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"
16 #include "vpx_mem/vpx_mem.h"
18 #include "vp9/common/vp9_common.h"
19 #include "vp9/common/vp9_entropy.h"
20 #include "vp9/common/vp9_entropymode.h"
21 #include "vp9/common/vp9_idct.h"
22 #include "vp9/common/vp9_mvref_common.h"
23 #include "vp9/common/vp9_pred_common.h"
24 #include "vp9/common/vp9_quant_common.h"
25 #include "vp9/common/vp9_reconinter.h"
26 #include "vp9/common/vp9_reconintra.h"
27 #include "vp9/common/vp9_seg_common.h"
28 #include "vp9/common/vp9_systemdependent.h"
30 #include "vp9/encoder/vp9_cost.h"
31 #include "vp9/encoder/vp9_encodemb.h"
32 #include "vp9/encoder/vp9_encodemv.h"
33 #include "vp9/encoder/vp9_encoder.h"
34 #include "vp9/encoder/vp9_mcomp.h"
35 #include "vp9/encoder/vp9_quantize.h"
36 #include "vp9/encoder/vp9_ratectrl.h"
37 #include "vp9/encoder/vp9_rd.h"
38 #include "vp9/encoder/vp9_rdopt.h"
39 #include "vp9/encoder/vp9_variance.h"
41 #define RD_THRESH_MAX_FACT 64
42 #define RD_THRESH_INC 1
44 #define LAST_FRAME_MODE_MASK ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | \
46 #define GOLDEN_FRAME_MODE_MASK ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | \
48 #define ALT_REF_MODE_MASK ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | \
51 #define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
53 #define MIN_EARLY_TERM_INDEX 3
57 MV_REFERENCE_FRAME ref_frame[2];
61 MV_REFERENCE_FRAME ref_frame[2];
64 struct rdcost_block_args {
66 ENTROPY_CONTEXT t_above[16];
67 ENTROPY_CONTEXT t_left[16];
77 int use_fast_coef_costing;
81 static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
82 {NEARESTMV, {LAST_FRAME, NONE}},
83 {NEARESTMV, {ALTREF_FRAME, NONE}},
84 {NEARESTMV, {GOLDEN_FRAME, NONE}},
86 {DC_PRED, {INTRA_FRAME, NONE}},
88 {NEWMV, {LAST_FRAME, NONE}},
89 {NEWMV, {ALTREF_FRAME, NONE}},
90 {NEWMV, {GOLDEN_FRAME, NONE}},
92 {NEARMV, {LAST_FRAME, NONE}},
93 {NEARMV, {ALTREF_FRAME, NONE}},
94 {NEARMV, {GOLDEN_FRAME, NONE}},
96 {ZEROMV, {LAST_FRAME, NONE}},
97 {ZEROMV, {GOLDEN_FRAME, NONE}},
98 {ZEROMV, {ALTREF_FRAME, NONE}},
100 {NEARESTMV, {LAST_FRAME, ALTREF_FRAME}},
101 {NEARESTMV, {GOLDEN_FRAME, ALTREF_FRAME}},
103 {TM_PRED, {INTRA_FRAME, NONE}},
105 {NEARMV, {LAST_FRAME, ALTREF_FRAME}},
106 {NEWMV, {LAST_FRAME, ALTREF_FRAME}},
107 {NEARMV, {GOLDEN_FRAME, ALTREF_FRAME}},
108 {NEWMV, {GOLDEN_FRAME, ALTREF_FRAME}},
110 {ZEROMV, {LAST_FRAME, ALTREF_FRAME}},
111 {ZEROMV, {GOLDEN_FRAME, ALTREF_FRAME}},
113 {H_PRED, {INTRA_FRAME, NONE}},
114 {V_PRED, {INTRA_FRAME, NONE}},
115 {D135_PRED, {INTRA_FRAME, NONE}},
116 {D207_PRED, {INTRA_FRAME, NONE}},
117 {D153_PRED, {INTRA_FRAME, NONE}},
118 {D63_PRED, {INTRA_FRAME, NONE}},
119 {D117_PRED, {INTRA_FRAME, NONE}},
120 {D45_PRED, {INTRA_FRAME, NONE}},
123 static const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
124 {{LAST_FRAME, NONE}},
125 {{GOLDEN_FRAME, NONE}},
126 {{ALTREF_FRAME, NONE}},
127 {{LAST_FRAME, ALTREF_FRAME}},
128 {{GOLDEN_FRAME, ALTREF_FRAME}},
129 {{INTRA_FRAME, NONE}},
132 static int raster_block_offset(BLOCK_SIZE plane_bsize,
133 int raster_block, int stride) {
134 const int bw = b_width_log2(plane_bsize);
135 const int y = 4 * (raster_block >> bw);
136 const int x = 4 * (raster_block & ((1 << bw) - 1));
137 return y * stride + x;
139 static int16_t* raster_block_offset_int16(BLOCK_SIZE plane_bsize,
140 int raster_block, int16_t *base) {
141 const int stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
142 return base + raster_block_offset(plane_bsize, raster_block, stride);
145 static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
146 int m, int n, int min_plane, int max_plane) {
149 for (i = min_plane; i < max_plane; ++i) {
150 struct macroblock_plane *const p = &x->plane[i];
151 struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
153 p->coeff = ctx->coeff_pbuf[i][m];
154 p->qcoeff = ctx->qcoeff_pbuf[i][m];
155 pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
156 p->eobs = ctx->eobs_pbuf[i][m];
158 ctx->coeff_pbuf[i][m] = ctx->coeff_pbuf[i][n];
159 ctx->qcoeff_pbuf[i][m] = ctx->qcoeff_pbuf[i][n];
160 ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
161 ctx->eobs_pbuf[i][m] = ctx->eobs_pbuf[i][n];
163 ctx->coeff_pbuf[i][n] = p->coeff;
164 ctx->qcoeff_pbuf[i][n] = p->qcoeff;
165 ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
166 ctx->eobs_pbuf[i][n] = p->eobs;
170 static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize,
171 MACROBLOCK *x, MACROBLOCKD *xd,
172 int *out_rate_sum, int64_t *out_dist_sum) {
173 // Note our transform coeffs are 8 times an orthogonal transform.
174 // Hence quantizer step is also 8 times. To get effective quantizer
175 // we need to divide by 8 before sending to modeling function.
177 int64_t rate_sum = 0;
178 int64_t dist_sum = 0;
179 const int ref = xd->mi[0].src_mi->mbmi.ref_frame[0];
181 unsigned int var = 0;
182 unsigned int sum_sse = 0;
187 x->pred_sse[ref] = 0;
189 for (i = 0; i < MAX_MB_PLANE; ++i) {
190 struct macroblock_plane *const p = &x->plane[i];
191 struct macroblockd_plane *const pd = &xd->plane[i];
192 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
193 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
194 const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
195 int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
196 int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
198 int lw = b_width_log2_lookup[unit_size] + 2;
199 int lh = b_height_log2_lookup[unit_size] + 2;
203 for (idy = 0; idy < bh; ++idy) {
204 for (idx = 0; idx < bw; ++idx) {
205 uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
206 uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
207 int block_idx = (idy << 1) + idx;
209 var = cpi->fn_ptr[unit_size].vf(src, p->src.stride,
210 dst, pd->dst.stride, &sse);
211 x->bsse[(i << 2) + block_idx] = sse;
214 if (!x->select_tx_size) {
215 if (x->bsse[(i << 2) + block_idx] < p->quant_thred[0] >> shift)
216 x->skip_txfm[(i << 2) + block_idx] = 1;
217 else if (var < p->quant_thred[1] >> shift)
218 x->skip_txfm[(i << 2) + block_idx] = 2;
220 x->skip_txfm[(i << 2) + block_idx] = 0;
224 x->pred_sse[ref] += sse;
228 // Fast approximate the modelling function.
229 if (cpi->oxcf.speed > 4) {
231 int64_t square_error = sse;
232 int quantizer = (pd->dequant[1] >> 3);
233 #if CONFIG_VP9_HIGHBITDEPTH
234 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
235 quantizer >>= (xd->bd - 8);
237 #endif // CONFIG_VP9_HIGHBITDEPTH
240 rate = (square_error * (280 - quantizer)) >> 8;
243 dist = (square_error * quantizer) >> 8;
247 #if CONFIG_VP9_HIGHBITDEPTH
248 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
249 vp9_model_rd_from_var_lapndz(sse, 1 << num_pels_log2_lookup[bs],
250 pd->dequant[1] >> (xd->bd - 5),
253 vp9_model_rd_from_var_lapndz(sse, 1 << num_pels_log2_lookup[bs],
254 pd->dequant[1] >> 3, &rate, &dist);
257 vp9_model_rd_from_var_lapndz(sum_sse, 1 << num_pels_log2_lookup[bs],
258 pd->dequant[1] >> 3, &rate, &dist);
259 #endif // CONFIG_VP9_HIGHBITDEPTH
265 *out_rate_sum = (int)rate_sum;
266 *out_dist_sum = dist_sum << 4;
269 int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
270 intptr_t block_size, int64_t *ssz) {
272 int64_t error = 0, sqcoeff = 0;
274 for (i = 0; i < block_size; i++) {
275 const int diff = coeff[i] - dqcoeff[i];
276 error += diff * diff;
277 sqcoeff += coeff[i] * coeff[i];
285 #if CONFIG_VP9_HIGHBITDEPTH
286 int64_t vp9_high_block_error_c(const tran_low_t *coeff,
287 const tran_low_t *dqcoeff,
289 int64_t *ssz, int bd) {
291 int64_t error = 0, sqcoeff = 0;
292 int shift = 2 * (bd - 8);
293 int rounding = shift > 0 ? 1 << (shift - 1) : 0;
295 for (i = 0; i < block_size; i++) {
296 const int64_t diff = coeff[i] - dqcoeff[i];
297 error += diff * diff;
298 sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
300 assert(error >= 0 && sqcoeff >= 0);
301 error = (error + rounding) >> shift;
302 sqcoeff = (sqcoeff + rounding) >> shift;
307 #endif // CONFIG_VP9_HIGHBITDEPTH
309 /* The trailing '0' is a terminator which is used inside cost_coeffs() to
310 * decide whether to include cost of a trailing EOB node or not (i.e. we
311 * can skip this if the last coefficient in this transform block, e.g. the
312 * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
314 static const int16_t band_counts[TX_SIZES][8] = {
315 { 1, 2, 3, 4, 3, 16 - 13, 0 },
316 { 1, 2, 3, 4, 11, 64 - 21, 0 },
317 { 1, 2, 3, 4, 11, 256 - 21, 0 },
318 { 1, 2, 3, 4, 11, 1024 - 21, 0 },
320 static INLINE int cost_coeffs(MACROBLOCK *x,
321 int plane, int block,
322 ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L,
324 const int16_t *scan, const int16_t *nb,
325 int use_fast_coef_costing) {
326 MACROBLOCKD *const xd = &x->e_mbd;
327 MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
328 const struct macroblock_plane *p = &x->plane[plane];
329 const struct macroblockd_plane *pd = &xd->plane[plane];
330 const PLANE_TYPE type = pd->plane_type;
331 const int16_t *band_count = &band_counts[tx_size][1];
332 const int eob = p->eobs[block];
333 const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
334 unsigned int (*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
335 x->token_costs[tx_size][type][is_inter_block(mbmi)];
336 uint8_t token_cache[32 * 32];
337 int pt = combine_entropy_contexts(*A, *L);
339 // Check for consistency of tx_size with mode info
340 assert(type == PLANE_TYPE_Y ? mbmi->tx_size == tx_size
341 : get_uv_tx_size(mbmi, pd) == tx_size);
345 cost = token_costs[0][0][pt][EOB_TOKEN];
348 int band_left = *band_count++;
352 int prev_t = vp9_dct_value_tokens_ptr[v].token;
353 cost = (*token_costs)[0][pt][prev_t] + vp9_dct_value_cost_ptr[v];
354 token_cache[0] = vp9_pt_energy_class[prev_t];
358 for (c = 1; c < eob; c++) {
359 const int rc = scan[c];
363 t = vp9_dct_value_tokens_ptr[v].token;
364 if (use_fast_coef_costing) {
365 cost += (*token_costs)[!prev_t][!prev_t][t] + vp9_dct_value_cost_ptr[v];
367 pt = get_coef_context(nb, token_cache, c);
368 cost += (*token_costs)[!prev_t][pt][t] + vp9_dct_value_cost_ptr[v];
369 token_cache[rc] = vp9_pt_energy_class[t];
373 band_left = *band_count++;
380 if (use_fast_coef_costing) {
381 cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
383 pt = get_coef_context(nb, token_cache, c);
384 cost += (*token_costs)[0][pt][EOB_TOKEN];
389 // is eob first coefficient;
395 #if CONFIG_VP9_HIGHBITDEPTH
396 static void dist_block(int plane, int block, TX_SIZE tx_size,
397 struct rdcost_block_args* args, int bd) {
399 static void dist_block(int plane, int block, TX_SIZE tx_size,
400 struct rdcost_block_args* args) {
401 #endif // CONFIG_VP9_HIGHBITDEPTH
402 const int ss_txfrm_size = tx_size << 1;
403 MACROBLOCK* const x = args->x;
404 MACROBLOCKD* const xd = &x->e_mbd;
405 const struct macroblock_plane *const p = &x->plane[plane];
406 const struct macroblockd_plane *const pd = &xd->plane[plane];
408 int shift = tx_size == TX_32X32 ? 0 : 2;
409 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
410 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
411 #if CONFIG_VP9_HIGHBITDEPTH
412 args->dist = vp9_high_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
413 &this_sse, bd) >> shift;
415 args->dist = vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
417 #endif // CONFIG_VP9_HIGHBITDEPTH
418 args->sse = this_sse >> shift;
420 if (x->skip_encode && !is_inter_block(&xd->mi[0].src_mi->mbmi)) {
421 // TODO(jingning): tune the model to better capture the distortion.
422 int64_t p = (pd->dequant[1] * pd->dequant[1] *
423 (1 << ss_txfrm_size)) >> (shift + 2);
424 #if CONFIG_VP9_HIGHBITDEPTH
425 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
426 p >>= ((xd->bd - 8) * 2);
428 #endif // CONFIG_VP9_HIGHBITDEPTH
429 args->dist += (p >> 4);
434 static void rate_block(int plane, int block, BLOCK_SIZE plane_bsize,
435 TX_SIZE tx_size, struct rdcost_block_args* args) {
437 txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x_idx, &y_idx);
439 args->rate = cost_coeffs(args->x, plane, block, args->t_above + x_idx,
440 args->t_left + y_idx, tx_size,
441 args->so->scan, args->so->neighbors,
442 args->use_fast_coef_costing);
445 static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
446 TX_SIZE tx_size, void *arg) {
447 struct rdcost_block_args *args = arg;
448 MACROBLOCK *const x = args->x;
449 MACROBLOCKD *const xd = &x->e_mbd;
450 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
451 int64_t rd1, rd2, rd;
456 if (!is_inter_block(mbmi)) {
457 vp9_encode_block_intra(x, plane, block, plane_bsize, tx_size, &mbmi->skip);
458 #if CONFIG_VP9_HIGHBITDEPTH
459 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
460 dist_block(plane, block, tx_size, args, xd->bd);
462 dist_block(plane, block, tx_size, args, 8);
465 dist_block(plane, block, tx_size, args);
466 #endif // CONFIG_VP9_HIGHBITDEPTH
467 } else if (max_txsize_lookup[plane_bsize] == tx_size) {
468 if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] == 0) {
469 // full forward transform and quantization
470 vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
471 #if CONFIG_VP9_HIGHBITDEPTH
472 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
473 dist_block(plane, block, tx_size, args, xd->bd);
475 dist_block(plane, block, tx_size, args, 8);
478 dist_block(plane, block, tx_size, args);
479 #endif // CONFIG_VP9_HIGHBITDEPTH
480 } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] == 2) {
481 // compute DC coefficient
482 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[plane].coeff, block);
483 tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
484 vp9_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
485 args->sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
486 args->dist = args->sse;
487 if (!x->plane[plane].eobs[block])
488 args->dist = args->sse - ((coeff[0] * coeff[0] -
489 (coeff[0] - dqcoeff[0]) * (coeff[0] - dqcoeff[0])) >> 2);
491 // skip forward transform
492 x->plane[plane].eobs[block] = 0;
493 args->sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
494 args->dist = args->sse;
497 // full forward transform and quantization
498 vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
499 #if CONFIG_VP9_HIGHBITDEPTH
500 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
501 dist_block(plane, block, tx_size, args, xd->bd);
503 dist_block(plane, block, tx_size, args, 8);
506 dist_block(plane, block, tx_size, args);
507 #endif // CONFIG_VP9_HIGHBITDEPTH
510 rate_block(plane, block, plane_bsize, tx_size, args);
511 rd1 = RDCOST(x->rdmult, x->rddiv, args->rate, args->dist);
512 rd2 = RDCOST(x->rdmult, x->rddiv, 0, args->sse);
514 // TODO(jingning): temporarily enabled only for luma component
517 x->zcoeff_blk[tx_size][block] = !x->plane[plane].eobs[block] ||
518 (rd1 > rd2 && !xd->lossless);
520 args->this_rate += args->rate;
521 args->this_dist += args->dist;
522 args->this_sse += args->sse;
525 if (args->this_rd > args->best_rd) {
531 static void txfm_rd_in_plane(MACROBLOCK *x,
532 int *rate, int64_t *distortion,
533 int *skippable, int64_t *sse,
534 int64_t ref_best_rd, int plane,
535 BLOCK_SIZE bsize, TX_SIZE tx_size,
536 int use_fast_coef_casting) {
537 MACROBLOCKD *const xd = &x->e_mbd;
538 const struct macroblockd_plane *const pd = &xd->plane[plane];
539 struct rdcost_block_args args;
542 args.best_rd = ref_best_rd;
543 args.use_fast_coef_costing = use_fast_coef_casting;
546 xd->mi[0].src_mi->mbmi.tx_size = tx_size;
548 vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
550 args.so = get_scan(xd, tx_size, pd->plane_type, 0);
552 vp9_foreach_transformed_block_in_plane(xd, bsize, plane,
553 block_rd_txfm, &args);
556 *distortion = INT64_MAX;
560 *distortion = args.this_dist;
561 *rate = args.this_rate;
562 *sse = args.this_sse;
563 *skippable = vp9_is_skippable_in_plane(x, bsize, plane);
567 static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x,
568 int *rate, int64_t *distortion,
569 int *skip, int64_t *sse,
572 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
573 VP9_COMMON *const cm = &cpi->common;
574 const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
575 MACROBLOCKD *const xd = &x->e_mbd;
576 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
578 mbmi->tx_size = MIN(max_tx_size, largest_tx_size);
580 txfm_rd_in_plane(x, rate, distortion, skip,
581 sse, ref_best_rd, 0, bs,
582 mbmi->tx_size, cpi->sf.use_fast_coef_costing);
585 static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
590 int64_t tx_cache[TX_MODES],
593 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
594 VP9_COMMON *const cm = &cpi->common;
595 MACROBLOCKD *const xd = &x->e_mbd;
596 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
597 vp9_prob skip_prob = vp9_get_skip_prob(cm, xd);
598 int r[TX_SIZES][2], s[TX_SIZES];
599 int64_t d[TX_SIZES], sse[TX_SIZES];
600 int64_t rd[TX_SIZES][2] = {{INT64_MAX, INT64_MAX},
601 {INT64_MAX, INT64_MAX},
602 {INT64_MAX, INT64_MAX},
603 {INT64_MAX, INT64_MAX}};
606 const TX_SIZE max_mode_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
607 int64_t best_rd = INT64_MAX;
608 TX_SIZE best_tx = max_tx_size;
610 const vp9_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc.tx_probs);
611 assert(skip_prob > 0);
612 s0 = vp9_cost_bit(skip_prob, 0);
613 s1 = vp9_cost_bit(skip_prob, 1);
615 for (n = max_tx_size; n >= 0; n--) {
616 txfm_rd_in_plane(x, &r[n][0], &d[n], &s[n],
617 &sse[n], ref_best_rd, 0, bs, n,
618 cpi->sf.use_fast_coef_costing);
620 if (r[n][0] < INT_MAX) {
621 for (m = 0; m <= n - (n == (int) max_tx_size); m++) {
623 r[n][1] += vp9_cost_zero(tx_probs[m]);
625 r[n][1] += vp9_cost_one(tx_probs[m]);
628 if (d[n] == INT64_MAX) {
629 rd[n][0] = rd[n][1] = INT64_MAX;
631 rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, d[n]);
633 rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
634 rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
637 // Early termination in transform size search.
638 if (cpi->sf.tx_size_search_breakout &&
639 (rd[n][1] == INT64_MAX ||
640 (n < (int) max_tx_size && rd[n][1] > rd[n + 1][1]) ||
644 if (rd[n][1] < best_rd) {
649 mbmi->tx_size = cm->tx_mode == TX_MODE_SELECT ?
650 best_tx : MIN(max_tx_size, max_mode_tx_size);
653 *distortion = d[mbmi->tx_size];
654 *rate = r[mbmi->tx_size][cm->tx_mode == TX_MODE_SELECT];
655 *skip = s[mbmi->tx_size];
656 *psse = sse[mbmi->tx_size];
658 tx_cache[ONLY_4X4] = rd[TX_4X4][0];
659 tx_cache[ALLOW_8X8] = rd[TX_8X8][0];
660 tx_cache[ALLOW_16X16] = rd[MIN(max_tx_size, TX_16X16)][0];
661 tx_cache[ALLOW_32X32] = rd[MIN(max_tx_size, TX_32X32)][0];
663 if (max_tx_size == TX_32X32 && best_tx == TX_32X32) {
664 tx_cache[TX_MODE_SELECT] = rd[TX_32X32][1];
665 } else if (max_tx_size >= TX_16X16 && best_tx == TX_16X16) {
666 tx_cache[TX_MODE_SELECT] = rd[TX_16X16][1];
667 } else if (rd[TX_8X8][1] < rd[TX_4X4][1]) {
668 tx_cache[TX_MODE_SELECT] = rd[TX_8X8][1];
670 tx_cache[TX_MODE_SELECT] = rd[TX_4X4][1];
674 static void super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
675 int64_t *distortion, int *skip,
676 int64_t *psse, BLOCK_SIZE bs,
677 int64_t txfm_cache[TX_MODES],
678 int64_t ref_best_rd) {
679 MACROBLOCKD *xd = &x->e_mbd;
681 int64_t *ret_sse = psse ? psse : &sse;
683 assert(bs == xd->mi[0].src_mi->mbmi.sb_type);
685 if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
686 vpx_memset(txfm_cache, 0, TX_MODES * sizeof(int64_t));
687 choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
690 choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse,
691 txfm_cache, ref_best_rd, bs);
695 static int conditional_skipintra(PREDICTION_MODE mode,
696 PREDICTION_MODE best_intra_mode) {
697 if (mode == D117_PRED &&
698 best_intra_mode != V_PRED &&
699 best_intra_mode != D135_PRED)
701 if (mode == D63_PRED &&
702 best_intra_mode != V_PRED &&
703 best_intra_mode != D45_PRED)
705 if (mode == D207_PRED &&
706 best_intra_mode != H_PRED &&
707 best_intra_mode != D45_PRED)
709 if (mode == D153_PRED &&
710 best_intra_mode != H_PRED &&
711 best_intra_mode != D135_PRED)
716 static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int ib,
717 PREDICTION_MODE *best_mode,
718 const int *bmode_costs,
719 ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
720 int *bestrate, int *bestratey,
721 int64_t *bestdistortion,
722 BLOCK_SIZE bsize, int64_t rd_thresh) {
723 PREDICTION_MODE mode;
724 MACROBLOCKD *const xd = &x->e_mbd;
725 int64_t best_rd = rd_thresh;
727 struct macroblock_plane *p = &x->plane[0];
728 struct macroblockd_plane *pd = &xd->plane[0];
729 const int src_stride = p->src.stride;
730 const int dst_stride = pd->dst.stride;
731 const uint8_t *src_init = &p->src.buf[raster_block_offset(BLOCK_8X8, ib,
733 uint8_t *dst_init = &pd->dst.buf[raster_block_offset(BLOCK_8X8, ib,
735 ENTROPY_CONTEXT ta[2], tempa[2];
736 ENTROPY_CONTEXT tl[2], templ[2];
738 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
739 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
741 uint8_t best_dst[8 * 8];
742 #if CONFIG_VP9_HIGHBITDEPTH
743 uint16_t best_dst16[8 * 8];
748 vpx_memcpy(ta, a, sizeof(ta));
749 vpx_memcpy(tl, l, sizeof(tl));
750 xd->mi[0].src_mi->mbmi.tx_size = TX_4X4;
752 #if CONFIG_VP9_HIGHBITDEPTH
753 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
754 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
757 int64_t distortion = 0;
758 int rate = bmode_costs[mode];
760 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
763 // Only do the oblique modes if the best so far is
764 // one of the neighboring directional modes
765 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
766 if (conditional_skipintra(mode, *best_mode))
770 vpx_memcpy(tempa, ta, sizeof(ta));
771 vpx_memcpy(templ, tl, sizeof(tl));
773 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
774 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
775 const int block = ib + idy * 2 + idx;
776 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
777 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
778 int16_t *const src_diff = raster_block_offset_int16(BLOCK_8X8, block,
780 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
781 xd->mi[0].src_mi->bmi[block].as_mode = mode;
782 vp9_predict_intra_block(xd, block, 1,
784 x->skip_encode ? src : dst,
785 x->skip_encode ? src_stride : dst_stride,
786 dst, dst_stride, idx, idy, 0);
787 vp9_high_subtract_block(4, 4, src_diff, 8, src, src_stride,
788 dst, dst_stride, xd->bd);
790 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
791 vp9_high_fwht4x4(src_diff, coeff, 8);
792 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
793 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
794 so->scan, so->neighbors,
795 cpi->sf.use_fast_coef_costing);
796 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
798 vp9_high_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block),
800 p->eobs[block], xd->bd);
803 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
804 const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
805 vp9_high_fht4x4(src_diff, coeff, 8, tx_type);
806 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
807 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
808 so->scan, so->neighbors,
809 cpi->sf.use_fast_coef_costing);
810 distortion += vp9_high_block_error(coeff,
811 BLOCK_OFFSET(pd->dqcoeff, block),
812 16, &unused, xd->bd) >> 2;
813 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
815 vp9_high_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
816 dst, dst_stride, p->eobs[block], xd->bd);
822 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
824 if (this_rd < best_rd) {
827 *bestdistortion = distortion;
830 vpx_memcpy(a, tempa, sizeof(tempa));
831 vpx_memcpy(l, templ, sizeof(templ));
832 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
833 vpx_memcpy(best_dst16 + idy * 8,
834 CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
835 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
841 if (best_rd >= rd_thresh || x->skip_encode)
844 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
845 vpx_memcpy(CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
846 best_dst16 + idy * 8,
847 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
852 #endif // CONFIG_VP9_HIGHBITDEPTH
854 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
857 int64_t distortion = 0;
858 int rate = bmode_costs[mode];
860 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
863 // Only do the oblique modes if the best so far is
864 // one of the neighboring directional modes
865 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
866 if (conditional_skipintra(mode, *best_mode))
870 vpx_memcpy(tempa, ta, sizeof(ta));
871 vpx_memcpy(templ, tl, sizeof(tl));
873 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
874 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
875 const int block = ib + idy * 2 + idx;
876 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
877 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
878 int16_t *const src_diff = raster_block_offset_int16(BLOCK_8X8, block,
880 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
881 xd->mi[0].src_mi->bmi[block].as_mode = mode;
882 vp9_predict_intra_block(xd, block, 1,
884 x->skip_encode ? src : dst,
885 x->skip_encode ? src_stride : dst_stride,
886 dst, dst_stride, idx, idy, 0);
887 vp9_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
890 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
891 vp9_fwht4x4(src_diff, coeff, 8);
892 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
893 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
894 so->scan, so->neighbors,
895 cpi->sf.use_fast_coef_costing);
896 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
898 vp9_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride,
902 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
903 const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
904 vp9_fht4x4(src_diff, coeff, 8, tx_type);
905 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
906 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
907 so->scan, so->neighbors,
908 cpi->sf.use_fast_coef_costing);
909 distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
911 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
913 vp9_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
914 dst, dst_stride, p->eobs[block]);
920 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
922 if (this_rd < best_rd) {
925 *bestdistortion = distortion;
928 vpx_memcpy(a, tempa, sizeof(tempa));
929 vpx_memcpy(l, templ, sizeof(templ));
930 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
931 vpx_memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
932 num_4x4_blocks_wide * 4);
938 if (best_rd >= rd_thresh || x->skip_encode)
941 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
942 vpx_memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
943 num_4x4_blocks_wide * 4);
948 static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
949 int *rate, int *rate_y,
953 const MACROBLOCKD *const xd = &mb->e_mbd;
954 MODE_INFO *const mic = xd->mi[0].src_mi;
955 const MODE_INFO *above_mi = xd->mi[-xd->mi_stride].src_mi;
956 const MODE_INFO *left_mi = xd->left_available ? xd->mi[-1].src_mi : NULL;
957 const BLOCK_SIZE bsize = xd->mi[0].src_mi->mbmi.sb_type;
958 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
959 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
962 int64_t total_distortion = 0;
964 int64_t total_rd = 0;
965 ENTROPY_CONTEXT t_above[4], t_left[4];
966 const int *bmode_costs = cpi->mbmode_cost;
968 vpx_memcpy(t_above, xd->plane[0].above_context, sizeof(t_above));
969 vpx_memcpy(t_left, xd->plane[0].left_context, sizeof(t_left));
971 // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
972 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
973 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
974 PREDICTION_MODE best_mode = DC_PRED;
975 int r = INT_MAX, ry = INT_MAX;
976 int64_t d = INT64_MAX, this_rd = INT64_MAX;
978 if (cpi->common.frame_type == KEY_FRAME) {
979 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i);
980 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i);
982 bmode_costs = cpi->y_mode_costs[A][L];
985 this_rd = rd_pick_intra4x4block(cpi, mb, i, &best_mode, bmode_costs,
986 t_above + idx, t_left + idy, &r, &ry, &d,
987 bsize, best_rd - total_rd);
988 if (this_rd >= best_rd - total_rd)
993 total_distortion += d;
996 mic->bmi[i].as_mode = best_mode;
997 for (j = 1; j < num_4x4_blocks_high; ++j)
998 mic->bmi[i + j * 2].as_mode = best_mode;
999 for (j = 1; j < num_4x4_blocks_wide; ++j)
1000 mic->bmi[i + j].as_mode = best_mode;
1002 if (total_rd >= best_rd)
1008 *rate_y = tot_rate_y;
1009 *distortion = total_distortion;
1010 mic->mbmi.mode = mic->bmi[3].as_mode;
1012 return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
1015 // This function is used only for intra_only frames
1016 static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x,
1017 int *rate, int *rate_tokenonly,
1018 int64_t *distortion, int *skippable,
1020 int64_t tx_cache[TX_MODES],
1022 PREDICTION_MODE mode;
1023 PREDICTION_MODE mode_selected = DC_PRED;
1024 MACROBLOCKD *const xd = &x->e_mbd;
1025 MODE_INFO *const mic = xd->mi[0].src_mi;
1026 int this_rate, this_rate_tokenonly, s;
1027 int64_t this_distortion, this_rd;
1028 TX_SIZE best_tx = TX_4X4;
1031 const MODE_INFO *above_mi = xd->mi[-xd->mi_stride].src_mi;
1032 const MODE_INFO *left_mi = xd->left_available ? xd->mi[-1].src_mi : NULL;
1033 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1034 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1035 bmode_costs = cpi->y_mode_costs[A][L];
1037 if (cpi->sf.tx_size_search_method == USE_FULL_RD)
1038 for (i = 0; i < TX_MODES; i++)
1039 tx_cache[i] = INT64_MAX;
1041 /* Y Search for intra prediction mode */
1042 for (mode = DC_PRED; mode <= TM_PRED; mode++) {
1043 int64_t local_tx_cache[TX_MODES];
1044 mic->mbmi.mode = mode;
1046 super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion,
1047 &s, NULL, bsize, local_tx_cache, best_rd);
1049 if (this_rate_tokenonly == INT_MAX)
1052 this_rate = this_rate_tokenonly + bmode_costs[mode];
1053 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1055 if (this_rd < best_rd) {
1056 mode_selected = mode;
1058 best_tx = mic->mbmi.tx_size;
1060 *rate_tokenonly = this_rate_tokenonly;
1061 *distortion = this_distortion;
1065 if (cpi->sf.tx_size_search_method == USE_FULL_RD && this_rd < INT64_MAX) {
1066 for (i = 0; i < TX_MODES && local_tx_cache[i] < INT64_MAX; i++) {
1067 const int64_t adj_rd = this_rd + local_tx_cache[i] -
1068 local_tx_cache[cpi->common.tx_mode];
1069 if (adj_rd < tx_cache[i]) {
1070 tx_cache[i] = adj_rd;
1076 mic->mbmi.mode = mode_selected;
1077 mic->mbmi.tx_size = best_tx;
1082 static void super_block_uvrd(const VP9_COMP *cpi, MACROBLOCK *x,
1083 int *rate, int64_t *distortion, int *skippable,
1084 int64_t *sse, BLOCK_SIZE bsize,
1085 int64_t ref_best_rd) {
1086 MACROBLOCKD *const xd = &x->e_mbd;
1087 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
1088 const TX_SIZE uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
1090 int pnrate = 0, pnskip = 1;
1091 int64_t pndist = 0, pnsse = 0;
1093 if (ref_best_rd < 0)
1096 if (is_inter_block(mbmi)) {
1098 for (plane = 1; plane < MAX_MB_PLANE; ++plane)
1099 vp9_subtract_plane(x, bsize, plane);
1107 for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
1108 txfm_rd_in_plane(x, &pnrate, &pndist, &pnskip, &pnsse,
1109 ref_best_rd, plane, bsize, uv_tx_size,
1110 cpi->sf.use_fast_coef_costing);
1111 if (pnrate == INT_MAX)
1114 *distortion += pndist;
1116 *skippable &= pnskip;
1122 *distortion = INT64_MAX;
1128 static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
1129 PICK_MODE_CONTEXT *ctx,
1130 int *rate, int *rate_tokenonly,
1131 int64_t *distortion, int *skippable,
1132 BLOCK_SIZE bsize, TX_SIZE max_tx_size) {
1133 MACROBLOCKD *xd = &x->e_mbd;
1134 PREDICTION_MODE mode;
1135 PREDICTION_MODE mode_selected = DC_PRED;
1136 int64_t best_rd = INT64_MAX, this_rd;
1137 int this_rate_tokenonly, this_rate, s;
1138 int64_t this_distortion, this_sse;
1140 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1141 if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode)))
1144 xd->mi[0].src_mi->mbmi.uv_mode = mode;
1146 super_block_uvrd(cpi, x, &this_rate_tokenonly,
1147 &this_distortion, &s, &this_sse, bsize, best_rd);
1148 if (this_rate_tokenonly == INT_MAX)
1150 this_rate = this_rate_tokenonly +
1151 cpi->intra_uv_mode_cost[cpi->common.frame_type][mode];
1152 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1154 if (this_rd < best_rd) {
1155 mode_selected = mode;
1158 *rate_tokenonly = this_rate_tokenonly;
1159 *distortion = this_distortion;
1161 if (!x->select_tx_size)
1162 swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
1166 xd->mi[0].src_mi->mbmi.uv_mode = mode_selected;
1170 static int64_t rd_sbuv_dcpred(const VP9_COMP *cpi, MACROBLOCK *x,
1171 int *rate, int *rate_tokenonly,
1172 int64_t *distortion, int *skippable,
1174 const VP9_COMMON *cm = &cpi->common;
1177 x->e_mbd.mi[0].src_mi->mbmi.uv_mode = DC_PRED;
1178 super_block_uvrd(cpi, x, rate_tokenonly, distortion,
1179 skippable, &unused, bsize, INT64_MAX);
1180 *rate = *rate_tokenonly + cpi->intra_uv_mode_cost[cm->frame_type][DC_PRED];
1181 return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
1184 static void choose_intra_uv_mode(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
1185 BLOCK_SIZE bsize, TX_SIZE max_tx_size,
1186 int *rate_uv, int *rate_uv_tokenonly,
1187 int64_t *dist_uv, int *skip_uv,
1188 PREDICTION_MODE *mode_uv) {
1189 MACROBLOCK *const x = &cpi->mb;
1191 // Use an estimated rd for uv_intra based on DC_PRED if the
1192 // appropriate speed flag is set.
1193 if (cpi->sf.use_uv_intra_rd_estimate) {
1194 rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv,
1195 skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
1196 // Else do a proper rd search for each possible transform size that may
1197 // be considered in the main rd loop.
1199 rd_pick_intra_sbuv_mode(cpi, x, ctx,
1200 rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
1201 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize, max_tx_size);
1203 *mode_uv = x->e_mbd.mi[0].src_mi->mbmi.uv_mode;
1206 static int cost_mv_ref(const VP9_COMP *cpi, PREDICTION_MODE mode,
1208 assert(is_inter_mode(mode));
1209 return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
1212 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
1215 int mi_row, int mi_col,
1216 int_mv single_newmv[MAX_REF_FRAMES],
1219 static int set_and_cost_bmi_mvs(VP9_COMP *cpi, MACROBLOCKD *xd, int i,
1220 PREDICTION_MODE mode, int_mv this_mv[2],
1221 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1222 int_mv seg_mvs[MAX_REF_FRAMES],
1223 int_mv *best_ref_mv[2], const int *mvjcost,
1225 MODE_INFO *const mic = xd->mi[0].src_mi;
1226 const MB_MODE_INFO *const mbmi = &mic->mbmi;
1229 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type];
1230 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
1231 const int is_compound = has_second_ref(mbmi);
1235 this_mv[0].as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
1236 thismvcost += vp9_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
1237 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1239 this_mv[1].as_int = seg_mvs[mbmi->ref_frame[1]].as_int;
1240 thismvcost += vp9_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
1241 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1246 this_mv[0].as_int = frame_mv[mode][mbmi->ref_frame[0]].as_int;
1248 this_mv[1].as_int = frame_mv[mode][mbmi->ref_frame[1]].as_int;
1251 this_mv[0].as_int = 0;
1253 this_mv[1].as_int = 0;
1259 mic->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
1261 mic->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
1263 mic->bmi[i].as_mode = mode;
1265 for (idy = 0; idy < num_4x4_blocks_high; ++idy)
1266 for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
1267 vpx_memcpy(&mic->bmi[i + idy * 2 + idx],
1268 &mic->bmi[i], sizeof(mic->bmi[i]));
1270 return cost_mv_ref(cpi, mode, mbmi->mode_context[mbmi->ref_frame[0]]) +
1274 static int64_t encode_inter_mb_segment(VP9_COMP *cpi,
1279 int64_t *distortion, int64_t *sse,
1280 ENTROPY_CONTEXT *ta,
1281 ENTROPY_CONTEXT *tl,
1282 int mi_row, int mi_col) {
1284 MACROBLOCKD *xd = &x->e_mbd;
1285 struct macroblockd_plane *const pd = &xd->plane[0];
1286 struct macroblock_plane *const p = &x->plane[0];
1287 MODE_INFO *const mi = xd->mi[0].src_mi;
1288 const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->mbmi.sb_type, pd);
1289 const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
1290 const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
1293 const uint8_t *const src = &p->src.buf[raster_block_offset(BLOCK_8X8, i,
1295 uint8_t *const dst = &pd->dst.buf[raster_block_offset(BLOCK_8X8, i,
1297 int64_t thisdistortion = 0, thissse = 0;
1298 int thisrate = 0, ref;
1299 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1300 const int is_compound = has_second_ref(&mi->mbmi);
1301 const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter);
1303 for (ref = 0; ref < 1 + is_compound; ++ref) {
1304 const uint8_t *pre = &pd->pre[ref].buf[raster_block_offset(BLOCK_8X8, i,
1305 pd->pre[ref].stride)];
1306 #if CONFIG_VP9_HIGHBITDEPTH
1307 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1308 vp9_high_build_inter_predictor(pre, pd->pre[ref].stride,
1309 dst, pd->dst.stride,
1310 &mi->bmi[i].as_mv[ref].as_mv,
1311 &xd->block_refs[ref]->sf, width, height, ref,
1312 kernel, MV_PRECISION_Q3,
1313 mi_col * MI_SIZE + 4 * (i % 2),
1314 mi_row * MI_SIZE + 4 * (i / 2), xd->bd);
1316 vp9_build_inter_predictor(pre, pd->pre[ref].stride,
1317 dst, pd->dst.stride,
1318 &mi->bmi[i].as_mv[ref].as_mv,
1319 &xd->block_refs[ref]->sf, width, height, ref,
1320 kernel, MV_PRECISION_Q3,
1321 mi_col * MI_SIZE + 4 * (i % 2),
1322 mi_row * MI_SIZE + 4 * (i / 2));
1325 vp9_build_inter_predictor(pre, pd->pre[ref].stride,
1326 dst, pd->dst.stride,
1327 &mi->bmi[i].as_mv[ref].as_mv,
1328 &xd->block_refs[ref]->sf, width, height, ref,
1329 kernel, MV_PRECISION_Q3,
1330 mi_col * MI_SIZE + 4 * (i % 2),
1331 mi_row * MI_SIZE + 4 * (i / 2));
1332 #endif // CONFIG_VP9_HIGHBITDEPTH
1335 #if CONFIG_VP9_HIGHBITDEPTH
1336 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1337 vp9_high_subtract_block(
1338 height, width, raster_block_offset_int16(BLOCK_8X8, i, p->src_diff), 8,
1339 src, p->src.stride, dst, pd->dst.stride, xd->bd);
1342 height, width, raster_block_offset_int16(BLOCK_8X8, i, p->src_diff), 8,
1343 src, p->src.stride, dst, pd->dst.stride);
1346 vp9_subtract_block(height, width,
1347 raster_block_offset_int16(BLOCK_8X8, i, p->src_diff), 8,
1348 src, p->src.stride, dst, pd->dst.stride);
1349 #endif // CONFIG_VP9_HIGHBITDEPTH
1352 for (idy = 0; idy < height / 4; ++idy) {
1353 for (idx = 0; idx < width / 4; ++idx) {
1354 int64_t ssz, rd, rd1, rd2;
1357 k += (idy * 2 + idx);
1358 coeff = BLOCK_OFFSET(p->coeff, k);
1359 x->fwd_txm4x4(raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
1361 vp9_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan);
1362 #if CONFIG_VP9_HIGHBITDEPTH
1363 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1364 thisdistortion += vp9_high_block_error(coeff,
1365 BLOCK_OFFSET(pd->dqcoeff, k),
1368 thisdistortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
1372 thisdistortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
1374 #endif // CONFIG_VP9_HIGHBITDEPTH
1376 thisrate += cost_coeffs(x, 0, k, ta + (k & 1), tl + (k >> 1), TX_4X4,
1377 so->scan, so->neighbors,
1378 cpi->sf.use_fast_coef_costing);
1379 rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
1380 rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
1387 *distortion = thisdistortion >> 2;
1388 *labelyrate = thisrate;
1389 *sse = thissse >> 2;
1391 return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
1402 ENTROPY_CONTEXT ta[2];
1403 ENTROPY_CONTEXT tl[2];
1415 PREDICTION_MODE modes[4];
1416 SEG_RDSTAT rdstat[4][INTER_MODES];
1420 static INLINE int mv_check_bounds(const MACROBLOCK *x, const MV *mv) {
1421 return (mv->row >> 3) < x->mv_row_min ||
1422 (mv->row >> 3) > x->mv_row_max ||
1423 (mv->col >> 3) < x->mv_col_min ||
1424 (mv->col >> 3) > x->mv_col_max;
1427 static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
1428 MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0].src_mi->mbmi;
1429 struct macroblock_plane *const p = &x->plane[0];
1430 struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
1432 p->src.buf = &p->src.buf[raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1433 assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
1434 pd->pre[0].buf = &pd->pre[0].buf[raster_block_offset(BLOCK_8X8, i,
1435 pd->pre[0].stride)];
1436 if (has_second_ref(mbmi))
1437 pd->pre[1].buf = &pd->pre[1].buf[raster_block_offset(BLOCK_8X8, i,
1438 pd->pre[1].stride)];
1441 static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
1442 struct buf_2d orig_pre[2]) {
1443 MB_MODE_INFO *mbmi = &x->e_mbd.mi[0].src_mi->mbmi;
1444 x->plane[0].src = orig_src;
1445 x->e_mbd.plane[0].pre[0] = orig_pre[0];
1446 if (has_second_ref(mbmi))
1447 x->e_mbd.plane[0].pre[1] = orig_pre[1];
1450 static INLINE int mv_has_subpel(const MV *mv) {
1451 return (mv->row & 0x0F) || (mv->col & 0x0F);
1454 // Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
1455 // TODO(aconverse): Find out if this is still productive then clean up or remove
1456 static int check_best_zero_mv(
1457 const VP9_COMP *cpi, const uint8_t mode_context[MAX_REF_FRAMES],
1458 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int this_mode,
1459 const MV_REFERENCE_FRAME ref_frames[2]) {
1460 if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
1461 frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
1462 (ref_frames[1] == NONE ||
1463 frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
1464 int rfc = mode_context[ref_frames[0]];
1465 int c1 = cost_mv_ref(cpi, NEARMV, rfc);
1466 int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
1467 int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
1469 if (this_mode == NEARMV) {
1470 if (c1 > c3) return 0;
1471 } else if (this_mode == NEARESTMV) {
1472 if (c2 > c3) return 0;
1474 assert(this_mode == ZEROMV);
1475 if (ref_frames[1] == NONE) {
1476 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
1477 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
1480 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
1481 frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
1482 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
1483 frame_mv[NEARMV][ref_frames[1]].as_int == 0))
1491 static int64_t rd_pick_best_sub8x8_mode(VP9_COMP *cpi, MACROBLOCK *x,
1492 const TileInfo * const tile,
1493 int_mv *best_ref_mv,
1494 int_mv *second_best_ref_mv,
1495 int64_t best_rd, int *returntotrate,
1497 int64_t *returndistortion,
1498 int *skippable, int64_t *psse,
1500 int_mv seg_mvs[4][MAX_REF_FRAMES],
1501 BEST_SEG_INFO *bsi_buf, int filter_idx,
1502 int mi_row, int mi_col) {
1504 BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
1505 MACROBLOCKD *xd = &x->e_mbd;
1506 MODE_INFO *mi = xd->mi[0].src_mi;
1507 MB_MODE_INFO *mbmi = &mi->mbmi;
1509 int k, br = 0, idx, idy;
1510 int64_t bd = 0, block_sse = 0;
1511 PREDICTION_MODE this_mode;
1512 VP9_COMMON *cm = &cpi->common;
1513 struct macroblock_plane *const p = &x->plane[0];
1514 struct macroblockd_plane *const pd = &xd->plane[0];
1515 const int label_count = 4;
1516 int64_t this_segment_rd = 0;
1517 int label_mv_thresh;
1518 int segmentyrate = 0;
1519 const BLOCK_SIZE bsize = mbmi->sb_type;
1520 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1521 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1522 ENTROPY_CONTEXT t_above[2], t_left[2];
1523 int subpelmv = 1, have_ref = 0;
1524 const int has_second_rf = has_second_ref(mbmi);
1525 const int inter_mode_mask = cpi->sf.inter_mode_mask[bsize];
1529 bsi->segment_rd = best_rd;
1530 bsi->ref_mv[0] = best_ref_mv;
1531 bsi->ref_mv[1] = second_best_ref_mv;
1532 bsi->mvp.as_int = best_ref_mv->as_int;
1533 bsi->mvthresh = mvthresh;
1535 for (i = 0; i < 4; i++)
1536 bsi->modes[i] = ZEROMV;
1538 vpx_memcpy(t_above, pd->above_context, sizeof(t_above));
1539 vpx_memcpy(t_left, pd->left_context, sizeof(t_left));
1541 // 64 makes this threshold really big effectively
1542 // making it so that we very rarely check mvs on
1543 // segments. setting this to 1 would make mv thresh
1544 // roughly equal to what it is for macroblocks
1545 label_mv_thresh = 1 * bsi->mvthresh / label_count;
1547 // Segmentation method overheads
1548 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1549 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1550 // TODO(jingning,rbultje): rewrite the rate-distortion optimization
1551 // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
1552 int_mv mode_mv[MB_MODE_COUNT][2];
1553 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1554 PREDICTION_MODE mode_selected = ZEROMV;
1555 int64_t best_rd = INT64_MAX;
1556 const int i = idy * 2 + idx;
1559 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1560 const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
1561 frame_mv[ZEROMV][frame].as_int = 0;
1562 vp9_append_sub8x8_mvs_for_idx(cm, xd, tile, i, ref, mi_row, mi_col,
1563 &frame_mv[NEARESTMV][frame],
1564 &frame_mv[NEARMV][frame]);
1567 // search for the best motion vector on this segment
1568 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1569 const struct buf_2d orig_src = x->plane[0].src;
1570 struct buf_2d orig_pre[2];
1572 mode_idx = INTER_OFFSET(this_mode);
1573 bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
1574 if (!(inter_mode_mask & (1 << this_mode)))
1577 if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv,
1578 this_mode, mbmi->ref_frame))
1581 vpx_memcpy(orig_pre, pd->pre, sizeof(orig_pre));
1582 vpx_memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
1583 sizeof(bsi->rdstat[i][mode_idx].ta));
1584 vpx_memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
1585 sizeof(bsi->rdstat[i][mode_idx].tl));
1587 // motion search for newmv (single predictor case only)
1588 if (!has_second_rf && this_mode == NEWMV &&
1589 seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV) {
1590 MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
1592 int thissme, bestsme = INT_MAX;
1593 int sadpb = x->sadperbit4;
1598 /* Is the best so far sufficiently good that we cant justify doing
1599 * and new motion search. */
1600 if (best_rd < label_mv_thresh)
1603 if (cpi->oxcf.mode != BEST) {
1604 // use previous block's result as next block's MV predictor.
1606 bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
1608 bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
1612 max_mv = x->max_mv_context[mbmi->ref_frame[0]];
1614 max_mv = MAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
1616 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
1617 // Take wtd average of the step_params based on the last frame's
1618 // max mv magnitude and the best ref mvs of the current block for
1619 // the given reference.
1620 step_param = (vp9_init_search_range(max_mv) +
1621 cpi->mv_step_param) / 2;
1623 step_param = cpi->mv_step_param;
1626 mvp_full.row = bsi->mvp.as_mv.row >> 3;
1627 mvp_full.col = bsi->mvp.as_mv.col >> 3;
1629 if (cpi->sf.adaptive_motion_search) {
1630 mvp_full.row = x->pred_mv[mbmi->ref_frame[0]].row >> 3;
1631 mvp_full.col = x->pred_mv[mbmi->ref_frame[0]].col >> 3;
1632 step_param = MAX(step_param, 8);
1635 // adjust src pointer for this block
1638 vp9_set_mv_search_range(x, &bsi->ref_mv[0]->as_mv);
1640 bestsme = vp9_full_pixel_search(
1641 cpi, x, bsize, &mvp_full, step_param, sadpb,
1642 cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? sad_list : NULL,
1643 &bsi->ref_mv[0]->as_mv, new_mv,
1646 // Should we do a full search (best quality only)
1647 if (cpi->oxcf.mode == BEST) {
1648 int_mv *const best_mv = &mi->bmi[i].as_mv[0];
1649 /* Check if mvp_full is within the range. */
1650 clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max,
1651 x->mv_row_min, x->mv_row_max);
1652 thissme = cpi->full_search_sad(x, &mvp_full,
1653 sadpb, 16, &cpi->fn_ptr[bsize],
1654 &bsi->ref_mv[0]->as_mv,
1656 sad_list[1] = sad_list[2] = sad_list[3] = sad_list[4] = INT_MAX;
1657 if (thissme < bestsme) {
1659 *new_mv = best_mv->as_mv;
1661 // The full search result is actually worse so re-instate the
1662 // previous best vector
1663 best_mv->as_mv = *new_mv;
1667 if (bestsme < INT_MAX) {
1669 cpi->find_fractional_mv_step(
1672 &bsi->ref_mv[0]->as_mv,
1673 cm->allow_high_precision_mv,
1674 x->errorperbit, &cpi->fn_ptr[bsize],
1675 cpi->sf.mv.subpel_force_stop,
1676 cpi->sf.mv.subpel_iters_per_step,
1677 cond_sad_list(cpi, sad_list),
1678 x->nmvjointcost, x->mvcost,
1680 &x->pred_sse[mbmi->ref_frame[0]],
1683 // save motion search result for use in compound prediction
1684 seg_mvs[i][mbmi->ref_frame[0]].as_mv = *new_mv;
1687 if (cpi->sf.adaptive_motion_search)
1688 x->pred_mv[mbmi->ref_frame[0]] = *new_mv;
1690 // restore src pointers
1691 mi_buf_restore(x, orig_src, orig_pre);
1694 if (has_second_rf) {
1695 if (seg_mvs[i][mbmi->ref_frame[1]].as_int == INVALID_MV ||
1696 seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV)
1700 if (has_second_rf && this_mode == NEWMV &&
1701 mbmi->interp_filter == EIGHTTAP) {
1702 // adjust src pointers
1704 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
1706 joint_motion_search(cpi, x, bsize, frame_mv[this_mode],
1707 mi_row, mi_col, seg_mvs[i],
1709 seg_mvs[i][mbmi->ref_frame[0]].as_int =
1710 frame_mv[this_mode][mbmi->ref_frame[0]].as_int;
1711 seg_mvs[i][mbmi->ref_frame[1]].as_int =
1712 frame_mv[this_mode][mbmi->ref_frame[1]].as_int;
1714 // restore src pointers
1715 mi_buf_restore(x, orig_src, orig_pre);
1718 bsi->rdstat[i][mode_idx].brate =
1719 set_and_cost_bmi_mvs(cpi, xd, i, this_mode, mode_mv[this_mode],
1720 frame_mv, seg_mvs[i], bsi->ref_mv,
1721 x->nmvjointcost, x->mvcost);
1723 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1724 bsi->rdstat[i][mode_idx].mvs[ref].as_int =
1725 mode_mv[this_mode][ref].as_int;
1726 if (num_4x4_blocks_wide > 1)
1727 bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int =
1728 mode_mv[this_mode][ref].as_int;
1729 if (num_4x4_blocks_high > 1)
1730 bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int =
1731 mode_mv[this_mode][ref].as_int;
1734 // Trap vectors that reach beyond the UMV borders
1735 if (mv_check_bounds(x, &mode_mv[this_mode][0].as_mv) ||
1737 mv_check_bounds(x, &mode_mv[this_mode][1].as_mv)))
1740 if (filter_idx > 0) {
1741 BEST_SEG_INFO *ref_bsi = bsi_buf;
1745 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1746 subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
1747 have_ref &= mode_mv[this_mode][ref].as_int ==
1748 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1751 if (filter_idx > 1 && !subpelmv && !have_ref) {
1752 ref_bsi = bsi_buf + 1;
1754 for (ref = 0; ref < 1 + has_second_rf; ++ref)
1755 have_ref &= mode_mv[this_mode][ref].as_int ==
1756 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1759 if (!subpelmv && have_ref &&
1760 ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1761 vpx_memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
1762 sizeof(SEG_RDSTAT));
1763 if (num_4x4_blocks_wide > 1)
1764 bsi->rdstat[i + 1][mode_idx].eobs =
1765 ref_bsi->rdstat[i + 1][mode_idx].eobs;
1766 if (num_4x4_blocks_high > 1)
1767 bsi->rdstat[i + 2][mode_idx].eobs =
1768 ref_bsi->rdstat[i + 2][mode_idx].eobs;
1770 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
1771 mode_selected = this_mode;
1772 best_rd = bsi->rdstat[i][mode_idx].brdcost;
1778 bsi->rdstat[i][mode_idx].brdcost =
1779 encode_inter_mb_segment(cpi, x,
1780 bsi->segment_rd - this_segment_rd, i,
1781 &bsi->rdstat[i][mode_idx].byrate,
1782 &bsi->rdstat[i][mode_idx].bdist,
1783 &bsi->rdstat[i][mode_idx].bsse,
1784 bsi->rdstat[i][mode_idx].ta,
1785 bsi->rdstat[i][mode_idx].tl,
1787 if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1788 bsi->rdstat[i][mode_idx].brdcost += RDCOST(x->rdmult, x->rddiv,
1789 bsi->rdstat[i][mode_idx].brate, 0);
1790 bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
1791 bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
1792 if (num_4x4_blocks_wide > 1)
1793 bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
1794 if (num_4x4_blocks_high > 1)
1795 bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
1798 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
1799 mode_selected = this_mode;
1800 best_rd = bsi->rdstat[i][mode_idx].brdcost;
1802 } /*for each 4x4 mode*/
1804 if (best_rd == INT64_MAX) {
1806 for (iy = i + 1; iy < 4; ++iy)
1807 for (midx = 0; midx < INTER_MODES; ++midx)
1808 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
1809 bsi->segment_rd = INT64_MAX;
1813 mode_idx = INTER_OFFSET(mode_selected);
1814 vpx_memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
1815 vpx_memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
1817 set_and_cost_bmi_mvs(cpi, xd, i, mode_selected, mode_mv[mode_selected],
1818 frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
1821 br += bsi->rdstat[i][mode_idx].brate;
1822 bd += bsi->rdstat[i][mode_idx].bdist;
1823 block_sse += bsi->rdstat[i][mode_idx].bsse;
1824 segmentyrate += bsi->rdstat[i][mode_idx].byrate;
1825 this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
1827 if (this_segment_rd > bsi->segment_rd) {
1829 for (iy = i + 1; iy < 4; ++iy)
1830 for (midx = 0; midx < INTER_MODES; ++midx)
1831 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
1832 bsi->segment_rd = INT64_MAX;
1836 } /* for each label */
1840 bsi->segment_yrate = segmentyrate;
1841 bsi->segment_rd = this_segment_rd;
1842 bsi->sse = block_sse;
1844 // update the coding decisions
1845 for (k = 0; k < 4; ++k)
1846 bsi->modes[k] = mi->bmi[k].as_mode;
1848 if (bsi->segment_rd > best_rd)
1850 /* set it to the best */
1851 for (i = 0; i < 4; i++) {
1852 mode_idx = INTER_OFFSET(bsi->modes[i]);
1853 mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
1854 if (has_second_ref(mbmi))
1855 mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
1856 x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
1857 mi->bmi[i].as_mode = bsi->modes[i];
1861 * used to set mbmi->mv.as_int
1863 *returntotrate = bsi->r;
1864 *returndistortion = bsi->d;
1865 *returnyrate = bsi->segment_yrate;
1866 *skippable = vp9_is_skippable_in_plane(x, BLOCK_8X8, 0);
1868 mbmi->mode = bsi->modes[3];
1870 return bsi->segment_rd;
1873 static void estimate_ref_frame_costs(const VP9_COMMON *cm,
1874 const MACROBLOCKD *xd,
1876 unsigned int *ref_costs_single,
1877 unsigned int *ref_costs_comp,
1878 vp9_prob *comp_mode_p) {
1879 int seg_ref_active = vp9_segfeature_active(&cm->seg, segment_id,
1881 if (seg_ref_active) {
1882 vpx_memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
1883 vpx_memset(ref_costs_comp, 0, MAX_REF_FRAMES * sizeof(*ref_costs_comp));
1886 vp9_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
1887 vp9_prob comp_inter_p = 128;
1889 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
1890 comp_inter_p = vp9_get_reference_mode_prob(cm, xd);
1891 *comp_mode_p = comp_inter_p;
1896 ref_costs_single[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1898 if (cm->reference_mode != COMPOUND_REFERENCE) {
1899 vp9_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1900 vp9_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1901 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
1903 if (cm->reference_mode == REFERENCE_MODE_SELECT)
1904 base_cost += vp9_cost_bit(comp_inter_p, 0);
1906 ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] =
1907 ref_costs_single[ALTREF_FRAME] = base_cost;
1908 ref_costs_single[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1909 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1910 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1911 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1912 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1914 ref_costs_single[LAST_FRAME] = 512;
1915 ref_costs_single[GOLDEN_FRAME] = 512;
1916 ref_costs_single[ALTREF_FRAME] = 512;
1918 if (cm->reference_mode != SINGLE_REFERENCE) {
1919 vp9_prob ref_comp_p = vp9_get_pred_prob_comp_ref_p(cm, xd);
1920 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
1922 if (cm->reference_mode == REFERENCE_MODE_SELECT)
1923 base_cost += vp9_cost_bit(comp_inter_p, 1);
1925 ref_costs_comp[LAST_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 0);
1926 ref_costs_comp[GOLDEN_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 1);
1928 ref_costs_comp[LAST_FRAME] = 512;
1929 ref_costs_comp[GOLDEN_FRAME] = 512;
1934 static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
1936 int64_t comp_pred_diff[REFERENCE_MODES],
1937 const int64_t tx_size_diff[TX_MODES],
1938 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS],
1940 MACROBLOCKD *const xd = &x->e_mbd;
1942 // Take a snapshot of the coding context so it can be
1943 // restored if we decide to encode this way
1944 ctx->skip = x->skip;
1945 ctx->skippable = skippable;
1946 ctx->best_mode_index = mode_index;
1947 ctx->mic = *xd->mi[0].src_mi;
1948 ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
1949 ctx->comp_pred_diff = (int)comp_pred_diff[COMPOUND_REFERENCE];
1950 ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
1952 vpx_memcpy(ctx->tx_rd_diff, tx_size_diff, sizeof(ctx->tx_rd_diff));
1953 vpx_memcpy(ctx->best_filter_diff, best_filter_diff,
1954 sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
1957 static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
1958 const TileInfo *const tile,
1959 MV_REFERENCE_FRAME ref_frame,
1960 BLOCK_SIZE block_size,
1961 int mi_row, int mi_col,
1962 int_mv frame_nearest_mv[MAX_REF_FRAMES],
1963 int_mv frame_near_mv[MAX_REF_FRAMES],
1964 struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
1965 const VP9_COMMON *cm = &cpi->common;
1966 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1967 MACROBLOCKD *const xd = &x->e_mbd;
1968 MODE_INFO *const mi = xd->mi[0].src_mi;
1969 int_mv *const candidates = mi->mbmi.ref_mvs[ref_frame];
1970 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
1972 // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
1973 // use the UV scaling factors.
1974 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
1976 // Gets an initial list of candidate vectors from neighbours and orders them
1977 vp9_find_mv_refs(cm, xd, tile, mi, ref_frame, candidates, mi_row, mi_col);
1979 // Candidate refinement carried out at encoder and decoder
1980 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1981 &frame_nearest_mv[ref_frame],
1982 &frame_near_mv[ref_frame]);
1984 // Further refinement that is encode side only to test the top few candidates
1985 // in full and choose the best as the centre point for subsequent searches.
1986 // The current implementation doesn't support scaling.
1987 if (!vp9_is_scaled(sf) && block_size >= BLOCK_8X8)
1988 vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
1989 ref_frame, block_size);
1992 static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
1994 int mi_row, int mi_col,
1995 int_mv *tmp_mv, int *rate_mv) {
1996 MACROBLOCKD *xd = &x->e_mbd;
1997 const VP9_COMMON *cm = &cpi->common;
1998 MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
1999 struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
2000 int bestsme = INT_MAX;
2002 int sadpb = x->sadperbit16;
2004 int ref = mbmi->ref_frame[0];
2005 MV ref_mv = mbmi->ref_mvs[ref][0].as_mv;
2007 int tmp_col_min = x->mv_col_min;
2008 int tmp_col_max = x->mv_col_max;
2009 int tmp_row_min = x->mv_row_min;
2010 int tmp_row_max = x->mv_row_max;
2013 const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi,
2017 pred_mv[0] = mbmi->ref_mvs[ref][0].as_mv;
2018 pred_mv[1] = mbmi->ref_mvs[ref][1].as_mv;
2019 pred_mv[2] = x->pred_mv[ref];
2021 if (scaled_ref_frame) {
2023 // Swap out the reference frame for a version that's been scaled to
2024 // match the resolution of the current frame, allowing the existing
2025 // motion search code to be used without additional modifications.
2026 for (i = 0; i < MAX_MB_PLANE; i++)
2027 backup_yv12[i] = xd->plane[i].pre[0];
2029 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
2032 vp9_set_mv_search_range(x, &ref_mv);
2034 // Work out the size of the first step in the mv step search.
2035 // 0 here is maximum length first step. 1 is MAX >> 1 etc.
2036 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
2037 // Take wtd average of the step_params based on the last frame's
2038 // max mv magnitude and that based on the best ref mvs of the current
2039 // block for the given reference.
2040 step_param = (vp9_init_search_range(x->max_mv_context[ref]) +
2041 cpi->mv_step_param) / 2;
2043 step_param = cpi->mv_step_param;
2046 if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64) {
2047 int boffset = 2 * (b_width_log2(BLOCK_64X64) - MIN(b_height_log2(bsize),
2048 b_width_log2(bsize)));
2049 step_param = MAX(step_param, boffset);
2052 if (cpi->sf.adaptive_motion_search) {
2053 int bwl = b_width_log2(bsize);
2054 int bhl = b_height_log2(bsize);
2056 int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
2061 for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
2062 if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
2063 x->pred_mv[ref].row = 0;
2064 x->pred_mv[ref].col = 0;
2065 tmp_mv->as_int = INVALID_MV;
2067 if (scaled_ref_frame) {
2069 for (i = 0; i < MAX_MB_PLANE; i++)
2070 xd->plane[i].pre[0] = backup_yv12[i];
2077 mvp_full = pred_mv[x->mv_best_ref_index[ref]];
2082 bestsme = vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
2083 cond_sad_list(cpi, sad_list),
2084 &ref_mv, &tmp_mv->as_mv, INT_MAX, 1);
2086 x->mv_col_min = tmp_col_min;
2087 x->mv_col_max = tmp_col_max;
2088 x->mv_row_min = tmp_row_min;
2089 x->mv_row_max = tmp_row_max;
2091 if (bestsme < INT_MAX) {
2092 int dis; /* TODO: use dis in distortion calculation later. */
2093 cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
2094 cm->allow_high_precision_mv,
2096 &cpi->fn_ptr[bsize],
2097 cpi->sf.mv.subpel_force_stop,
2098 cpi->sf.mv.subpel_iters_per_step,
2099 cond_sad_list(cpi, sad_list),
2100 x->nmvjointcost, x->mvcost,
2101 &dis, &x->pred_sse[ref], NULL, 0, 0);
2103 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
2104 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2106 if (cpi->sf.adaptive_motion_search)
2107 x->pred_mv[ref] = tmp_mv->as_mv;
2109 if (scaled_ref_frame) {
2111 for (i = 0; i < MAX_MB_PLANE; i++)
2112 xd->plane[i].pre[0] = backup_yv12[i];
2116 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
2119 int mi_row, int mi_col,
2120 int_mv single_newmv[MAX_REF_FRAMES],
2122 const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
2123 const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
2124 MACROBLOCKD *xd = &x->e_mbd;
2125 MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
2126 const int refs[2] = { mbmi->ref_frame[0],
2127 mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1] };
2130 // Prediction buffer from second frame.
2131 #if CONFIG_VP9_HIGHBITDEPTH
2132 uint8_t *second_pred;
2133 uint8_t *second_pred_alloc;
2135 uint8_t *second_pred = vpx_memalign(16, pw * ph * sizeof(uint8_t));
2136 #endif // CONFIG_VP9_HIGHBITDEPTH
2137 const InterpKernel *kernel = vp9_get_interp_kernel(mbmi->interp_filter);
2139 // Do joint motion search in compound mode to get more accurate mv.
2140 struct buf_2d backup_yv12[2][MAX_MB_PLANE];
2141 struct buf_2d scaled_first_yv12 = xd->plane[0].pre[0];
2142 int last_besterr[2] = {INT_MAX, INT_MAX};
2143 const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
2144 vp9_get_scaled_ref_frame(cpi, mbmi->ref_frame[0]),
2145 vp9_get_scaled_ref_frame(cpi, mbmi->ref_frame[1])
2147 #if CONFIG_VP9_HIGHBITDEPTH
2148 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2149 second_pred_alloc = vpx_memalign(16, pw * ph * sizeof(uint16_t));
2150 second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc);
2152 second_pred_alloc = vpx_memalign(16, pw * ph * sizeof(uint8_t));
2153 second_pred = second_pred_alloc;
2155 #endif // CONFIG_VP9_HIGHBITDEPTH
2157 for (ref = 0; ref < 2; ++ref) {
2158 ref_mv[ref] = mbmi->ref_mvs[refs[ref]][0];
2160 if (scaled_ref_frame[ref]) {
2162 // Swap out the reference frame for a version that's been scaled to
2163 // match the resolution of the current frame, allowing the existing
2164 // motion search code to be used without additional modifications.
2165 for (i = 0; i < MAX_MB_PLANE; i++)
2166 backup_yv12[ref][i] = xd->plane[i].pre[ref];
2167 vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
2171 frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
2174 // Allow joint search multiple times iteratively for each ref frame
2175 // and break out the search loop if it couldn't find better mv.
2176 for (ite = 0; ite < 4; ite++) {
2177 struct buf_2d ref_yv12[2];
2178 int bestsme = INT_MAX;
2179 int sadpb = x->sadperbit16;
2181 int search_range = 3;
2183 int tmp_col_min = x->mv_col_min;
2184 int tmp_col_max = x->mv_col_max;
2185 int tmp_row_min = x->mv_row_min;
2186 int tmp_row_max = x->mv_row_max;
2189 // Initialized here because of compiler problem in Visual Studio.
2190 ref_yv12[0] = xd->plane[0].pre[0];
2191 ref_yv12[1] = xd->plane[0].pre[1];
2193 // Get pred block from second frame.
2194 #if CONFIG_VP9_HIGHBITDEPTH
2195 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2196 vp9_high_build_inter_predictor(ref_yv12[!id].buf,
2197 ref_yv12[!id].stride,
2199 &frame_mv[refs[!id]].as_mv,
2200 &xd->block_refs[!id]->sf,
2202 kernel, MV_PRECISION_Q3,
2203 mi_col * MI_SIZE, mi_row * MI_SIZE,
2206 vp9_build_inter_predictor(ref_yv12[!id].buf,
2207 ref_yv12[!id].stride,
2209 &frame_mv[refs[!id]].as_mv,
2210 &xd->block_refs[!id]->sf,
2212 kernel, MV_PRECISION_Q3,
2213 mi_col * MI_SIZE, mi_row * MI_SIZE);
2216 vp9_build_inter_predictor(ref_yv12[!id].buf,
2217 ref_yv12[!id].stride,
2219 &frame_mv[refs[!id]].as_mv,
2220 &xd->block_refs[!id]->sf,
2222 kernel, MV_PRECISION_Q3,
2223 mi_col * MI_SIZE, mi_row * MI_SIZE);
2224 #endif // CONFIG_VP9_HIGHBITDEPTH
2226 // Compound motion search on first ref frame.
2228 xd->plane[0].pre[0] = ref_yv12[id];
2229 vp9_set_mv_search_range(x, &ref_mv[id].as_mv);
2231 // Use mv result from single mode as mvp.
2232 tmp_mv = frame_mv[refs[id]].as_mv;
2237 // Small-range full-pixel motion search
2238 bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb,
2240 &cpi->fn_ptr[bsize],
2241 &ref_mv[id].as_mv, second_pred);
2242 if (bestsme < INT_MAX)
2243 bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
2244 second_pred, &cpi->fn_ptr[bsize], 1);
2246 x->mv_col_min = tmp_col_min;
2247 x->mv_col_max = tmp_col_max;
2248 x->mv_row_min = tmp_row_min;
2249 x->mv_row_max = tmp_row_max;
2251 if (bestsme < INT_MAX) {
2252 int dis; /* TODO: use dis in distortion calculation later. */
2254 bestsme = cpi->find_fractional_mv_step(
2257 cpi->common.allow_high_precision_mv,
2259 &cpi->fn_ptr[bsize],
2260 0, cpi->sf.mv.subpel_iters_per_step,
2262 x->nmvjointcost, x->mvcost,
2263 &dis, &sse, second_pred,
2268 xd->plane[0].pre[0] = scaled_first_yv12;
2270 if (bestsme < last_besterr[id]) {
2271 frame_mv[refs[id]].as_mv = tmp_mv;
2272 last_besterr[id] = bestsme;
2280 for (ref = 0; ref < 2; ++ref) {
2281 if (scaled_ref_frame[ref]) {
2282 // restore the predictor
2284 for (i = 0; i < MAX_MB_PLANE; i++)
2285 xd->plane[i].pre[ref] = backup_yv12[ref][i];
2288 *rate_mv += vp9_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
2289 &mbmi->ref_mvs[refs[ref]][0].as_mv,
2290 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2293 #if CONFIG_VP9_HIGHBITDEPTH
2294 vpx_free(second_pred_alloc);
2296 vpx_free(second_pred);
2297 #endif // CONFIG_VP9_HIGHBITDEPTH
2300 static INLINE void restore_dst_buf(MACROBLOCKD *xd,
2301 uint8_t *orig_dst[MAX_MB_PLANE],
2302 int orig_dst_stride[MAX_MB_PLANE]) {
2304 for (i = 0; i < MAX_MB_PLANE; i++) {
2305 xd->plane[i].dst.buf = orig_dst[i];
2306 xd->plane[i].dst.stride = orig_dst_stride[i];
2310 static void rd_encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x,
2311 BLOCK_SIZE bsize, int *rate2,
2312 int64_t *distortion, int64_t *distortion_uv,
2313 int *disable_skip) {
2314 VP9_COMMON *cm = &cpi->common;
2315 MACROBLOCKD *xd = &x->e_mbd;
2316 const BLOCK_SIZE y_size = get_plane_block_size(bsize, &xd->plane[0]);
2317 const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
2318 unsigned int var, sse;
2319 // Skipping threshold for ac.
2320 unsigned int thresh_ac;
2321 // Skipping threshold for dc
2322 unsigned int thresh_dc;
2324 var = cpi->fn_ptr[y_size].vf(x->plane[0].src.buf, x->plane[0].src.stride,
2325 xd->plane[0].dst.buf,
2326 xd->plane[0].dst.stride, &sse);
2328 if (x->encode_breakout > 0) {
2329 // Set a maximum for threshold to avoid big PSNR loss in low bitrate
2330 // case. Use extreme low threshold for static frames to limit skipping.
2331 const unsigned int max_thresh = (cpi->allow_encode_breakout ==
2332 ENCODE_BREAKOUT_LIMITED) ? 128 : 36000;
2333 // The encode_breakout input
2334 const unsigned int min_thresh =
2335 MIN(((unsigned int)x->encode_breakout << 4), max_thresh);
2337 // Calculate threshold according to dequant value.
2338 thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9;
2339 #if CONFIG_VP9_HIGHBITDEPTH
2340 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2341 const int shift = 2 * xd->bd - 16;
2343 thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift);
2345 #endif // CONFIG_VP9_HIGHBITDEPTH
2346 thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
2348 // Adjust threshold according to partition size.
2349 thresh_ac >>= 8 - (b_width_log2(bsize) +
2350 b_height_log2(bsize));
2351 thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
2352 #if CONFIG_VP9_HIGHBITDEPTH
2353 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2354 const int shift = 2 * xd->bd - 16;
2356 thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift);
2358 #endif // CONFIG_VP9_HIGHBITDEPTH
2364 // Y skipping condition checking
2365 if (sse < thresh_ac || sse == 0) {
2366 // dc skipping checking
2367 if ((sse - var) < thresh_dc || sse == var) {
2368 unsigned int sse_u, sse_v;
2369 unsigned int var_u, var_v;
2371 var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
2372 x->plane[1].src.stride,
2373 xd->plane[1].dst.buf,
2374 xd->plane[1].dst.stride, &sse_u);
2376 // U skipping condition checking
2377 if ((sse_u * 4 < thresh_ac || sse_u == 0) &&
2378 (sse_u - var_u < thresh_dc || sse_u == var_u)) {
2379 var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
2380 x->plane[2].src.stride,
2381 xd->plane[2].dst.buf,
2382 xd->plane[2].dst.stride, &sse_v);
2384 // V skipping condition checking
2385 if ((sse_v * 4 < thresh_ac || sse_v == 0) &&
2386 (sse_v - var_v < thresh_dc || sse_v == var_v)) {
2389 // The cost of skip bit needs to be added.
2390 *rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2392 // Scaling factor for SSE from spatial domain to frequency domain
2393 // is 16. Adjust distortion accordingly.
2394 *distortion_uv = (sse_u + sse_v) << 4;
2395 *distortion = (sse << 4) + *distortion_uv;
2404 static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
2406 int64_t txfm_cache[],
2407 int *rate2, int64_t *distortion,
2409 int *rate_y, int64_t *distortion_y,
2410 int *rate_uv, int64_t *distortion_uv,
2412 int_mv (*mode_mv)[MAX_REF_FRAMES],
2413 int mi_row, int mi_col,
2414 int_mv single_newmv[MAX_REF_FRAMES],
2415 INTERP_FILTER (*single_filter)[MAX_REF_FRAMES],
2416 int (*single_skippable)[MAX_REF_FRAMES],
2418 const int64_t ref_best_rd) {
2419 VP9_COMMON *cm = &cpi->common;
2420 RD_OPT *rd_opt = &cpi->rd;
2421 MACROBLOCKD *xd = &x->e_mbd;
2422 MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
2423 const int is_comp_pred = has_second_ref(mbmi);
2424 const int this_mode = mbmi->mode;
2425 int_mv *frame_mv = mode_mv[this_mode];
2427 int refs[2] = { mbmi->ref_frame[0],
2428 (mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]) };
2430 #if CONFIG_VP9_HIGHBITDEPTH
2431 DECLARE_ALIGNED_ARRAY(16, uint16_t, tmp_buf16, MAX_MB_PLANE * 64 * 64);
2432 DECLARE_ALIGNED_ARRAY(16, uint8_t, tmp_buf8, MAX_MB_PLANE * 64 * 64);
2433 uint8_t *tmp_buf = tmp_buf8;
2435 DECLARE_ALIGNED_ARRAY(16, uint8_t, tmp_buf, MAX_MB_PLANE * 64 * 64);
2436 #endif // CONFIG_VP9_HIGHBITDEPTH
2437 int pred_exists = 0;
2439 int64_t rd, tmp_rd, best_rd = INT64_MAX;
2440 int best_needs_copy = 0;
2441 uint8_t *orig_dst[MAX_MB_PLANE];
2442 int orig_dst_stride[MAX_MB_PLANE];
2444 INTERP_FILTER best_filter = SWITCHABLE;
2445 uint8_t skip_txfm[MAX_MB_PLANE << 2] = {0};
2446 int64_t bsse[MAX_MB_PLANE << 2] = {0};
2448 int bsl = mi_width_log2_lookup[bsize];
2449 int pred_filter_search = cpi->sf.cb_pred_filter_search ?
2450 (((mi_row + mi_col) >> bsl) +
2451 get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
2453 #if CONFIG_VP9_HIGHBITDEPTH
2454 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2455 tmp_buf = CONVERT_TO_BYTEPTR(tmp_buf16);
2459 #endif // CONFIG_VP9_HIGHBITDEPTH
2461 if (pred_filter_search) {
2462 INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE;
2463 if (xd->up_available)
2464 af = xd->mi[-xd->mi_stride].src_mi->mbmi.interp_filter;
2465 if (xd->left_available)
2466 lf = xd->mi[-1].src_mi->mbmi.interp_filter;
2468 if ((this_mode != NEWMV) || (af == lf))
2473 if (frame_mv[refs[0]].as_int == INVALID_MV ||
2474 frame_mv[refs[1]].as_int == INVALID_MV)
2477 if (cpi->sf.adaptive_mode_search) {
2478 if (single_filter[this_mode][refs[0]] ==
2479 single_filter[this_mode][refs[1]])
2480 best_filter = single_filter[this_mode][refs[0]];
2484 if (this_mode == NEWMV) {
2487 // Initialize mv using single prediction mode result.
2488 frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
2489 frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
2491 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
2492 joint_motion_search(cpi, x, bsize, frame_mv,
2493 mi_row, mi_col, single_newmv, &rate_mv);
2495 rate_mv = vp9_mv_bit_cost(&frame_mv[refs[0]].as_mv,
2496 &mbmi->ref_mvs[refs[0]][0].as_mv,
2497 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2498 rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]].as_mv,
2499 &mbmi->ref_mvs[refs[1]][0].as_mv,
2500 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2505 single_motion_search(cpi, x, bsize, mi_row, mi_col,
2507 if (tmp_mv.as_int == INVALID_MV)
2510 frame_mv[refs[0]].as_int =
2511 xd->mi[0].src_mi->bmi[0].as_mv[0].as_int = tmp_mv.as_int;
2512 single_newmv[refs[0]].as_int = tmp_mv.as_int;
2516 for (i = 0; i < is_comp_pred + 1; ++i) {
2517 cur_mv[i] = frame_mv[refs[i]];
2518 // Clip "next_nearest" so that it does not extend to far out of image
2519 if (this_mode != NEWMV)
2520 clamp_mv2(&cur_mv[i].as_mv, xd);
2522 if (mv_check_bounds(x, &cur_mv[i].as_mv))
2524 mbmi->mv[i].as_int = cur_mv[i].as_int;
2527 // do first prediction into the destination buffer. Do the next
2528 // prediction into a temporary buffer. Then keep track of which one
2529 // of these currently holds the best predictor, and use the other
2530 // one for future predictions. In the end, copy from tmp_buf to
2531 // dst if necessary.
2532 for (i = 0; i < MAX_MB_PLANE; i++) {
2533 orig_dst[i] = xd->plane[i].dst.buf;
2534 orig_dst_stride[i] = xd->plane[i].dst.stride;
2537 /* We don't include the cost of the second reference here, because there
2538 * are only three options: Last/Golden, ARF/Last or Golden/ARF, or in other
2539 * words if you present them in that order, the second one is always known
2540 * if the first is known */
2541 *rate2 += cost_mv_ref(cpi, this_mode, mbmi->mode_context[refs[0]]);
2543 if (RDCOST(x->rdmult, x->rddiv, *rate2, 0) > ref_best_rd &&
2544 mbmi->mode != NEARESTMV)
2548 // Are all MVs integer pel for Y and UV
2549 intpel_mv = !mv_has_subpel(&mbmi->mv[0].as_mv);
2551 intpel_mv &= !mv_has_subpel(&mbmi->mv[1].as_mv);
2553 // Search for best switchable filter by checking the variance of
2554 // pred error irrespective of whether the filter will be used
2555 rd_opt->mask_filter = 0;
2556 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
2557 rd_opt->filter_cache[i] = INT64_MAX;
2559 if (cm->interp_filter != BILINEAR) {
2560 if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
2561 best_filter = EIGHTTAP;
2562 } else if (best_filter == SWITCHABLE) {
2564 int tmp_rate_sum = 0;
2565 int64_t tmp_dist_sum = 0;
2567 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
2570 mbmi->interp_filter = i;
2571 rs = vp9_get_switchable_rate(cpi);
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 rd_opt->filter_cache[i] = rd;
2577 rd_opt->filter_cache[SWITCHABLE_FILTERS] =
2578 MIN(rd_opt->filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2579 if (cm->interp_filter == SWITCHABLE)
2581 rd_opt->mask_filter = MAX(rd_opt->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 == mbmi->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);
2607 rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum);
2608 rd_opt->filter_cache[i] = rd;
2609 rd_opt->filter_cache[SWITCHABLE_FILTERS] =
2610 MIN(rd_opt->filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2611 if (cm->interp_filter == SWITCHABLE)
2613 rd_opt->mask_filter = MAX(rd_opt->mask_filter, rd);
2615 if (i == 0 && intpel_mv) {
2616 tmp_rate_sum = rate_sum;
2617 tmp_dist_sum = dist_sum;
2621 if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2622 if (rd / 2 > ref_best_rd) {
2623 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2627 newbest = i == 0 || rd < best_rd;
2631 best_filter = mbmi->interp_filter;
2632 if (cm->interp_filter == SWITCHABLE && i && !intpel_mv)
2633 best_needs_copy = !best_needs_copy;
2634 vpx_memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2635 vpx_memcpy(bsse, x->bsse, sizeof(bsse));
2638 if ((cm->interp_filter == SWITCHABLE && newbest) ||
2639 (cm->interp_filter != SWITCHABLE &&
2640 cm->interp_filter == mbmi->interp_filter)) {
2645 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2648 // Set the appropriate filter
2649 mbmi->interp_filter = cm->interp_filter != SWITCHABLE ?
2650 cm->interp_filter : best_filter;
2651 rs = cm->interp_filter == SWITCHABLE ? vp9_get_switchable_rate(cpi) : 0;
2654 if (best_needs_copy) {
2655 // again temporarily set the buffers to local memory to prevent a memcpy
2656 for (i = 0; i < MAX_MB_PLANE; i++) {
2657 xd->plane[i].dst.buf = tmp_buf + i * 64 * 64;
2658 xd->plane[i].dst.stride = 64;
2661 rd = tmp_rd + RDCOST(x->rdmult, x->rddiv, rs, 0);
2665 // Handles the special case when a filter that is not in the
2666 // switchable list (ex. bilinear) is indicated at the frame level, or
2667 // skip condition holds.
2668 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2669 model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist);
2670 rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
2671 vpx_memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2672 vpx_memcpy(bsse, x->bsse, sizeof(bsse));
2676 single_filter[this_mode][refs[0]] = mbmi->interp_filter;
2678 if (cpi->sf.adaptive_mode_search)
2680 if (single_skippable[this_mode][refs[0]] &&
2681 single_skippable[this_mode][refs[1]])
2682 vpx_memset(skip_txfm, 1, sizeof(skip_txfm));
2684 if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2685 // if current pred_error modeled rd is substantially more than the best
2686 // so far, do not bother doing full rd
2687 if (rd / 2 > ref_best_rd) {
2688 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2693 if (cm->interp_filter == SWITCHABLE)
2696 if (!is_comp_pred) {
2697 if (cpi->allow_encode_breakout)
2698 rd_encode_breakout_test(cpi, x, bsize, rate2, distortion, distortion_uv,
2702 vpx_memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm));
2703 vpx_memcpy(x->bsse, bsse, sizeof(bsse));
2706 int skippable_y, skippable_uv;
2707 int64_t sseuv = INT64_MAX;
2708 int64_t rdcosty = INT64_MAX;
2710 // Y cost and distortion
2711 vp9_subtract_plane(x, bsize, 0);
2712 super_block_yrd(cpi, x, rate_y, distortion_y, &skippable_y, psse,
2713 bsize, txfm_cache, ref_best_rd);
2715 if (*rate_y == INT_MAX) {
2717 *distortion = INT64_MAX;
2718 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2723 *distortion += *distortion_y;
2725 rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
2726 rdcosty = MIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
2728 super_block_uvrd(cpi, x, rate_uv, distortion_uv, &skippable_uv, &sseuv,
2729 bsize, ref_best_rd - rdcosty);
2730 if (*rate_uv == INT_MAX) {
2732 *distortion = INT64_MAX;
2733 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2739 *distortion += *distortion_uv;
2740 *skippable = skippable_y && skippable_uv;
2744 single_skippable[this_mode][refs[0]] = *skippable;
2746 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2747 return 0; // The rate-distortion cost will be re-calculated by caller.
2750 void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
2751 int *returnrate, int64_t *returndist,
2753 PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
2754 VP9_COMMON *const cm = &cpi->common;
2755 MACROBLOCKD *const xd = &x->e_mbd;
2756 struct macroblockd_plane *const pd = xd->plane;
2757 int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
2758 int y_skip = 0, uv_skip = 0;
2759 int64_t dist_y = 0, dist_uv = 0, tx_cache[TX_MODES] = { 0 };
2760 TX_SIZE max_uv_tx_size;
2763 xd->mi[0].src_mi->mbmi.ref_frame[0] = INTRA_FRAME;
2765 if (bsize >= BLOCK_8X8) {
2766 if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2767 &dist_y, &y_skip, bsize, tx_cache,
2768 best_rd) >= best_rd) {
2769 *returnrate = INT_MAX;
2774 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2775 &dist_y, best_rd) >= best_rd) {
2776 *returnrate = INT_MAX;
2780 max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0].src_mi->mbmi.tx_size, bsize,
2781 pd[1].subsampling_x,
2782 pd[1].subsampling_y);
2783 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
2784 &dist_uv, &uv_skip, MAX(BLOCK_8X8, bsize),
2787 if (y_skip && uv_skip) {
2788 *returnrate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
2789 vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2790 *returndist = dist_y + dist_uv;
2791 vp9_zero(ctx->tx_rd_diff);
2794 *returnrate = rate_y + rate_uv + vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
2795 *returndist = dist_y + dist_uv;
2796 if (cpi->sf.tx_size_search_method == USE_FULL_RD)
2797 for (i = 0; i < TX_MODES; i++) {
2798 if (tx_cache[i] < INT64_MAX && tx_cache[cm->tx_mode] < INT64_MAX)
2799 ctx->tx_rd_diff[i] = tx_cache[i] - tx_cache[cm->tx_mode];
2801 ctx->tx_rd_diff[i] = 0;
2805 ctx->mic = *xd->mi[0].src_mi;
2808 static void update_rd_thresh_fact(VP9_COMP *cpi, int bsize,
2809 int best_mode_index) {
2810 if (cpi->sf.adaptive_rd_thresh > 0) {
2811 const int top_mode = bsize < BLOCK_8X8 ? MAX_REFS : MAX_MODES;
2813 for (mode = 0; mode < top_mode; ++mode) {
2814 const BLOCK_SIZE min_size = MAX(bsize - 1, BLOCK_4X4);
2815 const BLOCK_SIZE max_size = MIN(bsize + 2, BLOCK_64X64);
2817 for (bs = min_size; bs <= max_size; ++bs) {
2818 int *const fact = &cpi->rd.thresh_freq_fact[bs][mode];
2819 if (mode == best_mode_index) {
2820 *fact -= (*fact >> 4);
2822 *fact = MIN(*fact + RD_THRESH_INC,
2823 cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
2830 int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
2831 const TileInfo *const tile,
2832 int mi_row, int mi_col,
2834 int64_t *returndistortion,
2836 PICK_MODE_CONTEXT *ctx,
2837 int64_t best_rd_so_far) {
2838 VP9_COMMON *const cm = &cpi->common;
2839 RD_OPT *const rd_opt = &cpi->rd;
2840 MACROBLOCKD *const xd = &x->e_mbd;
2841 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
2842 const struct segmentation *const seg = &cm->seg;
2843 struct macroblockd_plane *const pd = xd->plane;
2844 PREDICTION_MODE this_mode;
2845 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
2846 unsigned char segment_id = mbmi->segment_id;
2847 int comp_pred, i, k;
2848 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
2849 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2850 int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
2851 INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES];
2852 int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES];
2853 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2855 int64_t best_rd = best_rd_so_far;
2856 int64_t best_tx_rd[TX_MODES];
2857 int64_t best_tx_diff[TX_MODES];
2858 int64_t best_pred_diff[REFERENCE_MODES];
2859 int64_t best_pred_rd[REFERENCE_MODES];
2860 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
2861 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
2862 MB_MODE_INFO best_mbmode;
2863 int best_mode_skippable = 0;
2864 int midx, best_mode_index = -1;
2865 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
2866 vp9_prob comp_mode_p;
2867 int64_t best_intra_rd = INT64_MAX;
2868 unsigned int best_pred_sse = UINT_MAX;
2869 PREDICTION_MODE best_intra_mode = DC_PRED;
2870 int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
2871 int64_t dist_uv[TX_SIZES];
2872 int skip_uv[TX_SIZES];
2873 PREDICTION_MODE mode_uv[TX_SIZES];
2874 const int intra_cost_penalty = vp9_get_intra_cost_penalty(
2875 cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
2877 uint8_t ref_frame_skip_mask[2] = { 0 };
2878 uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 };
2879 int mode_skip_start = cpi->sf.mode_skip_start + 1;
2880 const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
2881 const int *const rd_thresh_freq_fact = rd_opt->thresh_freq_fact[bsize];
2882 int64_t mode_threshold[MAX_MODES];
2883 int *mode_map = rd_opt->mode_map[bsize];
2884 const int mode_search_skip_flags = cpi->sf.mode_search_skip_flags;
2885 vp9_zero(best_mbmode);
2887 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
2889 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
2892 for (i = 0; i < REFERENCE_MODES; ++i)
2893 best_pred_rd[i] = INT64_MAX;
2894 for (i = 0; i < TX_MODES; i++)
2895 best_tx_rd[i] = INT64_MAX;
2896 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
2897 best_filter_rd[i] = INT64_MAX;
2898 for (i = 0; i < TX_SIZES; i++)
2899 rate_uv_intra[i] = INT_MAX;
2900 for (i = 0; i < MAX_REF_FRAMES; ++i)
2901 x->pred_sse[i] = INT_MAX;
2902 for (i = 0; i < MB_MODE_COUNT; ++i) {
2903 for (k = 0; k < MAX_REF_FRAMES; ++k) {
2904 single_inter_filter[i][k] = SWITCHABLE;
2905 single_skippable[i][k] = 0;
2909 *returnrate = INT_MAX;
2911 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
2912 x->pred_mv_sad[ref_frame] = INT_MAX;
2913 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
2914 setup_buffer_inter(cpi, x, tile, ref_frame, bsize, mi_row, mi_col,
2915 frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
2917 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
2918 frame_mv[ZEROMV][ref_frame].as_int = 0;
2921 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
2922 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) {
2923 // Skip checking missing references in both single and compound reference
2924 // modes. Note that a mode will be skipped iff both reference frames
2926 ref_frame_skip_mask[0] |= (1 << ref_frame);
2927 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
2928 } else if (cpi->sf.reference_masking) {
2929 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
2930 // Skip fixed mv modes for poor references
2931 if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
2932 mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
2937 // If the segment reference frame feature is enabled....
2938 // then do nothing if the current ref frame is not allowed..
2939 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
2940 vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
2941 ref_frame_skip_mask[0] |= (1 << ref_frame);
2942 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
2946 // Disable this drop out case if the ref frame
2947 // segment level feature is enabled for this segment. This is to
2948 // prevent the possibility that we end up unable to pick any mode.
2949 if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
2950 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
2951 // unless ARNR filtering is enabled in which case we want
2952 // an unfiltered alternative. We allow near/nearest as well
2953 // because they may result in zero-zero MVs but be cheaper.
2954 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
2955 ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME);
2956 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
2957 mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
2958 if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
2959 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
2960 if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
2961 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
2965 if (cpi->rc.is_src_frame_alt_ref) {
2966 if (cpi->sf.alt_ref_search_fp) {
2967 mode_skip_mask[ALTREF_FRAME] = 0;
2968 ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME);
2969 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
2973 if (cpi->sf.alt_ref_search_fp)
2974 if (!cm->show_frame)
2975 if (x->pred_mv_sad[ALTREF_FRAME] > (x->pred_mv_sad[GOLDEN_FRAME] << 1))
2976 mode_skip_mask[ALTREF_FRAME] |= INTER_ALL;
2978 if (bsize > cpi->sf.max_intra_bsize) {
2979 ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
2980 ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
2983 mode_skip_mask[INTRA_FRAME] |=
2984 ~(cpi->sf.intra_y_mode_mask[max_txsize_lookup[bsize]]);
2986 for (i = 0; i < MAX_MODES; ++i)
2987 mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
2989 midx = cpi->sf.schedule_mode_search ? mode_skip_start : 0;
2991 uint8_t end_pos = 0;
2992 for (i = 5; i < midx; ++i) {
2993 if (mode_threshold[mode_map[i - 1]] > mode_threshold[mode_map[i]]) {
2994 uint8_t tmp = mode_map[i];
2995 mode_map[i] = mode_map[i - 1];
2996 mode_map[i - 1] = tmp;
3003 for (midx = 0; midx < MAX_MODES; ++midx) {
3004 int mode_index = mode_map[midx];
3005 int mode_excluded = 0;
3006 int64_t this_rd = INT64_MAX;
3007 int disable_skip = 0;
3008 int compmode_cost = 0;
3009 int rate2 = 0, rate_y = 0, rate_uv = 0;
3010 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3012 int64_t tx_cache[TX_MODES];
3014 int64_t total_sse = INT64_MAX;
3017 this_mode = vp9_mode_order[mode_index].mode;
3018 ref_frame = vp9_mode_order[mode_index].ref_frame[0];
3019 second_ref_frame = vp9_mode_order[mode_index].ref_frame[1];
3021 // Look at the reference frame of the best mode so far and set the
3022 // skip mask to look at a subset of the remaining modes.
3023 if (midx == mode_skip_start && best_mode_index >= 0) {
3024 switch (best_mbmode.ref_frame[0]) {
3028 ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK;
3029 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3032 ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
3033 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3036 ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK;
3039 case MAX_REF_FRAMES:
3040 assert(0 && "Invalid Reference frame");
3045 if (ref_frame_skip_mask[0] & (1 << ref_frame) &&
3046 ref_frame_skip_mask[1] & (1 << MAX(0, second_ref_frame)))
3049 if (mode_skip_mask[ref_frame] & (1 << this_mode))
3052 // Test best rd so far against threshold for trying this mode.
3053 if (best_mode_skippable && cpi->sf.schedule_mode_search)
3054 mode_threshold[mode_index] <<= 1;
3056 if (best_rd < mode_threshold[mode_index])
3059 if (cpi->sf.motion_field_mode_search) {
3060 const int mi_width = MIN(num_8x8_blocks_wide_lookup[bsize],
3061 tile->mi_col_end - mi_col);
3062 const int mi_height = MIN(num_8x8_blocks_high_lookup[bsize],
3063 tile->mi_row_end - mi_row);
3064 const int bsl = mi_width_log2(bsize);
3065 int cb_partition_search_ctrl = (((mi_row + mi_col) >> bsl)
3066 + get_chessboard_index(cm->current_video_frame)) & 0x1;
3067 MB_MODE_INFO *ref_mbmi;
3068 int const_motion = 1;
3069 int skip_ref_frame = !cb_partition_search_ctrl;
3070 MV_REFERENCE_FRAME rf = NONE;
3072 ref_mv.as_int = INVALID_MV;
3074 if ((mi_row - 1) >= tile->mi_row_start) {
3075 ref_mv = xd->mi[-xd->mi_stride].src_mi->mbmi.mv[0];
3076 rf = xd->mi[-xd->mi_stride].src_mi->mbmi.ref_frame[0];
3077 for (i = 0; i < mi_width; ++i) {
3078 ref_mbmi = &xd->mi[-xd->mi_stride + i].src_mi->mbmi;
3079 const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
3080 (ref_frame == ref_mbmi->ref_frame[0]);
3081 skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
3085 if ((mi_col - 1) >= tile->mi_col_start) {
3086 if (ref_mv.as_int == INVALID_MV)
3087 ref_mv = xd->mi[-1].src_mi->mbmi.mv[0];
3089 rf = xd->mi[-1].src_mi->mbmi.ref_frame[0];
3090 for (i = 0; i < mi_height; ++i) {
3091 ref_mbmi = &xd->mi[i * xd->mi_stride - 1].src_mi->mbmi;
3092 const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
3093 (ref_frame == ref_mbmi->ref_frame[0]);
3094 skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
3098 if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV)
3099 if (rf > INTRA_FRAME)
3100 if (ref_frame != rf)
3104 if (this_mode == NEARMV || this_mode == ZEROMV)
3108 comp_pred = second_ref_frame > INTRA_FRAME;
3110 if (!cm->allow_comp_inter_inter)
3113 // Skip compound inter modes if ARF is not available.
3114 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
3117 // Do not allow compound prediction if the segment level reference frame
3118 // feature is in use as in this case there can only be one reference.
3119 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
3122 if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3123 best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
3126 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3128 if (ref_frame != INTRA_FRAME)
3129 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3132 if (ref_frame == INTRA_FRAME) {
3133 if (cpi->sf.adaptive_mode_search)
3134 if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
3137 if (this_mode != DC_PRED) {
3138 // Disable intra modes other than DC_PRED for blocks with low variance
3139 // Threshold for intra skipping based on source variance
3140 // TODO(debargha): Specialize the threshold for super block sizes
3141 const unsigned int skip_intra_var_thresh = 64;
3142 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
3143 x->source_variance < skip_intra_var_thresh)
3145 // Only search the oblique modes if the best so far is
3146 // one of the neighboring directional modes
3147 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
3148 (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
3149 if (best_mode_index >= 0 &&
3150 best_mbmode.ref_frame[0] > INTRA_FRAME)
3153 if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
3154 if (conditional_skipintra(this_mode, best_intra_mode))
3159 const MV_REFERENCE_FRAME ref_frames[2] = {ref_frame, second_ref_frame};
3160 if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv,
3161 this_mode, ref_frames))
3165 mbmi->mode = this_mode;
3166 mbmi->uv_mode = DC_PRED;
3167 mbmi->ref_frame[0] = ref_frame;
3168 mbmi->ref_frame[1] = second_ref_frame;
3169 // Evaluate all sub-pel filters irrespective of whether we can use
3170 // them for this frame.
3171 mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
3172 : cm->interp_filter;
3173 mbmi->mv[0].as_int = mbmi->mv[1].as_int = 0;
3176 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3178 // Select prediction reference frames.
3179 for (i = 0; i < MAX_MB_PLANE; i++) {
3180 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3182 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3185 for (i = 0; i < TX_MODES; ++i)
3186 tx_cache[i] = INT64_MAX;
3188 if (ref_frame == INTRA_FRAME) {
3190 super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable,
3191 NULL, bsize, tx_cache, best_rd);
3193 if (rate_y == INT_MAX)
3196 uv_tx = get_uv_tx_size_impl(mbmi->tx_size, bsize, pd[1].subsampling_x,
3197 pd[1].subsampling_y);
3198 if (rate_uv_intra[uv_tx] == INT_MAX) {
3199 choose_intra_uv_mode(cpi, ctx, bsize, uv_tx,
3200 &rate_uv_intra[uv_tx], &rate_uv_tokenonly[uv_tx],
3201 &dist_uv[uv_tx], &skip_uv[uv_tx], &mode_uv[uv_tx]);
3204 rate_uv = rate_uv_tokenonly[uv_tx];
3205 distortion_uv = dist_uv[uv_tx];
3206 skippable = skippable && skip_uv[uv_tx];
3207 mbmi->uv_mode = mode_uv[uv_tx];
3209 rate2 = rate_y + cpi->mbmode_cost[mbmi->mode] + rate_uv_intra[uv_tx];
3210 if (this_mode != DC_PRED && this_mode != TM_PRED)
3211 rate2 += intra_cost_penalty;
3212 distortion2 = distortion_y + distortion_uv;
3214 this_rd = handle_inter_mode(cpi, x, bsize,
3216 &rate2, &distortion2, &skippable,
3217 &rate_y, &distortion_y,
3218 &rate_uv, &distortion_uv,
3219 &disable_skip, frame_mv,
3221 single_newmv, single_inter_filter,
3222 single_skippable, &total_sse, best_rd);
3223 if (this_rd == INT64_MAX)
3226 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
3228 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3229 rate2 += compmode_cost;
3232 // Estimate the reference frame signaling cost and add it
3233 // to the rolling cost variable.
3235 rate2 += ref_costs_comp[ref_frame];
3237 rate2 += ref_costs_single[ref_frame];
3240 if (!disable_skip) {
3242 // Back out the coefficient coding costs
3243 rate2 -= (rate_y + rate_uv);
3245 // Cost the skip mb case
3246 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
3247 } else if (ref_frame != INTRA_FRAME && !xd->lossless) {
3248 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
3249 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
3250 // Add in the cost of the no skip flag.
3251 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
3253 // FIXME(rbultje) make this work for splitmv also
3254 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
3255 distortion2 = total_sse;
3256 assert(total_sse >= 0);
3257 rate2 -= (rate_y + rate_uv);
3261 // Add in the cost of the no skip flag.
3262 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
3265 // Calculate the final RD estimate for this mode.
3266 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3269 if (ref_frame == INTRA_FRAME) {
3270 // Keep record of best intra rd
3271 if (this_rd < best_intra_rd) {
3272 best_intra_rd = this_rd;
3273 best_intra_mode = mbmi->mode;
3277 if (!disable_skip && ref_frame == INTRA_FRAME) {
3278 for (i = 0; i < REFERENCE_MODES; ++i)
3279 best_pred_rd[i] = MIN(best_pred_rd[i], this_rd);
3280 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3281 best_filter_rd[i] = MIN(best_filter_rd[i], this_rd);
3284 // Did this mode help.. i.e. is it the new best mode
3285 if (this_rd < best_rd || x->skip) {
3286 int max_plane = MAX_MB_PLANE;
3287 if (!mode_excluded) {
3288 // Note index of best mode so far
3289 best_mode_index = mode_index;
3291 if (ref_frame == INTRA_FRAME) {
3292 /* required for left and above block mv */
3293 mbmi->mv[0].as_int = 0;
3296 best_pred_sse = x->pred_sse[ref_frame];
3299 *returnrate = rate2;
3300 *returndistortion = distortion2;
3302 best_mbmode = *mbmi;
3303 best_skip2 = this_skip2;
3304 best_mode_skippable = skippable;
3306 if (!x->select_tx_size)
3307 swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
3308 vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size],
3309 sizeof(uint8_t) * ctx->num_4x4_blk);
3311 // TODO(debargha): enhance this test with a better distortion prediction
3312 // based on qp, activity mask and history
3313 if ((mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
3314 (mode_index > MIN_EARLY_TERM_INDEX)) {
3315 int qstep = xd->plane[0].dequant[1];
3316 // TODO(debargha): Enhance this by specializing for each mode_index
3318 #if CONFIG_VP9_HIGHBITDEPTH
3319 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
3320 qstep >>= (xd->bd - 8);
3322 #endif // CONFIG_VP9_HIGHBITDEPTH
3323 if (x->source_variance < UINT_MAX) {
3324 const int var_adjust = (x->source_variance < 16);
3325 scale -= var_adjust;
3327 if (ref_frame > INTRA_FRAME &&
3328 distortion2 * scale < qstep * qstep) {
3335 /* keep record of best compound/single-only prediction */
3336 if (!disable_skip && ref_frame != INTRA_FRAME) {
3337 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
3339 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3340 single_rate = rate2 - compmode_cost;
3341 hybrid_rate = rate2;
3343 single_rate = rate2;
3344 hybrid_rate = rate2 + compmode_cost;
3347 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
3348 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
3351 if (single_rd < best_pred_rd[SINGLE_REFERENCE])
3352 best_pred_rd[SINGLE_REFERENCE] = single_rd;
3354 if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
3355 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
3357 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
3358 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
3360 /* keep record of best filter type */
3361 if (!mode_excluded && cm->interp_filter != BILINEAR) {
3362 int64_t ref = rd_opt->filter_cache[cm->interp_filter == SWITCHABLE ?
3363 SWITCHABLE_FILTERS : cm->interp_filter];
3365 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3367 if (ref == INT64_MAX)
3369 else if (rd_opt->filter_cache[i] == INT64_MAX)
3370 // when early termination is triggered, the encoder does not have
3371 // access to the rate-distortion cost. it only knows that the cost
3372 // should be above the maximum valid value. hence it takes the known
3373 // maximum plus an arbitrary constant as the rate-distortion cost.
3374 adj_rd = rd_opt->mask_filter - ref + 10;
3376 adj_rd = rd_opt->filter_cache[i] - ref;
3379 best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd);
3384 /* keep record of best txfm size */
3385 if (bsize < BLOCK_32X32) {
3386 if (bsize < BLOCK_16X16)
3387 tx_cache[ALLOW_16X16] = tx_cache[ALLOW_8X8];
3389 tx_cache[ALLOW_32X32] = tx_cache[ALLOW_16X16];
3391 if (!mode_excluded && this_rd != INT64_MAX) {
3392 for (i = 0; i < TX_MODES && tx_cache[i] < INT64_MAX; i++) {
3393 int64_t adj_rd = INT64_MAX;
3394 adj_rd = this_rd + tx_cache[i] - tx_cache[cm->tx_mode];
3396 if (adj_rd < best_tx_rd[i])
3397 best_tx_rd[i] = adj_rd;
3404 if (x->skip && !comp_pred)
3408 // The inter modes' rate costs are not calculated precisely in some cases.
3409 // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and
3410 // ZEROMV. Here, checks are added for those cases, and the mode decisions
3412 if (best_mbmode.mode == NEWMV) {
3413 const MV_REFERENCE_FRAME refs[2] = {best_mbmode.ref_frame[0],
3414 best_mbmode.ref_frame[1]};
3415 int comp_pred_mode = refs[1] > INTRA_FRAME;
3417 if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3418 ((comp_pred_mode && frame_mv[NEARESTMV][refs[1]].as_int ==
3419 best_mbmode.mv[1].as_int) || !comp_pred_mode))
3420 best_mbmode.mode = NEARESTMV;
3421 else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3422 ((comp_pred_mode && frame_mv[NEARMV][refs[1]].as_int ==
3423 best_mbmode.mv[1].as_int) || !comp_pred_mode))
3424 best_mbmode.mode = NEARMV;
3425 else if (best_mbmode.mv[0].as_int == 0 &&
3426 ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) || !comp_pred_mode))
3427 best_mbmode.mode = ZEROMV;
3430 if (best_mode_index < 0 || best_rd >= best_rd_so_far)
3433 // If we used an estimate for the uv intra rd in the loop above...
3434 if (cpi->sf.use_uv_intra_rd_estimate) {
3435 // Do Intra UV best rd mode selection if best mode choice above was intra.
3436 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
3438 *mbmi = best_mbmode;
3439 uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
3440 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
3441 &rate_uv_tokenonly[uv_tx_size],
3442 &dist_uv[uv_tx_size],
3443 &skip_uv[uv_tx_size],
3444 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
3449 assert((cm->interp_filter == SWITCHABLE) ||
3450 (cm->interp_filter == best_mbmode.interp_filter) ||
3451 !is_inter_block(&best_mbmode));
3453 if (!cpi->rc.is_src_frame_alt_ref)
3454 update_rd_thresh_fact(cpi, bsize, best_mode_index);
3457 *mbmi = best_mbmode;
3458 x->skip |= best_skip2;
3460 for (i = 0; i < REFERENCE_MODES; ++i) {
3461 if (best_pred_rd[i] == INT64_MAX)
3462 best_pred_diff[i] = INT_MIN;
3464 best_pred_diff[i] = best_rd - best_pred_rd[i];
3468 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3469 if (best_filter_rd[i] == INT64_MAX)
3470 best_filter_diff[i] = 0;
3472 best_filter_diff[i] = best_rd - best_filter_rd[i];
3474 if (cm->interp_filter == SWITCHABLE)
3475 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
3476 for (i = 0; i < TX_MODES; i++) {
3477 if (best_tx_rd[i] == INT64_MAX)
3478 best_tx_diff[i] = 0;
3480 best_tx_diff[i] = best_rd - best_tx_rd[i];
3483 vp9_zero(best_filter_diff);
3484 vp9_zero(best_tx_diff);
3487 // TODO(yunqingwang): Moving this line in front of the above best_filter_diff
3488 // updating code causes PSNR loss. Need to figure out the confliction.
3489 x->skip |= best_mode_skippable;
3491 store_coding_context(x, ctx, best_mode_index, best_pred_diff,
3492 best_tx_diff, best_filter_diff, best_mode_skippable);
3497 int64_t vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi, MACROBLOCK *x,
3499 int64_t *returndistortion,
3501 PICK_MODE_CONTEXT *ctx,
3502 int64_t best_rd_so_far) {
3503 VP9_COMMON *const cm = &cpi->common;
3504 RD_OPT *const rd_opt = &cpi->rd;
3505 MACROBLOCKD *const xd = &x->e_mbd;
3506 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
3507 unsigned char segment_id = mbmi->segment_id;
3508 const int comp_pred = 0;
3510 int64_t best_tx_diff[TX_MODES];
3511 int64_t best_pred_diff[REFERENCE_MODES];
3512 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3513 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3514 vp9_prob comp_mode_p;
3515 INTERP_FILTER best_filter = SWITCHABLE;
3516 int64_t this_rd = INT64_MAX;
3518 const int64_t distortion2 = 0;
3520 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3522 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3525 for (i = 0; i < MAX_REF_FRAMES; ++i)
3526 x->pred_sse[i] = INT_MAX;
3527 for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i)
3528 x->pred_mv_sad[i] = INT_MAX;
3530 *returnrate = INT_MAX;
3532 assert(vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
3534 mbmi->mode = ZEROMV;
3535 mbmi->uv_mode = DC_PRED;
3536 mbmi->ref_frame[0] = LAST_FRAME;
3537 mbmi->ref_frame[1] = NONE;
3538 mbmi->mv[0].as_int = 0;
3541 // Search for best switchable filter by checking the variance of
3542 // pred error irrespective of whether the filter will be used
3543 rd_opt->mask_filter = 0;
3544 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3545 rd_opt->filter_cache[i] = INT64_MAX;
3547 if (cm->interp_filter != BILINEAR) {
3548 best_filter = EIGHTTAP;
3549 if (cm->interp_filter == SWITCHABLE &&
3550 x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
3552 int best_rs = INT_MAX;
3553 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
3554 mbmi->interp_filter = i;
3555 rs = vp9_get_switchable_rate(cpi);
3558 best_filter = mbmi->interp_filter;
3563 // Set the appropriate filter
3564 if (cm->interp_filter == SWITCHABLE) {
3565 mbmi->interp_filter = best_filter;
3566 rate2 += vp9_get_switchable_rate(cpi);
3568 mbmi->interp_filter = cm->interp_filter;
3571 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3572 rate2 += vp9_cost_bit(comp_mode_p, comp_pred);
3574 // Estimate the reference frame signaling cost and add it
3575 // to the rolling cost variable.
3576 rate2 += ref_costs_single[LAST_FRAME];
3577 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3579 *returnrate = rate2;
3580 *returndistortion = distortion2;
3582 if (this_rd >= best_rd_so_far)
3585 assert((cm->interp_filter == SWITCHABLE) ||
3586 (cm->interp_filter == mbmi->interp_filter));
3588 update_rd_thresh_fact(cpi, bsize, THR_ZEROMV);
3590 vp9_zero(best_pred_diff);
3591 vp9_zero(best_filter_diff);
3592 vp9_zero(best_tx_diff);
3594 if (!x->select_tx_size)
3595 swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE);
3596 store_coding_context(x, ctx, THR_ZEROMV,
3597 best_pred_diff, best_tx_diff, best_filter_diff, 0);
3602 int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
3603 const TileInfo *const tile,
3604 int mi_row, int mi_col,
3606 int64_t *returndistortion,
3608 PICK_MODE_CONTEXT *ctx,
3609 int64_t best_rd_so_far) {
3610 VP9_COMMON *const cm = &cpi->common;
3611 RD_OPT *const rd_opt = &cpi->rd;
3612 MACROBLOCKD *const xd = &x->e_mbd;
3613 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
3614 const struct segmentation *const seg = &cm->seg;
3615 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3616 unsigned char segment_id = mbmi->segment_id;
3618 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3619 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3620 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
3622 int64_t best_rd = best_rd_so_far;
3623 int64_t best_yrd = best_rd_so_far; // FIXME(rbultje) more precise
3624 static const int64_t best_tx_diff[TX_MODES] = { 0 };
3625 int64_t best_pred_diff[REFERENCE_MODES];
3626 int64_t best_pred_rd[REFERENCE_MODES];
3627 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3628 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3629 MB_MODE_INFO best_mbmode;
3630 int ref_index, best_ref_index = 0;
3631 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3632 vp9_prob comp_mode_p;
3633 INTERP_FILTER tmp_best_filter = SWITCHABLE;
3634 int rate_uv_intra, rate_uv_tokenonly;
3637 PREDICTION_MODE mode_uv = DC_PRED;
3638 const int intra_cost_penalty = vp9_get_intra_cost_penalty(
3639 cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
3640 int_mv seg_mvs[4][MAX_REF_FRAMES];
3641 b_mode_info best_bmodes[4];
3643 int ref_frame_skip_mask[2] = { 0 };
3645 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3646 vpx_memset(x->zcoeff_blk[TX_4X4], 0, 4);
3647 vp9_zero(best_mbmode);
3649 for (i = 0; i < 4; i++) {
3651 for (j = 0; j < MAX_REF_FRAMES; j++)
3652 seg_mvs[i][j].as_int = INVALID_MV;
3655 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3658 for (i = 0; i < REFERENCE_MODES; ++i)
3659 best_pred_rd[i] = INT64_MAX;
3660 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3661 best_filter_rd[i] = INT64_MAX;
3662 rate_uv_intra = INT_MAX;
3664 *returnrate = INT_MAX;
3666 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
3667 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3668 setup_buffer_inter(cpi, x, tile,
3669 ref_frame, bsize, mi_row, mi_col,
3670 frame_mv[NEARESTMV], frame_mv[NEARMV],
3673 ref_frame_skip_mask[0] |= (1 << ref_frame);
3674 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3676 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3677 frame_mv[ZEROMV][ref_frame].as_int = 0;
3680 for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
3681 int mode_excluded = 0;
3682 int64_t this_rd = INT64_MAX;
3683 int disable_skip = 0;
3684 int compmode_cost = 0;
3685 int rate2 = 0, rate_y = 0, rate_uv = 0;
3686 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3690 int64_t total_sse = INT_MAX;
3693 ref_frame = vp9_ref_order[ref_index].ref_frame[0];
3694 second_ref_frame = vp9_ref_order[ref_index].ref_frame[1];
3696 // Look at the reference frame of the best mode so far and set the
3697 // skip mask to look at a subset of the remaining modes.
3698 if (ref_index > 2 && cpi->sf.mode_skip_start < MAX_MODES) {
3699 if (ref_index == 3) {
3700 switch (best_mbmode.ref_frame[0]) {
3704 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME);
3705 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3708 ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME);
3709 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3712 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME);
3715 case MAX_REF_FRAMES:
3716 assert(0 && "Invalid Reference frame");
3722 if (ref_frame_skip_mask[0] & (1 << ref_frame) &&
3723 ref_frame_skip_mask[1] & (1 << MAX(0, second_ref_frame)))
3726 // Test best rd so far against threshold for trying this mode.
3727 if (rd_less_than_thresh(best_rd,
3728 rd_opt->threshes[segment_id][bsize][ref_index],
3729 rd_opt->thresh_freq_fact[bsize][ref_index]))
3732 comp_pred = second_ref_frame > INTRA_FRAME;
3734 if (!cm->allow_comp_inter_inter)
3736 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
3738 // Do not allow compound prediction if the segment level reference frame
3739 // feature is in use as in this case there can only be one reference.
3740 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
3743 if ((cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3744 best_mbmode.ref_frame[0] == INTRA_FRAME)
3748 // TODO(jingning, jkoleszar): scaling reference frame not supported for
3750 if (ref_frame > INTRA_FRAME &&
3751 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
3754 if (second_ref_frame > INTRA_FRAME &&
3755 vp9_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf))
3759 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3760 else if (ref_frame != INTRA_FRAME)
3761 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3763 // If the segment reference frame feature is enabled....
3764 // then do nothing if the current ref frame is not allowed..
3765 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3766 vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3768 // Disable this drop out case if the ref frame
3769 // segment level feature is enabled for this segment. This is to
3770 // prevent the possibility that we end up unable to pick any mode.
3771 } else if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3772 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3773 // unless ARNR filtering is enabled in which case we want
3774 // an unfiltered alternative. We allow near/nearest as well
3775 // because they may result in zero-zero MVs but be cheaper.
3776 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
3780 mbmi->tx_size = TX_4X4;
3781 mbmi->uv_mode = DC_PRED;
3782 mbmi->ref_frame[0] = ref_frame;
3783 mbmi->ref_frame[1] = second_ref_frame;
3784 // Evaluate all sub-pel filters irrespective of whether we can use
3785 // them for this frame.
3786 mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
3787 : cm->interp_filter;
3789 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3791 // Select prediction reference frames.
3792 for (i = 0; i < MAX_MB_PLANE; i++) {
3793 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3795 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3798 if (ref_frame == INTRA_FRAME) {
3800 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y,
3801 &distortion_y, best_rd) >= best_rd)
3804 rate2 += intra_cost_penalty;
3805 distortion2 += distortion_y;
3807 if (rate_uv_intra == INT_MAX) {
3808 choose_intra_uv_mode(cpi, ctx, bsize, TX_4X4,
3814 rate2 += rate_uv_intra;
3815 rate_uv = rate_uv_tokenonly;
3816 distortion2 += dist_uv;
3817 distortion_uv = dist_uv;
3818 mbmi->uv_mode = mode_uv;
3822 int64_t this_rd_thresh;
3823 int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
3824 int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
3825 int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
3826 int tmp_best_skippable = 0;
3827 int switchable_filter_index;
3828 int_mv *second_ref = comp_pred ?
3829 &mbmi->ref_mvs[second_ref_frame][0] : NULL;
3830 b_mode_info tmp_best_bmodes[16];
3831 MB_MODE_INFO tmp_best_mbmode;
3832 BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
3833 int pred_exists = 0;
3836 this_rd_thresh = (ref_frame == LAST_FRAME) ?
3837 rd_opt->threshes[segment_id][bsize][THR_LAST] :
3838 rd_opt->threshes[segment_id][bsize][THR_ALTR];
3839 this_rd_thresh = (ref_frame == GOLDEN_FRAME) ?
3840 rd_opt->threshes[segment_id][bsize][THR_GOLD] : this_rd_thresh;
3841 rd_opt->mask_filter = 0;
3842 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3843 rd_opt->filter_cache[i] = INT64_MAX;
3845 if (cm->interp_filter != BILINEAR) {
3846 tmp_best_filter = EIGHTTAP;
3847 if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
3848 tmp_best_filter = EIGHTTAP;
3849 } else if (cpi->sf.adaptive_pred_interp_filter == 1 &&
3850 ctx->pred_interp_filter < SWITCHABLE) {
3851 tmp_best_filter = ctx->pred_interp_filter;
3852 } else if (cpi->sf.adaptive_pred_interp_filter == 2) {
3853 tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE ?
3854 ctx->pred_interp_filter : 0;
3856 for (switchable_filter_index = 0;
3857 switchable_filter_index < SWITCHABLE_FILTERS;
3858 ++switchable_filter_index) {
3861 mbmi->interp_filter = switchable_filter_index;
3862 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x, tile,
3863 &mbmi->ref_mvs[ref_frame][0],
3864 second_ref, best_yrd, &rate,
3865 &rate_y, &distortion,
3866 &skippable, &total_sse,
3867 (int) this_rd_thresh, seg_mvs,
3868 bsi, switchable_filter_index,
3871 if (tmp_rd == INT64_MAX)
3873 rs = vp9_get_switchable_rate(cpi);
3874 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
3875 rd_opt->filter_cache[switchable_filter_index] = tmp_rd;
3876 rd_opt->filter_cache[SWITCHABLE_FILTERS] =
3877 MIN(rd_opt->filter_cache[SWITCHABLE_FILTERS],
3879 if (cm->interp_filter == SWITCHABLE)
3882 rd_opt->mask_filter = MAX(rd_opt->mask_filter, tmp_rd);
3884 newbest = (tmp_rd < tmp_best_rd);
3886 tmp_best_filter = mbmi->interp_filter;
3887 tmp_best_rd = tmp_rd;
3889 if ((newbest && cm->interp_filter == SWITCHABLE) ||
3890 (mbmi->interp_filter == cm->interp_filter &&
3891 cm->interp_filter != SWITCHABLE)) {
3892 tmp_best_rdu = tmp_rd;
3893 tmp_best_rate = rate;
3894 tmp_best_ratey = rate_y;
3895 tmp_best_distortion = distortion;
3896 tmp_best_sse = total_sse;
3897 tmp_best_skippable = skippable;
3898 tmp_best_mbmode = *mbmi;
3899 for (i = 0; i < 4; i++) {
3900 tmp_best_bmodes[i] = xd->mi[0].src_mi->bmi[i];
3901 x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
3904 if (switchable_filter_index == 0 &&
3905 cpi->sf.use_rd_breakout &&
3906 best_rd < INT64_MAX) {
3907 if (tmp_best_rdu / 2 > best_rd) {
3908 // skip searching the other filters if the first is
3909 // already substantially larger than the best so far
3910 tmp_best_filter = mbmi->interp_filter;
3911 tmp_best_rdu = INT64_MAX;
3916 } // switchable_filter_index loop
3920 if (tmp_best_rdu == INT64_MAX && pred_exists)
3923 mbmi->interp_filter = (cm->interp_filter == SWITCHABLE ?
3924 tmp_best_filter : cm->interp_filter);
3926 // Handles the special case when a filter that is not in the
3927 // switchable list (bilinear, 6-tap) is indicated at the frame level
3928 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x, tile,
3929 &mbmi->ref_mvs[ref_frame][0],
3930 second_ref, best_yrd, &rate, &rate_y,
3931 &distortion, &skippable, &total_sse,
3932 (int) this_rd_thresh, seg_mvs, bsi, 0,
3934 if (tmp_rd == INT64_MAX)
3937 total_sse = tmp_best_sse;
3938 rate = tmp_best_rate;
3939 rate_y = tmp_best_ratey;
3940 distortion = tmp_best_distortion;
3941 skippable = tmp_best_skippable;
3942 *mbmi = tmp_best_mbmode;
3943 for (i = 0; i < 4; i++)
3944 xd->mi[0].src_mi->bmi[i] = tmp_best_bmodes[i];
3948 distortion2 += distortion;
3950 if (cm->interp_filter == SWITCHABLE)
3951 rate2 += vp9_get_switchable_rate(cpi);
3954 mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
3955 : cm->reference_mode == COMPOUND_REFERENCE;
3957 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
3959 tmp_best_rdu = best_rd -
3960 MIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
3961 RDCOST(x->rdmult, x->rddiv, 0, total_sse));
3963 if (tmp_best_rdu > 0) {
3964 // If even the 'Y' rd value of split is higher than best so far
3965 // then dont bother looking at UV
3966 vp9_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col,
3968 super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
3969 &uv_sse, BLOCK_8X8, tmp_best_rdu);
3970 if (rate_uv == INT_MAX)
3973 distortion2 += distortion_uv;
3974 skippable = skippable && uv_skippable;
3975 total_sse += uv_sse;
3979 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3980 rate2 += compmode_cost;
3982 // Estimate the reference frame signaling cost and add it
3983 // to the rolling cost variable.
3984 if (second_ref_frame > INTRA_FRAME) {
3985 rate2 += ref_costs_comp[ref_frame];
3987 rate2 += ref_costs_single[ref_frame];
3990 if (!disable_skip) {
3991 // Skip is never coded at the segment level for sub8x8 blocks and instead
3992 // always coded in the bitstream at the mode info level.
3994 if (ref_frame != INTRA_FRAME && !xd->lossless) {
3995 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
3996 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
3997 // Add in the cost of the no skip flag.
3998 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
4000 // FIXME(rbultje) make this work for splitmv also
4001 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
4002 distortion2 = total_sse;
4003 assert(total_sse >= 0);
4004 rate2 -= (rate_y + rate_uv);
4010 // Add in the cost of the no skip flag.
4011 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
4014 // Calculate the final RD estimate for this mode.
4015 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
4018 if (!disable_skip && ref_frame == INTRA_FRAME) {
4019 for (i = 0; i < REFERENCE_MODES; ++i)
4020 best_pred_rd[i] = MIN(best_pred_rd[i], this_rd);
4021 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
4022 best_filter_rd[i] = MIN(best_filter_rd[i], this_rd);
4025 // Did this mode help.. i.e. is it the new best mode
4026 if (this_rd < best_rd || x->skip) {
4027 if (!mode_excluded) {
4028 int max_plane = MAX_MB_PLANE;
4029 // Note index of best mode so far
4030 best_ref_index = ref_index;
4032 if (ref_frame == INTRA_FRAME) {
4033 /* required for left and above block mv */
4034 mbmi->mv[0].as_int = 0;
4038 *returnrate = rate2;
4039 *returndistortion = distortion2;
4041 best_yrd = best_rd -
4042 RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
4043 best_mbmode = *mbmi;
4044 best_skip2 = this_skip2;
4045 if (!x->select_tx_size)
4046 swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
4047 vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
4048 sizeof(uint8_t) * ctx->num_4x4_blk);
4050 for (i = 0; i < 4; i++)
4051 best_bmodes[i] = xd->mi[0].src_mi->bmi[i];
4053 // TODO(debargha): enhance this test with a better distortion prediction
4054 // based on qp, activity mask and history
4055 if ((cpi->sf.mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
4056 (ref_index > MIN_EARLY_TERM_INDEX)) {
4057 int qstep = xd->plane[0].dequant[1];
4058 // TODO(debargha): Enhance this by specializing for each mode_index
4060 #if CONFIG_VP9_HIGHBITDEPTH
4061 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
4062 qstep >>= (xd->bd - 8);
4064 #endif // CONFIG_VP9_HIGHBITDEPTH
4065 if (x->source_variance < UINT_MAX) {
4066 const int var_adjust = (x->source_variance < 16);
4067 scale -= var_adjust;
4069 if (ref_frame > INTRA_FRAME &&
4070 distortion2 * scale < qstep * qstep) {
4077 /* keep record of best compound/single-only prediction */
4078 if (!disable_skip && ref_frame != INTRA_FRAME) {
4079 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
4081 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4082 single_rate = rate2 - compmode_cost;
4083 hybrid_rate = rate2;
4085 single_rate = rate2;
4086 hybrid_rate = rate2 + compmode_cost;
4089 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
4090 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
4092 if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
4093 best_pred_rd[SINGLE_REFERENCE] = single_rd;
4094 else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
4095 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
4097 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
4098 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
4101 /* keep record of best filter type */
4102 if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
4103 cm->interp_filter != BILINEAR) {
4104 int64_t ref = rd_opt->filter_cache[cm->interp_filter == SWITCHABLE ?
4105 SWITCHABLE_FILTERS : cm->interp_filter];
4107 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4108 if (ref == INT64_MAX)
4110 else if (rd_opt->filter_cache[i] == INT64_MAX)
4111 // when early termination is triggered, the encoder does not have
4112 // access to the rate-distortion cost. it only knows that the cost
4113 // should be above the maximum valid value. hence it takes the known
4114 // maximum plus an arbitrary constant as the rate-distortion cost.
4115 adj_rd = rd_opt->mask_filter - ref + 10;
4117 adj_rd = rd_opt->filter_cache[i] - ref;
4120 best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd);
4127 if (x->skip && !comp_pred)
4131 if (best_rd >= best_rd_so_far)
4134 // If we used an estimate for the uv intra rd in the loop above...
4135 if (cpi->sf.use_uv_intra_rd_estimate) {
4136 // Do Intra UV best rd mode selection if best mode choice above was intra.
4137 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
4138 *mbmi = best_mbmode;
4139 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra,
4147 if (best_rd == INT64_MAX) {
4148 *returnrate = INT_MAX;
4149 *returndistortion = INT64_MAX;
4153 assert((cm->interp_filter == SWITCHABLE) ||
4154 (cm->interp_filter == best_mbmode.interp_filter) ||
4155 !is_inter_block(&best_mbmode));
4157 update_rd_thresh_fact(cpi, bsize, best_ref_index);
4160 *mbmi = best_mbmode;
4161 x->skip |= best_skip2;
4162 if (!is_inter_block(&best_mbmode)) {
4163 for (i = 0; i < 4; i++)
4164 xd->mi[0].src_mi->bmi[i].as_mode = best_bmodes[i].as_mode;
4166 for (i = 0; i < 4; ++i)
4167 vpx_memcpy(&xd->mi[0].src_mi->bmi[i], &best_bmodes[i],
4168 sizeof(b_mode_info));
4170 mbmi->mv[0].as_int = xd->mi[0].src_mi->bmi[3].as_mv[0].as_int;
4171 mbmi->mv[1].as_int = xd->mi[0].src_mi->bmi[3].as_mv[1].as_int;
4174 for (i = 0; i < REFERENCE_MODES; ++i) {
4175 if (best_pred_rd[i] == INT64_MAX)
4176 best_pred_diff[i] = INT_MIN;
4178 best_pred_diff[i] = best_rd - best_pred_rd[i];
4182 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4183 if (best_filter_rd[i] == INT64_MAX)
4184 best_filter_diff[i] = 0;
4186 best_filter_diff[i] = best_rd - best_filter_rd[i];
4188 if (cm->interp_filter == SWITCHABLE)
4189 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
4191 vp9_zero(best_filter_diff);
4194 store_coding_context(x, ctx, best_ref_index,
4195 best_pred_diff, best_tx_diff, best_filter_diff, 0);