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"
45 PRED_BUFFER *best_pred;
46 PREDICTION_MODE best_mode;
48 TX_SIZE best_intra_tx_size;
49 MV_REFERENCE_FRAME best_ref_frame;
50 MV_REFERENCE_FRAME best_second_ref_frame;
51 uint8_t best_mode_skip_txfm;
52 INTERP_FILTER best_pred_filter;
55 static const int pos_shift_16x16[4][4] = {
56 { 9, 10, 13, 14 }, { 11, 12, 15, 16 }, { 17, 18, 21, 22 }, { 19, 20, 23, 24 }
59 static int mv_refs_rt(VP9_COMP *cpi, const VP9_COMMON *cm, const MACROBLOCK *x,
60 const MACROBLOCKD *xd, const TileInfo *const tile,
61 MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
62 int_mv *mv_ref_list, int_mv *base_mv, int mi_row,
63 int mi_col, int use_base_mv) {
64 const int *ref_sign_bias = cm->ref_frame_sign_bias;
65 int i, refmv_count = 0;
67 const POSITION *const mv_ref_search = mv_ref_blocks[mi->sb_type];
69 int different_ref_found = 0;
70 int context_counter = 0;
73 // Blank the reference vector list
74 memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
76 // The nearest 2 blocks are treated differently
77 // if the size < 8x8 we get the mv from the bmi substructure,
78 // and we also need to keep a mode count.
79 for (i = 0; i < 2; ++i) {
80 const POSITION *const mv_ref = &mv_ref_search[i];
81 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
82 const MODE_INFO *const candidate_mi =
83 xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
84 // Keep counts for entropy encoding.
85 context_counter += mode_2_counter[candidate_mi->mode];
86 different_ref_found = 1;
88 if (candidate_mi->ref_frame[0] == ref_frame)
89 ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1),
90 refmv_count, mv_ref_list, Done);
96 // Check the rest of the neighbors in much the same way
97 // as before except we don't need to keep track of sub blocks or
99 for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
100 const POSITION *const mv_ref = &mv_ref_search[i];
101 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
102 const MODE_INFO *const candidate_mi =
103 xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
104 different_ref_found = 1;
106 if (candidate_mi->ref_frame[0] == ref_frame)
107 ADD_MV_REF_LIST(candidate_mi->mv[0], refmv_count, mv_ref_list, Done);
111 // Since we couldn't find 2 mvs from the same reference frame
112 // go back through the neighbors and find motion vectors from
113 // different reference frames.
114 if (different_ref_found && !refmv_count) {
115 for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
116 const POSITION *mv_ref = &mv_ref_search[i];
117 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
118 const MODE_INFO *const candidate_mi =
119 xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
121 // If the candidate is INTRA we don't want to consider its mv.
122 IF_DIFF_REF_FRAME_ADD_MV(candidate_mi, ref_frame, ref_sign_bias,
123 refmv_count, mv_ref_list, Done);
128 !cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
129 ref_frame == LAST_FRAME) {
130 // Get base layer mv.
133 ->mvs[(mi_col >> 1) + (mi_row >> 1) * (cm->mi_cols >> 1)];
134 if (candidate->mv[0].as_int != INVALID_MV) {
135 base_mv->as_mv.row = (candidate->mv[0].as_mv.row * 2);
136 base_mv->as_mv.col = (candidate->mv[0].as_mv.col * 2);
137 clamp_mv_ref(&base_mv->as_mv, xd);
139 base_mv->as_int = INVALID_MV;
145 x->mbmi_ext->mode_context[ref_frame] = counter_to_context[context_counter];
148 for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
149 clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
154 static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
155 BLOCK_SIZE bsize, int mi_row, int mi_col,
156 int_mv *tmp_mv, int *rate_mv,
157 int64_t best_rd_sofar, int use_base_mv) {
158 MACROBLOCKD *xd = &x->e_mbd;
159 MODE_INFO *mi = xd->mi[0];
160 struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } };
161 const int step_param = cpi->sf.mv.fullpel_search_step_param;
162 const int sadpb = x->sadperbit16;
164 const int ref = mi->ref_frame[0];
165 const MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
169 const MvLimits tmp_mv_limits = x->mv_limits;
172 int search_subpel = 1;
173 const YV12_BUFFER_CONFIG *scaled_ref_frame =
174 vp9_get_scaled_ref_frame(cpi, ref);
175 if (scaled_ref_frame) {
177 // Swap out the reference frame for a version that's been scaled to
178 // match the resolution of the current frame, allowing the existing
179 // motion search code to be used without additional modifications.
180 for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
181 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
183 vp9_set_mv_search_range(&x->mv_limits, &ref_mv);
185 // Limit motion vector for large lightning change.
186 if (cpi->oxcf.speed > 5 && x->lowvar_highsumdiff) {
187 x->mv_limits.col_min = VPXMAX(x->mv_limits.col_min, -10);
188 x->mv_limits.row_min = VPXMAX(x->mv_limits.row_min, -10);
189 x->mv_limits.col_max = VPXMIN(x->mv_limits.col_max, 10);
190 x->mv_limits.row_max = VPXMIN(x->mv_limits.row_max, 10);
193 assert(x->mv_best_ref_index[ref] <= 2);
194 if (x->mv_best_ref_index[ref] < 2)
195 mvp_full = x->mbmi_ext->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
197 mvp_full = x->pred_mv[ref];
205 center_mv = tmp_mv->as_mv;
207 if (x->sb_use_mv_part) {
208 tmp_mv->as_mv.row = x->sb_mvrow_part >> 3;
209 tmp_mv->as_mv.col = x->sb_mvcol_part >> 3;
211 vp9_full_pixel_search(
212 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method, sadpb,
213 cond_cost_list(cpi, cost_list), ¢er_mv, &tmp_mv->as_mv, INT_MAX, 0);
216 x->mv_limits = tmp_mv_limits;
218 // calculate the bit cost on motion vector
219 mvp_full.row = tmp_mv->as_mv.row * 8;
220 mvp_full.col = tmp_mv->as_mv.col * 8;
222 *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv, x->nmvjointcost, x->mvcost,
226 cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref]][INTER_OFFSET(NEWMV)];
228 !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) > best_rd_sofar);
230 // For SVC on non-reference frame, avoid subpel for (0, 0) motion.
231 if (cpi->use_svc && cpi->svc.non_reference_frame) {
232 if (mvp_full.row == 0 && mvp_full.col == 0) search_subpel = 0;
235 if (rv && search_subpel) {
236 int subpel_force_stop = cpi->sf.mv.subpel_force_stop;
237 if (use_base_mv && cpi->sf.base_mv_aggressive) subpel_force_stop = 2;
238 if (cpi->sf.mv.enable_adaptive_subpel_force_stop) {
239 int mv_thresh = cpi->sf.mv.adapt_subpel_force_stop.mv_thresh;
240 if (abs(tmp_mv->as_mv.row) >= mv_thresh ||
241 abs(tmp_mv->as_mv.col) >= mv_thresh)
242 subpel_force_stop = cpi->sf.mv.adapt_subpel_force_stop.force_stop_above;
244 subpel_force_stop = cpi->sf.mv.adapt_subpel_force_stop.force_stop_below;
246 cpi->find_fractional_mv_step(
247 x, &tmp_mv->as_mv, &ref_mv, cpi->common.allow_high_precision_mv,
248 x->errorperbit, &cpi->fn_ptr[bsize], subpel_force_stop,
249 cpi->sf.mv.subpel_search_level, cond_cost_list(cpi, cost_list),
250 x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, 0, 0,
251 cpi->sf.use_accurate_subpel_search);
252 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost,
253 x->mvcost, MV_COST_WEIGHT);
256 if (scaled_ref_frame) {
258 for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
263 static void block_variance(const uint8_t *src, int src_stride,
264 const uint8_t *ref, int ref_stride, int w, int h,
265 unsigned int *sse, int *sum, int block_size,
266 #if CONFIG_VP9_HIGHBITDEPTH
267 int use_highbitdepth, vpx_bit_depth_t bd,
269 uint32_t *sse8x8, int *sum8x8, uint32_t *var8x8) {
275 for (i = 0; i < h; i += block_size) {
276 for (j = 0; j < w; j += block_size) {
277 #if CONFIG_VP9_HIGHBITDEPTH
278 if (use_highbitdepth) {
281 vpx_highbd_8_get8x8var(src + src_stride * i + j, src_stride,
282 ref + ref_stride * i + j, ref_stride,
283 &sse8x8[k], &sum8x8[k]);
286 vpx_highbd_10_get8x8var(src + src_stride * i + j, src_stride,
287 ref + ref_stride * i + j, ref_stride,
288 &sse8x8[k], &sum8x8[k]);
291 vpx_highbd_12_get8x8var(src + src_stride * i + j, src_stride,
292 ref + ref_stride * i + j, ref_stride,
293 &sse8x8[k], &sum8x8[k]);
297 vpx_get8x8var(src + src_stride * i + j, src_stride,
298 ref + ref_stride * i + j, ref_stride, &sse8x8[k],
302 vpx_get8x8var(src + src_stride * i + j, src_stride,
303 ref + ref_stride * i + j, ref_stride, &sse8x8[k],
308 var8x8[k] = sse8x8[k] - (uint32_t)(((int64_t)sum8x8[k] * sum8x8[k]) >> 6);
314 static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
315 unsigned int *sse_i, int *sum_i,
316 unsigned int *var_o, unsigned int *sse_o,
318 const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
319 const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
320 const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
323 for (i = 0; i < nh; i += 2) {
324 for (j = 0; j < nw; j += 2) {
325 sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
326 sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
327 sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
328 sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
329 var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >>
330 (b_width_log2_lookup[unit_size] +
331 b_height_log2_lookup[unit_size] + 6));
337 // Adjust the ac_thr according to speed, width, height and normalized sum
338 static int ac_thr_factor(const int speed, const int width, const int height,
339 const int norm_sum) {
340 if (speed >= 8 && norm_sum < 5) {
341 if (width <= 640 && height <= 480)
349 static TX_SIZE calculate_tx_size(VP9_COMP *const cpi, BLOCK_SIZE bsize,
350 MACROBLOCKD *const xd, unsigned int var,
351 unsigned int sse, int64_t ac_thr) {
353 if (cpi->common.tx_mode == TX_MODE_SELECT) {
354 if (sse > (var << 2))
355 tx_size = VPXMIN(max_txsize_lookup[bsize],
356 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
360 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
361 cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
363 else if (tx_size > TX_16X16)
366 // For screen-content force 4X4 tx_size over 8X8, for large variance.
367 if (cpi->oxcf.content == VP9E_CONTENT_SCREEN && tx_size == TX_8X8 &&
368 bsize <= BLOCK_16X16 && var > (ac_thr << 6))
371 tx_size = VPXMIN(max_txsize_lookup[bsize],
372 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
378 static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize,
379 MACROBLOCK *x, MACROBLOCKD *xd,
380 int *out_rate_sum, int64_t *out_dist_sum,
381 unsigned int *var_y, unsigned int *sse_y,
382 int mi_row, int mi_col, int *early_term,
383 int *flag_preduv_computed) {
384 // Note our transform coeffs are 8 times an orthogonal transform.
385 // Hence quantizer step is also 8 times. To get effective quantizer
386 // we need to divide by 8 before sending to modeling function.
390 struct macroblock_plane *const p = &x->plane[0];
391 struct macroblockd_plane *const pd = &xd->plane[0];
392 const uint32_t dc_quant = pd->dequant[0];
393 const uint32_t ac_quant = pd->dequant[1];
394 int64_t dc_thr = dc_quant * dc_quant >> 6;
395 int64_t ac_thr = ac_quant * ac_quant >> 6;
400 const int bw = b_width_log2_lookup[bsize];
401 const int bh = b_height_log2_lookup[bsize];
402 const int num8x8 = 1 << (bw + bh - 2);
403 unsigned int sse8x8[64] = { 0 };
404 int sum8x8[64] = { 0 };
405 unsigned int var8x8[64] = { 0 };
408 #if CONFIG_VP9_HIGHBITDEPTH
409 const vpx_bit_depth_t bd = cpi->common.bit_depth;
411 // Calculate variance for whole partition, and also save 8x8 blocks' variance
412 // to be used in following transform skipping test.
413 block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
414 4 << bw, 4 << bh, &sse, &sum, 8,
415 #if CONFIG_VP9_HIGHBITDEPTH
416 cpi->common.use_highbitdepth, bd,
418 sse8x8, sum8x8, var8x8);
419 var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
424 #if CONFIG_VP9_TEMPORAL_DENOISING
425 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
427 ac_thr = vp9_scale_acskip_thresh(ac_thr, cpi->denoiser.denoising_level,
428 (abs(sum) >> (bw + bh)),
429 cpi->svc.temporal_layer_id);
431 ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
432 cpi->common.height, abs(sum) >> (bw + bh));
434 ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
435 cpi->common.height, abs(sum) >> (bw + bh));
438 tx_size = calculate_tx_size(cpi, bsize, xd, var, sse, ac_thr);
439 // The code below for setting skip flag assumes tranform size of at least 8x8,
440 // so force this lower limit on transform.
441 if (tx_size < TX_8X8) tx_size = TX_8X8;
442 xd->mi[0]->tx_size = tx_size;
444 if (cpi->oxcf.content == VP9E_CONTENT_SCREEN && x->zero_temp_sad_source &&
445 x->source_variance == 0)
446 dc_thr = dc_thr << 1;
448 // Evaluate if the partition block is a skippable block in Y plane.
450 unsigned int sse16x16[16] = { 0 };
451 int sum16x16[16] = { 0 };
452 unsigned int var16x16[16] = { 0 };
453 const int num16x16 = num8x8 >> 2;
455 unsigned int sse32x32[4] = { 0 };
456 int sum32x32[4] = { 0 };
457 unsigned int var32x32[4] = { 0 };
458 const int num32x32 = num8x8 >> 4;
462 const int num = (tx_size == TX_8X8)
464 : ((tx_size == TX_16X16) ? num16x16 : num32x32);
465 const unsigned int *sse_tx =
466 (tx_size == TX_8X8) ? sse8x8
467 : ((tx_size == TX_16X16) ? sse16x16 : sse32x32);
468 const unsigned int *var_tx =
469 (tx_size == TX_8X8) ? var8x8
470 : ((tx_size == TX_16X16) ? var16x16 : var32x32);
472 // Calculate variance if tx_size > TX_8X8
473 if (tx_size >= TX_16X16)
474 calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
476 if (tx_size == TX_32X32)
477 calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32,
481 x->skip_txfm[0] = SKIP_TXFM_NONE;
482 for (k = 0; k < num; k++)
483 // Check if all ac coefficients can be quantized to zero.
484 if (!(var_tx[k] < ac_thr || var == 0)) {
489 for (k = 0; k < num; k++)
490 // Check if dc coefficient can be quantized to zero.
491 if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
497 x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
499 if (dc_test) x->skip_txfm[0] = SKIP_TXFM_AC_DC;
500 } else if (dc_test) {
505 if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
506 int skip_uv[2] = { 0 };
507 unsigned int var_uv[2];
508 unsigned int sse_uv[2];
511 *out_dist_sum = sse << 4;
513 // Transform skipping test in UV planes.
514 for (i = 1; i <= 2; i++) {
515 if (cpi->oxcf.speed < 8 || x->color_sensitivity[i - 1]) {
516 struct macroblock_plane *const p = &x->plane[i];
517 struct macroblockd_plane *const pd = &xd->plane[i];
518 const TX_SIZE uv_tx_size = get_uv_tx_size(xd->mi[0], pd);
519 const BLOCK_SIZE unit_size = txsize_to_bsize[uv_tx_size];
520 const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, pd);
521 const int uv_bw = b_width_log2_lookup[uv_bsize];
522 const int uv_bh = b_height_log2_lookup[uv_bsize];
523 const int sf = (uv_bw - b_width_log2_lookup[unit_size]) +
524 (uv_bh - b_height_log2_lookup[unit_size]);
525 const uint32_t uv_dc_thr = pd->dequant[0] * pd->dequant[0] >> (6 - sf);
526 const uint32_t uv_ac_thr = pd->dequant[1] * pd->dequant[1] >> (6 - sf);
529 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, i);
530 flag_preduv_computed[i - 1] = 1;
531 var_uv[j] = cpi->fn_ptr[uv_bsize].vf(
532 p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, &sse_uv[j]);
534 if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
535 (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
544 // If the transform in YUV planes are skippable, the mode search checks
545 // fewer inter modes and doesn't check intra modes.
546 if (skip_uv[0] & skip_uv[1]) {
553 #if CONFIG_VP9_HIGHBITDEPTH
554 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
555 dc_quant >> (xd->bd - 5), &rate, &dist);
557 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
558 dc_quant >> 3, &rate, &dist);
559 #endif // CONFIG_VP9_HIGHBITDEPTH
563 *out_rate_sum = rate >> 1;
564 *out_dist_sum = dist << 3;
567 *out_dist_sum = (sse - var) << 4;
570 #if CONFIG_VP9_HIGHBITDEPTH
571 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
572 ac_quant >> (xd->bd - 5), &rate, &dist);
574 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
576 #endif // CONFIG_VP9_HIGHBITDEPTH
578 *out_rate_sum += rate;
579 *out_dist_sum += dist << 4;
582 static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
583 MACROBLOCKD *xd, int *out_rate_sum,
584 int64_t *out_dist_sum, unsigned int *var_y,
585 unsigned int *sse_y) {
586 // Note our transform coeffs are 8 times an orthogonal transform.
587 // Hence quantizer step is also 8 times. To get effective quantizer
588 // we need to divide by 8 before sending to modeling function.
592 struct macroblock_plane *const p = &x->plane[0];
593 struct macroblockd_plane *const pd = &xd->plane[0];
594 const int64_t dc_thr = p->quant_thred[0] >> 6;
595 const int64_t ac_thr = p->quant_thred[1] >> 6;
596 const uint32_t dc_quant = pd->dequant[0];
597 const uint32_t ac_quant = pd->dequant[1];
598 unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
599 pd->dst.buf, pd->dst.stride, &sse);
605 xd->mi[0]->tx_size = calculate_tx_size(cpi, bsize, xd, var, sse, ac_thr);
607 // Evaluate if the partition block is a skippable block in Y plane.
609 const BLOCK_SIZE unit_size = txsize_to_bsize[xd->mi[0]->tx_size];
610 const unsigned int num_blk_log2 =
611 (b_width_log2_lookup[bsize] - b_width_log2_lookup[unit_size]) +
612 (b_height_log2_lookup[bsize] - b_height_log2_lookup[unit_size]);
613 const unsigned int sse_tx = sse >> num_blk_log2;
614 const unsigned int var_tx = var >> num_blk_log2;
616 x->skip_txfm[0] = SKIP_TXFM_NONE;
617 // Check if all ac coefficients can be quantized to zero.
618 if (var_tx < ac_thr || var == 0) {
619 x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
620 // Check if dc coefficient can be quantized to zero.
621 if (sse_tx - var_tx < dc_thr || sse == var)
622 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
624 if (sse_tx - var_tx < dc_thr || sse == var) skip_dc = 1;
628 if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
630 *out_dist_sum = sse << 4;
635 #if CONFIG_VP9_HIGHBITDEPTH
636 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
637 dc_quant >> (xd->bd - 5), &rate, &dist);
639 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
640 dc_quant >> 3, &rate, &dist);
641 #endif // CONFIG_VP9_HIGHBITDEPTH
645 *out_rate_sum = rate >> 1;
646 *out_dist_sum = dist << 3;
649 *out_dist_sum = (sse - var) << 4;
652 #if CONFIG_VP9_HIGHBITDEPTH
653 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
654 ac_quant >> (xd->bd - 5), &rate, &dist);
656 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
658 #endif // CONFIG_VP9_HIGHBITDEPTH
660 *out_rate_sum += rate;
661 *out_dist_sum += dist << 4;
664 static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
665 int *skippable, int64_t *sse, BLOCK_SIZE bsize,
666 TX_SIZE tx_size, int rd_computed) {
667 MACROBLOCKD *xd = &x->e_mbd;
668 const struct macroblockd_plane *pd = &xd->plane[0];
669 struct macroblock_plane *const p = &x->plane[0];
670 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
671 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
672 const int step = 1 << (tx_size << 1);
673 const int block_step = (1 << tx_size);
675 const int max_blocks_wide =
676 num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5);
677 const int max_blocks_high =
678 num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5);
680 const int bw = 4 * num_4x4_w;
681 const int bh = 4 * num_4x4_h;
683 #if CONFIG_VP9_HIGHBITDEPTH
684 // TODO(jingning): Implement the high bit-depth Hadamard transforms and
685 // remove this check condition.
686 // TODO(marpan): Use this path (model_rd) for 8bit under certain conditions
687 // for now, as the vp9_quantize_fp below for highbitdepth build is slow.
689 (cpi->oxcf.speed > 5 && cpi->common.frame_type != KEY_FRAME &&
690 bsize < BLOCK_32X32)) {
691 unsigned int var_y, sse_y;
694 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist,
702 if (cpi->sf.use_simple_block_yrd && cpi->common.frame_type != KEY_FRAME &&
703 (bsize < BLOCK_32X32 ||
705 (bsize < BLOCK_32X32 || cpi->svc.temporal_layer_id > 0)))) {
706 unsigned int var_y, sse_y;
709 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist,
718 // The max tx_size passed in is TX_16X16.
719 assert(tx_size != TX_32X32);
721 vpx_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
722 pd->dst.buf, pd->dst.stride);
724 // Keep track of the row and column of the blocks we use so that we know
725 // if we are in the unrestricted motion border.
726 for (r = 0; r < max_blocks_high; r += block_step) {
727 for (c = 0; c < num_4x4_w; c += block_step) {
728 if (c < max_blocks_wide) {
729 const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
730 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
731 tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
732 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
733 uint16_t *const eob = &p->eobs[block];
734 const int diff_stride = bw;
735 const int16_t *src_diff;
736 src_diff = &p->src_diff[(r * diff_stride + c) << 2];
740 vpx_hadamard_16x16(src_diff, diff_stride, coeff);
741 vp9_quantize_fp(coeff, 256, x->skip_block, p->round_fp, p->quant_fp,
742 qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
746 vpx_hadamard_8x8(src_diff, diff_stride, coeff);
747 vp9_quantize_fp(coeff, 64, x->skip_block, p->round_fp, p->quant_fp,
748 qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
752 assert(tx_size == TX_4X4);
753 x->fwd_txfm4x4(src_diff, coeff, diff_stride);
754 vp9_quantize_fp(coeff, 16, x->skip_block, p->round_fp, p->quant_fp,
755 qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
759 *skippable &= (*eob == 0);
767 if (*sse < INT64_MAX) {
768 *sse = (*sse << 6) >> 2;
770 this_rdc->dist = *sse;
777 for (r = 0; r < max_blocks_high; r += block_step) {
778 for (c = 0; c < num_4x4_w; c += block_step) {
779 if (c < max_blocks_wide) {
780 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
781 tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
782 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
783 uint16_t *const eob = &p->eobs[block];
786 this_rdc->rate += (int)abs(qcoeff[0]);
788 this_rdc->rate += vpx_satd(qcoeff, step << 4);
790 this_rdc->dist += vp9_block_error_fp(coeff, dqcoeff, step << 4) >> 2;
796 // If skippable is set, rate gets clobbered later.
797 this_rdc->rate <<= (2 + VP9_PROB_COST_SHIFT);
798 this_rdc->rate += (eob_cost << VP9_PROB_COST_SHIFT);
801 static void model_rd_for_sb_uv(VP9_COMP *cpi, BLOCK_SIZE plane_bsize,
802 MACROBLOCK *x, MACROBLOCKD *xd,
803 RD_COST *this_rdc, unsigned int *var_y,
804 unsigned int *sse_y, int start_plane,
806 // Note our transform coeffs are 8 times an orthogonal transform.
807 // Hence quantizer step is also 8 times. To get effective quantizer
808 // we need to divide by 8 before sending to modeling function.
813 #if CONFIG_VP9_HIGHBITDEPTH
814 uint64_t tot_var = *var_y;
815 uint64_t tot_sse = *sse_y;
817 uint32_t tot_var = *var_y;
818 uint32_t tot_sse = *sse_y;
824 for (i = start_plane; i <= stop_plane; ++i) {
825 struct macroblock_plane *const p = &x->plane[i];
826 struct macroblockd_plane *const pd = &xd->plane[i];
827 const uint32_t dc_quant = pd->dequant[0];
828 const uint32_t ac_quant = pd->dequant[1];
829 const BLOCK_SIZE bs = plane_bsize;
831 if (!x->color_sensitivity[i - 1]) continue;
833 var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf,
834 pd->dst.stride, &sse);
839 #if CONFIG_VP9_HIGHBITDEPTH
840 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
841 dc_quant >> (xd->bd - 5), &rate, &dist);
843 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
844 dc_quant >> 3, &rate, &dist);
845 #endif // CONFIG_VP9_HIGHBITDEPTH
847 this_rdc->rate += rate >> 1;
848 this_rdc->dist += dist << 3;
850 #if CONFIG_VP9_HIGHBITDEPTH
851 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
852 ac_quant >> (xd->bd - 5), &rate, &dist);
854 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs], ac_quant >> 3,
856 #endif // CONFIG_VP9_HIGHBITDEPTH
858 this_rdc->rate += rate;
859 this_rdc->dist += dist << 4;
862 #if CONFIG_VP9_HIGHBITDEPTH
863 *var_y = tot_var > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_var;
864 *sse_y = tot_sse > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_sse;
871 static int get_pred_buffer(PRED_BUFFER *p, int len) {
874 for (i = 0; i < len; i++) {
883 static void free_pred_buffer(PRED_BUFFER *p) {
884 if (p != NULL) p->in_use = 0;
887 static void encode_breakout_test(
888 VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, int mi_row, int mi_col,
889 MV_REFERENCE_FRAME ref_frame, PREDICTION_MODE this_mode, unsigned int var_y,
890 unsigned int sse_y, struct buf_2d yv12_mb[][MAX_MB_PLANE], int *rate,
891 int64_t *dist, int *flag_preduv_computed) {
892 MACROBLOCKD *xd = &x->e_mbd;
893 MODE_INFO *const mi = xd->mi[0];
894 const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
895 unsigned int var = var_y, sse = sse_y;
896 // Skipping threshold for ac.
897 unsigned int thresh_ac;
898 // Skipping threshold for dc.
899 unsigned int thresh_dc;
901 if (cpi->use_svc && ref_frame == GOLDEN_FRAME) return;
902 if (mi->mv[0].as_mv.row > 64 || mi->mv[0].as_mv.row < -64 ||
903 mi->mv[0].as_mv.col > 64 || mi->mv[0].as_mv.col < -64)
905 if (x->encode_breakout > 0 && motion_low == 1) {
906 // Set a maximum for threshold to avoid big PSNR loss in low bit rate
907 // case. Use extreme low threshold for static frames to limit
909 const unsigned int max_thresh = 36000;
910 // The encode_breakout input
911 const unsigned int min_thresh =
912 VPXMIN(((unsigned int)x->encode_breakout << 4), max_thresh);
913 #if CONFIG_VP9_HIGHBITDEPTH
914 const int shift = (xd->bd << 1) - 16;
917 // Calculate threshold according to dequant value.
918 thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) >> 3;
919 #if CONFIG_VP9_HIGHBITDEPTH
920 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
921 thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift);
923 #endif // CONFIG_VP9_HIGHBITDEPTH
924 thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
926 // Adjust ac threshold according to partition size.
928 8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
930 thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
931 #if CONFIG_VP9_HIGHBITDEPTH
932 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
933 thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift);
935 #endif // CONFIG_VP9_HIGHBITDEPTH
941 // Y skipping condition checking for ac and dc.
942 if (var <= thresh_ac && (sse - var) <= thresh_dc) {
943 unsigned int sse_u, sse_v;
944 unsigned int var_u, var_v;
945 unsigned int thresh_ac_uv = thresh_ac;
946 unsigned int thresh_dc_uv = thresh_dc;
952 if (!flag_preduv_computed[0] || !flag_preduv_computed[1]) {
953 xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
954 xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
955 vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
958 var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf, x->plane[1].src.stride,
959 xd->plane[1].dst.buf,
960 xd->plane[1].dst.stride, &sse_u);
962 // U skipping condition checking
963 if (((var_u << 2) <= thresh_ac_uv) && (sse_u - var_u <= thresh_dc_uv)) {
964 var_v = cpi->fn_ptr[uv_size].vf(
965 x->plane[2].src.buf, x->plane[2].src.stride, xd->plane[2].dst.buf,
966 xd->plane[2].dst.stride, &sse_v);
968 // V skipping condition checking
969 if (((var_v << 2) <= thresh_ac_uv) && (sse_v - var_v <= thresh_dc_uv)) {
972 // The cost of skip bit needs to be added.
973 *rate = cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
974 [INTER_OFFSET(this_mode)];
976 // More on this part of rate
977 // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
979 // Scaling factor for SSE from spatial domain to frequency
980 // domain is 16. Adjust distortion accordingly.
981 // TODO(yunqingwang): In this function, only y-plane dist is
983 *dist = (sse << 4); // + ((sse_u + sse_v) << 4);
985 // *disable_skip = 1;
991 struct estimate_block_intra_args {
994 PREDICTION_MODE mode;
999 static void estimate_block_intra(int plane, int block, int row, int col,
1000 BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
1002 struct estimate_block_intra_args *const args = arg;
1003 VP9_COMP *const cpi = args->cpi;
1004 MACROBLOCK *const x = args->x;
1005 MACROBLOCKD *const xd = &x->e_mbd;
1006 struct macroblock_plane *const p = &x->plane[0];
1007 struct macroblockd_plane *const pd = &xd->plane[0];
1008 const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
1009 uint8_t *const src_buf_base = p->src.buf;
1010 uint8_t *const dst_buf_base = pd->dst.buf;
1011 const int src_stride = p->src.stride;
1012 const int dst_stride = pd->dst.stride;
1017 p->src.buf = &src_buf_base[4 * (row * src_stride + col)];
1018 pd->dst.buf = &dst_buf_base[4 * (row * dst_stride + col)];
1019 // Use source buffer as an approximation for the fully reconstructed buffer.
1020 vp9_predict_intra_block(xd, b_width_log2_lookup[plane_bsize], tx_size,
1021 args->mode, x->skip_encode ? p->src.buf : pd->dst.buf,
1022 x->skip_encode ? src_stride : dst_stride, pd->dst.buf,
1023 dst_stride, col, row, plane);
1026 int64_t this_sse = INT64_MAX;
1027 // TODO(jingning): This needs further refactoring.
1028 block_yrd(cpi, x, &this_rdc, &args->skippable, &this_sse, bsize_tx,
1029 VPXMIN(tx_size, TX_16X16), 0);
1031 unsigned int var = 0;
1032 unsigned int sse = 0;
1033 model_rd_for_sb_uv(cpi, plane_bsize, x, xd, &this_rdc, &var, &sse, plane,
1037 p->src.buf = src_buf_base;
1038 pd->dst.buf = dst_buf_base;
1039 args->rdc->rate += this_rdc.rate;
1040 args->rdc->dist += this_rdc.dist;
1043 static const THR_MODES mode_idx[MAX_REF_FRAMES][4] = {
1044 { THR_DC, THR_V_PRED, THR_H_PRED, THR_TM },
1045 { THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV },
1046 { THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG },
1047 { THR_NEARESTA, THR_NEARA, THR_ZEROA, THR_NEWA },
1050 static const PREDICTION_MODE intra_mode_list[] = { DC_PRED, V_PRED, H_PRED,
1053 static int mode_offset(const PREDICTION_MODE mode) {
1054 if (mode >= NEARESTMV) {
1055 return INTER_OFFSET(mode);
1058 case DC_PRED: return 0;
1059 case V_PRED: return 1;
1060 case H_PRED: return 2;
1061 case TM_PRED: return 3;
1067 static INLINE int rd_less_than_thresh_row_mt(int64_t best_rd, int thresh,
1068 const int *const thresh_fact) {
1069 int is_rd_less_than_thresh;
1070 is_rd_less_than_thresh =
1071 best_rd < ((int64_t)thresh * (*thresh_fact) >> 5) || thresh == INT_MAX;
1072 return is_rd_less_than_thresh;
1075 static INLINE void update_thresh_freq_fact_row_mt(
1076 VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance,
1077 int thresh_freq_fact_idx, MV_REFERENCE_FRAME ref_frame,
1078 THR_MODES best_mode_idx, PREDICTION_MODE mode) {
1079 THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1080 int freq_fact_idx = thresh_freq_fact_idx + thr_mode_idx;
1081 int *freq_fact = &tile_data->row_base_thresh_freq_fact[freq_fact_idx];
1082 if (thr_mode_idx == best_mode_idx)
1083 *freq_fact -= (*freq_fact >> 4);
1084 else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
1085 ref_frame == LAST_FRAME && source_variance < 5) {
1086 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
1088 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
1089 cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1093 static INLINE void update_thresh_freq_fact(
1094 VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance,
1095 BLOCK_SIZE bsize, MV_REFERENCE_FRAME ref_frame, THR_MODES best_mode_idx,
1096 PREDICTION_MODE mode) {
1097 THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1098 int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
1099 if (thr_mode_idx == best_mode_idx)
1100 *freq_fact -= (*freq_fact >> 4);
1101 else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
1102 ref_frame == LAST_FRAME && source_variance < 5) {
1103 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
1105 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
1106 cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1110 void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
1111 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1112 MACROBLOCKD *const xd = &x->e_mbd;
1113 MODE_INFO *const mi = xd->mi[0];
1114 RD_COST this_rdc, best_rdc;
1115 PREDICTION_MODE this_mode;
1116 struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
1117 const TX_SIZE intra_tx_size =
1118 VPXMIN(max_txsize_lookup[bsize],
1119 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
1120 MODE_INFO *const mic = xd->mi[0];
1122 const MODE_INFO *above_mi = xd->above_mi;
1123 const MODE_INFO *left_mi = xd->left_mi;
1124 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1125 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1126 bmode_costs = cpi->y_mode_costs[A][L];
1129 vp9_rd_cost_reset(&best_rdc);
1130 vp9_rd_cost_reset(&this_rdc);
1132 mi->ref_frame[0] = INTRA_FRAME;
1133 // Initialize interp_filter here so we do not have to check for inter block
1134 // modes in get_pred_context_switchable_interp()
1135 mi->interp_filter = SWITCHABLE_FILTERS;
1137 mi->mv[0].as_int = INVALID_MV;
1138 mi->uv_mode = DC_PRED;
1139 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
1141 // Change the limit of this loop to add other intra prediction
1143 for (this_mode = DC_PRED; this_mode <= H_PRED; ++this_mode) {
1144 this_rdc.dist = this_rdc.rate = 0;
1145 args.mode = this_mode;
1147 args.rdc = &this_rdc;
1148 mi->tx_size = intra_tx_size;
1149 vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
1151 if (args.skippable) {
1152 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1153 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
1155 x->skip_txfm[0] = SKIP_TXFM_NONE;
1156 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
1158 this_rdc.rate += bmode_costs[this_mode];
1159 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1161 if (this_rdc.rdcost < best_rdc.rdcost) {
1162 best_rdc = this_rdc;
1163 mi->mode = this_mode;
1167 *rd_cost = best_rdc;
1170 static void init_ref_frame_cost(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1171 int ref_frame_cost[MAX_REF_FRAMES]) {
1172 vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
1173 vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1174 vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1176 ref_frame_cost[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1177 ref_frame_cost[LAST_FRAME] = ref_frame_cost[GOLDEN_FRAME] =
1178 ref_frame_cost[ALTREF_FRAME] = vp9_cost_bit(intra_inter_p, 1);
1180 ref_frame_cost[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1181 ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1182 ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1183 ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1184 ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1188 MV_REFERENCE_FRAME ref_frame;
1189 PREDICTION_MODE pred_mode;
1192 #define RT_INTER_MODES 12
1193 static const REF_MODE ref_mode_set[RT_INTER_MODES] = {
1194 { LAST_FRAME, ZEROMV }, { LAST_FRAME, NEARESTMV },
1195 { GOLDEN_FRAME, ZEROMV }, { LAST_FRAME, NEARMV },
1196 { LAST_FRAME, NEWMV }, { GOLDEN_FRAME, NEARESTMV },
1197 { GOLDEN_FRAME, NEARMV }, { GOLDEN_FRAME, NEWMV },
1198 { ALTREF_FRAME, ZEROMV }, { ALTREF_FRAME, NEARESTMV },
1199 { ALTREF_FRAME, NEARMV }, { ALTREF_FRAME, NEWMV }
1202 #define RT_INTER_MODES_SVC 8
1203 static const REF_MODE ref_mode_set_svc[RT_INTER_MODES_SVC] = {
1204 { LAST_FRAME, ZEROMV }, { LAST_FRAME, NEARESTMV },
1205 { LAST_FRAME, NEARMV }, { GOLDEN_FRAME, ZEROMV },
1206 { GOLDEN_FRAME, NEARESTMV }, { GOLDEN_FRAME, NEARMV },
1207 { LAST_FRAME, NEWMV }, { GOLDEN_FRAME, NEWMV }
1210 static INLINE void find_predictors(
1211 VP9_COMP *cpi, MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame,
1212 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1213 int const_motion[MAX_REF_FRAMES], int *ref_frame_skip_mask,
1214 const int flag_list[4], TileDataEnc *tile_data, int mi_row, int mi_col,
1215 struct buf_2d yv12_mb[4][MAX_MB_PLANE], BLOCK_SIZE bsize,
1216 int force_skip_low_temp_var, int comp_pred_allowed) {
1217 VP9_COMMON *const cm = &cpi->common;
1218 MACROBLOCKD *const xd = &x->e_mbd;
1219 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1220 TileInfo *const tile_info = &tile_data->tile_info;
1221 // TODO(jingning) placeholder for inter-frame non-RD mode decision.
1222 x->pred_mv_sad[ref_frame] = INT_MAX;
1223 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
1224 frame_mv[ZEROMV][ref_frame].as_int = 0;
1225 // this needs various further optimizations. to be continued..
1226 if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
1227 int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
1228 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
1229 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
1230 if (cm->use_prev_frame_mvs || comp_pred_allowed) {
1231 vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
1232 x->mbmi_ext->mode_context);
1234 const_motion[ref_frame] =
1235 mv_refs_rt(cpi, cm, x, xd, tile_info, xd->mi[0], ref_frame,
1236 candidates, &frame_mv[NEWMV][ref_frame], mi_row, mi_col,
1237 (int)(cpi->svc.use_base_mv && cpi->svc.spatial_layer_id));
1239 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1240 &frame_mv[NEARESTMV][ref_frame],
1241 &frame_mv[NEARMV][ref_frame]);
1242 // Early exit for golden frame if force_skip_low_temp_var is set.
1243 if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8 &&
1244 !(force_skip_low_temp_var && ref_frame == GOLDEN_FRAME)) {
1245 vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
1249 *ref_frame_skip_mask |= (1 << ref_frame);
1253 static void vp9_NEWMV_diff_bias(const NOISE_ESTIMATE *ne, MACROBLOCKD *xd,
1254 PREDICTION_MODE this_mode, RD_COST *this_rdc,
1255 BLOCK_SIZE bsize, int mv_row, int mv_col,
1256 int is_last_frame, int lowvar_highsumdiff,
1258 // Bias against MVs associated with NEWMV mode that are very different from
1259 // top/left neighbors.
1260 if (this_mode == NEWMV) {
1261 int al_mv_average_row;
1262 int al_mv_average_col;
1263 int left_row, left_col;
1264 int row_diff, col_diff;
1265 int above_mv_valid = 0;
1266 int left_mv_valid = 0;
1271 above_mv_valid = xd->above_mi->mv[0].as_int != INVALID_MV;
1272 above_row = xd->above_mi->mv[0].as_mv.row;
1273 above_col = xd->above_mi->mv[0].as_mv.col;
1276 left_mv_valid = xd->left_mi->mv[0].as_int != INVALID_MV;
1277 left_row = xd->left_mi->mv[0].as_mv.row;
1278 left_col = xd->left_mi->mv[0].as_mv.col;
1280 if (above_mv_valid && left_mv_valid) {
1281 al_mv_average_row = (above_row + left_row + 1) >> 1;
1282 al_mv_average_col = (above_col + left_col + 1) >> 1;
1283 } else if (above_mv_valid) {
1284 al_mv_average_row = above_row;
1285 al_mv_average_col = above_col;
1286 } else if (left_mv_valid) {
1287 al_mv_average_row = left_row;
1288 al_mv_average_col = left_col;
1290 al_mv_average_row = al_mv_average_col = 0;
1292 row_diff = (al_mv_average_row - mv_row);
1293 col_diff = (al_mv_average_col - mv_col);
1294 if (row_diff > 48 || row_diff < -48 || col_diff > 48 || col_diff < -48) {
1295 if (bsize > BLOCK_32X32)
1296 this_rdc->rdcost = this_rdc->rdcost << 1;
1298 this_rdc->rdcost = 3 * this_rdc->rdcost >> 1;
1301 // If noise estimation is enabled, and estimated level is above threshold,
1302 // add a bias to LAST reference with small motion, for large blocks.
1303 if (ne->enabled && ne->level >= kMedium && bsize >= BLOCK_32X32 &&
1304 is_last_frame && mv_row < 8 && mv_row > -8 && mv_col < 8 && mv_col > -8)
1305 this_rdc->rdcost = 7 * (this_rdc->rdcost >> 3);
1306 else if (lowvar_highsumdiff && !is_skin && bsize >= BLOCK_16X16 &&
1307 is_last_frame && mv_row < 16 && mv_row > -16 && mv_col < 16 &&
1309 this_rdc->rdcost = 7 * (this_rdc->rdcost >> 3);
1312 #if CONFIG_VP9_TEMPORAL_DENOISING
1313 static void vp9_pickmode_ctx_den_update(
1314 VP9_PICKMODE_CTX_DEN *ctx_den, int64_t zero_last_cost_orig,
1315 int ref_frame_cost[MAX_REF_FRAMES],
1316 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int reuse_inter_pred,
1317 BEST_PICKMODE *bp) {
1318 ctx_den->zero_last_cost_orig = zero_last_cost_orig;
1319 ctx_den->ref_frame_cost = ref_frame_cost;
1320 ctx_den->frame_mv = frame_mv;
1321 ctx_den->reuse_inter_pred = reuse_inter_pred;
1322 ctx_den->best_tx_size = bp->best_tx_size;
1323 ctx_den->best_mode = bp->best_mode;
1324 ctx_den->best_ref_frame = bp->best_ref_frame;
1325 ctx_den->best_pred_filter = bp->best_pred_filter;
1326 ctx_den->best_mode_skip_txfm = bp->best_mode_skip_txfm;
1329 static void recheck_zeromv_after_denoising(
1330 VP9_COMP *cpi, MODE_INFO *const mi, MACROBLOCK *x, MACROBLOCKD *const xd,
1331 VP9_DENOISER_DECISION decision, VP9_PICKMODE_CTX_DEN *ctx_den,
1332 struct buf_2d yv12_mb[4][MAX_MB_PLANE], RD_COST *best_rdc, BLOCK_SIZE bsize,
1333 int mi_row, int mi_col) {
1334 // If INTRA or GOLDEN reference was selected, re-evaluate ZEROMV on
1335 // denoised result. Only do this under noise conditions, and if rdcost of
1336 // ZEROMV onoriginal source is not significantly higher than rdcost of best
1338 if (cpi->noise_estimate.enabled && cpi->noise_estimate.level > kLow &&
1339 ctx_den->zero_last_cost_orig < (best_rdc->rdcost << 3) &&
1340 ((ctx_den->best_ref_frame == INTRA_FRAME && decision >= FILTER_BLOCK) ||
1341 (ctx_den->best_ref_frame == GOLDEN_FRAME &&
1342 cpi->svc.number_spatial_layers == 1 &&
1343 decision == FILTER_ZEROMV_BLOCK))) {
1344 // Check if we should pick ZEROMV on denoised signal.
1347 uint32_t var_y = UINT_MAX;
1348 uint32_t sse_y = UINT_MAX;
1351 mi->ref_frame[0] = LAST_FRAME;
1352 mi->ref_frame[1] = NONE;
1353 mi->mv[0].as_int = 0;
1354 mi->interp_filter = EIGHTTAP;
1355 if (cpi->sf.default_interp_filter == BILINEAR) mi->interp_filter = BILINEAR;
1356 xd->plane[0].pre[0] = yv12_mb[LAST_FRAME][0];
1357 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1358 model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
1359 this_rdc.rate = rate + ctx_den->ref_frame_cost[LAST_FRAME] +
1360 cpi->inter_mode_cost[x->mbmi_ext->mode_context[LAST_FRAME]]
1361 [INTER_OFFSET(ZEROMV)];
1362 this_rdc.dist = dist;
1363 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, rate, dist);
1364 // Don't switch to ZEROMV if the rdcost for ZEROMV on denoised source
1365 // is higher than best_ref mode (on original source).
1366 if (this_rdc.rdcost > best_rdc->rdcost) {
1367 this_rdc = *best_rdc;
1368 mi->mode = ctx_den->best_mode;
1369 mi->ref_frame[0] = ctx_den->best_ref_frame;
1370 mi->interp_filter = ctx_den->best_pred_filter;
1371 if (ctx_den->best_ref_frame == INTRA_FRAME) {
1372 mi->mv[0].as_int = INVALID_MV;
1373 mi->interp_filter = SWITCHABLE_FILTERS;
1374 } else if (ctx_den->best_ref_frame == GOLDEN_FRAME) {
1376 ctx_den->frame_mv[ctx_den->best_mode][ctx_den->best_ref_frame]
1378 if (ctx_den->reuse_inter_pred) {
1379 xd->plane[0].pre[0] = yv12_mb[GOLDEN_FRAME][0];
1380 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1383 mi->tx_size = ctx_den->best_tx_size;
1384 x->skip_txfm[0] = ctx_den->best_mode_skip_txfm;
1386 ctx_den->best_ref_frame = LAST_FRAME;
1387 *best_rdc = this_rdc;
1391 #endif // CONFIG_VP9_TEMPORAL_DENOISING
1393 static INLINE int get_force_skip_low_temp_var(uint8_t *variance_low, int mi_row,
1394 int mi_col, BLOCK_SIZE bsize) {
1395 const int i = (mi_row & 0x7) >> 1;
1396 const int j = (mi_col & 0x7) >> 1;
1397 int force_skip_low_temp_var = 0;
1398 // Set force_skip_low_temp_var based on the block size and block offset.
1399 if (bsize == BLOCK_64X64) {
1400 force_skip_low_temp_var = variance_low[0];
1401 } else if (bsize == BLOCK_64X32) {
1402 if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1403 force_skip_low_temp_var = variance_low[1];
1404 } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1405 force_skip_low_temp_var = variance_low[2];
1407 } else if (bsize == BLOCK_32X64) {
1408 if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1409 force_skip_low_temp_var = variance_low[3];
1410 } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1411 force_skip_low_temp_var = variance_low[4];
1413 } else if (bsize == BLOCK_32X32) {
1414 if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1415 force_skip_low_temp_var = variance_low[5];
1416 } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1417 force_skip_low_temp_var = variance_low[6];
1418 } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1419 force_skip_low_temp_var = variance_low[7];
1420 } else if ((mi_col & 0x7) && (mi_row & 0x7)) {
1421 force_skip_low_temp_var = variance_low[8];
1423 } else if (bsize == BLOCK_16X16) {
1424 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]];
1425 } else if (bsize == BLOCK_32X16) {
1426 // The col shift index for the second 16x16 block.
1427 const int j2 = ((mi_col + 2) & 0x7) >> 1;
1428 // Only if each 16x16 block inside has low temporal variance.
1429 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1430 variance_low[pos_shift_16x16[i][j2]];
1431 } else if (bsize == BLOCK_16X32) {
1432 // The row shift index for the second 16x16 block.
1433 const int i2 = ((mi_row + 2) & 0x7) >> 1;
1434 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1435 variance_low[pos_shift_16x16[i2][j]];
1437 return force_skip_low_temp_var;
1440 static void search_filter_ref(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
1441 int mi_row, int mi_col, PRED_BUFFER *tmp,
1442 BLOCK_SIZE bsize, int reuse_inter_pred,
1443 PRED_BUFFER **this_mode_pred, unsigned int *var_y,
1444 unsigned int *sse_y) {
1445 MACROBLOCKD *const xd = &x->e_mbd;
1446 MODE_INFO *const mi = xd->mi[0];
1447 struct macroblockd_plane *const pd = &xd->plane[0];
1448 const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
1450 int pf_rate[3] = { 0 };
1451 int64_t pf_dist[3] = { 0 };
1452 int curr_rate[3] = { 0 };
1453 unsigned int pf_var[3] = { 0 };
1454 unsigned int pf_sse[3] = { 0 };
1455 TX_SIZE pf_tx_size[3] = { 0 };
1456 int64_t best_cost = INT64_MAX;
1457 INTERP_FILTER best_filter = SWITCHABLE, filter;
1458 PRED_BUFFER *current_pred = *this_mode_pred;
1459 uint8_t skip_txfm = SKIP_TXFM_NONE;
1461 for (filter = EIGHTTAP; filter <= EIGHTTAP_SMOOTH; ++filter) {
1463 mi->interp_filter = filter;
1464 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1465 model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter], &pf_dist[filter],
1466 &pf_var[filter], &pf_sse[filter]);
1467 curr_rate[filter] = pf_rate[filter];
1468 pf_rate[filter] += vp9_get_switchable_rate(cpi, xd);
1469 cost = RDCOST(x->rdmult, x->rddiv, pf_rate[filter], pf_dist[filter]);
1470 pf_tx_size[filter] = mi->tx_size;
1471 if (cost < best_cost) {
1472 best_filter = filter;
1474 skip_txfm = x->skip_txfm[0];
1476 if (reuse_inter_pred) {
1477 if (*this_mode_pred != current_pred) {
1478 free_pred_buffer(*this_mode_pred);
1479 *this_mode_pred = current_pred;
1481 current_pred = &tmp[get_pred_buffer(tmp, 3)];
1482 pd->dst.buf = current_pred->data;
1483 pd->dst.stride = bw;
1488 if (reuse_inter_pred && *this_mode_pred != current_pred)
1489 free_pred_buffer(current_pred);
1491 mi->interp_filter = best_filter;
1492 mi->tx_size = pf_tx_size[best_filter];
1493 this_rdc->rate = curr_rate[best_filter];
1494 this_rdc->dist = pf_dist[best_filter];
1495 *var_y = pf_var[best_filter];
1496 *sse_y = pf_sse[best_filter];
1497 x->skip_txfm[0] = skip_txfm;
1498 if (reuse_inter_pred) {
1499 pd->dst.buf = (*this_mode_pred)->data;
1500 pd->dst.stride = (*this_mode_pred)->stride;
1504 static int search_new_mv(VP9_COMP *cpi, MACROBLOCK *x,
1505 int_mv frame_mv[][MAX_REF_FRAMES],
1506 MV_REFERENCE_FRAME ref_frame, int gf_temporal_ref,
1507 BLOCK_SIZE bsize, int mi_row, int mi_col,
1508 int best_pred_sad, int *rate_mv,
1509 unsigned int best_sse_sofar, RD_COST *best_rdc) {
1510 SVC *const svc = &cpi->svc;
1511 MACROBLOCKD *const xd = &x->e_mbd;
1512 MODE_INFO *const mi = xd->mi[0];
1513 SPEED_FEATURES *const sf = &cpi->sf;
1515 if (ref_frame > LAST_FRAME && gf_temporal_ref &&
1516 cpi->oxcf.rc_mode == VPX_CBR) {
1519 int cost_list[5] = { INT_MAX, INT_MAX, INT_MAX, INT_MAX, INT_MAX };
1521 if (bsize < BLOCK_16X16) return -1;
1523 tmp_sad = vp9_int_pro_motion_estimation(
1524 cpi, x, bsize, mi_row, mi_col,
1525 &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv);
1527 if (tmp_sad > x->pred_mv_sad[LAST_FRAME]) return -1;
1528 if (tmp_sad + (num_pels_log2_lookup[bsize] << 4) > best_pred_sad) return -1;
1530 frame_mv[NEWMV][ref_frame].as_int = mi->mv[0].as_int;
1531 *rate_mv = vp9_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv,
1532 &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1533 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1534 frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
1535 frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
1537 cpi->find_fractional_mv_step(
1538 x, &frame_mv[NEWMV][ref_frame].as_mv,
1539 &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1540 cpi->common.allow_high_precision_mv, x->errorperbit,
1541 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
1542 cpi->sf.mv.subpel_search_level, cond_cost_list(cpi, cost_list),
1543 x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref_frame], NULL, 0, 0,
1544 cpi->sf.use_accurate_subpel_search);
1545 } else if (svc->use_base_mv && svc->spatial_layer_id) {
1546 if (frame_mv[NEWMV][ref_frame].as_int != INVALID_MV) {
1547 const int pre_stride = xd->plane[0].pre[0].stride;
1548 unsigned int base_mv_sse = UINT_MAX;
1549 int scale = (cpi->rc.avg_frame_low_motion > 60) ? 2 : 4;
1550 const uint8_t *const pre_buf =
1551 xd->plane[0].pre[0].buf +
1552 (frame_mv[NEWMV][ref_frame].as_mv.row >> 3) * pre_stride +
1553 (frame_mv[NEWMV][ref_frame].as_mv.col >> 3);
1554 cpi->fn_ptr[bsize].vf(x->plane[0].src.buf, x->plane[0].src.stride,
1555 pre_buf, pre_stride, &base_mv_sse);
1557 // Exit NEWMV search if base_mv is (0,0) && bsize < BLOCK_16x16,
1558 // for SVC encoding.
1559 if (cpi->use_svc && svc->use_base_mv && bsize < BLOCK_16X16 &&
1560 frame_mv[NEWMV][ref_frame].as_mv.row == 0 &&
1561 frame_mv[NEWMV][ref_frame].as_mv.col == 0)
1564 // Exit NEWMV search if base_mv_sse is large.
1565 if (sf->base_mv_aggressive && base_mv_sse > (best_sse_sofar << scale))
1567 if (base_mv_sse < (best_sse_sofar << 1)) {
1568 // Base layer mv is good.
1569 // Exit NEWMV search if the base_mv is (0, 0) and sse is low, since
1570 // (0, 0) mode is already tested.
1571 unsigned int base_mv_sse_normalized =
1573 (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
1574 if (sf->base_mv_aggressive && base_mv_sse <= best_sse_sofar &&
1575 base_mv_sse_normalized < 400 &&
1576 frame_mv[NEWMV][ref_frame].as_mv.row == 0 &&
1577 frame_mv[NEWMV][ref_frame].as_mv.col == 0)
1579 if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1580 &frame_mv[NEWMV][ref_frame], rate_mv,
1581 best_rdc->rdcost, 1)) {
1584 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1585 &frame_mv[NEWMV][ref_frame], rate_mv,
1586 best_rdc->rdcost, 0)) {
1589 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1590 &frame_mv[NEWMV][ref_frame], rate_mv,
1591 best_rdc->rdcost, 0)) {
1594 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1595 &frame_mv[NEWMV][ref_frame], rate_mv,
1596 best_rdc->rdcost, 0)) {
1603 static INLINE void init_best_pickmode(BEST_PICKMODE *bp) {
1604 bp->best_mode = ZEROMV;
1605 bp->best_ref_frame = LAST_FRAME;
1606 bp->best_tx_size = TX_SIZES;
1607 bp->best_intra_tx_size = TX_SIZES;
1608 bp->best_pred_filter = EIGHTTAP;
1609 bp->best_mode_skip_txfm = SKIP_TXFM_NONE;
1610 bp->best_second_ref_frame = NONE;
1611 bp->best_pred = NULL;
1614 void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data,
1615 int mi_row, int mi_col, RD_COST *rd_cost,
1616 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1617 VP9_COMMON *const cm = &cpi->common;
1618 SPEED_FEATURES *const sf = &cpi->sf;
1619 SVC *const svc = &cpi->svc;
1620 MACROBLOCKD *const xd = &x->e_mbd;
1621 MODE_INFO *const mi = xd->mi[0];
1622 struct macroblockd_plane *const pd = &xd->plane[0];
1624 BEST_PICKMODE best_pickmode;
1626 MV_REFERENCE_FRAME ref_frame;
1627 MV_REFERENCE_FRAME usable_ref_frame, second_ref_frame;
1628 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1629 uint8_t mode_checked[MB_MODE_COUNT][MAX_REF_FRAMES];
1630 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1631 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1633 RD_COST this_rdc, best_rdc;
1634 // var_y and sse_y are saved to be used in skipping checking
1635 unsigned int var_y = UINT_MAX;
1636 unsigned int sse_y = UINT_MAX;
1637 const int intra_cost_penalty =
1638 vp9_get_intra_cost_penalty(cpi, bsize, cm->base_qindex, cm->y_dc_delta_q);
1639 int64_t inter_mode_thresh =
1640 RDCOST(x->rdmult, x->rddiv, intra_cost_penalty, 0);
1641 const int *const rd_threshes = cpi->rd.threshes[mi->segment_id][bsize];
1642 const int sb_row = mi_row >> MI_BLOCK_SIZE_LOG2;
1643 int thresh_freq_fact_idx = (sb_row * BLOCK_SIZES + bsize) * MAX_MODES;
1644 const int *const rd_thresh_freq_fact =
1645 (cpi->sf.adaptive_rd_thresh_row_mt)
1646 ? &(tile_data->row_base_thresh_freq_fact[thresh_freq_fact_idx])
1647 : tile_data->thresh_freq_fact[bsize];
1649 INTERP_FILTER filter_ref;
1650 const int bsl = mi_width_log2_lookup[bsize];
1651 const int pred_filter_search =
1652 cm->interp_filter == SWITCHABLE
1653 ? (((mi_row + mi_col) >> bsl) +
1654 get_chessboard_index(cm->current_video_frame)) &
1657 int const_motion[MAX_REF_FRAMES] = { 0 };
1658 const int bh = num_4x4_blocks_high_lookup[bsize] << 2;
1659 const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
1660 // For speed 6, the result of interp filter is reused later in actual encoding
1662 // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
1664 DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 64 * 64]);
1665 #if CONFIG_VP9_HIGHBITDEPTH
1666 DECLARE_ALIGNED(16, uint16_t, pred_buf_16[3 * 64 * 64]);
1668 struct buf_2d orig_dst = pd->dst;
1669 PRED_BUFFER *this_mode_pred = NULL;
1670 const int pixels_in_block = bh * bw;
1671 int reuse_inter_pred = cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready;
1672 int ref_frame_skip_mask = 0;
1674 int best_pred_sad = INT_MAX;
1675 int best_early_term = 0;
1676 int ref_frame_cost[MAX_REF_FRAMES];
1677 int svc_force_zero_mode[3] = { 0 };
1678 int perform_intra_pred = 1;
1679 int use_golden_nonzeromv = 1;
1680 int force_skip_low_temp_var = 0;
1681 int skip_ref_find_pred[4] = { 0 };
1682 unsigned int sse_zeromv_normalized = UINT_MAX;
1683 unsigned int best_sse_sofar = UINT_MAX;
1684 int gf_temporal_ref = 0;
1685 #if CONFIG_VP9_TEMPORAL_DENOISING
1686 VP9_PICKMODE_CTX_DEN ctx_den;
1687 int64_t zero_last_cost_orig = INT64_MAX;
1688 int denoise_svc_pickmode = 1;
1690 INTERP_FILTER filter_gf_svc = EIGHTTAP;
1691 MV_REFERENCE_FRAME inter_layer_ref = GOLDEN_FRAME;
1692 const struct segmentation *const seg = &cm->seg;
1694 int num_inter_modes = (cpi->use_svc) ? RT_INTER_MODES_SVC : RT_INTER_MODES;
1695 int flag_svc_subpel = 0;
1699 unsigned int thresh_svc_skip_golden = 500;
1700 int scene_change_detected =
1701 cpi->rc.high_source_sad ||
1702 (cpi->use_svc && cpi->svc.high_source_sad_superframe);
1704 init_best_pickmode(&best_pickmode);
1706 x->source_variance = UINT_MAX;
1707 if (cpi->sf.default_interp_filter == BILINEAR) {
1708 best_pickmode.best_pred_filter = BILINEAR;
1709 filter_gf_svc = BILINEAR;
1711 if (cpi->use_svc && svc->spatial_layer_id > 0) {
1713 LAYER_IDS_TO_IDX(svc->spatial_layer_id - 1, svc->temporal_layer_id,
1714 svc->number_temporal_layers);
1715 LAYER_CONTEXT *const lc = &svc->layer_context[layer];
1716 if (lc->scaling_factor_num == lc->scaling_factor_den) no_scaling = 1;
1718 if (svc->spatial_layer_id > 0 &&
1719 (svc->high_source_sad_superframe || no_scaling))
1720 thresh_svc_skip_golden = 0;
1721 // Lower the skip threshold if lower spatial layer is better quality relative
1722 // to current layer.
1723 else if (svc->spatial_layer_id > 0 && cm->base_qindex > 150 &&
1724 cm->base_qindex > svc->lower_layer_qindex + 15)
1725 thresh_svc_skip_golden = 100;
1726 // Increase skip threshold if lower spatial layer is lower quality relative
1727 // to current layer.
1728 else if (svc->spatial_layer_id > 0 && cm->base_qindex < 140 &&
1729 cm->base_qindex < svc->lower_layer_qindex - 20)
1730 thresh_svc_skip_golden = 1000;
1732 if (!cpi->use_svc ||
1733 (svc->use_gf_temporal_ref_current_layer &&
1734 !svc->layer_context[svc->temporal_layer_id].is_key_frame)) {
1735 gf_temporal_ref = 1;
1736 if (cpi->rc.avg_frame_low_motion > 70)
1737 thresh_svc_skip_golden = 500;
1739 thresh_svc_skip_golden = 0;
1742 init_ref_frame_cost(cm, xd, ref_frame_cost);
1743 memset(&mode_checked[0][0], 0, MB_MODE_COUNT * MAX_REF_FRAMES);
1745 if (reuse_inter_pred) {
1747 for (i = 0; i < 3; i++) {
1748 #if CONFIG_VP9_HIGHBITDEPTH
1749 if (cm->use_highbitdepth)
1750 tmp[i].data = CONVERT_TO_BYTEPTR(&pred_buf_16[pixels_in_block * i]);
1752 tmp[i].data = &pred_buf[pixels_in_block * i];
1754 tmp[i].data = &pred_buf[pixels_in_block * i];
1755 #endif // CONFIG_VP9_HIGHBITDEPTH
1759 tmp[3].data = pd->dst.buf;
1760 tmp[3].stride = pd->dst.stride;
1764 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1767 // Instead of using vp9_get_pred_context_switchable_interp(xd) to assign
1768 // filter_ref, we use a less strict condition on assigning filter_ref.
1769 // This is to reduce the probabily of entering the flow of not assigning
1770 // filter_ref and then skip filter search.
1771 filter_ref = cm->interp_filter;
1772 if (cpi->sf.default_interp_filter != BILINEAR) {
1773 if (xd->above_mi && is_inter_block(xd->above_mi))
1774 filter_ref = xd->above_mi->interp_filter;
1775 else if (xd->left_mi && is_inter_block(xd->left_mi))
1776 filter_ref = xd->left_mi->interp_filter;
1779 // initialize mode decisions
1780 vp9_rd_cost_reset(&best_rdc);
1781 vp9_rd_cost_reset(rd_cost);
1782 mi->sb_type = bsize;
1783 mi->ref_frame[0] = NONE;
1784 mi->ref_frame[1] = NONE;
1787 VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[cm->tx_mode]);
1789 if (sf->short_circuit_flat_blocks || sf->limit_newmv_early_exit) {
1790 #if CONFIG_VP9_HIGHBITDEPTH
1791 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
1792 x->source_variance = vp9_high_get_sby_perpixel_variance(
1793 cpi, &x->plane[0].src, bsize, xd->bd);
1795 #endif // CONFIG_VP9_HIGHBITDEPTH
1796 x->source_variance =
1797 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1799 if (cpi->oxcf.content == VP9E_CONTENT_SCREEN && mi->segment_id > 0 &&
1800 x->zero_temp_sad_source && x->source_variance == 0) {
1802 vp9_init_plane_quantizers(cpi, x);
1806 #if CONFIG_VP9_TEMPORAL_DENOISING
1807 if (cpi->oxcf.noise_sensitivity > 0) {
1810 LAYER_IDS_TO_IDX(svc->spatial_layer_id, svc->temporal_layer_id,
1811 svc->number_temporal_layers);
1812 LAYER_CONTEXT *lc = &svc->layer_context[layer];
1813 denoise_svc_pickmode = denoise_svc(cpi) && !lc->is_key_frame;
1815 if (cpi->denoiser.denoising_level > kDenLowLow && denoise_svc_pickmode)
1816 vp9_denoiser_reset_frame_stats(ctx);
1820 if (cpi->rc.frames_since_golden == 0 && gf_temporal_ref &&
1821 !cpi->rc.alt_ref_gf_group && !cpi->rc.last_frame_is_src_altref) {
1822 usable_ref_frame = LAST_FRAME;
1824 usable_ref_frame = GOLDEN_FRAME;
1827 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
1828 if (cpi->rc.alt_ref_gf_group || cpi->rc.is_src_frame_alt_ref)
1829 usable_ref_frame = ALTREF_FRAME;
1831 if (cpi->rc.is_src_frame_alt_ref) {
1832 skip_ref_find_pred[LAST_FRAME] = 1;
1833 skip_ref_find_pred[GOLDEN_FRAME] = 1;
1835 if (!cm->show_frame) {
1836 if (cpi->rc.frames_since_key == 1) {
1837 usable_ref_frame = LAST_FRAME;
1838 skip_ref_find_pred[GOLDEN_FRAME] = 1;
1839 skip_ref_find_pred[ALTREF_FRAME] = 1;
1844 // For svc mode, on spatial_layer_id > 0: if the reference has different scale
1845 // constrain the inter mode to only test zero motion.
1846 if (cpi->use_svc && svc->force_zero_mode_spatial_ref &&
1847 svc->spatial_layer_id > 0 && !gf_temporal_ref) {
1848 if (cpi->ref_frame_flags & flag_list[LAST_FRAME]) {
1849 struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf;
1850 if (vp9_is_scaled(sf)) {
1851 svc_force_zero_mode[LAST_FRAME - 1] = 1;
1852 inter_layer_ref = LAST_FRAME;
1855 if (cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) {
1856 struct scale_factors *const sf = &cm->frame_refs[GOLDEN_FRAME - 1].sf;
1857 if (vp9_is_scaled(sf)) {
1858 svc_force_zero_mode[GOLDEN_FRAME - 1] = 1;
1859 inter_layer_ref = GOLDEN_FRAME;
1864 if (cpi->sf.short_circuit_low_temp_var) {
1865 force_skip_low_temp_var =
1866 get_force_skip_low_temp_var(&x->variance_low[0], mi_row, mi_col, bsize);
1867 // If force_skip_low_temp_var is set, and for short circuit mode = 1 and 3,
1868 // skip golden reference.
1869 if ((cpi->sf.short_circuit_low_temp_var == 1 ||
1870 cpi->sf.short_circuit_low_temp_var == 3) &&
1871 force_skip_low_temp_var) {
1872 usable_ref_frame = LAST_FRAME;
1876 if (sf->disable_golden_ref && (x->content_state_sb != kVeryHighSad ||
1877 cpi->rc.avg_frame_low_motion < 60))
1878 usable_ref_frame = LAST_FRAME;
1880 if (!((cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) &&
1881 !svc_force_zero_mode[GOLDEN_FRAME - 1] && !force_skip_low_temp_var))
1882 use_golden_nonzeromv = 0;
1884 if (cpi->oxcf.speed >= 8 && !cpi->use_svc &&
1885 ((cpi->rc.frames_since_golden + 1) < x->last_sb_high_content ||
1886 x->last_sb_high_content > 40 || cpi->rc.frames_since_golden > 120))
1887 usable_ref_frame = LAST_FRAME;
1889 // Compound prediction modes: (0,0) on LAST/GOLDEN and ARF.
1890 if (cm->reference_mode == REFERENCE_MODE_SELECT &&
1891 cpi->sf.use_compound_nonrd_pickmode && usable_ref_frame == ALTREF_FRAME)
1894 // If the segment reference frame feature is enabled and it's set to GOLDEN
1895 // reference, then make sure we don't skip checking GOLDEN, this is to
1896 // prevent possibility of not picking any mode.
1897 if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
1898 get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) == GOLDEN_FRAME) {
1899 usable_ref_frame = GOLDEN_FRAME;
1900 skip_ref_find_pred[GOLDEN_FRAME] = 0;
1901 thresh_svc_skip_golden = 0;
1904 for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {
1905 // Skip find_predictor if the reference frame is not in the
1906 // ref_frame_flags (i.e., not used as a reference for this frame).
1907 skip_ref_find_pred[ref_frame] =
1908 !(cpi->ref_frame_flags & flag_list[ref_frame]);
1909 if (!skip_ref_find_pred[ref_frame]) {
1910 find_predictors(cpi, x, ref_frame, frame_mv, const_motion,
1911 &ref_frame_skip_mask, flag_list, tile_data, mi_row,
1912 mi_col, yv12_mb, bsize, force_skip_low_temp_var,
1917 if (cpi->use_svc || cpi->oxcf.speed <= 7 || bsize < BLOCK_32X32)
1918 x->sb_use_mv_part = 0;
1920 // Set the flag_svc_subpel to 1 for SVC if the lower spatial layer used
1921 // an averaging filter for downsampling (phase = 8). If so, we will test
1922 // a nonzero motion mode on the spatial reference.
1923 // The nonzero motion is half pixel shifted to left and top (-4, -4).
1924 if (cpi->use_svc && svc->spatial_layer_id > 0 &&
1925 svc_force_zero_mode[inter_layer_ref - 1] &&
1926 svc->downsample_filter_phase[svc->spatial_layer_id - 1] == 8 &&
1930 flag_svc_subpel = 1;
1933 for (idx = 0; idx < num_inter_modes + comp_modes; ++idx) {
1940 int this_early_term = 0;
1941 int rd_computed = 0;
1942 int flag_preduv_computed[2] = { 0 };
1943 int inter_mv_mode = 0;
1944 int skip_this_mv = 0;
1946 int force_mv_inter_layer = 0;
1947 PREDICTION_MODE this_mode;
1948 second_ref_frame = NONE;
1950 if (idx < num_inter_modes) {
1951 this_mode = ref_mode_set[idx].pred_mode;
1952 ref_frame = ref_mode_set[idx].ref_frame;
1955 this_mode = ref_mode_set_svc[idx].pred_mode;
1956 ref_frame = ref_mode_set_svc[idx].ref_frame;
1959 // Add (0,0) compound modes.
1961 ref_frame = LAST_FRAME;
1962 if (idx == num_inter_modes + comp_modes - 1) ref_frame = GOLDEN_FRAME;
1963 second_ref_frame = ALTREF_FRAME;
1967 if (ref_frame > usable_ref_frame) continue;
1968 if (skip_ref_find_pred[ref_frame]) continue;
1970 if (svc->previous_frame_is_intra_only) {
1971 if (ref_frame != LAST_FRAME || frame_mv[this_mode][ref_frame].as_int != 0)
1975 // If the segment reference frame feature is enabled then do nothing if the
1976 // current ref frame is not allowed.
1977 if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
1978 get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
1981 if (flag_svc_subpel && ref_frame == inter_layer_ref) {
1982 force_mv_inter_layer = 1;
1983 // Only test mode if NEARESTMV/NEARMV is (svc_mv_col, svc_mv_row),
1984 // otherwise set NEWMV to (svc_mv_col, svc_mv_row).
1985 if (this_mode == NEWMV) {
1986 frame_mv[this_mode][ref_frame].as_mv.col = svc_mv_col;
1987 frame_mv[this_mode][ref_frame].as_mv.row = svc_mv_row;
1988 } else if (frame_mv[this_mode][ref_frame].as_mv.col != svc_mv_col ||
1989 frame_mv[this_mode][ref_frame].as_mv.row != svc_mv_row) {
1995 if (!cpi->allow_comp_inter_inter) continue;
1996 // Skip compound inter modes if ARF is not available.
1997 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) continue;
1998 // Do not allow compound prediction if the segment level reference frame
1999 // feature is in use as in this case there can only be one reference.
2000 if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME)) continue;
2003 // For SVC, skip the golden (spatial) reference search if sse of zeromv_last
2004 // is below threshold.
2005 if (cpi->use_svc && ref_frame == GOLDEN_FRAME &&
2006 sse_zeromv_normalized < thresh_svc_skip_golden)
2009 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) continue;
2011 if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
2012 (frame_mv[this_mode][ref_frame].as_int != 0 ||
2013 (cpi->oxcf.content == VP9E_CONTENT_SCREEN && !svc->spatial_layer_id &&
2014 !x->zero_temp_sad_source))) {
2018 if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode))) continue;
2020 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
2021 if (cpi->rc.is_src_frame_alt_ref &&
2022 (ref_frame != ALTREF_FRAME ||
2023 frame_mv[this_mode][ref_frame].as_int != 0))
2026 if (!cm->show_frame && ref_frame == ALTREF_FRAME &&
2027 frame_mv[this_mode][ref_frame].as_int != 0)
2030 if (cpi->rc.alt_ref_gf_group && cm->show_frame &&
2031 cpi->rc.frames_since_golden > (cpi->rc.baseline_gf_interval >> 1) &&
2032 ref_frame == GOLDEN_FRAME &&
2033 frame_mv[this_mode][ref_frame].as_int != 0)
2036 if (cpi->rc.alt_ref_gf_group && cm->show_frame &&
2037 cpi->rc.frames_since_golden > 0 &&
2038 cpi->rc.frames_since_golden < (cpi->rc.baseline_gf_interval >> 1) &&
2039 ref_frame == ALTREF_FRAME &&
2040 frame_mv[this_mode][ref_frame].as_int != 0)
2044 if (const_motion[ref_frame] && this_mode == NEARMV) continue;
2046 // Skip non-zeromv mode search for golden frame if force_skip_low_temp_var
2047 // is set. If nearestmv for golden frame is 0, zeromv mode will be skipped
2049 if (!force_mv_inter_layer && force_skip_low_temp_var &&
2050 ref_frame == GOLDEN_FRAME &&
2051 frame_mv[this_mode][ref_frame].as_int != 0) {
2055 if (x->content_state_sb != kVeryHighSad &&
2056 (cpi->sf.short_circuit_low_temp_var >= 2 ||
2057 (cpi->sf.short_circuit_low_temp_var == 1 && bsize == BLOCK_64X64)) &&
2058 force_skip_low_temp_var && ref_frame == LAST_FRAME &&
2059 this_mode == NEWMV) {
2064 if (!force_mv_inter_layer && svc_force_zero_mode[ref_frame - 1] &&
2065 frame_mv[this_mode][ref_frame].as_int != 0)
2069 // Disable this drop out case if the ref frame segment level feature is
2070 // enabled for this segment. This is to prevent the possibility that we end
2071 // up unable to pick any mode.
2072 if (!segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME)) {
2073 if (sf->reference_masking &&
2074 !(frame_mv[this_mode][ref_frame].as_int == 0 &&
2075 ref_frame == LAST_FRAME)) {
2076 if (usable_ref_frame < ALTREF_FRAME) {
2077 if (!force_skip_low_temp_var && usable_ref_frame > LAST_FRAME) {
2078 i = (ref_frame == LAST_FRAME) ? GOLDEN_FRAME : LAST_FRAME;
2079 if ((cpi->ref_frame_flags & flag_list[i]))
2080 if (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[i] << 1))
2081 ref_frame_skip_mask |= (1 << ref_frame);
2083 } else if (!cpi->rc.is_src_frame_alt_ref &&
2084 !(frame_mv[this_mode][ref_frame].as_int == 0 &&
2085 ref_frame == ALTREF_FRAME)) {
2086 int ref1 = (ref_frame == GOLDEN_FRAME) ? LAST_FRAME : GOLDEN_FRAME;
2087 int ref2 = (ref_frame == ALTREF_FRAME) ? LAST_FRAME : ALTREF_FRAME;
2088 if (((cpi->ref_frame_flags & flag_list[ref1]) &&
2089 (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref1] << 1))) ||
2090 ((cpi->ref_frame_flags & flag_list[ref2]) &&
2091 (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref2] << 1))))
2092 ref_frame_skip_mask |= (1 << ref_frame);
2095 if (ref_frame_skip_mask & (1 << ref_frame)) continue;
2098 // Select prediction reference frames.
2099 for (i = 0; i < MAX_MB_PLANE; i++) {
2100 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
2101 if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
2104 mi->ref_frame[0] = ref_frame;
2105 mi->ref_frame[1] = second_ref_frame;
2106 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
2108 mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
2109 mode_rd_thresh = best_pickmode.best_mode_skip_txfm
2110 ? rd_threshes[mode_index] << 1
2111 : rd_threshes[mode_index];
2113 // Increase mode_rd_thresh value for GOLDEN_FRAME for improved encoding
2114 // speed with little/no subjective quality loss.
2115 if (cpi->sf.bias_golden && ref_frame == GOLDEN_FRAME &&
2116 cpi->rc.frames_since_golden > 4)
2117 mode_rd_thresh = mode_rd_thresh << 3;
2119 if ((cpi->sf.adaptive_rd_thresh_row_mt &&
2120 rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
2121 &rd_thresh_freq_fact[mode_index])) ||
2122 (!cpi->sf.adaptive_rd_thresh_row_mt &&
2123 rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
2124 &rd_thresh_freq_fact[mode_index])))
2125 if (frame_mv[this_mode][ref_frame].as_int != 0) continue;
2127 if (this_mode == NEWMV && !force_mv_inter_layer) {
2128 if (search_new_mv(cpi, x, frame_mv, ref_frame, gf_temporal_ref, bsize,
2129 mi_row, mi_col, best_pred_sad, &rate_mv, best_sse_sofar,
2134 // TODO(jianj): Skipping the testing of (duplicate) non-zero motion vector
2135 // causes some regression, leave it for duplicate zero-mv for now, until
2136 // regression issue is resolved.
2137 for (inter_mv_mode = NEARESTMV; inter_mv_mode <= NEWMV; inter_mv_mode++) {
2138 if (inter_mv_mode == this_mode || comp_pred) continue;
2139 if (mode_checked[inter_mv_mode][ref_frame] &&
2140 frame_mv[this_mode][ref_frame].as_int ==
2141 frame_mv[inter_mv_mode][ref_frame].as_int &&
2142 frame_mv[inter_mv_mode][ref_frame].as_int == 0) {
2148 if (skip_this_mv) continue;
2150 // If use_golden_nonzeromv is false, NEWMV mode is skipped for golden, no
2151 // need to compute best_pred_sad which is only used to skip golden NEWMV.
2152 if (use_golden_nonzeromv && this_mode == NEWMV && ref_frame == LAST_FRAME &&
2153 frame_mv[NEWMV][LAST_FRAME].as_int != INVALID_MV) {
2154 const int pre_stride = xd->plane[0].pre[0].stride;
2155 const uint8_t *const pre_buf =
2156 xd->plane[0].pre[0].buf +
2157 (frame_mv[NEWMV][LAST_FRAME].as_mv.row >> 3) * pre_stride +
2158 (frame_mv[NEWMV][LAST_FRAME].as_mv.col >> 3);
2159 best_pred_sad = cpi->fn_ptr[bsize].sdf(
2160 x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride);
2161 x->pred_mv_sad[LAST_FRAME] = best_pred_sad;
2164 if (this_mode != NEARESTMV && !comp_pred &&
2165 frame_mv[this_mode][ref_frame].as_int ==
2166 frame_mv[NEARESTMV][ref_frame].as_int)
2169 mi->mode = this_mode;
2170 mi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
2171 mi->mv[1].as_int = 0;
2173 // Search for the best prediction filter type, when the resulting
2174 // motion vector is at sub-pixel accuracy level for luma component, i.e.,
2175 // the last three bits are all zeros.
2176 if (reuse_inter_pred) {
2177 if (!this_mode_pred) {
2178 this_mode_pred = &tmp[3];
2180 this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
2181 pd->dst.buf = this_mode_pred->data;
2182 pd->dst.stride = bw;
2186 if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&
2187 pred_filter_search &&
2188 (ref_frame == LAST_FRAME ||
2189 (ref_frame == GOLDEN_FRAME && !force_mv_inter_layer &&
2190 (cpi->use_svc || cpi->oxcf.rc_mode == VPX_VBR))) &&
2191 (((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07) != 0)) {
2193 search_filter_ref(cpi, x, &this_rdc, mi_row, mi_col, tmp, bsize,
2194 reuse_inter_pred, &this_mode_pred, &var_y, &sse_y);
2196 // For low motion content use x->sb_is_skin in addition to VeryHighSad
2197 // for setting large_block.
2198 const int large_block =
2199 (x->content_state_sb == kVeryHighSad ||
2200 (x->sb_is_skin && cpi->rc.avg_frame_low_motion > 70) ||
2201 cpi->oxcf.speed < 7)
2202 ? bsize > BLOCK_32X32
2203 : bsize >= BLOCK_32X32;
2204 mi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP : filter_ref;
2206 if (cpi->use_svc && ref_frame == GOLDEN_FRAME &&
2207 svc_force_zero_mode[ref_frame - 1])
2208 mi->interp_filter = filter_gf_svc;
2210 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
2212 // For large partition blocks, extra testing is done.
2213 if (cpi->oxcf.rc_mode == VPX_CBR && large_block &&
2214 !cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) &&
2216 model_rd_for_sb_y_large(cpi, bsize, x, xd, &this_rdc.rate,
2217 &this_rdc.dist, &var_y, &sse_y, mi_row, mi_col,
2218 &this_early_term, flag_preduv_computed);
2221 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
2224 // Save normalized sse (between current and last frame) for (0, 0) motion.
2225 if (cpi->use_svc && ref_frame == LAST_FRAME &&
2226 frame_mv[this_mode][ref_frame].as_int == 0) {
2227 sse_zeromv_normalized =
2228 sse_y >> (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
2230 if (sse_y < best_sse_sofar) best_sse_sofar = sse_y;
2233 if (!this_early_term) {
2234 this_sse = (int64_t)sse_y;
2235 block_yrd(cpi, x, &this_rdc, &is_skippable, &this_sse, bsize,
2236 VPXMIN(mi->tx_size, TX_16X16), rd_computed);
2238 x->skip_txfm[0] = is_skippable;
2240 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2242 if (RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist) <
2243 RDCOST(x->rdmult, x->rddiv, 0, this_sse)) {
2244 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
2246 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2247 this_rdc.dist = this_sse;
2248 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
2252 if (cm->interp_filter == SWITCHABLE) {
2253 if ((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07)
2254 this_rdc.rate += vp9_get_switchable_rate(cpi, xd);
2257 this_rdc.rate += cm->interp_filter == SWITCHABLE
2258 ? vp9_get_switchable_rate(cpi, xd)
2260 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2263 if (!this_early_term &&
2264 (x->color_sensitivity[0] || x->color_sensitivity[1])) {
2266 const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, &xd->plane[1]);
2267 if (x->color_sensitivity[0] && !flag_preduv_computed[0]) {
2268 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 1);
2269 flag_preduv_computed[0] = 1;
2271 if (x->color_sensitivity[1] && !flag_preduv_computed[1]) {
2272 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 2);
2273 flag_preduv_computed[1] = 1;
2275 model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &rdc_uv, &var_y, &sse_y, 1, 2);
2276 this_rdc.rate += rdc_uv.rate;
2277 this_rdc.dist += rdc_uv.dist;
2280 this_rdc.rate += rate_mv;
2281 this_rdc.rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2282 [INTER_OFFSET(this_mode)];
2283 // TODO(marpan): Add costing for compound mode.
2284 this_rdc.rate += ref_frame_cost[ref_frame];
2285 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2287 // Bias against NEWMV that is very different from its neighbors, and bias
2288 // to small motion-lastref for noisy input.
2289 if (cpi->oxcf.rc_mode == VPX_CBR && cpi->oxcf.speed >= 5 &&
2290 cpi->oxcf.content != VP9E_CONTENT_SCREEN) {
2291 vp9_NEWMV_diff_bias(&cpi->noise_estimate, xd, this_mode, &this_rdc, bsize,
2292 frame_mv[this_mode][ref_frame].as_mv.row,
2293 frame_mv[this_mode][ref_frame].as_mv.col,
2294 ref_frame == LAST_FRAME, x->lowvar_highsumdiff,
2298 // Skipping checking: test to see if this block can be reconstructed by
2300 if (cpi->allow_encode_breakout && !xd->lossless && !scene_change_detected) {
2301 encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame, this_mode,
2302 var_y, sse_y, yv12_mb, &this_rdc.rate,
2303 &this_rdc.dist, flag_preduv_computed);
2305 this_rdc.rate += rate_mv;
2307 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2311 #if CONFIG_VP9_TEMPORAL_DENOISING
2312 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc_pickmode &&
2313 cpi->denoiser.denoising_level > kDenLowLow) {
2314 vp9_denoiser_update_frame_stats(mi, sse_y, this_mode, ctx);
2315 // Keep track of zero_last cost.
2316 if (ref_frame == LAST_FRAME && frame_mv[this_mode][ref_frame].as_int == 0)
2317 zero_last_cost_orig = this_rdc.rdcost;
2323 mode_checked[this_mode][ref_frame] = 1;
2325 if (this_rdc.rdcost < best_rdc.rdcost || x->skip) {
2326 best_rdc = this_rdc;
2327 best_early_term = this_early_term;
2328 best_pickmode.best_mode = this_mode;
2329 best_pickmode.best_pred_filter = mi->interp_filter;
2330 best_pickmode.best_tx_size = mi->tx_size;
2331 best_pickmode.best_ref_frame = ref_frame;
2332 best_pickmode.best_mode_skip_txfm = x->skip_txfm[0];
2333 best_pickmode.best_second_ref_frame = second_ref_frame;
2335 if (reuse_inter_pred) {
2336 free_pred_buffer(best_pickmode.best_pred);
2337 best_pickmode.best_pred = this_mode_pred;
2340 if (reuse_inter_pred) free_pred_buffer(this_mode_pred);
2345 // If early termination flag is 1 and at least 2 modes are checked,
2346 // the mode search is terminated.
2347 if (best_early_term && idx > 0 && !scene_change_detected) {
2353 mi->mode = best_pickmode.best_mode;
2354 mi->interp_filter = best_pickmode.best_pred_filter;
2355 mi->tx_size = best_pickmode.best_tx_size;
2356 mi->ref_frame[0] = best_pickmode.best_ref_frame;
2358 frame_mv[best_pickmode.best_mode][best_pickmode.best_ref_frame].as_int;
2359 xd->mi[0]->bmi[0].as_mv[0].as_int = mi->mv[0].as_int;
2360 x->skip_txfm[0] = best_pickmode.best_mode_skip_txfm;
2361 mi->ref_frame[1] = best_pickmode.best_second_ref_frame;
2363 // For spatial enhancemanent layer: perform intra prediction only if base
2364 // layer is chosen as the reference. Always perform intra prediction if
2365 // LAST is the only reference, or is_key_frame is set, or on base
2367 if (svc->spatial_layer_id && !gf_temporal_ref) {
2368 perform_intra_pred =
2369 svc->temporal_layer_id == 0 ||
2370 svc->layer_context[svc->temporal_layer_id].is_key_frame ||
2371 !(cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) ||
2372 (!svc->layer_context[svc->temporal_layer_id].is_key_frame &&
2373 svc_force_zero_mode[best_pickmode.best_ref_frame - 1]);
2374 inter_mode_thresh = (inter_mode_thresh << 1) + inter_mode_thresh;
2376 if ((cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR &&
2377 cpi->rc.is_src_frame_alt_ref) ||
2378 svc->previous_frame_is_intra_only)
2379 perform_intra_pred = 0;
2381 // If the segment reference frame feature is enabled and set then
2382 // skip the intra prediction.
2383 if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
2384 get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) > 0)
2385 perform_intra_pred = 0;
2387 // Perform intra prediction search, if the best SAD is above a certain
2389 if (best_rdc.rdcost == INT64_MAX ||
2390 (scene_change_detected && perform_intra_pred) ||
2391 ((!force_skip_low_temp_var || bsize < BLOCK_32X32 ||
2392 x->content_state_sb == kVeryHighSad) &&
2393 perform_intra_pred && !x->skip && best_rdc.rdcost > inter_mode_thresh &&
2394 bsize <= cpi->sf.max_intra_bsize && !x->skip_low_source_sad &&
2395 !x->lowvar_highsumdiff)) {
2396 struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
2398 PRED_BUFFER *const best_pred = best_pickmode.best_pred;
2399 TX_SIZE intra_tx_size =
2400 VPXMIN(max_txsize_lookup[bsize],
2401 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
2402 if (cpi->oxcf.content != VP9E_CONTENT_SCREEN && intra_tx_size > TX_16X16)
2403 intra_tx_size = TX_16X16;
2405 if (reuse_inter_pred && best_pred != NULL) {
2406 if (best_pred->data == orig_dst.buf) {
2407 this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
2408 #if CONFIG_VP9_HIGHBITDEPTH
2409 if (cm->use_highbitdepth)
2410 vpx_highbd_convolve_copy(
2411 CONVERT_TO_SHORTPTR(best_pred->data), best_pred->stride,
2412 CONVERT_TO_SHORTPTR(this_mode_pred->data), this_mode_pred->stride,
2413 NULL, 0, 0, 0, 0, bw, bh, xd->bd);
2415 vpx_convolve_copy(best_pred->data, best_pred->stride,
2416 this_mode_pred->data, this_mode_pred->stride, NULL,
2417 0, 0, 0, 0, bw, bh);
2419 vpx_convolve_copy(best_pred->data, best_pred->stride,
2420 this_mode_pred->data, this_mode_pred->stride, NULL, 0,
2422 #endif // CONFIG_VP9_HIGHBITDEPTH
2423 best_pickmode.best_pred = this_mode_pred;
2428 for (i = 0; i < 4; ++i) {
2429 const PREDICTION_MODE this_mode = intra_mode_list[i];
2430 THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)];
2431 int mode_rd_thresh = rd_threshes[mode_index];
2432 if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
2433 this_mode != DC_PRED) {
2437 if (!((1 << this_mode) & cpi->sf.intra_y_mode_bsize_mask[bsize]))
2440 if ((cpi->sf.adaptive_rd_thresh_row_mt &&
2441 rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
2442 &rd_thresh_freq_fact[mode_index])) ||
2443 (!cpi->sf.adaptive_rd_thresh_row_mt &&
2444 rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
2445 &rd_thresh_freq_fact[mode_index])))
2448 mi->mode = this_mode;
2449 mi->ref_frame[0] = INTRA_FRAME;
2450 this_rdc.dist = this_rdc.rate = 0;
2451 args.mode = this_mode;
2453 args.rdc = &this_rdc;
2454 mi->tx_size = intra_tx_size;
2455 vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
2457 // Check skip cost here since skippable is not set for for uv, this
2458 // mirrors the behavior used by inter
2459 if (args.skippable) {
2460 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
2461 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
2463 x->skip_txfm[0] = SKIP_TXFM_NONE;
2464 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
2466 // Inter and intra RD will mismatch in scale for non-screen content.
2467 if (cpi->oxcf.content == VP9E_CONTENT_SCREEN) {
2468 if (x->color_sensitivity[0])
2469 vp9_foreach_transformed_block_in_plane(xd, bsize, 1,
2470 estimate_block_intra, &args);
2471 if (x->color_sensitivity[1])
2472 vp9_foreach_transformed_block_in_plane(xd, bsize, 2,
2473 estimate_block_intra, &args);
2475 this_rdc.rate += cpi->mbmode_cost[this_mode];
2476 this_rdc.rate += ref_frame_cost[INTRA_FRAME];
2477 this_rdc.rate += intra_cost_penalty;
2479 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2481 if (this_rdc.rdcost < best_rdc.rdcost) {
2482 best_rdc = this_rdc;
2483 best_pickmode.best_mode = this_mode;
2484 best_pickmode.best_intra_tx_size = mi->tx_size;
2485 best_pickmode.best_ref_frame = INTRA_FRAME;
2486 best_pickmode.best_second_ref_frame = NONE;
2487 mi->uv_mode = this_mode;
2488 mi->mv[0].as_int = INVALID_MV;
2489 mi->mv[1].as_int = INVALID_MV;
2490 best_pickmode.best_mode_skip_txfm = x->skip_txfm[0];
2494 // Reset mb_mode_info to the best inter mode.
2495 if (best_pickmode.best_ref_frame != INTRA_FRAME) {
2496 mi->tx_size = best_pickmode.best_tx_size;
2498 mi->tx_size = best_pickmode.best_intra_tx_size;
2503 mi->mode = best_pickmode.best_mode;
2504 mi->ref_frame[0] = best_pickmode.best_ref_frame;
2505 mi->ref_frame[1] = best_pickmode.best_second_ref_frame;
2506 x->skip_txfm[0] = best_pickmode.best_mode_skip_txfm;
2508 if (!is_inter_block(mi)) {
2509 mi->interp_filter = SWITCHABLE_FILTERS;
2512 if (reuse_inter_pred && best_pickmode.best_pred != NULL) {
2513 PRED_BUFFER *const best_pred = best_pickmode.best_pred;
2514 if (best_pred->data != orig_dst.buf && is_inter_mode(mi->mode)) {
2515 #if CONFIG_VP9_HIGHBITDEPTH
2516 if (cm->use_highbitdepth)
2517 vpx_highbd_convolve_copy(
2518 CONVERT_TO_SHORTPTR(best_pred->data), best_pred->stride,
2519 CONVERT_TO_SHORTPTR(pd->dst.buf), pd->dst.stride, NULL, 0, 0, 0, 0,
2522 vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2523 pd->dst.stride, NULL, 0, 0, 0, 0, bw, bh);
2525 vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2526 pd->dst.stride, NULL, 0, 0, 0, 0, bw, bh);
2527 #endif // CONFIG_VP9_HIGHBITDEPTH
2531 #if CONFIG_VP9_TEMPORAL_DENOISING
2532 if (cpi->oxcf.noise_sensitivity > 0 && cpi->resize_pending == 0 &&
2533 denoise_svc_pickmode && cpi->denoiser.denoising_level > kDenLowLow &&
2534 cpi->denoiser.reset == 0) {
2535 VP9_DENOISER_DECISION decision = COPY_BLOCK;
2536 ctx->sb_skip_denoising = 0;
2537 // TODO(marpan): There is an issue with denoising when the
2538 // superblock partitioning scheme is based on the pickmode.
2539 // Remove this condition when the issue is resolved.
2540 if (x->sb_pickmode_part) ctx->sb_skip_denoising = 1;
2541 vp9_pickmode_ctx_den_update(&ctx_den, zero_last_cost_orig, ref_frame_cost,
2542 frame_mv, reuse_inter_pred, &best_pickmode);
2543 vp9_denoiser_denoise(cpi, x, mi_row, mi_col, bsize, ctx, &decision,
2545 recheck_zeromv_after_denoising(cpi, mi, x, xd, decision, &ctx_den, yv12_mb,
2546 &best_rdc, bsize, mi_row, mi_col);
2547 best_pickmode.best_ref_frame = ctx_den.best_ref_frame;
2551 if (best_pickmode.best_ref_frame == ALTREF_FRAME ||
2552 best_pickmode.best_second_ref_frame == ALTREF_FRAME)
2553 x->arf_frame_usage++;
2554 else if (best_pickmode.best_ref_frame != INTRA_FRAME)
2555 x->lastgolden_frame_usage++;
2557 if (cpi->sf.adaptive_rd_thresh) {
2558 THR_MODES best_mode_idx =
2559 mode_idx[best_pickmode.best_ref_frame][mode_offset(mi->mode)];
2561 if (best_pickmode.best_ref_frame == INTRA_FRAME) {
2562 // Only consider the modes that are included in the intra_mode_list.
2563 int intra_modes = sizeof(intra_mode_list) / sizeof(PREDICTION_MODE);
2566 // TODO(yunqingwang): Check intra mode mask and only update freq_fact
2567 // for those valid modes.
2568 for (i = 0; i < intra_modes; i++) {
2569 if (cpi->sf.adaptive_rd_thresh_row_mt)
2570 update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance,
2571 thresh_freq_fact_idx, INTRA_FRAME,
2572 best_mode_idx, intra_mode_list[i]);
2574 update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2575 INTRA_FRAME, best_mode_idx,
2576 intra_mode_list[i]);
2579 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2580 PREDICTION_MODE this_mode;
2581 if (best_pickmode.best_ref_frame != ref_frame) continue;
2582 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2583 if (cpi->sf.adaptive_rd_thresh_row_mt)
2584 update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance,
2585 thresh_freq_fact_idx, ref_frame,
2586 best_mode_idx, this_mode);
2588 update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2589 ref_frame, best_mode_idx, this_mode);
2595 *rd_cost = best_rdc;
2598 void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x, int mi_row,
2599 int mi_col, RD_COST *rd_cost, BLOCK_SIZE bsize,
2600 PICK_MODE_CONTEXT *ctx) {
2601 VP9_COMMON *const cm = &cpi->common;
2602 SPEED_FEATURES *const sf = &cpi->sf;
2603 MACROBLOCKD *const xd = &x->e_mbd;
2604 MODE_INFO *const mi = xd->mi[0];
2605 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2606 const struct segmentation *const seg = &cm->seg;
2607 MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE;
2608 MV_REFERENCE_FRAME best_ref_frame = NONE;
2609 unsigned char segment_id = mi->segment_id;
2610 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2611 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2613 int64_t best_rd = INT64_MAX;
2614 b_mode_info bsi[MAX_REF_FRAMES][4];
2615 int ref_frame_skip_mask = 0;
2616 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
2617 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
2620 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
2621 ctx->pred_pixel_ready = 0;
2623 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2624 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2626 x->pred_mv_sad[ref_frame] = INT_MAX;
2628 if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
2629 int_mv *const candidates = mbmi_ext->ref_mvs[ref_frame];
2630 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2631 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf,
2633 vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
2634 mbmi_ext->mode_context);
2636 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2637 &dummy_mv[0], &dummy_mv[1]);
2639 ref_frame_skip_mask |= (1 << ref_frame);
2643 mi->sb_type = bsize;
2644 mi->tx_size = TX_4X4;
2645 mi->uv_mode = DC_PRED;
2646 mi->ref_frame[0] = LAST_FRAME;
2647 mi->ref_frame[1] = NONE;
2649 cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
2651 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2652 int64_t this_rd = 0;
2655 if (ref_frame_skip_mask & (1 << ref_frame)) continue;
2657 #if CONFIG_BETTER_HW_COMPATIBILITY
2658 if ((bsize == BLOCK_8X4 || bsize == BLOCK_4X8) && ref_frame > INTRA_FRAME &&
2659 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2663 // TODO(jingning, agrange): Scaling reference frame not supported for
2664 // sub8x8 blocks. Is this supported now?
2665 if (ref_frame > INTRA_FRAME &&
2666 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2669 // If the segment reference frame feature is enabled....
2670 // then do nothing if the current ref frame is not allowed..
2671 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
2672 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
2675 mi->ref_frame[0] = ref_frame;
2677 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
2679 // Select prediction reference frames.
2680 for (plane = 0; plane < MAX_MB_PLANE; plane++)
2681 xd->plane[plane].pre[0] = yv12_mb[ref_frame][plane];
2683 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2684 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2685 int_mv b_mv[MB_MODE_COUNT];
2686 int64_t b_best_rd = INT64_MAX;
2687 const int i = idy * 2 + idx;
2688 PREDICTION_MODE this_mode;
2690 unsigned int var_y, sse_y;
2692 struct macroblock_plane *p = &x->plane[0];
2693 struct macroblockd_plane *pd = &xd->plane[0];
2695 const struct buf_2d orig_src = p->src;
2696 const struct buf_2d orig_dst = pd->dst;
2697 struct buf_2d orig_pre[2];
2698 memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));
2700 // set buffer pointers for sub8x8 motion search.
2702 &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
2704 &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
2707 .buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
2709 b_mv[ZEROMV].as_int = 0;
2710 b_mv[NEWMV].as_int = INVALID_MV;
2711 vp9_append_sub8x8_mvs_for_idx(cm, xd, i, 0, mi_row, mi_col,
2712 &b_mv[NEARESTMV], &b_mv[NEARMV],
2713 mbmi_ext->mode_context);
2715 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2717 xd->mi[0]->bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
2719 if (this_mode == NEWMV) {
2720 const int step_param = cpi->sf.mv.fullpel_search_step_param;
2724 const MvLimits tmp_mv_limits = x->mv_limits;
2725 uint32_t dummy_dist;
2728 mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3;
2729 mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3;
2731 mvp_full.row = xd->mi[0]->bmi[0].as_mv[0].as_mv.row >> 3;
2732 mvp_full.col = xd->mi[0]->bmi[0].as_mv[0].as_mv.col >> 3;
2735 vp9_set_mv_search_range(&x->mv_limits,
2736 &mbmi_ext->ref_mvs[ref_frame][0].as_mv);
2738 vp9_full_pixel_search(
2739 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method,
2740 x->sadperbit4, cond_cost_list(cpi, cost_list),
2741 &mbmi_ext->ref_mvs[ref_frame][0].as_mv, &tmp_mv, INT_MAX, 0);
2743 x->mv_limits = tmp_mv_limits;
2745 // calculate the bit cost on motion vector
2746 mvp_full.row = tmp_mv.row * 8;
2747 mvp_full.col = tmp_mv.col * 8;
2749 b_rate += vp9_mv_bit_cost(
2750 &mvp_full, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2751 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2753 b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2754 [INTER_OFFSET(NEWMV)];
2755 if (RDCOST(x->rdmult, x->rddiv, b_rate, 0) > b_best_rd) continue;
2757 cpi->find_fractional_mv_step(
2758 x, &tmp_mv, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2759 cpi->common.allow_high_precision_mv, x->errorperbit,
2760 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
2761 cpi->sf.mv.subpel_search_level, cond_cost_list(cpi, cost_list),
2762 x->nmvjointcost, x->mvcost, &dummy_dist,
2763 &x->pred_sse[ref_frame], NULL, 0, 0,
2764 cpi->sf.use_accurate_subpel_search);
2766 xd->mi[0]->bmi[i].as_mv[0].as_mv = tmp_mv;
2768 b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2769 [INTER_OFFSET(this_mode)];
2772 #if CONFIG_VP9_HIGHBITDEPTH
2773 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2774 vp9_highbd_build_inter_predictor(
2775 CONVERT_TO_SHORTPTR(pd->pre[0].buf), pd->pre[0].stride,
2776 CONVERT_TO_SHORTPTR(pd->dst.buf), pd->dst.stride,
2777 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2778 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2779 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2780 mi_col * MI_SIZE + 4 * (i & 0x01),
2781 mi_row * MI_SIZE + 4 * (i >> 1), xd->bd);
2784 vp9_build_inter_predictor(
2785 pd->pre[0].buf, pd->pre[0].stride, pd->dst.buf, pd->dst.stride,
2786 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2787 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2788 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2789 mi_col * MI_SIZE + 4 * (i & 0x01),
2790 mi_row * MI_SIZE + 4 * (i >> 1));
2792 #if CONFIG_VP9_HIGHBITDEPTH
2796 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
2799 this_rdc.rate += b_rate;
2801 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2802 if (this_rdc.rdcost < b_best_rd) {
2803 b_best_rd = this_rdc.rdcost;
2804 bsi[ref_frame][i].as_mode = this_mode;
2805 bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0]->bmi[i].as_mv[0].as_mv;
2809 // restore source and prediction buffer pointers.
2811 pd->pre[0] = orig_pre[0];
2813 this_rd += b_best_rd;
2815 xd->mi[0]->bmi[i] = bsi[ref_frame][i];
2816 if (num_4x4_blocks_wide > 1) xd->mi[0]->bmi[i + 1] = xd->mi[0]->bmi[i];
2817 if (num_4x4_blocks_high > 1) xd->mi[0]->bmi[i + 2] = xd->mi[0]->bmi[i];
2819 } // loop through sub8x8 blocks
2821 if (this_rd < best_rd) {
2823 best_ref_frame = ref_frame;
2825 } // reference frames
2827 mi->tx_size = TX_4X4;
2828 mi->ref_frame[0] = best_ref_frame;
2829 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2830 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2831 const int block = idy * 2 + idx;
2832 xd->mi[0]->bmi[block] = bsi[best_ref_frame][block];
2833 if (num_4x4_blocks_wide > 1)
2834 xd->mi[0]->bmi[block + 1] = bsi[best_ref_frame][block];
2835 if (num_4x4_blocks_high > 1)
2836 xd->mi[0]->bmi[block + 2] = bsi[best_ref_frame][block];
2839 mi->mode = xd->mi[0]->bmi[3].as_mode;
2840 ctx->mic = *(xd->mi[0]);
2841 ctx->mbmi_ext = *x->mbmi_ext;
2842 ctx->skip_txfm[0] = SKIP_TXFM_NONE;
2844 // Dummy assignment for speed -5. No effect in speed -6.
2845 rd_cost->rdcost = best_rd;