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/vpx_codec.h"
20 #include "vpx_dsp/vpx_dsp_common.h"
21 #include "vpx_mem/vpx_mem.h"
22 #include "vpx_ports/mem.h"
24 #include "vp9/common/vp9_blockd.h"
25 #include "vp9/common/vp9_common.h"
26 #include "vp9/common/vp9_mvref_common.h"
27 #include "vp9/common/vp9_pred_common.h"
28 #include "vp9/common/vp9_reconinter.h"
29 #include "vp9/common/vp9_reconintra.h"
30 #include "vp9/common/vp9_scan.h"
32 #include "vp9/encoder/vp9_cost.h"
33 #include "vp9/encoder/vp9_encoder.h"
34 #include "vp9/encoder/vp9_pickmode.h"
35 #include "vp9/encoder/vp9_ratectrl.h"
36 #include "vp9/encoder/vp9_rd.h"
44 static const int pos_shift_16x16[4][4] = {
45 { 9, 10, 13, 14 }, { 11, 12, 15, 16 }, { 17, 18, 21, 22 }, { 19, 20, 23, 24 }
48 static int mv_refs_rt(VP9_COMP *cpi, const VP9_COMMON *cm, const MACROBLOCK *x,
49 const MACROBLOCKD *xd, const TileInfo *const tile,
50 MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
51 int_mv *mv_ref_list, int_mv *base_mv, int mi_row,
52 int mi_col, int use_base_mv) {
53 const int *ref_sign_bias = cm->ref_frame_sign_bias;
54 int i, refmv_count = 0;
56 const POSITION *const mv_ref_search = mv_ref_blocks[mi->sb_type];
58 int different_ref_found = 0;
59 int context_counter = 0;
62 // Blank the reference vector list
63 memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
65 // The nearest 2 blocks are treated differently
66 // if the size < 8x8 we get the mv from the bmi substructure,
67 // and we also need to keep a mode count.
68 for (i = 0; i < 2; ++i) {
69 const POSITION *const mv_ref = &mv_ref_search[i];
70 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
71 const MODE_INFO *const candidate_mi =
72 xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
73 // Keep counts for entropy encoding.
74 context_counter += mode_2_counter[candidate_mi->mode];
75 different_ref_found = 1;
77 if (candidate_mi->ref_frame[0] == ref_frame)
78 ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1),
79 refmv_count, mv_ref_list, Done);
85 // Check the rest of the neighbors in much the same way
86 // as before except we don't need to keep track of sub blocks or
88 for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
89 const POSITION *const mv_ref = &mv_ref_search[i];
90 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
91 const MODE_INFO *const candidate_mi =
92 xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
93 different_ref_found = 1;
95 if (candidate_mi->ref_frame[0] == ref_frame)
96 ADD_MV_REF_LIST(candidate_mi->mv[0], refmv_count, mv_ref_list, Done);
100 // Since we couldn't find 2 mvs from the same reference frame
101 // go back through the neighbors and find motion vectors from
102 // different reference frames.
103 if (different_ref_found && !refmv_count) {
104 for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
105 const POSITION *mv_ref = &mv_ref_search[i];
106 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
107 const MODE_INFO *const candidate_mi =
108 xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
110 // If the candidate is INTRA we don't want to consider its mv.
111 IF_DIFF_REF_FRAME_ADD_MV(candidate_mi, ref_frame, ref_sign_bias,
112 refmv_count, mv_ref_list, Done);
117 !cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
118 ref_frame == LAST_FRAME) {
119 // Get base layer mv.
122 ->mvs[(mi_col >> 1) + (mi_row >> 1) * (cm->mi_cols >> 1)];
123 if (candidate->mv[0].as_int != INVALID_MV) {
124 base_mv->as_mv.row = (candidate->mv[0].as_mv.row * 2);
125 base_mv->as_mv.col = (candidate->mv[0].as_mv.col * 2);
126 clamp_mv_ref(&base_mv->as_mv, xd);
128 base_mv->as_int = INVALID_MV;
134 x->mbmi_ext->mode_context[ref_frame] = counter_to_context[context_counter];
137 for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
138 clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
143 static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
144 BLOCK_SIZE bsize, int mi_row, int mi_col,
145 int_mv *tmp_mv, int *rate_mv,
146 int64_t best_rd_sofar, int use_base_mv) {
147 MACROBLOCKD *xd = &x->e_mbd;
148 MODE_INFO *mi = xd->mi[0];
149 struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } };
150 const int step_param = cpi->sf.mv.fullpel_search_step_param;
151 const int sadpb = x->sadperbit16;
153 const int ref = mi->ref_frame[0];
154 const MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
158 const MvLimits tmp_mv_limits = x->mv_limits;
161 const YV12_BUFFER_CONFIG *scaled_ref_frame =
162 vp9_get_scaled_ref_frame(cpi, ref);
163 if (scaled_ref_frame) {
165 // Swap out the reference frame for a version that's been scaled to
166 // match the resolution of the current frame, allowing the existing
167 // motion search code to be used without additional modifications.
168 for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
169 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
171 vp9_set_mv_search_range(&x->mv_limits, &ref_mv);
173 assert(x->mv_best_ref_index[ref] <= 2);
174 if (x->mv_best_ref_index[ref] < 2)
175 mvp_full = x->mbmi_ext->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
177 mvp_full = x->pred_mv[ref];
185 center_mv = tmp_mv->as_mv;
187 vp9_full_pixel_search(
188 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method, sadpb,
189 cond_cost_list(cpi, cost_list), ¢er_mv, &tmp_mv->as_mv, INT_MAX, 0);
191 x->mv_limits = tmp_mv_limits;
193 // calculate the bit cost on motion vector
194 mvp_full.row = tmp_mv->as_mv.row * 8;
195 mvp_full.col = tmp_mv->as_mv.col * 8;
197 *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv, x->nmvjointcost, x->mvcost,
201 cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref]][INTER_OFFSET(NEWMV)];
203 !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) > best_rd_sofar);
206 const int subpel_force_stop = use_base_mv && cpi->sf.base_mv_aggressive
208 : cpi->sf.mv.subpel_force_stop;
209 cpi->find_fractional_mv_step(
210 x, &tmp_mv->as_mv, &ref_mv, cpi->common.allow_high_precision_mv,
211 x->errorperbit, &cpi->fn_ptr[bsize], subpel_force_stop,
212 cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
213 x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, 0, 0);
214 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost,
215 x->mvcost, MV_COST_WEIGHT);
218 if (scaled_ref_frame) {
220 for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
225 static void block_variance(const uint8_t *src, int src_stride,
226 const uint8_t *ref, int ref_stride, int w, int h,
227 unsigned int *sse, int *sum, int block_size,
228 #if CONFIG_VP9_HIGHBITDEPTH
229 int use_highbitdepth, vpx_bit_depth_t bd,
231 uint32_t *sse8x8, int *sum8x8, uint32_t *var8x8) {
237 for (i = 0; i < h; i += block_size) {
238 for (j = 0; j < w; j += block_size) {
239 #if CONFIG_VP9_HIGHBITDEPTH
240 if (use_highbitdepth) {
243 vpx_highbd_8_get8x8var(src + src_stride * i + j, src_stride,
244 ref + ref_stride * i + j, ref_stride,
245 &sse8x8[k], &sum8x8[k]);
248 vpx_highbd_10_get8x8var(src + src_stride * i + j, src_stride,
249 ref + ref_stride * i + j, ref_stride,
250 &sse8x8[k], &sum8x8[k]);
253 vpx_highbd_12_get8x8var(src + src_stride * i + j, src_stride,
254 ref + ref_stride * i + j, ref_stride,
255 &sse8x8[k], &sum8x8[k]);
259 vpx_get8x8var(src + src_stride * i + j, src_stride,
260 ref + ref_stride * i + j, ref_stride, &sse8x8[k],
264 vpx_get8x8var(src + src_stride * i + j, src_stride,
265 ref + ref_stride * i + j, ref_stride, &sse8x8[k],
270 var8x8[k] = sse8x8[k] - (uint32_t)(((int64_t)sum8x8[k] * sum8x8[k]) >> 6);
276 static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
277 unsigned int *sse_i, int *sum_i,
278 unsigned int *var_o, unsigned int *sse_o,
280 const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
281 const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
282 const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
285 for (i = 0; i < nh; i += 2) {
286 for (j = 0; j < nw; j += 2) {
287 sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
288 sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
289 sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
290 sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
291 var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >>
292 (b_width_log2_lookup[unit_size] +
293 b_height_log2_lookup[unit_size] + 6));
299 // Adjust the ac_thr according to speed, width, height and normalized sum
300 static int ac_thr_factor(const int speed, const int width, const int height,
301 const int norm_sum) {
302 if (speed >= 8 && norm_sum < 5) {
303 if (width <= 640 && height <= 480)
311 static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize,
312 MACROBLOCK *x, MACROBLOCKD *xd,
313 int *out_rate_sum, int64_t *out_dist_sum,
314 unsigned int *var_y, unsigned int *sse_y,
315 int mi_row, int mi_col, int *early_term) {
316 // Note our transform coeffs are 8 times an orthogonal transform.
317 // Hence quantizer step is also 8 times. To get effective quantizer
318 // we need to divide by 8 before sending to modeling function.
322 struct macroblock_plane *const p = &x->plane[0];
323 struct macroblockd_plane *const pd = &xd->plane[0];
324 const uint32_t dc_quant = pd->dequant[0];
325 const uint32_t ac_quant = pd->dequant[1];
326 const int64_t dc_thr = dc_quant * dc_quant >> 6;
327 int64_t ac_thr = ac_quant * ac_quant >> 6;
332 const int bw = b_width_log2_lookup[bsize];
333 const int bh = b_height_log2_lookup[bsize];
334 const int num8x8 = 1 << (bw + bh - 2);
335 unsigned int sse8x8[64] = { 0 };
336 int sum8x8[64] = { 0 };
337 unsigned int var8x8[64] = { 0 };
340 #if CONFIG_VP9_HIGHBITDEPTH
341 const vpx_bit_depth_t bd = cpi->common.bit_depth;
343 // Calculate variance for whole partition, and also save 8x8 blocks' variance
344 // to be used in following transform skipping test.
345 block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
346 4 << bw, 4 << bh, &sse, &sum, 8,
347 #if CONFIG_VP9_HIGHBITDEPTH
348 cpi->common.use_highbitdepth, bd,
350 sse8x8, sum8x8, var8x8);
351 var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
356 #if CONFIG_VP9_TEMPORAL_DENOISING
357 if (cpi->oxcf.noise_sensitivity > 0)
358 ac_thr = vp9_scale_acskip_thresh(ac_thr, cpi->denoiser.denoising_level,
359 (abs(sum) >> (bw + bh)));
361 ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
362 cpi->common.height, abs(sum) >> (bw + bh));
364 ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
365 cpi->common.height, abs(sum) >> (bw + bh));
368 if (cpi->common.tx_mode == TX_MODE_SELECT) {
369 if (sse > (var << 2))
370 tx_size = VPXMIN(max_txsize_lookup[bsize],
371 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
375 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
376 cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
378 else if (tx_size > TX_16X16)
381 tx_size = VPXMIN(max_txsize_lookup[bsize],
382 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
385 assert(tx_size >= TX_8X8);
386 xd->mi[0]->tx_size = tx_size;
388 // Evaluate if the partition block is a skippable block in Y plane.
390 unsigned int sse16x16[16] = { 0 };
391 int sum16x16[16] = { 0 };
392 unsigned int var16x16[16] = { 0 };
393 const int num16x16 = num8x8 >> 2;
395 unsigned int sse32x32[4] = { 0 };
396 int sum32x32[4] = { 0 };
397 unsigned int var32x32[4] = { 0 };
398 const int num32x32 = num8x8 >> 4;
402 const int num = (tx_size == TX_8X8)
404 : ((tx_size == TX_16X16) ? num16x16 : num32x32);
405 const unsigned int *sse_tx =
406 (tx_size == TX_8X8) ? sse8x8
407 : ((tx_size == TX_16X16) ? sse16x16 : sse32x32);
408 const unsigned int *var_tx =
409 (tx_size == TX_8X8) ? var8x8
410 : ((tx_size == TX_16X16) ? var16x16 : var32x32);
412 // Calculate variance if tx_size > TX_8X8
413 if (tx_size >= TX_16X16)
414 calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
416 if (tx_size == TX_32X32)
417 calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32,
421 x->skip_txfm[0] = SKIP_TXFM_NONE;
422 for (k = 0; k < num; k++)
423 // Check if all ac coefficients can be quantized to zero.
424 if (!(var_tx[k] < ac_thr || var == 0)) {
429 for (k = 0; k < num; k++)
430 // Check if dc coefficient can be quantized to zero.
431 if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
437 x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
439 if (dc_test) x->skip_txfm[0] = SKIP_TXFM_AC_DC;
440 } else if (dc_test) {
445 if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
446 int skip_uv[2] = { 0 };
447 unsigned int var_uv[2];
448 unsigned int sse_uv[2];
451 *out_dist_sum = sse << 4;
453 // Transform skipping test in UV planes.
454 for (i = 1; i <= 2; i++) {
455 struct macroblock_plane *const p = &x->plane[i];
456 struct macroblockd_plane *const pd = &xd->plane[i];
457 const TX_SIZE uv_tx_size = get_uv_tx_size(xd->mi[0], pd);
458 const BLOCK_SIZE unit_size = txsize_to_bsize[uv_tx_size];
459 const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, pd);
460 const int uv_bw = b_width_log2_lookup[uv_bsize];
461 const int uv_bh = b_height_log2_lookup[uv_bsize];
462 const int sf = (uv_bw - b_width_log2_lookup[unit_size]) +
463 (uv_bh - b_height_log2_lookup[unit_size]);
464 const uint32_t uv_dc_thr = pd->dequant[0] * pd->dequant[0] >> (6 - sf);
465 const uint32_t uv_ac_thr = pd->dequant[1] * pd->dequant[1] >> (6 - sf);
468 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, i);
469 var_uv[j] = cpi->fn_ptr[uv_bsize].vf(
470 p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, &sse_uv[j]);
472 if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
473 (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
479 // If the transform in YUV planes are skippable, the mode search checks
480 // fewer inter modes and doesn't check intra modes.
481 if (skip_uv[0] & skip_uv[1]) {
489 #if CONFIG_VP9_HIGHBITDEPTH
490 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
491 dc_quant >> (xd->bd - 5), &rate, &dist);
493 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
494 dc_quant >> 3, &rate, &dist);
495 #endif // CONFIG_VP9_HIGHBITDEPTH
499 *out_rate_sum = rate >> 1;
500 *out_dist_sum = dist << 3;
503 *out_dist_sum = (sse - var) << 4;
506 #if CONFIG_VP9_HIGHBITDEPTH
507 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
508 ac_quant >> (xd->bd - 5), &rate, &dist);
510 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
512 #endif // CONFIG_VP9_HIGHBITDEPTH
514 *out_rate_sum += rate;
515 *out_dist_sum += dist << 4;
518 static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
519 MACROBLOCKD *xd, int *out_rate_sum,
520 int64_t *out_dist_sum, unsigned int *var_y,
521 unsigned int *sse_y) {
522 // Note our transform coeffs are 8 times an orthogonal transform.
523 // Hence quantizer step is also 8 times. To get effective quantizer
524 // we need to divide by 8 before sending to modeling function.
528 struct macroblock_plane *const p = &x->plane[0];
529 struct macroblockd_plane *const pd = &xd->plane[0];
530 const int64_t dc_thr = p->quant_thred[0] >> 6;
531 const int64_t ac_thr = p->quant_thred[1] >> 6;
532 const uint32_t dc_quant = pd->dequant[0];
533 const uint32_t ac_quant = pd->dequant[1];
534 unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
535 pd->dst.buf, pd->dst.stride, &sse);
541 if (cpi->common.tx_mode == TX_MODE_SELECT) {
542 if (sse > (var << 2))
544 VPXMIN(max_txsize_lookup[bsize],
545 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
547 xd->mi[0]->tx_size = TX_8X8;
549 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
550 cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
551 xd->mi[0]->tx_size = TX_8X8;
552 else if (xd->mi[0]->tx_size > TX_16X16)
553 xd->mi[0]->tx_size = TX_16X16;
556 VPXMIN(max_txsize_lookup[bsize],
557 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
560 // Evaluate if the partition block is a skippable block in Y plane.
562 const BLOCK_SIZE unit_size = txsize_to_bsize[xd->mi[0]->tx_size];
563 const unsigned int num_blk_log2 =
564 (b_width_log2_lookup[bsize] - b_width_log2_lookup[unit_size]) +
565 (b_height_log2_lookup[bsize] - b_height_log2_lookup[unit_size]);
566 const unsigned int sse_tx = sse >> num_blk_log2;
567 const unsigned int var_tx = var >> num_blk_log2;
569 x->skip_txfm[0] = SKIP_TXFM_NONE;
570 // Check if all ac coefficients can be quantized to zero.
571 if (var_tx < ac_thr || var == 0) {
572 x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
573 // Check if dc coefficient can be quantized to zero.
574 if (sse_tx - var_tx < dc_thr || sse == var)
575 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
577 if (sse_tx - var_tx < dc_thr || sse == var) skip_dc = 1;
581 if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
583 *out_dist_sum = sse << 4;
588 #if CONFIG_VP9_HIGHBITDEPTH
589 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
590 dc_quant >> (xd->bd - 5), &rate, &dist);
592 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
593 dc_quant >> 3, &rate, &dist);
594 #endif // CONFIG_VP9_HIGHBITDEPTH
598 *out_rate_sum = rate >> 1;
599 *out_dist_sum = dist << 3;
602 *out_dist_sum = (sse - var) << 4;
605 #if CONFIG_VP9_HIGHBITDEPTH
606 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
607 ac_quant >> (xd->bd - 5), &rate, &dist);
609 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
611 #endif // CONFIG_VP9_HIGHBITDEPTH
613 *out_rate_sum += rate;
614 *out_dist_sum += dist << 4;
617 static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
618 int *skippable, int64_t *sse, BLOCK_SIZE bsize,
620 MACROBLOCKD *xd = &x->e_mbd;
621 const struct macroblockd_plane *pd = &xd->plane[0];
622 struct macroblock_plane *const p = &x->plane[0];
623 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
624 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
625 const int step = 1 << (tx_size << 1);
626 const int block_step = (1 << tx_size);
628 const int max_blocks_wide =
629 num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5);
630 const int max_blocks_high =
631 num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5);
633 const int bw = 4 * num_4x4_w;
634 const int bh = 4 * num_4x4_h;
636 #if CONFIG_VP9_HIGHBITDEPTH
637 // TODO(jingning): Implement the high bit-depth Hadamard transforms and
638 // remove this check condition.
639 // TODO(marpan): Disable this for 8 bit once optimizations for the functions
640 // below are merged in.
641 // if (xd->bd != 8) {
642 unsigned int var_y, sse_y;
644 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist, &var_y,
654 // The max tx_size passed in is TX_16X16.
655 assert(tx_size != TX_32X32);
657 vpx_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
658 pd->dst.buf, pd->dst.stride);
660 // Keep track of the row and column of the blocks we use so that we know
661 // if we are in the unrestricted motion border.
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 const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
666 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
667 tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
668 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
669 uint16_t *const eob = &p->eobs[block];
670 const int diff_stride = bw;
671 const int16_t *src_diff;
672 src_diff = &p->src_diff[(r * diff_stride + c) << 2];
676 vpx_hadamard_16x16(src_diff, diff_stride, coeff);
677 vp9_quantize_fp(coeff, 256, x->skip_block, p->zbin, p->round_fp,
678 p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
679 pd->dequant, eob, scan_order->scan,
683 vpx_hadamard_8x8(src_diff, diff_stride, coeff);
684 vp9_quantize_fp(coeff, 64, x->skip_block, p->zbin, p->round_fp,
685 p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
686 pd->dequant, eob, scan_order->scan,
690 x->fwd_txm4x4(src_diff, coeff, diff_stride);
691 vp9_quantize_fp(coeff, 16, x->skip_block, p->zbin, p->round_fp,
692 p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
693 pd->dequant, eob, scan_order->scan,
696 default: assert(0); break;
698 *skippable &= (*eob == 0);
706 if (*sse < INT64_MAX) {
707 *sse = (*sse << 6) >> 2;
709 this_rdc->dist = *sse;
716 for (r = 0; r < max_blocks_high; r += block_step) {
717 for (c = 0; c < num_4x4_w; c += block_step) {
718 if (c < max_blocks_wide) {
719 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
720 tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
721 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
722 uint16_t *const eob = &p->eobs[block];
725 this_rdc->rate += (int)abs(qcoeff[0]);
727 this_rdc->rate += vpx_satd(qcoeff, step << 4);
729 this_rdc->dist += vp9_block_error_fp(coeff, dqcoeff, step << 4) >> 2;
735 // If skippable is set, rate gets clobbered later.
736 this_rdc->rate <<= (2 + VP9_PROB_COST_SHIFT);
737 this_rdc->rate += (eob_cost << VP9_PROB_COST_SHIFT);
740 static void model_rd_for_sb_uv(VP9_COMP *cpi, BLOCK_SIZE plane_bsize,
741 MACROBLOCK *x, MACROBLOCKD *xd,
742 RD_COST *this_rdc, unsigned int *var_y,
743 unsigned int *sse_y, int start_plane,
745 // Note our transform coeffs are 8 times an orthogonal transform.
746 // Hence quantizer step is also 8 times. To get effective quantizer
747 // we need to divide by 8 before sending to modeling function.
752 #if CONFIG_VP9_HIGHBITDEPTH
753 uint64_t tot_var = *var_y;
754 uint64_t tot_sse = *sse_y;
756 uint32_t tot_var = *var_y;
757 uint32_t tot_sse = *sse_y;
763 for (i = start_plane; i <= stop_plane; ++i) {
764 struct macroblock_plane *const p = &x->plane[i];
765 struct macroblockd_plane *const pd = &xd->plane[i];
766 const uint32_t dc_quant = pd->dequant[0];
767 const uint32_t ac_quant = pd->dequant[1];
768 const BLOCK_SIZE bs = plane_bsize;
770 if (!x->color_sensitivity[i - 1]) continue;
772 var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf,
773 pd->dst.stride, &sse);
778 #if CONFIG_VP9_HIGHBITDEPTH
779 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
780 dc_quant >> (xd->bd - 5), &rate, &dist);
782 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
783 dc_quant >> 3, &rate, &dist);
784 #endif // CONFIG_VP9_HIGHBITDEPTH
786 this_rdc->rate += rate >> 1;
787 this_rdc->dist += dist << 3;
789 #if CONFIG_VP9_HIGHBITDEPTH
790 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
791 ac_quant >> (xd->bd - 5), &rate, &dist);
793 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs], ac_quant >> 3,
795 #endif // CONFIG_VP9_HIGHBITDEPTH
797 this_rdc->rate += rate;
798 this_rdc->dist += dist << 4;
801 #if CONFIG_VP9_HIGHBITDEPTH
802 *var_y = tot_var > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_var;
803 *sse_y = tot_sse > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_sse;
810 static int get_pred_buffer(PRED_BUFFER *p, int len) {
813 for (i = 0; i < len; i++) {
822 static void free_pred_buffer(PRED_BUFFER *p) {
823 if (p != NULL) p->in_use = 0;
826 static void encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
827 int mi_row, int mi_col,
828 MV_REFERENCE_FRAME ref_frame,
829 PREDICTION_MODE this_mode, unsigned int var_y,
831 struct buf_2d yv12_mb[][MAX_MB_PLANE],
832 int *rate, int64_t *dist) {
833 MACROBLOCKD *xd = &x->e_mbd;
834 MODE_INFO *const mi = xd->mi[0];
835 const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
836 unsigned int var = var_y, sse = sse_y;
837 // Skipping threshold for ac.
838 unsigned int thresh_ac;
839 // Skipping threshold for dc.
840 unsigned int thresh_dc;
842 if (mi->mv[0].as_mv.row > 64 || mi->mv[0].as_mv.row < -64 ||
843 mi->mv[0].as_mv.col > 64 || mi->mv[0].as_mv.col < -64)
845 if (x->encode_breakout > 0 && motion_low == 1) {
846 // Set a maximum for threshold to avoid big PSNR loss in low bit rate
847 // case. Use extreme low threshold for static frames to limit
849 const unsigned int max_thresh = 36000;
850 // The encode_breakout input
851 const unsigned int min_thresh =
852 VPXMIN(((unsigned int)x->encode_breakout << 4), max_thresh);
853 #if CONFIG_VP9_HIGHBITDEPTH
854 const int shift = (xd->bd << 1) - 16;
857 // Calculate threshold according to dequant value.
858 thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) >> 3;
859 #if CONFIG_VP9_HIGHBITDEPTH
860 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
861 thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift);
863 #endif // CONFIG_VP9_HIGHBITDEPTH
864 thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
866 // Adjust ac threshold according to partition size.
868 8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
870 thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
871 #if CONFIG_VP9_HIGHBITDEPTH
872 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
873 thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift);
875 #endif // CONFIG_VP9_HIGHBITDEPTH
881 // Y skipping condition checking for ac and dc.
882 if (var <= thresh_ac && (sse - var) <= thresh_dc) {
883 unsigned int sse_u, sse_v;
884 unsigned int var_u, var_v;
885 unsigned int thresh_ac_uv = thresh_ac;
886 unsigned int thresh_dc_uv = thresh_dc;
892 // Skip UV prediction unless breakout is zero (lossless) to save
893 // computation with low impact on the result
894 if (x->encode_breakout == 0) {
895 xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
896 xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
897 vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
900 var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf, x->plane[1].src.stride,
901 xd->plane[1].dst.buf,
902 xd->plane[1].dst.stride, &sse_u);
904 // U skipping condition checking
905 if (((var_u << 2) <= thresh_ac_uv) && (sse_u - var_u <= thresh_dc_uv)) {
906 var_v = cpi->fn_ptr[uv_size].vf(
907 x->plane[2].src.buf, x->plane[2].src.stride, xd->plane[2].dst.buf,
908 xd->plane[2].dst.stride, &sse_v);
910 // V skipping condition checking
911 if (((var_v << 2) <= thresh_ac_uv) && (sse_v - var_v <= thresh_dc_uv)) {
914 // The cost of skip bit needs to be added.
915 *rate = cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
916 [INTER_OFFSET(this_mode)];
918 // More on this part of rate
919 // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
921 // Scaling factor for SSE from spatial domain to frequency
922 // domain is 16. Adjust distortion accordingly.
923 // TODO(yunqingwang): In this function, only y-plane dist is
925 *dist = (sse << 4); // + ((sse_u + sse_v) << 4);
927 // *disable_skip = 1;
933 struct estimate_block_intra_args {
936 PREDICTION_MODE mode;
941 static void estimate_block_intra(int plane, int block, int row, int col,
942 BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
944 struct estimate_block_intra_args *const args = arg;
945 VP9_COMP *const cpi = args->cpi;
946 MACROBLOCK *const x = args->x;
947 MACROBLOCKD *const xd = &x->e_mbd;
948 struct macroblock_plane *const p = &x->plane[0];
949 struct macroblockd_plane *const pd = &xd->plane[0];
950 const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
951 uint8_t *const src_buf_base = p->src.buf;
952 uint8_t *const dst_buf_base = pd->dst.buf;
953 const int src_stride = p->src.stride;
954 const int dst_stride = pd->dst.stride;
959 p->src.buf = &src_buf_base[4 * (row * src_stride + col)];
960 pd->dst.buf = &dst_buf_base[4 * (row * dst_stride + col)];
961 // Use source buffer as an approximation for the fully reconstructed buffer.
962 vp9_predict_intra_block(xd, b_width_log2_lookup[plane_bsize], tx_size,
963 args->mode, x->skip_encode ? p->src.buf : pd->dst.buf,
964 x->skip_encode ? src_stride : dst_stride, pd->dst.buf,
965 dst_stride, col, row, plane);
968 int64_t this_sse = INT64_MAX;
969 // TODO(jingning): This needs further refactoring.
970 block_yrd(cpi, x, &this_rdc, &args->skippable, &this_sse, bsize_tx,
971 VPXMIN(tx_size, TX_16X16));
973 unsigned int var = 0;
974 unsigned int sse = 0;
975 model_rd_for_sb_uv(cpi, plane_bsize, x, xd, &this_rdc, &var, &sse, plane,
979 p->src.buf = src_buf_base;
980 pd->dst.buf = dst_buf_base;
981 args->rdc->rate += this_rdc.rate;
982 args->rdc->dist += this_rdc.dist;
985 static const THR_MODES mode_idx[MAX_REF_FRAMES][4] = {
986 { THR_DC, THR_V_PRED, THR_H_PRED, THR_TM },
987 { THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV },
988 { THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG },
989 { THR_NEARESTA, THR_NEARA, THR_ZEROA, THR_NEWA },
992 static const PREDICTION_MODE intra_mode_list[] = { DC_PRED, V_PRED, H_PRED,
995 static int mode_offset(const PREDICTION_MODE mode) {
996 if (mode >= NEARESTMV) {
997 return INTER_OFFSET(mode);
1000 case DC_PRED: return 0;
1001 case V_PRED: return 1;
1002 case H_PRED: return 2;
1003 case TM_PRED: return 3;
1009 static INLINE void update_thresh_freq_fact(
1010 VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance,
1011 BLOCK_SIZE bsize, MV_REFERENCE_FRAME ref_frame, THR_MODES best_mode_idx,
1012 PREDICTION_MODE mode) {
1013 THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1014 int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
1015 if (thr_mode_idx == best_mode_idx)
1016 *freq_fact -= (*freq_fact >> 4);
1017 else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
1018 ref_frame == LAST_FRAME && source_variance < 5) {
1019 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
1021 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
1022 cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1026 void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
1027 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1028 MACROBLOCKD *const xd = &x->e_mbd;
1029 MODE_INFO *const mi = xd->mi[0];
1030 RD_COST this_rdc, best_rdc;
1031 PREDICTION_MODE this_mode;
1032 struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
1033 const TX_SIZE intra_tx_size =
1034 VPXMIN(max_txsize_lookup[bsize],
1035 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
1036 MODE_INFO *const mic = xd->mi[0];
1038 const MODE_INFO *above_mi = xd->above_mi;
1039 const MODE_INFO *left_mi = xd->left_mi;
1040 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1041 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1042 bmode_costs = cpi->y_mode_costs[A][L];
1045 vp9_rd_cost_reset(&best_rdc);
1046 vp9_rd_cost_reset(&this_rdc);
1048 mi->ref_frame[0] = INTRA_FRAME;
1049 // Initialize interp_filter here so we do not have to check for inter block
1050 // modes in get_pred_context_switchable_interp()
1051 mi->interp_filter = SWITCHABLE_FILTERS;
1053 mi->mv[0].as_int = INVALID_MV;
1054 mi->uv_mode = DC_PRED;
1055 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
1057 // Change the limit of this loop to add other intra prediction
1059 for (this_mode = DC_PRED; this_mode <= H_PRED; ++this_mode) {
1060 this_rdc.dist = this_rdc.rate = 0;
1061 args.mode = this_mode;
1063 args.rdc = &this_rdc;
1064 mi->tx_size = intra_tx_size;
1065 vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
1067 if (args.skippable) {
1068 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1069 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
1071 x->skip_txfm[0] = SKIP_TXFM_NONE;
1072 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
1074 this_rdc.rate += bmode_costs[this_mode];
1075 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1077 if (this_rdc.rdcost < best_rdc.rdcost) {
1078 best_rdc = this_rdc;
1079 mi->mode = this_mode;
1083 *rd_cost = best_rdc;
1086 static void init_ref_frame_cost(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1087 int ref_frame_cost[MAX_REF_FRAMES]) {
1088 vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
1089 vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1090 vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1092 ref_frame_cost[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1093 ref_frame_cost[LAST_FRAME] = ref_frame_cost[GOLDEN_FRAME] =
1094 ref_frame_cost[ALTREF_FRAME] = vp9_cost_bit(intra_inter_p, 1);
1096 ref_frame_cost[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1097 ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1098 ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1099 ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1100 ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1104 MV_REFERENCE_FRAME ref_frame;
1105 PREDICTION_MODE pred_mode;
1108 #define RT_INTER_MODES 12
1109 static const REF_MODE ref_mode_set[RT_INTER_MODES] = {
1110 { LAST_FRAME, ZEROMV }, { LAST_FRAME, NEARESTMV },
1111 { GOLDEN_FRAME, ZEROMV }, { LAST_FRAME, NEARMV },
1112 { LAST_FRAME, NEWMV }, { GOLDEN_FRAME, NEARESTMV },
1113 { GOLDEN_FRAME, NEARMV }, { GOLDEN_FRAME, NEWMV },
1114 { ALTREF_FRAME, ZEROMV }, { ALTREF_FRAME, NEARESTMV },
1115 { ALTREF_FRAME, NEARMV }, { ALTREF_FRAME, NEWMV }
1117 static const REF_MODE ref_mode_set_svc[RT_INTER_MODES] = {
1118 { LAST_FRAME, ZEROMV }, { GOLDEN_FRAME, ZEROMV },
1119 { LAST_FRAME, NEARESTMV }, { LAST_FRAME, NEARMV },
1120 { GOLDEN_FRAME, NEARESTMV }, { GOLDEN_FRAME, NEARMV },
1121 { LAST_FRAME, NEWMV }, { GOLDEN_FRAME, NEWMV }
1124 static int set_intra_cost_penalty(const VP9_COMP *const cpi, BLOCK_SIZE bsize) {
1125 const VP9_COMMON *const cm = &cpi->common;
1126 // Reduce the intra cost penalty for small blocks (<=16x16).
1128 (bsize <= BLOCK_16X16) ? ((bsize <= BLOCK_8X8) ? 4 : 2) : 0;
1129 if (cpi->noise_estimate.enabled && cpi->noise_estimate.level == kHigh)
1130 // Don't reduce intra cost penalty if estimated noise level is high.
1132 return vp9_get_intra_cost_penalty(cm->base_qindex, cm->y_dc_delta_q,
1137 static INLINE void find_predictors(
1138 VP9_COMP *cpi, MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame,
1139 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1140 int const_motion[MAX_REF_FRAMES], int *ref_frame_skip_mask,
1141 const int flag_list[4], TileDataEnc *tile_data, int mi_row, int mi_col,
1142 struct buf_2d yv12_mb[4][MAX_MB_PLANE], BLOCK_SIZE bsize,
1143 int force_skip_low_temp_var) {
1144 VP9_COMMON *const cm = &cpi->common;
1145 MACROBLOCKD *const xd = &x->e_mbd;
1146 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1147 TileInfo *const tile_info = &tile_data->tile_info;
1148 // TODO(jingning) placeholder for inter-frame non-RD mode decision.
1149 x->pred_mv_sad[ref_frame] = INT_MAX;
1150 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
1151 frame_mv[ZEROMV][ref_frame].as_int = 0;
1152 // this needs various further optimizations. to be continued..
1153 if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
1154 int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
1155 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
1156 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
1157 if (cm->use_prev_frame_mvs) {
1158 vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
1159 x->mbmi_ext->mode_context);
1161 const_motion[ref_frame] =
1162 mv_refs_rt(cpi, cm, x, xd, tile_info, xd->mi[0], ref_frame,
1163 candidates, &frame_mv[NEWMV][ref_frame], mi_row, mi_col,
1164 (int)(cpi->svc.use_base_mv && cpi->svc.spatial_layer_id));
1166 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1167 &frame_mv[NEARESTMV][ref_frame],
1168 &frame_mv[NEARMV][ref_frame]);
1169 // Early exit for golden frame if force_skip_low_temp_var is set.
1170 if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8 &&
1171 !(force_skip_low_temp_var && ref_frame == GOLDEN_FRAME)) {
1172 vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
1176 *ref_frame_skip_mask |= (1 << ref_frame);
1180 static void vp9_NEWMV_diff_bias(const NOISE_ESTIMATE *ne, MACROBLOCKD *xd,
1181 PREDICTION_MODE this_mode, RD_COST *this_rdc,
1182 BLOCK_SIZE bsize, int mv_row, int mv_col,
1183 int is_last_frame) {
1184 // Bias against MVs associated with NEWMV mode that are very different from
1185 // top/left neighbors.
1186 if (this_mode == NEWMV) {
1187 int al_mv_average_row;
1188 int al_mv_average_col;
1189 int left_row, left_col;
1190 int row_diff, col_diff;
1191 int above_mv_valid = 0;
1192 int left_mv_valid = 0;
1197 above_mv_valid = xd->above_mi->mv[0].as_int != INVALID_MV;
1198 above_row = xd->above_mi->mv[0].as_mv.row;
1199 above_col = xd->above_mi->mv[0].as_mv.col;
1202 left_mv_valid = xd->left_mi->mv[0].as_int != INVALID_MV;
1203 left_row = xd->left_mi->mv[0].as_mv.row;
1204 left_col = xd->left_mi->mv[0].as_mv.col;
1206 if (above_mv_valid && left_mv_valid) {
1207 al_mv_average_row = (above_row + left_row + 1) >> 1;
1208 al_mv_average_col = (above_col + left_col + 1) >> 1;
1209 } else if (above_mv_valid) {
1210 al_mv_average_row = above_row;
1211 al_mv_average_col = above_col;
1212 } else if (left_mv_valid) {
1213 al_mv_average_row = left_row;
1214 al_mv_average_col = left_col;
1216 al_mv_average_row = al_mv_average_col = 0;
1218 row_diff = (al_mv_average_row - mv_row);
1219 col_diff = (al_mv_average_col - mv_col);
1220 if (row_diff > 48 || row_diff < -48 || col_diff > 48 || col_diff < -48) {
1221 if (bsize > BLOCK_32X32)
1222 this_rdc->rdcost = this_rdc->rdcost << 1;
1224 this_rdc->rdcost = 3 * this_rdc->rdcost >> 1;
1227 // If noise estimation is enabled, and estimated level is above threshold,
1228 // add a bias to LAST reference with small motion, for large blocks.
1229 if (ne->enabled && ne->level >= kMedium && bsize >= BLOCK_32X32 &&
1230 is_last_frame && mv_row < 8 && mv_row > -8 && mv_col < 8 && mv_col > -8) {
1231 this_rdc->rdcost = 7 * this_rdc->rdcost >> 3;
1235 #if CONFIG_VP9_TEMPORAL_DENOISING
1236 static void vp9_pickmode_ctx_den_update(
1237 VP9_PICKMODE_CTX_DEN *ctx_den, int64_t zero_last_cost_orig,
1238 int ref_frame_cost[MAX_REF_FRAMES],
1239 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int reuse_inter_pred,
1240 TX_SIZE best_tx_size, PREDICTION_MODE best_mode,
1241 MV_REFERENCE_FRAME best_ref_frame, INTERP_FILTER best_pred_filter,
1242 uint8_t best_mode_skip_txfm) {
1243 ctx_den->zero_last_cost_orig = zero_last_cost_orig;
1244 ctx_den->ref_frame_cost = ref_frame_cost;
1245 ctx_den->frame_mv = frame_mv;
1246 ctx_den->reuse_inter_pred = reuse_inter_pred;
1247 ctx_den->best_tx_size = best_tx_size;
1248 ctx_den->best_mode = best_mode;
1249 ctx_den->best_ref_frame = best_ref_frame;
1250 ctx_den->best_pred_filter = best_pred_filter;
1251 ctx_den->best_mode_skip_txfm = best_mode_skip_txfm;
1254 static void recheck_zeromv_after_denoising(
1255 VP9_COMP *cpi, MODE_INFO *const mi, MACROBLOCK *x, MACROBLOCKD *const xd,
1256 VP9_DENOISER_DECISION decision, VP9_PICKMODE_CTX_DEN *ctx_den,
1257 struct buf_2d yv12_mb[4][MAX_MB_PLANE], RD_COST *best_rdc, BLOCK_SIZE bsize,
1258 int mi_row, int mi_col) {
1259 // If INTRA or GOLDEN reference was selected, re-evaluate ZEROMV on
1260 // denoised result. Only do this under noise conditions, and if rdcost of
1261 // ZEROMV onoriginal source is not significantly higher than rdcost of best
1263 if (cpi->noise_estimate.enabled && cpi->noise_estimate.level > kLow &&
1264 ctx_den->zero_last_cost_orig < (best_rdc->rdcost << 3) &&
1265 ((ctx_den->best_ref_frame == INTRA_FRAME && decision >= FILTER_BLOCK) ||
1266 (ctx_den->best_ref_frame == GOLDEN_FRAME &&
1267 cpi->svc.number_spatial_layers == 1 &&
1268 decision == FILTER_ZEROMV_BLOCK))) {
1269 // Check if we should pick ZEROMV on denoised signal.
1272 uint32_t var_y = UINT_MAX;
1273 uint32_t sse_y = UINT_MAX;
1276 mi->ref_frame[0] = LAST_FRAME;
1277 mi->ref_frame[1] = NONE;
1278 mi->mv[0].as_int = 0;
1279 mi->interp_filter = EIGHTTAP;
1280 xd->plane[0].pre[0] = yv12_mb[LAST_FRAME][0];
1281 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1282 model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
1283 this_rdc.rate = rate + ctx_den->ref_frame_cost[LAST_FRAME] +
1284 cpi->inter_mode_cost[x->mbmi_ext->mode_context[LAST_FRAME]]
1285 [INTER_OFFSET(ZEROMV)];
1286 this_rdc.dist = dist;
1287 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, rate, dist);
1288 // Don't switch to ZEROMV if the rdcost for ZEROMV on denoised source
1289 // is higher than best_ref mode (on original source).
1290 if (this_rdc.rdcost > best_rdc->rdcost) {
1291 this_rdc = *best_rdc;
1292 mi->mode = ctx_den->best_mode;
1293 mi->ref_frame[0] = ctx_den->best_ref_frame;
1294 mi->interp_filter = ctx_den->best_pred_filter;
1295 if (ctx_den->best_ref_frame == INTRA_FRAME) {
1296 mi->mv[0].as_int = INVALID_MV;
1297 mi->interp_filter = SWITCHABLE_FILTERS;
1298 } else if (ctx_den->best_ref_frame == GOLDEN_FRAME) {
1300 ctx_den->frame_mv[ctx_den->best_mode][ctx_den->best_ref_frame]
1302 if (ctx_den->reuse_inter_pred) {
1303 xd->plane[0].pre[0] = yv12_mb[GOLDEN_FRAME][0];
1304 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1307 mi->tx_size = ctx_den->best_tx_size;
1308 x->skip_txfm[0] = ctx_den->best_mode_skip_txfm;
1310 ctx_den->best_ref_frame = LAST_FRAME;
1311 *best_rdc = this_rdc;
1315 #endif // CONFIG_VP9_TEMPORAL_DENOISING
1317 static INLINE int get_force_skip_low_temp_var(uint8_t *variance_low, int mi_row,
1318 int mi_col, BLOCK_SIZE bsize) {
1319 const int i = (mi_row & 0x7) >> 1;
1320 const int j = (mi_col & 0x7) >> 1;
1321 int force_skip_low_temp_var = 0;
1322 // Set force_skip_low_temp_var based on the block size and block offset.
1323 if (bsize == BLOCK_64X64) {
1324 force_skip_low_temp_var = variance_low[0];
1325 } else if (bsize == BLOCK_64X32) {
1326 if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1327 force_skip_low_temp_var = variance_low[1];
1328 } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1329 force_skip_low_temp_var = variance_low[2];
1331 } else if (bsize == BLOCK_32X64) {
1332 if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1333 force_skip_low_temp_var = variance_low[3];
1334 } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1335 force_skip_low_temp_var = variance_low[4];
1337 } else if (bsize == BLOCK_32X32) {
1338 if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1339 force_skip_low_temp_var = variance_low[5];
1340 } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1341 force_skip_low_temp_var = variance_low[6];
1342 } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1343 force_skip_low_temp_var = variance_low[7];
1344 } else if ((mi_col & 0x7) && (mi_row & 0x7)) {
1345 force_skip_low_temp_var = variance_low[8];
1347 } else if (bsize == BLOCK_16X16) {
1348 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]];
1349 } else if (bsize == BLOCK_32X16) {
1350 // The col shift index for the second 16x16 block.
1351 const int j2 = ((mi_col + 2) & 0x7) >> 1;
1352 // Only if each 16x16 block inside has low temporal variance.
1353 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1354 variance_low[pos_shift_16x16[i][j2]];
1355 } else if (bsize == BLOCK_16X32) {
1356 // The row shift index for the second 16x16 block.
1357 const int i2 = ((mi_row + 2) & 0x7) >> 1;
1358 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1359 variance_low[pos_shift_16x16[i2][j]];
1361 return force_skip_low_temp_var;
1364 void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data,
1365 int mi_row, int mi_col, RD_COST *rd_cost,
1366 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1367 VP9_COMMON *const cm = &cpi->common;
1368 SPEED_FEATURES *const sf = &cpi->sf;
1369 const SVC *const svc = &cpi->svc;
1370 MACROBLOCKD *const xd = &x->e_mbd;
1371 MODE_INFO *const mi = xd->mi[0];
1372 struct macroblockd_plane *const pd = &xd->plane[0];
1373 PREDICTION_MODE best_mode = ZEROMV;
1374 MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
1375 MV_REFERENCE_FRAME usable_ref_frame;
1376 TX_SIZE best_tx_size = TX_SIZES;
1377 INTERP_FILTER best_pred_filter = EIGHTTAP;
1378 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1379 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1380 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1382 RD_COST this_rdc, best_rdc;
1383 uint8_t skip_txfm = SKIP_TXFM_NONE, best_mode_skip_txfm = SKIP_TXFM_NONE;
1384 // var_y and sse_y are saved to be used in skipping checking
1385 unsigned int var_y = UINT_MAX;
1386 unsigned int sse_y = UINT_MAX;
1387 const int intra_cost_penalty = set_intra_cost_penalty(cpi, bsize);
1388 int64_t inter_mode_thresh =
1389 RDCOST(x->rdmult, x->rddiv, intra_cost_penalty, 0);
1390 const int *const rd_threshes = cpi->rd.threshes[mi->segment_id][bsize];
1391 const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
1392 INTERP_FILTER filter_ref;
1393 const int bsl = mi_width_log2_lookup[bsize];
1394 const int pred_filter_search =
1395 cm->interp_filter == SWITCHABLE
1396 ? (((mi_row + mi_col) >> bsl) +
1397 get_chessboard_index(cm->current_video_frame)) &
1400 int const_motion[MAX_REF_FRAMES] = { 0 };
1401 const int bh = num_4x4_blocks_high_lookup[bsize] << 2;
1402 const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
1403 // For speed 6, the result of interp filter is reused later in actual encoding
1405 // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
1407 DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 64 * 64]);
1408 #if CONFIG_VP9_HIGHBITDEPTH
1409 DECLARE_ALIGNED(16, uint16_t, pred_buf_16[3 * 64 * 64]);
1411 struct buf_2d orig_dst = pd->dst;
1412 PRED_BUFFER *best_pred = NULL;
1413 PRED_BUFFER *this_mode_pred = NULL;
1414 const int pixels_in_block = bh * bw;
1415 int reuse_inter_pred = cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready;
1416 int ref_frame_skip_mask = 0;
1418 int best_pred_sad = INT_MAX;
1419 int best_early_term = 0;
1420 int ref_frame_cost[MAX_REF_FRAMES];
1421 int svc_force_zero_mode[3] = { 0 };
1422 int perform_intra_pred = 1;
1423 int use_golden_nonzeromv = 1;
1424 int force_skip_low_temp_var = 0;
1425 int skip_ref_find_pred[4] = { 0 };
1426 #if CONFIG_VP9_TEMPORAL_DENOISING
1427 VP9_PICKMODE_CTX_DEN ctx_den;
1428 int64_t zero_last_cost_orig = INT64_MAX;
1431 init_ref_frame_cost(cm, xd, ref_frame_cost);
1433 if (reuse_inter_pred) {
1435 for (i = 0; i < 3; i++) {
1436 #if CONFIG_VP9_HIGHBITDEPTH
1437 if (cm->use_highbitdepth)
1438 tmp[i].data = CONVERT_TO_BYTEPTR(&pred_buf_16[pixels_in_block * i]);
1440 tmp[i].data = &pred_buf[pixels_in_block * i];
1442 tmp[i].data = &pred_buf[pixels_in_block * i];
1443 #endif // CONFIG_VP9_HIGHBITDEPTH
1447 tmp[3].data = pd->dst.buf;
1448 tmp[3].stride = pd->dst.stride;
1452 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1455 // Instead of using vp9_get_pred_context_switchable_interp(xd) to assign
1456 // filter_ref, we use a less strict condition on assigning filter_ref.
1457 // This is to reduce the probabily of entering the flow of not assigning
1458 // filter_ref and then skip filter search.
1459 if (xd->above_mi && is_inter_block(xd->above_mi))
1460 filter_ref = xd->above_mi->interp_filter;
1461 else if (xd->left_mi && is_inter_block(xd->left_mi))
1462 filter_ref = xd->left_mi->interp_filter;
1464 filter_ref = cm->interp_filter;
1466 // initialize mode decisions
1467 vp9_rd_cost_reset(&best_rdc);
1468 vp9_rd_cost_reset(rd_cost);
1469 mi->sb_type = bsize;
1470 mi->ref_frame[0] = NONE;
1471 mi->ref_frame[1] = NONE;
1474 VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[cm->tx_mode]);
1476 if (sf->short_circuit_flat_blocks || sf->limit_newmv_early_exit) {
1477 #if CONFIG_VP9_HIGHBITDEPTH
1478 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
1479 x->source_variance = vp9_high_get_sby_perpixel_variance(
1480 cpi, &x->plane[0].src, bsize, xd->bd);
1482 #endif // CONFIG_VP9_HIGHBITDEPTH
1483 x->source_variance =
1484 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1487 #if CONFIG_VP9_TEMPORAL_DENOISING
1488 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
1489 cpi->denoiser.denoising_level > kDenLowLow) {
1490 vp9_denoiser_reset_frame_stats(ctx);
1494 if (cpi->rc.frames_since_golden == 0 && !cpi->use_svc) {
1495 usable_ref_frame = LAST_FRAME;
1497 usable_ref_frame = GOLDEN_FRAME;
1500 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
1501 if (cpi->rc.alt_ref_gf_group || cpi->rc.is_src_frame_alt_ref)
1502 usable_ref_frame = ALTREF_FRAME;
1504 if (cpi->rc.is_src_frame_alt_ref) {
1505 skip_ref_find_pred[LAST_FRAME] = 1;
1506 skip_ref_find_pred[GOLDEN_FRAME] = 1;
1510 // For svc mode, on spatial_layer_id > 0: if the reference has different scale
1511 // constrain the inter mode to only test zero motion.
1512 if (cpi->use_svc && svc->force_zero_mode_spatial_ref &&
1513 cpi->svc.spatial_layer_id > 0) {
1514 if (cpi->ref_frame_flags & flag_list[LAST_FRAME]) {
1515 struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf;
1516 if (vp9_is_scaled(sf)) svc_force_zero_mode[LAST_FRAME - 1] = 1;
1518 if (cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) {
1519 struct scale_factors *const sf = &cm->frame_refs[GOLDEN_FRAME - 1].sf;
1520 if (vp9_is_scaled(sf)) svc_force_zero_mode[GOLDEN_FRAME - 1] = 1;
1524 if (cpi->sf.short_circuit_low_temp_var) {
1525 force_skip_low_temp_var =
1526 get_force_skip_low_temp_var(&x->variance_low[0], mi_row, mi_col, bsize);
1527 // If force_skip_low_temp_var is set, and for short circuit mode = 1 and 3,
1528 // skip golden reference.
1529 if ((cpi->sf.short_circuit_low_temp_var == 1 ||
1530 cpi->sf.short_circuit_low_temp_var == 3) &&
1531 force_skip_low_temp_var) {
1532 usable_ref_frame = LAST_FRAME;
1536 if (!((cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) &&
1537 !svc_force_zero_mode[GOLDEN_FRAME - 1] && !force_skip_low_temp_var))
1538 use_golden_nonzeromv = 0;
1540 for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {
1541 if (!skip_ref_find_pred[ref_frame]) {
1542 find_predictors(cpi, x, ref_frame, frame_mv, const_motion,
1543 &ref_frame_skip_mask, flag_list, tile_data, mi_row,
1544 mi_col, yv12_mb, bsize, force_skip_low_temp_var);
1548 for (idx = 0; idx < RT_INTER_MODES; ++idx) {
1555 int this_early_term = 0;
1556 PREDICTION_MODE this_mode = ref_mode_set[idx].pred_mode;
1558 ref_frame = ref_mode_set[idx].ref_frame;
1561 this_mode = ref_mode_set_svc[idx].pred_mode;
1562 ref_frame = ref_mode_set_svc[idx].ref_frame;
1564 if (ref_frame > usable_ref_frame) continue;
1565 if (skip_ref_find_pred[ref_frame]) continue;
1567 if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
1568 this_mode != NEARESTMV) {
1572 if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode))) continue;
1574 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
1575 if (cpi->rc.is_src_frame_alt_ref &&
1576 (ref_frame != ALTREF_FRAME ||
1577 frame_mv[this_mode][ref_frame].as_int != 0))
1580 if (cpi->rc.alt_ref_gf_group &&
1581 cpi->rc.frames_since_golden > (cpi->rc.baseline_gf_interval >> 1) &&
1582 ref_frame == GOLDEN_FRAME &&
1583 frame_mv[this_mode][ref_frame].as_int != 0)
1586 if (cpi->rc.alt_ref_gf_group &&
1587 cpi->rc.frames_since_golden < (cpi->rc.baseline_gf_interval >> 1) &&
1588 ref_frame == ALTREF_FRAME &&
1589 frame_mv[this_mode][ref_frame].as_int != 0)
1593 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) continue;
1595 if (const_motion[ref_frame] && this_mode == NEARMV) continue;
1597 // Skip non-zeromv mode search for golden frame if force_skip_low_temp_var
1598 // is set. If nearestmv for golden frame is 0, zeromv mode will be skipped
1600 if (force_skip_low_temp_var && ref_frame == GOLDEN_FRAME &&
1601 frame_mv[this_mode][ref_frame].as_int != 0) {
1605 if ((cpi->sf.short_circuit_low_temp_var >= 2 ||
1606 (cpi->sf.short_circuit_low_temp_var == 1 && bsize == BLOCK_64X64)) &&
1607 force_skip_low_temp_var && ref_frame == LAST_FRAME &&
1608 this_mode == NEWMV) {
1613 if (svc_force_zero_mode[ref_frame - 1] &&
1614 frame_mv[this_mode][ref_frame].as_int != 0)
1618 if (sf->reference_masking &&
1619 !(frame_mv[this_mode][ref_frame].as_int == 0 &&
1620 ref_frame == LAST_FRAME)) {
1621 if (usable_ref_frame < ALTREF_FRAME) {
1622 if (!force_skip_low_temp_var && usable_ref_frame > LAST_FRAME) {
1623 i = (ref_frame == LAST_FRAME) ? GOLDEN_FRAME : LAST_FRAME;
1624 if ((cpi->ref_frame_flags & flag_list[i]))
1625 if (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[i] << 1))
1626 ref_frame_skip_mask |= (1 << ref_frame);
1628 } else if (!cpi->rc.is_src_frame_alt_ref &&
1629 !(frame_mv[this_mode][ref_frame].as_int == 0 &&
1630 ref_frame == ALTREF_FRAME)) {
1631 int ref1 = (ref_frame == GOLDEN_FRAME) ? LAST_FRAME : GOLDEN_FRAME;
1632 int ref2 = (ref_frame == ALTREF_FRAME) ? LAST_FRAME : ALTREF_FRAME;
1633 if (((cpi->ref_frame_flags & flag_list[ref1]) &&
1634 (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref1] << 1))) ||
1635 ((cpi->ref_frame_flags & flag_list[ref2]) &&
1636 (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref2] << 1))))
1637 ref_frame_skip_mask |= (1 << ref_frame);
1640 if (ref_frame_skip_mask & (1 << ref_frame)) continue;
1642 // Select prediction reference frames.
1643 for (i = 0; i < MAX_MB_PLANE; i++)
1644 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
1646 mi->ref_frame[0] = ref_frame;
1647 set_ref_ptrs(cm, xd, ref_frame, NONE);
1649 mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
1650 mode_rd_thresh = best_mode_skip_txfm ? rd_threshes[mode_index] << 1
1651 : rd_threshes[mode_index];
1653 // Increase mode_rd_thresh value for GOLDEN_FRAME for improved encoding
1654 // speed with little/no subjective quality loss.
1655 if (cpi->sf.bias_golden && ref_frame == GOLDEN_FRAME &&
1656 cpi->rc.frames_since_golden > 4)
1657 mode_rd_thresh = mode_rd_thresh << 3;
1659 if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
1660 #if CONFIG_MULTITHREAD
1661 tile_data->enc_row_mt_mutex,
1663 &rd_thresh_freq_fact[mode_index]))
1666 if (this_mode == NEWMV) {
1667 if (ref_frame > LAST_FRAME && !cpi->use_svc &&
1668 cpi->oxcf.rc_mode == VPX_CBR) {
1673 if (bsize < BLOCK_16X16) continue;
1675 tmp_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
1677 if (tmp_sad > x->pred_mv_sad[LAST_FRAME]) continue;
1678 if (tmp_sad + (num_pels_log2_lookup[bsize] << 4) > best_pred_sad)
1681 frame_mv[NEWMV][ref_frame].as_int = mi->mv[0].as_int;
1682 rate_mv = vp9_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv,
1683 &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1684 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1685 frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
1686 frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
1688 cpi->find_fractional_mv_step(
1689 x, &frame_mv[NEWMV][ref_frame].as_mv,
1690 &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1691 cpi->common.allow_high_precision_mv, x->errorperbit,
1692 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
1693 cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
1694 x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref_frame], NULL, 0,
1696 } else if (svc->use_base_mv && svc->spatial_layer_id) {
1697 if (frame_mv[NEWMV][ref_frame].as_int != INVALID_MV) {
1698 const int pre_stride = xd->plane[0].pre[0].stride;
1699 int base_mv_sad = INT_MAX;
1700 const float base_mv_bias = sf->base_mv_aggressive ? 1.5f : 1.0f;
1701 const uint8_t *const pre_buf =
1702 xd->plane[0].pre[0].buf +
1703 (frame_mv[NEWMV][ref_frame].as_mv.row >> 3) * pre_stride +
1704 (frame_mv[NEWMV][ref_frame].as_mv.col >> 3);
1705 base_mv_sad = cpi->fn_ptr[bsize].sdf(
1706 x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride);
1708 if (base_mv_sad < (int)(base_mv_bias * x->pred_mv_sad[ref_frame])) {
1709 // Base layer mv is good.
1710 if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1711 &frame_mv[NEWMV][ref_frame], &rate_mv,
1712 best_rdc.rdcost, 1)) {
1715 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1716 &frame_mv[NEWMV][ref_frame],
1717 &rate_mv, best_rdc.rdcost, 0)) {
1720 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1721 &frame_mv[NEWMV][ref_frame],
1722 &rate_mv, best_rdc.rdcost, 0)) {
1725 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1726 &frame_mv[NEWMV][ref_frame], &rate_mv,
1727 best_rdc.rdcost, 0)) {
1732 // If use_golden_nonzeromv is false, NEWMV mode is skipped for golden, no
1733 // need to compute best_pred_sad which is only used to skip golden NEWMV.
1734 if (use_golden_nonzeromv && this_mode == NEWMV && ref_frame == LAST_FRAME &&
1735 frame_mv[NEWMV][LAST_FRAME].as_int != INVALID_MV) {
1736 const int pre_stride = xd->plane[0].pre[0].stride;
1737 const uint8_t *const pre_buf =
1738 xd->plane[0].pre[0].buf +
1739 (frame_mv[NEWMV][LAST_FRAME].as_mv.row >> 3) * pre_stride +
1740 (frame_mv[NEWMV][LAST_FRAME].as_mv.col >> 3);
1741 best_pred_sad = cpi->fn_ptr[bsize].sdf(
1742 x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride);
1743 x->pred_mv_sad[LAST_FRAME] = best_pred_sad;
1746 if (this_mode != NEARESTMV &&
1747 frame_mv[this_mode][ref_frame].as_int ==
1748 frame_mv[NEARESTMV][ref_frame].as_int)
1751 mi->mode = this_mode;
1752 mi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
1754 // Search for the best prediction filter type, when the resulting
1755 // motion vector is at sub-pixel accuracy level for luma component, i.e.,
1756 // the last three bits are all zeros.
1757 if (reuse_inter_pred) {
1758 if (!this_mode_pred) {
1759 this_mode_pred = &tmp[3];
1761 this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
1762 pd->dst.buf = this_mode_pred->data;
1763 pd->dst.stride = bw;
1767 if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&
1768 pred_filter_search &&
1769 (ref_frame == LAST_FRAME ||
1770 (ref_frame == GOLDEN_FRAME &&
1771 (cpi->use_svc || cpi->oxcf.rc_mode == VPX_VBR))) &&
1772 (((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07) != 0)) {
1775 unsigned int pf_var[3];
1776 unsigned int pf_sse[3];
1777 TX_SIZE pf_tx_size[3];
1778 int64_t best_cost = INT64_MAX;
1779 INTERP_FILTER best_filter = SWITCHABLE, filter;
1780 PRED_BUFFER *current_pred = this_mode_pred;
1782 for (filter = EIGHTTAP; filter <= EIGHTTAP_SMOOTH; ++filter) {
1784 mi->interp_filter = filter;
1785 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1786 model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter], &pf_dist[filter],
1787 &pf_var[filter], &pf_sse[filter]);
1788 pf_rate[filter] += vp9_get_switchable_rate(cpi, xd);
1789 cost = RDCOST(x->rdmult, x->rddiv, pf_rate[filter], pf_dist[filter]);
1790 pf_tx_size[filter] = mi->tx_size;
1791 if (cost < best_cost) {
1792 best_filter = filter;
1794 skip_txfm = x->skip_txfm[0];
1796 if (reuse_inter_pred) {
1797 if (this_mode_pred != current_pred) {
1798 free_pred_buffer(this_mode_pred);
1799 this_mode_pred = current_pred;
1801 current_pred = &tmp[get_pred_buffer(tmp, 3)];
1802 pd->dst.buf = current_pred->data;
1803 pd->dst.stride = bw;
1808 if (reuse_inter_pred && this_mode_pred != current_pred)
1809 free_pred_buffer(current_pred);
1811 mi->interp_filter = best_filter;
1812 mi->tx_size = pf_tx_size[best_filter];
1813 this_rdc.rate = pf_rate[best_filter];
1814 this_rdc.dist = pf_dist[best_filter];
1815 var_y = pf_var[best_filter];
1816 sse_y = pf_sse[best_filter];
1817 x->skip_txfm[0] = skip_txfm;
1818 if (reuse_inter_pred) {
1819 pd->dst.buf = this_mode_pred->data;
1820 pd->dst.stride = this_mode_pred->stride;
1823 const int large_block = (x->sb_is_skin || cpi->oxcf.speed < 7)
1824 ? bsize > BLOCK_32X32
1825 : bsize >= BLOCK_32X32;
1826 mi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP : filter_ref;
1827 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1829 // For large partition blocks, extra testing is done.
1830 if (cpi->oxcf.rc_mode == VPX_CBR && large_block &&
1831 !cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) &&
1833 model_rd_for_sb_y_large(cpi, bsize, x, xd, &this_rdc.rate,
1834 &this_rdc.dist, &var_y, &sse_y, mi_row, mi_col,
1837 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
1842 if (!this_early_term) {
1843 this_sse = (int64_t)sse_y;
1844 block_yrd(cpi, x, &this_rdc, &is_skippable, &this_sse, bsize,
1845 VPXMIN(mi->tx_size, TX_16X16));
1846 x->skip_txfm[0] = is_skippable;
1848 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1850 if (RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist) <
1851 RDCOST(x->rdmult, x->rddiv, 0, this_sse)) {
1852 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
1854 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1855 this_rdc.dist = this_sse;
1856 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1860 if (cm->interp_filter == SWITCHABLE) {
1861 if ((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07)
1862 this_rdc.rate += vp9_get_switchable_rate(cpi, xd);
1865 this_rdc.rate += cm->interp_filter == SWITCHABLE
1866 ? vp9_get_switchable_rate(cpi, xd)
1868 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1871 if (x->color_sensitivity[0] || x->color_sensitivity[1]) {
1873 const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, &xd->plane[1]);
1874 if (x->color_sensitivity[0])
1875 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 1);
1876 if (x->color_sensitivity[1])
1877 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 2);
1878 model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &rdc_uv, &var_y, &sse_y, 1, 2);
1879 this_rdc.rate += rdc_uv.rate;
1880 this_rdc.dist += rdc_uv.dist;
1883 this_rdc.rate += rate_mv;
1884 this_rdc.rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
1885 [INTER_OFFSET(this_mode)];
1886 this_rdc.rate += ref_frame_cost[ref_frame];
1887 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1889 // Bias against NEWMV that is very different from its neighbors, and bias
1890 // to small motion-lastref for noisy input.
1891 if (cpi->oxcf.rc_mode == VPX_CBR && cpi->oxcf.speed >= 5 &&
1892 cpi->oxcf.content != VP9E_CONTENT_SCREEN) {
1893 vp9_NEWMV_diff_bias(&cpi->noise_estimate, xd, this_mode, &this_rdc, bsize,
1894 frame_mv[this_mode][ref_frame].as_mv.row,
1895 frame_mv[this_mode][ref_frame].as_mv.col,
1896 ref_frame == LAST_FRAME);
1899 // Skipping checking: test to see if this block can be reconstructed by
1901 if (cpi->allow_encode_breakout) {
1902 encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame, this_mode,
1903 var_y, sse_y, yv12_mb, &this_rdc.rate,
1906 this_rdc.rate += rate_mv;
1908 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1912 #if CONFIG_VP9_TEMPORAL_DENOISING
1913 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
1914 cpi->denoiser.denoising_level > kDenLowLow) {
1915 vp9_denoiser_update_frame_stats(mi, sse_y, this_mode, ctx);
1916 // Keep track of zero_last cost.
1917 if (ref_frame == LAST_FRAME && frame_mv[this_mode][ref_frame].as_int == 0)
1918 zero_last_cost_orig = this_rdc.rdcost;
1924 if (this_rdc.rdcost < best_rdc.rdcost || x->skip) {
1925 best_rdc = this_rdc;
1926 best_mode = this_mode;
1927 best_pred_filter = mi->interp_filter;
1928 best_tx_size = mi->tx_size;
1929 best_ref_frame = ref_frame;
1930 best_mode_skip_txfm = x->skip_txfm[0];
1931 best_early_term = this_early_term;
1933 if (reuse_inter_pred) {
1934 free_pred_buffer(best_pred);
1935 best_pred = this_mode_pred;
1938 if (reuse_inter_pred) free_pred_buffer(this_mode_pred);
1943 // If early termination flag is 1 and at least 2 modes are checked,
1944 // the mode search is terminated.
1945 if (best_early_term && idx > 0) {
1951 mi->mode = best_mode;
1952 mi->interp_filter = best_pred_filter;
1953 mi->tx_size = best_tx_size;
1954 mi->ref_frame[0] = best_ref_frame;
1955 mi->mv[0].as_int = frame_mv[best_mode][best_ref_frame].as_int;
1956 xd->mi[0]->bmi[0].as_mv[0].as_int = mi->mv[0].as_int;
1957 x->skip_txfm[0] = best_mode_skip_txfm;
1959 // For spatial enhancemanent layer: perform intra prediction only if base
1960 // layer is chosen as the reference. Always perform intra prediction if
1961 // LAST is the only reference or is_key_frame is set.
1962 if (cpi->svc.spatial_layer_id) {
1963 perform_intra_pred =
1964 cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame ||
1965 !(cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) ||
1966 (!cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
1967 svc_force_zero_mode[best_ref_frame - 1]);
1968 inter_mode_thresh = (inter_mode_thresh << 1) + inter_mode_thresh;
1970 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR &&
1971 cpi->rc.is_src_frame_alt_ref)
1972 perform_intra_pred = 0;
1973 // Perform intra prediction search, if the best SAD is above a certain
1975 if ((!force_skip_low_temp_var || bsize < BLOCK_32X32) && perform_intra_pred &&
1976 (best_rdc.rdcost == INT64_MAX ||
1977 (!x->skip && best_rdc.rdcost > inter_mode_thresh &&
1978 bsize <= cpi->sf.max_intra_bsize)) &&
1979 !x->skip_low_source_sad) {
1980 struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
1982 TX_SIZE best_intra_tx_size = TX_SIZES;
1983 TX_SIZE intra_tx_size =
1984 VPXMIN(max_txsize_lookup[bsize],
1985 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
1986 if (cpi->oxcf.content != VP9E_CONTENT_SCREEN && intra_tx_size > TX_16X16)
1987 intra_tx_size = TX_16X16;
1989 if (reuse_inter_pred && best_pred != NULL) {
1990 if (best_pred->data == orig_dst.buf) {
1991 this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
1992 #if CONFIG_VP9_HIGHBITDEPTH
1993 if (cm->use_highbitdepth)
1994 vpx_highbd_convolve_copy(best_pred->data, best_pred->stride,
1995 this_mode_pred->data, this_mode_pred->stride,
1996 NULL, 0, NULL, 0, bw, bh, xd->bd);
1998 vpx_convolve_copy(best_pred->data, best_pred->stride,
1999 this_mode_pred->data, this_mode_pred->stride, NULL,
2000 0, NULL, 0, bw, bh);
2002 vpx_convolve_copy(best_pred->data, best_pred->stride,
2003 this_mode_pred->data, this_mode_pred->stride, NULL, 0,
2005 #endif // CONFIG_VP9_HIGHBITDEPTH
2006 best_pred = this_mode_pred;
2011 for (i = 0; i < 4; ++i) {
2012 const PREDICTION_MODE this_mode = intra_mode_list[i];
2013 THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)];
2014 int mode_rd_thresh = rd_threshes[mode_index];
2015 if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
2016 this_mode != DC_PRED) {
2020 if (!((1 << this_mode) & cpi->sf.intra_y_mode_bsize_mask[bsize]))
2023 if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
2024 #if CONFIG_MULTITHREAD
2025 tile_data->enc_row_mt_mutex,
2027 &rd_thresh_freq_fact[mode_index]))
2030 mi->mode = this_mode;
2031 mi->ref_frame[0] = INTRA_FRAME;
2032 this_rdc.dist = this_rdc.rate = 0;
2033 args.mode = this_mode;
2035 args.rdc = &this_rdc;
2036 mi->tx_size = intra_tx_size;
2037 vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
2039 // Check skip cost here since skippable is not set for for uv, this
2040 // mirrors the behavior used by inter
2041 if (args.skippable) {
2042 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
2043 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
2045 x->skip_txfm[0] = SKIP_TXFM_NONE;
2046 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
2048 // Inter and intra RD will mismatch in scale for non-screen content.
2049 if (cpi->oxcf.content == VP9E_CONTENT_SCREEN) {
2050 if (x->color_sensitivity[0])
2051 vp9_foreach_transformed_block_in_plane(xd, bsize, 1,
2052 estimate_block_intra, &args);
2053 if (x->color_sensitivity[1])
2054 vp9_foreach_transformed_block_in_plane(xd, bsize, 2,
2055 estimate_block_intra, &args);
2057 this_rdc.rate += cpi->mbmode_cost[this_mode];
2058 this_rdc.rate += ref_frame_cost[INTRA_FRAME];
2059 this_rdc.rate += intra_cost_penalty;
2061 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2063 if (this_rdc.rdcost < best_rdc.rdcost) {
2064 best_rdc = this_rdc;
2065 best_mode = this_mode;
2066 best_intra_tx_size = mi->tx_size;
2067 best_ref_frame = INTRA_FRAME;
2068 mi->uv_mode = this_mode;
2069 mi->mv[0].as_int = INVALID_MV;
2070 best_mode_skip_txfm = x->skip_txfm[0];
2074 // Reset mb_mode_info to the best inter mode.
2075 if (best_ref_frame != INTRA_FRAME) {
2076 mi->tx_size = best_tx_size;
2078 mi->tx_size = best_intra_tx_size;
2083 mi->mode = best_mode;
2084 mi->ref_frame[0] = best_ref_frame;
2085 x->skip_txfm[0] = best_mode_skip_txfm;
2087 if (!is_inter_block(mi)) {
2088 mi->interp_filter = SWITCHABLE_FILTERS;
2091 if (reuse_inter_pred && best_pred != NULL) {
2092 if (best_pred->data != orig_dst.buf && is_inter_mode(mi->mode)) {
2093 #if CONFIG_VP9_HIGHBITDEPTH
2094 if (cm->use_highbitdepth)
2095 vpx_highbd_convolve_copy(best_pred->data, best_pred->stride,
2096 pd->dst.buf, pd->dst.stride, NULL, 0, NULL, 0,
2099 vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2100 pd->dst.stride, NULL, 0, NULL, 0, bw, bh);
2102 vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2103 pd->dst.stride, NULL, 0, NULL, 0, bw, bh);
2104 #endif // CONFIG_VP9_HIGHBITDEPTH
2108 #if CONFIG_VP9_TEMPORAL_DENOISING
2109 if (cpi->oxcf.noise_sensitivity > 0 && cpi->resize_pending == 0 &&
2110 denoise_svc(cpi) && cpi->denoiser.denoising_level > kDenLowLow &&
2111 cpi->denoiser.reset == 0) {
2112 VP9_DENOISER_DECISION decision = COPY_BLOCK;
2113 vp9_pickmode_ctx_den_update(&ctx_den, zero_last_cost_orig, ref_frame_cost,
2114 frame_mv, reuse_inter_pred, best_tx_size,
2115 best_mode, best_ref_frame, best_pred_filter,
2116 best_mode_skip_txfm);
2117 vp9_denoiser_denoise(cpi, x, mi_row, mi_col, bsize, ctx, &decision);
2118 recheck_zeromv_after_denoising(cpi, mi, x, xd, decision, &ctx_den, yv12_mb,
2119 &best_rdc, bsize, mi_row, mi_col);
2120 best_ref_frame = ctx_den.best_ref_frame;
2124 if (cpi->sf.adaptive_rd_thresh) {
2125 THR_MODES best_mode_idx = mode_idx[best_ref_frame][mode_offset(mi->mode)];
2127 if (best_ref_frame == INTRA_FRAME) {
2128 // Only consider the modes that are included in the intra_mode_list.
2129 int intra_modes = sizeof(intra_mode_list) / sizeof(PREDICTION_MODE);
2132 // TODO(yunqingwang): Check intra mode mask and only update freq_fact
2133 // for those valid modes.
2134 for (i = 0; i < intra_modes; i++) {
2135 update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2136 INTRA_FRAME, best_mode_idx, intra_mode_list[i]);
2139 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2140 PREDICTION_MODE this_mode;
2141 if (best_ref_frame != ref_frame) continue;
2142 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2143 update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2144 ref_frame, best_mode_idx, this_mode);
2150 *rd_cost = best_rdc;
2153 void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x, int mi_row,
2154 int mi_col, RD_COST *rd_cost, BLOCK_SIZE bsize,
2155 PICK_MODE_CONTEXT *ctx) {
2156 VP9_COMMON *const cm = &cpi->common;
2157 SPEED_FEATURES *const sf = &cpi->sf;
2158 MACROBLOCKD *const xd = &x->e_mbd;
2159 MODE_INFO *const mi = xd->mi[0];
2160 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2161 const struct segmentation *const seg = &cm->seg;
2162 MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE;
2163 MV_REFERENCE_FRAME best_ref_frame = NONE;
2164 unsigned char segment_id = mi->segment_id;
2165 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2166 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2168 int64_t best_rd = INT64_MAX;
2169 b_mode_info bsi[MAX_REF_FRAMES][4];
2170 int ref_frame_skip_mask = 0;
2171 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
2172 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
2175 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
2176 ctx->pred_pixel_ready = 0;
2178 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2179 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2181 x->pred_mv_sad[ref_frame] = INT_MAX;
2183 if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
2184 int_mv *const candidates = mbmi_ext->ref_mvs[ref_frame];
2185 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2186 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf,
2188 vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
2189 mbmi_ext->mode_context);
2191 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2192 &dummy_mv[0], &dummy_mv[1]);
2194 ref_frame_skip_mask |= (1 << ref_frame);
2198 mi->sb_type = bsize;
2199 mi->tx_size = TX_4X4;
2200 mi->uv_mode = DC_PRED;
2201 mi->ref_frame[0] = LAST_FRAME;
2202 mi->ref_frame[1] = NONE;
2204 cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
2206 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2207 int64_t this_rd = 0;
2210 if (ref_frame_skip_mask & (1 << ref_frame)) continue;
2212 #if CONFIG_BETTER_HW_COMPATIBILITY
2213 if ((bsize == BLOCK_8X4 || bsize == BLOCK_4X8) && ref_frame > INTRA_FRAME &&
2214 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2218 // TODO(jingning, agrange): Scaling reference frame not supported for
2219 // sub8x8 blocks. Is this supported now?
2220 if (ref_frame > INTRA_FRAME &&
2221 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2224 // If the segment reference frame feature is enabled....
2225 // then do nothing if the current ref frame is not allowed..
2226 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
2227 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
2230 mi->ref_frame[0] = ref_frame;
2232 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
2234 // Select prediction reference frames.
2235 for (plane = 0; plane < MAX_MB_PLANE; plane++)
2236 xd->plane[plane].pre[0] = yv12_mb[ref_frame][plane];
2238 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2239 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2240 int_mv b_mv[MB_MODE_COUNT];
2241 int64_t b_best_rd = INT64_MAX;
2242 const int i = idy * 2 + idx;
2243 PREDICTION_MODE this_mode;
2245 unsigned int var_y, sse_y;
2247 struct macroblock_plane *p = &x->plane[0];
2248 struct macroblockd_plane *pd = &xd->plane[0];
2250 const struct buf_2d orig_src = p->src;
2251 const struct buf_2d orig_dst = pd->dst;
2252 struct buf_2d orig_pre[2];
2253 memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));
2255 // set buffer pointers for sub8x8 motion search.
2257 &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
2259 &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
2262 .buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
2264 b_mv[ZEROMV].as_int = 0;
2265 b_mv[NEWMV].as_int = INVALID_MV;
2266 vp9_append_sub8x8_mvs_for_idx(cm, xd, i, 0, mi_row, mi_col,
2267 &b_mv[NEARESTMV], &b_mv[NEARMV],
2268 mbmi_ext->mode_context);
2270 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2272 xd->mi[0]->bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
2274 if (this_mode == NEWMV) {
2275 const int step_param = cpi->sf.mv.fullpel_search_step_param;
2279 const MvLimits tmp_mv_limits = x->mv_limits;
2280 uint32_t dummy_dist;
2283 mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3;
2284 mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3;
2286 mvp_full.row = xd->mi[0]->bmi[0].as_mv[0].as_mv.row >> 3;
2287 mvp_full.col = xd->mi[0]->bmi[0].as_mv[0].as_mv.col >> 3;
2290 vp9_set_mv_search_range(&x->mv_limits,
2291 &mbmi_ext->ref_mvs[0]->as_mv);
2293 vp9_full_pixel_search(
2294 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method,
2295 x->sadperbit4, cond_cost_list(cpi, cost_list),
2296 &mbmi_ext->ref_mvs[ref_frame][0].as_mv, &tmp_mv, INT_MAX, 0);
2298 x->mv_limits = tmp_mv_limits;
2300 // calculate the bit cost on motion vector
2301 mvp_full.row = tmp_mv.row * 8;
2302 mvp_full.col = tmp_mv.col * 8;
2304 b_rate += vp9_mv_bit_cost(
2305 &mvp_full, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2306 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2308 b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2309 [INTER_OFFSET(NEWMV)];
2310 if (RDCOST(x->rdmult, x->rddiv, b_rate, 0) > b_best_rd) continue;
2312 cpi->find_fractional_mv_step(
2313 x, &tmp_mv, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2314 cpi->common.allow_high_precision_mv, x->errorperbit,
2315 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
2316 cpi->sf.mv.subpel_iters_per_step,
2317 cond_cost_list(cpi, cost_list), x->nmvjointcost, x->mvcost,
2318 &dummy_dist, &x->pred_sse[ref_frame], NULL, 0, 0);
2320 xd->mi[0]->bmi[i].as_mv[0].as_mv = tmp_mv;
2322 b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2323 [INTER_OFFSET(this_mode)];
2326 #if CONFIG_VP9_HIGHBITDEPTH
2327 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2328 vp9_highbd_build_inter_predictor(
2329 pd->pre[0].buf, pd->pre[0].stride, pd->dst.buf, pd->dst.stride,
2330 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2331 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2332 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2333 mi_col * MI_SIZE + 4 * (i & 0x01),
2334 mi_row * MI_SIZE + 4 * (i >> 1), xd->bd);
2337 vp9_build_inter_predictor(
2338 pd->pre[0].buf, pd->pre[0].stride, pd->dst.buf, pd->dst.stride,
2339 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2340 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2341 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2342 mi_col * MI_SIZE + 4 * (i & 0x01),
2343 mi_row * MI_SIZE + 4 * (i >> 1));
2345 #if CONFIG_VP9_HIGHBITDEPTH
2349 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
2352 this_rdc.rate += b_rate;
2354 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2355 if (this_rdc.rdcost < b_best_rd) {
2356 b_best_rd = this_rdc.rdcost;
2357 bsi[ref_frame][i].as_mode = this_mode;
2358 bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0]->bmi[i].as_mv[0].as_mv;
2362 // restore source and prediction buffer pointers.
2364 pd->pre[0] = orig_pre[0];
2366 this_rd += b_best_rd;
2368 xd->mi[0]->bmi[i] = bsi[ref_frame][i];
2369 if (num_4x4_blocks_wide > 1) xd->mi[0]->bmi[i + 1] = xd->mi[0]->bmi[i];
2370 if (num_4x4_blocks_high > 1) xd->mi[0]->bmi[i + 2] = xd->mi[0]->bmi[i];
2372 } // loop through sub8x8 blocks
2374 if (this_rd < best_rd) {
2376 best_ref_frame = ref_frame;
2378 } // reference frames
2380 mi->tx_size = TX_4X4;
2381 mi->ref_frame[0] = best_ref_frame;
2382 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2383 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2384 const int block = idy * 2 + idx;
2385 xd->mi[0]->bmi[block] = bsi[best_ref_frame][block];
2386 if (num_4x4_blocks_wide > 1)
2387 xd->mi[0]->bmi[block + 1] = bsi[best_ref_frame][block];
2388 if (num_4x4_blocks_high > 1)
2389 xd->mi[0]->bmi[block + 2] = bsi[best_ref_frame][block];
2392 mi->mode = xd->mi[0]->bmi[3].as_mode;
2393 ctx->mic = *(xd->mi[0]);
2394 ctx->mbmi_ext = *x->mbmi_ext;
2395 ctx->skip_txfm[0] = SKIP_TXFM_NONE;
2397 // Dummy assignment for speed -5. No effect in speed -6.
2398 rd_cost->rdcost = best_rd;