip) {
2778 rd_cost->rate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
2779 vp10_cost_bit(vp10_get_skip_prob(cm, xd), 1);
2780 rd_cost->dist = dist_y + dist_uv;
2782 rd_cost->rate = rate_y + rate_uv +
2783 vp10_cost_bit(vp10_get_skip_prob(cm, xd), 0);
2784 rd_cost->dist = dist_y + dist_uv;
2787 ctx->mic = *xd->mi[0];
2788 ctx->mbmi_ext = *x->mbmi_ext;
2789 rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
2792 // This function is designed to apply a bias or adjustment to an rd value based
2793 // on the relative variance of the source and reconstruction.
2794 #define LOW_VAR_THRESH 16
2795 #define VLOW_ADJ_MAX 25
2796 #define VHIGH_ADJ_MAX 8
2797 static void rd_variance_adjustment(VP10_COMP *cpi,
2801 MV_REFERENCE_FRAME ref_frame,
2802 unsigned int source_variance) {
2803 MACROBLOCKD *const xd = &x->e_mbd;
2804 unsigned int recon_variance;
2805 unsigned int absvar_diff = 0;
2806 int64_t var_error = 0;
2807 int64_t var_factor = 0;
2809 if (*this_rd == INT64_MAX)
2812 #if CONFIG_VP9_HIGHBITDEPTH
2813 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2815 vp10_high_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize, xd->bd);
2818 vp10_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
2822 vp10_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
2823 #endif // CONFIG_VP9_HIGHBITDEPTH
2825 if ((source_variance + recon_variance) > LOW_VAR_THRESH) {
2826 absvar_diff = (source_variance > recon_variance)
2827 ? (source_variance - recon_variance)
2828 : (recon_variance - source_variance);
2830 var_error = ((int64_t)200 * source_variance * recon_variance) /
2831 (((int64_t)source_variance * source_variance) +
2832 ((int64_t)recon_variance * recon_variance));
2833 var_error = 100 - var_error;
2836 // Source variance above a threshold and ref frame is intra.
2837 // This case is targeted mainly at discouraging intra modes that give rise
2838 // to a predictor with a low spatial complexity compared to the source.
2839 if ((source_variance > LOW_VAR_THRESH) && (ref_frame == INTRA_FRAME) &&
2840 (source_variance > recon_variance)) {
2841 var_factor = VPXMIN(absvar_diff, VPXMIN(VLOW_ADJ_MAX, var_error));
2842 // A second possible case of interest is where the source variance
2843 // is very low and we wish to discourage false texture or motion trails.
2844 } else if ((source_variance < (LOW_VAR_THRESH >> 1)) &&
2845 (recon_variance > source_variance)) {
2846 var_factor = VPXMIN(absvar_diff, VPXMIN(VHIGH_ADJ_MAX, var_error));
2848 *this_rd += (*this_rd * var_factor) / 100;
2852 // Do we have an internal image edge (e.g. formatting bars).
2853 int vp10_internal_image_edge(VP10_COMP *cpi) {
2854 return (cpi->oxcf.pass == 2) &&
2855 ((cpi->twopass.this_frame_stats.inactive_zone_rows > 0) ||
2856 (cpi->twopass.this_frame_stats.inactive_zone_cols > 0));
2859 // Checks to see if a super block is on a horizontal image edge.
2860 // In most cases this is the "real" edge unless there are formatting
2861 // bars embedded in the stream.
2862 int vp10_active_h_edge(VP10_COMP *cpi, int mi_row, int mi_step) {
2864 int bottom_edge = cpi->common.mi_rows;
2865 int is_active_h_edge = 0;
2867 // For two pass account for any formatting bars detected.
2868 if (cpi->oxcf.pass == 2) {
2869 TWO_PASS *twopass = &cpi->twopass;
2871 // The inactive region is specified in MBs not mi units.
2872 // The image edge is in the following MB row.
2873 top_edge += (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
2875 bottom_edge -= (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
2876 bottom_edge = VPXMAX(top_edge, bottom_edge);
2879 if (((top_edge >= mi_row) && (top_edge < (mi_row + mi_step))) ||
2880 ((bottom_edge >= mi_row) && (bottom_edge < (mi_row + mi_step)))) {
2881 is_active_h_edge = 1;
2883 return is_active_h_edge;
2886 // Checks to see if a super block is on a vertical image edge.
2887 // In most cases this is the "real" edge unless there are formatting
2888 // bars embedded in the stream.
2889 int vp10_active_v_edge(VP10_COMP *cpi, int mi_col, int mi_step) {
2891 int right_edge = cpi->common.mi_cols;
2892 int is_active_v_edge = 0;
2894 // For two pass account for any formatting bars detected.
2895 if (cpi->oxcf.pass == 2) {
2896 TWO_PASS *twopass = &cpi->twopass;
2898 // The inactive region is specified in MBs not mi units.
2899 // The image edge is in the following MB row.
2900 left_edge += (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
2902 right_edge -= (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
2903 right_edge = VPXMAX(left_edge, right_edge);
2906 if (((left_edge >= mi_col) && (left_edge < (mi_col + mi_step))) ||
2907 ((right_edge >= mi_col) && (right_edge < (mi_col + mi_step)))) {
2908 is_active_v_edge = 1;
2910 return is_active_v_edge;
2913 // Checks to see if a super block is at the edge of the active image.
2914 // In most cases this is the "real" edge unless there are formatting
2915 // bars embedded in the stream.
2916 int vp10_active_edge_sb(VP10_COMP *cpi,
2917 int mi_row, int mi_col) {
2918 return vp10_active_h_edge(cpi, mi_row, MI_BLOCK_SIZE) ||
2919 vp10_active_v_edge(cpi, mi_col, MI_BLOCK_SIZE);
2922 void vp10_rd_pick_inter_mode_sb(VP10_COMP *cpi,
2923 TileDataEnc *tile_data,
2925 int mi_row, int mi_col,
2926 RD_COST *rd_cost, BLOCK_SIZE bsize,
2927 PICK_MODE_CONTEXT *ctx,
2928 int64_t best_rd_so_far) {
2929 VP10_COMMON *const cm = &cpi->common;
2930 RD_OPT *const rd_opt = &cpi->rd;
2931 SPEED_FEATURES *const sf = &cpi->sf;
2932 MACROBLOCKD *const xd = &x->e_mbd;
2933 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
2934 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2935 const struct segmentation *const seg = &cm->seg;
2936 PREDICTION_MODE this_mode;
2937 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
2938 unsigned char segment_id = mbmi->segment_id;
2939 int comp_pred, i, k;
2940 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
2941 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2942 int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
2943 INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES];
2944 int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES];
2945 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2947 int64_t best_rd = best_rd_so_far;
2948 int64_t best_pred_diff[REFERENCE_MODES];
2949 int64_t best_pred_rd[REFERENCE_MODES];
2950 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
2951 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
2952 MB_MODE_INFO best_mbmode;
2953 int best_mode_skippable = 0;
2954 int midx, best_mode_index = -1;
2955 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
2956 vpx_prob comp_mode_p;
2957 int64_t best_intra_rd = INT64_MAX;
2958 unsigned int best_pred_sse = UINT_MAX;
2959 PREDICTION_MODE best_intra_mode = DC_PRED;
2960 int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
2961 int64_t dist_uv[TX_SIZES];
2962 int skip_uv[TX_SIZES];
2963 PREDICTION_MODE mode_uv[TX_SIZES];
2964 const int intra_cost_penalty = vp10_get_intra_cost_penalty(
2965 cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
2967 uint8_t ref_frame_skip_mask[2] = { 0 };
2968 uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 };
2969 int mode_skip_start = sf->mode_skip_start + 1;
2970 const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
2971 const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
2972 int64_t mode_threshold[MAX_MODES];
2973 int *mode_map = tile_data->mode_map[bsize];
2974 const int mode_search_skip_flags = sf->mode_search_skip_flags;
2975 int64_t mask_filter = 0;
2976 int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
2978 vp10_zero(best_mbmode);
2980 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
2981 filter_cache[i] = INT64_MAX;
2983 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
2986 for (i = 0; i < REFERENCE_MODES; ++i)
2987 best_pred_rd[i] = INT64_MAX;
2988 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
2989 best_filter_rd[i] = INT64_MAX;
2990 for (i = 0; i < TX_SIZES; i++)
2991 rate_uv_intra[i] = INT_MAX;
2992 for (i = 0; i < MAX_REF_FRAMES; ++i)
2993 x->pred_sse[i] = INT_MAX;
2994 for (i = 0; i < MB_MODE_COUNT; ++i) {
2995 for (k = 0; k < MAX_REF_FRAMES; ++k) {
2996 single_inter_filter[i][k] = SWITCHABLE;
2997 single_skippable[i][k] = 0;
3001 rd_cost->rate = INT_MAX;
3003 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3004 x->pred_mv_sad[ref_frame] = INT_MAX;
3005 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3006 assert(get_ref_frame_buffer(cpi, ref_frame) != NULL);
3007 setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
3008 frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
3010 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3011 frame_mv[ZEROMV][ref_frame].as_int = 0;
3014 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3015 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) {
3016 // Skip checking missing references in both single and compound reference
3017 // modes. Note that a mode will be skipped iff both reference frames
3019 ref_frame_skip_mask[0] |= (1 << ref_frame);
3020 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3022 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
3023 // Skip fixed mv modes for poor references
3024 if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
3025 mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
3030 // If the segment reference frame feature is enabled....
3031 // then do nothing if the current ref frame is not allowed..
3032 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3033 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3034 ref_frame_skip_mask[0] |= (1 << ref_frame);
3035 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3039 // Disable this drop out case if the ref frame
3040 // segment level feature is enabled for this segment. This is to
3041 // prevent the possibility that we end up unable to pick any mode.
3042 if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3043 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3044 // unless ARNR filtering is enabled in which case we want
3045 // an unfiltered alternative. We allow near/nearest as well
3046 // because they may result in zero-zero MVs but be cheaper.
3047 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
3048 ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME);
3049 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3050 mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
3051 if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
3052 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
3053 if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
3054 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
3058 if (cpi->rc.is_src_frame_alt_ref) {
3059 if (sf->alt_ref_search_fp) {
3060 mode_skip_mask[ALTREF_FRAME] = 0;
3061 ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME);
3062 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3066 if (sf->alt_ref_search_fp)
3067 if (!cm->show_frame && x->pred_mv_sad[GOLDEN_FRAME] < INT_MAX)
3068 if (x->pred_mv_sad[ALTREF_FRAME] > (x->pred_mv_sad[GOLDEN_FRAME] << 1))
3069 mode_skip_mask[ALTREF_FRAME] |= INTER_ALL;
3071 if (sf->adaptive_mode_search) {
3072 if (cm->show_frame && !cpi->rc.is_src_frame_alt_ref &&
3073 cpi->rc.frames_since_golden >= 3)
3074 if (x->pred_mv_sad[GOLDEN_FRAME] > (x->pred_mv_sad[LAST_FRAME] << 1))
3075 mode_skip_mask[GOLDEN_FRAME] |= INTER_ALL;
3078 if (bsize > sf->max_intra_bsize) {
3079 ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
3080 ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
3083 mode_skip_mask[INTRA_FRAME] |=
3084 ~(sf->intra_y_mode_mask[max_txsize_lookup[bsize]]);
3086 for (i = 0; i <= LAST_NEW_MV_INDEX; ++i)
3087 mode_threshold[i] = 0;
3088 for (i = LAST_NEW_MV_INDEX + 1; i < MAX_MODES; ++i)
3089 mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
3091 midx = sf->schedule_mode_search ? mode_skip_start : 0;
3093 uint8_t end_pos = 0;
3094 for (i = 5; i < midx; ++i) {
3095 if (mode_threshold[mode_map[i - 1]] > mode_threshold[mode_map[i]]) {
3096 uint8_t tmp = mode_map[i];
3097 mode_map[i] = mode_map[i - 1];
3098 mode_map[i - 1] = tmp;
3105 for (midx = 0; midx < MAX_MODES; ++midx) {
3106 int mode_index = mode_map[midx];
3107 int mode_excluded = 0;
3108 int64_t this_rd = INT64_MAX;
3109 int disable_skip = 0;
3110 int compmode_cost = 0;
3111 int rate2 = 0, rate_y = 0, rate_uv = 0;
3112 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3115 int64_t total_sse = INT64_MAX;
3118 this_mode = vp10_mode_order[mode_index].mode;
3119 ref_frame = vp10_mode_order[mode_index].ref_frame[0];
3120 second_ref_frame = vp10_mode_order[mode_index].ref_frame[1];
3122 // Look at the reference frame of the best mode so far and set the
3123 // skip mask to look at a subset of the remaining modes.
3124 if (midx == mode_skip_start && best_mode_index >= 0) {
3125 switch (best_mbmode.ref_frame[0]) {
3129 ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK;
3130 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3133 ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
3134 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3137 ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK;
3140 case MAX_REF_FRAMES:
3141 assert(0 && "Invalid Reference frame");
3146 if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3147 (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
3150 if (mode_skip_mask[ref_frame] & (1 << this_mode))
3153 // Test best rd so far against threshold for trying this mode.
3154 if (best_mode_skippable && sf->schedule_mode_search)
3155 mode_threshold[mode_index] <<= 1;
3157 if (best_rd < mode_threshold[mode_index])
3160 comp_pred = second_ref_frame > INTRA_FRAME;
3162 if (!cpi->allow_comp_inter_inter)
3165 // Skip compound inter modes if ARF is not available.
3166 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
3169 // Do not allow compound prediction if the segment level reference frame
3170 // feature is in use as in this case there can only be one reference.
3171 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
3174 if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3175 best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
3178 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3180 if (ref_frame != INTRA_FRAME)
3181 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3184 if (ref_frame == INTRA_FRAME) {
3185 if (sf->adaptive_mode_search)
3186 if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
3189 if (this_mode != DC_PRED) {
3190 // Disable intra modes other than DC_PRED for blocks with low variance
3191 // Threshold for intra skipping based on source variance
3192 // TODO(debargha): Specialize the threshold for super block sizes
3193 const unsigned int skip_intra_var_thresh = 64;
3194 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
3195 x->source_variance < skip_intra_var_thresh)
3197 // Only search the oblique modes if the best so far is
3198 // one of the neighboring directional modes
3199 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
3200 (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
3201 if (best_mode_index >= 0 &&
3202 best_mbmode.ref_frame[0] > INTRA_FRAME)
3205 if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
3206 if (conditional_skipintra(this_mode, best_intra_mode))
3211 const MV_REFERENCE_FRAME ref_frames[2] = {ref_frame, second_ref_frame};
3212 if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
3213 this_mode, ref_frames))
3217 mbmi->mode = this_mode;
3218 mbmi->uv_mode = DC_PRED;
3219 mbmi->ref_frame[0] = ref_frame;
3220 mbmi->ref_frame[1] = second_ref_frame;
3221 // Evaluate all sub-pel filters irrespective of whether we can use
3222 // them for this frame.
3223 mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
3224 : cm->interp_filter;
3225 mbmi->mv[0].as_int = mbmi->mv[1].as_int = 0;
3228 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3230 // Select prediction reference frames.
3231 for (i = 0; i < MAX_MB_PLANE; i++) {
3232 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3234 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3237 if (ref_frame == INTRA_FRAME) {
3239 struct macroblockd_plane *const pd = &xd->plane[1];
3240 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
3241 super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable,
3242 NULL, bsize, best_rd);
3243 if (rate_y == INT_MAX)
3246 uv_tx = get_uv_tx_size_impl(mbmi->tx_size, bsize, pd->subsampling_x,
3248 if (rate_uv_intra[uv_tx] == INT_MAX) {
3249 choose_intra_uv_mode(cpi, x, ctx, bsize, uv_tx,
3250 &rate_uv_intra[uv_tx], &rate_uv_tokenonly[uv_tx],
3251 &dist_uv[uv_tx], &skip_uv[uv_tx], &mode_uv[uv_tx]);
3254 rate_uv = rate_uv_tokenonly[uv_tx];
3255 distortion_uv = dist_uv[uv_tx];
3256 skippable = skippable && skip_uv[uv_tx];
3257 mbmi->uv_mode = mode_uv[uv_tx];
3259 rate2 = rate_y + cpi->mbmode_cost[mbmi->mode] + rate_uv_intra[uv_tx];
3260 if (this_mode != DC_PRED && this_mode != TM_PRED)
3261 rate2 += intra_cost_penalty;
3262 distortion2 = distortion_y + distortion_uv;
3264 this_rd = handle_inter_mode(cpi, x, bsize,
3265 &rate2, &distortion2, &skippable,
3267 &disable_skip, frame_mv,
3269 single_newmv, single_inter_filter,
3270 single_skippable, &total_sse, best_rd,
3271 &mask_filter, filter_cache);
3272 if (this_rd == INT64_MAX)
3275 compmode_cost = vp10_cost_bit(comp_mode_p, comp_pred);
3277 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3278 rate2 += compmode_cost;
3281 // Estimate the reference frame signaling cost and add it
3282 // to the rolling cost variable.
3284 rate2 += ref_costs_comp[ref_frame];
3286 rate2 += ref_costs_single[ref_frame];
3289 if (!disable_skip) {
3291 // Back out the coefficient coding costs
3292 rate2 -= (rate_y + rate_uv);
3294 // Cost the skip mb case
3295 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 1);
3296 } else if (ref_frame != INTRA_FRAME && !xd->lossless[mbmi->segment_id]) {
3297 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
3298 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
3299 // Add in the cost of the no skip flag.
3300 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 0);
3302 // FIXME(rbultje) make this work for splitmv also
3303 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 1);
3304 distortion2 = total_sse;
3305 assert(total_sse >= 0);
3306 rate2 -= (rate_y + rate_uv);
3310 // Add in the cost of the no skip flag.
3311 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 0);
3314 // Calculate the final RD estimate for this mode.
3315 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3318 // Apply an adjustment to the rd value based on the similarity of the
3319 // source variance and reconstructed variance.
3320 rd_variance_adjustment(cpi, x, bsize, &this_rd,
3321 ref_frame, x->source_variance);
3323 if (ref_frame == INTRA_FRAME) {
3324 // Keep record of best intra rd
3325 if (this_rd < best_intra_rd) {
3326 best_intra_rd = this_rd;
3327 best_intra_mode = mbmi->mode;
3331 if (!disable_skip && ref_frame == INTRA_FRAME) {
3332 for (i = 0; i < REFERENCE_MODES; ++i)
3333 best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
3334 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3335 best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
3338 // Did this mode help.. i.e. is it the new best mode
3339 if (this_rd < best_rd || x->skip) {
3340 int max_plane = MAX_MB_PLANE;
3341 if (!mode_excluded) {
3342 // Note index of best mode so far
3343 best_mode_index = mode_index;
3345 if (ref_frame == INTRA_FRAME) {
3346 /* required for left and above block mv */
3347 mbmi->mv[0].as_int = 0;
3350 best_pred_sse = x->pred_sse[ref_frame];
3353 rd_cost->rate = rate2;
3354 rd_cost->dist = distortion2;
3355 rd_cost->rdcost = this_rd;
3357 best_mbmode = *mbmi;
3358 best_skip2 = this_skip2;
3359 best_mode_skippable = skippable;
3361 if (!x->select_tx_size)
3362 swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
3363 memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size],
3364 sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
3366 // TODO(debargha): enhance this test with a better distortion prediction
3367 // based on qp, activity mask and history
3368 if ((mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
3369 (mode_index > MIN_EARLY_TERM_INDEX)) {
3370 int qstep = xd->plane[0].dequant[1];
3371 // TODO(debargha): Enhance this by specializing for each mode_index
3373 #if CONFIG_VP9_HIGHBITDEPTH
3374 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
3375 qstep >>= (xd->bd - 8);
3377 #endif // CONFIG_VP9_HIGHBITDEPTH
3378 if (x->source_variance < UINT_MAX) {
3379 const int var_adjust = (x->source_variance < 16);
3380 scale -= var_adjust;
3382 if (ref_frame > INTRA_FRAME &&
3383 distortion2 * scale < qstep * qstep) {
3390 /* keep record of best compound/single-only prediction */
3391 if (!disable_skip && ref_frame != INTRA_FRAME) {
3392 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
3394 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3395 single_rate = rate2 - compmode_cost;
3396 hybrid_rate = rate2;
3398 single_rate = rate2;
3399 hybrid_rate = rate2 + compmode_cost;
3402 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
3403 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
3406 if (single_rd < best_pred_rd[SINGLE_REFERENCE])
3407 best_pred_rd[SINGLE_REFERENCE] = single_rd;
3409 if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
3410 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
3412 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
3413 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
3415 /* keep record of best filter type */
3416 if (!mode_excluded && cm->interp_filter != BILINEAR) {
3417 int64_t ref = filter_cache[cm->interp_filter == SWITCHABLE ?
3418 SWITCHABLE_FILTERS : cm->interp_filter];
3420 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3422 if (ref == INT64_MAX)
3424 else if (filter_cache[i] == INT64_MAX)
3425 // when early termination is triggered, the encoder does not have
3426 // access to the rate-distortion cost. it only knows that the cost
3427 // should be above the maximum valid value. hence it takes the known
3428 // maximum plus an arbitrary constant as the rate-distortion cost.
3429 adj_rd = mask_filter - ref + 10;
3431 adj_rd = filter_cache[i] - ref;
3434 best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
3442 if (x->skip && !comp_pred)
3446 // The inter modes' rate costs are not calculated precisely in some cases.
3447 // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and
3448 // ZEROMV. Here, checks are added for those cases, and the mode decisions
3450 if (best_mbmode.mode == NEWMV) {
3451 const MV_REFERENCE_FRAME refs[2] = {best_mbmode.ref_frame[0],
3452 best_mbmode.ref_frame[1]};
3453 int comp_pred_mode = refs[1] > INTRA_FRAME;
3455 if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3456 ((comp_pred_mode && frame_mv[NEARESTMV][refs[1]].as_int ==
3457 best_mbmode.mv[1].as_int) || !comp_pred_mode))
3458 best_mbmode.mode = NEARESTMV;
3459 else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3460 ((comp_pred_mode && frame_mv[NEARMV][refs[1]].as_int ==
3461 best_mbmode.mv[1].as_int) || !comp_pred_mode))
3462 best_mbmode.mode = NEARMV;
3463 else if (best_mbmode.mv[0].as_int == 0 &&
3464 ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) || !comp_pred_mode))
3465 best_mbmode.mode = ZEROMV;
3468 if (best_mode_index < 0 || best_rd >= best_rd_so_far) {
3469 rd_cost->rate = INT_MAX;
3470 rd_cost->rdcost = INT64_MAX;
3474 // If we used an estimate for the uv intra rd in the loop above...
3475 if (sf->use_uv_intra_rd_estimate) {
3476 // Do Intra UV best rd mode selection if best mode choice above was intra.
3477 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
3479 *mbmi = best_mbmode;
3480 uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
3481 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
3482 &rate_uv_tokenonly[uv_tx_size],
3483 &dist_uv[uv_tx_size],
3484 &skip_uv[uv_tx_size],
3485 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
3490 assert((cm->interp_filter == SWITCHABLE) ||
3491 (cm->interp_filter == best_mbmode.interp_filter) ||
3492 !is_inter_block(&best_mbmode));
3494 if (!cpi->rc.is_src_frame_alt_ref)
3495 vp10_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3496 sf->adaptive_rd_thresh, bsize, best_mode_index);
3499 *mbmi = best_mbmode;
3500 x->skip |= best_skip2;
3502 for (i = 0; i < REFERENCE_MODES; ++i) {
3503 if (best_pred_rd[i] == INT64_MAX)
3504 best_pred_diff[i] = INT_MIN;
3506 best_pred_diff[i] = best_rd - best_pred_rd[i];
3510 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3511 if (best_filter_rd[i] == INT64_MAX)
3512 best_filter_diff[i] = 0;
3514 best_filter_diff[i] = best_rd - best_filter_rd[i];
3516 if (cm->interp_filter == SWITCHABLE)
3517 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
3519 vp10_zero(best_filter_diff);
3522 // TODO(yunqingwang): Moving this line in front of the above best_filter_diff
3523 // updating code causes PSNR loss. Need to figure out the confliction.
3524 x->skip |= best_mode_skippable;
3526 if (!x->skip && !x->select_tx_size) {
3527 int has_high_freq_coeff = 0;
3529 int max_plane = is_inter_block(&xd->mi[0]->mbmi)
3531 for (plane = 0; plane < max_plane; ++plane) {
3532 x->plane[plane].eobs = ctx->eobs_pbuf[plane][1];
3533 has_high_freq_coeff |= vp10_has_high_freq_in_plane(x, bsize, plane);
3536 for (plane = max_plane; plane < MAX_MB_PLANE; ++plane) {
3537 x->plane[plane].eobs = ctx->eobs_pbuf[plane][2];
3538 has_high_freq_coeff |= vp10_has_high_freq_in_plane(x, bsize, plane);
3541 best_mode_skippable |= !has_high_freq_coeff;
3544 assert(best_mode_index >= 0);
3546 store_coding_context(x, ctx, best_mode_index, best_pred_diff,
3547 best_filter_diff, best_mode_skippable);
3550 void vp10_rd_pick_inter_mode_sb_seg_skip(VP10_COMP *cpi,
3551 TileDataEnc *tile_data,
3555 PICK_MODE_CONTEXT *ctx,
3556 int64_t best_rd_so_far) {
3557 VP10_COMMON *const cm = &cpi->common;
3558 MACROBLOCKD *const xd = &x->e_mbd;
3559 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
3560 unsigned char segment_id = mbmi->segment_id;
3561 const int comp_pred = 0;
3563 int64_t best_pred_diff[REFERENCE_MODES];
3564 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3565 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3566 vpx_prob comp_mode_p;
3567 INTERP_FILTER best_filter = SWITCHABLE;
3568 int64_t this_rd = INT64_MAX;
3570 const int64_t distortion2 = 0;
3572 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3575 for (i = 0; i < MAX_REF_FRAMES; ++i)
3576 x->pred_sse[i] = INT_MAX;
3577 for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i)
3578 x->pred_mv_sad[i] = INT_MAX;
3580 rd_cost->rate = INT_MAX;
3582 assert(segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
3584 mbmi->mode = ZEROMV;
3585 mbmi->uv_mode = DC_PRED;
3586 mbmi->ref_frame[0] = LAST_FRAME;
3587 mbmi->ref_frame[1] = NONE;
3588 mbmi->mv[0].as_int = 0;
3591 if (cm->interp_filter != BILINEAR) {
3592 best_filter = EIGHTTAP;
3593 if (cm->interp_filter == SWITCHABLE &&
3594 x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
3596 int best_rs = INT_MAX;
3597 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
3598 mbmi->interp_filter = i;
3599 rs = vp10_get_switchable_rate(cpi, xd);
3602 best_filter = mbmi->interp_filter;
3607 // Set the appropriate filter
3608 if (cm->interp_filter == SWITCHABLE) {
3609 mbmi->interp_filter = best_filter;
3610 rate2 += vp10_get_switchable_rate(cpi, xd);
3612 mbmi->interp_filter = cm->interp_filter;
3615 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3616 rate2 += vp10_cost_bit(comp_mode_p, comp_pred);
3618 // Estimate the reference frame signaling cost and add it
3619 // to the rolling cost variable.
3620 rate2 += ref_costs_single[LAST_FRAME];
3621 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3623 rd_cost->rate = rate2;
3624 rd_cost->dist = distortion2;
3625 rd_cost->rdcost = this_rd;
3627 if (this_rd >= best_rd_so_far) {
3628 rd_cost->rate = INT_MAX;
3629 rd_cost->rdcost = INT64_MAX;
3633 assert((cm->interp_filter == SWITCHABLE) ||
3634 (cm->interp_filter == mbmi->interp_filter));
3636 vp10_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3637 cpi->sf.adaptive_rd_thresh, bsize, THR_ZEROMV);
3639 vp10_zero(best_pred_diff);
3640 vp10_zero(best_filter_diff);
3642 if (!x->select_tx_size)
3643 swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE);
3644 store_coding_context(x, ctx, THR_ZEROMV,
3645 best_pred_diff, best_filter_diff, 0);
3648 void vp10_rd_pick_inter_mode_sub8x8(VP10_COMP *cpi,
3649 TileDataEnc *tile_data,
3651 int mi_row, int mi_col,
3654 PICK_MODE_CONTEXT *ctx,
3655 int64_t best_rd_so_far) {
3656 VP10_COMMON *const cm = &cpi->common;
3657 RD_OPT *const rd_opt = &cpi->rd;
3658 SPEED_FEATURES *const sf = &cpi->sf;
3659 MACROBLOCKD *const xd = &x->e_mbd;
3660 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
3661 const struct segmentation *const seg = &cm->seg;
3662 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3663 unsigned char segment_id = mbmi->segment_id;
3665 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3666 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3667 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
3669 int64_t best_rd = best_rd_so_far;
3670 int64_t best_yrd = best_rd_so_far; // FIXME(rbultje) more precise
3671 int64_t best_pred_diff[REFERENCE_MODES];
3672 int64_t best_pred_rd[REFERENCE_MODES];
3673 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3674 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3675 MB_MODE_INFO best_mbmode;
3676 int ref_index, best_ref_index = 0;
3677 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3678 vpx_prob comp_mode_p;
3679 INTERP_FILTER tmp_best_filter = SWITCHABLE;
3680 int rate_uv_intra, rate_uv_tokenonly;
3683 PREDICTION_MODE mode_uv = DC_PRED;
3684 const int intra_cost_penalty = vp10_get_intra_cost_penalty(
3685 cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
3686 int_mv seg_mvs[4][MAX_REF_FRAMES];
3687 b_mode_info best_bmodes[4];
3689 int ref_frame_skip_mask[2] = { 0 };
3690 int64_t mask_filter = 0;
3691 int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
3692 int internal_active_edge =
3693 vp10_active_edge_sb(cpi, mi_row, mi_col) && vp10_internal_image_edge(cpi);
3695 memset(x->zcoeff_blk[TX_4X4], 0, 4);
3696 vp10_zero(best_mbmode);
3698 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3699 filter_cache[i] = INT64_MAX;
3701 for (i = 0; i < 4; i++) {
3703 for (j = 0; j < MAX_REF_FRAMES; j++)
3704 seg_mvs[i][j].as_int = INVALID_MV;
3707 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3710 for (i = 0; i < REFERENCE_MODES; ++i)
3711 best_pred_rd[i] = INT64_MAX;
3712 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3713 best_filter_rd[i] = INT64_MAX;
3714 rate_uv_intra = INT_MAX;
3716 rd_cost->rate = INT_MAX;
3718 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
3719 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3720 setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
3721 frame_mv[NEARESTMV], frame_mv[NEARMV],
3724 ref_frame_skip_mask[0] |= (1 << ref_frame);
3725 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3727 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3728 frame_mv[ZEROMV][ref_frame].as_int = 0;
3731 for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
3732 int mode_excluded = 0;
3733 int64_t this_rd = INT64_MAX;
3734 int disable_skip = 0;
3735 int compmode_cost = 0;
3736 int rate2 = 0, rate_y = 0, rate_uv = 0;
3737 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3741 int64_t total_sse = INT_MAX;
3744 ref_frame = vp10_ref_order[ref_index].ref_frame[0];
3745 second_ref_frame = vp10_ref_order[ref_index].ref_frame[1];
3747 // Look at the reference frame of the best mode so far and set the
3748 // skip mask to look at a subset of the remaining modes.
3749 if (ref_index > 2 && sf->mode_skip_start < MAX_MODES) {
3750 if (ref_index == 3) {
3751 switch (best_mbmode.ref_frame[0]) {
3755 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME);
3756 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3759 ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME);
3760 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3763 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME);
3766 case MAX_REF_FRAMES:
3767 assert(0 && "Invalid Reference frame");
3773 if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3774 (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
3777 // Test best rd so far against threshold for trying this mode.
3778 if (!internal_active_edge &&
3779 rd_less_than_thresh(best_rd,
3780 rd_opt->threshes[segment_id][bsize][ref_index],
3781 tile_data->thresh_freq_fact[bsize][ref_index]))
3784 comp_pred = second_ref_frame > INTRA_FRAME;
3786 if (!cpi->allow_comp_inter_inter)
3788 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
3790 // Do not allow compound prediction if the segment level reference frame
3791 // feature is in use as in this case there can only be one reference.
3792 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
3795 if ((sf->mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3796 best_mbmode.ref_frame[0] == INTRA_FRAME)
3800 // TODO(jingning, jkoleszar): scaling reference frame not supported for
3802 if (ref_frame > INTRA_FRAME &&
3803 vp10_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
3806 if (second_ref_frame > INTRA_FRAME &&
3807 vp10_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf))
3811 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3812 else if (ref_frame != INTRA_FRAME)
3813 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3815 // If the segment reference frame feature is enabled....
3816 // then do nothing if the current ref frame is not allowed..
3817 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3818 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3820 // Disable this drop out case if the ref frame
3821 // segment level feature is enabled for this segment. This is to
3822 // prevent the possibility that we end up unable to pick any mode.
3823 } else if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3824 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3825 // unless ARNR filtering is enabled in which case we want
3826 // an unfiltered alternative. We allow near/nearest as well
3827 // because they may result in zero-zero MVs but be cheaper.
3828 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
3832 mbmi->tx_size = TX_4X4;
3833 mbmi->uv_mode = DC_PRED;
3834 mbmi->ref_frame[0] = ref_frame;
3835 mbmi->ref_frame[1] = second_ref_frame;
3836 // Evaluate all sub-pel filters irrespective of whether we can use
3837 // them for this frame.
3838 mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
3839 : cm->interp_filter;
3841 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3843 // Select prediction reference frames.
3844 for (i = 0; i < MAX_MB_PLANE; i++) {
3845 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3847 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3850 if (ref_frame == INTRA_FRAME) {
3852 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y,
3853 &distortion_y, best_rd) >= best_rd)
3856 rate2 += intra_cost_penalty;
3857 distortion2 += distortion_y;
3859 if (rate_uv_intra == INT_MAX) {
3860 choose_intra_uv_mode(cpi, x, ctx, bsize, TX_4X4,
3866 rate2 += rate_uv_intra;
3867 rate_uv = rate_uv_tokenonly;
3868 distortion2 += dist_uv;
3869 distortion_uv = dist_uv;
3870 mbmi->uv_mode = mode_uv;
3874 int64_t this_rd_thresh;
3875 int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
3876 int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
3877 int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
3878 int tmp_best_skippable = 0;
3879 int switchable_filter_index;
3880 int_mv *second_ref = comp_pred ?
3881 &x->mbmi_ext->ref_mvs[second_ref_frame][0] : NULL;
3882 b_mode_info tmp_best_bmodes[16];
3883 MB_MODE_INFO tmp_best_mbmode;
3884 BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
3885 int pred_exists = 0;
3888 this_rd_thresh = (ref_frame == LAST_FRAME) ?
3889 rd_opt->threshes[segment_id][bsize][THR_LAST] :
3890 rd_opt->threshes[segment_id][bsize][THR_ALTR];
3891 this_rd_thresh = (ref_frame == GOLDEN_FRAME) ?
3892 rd_opt->threshes[segment_id][bsize][THR_GOLD] : this_rd_thresh;
3893 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3894 filter_cache[i] = INT64_MAX;
3896 if (cm->interp_filter != BILINEAR) {
3897 tmp_best_filter = EIGHTTAP;
3898 if (x->source_variance < sf->disable_filter_search_var_thresh) {
3899 tmp_best_filter = EIGHTTAP;
3900 } else if (sf->adaptive_pred_interp_filter == 1 &&
3901 ctx->pred_interp_filter < SWITCHABLE) {
3902 tmp_best_filter = ctx->pred_interp_filter;
3903 } else if (sf->adaptive_pred_interp_filter == 2) {
3904 tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE ?
3905 ctx->pred_interp_filter : 0;
3907 for (switchable_filter_index = 0;
3908 switchable_filter_index < SWITCHABLE_FILTERS;
3909 ++switchable_filter_index) {
3912 MB_MODE_INFO_EXT *mbmi_ext = x->mbmi_ext;
3913 mbmi->interp_filter = switchable_filter_index;
3914 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x,
3915 &mbmi_ext->ref_mvs[ref_frame][0],
3916 second_ref, best_yrd, &rate,
3917 &rate_y, &distortion,
3918 &skippable, &total_sse,
3919 (int) this_rd_thresh, seg_mvs,
3920 bsi, switchable_filter_index,
3923 if (tmp_rd == INT64_MAX)
3925 rs = vp10_get_switchable_rate(cpi, xd);
3926 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
3927 filter_cache[switchable_filter_index] = tmp_rd;
3928 filter_cache[SWITCHABLE_FILTERS] =
3929 VPXMIN(filter_cache[SWITCHABLE_FILTERS], tmp_rd + rs_rd);
3930 if (cm->interp_filter == SWITCHABLE)
3933 mask_filter = VPXMAX(mask_filter, tmp_rd);
3935 newbest = (tmp_rd < tmp_best_rd);
3937 tmp_best_filter = mbmi->interp_filter;
3938 tmp_best_rd = tmp_rd;
3940 if ((newbest && cm->interp_filter == SWITCHABLE) ||
3941 (mbmi->interp_filter == cm->interp_filter &&
3942 cm->interp_filter != SWITCHABLE)) {
3943 tmp_best_rdu = tmp_rd;
3944 tmp_best_rate = rate;
3945 tmp_best_ratey = rate_y;
3946 tmp_best_distortion = distortion;
3947 tmp_best_sse = total_sse;
3948 tmp_best_skippable = skippable;
3949 tmp_best_mbmode = *mbmi;
3950 for (i = 0; i < 4; i++) {
3951 tmp_best_bmodes[i] = xd->mi[0]->bmi[i];
3952 x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
3955 if (switchable_filter_index == 0 &&
3956 sf->use_rd_breakout &&
3957 best_rd < INT64_MAX) {
3958 if (tmp_best_rdu / 2 > best_rd) {
3959 // skip searching the other filters if the first is
3960 // already substantially larger than the best so far
3961 tmp_best_filter = mbmi->interp_filter;
3962 tmp_best_rdu = INT64_MAX;
3967 } // switchable_filter_index loop
3971 if (tmp_best_rdu == INT64_MAX && pred_exists)
3974 mbmi->interp_filter = (cm->interp_filter == SWITCHABLE ?
3975 tmp_best_filter : cm->interp_filter);
3977 // Handles the special case when a filter that is not in the
3978 // switchable list (bilinear, 6-tap) is indicated at the frame level
3979 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x,
3980 &x->mbmi_ext->ref_mvs[ref_frame][0],
3981 second_ref, best_yrd, &rate, &rate_y,
3982 &distortion, &skippable, &total_sse,
3983 (int) this_rd_thresh, seg_mvs, bsi, 0,
3985 if (tmp_rd == INT64_MAX)
3988 total_sse = tmp_best_sse;
3989 rate = tmp_best_rate;
3990 rate_y = tmp_best_ratey;
3991 distortion = tmp_best_distortion;
3992 skippable = tmp_best_skippable;
3993 *mbmi = tmp_best_mbmode;
3994 for (i = 0; i < 4; i++)
3995 xd->mi[0]->bmi[i] = tmp_best_bmodes[i];
3999 distortion2 += distortion;
4001 if (cm->interp_filter == SWITCHABLE)
4002 rate2 += vp10_get_switchable_rate(cpi, xd);
4005 mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
4006 : cm->reference_mode == COMPOUND_REFERENCE;
4008 compmode_cost = vp10_cost_bit(comp_mode_p, comp_pred);
4010 tmp_best_rdu = best_rd -
4011 VPXMIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
4012 RDCOST(x->rdmult, x->rddiv, 0, total_sse));
4014 if (tmp_best_rdu > 0) {
4015 // If even the 'Y' rd value of split is higher than best so far
4016 // then dont bother looking at UV
4017 vp10_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col,
4019 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
4020 if (!super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
4021 &uv_sse, BLOCK_8X8, tmp_best_rdu))
4025 distortion2 += distortion_uv;
4026 skippable = skippable && uv_skippable;
4027 total_sse += uv_sse;
4031 if (cm->reference_mode == REFERENCE_MODE_SELECT)
4032 rate2 += compmode_cost;
4034 // Estimate the reference frame signaling cost and add it
4035 // to the rolling cost variable.
4036 if (second_ref_frame > INTRA_FRAME) {
4037 rate2 += ref_costs_comp[ref_frame];
4039 rate2 += ref_costs_single[ref_frame];
4042 if (!disable_skip) {
4043 // Skip is never coded at the segment level for sub8x8 blocks and instead
4044 // always coded in the bitstream at the mode info level.
4046 if (ref_frame != INTRA_FRAME && !xd->lossless[mbmi->segment_id]) {
4047 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
4048 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
4049 // Add in the cost of the no skip flag.
4050 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 0);
4052 // FIXME(rbultje) make this work for splitmv also
4053 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 1);
4054 distortion2 = total_sse;
4055 assert(total_sse >= 0);
4056 rate2 -= (rate_y + rate_uv);
4062 // Add in the cost of the no skip flag.
4063 rate2 += vp10_cost_bit(vp10_get_skip_prob(cm, xd), 0);
4066 // Calculate the final RD estimate for this mode.
4067 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
4070 if (!disable_skip && ref_frame == INTRA_FRAME) {
4071 for (i = 0; i < REFERENCE_MODES; ++i)
4072 best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
4073 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
4074 best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
4077 // Did this mode help.. i.e. is it the new best mode
4078 if (this_rd < best_rd || x->skip) {
4079 if (!mode_excluded) {
4080 int max_plane = MAX_MB_PLANE;
4081 // Note index of best mode so far
4082 best_ref_index = ref_index;
4084 if (ref_frame == INTRA_FRAME) {
4085 /* required for left and above block mv */
4086 mbmi->mv[0].as_int = 0;
4090 rd_cost->rate = rate2;
4091 rd_cost->dist = distortion2;
4092 rd_cost->rdcost = this_rd;
4094 best_yrd = best_rd -
4095 RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
4096 best_mbmode = *mbmi;
4097 best_skip2 = this_skip2;
4098 if (!x->select_tx_size)
4099 swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
4100 memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
4101 sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
4103 for (i = 0; i < 4; i++)
4104 best_bmodes[i] = xd->mi[0]->bmi[i];
4106 // TODO(debargha): enhance this test with a better distortion prediction
4107 // based on qp, activity mask and history
4108 if ((sf->mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
4109 (ref_index > MIN_EARLY_TERM_INDEX)) {
4110 int qstep = xd->plane[0].dequant[1];
4111 // TODO(debargha): Enhance this by specializing for each mode_index
4113 #if CONFIG_VP9_HIGHBITDEPTH
4114 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
4115 qstep >>= (xd->bd - 8);
4117 #endif // CONFIG_VP9_HIGHBITDEPTH
4118 if (x->source_variance < UINT_MAX) {
4119 const int var_adjust = (x->source_variance < 16);
4120 scale -= var_adjust;
4122 if (ref_frame > INTRA_FRAME &&
4123 distortion2 * scale < qstep * qstep) {
4130 /* keep record of best compound/single-only prediction */
4131 if (!disable_skip && ref_frame != INTRA_FRAME) {
4132 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
4134 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4135 single_rate = rate2 - compmode_cost;
4136 hybrid_rate = rate2;
4138 single_rate = rate2;
4139 hybrid_rate = rate2 + compmode_cost;
4142 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
4143 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
4145 if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
4146 best_pred_rd[SINGLE_REFERENCE] = single_rd;
4147 else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
4148 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
4150 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
4151 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
4154 /* keep record of best filter type */
4155 if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
4156 cm->interp_filter != BILINEAR) {
4157 int64_t ref = filter_cache[cm->interp_filter == SWITCHABLE ?
4158 SWITCHABLE_FILTERS : cm->interp_filter];
4160 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4161 if (ref == INT64_MAX)
4163 else if (filter_cache[i] == INT64_MAX)
4164 // when early termination is triggered, the encoder does not have
4165 // access to the rate-distortion cost. it only knows that the cost
4166 // should be above the maximum valid value. hence it takes the known
4167 // maximum plus an arbitrary constant as the rate-distortion cost.
4168 adj_rd = mask_filter - ref + 10;
4170 adj_rd = filter_cache[i] - ref;
4173 best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
4180 if (x->skip && !comp_pred)
4184 if (best_rd >= best_rd_so_far) {
4185 rd_cost->rate = INT_MAX;
4186 rd_cost->rdcost = INT64_MAX;
4190 // If we used an estimate for the uv intra rd in the loop above...
4191 if (sf->use_uv_intra_rd_estimate) {
4192 // Do Intra UV best rd mode selection if best mode choice above was intra.
4193 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
4194 *mbmi = best_mbmode;
4195 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra,
4203 if (best_rd == INT64_MAX) {
4204 rd_cost->rate = INT_MAX;
4205 rd_cost->dist = INT64_MAX;
4206 rd_cost->rdcost = INT64_MAX;
4210 assert((cm->interp_filter == SWITCHABLE) ||
4211 (cm->interp_filter == best_mbmode.interp_filter) ||
4212 !is_inter_block(&best_mbmode));
4214 vp10_update_rd_thresh_fact(tile_data->thresh_freq_fact,
4215 sf->adaptive_rd_thresh, bsize, best_ref_index);
4218 *mbmi = best_mbmode;
4219 x->skip |= best_skip2;
4220 if (!is_inter_block(&best_mbmode)) {
4221 for (i = 0; i < 4; i++)
4222 xd->mi[0]->bmi[i].as_mode = best_bmodes[i].as_mode;
4224 for (i = 0; i < 4; ++i)
4225 memcpy(&xd->mi[0]->bmi[i], &best_bmodes[i], sizeof(b_mode_info));
4227 mbmi->mv[0].as_int = xd->mi[0]->bmi[3].as_mv[0].as_int;
4228 mbmi->mv[1].as_int = xd->mi[0]->bmi[3].as_mv[1].as_int;
4231 for (i = 0; i < REFERENCE_MODES; ++i) {
4232 if (best_pred_rd[i] == INT64_MAX)
4233 best_pred_diff[i] = INT_MIN;
4235 best_pred_diff[i] = best_rd - best_pred_rd[i];
4239 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4240 if (best_filter_rd[i] == INT64_MAX)
4241 best_filter_diff[i] = 0;
4243 best_filter_diff[i] = best_rd - best_filter_rd[i];
4245 if (cm->interp_filter == SWITCHABLE)
4246 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
4248 vp10_zero(best_filter_diff);
4251 store_coding_context(x, ctx, best_ref_index,
4252 best_pred_diff, best_filter_diff, 0);