2 * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
16 #include "./vp9_rtcd.h"
17 #include "./vpx_dsp_rtcd.h"
19 #include "vpx/vpx_codec.h"
20 #include "vpx_dsp/vpx_dsp_common.h"
21 #include "vpx_mem/vpx_mem.h"
22 #include "vpx_ports/mem.h"
24 #include "vp9/common/vp9_blockd.h"
25 #include "vp9/common/vp9_common.h"
26 #include "vp9/common/vp9_mvref_common.h"
27 #include "vp9/common/vp9_pred_common.h"
28 #include "vp9/common/vp9_reconinter.h"
29 #include "vp9/common/vp9_reconintra.h"
30 #include "vp9/common/vp9_scan.h"
32 #include "vp9/encoder/vp9_cost.h"
33 #include "vp9/encoder/vp9_encoder.h"
34 #include "vp9/encoder/vp9_pickmode.h"
35 #include "vp9/encoder/vp9_ratectrl.h"
36 #include "vp9/encoder/vp9_rd.h"
44 static const int pos_shift_16x16[4][4] = {
45 { 9, 10, 13, 14 }, { 11, 12, 15, 16 }, { 17, 18, 21, 22 }, { 19, 20, 23, 24 }
48 static int mv_refs_rt(VP9_COMP *cpi, const VP9_COMMON *cm, const MACROBLOCK *x,
49 const MACROBLOCKD *xd, const TileInfo *const tile,
50 MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
51 int_mv *mv_ref_list, int_mv *base_mv, int mi_row,
52 int mi_col, int use_base_mv) {
53 const int *ref_sign_bias = cm->ref_frame_sign_bias;
54 int i, refmv_count = 0;
56 const POSITION *const mv_ref_search = mv_ref_blocks[mi->sb_type];
58 int different_ref_found = 0;
59 int context_counter = 0;
62 // Blank the reference vector list
63 memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
65 // The nearest 2 blocks are treated differently
66 // if the size < 8x8 we get the mv from the bmi substructure,
67 // and we also need to keep a mode count.
68 for (i = 0; i < 2; ++i) {
69 const POSITION *const mv_ref = &mv_ref_search[i];
70 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
71 const MODE_INFO *const candidate_mi =
72 xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
73 // Keep counts for entropy encoding.
74 context_counter += mode_2_counter[candidate_mi->mode];
75 different_ref_found = 1;
77 if (candidate_mi->ref_frame[0] == ref_frame)
78 ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1),
79 refmv_count, mv_ref_list, Done);
85 // Check the rest of the neighbors in much the same way
86 // as before except we don't need to keep track of sub blocks or
88 for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
89 const POSITION *const mv_ref = &mv_ref_search[i];
90 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
91 const MODE_INFO *const candidate_mi =
92 xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
93 different_ref_found = 1;
95 if (candidate_mi->ref_frame[0] == ref_frame)
96 ADD_MV_REF_LIST(candidate_mi->mv[0], refmv_count, mv_ref_list, Done);
100 // Since we couldn't find 2 mvs from the same reference frame
101 // go back through the neighbors and find motion vectors from
102 // different reference frames.
103 if (different_ref_found && !refmv_count) {
104 for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
105 const POSITION *mv_ref = &mv_ref_search[i];
106 if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
107 const MODE_INFO *const candidate_mi =
108 xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
110 // If the candidate is INTRA we don't want to consider its mv.
111 IF_DIFF_REF_FRAME_ADD_MV(candidate_mi, ref_frame, ref_sign_bias,
112 refmv_count, mv_ref_list, Done);
117 !cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
118 ref_frame == LAST_FRAME) {
119 // Get base layer mv.
122 ->mvs[(mi_col >> 1) + (mi_row >> 1) * (cm->mi_cols >> 1)];
123 if (candidate->mv[0].as_int != INVALID_MV) {
124 base_mv->as_mv.row = (candidate->mv[0].as_mv.row * 2);
125 base_mv->as_mv.col = (candidate->mv[0].as_mv.col * 2);
126 clamp_mv_ref(&base_mv->as_mv, xd);
128 base_mv->as_int = INVALID_MV;
134 x->mbmi_ext->mode_context[ref_frame] = counter_to_context[context_counter];
137 for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
138 clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
143 static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
144 BLOCK_SIZE bsize, int mi_row, int mi_col,
145 int_mv *tmp_mv, int *rate_mv,
146 int64_t best_rd_sofar, int use_base_mv) {
147 MACROBLOCKD *xd = &x->e_mbd;
148 MODE_INFO *mi = xd->mi[0];
149 struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } };
150 const int step_param = cpi->sf.mv.fullpel_search_step_param;
151 const int sadpb = x->sadperbit16;
153 const int ref = mi->ref_frame[0];
154 const MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
158 const MvLimits tmp_mv_limits = x->mv_limits;
161 const YV12_BUFFER_CONFIG *scaled_ref_frame =
162 vp9_get_scaled_ref_frame(cpi, ref);
163 if (scaled_ref_frame) {
165 // Swap out the reference frame for a version that's been scaled to
166 // match the resolution of the current frame, allowing the existing
167 // motion search code to be used without additional modifications.
168 for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
169 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
171 vp9_set_mv_search_range(&x->mv_limits, &ref_mv);
173 assert(x->mv_best_ref_index[ref] <= 2);
174 if (x->mv_best_ref_index[ref] < 2)
175 mvp_full = x->mbmi_ext->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
177 mvp_full = x->pred_mv[ref];
185 center_mv = tmp_mv->as_mv;
187 vp9_full_pixel_search(
188 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method, sadpb,
189 cond_cost_list(cpi, cost_list), ¢er_mv, &tmp_mv->as_mv, INT_MAX, 0);
191 x->mv_limits = tmp_mv_limits;
193 // calculate the bit cost on motion vector
194 mvp_full.row = tmp_mv->as_mv.row * 8;
195 mvp_full.col = tmp_mv->as_mv.col * 8;
197 *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv, x->nmvjointcost, x->mvcost,
201 cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref]][INTER_OFFSET(NEWMV)];
203 !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) > best_rd_sofar);
206 const int subpel_force_stop = use_base_mv && cpi->sf.base_mv_aggressive
208 : cpi->sf.mv.subpel_force_stop;
209 cpi->find_fractional_mv_step(
210 x, &tmp_mv->as_mv, &ref_mv, cpi->common.allow_high_precision_mv,
211 x->errorperbit, &cpi->fn_ptr[bsize], subpel_force_stop,
212 cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
213 x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, 0, 0);
214 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost,
215 x->mvcost, MV_COST_WEIGHT);
218 if (scaled_ref_frame) {
220 for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
225 static void block_variance(const uint8_t *src, int src_stride,
226 const uint8_t *ref, int ref_stride, int w, int h,
227 unsigned int *sse, int *sum, int block_size,
228 #if CONFIG_VP9_HIGHBITDEPTH
229 int use_highbitdepth, vpx_bit_depth_t bd,
231 uint32_t *sse8x8, int *sum8x8, uint32_t *var8x8) {
237 for (i = 0; i < h; i += block_size) {
238 for (j = 0; j < w; j += block_size) {
239 #if CONFIG_VP9_HIGHBITDEPTH
240 if (use_highbitdepth) {
243 vpx_highbd_8_get8x8var(src + src_stride * i + j, src_stride,
244 ref + ref_stride * i + j, ref_stride,
245 &sse8x8[k], &sum8x8[k]);
248 vpx_highbd_10_get8x8var(src + src_stride * i + j, src_stride,
249 ref + ref_stride * i + j, ref_stride,
250 &sse8x8[k], &sum8x8[k]);
253 vpx_highbd_12_get8x8var(src + src_stride * i + j, src_stride,
254 ref + ref_stride * i + j, ref_stride,
255 &sse8x8[k], &sum8x8[k]);
259 vpx_get8x8var(src + src_stride * i + j, src_stride,
260 ref + ref_stride * i + j, ref_stride, &sse8x8[k],
264 vpx_get8x8var(src + src_stride * i + j, src_stride,
265 ref + ref_stride * i + j, ref_stride, &sse8x8[k],
270 var8x8[k] = sse8x8[k] - (uint32_t)(((int64_t)sum8x8[k] * sum8x8[k]) >> 6);
276 static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
277 unsigned int *sse_i, int *sum_i,
278 unsigned int *var_o, unsigned int *sse_o,
280 const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
281 const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
282 const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
285 for (i = 0; i < nh; i += 2) {
286 for (j = 0; j < nw; j += 2) {
287 sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
288 sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
289 sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
290 sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
291 var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >>
292 (b_width_log2_lookup[unit_size] +
293 b_height_log2_lookup[unit_size] + 6));
299 // Adjust the ac_thr according to speed, width, height and normalized sum
300 static int ac_thr_factor(const int speed, const int width, const int height,
301 const int norm_sum) {
302 if (speed >= 8 && norm_sum < 5) {
303 if (width <= 640 && height <= 480)
311 static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize,
312 MACROBLOCK *x, MACROBLOCKD *xd,
313 int *out_rate_sum, int64_t *out_dist_sum,
314 unsigned int *var_y, unsigned int *sse_y,
315 int mi_row, int mi_col, int *early_term) {
316 // Note our transform coeffs are 8 times an orthogonal transform.
317 // Hence quantizer step is also 8 times. To get effective quantizer
318 // we need to divide by 8 before sending to modeling function.
322 struct macroblock_plane *const p = &x->plane[0];
323 struct macroblockd_plane *const pd = &xd->plane[0];
324 const uint32_t dc_quant = pd->dequant[0];
325 const uint32_t ac_quant = pd->dequant[1];
326 const int64_t dc_thr = dc_quant * dc_quant >> 6;
327 int64_t ac_thr = ac_quant * ac_quant >> 6;
332 const int bw = b_width_log2_lookup[bsize];
333 const int bh = b_height_log2_lookup[bsize];
334 const int num8x8 = 1 << (bw + bh - 2);
335 unsigned int sse8x8[64] = { 0 };
336 int sum8x8[64] = { 0 };
337 unsigned int var8x8[64] = { 0 };
340 #if CONFIG_VP9_HIGHBITDEPTH
341 const vpx_bit_depth_t bd = cpi->common.bit_depth;
343 // Calculate variance for whole partition, and also save 8x8 blocks' variance
344 // to be used in following transform skipping test.
345 block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
346 4 << bw, 4 << bh, &sse, &sum, 8,
347 #if CONFIG_VP9_HIGHBITDEPTH
348 cpi->common.use_highbitdepth, bd,
350 sse8x8, sum8x8, var8x8);
351 var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
356 #if CONFIG_VP9_TEMPORAL_DENOISING
357 if (cpi->oxcf.noise_sensitivity > 0 && cpi->oxcf.speed > 5)
358 ac_thr = vp9_scale_acskip_thresh(ac_thr, cpi->denoiser.denoising_level,
359 (abs(sum) >> (bw + bh)));
361 ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
362 cpi->common.height, abs(sum) >> (bw + bh));
364 ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
365 cpi->common.height, abs(sum) >> (bw + bh));
368 if (cpi->common.tx_mode == TX_MODE_SELECT) {
369 if (sse > (var << 2))
370 tx_size = VPXMIN(max_txsize_lookup[bsize],
371 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
375 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
376 cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
378 else if (tx_size > TX_16X16)
381 tx_size = VPXMIN(max_txsize_lookup[bsize],
382 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
385 assert(tx_size >= TX_8X8);
386 xd->mi[0]->tx_size = tx_size;
388 // Evaluate if the partition block is a skippable block in Y plane.
390 unsigned int sse16x16[16] = { 0 };
391 int sum16x16[16] = { 0 };
392 unsigned int var16x16[16] = { 0 };
393 const int num16x16 = num8x8 >> 2;
395 unsigned int sse32x32[4] = { 0 };
396 int sum32x32[4] = { 0 };
397 unsigned int var32x32[4] = { 0 };
398 const int num32x32 = num8x8 >> 4;
402 const int num = (tx_size == TX_8X8)
404 : ((tx_size == TX_16X16) ? num16x16 : num32x32);
405 const unsigned int *sse_tx =
406 (tx_size == TX_8X8) ? sse8x8
407 : ((tx_size == TX_16X16) ? sse16x16 : sse32x32);
408 const unsigned int *var_tx =
409 (tx_size == TX_8X8) ? var8x8
410 : ((tx_size == TX_16X16) ? var16x16 : var32x32);
412 // Calculate variance if tx_size > TX_8X8
413 if (tx_size >= TX_16X16)
414 calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
416 if (tx_size == TX_32X32)
417 calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32,
421 x->skip_txfm[0] = SKIP_TXFM_NONE;
422 for (k = 0; k < num; k++)
423 // Check if all ac coefficients can be quantized to zero.
424 if (!(var_tx[k] < ac_thr || var == 0)) {
429 for (k = 0; k < num; k++)
430 // Check if dc coefficient can be quantized to zero.
431 if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
437 x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
439 if (dc_test) x->skip_txfm[0] = SKIP_TXFM_AC_DC;
440 } else if (dc_test) {
445 if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
446 int skip_uv[2] = { 0 };
447 unsigned int var_uv[2];
448 unsigned int sse_uv[2];
451 *out_dist_sum = sse << 4;
453 // Transform skipping test in UV planes.
454 for (i = 1; i <= 2; i++) {
455 struct macroblock_plane *const p = &x->plane[i];
456 struct macroblockd_plane *const pd = &xd->plane[i];
457 const TX_SIZE uv_tx_size = get_uv_tx_size(xd->mi[0], pd);
458 const BLOCK_SIZE unit_size = txsize_to_bsize[uv_tx_size];
459 const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, pd);
460 const int uv_bw = b_width_log2_lookup[uv_bsize];
461 const int uv_bh = b_height_log2_lookup[uv_bsize];
462 const int sf = (uv_bw - b_width_log2_lookup[unit_size]) +
463 (uv_bh - b_height_log2_lookup[unit_size]);
464 const uint32_t uv_dc_thr = pd->dequant[0] * pd->dequant[0] >> (6 - sf);
465 const uint32_t uv_ac_thr = pd->dequant[1] * pd->dequant[1] >> (6 - sf);
468 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, i);
469 var_uv[j] = cpi->fn_ptr[uv_bsize].vf(
470 p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, &sse_uv[j]);
472 if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
473 (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
479 // If the transform in YUV planes are skippable, the mode search checks
480 // fewer inter modes and doesn't check intra modes.
481 if (skip_uv[0] & skip_uv[1]) {
489 #if CONFIG_VP9_HIGHBITDEPTH
490 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
491 dc_quant >> (xd->bd - 5), &rate, &dist);
493 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
494 dc_quant >> 3, &rate, &dist);
495 #endif // CONFIG_VP9_HIGHBITDEPTH
499 *out_rate_sum = rate >> 1;
500 *out_dist_sum = dist << 3;
503 *out_dist_sum = (sse - var) << 4;
506 #if CONFIG_VP9_HIGHBITDEPTH
507 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
508 ac_quant >> (xd->bd - 5), &rate, &dist);
510 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
512 #endif // CONFIG_VP9_HIGHBITDEPTH
514 *out_rate_sum += rate;
515 *out_dist_sum += dist << 4;
518 static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
519 MACROBLOCKD *xd, int *out_rate_sum,
520 int64_t *out_dist_sum, unsigned int *var_y,
521 unsigned int *sse_y) {
522 // Note our transform coeffs are 8 times an orthogonal transform.
523 // Hence quantizer step is also 8 times. To get effective quantizer
524 // we need to divide by 8 before sending to modeling function.
528 struct macroblock_plane *const p = &x->plane[0];
529 struct macroblockd_plane *const pd = &xd->plane[0];
530 const int64_t dc_thr = p->quant_thred[0] >> 6;
531 const int64_t ac_thr = p->quant_thred[1] >> 6;
532 const uint32_t dc_quant = pd->dequant[0];
533 const uint32_t ac_quant = pd->dequant[1];
534 unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
535 pd->dst.buf, pd->dst.stride, &sse);
541 if (cpi->common.tx_mode == TX_MODE_SELECT) {
542 if (sse > (var << 2))
544 VPXMIN(max_txsize_lookup[bsize],
545 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
547 xd->mi[0]->tx_size = TX_8X8;
549 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
550 cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
551 xd->mi[0]->tx_size = TX_8X8;
552 else if (xd->mi[0]->tx_size > TX_16X16)
553 xd->mi[0]->tx_size = TX_16X16;
556 VPXMIN(max_txsize_lookup[bsize],
557 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
560 // Evaluate if the partition block is a skippable block in Y plane.
562 const BLOCK_SIZE unit_size = txsize_to_bsize[xd->mi[0]->tx_size];
563 const unsigned int num_blk_log2 =
564 (b_width_log2_lookup[bsize] - b_width_log2_lookup[unit_size]) +
565 (b_height_log2_lookup[bsize] - b_height_log2_lookup[unit_size]);
566 const unsigned int sse_tx = sse >> num_blk_log2;
567 const unsigned int var_tx = var >> num_blk_log2;
569 x->skip_txfm[0] = SKIP_TXFM_NONE;
570 // Check if all ac coefficients can be quantized to zero.
571 if (var_tx < ac_thr || var == 0) {
572 x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
573 // Check if dc coefficient can be quantized to zero.
574 if (sse_tx - var_tx < dc_thr || sse == var)
575 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
577 if (sse_tx - var_tx < dc_thr || sse == var) skip_dc = 1;
581 if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
583 *out_dist_sum = sse << 4;
588 #if CONFIG_VP9_HIGHBITDEPTH
589 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
590 dc_quant >> (xd->bd - 5), &rate, &dist);
592 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
593 dc_quant >> 3, &rate, &dist);
594 #endif // CONFIG_VP9_HIGHBITDEPTH
598 *out_rate_sum = rate >> 1;
599 *out_dist_sum = dist << 3;
602 *out_dist_sum = (sse - var) << 4;
605 #if CONFIG_VP9_HIGHBITDEPTH
606 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
607 ac_quant >> (xd->bd - 5), &rate, &dist);
609 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
611 #endif // CONFIG_VP9_HIGHBITDEPTH
613 *out_rate_sum += rate;
614 *out_dist_sum += dist << 4;
617 static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
618 int *skippable, int64_t *sse, BLOCK_SIZE bsize,
620 MACROBLOCKD *xd = &x->e_mbd;
621 const struct macroblockd_plane *pd = &xd->plane[0];
622 struct macroblock_plane *const p = &x->plane[0];
623 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
624 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
625 const int step = 1 << (tx_size << 1);
626 const int block_step = (1 << tx_size);
628 const int max_blocks_wide =
629 num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5);
630 const int max_blocks_high =
631 num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5);
633 const int bw = 4 * num_4x4_w;
634 const int bh = 4 * num_4x4_h;
636 #if CONFIG_VP9_HIGHBITDEPTH
637 // TODO(jingning): Implement the high bit-depth Hadamard transforms and
638 // remove this check condition.
639 // TODO(marpan): Use this path (model_rd) for 8bit under certain conditions
640 // for now, as the vp9_quantize_fp below for highbitdepth build is slow.
642 (cpi->oxcf.speed > 5 && cpi->common.frame_type != KEY_FRAME &&
643 bsize < BLOCK_32X32)) {
644 unsigned int var_y, sse_y;
646 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist,
654 if (cpi->sf.use_simple_block_yrd && cpi->common.frame_type != KEY_FRAME &&
655 bsize < BLOCK_32X32) {
656 unsigned int var_y, sse_y;
658 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist,
667 // The max tx_size passed in is TX_16X16.
668 assert(tx_size != TX_32X32);
670 vpx_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
671 pd->dst.buf, pd->dst.stride);
673 // Keep track of the row and column of the blocks we use so that we know
674 // if we are in the unrestricted motion border.
675 for (r = 0; r < max_blocks_high; r += block_step) {
676 for (c = 0; c < num_4x4_w; c += block_step) {
677 if (c < max_blocks_wide) {
678 const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
679 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
680 tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
681 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
682 uint16_t *const eob = &p->eobs[block];
683 const int diff_stride = bw;
684 const int16_t *src_diff;
685 src_diff = &p->src_diff[(r * diff_stride + c) << 2];
689 vpx_hadamard_16x16(src_diff, diff_stride, coeff);
690 vp9_quantize_fp(coeff, 256, x->skip_block, p->round_fp, p->quant_fp,
691 qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
695 vpx_hadamard_8x8(src_diff, diff_stride, coeff);
696 vp9_quantize_fp(coeff, 64, x->skip_block, p->round_fp, p->quant_fp,
697 qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
701 x->fwd_txm4x4(src_diff, coeff, diff_stride);
702 vp9_quantize_fp(coeff, 16, x->skip_block, p->round_fp, p->quant_fp,
703 qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
706 default: assert(0); break;
708 *skippable &= (*eob == 0);
716 if (*sse < INT64_MAX) {
717 *sse = (*sse << 6) >> 2;
719 this_rdc->dist = *sse;
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 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
730 tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
731 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
732 uint16_t *const eob = &p->eobs[block];
735 this_rdc->rate += (int)abs(qcoeff[0]);
737 this_rdc->rate += vpx_satd(qcoeff, step << 4);
739 this_rdc->dist += vp9_block_error_fp(coeff, dqcoeff, step << 4) >> 2;
745 // If skippable is set, rate gets clobbered later.
746 this_rdc->rate <<= (2 + VP9_PROB_COST_SHIFT);
747 this_rdc->rate += (eob_cost << VP9_PROB_COST_SHIFT);
750 static void model_rd_for_sb_uv(VP9_COMP *cpi, BLOCK_SIZE plane_bsize,
751 MACROBLOCK *x, MACROBLOCKD *xd,
752 RD_COST *this_rdc, unsigned int *var_y,
753 unsigned int *sse_y, int start_plane,
755 // Note our transform coeffs are 8 times an orthogonal transform.
756 // Hence quantizer step is also 8 times. To get effective quantizer
757 // we need to divide by 8 before sending to modeling function.
762 #if CONFIG_VP9_HIGHBITDEPTH
763 uint64_t tot_var = *var_y;
764 uint64_t tot_sse = *sse_y;
766 uint32_t tot_var = *var_y;
767 uint32_t tot_sse = *sse_y;
773 for (i = start_plane; i <= stop_plane; ++i) {
774 struct macroblock_plane *const p = &x->plane[i];
775 struct macroblockd_plane *const pd = &xd->plane[i];
776 const uint32_t dc_quant = pd->dequant[0];
777 const uint32_t ac_quant = pd->dequant[1];
778 const BLOCK_SIZE bs = plane_bsize;
780 if (!x->color_sensitivity[i - 1]) continue;
782 var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf,
783 pd->dst.stride, &sse);
788 #if CONFIG_VP9_HIGHBITDEPTH
789 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
790 dc_quant >> (xd->bd - 5), &rate, &dist);
792 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
793 dc_quant >> 3, &rate, &dist);
794 #endif // CONFIG_VP9_HIGHBITDEPTH
796 this_rdc->rate += rate >> 1;
797 this_rdc->dist += dist << 3;
799 #if CONFIG_VP9_HIGHBITDEPTH
800 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
801 ac_quant >> (xd->bd - 5), &rate, &dist);
803 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs], ac_quant >> 3,
805 #endif // CONFIG_VP9_HIGHBITDEPTH
807 this_rdc->rate += rate;
808 this_rdc->dist += dist << 4;
811 #if CONFIG_VP9_HIGHBITDEPTH
812 *var_y = tot_var > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_var;
813 *sse_y = tot_sse > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_sse;
820 static int get_pred_buffer(PRED_BUFFER *p, int len) {
823 for (i = 0; i < len; i++) {
832 static void free_pred_buffer(PRED_BUFFER *p) {
833 if (p != NULL) p->in_use = 0;
836 static void encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
837 int mi_row, int mi_col,
838 MV_REFERENCE_FRAME ref_frame,
839 PREDICTION_MODE this_mode, unsigned int var_y,
841 struct buf_2d yv12_mb[][MAX_MB_PLANE],
842 int *rate, int64_t *dist) {
843 MACROBLOCKD *xd = &x->e_mbd;
844 MODE_INFO *const mi = xd->mi[0];
845 const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
846 unsigned int var = var_y, sse = sse_y;
847 // Skipping threshold for ac.
848 unsigned int thresh_ac;
849 // Skipping threshold for dc.
850 unsigned int thresh_dc;
852 if (mi->mv[0].as_mv.row > 64 || mi->mv[0].as_mv.row < -64 ||
853 mi->mv[0].as_mv.col > 64 || mi->mv[0].as_mv.col < -64)
855 if (x->encode_breakout > 0 && motion_low == 1) {
856 // Set a maximum for threshold to avoid big PSNR loss in low bit rate
857 // case. Use extreme low threshold for static frames to limit
859 const unsigned int max_thresh = 36000;
860 // The encode_breakout input
861 const unsigned int min_thresh =
862 VPXMIN(((unsigned int)x->encode_breakout << 4), max_thresh);
863 #if CONFIG_VP9_HIGHBITDEPTH
864 const int shift = (xd->bd << 1) - 16;
867 // Calculate threshold according to dequant value.
868 thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) >> 3;
869 #if CONFIG_VP9_HIGHBITDEPTH
870 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
871 thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift);
873 #endif // CONFIG_VP9_HIGHBITDEPTH
874 thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
876 // Adjust ac threshold according to partition size.
878 8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
880 thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
881 #if CONFIG_VP9_HIGHBITDEPTH
882 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
883 thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift);
885 #endif // CONFIG_VP9_HIGHBITDEPTH
891 // Y skipping condition checking for ac and dc.
892 if (var <= thresh_ac && (sse - var) <= thresh_dc) {
893 unsigned int sse_u, sse_v;
894 unsigned int var_u, var_v;
895 unsigned int thresh_ac_uv = thresh_ac;
896 unsigned int thresh_dc_uv = thresh_dc;
902 // Skip UV prediction unless breakout is zero (lossless) to save
903 // computation with low impact on the result
904 if (x->encode_breakout == 0) {
905 xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
906 xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
907 vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
910 var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf, x->plane[1].src.stride,
911 xd->plane[1].dst.buf,
912 xd->plane[1].dst.stride, &sse_u);
914 // U skipping condition checking
915 if (((var_u << 2) <= thresh_ac_uv) && (sse_u - var_u <= thresh_dc_uv)) {
916 var_v = cpi->fn_ptr[uv_size].vf(
917 x->plane[2].src.buf, x->plane[2].src.stride, xd->plane[2].dst.buf,
918 xd->plane[2].dst.stride, &sse_v);
920 // V skipping condition checking
921 if (((var_v << 2) <= thresh_ac_uv) && (sse_v - var_v <= thresh_dc_uv)) {
924 // The cost of skip bit needs to be added.
925 *rate = cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
926 [INTER_OFFSET(this_mode)];
928 // More on this part of rate
929 // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
931 // Scaling factor for SSE from spatial domain to frequency
932 // domain is 16. Adjust distortion accordingly.
933 // TODO(yunqingwang): In this function, only y-plane dist is
935 *dist = (sse << 4); // + ((sse_u + sse_v) << 4);
937 // *disable_skip = 1;
943 struct estimate_block_intra_args {
946 PREDICTION_MODE mode;
951 static void estimate_block_intra(int plane, int block, int row, int col,
952 BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
954 struct estimate_block_intra_args *const args = arg;
955 VP9_COMP *const cpi = args->cpi;
956 MACROBLOCK *const x = args->x;
957 MACROBLOCKD *const xd = &x->e_mbd;
958 struct macroblock_plane *const p = &x->plane[0];
959 struct macroblockd_plane *const pd = &xd->plane[0];
960 const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
961 uint8_t *const src_buf_base = p->src.buf;
962 uint8_t *const dst_buf_base = pd->dst.buf;
963 const int src_stride = p->src.stride;
964 const int dst_stride = pd->dst.stride;
969 p->src.buf = &src_buf_base[4 * (row * src_stride + col)];
970 pd->dst.buf = &dst_buf_base[4 * (row * dst_stride + col)];
971 // Use source buffer as an approximation for the fully reconstructed buffer.
972 vp9_predict_intra_block(xd, b_width_log2_lookup[plane_bsize], tx_size,
973 args->mode, x->skip_encode ? p->src.buf : pd->dst.buf,
974 x->skip_encode ? src_stride : dst_stride, pd->dst.buf,
975 dst_stride, col, row, plane);
978 int64_t this_sse = INT64_MAX;
979 // TODO(jingning): This needs further refactoring.
980 block_yrd(cpi, x, &this_rdc, &args->skippable, &this_sse, bsize_tx,
981 VPXMIN(tx_size, TX_16X16));
983 unsigned int var = 0;
984 unsigned int sse = 0;
985 model_rd_for_sb_uv(cpi, plane_bsize, x, xd, &this_rdc, &var, &sse, plane,
989 p->src.buf = src_buf_base;
990 pd->dst.buf = dst_buf_base;
991 args->rdc->rate += this_rdc.rate;
992 args->rdc->dist += this_rdc.dist;
995 static const THR_MODES mode_idx[MAX_REF_FRAMES][4] = {
996 { THR_DC, THR_V_PRED, THR_H_PRED, THR_TM },
997 { THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV },
998 { THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG },
999 { THR_NEARESTA, THR_NEARA, THR_ZEROA, THR_NEWA },
1002 static const PREDICTION_MODE intra_mode_list[] = { DC_PRED, V_PRED, H_PRED,
1005 static int mode_offset(const PREDICTION_MODE mode) {
1006 if (mode >= NEARESTMV) {
1007 return INTER_OFFSET(mode);
1010 case DC_PRED: return 0;
1011 case V_PRED: return 1;
1012 case H_PRED: return 2;
1013 case TM_PRED: return 3;
1019 static INLINE int rd_less_than_thresh_row_mt(int64_t best_rd, int thresh,
1020 const int *const thresh_fact) {
1021 int is_rd_less_than_thresh;
1022 is_rd_less_than_thresh =
1023 best_rd < ((int64_t)thresh * (*thresh_fact) >> 5) || thresh == INT_MAX;
1024 return is_rd_less_than_thresh;
1027 static INLINE void update_thresh_freq_fact_row_mt(
1028 VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance,
1029 int thresh_freq_fact_idx, MV_REFERENCE_FRAME ref_frame,
1030 THR_MODES best_mode_idx, PREDICTION_MODE mode) {
1031 THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1032 int freq_fact_idx = thresh_freq_fact_idx + thr_mode_idx;
1033 int *freq_fact = &tile_data->row_base_thresh_freq_fact[freq_fact_idx];
1034 if (thr_mode_idx == best_mode_idx)
1035 *freq_fact -= (*freq_fact >> 4);
1036 else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
1037 ref_frame == LAST_FRAME && source_variance < 5) {
1038 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
1040 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
1041 cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1045 static INLINE void update_thresh_freq_fact(
1046 VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance,
1047 BLOCK_SIZE bsize, MV_REFERENCE_FRAME ref_frame, THR_MODES best_mode_idx,
1048 PREDICTION_MODE mode) {
1049 THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1050 int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
1051 if (thr_mode_idx == best_mode_idx)
1052 *freq_fact -= (*freq_fact >> 4);
1053 else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
1054 ref_frame == LAST_FRAME && source_variance < 5) {
1055 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
1057 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
1058 cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1062 void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
1063 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1064 MACROBLOCKD *const xd = &x->e_mbd;
1065 MODE_INFO *const mi = xd->mi[0];
1066 RD_COST this_rdc, best_rdc;
1067 PREDICTION_MODE this_mode;
1068 struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
1069 const TX_SIZE intra_tx_size =
1070 VPXMIN(max_txsize_lookup[bsize],
1071 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
1072 MODE_INFO *const mic = xd->mi[0];
1074 const MODE_INFO *above_mi = xd->above_mi;
1075 const MODE_INFO *left_mi = xd->left_mi;
1076 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1077 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1078 bmode_costs = cpi->y_mode_costs[A][L];
1081 vp9_rd_cost_reset(&best_rdc);
1082 vp9_rd_cost_reset(&this_rdc);
1084 mi->ref_frame[0] = INTRA_FRAME;
1085 // Initialize interp_filter here so we do not have to check for inter block
1086 // modes in get_pred_context_switchable_interp()
1087 mi->interp_filter = SWITCHABLE_FILTERS;
1089 mi->mv[0].as_int = INVALID_MV;
1090 mi->uv_mode = DC_PRED;
1091 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
1093 // Change the limit of this loop to add other intra prediction
1095 for (this_mode = DC_PRED; this_mode <= H_PRED; ++this_mode) {
1096 this_rdc.dist = this_rdc.rate = 0;
1097 args.mode = this_mode;
1099 args.rdc = &this_rdc;
1100 mi->tx_size = intra_tx_size;
1101 vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
1103 if (args.skippable) {
1104 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1105 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
1107 x->skip_txfm[0] = SKIP_TXFM_NONE;
1108 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
1110 this_rdc.rate += bmode_costs[this_mode];
1111 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1113 if (this_rdc.rdcost < best_rdc.rdcost) {
1114 best_rdc = this_rdc;
1115 mi->mode = this_mode;
1119 *rd_cost = best_rdc;
1122 static void init_ref_frame_cost(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1123 int ref_frame_cost[MAX_REF_FRAMES]) {
1124 vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
1125 vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1126 vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1128 ref_frame_cost[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1129 ref_frame_cost[LAST_FRAME] = ref_frame_cost[GOLDEN_FRAME] =
1130 ref_frame_cost[ALTREF_FRAME] = vp9_cost_bit(intra_inter_p, 1);
1132 ref_frame_cost[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1133 ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1134 ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1135 ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1136 ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1140 MV_REFERENCE_FRAME ref_frame;
1141 PREDICTION_MODE pred_mode;
1144 #define RT_INTER_MODES 12
1145 static const REF_MODE ref_mode_set[RT_INTER_MODES] = {
1146 { LAST_FRAME, ZEROMV }, { LAST_FRAME, NEARESTMV },
1147 { GOLDEN_FRAME, ZEROMV }, { LAST_FRAME, NEARMV },
1148 { LAST_FRAME, NEWMV }, { GOLDEN_FRAME, NEARESTMV },
1149 { GOLDEN_FRAME, NEARMV }, { GOLDEN_FRAME, NEWMV },
1150 { ALTREF_FRAME, ZEROMV }, { ALTREF_FRAME, NEARESTMV },
1151 { ALTREF_FRAME, NEARMV }, { ALTREF_FRAME, NEWMV }
1153 static const REF_MODE ref_mode_set_svc[RT_INTER_MODES] = {
1154 { LAST_FRAME, ZEROMV }, { GOLDEN_FRAME, ZEROMV },
1155 { LAST_FRAME, NEARESTMV }, { LAST_FRAME, NEARMV },
1156 { GOLDEN_FRAME, NEARESTMV }, { GOLDEN_FRAME, NEARMV },
1157 { LAST_FRAME, NEWMV }, { GOLDEN_FRAME, NEWMV }
1160 static int set_intra_cost_penalty(const VP9_COMP *const cpi, BLOCK_SIZE bsize) {
1161 const VP9_COMMON *const cm = &cpi->common;
1162 // Reduce the intra cost penalty for small blocks (<=16x16).
1164 (bsize <= BLOCK_16X16) ? ((bsize <= BLOCK_8X8) ? 4 : 2) : 0;
1165 if (cpi->noise_estimate.enabled && cpi->noise_estimate.level == kHigh)
1166 // Don't reduce intra cost penalty if estimated noise level is high.
1168 return vp9_get_intra_cost_penalty(cm->base_qindex, cm->y_dc_delta_q,
1173 static INLINE void find_predictors(
1174 VP9_COMP *cpi, MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame,
1175 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1176 int const_motion[MAX_REF_FRAMES], int *ref_frame_skip_mask,
1177 const int flag_list[4], TileDataEnc *tile_data, int mi_row, int mi_col,
1178 struct buf_2d yv12_mb[4][MAX_MB_PLANE], BLOCK_SIZE bsize,
1179 int force_skip_low_temp_var) {
1180 VP9_COMMON *const cm = &cpi->common;
1181 MACROBLOCKD *const xd = &x->e_mbd;
1182 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1183 TileInfo *const tile_info = &tile_data->tile_info;
1184 // TODO(jingning) placeholder for inter-frame non-RD mode decision.
1185 x->pred_mv_sad[ref_frame] = INT_MAX;
1186 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
1187 frame_mv[ZEROMV][ref_frame].as_int = 0;
1188 // this needs various further optimizations. to be continued..
1189 if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
1190 int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
1191 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
1192 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
1193 if (cm->use_prev_frame_mvs) {
1194 vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
1195 x->mbmi_ext->mode_context);
1197 const_motion[ref_frame] =
1198 mv_refs_rt(cpi, cm, x, xd, tile_info, xd->mi[0], ref_frame,
1199 candidates, &frame_mv[NEWMV][ref_frame], mi_row, mi_col,
1200 (int)(cpi->svc.use_base_mv && cpi->svc.spatial_layer_id));
1202 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1203 &frame_mv[NEARESTMV][ref_frame],
1204 &frame_mv[NEARMV][ref_frame]);
1205 // Early exit for golden frame if force_skip_low_temp_var is set.
1206 if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8 &&
1207 !(force_skip_low_temp_var && ref_frame == GOLDEN_FRAME)) {
1208 vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
1212 *ref_frame_skip_mask |= (1 << ref_frame);
1216 static void vp9_NEWMV_diff_bias(const NOISE_ESTIMATE *ne, MACROBLOCKD *xd,
1217 PREDICTION_MODE this_mode, RD_COST *this_rdc,
1218 BLOCK_SIZE bsize, int mv_row, int mv_col,
1219 int is_last_frame) {
1220 // Bias against MVs associated with NEWMV mode that are very different from
1221 // top/left neighbors.
1222 if (this_mode == NEWMV) {
1223 int al_mv_average_row;
1224 int al_mv_average_col;
1225 int left_row, left_col;
1226 int row_diff, col_diff;
1227 int above_mv_valid = 0;
1228 int left_mv_valid = 0;
1233 above_mv_valid = xd->above_mi->mv[0].as_int != INVALID_MV;
1234 above_row = xd->above_mi->mv[0].as_mv.row;
1235 above_col = xd->above_mi->mv[0].as_mv.col;
1238 left_mv_valid = xd->left_mi->mv[0].as_int != INVALID_MV;
1239 left_row = xd->left_mi->mv[0].as_mv.row;
1240 left_col = xd->left_mi->mv[0].as_mv.col;
1242 if (above_mv_valid && left_mv_valid) {
1243 al_mv_average_row = (above_row + left_row + 1) >> 1;
1244 al_mv_average_col = (above_col + left_col + 1) >> 1;
1245 } else if (above_mv_valid) {
1246 al_mv_average_row = above_row;
1247 al_mv_average_col = above_col;
1248 } else if (left_mv_valid) {
1249 al_mv_average_row = left_row;
1250 al_mv_average_col = left_col;
1252 al_mv_average_row = al_mv_average_col = 0;
1254 row_diff = (al_mv_average_row - mv_row);
1255 col_diff = (al_mv_average_col - mv_col);
1256 if (row_diff > 48 || row_diff < -48 || col_diff > 48 || col_diff < -48) {
1257 if (bsize > BLOCK_32X32)
1258 this_rdc->rdcost = this_rdc->rdcost << 1;
1260 this_rdc->rdcost = 3 * this_rdc->rdcost >> 1;
1263 // If noise estimation is enabled, and estimated level is above threshold,
1264 // add a bias to LAST reference with small motion, for large blocks.
1265 if (ne->enabled && ne->level >= kMedium && bsize >= BLOCK_32X32 &&
1266 is_last_frame && mv_row < 8 && mv_row > -8 && mv_col < 8 && mv_col > -8) {
1267 this_rdc->rdcost = 7 * this_rdc->rdcost >> 3;
1271 #if CONFIG_VP9_TEMPORAL_DENOISING
1272 static void vp9_pickmode_ctx_den_update(
1273 VP9_PICKMODE_CTX_DEN *ctx_den, int64_t zero_last_cost_orig,
1274 int ref_frame_cost[MAX_REF_FRAMES],
1275 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int reuse_inter_pred,
1276 TX_SIZE best_tx_size, PREDICTION_MODE best_mode,
1277 MV_REFERENCE_FRAME best_ref_frame, INTERP_FILTER best_pred_filter,
1278 uint8_t best_mode_skip_txfm) {
1279 ctx_den->zero_last_cost_orig = zero_last_cost_orig;
1280 ctx_den->ref_frame_cost = ref_frame_cost;
1281 ctx_den->frame_mv = frame_mv;
1282 ctx_den->reuse_inter_pred = reuse_inter_pred;
1283 ctx_den->best_tx_size = best_tx_size;
1284 ctx_den->best_mode = best_mode;
1285 ctx_den->best_ref_frame = best_ref_frame;
1286 ctx_den->best_pred_filter = best_pred_filter;
1287 ctx_den->best_mode_skip_txfm = best_mode_skip_txfm;
1290 static void recheck_zeromv_after_denoising(
1291 VP9_COMP *cpi, MODE_INFO *const mi, MACROBLOCK *x, MACROBLOCKD *const xd,
1292 VP9_DENOISER_DECISION decision, VP9_PICKMODE_CTX_DEN *ctx_den,
1293 struct buf_2d yv12_mb[4][MAX_MB_PLANE], RD_COST *best_rdc, BLOCK_SIZE bsize,
1294 int mi_row, int mi_col) {
1295 // If INTRA or GOLDEN reference was selected, re-evaluate ZEROMV on
1296 // denoised result. Only do this under noise conditions, and if rdcost of
1297 // ZEROMV onoriginal source is not significantly higher than rdcost of best
1299 if (cpi->noise_estimate.enabled && cpi->noise_estimate.level > kLow &&
1300 ctx_den->zero_last_cost_orig < (best_rdc->rdcost << 3) &&
1301 ((ctx_den->best_ref_frame == INTRA_FRAME && decision >= FILTER_BLOCK) ||
1302 (ctx_den->best_ref_frame == GOLDEN_FRAME &&
1303 cpi->svc.number_spatial_layers == 1 &&
1304 decision == FILTER_ZEROMV_BLOCK))) {
1305 // Check if we should pick ZEROMV on denoised signal.
1308 uint32_t var_y = UINT_MAX;
1309 uint32_t sse_y = UINT_MAX;
1312 mi->ref_frame[0] = LAST_FRAME;
1313 mi->ref_frame[1] = NONE;
1314 mi->mv[0].as_int = 0;
1315 mi->interp_filter = EIGHTTAP;
1316 xd->plane[0].pre[0] = yv12_mb[LAST_FRAME][0];
1317 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1318 model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
1319 this_rdc.rate = rate + ctx_den->ref_frame_cost[LAST_FRAME] +
1320 cpi->inter_mode_cost[x->mbmi_ext->mode_context[LAST_FRAME]]
1321 [INTER_OFFSET(ZEROMV)];
1322 this_rdc.dist = dist;
1323 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, rate, dist);
1324 // Don't switch to ZEROMV if the rdcost for ZEROMV on denoised source
1325 // is higher than best_ref mode (on original source).
1326 if (this_rdc.rdcost > best_rdc->rdcost) {
1327 this_rdc = *best_rdc;
1328 mi->mode = ctx_den->best_mode;
1329 mi->ref_frame[0] = ctx_den->best_ref_frame;
1330 mi->interp_filter = ctx_den->best_pred_filter;
1331 if (ctx_den->best_ref_frame == INTRA_FRAME) {
1332 mi->mv[0].as_int = INVALID_MV;
1333 mi->interp_filter = SWITCHABLE_FILTERS;
1334 } else if (ctx_den->best_ref_frame == GOLDEN_FRAME) {
1336 ctx_den->frame_mv[ctx_den->best_mode][ctx_den->best_ref_frame]
1338 if (ctx_den->reuse_inter_pred) {
1339 xd->plane[0].pre[0] = yv12_mb[GOLDEN_FRAME][0];
1340 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1343 mi->tx_size = ctx_den->best_tx_size;
1344 x->skip_txfm[0] = ctx_den->best_mode_skip_txfm;
1346 ctx_den->best_ref_frame = LAST_FRAME;
1347 *best_rdc = this_rdc;
1351 #endif // CONFIG_VP9_TEMPORAL_DENOISING
1353 static INLINE int get_force_skip_low_temp_var(uint8_t *variance_low, int mi_row,
1354 int mi_col, BLOCK_SIZE bsize) {
1355 const int i = (mi_row & 0x7) >> 1;
1356 const int j = (mi_col & 0x7) >> 1;
1357 int force_skip_low_temp_var = 0;
1358 // Set force_skip_low_temp_var based on the block size and block offset.
1359 if (bsize == BLOCK_64X64) {
1360 force_skip_low_temp_var = variance_low[0];
1361 } else if (bsize == BLOCK_64X32) {
1362 if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1363 force_skip_low_temp_var = variance_low[1];
1364 } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1365 force_skip_low_temp_var = variance_low[2];
1367 } else if (bsize == BLOCK_32X64) {
1368 if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1369 force_skip_low_temp_var = variance_low[3];
1370 } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1371 force_skip_low_temp_var = variance_low[4];
1373 } else if (bsize == BLOCK_32X32) {
1374 if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1375 force_skip_low_temp_var = variance_low[5];
1376 } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1377 force_skip_low_temp_var = variance_low[6];
1378 } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1379 force_skip_low_temp_var = variance_low[7];
1380 } else if ((mi_col & 0x7) && (mi_row & 0x7)) {
1381 force_skip_low_temp_var = variance_low[8];
1383 } else if (bsize == BLOCK_16X16) {
1384 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]];
1385 } else if (bsize == BLOCK_32X16) {
1386 // The col shift index for the second 16x16 block.
1387 const int j2 = ((mi_col + 2) & 0x7) >> 1;
1388 // Only if each 16x16 block inside has low temporal variance.
1389 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1390 variance_low[pos_shift_16x16[i][j2]];
1391 } else if (bsize == BLOCK_16X32) {
1392 // The row shift index for the second 16x16 block.
1393 const int i2 = ((mi_row + 2) & 0x7) >> 1;
1394 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1395 variance_low[pos_shift_16x16[i2][j]];
1397 return force_skip_low_temp_var;
1400 void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data,
1401 int mi_row, int mi_col, RD_COST *rd_cost,
1402 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1403 VP9_COMMON *const cm = &cpi->common;
1404 SPEED_FEATURES *const sf = &cpi->sf;
1405 const SVC *const svc = &cpi->svc;
1406 MACROBLOCKD *const xd = &x->e_mbd;
1407 MODE_INFO *const mi = xd->mi[0];
1408 struct macroblockd_plane *const pd = &xd->plane[0];
1409 PREDICTION_MODE best_mode = ZEROMV;
1410 MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
1411 MV_REFERENCE_FRAME usable_ref_frame;
1412 TX_SIZE best_tx_size = TX_SIZES;
1413 INTERP_FILTER best_pred_filter = EIGHTTAP;
1414 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1415 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1416 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1418 RD_COST this_rdc, best_rdc;
1419 uint8_t skip_txfm = SKIP_TXFM_NONE, best_mode_skip_txfm = SKIP_TXFM_NONE;
1420 // var_y and sse_y are saved to be used in skipping checking
1421 unsigned int var_y = UINT_MAX;
1422 unsigned int sse_y = UINT_MAX;
1423 const int intra_cost_penalty = set_intra_cost_penalty(cpi, bsize);
1424 int64_t inter_mode_thresh =
1425 RDCOST(x->rdmult, x->rddiv, intra_cost_penalty, 0);
1426 const int *const rd_threshes = cpi->rd.threshes[mi->segment_id][bsize];
1427 const int sb_row = mi_row >> MI_BLOCK_SIZE_LOG2;
1428 int thresh_freq_fact_idx = (sb_row * BLOCK_SIZES + bsize) * MAX_MODES;
1429 const int *const rd_thresh_freq_fact =
1430 (cpi->sf.adaptive_rd_thresh_row_mt)
1431 ? &(tile_data->row_base_thresh_freq_fact[thresh_freq_fact_idx])
1432 : tile_data->thresh_freq_fact[bsize];
1434 INTERP_FILTER filter_ref;
1435 const int bsl = mi_width_log2_lookup[bsize];
1436 const int pred_filter_search =
1437 cm->interp_filter == SWITCHABLE
1438 ? (((mi_row + mi_col) >> bsl) +
1439 get_chessboard_index(cm->current_video_frame)) &
1442 int const_motion[MAX_REF_FRAMES] = { 0 };
1443 const int bh = num_4x4_blocks_high_lookup[bsize] << 2;
1444 const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
1445 // For speed 6, the result of interp filter is reused later in actual encoding
1447 // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
1449 DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 64 * 64]);
1450 #if CONFIG_VP9_HIGHBITDEPTH
1451 DECLARE_ALIGNED(16, uint16_t, pred_buf_16[3 * 64 * 64]);
1453 struct buf_2d orig_dst = pd->dst;
1454 PRED_BUFFER *best_pred = NULL;
1455 PRED_BUFFER *this_mode_pred = NULL;
1456 const int pixels_in_block = bh * bw;
1457 int reuse_inter_pred = cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready;
1458 int ref_frame_skip_mask = 0;
1460 int best_pred_sad = INT_MAX;
1461 int best_early_term = 0;
1462 int ref_frame_cost[MAX_REF_FRAMES];
1463 int svc_force_zero_mode[3] = { 0 };
1464 int perform_intra_pred = 1;
1465 int use_golden_nonzeromv = 1;
1466 int force_skip_low_temp_var = 0;
1467 int skip_ref_find_pred[4] = { 0 };
1468 #if CONFIG_VP9_TEMPORAL_DENOISING
1469 VP9_PICKMODE_CTX_DEN ctx_den;
1470 int64_t zero_last_cost_orig = INT64_MAX;
1471 int denoise_svc_pickmode = 1;
1474 init_ref_frame_cost(cm, xd, ref_frame_cost);
1476 if (reuse_inter_pred) {
1478 for (i = 0; i < 3; i++) {
1479 #if CONFIG_VP9_HIGHBITDEPTH
1480 if (cm->use_highbitdepth)
1481 tmp[i].data = CONVERT_TO_BYTEPTR(&pred_buf_16[pixels_in_block * i]);
1483 tmp[i].data = &pred_buf[pixels_in_block * i];
1485 tmp[i].data = &pred_buf[pixels_in_block * i];
1486 #endif // CONFIG_VP9_HIGHBITDEPTH
1490 tmp[3].data = pd->dst.buf;
1491 tmp[3].stride = pd->dst.stride;
1495 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1498 // Instead of using vp9_get_pred_context_switchable_interp(xd) to assign
1499 // filter_ref, we use a less strict condition on assigning filter_ref.
1500 // This is to reduce the probabily of entering the flow of not assigning
1501 // filter_ref and then skip filter search.
1502 if (xd->above_mi && is_inter_block(xd->above_mi))
1503 filter_ref = xd->above_mi->interp_filter;
1504 else if (xd->left_mi && is_inter_block(xd->left_mi))
1505 filter_ref = xd->left_mi->interp_filter;
1507 filter_ref = cm->interp_filter;
1509 // initialize mode decisions
1510 vp9_rd_cost_reset(&best_rdc);
1511 vp9_rd_cost_reset(rd_cost);
1512 mi->sb_type = bsize;
1513 mi->ref_frame[0] = NONE;
1514 mi->ref_frame[1] = NONE;
1517 VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[cm->tx_mode]);
1519 if (sf->short_circuit_flat_blocks || sf->limit_newmv_early_exit) {
1520 #if CONFIG_VP9_HIGHBITDEPTH
1521 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
1522 x->source_variance = vp9_high_get_sby_perpixel_variance(
1523 cpi, &x->plane[0].src, bsize, xd->bd);
1525 #endif // CONFIG_VP9_HIGHBITDEPTH
1526 x->source_variance =
1527 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1530 #if CONFIG_VP9_TEMPORAL_DENOISING
1531 if (cpi->oxcf.noise_sensitivity > 0) {
1533 int layer = LAYER_IDS_TO_IDX(cpi->svc.spatial_layer_id,
1534 cpi->svc.temporal_layer_id,
1535 cpi->svc.number_temporal_layers);
1536 LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer];
1537 denoise_svc_pickmode = denoise_svc(cpi) && !lc->is_key_frame;
1539 if (cpi->denoiser.denoising_level > kDenLowLow && denoise_svc_pickmode)
1540 vp9_denoiser_reset_frame_stats(ctx);
1544 if (cpi->rc.frames_since_golden == 0 && !cpi->use_svc) {
1545 usable_ref_frame = LAST_FRAME;
1547 usable_ref_frame = GOLDEN_FRAME;
1550 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
1551 if (cpi->rc.alt_ref_gf_group || cpi->rc.is_src_frame_alt_ref)
1552 usable_ref_frame = ALTREF_FRAME;
1554 if (cpi->rc.is_src_frame_alt_ref) {
1555 skip_ref_find_pred[LAST_FRAME] = 1;
1556 skip_ref_find_pred[GOLDEN_FRAME] = 1;
1560 // For svc mode, on spatial_layer_id > 0: if the reference has different scale
1561 // constrain the inter mode to only test zero motion.
1562 if (cpi->use_svc && svc->force_zero_mode_spatial_ref &&
1563 cpi->svc.spatial_layer_id > 0) {
1564 if (cpi->ref_frame_flags & flag_list[LAST_FRAME]) {
1565 struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf;
1566 if (vp9_is_scaled(sf)) svc_force_zero_mode[LAST_FRAME - 1] = 1;
1568 if (cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) {
1569 struct scale_factors *const sf = &cm->frame_refs[GOLDEN_FRAME - 1].sf;
1570 if (vp9_is_scaled(sf)) svc_force_zero_mode[GOLDEN_FRAME - 1] = 1;
1574 if (cpi->sf.short_circuit_low_temp_var) {
1575 force_skip_low_temp_var =
1576 get_force_skip_low_temp_var(&x->variance_low[0], mi_row, mi_col, bsize);
1577 // If force_skip_low_temp_var is set, and for short circuit mode = 1 and 3,
1578 // skip golden reference.
1579 if ((cpi->sf.short_circuit_low_temp_var == 1 ||
1580 cpi->sf.short_circuit_low_temp_var == 3) &&
1581 force_skip_low_temp_var) {
1582 usable_ref_frame = LAST_FRAME;
1586 if (!((cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) &&
1587 !svc_force_zero_mode[GOLDEN_FRAME - 1] && !force_skip_low_temp_var))
1588 use_golden_nonzeromv = 0;
1590 if (cpi->oxcf.speed >= 8 && !cpi->use_svc &&
1591 ((cpi->rc.frames_since_golden + 1) < x->last_sb_high_content ||
1592 x->last_sb_high_content > 40))
1593 usable_ref_frame = LAST_FRAME;
1595 for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {
1596 if (!skip_ref_find_pred[ref_frame]) {
1597 find_predictors(cpi, x, ref_frame, frame_mv, const_motion,
1598 &ref_frame_skip_mask, flag_list, tile_data, mi_row,
1599 mi_col, yv12_mb, bsize, force_skip_low_temp_var);
1603 for (idx = 0; idx < RT_INTER_MODES; ++idx) {
1610 int this_early_term = 0;
1611 PREDICTION_MODE this_mode = ref_mode_set[idx].pred_mode;
1613 ref_frame = ref_mode_set[idx].ref_frame;
1616 this_mode = ref_mode_set_svc[idx].pred_mode;
1617 ref_frame = ref_mode_set_svc[idx].ref_frame;
1619 if (ref_frame > usable_ref_frame) continue;
1620 if (skip_ref_find_pred[ref_frame]) continue;
1622 if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
1623 this_mode != NEARESTMV) {
1627 if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode))) continue;
1629 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
1630 if (cpi->rc.is_src_frame_alt_ref &&
1631 (ref_frame != ALTREF_FRAME ||
1632 frame_mv[this_mode][ref_frame].as_int != 0))
1635 if (cpi->rc.alt_ref_gf_group &&
1636 cpi->rc.frames_since_golden > (cpi->rc.baseline_gf_interval >> 1) &&
1637 ref_frame == GOLDEN_FRAME &&
1638 frame_mv[this_mode][ref_frame].as_int != 0)
1641 if (cpi->rc.alt_ref_gf_group &&
1642 cpi->rc.frames_since_golden < (cpi->rc.baseline_gf_interval >> 1) &&
1643 ref_frame == ALTREF_FRAME &&
1644 frame_mv[this_mode][ref_frame].as_int != 0)
1648 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) continue;
1650 if (const_motion[ref_frame] && this_mode == NEARMV) continue;
1652 // Skip non-zeromv mode search for golden frame if force_skip_low_temp_var
1653 // is set. If nearestmv for golden frame is 0, zeromv mode will be skipped
1655 if (force_skip_low_temp_var && ref_frame == GOLDEN_FRAME &&
1656 frame_mv[this_mode][ref_frame].as_int != 0) {
1660 if ((cpi->sf.short_circuit_low_temp_var >= 2 ||
1661 (cpi->sf.short_circuit_low_temp_var == 1 && bsize == BLOCK_64X64)) &&
1662 force_skip_low_temp_var && ref_frame == LAST_FRAME &&
1663 this_mode == NEWMV) {
1668 if (svc_force_zero_mode[ref_frame - 1] &&
1669 frame_mv[this_mode][ref_frame].as_int != 0)
1673 if (sf->reference_masking &&
1674 !(frame_mv[this_mode][ref_frame].as_int == 0 &&
1675 ref_frame == LAST_FRAME)) {
1676 if (usable_ref_frame < ALTREF_FRAME) {
1677 if (!force_skip_low_temp_var && usable_ref_frame > LAST_FRAME) {
1678 i = (ref_frame == LAST_FRAME) ? GOLDEN_FRAME : LAST_FRAME;
1679 if ((cpi->ref_frame_flags & flag_list[i]))
1680 if (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[i] << 1))
1681 ref_frame_skip_mask |= (1 << ref_frame);
1683 } else if (!cpi->rc.is_src_frame_alt_ref &&
1684 !(frame_mv[this_mode][ref_frame].as_int == 0 &&
1685 ref_frame == ALTREF_FRAME)) {
1686 int ref1 = (ref_frame == GOLDEN_FRAME) ? LAST_FRAME : GOLDEN_FRAME;
1687 int ref2 = (ref_frame == ALTREF_FRAME) ? LAST_FRAME : ALTREF_FRAME;
1688 if (((cpi->ref_frame_flags & flag_list[ref1]) &&
1689 (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref1] << 1))) ||
1690 ((cpi->ref_frame_flags & flag_list[ref2]) &&
1691 (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref2] << 1))))
1692 ref_frame_skip_mask |= (1 << ref_frame);
1695 if (ref_frame_skip_mask & (1 << ref_frame)) continue;
1697 // Select prediction reference frames.
1698 for (i = 0; i < MAX_MB_PLANE; i++)
1699 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
1701 mi->ref_frame[0] = ref_frame;
1702 set_ref_ptrs(cm, xd, ref_frame, NONE);
1704 mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
1705 mode_rd_thresh = best_mode_skip_txfm ? rd_threshes[mode_index] << 1
1706 : rd_threshes[mode_index];
1708 // Increase mode_rd_thresh value for GOLDEN_FRAME for improved encoding
1709 // speed with little/no subjective quality loss.
1710 if (cpi->sf.bias_golden && ref_frame == GOLDEN_FRAME &&
1711 cpi->rc.frames_since_golden > 4)
1712 mode_rd_thresh = mode_rd_thresh << 3;
1714 if ((cpi->sf.adaptive_rd_thresh_row_mt &&
1715 rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
1716 &rd_thresh_freq_fact[mode_index])) ||
1717 (!cpi->sf.adaptive_rd_thresh_row_mt &&
1718 rd_less_than_thresh(
1719 best_rdc.rdcost, mode_rd_thresh,
1720 #if CONFIG_MULTITHREAD
1721 // Synchronization of this function
1722 // is only necessary when
1723 // adaptive_rd_thresh is > 0.
1724 cpi->sf.adaptive_rd_thresh ? tile_data->enc_row_mt_mutex : NULL,
1726 &rd_thresh_freq_fact[mode_index])))
1729 if (this_mode == NEWMV) {
1730 if (ref_frame > LAST_FRAME && !cpi->use_svc &&
1731 cpi->oxcf.rc_mode == VPX_CBR) {
1736 if (bsize < BLOCK_16X16) continue;
1738 tmp_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
1740 if (tmp_sad > x->pred_mv_sad[LAST_FRAME]) continue;
1741 if (tmp_sad + (num_pels_log2_lookup[bsize] << 4) > best_pred_sad)
1744 frame_mv[NEWMV][ref_frame].as_int = mi->mv[0].as_int;
1745 rate_mv = vp9_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv,
1746 &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1747 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1748 frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
1749 frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
1751 cpi->find_fractional_mv_step(
1752 x, &frame_mv[NEWMV][ref_frame].as_mv,
1753 &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1754 cpi->common.allow_high_precision_mv, x->errorperbit,
1755 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
1756 cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
1757 x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref_frame], NULL, 0,
1759 } else if (svc->use_base_mv && svc->spatial_layer_id) {
1760 if (frame_mv[NEWMV][ref_frame].as_int != INVALID_MV) {
1761 const int pre_stride = xd->plane[0].pre[0].stride;
1762 int base_mv_sad = INT_MAX;
1763 const float base_mv_bias = sf->base_mv_aggressive ? 1.5f : 1.0f;
1764 const uint8_t *const pre_buf =
1765 xd->plane[0].pre[0].buf +
1766 (frame_mv[NEWMV][ref_frame].as_mv.row >> 3) * pre_stride +
1767 (frame_mv[NEWMV][ref_frame].as_mv.col >> 3);
1768 base_mv_sad = cpi->fn_ptr[bsize].sdf(
1769 x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride);
1771 if (base_mv_sad < (int)(base_mv_bias * x->pred_mv_sad[ref_frame])) {
1772 // Base layer mv is good.
1773 if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1774 &frame_mv[NEWMV][ref_frame], &rate_mv,
1775 best_rdc.rdcost, 1)) {
1778 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1779 &frame_mv[NEWMV][ref_frame],
1780 &rate_mv, best_rdc.rdcost, 0)) {
1783 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1784 &frame_mv[NEWMV][ref_frame],
1785 &rate_mv, best_rdc.rdcost, 0)) {
1788 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1789 &frame_mv[NEWMV][ref_frame], &rate_mv,
1790 best_rdc.rdcost, 0)) {
1795 // If use_golden_nonzeromv is false, NEWMV mode is skipped for golden, no
1796 // need to compute best_pred_sad which is only used to skip golden NEWMV.
1797 if (use_golden_nonzeromv && this_mode == NEWMV && ref_frame == LAST_FRAME &&
1798 frame_mv[NEWMV][LAST_FRAME].as_int != INVALID_MV) {
1799 const int pre_stride = xd->plane[0].pre[0].stride;
1800 const uint8_t *const pre_buf =
1801 xd->plane[0].pre[0].buf +
1802 (frame_mv[NEWMV][LAST_FRAME].as_mv.row >> 3) * pre_stride +
1803 (frame_mv[NEWMV][LAST_FRAME].as_mv.col >> 3);
1804 best_pred_sad = cpi->fn_ptr[bsize].sdf(
1805 x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride);
1806 x->pred_mv_sad[LAST_FRAME] = best_pred_sad;
1809 if (this_mode != NEARESTMV &&
1810 frame_mv[this_mode][ref_frame].as_int ==
1811 frame_mv[NEARESTMV][ref_frame].as_int)
1814 mi->mode = this_mode;
1815 mi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
1817 // Search for the best prediction filter type, when the resulting
1818 // motion vector is at sub-pixel accuracy level for luma component, i.e.,
1819 // the last three bits are all zeros.
1820 if (reuse_inter_pred) {
1821 if (!this_mode_pred) {
1822 this_mode_pred = &tmp[3];
1824 this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
1825 pd->dst.buf = this_mode_pred->data;
1826 pd->dst.stride = bw;
1830 if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&
1831 pred_filter_search &&
1832 (ref_frame == LAST_FRAME ||
1833 (ref_frame == GOLDEN_FRAME &&
1834 (cpi->use_svc || cpi->oxcf.rc_mode == VPX_VBR))) &&
1835 (((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07) != 0)) {
1838 unsigned int pf_var[3];
1839 unsigned int pf_sse[3];
1840 TX_SIZE pf_tx_size[3];
1841 int64_t best_cost = INT64_MAX;
1842 INTERP_FILTER best_filter = SWITCHABLE, filter;
1843 PRED_BUFFER *current_pred = this_mode_pred;
1845 for (filter = EIGHTTAP; filter <= EIGHTTAP_SMOOTH; ++filter) {
1847 mi->interp_filter = filter;
1848 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1849 model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter], &pf_dist[filter],
1850 &pf_var[filter], &pf_sse[filter]);
1851 pf_rate[filter] += vp9_get_switchable_rate(cpi, xd);
1852 cost = RDCOST(x->rdmult, x->rddiv, pf_rate[filter], pf_dist[filter]);
1853 pf_tx_size[filter] = mi->tx_size;
1854 if (cost < best_cost) {
1855 best_filter = filter;
1857 skip_txfm = x->skip_txfm[0];
1859 if (reuse_inter_pred) {
1860 if (this_mode_pred != current_pred) {
1861 free_pred_buffer(this_mode_pred);
1862 this_mode_pred = current_pred;
1864 current_pred = &tmp[get_pred_buffer(tmp, 3)];
1865 pd->dst.buf = current_pred->data;
1866 pd->dst.stride = bw;
1871 if (reuse_inter_pred && this_mode_pred != current_pred)
1872 free_pred_buffer(current_pred);
1874 mi->interp_filter = best_filter;
1875 mi->tx_size = pf_tx_size[best_filter];
1876 this_rdc.rate = pf_rate[best_filter];
1877 this_rdc.dist = pf_dist[best_filter];
1878 var_y = pf_var[best_filter];
1879 sse_y = pf_sse[best_filter];
1880 x->skip_txfm[0] = skip_txfm;
1881 if (reuse_inter_pred) {
1882 pd->dst.buf = this_mode_pred->data;
1883 pd->dst.stride = this_mode_pred->stride;
1886 const int large_block = (x->sb_is_skin || cpi->oxcf.speed < 7)
1887 ? bsize > BLOCK_32X32
1888 : bsize >= BLOCK_32X32;
1889 mi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP : filter_ref;
1890 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1892 // For large partition blocks, extra testing is done.
1893 if (cpi->oxcf.rc_mode == VPX_CBR && large_block &&
1894 !cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) &&
1896 model_rd_for_sb_y_large(cpi, bsize, x, xd, &this_rdc.rate,
1897 &this_rdc.dist, &var_y, &sse_y, mi_row, mi_col,
1900 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
1905 if (!this_early_term) {
1906 this_sse = (int64_t)sse_y;
1907 block_yrd(cpi, x, &this_rdc, &is_skippable, &this_sse, bsize,
1908 VPXMIN(mi->tx_size, TX_16X16));
1909 x->skip_txfm[0] = is_skippable;
1911 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1913 if (RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist) <
1914 RDCOST(x->rdmult, x->rddiv, 0, this_sse)) {
1915 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
1917 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1918 this_rdc.dist = this_sse;
1919 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1923 if (cm->interp_filter == SWITCHABLE) {
1924 if ((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07)
1925 this_rdc.rate += vp9_get_switchable_rate(cpi, xd);
1928 this_rdc.rate += cm->interp_filter == SWITCHABLE
1929 ? vp9_get_switchable_rate(cpi, xd)
1931 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1934 if (x->color_sensitivity[0] || x->color_sensitivity[1]) {
1936 const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, &xd->plane[1]);
1937 if (x->color_sensitivity[0])
1938 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 1);
1939 if (x->color_sensitivity[1])
1940 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 2);
1941 model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &rdc_uv, &var_y, &sse_y, 1, 2);
1942 this_rdc.rate += rdc_uv.rate;
1943 this_rdc.dist += rdc_uv.dist;
1946 this_rdc.rate += rate_mv;
1947 this_rdc.rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
1948 [INTER_OFFSET(this_mode)];
1949 this_rdc.rate += ref_frame_cost[ref_frame];
1950 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1952 // Bias against NEWMV that is very different from its neighbors, and bias
1953 // to small motion-lastref for noisy input.
1954 if (cpi->oxcf.rc_mode == VPX_CBR && cpi->oxcf.speed >= 5 &&
1955 cpi->oxcf.content != VP9E_CONTENT_SCREEN) {
1956 vp9_NEWMV_diff_bias(&cpi->noise_estimate, xd, this_mode, &this_rdc, bsize,
1957 frame_mv[this_mode][ref_frame].as_mv.row,
1958 frame_mv[this_mode][ref_frame].as_mv.col,
1959 ref_frame == LAST_FRAME);
1962 // Skipping checking: test to see if this block can be reconstructed by
1964 if (cpi->allow_encode_breakout) {
1965 encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame, this_mode,
1966 var_y, sse_y, yv12_mb, &this_rdc.rate,
1969 this_rdc.rate += rate_mv;
1971 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1975 #if CONFIG_VP9_TEMPORAL_DENOISING
1976 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc_pickmode &&
1977 cpi->denoiser.denoising_level > kDenLowLow) {
1978 vp9_denoiser_update_frame_stats(mi, sse_y, this_mode, ctx);
1979 // Keep track of zero_last cost.
1980 if (ref_frame == LAST_FRAME && frame_mv[this_mode][ref_frame].as_int == 0)
1981 zero_last_cost_orig = this_rdc.rdcost;
1987 if (this_rdc.rdcost < best_rdc.rdcost || x->skip) {
1988 best_rdc = this_rdc;
1989 best_mode = this_mode;
1990 best_pred_filter = mi->interp_filter;
1991 best_tx_size = mi->tx_size;
1992 best_ref_frame = ref_frame;
1993 best_mode_skip_txfm = x->skip_txfm[0];
1994 best_early_term = this_early_term;
1996 if (reuse_inter_pred) {
1997 free_pred_buffer(best_pred);
1998 best_pred = this_mode_pred;
2001 if (reuse_inter_pred) free_pred_buffer(this_mode_pred);
2006 // If early termination flag is 1 and at least 2 modes are checked,
2007 // the mode search is terminated.
2008 if (best_early_term && idx > 0) {
2014 mi->mode = best_mode;
2015 mi->interp_filter = best_pred_filter;
2016 mi->tx_size = best_tx_size;
2017 mi->ref_frame[0] = best_ref_frame;
2018 mi->mv[0].as_int = frame_mv[best_mode][best_ref_frame].as_int;
2019 xd->mi[0]->bmi[0].as_mv[0].as_int = mi->mv[0].as_int;
2020 x->skip_txfm[0] = best_mode_skip_txfm;
2022 // For spatial enhancemanent layer: perform intra prediction only if base
2023 // layer is chosen as the reference. Always perform intra prediction if
2024 // LAST is the only reference or is_key_frame is set.
2025 if (cpi->svc.spatial_layer_id) {
2026 perform_intra_pred =
2027 cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame ||
2028 !(cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) ||
2029 (!cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
2030 svc_force_zero_mode[best_ref_frame - 1]);
2031 inter_mode_thresh = (inter_mode_thresh << 1) + inter_mode_thresh;
2033 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR &&
2034 cpi->rc.is_src_frame_alt_ref)
2035 perform_intra_pred = 0;
2036 // Perform intra prediction search, if the best SAD is above a certain
2038 if (best_rdc.rdcost == INT64_MAX ||
2039 ((!force_skip_low_temp_var || bsize < BLOCK_32X32) &&
2040 perform_intra_pred && !x->skip && best_rdc.rdcost > inter_mode_thresh &&
2041 bsize <= cpi->sf.max_intra_bsize && !x->skip_low_source_sad)) {
2042 struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
2044 TX_SIZE best_intra_tx_size = TX_SIZES;
2045 TX_SIZE intra_tx_size =
2046 VPXMIN(max_txsize_lookup[bsize],
2047 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
2048 if (cpi->oxcf.content != VP9E_CONTENT_SCREEN && intra_tx_size > TX_16X16)
2049 intra_tx_size = TX_16X16;
2051 if (reuse_inter_pred && best_pred != NULL) {
2052 if (best_pred->data == orig_dst.buf) {
2053 this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
2054 #if CONFIG_VP9_HIGHBITDEPTH
2055 if (cm->use_highbitdepth)
2056 vpx_highbd_convolve_copy(
2057 CAST_TO_BYTEPTR(CONVERT_TO_SHORTPTR(best_pred->data)),
2059 CAST_TO_BYTEPTR(CONVERT_TO_SHORTPTR(this_mode_pred->data)),
2060 this_mode_pred->stride, NULL, 0, NULL, 0, bw, bh, xd->bd);
2062 vpx_convolve_copy(best_pred->data, best_pred->stride,
2063 this_mode_pred->data, this_mode_pred->stride, NULL,
2064 0, NULL, 0, bw, bh);
2066 vpx_convolve_copy(best_pred->data, best_pred->stride,
2067 this_mode_pred->data, this_mode_pred->stride, NULL, 0,
2069 #endif // CONFIG_VP9_HIGHBITDEPTH
2070 best_pred = this_mode_pred;
2075 for (i = 0; i < 4; ++i) {
2076 const PREDICTION_MODE this_mode = intra_mode_list[i];
2077 THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)];
2078 int mode_rd_thresh = rd_threshes[mode_index];
2079 if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
2080 this_mode != DC_PRED) {
2084 if (!((1 << this_mode) & cpi->sf.intra_y_mode_bsize_mask[bsize]))
2087 if ((cpi->sf.adaptive_rd_thresh_row_mt &&
2088 rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
2089 &rd_thresh_freq_fact[mode_index])) ||
2090 (!cpi->sf.adaptive_rd_thresh_row_mt &&
2091 rd_less_than_thresh(
2092 best_rdc.rdcost, mode_rd_thresh,
2093 #if CONFIG_MULTITHREAD
2094 // Synchronization of this function
2095 // is only necessary when
2096 // adaptive_rd_thresh is > 0.
2097 cpi->sf.adaptive_rd_thresh ? tile_data->enc_row_mt_mutex : NULL,
2099 &rd_thresh_freq_fact[mode_index])))
2102 mi->mode = this_mode;
2103 mi->ref_frame[0] = INTRA_FRAME;
2104 this_rdc.dist = this_rdc.rate = 0;
2105 args.mode = this_mode;
2107 args.rdc = &this_rdc;
2108 mi->tx_size = intra_tx_size;
2109 vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
2111 // Check skip cost here since skippable is not set for for uv, this
2112 // mirrors the behavior used by inter
2113 if (args.skippable) {
2114 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
2115 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
2117 x->skip_txfm[0] = SKIP_TXFM_NONE;
2118 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
2120 // Inter and intra RD will mismatch in scale for non-screen content.
2121 if (cpi->oxcf.content == VP9E_CONTENT_SCREEN) {
2122 if (x->color_sensitivity[0])
2123 vp9_foreach_transformed_block_in_plane(xd, bsize, 1,
2124 estimate_block_intra, &args);
2125 if (x->color_sensitivity[1])
2126 vp9_foreach_transformed_block_in_plane(xd, bsize, 2,
2127 estimate_block_intra, &args);
2129 this_rdc.rate += cpi->mbmode_cost[this_mode];
2130 this_rdc.rate += ref_frame_cost[INTRA_FRAME];
2131 this_rdc.rate += intra_cost_penalty;
2133 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2135 if (this_rdc.rdcost < best_rdc.rdcost) {
2136 best_rdc = this_rdc;
2137 best_mode = this_mode;
2138 best_intra_tx_size = mi->tx_size;
2139 best_ref_frame = INTRA_FRAME;
2140 mi->uv_mode = this_mode;
2141 mi->mv[0].as_int = INVALID_MV;
2142 best_mode_skip_txfm = x->skip_txfm[0];
2146 // Reset mb_mode_info to the best inter mode.
2147 if (best_ref_frame != INTRA_FRAME) {
2148 mi->tx_size = best_tx_size;
2150 mi->tx_size = best_intra_tx_size;
2155 mi->mode = best_mode;
2156 mi->ref_frame[0] = best_ref_frame;
2157 x->skip_txfm[0] = best_mode_skip_txfm;
2159 if (!is_inter_block(mi)) {
2160 mi->interp_filter = SWITCHABLE_FILTERS;
2163 if (reuse_inter_pred && best_pred != NULL) {
2164 if (best_pred->data != orig_dst.buf && is_inter_mode(mi->mode)) {
2165 #if CONFIG_VP9_HIGHBITDEPTH
2166 if (cm->use_highbitdepth)
2167 vpx_highbd_convolve_copy(
2168 CAST_TO_BYTEPTR(CONVERT_TO_SHORTPTR(best_pred->data)),
2170 CAST_TO_BYTEPTR(CONVERT_TO_SHORTPTR(pd->dst.buf)), pd->dst.stride,
2171 NULL, 0, NULL, 0, bw, bh, xd->bd);
2173 vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2174 pd->dst.stride, NULL, 0, NULL, 0, bw, bh);
2176 vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2177 pd->dst.stride, NULL, 0, NULL, 0, bw, bh);
2178 #endif // CONFIG_VP9_HIGHBITDEPTH
2182 #if CONFIG_VP9_TEMPORAL_DENOISING
2183 if (cpi->oxcf.noise_sensitivity > 0 && cpi->resize_pending == 0 &&
2184 denoise_svc_pickmode && cpi->denoiser.denoising_level > kDenLowLow &&
2185 cpi->denoiser.reset == 0) {
2186 VP9_DENOISER_DECISION decision = COPY_BLOCK;
2187 vp9_pickmode_ctx_den_update(&ctx_den, zero_last_cost_orig, ref_frame_cost,
2188 frame_mv, reuse_inter_pred, best_tx_size,
2189 best_mode, best_ref_frame, best_pred_filter,
2190 best_mode_skip_txfm);
2191 vp9_denoiser_denoise(cpi, x, mi_row, mi_col, bsize, ctx, &decision);
2192 recheck_zeromv_after_denoising(cpi, mi, x, xd, decision, &ctx_den, yv12_mb,
2193 &best_rdc, bsize, mi_row, mi_col);
2194 best_ref_frame = ctx_den.best_ref_frame;
2198 if (cpi->sf.adaptive_rd_thresh) {
2199 THR_MODES best_mode_idx = mode_idx[best_ref_frame][mode_offset(mi->mode)];
2201 if (best_ref_frame == INTRA_FRAME) {
2202 // Only consider the modes that are included in the intra_mode_list.
2203 int intra_modes = sizeof(intra_mode_list) / sizeof(PREDICTION_MODE);
2206 // TODO(yunqingwang): Check intra mode mask and only update freq_fact
2207 // for those valid modes.
2208 for (i = 0; i < intra_modes; i++) {
2209 if (cpi->sf.adaptive_rd_thresh_row_mt)
2210 update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance,
2211 thresh_freq_fact_idx, INTRA_FRAME,
2212 best_mode_idx, intra_mode_list[i]);
2214 update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2215 INTRA_FRAME, best_mode_idx,
2216 intra_mode_list[i]);
2219 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2220 PREDICTION_MODE this_mode;
2221 if (best_ref_frame != ref_frame) continue;
2222 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2223 if (cpi->sf.adaptive_rd_thresh_row_mt)
2224 update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance,
2225 thresh_freq_fact_idx, ref_frame,
2226 best_mode_idx, this_mode);
2228 update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2229 ref_frame, best_mode_idx, this_mode);
2235 *rd_cost = best_rdc;
2238 void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x, int mi_row,
2239 int mi_col, RD_COST *rd_cost, BLOCK_SIZE bsize,
2240 PICK_MODE_CONTEXT *ctx) {
2241 VP9_COMMON *const cm = &cpi->common;
2242 SPEED_FEATURES *const sf = &cpi->sf;
2243 MACROBLOCKD *const xd = &x->e_mbd;
2244 MODE_INFO *const mi = xd->mi[0];
2245 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2246 const struct segmentation *const seg = &cm->seg;
2247 MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE;
2248 MV_REFERENCE_FRAME best_ref_frame = NONE;
2249 unsigned char segment_id = mi->segment_id;
2250 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2251 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2253 int64_t best_rd = INT64_MAX;
2254 b_mode_info bsi[MAX_REF_FRAMES][4];
2255 int ref_frame_skip_mask = 0;
2256 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
2257 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
2260 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
2261 ctx->pred_pixel_ready = 0;
2263 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2264 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2266 x->pred_mv_sad[ref_frame] = INT_MAX;
2268 if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
2269 int_mv *const candidates = mbmi_ext->ref_mvs[ref_frame];
2270 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2271 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf,
2273 vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
2274 mbmi_ext->mode_context);
2276 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2277 &dummy_mv[0], &dummy_mv[1]);
2279 ref_frame_skip_mask |= (1 << ref_frame);
2283 mi->sb_type = bsize;
2284 mi->tx_size = TX_4X4;
2285 mi->uv_mode = DC_PRED;
2286 mi->ref_frame[0] = LAST_FRAME;
2287 mi->ref_frame[1] = NONE;
2289 cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
2291 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2292 int64_t this_rd = 0;
2295 if (ref_frame_skip_mask & (1 << ref_frame)) continue;
2297 #if CONFIG_BETTER_HW_COMPATIBILITY
2298 if ((bsize == BLOCK_8X4 || bsize == BLOCK_4X8) && ref_frame > INTRA_FRAME &&
2299 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2303 // TODO(jingning, agrange): Scaling reference frame not supported for
2304 // sub8x8 blocks. Is this supported now?
2305 if (ref_frame > INTRA_FRAME &&
2306 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2309 // If the segment reference frame feature is enabled....
2310 // then do nothing if the current ref frame is not allowed..
2311 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
2312 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
2315 mi->ref_frame[0] = ref_frame;
2317 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
2319 // Select prediction reference frames.
2320 for (plane = 0; plane < MAX_MB_PLANE; plane++)
2321 xd->plane[plane].pre[0] = yv12_mb[ref_frame][plane];
2323 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2324 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2325 int_mv b_mv[MB_MODE_COUNT];
2326 int64_t b_best_rd = INT64_MAX;
2327 const int i = idy * 2 + idx;
2328 PREDICTION_MODE this_mode;
2330 unsigned int var_y, sse_y;
2332 struct macroblock_plane *p = &x->plane[0];
2333 struct macroblockd_plane *pd = &xd->plane[0];
2335 const struct buf_2d orig_src = p->src;
2336 const struct buf_2d orig_dst = pd->dst;
2337 struct buf_2d orig_pre[2];
2338 memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));
2340 // set buffer pointers for sub8x8 motion search.
2342 &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
2344 &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
2347 .buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
2349 b_mv[ZEROMV].as_int = 0;
2350 b_mv[NEWMV].as_int = INVALID_MV;
2351 vp9_append_sub8x8_mvs_for_idx(cm, xd, i, 0, mi_row, mi_col,
2352 &b_mv[NEARESTMV], &b_mv[NEARMV],
2353 mbmi_ext->mode_context);
2355 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2357 xd->mi[0]->bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
2359 if (this_mode == NEWMV) {
2360 const int step_param = cpi->sf.mv.fullpel_search_step_param;
2364 const MvLimits tmp_mv_limits = x->mv_limits;
2365 uint32_t dummy_dist;
2368 mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3;
2369 mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3;
2371 mvp_full.row = xd->mi[0]->bmi[0].as_mv[0].as_mv.row >> 3;
2372 mvp_full.col = xd->mi[0]->bmi[0].as_mv[0].as_mv.col >> 3;
2375 vp9_set_mv_search_range(&x->mv_limits,
2376 &mbmi_ext->ref_mvs[ref_frame][0].as_mv);
2378 vp9_full_pixel_search(
2379 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method,
2380 x->sadperbit4, cond_cost_list(cpi, cost_list),
2381 &mbmi_ext->ref_mvs[ref_frame][0].as_mv, &tmp_mv, INT_MAX, 0);
2383 x->mv_limits = tmp_mv_limits;
2385 // calculate the bit cost on motion vector
2386 mvp_full.row = tmp_mv.row * 8;
2387 mvp_full.col = tmp_mv.col * 8;
2389 b_rate += vp9_mv_bit_cost(
2390 &mvp_full, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2391 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2393 b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2394 [INTER_OFFSET(NEWMV)];
2395 if (RDCOST(x->rdmult, x->rddiv, b_rate, 0) > b_best_rd) continue;
2397 cpi->find_fractional_mv_step(
2398 x, &tmp_mv, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2399 cpi->common.allow_high_precision_mv, x->errorperbit,
2400 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
2401 cpi->sf.mv.subpel_iters_per_step,
2402 cond_cost_list(cpi, cost_list), x->nmvjointcost, x->mvcost,
2403 &dummy_dist, &x->pred_sse[ref_frame], NULL, 0, 0);
2405 xd->mi[0]->bmi[i].as_mv[0].as_mv = tmp_mv;
2407 b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2408 [INTER_OFFSET(this_mode)];
2411 #if CONFIG_VP9_HIGHBITDEPTH
2412 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2413 vp9_highbd_build_inter_predictor(
2414 pd->pre[0].buf, pd->pre[0].stride, pd->dst.buf, pd->dst.stride,
2415 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2416 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2417 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2418 mi_col * MI_SIZE + 4 * (i & 0x01),
2419 mi_row * MI_SIZE + 4 * (i >> 1), xd->bd);
2422 vp9_build_inter_predictor(
2423 pd->pre[0].buf, pd->pre[0].stride, pd->dst.buf, pd->dst.stride,
2424 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2425 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2426 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2427 mi_col * MI_SIZE + 4 * (i & 0x01),
2428 mi_row * MI_SIZE + 4 * (i >> 1));
2430 #if CONFIG_VP9_HIGHBITDEPTH
2434 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
2437 this_rdc.rate += b_rate;
2439 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2440 if (this_rdc.rdcost < b_best_rd) {
2441 b_best_rd = this_rdc.rdcost;
2442 bsi[ref_frame][i].as_mode = this_mode;
2443 bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0]->bmi[i].as_mv[0].as_mv;
2447 // restore source and prediction buffer pointers.
2449 pd->pre[0] = orig_pre[0];
2451 this_rd += b_best_rd;
2453 xd->mi[0]->bmi[i] = bsi[ref_frame][i];
2454 if (num_4x4_blocks_wide > 1) xd->mi[0]->bmi[i + 1] = xd->mi[0]->bmi[i];
2455 if (num_4x4_blocks_high > 1) xd->mi[0]->bmi[i + 2] = xd->mi[0]->bmi[i];
2457 } // loop through sub8x8 blocks
2459 if (this_rd < best_rd) {
2461 best_ref_frame = ref_frame;
2463 } // reference frames
2465 mi->tx_size = TX_4X4;
2466 mi->ref_frame[0] = best_ref_frame;
2467 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2468 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2469 const int block = idy * 2 + idx;
2470 xd->mi[0]->bmi[block] = bsi[best_ref_frame][block];
2471 if (num_4x4_blocks_wide > 1)
2472 xd->mi[0]->bmi[block + 1] = bsi[best_ref_frame][block];
2473 if (num_4x4_blocks_high > 1)
2474 xd->mi[0]->bmi[block + 2] = bsi[best_ref_frame][block];
2477 mi->mode = xd->mi[0]->bmi[3].as_mode;
2478 ctx->mic = *(xd->mi[0]);
2479 ctx->mbmi_ext = *x->mbmi_ext;
2480 ctx->skip_txfm[0] = SKIP_TXFM_NONE;
2482 // Dummy assignment for speed -5. No effect in speed -6.
2483 rd_cost->rdcost = best_rd;