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 "./vp9_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 "vp9/common/vp9_common.h"
23 #include "vp9/common/vp9_entropy.h"
24 #include "vp9/common/vp9_entropymode.h"
25 #include "vp9/common/vp9_idct.h"
26 #include "vp9/common/vp9_mvref_common.h"
27 #include "vp9/common/vp9_pred_common.h"
28 #include "vp9/common/vp9_quant_common.h"
29 #include "vp9/common/vp9_reconinter.h"
30 #include "vp9/common/vp9_reconintra.h"
31 #include "vp9/common/vp9_scan.h"
32 #include "vp9/common/vp9_seg_common.h"
34 #include "vp9/encoder/vp9_cost.h"
35 #include "vp9/encoder/vp9_encodemb.h"
36 #include "vp9/encoder/vp9_encodemv.h"
37 #include "vp9/encoder/vp9_encoder.h"
38 #include "vp9/encoder/vp9_mcomp.h"
39 #include "vp9/encoder/vp9_quantize.h"
40 #include "vp9/encoder/vp9_ratectrl.h"
41 #include "vp9/encoder/vp9_rd.h"
42 #include "vp9/encoder/vp9_rdopt.h"
43 #include "vp9/encoder/vp9_aq_variance.h"
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 vp9_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 vp9_ref_order[MAX_REFS] = {
125 {{LAST_FRAME, NONE}},
126 {{GOLDEN_FRAME, NONE}},
127 {{ALTREF_FRAME, NONE}},
128 {{LAST_FRAME, ALTREF_FRAME}},
129 {{GOLDEN_FRAME, ALTREF_FRAME}},
130 {{INTRA_FRAME, NONE}},
133 static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
134 int m, int n, int min_plane, int max_plane) {
137 for (i = min_plane; i < max_plane; ++i) {
138 struct macroblock_plane *const p = &x->plane[i];
139 struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
141 p->coeff = ctx->coeff_pbuf[i][m];
142 p->qcoeff = ctx->qcoeff_pbuf[i][m];
143 pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
144 p->eobs = ctx->eobs_pbuf[i][m];
146 ctx->coeff_pbuf[i][m] = ctx->coeff_pbuf[i][n];
147 ctx->qcoeff_pbuf[i][m] = ctx->qcoeff_pbuf[i][n];
148 ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
149 ctx->eobs_pbuf[i][m] = ctx->eobs_pbuf[i][n];
151 ctx->coeff_pbuf[i][n] = p->coeff;
152 ctx->qcoeff_pbuf[i][n] = p->qcoeff;
153 ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
154 ctx->eobs_pbuf[i][n] = p->eobs;
158 static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize,
159 MACROBLOCK *x, MACROBLOCKD *xd,
160 int *out_rate_sum, int64_t *out_dist_sum,
161 int *skip_txfm_sb, int64_t *skip_sse_sb) {
162 // Note our transform coeffs are 8 times an orthogonal transform.
163 // Hence quantizer step is also 8 times. To get effective quantizer
164 // we need to divide by 8 before sending to modeling function.
166 int64_t rate_sum = 0;
167 int64_t dist_sum = 0;
168 const int ref = xd->mi[0]->ref_frame[0];
170 unsigned int var = 0;
171 unsigned int sum_sse = 0;
172 int64_t total_sse = 0;
177 const int dequant_shift =
178 #if CONFIG_VP9_HIGHBITDEPTH
179 (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ?
181 #endif // CONFIG_VP9_HIGHBITDEPTH
184 x->pred_sse[ref] = 0;
186 for (i = 0; i < MAX_MB_PLANE; ++i) {
187 struct macroblock_plane *const p = &x->plane[i];
188 struct macroblockd_plane *const pd = &xd->plane[i];
189 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
190 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
191 const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
192 const int64_t dc_thr = p->quant_thred[0] >> shift;
193 const int64_t ac_thr = p->quant_thred[1] >> shift;
194 // The low thresholds are used to measure if the prediction errors are
195 // low enough so that we can skip the mode search.
196 const int64_t low_dc_thr = VPXMIN(50, dc_thr >> 2);
197 const int64_t low_ac_thr = VPXMIN(80, ac_thr >> 2);
198 int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
199 int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
201 int lw = b_width_log2_lookup[unit_size] + 2;
202 int lh = b_height_log2_lookup[unit_size] + 2;
206 for (idy = 0; idy < bh; ++idy) {
207 for (idx = 0; idx < bw; ++idx) {
208 uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
209 uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
210 int block_idx = (idy << 1) + idx;
211 int low_err_skip = 0;
213 var = cpi->fn_ptr[unit_size].vf(src, p->src.stride,
214 dst, pd->dst.stride, &sse);
215 x->bsse[(i << 2) + block_idx] = sse;
218 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_NONE;
219 if (!x->select_tx_size) {
220 // Check if all ac coefficients can be quantized to zero.
221 if (var < ac_thr || var == 0) {
222 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_ONLY;
224 // Check if dc coefficient can be quantized to zero.
225 if (sse - var < dc_thr || sse == var) {
226 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_DC;
228 if (!sse || (var < low_ac_thr && sse - var < low_dc_thr))
234 if (skip_flag && !low_err_skip)
238 x->pred_sse[ref] += sse;
242 total_sse += sum_sse;
244 // Fast approximate the modelling function.
245 if (cpi->sf.simple_model_rd_from_var) {
247 const int64_t square_error = sum_sse;
248 int quantizer = (pd->dequant[1] >> dequant_shift);
251 rate = (square_error * (280 - quantizer)) >> (16 - VP9_PROB_COST_SHIFT);
254 dist = (square_error * quantizer) >> 8;
258 vp9_model_rd_from_var_lapndz(sum_sse, num_pels_log2_lookup[bs],
259 pd->dequant[1] >> dequant_shift,
266 *skip_txfm_sb = skip_flag;
267 *skip_sse_sb = total_sse << 4;
268 *out_rate_sum = (int)rate_sum;
269 *out_dist_sum = dist_sum << 4;
272 #if CONFIG_VP9_HIGHBITDEPTH
273 int64_t vp9_highbd_block_error_c(const tran_low_t *coeff,
274 const tran_low_t *dqcoeff,
276 int64_t *ssz, int bd) {
278 int64_t error = 0, sqcoeff = 0;
279 int shift = 2 * (bd - 8);
280 int rounding = shift > 0 ? 1 << (shift - 1) : 0;
282 for (i = 0; i < block_size; i++) {
283 const int64_t diff = coeff[i] - dqcoeff[i];
284 error += diff * diff;
285 sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
287 assert(error >= 0 && sqcoeff >= 0);
288 error = (error + rounding) >> shift;
289 sqcoeff = (sqcoeff + rounding) >> shift;
295 int64_t vp9_highbd_block_error_8bit_c(const tran_low_t *coeff,
296 const tran_low_t *dqcoeff,
299 // Note that the C versions of these 2 functions (vp9_block_error and
300 // vp9_highbd_block_error_8bit are the same, but the optimized assembly
301 // routines are not compatible in the non high bitdepth configuration, so
302 // they still cannot share the same name.
303 return vp9_block_error_c(coeff, dqcoeff, block_size, ssz);
306 static int64_t vp9_highbd_block_error_dispatch(const tran_low_t *coeff,
307 const tran_low_t *dqcoeff,
309 int64_t *ssz, int bd) {
311 return vp9_highbd_block_error_8bit(coeff, dqcoeff, block_size, ssz);
313 return vp9_highbd_block_error(coeff, dqcoeff, block_size, ssz, bd);
316 #endif // CONFIG_VP9_HIGHBITDEPTH
318 int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
319 intptr_t block_size, int64_t *ssz) {
321 int64_t error = 0, sqcoeff = 0;
323 for (i = 0; i < block_size; i++) {
324 const int diff = coeff[i] - dqcoeff[i];
325 error += diff * diff;
326 sqcoeff += coeff[i] * coeff[i];
333 int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff,
338 for (i = 0; i < block_size; i++) {
339 const int diff = coeff[i] - dqcoeff[i];
340 error += diff * diff;
346 /* The trailing '0' is a terminator which is used inside cost_coeffs() to
347 * decide whether to include cost of a trailing EOB node or not (i.e. we
348 * can skip this if the last coefficient in this transform block, e.g. the
349 * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
351 static const int16_t band_counts[TX_SIZES][8] = {
352 { 1, 2, 3, 4, 3, 16 - 13, 0 },
353 { 1, 2, 3, 4, 11, 64 - 21, 0 },
354 { 1, 2, 3, 4, 11, 256 - 21, 0 },
355 { 1, 2, 3, 4, 11, 1024 - 21, 0 },
357 static int cost_coeffs(MACROBLOCK *x,
358 int plane, int block,
359 ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L,
361 const int16_t *scan, const int16_t *nb,
362 int use_fast_coef_costing) {
363 MACROBLOCKD *const xd = &x->e_mbd;
364 MODE_INFO *mi = xd->mi[0];
365 const struct macroblock_plane *p = &x->plane[plane];
366 const PLANE_TYPE type = get_plane_type(plane);
367 const int16_t *band_count = &band_counts[tx_size][1];
368 const int eob = p->eobs[block];
369 const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
370 unsigned int (*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
371 x->token_costs[tx_size][type][is_inter_block(mi)];
372 uint8_t token_cache[32 * 32];
373 int pt = combine_entropy_contexts(*A, *L);
375 #if CONFIG_VP9_HIGHBITDEPTH
376 const int *cat6_high_cost = vp9_get_high_cost_table(xd->bd);
378 const int *cat6_high_cost = vp9_get_high_cost_table(8);
381 // Check for consistency of tx_size with mode info
382 assert(type == PLANE_TYPE_Y ? mi->tx_size == tx_size :
383 get_uv_tx_size(mi, &xd->plane[plane]) == tx_size);
387 cost = token_costs[0][0][pt][EOB_TOKEN];
390 int band_left = *band_count++;
396 vp9_get_token_extra(v, &prev_t, &e);
397 cost = (*token_costs)[0][pt][prev_t] +
398 vp9_get_cost(prev_t, e, cat6_high_cost);
400 token_cache[0] = vp9_pt_energy_class[prev_t];
404 for (c = 1; c < eob; c++) {
405 const int rc = scan[c];
409 vp9_get_token_extra(v, &t, &e);
410 if (use_fast_coef_costing) {
411 cost += (*token_costs)[!prev_t][!prev_t][t] +
412 vp9_get_cost(t, e, cat6_high_cost);
414 pt = get_coef_context(nb, token_cache, c);
415 cost += (*token_costs)[!prev_t][pt][t] +
416 vp9_get_cost(t, e, cat6_high_cost);
417 token_cache[rc] = vp9_pt_energy_class[t];
421 band_left = *band_count++;
428 if (use_fast_coef_costing) {
429 cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
431 pt = get_coef_context(nb, token_cache, c);
432 cost += (*token_costs)[0][pt][EOB_TOKEN];
437 // is eob first coefficient;
443 static void dist_block(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
444 int64_t *out_dist, int64_t *out_sse) {
445 const int ss_txfrm_size = tx_size << 1;
446 MACROBLOCKD* const xd = &x->e_mbd;
447 const struct macroblock_plane *const p = &x->plane[plane];
448 const struct macroblockd_plane *const pd = &xd->plane[plane];
450 int shift = tx_size == TX_32X32 ? 0 : 2;
451 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
452 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
453 #if CONFIG_VP9_HIGHBITDEPTH
454 const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
455 *out_dist = vp9_highbd_block_error_dispatch(coeff, dqcoeff,
457 &this_sse, bd) >> shift;
459 *out_dist = vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
461 #endif // CONFIG_VP9_HIGHBITDEPTH
462 *out_sse = this_sse >> shift;
464 if (x->skip_encode && !is_inter_block(xd->mi[0])) {
465 // TODO(jingning): tune the model to better capture the distortion.
466 int64_t p = (pd->dequant[1] * pd->dequant[1] *
467 (1 << ss_txfrm_size)) >>
468 #if CONFIG_VP9_HIGHBITDEPTH
469 (shift + 2 + (bd - 8) * 2);
472 #endif // CONFIG_VP9_HIGHBITDEPTH
473 *out_dist += (p >> 4);
478 static int rate_block(int plane, int block, BLOCK_SIZE plane_bsize,
479 TX_SIZE tx_size, struct rdcost_block_args* args) {
481 txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x_idx, &y_idx);
483 return cost_coeffs(args->x, plane, block, args->t_above + x_idx,
484 args->t_left + y_idx, tx_size,
485 args->so->scan, args->so->neighbors,
486 args->use_fast_coef_costing);
489 static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
490 TX_SIZE tx_size, void *arg) {
491 struct rdcost_block_args *args = arg;
492 MACROBLOCK *const x = args->x;
493 MACROBLOCKD *const xd = &x->e_mbd;
494 MODE_INFO *const mi = xd->mi[0];
495 int64_t rd1, rd2, rd;
500 if (args->exit_early)
503 if (!is_inter_block(mi)) {
504 struct encode_b_args arg = {x, NULL, &mi->skip};
505 vp9_encode_block_intra(plane, block, plane_bsize, tx_size, &arg);
506 dist_block(x, plane, block, tx_size, &dist, &sse);
507 } else if (max_txsize_lookup[plane_bsize] == tx_size) {
508 if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
510 // full forward transform and quantization
511 vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
512 dist_block(x, plane, block, tx_size, &dist, &sse);
513 } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
515 // compute DC coefficient
516 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[plane].coeff, block);
517 tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
518 vp9_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
519 sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
521 if (x->plane[plane].eobs[block]) {
522 const int64_t orig_sse = (int64_t)coeff[0] * coeff[0];
523 const int64_t resd_sse = coeff[0] - dqcoeff[0];
524 int64_t dc_correct = orig_sse - resd_sse * resd_sse;
525 #if CONFIG_VP9_HIGHBITDEPTH
526 dc_correct >>= ((xd->bd - 8) * 2);
528 if (tx_size != TX_32X32)
531 dist = VPXMAX(0, sse - dc_correct);
535 // skip forward transform
536 x->plane[plane].eobs[block] = 0;
537 sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
541 // full forward transform and quantization
542 vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
543 dist_block(x, plane, block, tx_size, &dist, &sse);
546 rd = RDCOST(x->rdmult, x->rddiv, 0, dist);
547 if (args->this_rd + rd > args->best_rd) {
548 args->exit_early = 1;
552 rate = rate_block(plane, block, plane_bsize, tx_size, args);
553 rd1 = RDCOST(x->rdmult, x->rddiv, rate, dist);
554 rd2 = RDCOST(x->rdmult, x->rddiv, 0, sse);
556 // TODO(jingning): temporarily enabled only for luma component
557 rd = VPXMIN(rd1, rd2);
559 x->zcoeff_blk[tx_size][block] = !x->plane[plane].eobs[block] ||
560 (rd1 > rd2 && !xd->lossless);
562 args->this_rate += rate;
563 args->this_dist += dist;
564 args->this_sse += sse;
567 if (args->this_rd > args->best_rd) {
568 args->exit_early = 1;
572 args->skippable &= !x->plane[plane].eobs[block];
575 static void txfm_rd_in_plane(MACROBLOCK *x,
576 int *rate, int64_t *distortion,
577 int *skippable, int64_t *sse,
578 int64_t ref_best_rd, int plane,
579 BLOCK_SIZE bsize, TX_SIZE tx_size,
580 int use_fast_coef_casting) {
581 MACROBLOCKD *const xd = &x->e_mbd;
582 const struct macroblockd_plane *const pd = &xd->plane[plane];
583 struct rdcost_block_args args;
586 args.best_rd = ref_best_rd;
587 args.use_fast_coef_costing = use_fast_coef_casting;
591 xd->mi[0]->tx_size = tx_size;
593 vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
595 args.so = get_scan(xd, tx_size, get_plane_type(plane), 0);
597 vp9_foreach_transformed_block_in_plane(xd, bsize, plane,
598 block_rd_txfm, &args);
599 if (args.exit_early) {
601 *distortion = INT64_MAX;
605 *distortion = args.this_dist;
606 *rate = args.this_rate;
607 *sse = args.this_sse;
608 *skippable = args.skippable;
612 static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x,
613 int *rate, int64_t *distortion,
614 int *skip, int64_t *sse,
617 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
618 VP9_COMMON *const cm = &cpi->common;
619 const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
620 MACROBLOCKD *const xd = &x->e_mbd;
621 MODE_INFO *const mi = xd->mi[0];
623 mi->tx_size = VPXMIN(max_tx_size, largest_tx_size);
625 txfm_rd_in_plane(x, rate, distortion, skip,
626 sse, ref_best_rd, 0, bs,
627 mi->tx_size, cpi->sf.use_fast_coef_costing);
630 static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
637 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
638 VP9_COMMON *const cm = &cpi->common;
639 MACROBLOCKD *const xd = &x->e_mbd;
640 MODE_INFO *const mi = xd->mi[0];
641 vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
642 int r[TX_SIZES][2], s[TX_SIZES];
643 int64_t d[TX_SIZES], sse[TX_SIZES];
644 int64_t rd[TX_SIZES][2] = {{INT64_MAX, INT64_MAX},
645 {INT64_MAX, INT64_MAX},
646 {INT64_MAX, INT64_MAX},
647 {INT64_MAX, INT64_MAX}};
650 int64_t best_rd = INT64_MAX;
651 TX_SIZE best_tx = max_tx_size;
652 int start_tx, end_tx;
654 const vpx_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc->tx_probs);
655 assert(skip_prob > 0);
656 s0 = vp9_cost_bit(skip_prob, 0);
657 s1 = vp9_cost_bit(skip_prob, 1);
659 if (cm->tx_mode == TX_MODE_SELECT) {
660 start_tx = max_tx_size;
663 TX_SIZE chosen_tx_size = VPXMIN(max_tx_size,
664 tx_mode_to_biggest_tx_size[cm->tx_mode]);
665 start_tx = chosen_tx_size;
666 end_tx = chosen_tx_size;
669 for (n = start_tx; n >= end_tx; n--) {
671 for (m = 0; m <= n - (n == (int) max_tx_size); m++) {
673 r_tx_size += vp9_cost_zero(tx_probs[m]);
675 r_tx_size += vp9_cost_one(tx_probs[m]);
677 txfm_rd_in_plane(x, &r[n][0], &d[n], &s[n],
678 &sse[n], ref_best_rd, 0, bs, n,
679 cpi->sf.use_fast_coef_costing);
681 if (r[n][0] < INT_MAX) {
682 r[n][1] += r_tx_size;
684 if (d[n] == INT64_MAX || r[n][0] == INT_MAX) {
685 rd[n][0] = rd[n][1] = INT64_MAX;
687 if (is_inter_block(mi)) {
688 rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
689 r[n][1] -= r_tx_size;
691 rd[n][0] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
692 rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1 + r_tx_size, sse[n]);
695 rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
696 rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
699 if (is_inter_block(mi) && !xd->lossless && !s[n] && sse[n] != INT64_MAX) {
700 rd[n][0] = VPXMIN(rd[n][0], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
701 rd[n][1] = VPXMIN(rd[n][1], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
704 // Early termination in transform size search.
705 if (cpi->sf.tx_size_search_breakout &&
706 (rd[n][1] == INT64_MAX ||
707 (n < (int) max_tx_size && rd[n][1] > rd[n + 1][1]) ||
711 if (rd[n][1] < best_rd) {
716 mi->tx_size = best_tx;
718 *distortion = d[mi->tx_size];
719 *rate = r[mi->tx_size][cm->tx_mode == TX_MODE_SELECT];
720 *skip = s[mi->tx_size];
721 *psse = sse[mi->tx_size];
724 static void super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
725 int64_t *distortion, int *skip,
726 int64_t *psse, BLOCK_SIZE bs,
727 int64_t ref_best_rd) {
728 MACROBLOCKD *xd = &x->e_mbd;
730 int64_t *ret_sse = psse ? psse : &sse;
732 assert(bs == xd->mi[0]->sb_type);
734 if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
735 choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
738 choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse,
743 static int conditional_skipintra(PREDICTION_MODE mode,
744 PREDICTION_MODE best_intra_mode) {
745 if (mode == D117_PRED &&
746 best_intra_mode != V_PRED &&
747 best_intra_mode != D135_PRED)
749 if (mode == D63_PRED &&
750 best_intra_mode != V_PRED &&
751 best_intra_mode != D45_PRED)
753 if (mode == D207_PRED &&
754 best_intra_mode != H_PRED &&
755 best_intra_mode != D45_PRED)
757 if (mode == D153_PRED &&
758 best_intra_mode != H_PRED &&
759 best_intra_mode != D135_PRED)
764 static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x,
766 PREDICTION_MODE *best_mode,
767 const int *bmode_costs,
768 ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
769 int *bestrate, int *bestratey,
770 int64_t *bestdistortion,
771 BLOCK_SIZE bsize, int64_t rd_thresh) {
772 PREDICTION_MODE mode;
773 MACROBLOCKD *const xd = &x->e_mbd;
774 int64_t best_rd = rd_thresh;
775 struct macroblock_plane *p = &x->plane[0];
776 struct macroblockd_plane *pd = &xd->plane[0];
777 const int src_stride = p->src.stride;
778 const int dst_stride = pd->dst.stride;
779 const uint8_t *src_init = &p->src.buf[row * 4 * src_stride + col * 4];
780 uint8_t *dst_init = &pd->dst.buf[row * 4 * src_stride + col * 4];
781 ENTROPY_CONTEXT ta[2], tempa[2];
782 ENTROPY_CONTEXT tl[2], templ[2];
783 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
784 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
786 uint8_t best_dst[8 * 8];
787 #if CONFIG_VP9_HIGHBITDEPTH
788 uint16_t best_dst16[8 * 8];
791 memcpy(ta, a, sizeof(ta));
792 memcpy(tl, l, sizeof(tl));
793 xd->mi[0]->tx_size = TX_4X4;
795 #if CONFIG_VP9_HIGHBITDEPTH
796 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
797 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
800 int64_t distortion = 0;
801 int rate = bmode_costs[mode];
803 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
806 // Only do the oblique modes if the best so far is
807 // one of the neighboring directional modes
808 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
809 if (conditional_skipintra(mode, *best_mode))
813 memcpy(tempa, ta, sizeof(ta));
814 memcpy(templ, tl, sizeof(tl));
816 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
817 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
818 const int block = (row + idy) * 2 + (col + idx);
819 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
820 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
821 int16_t *const src_diff = vp9_raster_block_offset_int16(BLOCK_8X8,
824 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
825 xd->mi[0]->bmi[block].as_mode = mode;
826 vp9_predict_intra_block(xd, 1, TX_4X4, mode,
827 x->skip_encode ? src : dst,
828 x->skip_encode ? src_stride : dst_stride,
830 col + idx, row + idy, 0);
831 vpx_highbd_subtract_block(4, 4, src_diff, 8, src, src_stride,
832 dst, dst_stride, xd->bd);
834 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
835 vp9_highbd_fwht4x4(src_diff, coeff, 8);
836 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
837 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
838 so->scan, so->neighbors,
839 cpi->sf.use_fast_coef_costing);
840 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
842 vp9_highbd_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block),
844 p->eobs[block], xd->bd);
847 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
848 const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
849 if (tx_type == DCT_DCT)
850 vpx_highbd_fdct4x4(src_diff, coeff, 8);
852 vp9_highbd_fht4x4(src_diff, coeff, 8, tx_type);
853 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
854 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
855 so->scan, so->neighbors,
856 cpi->sf.use_fast_coef_costing);
857 distortion += vp9_highbd_block_error_dispatch(
858 coeff, BLOCK_OFFSET(pd->dqcoeff, block),
859 16, &unused, xd->bd) >> 2;
860 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
862 vp9_highbd_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
863 dst, dst_stride, p->eobs[block], xd->bd);
869 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
871 if (this_rd < best_rd) {
874 *bestdistortion = distortion;
877 memcpy(a, tempa, sizeof(tempa));
878 memcpy(l, templ, sizeof(templ));
879 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
880 memcpy(best_dst16 + idy * 8,
881 CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
882 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
888 if (best_rd >= rd_thresh || x->skip_encode)
891 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
892 memcpy(CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
893 best_dst16 + idy * 8,
894 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
899 #endif // CONFIG_VP9_HIGHBITDEPTH
901 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
904 int64_t distortion = 0;
905 int rate = bmode_costs[mode];
907 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
910 // Only do the oblique modes if the best so far is
911 // one of the neighboring directional modes
912 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
913 if (conditional_skipintra(mode, *best_mode))
917 memcpy(tempa, ta, sizeof(ta));
918 memcpy(templ, tl, sizeof(tl));
920 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
921 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
922 const int block = (row + idy) * 2 + (col + idx);
923 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
924 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
925 int16_t *const src_diff =
926 vp9_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
927 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
928 xd->mi[0]->bmi[block].as_mode = mode;
929 vp9_predict_intra_block(xd, 1, TX_4X4, mode,
930 x->skip_encode ? src : dst,
931 x->skip_encode ? src_stride : dst_stride,
932 dst, dst_stride, col + idx, row + idy, 0);
933 vpx_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
936 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
937 vp9_fwht4x4(src_diff, coeff, 8);
938 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
939 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
940 so->scan, so->neighbors,
941 cpi->sf.use_fast_coef_costing);
942 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
944 vp9_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride,
948 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
949 const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
950 vp9_fht4x4(src_diff, coeff, 8, tx_type);
951 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
952 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
953 so->scan, so->neighbors,
954 cpi->sf.use_fast_coef_costing);
955 #if CONFIG_VP9_HIGHBITDEPTH
956 distortion += vp9_highbd_block_error_8bit(
957 coeff, BLOCK_OFFSET(pd->dqcoeff, block), 16, &unused) >> 2;
959 distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
962 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
964 vp9_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
965 dst, dst_stride, p->eobs[block]);
971 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
973 if (this_rd < best_rd) {
976 *bestdistortion = distortion;
979 memcpy(a, tempa, sizeof(tempa));
980 memcpy(l, templ, sizeof(templ));
981 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
982 memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
983 num_4x4_blocks_wide * 4);
989 if (best_rd >= rd_thresh || x->skip_encode)
992 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
993 memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
994 num_4x4_blocks_wide * 4);
999 static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
1000 int *rate, int *rate_y,
1001 int64_t *distortion,
1004 const MACROBLOCKD *const xd = &mb->e_mbd;
1005 MODE_INFO *const mic = xd->mi[0];
1006 const MODE_INFO *above_mi = xd->above_mi;
1007 const MODE_INFO *left_mi = xd->left_mi;
1008 const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
1009 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1010 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1013 int64_t total_distortion = 0;
1015 int64_t total_rd = 0;
1016 ENTROPY_CONTEXT t_above[4], t_left[4];
1017 const int *bmode_costs = cpi->mbmode_cost;
1019 memcpy(t_above, xd->plane[0].above_context, sizeof(t_above));
1020 memcpy(t_left, xd->plane[0].left_context, sizeof(t_left));
1022 // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
1023 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1024 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1025 PREDICTION_MODE best_mode = DC_PRED;
1026 int r = INT_MAX, ry = INT_MAX;
1027 int64_t d = INT64_MAX, this_rd = INT64_MAX;
1029 if (cpi->common.frame_type == KEY_FRAME) {
1030 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i);
1031 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i);
1033 bmode_costs = cpi->y_mode_costs[A][L];
1036 this_rd = rd_pick_intra4x4block(cpi, mb, idy, idx, &best_mode,
1037 bmode_costs, t_above + idx, t_left + idy,
1038 &r, &ry, &d, bsize, best_rd - total_rd);
1039 if (this_rd >= best_rd - total_rd)
1042 total_rd += this_rd;
1044 total_distortion += d;
1047 mic->bmi[i].as_mode = best_mode;
1048 for (j = 1; j < num_4x4_blocks_high; ++j)
1049 mic->bmi[i + j * 2].as_mode = best_mode;
1050 for (j = 1; j < num_4x4_blocks_wide; ++j)
1051 mic->bmi[i + j].as_mode = best_mode;
1053 if (total_rd >= best_rd)
1059 *rate_y = tot_rate_y;
1060 *distortion = total_distortion;
1061 mic->mode = mic->bmi[3].as_mode;
1063 return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
1066 // This function is used only for intra_only frames
1067 static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x,
1068 int *rate, int *rate_tokenonly,
1069 int64_t *distortion, int *skippable,
1072 PREDICTION_MODE mode;
1073 PREDICTION_MODE mode_selected = DC_PRED;
1074 MACROBLOCKD *const xd = &x->e_mbd;
1075 MODE_INFO *const mic = xd->mi[0];
1076 int this_rate, this_rate_tokenonly, s;
1077 int64_t this_distortion, this_rd;
1078 TX_SIZE best_tx = TX_4X4;
1080 const MODE_INFO *above_mi = xd->above_mi;
1081 const MODE_INFO *left_mi = xd->left_mi;
1082 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1083 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1084 bmode_costs = cpi->y_mode_costs[A][L];
1086 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1087 /* Y Search for intra prediction mode */
1088 for (mode = DC_PRED; mode <= TM_PRED; mode++) {
1089 if (cpi->sf.use_nonrd_pick_mode) {
1090 // These speed features are turned on in hybrid non-RD and RD mode
1091 // for key frame coding in the context of real-time setting.
1092 if (conditional_skipintra(mode, mode_selected))
1100 super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion,
1101 &s, NULL, bsize, best_rd);
1103 if (this_rate_tokenonly == INT_MAX)
1106 this_rate = this_rate_tokenonly + bmode_costs[mode];
1107 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1109 if (this_rd < best_rd) {
1110 mode_selected = mode;
1112 best_tx = mic->tx_size;
1114 *rate_tokenonly = this_rate_tokenonly;
1115 *distortion = this_distortion;
1120 mic->mode = mode_selected;
1121 mic->tx_size = best_tx;
1126 // Return value 0: early termination triggered, no valid rd cost available;
1127 // 1: rd cost values are valid.
1128 static int super_block_uvrd(const VP9_COMP *cpi, MACROBLOCK *x,
1129 int *rate, int64_t *distortion, int *skippable,
1130 int64_t *sse, BLOCK_SIZE bsize,
1131 int64_t ref_best_rd) {
1132 MACROBLOCKD *const xd = &x->e_mbd;
1133 MODE_INFO *const mi = xd->mi[0];
1134 const TX_SIZE uv_tx_size = get_uv_tx_size(mi, &xd->plane[1]);
1136 int pnrate = 0, pnskip = 1;
1137 int64_t pndist = 0, pnsse = 0;
1138 int is_cost_valid = 1;
1140 if (ref_best_rd < 0)
1143 if (is_inter_block(mi) && is_cost_valid) {
1145 for (plane = 1; plane < MAX_MB_PLANE; ++plane)
1146 vp9_subtract_plane(x, bsize, plane);
1154 for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
1155 txfm_rd_in_plane(x, &pnrate, &pndist, &pnskip, &pnsse,
1156 ref_best_rd, plane, bsize, uv_tx_size,
1157 cpi->sf.use_fast_coef_costing);
1158 if (pnrate == INT_MAX) {
1163 *distortion += pndist;
1165 *skippable &= pnskip;
1168 if (!is_cost_valid) {
1171 *distortion = INT64_MAX;
1176 return is_cost_valid;
1179 static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
1180 PICK_MODE_CONTEXT *ctx,
1181 int *rate, int *rate_tokenonly,
1182 int64_t *distortion, int *skippable,
1183 BLOCK_SIZE bsize, TX_SIZE max_tx_size) {
1184 MACROBLOCKD *xd = &x->e_mbd;
1185 PREDICTION_MODE mode;
1186 PREDICTION_MODE mode_selected = DC_PRED;
1187 int64_t best_rd = INT64_MAX, this_rd;
1188 int this_rate_tokenonly, this_rate, s;
1189 int64_t this_distortion, this_sse;
1191 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1192 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1193 if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode)))
1196 xd->mi[0]->uv_mode = mode;
1198 if (!super_block_uvrd(cpi, x, &this_rate_tokenonly,
1199 &this_distortion, &s, &this_sse, bsize, best_rd))
1201 this_rate = this_rate_tokenonly +
1202 cpi->intra_uv_mode_cost[cpi->common.frame_type]
1203 [xd->mi[0]->mode][mode];
1204 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1206 if (this_rd < best_rd) {
1207 mode_selected = mode;
1210 *rate_tokenonly = this_rate_tokenonly;
1211 *distortion = this_distortion;
1213 if (!x->select_tx_size)
1214 swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
1218 xd->mi[0]->uv_mode = mode_selected;
1222 static int64_t rd_sbuv_dcpred(const VP9_COMP *cpi, MACROBLOCK *x,
1223 int *rate, int *rate_tokenonly,
1224 int64_t *distortion, int *skippable,
1226 const VP9_COMMON *cm = &cpi->common;
1229 x->e_mbd.mi[0]->uv_mode = DC_PRED;
1230 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1231 super_block_uvrd(cpi, x, rate_tokenonly, distortion,
1232 skippable, &unused, bsize, INT64_MAX);
1233 *rate = *rate_tokenonly +
1234 cpi->intra_uv_mode_cost[cm->frame_type]
1235 [x->e_mbd.mi[0]->mode][DC_PRED];
1236 return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
1239 static void choose_intra_uv_mode(VP9_COMP *cpi, MACROBLOCK *const x,
1240 PICK_MODE_CONTEXT *ctx,
1241 BLOCK_SIZE bsize, TX_SIZE max_tx_size,
1242 int *rate_uv, int *rate_uv_tokenonly,
1243 int64_t *dist_uv, int *skip_uv,
1244 PREDICTION_MODE *mode_uv) {
1245 // Use an estimated rd for uv_intra based on DC_PRED if the
1246 // appropriate speed flag is set.
1247 if (cpi->sf.use_uv_intra_rd_estimate) {
1248 rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv,
1249 skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
1250 // Else do a proper rd search for each possible transform size that may
1251 // be considered in the main rd loop.
1253 rd_pick_intra_sbuv_mode(cpi, x, ctx,
1254 rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
1255 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize, max_tx_size);
1257 *mode_uv = x->e_mbd.mi[0]->uv_mode;
1260 static int cost_mv_ref(const VP9_COMP *cpi, PREDICTION_MODE mode,
1262 assert(is_inter_mode(mode));
1263 return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
1266 static int set_and_cost_bmi_mvs(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
1268 PREDICTION_MODE mode, int_mv this_mv[2],
1269 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1270 int_mv seg_mvs[MAX_REF_FRAMES],
1271 int_mv *best_ref_mv[2], const int *mvjcost,
1273 MODE_INFO *const mi = xd->mi[0];
1274 const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1277 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mi->sb_type];
1278 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mi->sb_type];
1279 const int is_compound = has_second_ref(mi);
1283 this_mv[0].as_int = seg_mvs[mi->ref_frame[0]].as_int;
1284 thismvcost += vp9_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
1285 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1287 this_mv[1].as_int = seg_mvs[mi->ref_frame[1]].as_int;
1288 thismvcost += vp9_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
1289 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1294 this_mv[0].as_int = frame_mv[mode][mi->ref_frame[0]].as_int;
1296 this_mv[1].as_int = frame_mv[mode][mi->ref_frame[1]].as_int;
1299 this_mv[0].as_int = 0;
1301 this_mv[1].as_int = 0;
1307 mi->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
1309 mi->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
1311 mi->bmi[i].as_mode = mode;
1313 for (idy = 0; idy < num_4x4_blocks_high; ++idy)
1314 for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
1315 memmove(&mi->bmi[i + idy * 2 + idx], &mi->bmi[i], sizeof(mi->bmi[i]));
1317 return cost_mv_ref(cpi, mode, mbmi_ext->mode_context[mi->ref_frame[0]]) +
1321 static int64_t encode_inter_mb_segment(VP9_COMP *cpi,
1326 int64_t *distortion, int64_t *sse,
1327 ENTROPY_CONTEXT *ta,
1328 ENTROPY_CONTEXT *tl,
1329 int mi_row, int mi_col) {
1331 MACROBLOCKD *xd = &x->e_mbd;
1332 struct macroblockd_plane *const pd = &xd->plane[0];
1333 struct macroblock_plane *const p = &x->plane[0];
1334 MODE_INFO *const mi = xd->mi[0];
1335 const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->sb_type, pd);
1336 const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
1337 const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
1340 const uint8_t *const src =
1341 &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1342 uint8_t *const dst = &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i,
1344 int64_t thisdistortion = 0, thissse = 0;
1345 int thisrate = 0, ref;
1346 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1347 const int is_compound = has_second_ref(mi);
1348 const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
1350 for (ref = 0; ref < 1 + is_compound; ++ref) {
1351 const int bw = b_width_log2_lookup[BLOCK_8X8];
1352 const int h = 4 * (i >> bw);
1353 const int w = 4 * (i & ((1 << bw) - 1));
1354 const struct scale_factors *sf = &xd->block_refs[ref]->sf;
1355 int y_stride = pd->pre[ref].stride;
1356 uint8_t *pre = pd->pre[ref].buf + (h * pd->pre[ref].stride + w);
1358 if (vp9_is_scaled(sf)) {
1359 const int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
1360 const int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
1362 y_stride = xd->block_refs[ref]->buf->y_stride;
1363 pre = xd->block_refs[ref]->buf->y_buffer;
1364 pre += scaled_buffer_offset(x_start + w, y_start + h,
1367 #if CONFIG_VP9_HIGHBITDEPTH
1368 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1369 vp9_highbd_build_inter_predictor(pre, y_stride,
1370 dst, pd->dst.stride,
1371 &mi->bmi[i].as_mv[ref].as_mv,
1372 &xd->block_refs[ref]->sf, width, height,
1373 ref, kernel, MV_PRECISION_Q3,
1374 mi_col * MI_SIZE + 4 * (i % 2),
1375 mi_row * MI_SIZE + 4 * (i / 2), xd->bd);
1377 vp9_build_inter_predictor(pre, y_stride,
1378 dst, pd->dst.stride,
1379 &mi->bmi[i].as_mv[ref].as_mv,
1380 &xd->block_refs[ref]->sf, width, height, ref,
1381 kernel, MV_PRECISION_Q3,
1382 mi_col * MI_SIZE + 4 * (i % 2),
1383 mi_row * MI_SIZE + 4 * (i / 2));
1386 vp9_build_inter_predictor(pre, y_stride,
1387 dst, pd->dst.stride,
1388 &mi->bmi[i].as_mv[ref].as_mv,
1389 &xd->block_refs[ref]->sf, width, height, ref,
1390 kernel, MV_PRECISION_Q3,
1391 mi_col * MI_SIZE + 4 * (i % 2),
1392 mi_row * MI_SIZE + 4 * (i / 2));
1393 #endif // CONFIG_VP9_HIGHBITDEPTH
1396 #if CONFIG_VP9_HIGHBITDEPTH
1397 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1398 vpx_highbd_subtract_block(
1399 height, width, vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1400 8, src, p->src.stride, dst, pd->dst.stride, xd->bd);
1403 height, width, vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1404 8, src, p->src.stride, dst, pd->dst.stride);
1407 vpx_subtract_block(height, width,
1408 vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1409 8, src, p->src.stride, dst, pd->dst.stride);
1410 #endif // CONFIG_VP9_HIGHBITDEPTH
1413 for (idy = 0; idy < height / 4; ++idy) {
1414 for (idx = 0; idx < width / 4; ++idx) {
1415 #if CONFIG_VP9_HIGHBITDEPTH
1416 const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
1418 int64_t ssz, rd, rd1, rd2;
1421 k += (idy * 2 + idx);
1422 coeff = BLOCK_OFFSET(p->coeff, k);
1423 x->fwd_txm4x4(vp9_raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
1425 vp9_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan);
1426 #if CONFIG_VP9_HIGHBITDEPTH
1427 thisdistortion += vp9_highbd_block_error_dispatch(
1428 coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz, bd);
1430 thisdistortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
1432 #endif // CONFIG_VP9_HIGHBITDEPTH
1434 thisrate += cost_coeffs(x, 0, k, ta + (k & 1), tl + (k >> 1), TX_4X4,
1435 so->scan, so->neighbors,
1436 cpi->sf.use_fast_coef_costing);
1437 rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
1438 rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
1439 rd = VPXMIN(rd1, rd2);
1445 *distortion = thisdistortion >> 2;
1446 *labelyrate = thisrate;
1447 *sse = thissse >> 2;
1449 return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
1460 ENTROPY_CONTEXT ta[2];
1461 ENTROPY_CONTEXT tl[2];
1473 PREDICTION_MODE modes[4];
1474 SEG_RDSTAT rdstat[4][INTER_MODES];
1478 static INLINE int mv_check_bounds(const MACROBLOCK *x, const MV *mv) {
1479 return (mv->row >> 3) < x->mv_row_min ||
1480 (mv->row >> 3) > x->mv_row_max ||
1481 (mv->col >> 3) < x->mv_col_min ||
1482 (mv->col >> 3) > x->mv_col_max;
1485 static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
1486 MODE_INFO *const mi = x->e_mbd.mi[0];
1487 struct macroblock_plane *const p = &x->plane[0];
1488 struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
1490 p->src.buf = &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i,
1492 assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
1493 pd->pre[0].buf = &pd->pre[0].buf[vp9_raster_block_offset(BLOCK_8X8, i,
1494 pd->pre[0].stride)];
1495 if (has_second_ref(mi))
1496 pd->pre[1].buf = &pd->pre[1].buf[vp9_raster_block_offset(BLOCK_8X8, i,
1497 pd->pre[1].stride)];
1500 static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
1501 struct buf_2d orig_pre[2]) {
1502 MODE_INFO *mi = x->e_mbd.mi[0];
1503 x->plane[0].src = orig_src;
1504 x->e_mbd.plane[0].pre[0] = orig_pre[0];
1505 if (has_second_ref(mi))
1506 x->e_mbd.plane[0].pre[1] = orig_pre[1];
1509 static INLINE int mv_has_subpel(const MV *mv) {
1510 return (mv->row & 0x0F) || (mv->col & 0x0F);
1513 // Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
1514 // TODO(aconverse): Find out if this is still productive then clean up or remove
1515 static int check_best_zero_mv(
1516 const VP9_COMP *cpi, const uint8_t mode_context[MAX_REF_FRAMES],
1517 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int this_mode,
1518 const MV_REFERENCE_FRAME ref_frames[2]) {
1519 if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
1520 frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
1521 (ref_frames[1] == NONE ||
1522 frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
1523 int rfc = mode_context[ref_frames[0]];
1524 int c1 = cost_mv_ref(cpi, NEARMV, rfc);
1525 int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
1526 int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
1528 if (this_mode == NEARMV) {
1529 if (c1 > c3) return 0;
1530 } else if (this_mode == NEARESTMV) {
1531 if (c2 > c3) return 0;
1533 assert(this_mode == ZEROMV);
1534 if (ref_frames[1] == NONE) {
1535 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
1536 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
1539 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
1540 frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
1541 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
1542 frame_mv[NEARMV][ref_frames[1]].as_int == 0))
1550 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
1553 int mi_row, int mi_col,
1554 int_mv single_newmv[MAX_REF_FRAMES],
1556 const VP9_COMMON *const cm = &cpi->common;
1557 const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
1558 const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
1559 MACROBLOCKD *xd = &x->e_mbd;
1560 MODE_INFO *mi = xd->mi[0];
1561 const int refs[2] = {mi->ref_frame[0],
1562 mi->ref_frame[1] < 0 ? 0 : mi->ref_frame[1]};
1565 const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
1566 struct scale_factors sf;
1568 // Do joint motion search in compound mode to get more accurate mv.
1569 struct buf_2d backup_yv12[2][MAX_MB_PLANE];
1570 int last_besterr[2] = {INT_MAX, INT_MAX};
1571 const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
1572 vp9_get_scaled_ref_frame(cpi, mi->ref_frame[0]),
1573 vp9_get_scaled_ref_frame(cpi, mi->ref_frame[1])
1576 // Prediction buffer from second frame.
1577 #if CONFIG_VP9_HIGHBITDEPTH
1578 DECLARE_ALIGNED(16, uint16_t, second_pred_alloc_16[64 * 64]);
1579 uint8_t *second_pred;
1581 DECLARE_ALIGNED(16, uint8_t, second_pred[64 * 64]);
1582 #endif // CONFIG_VP9_HIGHBITDEPTH
1584 for (ref = 0; ref < 2; ++ref) {
1585 ref_mv[ref] = x->mbmi_ext->ref_mvs[refs[ref]][0];
1587 if (scaled_ref_frame[ref]) {
1589 // Swap out the reference frame for a version that's been scaled to
1590 // match the resolution of the current frame, allowing the existing
1591 // motion search code to be used without additional modifications.
1592 for (i = 0; i < MAX_MB_PLANE; i++)
1593 backup_yv12[ref][i] = xd->plane[i].pre[ref];
1594 vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
1598 frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
1601 // Since we have scaled the reference frames to match the size of the current
1602 // frame we must use a unit scaling factor during mode selection.
1603 #if CONFIG_VP9_HIGHBITDEPTH
1604 vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height,
1605 cm->width, cm->height,
1606 cm->use_highbitdepth);
1608 vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height,
1609 cm->width, cm->height);
1610 #endif // CONFIG_VP9_HIGHBITDEPTH
1612 // Allow joint search multiple times iteratively for each reference frame
1613 // and break out of the search loop if it couldn't find a better mv.
1614 for (ite = 0; ite < 4; ite++) {
1615 struct buf_2d ref_yv12[2];
1616 int bestsme = INT_MAX;
1617 int sadpb = x->sadperbit16;
1619 int search_range = 3;
1621 int tmp_col_min = x->mv_col_min;
1622 int tmp_col_max = x->mv_col_max;
1623 int tmp_row_min = x->mv_row_min;
1624 int tmp_row_max = x->mv_row_max;
1625 int id = ite % 2; // Even iterations search in the first reference frame,
1626 // odd iterations search in the second. The predictor
1627 // found for the 'other' reference frame is factored in.
1629 // Initialized here because of compiler problem in Visual Studio.
1630 ref_yv12[0] = xd->plane[0].pre[0];
1631 ref_yv12[1] = xd->plane[0].pre[1];
1633 // Get the prediction block from the 'other' reference frame.
1634 #if CONFIG_VP9_HIGHBITDEPTH
1635 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1636 second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc_16);
1637 vp9_highbd_build_inter_predictor(ref_yv12[!id].buf,
1638 ref_yv12[!id].stride,
1640 &frame_mv[refs[!id]].as_mv,
1642 kernel, MV_PRECISION_Q3,
1643 mi_col * MI_SIZE, mi_row * MI_SIZE,
1646 second_pred = (uint8_t *)second_pred_alloc_16;
1647 vp9_build_inter_predictor(ref_yv12[!id].buf,
1648 ref_yv12[!id].stride,
1650 &frame_mv[refs[!id]].as_mv,
1652 kernel, MV_PRECISION_Q3,
1653 mi_col * MI_SIZE, mi_row * MI_SIZE);
1656 vp9_build_inter_predictor(ref_yv12[!id].buf,
1657 ref_yv12[!id].stride,
1659 &frame_mv[refs[!id]].as_mv,
1661 kernel, MV_PRECISION_Q3,
1662 mi_col * MI_SIZE, mi_row * MI_SIZE);
1663 #endif // CONFIG_VP9_HIGHBITDEPTH
1665 // Do compound motion search on the current reference frame.
1667 xd->plane[0].pre[0] = ref_yv12[id];
1668 vp9_set_mv_search_range(x, &ref_mv[id].as_mv);
1670 // Use the mv result from the single mode as mv predictor.
1671 tmp_mv = frame_mv[refs[id]].as_mv;
1676 // Small-range full-pixel motion search.
1677 bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb,
1679 &cpi->fn_ptr[bsize],
1680 &ref_mv[id].as_mv, second_pred);
1681 if (bestsme < INT_MAX)
1682 bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
1683 second_pred, &cpi->fn_ptr[bsize], 1);
1685 x->mv_col_min = tmp_col_min;
1686 x->mv_col_max = tmp_col_max;
1687 x->mv_row_min = tmp_row_min;
1688 x->mv_row_max = tmp_row_max;
1690 if (bestsme < INT_MAX) {
1691 int dis; /* TODO: use dis in distortion calculation later. */
1693 bestsme = cpi->find_fractional_mv_step(
1696 cpi->common.allow_high_precision_mv,
1698 &cpi->fn_ptr[bsize],
1699 0, cpi->sf.mv.subpel_iters_per_step,
1701 x->nmvjointcost, x->mvcost,
1702 &dis, &sse, second_pred,
1706 // Restore the pointer to the first (possibly scaled) prediction buffer.
1708 xd->plane[0].pre[0] = ref_yv12[0];
1710 if (bestsme < last_besterr[id]) {
1711 frame_mv[refs[id]].as_mv = tmp_mv;
1712 last_besterr[id] = bestsme;
1720 for (ref = 0; ref < 2; ++ref) {
1721 if (scaled_ref_frame[ref]) {
1722 // Restore the prediction frame pointers to their unscaled versions.
1724 for (i = 0; i < MAX_MB_PLANE; i++)
1725 xd->plane[i].pre[ref] = backup_yv12[ref][i];
1728 *rate_mv += vp9_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
1729 &x->mbmi_ext->ref_mvs[refs[ref]][0].as_mv,
1730 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1734 static int64_t rd_pick_best_sub8x8_mode(VP9_COMP *cpi, MACROBLOCK *x,
1735 int_mv *best_ref_mv,
1736 int_mv *second_best_ref_mv,
1737 int64_t best_rd, int *returntotrate,
1739 int64_t *returndistortion,
1740 int *skippable, int64_t *psse,
1742 int_mv seg_mvs[4][MAX_REF_FRAMES],
1743 BEST_SEG_INFO *bsi_buf, int filter_idx,
1744 int mi_row, int mi_col) {
1746 BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
1747 MACROBLOCKD *xd = &x->e_mbd;
1748 MODE_INFO *mi = xd->mi[0];
1750 int k, br = 0, idx, idy;
1751 int64_t bd = 0, block_sse = 0;
1752 PREDICTION_MODE this_mode;
1753 VP9_COMMON *cm = &cpi->common;
1754 struct macroblock_plane *const p = &x->plane[0];
1755 struct macroblockd_plane *const pd = &xd->plane[0];
1756 const int label_count = 4;
1757 int64_t this_segment_rd = 0;
1758 int label_mv_thresh;
1759 int segmentyrate = 0;
1760 const BLOCK_SIZE bsize = mi->sb_type;
1761 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1762 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1763 ENTROPY_CONTEXT t_above[2], t_left[2];
1764 int subpelmv = 1, have_ref = 0;
1765 SPEED_FEATURES *const sf = &cpi->sf;
1766 const int has_second_rf = has_second_ref(mi);
1767 const int inter_mode_mask = sf->inter_mode_mask[bsize];
1768 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1772 bsi->segment_rd = best_rd;
1773 bsi->ref_mv[0] = best_ref_mv;
1774 bsi->ref_mv[1] = second_best_ref_mv;
1775 bsi->mvp.as_int = best_ref_mv->as_int;
1776 bsi->mvthresh = mvthresh;
1778 for (i = 0; i < 4; i++)
1779 bsi->modes[i] = ZEROMV;
1781 memcpy(t_above, pd->above_context, sizeof(t_above));
1782 memcpy(t_left, pd->left_context, sizeof(t_left));
1784 // 64 makes this threshold really big effectively
1785 // making it so that we very rarely check mvs on
1786 // segments. setting this to 1 would make mv thresh
1787 // roughly equal to what it is for macroblocks
1788 label_mv_thresh = 1 * bsi->mvthresh / label_count;
1790 // Segmentation method overheads
1791 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1792 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1793 // TODO(jingning,rbultje): rewrite the rate-distortion optimization
1794 // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
1795 int_mv mode_mv[MB_MODE_COUNT][2];
1796 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1797 PREDICTION_MODE mode_selected = ZEROMV;
1798 int64_t best_rd = INT64_MAX;
1799 const int i = idy * 2 + idx;
1802 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1803 const MV_REFERENCE_FRAME frame = mi->ref_frame[ref];
1804 frame_mv[ZEROMV][frame].as_int = 0;
1805 vp9_append_sub8x8_mvs_for_idx(cm, xd, i, ref, mi_row, mi_col,
1806 &frame_mv[NEARESTMV][frame],
1807 &frame_mv[NEARMV][frame],
1808 mbmi_ext->mode_context);
1811 // search for the best motion vector on this segment
1812 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1813 const struct buf_2d orig_src = x->plane[0].src;
1814 struct buf_2d orig_pre[2];
1816 mode_idx = INTER_OFFSET(this_mode);
1817 bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
1818 if (!(inter_mode_mask & (1 << this_mode)))
1821 if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
1822 this_mode, mi->ref_frame))
1825 memcpy(orig_pre, pd->pre, sizeof(orig_pre));
1826 memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
1827 sizeof(bsi->rdstat[i][mode_idx].ta));
1828 memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
1829 sizeof(bsi->rdstat[i][mode_idx].tl));
1831 // motion search for newmv (single predictor case only)
1832 if (!has_second_rf && this_mode == NEWMV &&
1833 seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV) {
1834 MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
1836 int bestsme = INT_MAX;
1837 int sadpb = x->sadperbit4;
1842 /* Is the best so far sufficiently good that we cant justify doing
1843 * and new motion search. */
1844 if (best_rd < label_mv_thresh)
1847 if (cpi->oxcf.mode != BEST) {
1848 // use previous block's result as next block's MV predictor.
1850 bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
1852 bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
1856 max_mv = x->max_mv_context[mi->ref_frame[0]];
1859 VPXMAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
1861 if (sf->mv.auto_mv_step_size && cm->show_frame) {
1862 // Take wtd average of the step_params based on the last frame's
1863 // max mv magnitude and the best ref mvs of the current block for
1864 // the given reference.
1865 step_param = (vp9_init_search_range(max_mv) +
1866 cpi->mv_step_param) / 2;
1868 step_param = cpi->mv_step_param;
1871 mvp_full.row = bsi->mvp.as_mv.row >> 3;
1872 mvp_full.col = bsi->mvp.as_mv.col >> 3;
1874 if (sf->adaptive_motion_search) {
1875 mvp_full.row = x->pred_mv[mi->ref_frame[0]].row >> 3;
1876 mvp_full.col = x->pred_mv[mi->ref_frame[0]].col >> 3;
1877 step_param = VPXMAX(step_param, 8);
1880 // adjust src pointer for this block
1883 vp9_set_mv_search_range(x, &bsi->ref_mv[0]->as_mv);
1885 bestsme = vp9_full_pixel_search(
1886 cpi, x, bsize, &mvp_full, step_param, sadpb,
1887 sf->mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
1888 &bsi->ref_mv[0]->as_mv, new_mv,
1891 if (bestsme < INT_MAX) {
1893 cpi->find_fractional_mv_step(
1896 &bsi->ref_mv[0]->as_mv,
1897 cm->allow_high_precision_mv,
1898 x->errorperbit, &cpi->fn_ptr[bsize],
1899 sf->mv.subpel_force_stop,
1900 sf->mv.subpel_iters_per_step,
1901 cond_cost_list(cpi, cost_list),
1902 x->nmvjointcost, x->mvcost,
1904 &x->pred_sse[mi->ref_frame[0]],
1907 // save motion search result for use in compound prediction
1908 seg_mvs[i][mi->ref_frame[0]].as_mv = *new_mv;
1911 if (sf->adaptive_motion_search)
1912 x->pred_mv[mi->ref_frame[0]] = *new_mv;
1914 // restore src pointers
1915 mi_buf_restore(x, orig_src, orig_pre);
1918 if (has_second_rf) {
1919 if (seg_mvs[i][mi->ref_frame[1]].as_int == INVALID_MV ||
1920 seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV)
1924 if (has_second_rf && this_mode == NEWMV &&
1925 mi->interp_filter == EIGHTTAP) {
1926 // adjust src pointers
1928 if (sf->comp_inter_joint_search_thresh <= bsize) {
1930 joint_motion_search(cpi, x, bsize, frame_mv[this_mode],
1931 mi_row, mi_col, seg_mvs[i],
1933 seg_mvs[i][mi->ref_frame[0]].as_int =
1934 frame_mv[this_mode][mi->ref_frame[0]].as_int;
1935 seg_mvs[i][mi->ref_frame[1]].as_int =
1936 frame_mv[this_mode][mi->ref_frame[1]].as_int;
1938 // restore src pointers
1939 mi_buf_restore(x, orig_src, orig_pre);
1942 bsi->rdstat[i][mode_idx].brate =
1943 set_and_cost_bmi_mvs(cpi, x, xd, i, this_mode, mode_mv[this_mode],
1944 frame_mv, seg_mvs[i], bsi->ref_mv,
1945 x->nmvjointcost, x->mvcost);
1947 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1948 bsi->rdstat[i][mode_idx].mvs[ref].as_int =
1949 mode_mv[this_mode][ref].as_int;
1950 if (num_4x4_blocks_wide > 1)
1951 bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int =
1952 mode_mv[this_mode][ref].as_int;
1953 if (num_4x4_blocks_high > 1)
1954 bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int =
1955 mode_mv[this_mode][ref].as_int;
1958 // Trap vectors that reach beyond the UMV borders
1959 if (mv_check_bounds(x, &mode_mv[this_mode][0].as_mv) ||
1961 mv_check_bounds(x, &mode_mv[this_mode][1].as_mv)))
1964 if (filter_idx > 0) {
1965 BEST_SEG_INFO *ref_bsi = bsi_buf;
1969 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1970 subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
1971 have_ref &= mode_mv[this_mode][ref].as_int ==
1972 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1975 if (filter_idx > 1 && !subpelmv && !have_ref) {
1976 ref_bsi = bsi_buf + 1;
1978 for (ref = 0; ref < 1 + has_second_rf; ++ref)
1979 have_ref &= mode_mv[this_mode][ref].as_int ==
1980 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1983 if (!subpelmv && have_ref &&
1984 ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1985 memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
1986 sizeof(SEG_RDSTAT));
1987 if (num_4x4_blocks_wide > 1)
1988 bsi->rdstat[i + 1][mode_idx].eobs =
1989 ref_bsi->rdstat[i + 1][mode_idx].eobs;
1990 if (num_4x4_blocks_high > 1)
1991 bsi->rdstat[i + 2][mode_idx].eobs =
1992 ref_bsi->rdstat[i + 2][mode_idx].eobs;
1994 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
1995 mode_selected = this_mode;
1996 best_rd = bsi->rdstat[i][mode_idx].brdcost;
2002 bsi->rdstat[i][mode_idx].brdcost =
2003 encode_inter_mb_segment(cpi, x,
2004 bsi->segment_rd - this_segment_rd, i,
2005 &bsi->rdstat[i][mode_idx].byrate,
2006 &bsi->rdstat[i][mode_idx].bdist,
2007 &bsi->rdstat[i][mode_idx].bsse,
2008 bsi->rdstat[i][mode_idx].ta,
2009 bsi->rdstat[i][mode_idx].tl,
2011 if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
2012 bsi->rdstat[i][mode_idx].brdcost += RDCOST(x->rdmult, x->rddiv,
2013 bsi->rdstat[i][mode_idx].brate, 0);
2014 bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
2015 bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
2016 if (num_4x4_blocks_wide > 1)
2017 bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
2018 if (num_4x4_blocks_high > 1)
2019 bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
2022 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
2023 mode_selected = this_mode;
2024 best_rd = bsi->rdstat[i][mode_idx].brdcost;
2026 } /*for each 4x4 mode*/
2028 if (best_rd == INT64_MAX) {
2030 for (iy = i + 1; iy < 4; ++iy)
2031 for (midx = 0; midx < INTER_MODES; ++midx)
2032 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2033 bsi->segment_rd = INT64_MAX;
2037 mode_idx = INTER_OFFSET(mode_selected);
2038 memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
2039 memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
2041 set_and_cost_bmi_mvs(cpi, x, xd, i, mode_selected, mode_mv[mode_selected],
2042 frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
2045 br += bsi->rdstat[i][mode_idx].brate;
2046 bd += bsi->rdstat[i][mode_idx].bdist;
2047 block_sse += bsi->rdstat[i][mode_idx].bsse;
2048 segmentyrate += bsi->rdstat[i][mode_idx].byrate;
2049 this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
2051 if (this_segment_rd > bsi->segment_rd) {
2053 for (iy = i + 1; iy < 4; ++iy)
2054 for (midx = 0; midx < INTER_MODES; ++midx)
2055 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2056 bsi->segment_rd = INT64_MAX;
2060 } /* for each label */
2064 bsi->segment_yrate = segmentyrate;
2065 bsi->segment_rd = this_segment_rd;
2066 bsi->sse = block_sse;
2068 // update the coding decisions
2069 for (k = 0; k < 4; ++k)
2070 bsi->modes[k] = mi->bmi[k].as_mode;
2072 if (bsi->segment_rd > best_rd)
2074 /* set it to the best */
2075 for (i = 0; i < 4; i++) {
2076 mode_idx = INTER_OFFSET(bsi->modes[i]);
2077 mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
2078 if (has_second_ref(mi))
2079 mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
2080 x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
2081 mi->bmi[i].as_mode = bsi->modes[i];
2085 * used to set mbmi->mv.as_int
2087 *returntotrate = bsi->r;
2088 *returndistortion = bsi->d;
2089 *returnyrate = bsi->segment_yrate;
2090 *skippable = vp9_is_skippable_in_plane(x, BLOCK_8X8, 0);
2092 mi->mode = bsi->modes[3];
2094 return bsi->segment_rd;
2097 static void estimate_ref_frame_costs(const VP9_COMMON *cm,
2098 const MACROBLOCKD *xd,
2100 unsigned int *ref_costs_single,
2101 unsigned int *ref_costs_comp,
2102 vpx_prob *comp_mode_p) {
2103 int seg_ref_active = segfeature_active(&cm->seg, segment_id,
2105 if (seg_ref_active) {
2106 memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
2107 memset(ref_costs_comp, 0, MAX_REF_FRAMES * sizeof(*ref_costs_comp));
2110 vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
2111 vpx_prob comp_inter_p = 128;
2113 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
2114 comp_inter_p = vp9_get_reference_mode_prob(cm, xd);
2115 *comp_mode_p = comp_inter_p;
2120 ref_costs_single[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
2122 if (cm->reference_mode != COMPOUND_REFERENCE) {
2123 vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
2124 vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
2125 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
2127 if (cm->reference_mode == REFERENCE_MODE_SELECT)
2128 base_cost += vp9_cost_bit(comp_inter_p, 0);
2130 ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] =
2131 ref_costs_single[ALTREF_FRAME] = base_cost;
2132 ref_costs_single[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
2133 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
2134 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
2135 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
2136 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
2138 ref_costs_single[LAST_FRAME] = 512;
2139 ref_costs_single[GOLDEN_FRAME] = 512;
2140 ref_costs_single[ALTREF_FRAME] = 512;
2142 if (cm->reference_mode != SINGLE_REFERENCE) {
2143 vpx_prob ref_comp_p = vp9_get_pred_prob_comp_ref_p(cm, xd);
2144 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
2146 if (cm->reference_mode == REFERENCE_MODE_SELECT)
2147 base_cost += vp9_cost_bit(comp_inter_p, 1);
2149 ref_costs_comp[LAST_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 0);
2150 ref_costs_comp[GOLDEN_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 1);
2152 ref_costs_comp[LAST_FRAME] = 512;
2153 ref_costs_comp[GOLDEN_FRAME] = 512;
2158 static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
2160 int64_t comp_pred_diff[REFERENCE_MODES],
2161 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS],
2163 MACROBLOCKD *const xd = &x->e_mbd;
2165 // Take a snapshot of the coding context so it can be
2166 // restored if we decide to encode this way
2167 ctx->skip = x->skip;
2168 ctx->skippable = skippable;
2169 ctx->best_mode_index = mode_index;
2170 ctx->mic = *xd->mi[0];
2171 ctx->mbmi_ext = *x->mbmi_ext;
2172 ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
2173 ctx->comp_pred_diff = (int)comp_pred_diff[COMPOUND_REFERENCE];
2174 ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
2176 memcpy(ctx->best_filter_diff, best_filter_diff,
2177 sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
2180 static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
2181 MV_REFERENCE_FRAME ref_frame,
2182 BLOCK_SIZE block_size,
2183 int mi_row, int mi_col,
2184 int_mv frame_nearest_mv[MAX_REF_FRAMES],
2185 int_mv frame_near_mv[MAX_REF_FRAMES],
2186 struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
2187 const VP9_COMMON *cm = &cpi->common;
2188 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2189 MACROBLOCKD *const xd = &x->e_mbd;
2190 MODE_INFO *const mi = xd->mi[0];
2191 int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
2192 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2193 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2195 assert(yv12 != NULL);
2197 // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
2198 // use the UV scaling factors.
2199 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
2201 // Gets an initial list of candidate vectors from neighbours and orders them
2202 vp9_find_mv_refs(cm, xd, mi, ref_frame, candidates, mi_row, mi_col,
2203 mbmi_ext->mode_context);
2205 // Candidate refinement carried out at encoder and decoder
2206 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2207 &frame_nearest_mv[ref_frame],
2208 &frame_near_mv[ref_frame]);
2210 // Further refinement that is encode side only to test the top few candidates
2211 // in full and choose the best as the centre point for subsequent searches.
2212 // The current implementation doesn't support scaling.
2213 if (!vp9_is_scaled(sf) && block_size >= BLOCK_8X8)
2214 vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
2215 ref_frame, block_size);
2218 static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
2220 int mi_row, int mi_col,
2221 int_mv *tmp_mv, int *rate_mv) {
2222 MACROBLOCKD *xd = &x->e_mbd;
2223 const VP9_COMMON *cm = &cpi->common;
2224 MODE_INFO *mi = xd->mi[0];
2225 struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
2226 int bestsme = INT_MAX;
2228 int sadpb = x->sadperbit16;
2230 int ref = mi->ref_frame[0];
2231 MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2233 int tmp_col_min = x->mv_col_min;
2234 int tmp_col_max = x->mv_col_max;
2235 int tmp_row_min = x->mv_row_min;
2236 int tmp_row_max = x->mv_row_max;
2239 const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi,
2243 pred_mv[0] = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2244 pred_mv[1] = x->mbmi_ext->ref_mvs[ref][1].as_mv;
2245 pred_mv[2] = x->pred_mv[ref];
2247 if (scaled_ref_frame) {
2249 // Swap out the reference frame for a version that's been scaled to
2250 // match the resolution of the current frame, allowing the existing
2251 // motion search code to be used without additional modifications.
2252 for (i = 0; i < MAX_MB_PLANE; i++)
2253 backup_yv12[i] = xd->plane[i].pre[0];
2255 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
2258 vp9_set_mv_search_range(x, &ref_mv);
2260 // Work out the size of the first step in the mv step search.
2261 // 0 here is maximum length first step. 1 is VPXMAX >> 1 etc.
2262 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
2263 // Take wtd average of the step_params based on the last frame's
2264 // max mv magnitude and that based on the best ref mvs of the current
2265 // block for the given reference.
2266 step_param = (vp9_init_search_range(x->max_mv_context[ref]) +
2267 cpi->mv_step_param) / 2;
2269 step_param = cpi->mv_step_param;
2272 if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64) {
2274 2 * (b_width_log2_lookup[BLOCK_64X64] -
2275 VPXMIN(b_height_log2_lookup[bsize], b_width_log2_lookup[bsize]));
2276 step_param = VPXMAX(step_param, boffset);
2279 if (cpi->sf.adaptive_motion_search) {
2280 int bwl = b_width_log2_lookup[bsize];
2281 int bhl = b_height_log2_lookup[bsize];
2282 int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
2287 // prev_mv_sad is not setup for dynamically scaled frames.
2288 if (cpi->oxcf.resize_mode != RESIZE_DYNAMIC) {
2290 for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
2291 if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
2292 x->pred_mv[ref].row = 0;
2293 x->pred_mv[ref].col = 0;
2294 tmp_mv->as_int = INVALID_MV;
2296 if (scaled_ref_frame) {
2298 for (i = 0; i < MAX_MB_PLANE; ++i)
2299 xd->plane[i].pre[0] = backup_yv12[i];
2307 mvp_full = pred_mv[x->mv_best_ref_index[ref]];
2312 bestsme = vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
2313 cond_cost_list(cpi, cost_list),
2314 &ref_mv, &tmp_mv->as_mv, INT_MAX, 1);
2316 x->mv_col_min = tmp_col_min;
2317 x->mv_col_max = tmp_col_max;
2318 x->mv_row_min = tmp_row_min;
2319 x->mv_row_max = tmp_row_max;
2321 if (bestsme < INT_MAX) {
2322 int dis; /* TODO: use dis in distortion calculation later. */
2323 cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
2324 cm->allow_high_precision_mv,
2326 &cpi->fn_ptr[bsize],
2327 cpi->sf.mv.subpel_force_stop,
2328 cpi->sf.mv.subpel_iters_per_step,
2329 cond_cost_list(cpi, cost_list),
2330 x->nmvjointcost, x->mvcost,
2331 &dis, &x->pred_sse[ref], NULL, 0, 0);
2333 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
2334 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2336 if (cpi->sf.adaptive_motion_search)
2337 x->pred_mv[ref] = tmp_mv->as_mv;
2339 if (scaled_ref_frame) {
2341 for (i = 0; i < MAX_MB_PLANE; i++)
2342 xd->plane[i].pre[0] = backup_yv12[i];
2348 static INLINE void restore_dst_buf(MACROBLOCKD *xd,
2349 uint8_t *orig_dst[MAX_MB_PLANE],
2350 int orig_dst_stride[MAX_MB_PLANE]) {
2352 for (i = 0; i < MAX_MB_PLANE; i++) {
2353 xd->plane[i].dst.buf = orig_dst[i];
2354 xd->plane[i].dst.stride = orig_dst_stride[i];
2358 // In some situations we want to discount tha pparent cost of a new motion
2359 // vector. Where there is a subtle motion field and especially where there is
2360 // low spatial complexity then it can be hard to cover the cost of a new motion
2361 // vector in a single block, even if that motion vector reduces distortion.
2362 // However, once established that vector may be usable through the nearest and
2363 // near mv modes to reduce distortion in subsequent blocks and also improve
2365 static int discount_newmv_test(const VP9_COMP *cpi,
2368 int_mv (*mode_mv)[MAX_REF_FRAMES],
2370 return (!cpi->rc.is_src_frame_alt_ref &&
2371 (this_mode == NEWMV) &&
2372 (this_mv.as_int != 0) &&
2373 ((mode_mv[NEARESTMV][ref_frame].as_int == 0) ||
2374 (mode_mv[NEARESTMV][ref_frame].as_int == INVALID_MV)) &&
2375 ((mode_mv[NEARMV][ref_frame].as_int == 0) ||
2376 (mode_mv[NEARMV][ref_frame].as_int == INVALID_MV)));
2379 static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
2381 int *rate2, int64_t *distortion,
2383 int *rate_y, int *rate_uv,
2385 int_mv (*mode_mv)[MAX_REF_FRAMES],
2386 int mi_row, int mi_col,
2387 int_mv single_newmv[MAX_REF_FRAMES],
2388 INTERP_FILTER (*single_filter)[MAX_REF_FRAMES],
2389 int (*single_skippable)[MAX_REF_FRAMES],
2391 const int64_t ref_best_rd,
2392 int64_t *mask_filter,
2393 int64_t filter_cache[]) {
2394 VP9_COMMON *cm = &cpi->common;
2395 MACROBLOCKD *xd = &x->e_mbd;
2396 MODE_INFO *mi = xd->mi[0];
2397 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2398 const int is_comp_pred = has_second_ref(mi);
2399 const int this_mode = mi->mode;
2400 int_mv *frame_mv = mode_mv[this_mode];
2402 int refs[2] = { mi->ref_frame[0],
2403 (mi->ref_frame[1] < 0 ? 0 : mi->ref_frame[1]) };
2405 #if CONFIG_VP9_HIGHBITDEPTH
2406 DECLARE_ALIGNED(16, uint16_t, tmp_buf16[MAX_MB_PLANE * 64 * 64]);
2409 DECLARE_ALIGNED(16, uint8_t, tmp_buf[MAX_MB_PLANE * 64 * 64]);
2410 #endif // CONFIG_VP9_HIGHBITDEPTH
2411 int pred_exists = 0;
2413 int64_t rd, tmp_rd, best_rd = INT64_MAX;
2414 int best_needs_copy = 0;
2415 uint8_t *orig_dst[MAX_MB_PLANE];
2416 int orig_dst_stride[MAX_MB_PLANE];
2418 INTERP_FILTER best_filter = SWITCHABLE;
2419 uint8_t skip_txfm[MAX_MB_PLANE << 2] = {0};
2420 int64_t bsse[MAX_MB_PLANE << 2] = {0};
2422 int bsl = mi_width_log2_lookup[bsize];
2423 int pred_filter_search = cpi->sf.cb_pred_filter_search ?
2424 (((mi_row + mi_col) >> bsl) +
2425 get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
2427 int skip_txfm_sb = 0;
2428 int64_t skip_sse_sb = INT64_MAX;
2429 int64_t distortion_y = 0, distortion_uv = 0;
2431 #if CONFIG_VP9_HIGHBITDEPTH
2432 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2433 tmp_buf = CONVERT_TO_BYTEPTR(tmp_buf16);
2435 tmp_buf = (uint8_t *)tmp_buf16;
2437 #endif // CONFIG_VP9_HIGHBITDEPTH
2439 if (pred_filter_search) {
2440 INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE;
2441 if (xd->up_available)
2442 af = xd->mi[-xd->mi_stride]->interp_filter;
2443 if (xd->left_available)
2444 lf = xd->mi[-1]->interp_filter;
2446 if ((this_mode != NEWMV) || (af == lf))
2451 if (frame_mv[refs[0]].as_int == INVALID_MV ||
2452 frame_mv[refs[1]].as_int == INVALID_MV)
2455 if (cpi->sf.adaptive_mode_search) {
2456 if (single_filter[this_mode][refs[0]] ==
2457 single_filter[this_mode][refs[1]])
2458 best_filter = single_filter[this_mode][refs[0]];
2462 if (this_mode == NEWMV) {
2465 // Initialize mv using single prediction mode result.
2466 frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
2467 frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
2469 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
2470 joint_motion_search(cpi, x, bsize, frame_mv,
2471 mi_row, mi_col, single_newmv, &rate_mv);
2473 rate_mv = vp9_mv_bit_cost(&frame_mv[refs[0]].as_mv,
2474 &x->mbmi_ext->ref_mvs[refs[0]][0].as_mv,
2475 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2476 rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]].as_mv,
2477 &x->mbmi_ext->ref_mvs[refs[1]][0].as_mv,
2478 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2483 single_motion_search(cpi, x, bsize, mi_row, mi_col,
2485 if (tmp_mv.as_int == INVALID_MV)
2488 frame_mv[refs[0]].as_int =
2489 xd->mi[0]->bmi[0].as_mv[0].as_int = tmp_mv.as_int;
2490 single_newmv[refs[0]].as_int = tmp_mv.as_int;
2492 // Estimate the rate implications of a new mv but discount this
2493 // under certain circumstances where we want to help initiate a weak
2494 // motion field, where the distortion gain for a single block may not
2495 // be enough to overcome the cost of a new mv.
2496 if (discount_newmv_test(cpi, this_mode, tmp_mv, mode_mv, refs[0])) {
2497 *rate2 += VPXMAX((rate_mv / NEW_MV_DISCOUNT_FACTOR), 1);
2504 for (i = 0; i < is_comp_pred + 1; ++i) {
2505 cur_mv[i] = frame_mv[refs[i]];
2506 // Clip "next_nearest" so that it does not extend to far out of image
2507 if (this_mode != NEWMV)
2508 clamp_mv2(&cur_mv[i].as_mv, xd);
2510 if (mv_check_bounds(x, &cur_mv[i].as_mv))
2512 mi->mv[i].as_int = cur_mv[i].as_int;
2515 // do first prediction into the destination buffer. Do the next
2516 // prediction into a temporary buffer. Then keep track of which one
2517 // of these currently holds the best predictor, and use the other
2518 // one for future predictions. In the end, copy from tmp_buf to
2519 // dst if necessary.
2520 for (i = 0; i < MAX_MB_PLANE; i++) {
2521 orig_dst[i] = xd->plane[i].dst.buf;
2522 orig_dst_stride[i] = xd->plane[i].dst.stride;
2525 // We don't include the cost of the second reference here, because there
2526 // are only two options: Last/ARF or Golden/ARF; The second one is always
2527 // known, which is ARF.
2529 // Under some circumstances we discount the cost of new mv mode to encourage
2530 // initiation of a motion field.
2531 if (discount_newmv_test(cpi, this_mode, frame_mv[refs[0]],
2532 mode_mv, refs[0])) {
2533 *rate2 += VPXMIN(cost_mv_ref(cpi, this_mode,
2534 mbmi_ext->mode_context[refs[0]]),
2535 cost_mv_ref(cpi, NEARESTMV,
2536 mbmi_ext->mode_context[refs[0]]));
2538 *rate2 += cost_mv_ref(cpi, this_mode, mbmi_ext->mode_context[refs[0]]);
2541 if (RDCOST(x->rdmult, x->rddiv, *rate2, 0) > ref_best_rd &&
2542 mi->mode != NEARESTMV)
2546 // Are all MVs integer pel for Y and UV
2547 intpel_mv = !mv_has_subpel(&mi->mv[0].as_mv);
2549 intpel_mv &= !mv_has_subpel(&mi->mv[1].as_mv);
2551 // Search for best switchable filter by checking the variance of
2552 // pred error irrespective of whether the filter will be used
2553 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
2554 filter_cache[i] = INT64_MAX;
2556 if (cm->interp_filter != BILINEAR) {
2557 if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
2558 best_filter = EIGHTTAP;
2559 } else if (best_filter == SWITCHABLE) {
2561 int tmp_rate_sum = 0;
2562 int64_t tmp_dist_sum = 0;
2564 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
2567 int tmp_skip_sb = 0;
2568 int64_t tmp_skip_sse = INT64_MAX;
2570 mi->interp_filter = i;
2571 rs = vp9_get_switchable_rate(cpi, xd);
2572 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
2574 if (i > 0 && intpel_mv) {
2575 rd = RDCOST(x->rdmult, x->rddiv, tmp_rate_sum, tmp_dist_sum);
2576 filter_cache[i] = rd;
2577 filter_cache[SWITCHABLE_FILTERS] =
2578 VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2579 if (cm->interp_filter == SWITCHABLE)
2581 *mask_filter = VPXMAX(*mask_filter, rd);
2584 int64_t dist_sum = 0;
2585 if (i > 0 && cpi->sf.adaptive_interp_filter_search &&
2586 (cpi->sf.interp_filter_search_mask & (1 << i))) {
2588 dist_sum = INT64_MAX;
2592 if ((cm->interp_filter == SWITCHABLE &&
2593 (!i || best_needs_copy)) ||
2594 (cm->interp_filter != SWITCHABLE &&
2595 (cm->interp_filter == mi->interp_filter ||
2596 (i == 0 && intpel_mv)))) {
2597 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2599 for (j = 0; j < MAX_MB_PLANE; j++) {
2600 xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
2601 xd->plane[j].dst.stride = 64;
2604 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2605 model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum,
2606 &tmp_skip_sb, &tmp_skip_sse);
2608 rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum);
2609 filter_cache[i] = rd;
2610 filter_cache[SWITCHABLE_FILTERS] =
2611 VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2612 if (cm->interp_filter == SWITCHABLE)
2614 *mask_filter = VPXMAX(*mask_filter, rd);
2616 if (i == 0 && intpel_mv) {
2617 tmp_rate_sum = rate_sum;
2618 tmp_dist_sum = dist_sum;
2622 if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2623 if (rd / 2 > ref_best_rd) {
2624 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2628 newbest = i == 0 || rd < best_rd;
2632 best_filter = mi->interp_filter;
2633 if (cm->interp_filter == SWITCHABLE && i && !intpel_mv)
2634 best_needs_copy = !best_needs_copy;
2637 if ((cm->interp_filter == SWITCHABLE && newbest) ||
2638 (cm->interp_filter != SWITCHABLE &&
2639 cm->interp_filter == mi->interp_filter)) {
2643 skip_txfm_sb = tmp_skip_sb;
2644 skip_sse_sb = tmp_skip_sse;
2645 memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2646 memcpy(bsse, x->bsse, sizeof(bsse));
2649 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2652 // Set the appropriate filter
2653 mi->interp_filter = cm->interp_filter != SWITCHABLE ?
2654 cm->interp_filter : best_filter;
2655 rs = cm->interp_filter == SWITCHABLE ? vp9_get_switchable_rate(cpi, xd) : 0;
2658 if (best_needs_copy) {
2659 // again temporarily set the buffers to local memory to prevent a memcpy
2660 for (i = 0; i < MAX_MB_PLANE; i++) {
2661 xd->plane[i].dst.buf = tmp_buf + i * 64 * 64;
2662 xd->plane[i].dst.stride = 64;
2665 rd = tmp_rd + RDCOST(x->rdmult, x->rddiv, rs, 0);
2669 // Handles the special case when a filter that is not in the
2670 // switchable list (ex. bilinear) is indicated at the frame level, or
2671 // skip condition holds.
2672 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2673 model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist,
2674 &skip_txfm_sb, &skip_sse_sb);
2675 rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
2676 memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2677 memcpy(bsse, x->bsse, sizeof(bsse));
2681 single_filter[this_mode][refs[0]] = mi->interp_filter;
2683 if (cpi->sf.adaptive_mode_search)
2685 if (single_skippable[this_mode][refs[0]] &&
2686 single_skippable[this_mode][refs[1]])
2687 memset(skip_txfm, SKIP_TXFM_AC_DC, sizeof(skip_txfm));
2689 if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2690 // if current pred_error modeled rd is substantially more than the best
2691 // so far, do not bother doing full rd
2692 if (rd / 2 > ref_best_rd) {
2693 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2698 if (cm->interp_filter == SWITCHABLE)
2701 memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm));
2702 memcpy(x->bsse, bsse, sizeof(bsse));
2704 if (!skip_txfm_sb) {
2705 int skippable_y, skippable_uv;
2706 int64_t sseuv = INT64_MAX;
2707 int64_t rdcosty = INT64_MAX;
2709 // Y cost and distortion
2710 vp9_subtract_plane(x, bsize, 0);
2711 super_block_yrd(cpi, x, rate_y, &distortion_y, &skippable_y, psse,
2712 bsize, ref_best_rd);
2714 if (*rate_y == INT_MAX) {
2716 *distortion = INT64_MAX;
2717 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2722 *distortion += distortion_y;
2724 rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
2725 rdcosty = VPXMIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
2727 if (!super_block_uvrd(cpi, x, rate_uv, &distortion_uv, &skippable_uv,
2728 &sseuv, bsize, ref_best_rd - rdcosty)) {
2730 *distortion = INT64_MAX;
2731 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2737 *distortion += distortion_uv;
2738 *skippable = skippable_y && skippable_uv;
2743 // The cost of skip bit needs to be added.
2744 *rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2746 *distortion = skip_sse_sb;
2750 single_skippable[this_mode][refs[0]] = *skippable;
2752 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2753 return 0; // The rate-distortion cost will be re-calculated by caller.
2756 void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
2757 RD_COST *rd_cost, BLOCK_SIZE bsize,
2758 PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
2759 VP9_COMMON *const cm = &cpi->common;
2760 MACROBLOCKD *const xd = &x->e_mbd;
2761 struct macroblockd_plane *const pd = xd->plane;
2762 int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
2763 int y_skip = 0, uv_skip = 0;
2764 int64_t dist_y = 0, dist_uv = 0;
2765 TX_SIZE max_uv_tx_size;
2768 xd->mi[0]->ref_frame[0] = INTRA_FRAME;
2769 xd->mi[0]->ref_frame[1] = NONE;
2771 if (bsize >= BLOCK_8X8) {
2772 if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2773 &dist_y, &y_skip, bsize,
2774 best_rd) >= best_rd) {
2775 rd_cost->rate = INT_MAX;
2780 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2781 &dist_y, best_rd) >= best_rd) {
2782 rd_cost->rate = INT_MAX;
2786 max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0]->tx_size, bsize,
2787 pd[1].subsampling_x,
2788 pd[1].subsampling_y);
2789 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
2790 &dist_uv, &uv_skip, VPXMAX(BLOCK_8X8, bsize),
2793 if (y_skip && uv_skip) {
2794 rd_cost->rate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
2795 vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2796 rd_cost->dist = dist_y + dist_uv;
2798 rd_cost->rate = rate_y + rate_uv +
2799 vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
2800 rd_cost->dist = dist_y + dist_uv;
2803 ctx->mic = *xd->mi[0];
2804 ctx->mbmi_ext = *x->mbmi_ext;
2805 rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
2808 // This function is designed to apply a bias or adjustment to an rd value based
2809 // on the relative variance of the source and reconstruction.
2810 #define LOW_VAR_THRESH 16
2811 #define VLOW_ADJ_MAX 25
2812 #define VHIGH_ADJ_MAX 8
2813 static void rd_variance_adjustment(VP9_COMP *cpi,
2817 MV_REFERENCE_FRAME ref_frame,
2818 unsigned int source_variance) {
2819 MACROBLOCKD *const xd = &x->e_mbd;
2820 unsigned int recon_variance;
2821 unsigned int absvar_diff = 0;
2822 int64_t var_error = 0;
2823 int64_t var_factor = 0;
2825 if (*this_rd == INT64_MAX)
2828 #if CONFIG_VP9_HIGHBITDEPTH
2829 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2831 vp9_high_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize, xd->bd);
2834 vp9_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
2838 vp9_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
2839 #endif // CONFIG_VP9_HIGHBITDEPTH
2841 if ((source_variance + recon_variance) > LOW_VAR_THRESH) {
2842 absvar_diff = (source_variance > recon_variance)
2843 ? (source_variance - recon_variance)
2844 : (recon_variance - source_variance);
2846 var_error = ((int64_t)200 * source_variance * recon_variance) /
2847 (((int64_t)source_variance * source_variance) +
2848 ((int64_t)recon_variance * recon_variance));
2849 var_error = 100 - var_error;
2852 // Source variance above a threshold and ref frame is intra.
2853 // This case is targeted mainly at discouraging intra modes that give rise
2854 // to a predictor with a low spatial complexity compared to the source.
2855 if ((source_variance > LOW_VAR_THRESH) && (ref_frame == INTRA_FRAME) &&
2856 (source_variance > recon_variance)) {
2857 var_factor = VPXMIN(absvar_diff, VPXMIN(VLOW_ADJ_MAX, var_error));
2858 // A second possible case of interest is where the source variance
2859 // is very low and we wish to discourage false texture or motion trails.
2860 } else if ((source_variance < (LOW_VAR_THRESH >> 1)) &&
2861 (recon_variance > source_variance)) {
2862 var_factor = VPXMIN(absvar_diff, VPXMIN(VHIGH_ADJ_MAX, var_error));
2864 *this_rd += (*this_rd * var_factor) / 100;
2868 // Do we have an internal image edge (e.g. formatting bars).
2869 int vp9_internal_image_edge(VP9_COMP *cpi) {
2870 return (cpi->oxcf.pass == 2) &&
2871 ((cpi->twopass.this_frame_stats.inactive_zone_rows > 0) ||
2872 (cpi->twopass.this_frame_stats.inactive_zone_cols > 0));
2875 // Checks to see if a super block is on a horizontal image edge.
2876 // In most cases this is the "real" edge unless there are formatting
2877 // bars embedded in the stream.
2878 int vp9_active_h_edge(VP9_COMP *cpi, int mi_row, int mi_step) {
2880 int bottom_edge = cpi->common.mi_rows;
2881 int is_active_h_edge = 0;
2883 // For two pass account for any formatting bars detected.
2884 if (cpi->oxcf.pass == 2) {
2885 TWO_PASS *twopass = &cpi->twopass;
2887 // The inactive region is specified in MBs not mi units.
2888 // The image edge is in the following MB row.
2889 top_edge += (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
2891 bottom_edge -= (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
2892 bottom_edge = VPXMAX(top_edge, bottom_edge);
2895 if (((top_edge >= mi_row) && (top_edge < (mi_row + mi_step))) ||
2896 ((bottom_edge >= mi_row) && (bottom_edge < (mi_row + mi_step)))) {
2897 is_active_h_edge = 1;
2899 return is_active_h_edge;
2902 // Checks to see if a super block is on a vertical image edge.
2903 // In most cases this is the "real" edge unless there are formatting
2904 // bars embedded in the stream.
2905 int vp9_active_v_edge(VP9_COMP *cpi, int mi_col, int mi_step) {
2907 int right_edge = cpi->common.mi_cols;
2908 int is_active_v_edge = 0;
2910 // For two pass account for any formatting bars detected.
2911 if (cpi->oxcf.pass == 2) {
2912 TWO_PASS *twopass = &cpi->twopass;
2914 // The inactive region is specified in MBs not mi units.
2915 // The image edge is in the following MB row.
2916 left_edge += (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
2918 right_edge -= (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
2919 right_edge = VPXMAX(left_edge, right_edge);
2922 if (((left_edge >= mi_col) && (left_edge < (mi_col + mi_step))) ||
2923 ((right_edge >= mi_col) && (right_edge < (mi_col + mi_step)))) {
2924 is_active_v_edge = 1;
2926 return is_active_v_edge;
2929 // Checks to see if a super block is at the edge of the active image.
2930 // In most cases this is the "real" edge unless there are formatting
2931 // bars embedded in the stream.
2932 int vp9_active_edge_sb(VP9_COMP *cpi,
2933 int mi_row, int mi_col) {
2934 return vp9_active_h_edge(cpi, mi_row, MI_BLOCK_SIZE) ||
2935 vp9_active_v_edge(cpi, mi_col, MI_BLOCK_SIZE);
2938 void vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi,
2939 TileDataEnc *tile_data,
2941 int mi_row, int mi_col,
2942 RD_COST *rd_cost, BLOCK_SIZE bsize,
2943 PICK_MODE_CONTEXT *ctx,
2944 int64_t best_rd_so_far) {
2945 VP9_COMMON *const cm = &cpi->common;
2946 TileInfo *const tile_info = &tile_data->tile_info;
2947 RD_OPT *const rd_opt = &cpi->rd;
2948 SPEED_FEATURES *const sf = &cpi->sf;
2949 MACROBLOCKD *const xd = &x->e_mbd;
2950 MODE_INFO *const mi = xd->mi[0];
2951 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2952 const struct segmentation *const seg = &cm->seg;
2953 PREDICTION_MODE this_mode;
2954 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
2955 unsigned char segment_id = mi->segment_id;
2956 int comp_pred, i, k;
2957 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
2958 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2959 int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
2960 INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES];
2961 int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES];
2962 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2964 int64_t best_rd = best_rd_so_far;
2965 int64_t best_pred_diff[REFERENCE_MODES];
2966 int64_t best_pred_rd[REFERENCE_MODES];
2967 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
2968 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
2969 MODE_INFO best_mbmode;
2970 int best_mode_skippable = 0;
2971 int midx, best_mode_index = -1;
2972 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
2973 vpx_prob comp_mode_p;
2974 int64_t best_intra_rd = INT64_MAX;
2975 unsigned int best_pred_sse = UINT_MAX;
2976 PREDICTION_MODE best_intra_mode = DC_PRED;
2977 int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
2978 int64_t dist_uv[TX_SIZES];
2979 int skip_uv[TX_SIZES];
2980 PREDICTION_MODE mode_uv[TX_SIZES];
2981 const int intra_cost_penalty = vp9_get_intra_cost_penalty(
2982 cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
2984 uint8_t ref_frame_skip_mask[2] = { 0 };
2985 uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 };
2986 int mode_skip_start = sf->mode_skip_start + 1;
2987 const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
2988 const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
2989 int64_t mode_threshold[MAX_MODES];
2990 int *mode_map = tile_data->mode_map[bsize];
2991 const int mode_search_skip_flags = sf->mode_search_skip_flags;
2992 int64_t mask_filter = 0;
2993 int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
2995 vp9_zero(best_mbmode);
2997 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
2999 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3000 filter_cache[i] = INT64_MAX;
3002 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3005 for (i = 0; i < REFERENCE_MODES; ++i)
3006 best_pred_rd[i] = INT64_MAX;
3007 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3008 best_filter_rd[i] = INT64_MAX;
3009 for (i = 0; i < TX_SIZES; i++)
3010 rate_uv_intra[i] = INT_MAX;
3011 for (i = 0; i < MAX_REF_FRAMES; ++i)
3012 x->pred_sse[i] = INT_MAX;
3013 for (i = 0; i < MB_MODE_COUNT; ++i) {
3014 for (k = 0; k < MAX_REF_FRAMES; ++k) {
3015 single_inter_filter[i][k] = SWITCHABLE;
3016 single_skippable[i][k] = 0;
3020 rd_cost->rate = INT_MAX;
3022 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3023 x->pred_mv_sad[ref_frame] = INT_MAX;
3024 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3025 assert(get_ref_frame_buffer(cpi, ref_frame) != NULL);
3026 setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
3027 frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
3029 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3030 frame_mv[ZEROMV][ref_frame].as_int = 0;
3033 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3034 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) {
3035 // Skip checking missing references in both single and compound reference
3036 // modes. Note that a mode will be skipped if both reference frames
3038 ref_frame_skip_mask[0] |= (1 << ref_frame);
3039 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3040 } else if (sf->reference_masking) {
3041 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
3042 // Skip fixed mv modes for poor references
3043 if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
3044 mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
3049 // If the segment reference frame feature is enabled....
3050 // then do nothing if the current ref frame is not allowed..
3051 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3052 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3053 ref_frame_skip_mask[0] |= (1 << ref_frame);
3054 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3058 // Disable this drop out case if the ref frame
3059 // segment level feature is enabled for this segment. This is to
3060 // prevent the possibility that we end up unable to pick any mode.
3061 if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3062 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3063 // unless ARNR filtering is enabled in which case we want
3064 // an unfiltered alternative. We allow near/nearest as well
3065 // because they may result in zero-zero MVs but be cheaper.
3066 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
3067 ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME);
3068 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3069 mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
3070 if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
3071 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
3072 if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
3073 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
3077 if (cpi->rc.is_src_frame_alt_ref) {
3078 if (sf->alt_ref_search_fp) {
3079 mode_skip_mask[ALTREF_FRAME] = 0;
3080 ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME);
3081 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3085 if (sf->alt_ref_search_fp)
3086 if (!cm->show_frame && x->pred_mv_sad[GOLDEN_FRAME] < INT_MAX)
3087 if (x->pred_mv_sad[ALTREF_FRAME] > (x->pred_mv_sad[GOLDEN_FRAME] << 1))
3088 mode_skip_mask[ALTREF_FRAME] |= INTER_ALL;
3090 if (sf->adaptive_mode_search) {
3091 if (cm->show_frame && !cpi->rc.is_src_frame_alt_ref &&
3092 cpi->rc.frames_since_golden >= 3)
3093 if (x->pred_mv_sad[GOLDEN_FRAME] > (x->pred_mv_sad[LAST_FRAME] << 1))
3094 mode_skip_mask[GOLDEN_FRAME] |= INTER_ALL;
3097 if (bsize > sf->max_intra_bsize) {
3098 ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
3099 ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
3102 mode_skip_mask[INTRA_FRAME] |=
3103 ~(sf->intra_y_mode_mask[max_txsize_lookup[bsize]]);
3105 for (i = 0; i <= LAST_NEW_MV_INDEX; ++i)
3106 mode_threshold[i] = 0;
3107 for (i = LAST_NEW_MV_INDEX + 1; i < MAX_MODES; ++i)
3108 mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
3110 midx = sf->schedule_mode_search ? mode_skip_start : 0;
3112 uint8_t end_pos = 0;
3113 for (i = 5; i < midx; ++i) {
3114 if (mode_threshold[mode_map[i - 1]] > mode_threshold[mode_map[i]]) {
3115 uint8_t tmp = mode_map[i];
3116 mode_map[i] = mode_map[i - 1];
3117 mode_map[i - 1] = tmp;
3124 for (midx = 0; midx < MAX_MODES; ++midx) {
3125 int mode_index = mode_map[midx];
3126 int mode_excluded = 0;
3127 int64_t this_rd = INT64_MAX;
3128 int disable_skip = 0;
3129 int compmode_cost = 0;
3130 int rate2 = 0, rate_y = 0, rate_uv = 0;
3131 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3134 int64_t total_sse = INT64_MAX;
3137 this_mode = vp9_mode_order[mode_index].mode;
3138 ref_frame = vp9_mode_order[mode_index].ref_frame[0];
3139 second_ref_frame = vp9_mode_order[mode_index].ref_frame[1];
3141 // Look at the reference frame of the best mode so far and set the
3142 // skip mask to look at a subset of the remaining modes.
3143 if (midx == mode_skip_start && best_mode_index >= 0) {
3144 switch (best_mbmode.ref_frame[0]) {
3148 ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK;
3149 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3152 ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
3153 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3156 ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK;
3159 case MAX_REF_FRAMES:
3160 assert(0 && "Invalid Reference frame");
3165 if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3166 (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
3169 if (mode_skip_mask[ref_frame] & (1 << this_mode))
3172 // Test best rd so far against threshold for trying this mode.
3173 if (best_mode_skippable && sf->schedule_mode_search)
3174 mode_threshold[mode_index] <<= 1;
3176 if (best_rd < mode_threshold[mode_index])
3179 if (sf->motion_field_mode_search) {
3180 const int mi_width = VPXMIN(num_8x8_blocks_wide_lookup[bsize],
3181 tile_info->mi_col_end - mi_col);
3182 const int mi_height = VPXMIN(num_8x8_blocks_high_lookup[bsize],
3183 tile_info->mi_row_end - mi_row);
3184 const int bsl = mi_width_log2_lookup[bsize];
3185 int cb_partition_search_ctrl = (((mi_row + mi_col) >> bsl)
3186 + get_chessboard_index(cm->current_video_frame)) & 0x1;
3188 int const_motion = 1;
3189 int skip_ref_frame = !cb_partition_search_ctrl;
3190 MV_REFERENCE_FRAME rf = NONE;
3192 ref_mv.as_int = INVALID_MV;
3194 if ((mi_row - 1) >= tile_info->mi_row_start) {
3195 ref_mv = xd->mi[-xd->mi_stride]->mv[0];
3196 rf = xd->mi[-xd->mi_stride]->ref_frame[0];
3197 for (i = 0; i < mi_width; ++i) {
3198 ref_mi = xd->mi[-xd->mi_stride + i];
3199 const_motion &= (ref_mv.as_int == ref_mi->mv[0].as_int) &&
3200 (ref_frame == ref_mi->ref_frame[0]);
3201 skip_ref_frame &= (rf == ref_mi->ref_frame[0]);
3205 if ((mi_col - 1) >= tile_info->mi_col_start) {
3206 if (ref_mv.as_int == INVALID_MV)
3207 ref_mv = xd->mi[-1]->mv[0];
3209 rf = xd->mi[-1]->ref_frame[0];
3210 for (i = 0; i < mi_height; ++i) {
3211 ref_mi = xd->mi[i * xd->mi_stride - 1];
3212 const_motion &= (ref_mv.as_int == ref_mi->mv[0].as_int) &&
3213 (ref_frame == ref_mi->ref_frame[0]);
3214 skip_ref_frame &= (rf == ref_mi->ref_frame[0]);
3218 if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV)
3219 if (rf > INTRA_FRAME)
3220 if (ref_frame != rf)
3224 if (this_mode == NEARMV || this_mode == ZEROMV)
3228 comp_pred = second_ref_frame > INTRA_FRAME;
3230 if (!cpi->allow_comp_inter_inter)
3233 // Skip compound inter modes if ARF is not available.
3234 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
3237 // Do not allow compound prediction if the segment level reference frame
3238 // feature is in use as in this case there can only be one reference.
3239 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
3242 if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3243 best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
3246 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3248 if (ref_frame != INTRA_FRAME)
3249 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3252 if (ref_frame == INTRA_FRAME) {
3253 if (sf->adaptive_mode_search)
3254 if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
3257 if (this_mode != DC_PRED) {
3258 // Disable intra modes other than DC_PRED for blocks with low variance
3259 // Threshold for intra skipping based on source variance
3260 // TODO(debargha): Specialize the threshold for super block sizes
3261 const unsigned int skip_intra_var_thresh = 64;
3262 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
3263 x->source_variance < skip_intra_var_thresh)
3265 // Only search the oblique modes if the best so far is
3266 // one of the neighboring directional modes
3267 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
3268 (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
3269 if (best_mode_index >= 0 &&
3270 best_mbmode.ref_frame[0] > INTRA_FRAME)
3273 if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
3274 if (conditional_skipintra(this_mode, best_intra_mode))
3279 const MV_REFERENCE_FRAME ref_frames[2] = {ref_frame, second_ref_frame};
3280 if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
3281 this_mode, ref_frames))
3285 mi->mode = this_mode;
3286 mi->uv_mode = DC_PRED;
3287 mi->ref_frame[0] = ref_frame;
3288 mi->ref_frame[1] = second_ref_frame;
3289 // Evaluate all sub-pel filters irrespective of whether we can use
3290 // them for this frame.
3291 mi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
3292 : cm->interp_filter;
3293 mi->mv[0].as_int = mi->mv[1].as_int = 0;
3296 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3298 // Select prediction reference frames.
3299 for (i = 0; i < MAX_MB_PLANE; i++) {
3300 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3302 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3305 if (ref_frame == INTRA_FRAME) {
3307 struct macroblockd_plane *const pd = &xd->plane[1];
3308 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
3309 super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable,
3310 NULL, bsize, best_rd);
3311 if (rate_y == INT_MAX)
3314 uv_tx = get_uv_tx_size_impl(mi->tx_size, bsize, pd->subsampling_x,
3316 if (rate_uv_intra[uv_tx] == INT_MAX) {
3317 choose_intra_uv_mode(cpi, x, ctx, bsize, uv_tx,
3318 &rate_uv_intra[uv_tx], &rate_uv_tokenonly[uv_tx],
3319 &dist_uv[uv_tx], &skip_uv[uv_tx], &mode_uv[uv_tx]);
3322 rate_uv = rate_uv_tokenonly[uv_tx];
3323 distortion_uv = dist_uv[uv_tx];
3324 skippable = skippable && skip_uv[uv_tx];
3325 mi->uv_mode = mode_uv[uv_tx];
3327 rate2 = rate_y + cpi->mbmode_cost[mi->mode] + rate_uv_intra[uv_tx];
3328 if (this_mode != DC_PRED && this_mode != TM_PRED)
3329 rate2 += intra_cost_penalty;
3330 distortion2 = distortion_y + distortion_uv;
3332 this_rd = handle_inter_mode(cpi, x, bsize,
3333 &rate2, &distortion2, &skippable,
3335 &disable_skip, frame_mv,
3337 single_newmv, single_inter_filter,
3338 single_skippable, &total_sse, best_rd,
3339 &mask_filter, filter_cache);
3340 if (this_rd == INT64_MAX)
3343 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
3345 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3346 rate2 += compmode_cost;
3349 // Estimate the reference frame signaling cost and add it
3350 // to the rolling cost variable.
3352 rate2 += ref_costs_comp[ref_frame];
3354 rate2 += ref_costs_single[ref_frame];
3357 if (!disable_skip) {
3359 // Back out the coefficient coding costs
3360 rate2 -= (rate_y + rate_uv);
3362 // Cost the skip mb case
3363 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
3364 } else if (ref_frame != INTRA_FRAME && !xd->lossless) {
3365 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
3366 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
3367 // Add in the cost of the no skip flag.
3368 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
3370 // FIXME(rbultje) make this work for splitmv also
3371 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
3372 distortion2 = total_sse;
3373 assert(total_sse >= 0);
3374 rate2 -= (rate_y + rate_uv);
3378 // Add in the cost of the no skip flag.
3379 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
3382 // Calculate the final RD estimate for this mode.
3383 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3386 // Apply an adjustment to the rd value based on the similarity of the
3387 // source variance and reconstructed variance.
3388 rd_variance_adjustment(cpi, x, bsize, &this_rd,
3389 ref_frame, x->source_variance);
3391 if (ref_frame == INTRA_FRAME) {
3392 // Keep record of best intra rd
3393 if (this_rd < best_intra_rd) {
3394 best_intra_rd = this_rd;
3395 best_intra_mode = mi->mode;
3399 if (!disable_skip && ref_frame == INTRA_FRAME) {
3400 for (i = 0; i < REFERENCE_MODES; ++i)
3401 best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
3402 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3403 best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
3406 // Did this mode help.. i.e. is it the new best mode
3407 if (this_rd < best_rd || x->skip) {
3408 int max_plane = MAX_MB_PLANE;
3409 if (!mode_excluded) {
3410 // Note index of best mode so far
3411 best_mode_index = mode_index;
3413 if (ref_frame == INTRA_FRAME) {
3414 /* required for left and above block mv */
3415 mi->mv[0].as_int = 0;
3418 best_pred_sse = x->pred_sse[ref_frame];
3421 rd_cost->rate = rate2;
3422 rd_cost->dist = distortion2;
3423 rd_cost->rdcost = this_rd;
3426 best_skip2 = this_skip2;
3427 best_mode_skippable = skippable;
3429 if (!x->select_tx_size)
3430 swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
3431 memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mi->tx_size],
3432 sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
3434 // TODO(debargha): enhance this test with a better distortion prediction
3435 // based on qp, activity mask and history
3436 if ((mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
3437 (mode_index > MIN_EARLY_TERM_INDEX)) {
3438 int qstep = xd->plane[0].dequant[1];
3439 // TODO(debargha): Enhance this by specializing for each mode_index
3441 #if CONFIG_VP9_HIGHBITDEPTH
3442 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
3443 qstep >>= (xd->bd - 8);
3445 #endif // CONFIG_VP9_HIGHBITDEPTH
3446 if (x->source_variance < UINT_MAX) {
3447 const int var_adjust = (x->source_variance < 16);
3448 scale -= var_adjust;
3450 if (ref_frame > INTRA_FRAME &&
3451 distortion2 * scale < qstep * qstep) {
3458 /* keep record of best compound/single-only prediction */
3459 if (!disable_skip && ref_frame != INTRA_FRAME) {
3460 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
3462 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3463 single_rate = rate2 - compmode_cost;
3464 hybrid_rate = rate2;
3466 single_rate = rate2;
3467 hybrid_rate = rate2 + compmode_cost;
3470 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
3471 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
3474 if (single_rd < best_pred_rd[SINGLE_REFERENCE])
3475 best_pred_rd[SINGLE_REFERENCE] = single_rd;
3477 if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
3478 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
3480 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
3481 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
3483 /* keep record of best filter type */
3484 if (!mode_excluded && cm->interp_filter != BILINEAR) {
3485 int64_t ref = filter_cache[cm->interp_filter == SWITCHABLE ?
3486 SWITCHABLE_FILTERS : cm->interp_filter];
3488 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3490 if (ref == INT64_MAX)
3492 else if (filter_cache[i] == INT64_MAX)
3493 // when early termination is triggered, the encoder does not have
3494 // access to the rate-distortion cost. it only knows that the cost
3495 // should be above the maximum valid value. hence it takes the known
3496 // maximum plus an arbitrary constant as the rate-distortion cost.
3497 adj_rd = mask_filter - ref + 10;
3499 adj_rd = filter_cache[i] - ref;
3502 best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
3510 if (x->skip && !comp_pred)
3514 // The inter modes' rate costs are not calculated precisely in some cases.
3515 // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and
3516 // ZEROMV. Here, checks are added for those cases, and the mode decisions
3518 if (best_mbmode.mode == NEWMV) {
3519 const MV_REFERENCE_FRAME refs[2] = {best_mbmode.ref_frame[0],
3520 best_mbmode.ref_frame[1]};
3521 int comp_pred_mode = refs[1] > INTRA_FRAME;
3523 if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3524 ((comp_pred_mode && frame_mv[NEARESTMV][refs[1]].as_int ==
3525 best_mbmode.mv[1].as_int) || !comp_pred_mode))
3526 best_mbmode.mode = NEARESTMV;
3527 else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3528 ((comp_pred_mode && frame_mv[NEARMV][refs[1]].as_int ==
3529 best_mbmode.mv[1].as_int) || !comp_pred_mode))
3530 best_mbmode.mode = NEARMV;
3531 else if (best_mbmode.mv[0].as_int == 0 &&
3532 ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) || !comp_pred_mode))
3533 best_mbmode.mode = ZEROMV;
3536 if (best_mode_index < 0 || best_rd >= best_rd_so_far) {
3537 rd_cost->rate = INT_MAX;
3538 rd_cost->rdcost = INT64_MAX;
3542 // If we used an estimate for the uv intra rd in the loop above...
3543 if (sf->use_uv_intra_rd_estimate) {
3544 // Do Intra UV best rd mode selection if best mode choice above was intra.
3545 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
3548 uv_tx_size = get_uv_tx_size(mi, &xd->plane[1]);
3549 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
3550 &rate_uv_tokenonly[uv_tx_size],
3551 &dist_uv[uv_tx_size],
3552 &skip_uv[uv_tx_size],
3553 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
3558 assert((cm->interp_filter == SWITCHABLE) ||
3559 (cm->interp_filter == best_mbmode.interp_filter) ||
3560 !is_inter_block(&best_mbmode));
3562 if (!cpi->rc.is_src_frame_alt_ref)
3563 vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3564 sf->adaptive_rd_thresh, bsize, best_mode_index);
3568 x->skip |= best_skip2;
3570 for (i = 0; i < REFERENCE_MODES; ++i) {
3571 if (best_pred_rd[i] == INT64_MAX)
3572 best_pred_diff[i] = INT_MIN;
3574 best_pred_diff[i] = best_rd - best_pred_rd[i];
3578 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3579 if (best_filter_rd[i] == INT64_MAX)
3580 best_filter_diff[i] = 0;
3582 best_filter_diff[i] = best_rd - best_filter_rd[i];
3584 if (cm->interp_filter == SWITCHABLE)
3585 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
3587 vp9_zero(best_filter_diff);
3590 // TODO(yunqingwang): Moving this line in front of the above best_filter_diff
3591 // updating code causes PSNR loss. Need to figure out the confliction.
3592 x->skip |= best_mode_skippable;
3594 if (!x->skip && !x->select_tx_size) {
3595 int has_high_freq_coeff = 0;
3597 int max_plane = is_inter_block(xd->mi[0])
3599 for (plane = 0; plane < max_plane; ++plane) {
3600 x->plane[plane].eobs = ctx->eobs_pbuf[plane][1];
3601 has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
3604 for (plane = max_plane; plane < MAX_MB_PLANE; ++plane) {
3605 x->plane[plane].eobs = ctx->eobs_pbuf[plane][2];
3606 has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
3609 best_mode_skippable |= !has_high_freq_coeff;
3612 assert(best_mode_index >= 0);
3614 store_coding_context(x, ctx, best_mode_index, best_pred_diff,
3615 best_filter_diff, best_mode_skippable);
3618 void vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi,
3619 TileDataEnc *tile_data,
3623 PICK_MODE_CONTEXT *ctx,
3624 int64_t best_rd_so_far) {
3625 VP9_COMMON *const cm = &cpi->common;
3626 MACROBLOCKD *const xd = &x->e_mbd;
3627 MODE_INFO *const mi = xd->mi[0];
3628 unsigned char segment_id = mi->segment_id;
3629 const int comp_pred = 0;
3631 int64_t best_pred_diff[REFERENCE_MODES];
3632 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3633 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3634 vpx_prob comp_mode_p;
3635 INTERP_FILTER best_filter = SWITCHABLE;
3636 int64_t this_rd = INT64_MAX;
3638 const int64_t distortion2 = 0;
3640 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3642 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3645 for (i = 0; i < MAX_REF_FRAMES; ++i)
3646 x->pred_sse[i] = INT_MAX;
3647 for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i)
3648 x->pred_mv_sad[i] = INT_MAX;
3650 rd_cost->rate = INT_MAX;
3652 assert(segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
3655 mi->uv_mode = DC_PRED;
3656 mi->ref_frame[0] = LAST_FRAME;
3657 mi->ref_frame[1] = NONE;
3658 mi->mv[0].as_int = 0;
3661 if (cm->interp_filter != BILINEAR) {
3662 best_filter = EIGHTTAP;
3663 if (cm->interp_filter == SWITCHABLE &&
3664 x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
3666 int best_rs = INT_MAX;
3667 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
3668 mi->interp_filter = i;
3669 rs = vp9_get_switchable_rate(cpi, xd);
3672 best_filter = mi->interp_filter;
3677 // Set the appropriate filter
3678 if (cm->interp_filter == SWITCHABLE) {
3679 mi->interp_filter = best_filter;
3680 rate2 += vp9_get_switchable_rate(cpi, xd);
3682 mi->interp_filter = cm->interp_filter;
3685 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3686 rate2 += vp9_cost_bit(comp_mode_p, comp_pred);
3688 // Estimate the reference frame signaling cost and add it
3689 // to the rolling cost variable.
3690 rate2 += ref_costs_single[LAST_FRAME];
3691 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3693 rd_cost->rate = rate2;
3694 rd_cost->dist = distortion2;
3695 rd_cost->rdcost = this_rd;
3697 if (this_rd >= best_rd_so_far) {
3698 rd_cost->rate = INT_MAX;
3699 rd_cost->rdcost = INT64_MAX;
3703 assert((cm->interp_filter == SWITCHABLE) ||
3704 (cm->interp_filter == mi->interp_filter));
3706 vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3707 cpi->sf.adaptive_rd_thresh, bsize, THR_ZEROMV);
3709 vp9_zero(best_pred_diff);
3710 vp9_zero(best_filter_diff);
3712 if (!x->select_tx_size)
3713 swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE);
3714 store_coding_context(x, ctx, THR_ZEROMV,
3715 best_pred_diff, best_filter_diff, 0);
3718 void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi,
3719 TileDataEnc *tile_data,
3721 int mi_row, int mi_col,
3724 PICK_MODE_CONTEXT *ctx,
3725 int64_t best_rd_so_far) {
3726 VP9_COMMON *const cm = &cpi->common;
3727 RD_OPT *const rd_opt = &cpi->rd;
3728 SPEED_FEATURES *const sf = &cpi->sf;
3729 MACROBLOCKD *const xd = &x->e_mbd;
3730 MODE_INFO *const mi = xd->mi[0];
3731 const struct segmentation *const seg = &cm->seg;
3732 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3733 unsigned char segment_id = mi->segment_id;
3735 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3736 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3737 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
3739 int64_t best_rd = best_rd_so_far;
3740 int64_t best_yrd = best_rd_so_far; // FIXME(rbultje) more precise
3741 int64_t best_pred_diff[REFERENCE_MODES];
3742 int64_t best_pred_rd[REFERENCE_MODES];
3743 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3744 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3745 MODE_INFO best_mbmode;
3746 int ref_index, best_ref_index = 0;
3747 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3748 vpx_prob comp_mode_p;
3749 INTERP_FILTER tmp_best_filter = SWITCHABLE;
3750 int rate_uv_intra, rate_uv_tokenonly;
3753 PREDICTION_MODE mode_uv = DC_PRED;
3754 const int intra_cost_penalty = vp9_get_intra_cost_penalty(
3755 cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
3756 int_mv seg_mvs[4][MAX_REF_FRAMES];
3757 b_mode_info best_bmodes[4];
3759 int ref_frame_skip_mask[2] = { 0 };
3760 int64_t mask_filter = 0;
3761 int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
3762 int internal_active_edge =
3763 vp9_active_edge_sb(cpi, mi_row, mi_col) && vp9_internal_image_edge(cpi);
3765 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3766 memset(x->zcoeff_blk[TX_4X4], 0, 4);
3767 vp9_zero(best_mbmode);
3769 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3770 filter_cache[i] = INT64_MAX;
3772 for (i = 0; i < 4; i++) {
3774 for (j = 0; j < MAX_REF_FRAMES; j++)
3775 seg_mvs[i][j].as_int = INVALID_MV;
3778 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3781 for (i = 0; i < REFERENCE_MODES; ++i)
3782 best_pred_rd[i] = INT64_MAX;
3783 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3784 best_filter_rd[i] = INT64_MAX;
3785 rate_uv_intra = INT_MAX;
3787 rd_cost->rate = INT_MAX;
3789 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
3790 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3791 setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
3792 frame_mv[NEARESTMV], frame_mv[NEARMV],
3795 ref_frame_skip_mask[0] |= (1 << ref_frame);
3796 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3798 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3799 frame_mv[ZEROMV][ref_frame].as_int = 0;
3802 for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
3803 int mode_excluded = 0;
3804 int64_t this_rd = INT64_MAX;
3805 int disable_skip = 0;
3806 int compmode_cost = 0;
3807 int rate2 = 0, rate_y = 0, rate_uv = 0;
3808 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3812 int64_t total_sse = INT_MAX;
3814 struct buf_2d backup_yv12[2][MAX_MB_PLANE];
3816 ref_frame = vp9_ref_order[ref_index].ref_frame[0];
3817 second_ref_frame = vp9_ref_order[ref_index].ref_frame[1];
3819 #if CONFIG_BETTER_HW_COMPATIBILITY
3820 // forbid 8X4 and 4X8 partitions if any reference frame is scaled.
3821 if (bsize == BLOCK_8X4 || bsize == BLOCK_4X8) {
3822 int ref_scaled = vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf);
3823 if (second_ref_frame > INTRA_FRAME)
3824 ref_scaled += vp9_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf);
3829 // Look at the reference frame of the best mode so far and set the
3830 // skip mask to look at a subset of the remaining modes.
3831 if (ref_index > 2 && sf->mode_skip_start < MAX_MODES) {
3832 if (ref_index == 3) {
3833 switch (best_mbmode.ref_frame[0]) {
3837 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME);
3838 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3841 ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME);
3842 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3845 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME);
3848 case MAX_REF_FRAMES:
3849 assert(0 && "Invalid Reference frame");
3855 if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3856 (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
3859 // Test best rd so far against threshold for trying this mode.
3860 if (!internal_active_edge &&
3861 rd_less_than_thresh(best_rd,
3862 rd_opt->threshes[segment_id][bsize][ref_index],
3863 tile_data->thresh_freq_fact[bsize][ref_index]))
3866 comp_pred = second_ref_frame > INTRA_FRAME;
3868 if (!cpi->allow_comp_inter_inter)
3870 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
3872 // Do not allow compound prediction if the segment level reference frame
3873 // feature is in use as in this case there can only be one reference.
3874 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
3877 if ((sf->mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3878 best_mbmode.ref_frame[0] == INTRA_FRAME)
3883 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3884 else if (ref_frame != INTRA_FRAME)
3885 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3887 // If the segment reference frame feature is enabled....
3888 // then do nothing if the current ref frame is not allowed..
3889 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3890 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3892 // Disable this drop out case if the ref frame
3893 // segment level feature is enabled for this segment. This is to
3894 // prevent the possibility that we end up unable to pick any mode.
3895 } else if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3896 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3897 // unless ARNR filtering is enabled in which case we want
3898 // an unfiltered alternative. We allow near/nearest as well
3899 // because they may result in zero-zero MVs but be cheaper.
3900 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
3904 mi->tx_size = TX_4X4;
3905 mi->uv_mode = DC_PRED;
3906 mi->ref_frame[0] = ref_frame;
3907 mi->ref_frame[1] = second_ref_frame;
3908 // Evaluate all sub-pel filters irrespective of whether we can use
3909 // them for this frame.
3910 mi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
3911 : cm->interp_filter;
3913 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3915 // Select prediction reference frames.
3916 for (i = 0; i < MAX_MB_PLANE; i++) {
3917 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3919 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3922 if (ref_frame == INTRA_FRAME) {
3924 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y,
3925 &distortion_y, best_rd) >= best_rd)
3928 rate2 += intra_cost_penalty;
3929 distortion2 += distortion_y;
3931 if (rate_uv_intra == INT_MAX) {
3932 choose_intra_uv_mode(cpi, x, ctx, bsize, TX_4X4,
3938 rate2 += rate_uv_intra;
3939 rate_uv = rate_uv_tokenonly;
3940 distortion2 += dist_uv;
3941 distortion_uv = dist_uv;
3942 mi->uv_mode = mode_uv;
3946 int64_t this_rd_thresh;
3947 int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
3948 int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
3949 int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
3950 int tmp_best_skippable = 0;
3951 int switchable_filter_index;
3952 int_mv *second_ref = comp_pred ?
3953 &x->mbmi_ext->ref_mvs[second_ref_frame][0] : NULL;
3954 b_mode_info tmp_best_bmodes[16];
3955 MODE_INFO tmp_best_mbmode;
3956 BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
3957 int pred_exists = 0;
3960 YV12_BUFFER_CONFIG *scaled_ref_frame[2] = {NULL, NULL};
3963 for (ref = 0; ref < 2; ++ref) {
3964 scaled_ref_frame[ref] = mi->ref_frame[ref] > INTRA_FRAME ?
3965 vp9_get_scaled_ref_frame(cpi, mi->ref_frame[ref]) : NULL;
3967 if (scaled_ref_frame[ref]) {
3969 // Swap out the reference frame for a version that's been scaled to
3970 // match the resolution of the current frame, allowing the existing
3971 // motion search code to be used without additional modifications.
3972 for (i = 0; i < MAX_MB_PLANE; i++)
3973 backup_yv12[ref][i] = xd->plane[i].pre[ref];
3974 vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
3979 this_rd_thresh = (ref_frame == LAST_FRAME) ?
3980 rd_opt->threshes[segment_id][bsize][THR_LAST] :
3981 rd_opt->threshes[segment_id][bsize][THR_ALTR];
3982 this_rd_thresh = (ref_frame == GOLDEN_FRAME) ?
3983 rd_opt->threshes[segment_id][bsize][THR_GOLD] : this_rd_thresh;
3984 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3985 filter_cache[i] = INT64_MAX;
3987 if (cm->interp_filter != BILINEAR) {
3988 tmp_best_filter = EIGHTTAP;
3989 if (x->source_variance < sf->disable_filter_search_var_thresh) {
3990 tmp_best_filter = EIGHTTAP;
3991 } else if (sf->adaptive_pred_interp_filter == 1 &&
3992 ctx->pred_interp_filter < SWITCHABLE) {
3993 tmp_best_filter = ctx->pred_interp_filter;
3994 } else if (sf->adaptive_pred_interp_filter == 2) {
3995 tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE ?
3996 ctx->pred_interp_filter : 0;
3998 for (switchable_filter_index = 0;
3999 switchable_filter_index < SWITCHABLE_FILTERS;
4000 ++switchable_filter_index) {
4003 MB_MODE_INFO_EXT *mbmi_ext = x->mbmi_ext;
4004 mi->interp_filter = switchable_filter_index;
4005 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x,
4006 &mbmi_ext->ref_mvs[ref_frame][0],
4007 second_ref, best_yrd, &rate,
4008 &rate_y, &distortion,
4009 &skippable, &total_sse,
4010 (int) this_rd_thresh, seg_mvs,
4011 bsi, switchable_filter_index,
4014 if (tmp_rd == INT64_MAX)
4016 rs = vp9_get_switchable_rate(cpi, xd);
4017 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
4018 filter_cache[switchable_filter_index] = tmp_rd;
4019 filter_cache[SWITCHABLE_FILTERS] =
4020 VPXMIN(filter_cache[SWITCHABLE_FILTERS], tmp_rd + rs_rd);
4021 if (cm->interp_filter == SWITCHABLE)
4024 mask_filter = VPXMAX(mask_filter, tmp_rd);
4026 newbest = (tmp_rd < tmp_best_rd);
4028 tmp_best_filter = mi->interp_filter;
4029 tmp_best_rd = tmp_rd;
4031 if ((newbest && cm->interp_filter == SWITCHABLE) ||
4032 (mi->interp_filter == cm->interp_filter &&
4033 cm->interp_filter != SWITCHABLE)) {
4034 tmp_best_rdu = tmp_rd;
4035 tmp_best_rate = rate;
4036 tmp_best_ratey = rate_y;
4037 tmp_best_distortion = distortion;
4038 tmp_best_sse = total_sse;
4039 tmp_best_skippable = skippable;
4040 tmp_best_mbmode = *mi;
4041 for (i = 0; i < 4; i++) {
4042 tmp_best_bmodes[i] = xd->mi[0]->bmi[i];
4043 x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
4046 if (switchable_filter_index == 0 &&
4047 sf->use_rd_breakout &&
4048 best_rd < INT64_MAX) {
4049 if (tmp_best_rdu / 2 > best_rd) {
4050 // skip searching the other filters if the first is
4051 // already substantially larger than the best so far
4052 tmp_best_filter = mi->interp_filter;
4053 tmp_best_rdu = INT64_MAX;
4058 } // switchable_filter_index loop
4062 if (tmp_best_rdu == INT64_MAX && pred_exists)
4065 mi->interp_filter = (cm->interp_filter == SWITCHABLE ?
4066 tmp_best_filter : cm->interp_filter);
4068 // Handles the special case when a filter that is not in the
4069 // switchable list (bilinear, 6-tap) is indicated at the frame level
4070 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x,
4071 &x->mbmi_ext->ref_mvs[ref_frame][0],
4072 second_ref, best_yrd, &rate, &rate_y,
4073 &distortion, &skippable, &total_sse,
4074 (int) this_rd_thresh, seg_mvs, bsi, 0,
4076 if (tmp_rd == INT64_MAX)
4079 total_sse = tmp_best_sse;
4080 rate = tmp_best_rate;
4081 rate_y = tmp_best_ratey;
4082 distortion = tmp_best_distortion;
4083 skippable = tmp_best_skippable;
4084 *mi = tmp_best_mbmode;
4085 for (i = 0; i < 4; i++)
4086 xd->mi[0]->bmi[i] = tmp_best_bmodes[i];
4090 distortion2 += distortion;
4092 if (cm->interp_filter == SWITCHABLE)
4093 rate2 += vp9_get_switchable_rate(cpi, xd);
4096 mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
4097 : cm->reference_mode == COMPOUND_REFERENCE;
4099 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
4102 best_rd - VPXMIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
4103 RDCOST(x->rdmult, x->rddiv, 0, total_sse));
4105 if (tmp_best_rdu > 0) {
4106 // If even the 'Y' rd value of split is higher than best so far
4107 // then dont bother looking at UV
4108 vp9_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col,
4110 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
4111 if (!super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
4112 &uv_sse, BLOCK_8X8, tmp_best_rdu)) {
4113 for (ref = 0; ref < 2; ++ref) {
4114 if (scaled_ref_frame[ref]) {
4116 for (i = 0; i < MAX_MB_PLANE; ++i)
4117 xd->plane[i].pre[ref] = backup_yv12[ref][i];
4124 distortion2 += distortion_uv;
4125 skippable = skippable && uv_skippable;
4126 total_sse += uv_sse;
4129 for (ref = 0; ref < 2; ++ref) {
4130 if (scaled_ref_frame[ref]) {
4131 // Restore the prediction frame pointers to their unscaled versions.
4133 for (i = 0; i < MAX_MB_PLANE; ++i)
4134 xd->plane[i].pre[ref] = backup_yv12[ref][i];
4139 if (cm->reference_mode == REFERENCE_MODE_SELECT)
4140 rate2 += compmode_cost;
4142 // Estimate the reference frame signaling cost and add it
4143 // to the rolling cost variable.
4144 if (second_ref_frame > INTRA_FRAME) {
4145 rate2 += ref_costs_comp[ref_frame];
4147 rate2 += ref_costs_single[ref_frame];
4150 if (!disable_skip) {
4151 // Skip is never coded at the segment level for sub8x8 blocks and instead
4152 // always coded in the bitstream at the mode info level.
4154 if (ref_frame != INTRA_FRAME && !xd->lossless) {
4155 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
4156 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
4157 // Add in the cost of the no skip flag.
4158 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
4160 // FIXME(rbultje) make this work for splitmv also
4161 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
4162 distortion2 = total_sse;
4163 assert(total_sse >= 0);
4164 rate2 -= (rate_y + rate_uv);
4170 // Add in the cost of the no skip flag.
4171 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
4174 // Calculate the final RD estimate for this mode.
4175 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
4178 if (!disable_skip && ref_frame == INTRA_FRAME) {
4179 for (i = 0; i < REFERENCE_MODES; ++i)
4180 best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
4181 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
4182 best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
4185 // Did this mode help.. i.e. is it the new best mode
4186 if (this_rd < best_rd || x->skip) {
4187 if (!mode_excluded) {
4188 int max_plane = MAX_MB_PLANE;
4189 // Note index of best mode so far
4190 best_ref_index = ref_index;
4192 if (ref_frame == INTRA_FRAME) {
4193 /* required for left and above block mv */
4194 mi->mv[0].as_int = 0;
4198 rd_cost->rate = rate2;
4199 rd_cost->dist = distortion2;
4200 rd_cost->rdcost = this_rd;
4202 best_yrd = best_rd -
4203 RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
4205 best_skip2 = this_skip2;
4206 if (!x->select_tx_size)
4207 swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
4208 memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
4209 sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
4211 for (i = 0; i < 4; i++)
4212 best_bmodes[i] = xd->mi[0]->bmi[i];
4214 // TODO(debargha): enhance this test with a better distortion prediction
4215 // based on qp, activity mask and history
4216 if ((sf->mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
4217 (ref_index > MIN_EARLY_TERM_INDEX)) {
4218 int qstep = xd->plane[0].dequant[1];
4219 // TODO(debargha): Enhance this by specializing for each mode_index
4221 #if CONFIG_VP9_HIGHBITDEPTH
4222 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
4223 qstep >>= (xd->bd - 8);
4225 #endif // CONFIG_VP9_HIGHBITDEPTH
4226 if (x->source_variance < UINT_MAX) {
4227 const int var_adjust = (x->source_variance < 16);
4228 scale -= var_adjust;
4230 if (ref_frame > INTRA_FRAME &&
4231 distortion2 * scale < qstep * qstep) {
4238 /* keep record of best compound/single-only prediction */
4239 if (!disable_skip && ref_frame != INTRA_FRAME) {
4240 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
4242 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4243 single_rate = rate2 - compmode_cost;
4244 hybrid_rate = rate2;
4246 single_rate = rate2;
4247 hybrid_rate = rate2 + compmode_cost;
4250 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
4251 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
4253 if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
4254 best_pred_rd[SINGLE_REFERENCE] = single_rd;
4255 else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
4256 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
4258 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
4259 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
4262 /* keep record of best filter type */
4263 if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
4264 cm->interp_filter != BILINEAR) {
4265 int64_t ref = filter_cache[cm->interp_filter == SWITCHABLE ?
4266 SWITCHABLE_FILTERS : cm->interp_filter];
4268 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4269 if (ref == INT64_MAX)
4271 else if (filter_cache[i] == INT64_MAX)
4272 // when early termination is triggered, the encoder does not have
4273 // access to the rate-distortion cost. it only knows that the cost
4274 // should be above the maximum valid value. hence it takes the known
4275 // maximum plus an arbitrary constant as the rate-distortion cost.
4276 adj_rd = mask_filter - ref + 10;
4278 adj_rd = filter_cache[i] - ref;
4281 best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
4288 if (x->skip && !comp_pred)
4292 if (best_rd >= best_rd_so_far) {
4293 rd_cost->rate = INT_MAX;
4294 rd_cost->rdcost = INT64_MAX;
4298 // If we used an estimate for the uv intra rd in the loop above...
4299 if (sf->use_uv_intra_rd_estimate) {
4300 // Do Intra UV best rd mode selection if best mode choice above was intra.
4301 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
4303 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra,
4311 if (best_rd == INT64_MAX) {
4312 rd_cost->rate = INT_MAX;
4313 rd_cost->dist = INT64_MAX;
4314 rd_cost->rdcost = INT64_MAX;
4318 assert((cm->interp_filter == SWITCHABLE) ||
4319 (cm->interp_filter == best_mbmode.interp_filter) ||
4320 !is_inter_block(&best_mbmode));
4322 vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
4323 sf->adaptive_rd_thresh, bsize, best_ref_index);
4327 x->skip |= best_skip2;
4328 if (!is_inter_block(&best_mbmode)) {
4329 for (i = 0; i < 4; i++)
4330 xd->mi[0]->bmi[i].as_mode = best_bmodes[i].as_mode;
4332 for (i = 0; i < 4; ++i)
4333 memcpy(&xd->mi[0]->bmi[i], &best_bmodes[i], sizeof(b_mode_info));
4335 mi->mv[0].as_int = xd->mi[0]->bmi[3].as_mv[0].as_int;
4336 mi->mv[1].as_int = xd->mi[0]->bmi[3].as_mv[1].as_int;
4339 for (i = 0; i < REFERENCE_MODES; ++i) {
4340 if (best_pred_rd[i] == INT64_MAX)
4341 best_pred_diff[i] = INT_MIN;
4343 best_pred_diff[i] = best_rd - best_pred_rd[i];
4347 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4348 if (best_filter_rd[i] == INT64_MAX)
4349 best_filter_diff[i] = 0;
4351 best_filter_diff[i] = best_rd - best_filter_rd[i];
4353 if (cm->interp_filter == SWITCHABLE)
4354 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
4356 vp9_zero(best_filter_diff);
4359 store_coding_context(x, ctx, best_ref_index,
4360 best_pred_diff, best_filter_diff, 0);