2 * Copyright (c) 2014 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.
16 #include "./vp9_rtcd.h"
17 #include "./vpx_dsp_rtcd.h"
19 #include "vpx_mem/vpx_mem.h"
20 #include "vpx_ports/mem.h"
22 #include "vp9/common/vp9_blockd.h"
23 #include "vp9/common/vp9_common.h"
24 #include "vp9/common/vp9_mvref_common.h"
25 #include "vp9/common/vp9_pred_common.h"
26 #include "vp9/common/vp9_reconinter.h"
27 #include "vp9/common/vp9_reconintra.h"
28 #include "vp9/common/vp9_scan.h"
30 #include "vp9/encoder/vp9_cost.h"
31 #include "vp9/encoder/vp9_encoder.h"
32 #include "vp9/encoder/vp9_pickmode.h"
33 #include "vp9/encoder/vp9_ratectrl.h"
34 #include "vp9/encoder/vp9_rd.h"
42 static int mv_refs_rt(const VP9_COMMON *cm, const MACROBLOCKD *xd,
43 const TileInfo *const tile,
44 MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
46 int mi_row, int mi_col) {
47 const int *ref_sign_bias = cm->ref_frame_sign_bias;
48 int i, refmv_count = 0;
50 const POSITION *const mv_ref_search = mv_ref_blocks[mi->mbmi.sb_type];
52 int different_ref_found = 0;
53 int context_counter = 0;
56 // Blank the reference vector list
57 memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
59 // The nearest 2 blocks are treated differently
60 // if the size < 8x8 we get the mv from the bmi substructure,
61 // and we also need to keep a mode count.
62 for (i = 0; i < 2; ++i) {
63 const POSITION *const mv_ref = &mv_ref_search[i];
64 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
65 const MODE_INFO *const candidate_mi = xd->mi[mv_ref->col + mv_ref->row *
67 const MB_MODE_INFO *const candidate = &candidate_mi->mbmi;
68 // Keep counts for entropy encoding.
69 context_counter += mode_2_counter[candidate->mode];
70 different_ref_found = 1;
72 if (candidate->ref_frame[0] == ref_frame)
73 ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1),
74 refmv_count, mv_ref_list, Done);
80 // Check the rest of the neighbors in much the same way
81 // as before except we don't need to keep track of sub blocks or
83 for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
84 const POSITION *const mv_ref = &mv_ref_search[i];
85 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
86 const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row *
88 different_ref_found = 1;
90 if (candidate->ref_frame[0] == ref_frame)
91 ADD_MV_REF_LIST(candidate->mv[0], refmv_count, mv_ref_list, Done);
95 // Since we couldn't find 2 mvs from the same reference frame
96 // go back through the neighbors and find motion vectors from
97 // different reference frames.
98 if (different_ref_found && !refmv_count) {
99 for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
100 const POSITION *mv_ref = &mv_ref_search[i];
101 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
102 const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row
103 * xd->mi_stride]->mbmi;
105 // If the candidate is INTRA we don't want to consider its mv.
106 IF_DIFF_REF_FRAME_ADD_MV(candidate, ref_frame, ref_sign_bias,
107 refmv_count, mv_ref_list, Done);
114 mi->mbmi.mode_context[ref_frame] = counter_to_context[context_counter];
117 for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
118 clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
123 static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
124 BLOCK_SIZE bsize, int mi_row, int mi_col,
125 int_mv *tmp_mv, int *rate_mv,
126 int64_t best_rd_sofar) {
127 MACROBLOCKD *xd = &x->e_mbd;
128 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
129 struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
130 const int step_param = cpi->sf.mv.fullpel_search_step_param;
131 const int sadpb = x->sadperbit16;
133 const int ref = mbmi->ref_frame[0];
134 const MV ref_mv = mbmi->ref_mvs[ref][0].as_mv;
137 const int tmp_col_min = x->mv_col_min;
138 const int tmp_col_max = x->mv_col_max;
139 const int tmp_row_min = x->mv_row_min;
140 const int tmp_row_max = x->mv_row_max;
143 const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi,
145 if (scaled_ref_frame) {
147 // Swap out the reference frame for a version that's been scaled to
148 // match the resolution of the current frame, allowing the existing
149 // motion search code to be used without additional modifications.
150 for (i = 0; i < MAX_MB_PLANE; i++)
151 backup_yv12[i] = xd->plane[i].pre[0];
152 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
154 vp9_set_mv_search_range(x, &ref_mv);
156 assert(x->mv_best_ref_index[ref] <= 2);
157 if (x->mv_best_ref_index[ref] < 2)
158 mvp_full = mbmi->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
160 mvp_full = x->pred_mv[ref];
165 vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
166 cond_cost_list(cpi, cost_list),
167 &ref_mv, &tmp_mv->as_mv, INT_MAX, 0);
169 x->mv_col_min = tmp_col_min;
170 x->mv_col_max = tmp_col_max;
171 x->mv_row_min = tmp_row_min;
172 x->mv_row_max = tmp_row_max;
174 // calculate the bit cost on motion vector
175 mvp_full.row = tmp_mv->as_mv.row * 8;
176 mvp_full.col = tmp_mv->as_mv.col * 8;
178 *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv,
179 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
181 rate_mode = cpi->inter_mode_cost[mbmi->mode_context[ref]]
182 [INTER_OFFSET(NEWMV)];
183 rv = !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) >
187 cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
188 cpi->common.allow_high_precision_mv,
191 cpi->sf.mv.subpel_force_stop,
192 cpi->sf.mv.subpel_iters_per_step,
193 cond_cost_list(cpi, cost_list),
194 x->nmvjointcost, x->mvcost,
195 &dis, &x->pred_sse[ref], NULL, 0, 0);
196 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
197 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
200 if (scaled_ref_frame) {
202 for (i = 0; i < MAX_MB_PLANE; i++)
203 xd->plane[i].pre[0] = backup_yv12[i];
208 static void block_variance(const uint8_t *src, int src_stride,
209 const uint8_t *ref, int ref_stride,
210 int w, int h, unsigned int *sse, int *sum,
211 int block_size, unsigned int *sse8x8,
212 int *sum8x8, unsigned int *var8x8) {
218 for (i = 0; i < h; i += block_size) {
219 for (j = 0; j < w; j += block_size) {
220 vpx_get8x8var(src + src_stride * i + j, src_stride,
221 ref + ref_stride * i + j, ref_stride,
222 &sse8x8[k], &sum8x8[k]);
225 var8x8[k] = sse8x8[k] - (((unsigned int)sum8x8[k] * sum8x8[k]) >> 6);
231 static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
232 unsigned int *sse_i, int *sum_i,
233 unsigned int *var_o, unsigned int *sse_o,
235 const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
236 const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
237 const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
240 for (i = 0; i < nh; i += 2) {
241 for (j = 0; j < nw; j += 2) {
242 sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
243 sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
244 sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
245 sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
246 var_o[k] = sse_o[k] - (((unsigned int)sum_o[k] * sum_o[k]) >>
247 (b_width_log2_lookup[unit_size] +
248 b_height_log2_lookup[unit_size] + 6));
254 static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize,
255 MACROBLOCK *x, MACROBLOCKD *xd,
256 int *out_rate_sum, int64_t *out_dist_sum,
257 unsigned int *var_y, unsigned int *sse_y,
258 int mi_row, int mi_col, int *early_term) {
259 // Note our transform coeffs are 8 times an orthogonal transform.
260 // Hence quantizer step is also 8 times. To get effective quantizer
261 // we need to divide by 8 before sending to modeling function.
265 struct macroblock_plane *const p = &x->plane[0];
266 struct macroblockd_plane *const pd = &xd->plane[0];
267 const uint32_t dc_quant = pd->dequant[0];
268 const uint32_t ac_quant = pd->dequant[1];
269 const int64_t dc_thr = dc_quant * dc_quant >> 6;
270 const int64_t ac_thr = ac_quant * ac_quant >> 6;
275 const int bw = b_width_log2_lookup[bsize];
276 const int bh = b_height_log2_lookup[bsize];
277 const int num8x8 = 1 << (bw + bh - 2);
278 unsigned int sse8x8[64] = {0};
279 int sum8x8[64] = {0};
280 unsigned int var8x8[64] = {0};
284 // Calculate variance for whole partition, and also save 8x8 blocks' variance
285 // to be used in following transform skipping test.
286 block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
287 4 << bw, 4 << bh, &sse, &sum, 8, sse8x8, sum8x8, var8x8);
288 var = sse - (((int64_t)sum * sum) >> (bw + bh + 4));
293 if (cpi->common.tx_mode == TX_MODE_SELECT) {
294 if (sse > (var << 2))
295 tx_size = MIN(max_txsize_lookup[bsize],
296 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
300 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
301 cyclic_refresh_segment_id_boosted(xd->mi[0]->mbmi.segment_id))
303 else if (tx_size > TX_16X16)
306 tx_size = MIN(max_txsize_lookup[bsize],
307 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
310 assert(tx_size >= TX_8X8);
311 xd->mi[0]->mbmi.tx_size = tx_size;
313 // Evaluate if the partition block is a skippable block in Y plane.
315 unsigned int sse16x16[16] = {0};
316 int sum16x16[16] = {0};
317 unsigned int var16x16[16] = {0};
318 const int num16x16 = num8x8 >> 2;
320 unsigned int sse32x32[4] = {0};
321 int sum32x32[4] = {0};
322 unsigned int var32x32[4] = {0};
323 const int num32x32 = num8x8 >> 4;
327 const int num = (tx_size == TX_8X8) ? num8x8 :
328 ((tx_size == TX_16X16) ? num16x16 : num32x32);
329 const unsigned int *sse_tx = (tx_size == TX_8X8) ? sse8x8 :
330 ((tx_size == TX_16X16) ? sse16x16 : sse32x32);
331 const unsigned int *var_tx = (tx_size == TX_8X8) ? var8x8 :
332 ((tx_size == TX_16X16) ? var16x16 : var32x32);
334 // Calculate variance if tx_size > TX_8X8
335 if (tx_size >= TX_16X16)
336 calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
338 if (tx_size == TX_32X32)
339 calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32,
344 for (k = 0; k < num; k++)
345 // Check if all ac coefficients can be quantized to zero.
346 if (!(var_tx[k] < ac_thr || var == 0)) {
351 for (k = 0; k < num; k++)
352 // Check if dc coefficient can be quantized to zero.
353 if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
363 } else if (dc_test) {
368 if (x->skip_txfm[0] == 1) {
369 int skip_uv[2] = {0};
370 unsigned int var_uv[2];
371 unsigned int sse_uv[2];
374 *out_dist_sum = sse << 4;
376 // Transform skipping test in UV planes.
377 for (i = 1; i <= 2; i++) {
378 struct macroblock_plane *const p = &x->plane[i];
379 struct macroblockd_plane *const pd = &xd->plane[i];
380 const TX_SIZE uv_tx_size = get_uv_tx_size(&xd->mi[0]->mbmi, pd);
381 const BLOCK_SIZE unit_size = txsize_to_bsize[uv_tx_size];
382 const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, pd);
383 const int uv_bw = b_width_log2_lookup[uv_bsize];
384 const int uv_bh = b_height_log2_lookup[uv_bsize];
385 const int sf = (uv_bw - b_width_log2_lookup[unit_size]) +
386 (uv_bh - b_height_log2_lookup[unit_size]);
387 const uint32_t uv_dc_thr = pd->dequant[0] * pd->dequant[0] >> (6 - sf);
388 const uint32_t uv_ac_thr = pd->dequant[1] * pd->dequant[1] >> (6 - sf);
391 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, i);
392 var_uv[j] = cpi->fn_ptr[uv_bsize].vf(p->src.buf, p->src.stride,
393 pd->dst.buf, pd->dst.stride, &sse_uv[j]);
395 if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
396 (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
402 // If the transform in YUV planes are skippable, the mode search checks
403 // fewer inter modes and doesn't check intra modes.
404 if (skip_uv[0] & skip_uv[1]) {
412 #if CONFIG_VP9_HIGHBITDEPTH
413 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
414 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
415 dc_quant >> (xd->bd - 5), &rate, &dist);
417 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
418 dc_quant >> 3, &rate, &dist);
421 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
422 dc_quant >> 3, &rate, &dist);
423 #endif // CONFIG_VP9_HIGHBITDEPTH
427 *out_rate_sum = rate >> 1;
428 *out_dist_sum = dist << 3;
431 *out_dist_sum = (sse - var) << 4;
434 #if CONFIG_VP9_HIGHBITDEPTH
435 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
436 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
437 ac_quant >> (xd->bd - 5), &rate, &dist);
439 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
440 ac_quant >> 3, &rate, &dist);
443 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
444 ac_quant >> 3, &rate, &dist);
445 #endif // CONFIG_VP9_HIGHBITDEPTH
447 *out_rate_sum += rate;
448 *out_dist_sum += dist << 4;
451 static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize,
452 MACROBLOCK *x, MACROBLOCKD *xd,
453 int *out_rate_sum, int64_t *out_dist_sum,
454 unsigned int *var_y, unsigned int *sse_y) {
455 // Note our transform coeffs are 8 times an orthogonal transform.
456 // Hence quantizer step is also 8 times. To get effective quantizer
457 // we need to divide by 8 before sending to modeling function.
461 struct macroblock_plane *const p = &x->plane[0];
462 struct macroblockd_plane *const pd = &xd->plane[0];
463 const int64_t dc_thr = p->quant_thred[0] >> 6;
464 const int64_t ac_thr = p->quant_thred[1] >> 6;
465 const uint32_t dc_quant = pd->dequant[0];
466 const uint32_t ac_quant = pd->dequant[1];
467 unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
468 pd->dst.buf, pd->dst.stride, &sse);
474 if (cpi->common.tx_mode == TX_MODE_SELECT) {
475 if (sse > (var << 2))
476 xd->mi[0]->mbmi.tx_size =
477 MIN(max_txsize_lookup[bsize],
478 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
480 xd->mi[0]->mbmi.tx_size = TX_8X8;
482 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
483 cyclic_refresh_segment_id_boosted(xd->mi[0]->mbmi.segment_id))
484 xd->mi[0]->mbmi.tx_size = TX_8X8;
485 else if (xd->mi[0]->mbmi.tx_size > TX_16X16)
486 xd->mi[0]->mbmi.tx_size = TX_16X16;
488 xd->mi[0]->mbmi.tx_size =
489 MIN(max_txsize_lookup[bsize],
490 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
493 // Evaluate if the partition block is a skippable block in Y plane.
495 const BLOCK_SIZE unit_size =
496 txsize_to_bsize[xd->mi[0]->mbmi.tx_size];
497 const unsigned int num_blk_log2 =
498 (b_width_log2_lookup[bsize] - b_width_log2_lookup[unit_size]) +
499 (b_height_log2_lookup[bsize] - b_height_log2_lookup[unit_size]);
500 const unsigned int sse_tx = sse >> num_blk_log2;
501 const unsigned int var_tx = var >> num_blk_log2;
504 // Check if all ac coefficients can be quantized to zero.
505 if (var_tx < ac_thr || var == 0) {
507 // Check if dc coefficient can be quantized to zero.
508 if (sse_tx - var_tx < dc_thr || sse == var)
511 if (sse_tx - var_tx < dc_thr || sse == var)
516 if (x->skip_txfm[0] == 1) {
518 *out_dist_sum = sse << 4;
523 #if CONFIG_VP9_HIGHBITDEPTH
524 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
525 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
526 dc_quant >> (xd->bd - 5), &rate, &dist);
528 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
529 dc_quant >> 3, &rate, &dist);
532 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
533 dc_quant >> 3, &rate, &dist);
534 #endif // CONFIG_VP9_HIGHBITDEPTH
538 *out_rate_sum = rate >> 1;
539 *out_dist_sum = dist << 3;
542 *out_dist_sum = (sse - var) << 4;
545 #if CONFIG_VP9_HIGHBITDEPTH
546 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
547 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
548 ac_quant >> (xd->bd - 5), &rate, &dist);
550 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
551 ac_quant >> 3, &rate, &dist);
554 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
555 ac_quant >> 3, &rate, &dist);
556 #endif // CONFIG_VP9_HIGHBITDEPTH
558 *out_rate_sum += rate;
559 *out_dist_sum += dist << 4;
562 #if CONFIG_VP9_HIGHBITDEPTH
563 static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate, int64_t *dist,
564 int *skippable, int64_t *sse, int plane,
565 BLOCK_SIZE bsize, TX_SIZE tx_size) {
566 MACROBLOCKD *xd = &x->e_mbd;
567 unsigned int var_y, sse_y;
570 model_rd_for_sb_y(cpi, bsize, x, xd, rate, dist, &var_y, &sse_y);
576 static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate, int64_t *dist,
577 int *skippable, int64_t *sse, int plane,
578 BLOCK_SIZE bsize, TX_SIZE tx_size) {
579 MACROBLOCKD *xd = &x->e_mbd;
580 const struct macroblockd_plane *pd = &xd->plane[plane];
581 const struct macroblock_plane *const p = &x->plane[plane];
582 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
583 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
584 const int step = 1 << (tx_size << 1);
585 const int block_step = (1 << tx_size);
587 int shift = tx_size == TX_32X32 ? 0 : 2;
588 const int max_blocks_wide = num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 :
589 xd->mb_to_right_edge >> (5 + pd->subsampling_x));
590 const int max_blocks_high = num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 :
591 xd->mb_to_bottom_edge >> (5 + pd->subsampling_y));
595 vp9_subtract_plane(x, bsize, plane);
597 // Keep track of the row and column of the blocks we use so that we know
598 // if we are in the unrestricted motion border.
599 for (r = 0; r < max_blocks_high; r += block_step) {
600 for (c = 0; c < num_4x4_w; c += block_step) {
601 if (c < max_blocks_wide) {
602 const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
603 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
604 tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
605 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
606 uint16_t *const eob = &p->eobs[block];
607 const int diff_stride = 4 * num_4x4_blocks_wide_lookup[bsize];
608 const int16_t *src_diff;
609 src_diff = &p->src_diff[(r * diff_stride + c) << 2];
613 vp9_fdct32x32_rd(src_diff, coeff, diff_stride);
614 vp9_quantize_fp_32x32(coeff, 1024, x->skip_block, p->zbin,
615 p->round_fp, p->quant_fp, p->quant_shift,
616 qcoeff, dqcoeff, pd->dequant, eob,
617 scan_order->scan, scan_order->iscan);
620 vp9_hadamard_16x16(src_diff, diff_stride, (int16_t *)coeff);
621 vp9_quantize_fp(coeff, 256, x->skip_block, p->zbin, p->round_fp,
622 p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
624 scan_order->scan, scan_order->iscan);
627 vp9_hadamard_8x8(src_diff, diff_stride, (int16_t *)coeff);
628 vp9_quantize_fp(coeff, 64, x->skip_block, p->zbin, p->round_fp,
629 p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
631 scan_order->scan, scan_order->iscan);
634 x->fwd_txm4x4(src_diff, coeff, diff_stride);
635 vp9_quantize_fp(coeff, 16, x->skip_block, p->zbin, p->round_fp,
636 p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
638 scan_order->scan, scan_order->iscan);
644 *skippable &= (*eob == 0);
651 if (*skippable && *sse < INT64_MAX) {
653 *dist = (*sse << 6) >> shift;
661 *sse = (*sse << 6) >> shift;
662 for (r = 0; r < max_blocks_high; r += block_step) {
663 for (c = 0; c < num_4x4_w; c += block_step) {
664 if (c < max_blocks_wide) {
665 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
666 tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
667 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
668 uint16_t *const eob = &p->eobs[block];
671 *rate += (int)abs(qcoeff[0]);
673 *rate += (int)vp9_satd((const int16_t *)qcoeff, step << 4);
675 *dist += vp9_block_error_fp(coeff, dqcoeff, step << 4) >> shift;
681 if (*skippable == 0) {
683 *rate += (eob_cost << 8);
688 static void model_rd_for_sb_uv(VP9_COMP *cpi, BLOCK_SIZE bsize,
689 MACROBLOCK *x, MACROBLOCKD *xd,
690 int *out_rate_sum, int64_t *out_dist_sum,
691 unsigned int *var_y, unsigned int *sse_y) {
692 // Note our transform coeffs are 8 times an orthogonal transform.
693 // Hence quantizer step is also 8 times. To get effective quantizer
694 // we need to divide by 8 before sending to modeling function.
703 for (i = 1; i <= 2; ++i) {
704 struct macroblock_plane *const p = &x->plane[i];
705 struct macroblockd_plane *const pd = &xd->plane[i];
706 const uint32_t dc_quant = pd->dequant[0];
707 const uint32_t ac_quant = pd->dequant[1];
708 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
711 if (!x->color_sensitivity[i - 1])
714 var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride,
715 pd->dst.buf, pd->dst.stride, &sse);
719 #if CONFIG_VP9_HIGHBITDEPTH
720 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
721 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
722 dc_quant >> (xd->bd - 5), &rate, &dist);
724 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
725 dc_quant >> 3, &rate, &dist);
728 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
729 dc_quant >> 3, &rate, &dist);
730 #endif // CONFIG_VP9_HIGHBITDEPTH
732 *out_rate_sum += rate >> 1;
733 *out_dist_sum += dist << 3;
735 #if CONFIG_VP9_HIGHBITDEPTH
736 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
737 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
738 ac_quant >> (xd->bd - 5), &rate, &dist);
740 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
741 ac_quant >> 3, &rate, &dist);
744 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
745 ac_quant >> 3, &rate, &dist);
746 #endif // CONFIG_VP9_HIGHBITDEPTH
748 *out_rate_sum += rate;
749 *out_dist_sum += dist << 4;
753 static int get_pred_buffer(PRED_BUFFER *p, int len) {
756 for (i = 0; i < len; i++) {
765 static void free_pred_buffer(PRED_BUFFER *p) {
770 static void encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x,
771 BLOCK_SIZE bsize, int mi_row, int mi_col,
772 MV_REFERENCE_FRAME ref_frame,
773 PREDICTION_MODE this_mode,
774 unsigned int var_y, unsigned int sse_y,
775 struct buf_2d yv12_mb[][MAX_MB_PLANE],
776 int *rate, int64_t *dist) {
777 MACROBLOCKD *xd = &x->e_mbd;
778 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
780 const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
781 unsigned int var = var_y, sse = sse_y;
782 // Skipping threshold for ac.
783 unsigned int thresh_ac;
784 // Skipping threshold for dc.
785 unsigned int thresh_dc;
786 if (x->encode_breakout > 0) {
787 // Set a maximum for threshold to avoid big PSNR loss in low bit rate
788 // case. Use extreme low threshold for static frames to limit
790 const unsigned int max_thresh = 36000;
791 // The encode_breakout input
792 const unsigned int min_thresh =
793 MIN(((unsigned int)x->encode_breakout << 4), max_thresh);
794 #if CONFIG_VP9_HIGHBITDEPTH
795 const int shift = (xd->bd << 1) - 16;
798 // Calculate threshold according to dequant value.
799 thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) >> 3;
800 #if CONFIG_VP9_HIGHBITDEPTH
801 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
802 thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift);
804 #endif // CONFIG_VP9_HIGHBITDEPTH
805 thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
807 // Adjust ac threshold according to partition size.
809 8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
811 thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
812 #if CONFIG_VP9_HIGHBITDEPTH
813 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
814 thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift);
816 #endif // CONFIG_VP9_HIGHBITDEPTH
822 // Y skipping condition checking for ac and dc.
823 if (var <= thresh_ac && (sse - var) <= thresh_dc) {
824 unsigned int sse_u, sse_v;
825 unsigned int var_u, var_v;
827 // Skip UV prediction unless breakout is zero (lossless) to save
828 // computation with low impact on the result
829 if (x->encode_breakout == 0) {
830 xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
831 xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
832 vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
835 var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
836 x->plane[1].src.stride,
837 xd->plane[1].dst.buf,
838 xd->plane[1].dst.stride, &sse_u);
840 // U skipping condition checking
841 if (((var_u << 2) <= thresh_ac) && (sse_u - var_u <= thresh_dc)) {
842 var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
843 x->plane[2].src.stride,
844 xd->plane[2].dst.buf,
845 xd->plane[2].dst.stride, &sse_v);
847 // V skipping condition checking
848 if (((var_v << 2) <= thresh_ac) && (sse_v - var_v <= thresh_dc)) {
851 // The cost of skip bit needs to be added.
852 *rate = cpi->inter_mode_cost[mbmi->mode_context[ref_frame]]
853 [INTER_OFFSET(this_mode)];
855 // More on this part of rate
856 // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
858 // Scaling factor for SSE from spatial domain to frequency
859 // domain is 16. Adjust distortion accordingly.
860 // TODO(yunqingwang): In this function, only y-plane dist is
862 *dist = (sse << 4); // + ((sse_u + sse_v) << 4);
864 // *disable_skip = 1;
870 struct estimate_block_intra_args {
873 PREDICTION_MODE mode;
878 static void estimate_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
879 TX_SIZE tx_size, void *arg) {
880 struct estimate_block_intra_args* const args = arg;
881 VP9_COMP *const cpi = args->cpi;
882 MACROBLOCK *const x = args->x;
883 MACROBLOCKD *const xd = &x->e_mbd;
884 struct macroblock_plane *const p = &x->plane[0];
885 struct macroblockd_plane *const pd = &xd->plane[0];
886 const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
887 uint8_t *const src_buf_base = p->src.buf;
888 uint8_t *const dst_buf_base = pd->dst.buf;
889 const int src_stride = p->src.stride;
890 const int dst_stride = pd->dst.stride;
894 int64_t this_sse = INT64_MAX;
897 txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
901 p->src.buf = &src_buf_base[4 * (j * src_stride + i)];
902 pd->dst.buf = &dst_buf_base[4 * (j * dst_stride + i)];
903 // Use source buffer as an approximation for the fully reconstructed buffer.
904 vp9_predict_intra_block(xd, block >> (2 * tx_size),
905 b_width_log2_lookup[plane_bsize],
907 x->skip_encode ? p->src.buf : pd->dst.buf,
908 x->skip_encode ? src_stride : dst_stride,
909 pd->dst.buf, dst_stride,
912 // TODO(jingning): This needs further refactoring.
913 block_yrd(cpi, x, &rate, &dist, &is_skippable, &this_sse, 0,
914 bsize_tx, MIN(tx_size, TX_16X16));
915 x->skip_txfm[0] = is_skippable;
916 rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), is_skippable);
918 p->src.buf = src_buf_base;
919 pd->dst.buf = dst_buf_base;
924 static const THR_MODES mode_idx[MAX_REF_FRAMES - 1][4] = {
925 {THR_DC, THR_V_PRED, THR_H_PRED, THR_TM},
926 {THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV},
927 {THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG},
930 static const PREDICTION_MODE intra_mode_list[] = {
931 DC_PRED, V_PRED, H_PRED, TM_PRED
934 static int mode_offset(const PREDICTION_MODE mode) {
935 if (mode >= NEARESTMV) {
936 return INTER_OFFSET(mode);
953 static INLINE void update_thresh_freq_fact(VP9_COMP *cpi,
954 TileDataEnc *tile_data,
956 MV_REFERENCE_FRAME ref_frame,
957 THR_MODES best_mode_idx,
958 PREDICTION_MODE mode) {
959 THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
960 int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
961 if (thr_mode_idx == best_mode_idx)
962 *freq_fact -= (*freq_fact >> 4);
964 *freq_fact = MIN(*freq_fact + RD_THRESH_INC,
965 cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
968 void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
969 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
970 MACROBLOCKD *const xd = &x->e_mbd;
971 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
972 RD_COST this_rdc, best_rdc;
973 PREDICTION_MODE this_mode;
974 struct estimate_block_intra_args args = { cpi, x, DC_PRED, 0, 0 };
975 const TX_SIZE intra_tx_size =
976 MIN(max_txsize_lookup[bsize],
977 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
978 MODE_INFO *const mic = xd->mi[0];
980 const MODE_INFO *above_mi = xd->mi[-xd->mi_stride];
981 const MODE_INFO *left_mi = xd->left_available ? xd->mi[-1] : NULL;
982 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
983 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
984 bmode_costs = cpi->y_mode_costs[A][L];
987 vp9_rd_cost_reset(&best_rdc);
988 vp9_rd_cost_reset(&this_rdc);
990 mbmi->ref_frame[0] = INTRA_FRAME;
991 mbmi->mv[0].as_int = INVALID_MV;
992 mbmi->uv_mode = DC_PRED;
993 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
995 // Change the limit of this loop to add other intra prediction
997 for (this_mode = DC_PRED; this_mode <= H_PRED; ++this_mode) {
998 args.mode = this_mode;
1001 mbmi->tx_size = intra_tx_size;
1002 vp9_foreach_transformed_block_in_plane(xd, bsize, 0,
1003 estimate_block_intra, &args);
1004 this_rdc.rate = args.rate;
1005 this_rdc.dist = args.dist;
1006 this_rdc.rate += bmode_costs[this_mode];
1007 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
1008 this_rdc.rate, this_rdc.dist);
1010 if (this_rdc.rdcost < best_rdc.rdcost) {
1011 best_rdc = this_rdc;
1012 mbmi->mode = this_mode;
1016 *rd_cost = best_rdc;
1019 static void init_ref_frame_cost(VP9_COMMON *const cm,
1020 MACROBLOCKD *const xd,
1021 int ref_frame_cost[MAX_REF_FRAMES]) {
1022 vp9_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
1023 vp9_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1024 vp9_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1026 ref_frame_cost[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1027 ref_frame_cost[LAST_FRAME] = ref_frame_cost[GOLDEN_FRAME] =
1028 ref_frame_cost[ALTREF_FRAME] = vp9_cost_bit(intra_inter_p, 1);
1030 ref_frame_cost[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1031 ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1032 ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1033 ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1034 ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1038 MV_REFERENCE_FRAME ref_frame;
1039 PREDICTION_MODE pred_mode;
1042 #define RT_INTER_MODES 8
1043 static const REF_MODE ref_mode_set[RT_INTER_MODES] = {
1044 {LAST_FRAME, ZEROMV},
1045 {LAST_FRAME, NEARESTMV},
1046 {GOLDEN_FRAME, ZEROMV},
1047 {LAST_FRAME, NEARMV},
1048 {LAST_FRAME, NEWMV},
1049 {GOLDEN_FRAME, NEARESTMV},
1050 {GOLDEN_FRAME, NEARMV},
1051 {GOLDEN_FRAME, NEWMV}
1054 // TODO(jingning) placeholder for inter-frame non-RD mode decision.
1055 // this needs various further optimizations. to be continued..
1056 void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
1057 TileDataEnc *tile_data,
1058 int mi_row, int mi_col, RD_COST *rd_cost,
1059 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1060 VP9_COMMON *const cm = &cpi->common;
1061 TileInfo *const tile_info = &tile_data->tile_info;
1062 MACROBLOCKD *const xd = &x->e_mbd;
1063 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1064 struct macroblockd_plane *const pd = &xd->plane[0];
1065 PREDICTION_MODE best_mode = ZEROMV;
1066 MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
1067 MV_REFERENCE_FRAME usable_ref_frame;
1068 TX_SIZE best_tx_size = TX_SIZES;
1069 INTERP_FILTER best_pred_filter = EIGHTTAP;
1070 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1071 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1072 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1074 RD_COST this_rdc, best_rdc;
1075 uint8_t skip_txfm = 0, best_mode_skip_txfm = 0;
1076 // var_y and sse_y are saved to be used in skipping checking
1077 unsigned int var_y = UINT_MAX;
1078 unsigned int sse_y = UINT_MAX;
1079 // Reduce the intra cost penalty for small blocks (<=16x16).
1080 const int reduction_fac = (bsize <= BLOCK_16X16) ?
1081 ((bsize <= BLOCK_8X8) ? 4 : 2) : 0;
1082 const int intra_cost_penalty = vp9_get_intra_cost_penalty(
1083 cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth) >> reduction_fac;
1084 const int64_t inter_mode_thresh = RDCOST(x->rdmult, x->rddiv,
1085 intra_cost_penalty, 0);
1086 const int *const rd_threshes = cpi->rd.threshes[mbmi->segment_id][bsize];
1087 const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
1088 INTERP_FILTER filter_ref;
1089 const int bsl = mi_width_log2_lookup[bsize];
1090 const int pred_filter_search = cm->interp_filter == SWITCHABLE ?
1091 (((mi_row + mi_col) >> bsl) +
1092 get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
1093 int const_motion[MAX_REF_FRAMES] = { 0 };
1094 const int bh = num_4x4_blocks_high_lookup[bsize] << 2;
1095 const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
1096 // For speed 6, the result of interp filter is reused later in actual encoding
1098 // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
1100 DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 64 * 64]);
1101 #if CONFIG_VP9_HIGHBITDEPTH
1102 DECLARE_ALIGNED(16, uint16_t, pred_buf_16[3 * 64 * 64]);
1104 struct buf_2d orig_dst = pd->dst;
1105 PRED_BUFFER *best_pred = NULL;
1106 PRED_BUFFER *this_mode_pred = NULL;
1107 const int pixels_in_block = bh * bw;
1108 int reuse_inter_pred = cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready;
1109 int ref_frame_skip_mask = 0;
1111 int best_pred_sad = INT_MAX;
1112 int best_early_term = 0;
1113 int ref_frame_cost[MAX_REF_FRAMES];
1115 init_ref_frame_cost(cm, xd, ref_frame_cost);
1117 if (reuse_inter_pred) {
1119 for (i = 0; i < 3; i++) {
1120 #if CONFIG_VP9_HIGHBITDEPTH
1121 if (cm->use_highbitdepth)
1122 tmp[i].data = CONVERT_TO_BYTEPTR(&pred_buf_16[pixels_in_block * i]);
1124 tmp[i].data = &pred_buf[pixels_in_block * i];
1126 tmp[i].data = &pred_buf[pixels_in_block * i];
1127 #endif // CONFIG_VP9_HIGHBITDEPTH
1131 tmp[3].data = pd->dst.buf;
1132 tmp[3].stride = pd->dst.stride;
1136 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1139 if (xd->up_available)
1140 filter_ref = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
1141 else if (xd->left_available)
1142 filter_ref = xd->mi[-1]->mbmi.interp_filter;
1144 filter_ref = cm->interp_filter;
1146 // initialize mode decisions
1147 vp9_rd_cost_reset(&best_rdc);
1148 vp9_rd_cost_reset(rd_cost);
1149 mbmi->sb_type = bsize;
1150 mbmi->ref_frame[0] = NONE;
1151 mbmi->ref_frame[1] = NONE;
1152 mbmi->tx_size = MIN(max_txsize_lookup[bsize],
1153 tx_mode_to_biggest_tx_size[cm->tx_mode]);
1155 #if CONFIG_VP9_TEMPORAL_DENOISING
1156 vp9_denoiser_reset_frame_stats(ctx);
1159 if (cpi->rc.frames_since_golden == 0) {
1160 usable_ref_frame = LAST_FRAME;
1162 usable_ref_frame = GOLDEN_FRAME;
1165 for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {
1166 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1168 x->pred_mv_sad[ref_frame] = INT_MAX;
1169 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
1170 frame_mv[ZEROMV][ref_frame].as_int = 0;
1172 if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
1173 int_mv *const candidates = mbmi->ref_mvs[ref_frame];
1174 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
1176 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col,
1179 if (cm->use_prev_frame_mvs)
1180 vp9_find_mv_refs(cm, xd, tile_info, xd->mi[0], ref_frame,
1181 candidates, mi_row, mi_col, NULL, NULL);
1183 const_motion[ref_frame] = mv_refs_rt(cm, xd, tile_info,
1185 ref_frame, candidates,
1188 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1189 &frame_mv[NEARESTMV][ref_frame],
1190 &frame_mv[NEARMV][ref_frame]);
1192 if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8)
1193 vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
1196 ref_frame_skip_mask |= (1 << ref_frame);
1200 for (idx = 0; idx < RT_INTER_MODES; ++idx) {
1205 PREDICTION_MODE this_mode = ref_mode_set[idx].pred_mode;
1208 int this_early_term = 0;
1210 if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode)))
1213 ref_frame = ref_mode_set[idx].ref_frame;
1214 if (!(cpi->ref_frame_flags & flag_list[ref_frame]))
1216 if (const_motion[ref_frame] && this_mode == NEARMV)
1219 i = (ref_frame == LAST_FRAME) ? GOLDEN_FRAME : LAST_FRAME;
1220 if (cpi->ref_frame_flags & flag_list[i])
1221 if (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[i] << 1))
1222 ref_frame_skip_mask |= (1 << ref_frame);
1223 if (ref_frame_skip_mask & (1 << ref_frame))
1226 // Select prediction reference frames.
1227 for (i = 0; i < MAX_MB_PLANE; i++)
1228 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
1230 mbmi->ref_frame[0] = ref_frame;
1231 set_ref_ptrs(cm, xd, ref_frame, NONE);
1233 mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
1234 mode_rd_thresh = best_mode_skip_txfm ?
1235 rd_threshes[mode_index] << 1 : rd_threshes[mode_index];
1236 if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
1237 rd_thresh_freq_fact[mode_index]))
1240 if (this_mode == NEWMV) {
1241 if (ref_frame > LAST_FRAME) {
1243 int dis, cost_list[5];
1245 if (bsize < BLOCK_16X16)
1248 tmp_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
1250 if (tmp_sad > x->pred_mv_sad[LAST_FRAME])
1252 if (tmp_sad + (num_pels_log2_lookup[bsize] << 4) > best_pred_sad)
1255 frame_mv[NEWMV][ref_frame].as_int = mbmi->mv[0].as_int;
1256 rate_mv = vp9_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv,
1257 &mbmi->ref_mvs[ref_frame][0].as_mv,
1258 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1259 frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
1260 frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
1262 cpi->find_fractional_mv_step(x, &frame_mv[NEWMV][ref_frame].as_mv,
1263 &mbmi->ref_mvs[ref_frame][0].as_mv,
1264 cpi->common.allow_high_precision_mv,
1266 &cpi->fn_ptr[bsize],
1267 cpi->sf.mv.subpel_force_stop,
1268 cpi->sf.mv.subpel_iters_per_step,
1269 cond_cost_list(cpi, cost_list),
1270 x->nmvjointcost, x->mvcost, &dis,
1271 &x->pred_sse[ref_frame], NULL, 0, 0);
1272 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1273 &frame_mv[NEWMV][ref_frame], &rate_mv, best_rdc.rdcost)) {
1278 if (this_mode == NEWMV && ref_frame == LAST_FRAME &&
1279 frame_mv[NEWMV][LAST_FRAME].as_int != INVALID_MV) {
1280 const int pre_stride = xd->plane[0].pre[0].stride;
1281 const uint8_t * const pre_buf = xd->plane[0].pre[0].buf +
1282 (frame_mv[NEWMV][LAST_FRAME].as_mv.row >> 3) * pre_stride +
1283 (frame_mv[NEWMV][LAST_FRAME].as_mv.col >> 3);
1284 best_pred_sad = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf,
1285 x->plane[0].src.stride,
1286 pre_buf, pre_stride);
1287 x->pred_mv_sad[LAST_FRAME] = best_pred_sad;
1290 if (this_mode != NEARESTMV &&
1291 frame_mv[this_mode][ref_frame].as_int ==
1292 frame_mv[NEARESTMV][ref_frame].as_int)
1295 mbmi->mode = this_mode;
1296 mbmi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
1298 // Search for the best prediction filter type, when the resulting
1299 // motion vector is at sub-pixel accuracy level for luma component, i.e.,
1300 // the last three bits are all zeros.
1301 if (reuse_inter_pred) {
1302 if (!this_mode_pred) {
1303 this_mode_pred = &tmp[3];
1305 this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
1306 pd->dst.buf = this_mode_pred->data;
1307 pd->dst.stride = bw;
1311 if ((this_mode == NEWMV || filter_ref == SWITCHABLE) && pred_filter_search
1312 && (ref_frame == LAST_FRAME)
1313 && (((mbmi->mv[0].as_mv.row | mbmi->mv[0].as_mv.col) & 0x07) != 0)) {
1316 unsigned int pf_var[3];
1317 unsigned int pf_sse[3];
1318 TX_SIZE pf_tx_size[3];
1319 int64_t best_cost = INT64_MAX;
1320 INTERP_FILTER best_filter = SWITCHABLE, filter;
1321 PRED_BUFFER *current_pred = this_mode_pred;
1323 for (filter = EIGHTTAP; filter <= EIGHTTAP_SMOOTH; ++filter) {
1325 mbmi->interp_filter = filter;
1326 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1327 model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter], &pf_dist[filter],
1328 &pf_var[filter], &pf_sse[filter]);
1329 pf_rate[filter] += vp9_get_switchable_rate(cpi, xd);
1330 cost = RDCOST(x->rdmult, x->rddiv, pf_rate[filter], pf_dist[filter]);
1331 pf_tx_size[filter] = mbmi->tx_size;
1332 if (cost < best_cost) {
1333 best_filter = filter;
1335 skip_txfm = x->skip_txfm[0];
1337 if (reuse_inter_pred) {
1338 if (this_mode_pred != current_pred) {
1339 free_pred_buffer(this_mode_pred);
1340 this_mode_pred = current_pred;
1343 if (filter < EIGHTTAP_SHARP) {
1344 current_pred = &tmp[get_pred_buffer(tmp, 3)];
1345 pd->dst.buf = current_pred->data;
1346 pd->dst.stride = bw;
1352 if (reuse_inter_pred && this_mode_pred != current_pred)
1353 free_pred_buffer(current_pred);
1355 mbmi->interp_filter = best_filter;
1356 mbmi->tx_size = pf_tx_size[best_filter];
1357 this_rdc.rate = pf_rate[best_filter];
1358 this_rdc.dist = pf_dist[best_filter];
1359 var_y = pf_var[best_filter];
1360 sse_y = pf_sse[best_filter];
1361 x->skip_txfm[0] = skip_txfm;
1362 if (reuse_inter_pred) {
1363 pd->dst.buf = this_mode_pred->data;
1364 pd->dst.stride = this_mode_pred->stride;
1367 mbmi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP : filter_ref;
1368 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1370 // For large partition blocks, extra testing is done.
1371 if (bsize > BLOCK_32X32 &&
1372 !cyclic_refresh_segment_id_boosted(xd->mi[0]->mbmi.segment_id) &&
1374 model_rd_for_sb_y_large(cpi, bsize, x, xd, &this_rdc.rate,
1375 &this_rdc.dist, &var_y, &sse_y, mi_row, mi_col,
1378 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
1383 if (!this_early_term) {
1384 this_sse = (int64_t)sse_y;
1385 block_yrd(cpi, x, &this_rdc.rate, &this_rdc.dist, &is_skippable,
1386 &this_sse, 0, bsize, MIN(mbmi->tx_size, TX_16X16));
1387 x->skip_txfm[0] = is_skippable;
1389 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1391 if (RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist) <
1392 RDCOST(x->rdmult, x->rddiv, 0, this_sse)) {
1393 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
1395 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1396 this_rdc.dist = this_sse;
1397 x->skip_txfm[0] = 1;
1401 if (cm->interp_filter == SWITCHABLE) {
1402 if ((mbmi->mv[0].as_mv.row | mbmi->mv[0].as_mv.col) & 0x07)
1403 this_rdc.rate += vp9_get_switchable_rate(cpi, xd);
1406 this_rdc.rate += cm->interp_filter == SWITCHABLE ?
1407 vp9_get_switchable_rate(cpi, xd) : 0;
1408 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1411 if (x->color_sensitivity[0] || x->color_sensitivity[1]) {
1413 int64_t uv_dist = 0;
1414 if (x->color_sensitivity[0])
1415 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 1);
1416 if (x->color_sensitivity[1])
1417 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 2);
1418 model_rd_for_sb_uv(cpi, bsize, x, xd, &uv_rate, &uv_dist,
1420 this_rdc.rate += uv_rate;
1421 this_rdc.dist += uv_dist;
1424 this_rdc.rate += rate_mv;
1426 cpi->inter_mode_cost[mbmi->mode_context[ref_frame]][INTER_OFFSET(
1428 this_rdc.rate += ref_frame_cost[ref_frame];
1429 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1431 // Skipping checking: test to see if this block can be reconstructed by
1433 if (cpi->allow_encode_breakout) {
1434 encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame, this_mode,
1435 var_y, sse_y, yv12_mb, &this_rdc.rate,
1438 this_rdc.rate += rate_mv;
1439 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate,
1444 #if CONFIG_VP9_TEMPORAL_DENOISING
1445 if (cpi->oxcf.noise_sensitivity > 0)
1446 vp9_denoiser_update_frame_stats(mbmi, sse_y, this_mode, ctx);
1451 if (this_rdc.rdcost < best_rdc.rdcost || x->skip) {
1452 best_rdc = this_rdc;
1453 best_mode = this_mode;
1454 best_pred_filter = mbmi->interp_filter;
1455 best_tx_size = mbmi->tx_size;
1456 best_ref_frame = ref_frame;
1457 best_mode_skip_txfm = x->skip_txfm[0];
1458 best_early_term = this_early_term;
1460 if (reuse_inter_pred) {
1461 free_pred_buffer(best_pred);
1462 best_pred = this_mode_pred;
1465 if (reuse_inter_pred)
1466 free_pred_buffer(this_mode_pred);
1472 // If early termination flag is 1 and at least 2 modes are checked,
1473 // the mode search is terminated.
1474 if (best_early_term && idx > 0) {
1480 mbmi->mode = best_mode;
1481 mbmi->interp_filter = best_pred_filter;
1482 mbmi->tx_size = best_tx_size;
1483 mbmi->ref_frame[0] = best_ref_frame;
1484 mbmi->mv[0].as_int = frame_mv[best_mode][best_ref_frame].as_int;
1485 xd->mi[0]->bmi[0].as_mv[0].as_int = mbmi->mv[0].as_int;
1486 x->skip_txfm[0] = best_mode_skip_txfm;
1488 // Perform intra prediction search, if the best SAD is above a certain
1490 if (best_rdc.rdcost == INT64_MAX ||
1491 (!x->skip && best_rdc.rdcost > inter_mode_thresh &&
1492 bsize <= cpi->sf.max_intra_bsize)) {
1493 struct estimate_block_intra_args args = { cpi, x, DC_PRED, 0, 0 };
1494 const TX_SIZE intra_tx_size =
1495 MIN(max_txsize_lookup[bsize],
1496 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
1498 TX_SIZE best_intra_tx_size = TX_SIZES;
1500 if (reuse_inter_pred && best_pred != NULL) {
1501 if (best_pred->data == orig_dst.buf) {
1502 this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
1503 #if CONFIG_VP9_HIGHBITDEPTH
1504 if (cm->use_highbitdepth)
1505 vp9_highbd_convolve_copy(best_pred->data, best_pred->stride,
1506 this_mode_pred->data, this_mode_pred->stride,
1507 NULL, 0, NULL, 0, bw, bh, xd->bd);
1509 vp9_convolve_copy(best_pred->data, best_pred->stride,
1510 this_mode_pred->data, this_mode_pred->stride,
1511 NULL, 0, NULL, 0, bw, bh);
1513 vp9_convolve_copy(best_pred->data, best_pred->stride,
1514 this_mode_pred->data, this_mode_pred->stride,
1515 NULL, 0, NULL, 0, bw, bh);
1516 #endif // CONFIG_VP9_HIGHBITDEPTH
1517 best_pred = this_mode_pred;
1522 for (i = 0; i < 4; ++i) {
1523 const PREDICTION_MODE this_mode = intra_mode_list[i];
1524 THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)];
1525 int mode_rd_thresh = rd_threshes[mode_index];
1527 if (!((1 << this_mode) & cpi->sf.intra_y_mode_bsize_mask[bsize]))
1530 if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
1531 rd_thresh_freq_fact[mode_index]))
1534 mbmi->mode = this_mode;
1535 mbmi->ref_frame[0] = INTRA_FRAME;
1536 args.mode = this_mode;
1539 mbmi->tx_size = intra_tx_size;
1540 vp9_foreach_transformed_block_in_plane(xd, bsize, 0,
1541 estimate_block_intra, &args);
1542 this_rdc.rate = args.rate;
1543 this_rdc.dist = args.dist;
1544 this_rdc.rate += cpi->mbmode_cost[this_mode];
1545 this_rdc.rate += ref_frame_cost[INTRA_FRAME];
1546 this_rdc.rate += intra_cost_penalty;
1547 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
1548 this_rdc.rate, this_rdc.dist);
1550 if (this_rdc.rdcost < best_rdc.rdcost) {
1551 best_rdc = this_rdc;
1552 best_mode = this_mode;
1553 best_intra_tx_size = mbmi->tx_size;
1554 best_ref_frame = INTRA_FRAME;
1555 mbmi->uv_mode = this_mode;
1556 mbmi->mv[0].as_int = INVALID_MV;
1557 best_mode_skip_txfm = x->skip_txfm[0];
1561 // Reset mb_mode_info to the best inter mode.
1562 if (best_ref_frame != INTRA_FRAME) {
1563 mbmi->tx_size = best_tx_size;
1565 mbmi->tx_size = best_intra_tx_size;
1570 mbmi->mode = best_mode;
1571 mbmi->ref_frame[0] = best_ref_frame;
1572 x->skip_txfm[0] = best_mode_skip_txfm;
1574 if (reuse_inter_pred && best_pred != NULL) {
1575 if (best_pred->data != orig_dst.buf && is_inter_mode(mbmi->mode)) {
1576 #if CONFIG_VP9_HIGHBITDEPTH
1577 if (cm->use_highbitdepth)
1578 vp9_highbd_convolve_copy(best_pred->data, best_pred->stride,
1579 pd->dst.buf, pd->dst.stride, NULL, 0,
1580 NULL, 0, bw, bh, xd->bd);
1582 vp9_convolve_copy(best_pred->data, best_pred->stride,
1583 pd->dst.buf, pd->dst.stride, NULL, 0,
1586 vp9_convolve_copy(best_pred->data, best_pred->stride,
1587 pd->dst.buf, pd->dst.stride, NULL, 0,
1589 #endif // CONFIG_VP9_HIGHBITDEPTH
1593 if (cpi->sf.adaptive_rd_thresh) {
1594 THR_MODES best_mode_idx = mode_idx[best_ref_frame][mode_offset(mbmi->mode)];
1596 if (best_ref_frame == INTRA_FRAME) {
1597 // Only consider the modes that are included in the intra_mode_list.
1598 int intra_modes = sizeof(intra_mode_list)/sizeof(PREDICTION_MODE);
1601 // TODO(yunqingwang): Check intra mode mask and only update freq_fact
1602 // for those valid modes.
1603 for (i = 0; i < intra_modes; i++) {
1604 update_thresh_freq_fact(cpi, tile_data, bsize, INTRA_FRAME,
1605 best_mode_idx, intra_mode_list[i]);
1608 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
1609 PREDICTION_MODE this_mode;
1610 if (best_ref_frame != ref_frame) continue;
1611 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1612 update_thresh_freq_fact(cpi, tile_data, bsize, ref_frame,
1613 best_mode_idx, this_mode);
1619 *rd_cost = best_rdc;
1622 void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
1623 TileDataEnc *tile_data,
1624 int mi_row, int mi_col, RD_COST *rd_cost,
1625 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1626 VP9_COMMON *const cm = &cpi->common;
1627 TileInfo *const tile_info = &tile_data->tile_info;
1628 SPEED_FEATURES *const sf = &cpi->sf;
1629 MACROBLOCKD *const xd = &x->e_mbd;
1630 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1631 const struct segmentation *const seg = &cm->seg;
1632 MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE;
1633 MV_REFERENCE_FRAME best_ref_frame = NONE;
1634 unsigned char segment_id = mbmi->segment_id;
1635 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1636 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1638 int64_t best_rd = INT64_MAX;
1639 b_mode_info bsi[MAX_REF_FRAMES][4];
1640 int ref_frame_skip_mask = 0;
1641 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1642 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1645 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1646 ctx->pred_pixel_ready = 0;
1648 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
1649 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1651 x->pred_mv_sad[ref_frame] = INT_MAX;
1653 if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
1654 int_mv *const candidates = mbmi->ref_mvs[ref_frame];
1655 const struct scale_factors *const sf =
1656 &cm->frame_refs[ref_frame - 1].sf;
1657 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col,
1659 vp9_find_mv_refs(cm, xd, tile_info, xd->mi[0], ref_frame,
1660 candidates, mi_row, mi_col, NULL, NULL);
1662 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1663 &dummy_mv[0], &dummy_mv[1]);
1665 ref_frame_skip_mask |= (1 << ref_frame);
1669 mbmi->sb_type = bsize;
1670 mbmi->tx_size = TX_4X4;
1671 mbmi->uv_mode = DC_PRED;
1672 mbmi->ref_frame[0] = LAST_FRAME;
1673 mbmi->ref_frame[1] = NONE;
1674 mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
1675 : cm->interp_filter;
1677 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
1678 int64_t this_rd = 0;
1681 if (ref_frame_skip_mask & (1 << ref_frame))
1684 // TODO(jingning, agrange): Scaling reference frame not supported for
1685 // sub8x8 blocks. Is this supported now?
1686 if (ref_frame > INTRA_FRAME &&
1687 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
1690 // If the segment reference frame feature is enabled....
1691 // then do nothing if the current ref frame is not allowed..
1692 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
1693 vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
1696 mbmi->ref_frame[0] = ref_frame;
1698 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
1700 // Select prediction reference frames.
1701 for (plane = 0; plane < MAX_MB_PLANE; plane++)
1702 xd->plane[plane].pre[0] = yv12_mb[ref_frame][plane];
1704 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1705 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1706 int_mv b_mv[MB_MODE_COUNT];
1707 int64_t b_best_rd = INT64_MAX;
1708 const int i = idy * 2 + idx;
1709 PREDICTION_MODE this_mode;
1711 unsigned int var_y, sse_y;
1713 struct macroblock_plane *p = &x->plane[0];
1714 struct macroblockd_plane *pd = &xd->plane[0];
1716 const struct buf_2d orig_src = p->src;
1717 const struct buf_2d orig_dst = pd->dst;
1718 struct buf_2d orig_pre[2];
1719 memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));
1721 // set buffer pointers for sub8x8 motion search.
1723 &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1725 &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
1727 &pd->pre[0].buf[vp9_raster_block_offset(BLOCK_8X8,
1728 i, pd->pre[0].stride)];
1730 b_mv[ZEROMV].as_int = 0;
1731 b_mv[NEWMV].as_int = INVALID_MV;
1732 vp9_append_sub8x8_mvs_for_idx(cm, xd, tile_info, i, 0, mi_row, mi_col,
1736 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1738 xd->mi[0]->bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
1740 if (this_mode == NEWMV) {
1741 const int step_param = cpi->sf.mv.fullpel_search_step_param;
1745 const int tmp_col_min = x->mv_col_min;
1746 const int tmp_col_max = x->mv_col_max;
1747 const int tmp_row_min = x->mv_row_min;
1748 const int tmp_row_max = x->mv_row_max;
1752 mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3;
1753 mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3;
1755 mvp_full.row = xd->mi[0]->bmi[0].as_mv[0].as_mv.row >> 3;
1756 mvp_full.col = xd->mi[0]->bmi[0].as_mv[0].as_mv.col >> 3;
1759 vp9_set_mv_search_range(x, &mbmi->ref_mvs[0]->as_mv);
1761 vp9_full_pixel_search(
1762 cpi, x, bsize, &mvp_full, step_param, x->sadperbit4,
1763 cond_cost_list(cpi, cost_list),
1764 &mbmi->ref_mvs[ref_frame][0].as_mv, &tmp_mv,
1767 x->mv_col_min = tmp_col_min;
1768 x->mv_col_max = tmp_col_max;
1769 x->mv_row_min = tmp_row_min;
1770 x->mv_row_max = tmp_row_max;
1772 // calculate the bit cost on motion vector
1773 mvp_full.row = tmp_mv.row * 8;
1774 mvp_full.col = tmp_mv.col * 8;
1776 b_rate += vp9_mv_bit_cost(&mvp_full,
1777 &mbmi->ref_mvs[ref_frame][0].as_mv,
1778 x->nmvjointcost, x->mvcost,
1781 b_rate += cpi->inter_mode_cost[mbmi->mode_context[ref_frame]]
1782 [INTER_OFFSET(NEWMV)];
1783 if (RDCOST(x->rdmult, x->rddiv, b_rate, 0) > b_best_rd)
1786 cpi->find_fractional_mv_step(x, &tmp_mv,
1787 &mbmi->ref_mvs[ref_frame][0].as_mv,
1788 cpi->common.allow_high_precision_mv,
1790 &cpi->fn_ptr[bsize],
1791 cpi->sf.mv.subpel_force_stop,
1792 cpi->sf.mv.subpel_iters_per_step,
1793 cond_cost_list(cpi, cost_list),
1794 x->nmvjointcost, x->mvcost,
1796 &x->pred_sse[ref_frame], NULL, 0, 0);
1798 xd->mi[0]->bmi[i].as_mv[0].as_mv = tmp_mv;
1800 b_rate += cpi->inter_mode_cost[mbmi->mode_context[ref_frame]]
1801 [INTER_OFFSET(this_mode)];
1804 #if CONFIG_VP9_HIGHBITDEPTH
1805 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1806 vp9_highbd_build_inter_predictor(pd->pre[0].buf, pd->pre[0].stride,
1807 pd->dst.buf, pd->dst.stride,
1808 &xd->mi[0]->bmi[i].as_mv[0].as_mv,
1809 &xd->block_refs[0]->sf,
1810 4 * num_4x4_blocks_wide,
1811 4 * num_4x4_blocks_high, 0,
1812 vp9_get_interp_kernel(mbmi->interp_filter),
1814 mi_col * MI_SIZE + 4 * (i & 0x01),
1815 mi_row * MI_SIZE + 4 * (i >> 1), xd->bd);
1818 vp9_build_inter_predictor(pd->pre[0].buf, pd->pre[0].stride,
1819 pd->dst.buf, pd->dst.stride,
1820 &xd->mi[0]->bmi[i].as_mv[0].as_mv,
1821 &xd->block_refs[0]->sf,
1822 4 * num_4x4_blocks_wide,
1823 4 * num_4x4_blocks_high, 0,
1824 vp9_get_interp_kernel(mbmi->interp_filter),
1826 mi_col * MI_SIZE + 4 * (i & 0x01),
1827 mi_row * MI_SIZE + 4 * (i >> 1));
1829 #if CONFIG_VP9_HIGHBITDEPTH
1833 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
1836 this_rdc.rate += b_rate;
1837 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
1838 this_rdc.rate, this_rdc.dist);
1839 if (this_rdc.rdcost < b_best_rd) {
1840 b_best_rd = this_rdc.rdcost;
1841 bsi[ref_frame][i].as_mode = this_mode;
1842 bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0]->bmi[i].as_mv[0].as_mv;
1846 // restore source and prediction buffer pointers.
1848 pd->pre[0] = orig_pre[0];
1850 this_rd += b_best_rd;
1852 xd->mi[0]->bmi[i] = bsi[ref_frame][i];
1853 if (num_4x4_blocks_wide > 1)
1854 xd->mi[0]->bmi[i + 1] = xd->mi[0]->bmi[i];
1855 if (num_4x4_blocks_high > 1)
1856 xd->mi[0]->bmi[i + 2] = xd->mi[0]->bmi[i];
1858 } // loop through sub8x8 blocks
1860 if (this_rd < best_rd) {
1862 best_ref_frame = ref_frame;
1864 } // reference frames
1866 mbmi->tx_size = TX_4X4;
1867 mbmi->ref_frame[0] = best_ref_frame;
1868 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1869 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1870 const int block = idy * 2 + idx;
1871 xd->mi[0]->bmi[block] = bsi[best_ref_frame][block];
1872 if (num_4x4_blocks_wide > 1)
1873 xd->mi[0]->bmi[block + 1] = bsi[best_ref_frame][block];
1874 if (num_4x4_blocks_high > 1)
1875 xd->mi[0]->bmi[block + 2] = bsi[best_ref_frame][block];
1878 mbmi->mode = xd->mi[0]->bmi[3].as_mode;
1879 ctx->mic = *(xd->mi[0]);
1880 ctx->skip_txfm[0] = 0;
1882 // Dummy assignment for speed -5. No effect in speed -6.
1883 rd_cost->rdcost = best_rd;