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 int search_subpel = 1;
162 const YV12_BUFFER_CONFIG *scaled_ref_frame =
163 vp9_get_scaled_ref_frame(cpi, ref);
164 if (scaled_ref_frame) {
166 // Swap out the reference frame for a version that's been scaled to
167 // match the resolution of the current frame, allowing the existing
168 // motion search code to be used without additional modifications.
169 for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
170 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
172 vp9_set_mv_search_range(&x->mv_limits, &ref_mv);
174 // Limit motion vector for large lightning change.
175 if (cpi->oxcf.speed > 5 && x->lowvar_highsumdiff) {
176 x->mv_limits.col_min = VPXMAX(x->mv_limits.col_min, -10);
177 x->mv_limits.row_min = VPXMAX(x->mv_limits.row_min, -10);
178 x->mv_limits.col_max = VPXMIN(x->mv_limits.col_max, 10);
179 x->mv_limits.row_max = VPXMIN(x->mv_limits.row_max, 10);
182 assert(x->mv_best_ref_index[ref] <= 2);
183 if (x->mv_best_ref_index[ref] < 2)
184 mvp_full = x->mbmi_ext->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
186 mvp_full = x->pred_mv[ref];
194 center_mv = tmp_mv->as_mv;
196 vp9_full_pixel_search(
197 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method, sadpb,
198 cond_cost_list(cpi, cost_list), ¢er_mv, &tmp_mv->as_mv, INT_MAX, 0);
200 x->mv_limits = tmp_mv_limits;
202 // calculate the bit cost on motion vector
203 mvp_full.row = tmp_mv->as_mv.row * 8;
204 mvp_full.col = tmp_mv->as_mv.col * 8;
206 *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv, x->nmvjointcost, x->mvcost,
210 cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref]][INTER_OFFSET(NEWMV)];
212 !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) > best_rd_sofar);
214 // For SVC on non-reference frame, avoid subpel for (0, 0) motion.
215 if (cpi->use_svc && cpi->svc.non_reference_frame) {
216 if (mvp_full.row == 0 && mvp_full.col == 0) search_subpel = 0;
219 if (rv && search_subpel) {
220 const int subpel_force_stop = cpi->sf.mv.subpel_force_stop;
221 cpi->find_fractional_mv_step(
222 x, &tmp_mv->as_mv, &ref_mv, cpi->common.allow_high_precision_mv,
223 x->errorperbit, &cpi->fn_ptr[bsize], subpel_force_stop,
224 cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
225 x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, 0, 0);
226 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost,
227 x->mvcost, MV_COST_WEIGHT);
230 if (scaled_ref_frame) {
232 for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
237 static void block_variance(const uint8_t *src, int src_stride,
238 const uint8_t *ref, int ref_stride, int w, int h,
239 unsigned int *sse, int *sum, int block_size,
240 #if CONFIG_VP9_HIGHBITDEPTH
241 int use_highbitdepth, vpx_bit_depth_t bd,
243 uint32_t *sse8x8, int *sum8x8, uint32_t *var8x8) {
249 for (i = 0; i < h; i += block_size) {
250 for (j = 0; j < w; j += block_size) {
251 #if CONFIG_VP9_HIGHBITDEPTH
252 if (use_highbitdepth) {
255 vpx_highbd_8_get8x8var(src + src_stride * i + j, src_stride,
256 ref + ref_stride * i + j, ref_stride,
257 &sse8x8[k], &sum8x8[k]);
260 vpx_highbd_10_get8x8var(src + src_stride * i + j, src_stride,
261 ref + ref_stride * i + j, ref_stride,
262 &sse8x8[k], &sum8x8[k]);
265 vpx_highbd_12_get8x8var(src + src_stride * i + j, src_stride,
266 ref + ref_stride * i + j, ref_stride,
267 &sse8x8[k], &sum8x8[k]);
271 vpx_get8x8var(src + src_stride * i + j, src_stride,
272 ref + ref_stride * i + j, ref_stride, &sse8x8[k],
276 vpx_get8x8var(src + src_stride * i + j, src_stride,
277 ref + ref_stride * i + j, ref_stride, &sse8x8[k],
282 var8x8[k] = sse8x8[k] - (uint32_t)(((int64_t)sum8x8[k] * sum8x8[k]) >> 6);
288 static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
289 unsigned int *sse_i, int *sum_i,
290 unsigned int *var_o, unsigned int *sse_o,
292 const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
293 const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
294 const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
297 for (i = 0; i < nh; i += 2) {
298 for (j = 0; j < nw; j += 2) {
299 sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
300 sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
301 sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
302 sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
303 var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >>
304 (b_width_log2_lookup[unit_size] +
305 b_height_log2_lookup[unit_size] + 6));
311 // Adjust the ac_thr according to speed, width, height and normalized sum
312 static int ac_thr_factor(const int speed, const int width, const int height,
313 const int norm_sum) {
314 if (speed >= 8 && norm_sum < 5) {
315 if (width <= 640 && height <= 480)
323 static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize,
324 MACROBLOCK *x, MACROBLOCKD *xd,
325 int *out_rate_sum, int64_t *out_dist_sum,
326 unsigned int *var_y, unsigned int *sse_y,
327 int mi_row, int mi_col, int *early_term) {
328 // Note our transform coeffs are 8 times an orthogonal transform.
329 // Hence quantizer step is also 8 times. To get effective quantizer
330 // we need to divide by 8 before sending to modeling function.
334 struct macroblock_plane *const p = &x->plane[0];
335 struct macroblockd_plane *const pd = &xd->plane[0];
336 const uint32_t dc_quant = pd->dequant[0];
337 const uint32_t ac_quant = pd->dequant[1];
338 const int64_t dc_thr = dc_quant * dc_quant >> 6;
339 int64_t ac_thr = ac_quant * ac_quant >> 6;
344 const int bw = b_width_log2_lookup[bsize];
345 const int bh = b_height_log2_lookup[bsize];
346 const int num8x8 = 1 << (bw + bh - 2);
347 unsigned int sse8x8[64] = { 0 };
348 int sum8x8[64] = { 0 };
349 unsigned int var8x8[64] = { 0 };
352 #if CONFIG_VP9_HIGHBITDEPTH
353 const vpx_bit_depth_t bd = cpi->common.bit_depth;
355 // Calculate variance for whole partition, and also save 8x8 blocks' variance
356 // to be used in following transform skipping test.
357 block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
358 4 << bw, 4 << bh, &sse, &sum, 8,
359 #if CONFIG_VP9_HIGHBITDEPTH
360 cpi->common.use_highbitdepth, bd,
362 sse8x8, sum8x8, var8x8);
363 var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
368 #if CONFIG_VP9_TEMPORAL_DENOISING
369 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
371 ac_thr = vp9_scale_acskip_thresh(ac_thr, cpi->denoiser.denoising_level,
372 (abs(sum) >> (bw + bh)),
373 cpi->svc.temporal_layer_id);
375 ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
376 cpi->common.height, abs(sum) >> (bw + bh));
378 ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
379 cpi->common.height, abs(sum) >> (bw + bh));
382 if (cpi->common.tx_mode == TX_MODE_SELECT) {
383 if (sse > (var << 2))
384 tx_size = VPXMIN(max_txsize_lookup[bsize],
385 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
389 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
390 cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
392 else if (tx_size > TX_16X16)
395 tx_size = VPXMIN(max_txsize_lookup[bsize],
396 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
399 assert(tx_size >= TX_8X8);
400 xd->mi[0]->tx_size = tx_size;
402 // Evaluate if the partition block is a skippable block in Y plane.
404 unsigned int sse16x16[16] = { 0 };
405 int sum16x16[16] = { 0 };
406 unsigned int var16x16[16] = { 0 };
407 const int num16x16 = num8x8 >> 2;
409 unsigned int sse32x32[4] = { 0 };
410 int sum32x32[4] = { 0 };
411 unsigned int var32x32[4] = { 0 };
412 const int num32x32 = num8x8 >> 4;
416 const int num = (tx_size == TX_8X8)
418 : ((tx_size == TX_16X16) ? num16x16 : num32x32);
419 const unsigned int *sse_tx =
420 (tx_size == TX_8X8) ? sse8x8
421 : ((tx_size == TX_16X16) ? sse16x16 : sse32x32);
422 const unsigned int *var_tx =
423 (tx_size == TX_8X8) ? var8x8
424 : ((tx_size == TX_16X16) ? var16x16 : var32x32);
426 // Calculate variance if tx_size > TX_8X8
427 if (tx_size >= TX_16X16)
428 calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
430 if (tx_size == TX_32X32)
431 calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32,
435 x->skip_txfm[0] = SKIP_TXFM_NONE;
436 for (k = 0; k < num; k++)
437 // Check if all ac coefficients can be quantized to zero.
438 if (!(var_tx[k] < ac_thr || var == 0)) {
443 for (k = 0; k < num; k++)
444 // Check if dc coefficient can be quantized to zero.
445 if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
451 x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
453 if (dc_test) x->skip_txfm[0] = SKIP_TXFM_AC_DC;
454 } else if (dc_test) {
459 if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
460 int skip_uv[2] = { 0 };
461 unsigned int var_uv[2];
462 unsigned int sse_uv[2];
465 *out_dist_sum = sse << 4;
467 // Transform skipping test in UV planes.
468 for (i = 1; i <= 2; i++) {
469 if (cpi->oxcf.speed < 8 || x->color_sensitivity[i - 1]) {
470 struct macroblock_plane *const p = &x->plane[i];
471 struct macroblockd_plane *const pd = &xd->plane[i];
472 const TX_SIZE uv_tx_size = get_uv_tx_size(xd->mi[0], pd);
473 const BLOCK_SIZE unit_size = txsize_to_bsize[uv_tx_size];
474 const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, pd);
475 const int uv_bw = b_width_log2_lookup[uv_bsize];
476 const int uv_bh = b_height_log2_lookup[uv_bsize];
477 const int sf = (uv_bw - b_width_log2_lookup[unit_size]) +
478 (uv_bh - b_height_log2_lookup[unit_size]);
479 const uint32_t uv_dc_thr = pd->dequant[0] * pd->dequant[0] >> (6 - sf);
480 const uint32_t uv_ac_thr = pd->dequant[1] * pd->dequant[1] >> (6 - sf);
483 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, i);
484 var_uv[j] = cpi->fn_ptr[uv_bsize].vf(
485 p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, &sse_uv[j]);
487 if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
488 (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
497 // If the transform in YUV planes are skippable, the mode search checks
498 // fewer inter modes and doesn't check intra modes.
499 if (skip_uv[0] & skip_uv[1]) {
506 #if CONFIG_VP9_HIGHBITDEPTH
507 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
508 dc_quant >> (xd->bd - 5), &rate, &dist);
510 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
511 dc_quant >> 3, &rate, &dist);
512 #endif // CONFIG_VP9_HIGHBITDEPTH
516 *out_rate_sum = rate >> 1;
517 *out_dist_sum = dist << 3;
520 *out_dist_sum = (sse - var) << 4;
523 #if CONFIG_VP9_HIGHBITDEPTH
524 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
525 ac_quant >> (xd->bd - 5), &rate, &dist);
527 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
529 #endif // CONFIG_VP9_HIGHBITDEPTH
531 *out_rate_sum += rate;
532 *out_dist_sum += dist << 4;
535 static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
536 MACROBLOCKD *xd, int *out_rate_sum,
537 int64_t *out_dist_sum, unsigned int *var_y,
538 unsigned int *sse_y) {
539 // Note our transform coeffs are 8 times an orthogonal transform.
540 // Hence quantizer step is also 8 times. To get effective quantizer
541 // we need to divide by 8 before sending to modeling function.
545 struct macroblock_plane *const p = &x->plane[0];
546 struct macroblockd_plane *const pd = &xd->plane[0];
547 const int64_t dc_thr = p->quant_thred[0] >> 6;
548 const int64_t ac_thr = p->quant_thred[1] >> 6;
549 const uint32_t dc_quant = pd->dequant[0];
550 const uint32_t ac_quant = pd->dequant[1];
551 unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
552 pd->dst.buf, pd->dst.stride, &sse);
558 if (cpi->common.tx_mode == TX_MODE_SELECT) {
559 if (sse > (var << 2))
561 VPXMIN(max_txsize_lookup[bsize],
562 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
564 xd->mi[0]->tx_size = TX_8X8;
566 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
567 cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
568 xd->mi[0]->tx_size = TX_8X8;
569 else if (xd->mi[0]->tx_size > TX_16X16)
570 xd->mi[0]->tx_size = TX_16X16;
573 VPXMIN(max_txsize_lookup[bsize],
574 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
577 // Evaluate if the partition block is a skippable block in Y plane.
579 const BLOCK_SIZE unit_size = txsize_to_bsize[xd->mi[0]->tx_size];
580 const unsigned int num_blk_log2 =
581 (b_width_log2_lookup[bsize] - b_width_log2_lookup[unit_size]) +
582 (b_height_log2_lookup[bsize] - b_height_log2_lookup[unit_size]);
583 const unsigned int sse_tx = sse >> num_blk_log2;
584 const unsigned int var_tx = var >> num_blk_log2;
586 x->skip_txfm[0] = SKIP_TXFM_NONE;
587 // Check if all ac coefficients can be quantized to zero.
588 if (var_tx < ac_thr || var == 0) {
589 x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
590 // Check if dc coefficient can be quantized to zero.
591 if (sse_tx - var_tx < dc_thr || sse == var)
592 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
594 if (sse_tx - var_tx < dc_thr || sse == var) skip_dc = 1;
598 if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
600 *out_dist_sum = sse << 4;
605 #if CONFIG_VP9_HIGHBITDEPTH
606 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
607 dc_quant >> (xd->bd - 5), &rate, &dist);
609 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
610 dc_quant >> 3, &rate, &dist);
611 #endif // CONFIG_VP9_HIGHBITDEPTH
615 *out_rate_sum = rate >> 1;
616 *out_dist_sum = dist << 3;
619 *out_dist_sum = (sse - var) << 4;
622 #if CONFIG_VP9_HIGHBITDEPTH
623 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
624 ac_quant >> (xd->bd - 5), &rate, &dist);
626 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
628 #endif // CONFIG_VP9_HIGHBITDEPTH
630 *out_rate_sum += rate;
631 *out_dist_sum += dist << 4;
634 static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
635 int *skippable, int64_t *sse, BLOCK_SIZE bsize,
636 TX_SIZE tx_size, int rd_computed) {
637 MACROBLOCKD *xd = &x->e_mbd;
638 const struct macroblockd_plane *pd = &xd->plane[0];
639 struct macroblock_plane *const p = &x->plane[0];
640 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
641 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
642 const int step = 1 << (tx_size << 1);
643 const int block_step = (1 << tx_size);
645 const int max_blocks_wide =
646 num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5);
647 const int max_blocks_high =
648 num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5);
650 const int bw = 4 * num_4x4_w;
651 const int bh = 4 * num_4x4_h;
653 #if CONFIG_VP9_HIGHBITDEPTH
654 // TODO(jingning): Implement the high bit-depth Hadamard transforms and
655 // remove this check condition.
656 // TODO(marpan): Use this path (model_rd) for 8bit under certain conditions
657 // for now, as the vp9_quantize_fp below for highbitdepth build is slow.
659 (cpi->oxcf.speed > 5 && cpi->common.frame_type != KEY_FRAME &&
660 bsize < BLOCK_32X32)) {
661 unsigned int var_y, sse_y;
664 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist,
672 if (cpi->sf.use_simple_block_yrd && cpi->common.frame_type != KEY_FRAME &&
673 bsize < BLOCK_32X32) {
674 unsigned int var_y, sse_y;
677 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist,
686 // The max tx_size passed in is TX_16X16.
687 assert(tx_size != TX_32X32);
689 vpx_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
690 pd->dst.buf, pd->dst.stride);
692 // Keep track of the row and column of the blocks we use so that we know
693 // if we are in the unrestricted motion border.
694 for (r = 0; r < max_blocks_high; r += block_step) {
695 for (c = 0; c < num_4x4_w; c += block_step) {
696 if (c < max_blocks_wide) {
697 const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
698 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
699 tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
700 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
701 uint16_t *const eob = &p->eobs[block];
702 const int diff_stride = bw;
703 const int16_t *src_diff;
704 src_diff = &p->src_diff[(r * diff_stride + c) << 2];
708 vpx_hadamard_16x16(src_diff, diff_stride, coeff);
709 vp9_quantize_fp(coeff, 256, x->skip_block, p->round_fp, p->quant_fp,
710 qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
714 vpx_hadamard_8x8(src_diff, diff_stride, coeff);
715 vp9_quantize_fp(coeff, 64, x->skip_block, p->round_fp, p->quant_fp,
716 qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
720 x->fwd_txm4x4(src_diff, coeff, diff_stride);
721 vp9_quantize_fp(coeff, 16, x->skip_block, p->round_fp, p->quant_fp,
722 qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
725 default: assert(0); break;
727 *skippable &= (*eob == 0);
735 if (*sse < INT64_MAX) {
736 *sse = (*sse << 6) >> 2;
738 this_rdc->dist = *sse;
745 for (r = 0; r < max_blocks_high; r += block_step) {
746 for (c = 0; c < num_4x4_w; c += block_step) {
747 if (c < max_blocks_wide) {
748 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
749 tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
750 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
751 uint16_t *const eob = &p->eobs[block];
754 this_rdc->rate += (int)abs(qcoeff[0]);
756 this_rdc->rate += vpx_satd(qcoeff, step << 4);
758 this_rdc->dist += vp9_block_error_fp(coeff, dqcoeff, step << 4) >> 2;
764 // If skippable is set, rate gets clobbered later.
765 this_rdc->rate <<= (2 + VP9_PROB_COST_SHIFT);
766 this_rdc->rate += (eob_cost << VP9_PROB_COST_SHIFT);
769 static void model_rd_for_sb_uv(VP9_COMP *cpi, BLOCK_SIZE plane_bsize,
770 MACROBLOCK *x, MACROBLOCKD *xd,
771 RD_COST *this_rdc, unsigned int *var_y,
772 unsigned int *sse_y, int start_plane,
774 // Note our transform coeffs are 8 times an orthogonal transform.
775 // Hence quantizer step is also 8 times. To get effective quantizer
776 // we need to divide by 8 before sending to modeling function.
781 #if CONFIG_VP9_HIGHBITDEPTH
782 uint64_t tot_var = *var_y;
783 uint64_t tot_sse = *sse_y;
785 uint32_t tot_var = *var_y;
786 uint32_t tot_sse = *sse_y;
792 for (i = start_plane; i <= stop_plane; ++i) {
793 struct macroblock_plane *const p = &x->plane[i];
794 struct macroblockd_plane *const pd = &xd->plane[i];
795 const uint32_t dc_quant = pd->dequant[0];
796 const uint32_t ac_quant = pd->dequant[1];
797 const BLOCK_SIZE bs = plane_bsize;
799 if (!x->color_sensitivity[i - 1]) continue;
801 var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf,
802 pd->dst.stride, &sse);
807 #if CONFIG_VP9_HIGHBITDEPTH
808 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
809 dc_quant >> (xd->bd - 5), &rate, &dist);
811 vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
812 dc_quant >> 3, &rate, &dist);
813 #endif // CONFIG_VP9_HIGHBITDEPTH
815 this_rdc->rate += rate >> 1;
816 this_rdc->dist += dist << 3;
818 #if CONFIG_VP9_HIGHBITDEPTH
819 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
820 ac_quant >> (xd->bd - 5), &rate, &dist);
822 vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs], ac_quant >> 3,
824 #endif // CONFIG_VP9_HIGHBITDEPTH
826 this_rdc->rate += rate;
827 this_rdc->dist += dist << 4;
830 #if CONFIG_VP9_HIGHBITDEPTH
831 *var_y = tot_var > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_var;
832 *sse_y = tot_sse > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_sse;
839 static int get_pred_buffer(PRED_BUFFER *p, int len) {
842 for (i = 0; i < len; i++) {
851 static void free_pred_buffer(PRED_BUFFER *p) {
852 if (p != NULL) p->in_use = 0;
855 static void encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
856 int mi_row, int mi_col,
857 MV_REFERENCE_FRAME ref_frame,
858 PREDICTION_MODE this_mode, unsigned int var_y,
860 struct buf_2d yv12_mb[][MAX_MB_PLANE],
861 int *rate, int64_t *dist) {
862 MACROBLOCKD *xd = &x->e_mbd;
863 MODE_INFO *const mi = xd->mi[0];
864 const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
865 unsigned int var = var_y, sse = sse_y;
866 // Skipping threshold for ac.
867 unsigned int thresh_ac;
868 // Skipping threshold for dc.
869 unsigned int thresh_dc;
871 if (mi->mv[0].as_mv.row > 64 || mi->mv[0].as_mv.row < -64 ||
872 mi->mv[0].as_mv.col > 64 || mi->mv[0].as_mv.col < -64)
874 if (x->encode_breakout > 0 && motion_low == 1) {
875 // Set a maximum for threshold to avoid big PSNR loss in low bit rate
876 // case. Use extreme low threshold for static frames to limit
878 const unsigned int max_thresh = 36000;
879 // The encode_breakout input
880 const unsigned int min_thresh =
881 VPXMIN(((unsigned int)x->encode_breakout << 4), max_thresh);
882 #if CONFIG_VP9_HIGHBITDEPTH
883 const int shift = (xd->bd << 1) - 16;
886 // Calculate threshold according to dequant value.
887 thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) >> 3;
888 #if CONFIG_VP9_HIGHBITDEPTH
889 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
890 thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift);
892 #endif // CONFIG_VP9_HIGHBITDEPTH
893 thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
895 // Adjust ac threshold according to partition size.
897 8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
899 thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
900 #if CONFIG_VP9_HIGHBITDEPTH
901 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
902 thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift);
904 #endif // CONFIG_VP9_HIGHBITDEPTH
910 // Y skipping condition checking for ac and dc.
911 if (var <= thresh_ac && (sse - var) <= thresh_dc) {
912 unsigned int sse_u, sse_v;
913 unsigned int var_u, var_v;
914 unsigned int thresh_ac_uv = thresh_ac;
915 unsigned int thresh_dc_uv = thresh_dc;
921 // Skip UV prediction unless breakout is zero (lossless) to save
922 // computation with low impact on the result
923 if (x->encode_breakout == 0) {
924 xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
925 xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
926 vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
929 var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf, x->plane[1].src.stride,
930 xd->plane[1].dst.buf,
931 xd->plane[1].dst.stride, &sse_u);
933 // U skipping condition checking
934 if (((var_u << 2) <= thresh_ac_uv) && (sse_u - var_u <= thresh_dc_uv)) {
935 var_v = cpi->fn_ptr[uv_size].vf(
936 x->plane[2].src.buf, x->plane[2].src.stride, xd->plane[2].dst.buf,
937 xd->plane[2].dst.stride, &sse_v);
939 // V skipping condition checking
940 if (((var_v << 2) <= thresh_ac_uv) && (sse_v - var_v <= thresh_dc_uv)) {
943 // The cost of skip bit needs to be added.
944 *rate = cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
945 [INTER_OFFSET(this_mode)];
947 // More on this part of rate
948 // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
950 // Scaling factor for SSE from spatial domain to frequency
951 // domain is 16. Adjust distortion accordingly.
952 // TODO(yunqingwang): In this function, only y-plane dist is
954 *dist = (sse << 4); // + ((sse_u + sse_v) << 4);
956 // *disable_skip = 1;
962 struct estimate_block_intra_args {
965 PREDICTION_MODE mode;
970 static void estimate_block_intra(int plane, int block, int row, int col,
971 BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
973 struct estimate_block_intra_args *const args = arg;
974 VP9_COMP *const cpi = args->cpi;
975 MACROBLOCK *const x = args->x;
976 MACROBLOCKD *const xd = &x->e_mbd;
977 struct macroblock_plane *const p = &x->plane[0];
978 struct macroblockd_plane *const pd = &xd->plane[0];
979 const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
980 uint8_t *const src_buf_base = p->src.buf;
981 uint8_t *const dst_buf_base = pd->dst.buf;
982 const int src_stride = p->src.stride;
983 const int dst_stride = pd->dst.stride;
988 p->src.buf = &src_buf_base[4 * (row * src_stride + col)];
989 pd->dst.buf = &dst_buf_base[4 * (row * dst_stride + col)];
990 // Use source buffer as an approximation for the fully reconstructed buffer.
991 vp9_predict_intra_block(xd, b_width_log2_lookup[plane_bsize], tx_size,
992 args->mode, x->skip_encode ? p->src.buf : pd->dst.buf,
993 x->skip_encode ? src_stride : dst_stride, pd->dst.buf,
994 dst_stride, col, row, plane);
997 int64_t this_sse = INT64_MAX;
998 // TODO(jingning): This needs further refactoring.
999 block_yrd(cpi, x, &this_rdc, &args->skippable, &this_sse, bsize_tx,
1000 VPXMIN(tx_size, TX_16X16), 0);
1002 unsigned int var = 0;
1003 unsigned int sse = 0;
1004 model_rd_for_sb_uv(cpi, plane_bsize, x, xd, &this_rdc, &var, &sse, plane,
1008 p->src.buf = src_buf_base;
1009 pd->dst.buf = dst_buf_base;
1010 args->rdc->rate += this_rdc.rate;
1011 args->rdc->dist += this_rdc.dist;
1014 static const THR_MODES mode_idx[MAX_REF_FRAMES][4] = {
1015 { THR_DC, THR_V_PRED, THR_H_PRED, THR_TM },
1016 { THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV },
1017 { THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG },
1018 { THR_NEARESTA, THR_NEARA, THR_ZEROA, THR_NEWA },
1021 static const PREDICTION_MODE intra_mode_list[] = { DC_PRED, V_PRED, H_PRED,
1024 static int mode_offset(const PREDICTION_MODE mode) {
1025 if (mode >= NEARESTMV) {
1026 return INTER_OFFSET(mode);
1029 case DC_PRED: return 0;
1030 case V_PRED: return 1;
1031 case H_PRED: return 2;
1032 case TM_PRED: return 3;
1038 static INLINE int rd_less_than_thresh_row_mt(int64_t best_rd, int thresh,
1039 const int *const thresh_fact) {
1040 int is_rd_less_than_thresh;
1041 is_rd_less_than_thresh =
1042 best_rd < ((int64_t)thresh * (*thresh_fact) >> 5) || thresh == INT_MAX;
1043 return is_rd_less_than_thresh;
1046 static INLINE void update_thresh_freq_fact_row_mt(
1047 VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance,
1048 int thresh_freq_fact_idx, MV_REFERENCE_FRAME ref_frame,
1049 THR_MODES best_mode_idx, PREDICTION_MODE mode) {
1050 THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1051 int freq_fact_idx = thresh_freq_fact_idx + thr_mode_idx;
1052 int *freq_fact = &tile_data->row_base_thresh_freq_fact[freq_fact_idx];
1053 if (thr_mode_idx == best_mode_idx)
1054 *freq_fact -= (*freq_fact >> 4);
1055 else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
1056 ref_frame == LAST_FRAME && source_variance < 5) {
1057 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
1059 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
1060 cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1064 static INLINE void update_thresh_freq_fact(
1065 VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance,
1066 BLOCK_SIZE bsize, MV_REFERENCE_FRAME ref_frame, THR_MODES best_mode_idx,
1067 PREDICTION_MODE mode) {
1068 THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1069 int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
1070 if (thr_mode_idx == best_mode_idx)
1071 *freq_fact -= (*freq_fact >> 4);
1072 else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
1073 ref_frame == LAST_FRAME && source_variance < 5) {
1074 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
1076 *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
1077 cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1081 void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
1082 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1083 MACROBLOCKD *const xd = &x->e_mbd;
1084 MODE_INFO *const mi = xd->mi[0];
1085 RD_COST this_rdc, best_rdc;
1086 PREDICTION_MODE this_mode;
1087 struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
1088 const TX_SIZE intra_tx_size =
1089 VPXMIN(max_txsize_lookup[bsize],
1090 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
1091 MODE_INFO *const mic = xd->mi[0];
1093 const MODE_INFO *above_mi = xd->above_mi;
1094 const MODE_INFO *left_mi = xd->left_mi;
1095 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1096 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1097 bmode_costs = cpi->y_mode_costs[A][L];
1100 vp9_rd_cost_reset(&best_rdc);
1101 vp9_rd_cost_reset(&this_rdc);
1103 mi->ref_frame[0] = INTRA_FRAME;
1104 // Initialize interp_filter here so we do not have to check for inter block
1105 // modes in get_pred_context_switchable_interp()
1106 mi->interp_filter = SWITCHABLE_FILTERS;
1108 mi->mv[0].as_int = INVALID_MV;
1109 mi->uv_mode = DC_PRED;
1110 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
1112 // Change the limit of this loop to add other intra prediction
1114 for (this_mode = DC_PRED; this_mode <= H_PRED; ++this_mode) {
1115 this_rdc.dist = this_rdc.rate = 0;
1116 args.mode = this_mode;
1118 args.rdc = &this_rdc;
1119 mi->tx_size = intra_tx_size;
1120 vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
1122 if (args.skippable) {
1123 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1124 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
1126 x->skip_txfm[0] = SKIP_TXFM_NONE;
1127 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
1129 this_rdc.rate += bmode_costs[this_mode];
1130 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1132 if (this_rdc.rdcost < best_rdc.rdcost) {
1133 best_rdc = this_rdc;
1134 mi->mode = this_mode;
1138 *rd_cost = best_rdc;
1141 static void init_ref_frame_cost(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1142 int ref_frame_cost[MAX_REF_FRAMES]) {
1143 vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
1144 vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1145 vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1147 ref_frame_cost[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1148 ref_frame_cost[LAST_FRAME] = ref_frame_cost[GOLDEN_FRAME] =
1149 ref_frame_cost[ALTREF_FRAME] = vp9_cost_bit(intra_inter_p, 1);
1151 ref_frame_cost[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1152 ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1153 ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1154 ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1155 ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1159 MV_REFERENCE_FRAME ref_frame;
1160 PREDICTION_MODE pred_mode;
1163 #define RT_INTER_MODES 12
1164 static const REF_MODE ref_mode_set[RT_INTER_MODES] = {
1165 { LAST_FRAME, ZEROMV }, { LAST_FRAME, NEARESTMV },
1166 { GOLDEN_FRAME, ZEROMV }, { LAST_FRAME, NEARMV },
1167 { LAST_FRAME, NEWMV }, { GOLDEN_FRAME, NEARESTMV },
1168 { GOLDEN_FRAME, NEARMV }, { GOLDEN_FRAME, NEWMV },
1169 { ALTREF_FRAME, ZEROMV }, { ALTREF_FRAME, NEARESTMV },
1170 { ALTREF_FRAME, NEARMV }, { ALTREF_FRAME, NEWMV }
1172 static const REF_MODE ref_mode_set_svc[RT_INTER_MODES] = {
1173 { LAST_FRAME, ZEROMV }, { LAST_FRAME, NEARESTMV },
1174 { LAST_FRAME, NEARMV }, { GOLDEN_FRAME, ZEROMV },
1175 { GOLDEN_FRAME, NEARESTMV }, { GOLDEN_FRAME, NEARMV },
1176 { LAST_FRAME, NEWMV }, { GOLDEN_FRAME, NEWMV }
1179 static int set_intra_cost_penalty(const VP9_COMP *const cpi, BLOCK_SIZE bsize) {
1180 const VP9_COMMON *const cm = &cpi->common;
1181 // Reduce the intra cost penalty for small blocks (<=16x16).
1183 (bsize <= BLOCK_16X16) ? ((bsize <= BLOCK_8X8) ? 4 : 2) : 0;
1184 if (cpi->noise_estimate.enabled && cpi->noise_estimate.level == kHigh)
1185 // Don't reduce intra cost penalty if estimated noise level is high.
1187 return vp9_get_intra_cost_penalty(cm->base_qindex, cm->y_dc_delta_q,
1192 static INLINE void find_predictors(
1193 VP9_COMP *cpi, MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame,
1194 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1195 int const_motion[MAX_REF_FRAMES], int *ref_frame_skip_mask,
1196 const int flag_list[4], TileDataEnc *tile_data, int mi_row, int mi_col,
1197 struct buf_2d yv12_mb[4][MAX_MB_PLANE], BLOCK_SIZE bsize,
1198 int force_skip_low_temp_var) {
1199 VP9_COMMON *const cm = &cpi->common;
1200 MACROBLOCKD *const xd = &x->e_mbd;
1201 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1202 TileInfo *const tile_info = &tile_data->tile_info;
1203 // TODO(jingning) placeholder for inter-frame non-RD mode decision.
1204 x->pred_mv_sad[ref_frame] = INT_MAX;
1205 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
1206 frame_mv[ZEROMV][ref_frame].as_int = 0;
1207 // this needs various further optimizations. to be continued..
1208 if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
1209 int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
1210 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
1211 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
1212 if (cm->use_prev_frame_mvs) {
1213 vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
1214 x->mbmi_ext->mode_context);
1216 const_motion[ref_frame] =
1217 mv_refs_rt(cpi, cm, x, xd, tile_info, xd->mi[0], ref_frame,
1218 candidates, &frame_mv[NEWMV][ref_frame], mi_row, mi_col,
1219 (int)(cpi->svc.use_base_mv && cpi->svc.spatial_layer_id));
1221 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1222 &frame_mv[NEARESTMV][ref_frame],
1223 &frame_mv[NEARMV][ref_frame]);
1224 // Early exit for golden frame if force_skip_low_temp_var is set.
1225 if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8 &&
1226 !(force_skip_low_temp_var && ref_frame == GOLDEN_FRAME)) {
1227 vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
1231 *ref_frame_skip_mask |= (1 << ref_frame);
1235 static void vp9_NEWMV_diff_bias(const NOISE_ESTIMATE *ne, MACROBLOCKD *xd,
1236 PREDICTION_MODE this_mode, RD_COST *this_rdc,
1237 BLOCK_SIZE bsize, int mv_row, int mv_col,
1238 int is_last_frame, int lowvar_highsumdiff,
1240 // Bias against MVs associated with NEWMV mode that are very different from
1241 // top/left neighbors.
1242 if (this_mode == NEWMV) {
1243 int al_mv_average_row;
1244 int al_mv_average_col;
1245 int left_row, left_col;
1246 int row_diff, col_diff;
1247 int above_mv_valid = 0;
1248 int left_mv_valid = 0;
1253 above_mv_valid = xd->above_mi->mv[0].as_int != INVALID_MV;
1254 above_row = xd->above_mi->mv[0].as_mv.row;
1255 above_col = xd->above_mi->mv[0].as_mv.col;
1258 left_mv_valid = xd->left_mi->mv[0].as_int != INVALID_MV;
1259 left_row = xd->left_mi->mv[0].as_mv.row;
1260 left_col = xd->left_mi->mv[0].as_mv.col;
1262 if (above_mv_valid && left_mv_valid) {
1263 al_mv_average_row = (above_row + left_row + 1) >> 1;
1264 al_mv_average_col = (above_col + left_col + 1) >> 1;
1265 } else if (above_mv_valid) {
1266 al_mv_average_row = above_row;
1267 al_mv_average_col = above_col;
1268 } else if (left_mv_valid) {
1269 al_mv_average_row = left_row;
1270 al_mv_average_col = left_col;
1272 al_mv_average_row = al_mv_average_col = 0;
1274 row_diff = (al_mv_average_row - mv_row);
1275 col_diff = (al_mv_average_col - mv_col);
1276 if (row_diff > 48 || row_diff < -48 || col_diff > 48 || col_diff < -48) {
1277 if (bsize > BLOCK_32X32)
1278 this_rdc->rdcost = this_rdc->rdcost << 1;
1280 this_rdc->rdcost = 3 * this_rdc->rdcost >> 1;
1283 // If noise estimation is enabled, and estimated level is above threshold,
1284 // add a bias to LAST reference with small motion, for large blocks.
1285 if (ne->enabled && ne->level >= kMedium && bsize >= BLOCK_32X32 &&
1286 is_last_frame && mv_row < 8 && mv_row > -8 && mv_col < 8 && mv_col > -8)
1287 this_rdc->rdcost = 7 * (this_rdc->rdcost >> 3);
1288 else if (lowvar_highsumdiff && !is_skin && bsize >= BLOCK_16X16 &&
1289 is_last_frame && mv_row < 16 && mv_row > -16 && mv_col < 16 &&
1291 this_rdc->rdcost = 7 * (this_rdc->rdcost >> 3);
1294 #if CONFIG_VP9_TEMPORAL_DENOISING
1295 static void vp9_pickmode_ctx_den_update(
1296 VP9_PICKMODE_CTX_DEN *ctx_den, int64_t zero_last_cost_orig,
1297 int ref_frame_cost[MAX_REF_FRAMES],
1298 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int reuse_inter_pred,
1299 TX_SIZE best_tx_size, PREDICTION_MODE best_mode,
1300 MV_REFERENCE_FRAME best_ref_frame, INTERP_FILTER best_pred_filter,
1301 uint8_t best_mode_skip_txfm) {
1302 ctx_den->zero_last_cost_orig = zero_last_cost_orig;
1303 ctx_den->ref_frame_cost = ref_frame_cost;
1304 ctx_den->frame_mv = frame_mv;
1305 ctx_den->reuse_inter_pred = reuse_inter_pred;
1306 ctx_den->best_tx_size = best_tx_size;
1307 ctx_den->best_mode = best_mode;
1308 ctx_den->best_ref_frame = best_ref_frame;
1309 ctx_den->best_pred_filter = best_pred_filter;
1310 ctx_den->best_mode_skip_txfm = best_mode_skip_txfm;
1313 static void recheck_zeromv_after_denoising(
1314 VP9_COMP *cpi, MODE_INFO *const mi, MACROBLOCK *x, MACROBLOCKD *const xd,
1315 VP9_DENOISER_DECISION decision, VP9_PICKMODE_CTX_DEN *ctx_den,
1316 struct buf_2d yv12_mb[4][MAX_MB_PLANE], RD_COST *best_rdc, BLOCK_SIZE bsize,
1317 int mi_row, int mi_col) {
1318 // If INTRA or GOLDEN reference was selected, re-evaluate ZEROMV on
1319 // denoised result. Only do this under noise conditions, and if rdcost of
1320 // ZEROMV onoriginal source is not significantly higher than rdcost of best
1322 if (cpi->noise_estimate.enabled && cpi->noise_estimate.level > kLow &&
1323 ctx_den->zero_last_cost_orig < (best_rdc->rdcost << 3) &&
1324 ((ctx_den->best_ref_frame == INTRA_FRAME && decision >= FILTER_BLOCK) ||
1325 (ctx_den->best_ref_frame == GOLDEN_FRAME &&
1326 cpi->svc.number_spatial_layers == 1 &&
1327 decision == FILTER_ZEROMV_BLOCK))) {
1328 // Check if we should pick ZEROMV on denoised signal.
1331 uint32_t var_y = UINT_MAX;
1332 uint32_t sse_y = UINT_MAX;
1335 mi->ref_frame[0] = LAST_FRAME;
1336 mi->ref_frame[1] = NONE;
1337 mi->mv[0].as_int = 0;
1338 mi->interp_filter = EIGHTTAP;
1339 xd->plane[0].pre[0] = yv12_mb[LAST_FRAME][0];
1340 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1341 model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
1342 this_rdc.rate = rate + ctx_den->ref_frame_cost[LAST_FRAME] +
1343 cpi->inter_mode_cost[x->mbmi_ext->mode_context[LAST_FRAME]]
1344 [INTER_OFFSET(ZEROMV)];
1345 this_rdc.dist = dist;
1346 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, rate, dist);
1347 // Don't switch to ZEROMV if the rdcost for ZEROMV on denoised source
1348 // is higher than best_ref mode (on original source).
1349 if (this_rdc.rdcost > best_rdc->rdcost) {
1350 this_rdc = *best_rdc;
1351 mi->mode = ctx_den->best_mode;
1352 mi->ref_frame[0] = ctx_den->best_ref_frame;
1353 mi->interp_filter = ctx_den->best_pred_filter;
1354 if (ctx_den->best_ref_frame == INTRA_FRAME) {
1355 mi->mv[0].as_int = INVALID_MV;
1356 mi->interp_filter = SWITCHABLE_FILTERS;
1357 } else if (ctx_den->best_ref_frame == GOLDEN_FRAME) {
1359 ctx_den->frame_mv[ctx_den->best_mode][ctx_den->best_ref_frame]
1361 if (ctx_den->reuse_inter_pred) {
1362 xd->plane[0].pre[0] = yv12_mb[GOLDEN_FRAME][0];
1363 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1366 mi->tx_size = ctx_den->best_tx_size;
1367 x->skip_txfm[0] = ctx_den->best_mode_skip_txfm;
1369 ctx_den->best_ref_frame = LAST_FRAME;
1370 *best_rdc = this_rdc;
1374 #endif // CONFIG_VP9_TEMPORAL_DENOISING
1376 static INLINE int get_force_skip_low_temp_var(uint8_t *variance_low, int mi_row,
1377 int mi_col, BLOCK_SIZE bsize) {
1378 const int i = (mi_row & 0x7) >> 1;
1379 const int j = (mi_col & 0x7) >> 1;
1380 int force_skip_low_temp_var = 0;
1381 // Set force_skip_low_temp_var based on the block size and block offset.
1382 if (bsize == BLOCK_64X64) {
1383 force_skip_low_temp_var = variance_low[0];
1384 } else if (bsize == BLOCK_64X32) {
1385 if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1386 force_skip_low_temp_var = variance_low[1];
1387 } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1388 force_skip_low_temp_var = variance_low[2];
1390 } else if (bsize == BLOCK_32X64) {
1391 if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1392 force_skip_low_temp_var = variance_low[3];
1393 } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1394 force_skip_low_temp_var = variance_low[4];
1396 } else if (bsize == BLOCK_32X32) {
1397 if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1398 force_skip_low_temp_var = variance_low[5];
1399 } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1400 force_skip_low_temp_var = variance_low[6];
1401 } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1402 force_skip_low_temp_var = variance_low[7];
1403 } else if ((mi_col & 0x7) && (mi_row & 0x7)) {
1404 force_skip_low_temp_var = variance_low[8];
1406 } else if (bsize == BLOCK_16X16) {
1407 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]];
1408 } else if (bsize == BLOCK_32X16) {
1409 // The col shift index for the second 16x16 block.
1410 const int j2 = ((mi_col + 2) & 0x7) >> 1;
1411 // Only if each 16x16 block inside has low temporal variance.
1412 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1413 variance_low[pos_shift_16x16[i][j2]];
1414 } else if (bsize == BLOCK_16X32) {
1415 // The row shift index for the second 16x16 block.
1416 const int i2 = ((mi_row + 2) & 0x7) >> 1;
1417 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1418 variance_low[pos_shift_16x16[i2][j]];
1420 return force_skip_low_temp_var;
1423 void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data,
1424 int mi_row, int mi_col, RD_COST *rd_cost,
1425 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1426 VP9_COMMON *const cm = &cpi->common;
1427 SPEED_FEATURES *const sf = &cpi->sf;
1428 const SVC *const svc = &cpi->svc;
1429 MACROBLOCKD *const xd = &x->e_mbd;
1430 MODE_INFO *const mi = xd->mi[0];
1431 struct macroblockd_plane *const pd = &xd->plane[0];
1432 PREDICTION_MODE best_mode = ZEROMV;
1433 MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
1434 MV_REFERENCE_FRAME usable_ref_frame;
1435 TX_SIZE best_tx_size = TX_SIZES;
1436 INTERP_FILTER best_pred_filter = EIGHTTAP;
1437 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1438 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1439 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1441 RD_COST this_rdc, best_rdc;
1442 uint8_t skip_txfm = SKIP_TXFM_NONE, best_mode_skip_txfm = SKIP_TXFM_NONE;
1443 // var_y and sse_y are saved to be used in skipping checking
1444 unsigned int var_y = UINT_MAX;
1445 unsigned int sse_y = UINT_MAX;
1446 const int intra_cost_penalty = set_intra_cost_penalty(cpi, bsize);
1447 int64_t inter_mode_thresh =
1448 RDCOST(x->rdmult, x->rddiv, intra_cost_penalty, 0);
1449 const int *const rd_threshes = cpi->rd.threshes[mi->segment_id][bsize];
1450 const int sb_row = mi_row >> MI_BLOCK_SIZE_LOG2;
1451 int thresh_freq_fact_idx = (sb_row * BLOCK_SIZES + bsize) * MAX_MODES;
1452 const int *const rd_thresh_freq_fact =
1453 (cpi->sf.adaptive_rd_thresh_row_mt)
1454 ? &(tile_data->row_base_thresh_freq_fact[thresh_freq_fact_idx])
1455 : tile_data->thresh_freq_fact[bsize];
1457 INTERP_FILTER filter_ref;
1458 const int bsl = mi_width_log2_lookup[bsize];
1459 const int pred_filter_search =
1460 cm->interp_filter == SWITCHABLE
1461 ? (((mi_row + mi_col) >> bsl) +
1462 get_chessboard_index(cm->current_video_frame)) &
1465 int const_motion[MAX_REF_FRAMES] = { 0 };
1466 const int bh = num_4x4_blocks_high_lookup[bsize] << 2;
1467 const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
1468 // For speed 6, the result of interp filter is reused later in actual encoding
1470 // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
1472 DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 64 * 64]);
1473 #if CONFIG_VP9_HIGHBITDEPTH
1474 DECLARE_ALIGNED(16, uint16_t, pred_buf_16[3 * 64 * 64]);
1476 struct buf_2d orig_dst = pd->dst;
1477 PRED_BUFFER *best_pred = NULL;
1478 PRED_BUFFER *this_mode_pred = NULL;
1479 const int pixels_in_block = bh * bw;
1480 int reuse_inter_pred = cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready;
1481 int ref_frame_skip_mask = 0;
1483 int best_pred_sad = INT_MAX;
1484 int best_early_term = 0;
1485 int ref_frame_cost[MAX_REF_FRAMES];
1486 int svc_force_zero_mode[3] = { 0 };
1487 int perform_intra_pred = 1;
1488 int use_golden_nonzeromv = 1;
1489 int force_skip_low_temp_var = 0;
1490 int skip_ref_find_pred[4] = { 0 };
1491 unsigned int sse_zeromv_normalized = UINT_MAX;
1492 unsigned int thresh_svc_skip_golden = 500;
1493 #if CONFIG_VP9_TEMPORAL_DENOISING
1494 VP9_PICKMODE_CTX_DEN ctx_den;
1495 int64_t zero_last_cost_orig = INT64_MAX;
1496 int denoise_svc_pickmode = 1;
1498 INTERP_FILTER filter_gf_svc = EIGHTTAP;
1500 init_ref_frame_cost(cm, xd, ref_frame_cost);
1502 if (reuse_inter_pred) {
1504 for (i = 0; i < 3; i++) {
1505 #if CONFIG_VP9_HIGHBITDEPTH
1506 if (cm->use_highbitdepth)
1507 tmp[i].data = CONVERT_TO_BYTEPTR(&pred_buf_16[pixels_in_block * i]);
1509 tmp[i].data = &pred_buf[pixels_in_block * i];
1511 tmp[i].data = &pred_buf[pixels_in_block * i];
1512 #endif // CONFIG_VP9_HIGHBITDEPTH
1516 tmp[3].data = pd->dst.buf;
1517 tmp[3].stride = pd->dst.stride;
1521 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1524 // Instead of using vp9_get_pred_context_switchable_interp(xd) to assign
1525 // filter_ref, we use a less strict condition on assigning filter_ref.
1526 // This is to reduce the probabily of entering the flow of not assigning
1527 // filter_ref and then skip filter search.
1528 if (xd->above_mi && is_inter_block(xd->above_mi))
1529 filter_ref = xd->above_mi->interp_filter;
1530 else if (xd->left_mi && is_inter_block(xd->left_mi))
1531 filter_ref = xd->left_mi->interp_filter;
1533 filter_ref = cm->interp_filter;
1535 // initialize mode decisions
1536 vp9_rd_cost_reset(&best_rdc);
1537 vp9_rd_cost_reset(rd_cost);
1538 mi->sb_type = bsize;
1539 mi->ref_frame[0] = NONE;
1540 mi->ref_frame[1] = NONE;
1543 VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[cm->tx_mode]);
1545 if (sf->short_circuit_flat_blocks || sf->limit_newmv_early_exit) {
1546 #if CONFIG_VP9_HIGHBITDEPTH
1547 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
1548 x->source_variance = vp9_high_get_sby_perpixel_variance(
1549 cpi, &x->plane[0].src, bsize, xd->bd);
1551 #endif // CONFIG_VP9_HIGHBITDEPTH
1552 x->source_variance =
1553 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1556 #if CONFIG_VP9_TEMPORAL_DENOISING
1557 if (cpi->oxcf.noise_sensitivity > 0) {
1559 int layer = LAYER_IDS_TO_IDX(cpi->svc.spatial_layer_id,
1560 cpi->svc.temporal_layer_id,
1561 cpi->svc.number_temporal_layers);
1562 LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer];
1563 denoise_svc_pickmode = denoise_svc(cpi) && !lc->is_key_frame;
1565 if (cpi->denoiser.denoising_level > kDenLowLow && denoise_svc_pickmode)
1566 vp9_denoiser_reset_frame_stats(ctx);
1570 if (cpi->rc.frames_since_golden == 0 && !cpi->use_svc) {
1571 usable_ref_frame = LAST_FRAME;
1573 usable_ref_frame = GOLDEN_FRAME;
1576 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
1577 if (cpi->rc.alt_ref_gf_group || cpi->rc.is_src_frame_alt_ref)
1578 usable_ref_frame = ALTREF_FRAME;
1580 if (cpi->rc.is_src_frame_alt_ref) {
1581 skip_ref_find_pred[LAST_FRAME] = 1;
1582 skip_ref_find_pred[GOLDEN_FRAME] = 1;
1586 // For svc mode, on spatial_layer_id > 0: if the reference has different scale
1587 // constrain the inter mode to only test zero motion.
1588 if (cpi->use_svc && svc->force_zero_mode_spatial_ref &&
1589 cpi->svc.spatial_layer_id > 0) {
1590 if (cpi->ref_frame_flags & flag_list[LAST_FRAME]) {
1591 struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf;
1592 if (vp9_is_scaled(sf)) svc_force_zero_mode[LAST_FRAME - 1] = 1;
1594 if (cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) {
1595 struct scale_factors *const sf = &cm->frame_refs[GOLDEN_FRAME - 1].sf;
1596 if (vp9_is_scaled(sf)) svc_force_zero_mode[GOLDEN_FRAME - 1] = 1;
1600 if (cpi->sf.short_circuit_low_temp_var) {
1601 force_skip_low_temp_var =
1602 get_force_skip_low_temp_var(&x->variance_low[0], mi_row, mi_col, bsize);
1603 // If force_skip_low_temp_var is set, and for short circuit mode = 1 and 3,
1604 // skip golden reference.
1605 if ((cpi->sf.short_circuit_low_temp_var == 1 ||
1606 cpi->sf.short_circuit_low_temp_var == 3) &&
1607 force_skip_low_temp_var) {
1608 usable_ref_frame = LAST_FRAME;
1612 if (!((cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) &&
1613 !svc_force_zero_mode[GOLDEN_FRAME - 1] && !force_skip_low_temp_var))
1614 use_golden_nonzeromv = 0;
1616 if (cpi->oxcf.speed >= 8 && !cpi->use_svc &&
1617 ((cpi->rc.frames_since_golden + 1) < x->last_sb_high_content ||
1618 x->last_sb_high_content > 40))
1619 usable_ref_frame = LAST_FRAME;
1621 for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {
1622 if (!skip_ref_find_pred[ref_frame]) {
1623 find_predictors(cpi, x, ref_frame, frame_mv, const_motion,
1624 &ref_frame_skip_mask, flag_list, tile_data, mi_row,
1625 mi_col, yv12_mb, bsize, force_skip_low_temp_var);
1629 for (idx = 0; idx < RT_INTER_MODES; ++idx) {
1636 int this_early_term = 0;
1637 int rd_computed = 0;
1639 PREDICTION_MODE this_mode = ref_mode_set[idx].pred_mode;
1641 ref_frame = ref_mode_set[idx].ref_frame;
1644 this_mode = ref_mode_set_svc[idx].pred_mode;
1645 ref_frame = ref_mode_set_svc[idx].ref_frame;
1647 if (ref_frame > usable_ref_frame) continue;
1648 if (skip_ref_find_pred[ref_frame]) continue;
1650 // For SVC, skip the golden (spatial) reference search if sse of zeromv_last
1651 // is below threshold.
1652 if (cpi->use_svc && ref_frame == GOLDEN_FRAME &&
1653 sse_zeromv_normalized < thresh_svc_skip_golden)
1656 if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
1657 this_mode != NEARESTMV) {
1661 if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode))) continue;
1663 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
1664 if (cpi->rc.is_src_frame_alt_ref &&
1665 (ref_frame != ALTREF_FRAME ||
1666 frame_mv[this_mode][ref_frame].as_int != 0))
1669 if (cpi->rc.alt_ref_gf_group &&
1670 cpi->rc.frames_since_golden > (cpi->rc.baseline_gf_interval >> 1) &&
1671 ref_frame == GOLDEN_FRAME &&
1672 frame_mv[this_mode][ref_frame].as_int != 0)
1675 if (cpi->rc.alt_ref_gf_group &&
1676 cpi->rc.frames_since_golden < (cpi->rc.baseline_gf_interval >> 1) &&
1677 ref_frame == ALTREF_FRAME &&
1678 frame_mv[this_mode][ref_frame].as_int != 0)
1682 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) continue;
1684 if (const_motion[ref_frame] && this_mode == NEARMV) continue;
1686 // Skip non-zeromv mode search for golden frame if force_skip_low_temp_var
1687 // is set. If nearestmv for golden frame is 0, zeromv mode will be skipped
1689 if (force_skip_low_temp_var && ref_frame == GOLDEN_FRAME &&
1690 frame_mv[this_mode][ref_frame].as_int != 0) {
1694 if ((cpi->sf.short_circuit_low_temp_var >= 2 ||
1695 (cpi->sf.short_circuit_low_temp_var == 1 && bsize == BLOCK_64X64)) &&
1696 force_skip_low_temp_var && ref_frame == LAST_FRAME &&
1697 this_mode == NEWMV) {
1702 if (svc_force_zero_mode[ref_frame - 1] &&
1703 frame_mv[this_mode][ref_frame].as_int != 0)
1707 if (sf->reference_masking &&
1708 !(frame_mv[this_mode][ref_frame].as_int == 0 &&
1709 ref_frame == LAST_FRAME)) {
1710 if (usable_ref_frame < ALTREF_FRAME) {
1711 if (!force_skip_low_temp_var && usable_ref_frame > LAST_FRAME) {
1712 i = (ref_frame == LAST_FRAME) ? GOLDEN_FRAME : LAST_FRAME;
1713 if ((cpi->ref_frame_flags & flag_list[i]))
1714 if (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[i] << 1))
1715 ref_frame_skip_mask |= (1 << ref_frame);
1717 } else if (!cpi->rc.is_src_frame_alt_ref &&
1718 !(frame_mv[this_mode][ref_frame].as_int == 0 &&
1719 ref_frame == ALTREF_FRAME)) {
1720 int ref1 = (ref_frame == GOLDEN_FRAME) ? LAST_FRAME : GOLDEN_FRAME;
1721 int ref2 = (ref_frame == ALTREF_FRAME) ? LAST_FRAME : ALTREF_FRAME;
1722 if (((cpi->ref_frame_flags & flag_list[ref1]) &&
1723 (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref1] << 1))) ||
1724 ((cpi->ref_frame_flags & flag_list[ref2]) &&
1725 (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref2] << 1))))
1726 ref_frame_skip_mask |= (1 << ref_frame);
1729 if (ref_frame_skip_mask & (1 << ref_frame)) continue;
1731 // Select prediction reference frames.
1732 for (i = 0; i < MAX_MB_PLANE; i++)
1733 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
1735 mi->ref_frame[0] = ref_frame;
1736 set_ref_ptrs(cm, xd, ref_frame, NONE);
1738 mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
1739 mode_rd_thresh = best_mode_skip_txfm ? rd_threshes[mode_index] << 1
1740 : rd_threshes[mode_index];
1742 // Increase mode_rd_thresh value for GOLDEN_FRAME for improved encoding
1743 // speed with little/no subjective quality loss.
1744 if (cpi->sf.bias_golden && ref_frame == GOLDEN_FRAME &&
1745 cpi->rc.frames_since_golden > 4)
1746 mode_rd_thresh = mode_rd_thresh << 3;
1748 if ((cpi->sf.adaptive_rd_thresh_row_mt &&
1749 rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
1750 &rd_thresh_freq_fact[mode_index])) ||
1751 (!cpi->sf.adaptive_rd_thresh_row_mt &&
1752 rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
1753 &rd_thresh_freq_fact[mode_index])))
1756 if (this_mode == NEWMV) {
1757 if (ref_frame > LAST_FRAME && !cpi->use_svc &&
1758 cpi->oxcf.rc_mode == VPX_CBR) {
1761 int cost_list[5] = { INT_MAX, INT_MAX, INT_MAX, INT_MAX, INT_MAX };
1763 if (bsize < BLOCK_16X16) continue;
1765 tmp_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
1767 if (tmp_sad > x->pred_mv_sad[LAST_FRAME]) continue;
1768 if (tmp_sad + (num_pels_log2_lookup[bsize] << 4) > best_pred_sad)
1771 frame_mv[NEWMV][ref_frame].as_int = mi->mv[0].as_int;
1772 rate_mv = vp9_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv,
1773 &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1774 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1775 frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
1776 frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
1778 cpi->find_fractional_mv_step(
1779 x, &frame_mv[NEWMV][ref_frame].as_mv,
1780 &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1781 cpi->common.allow_high_precision_mv, x->errorperbit,
1782 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
1783 cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
1784 x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref_frame], NULL, 0,
1786 } else if (svc->use_base_mv && svc->spatial_layer_id) {
1787 if (frame_mv[NEWMV][ref_frame].as_int != INVALID_MV) {
1788 const int pre_stride = xd->plane[0].pre[0].stride;
1789 int base_mv_sad = INT_MAX;
1790 const float base_mv_bias = sf->base_mv_aggressive ? 1.5f : 1.0f;
1791 const uint8_t *const pre_buf =
1792 xd->plane[0].pre[0].buf +
1793 (frame_mv[NEWMV][ref_frame].as_mv.row >> 3) * pre_stride +
1794 (frame_mv[NEWMV][ref_frame].as_mv.col >> 3);
1795 base_mv_sad = cpi->fn_ptr[bsize].sdf(
1796 x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride);
1798 if (base_mv_sad < (int)(base_mv_bias * x->pred_mv_sad[ref_frame])) {
1799 // Base layer mv is good.
1800 if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1801 &frame_mv[NEWMV][ref_frame], &rate_mv,
1802 best_rdc.rdcost, 1)) {
1805 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1806 &frame_mv[NEWMV][ref_frame],
1807 &rate_mv, best_rdc.rdcost, 0)) {
1810 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1811 &frame_mv[NEWMV][ref_frame],
1812 &rate_mv, best_rdc.rdcost, 0)) {
1815 } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1816 &frame_mv[NEWMV][ref_frame], &rate_mv,
1817 best_rdc.rdcost, 0)) {
1822 // If use_golden_nonzeromv is false, NEWMV mode is skipped for golden, no
1823 // need to compute best_pred_sad which is only used to skip golden NEWMV.
1824 if (use_golden_nonzeromv && this_mode == NEWMV && ref_frame == LAST_FRAME &&
1825 frame_mv[NEWMV][LAST_FRAME].as_int != INVALID_MV) {
1826 const int pre_stride = xd->plane[0].pre[0].stride;
1827 const uint8_t *const pre_buf =
1828 xd->plane[0].pre[0].buf +
1829 (frame_mv[NEWMV][LAST_FRAME].as_mv.row >> 3) * pre_stride +
1830 (frame_mv[NEWMV][LAST_FRAME].as_mv.col >> 3);
1831 best_pred_sad = cpi->fn_ptr[bsize].sdf(
1832 x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride);
1833 x->pred_mv_sad[LAST_FRAME] = best_pred_sad;
1836 if (this_mode != NEARESTMV &&
1837 frame_mv[this_mode][ref_frame].as_int ==
1838 frame_mv[NEARESTMV][ref_frame].as_int)
1841 mi->mode = this_mode;
1842 mi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
1844 // Search for the best prediction filter type, when the resulting
1845 // motion vector is at sub-pixel accuracy level for luma component, i.e.,
1846 // the last three bits are all zeros.
1847 if (reuse_inter_pred) {
1848 if (!this_mode_pred) {
1849 this_mode_pred = &tmp[3];
1851 this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
1852 pd->dst.buf = this_mode_pred->data;
1853 pd->dst.stride = bw;
1857 if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&
1858 pred_filter_search &&
1859 (ref_frame == LAST_FRAME ||
1860 (ref_frame == GOLDEN_FRAME &&
1861 (cpi->use_svc || cpi->oxcf.rc_mode == VPX_VBR))) &&
1862 (((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07) != 0)) {
1866 unsigned int pf_var[3];
1867 unsigned int pf_sse[3];
1868 TX_SIZE pf_tx_size[3];
1869 int64_t best_cost = INT64_MAX;
1870 INTERP_FILTER best_filter = SWITCHABLE, filter;
1871 PRED_BUFFER *current_pred = this_mode_pred;
1874 for (filter = EIGHTTAP; filter <= EIGHTTAP_SMOOTH; ++filter) {
1876 mi->interp_filter = filter;
1877 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1878 model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter], &pf_dist[filter],
1879 &pf_var[filter], &pf_sse[filter]);
1880 curr_rate[filter] = pf_rate[filter];
1881 pf_rate[filter] += vp9_get_switchable_rate(cpi, xd);
1882 cost = RDCOST(x->rdmult, x->rddiv, pf_rate[filter], pf_dist[filter]);
1883 pf_tx_size[filter] = mi->tx_size;
1884 if (cost < best_cost) {
1885 best_filter = filter;
1887 skip_txfm = x->skip_txfm[0];
1889 if (reuse_inter_pred) {
1890 if (this_mode_pred != current_pred) {
1891 free_pred_buffer(this_mode_pred);
1892 this_mode_pred = current_pred;
1894 current_pred = &tmp[get_pred_buffer(tmp, 3)];
1895 pd->dst.buf = current_pred->data;
1896 pd->dst.stride = bw;
1901 if (reuse_inter_pred && this_mode_pred != current_pred)
1902 free_pred_buffer(current_pred);
1904 mi->interp_filter = best_filter;
1905 mi->tx_size = pf_tx_size[best_filter];
1906 this_rdc.rate = curr_rate[best_filter];
1907 this_rdc.dist = pf_dist[best_filter];
1908 var_y = pf_var[best_filter];
1909 sse_y = pf_sse[best_filter];
1910 x->skip_txfm[0] = skip_txfm;
1911 if (reuse_inter_pred) {
1912 pd->dst.buf = this_mode_pred->data;
1913 pd->dst.stride = this_mode_pred->stride;
1916 const int large_block = (x->sb_is_skin || cpi->oxcf.speed < 7)
1917 ? bsize > BLOCK_32X32
1918 : bsize >= BLOCK_32X32;
1919 mi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP : filter_ref;
1921 if (cpi->use_svc && ref_frame == GOLDEN_FRAME &&
1922 svc_force_zero_mode[ref_frame - 1])
1923 mi->interp_filter = filter_gf_svc;
1925 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1927 // For large partition blocks, extra testing is done.
1928 if (cpi->oxcf.rc_mode == VPX_CBR && large_block &&
1929 !cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) &&
1931 model_rd_for_sb_y_large(cpi, bsize, x, xd, &this_rdc.rate,
1932 &this_rdc.dist, &var_y, &sse_y, mi_row, mi_col,
1936 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
1939 // Save normalized sse (between current and last frame) for (0, 0) motion.
1940 if (cpi->use_svc && ref_frame == LAST_FRAME &&
1941 frame_mv[this_mode][ref_frame].as_int == 0) {
1942 sse_zeromv_normalized =
1943 sse_y >> (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
1947 if (!this_early_term) {
1948 this_sse = (int64_t)sse_y;
1949 block_yrd(cpi, x, &this_rdc, &is_skippable, &this_sse, bsize,
1950 VPXMIN(mi->tx_size, TX_16X16), rd_computed);
1952 x->skip_txfm[0] = is_skippable;
1954 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1956 if (RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist) <
1957 RDCOST(x->rdmult, x->rddiv, 0, this_sse)) {
1958 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
1960 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1961 this_rdc.dist = this_sse;
1962 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1966 if (cm->interp_filter == SWITCHABLE) {
1967 if ((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07)
1968 this_rdc.rate += vp9_get_switchable_rate(cpi, xd);
1971 this_rdc.rate += cm->interp_filter == SWITCHABLE
1972 ? vp9_get_switchable_rate(cpi, xd)
1974 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1977 if (x->color_sensitivity[0] || x->color_sensitivity[1]) {
1979 const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, &xd->plane[1]);
1980 if (x->color_sensitivity[0])
1981 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 1);
1982 if (x->color_sensitivity[1])
1983 vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 2);
1984 model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &rdc_uv, &var_y, &sse_y, 1, 2);
1985 this_rdc.rate += rdc_uv.rate;
1986 this_rdc.dist += rdc_uv.dist;
1989 this_rdc.rate += rate_mv;
1990 this_rdc.rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
1991 [INTER_OFFSET(this_mode)];
1992 this_rdc.rate += ref_frame_cost[ref_frame];
1993 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1995 // Bias against NEWMV that is very different from its neighbors, and bias
1996 // to small motion-lastref for noisy input.
1997 if (cpi->oxcf.rc_mode == VPX_CBR && cpi->oxcf.speed >= 5 &&
1998 cpi->oxcf.content != VP9E_CONTENT_SCREEN) {
1999 vp9_NEWMV_diff_bias(&cpi->noise_estimate, xd, this_mode, &this_rdc, bsize,
2000 frame_mv[this_mode][ref_frame].as_mv.row,
2001 frame_mv[this_mode][ref_frame].as_mv.col,
2002 ref_frame == LAST_FRAME, x->lowvar_highsumdiff,
2006 // Skipping checking: test to see if this block can be reconstructed by
2008 if (cpi->allow_encode_breakout) {
2009 encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame, this_mode,
2010 var_y, sse_y, yv12_mb, &this_rdc.rate,
2013 this_rdc.rate += rate_mv;
2015 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2019 #if CONFIG_VP9_TEMPORAL_DENOISING
2020 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc_pickmode &&
2021 cpi->denoiser.denoising_level > kDenLowLow) {
2022 vp9_denoiser_update_frame_stats(mi, sse_y, this_mode, ctx);
2023 // Keep track of zero_last cost.
2024 if (ref_frame == LAST_FRAME && frame_mv[this_mode][ref_frame].as_int == 0)
2025 zero_last_cost_orig = this_rdc.rdcost;
2031 if (this_rdc.rdcost < best_rdc.rdcost || x->skip) {
2032 best_rdc = this_rdc;
2033 best_mode = this_mode;
2034 best_pred_filter = mi->interp_filter;
2035 best_tx_size = mi->tx_size;
2036 best_ref_frame = ref_frame;
2037 best_mode_skip_txfm = x->skip_txfm[0];
2038 best_early_term = this_early_term;
2040 if (reuse_inter_pred) {
2041 free_pred_buffer(best_pred);
2042 best_pred = this_mode_pred;
2045 if (reuse_inter_pred) free_pred_buffer(this_mode_pred);
2050 // If early termination flag is 1 and at least 2 modes are checked,
2051 // the mode search is terminated.
2052 if (best_early_term && idx > 0) {
2058 mi->mode = best_mode;
2059 mi->interp_filter = best_pred_filter;
2060 mi->tx_size = best_tx_size;
2061 mi->ref_frame[0] = best_ref_frame;
2062 mi->mv[0].as_int = frame_mv[best_mode][best_ref_frame].as_int;
2063 xd->mi[0]->bmi[0].as_mv[0].as_int = mi->mv[0].as_int;
2064 x->skip_txfm[0] = best_mode_skip_txfm;
2066 // For spatial enhancemanent layer: perform intra prediction only if base
2067 // layer is chosen as the reference. Always perform intra prediction if
2068 // LAST is the only reference or is_key_frame is set.
2069 if (cpi->svc.spatial_layer_id) {
2070 perform_intra_pred =
2071 cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame ||
2072 !(cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) ||
2073 (!cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
2074 svc_force_zero_mode[best_ref_frame - 1]);
2075 inter_mode_thresh = (inter_mode_thresh << 1) + inter_mode_thresh;
2077 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR &&
2078 cpi->rc.is_src_frame_alt_ref)
2079 perform_intra_pred = 0;
2080 // Perform intra prediction search, if the best SAD is above a certain
2082 if (best_rdc.rdcost == INT64_MAX ||
2083 ((!force_skip_low_temp_var || bsize < BLOCK_32X32) &&
2084 perform_intra_pred && !x->skip && best_rdc.rdcost > inter_mode_thresh &&
2085 bsize <= cpi->sf.max_intra_bsize && !x->skip_low_source_sad &&
2086 !x->lowvar_highsumdiff)) {
2087 struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
2089 TX_SIZE best_intra_tx_size = TX_SIZES;
2090 TX_SIZE intra_tx_size =
2091 VPXMIN(max_txsize_lookup[bsize],
2092 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
2093 if (cpi->oxcf.content != VP9E_CONTENT_SCREEN && intra_tx_size > TX_16X16)
2094 intra_tx_size = TX_16X16;
2096 if (reuse_inter_pred && best_pred != NULL) {
2097 if (best_pred->data == orig_dst.buf) {
2098 this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
2099 #if CONFIG_VP9_HIGHBITDEPTH
2100 if (cm->use_highbitdepth)
2101 vpx_highbd_convolve_copy(
2102 CONVERT_TO_SHORTPTR(best_pred->data), best_pred->stride,
2103 CONVERT_TO_SHORTPTR(this_mode_pred->data), this_mode_pred->stride,
2104 NULL, 0, NULL, 0, bw, bh, xd->bd);
2106 vpx_convolve_copy(best_pred->data, best_pred->stride,
2107 this_mode_pred->data, this_mode_pred->stride, NULL,
2108 0, NULL, 0, bw, bh);
2110 vpx_convolve_copy(best_pred->data, best_pred->stride,
2111 this_mode_pred->data, this_mode_pred->stride, NULL, 0,
2113 #endif // CONFIG_VP9_HIGHBITDEPTH
2114 best_pred = this_mode_pred;
2119 for (i = 0; i < 4; ++i) {
2120 const PREDICTION_MODE this_mode = intra_mode_list[i];
2121 THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)];
2122 int mode_rd_thresh = rd_threshes[mode_index];
2123 if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
2124 this_mode != DC_PRED) {
2128 if (!((1 << this_mode) & cpi->sf.intra_y_mode_bsize_mask[bsize]))
2131 if ((cpi->sf.adaptive_rd_thresh_row_mt &&
2132 rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
2133 &rd_thresh_freq_fact[mode_index])) ||
2134 (!cpi->sf.adaptive_rd_thresh_row_mt &&
2135 rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
2136 &rd_thresh_freq_fact[mode_index])))
2139 mi->mode = this_mode;
2140 mi->ref_frame[0] = INTRA_FRAME;
2141 this_rdc.dist = this_rdc.rate = 0;
2142 args.mode = this_mode;
2144 args.rdc = &this_rdc;
2145 mi->tx_size = intra_tx_size;
2146 vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
2148 // Check skip cost here since skippable is not set for for uv, this
2149 // mirrors the behavior used by inter
2150 if (args.skippable) {
2151 x->skip_txfm[0] = SKIP_TXFM_AC_DC;
2152 this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
2154 x->skip_txfm[0] = SKIP_TXFM_NONE;
2155 this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
2157 // Inter and intra RD will mismatch in scale for non-screen content.
2158 if (cpi->oxcf.content == VP9E_CONTENT_SCREEN) {
2159 if (x->color_sensitivity[0])
2160 vp9_foreach_transformed_block_in_plane(xd, bsize, 1,
2161 estimate_block_intra, &args);
2162 if (x->color_sensitivity[1])
2163 vp9_foreach_transformed_block_in_plane(xd, bsize, 2,
2164 estimate_block_intra, &args);
2166 this_rdc.rate += cpi->mbmode_cost[this_mode];
2167 this_rdc.rate += ref_frame_cost[INTRA_FRAME];
2168 this_rdc.rate += intra_cost_penalty;
2170 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2172 if (this_rdc.rdcost < best_rdc.rdcost) {
2173 best_rdc = this_rdc;
2174 best_mode = this_mode;
2175 best_intra_tx_size = mi->tx_size;
2176 best_ref_frame = INTRA_FRAME;
2177 mi->uv_mode = this_mode;
2178 mi->mv[0].as_int = INVALID_MV;
2179 best_mode_skip_txfm = x->skip_txfm[0];
2183 // Reset mb_mode_info to the best inter mode.
2184 if (best_ref_frame != INTRA_FRAME) {
2185 mi->tx_size = best_tx_size;
2187 mi->tx_size = best_intra_tx_size;
2192 mi->mode = best_mode;
2193 mi->ref_frame[0] = best_ref_frame;
2194 x->skip_txfm[0] = best_mode_skip_txfm;
2196 if (!is_inter_block(mi)) {
2197 mi->interp_filter = SWITCHABLE_FILTERS;
2200 if (reuse_inter_pred && best_pred != NULL) {
2201 if (best_pred->data != orig_dst.buf && is_inter_mode(mi->mode)) {
2202 #if CONFIG_VP9_HIGHBITDEPTH
2203 if (cm->use_highbitdepth)
2204 vpx_highbd_convolve_copy(
2205 CONVERT_TO_SHORTPTR(best_pred->data), best_pred->stride,
2206 CONVERT_TO_SHORTPTR(pd->dst.buf), pd->dst.stride, NULL, 0, NULL, 0,
2209 vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2210 pd->dst.stride, NULL, 0, NULL, 0, bw, bh);
2212 vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2213 pd->dst.stride, NULL, 0, NULL, 0, bw, bh);
2214 #endif // CONFIG_VP9_HIGHBITDEPTH
2218 #if CONFIG_VP9_TEMPORAL_DENOISING
2219 if (cpi->oxcf.noise_sensitivity > 0 && cpi->resize_pending == 0 &&
2220 denoise_svc_pickmode && cpi->denoiser.denoising_level > kDenLowLow &&
2221 cpi->denoiser.reset == 0) {
2222 VP9_DENOISER_DECISION decision = COPY_BLOCK;
2223 vp9_pickmode_ctx_den_update(&ctx_den, zero_last_cost_orig, ref_frame_cost,
2224 frame_mv, reuse_inter_pred, best_tx_size,
2225 best_mode, best_ref_frame, best_pred_filter,
2226 best_mode_skip_txfm);
2227 vp9_denoiser_denoise(cpi, x, mi_row, mi_col, bsize, ctx, &decision);
2228 recheck_zeromv_after_denoising(cpi, mi, x, xd, decision, &ctx_den, yv12_mb,
2229 &best_rdc, bsize, mi_row, mi_col);
2230 best_ref_frame = ctx_den.best_ref_frame;
2234 if (cpi->sf.adaptive_rd_thresh) {
2235 THR_MODES best_mode_idx = mode_idx[best_ref_frame][mode_offset(mi->mode)];
2237 if (best_ref_frame == INTRA_FRAME) {
2238 // Only consider the modes that are included in the intra_mode_list.
2239 int intra_modes = sizeof(intra_mode_list) / sizeof(PREDICTION_MODE);
2242 // TODO(yunqingwang): Check intra mode mask and only update freq_fact
2243 // for those valid modes.
2244 for (i = 0; i < intra_modes; i++) {
2245 if (cpi->sf.adaptive_rd_thresh_row_mt)
2246 update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance,
2247 thresh_freq_fact_idx, INTRA_FRAME,
2248 best_mode_idx, intra_mode_list[i]);
2250 update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2251 INTRA_FRAME, best_mode_idx,
2252 intra_mode_list[i]);
2255 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2256 PREDICTION_MODE this_mode;
2257 if (best_ref_frame != ref_frame) continue;
2258 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2259 if (cpi->sf.adaptive_rd_thresh_row_mt)
2260 update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance,
2261 thresh_freq_fact_idx, ref_frame,
2262 best_mode_idx, this_mode);
2264 update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2265 ref_frame, best_mode_idx, this_mode);
2271 *rd_cost = best_rdc;
2274 void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x, int mi_row,
2275 int mi_col, RD_COST *rd_cost, BLOCK_SIZE bsize,
2276 PICK_MODE_CONTEXT *ctx) {
2277 VP9_COMMON *const cm = &cpi->common;
2278 SPEED_FEATURES *const sf = &cpi->sf;
2279 MACROBLOCKD *const xd = &x->e_mbd;
2280 MODE_INFO *const mi = xd->mi[0];
2281 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2282 const struct segmentation *const seg = &cm->seg;
2283 MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE;
2284 MV_REFERENCE_FRAME best_ref_frame = NONE;
2285 unsigned char segment_id = mi->segment_id;
2286 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2287 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2289 int64_t best_rd = INT64_MAX;
2290 b_mode_info bsi[MAX_REF_FRAMES][4];
2291 int ref_frame_skip_mask = 0;
2292 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
2293 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
2296 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
2297 ctx->pred_pixel_ready = 0;
2299 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2300 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2302 x->pred_mv_sad[ref_frame] = INT_MAX;
2304 if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
2305 int_mv *const candidates = mbmi_ext->ref_mvs[ref_frame];
2306 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2307 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf,
2309 vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
2310 mbmi_ext->mode_context);
2312 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2313 &dummy_mv[0], &dummy_mv[1]);
2315 ref_frame_skip_mask |= (1 << ref_frame);
2319 mi->sb_type = bsize;
2320 mi->tx_size = TX_4X4;
2321 mi->uv_mode = DC_PRED;
2322 mi->ref_frame[0] = LAST_FRAME;
2323 mi->ref_frame[1] = NONE;
2325 cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
2327 for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2328 int64_t this_rd = 0;
2331 if (ref_frame_skip_mask & (1 << ref_frame)) continue;
2333 #if CONFIG_BETTER_HW_COMPATIBILITY
2334 if ((bsize == BLOCK_8X4 || bsize == BLOCK_4X8) && ref_frame > INTRA_FRAME &&
2335 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2339 // TODO(jingning, agrange): Scaling reference frame not supported for
2340 // sub8x8 blocks. Is this supported now?
2341 if (ref_frame > INTRA_FRAME &&
2342 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2345 // If the segment reference frame feature is enabled....
2346 // then do nothing if the current ref frame is not allowed..
2347 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
2348 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
2351 mi->ref_frame[0] = ref_frame;
2353 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
2355 // Select prediction reference frames.
2356 for (plane = 0; plane < MAX_MB_PLANE; plane++)
2357 xd->plane[plane].pre[0] = yv12_mb[ref_frame][plane];
2359 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2360 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2361 int_mv b_mv[MB_MODE_COUNT];
2362 int64_t b_best_rd = INT64_MAX;
2363 const int i = idy * 2 + idx;
2364 PREDICTION_MODE this_mode;
2366 unsigned int var_y, sse_y;
2368 struct macroblock_plane *p = &x->plane[0];
2369 struct macroblockd_plane *pd = &xd->plane[0];
2371 const struct buf_2d orig_src = p->src;
2372 const struct buf_2d orig_dst = pd->dst;
2373 struct buf_2d orig_pre[2];
2374 memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));
2376 // set buffer pointers for sub8x8 motion search.
2378 &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
2380 &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
2383 .buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
2385 b_mv[ZEROMV].as_int = 0;
2386 b_mv[NEWMV].as_int = INVALID_MV;
2387 vp9_append_sub8x8_mvs_for_idx(cm, xd, i, 0, mi_row, mi_col,
2388 &b_mv[NEARESTMV], &b_mv[NEARMV],
2389 mbmi_ext->mode_context);
2391 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2393 xd->mi[0]->bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
2395 if (this_mode == NEWMV) {
2396 const int step_param = cpi->sf.mv.fullpel_search_step_param;
2400 const MvLimits tmp_mv_limits = x->mv_limits;
2401 uint32_t dummy_dist;
2404 mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3;
2405 mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3;
2407 mvp_full.row = xd->mi[0]->bmi[0].as_mv[0].as_mv.row >> 3;
2408 mvp_full.col = xd->mi[0]->bmi[0].as_mv[0].as_mv.col >> 3;
2411 vp9_set_mv_search_range(&x->mv_limits,
2412 &mbmi_ext->ref_mvs[ref_frame][0].as_mv);
2414 vp9_full_pixel_search(
2415 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method,
2416 x->sadperbit4, cond_cost_list(cpi, cost_list),
2417 &mbmi_ext->ref_mvs[ref_frame][0].as_mv, &tmp_mv, INT_MAX, 0);
2419 x->mv_limits = tmp_mv_limits;
2421 // calculate the bit cost on motion vector
2422 mvp_full.row = tmp_mv.row * 8;
2423 mvp_full.col = tmp_mv.col * 8;
2425 b_rate += vp9_mv_bit_cost(
2426 &mvp_full, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2427 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2429 b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2430 [INTER_OFFSET(NEWMV)];
2431 if (RDCOST(x->rdmult, x->rddiv, b_rate, 0) > b_best_rd) continue;
2433 cpi->find_fractional_mv_step(
2434 x, &tmp_mv, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2435 cpi->common.allow_high_precision_mv, x->errorperbit,
2436 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
2437 cpi->sf.mv.subpel_iters_per_step,
2438 cond_cost_list(cpi, cost_list), x->nmvjointcost, x->mvcost,
2439 &dummy_dist, &x->pred_sse[ref_frame], NULL, 0, 0);
2441 xd->mi[0]->bmi[i].as_mv[0].as_mv = tmp_mv;
2443 b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2444 [INTER_OFFSET(this_mode)];
2447 #if CONFIG_VP9_HIGHBITDEPTH
2448 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2449 vp9_highbd_build_inter_predictor(
2450 CONVERT_TO_SHORTPTR(pd->pre[0].buf), pd->pre[0].stride,
2451 CONVERT_TO_SHORTPTR(pd->dst.buf), pd->dst.stride,
2452 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2453 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2454 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2455 mi_col * MI_SIZE + 4 * (i & 0x01),
2456 mi_row * MI_SIZE + 4 * (i >> 1), xd->bd);
2459 vp9_build_inter_predictor(
2460 pd->pre[0].buf, pd->pre[0].stride, pd->dst.buf, pd->dst.stride,
2461 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2462 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2463 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2464 mi_col * MI_SIZE + 4 * (i & 0x01),
2465 mi_row * MI_SIZE + 4 * (i >> 1));
2467 #if CONFIG_VP9_HIGHBITDEPTH
2471 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
2474 this_rdc.rate += b_rate;
2476 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2477 if (this_rdc.rdcost < b_best_rd) {
2478 b_best_rd = this_rdc.rdcost;
2479 bsi[ref_frame][i].as_mode = this_mode;
2480 bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0]->bmi[i].as_mv[0].as_mv;
2484 // restore source and prediction buffer pointers.
2486 pd->pre[0] = orig_pre[0];
2488 this_rd += b_best_rd;
2490 xd->mi[0]->bmi[i] = bsi[ref_frame][i];
2491 if (num_4x4_blocks_wide > 1) xd->mi[0]->bmi[i + 1] = xd->mi[0]->bmi[i];
2492 if (num_4x4_blocks_high > 1) xd->mi[0]->bmi[i + 2] = xd->mi[0]->bmi[i];
2494 } // loop through sub8x8 blocks
2496 if (this_rd < best_rd) {
2498 best_ref_frame = ref_frame;
2500 } // reference frames
2502 mi->tx_size = TX_4X4;
2503 mi->ref_frame[0] = best_ref_frame;
2504 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2505 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2506 const int block = idy * 2 + idx;
2507 xd->mi[0]->bmi[block] = bsi[best_ref_frame][block];
2508 if (num_4x4_blocks_wide > 1)
2509 xd->mi[0]->bmi[block + 1] = bsi[best_ref_frame][block];
2510 if (num_4x4_blocks_high > 1)
2511 xd->mi[0]->bmi[block + 2] = bsi[best_ref_frame][block];
2514 mi->mode = xd->mi[0]->bmi[3].as_mode;
2515 ctx->mic = *(xd->mi[0]);
2516 ctx->mbmi_ext = *x->mbmi_ext;
2517 ctx->skip_txfm[0] = SKIP_TXFM_NONE;
2519 // Dummy assignment for speed -5. No effect in speed -6.
2520 rd_cost->rdcost = best_rd;