1 /*
2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10
11 */
12
13 #include <assert.h>
14 #include <limits.h>
15 #include <math.h>
16 #include <stdio.h>
17
18 #include "av1/common/reconinter.h"
19 #include "av1/common/reconintra.h"
20
21 #include "av1/encoder/encodemv.h"
22 #include "av1/encoder/intra_mode_search.h"
23 #include "av1/encoder/model_rd.h"
24 #include "av1/encoder/motion_search_facade.h"
25 #include "av1/encoder/nonrd_opt.h"
26 #include "av1/encoder/reconinter_enc.h"
27 #include "av1/encoder/var_based_part.h"
28
early_term_inter_search_with_sse(int early_term_idx,BLOCK_SIZE bsize,int64_t this_sse,int64_t best_sse,PREDICTION_MODE this_mode)29 static INLINE int early_term_inter_search_with_sse(int early_term_idx,
30 BLOCK_SIZE bsize,
31 int64_t this_sse,
32 int64_t best_sse,
33 PREDICTION_MODE this_mode) {
34 // Aggressiveness to terminate inter mode search early is adjusted based on
35 // speed and block size.
36 static const double early_term_thresh[4][4] = { { 0.65, 0.65, 0.65, 0.7 },
37 { 0.6, 0.65, 0.85, 0.9 },
38 { 0.5, 0.5, 0.55, 0.6 },
39 { 0.6, 0.75, 0.85, 0.85 } };
40 static const double early_term_thresh_newmv_nearestmv[4] = { 0.3, 0.3, 0.3,
41 0.3 };
42
43 const int size_group = size_group_lookup[bsize];
44 assert(size_group < 4);
45 assert((early_term_idx > 0) && (early_term_idx < EARLY_TERM_INDICES));
46 const double threshold =
47 ((early_term_idx == EARLY_TERM_IDX_4) &&
48 (this_mode == NEWMV || this_mode == NEARESTMV))
49 ? early_term_thresh_newmv_nearestmv[size_group]
50 : early_term_thresh[early_term_idx - 1][size_group];
51
52 // Terminate inter mode search early based on best sse so far.
53 if ((early_term_idx > 0) && (threshold * this_sse > best_sse)) {
54 return 1;
55 }
56 return 0;
57 }
58
init_best_pickmode(BEST_PICKMODE * bp)59 static INLINE void init_best_pickmode(BEST_PICKMODE *bp) {
60 bp->best_sse = INT64_MAX;
61 bp->best_mode = NEARESTMV;
62 bp->best_ref_frame = LAST_FRAME;
63 bp->best_second_ref_frame = NONE_FRAME;
64 bp->best_tx_size = TX_8X8;
65 bp->tx_type = DCT_DCT;
66 bp->best_pred_filter = av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
67 bp->best_mode_skip_txfm = 0;
68 bp->best_mode_initial_skip_flag = 0;
69 bp->best_pred = NULL;
70 bp->best_motion_mode = SIMPLE_TRANSLATION;
71 bp->num_proj_ref = 0;
72 av1_zero(bp->wm_params);
73 av1_zero(bp->pmi);
74 }
75
76 // Copy best inter mode parameters to best_pickmode
update_search_state_nonrd(InterModeSearchStateNonrd * search_state,MB_MODE_INFO * const mi,TxfmSearchInfo * txfm_info,RD_STATS * nonskip_rdc,PICK_MODE_CONTEXT * ctx,PREDICTION_MODE this_best_mode,const int64_t sse_y)77 static INLINE void update_search_state_nonrd(
78 InterModeSearchStateNonrd *search_state, MB_MODE_INFO *const mi,
79 TxfmSearchInfo *txfm_info, RD_STATS *nonskip_rdc, PICK_MODE_CONTEXT *ctx,
80 PREDICTION_MODE this_best_mode, const int64_t sse_y) {
81 BEST_PICKMODE *const best_pickmode = &search_state->best_pickmode;
82
83 best_pickmode->best_sse = sse_y;
84 best_pickmode->best_mode = this_best_mode;
85 best_pickmode->best_motion_mode = mi->motion_mode;
86 best_pickmode->wm_params = mi->wm_params;
87 best_pickmode->num_proj_ref = mi->num_proj_ref;
88 best_pickmode->best_pred_filter = mi->interp_filters;
89 best_pickmode->best_tx_size = mi->tx_size;
90 best_pickmode->best_ref_frame = mi->ref_frame[0];
91 best_pickmode->best_second_ref_frame = mi->ref_frame[1];
92 best_pickmode->best_mode_skip_txfm = search_state->this_rdc.skip_txfm;
93 best_pickmode->best_mode_initial_skip_flag =
94 (nonskip_rdc->rate == INT_MAX && search_state->this_rdc.skip_txfm);
95 if (!best_pickmode->best_mode_skip_txfm) {
96 memcpy(ctx->blk_skip, txfm_info->blk_skip,
97 sizeof(txfm_info->blk_skip[0]) * ctx->num_4x4_blk);
98 }
99 }
100
subpel_select(AV1_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int_mv * mv,MV ref_mv,FULLPEL_MV start_mv,bool fullpel_performed_well)101 static INLINE int subpel_select(AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
102 int_mv *mv, MV ref_mv, FULLPEL_MV start_mv,
103 bool fullpel_performed_well) {
104 const int frame_lowmotion = cpi->rc.avg_frame_low_motion;
105 const int reduce_mv_pel_precision_highmotion =
106 cpi->sf.rt_sf.reduce_mv_pel_precision_highmotion;
107
108 // Reduce MV precision for higher int MV value & frame-level motion
109 if (reduce_mv_pel_precision_highmotion >= 3) {
110 int mv_thresh = 4;
111 const int is_low_resoln =
112 (cpi->common.width * cpi->common.height <= 320 * 240);
113 mv_thresh = (bsize > BLOCK_32X32) ? 2 : (bsize > BLOCK_16X16) ? 4 : 6;
114 if (frame_lowmotion > 0 && frame_lowmotion < 40) mv_thresh = 12;
115 mv_thresh = (is_low_resoln) ? mv_thresh >> 1 : mv_thresh;
116 if (abs(mv->as_fullmv.row) >= mv_thresh ||
117 abs(mv->as_fullmv.col) >= mv_thresh)
118 return HALF_PEL;
119 } else if (reduce_mv_pel_precision_highmotion >= 1) {
120 int mv_thresh;
121 const int th_vals[2][3] = { { 4, 8, 10 }, { 4, 6, 8 } };
122 const int th_idx = reduce_mv_pel_precision_highmotion - 1;
123 assert(th_idx >= 0 && th_idx < 2);
124 if (frame_lowmotion > 0 && frame_lowmotion < 40)
125 mv_thresh = 12;
126 else
127 mv_thresh = (bsize >= BLOCK_32X32) ? th_vals[th_idx][0]
128 : (bsize >= BLOCK_16X16) ? th_vals[th_idx][1]
129 : th_vals[th_idx][2];
130 if (abs(mv->as_fullmv.row) >= (mv_thresh << 1) ||
131 abs(mv->as_fullmv.col) >= (mv_thresh << 1))
132 return FULL_PEL;
133 else if (abs(mv->as_fullmv.row) >= mv_thresh ||
134 abs(mv->as_fullmv.col) >= mv_thresh)
135 return HALF_PEL;
136 }
137 // Reduce MV precision for relatively static (e.g. background), low-complex
138 // large areas
139 if (cpi->sf.rt_sf.reduce_mv_pel_precision_lowcomplex >= 2) {
140 const int qband = x->qindex >> (QINDEX_BITS - 2);
141 assert(qband < 4);
142 if (x->content_state_sb.source_sad_nonrd <= kVeryLowSad &&
143 bsize > BLOCK_16X16 && qband != 0) {
144 if (x->source_variance < 500)
145 return FULL_PEL;
146 else if (x->source_variance < 5000)
147 return HALF_PEL;
148 }
149 } else if (cpi->sf.rt_sf.reduce_mv_pel_precision_lowcomplex >= 1) {
150 if (fullpel_performed_well && ref_mv.row == 0 && ref_mv.col == 0 &&
151 start_mv.row == 0 && start_mv.col == 0)
152 return HALF_PEL;
153 }
154 return cpi->sf.mv_sf.subpel_force_stop;
155 }
156
use_aggressive_subpel_search_method(MACROBLOCK * x,bool use_adaptive_subpel_search,bool fullpel_performed_well)157 static bool use_aggressive_subpel_search_method(MACROBLOCK *x,
158 bool use_adaptive_subpel_search,
159 bool fullpel_performed_well) {
160 if (!use_adaptive_subpel_search) return false;
161 const int qband = x->qindex >> (QINDEX_BITS - 2);
162 assert(qband < 4);
163 if ((qband > 0) && (fullpel_performed_well ||
164 (x->content_state_sb.source_sad_nonrd <= kLowSad) ||
165 (x->source_variance < 100)))
166 return true;
167 return false;
168 }
169
170 /*!\brief Runs Motion Estimation for a specific block and specific ref frame.
171 *
172 * \ingroup nonrd_mode_search
173 * \callgraph
174 * \callergraph
175 * Finds the best Motion Vector by running Motion Estimation for a specific
176 * block and a specific reference frame. Exits early if RDCost of Full Pel part
177 * exceeds best RD Cost fund so far
178 * \param[in] cpi Top-level encoder structure
179 * \param[in] x Pointer to structure holding all the
180 * data for the current macroblock
181 * \param[in] bsize Current block size
182 * \param[in] tmp_mv Pointer to best found New MV
183 * \param[in] rate_mv Pointer to Rate of the best new MV
184 * \param[in] best_rd_sofar RD Cost of the best mode found so far
185 * \param[in] use_base_mv Flag, indicating that tmp_mv holds
186 * specific MV to start the search with
187 *
188 * \return Returns 0 if ME was terminated after Full Pel Search because too
189 * high RD Cost. Otherwise returns 1. Best New MV is placed into \c tmp_mv.
190 * Rate estimation for this vector is placed to \c rate_mv
191 */
combined_motion_search(AV1_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int_mv * tmp_mv,int * rate_mv,int64_t best_rd_sofar,int use_base_mv)192 static int combined_motion_search(AV1_COMP *cpi, MACROBLOCK *x,
193 BLOCK_SIZE bsize, int_mv *tmp_mv,
194 int *rate_mv, int64_t best_rd_sofar,
195 int use_base_mv) {
196 MACROBLOCKD *xd = &x->e_mbd;
197 const AV1_COMMON *cm = &cpi->common;
198 const SPEED_FEATURES *sf = &cpi->sf;
199 MB_MODE_INFO *mi = xd->mi[0];
200 int step_param = (sf->rt_sf.fullpel_search_step_param)
201 ? sf->rt_sf.fullpel_search_step_param
202 : cpi->mv_search_params.mv_step_param;
203 FULLPEL_MV start_mv;
204 const int ref = mi->ref_frame[0];
205 const MV ref_mv = av1_get_ref_mv(x, mi->ref_mv_idx).as_mv;
206 MV center_mv;
207 int dis;
208 int rv = 0;
209 int cost_list[5];
210 int search_subpel = 1;
211
212 start_mv = get_fullmv_from_mv(&ref_mv);
213
214 if (!use_base_mv)
215 center_mv = ref_mv;
216 else
217 center_mv = tmp_mv->as_mv;
218
219 const SEARCH_METHODS search_method =
220 av1_get_default_mv_search_method(x, &cpi->sf.mv_sf, bsize);
221 const search_site_config *src_search_sites =
222 av1_get_search_site_config(cpi, x, search_method);
223 FULLPEL_MOTION_SEARCH_PARAMS full_ms_params;
224 FULLPEL_MV_STATS best_mv_stats;
225 av1_make_default_fullpel_ms_params(&full_ms_params, cpi, x, bsize, ¢er_mv,
226 start_mv, src_search_sites, search_method,
227 /*fine_search_interval=*/0);
228
229 const unsigned int full_var_rd = av1_full_pixel_search(
230 start_mv, &full_ms_params, step_param, cond_cost_list(cpi, cost_list),
231 &tmp_mv->as_fullmv, &best_mv_stats, NULL);
232
233 // calculate the bit cost on motion vector
234 MV mvp_full = get_mv_from_fullmv(&tmp_mv->as_fullmv);
235
236 *rate_mv = av1_mv_bit_cost(&mvp_full, &ref_mv, x->mv_costs->nmv_joint_cost,
237 x->mv_costs->mv_cost_stack, MV_COST_WEIGHT);
238
239 // TODO(kyslov) Account for Rate Mode!
240 rv = !(RDCOST(x->rdmult, (*rate_mv), 0) > best_rd_sofar);
241
242 if (rv && search_subpel) {
243 SUBPEL_MOTION_SEARCH_PARAMS ms_params;
244 av1_make_default_subpel_ms_params(&ms_params, cpi, x, bsize, &ref_mv,
245 cost_list);
246 const bool fullpel_performed_well =
247 (bsize == BLOCK_64X64 && full_var_rd * 40 < 62267 * 7) ||
248 (bsize == BLOCK_32X32 && full_var_rd * 8 < 42380) ||
249 (bsize == BLOCK_16X16 && full_var_rd * 8 < 10127);
250 if (sf->rt_sf.reduce_mv_pel_precision_highmotion ||
251 sf->rt_sf.reduce_mv_pel_precision_lowcomplex)
252 ms_params.forced_stop = subpel_select(cpi, x, bsize, tmp_mv, ref_mv,
253 start_mv, fullpel_performed_well);
254
255 MV subpel_start_mv = get_mv_from_fullmv(&tmp_mv->as_fullmv);
256 assert(av1_is_subpelmv_in_range(&ms_params.mv_limits, subpel_start_mv));
257 // adaptively downgrade subpel search method based on block properties
258 if (use_aggressive_subpel_search_method(
259 x, sf->rt_sf.use_adaptive_subpel_search, fullpel_performed_well))
260 av1_find_best_sub_pixel_tree_pruned_more(
261 xd, cm, &ms_params, subpel_start_mv, &best_mv_stats, &tmp_mv->as_mv,
262 &dis, &x->pred_sse[ref], NULL);
263 else
264 cpi->mv_search_params.find_fractional_mv_step(
265 xd, cm, &ms_params, subpel_start_mv, &best_mv_stats, &tmp_mv->as_mv,
266 &dis, &x->pred_sse[ref], NULL);
267 *rate_mv =
268 av1_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->mv_costs->nmv_joint_cost,
269 x->mv_costs->mv_cost_stack, MV_COST_WEIGHT);
270 }
271 // The final MV can not be equal to the reference MV as this will trigger an
272 // assert later. This can happen if both NEAREST and NEAR modes were skipped.
273 rv = (tmp_mv->as_mv.col != ref_mv.col || tmp_mv->as_mv.row != ref_mv.row);
274 return rv;
275 }
276
277 /*!\brief Searches for the best New Motion Vector.
278 *
279 * \ingroup nonrd_mode_search
280 * \callgraph
281 * \callergraph
282 * Finds the best Motion Vector by doing Motion Estimation. Uses reduced
283 * complexity ME for non-LAST frames or calls \c combined_motion_search
284 * for LAST reference frame
285 * \param[in] cpi Top-level encoder structure
286 * \param[in] x Pointer to structure holding all the
287 * data for the current macroblock
288 * \param[in] frame_mv Array that holds MVs for all modes
289 * and ref frames
290 * \param[in] ref_frame Reference frame for which to find
291 * the best New MVs
292 * \param[in] gf_temporal_ref Flag, indicating temporal reference
293 * for GOLDEN frame
294 * \param[in] bsize Current block size
295 * \param[in] mi_row Row index in 4x4 units
296 * \param[in] mi_col Column index in 4x4 units
297 * \param[in] rate_mv Pointer to Rate of the best new MV
298 * \param[in] best_rdc Pointer to the RD Cost for the best
299 * mode found so far
300 *
301 * \return Returns -1 if the search was not done, otherwise returns 0.
302 * Best New MV is placed into \c frame_mv array, Rate estimation for this
303 * vector is placed to \c rate_mv
304 */
search_new_mv(AV1_COMP * cpi,MACROBLOCK * x,int_mv frame_mv[][REF_FRAMES],MV_REFERENCE_FRAME ref_frame,int gf_temporal_ref,BLOCK_SIZE bsize,int mi_row,int mi_col,int * rate_mv,RD_STATS * best_rdc)305 static int search_new_mv(AV1_COMP *cpi, MACROBLOCK *x,
306 int_mv frame_mv[][REF_FRAMES],
307 MV_REFERENCE_FRAME ref_frame, int gf_temporal_ref,
308 BLOCK_SIZE bsize, int mi_row, int mi_col, int *rate_mv,
309 RD_STATS *best_rdc) {
310 MACROBLOCKD *const xd = &x->e_mbd;
311 MB_MODE_INFO *const mi = xd->mi[0];
312 AV1_COMMON *cm = &cpi->common;
313 int_mv *this_ref_frm_newmv = &frame_mv[NEWMV][ref_frame];
314 unsigned int y_sad_zero;
315 if (ref_frame > LAST_FRAME && cpi->oxcf.rc_cfg.mode == AOM_CBR &&
316 gf_temporal_ref) {
317 int tmp_sad;
318 int dis;
319
320 if (bsize < BLOCK_16X16) return -1;
321
322 int me_search_size_col = block_size_wide[bsize] >> 1;
323 int me_search_size_row = block_size_high[bsize] >> 1;
324 tmp_sad = av1_int_pro_motion_estimation(
325 cpi, x, bsize, mi_row, mi_col,
326 &x->mbmi_ext.ref_mv_stack[ref_frame][0].this_mv.as_mv, &y_sad_zero,
327 me_search_size_col, me_search_size_row);
328
329 if (tmp_sad > x->pred_mv_sad[LAST_FRAME]) return -1;
330
331 this_ref_frm_newmv->as_int = mi->mv[0].as_int;
332 int_mv best_mv = mi->mv[0];
333 best_mv.as_mv.row >>= 3;
334 best_mv.as_mv.col >>= 3;
335 MV ref_mv = av1_get_ref_mv(x, 0).as_mv;
336 this_ref_frm_newmv->as_mv.row >>= 3;
337 this_ref_frm_newmv->as_mv.col >>= 3;
338
339 SUBPEL_MOTION_SEARCH_PARAMS ms_params;
340 av1_make_default_subpel_ms_params(&ms_params, cpi, x, bsize, &ref_mv, NULL);
341 if (cpi->sf.rt_sf.reduce_mv_pel_precision_highmotion ||
342 cpi->sf.rt_sf.reduce_mv_pel_precision_lowcomplex) {
343 FULLPEL_MV start_mv = { .row = 0, .col = 0 };
344 ms_params.forced_stop =
345 subpel_select(cpi, x, bsize, &best_mv, ref_mv, start_mv, false);
346 }
347 MV start_mv = get_mv_from_fullmv(&best_mv.as_fullmv);
348 assert(av1_is_subpelmv_in_range(&ms_params.mv_limits, start_mv));
349 cpi->mv_search_params.find_fractional_mv_step(
350 xd, cm, &ms_params, start_mv, NULL, &best_mv.as_mv, &dis,
351 &x->pred_sse[ref_frame], NULL);
352 this_ref_frm_newmv->as_int = best_mv.as_int;
353
354 // When NEWMV is same as ref_mv from the drl, it is preferred to code the
355 // MV as NEARESTMV or NEARMV. In this case, NEWMV needs to be skipped to
356 // avoid an assert failure at a later stage. The scenario can occur if
357 // NEARESTMV was not evaluated for ALTREF.
358 if (this_ref_frm_newmv->as_mv.col == ref_mv.col &&
359 this_ref_frm_newmv->as_mv.row == ref_mv.row)
360 return -1;
361
362 *rate_mv = av1_mv_bit_cost(&this_ref_frm_newmv->as_mv, &ref_mv,
363 x->mv_costs->nmv_joint_cost,
364 x->mv_costs->mv_cost_stack, MV_COST_WEIGHT);
365 } else if (!combined_motion_search(cpi, x, bsize, &frame_mv[NEWMV][ref_frame],
366 rate_mv, best_rdc->rdcost, 0)) {
367 return -1;
368 }
369
370 return 0;
371 }
372
estimate_single_ref_frame_costs(const AV1_COMMON * cm,const MACROBLOCKD * xd,const ModeCosts * mode_costs,int segment_id,BLOCK_SIZE bsize,unsigned int * ref_costs_single)373 static void estimate_single_ref_frame_costs(const AV1_COMMON *cm,
374 const MACROBLOCKD *xd,
375 const ModeCosts *mode_costs,
376 int segment_id, BLOCK_SIZE bsize,
377 unsigned int *ref_costs_single) {
378 int seg_ref_active =
379 segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME);
380 if (seg_ref_active) {
381 memset(ref_costs_single, 0, REF_FRAMES * sizeof(*ref_costs_single));
382 } else {
383 int intra_inter_ctx = av1_get_intra_inter_context(xd);
384 ref_costs_single[INTRA_FRAME] =
385 mode_costs->intra_inter_cost[intra_inter_ctx][0];
386 unsigned int base_cost = mode_costs->intra_inter_cost[intra_inter_ctx][1];
387 if (cm->current_frame.reference_mode == REFERENCE_MODE_SELECT &&
388 is_comp_ref_allowed(bsize)) {
389 const int comp_ref_type_ctx = av1_get_comp_reference_type_context(xd);
390 base_cost += mode_costs->comp_ref_type_cost[comp_ref_type_ctx][1];
391 }
392 ref_costs_single[LAST_FRAME] = base_cost;
393 ref_costs_single[GOLDEN_FRAME] = base_cost;
394 ref_costs_single[ALTREF_FRAME] = base_cost;
395 // add cost for last, golden, altref
396 ref_costs_single[LAST_FRAME] += mode_costs->single_ref_cost[0][0][0];
397 ref_costs_single[GOLDEN_FRAME] += mode_costs->single_ref_cost[0][0][1];
398 ref_costs_single[GOLDEN_FRAME] += mode_costs->single_ref_cost[0][1][0];
399 ref_costs_single[ALTREF_FRAME] += mode_costs->single_ref_cost[0][0][1];
400 ref_costs_single[ALTREF_FRAME] += mode_costs->single_ref_cost[0][2][0];
401 }
402 }
403
set_force_skip_flag(const AV1_COMP * const cpi,MACROBLOCK * const x,unsigned int sse,int * force_skip)404 static INLINE void set_force_skip_flag(const AV1_COMP *const cpi,
405 MACROBLOCK *const x, unsigned int sse,
406 int *force_skip) {
407 if (x->txfm_search_params.tx_mode_search_type == TX_MODE_SELECT &&
408 cpi->sf.rt_sf.tx_size_level_based_on_qstep &&
409 cpi->sf.rt_sf.tx_size_level_based_on_qstep >= 2) {
410 const int qstep = x->plane[AOM_PLANE_Y].dequant_QTX[1] >> (x->e_mbd.bd - 5);
411 const unsigned int qstep_sq = qstep * qstep;
412 // If the sse is low for low source variance blocks, mark those as
413 // transform skip.
414 // Note: Though qstep_sq is based on ac qstep, the threshold is kept
415 // low so that reliable early estimate of tx skip can be obtained
416 // through its comparison with sse.
417 if (sse < qstep_sq && x->source_variance < qstep_sq &&
418 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] == 0 &&
419 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] == 0)
420 *force_skip = 1;
421 }
422 }
423
424 #define CAP_TX_SIZE_FOR_BSIZE_GT32(tx_mode_search_type, bsize) \
425 (((tx_mode_search_type) != ONLY_4X4 && (bsize) > BLOCK_32X32) ? true : false)
426 #define TX_SIZE_FOR_BSIZE_GT32 (TX_16X16)
427
calculate_tx_size(const AV1_COMP * const cpi,BLOCK_SIZE bsize,MACROBLOCK * const x,unsigned int var,unsigned int sse,int * force_skip)428 static TX_SIZE calculate_tx_size(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
429 MACROBLOCK *const x, unsigned int var,
430 unsigned int sse, int *force_skip) {
431 MACROBLOCKD *const xd = &x->e_mbd;
432 TX_SIZE tx_size;
433 const TxfmSearchParams *txfm_params = &x->txfm_search_params;
434 if (txfm_params->tx_mode_search_type == TX_MODE_SELECT) {
435 int multiplier = 8;
436 unsigned int var_thresh = 0;
437 unsigned int is_high_var = 1;
438 // Use quantizer based thresholds to determine transform size.
439 if (cpi->sf.rt_sf.tx_size_level_based_on_qstep) {
440 const int qband = x->qindex >> (QINDEX_BITS - 2);
441 const int mult[4] = { 8, 7, 6, 5 };
442 assert(qband < 4);
443 multiplier = mult[qband];
444 const int qstep = x->plane[AOM_PLANE_Y].dequant_QTX[1] >> (xd->bd - 5);
445 const unsigned int qstep_sq = qstep * qstep;
446 var_thresh = qstep_sq * 2;
447 if (cpi->sf.rt_sf.tx_size_level_based_on_qstep >= 2) {
448 // If the sse is low for low source variance blocks, mark those as
449 // transform skip.
450 // Note: Though qstep_sq is based on ac qstep, the threshold is kept
451 // low so that reliable early estimate of tx skip can be obtained
452 // through its comparison with sse.
453 if (sse < qstep_sq && x->source_variance < qstep_sq &&
454 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] == 0 &&
455 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] == 0)
456 *force_skip = 1;
457 // Further lower transform size based on aq mode only if residual
458 // variance is high.
459 is_high_var = (var >= var_thresh);
460 }
461 }
462 // Choose larger transform size for blocks where dc component is dominant or
463 // the ac component is low.
464 if (sse > ((var * multiplier) >> 2) || (var < var_thresh))
465 tx_size =
466 AOMMIN(max_txsize_lookup[bsize],
467 tx_mode_to_biggest_tx_size[txfm_params->tx_mode_search_type]);
468 else
469 tx_size = TX_8X8;
470
471 if (cpi->oxcf.q_cfg.aq_mode == CYCLIC_REFRESH_AQ &&
472 cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) && is_high_var)
473 tx_size = TX_8X8;
474 else if (tx_size > TX_16X16)
475 tx_size = TX_16X16;
476 } else {
477 tx_size =
478 AOMMIN(max_txsize_lookup[bsize],
479 tx_mode_to_biggest_tx_size[txfm_params->tx_mode_search_type]);
480 }
481
482 if (CAP_TX_SIZE_FOR_BSIZE_GT32(txfm_params->tx_mode_search_type, bsize))
483 tx_size = TX_SIZE_FOR_BSIZE_GT32;
484
485 return AOMMIN(tx_size, TX_16X16);
486 }
487
block_variance(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,int w,int h,unsigned int * sse,int * sum,int block_size,uint32_t * sse8x8,int * sum8x8,uint32_t * var8x8)488 static void block_variance(const uint8_t *src, int src_stride,
489 const uint8_t *ref, int ref_stride, int w, int h,
490 unsigned int *sse, int *sum, int block_size,
491 uint32_t *sse8x8, int *sum8x8, uint32_t *var8x8) {
492 int k = 0;
493 *sse = 0;
494 *sum = 0;
495
496 // This function is called for block sizes >= BLOCK_32x32. As per the design
497 // the aom_get_var_sse_sum_8x8_quad() processes four 8x8 blocks (in a 8x32)
498 // per call. Hence the width and height of the block need to be at least 8 and
499 // 32 samples respectively.
500 assert(w >= 32);
501 assert(h >= 8);
502 for (int row = 0; row < h; row += block_size) {
503 for (int col = 0; col < w; col += 32) {
504 aom_get_var_sse_sum_8x8_quad(src + src_stride * row + col, src_stride,
505 ref + ref_stride * row + col, ref_stride,
506 &sse8x8[k], &sum8x8[k], sse, sum,
507 &var8x8[k]);
508 k += 4;
509 }
510 }
511 }
512
block_variance_16x16_dual(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,int w,int h,unsigned int * sse,int * sum,int block_size,uint32_t * sse16x16,uint32_t * var16x16)513 static void block_variance_16x16_dual(const uint8_t *src, int src_stride,
514 const uint8_t *ref, int ref_stride, int w,
515 int h, unsigned int *sse, int *sum,
516 int block_size, uint32_t *sse16x16,
517 uint32_t *var16x16) {
518 int k = 0;
519 *sse = 0;
520 *sum = 0;
521 // This function is called for block sizes >= BLOCK_32x32. As per the design
522 // the aom_get_var_sse_sum_16x16_dual() processes four 16x16 blocks (in a
523 // 16x32) per call. Hence the width and height of the block need to be at
524 // least 16 and 32 samples respectively.
525 assert(w >= 32);
526 assert(h >= 16);
527 for (int row = 0; row < h; row += block_size) {
528 for (int col = 0; col < w; col += 32) {
529 aom_get_var_sse_sum_16x16_dual(src + src_stride * row + col, src_stride,
530 ref + ref_stride * row + col, ref_stride,
531 &sse16x16[k], sse, sum, &var16x16[k]);
532 k += 2;
533 }
534 }
535 }
536
calculate_variance(int bw,int bh,TX_SIZE tx_size,unsigned int * sse_i,int * sum_i,unsigned int * var_o,unsigned int * sse_o,int * sum_o)537 static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
538 unsigned int *sse_i, int *sum_i,
539 unsigned int *var_o, unsigned int *sse_o,
540 int *sum_o) {
541 const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
542 const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
543 const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
544 int row, col, k = 0;
545
546 for (row = 0; row < nh; row += 2) {
547 for (col = 0; col < nw; col += 2) {
548 sse_o[k] = sse_i[row * nw + col] + sse_i[row * nw + col + 1] +
549 sse_i[(row + 1) * nw + col] + sse_i[(row + 1) * nw + col + 1];
550 sum_o[k] = sum_i[row * nw + col] + sum_i[row * nw + col + 1] +
551 sum_i[(row + 1) * nw + col] + sum_i[(row + 1) * nw + col + 1];
552 var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >>
553 (b_width_log2_lookup[unit_size] +
554 b_height_log2_lookup[unit_size] + 6));
555 k++;
556 }
557 }
558 }
559
560 // Adjust the ac_thr according to speed, width, height and normalized sum
ac_thr_factor(int speed,int width,int height,int norm_sum)561 static int ac_thr_factor(int speed, int width, int height, int norm_sum) {
562 if (speed >= 8 && norm_sum < 5) {
563 if (width <= 640 && height <= 480)
564 return 4;
565 else
566 return 2;
567 }
568 return 1;
569 }
570
571 // Sets early_term flag based on chroma planes prediction
set_early_term_based_on_uv_plane(AV1_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,MACROBLOCKD * xd,int mi_row,int mi_col,int * early_term,int num_blk,const unsigned int * sse_tx,const unsigned int * var_tx,int sum,unsigned int var,unsigned int sse)572 static INLINE void set_early_term_based_on_uv_plane(
573 AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, MACROBLOCKD *xd, int mi_row,
574 int mi_col, int *early_term, int num_blk, const unsigned int *sse_tx,
575 const unsigned int *var_tx, int sum, unsigned int var, unsigned int sse) {
576 AV1_COMMON *const cm = &cpi->common;
577 struct macroblock_plane *const p = &x->plane[AOM_PLANE_Y];
578 const uint32_t dc_quant = p->dequant_QTX[0];
579 const uint32_t ac_quant = p->dequant_QTX[1];
580 int64_t dc_thr = dc_quant * dc_quant >> 6;
581 int64_t ac_thr = ac_quant * ac_quant >> 6;
582 const int bw = b_width_log2_lookup[bsize];
583 const int bh = b_height_log2_lookup[bsize];
584 int ac_test = 1;
585 int dc_test = 1;
586 const int norm_sum = abs(sum) >> (bw + bh);
587
588 #if CONFIG_AV1_TEMPORAL_DENOISING
589 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
590 cpi->oxcf.speed > 5)
591 ac_thr = av1_scale_acskip_thresh(ac_thr, cpi->denoiser.denoising_level,
592 norm_sum, cpi->svc.temporal_layer_id);
593 else
594 ac_thr *= ac_thr_factor(cpi->oxcf.speed, cm->width, cm->height, norm_sum);
595 #else
596 ac_thr *= ac_thr_factor(cpi->oxcf.speed, cm->width, cm->height, norm_sum);
597
598 #endif
599
600 if (cpi->sf.rt_sf.increase_source_sad_thresh) {
601 dc_thr = dc_thr << 1;
602 ac_thr = ac_thr << 2;
603 }
604
605 for (int k = 0; k < num_blk; k++) {
606 // Check if all ac coefficients can be quantized to zero.
607 if (!(var_tx[k] < ac_thr || var == 0)) {
608 ac_test = 0;
609 break;
610 }
611 // Check if dc coefficient can be quantized to zero.
612 if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
613 dc_test = 0;
614 break;
615 }
616 }
617
618 // Check if chroma can be skipped based on ac and dc test flags.
619 if (ac_test && dc_test) {
620 int skip_uv[2] = { 0 };
621 unsigned int var_uv[2];
622 unsigned int sse_uv[2];
623 // Transform skipping test in UV planes.
624 for (int plane = AOM_PLANE_U; plane <= AOM_PLANE_V; plane++) {
625 int j = plane - 1;
626 skip_uv[j] = 1;
627 if (x->color_sensitivity[COLOR_SENS_IDX(plane)]) {
628 skip_uv[j] = 0;
629 struct macroblock_plane *const puv = &x->plane[plane];
630 struct macroblockd_plane *const puvd = &xd->plane[plane];
631 const BLOCK_SIZE uv_bsize = get_plane_block_size(
632 bsize, puvd->subsampling_x, puvd->subsampling_y);
633 // Adjust these thresholds for UV.
634 const int shift_ac = cpi->sf.rt_sf.increase_source_sad_thresh ? 5 : 3;
635 const int shift_dc = cpi->sf.rt_sf.increase_source_sad_thresh ? 4 : 3;
636 const int64_t uv_dc_thr =
637 (puv->dequant_QTX[0] * puv->dequant_QTX[0]) >> shift_dc;
638 const int64_t uv_ac_thr =
639 (puv->dequant_QTX[1] * puv->dequant_QTX[1]) >> shift_ac;
640 av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
641 plane, plane);
642 var_uv[j] = cpi->ppi->fn_ptr[uv_bsize].vf(puv->src.buf, puv->src.stride,
643 puvd->dst.buf,
644 puvd->dst.stride, &sse_uv[j]);
645 if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
646 (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
647 skip_uv[j] = 1;
648 else
649 break;
650 }
651 }
652 if (skip_uv[0] & skip_uv[1]) {
653 *early_term = 1;
654 }
655 }
656 }
657
calc_rate_dist_block_param(AV1_COMP * cpi,MACROBLOCK * x,RD_STATS * rd_stats,int calculate_rd,int * early_term,BLOCK_SIZE bsize,unsigned int sse)658 static INLINE void calc_rate_dist_block_param(AV1_COMP *cpi, MACROBLOCK *x,
659 RD_STATS *rd_stats,
660 int calculate_rd, int *early_term,
661 BLOCK_SIZE bsize,
662 unsigned int sse) {
663 if (calculate_rd) {
664 if (!*early_term) {
665 const int bw = block_size_wide[bsize];
666 const int bh = block_size_high[bsize];
667
668 model_rd_with_curvfit(cpi, x, bsize, AOM_PLANE_Y, rd_stats->sse, bw * bh,
669 &rd_stats->rate, &rd_stats->dist);
670 }
671
672 if (*early_term) {
673 rd_stats->rate = 0;
674 rd_stats->dist = sse << 4;
675 }
676 }
677 }
678
model_skip_for_sb_y_large_64(AV1_COMP * cpi,BLOCK_SIZE bsize,int mi_row,int mi_col,MACROBLOCK * x,MACROBLOCKD * xd,RD_STATS * rd_stats,int * early_term,int calculate_rd,int64_t best_sse,unsigned int * var_output,unsigned int var_prune_threshold)679 static void model_skip_for_sb_y_large_64(AV1_COMP *cpi, BLOCK_SIZE bsize,
680 int mi_row, int mi_col, MACROBLOCK *x,
681 MACROBLOCKD *xd, RD_STATS *rd_stats,
682 int *early_term, int calculate_rd,
683 int64_t best_sse,
684 unsigned int *var_output,
685 unsigned int var_prune_threshold) {
686 // Note our transform coeffs are 8 times an orthogonal transform.
687 // Hence quantizer step is also 8 times. To get effective quantizer
688 // we need to divide by 8 before sending to modeling function.
689 unsigned int sse;
690 struct macroblock_plane *const p = &x->plane[AOM_PLANE_Y];
691 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
692 int test_skip = 1;
693 unsigned int var;
694 int sum;
695 const int bw = b_width_log2_lookup[bsize];
696 const int bh = b_height_log2_lookup[bsize];
697 unsigned int sse16x16[64] = { 0 };
698 unsigned int var16x16[64] = { 0 };
699 assert(xd->mi[0]->tx_size == TX_16X16);
700 assert(bsize > BLOCK_32X32);
701
702 // Calculate variance for whole partition, and also save 16x16 blocks'
703 // variance to be used in following transform skipping test.
704 block_variance_16x16_dual(p->src.buf, p->src.stride, pd->dst.buf,
705 pd->dst.stride, 4 << bw, 4 << bh, &sse, &sum, 16,
706 sse16x16, var16x16);
707
708 var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
709 if (var_output) {
710 *var_output = var;
711 if (*var_output > var_prune_threshold) {
712 return;
713 }
714 }
715
716 rd_stats->sse = sse;
717 // Skipping test
718 *early_term = 0;
719 set_force_skip_flag(cpi, x, sse, early_term);
720 // The code below for setting skip flag assumes transform size of at least
721 // 8x8, so force this lower limit on transform.
722 MB_MODE_INFO *const mi = xd->mi[0];
723 if (!calculate_rd && cpi->sf.rt_sf.sse_early_term_inter_search &&
724 early_term_inter_search_with_sse(
725 cpi->sf.rt_sf.sse_early_term_inter_search, bsize, sse, best_sse,
726 mi->mode))
727 test_skip = 0;
728
729 if (*early_term) test_skip = 0;
730
731 // Evaluate if the partition block is a skippable block in Y plane.
732 if (test_skip) {
733 const unsigned int *sse_tx = sse16x16;
734 const unsigned int *var_tx = var16x16;
735 const unsigned int num_block = (1 << (bw + bh - 2)) >> 2;
736 set_early_term_based_on_uv_plane(cpi, x, bsize, xd, mi_row, mi_col,
737 early_term, num_block, sse_tx, var_tx, sum,
738 var, sse);
739 }
740 calc_rate_dist_block_param(cpi, x, rd_stats, calculate_rd, early_term, bsize,
741 sse);
742 }
743
model_skip_for_sb_y_large(AV1_COMP * cpi,BLOCK_SIZE bsize,int mi_row,int mi_col,MACROBLOCK * x,MACROBLOCKD * xd,RD_STATS * rd_stats,int * early_term,int calculate_rd,int64_t best_sse,unsigned int * var_output,unsigned int var_prune_threshold)744 static void model_skip_for_sb_y_large(AV1_COMP *cpi, BLOCK_SIZE bsize,
745 int mi_row, int mi_col, MACROBLOCK *x,
746 MACROBLOCKD *xd, RD_STATS *rd_stats,
747 int *early_term, int calculate_rd,
748 int64_t best_sse,
749 unsigned int *var_output,
750 unsigned int var_prune_threshold) {
751 if (x->force_zeromv_skip_for_blk) {
752 *early_term = 1;
753 rd_stats->rate = 0;
754 rd_stats->dist = 0;
755 rd_stats->sse = 0;
756 return;
757 }
758
759 // For block sizes greater than 32x32, the transform size is always 16x16.
760 // This function avoids calling calculate_variance() for tx_size 16x16 cases
761 // by directly populating variance at tx_size level from
762 // block_variance_16x16_dual() function.
763 const TxfmSearchParams *txfm_params = &x->txfm_search_params;
764 if (CAP_TX_SIZE_FOR_BSIZE_GT32(txfm_params->tx_mode_search_type, bsize)) {
765 xd->mi[0]->tx_size = TX_SIZE_FOR_BSIZE_GT32;
766 model_skip_for_sb_y_large_64(cpi, bsize, mi_row, mi_col, x, xd, rd_stats,
767 early_term, calculate_rd, best_sse, var_output,
768 var_prune_threshold);
769 return;
770 }
771
772 // Note our transform coeffs are 8 times an orthogonal transform.
773 // Hence quantizer step is also 8 times. To get effective quantizer
774 // we need to divide by 8 before sending to modeling function.
775 unsigned int sse;
776 struct macroblock_plane *const p = &x->plane[AOM_PLANE_Y];
777 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
778 int test_skip = 1;
779 unsigned int var;
780 int sum;
781
782 const int bw = b_width_log2_lookup[bsize];
783 const int bh = b_height_log2_lookup[bsize];
784 unsigned int sse8x8[256] = { 0 };
785 int sum8x8[256] = { 0 };
786 unsigned int var8x8[256] = { 0 };
787 TX_SIZE tx_size;
788
789 // Calculate variance for whole partition, and also save 8x8 blocks' variance
790 // to be used in following transform skipping test.
791 block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
792 4 << bw, 4 << bh, &sse, &sum, 8, sse8x8, sum8x8, var8x8);
793 var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
794 if (var_output) {
795 *var_output = var;
796 if (*var_output > var_prune_threshold) {
797 return;
798 }
799 }
800
801 rd_stats->sse = sse;
802 // Skipping test
803 *early_term = 0;
804 tx_size = calculate_tx_size(cpi, bsize, x, var, sse, early_term);
805 assert(tx_size <= TX_16X16);
806 // The code below for setting skip flag assumes transform size of at least
807 // 8x8, so force this lower limit on transform.
808 if (tx_size < TX_8X8) tx_size = TX_8X8;
809 xd->mi[0]->tx_size = tx_size;
810
811 MB_MODE_INFO *const mi = xd->mi[0];
812 if (!calculate_rd && cpi->sf.rt_sf.sse_early_term_inter_search &&
813 early_term_inter_search_with_sse(
814 cpi->sf.rt_sf.sse_early_term_inter_search, bsize, sse, best_sse,
815 mi->mode))
816 test_skip = 0;
817
818 if (*early_term) test_skip = 0;
819
820 // Evaluate if the partition block is a skippable block in Y plane.
821 if (test_skip) {
822 unsigned int sse16x16[64] = { 0 };
823 int sum16x16[64] = { 0 };
824 unsigned int var16x16[64] = { 0 };
825 const unsigned int *sse_tx = sse8x8;
826 const unsigned int *var_tx = var8x8;
827 unsigned int num_blks = 1 << (bw + bh - 2);
828
829 if (tx_size >= TX_16X16) {
830 calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
831 sum16x16);
832 sse_tx = sse16x16;
833 var_tx = var16x16;
834 num_blks = num_blks >> 2;
835 }
836 set_early_term_based_on_uv_plane(cpi, x, bsize, xd, mi_row, mi_col,
837 early_term, num_blks, sse_tx, var_tx, sum,
838 var, sse);
839 }
840 calc_rate_dist_block_param(cpi, x, rd_stats, calculate_rd, early_term, bsize,
841 sse);
842 }
843
model_rd_for_sb_y(const AV1_COMP * const cpi,BLOCK_SIZE bsize,MACROBLOCK * x,MACROBLOCKD * xd,RD_STATS * rd_stats,unsigned int * var_out,int calculate_rd,int * early_term)844 static void model_rd_for_sb_y(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
845 MACROBLOCK *x, MACROBLOCKD *xd,
846 RD_STATS *rd_stats, unsigned int *var_out,
847 int calculate_rd, int *early_term) {
848 if (x->force_zeromv_skip_for_blk && early_term != NULL) {
849 *early_term = 1;
850 rd_stats->rate = 0;
851 rd_stats->dist = 0;
852 rd_stats->sse = 0;
853 }
854
855 // Note our transform coeffs are 8 times an orthogonal transform.
856 // Hence quantizer step is also 8 times. To get effective quantizer
857 // we need to divide by 8 before sending to modeling function.
858 const int ref = xd->mi[0]->ref_frame[0];
859
860 assert(bsize < BLOCK_SIZES_ALL);
861
862 struct macroblock_plane *const p = &x->plane[AOM_PLANE_Y];
863 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
864 unsigned int sse;
865 int rate;
866 int64_t dist;
867
868 unsigned int var = cpi->ppi->fn_ptr[bsize].vf(
869 p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, &sse);
870 int force_skip = 0;
871 xd->mi[0]->tx_size = calculate_tx_size(cpi, bsize, x, var, sse, &force_skip);
872 if (var_out) {
873 *var_out = var;
874 }
875
876 if (calculate_rd && (!force_skip || ref == INTRA_FRAME)) {
877 const int bwide = block_size_wide[bsize];
878 const int bhigh = block_size_high[bsize];
879 model_rd_with_curvfit(cpi, x, bsize, AOM_PLANE_Y, sse, bwide * bhigh, &rate,
880 &dist);
881 } else {
882 rate = INT_MAX; // this will be overwritten later with av1_block_yrd
883 dist = INT_MAX;
884 }
885 rd_stats->sse = sse;
886 x->pred_sse[ref] = (unsigned int)AOMMIN(sse, UINT_MAX);
887
888 if (force_skip && ref > INTRA_FRAME) {
889 rate = 0;
890 dist = (int64_t)sse << 4;
891 }
892
893 assert(rate >= 0);
894
895 rd_stats->skip_txfm = (rate == 0);
896 rate = AOMMIN(rate, INT_MAX);
897 rd_stats->rate = rate;
898 rd_stats->dist = dist;
899 }
900
get_drl_cost(PREDICTION_MODE this_mode,int ref_mv_idx,const MB_MODE_INFO_EXT * mbmi_ext,const int (* const drl_mode_cost0)[2],int8_t ref_frame_type)901 static INLINE int get_drl_cost(PREDICTION_MODE this_mode, int ref_mv_idx,
902 const MB_MODE_INFO_EXT *mbmi_ext,
903 const int (*const drl_mode_cost0)[2],
904 int8_t ref_frame_type) {
905 int cost = 0;
906 if (this_mode == NEWMV || this_mode == NEW_NEWMV) {
907 for (int idx = 0; idx < 2; ++idx) {
908 if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
909 uint8_t drl_ctx = av1_drl_ctx(mbmi_ext->weight[ref_frame_type], idx);
910 cost += drl_mode_cost0[drl_ctx][ref_mv_idx != idx];
911 if (ref_mv_idx == idx) return cost;
912 }
913 }
914 return cost;
915 }
916
917 if (have_nearmv_in_inter_mode(this_mode)) {
918 for (int idx = 1; idx < 3; ++idx) {
919 if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
920 uint8_t drl_ctx = av1_drl_ctx(mbmi_ext->weight[ref_frame_type], idx);
921 cost += drl_mode_cost0[drl_ctx][ref_mv_idx != (idx - 1)];
922 if (ref_mv_idx == (idx - 1)) return cost;
923 }
924 }
925 return cost;
926 }
927 return cost;
928 }
929
cost_mv_ref(const ModeCosts * const mode_costs,PREDICTION_MODE mode,int16_t mode_context)930 static int cost_mv_ref(const ModeCosts *const mode_costs, PREDICTION_MODE mode,
931 int16_t mode_context) {
932 if (is_inter_compound_mode(mode)) {
933 return mode_costs
934 ->inter_compound_mode_cost[mode_context][INTER_COMPOUND_OFFSET(mode)];
935 }
936
937 int mode_cost = 0;
938 int16_t mode_ctx = mode_context & NEWMV_CTX_MASK;
939
940 assert(is_inter_mode(mode));
941
942 if (mode == NEWMV) {
943 mode_cost = mode_costs->newmv_mode_cost[mode_ctx][0];
944 return mode_cost;
945 } else {
946 mode_cost = mode_costs->newmv_mode_cost[mode_ctx][1];
947 mode_ctx = (mode_context >> GLOBALMV_OFFSET) & GLOBALMV_CTX_MASK;
948
949 if (mode == GLOBALMV) {
950 mode_cost += mode_costs->zeromv_mode_cost[mode_ctx][0];
951 return mode_cost;
952 } else {
953 mode_cost += mode_costs->zeromv_mode_cost[mode_ctx][1];
954 mode_ctx = (mode_context >> REFMV_OFFSET) & REFMV_CTX_MASK;
955 mode_cost += mode_costs->refmv_mode_cost[mode_ctx][mode != NEARESTMV];
956 return mode_cost;
957 }
958 }
959 }
960
newmv_diff_bias(MACROBLOCKD * xd,PREDICTION_MODE this_mode,RD_STATS * this_rdc,BLOCK_SIZE bsize,int mv_row,int mv_col,int speed,uint32_t spatial_variance,CONTENT_STATE_SB content_state_sb)961 static void newmv_diff_bias(MACROBLOCKD *xd, PREDICTION_MODE this_mode,
962 RD_STATS *this_rdc, BLOCK_SIZE bsize, int mv_row,
963 int mv_col, int speed, uint32_t spatial_variance,
964 CONTENT_STATE_SB content_state_sb) {
965 // Bias against MVs associated with NEWMV mode that are very different from
966 // top/left neighbors.
967 if (this_mode == NEWMV) {
968 int al_mv_average_row;
969 int al_mv_average_col;
970 int row_diff, col_diff;
971 int above_mv_valid = 0;
972 int left_mv_valid = 0;
973 int above_row = INVALID_MV_ROW_COL, above_col = INVALID_MV_ROW_COL;
974 int left_row = INVALID_MV_ROW_COL, left_col = INVALID_MV_ROW_COL;
975 if (bsize >= BLOCK_64X64 && content_state_sb.source_sad_nonrd != kHighSad &&
976 spatial_variance < 300 &&
977 (mv_row > 16 || mv_row < -16 || mv_col > 16 || mv_col < -16)) {
978 this_rdc->rdcost = this_rdc->rdcost << 2;
979 return;
980 }
981 if (xd->above_mbmi) {
982 above_mv_valid = xd->above_mbmi->mv[0].as_int != INVALID_MV;
983 above_row = xd->above_mbmi->mv[0].as_mv.row;
984 above_col = xd->above_mbmi->mv[0].as_mv.col;
985 }
986 if (xd->left_mbmi) {
987 left_mv_valid = xd->left_mbmi->mv[0].as_int != INVALID_MV;
988 left_row = xd->left_mbmi->mv[0].as_mv.row;
989 left_col = xd->left_mbmi->mv[0].as_mv.col;
990 }
991 if (above_mv_valid && left_mv_valid) {
992 al_mv_average_row = (above_row + left_row + 1) >> 1;
993 al_mv_average_col = (above_col + left_col + 1) >> 1;
994 } else if (above_mv_valid) {
995 al_mv_average_row = above_row;
996 al_mv_average_col = above_col;
997 } else if (left_mv_valid) {
998 al_mv_average_row = left_row;
999 al_mv_average_col = left_col;
1000 } else {
1001 al_mv_average_row = al_mv_average_col = 0;
1002 }
1003 row_diff = al_mv_average_row - mv_row;
1004 col_diff = al_mv_average_col - mv_col;
1005 if (row_diff > 80 || row_diff < -80 || col_diff > 80 || col_diff < -80) {
1006 if (bsize >= BLOCK_32X32)
1007 this_rdc->rdcost = this_rdc->rdcost << 1;
1008 else
1009 this_rdc->rdcost = 5 * this_rdc->rdcost >> 2;
1010 }
1011 } else {
1012 // Bias for speed >= 8 for low spatial variance.
1013 if (speed >= 8 && spatial_variance < 150 &&
1014 (mv_row > 64 || mv_row < -64 || mv_col > 64 || mv_col < -64))
1015 this_rdc->rdcost = 5 * this_rdc->rdcost >> 2;
1016 }
1017 }
1018
update_thresh_freq_fact(AV1_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,MV_REFERENCE_FRAME ref_frame,THR_MODES best_mode_idx,PREDICTION_MODE mode)1019 static INLINE void update_thresh_freq_fact(AV1_COMP *cpi, MACROBLOCK *x,
1020 BLOCK_SIZE bsize,
1021 MV_REFERENCE_FRAME ref_frame,
1022 THR_MODES best_mode_idx,
1023 PREDICTION_MODE mode) {
1024 const THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1025 const BLOCK_SIZE min_size = AOMMAX(bsize - 3, BLOCK_4X4);
1026 const BLOCK_SIZE max_size = AOMMIN(bsize + 6, BLOCK_128X128);
1027 for (BLOCK_SIZE bs = min_size; bs <= max_size; bs += 3) {
1028 int *freq_fact = &x->thresh_freq_fact[bs][thr_mode_idx];
1029 if (thr_mode_idx == best_mode_idx) {
1030 *freq_fact -= (*freq_fact >> 4);
1031 } else {
1032 *freq_fact =
1033 AOMMIN(*freq_fact + RD_THRESH_INC,
1034 cpi->sf.inter_sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1035 }
1036 }
1037 }
1038
1039 #if CONFIG_AV1_TEMPORAL_DENOISING
av1_pickmode_ctx_den_update(AV1_PICKMODE_CTX_DEN * ctx_den,int64_t zero_last_cost_orig,unsigned int ref_frame_cost[REF_FRAMES],int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES],int reuse_inter_pred,BEST_PICKMODE * bp)1040 static void av1_pickmode_ctx_den_update(
1041 AV1_PICKMODE_CTX_DEN *ctx_den, int64_t zero_last_cost_orig,
1042 unsigned int ref_frame_cost[REF_FRAMES],
1043 int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES], int reuse_inter_pred,
1044 BEST_PICKMODE *bp) {
1045 ctx_den->zero_last_cost_orig = zero_last_cost_orig;
1046 ctx_den->ref_frame_cost = ref_frame_cost;
1047 ctx_den->frame_mv = frame_mv;
1048 ctx_den->reuse_inter_pred = reuse_inter_pred;
1049 ctx_den->best_tx_size = bp->best_tx_size;
1050 ctx_den->best_mode = bp->best_mode;
1051 ctx_den->best_ref_frame = bp->best_ref_frame;
1052 ctx_den->best_pred_filter = bp->best_pred_filter;
1053 ctx_den->best_mode_skip_txfm = bp->best_mode_skip_txfm;
1054 }
1055
recheck_zeromv_after_denoising(AV1_COMP * cpi,MB_MODE_INFO * const mi,MACROBLOCK * x,MACROBLOCKD * const xd,AV1_DENOISER_DECISION decision,AV1_PICKMODE_CTX_DEN * ctx_den,struct buf_2d yv12_mb[4][MAX_MB_PLANE],RD_STATS * best_rdc,BEST_PICKMODE * best_pickmode,BLOCK_SIZE bsize,int mi_row,int mi_col)1056 static void recheck_zeromv_after_denoising(
1057 AV1_COMP *cpi, MB_MODE_INFO *const mi, MACROBLOCK *x, MACROBLOCKD *const xd,
1058 AV1_DENOISER_DECISION decision, AV1_PICKMODE_CTX_DEN *ctx_den,
1059 struct buf_2d yv12_mb[4][MAX_MB_PLANE], RD_STATS *best_rdc,
1060 BEST_PICKMODE *best_pickmode, BLOCK_SIZE bsize, int mi_row, int mi_col) {
1061 // If INTRA or GOLDEN reference was selected, re-evaluate ZEROMV on
1062 // denoised result. Only do this under noise conditions, and if rdcost of
1063 // ZEROMV on original source is not significantly higher than rdcost of best
1064 // mode.
1065 if (cpi->noise_estimate.enabled && cpi->noise_estimate.level > kLow &&
1066 ctx_den->zero_last_cost_orig < (best_rdc->rdcost << 3) &&
1067 ((ctx_den->best_ref_frame == INTRA_FRAME && decision >= FILTER_BLOCK) ||
1068 (ctx_den->best_ref_frame == GOLDEN_FRAME &&
1069 cpi->svc.number_spatial_layers == 1 &&
1070 decision == FILTER_ZEROMV_BLOCK))) {
1071 // Check if we should pick ZEROMV on denoised signal.
1072 AV1_COMMON *const cm = &cpi->common;
1073 RD_STATS this_rdc;
1074 const ModeCosts *mode_costs = &x->mode_costs;
1075 TxfmSearchInfo *txfm_info = &x->txfm_search_info;
1076 MB_MODE_INFO_EXT *const mbmi_ext = &x->mbmi_ext;
1077
1078 mi->mode = GLOBALMV;
1079 mi->ref_frame[0] = LAST_FRAME;
1080 mi->ref_frame[1] = NONE_FRAME;
1081 set_ref_ptrs(cm, xd, mi->ref_frame[0], NONE_FRAME);
1082 mi->mv[0].as_int = 0;
1083 mi->interp_filters = av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
1084 xd->plane[AOM_PLANE_Y].pre[0] = yv12_mb[LAST_FRAME][AOM_PLANE_Y];
1085 av1_enc_build_inter_predictor_y(xd, mi_row, mi_col);
1086 unsigned int var;
1087 model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc, &var, 1, NULL);
1088
1089 const int16_t mode_ctx =
1090 av1_mode_context_analyzer(mbmi_ext->mode_context, mi->ref_frame);
1091 this_rdc.rate += cost_mv_ref(mode_costs, GLOBALMV, mode_ctx);
1092
1093 this_rdc.rate += ctx_den->ref_frame_cost[LAST_FRAME];
1094 this_rdc.rdcost = RDCOST(x->rdmult, this_rdc.rate, this_rdc.dist);
1095 txfm_info->skip_txfm = this_rdc.skip_txfm;
1096 // Don't switch to ZEROMV if the rdcost for ZEROMV on denoised source
1097 // is higher than best_ref mode (on original source).
1098 if (this_rdc.rdcost > best_rdc->rdcost) {
1099 this_rdc = *best_rdc;
1100 mi->mode = best_pickmode->best_mode;
1101 mi->ref_frame[0] = best_pickmode->best_ref_frame;
1102 set_ref_ptrs(cm, xd, mi->ref_frame[0], NONE_FRAME);
1103 mi->interp_filters = best_pickmode->best_pred_filter;
1104 if (best_pickmode->best_ref_frame == INTRA_FRAME) {
1105 mi->mv[0].as_int = INVALID_MV;
1106 } else {
1107 mi->mv[0].as_int = ctx_den
1108 ->frame_mv[best_pickmode->best_mode]
1109 [best_pickmode->best_ref_frame]
1110 .as_int;
1111 if (ctx_den->reuse_inter_pred) {
1112 xd->plane[AOM_PLANE_Y].pre[0] = yv12_mb[GOLDEN_FRAME][AOM_PLANE_Y];
1113 av1_enc_build_inter_predictor_y(xd, mi_row, mi_col);
1114 }
1115 }
1116 mi->tx_size = best_pickmode->best_tx_size;
1117 txfm_info->skip_txfm = best_pickmode->best_mode_skip_txfm;
1118 } else {
1119 ctx_den->best_ref_frame = LAST_FRAME;
1120 *best_rdc = this_rdc;
1121 }
1122 }
1123 }
1124 #endif // CONFIG_AV1_TEMPORAL_DENOISING
1125
1126 /*!\brief Searches for the best interpolation filter
1127 *
1128 * \ingroup nonrd_mode_search
1129 * \callgraph
1130 * \callergraph
1131 * Iterates through subset of possible interpolation filters (EIGHTTAP_REGULAR,
1132 * EIGTHTAP_SMOOTH, MULTITAP_SHARP, depending on FILTER_SEARCH_SIZE) and selects
1133 * the one that gives lowest RD cost. RD cost is calculated using curvfit model.
1134 * Support for dual filters (different filters in the x & y directions) is
1135 * allowed if sf.interp_sf.disable_dual_filter = 0.
1136 *
1137 * \param[in] cpi Top-level encoder structure
1138 * \param[in] x Pointer to structure holding all the
1139 * data for the current macroblock
1140 * \param[in] this_rdc Pointer to calculated RD Cost
1141 * \param[in] inter_pred_params_sr Pointer to structure holding parameters of
1142 inter prediction for single reference
1143 * \param[in] mi_row Row index in 4x4 units
1144 * \param[in] mi_col Column index in 4x4 units
1145 * \param[in] tmp_buffer Pointer to a temporary buffer for
1146 * prediction re-use
1147 * \param[in] bsize Current block size
1148 * \param[in] reuse_inter_pred Flag, indicating prediction re-use
1149 * \param[out] this_mode_pred Pointer to store prediction buffer
1150 * for prediction re-use
1151 * \param[out] this_early_term Flag, indicating that transform can be
1152 * skipped
1153 * \param[out] var The residue variance of the current
1154 * predictor.
1155 * \param[in] use_model_yrd_large Flag, indicating special logic to handle
1156 * large blocks
1157 * \param[in] best_sse Best sse so far.
1158 * \param[in] is_single_pred Flag, indicating single mode.
1159 *
1160 * \remark Nothing is returned. Instead, calculated RD cost is placed to
1161 * \c this_rdc and best filter is placed to \c mi->interp_filters. In case
1162 * \c reuse_inter_pred flag is set, this function also outputs
1163 * \c this_mode_pred. Also \c this_early_temp is set if transform can be
1164 * skipped
1165 */
search_filter_ref(AV1_COMP * cpi,MACROBLOCK * x,RD_STATS * this_rdc,InterPredParams * inter_pred_params_sr,int mi_row,int mi_col,PRED_BUFFER * tmp_buffer,BLOCK_SIZE bsize,int reuse_inter_pred,PRED_BUFFER ** this_mode_pred,int * this_early_term,unsigned int * var,int use_model_yrd_large,int64_t best_sse,int is_single_pred)1166 static void search_filter_ref(AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *this_rdc,
1167 InterPredParams *inter_pred_params_sr, int mi_row,
1168 int mi_col, PRED_BUFFER *tmp_buffer,
1169 BLOCK_SIZE bsize, int reuse_inter_pred,
1170 PRED_BUFFER **this_mode_pred,
1171 int *this_early_term, unsigned int *var,
1172 int use_model_yrd_large, int64_t best_sse,
1173 int is_single_pred) {
1174 AV1_COMMON *const cm = &cpi->common;
1175 MACROBLOCKD *const xd = &x->e_mbd;
1176 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
1177 MB_MODE_INFO *const mi = xd->mi[0];
1178 const int bw = block_size_wide[bsize];
1179 int dim_factor =
1180 (cpi->sf.interp_sf.disable_dual_filter == 0) ? FILTER_SEARCH_SIZE : 1;
1181 RD_STATS pf_rd_stats[FILTER_SEARCH_SIZE * FILTER_SEARCH_SIZE] = { 0 };
1182 TX_SIZE pf_tx_size[FILTER_SEARCH_SIZE * FILTER_SEARCH_SIZE] = { 0 };
1183 PRED_BUFFER *current_pred = *this_mode_pred;
1184 int best_skip = 0;
1185 int best_early_term = 0;
1186 int64_t best_cost = INT64_MAX;
1187 int best_filter_index = -1;
1188
1189 SubpelParams subpel_params;
1190 // Initialize inter prediction params at mode level for single reference
1191 // mode.
1192 if (is_single_pred)
1193 init_inter_mode_params(&mi->mv[0].as_mv, inter_pred_params_sr,
1194 &subpel_params, xd->block_ref_scale_factors[0],
1195 pd->pre->width, pd->pre->height);
1196 for (int filter_idx = 0; filter_idx < FILTER_SEARCH_SIZE * FILTER_SEARCH_SIZE;
1197 ++filter_idx) {
1198 int64_t cost;
1199 if (cpi->sf.interp_sf.disable_dual_filter &&
1200 filters_ref_set[filter_idx].as_filters.x_filter !=
1201 filters_ref_set[filter_idx].as_filters.y_filter)
1202 continue;
1203
1204 mi->interp_filters.as_int = filters_ref_set[filter_idx].as_int;
1205 if (is_single_pred)
1206 av1_enc_build_inter_predictor_y_nonrd(xd, inter_pred_params_sr,
1207 &subpel_params);
1208 else
1209 av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
1210 AOM_PLANE_Y, AOM_PLANE_Y);
1211 unsigned int curr_var = UINT_MAX;
1212 if (use_model_yrd_large)
1213 model_skip_for_sb_y_large(cpi, bsize, mi_row, mi_col, x, xd,
1214 &pf_rd_stats[filter_idx], this_early_term, 1,
1215 best_sse, &curr_var, UINT_MAX);
1216 else
1217 model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rd_stats[filter_idx], &curr_var,
1218 1, NULL);
1219 pf_rd_stats[filter_idx].rate += av1_get_switchable_rate(
1220 x, xd, cm->features.interp_filter, cm->seq_params->enable_dual_filter);
1221 cost = RDCOST(x->rdmult, pf_rd_stats[filter_idx].rate,
1222 pf_rd_stats[filter_idx].dist);
1223 pf_tx_size[filter_idx] = mi->tx_size;
1224 if (cost < best_cost) {
1225 *var = curr_var;
1226 best_filter_index = filter_idx;
1227 best_cost = cost;
1228 best_skip = pf_rd_stats[filter_idx].skip_txfm;
1229 best_early_term = *this_early_term;
1230 if (reuse_inter_pred) {
1231 if (*this_mode_pred != current_pred) {
1232 free_pred_buffer(*this_mode_pred);
1233 *this_mode_pred = current_pred;
1234 }
1235 current_pred = &tmp_buffer[get_pred_buffer(tmp_buffer, 3)];
1236 pd->dst.buf = current_pred->data;
1237 pd->dst.stride = bw;
1238 }
1239 }
1240 }
1241 assert(best_filter_index >= 0 &&
1242 best_filter_index < dim_factor * FILTER_SEARCH_SIZE);
1243 if (reuse_inter_pred && *this_mode_pred != current_pred)
1244 free_pred_buffer(current_pred);
1245
1246 mi->interp_filters.as_int = filters_ref_set[best_filter_index].as_int;
1247 mi->tx_size = pf_tx_size[best_filter_index];
1248 this_rdc->rate = pf_rd_stats[best_filter_index].rate;
1249 this_rdc->dist = pf_rd_stats[best_filter_index].dist;
1250 this_rdc->sse = pf_rd_stats[best_filter_index].sse;
1251 this_rdc->skip_txfm = (best_skip || best_early_term);
1252 *this_early_term = best_early_term;
1253 if (reuse_inter_pred) {
1254 pd->dst.buf = (*this_mode_pred)->data;
1255 pd->dst.stride = (*this_mode_pred)->stride;
1256 } else if (best_filter_index < dim_factor * FILTER_SEARCH_SIZE - 1) {
1257 if (is_single_pred)
1258 av1_enc_build_inter_predictor_y_nonrd(xd, inter_pred_params_sr,
1259 &subpel_params);
1260 else
1261 av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
1262 AOM_PLANE_Y, AOM_PLANE_Y);
1263 }
1264 }
1265 #if !CONFIG_REALTIME_ONLY
1266
is_warped_mode_allowed(const AV1_COMP * cpi,MACROBLOCK * const x,const MB_MODE_INFO * mbmi)1267 static AOM_INLINE int is_warped_mode_allowed(const AV1_COMP *cpi,
1268 MACROBLOCK *const x,
1269 const MB_MODE_INFO *mbmi) {
1270 const FeatureFlags *const features = &cpi->common.features;
1271 const MACROBLOCKD *xd = &x->e_mbd;
1272
1273 if (cpi->sf.inter_sf.extra_prune_warped) return 0;
1274 if (has_second_ref(mbmi)) return 0;
1275 MOTION_MODE last_motion_mode_allowed = SIMPLE_TRANSLATION;
1276
1277 if (features->switchable_motion_mode) {
1278 // Determine which motion modes to search if more than SIMPLE_TRANSLATION
1279 // is allowed.
1280 last_motion_mode_allowed = motion_mode_allowed(
1281 xd->global_motion, xd, mbmi, features->allow_warped_motion);
1282 }
1283
1284 if (last_motion_mode_allowed == WARPED_CAUSAL) {
1285 return 1;
1286 }
1287
1288 return 0;
1289 }
1290
calc_num_proj_ref(AV1_COMP * cpi,MACROBLOCK * x,MB_MODE_INFO * mi)1291 static void calc_num_proj_ref(AV1_COMP *cpi, MACROBLOCK *x, MB_MODE_INFO *mi) {
1292 AV1_COMMON *const cm = &cpi->common;
1293 MACROBLOCKD *const xd = &x->e_mbd;
1294 const FeatureFlags *const features = &cm->features;
1295
1296 mi->num_proj_ref = 1;
1297 WARP_SAMPLE_INFO *const warp_sample_info =
1298 &x->warp_sample_info[mi->ref_frame[0]];
1299 int *pts0 = warp_sample_info->pts;
1300 int *pts_inref0 = warp_sample_info->pts_inref;
1301 MOTION_MODE last_motion_mode_allowed = SIMPLE_TRANSLATION;
1302
1303 if (features->switchable_motion_mode) {
1304 // Determine which motion modes to search if more than SIMPLE_TRANSLATION
1305 // is allowed.
1306 last_motion_mode_allowed = motion_mode_allowed(
1307 xd->global_motion, xd, mi, features->allow_warped_motion);
1308 }
1309
1310 if (last_motion_mode_allowed == WARPED_CAUSAL) {
1311 if (warp_sample_info->num < 0) {
1312 warp_sample_info->num = av1_findSamples(cm, xd, pts0, pts_inref0);
1313 }
1314 mi->num_proj_ref = warp_sample_info->num;
1315 }
1316 }
1317
search_motion_mode(AV1_COMP * cpi,MACROBLOCK * x,RD_STATS * this_rdc,int mi_row,int mi_col,BLOCK_SIZE bsize,int * this_early_term,int use_model_yrd_large,int * rate_mv,int64_t best_sse)1318 static void search_motion_mode(AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *this_rdc,
1319 int mi_row, int mi_col, BLOCK_SIZE bsize,
1320 int *this_early_term, int use_model_yrd_large,
1321 int *rate_mv, int64_t best_sse) {
1322 AV1_COMMON *const cm = &cpi->common;
1323 MACROBLOCKD *const xd = &x->e_mbd;
1324 const FeatureFlags *const features = &cm->features;
1325 MB_MODE_INFO *const mi = xd->mi[0];
1326 RD_STATS pf_rd_stats[MOTION_MODE_SEARCH_SIZE] = { 0 };
1327 int best_skip = 0;
1328 int best_early_term = 0;
1329 int64_t best_cost = INT64_MAX;
1330 int best_mode_index = -1;
1331 const int interp_filter = features->interp_filter;
1332
1333 const MOTION_MODE motion_modes[MOTION_MODE_SEARCH_SIZE] = {
1334 SIMPLE_TRANSLATION, WARPED_CAUSAL
1335 };
1336 int mode_search_size = is_warped_mode_allowed(cpi, x, mi) ? 2 : 1;
1337
1338 WARP_SAMPLE_INFO *const warp_sample_info =
1339 &x->warp_sample_info[mi->ref_frame[0]];
1340 int *pts0 = warp_sample_info->pts;
1341 int *pts_inref0 = warp_sample_info->pts_inref;
1342
1343 const int total_samples = mi->num_proj_ref;
1344 if (total_samples == 0) {
1345 // Do not search WARPED_CAUSAL if there are no samples to use to determine
1346 // warped parameters.
1347 mode_search_size = 1;
1348 }
1349
1350 const MB_MODE_INFO base_mbmi = *mi;
1351 MB_MODE_INFO best_mbmi;
1352
1353 for (int mode_index = 0; mode_index < mode_search_size; ++mode_index) {
1354 int64_t cost = INT64_MAX;
1355 MOTION_MODE motion_mode = motion_modes[mode_index];
1356 *mi = base_mbmi;
1357 mi->motion_mode = motion_mode;
1358 if (motion_mode == SIMPLE_TRANSLATION) {
1359 mi->interp_filters = av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
1360
1361 av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
1362 AOM_PLANE_Y, AOM_PLANE_Y);
1363 if (use_model_yrd_large)
1364 model_skip_for_sb_y_large(cpi, bsize, mi_row, mi_col, x, xd,
1365 &pf_rd_stats[mode_index], this_early_term, 1,
1366 best_sse, NULL, UINT_MAX);
1367 else
1368 model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rd_stats[mode_index], NULL, 1,
1369 NULL);
1370 pf_rd_stats[mode_index].rate +=
1371 av1_get_switchable_rate(x, xd, cm->features.interp_filter,
1372 cm->seq_params->enable_dual_filter);
1373 cost = RDCOST(x->rdmult, pf_rd_stats[mode_index].rate,
1374 pf_rd_stats[mode_index].dist);
1375 } else if (motion_mode == WARPED_CAUSAL) {
1376 int pts[SAMPLES_ARRAY_SIZE], pts_inref[SAMPLES_ARRAY_SIZE];
1377 const ModeCosts *mode_costs = &x->mode_costs;
1378 mi->wm_params.wmtype = DEFAULT_WMTYPE;
1379 mi->interp_filters =
1380 av1_broadcast_interp_filter(av1_unswitchable_filter(interp_filter));
1381
1382 memcpy(pts, pts0, total_samples * 2 * sizeof(*pts0));
1383 memcpy(pts_inref, pts_inref0, total_samples * 2 * sizeof(*pts_inref0));
1384 // Select the samples according to motion vector difference
1385 if (mi->num_proj_ref > 1) {
1386 mi->num_proj_ref = av1_selectSamples(&mi->mv[0].as_mv, pts, pts_inref,
1387 mi->num_proj_ref, bsize);
1388 }
1389
1390 // Compute the warped motion parameters with a least squares fit
1391 // using the collected samples
1392 if (!av1_find_projection(mi->num_proj_ref, pts, pts_inref, bsize,
1393 mi->mv[0].as_mv.row, mi->mv[0].as_mv.col,
1394 &mi->wm_params, mi_row, mi_col)) {
1395 if (mi->mode == NEWMV) {
1396 const int_mv mv0 = mi->mv[0];
1397 const WarpedMotionParams wm_params0 = mi->wm_params;
1398 const int num_proj_ref0 = mi->num_proj_ref;
1399
1400 const int_mv ref_mv = av1_get_ref_mv(x, 0);
1401 SUBPEL_MOTION_SEARCH_PARAMS ms_params;
1402 av1_make_default_subpel_ms_params(&ms_params, cpi, x, bsize,
1403 &ref_mv.as_mv, NULL);
1404
1405 // Refine MV in a small range.
1406 av1_refine_warped_mv(xd, cm, &ms_params, bsize, pts0, pts_inref0,
1407 total_samples, cpi->sf.mv_sf.warp_search_method,
1408 cpi->sf.mv_sf.warp_search_iters);
1409 if (mi->mv[0].as_int == ref_mv.as_int) {
1410 continue;
1411 }
1412
1413 if (mv0.as_int != mi->mv[0].as_int) {
1414 // Keep the refined MV and WM parameters.
1415 int tmp_rate_mv = av1_mv_bit_cost(
1416 &mi->mv[0].as_mv, &ref_mv.as_mv, x->mv_costs->nmv_joint_cost,
1417 x->mv_costs->mv_cost_stack, MV_COST_WEIGHT);
1418 *rate_mv = tmp_rate_mv;
1419 } else {
1420 // Restore the old MV and WM parameters.
1421 mi->mv[0] = mv0;
1422 mi->wm_params = wm_params0;
1423 mi->num_proj_ref = num_proj_ref0;
1424 }
1425 }
1426 // Build the warped predictor
1427 av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
1428 AOM_PLANE_Y, av1_num_planes(cm) - 1);
1429 if (use_model_yrd_large)
1430 model_skip_for_sb_y_large(cpi, bsize, mi_row, mi_col, x, xd,
1431 &pf_rd_stats[mode_index], this_early_term,
1432 1, best_sse, NULL, UINT_MAX);
1433 else
1434 model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rd_stats[mode_index], NULL,
1435 1, NULL);
1436
1437 pf_rd_stats[mode_index].rate +=
1438 mode_costs->motion_mode_cost[bsize][mi->motion_mode];
1439 cost = RDCOST(x->rdmult, pf_rd_stats[mode_index].rate,
1440 pf_rd_stats[mode_index].dist);
1441 } else {
1442 cost = INT64_MAX;
1443 }
1444 }
1445 if (cost < best_cost) {
1446 best_mode_index = mode_index;
1447 best_cost = cost;
1448 best_skip = pf_rd_stats[mode_index].skip_txfm;
1449 best_early_term = *this_early_term;
1450 best_mbmi = *mi;
1451 }
1452 }
1453 assert(best_mode_index >= 0 && best_mode_index < FILTER_SEARCH_SIZE);
1454
1455 *mi = best_mbmi;
1456 this_rdc->rate = pf_rd_stats[best_mode_index].rate;
1457 this_rdc->dist = pf_rd_stats[best_mode_index].dist;
1458 this_rdc->sse = pf_rd_stats[best_mode_index].sse;
1459 this_rdc->skip_txfm = (best_skip || best_early_term);
1460 *this_early_term = best_early_term;
1461 if (best_mode_index < FILTER_SEARCH_SIZE - 1) {
1462 av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
1463 AOM_PLANE_Y, AOM_PLANE_Y);
1464 }
1465 }
1466 #endif // !CONFIG_REALTIME_ONLY
1467
1468 #define COLLECT_NON_SQR_STAT 0
1469
1470 #if COLLECT_NONRD_PICK_MODE_STAT
1471
print_stage_time(const char * stage_name,int64_t stage_time,int64_t total_time)1472 static AOM_INLINE void print_stage_time(const char *stage_name,
1473 int64_t stage_time,
1474 int64_t total_time) {
1475 printf(" %s: %ld (%f%%)\n", stage_name, stage_time,
1476 100 * stage_time / (float)total_time);
1477 }
1478
print_time(const mode_search_stat_nonrd * const ms_stat,BLOCK_SIZE bsize,int mi_rows,int mi_cols,int mi_row,int mi_col)1479 static void print_time(const mode_search_stat_nonrd *const ms_stat,
1480 BLOCK_SIZE bsize, int mi_rows, int mi_cols, int mi_row,
1481 int mi_col) {
1482 if ((mi_row + mi_size_high[bsize] >= mi_rows) &&
1483 (mi_col + mi_size_wide[bsize] >= mi_cols)) {
1484 int64_t total_time = 0l;
1485 int32_t total_blocks = 0;
1486 for (BLOCK_SIZE bs = 0; bs < BLOCK_SIZES; bs++) {
1487 total_time += ms_stat->total_block_times[bs];
1488 total_blocks += ms_stat->num_blocks[bs];
1489 }
1490
1491 printf("\n");
1492 for (BLOCK_SIZE bs = 0; bs < BLOCK_SIZES; bs++) {
1493 if (ms_stat->num_blocks[bs] == 0) {
1494 continue;
1495 }
1496 if (!COLLECT_NON_SQR_STAT && block_size_wide[bs] != block_size_high[bs]) {
1497 continue;
1498 }
1499
1500 printf("BLOCK_%dX%d Num %d, Time: %ld (%f%%), Avg_time %f:\n",
1501 block_size_wide[bs], block_size_high[bs], ms_stat->num_blocks[bs],
1502 ms_stat->total_block_times[bs],
1503 100 * ms_stat->total_block_times[bs] / (float)total_time,
1504 (float)ms_stat->total_block_times[bs] / ms_stat->num_blocks[bs]);
1505 for (int j = 0; j < MB_MODE_COUNT; j++) {
1506 if (ms_stat->nonskipped_search_times[bs][j] == 0) {
1507 continue;
1508 }
1509
1510 int64_t total_mode_time = ms_stat->nonskipped_search_times[bs][j];
1511 printf(" Mode %d, %d/%d tps %f\n", j,
1512 ms_stat->num_nonskipped_searches[bs][j],
1513 ms_stat->num_searches[bs][j],
1514 ms_stat->num_nonskipped_searches[bs][j] > 0
1515 ? (float)ms_stat->nonskipped_search_times[bs][j] /
1516 ms_stat->num_nonskipped_searches[bs][j]
1517 : 0l);
1518 if (j >= INTER_MODE_START) {
1519 total_mode_time = ms_stat->ms_time[bs][j] + ms_stat->ifs_time[bs][j] +
1520 ms_stat->model_rd_time[bs][j] +
1521 ms_stat->txfm_time[bs][j];
1522 print_stage_time("Motion Search Time", ms_stat->ms_time[bs][j],
1523 total_time);
1524 print_stage_time("Filter Search Time", ms_stat->ifs_time[bs][j],
1525 total_time);
1526 print_stage_time("Model RD Time", ms_stat->model_rd_time[bs][j],
1527 total_time);
1528 print_stage_time("Tranfm Search Time", ms_stat->txfm_time[bs][j],
1529 total_time);
1530 }
1531 print_stage_time("Total Mode Time", total_mode_time, total_time);
1532 }
1533 printf("\n");
1534 }
1535 printf("Total time = %ld. Total blocks = %d\n", total_time, total_blocks);
1536 }
1537 }
1538 #endif // COLLECT_NONRD_PICK_MODE_STAT
1539
should_prune_intra_modes_using_neighbors(const MACROBLOCKD * xd,bool enable_intra_mode_pruning_using_neighbors,PREDICTION_MODE this_mode,PREDICTION_MODE above_mode,PREDICTION_MODE left_mode)1540 static bool should_prune_intra_modes_using_neighbors(
1541 const MACROBLOCKD *xd, bool enable_intra_mode_pruning_using_neighbors,
1542 PREDICTION_MODE this_mode, PREDICTION_MODE above_mode,
1543 PREDICTION_MODE left_mode) {
1544 if (!enable_intra_mode_pruning_using_neighbors) return false;
1545
1546 // Avoid pruning of DC_PRED as it is the most probable mode to win as per the
1547 // statistics generated for nonrd intra mode evaluations.
1548 if (this_mode == DC_PRED) return false;
1549
1550 // Enable the pruning for current mode only if it is not the winner mode of
1551 // both the neighboring blocks (left/top).
1552 return xd->up_available && this_mode != above_mode && xd->left_available &&
1553 this_mode != left_mode;
1554 }
1555
av1_nonrd_pick_intra_mode(AV1_COMP * cpi,MACROBLOCK * x,RD_STATS * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)1556 void av1_nonrd_pick_intra_mode(AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *rd_cost,
1557 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1558 AV1_COMMON *const cm = &cpi->common;
1559 MACROBLOCKD *const xd = &x->e_mbd;
1560 MB_MODE_INFO *const mi = xd->mi[0];
1561 RD_STATS this_rdc, best_rdc;
1562 struct estimate_block_intra_args args;
1563 init_estimate_block_intra_args(&args, cpi, x);
1564 const TxfmSearchParams *txfm_params = &x->txfm_search_params;
1565 mi->tx_size =
1566 AOMMIN(max_txsize_lookup[bsize],
1567 tx_mode_to_biggest_tx_size[txfm_params->tx_mode_search_type]);
1568 assert(IMPLIES(xd->lossless[mi->segment_id], mi->tx_size == TX_4X4));
1569 const BLOCK_SIZE tx_bsize = txsize_to_bsize[mi->tx_size];
1570
1571 // If the current block size is the same as the transform block size, enable
1572 // mode pruning based on the best SAD so far.
1573 if (cpi->sf.rt_sf.prune_intra_mode_using_best_sad_so_far && bsize == tx_bsize)
1574 args.prune_mode_based_on_sad = true;
1575
1576 int *bmode_costs;
1577 PREDICTION_MODE best_mode = DC_PRED;
1578 const MB_MODE_INFO *above_mi = xd->above_mbmi;
1579 const MB_MODE_INFO *left_mi = xd->left_mbmi;
1580 const PREDICTION_MODE A = av1_above_block_mode(above_mi);
1581 const PREDICTION_MODE L = av1_left_block_mode(left_mi);
1582 const int above_ctx = intra_mode_context[A];
1583 const int left_ctx = intra_mode_context[L];
1584 const unsigned int source_variance = x->source_variance;
1585 bmode_costs = x->mode_costs.y_mode_costs[above_ctx][left_ctx];
1586
1587 av1_invalid_rd_stats(&best_rdc);
1588 av1_invalid_rd_stats(&this_rdc);
1589
1590 init_mbmi_nonrd(mi, DC_PRED, INTRA_FRAME, NONE_FRAME, cm);
1591 mi->mv[0].as_int = mi->mv[1].as_int = INVALID_MV;
1592
1593 // Change the limit of this loop to add other intra prediction
1594 // mode tests.
1595 for (int mode_index = 0; mode_index < RTC_INTRA_MODES; ++mode_index) {
1596 PREDICTION_MODE this_mode = intra_mode_list[mode_index];
1597
1598 // As per the statistics generated for intra mode evaluation in the nonrd
1599 // path, it is found that the probability of H_PRED mode being the winner is
1600 // very low when the best mode so far is V_PRED (out of DC_PRED and V_PRED).
1601 // If V_PRED is the winner mode out of DC_PRED and V_PRED, it could imply
1602 // the presence of a vertically dominant pattern. Hence, H_PRED mode is not
1603 // evaluated.
1604 if (cpi->sf.rt_sf.prune_h_pred_using_best_mode_so_far &&
1605 this_mode == H_PRED && best_mode == V_PRED)
1606 continue;
1607
1608 if (should_prune_intra_modes_using_neighbors(
1609 xd, cpi->sf.rt_sf.enable_intra_mode_pruning_using_neighbors,
1610 this_mode, A, L)) {
1611 // Prune V_PRED and H_PRED if source variance of the block is less than
1612 // or equal to 50. The source variance threshold is obtained empirically.
1613 if ((this_mode == V_PRED || this_mode == H_PRED) && source_variance <= 50)
1614 continue;
1615
1616 // As per the statistics, probability of SMOOTH_PRED being the winner is
1617 // low when best mode so far is DC_PRED (out of DC_PRED, V_PRED and
1618 // H_PRED). Hence, SMOOTH_PRED mode is not evaluated.
1619 if (best_mode == DC_PRED && this_mode == SMOOTH_PRED) continue;
1620 }
1621
1622 this_rdc.dist = this_rdc.rate = 0;
1623 args.mode = this_mode;
1624 args.skippable = 1;
1625 args.rdc = &this_rdc;
1626 mi->mode = this_mode;
1627 av1_foreach_transformed_block_in_plane(xd, bsize, AOM_PLANE_Y,
1628 av1_estimate_block_intra, &args);
1629
1630 if (this_rdc.rate == INT_MAX) continue;
1631
1632 const int skip_ctx = av1_get_skip_txfm_context(xd);
1633 if (args.skippable) {
1634 this_rdc.rate = x->mode_costs.skip_txfm_cost[skip_ctx][1];
1635 } else {
1636 this_rdc.rate += x->mode_costs.skip_txfm_cost[skip_ctx][0];
1637 }
1638 this_rdc.rate += bmode_costs[this_mode];
1639 this_rdc.rdcost = RDCOST(x->rdmult, this_rdc.rate, this_rdc.dist);
1640
1641 if (this_rdc.rdcost < best_rdc.rdcost) {
1642 best_rdc = this_rdc;
1643 best_mode = this_mode;
1644 if (!this_rdc.skip_txfm) {
1645 memset(ctx->blk_skip, 0,
1646 sizeof(x->txfm_search_info.blk_skip[0]) * ctx->num_4x4_blk);
1647 }
1648 }
1649 }
1650
1651 mi->mode = best_mode;
1652 // Keep DC for UV since mode test is based on Y channel only.
1653 mi->uv_mode = UV_DC_PRED;
1654 *rd_cost = best_rdc;
1655
1656 // For lossless: always force the skip flags off.
1657 // Even though the blk_skip is set to 0 above in the rdcost comparison,
1658 // do it here again in case the above logic changes.
1659 if (is_lossless_requested(&cpi->oxcf.rc_cfg)) {
1660 x->txfm_search_info.skip_txfm = 0;
1661 memset(ctx->blk_skip, 0,
1662 sizeof(x->txfm_search_info.blk_skip[0]) * ctx->num_4x4_blk);
1663 }
1664
1665 #if CONFIG_INTERNAL_STATS
1666 store_coding_context_nonrd(x, ctx, mi->mode);
1667 #else
1668 store_coding_context_nonrd(x, ctx);
1669 #endif // CONFIG_INTERNAL_STATS
1670 }
1671
is_same_gf_and_last_scale(AV1_COMMON * cm)1672 static AOM_INLINE int is_same_gf_and_last_scale(AV1_COMMON *cm) {
1673 struct scale_factors *const sf_last = get_ref_scale_factors(cm, LAST_FRAME);
1674 struct scale_factors *const sf_golden =
1675 get_ref_scale_factors(cm, GOLDEN_FRAME);
1676 return ((sf_last->x_scale_fp == sf_golden->x_scale_fp) &&
1677 (sf_last->y_scale_fp == sf_golden->y_scale_fp));
1678 }
1679
get_ref_frame_use_mask(AV1_COMP * cpi,MACROBLOCK * x,MB_MODE_INFO * mi,int mi_row,int mi_col,BLOCK_SIZE bsize,int gf_temporal_ref,int use_ref_frame[],int * force_skip_low_temp_var)1680 static AOM_INLINE void get_ref_frame_use_mask(AV1_COMP *cpi, MACROBLOCK *x,
1681 MB_MODE_INFO *mi, int mi_row,
1682 int mi_col, BLOCK_SIZE bsize,
1683 int gf_temporal_ref,
1684 int use_ref_frame[],
1685 int *force_skip_low_temp_var) {
1686 AV1_COMMON *const cm = &cpi->common;
1687 const struct segmentation *const seg = &cm->seg;
1688 const int is_small_sb = (cm->seq_params->sb_size == BLOCK_64X64);
1689
1690 // When the ref_frame_config is used to set the reference frame structure
1691 // then the usage of alt_ref is determined by the ref_frame_flags
1692 // (and not the speed feature use_nonrd_altref_frame).
1693 int use_alt_ref_frame = cpi->ppi->rtc_ref.set_ref_frame_config ||
1694 cpi->sf.rt_sf.use_nonrd_altref_frame;
1695
1696 int use_golden_ref_frame = 1;
1697 int use_last_ref_frame = 1;
1698
1699 // When the ref_frame_config is used to set the reference frame structure:
1700 // check if LAST is used as a reference. And only remove golden and altref
1701 // references below if last is used as a reference.
1702 if (cpi->ppi->rtc_ref.set_ref_frame_config)
1703 use_last_ref_frame =
1704 cpi->ref_frame_flags & AOM_LAST_FLAG ? use_last_ref_frame : 0;
1705
1706 // frame_since_golden is not used when user sets the referene structure.
1707 if (!cpi->ppi->rtc_ref.set_ref_frame_config && use_last_ref_frame &&
1708 cpi->rc.frames_since_golden == 0 && gf_temporal_ref) {
1709 use_golden_ref_frame = 0;
1710 }
1711
1712 if (use_last_ref_frame && cpi->sf.rt_sf.short_circuit_low_temp_var &&
1713 x->nonrd_prune_ref_frame_search) {
1714 if (is_small_sb)
1715 *force_skip_low_temp_var = av1_get_force_skip_low_temp_var_small_sb(
1716 &x->part_search_info.variance_low[0], mi_row, mi_col, bsize);
1717 else
1718 *force_skip_low_temp_var = av1_get_force_skip_low_temp_var(
1719 &x->part_search_info.variance_low[0], mi_row, mi_col, bsize);
1720 // If force_skip_low_temp_var is set, skip golden reference.
1721 if (*force_skip_low_temp_var) {
1722 use_golden_ref_frame = 0;
1723 use_alt_ref_frame = 0;
1724 }
1725 }
1726
1727 if (use_last_ref_frame &&
1728 (x->nonrd_prune_ref_frame_search > 2 || x->force_zeromv_skip_for_blk ||
1729 (x->nonrd_prune_ref_frame_search > 1 && bsize > BLOCK_64X64))) {
1730 use_golden_ref_frame = 0;
1731 use_alt_ref_frame = 0;
1732 }
1733
1734 if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
1735 get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) == GOLDEN_FRAME) {
1736 use_golden_ref_frame = 1;
1737 use_alt_ref_frame = 0;
1738 }
1739
1740 // Skip golden/altref reference if color is set, on flat blocks with motion.
1741 // For screen: always skip golden/alt (if color_sensitivity_sb_g/alt is set)
1742 // except when x->nonrd_prune_ref_frame_search = 0. This latter flag
1743 // may be set in the variance partition when golden is a much better
1744 // reference than last, in which case it may not be worth skipping
1745 // golden/altref completely.
1746 // Condition on use_last_ref to make sure there remains at least one
1747 // reference.
1748 if (use_last_ref_frame &&
1749 ((cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
1750 x->nonrd_prune_ref_frame_search != 0) ||
1751 (x->source_variance < 200 &&
1752 x->content_state_sb.source_sad_nonrd >= kLowSad))) {
1753 if (x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_U)] == 1 ||
1754 x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_V)] == 1)
1755 use_golden_ref_frame = 0;
1756 if (x->color_sensitivity_sb_alt[COLOR_SENS_IDX(AOM_PLANE_U)] == 1 ||
1757 x->color_sensitivity_sb_alt[COLOR_SENS_IDX(AOM_PLANE_V)] == 1)
1758 use_alt_ref_frame = 0;
1759 }
1760
1761 // For non-screen: if golden and altref are not being selected as references
1762 // (use_golden_ref_frame/use_alt_ref_frame = 0) check to allow golden back
1763 // based on the sad of nearest/nearmv of LAST ref. If this block sad is large,
1764 // keep golden as reference. Only do this for the agrressive pruning mode and
1765 // avoid it when color is set for golden reference.
1766 if (cpi->oxcf.tune_cfg.content != AOM_CONTENT_SCREEN &&
1767 (cpi->ref_frame_flags & AOM_LAST_FLAG) && !use_golden_ref_frame &&
1768 !use_alt_ref_frame && x->pred_mv_sad[LAST_FRAME] != INT_MAX &&
1769 x->nonrd_prune_ref_frame_search > 2 &&
1770 x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_U)] == 0 &&
1771 x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_V)] == 0) {
1772 int thr = (cm->width * cm->height > RESOLUTION_288P) ? 100 : 150;
1773 int pred = x->pred_mv_sad[LAST_FRAME] >>
1774 (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
1775 if (pred > thr) use_golden_ref_frame = 1;
1776 }
1777
1778 use_alt_ref_frame =
1779 cpi->ref_frame_flags & AOM_ALT_FLAG ? use_alt_ref_frame : 0;
1780 use_golden_ref_frame =
1781 cpi->ref_frame_flags & AOM_GOLD_FLAG ? use_golden_ref_frame : 0;
1782
1783 // For spatial layers: enable golden ref if it is set by user and
1784 // corresponds to the lower spatial layer.
1785 if (cpi->svc.spatial_layer_id > 0 && (cpi->ref_frame_flags & AOM_GOLD_FLAG) &&
1786 x->content_state_sb.source_sad_nonrd < kHighSad) {
1787 const int buffslot_golden =
1788 cpi->ppi->rtc_ref.ref_idx[GOLDEN_FRAME - LAST_FRAME];
1789 if (cpi->ppi->rtc_ref.buffer_time_index[buffslot_golden] ==
1790 cpi->svc.current_superframe)
1791 use_golden_ref_frame = 1;
1792 }
1793
1794 use_ref_frame[ALTREF_FRAME] = use_alt_ref_frame;
1795 use_ref_frame[GOLDEN_FRAME] = use_golden_ref_frame;
1796 use_ref_frame[LAST_FRAME] = use_last_ref_frame;
1797 // Keep this assert on, as only 3 references are used in nonrd_pickmode
1798 // (LAST, GOLDEN, ALTREF), and if all 3 are not set by user then this
1799 // frame must be an intra-only frame and hence should never enter the
1800 // pickmode here for inter frames.
1801 assert(use_last_ref_frame || use_golden_ref_frame || use_alt_ref_frame);
1802 }
1803
is_filter_search_enabled_blk(AV1_COMP * cpi,MACROBLOCK * x,int mi_row,int mi_col,BLOCK_SIZE bsize,int segment_id,int cb_pred_filter_search,InterpFilter * filt_select)1804 static AOM_INLINE int is_filter_search_enabled_blk(
1805 AV1_COMP *cpi, MACROBLOCK *x, int mi_row, int mi_col, BLOCK_SIZE bsize,
1806 int segment_id, int cb_pred_filter_search, InterpFilter *filt_select) {
1807 const AV1_COMMON *const cm = &cpi->common;
1808 // filt search disabled
1809 if (!cpi->sf.rt_sf.use_nonrd_filter_search) return 0;
1810 // filt search purely based on mode properties
1811 if (!cb_pred_filter_search) return 1;
1812 MACROBLOCKD *const xd = &x->e_mbd;
1813 int enable_interp_search = 0;
1814 if (!(xd->left_mbmi && xd->above_mbmi)) {
1815 // neighbors info unavailable
1816 enable_interp_search = 2;
1817 } else if (!(is_inter_block(xd->left_mbmi) &&
1818 is_inter_block(xd->above_mbmi))) {
1819 // neighbor is INTRA
1820 enable_interp_search = 2;
1821 } else if (xd->left_mbmi->interp_filters.as_int !=
1822 xd->above_mbmi->interp_filters.as_int) {
1823 // filters are different
1824 enable_interp_search = 2;
1825 } else if ((cb_pred_filter_search == 1) &&
1826 (xd->left_mbmi->interp_filters.as_filters.x_filter !=
1827 EIGHTTAP_REGULAR)) {
1828 // not regular
1829 enable_interp_search = 2;
1830 } else {
1831 // enable prediction based on chessboard pattern
1832 if (xd->left_mbmi->interp_filters.as_filters.x_filter == EIGHTTAP_SMOOTH)
1833 *filt_select = EIGHTTAP_SMOOTH;
1834 const int bsl = mi_size_wide_log2[bsize];
1835 enable_interp_search =
1836 (bool)((((mi_row + mi_col) >> bsl) +
1837 get_chessboard_index(cm->current_frame.frame_number)) &
1838 0x1);
1839 if (cyclic_refresh_segment_id_boosted(segment_id)) enable_interp_search = 1;
1840 }
1841 return enable_interp_search;
1842 }
1843
skip_mode_by_threshold(PREDICTION_MODE mode,MV_REFERENCE_FRAME ref_frame,int_mv mv,int frames_since_golden,const int * const rd_threshes,const int * const rd_thresh_freq_fact,int64_t best_cost,int best_skip,int extra_shift)1844 static AOM_INLINE int skip_mode_by_threshold(
1845 PREDICTION_MODE mode, MV_REFERENCE_FRAME ref_frame, int_mv mv,
1846 int frames_since_golden, const int *const rd_threshes,
1847 const int *const rd_thresh_freq_fact, int64_t best_cost, int best_skip,
1848 int extra_shift) {
1849 int skip_this_mode = 0;
1850 const THR_MODES mode_index = mode_idx[ref_frame][INTER_OFFSET(mode)];
1851 int64_t mode_rd_thresh =
1852 best_skip ? ((int64_t)rd_threshes[mode_index]) << (extra_shift + 1)
1853 : ((int64_t)rd_threshes[mode_index]) << extra_shift;
1854
1855 // Increase mode_rd_thresh value for non-LAST for improved encoding
1856 // speed
1857 if (ref_frame != LAST_FRAME) {
1858 mode_rd_thresh = mode_rd_thresh << 1;
1859 if (ref_frame == GOLDEN_FRAME && frames_since_golden > 4)
1860 mode_rd_thresh = mode_rd_thresh << (extra_shift + 1);
1861 }
1862
1863 if (rd_less_than_thresh(best_cost, mode_rd_thresh,
1864 rd_thresh_freq_fact[mode_index]))
1865 if (mv.as_int != 0) skip_this_mode = 1;
1866
1867 return skip_this_mode;
1868 }
1869
skip_mode_by_low_temp(PREDICTION_MODE mode,MV_REFERENCE_FRAME ref_frame,BLOCK_SIZE bsize,CONTENT_STATE_SB content_state_sb,int_mv mv,int force_skip_low_temp_var)1870 static AOM_INLINE int skip_mode_by_low_temp(
1871 PREDICTION_MODE mode, MV_REFERENCE_FRAME ref_frame, BLOCK_SIZE bsize,
1872 CONTENT_STATE_SB content_state_sb, int_mv mv, int force_skip_low_temp_var) {
1873 // Skip non-zeromv mode search for non-LAST frame if force_skip_low_temp_var
1874 // is set. If nearestmv for golden frame is 0, zeromv mode will be skipped
1875 // later.
1876 if (force_skip_low_temp_var && ref_frame != LAST_FRAME && mv.as_int != 0) {
1877 return 1;
1878 }
1879
1880 if (content_state_sb.source_sad_nonrd != kHighSad && bsize >= BLOCK_64X64 &&
1881 force_skip_low_temp_var && mode == NEWMV) {
1882 return 1;
1883 }
1884 return 0;
1885 }
1886
skip_mode_by_bsize_and_ref_frame(PREDICTION_MODE mode,MV_REFERENCE_FRAME ref_frame,BLOCK_SIZE bsize,int extra_prune,unsigned int sse_zeromv_norm,int more_prune)1887 static AOM_INLINE int skip_mode_by_bsize_and_ref_frame(
1888 PREDICTION_MODE mode, MV_REFERENCE_FRAME ref_frame, BLOCK_SIZE bsize,
1889 int extra_prune, unsigned int sse_zeromv_norm, int more_prune) {
1890 const unsigned int thresh_skip_golden = 500;
1891
1892 if (ref_frame != LAST_FRAME && sse_zeromv_norm < thresh_skip_golden &&
1893 mode == NEWMV)
1894 return 1;
1895
1896 if (bsize == BLOCK_128X128 && mode == NEWMV) return 1;
1897
1898 // Skip testing non-LAST if this flag is set.
1899 if (extra_prune) {
1900 if (extra_prune > 1 && ref_frame != LAST_FRAME &&
1901 (bsize > BLOCK_16X16 && mode == NEWMV))
1902 return 1;
1903
1904 if (ref_frame != LAST_FRAME && mode == NEARMV) return 1;
1905
1906 if (more_prune && bsize >= BLOCK_32X32 && mode == NEARMV) return 1;
1907 }
1908 return 0;
1909 }
1910
set_block_source_sad(AV1_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,struct buf_2d * yv12_mb)1911 static void set_block_source_sad(AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
1912 struct buf_2d *yv12_mb) {
1913 struct macroblock_plane *const p = &x->plane[0];
1914 const int y_sad = cpi->ppi->fn_ptr[bsize].sdf(p->src.buf, p->src.stride,
1915 yv12_mb->buf, yv12_mb->stride);
1916 if (y_sad == 0) x->block_is_zero_sad = 1;
1917 }
1918
set_color_sensitivity(AV1_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int y_sad,unsigned int source_variance,struct buf_2d yv12_mb[MAX_MB_PLANE])1919 static void set_color_sensitivity(AV1_COMP *cpi, MACROBLOCK *x,
1920 BLOCK_SIZE bsize, int y_sad,
1921 unsigned int source_variance,
1922 struct buf_2d yv12_mb[MAX_MB_PLANE]) {
1923 const int subsampling_x = cpi->common.seq_params->subsampling_x;
1924 const int subsampling_y = cpi->common.seq_params->subsampling_y;
1925 const int source_sad_nonrd = x->content_state_sb.source_sad_nonrd;
1926 const int high_res = cpi->common.width * cpi->common.height >= 640 * 360;
1927 if (bsize == cpi->common.seq_params->sb_size) {
1928 // At superblock level color_sensitivity is already set to 0, 1, or 2.
1929 // 2 is middle/uncertain level. To avoid additional sad
1930 // computations when bsize = sb_size force level 2 to 1 (certain color)
1931 // for motion areas.
1932 if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] == 2) {
1933 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] =
1934 source_sad_nonrd >= kMedSad ? 1 : 0;
1935 }
1936 if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] == 2) {
1937 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] =
1938 source_sad_nonrd >= kMedSad ? 1 : 0;
1939 }
1940 return;
1941 }
1942 int shift = 3;
1943 unsigned int source_var_thr = 50;
1944 int uv_sad_thr = 100;
1945 if (source_sad_nonrd >= kMedSad && x->source_variance > 0 && high_res)
1946 shift = 4;
1947 if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN) {
1948 if (cpi->rc.high_source_sad) shift = 6;
1949 if (source_sad_nonrd > kMedSad) {
1950 source_var_thr = 1200;
1951 uv_sad_thr = 10;
1952 }
1953 }
1954 NOISE_LEVEL noise_level = kLow;
1955 int norm_sad =
1956 y_sad >> (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
1957 unsigned int thresh_spatial = (cpi->common.width > 1920) ? 5000 : 1000;
1958 // If the spatial source variance is high and the normalized y_sad
1959 // is low, then y-channel is likely good for mode estimation, so keep
1960 // color_sensitivity off. For low noise content for now, since there is
1961 // some bdrate regression for noisy color clip.
1962 if (cpi->noise_estimate.enabled)
1963 noise_level = av1_noise_estimate_extract_level(&cpi->noise_estimate);
1964 if (noise_level == kLow && source_variance > thresh_spatial &&
1965 cpi->oxcf.tune_cfg.content != AOM_CONTENT_SCREEN && norm_sad < 50) {
1966 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] = 0;
1967 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] = 0;
1968 return;
1969 }
1970 const int num_planes = av1_num_planes(&cpi->common);
1971
1972 for (int plane = AOM_PLANE_U; plane < num_planes; ++plane) {
1973 // Always check if level = 2. If level = 0 check again for
1974 // motion areas for higher resolns, where color artifacts
1975 // are more noticeable.
1976 if (x->color_sensitivity[COLOR_SENS_IDX(plane)] == 2 ||
1977 (x->color_sensitivity[COLOR_SENS_IDX(plane)] == 0 &&
1978 source_sad_nonrd >= kMedSad && high_res)) {
1979 struct macroblock_plane *const p = &x->plane[plane];
1980 const BLOCK_SIZE bs =
1981 get_plane_block_size(bsize, subsampling_x, subsampling_y);
1982
1983 const int uv_sad = cpi->ppi->fn_ptr[bs].sdf(
1984 p->src.buf, p->src.stride, yv12_mb[plane].buf, yv12_mb[plane].stride);
1985
1986 const int norm_uv_sad =
1987 uv_sad >> (b_width_log2_lookup[bs] + b_height_log2_lookup[bs]);
1988 x->color_sensitivity[COLOR_SENS_IDX(plane)] =
1989 uv_sad > (y_sad >> shift) && norm_uv_sad > 40;
1990 if (source_variance < source_var_thr && norm_uv_sad > uv_sad_thr)
1991 x->color_sensitivity[COLOR_SENS_IDX(plane)] = 1;
1992 }
1993 }
1994 }
1995
setup_compound_prediction(const AV1_COMMON * cm,MACROBLOCK * x,struct buf_2d yv12_mb[8][MAX_MB_PLANE],const int * use_ref_frame_mask,const MV_REFERENCE_FRAME * rf,int * ref_mv_idx)1996 static void setup_compound_prediction(const AV1_COMMON *cm, MACROBLOCK *x,
1997 struct buf_2d yv12_mb[8][MAX_MB_PLANE],
1998 const int *use_ref_frame_mask,
1999 const MV_REFERENCE_FRAME *rf,
2000 int *ref_mv_idx) {
2001 MACROBLOCKD *const xd = &x->e_mbd;
2002 MB_MODE_INFO *const mbmi = xd->mi[0];
2003 MB_MODE_INFO_EXT *const mbmi_ext = &x->mbmi_ext;
2004 MV_REFERENCE_FRAME ref_frame_comp;
2005 if (!use_ref_frame_mask[rf[1]]) {
2006 // Need to setup pred_block, if it hasn't been done in find_predictors.
2007 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_yv12_buf(cm, rf[1]);
2008 const int num_planes = av1_num_planes(cm);
2009 if (yv12 != NULL) {
2010 const struct scale_factors *const sf =
2011 get_ref_scale_factors_const(cm, rf[1]);
2012 av1_setup_pred_block(xd, yv12_mb[rf[1]], yv12, sf, sf, num_planes);
2013 }
2014 }
2015 ref_frame_comp = av1_ref_frame_type(rf);
2016 mbmi_ext->mode_context[ref_frame_comp] = 0;
2017 mbmi_ext->ref_mv_count[ref_frame_comp] = UINT8_MAX;
2018 av1_find_mv_refs(cm, xd, mbmi, ref_frame_comp, mbmi_ext->ref_mv_count,
2019 xd->ref_mv_stack, xd->weight, NULL, mbmi_ext->global_mvs,
2020 mbmi_ext->mode_context);
2021 av1_copy_usable_ref_mv_stack_and_weight(xd, mbmi_ext, ref_frame_comp);
2022 *ref_mv_idx = mbmi->ref_mv_idx + 1;
2023 }
2024
set_compound_mode(MACROBLOCK * x,MV_REFERENCE_FRAME ref_frame,MV_REFERENCE_FRAME ref_frame2,int ref_mv_idx,int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES],PREDICTION_MODE this_mode)2025 static void set_compound_mode(MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame,
2026 MV_REFERENCE_FRAME ref_frame2, int ref_mv_idx,
2027 int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES],
2028 PREDICTION_MODE this_mode) {
2029 MACROBLOCKD *const xd = &x->e_mbd;
2030 MB_MODE_INFO *const mi = xd->mi[0];
2031 mi->ref_frame[0] = ref_frame;
2032 mi->ref_frame[1] = ref_frame2;
2033 mi->compound_idx = 1;
2034 mi->comp_group_idx = 0;
2035 mi->interinter_comp.type = COMPOUND_AVERAGE;
2036 MV_REFERENCE_FRAME ref_frame_comp = av1_ref_frame_type(mi->ref_frame);
2037 if (this_mode == GLOBAL_GLOBALMV) {
2038 frame_mv[this_mode][ref_frame].as_int = 0;
2039 frame_mv[this_mode][ref_frame2].as_int = 0;
2040 } else if (this_mode == NEAREST_NEARESTMV) {
2041 frame_mv[this_mode][ref_frame].as_int =
2042 xd->ref_mv_stack[ref_frame_comp][0].this_mv.as_int;
2043 frame_mv[this_mode][ref_frame2].as_int =
2044 xd->ref_mv_stack[ref_frame_comp][0].comp_mv.as_int;
2045 } else if (this_mode == NEAR_NEARMV) {
2046 frame_mv[this_mode][ref_frame].as_int =
2047 xd->ref_mv_stack[ref_frame_comp][ref_mv_idx].this_mv.as_int;
2048 frame_mv[this_mode][ref_frame2].as_int =
2049 xd->ref_mv_stack[ref_frame_comp][ref_mv_idx].comp_mv.as_int;
2050 }
2051 }
2052
2053 // Prune compound mode if the single mode variance is lower than a fixed
2054 // percentage of the median value.
skip_comp_based_on_var(const unsigned int (* single_vars)[REF_FRAMES],BLOCK_SIZE bsize)2055 static bool skip_comp_based_on_var(
2056 const unsigned int (*single_vars)[REF_FRAMES], BLOCK_SIZE bsize) {
2057 unsigned int best_var = UINT_MAX;
2058 for (int cur_mode_idx = 0; cur_mode_idx < RTC_INTER_MODES; cur_mode_idx++) {
2059 for (int ref_idx = 0; ref_idx < REF_FRAMES; ref_idx++) {
2060 best_var = AOMMIN(best_var, single_vars[cur_mode_idx][ref_idx]);
2061 }
2062 }
2063 const unsigned int thresh_64 = (unsigned int)(0.57356805f * 8659);
2064 const unsigned int thresh_32 = (unsigned int)(0.23964763f * 4281);
2065
2066 // Currently, the thresh for 128 and 16 are not well-tuned. We are using the
2067 // results from 64 and 32 as an heuristic.
2068 switch (bsize) {
2069 case BLOCK_128X128: return best_var < 4 * thresh_64;
2070 case BLOCK_64X64: return best_var < thresh_64;
2071 case BLOCK_32X32: return best_var < thresh_32;
2072 case BLOCK_16X16: return best_var < thresh_32 / 4;
2073 default: return false;
2074 }
2075 }
2076
fill_single_inter_mode_costs(int (* single_inter_mode_costs)[REF_FRAMES],int num_inter_modes,const REF_MODE * reference_mode_set,const ModeCosts * mode_costs,const int16_t * mode_context)2077 static AOM_FORCE_INLINE void fill_single_inter_mode_costs(
2078 int (*single_inter_mode_costs)[REF_FRAMES], int num_inter_modes,
2079 const REF_MODE *reference_mode_set, const ModeCosts *mode_costs,
2080 const int16_t *mode_context) {
2081 bool ref_frame_used[REF_FRAMES] = { false };
2082 for (int idx = 0; idx < num_inter_modes; idx++) {
2083 ref_frame_used[reference_mode_set[idx].ref_frame] = true;
2084 }
2085
2086 for (int this_ref_frame = LAST_FRAME; this_ref_frame < REF_FRAMES;
2087 this_ref_frame++) {
2088 if (!ref_frame_used[this_ref_frame]) {
2089 continue;
2090 }
2091
2092 const MV_REFERENCE_FRAME rf[2] = { this_ref_frame, NONE_FRAME };
2093 const int16_t mode_ctx = av1_mode_context_analyzer(mode_context, rf);
2094 for (PREDICTION_MODE this_mode = NEARESTMV; this_mode <= NEWMV;
2095 this_mode++) {
2096 single_inter_mode_costs[INTER_OFFSET(this_mode)][this_ref_frame] =
2097 cost_mv_ref(mode_costs, this_mode, mode_ctx);
2098 }
2099 }
2100 }
2101
is_globalmv_better(PREDICTION_MODE this_mode,MV_REFERENCE_FRAME ref_frame,int rate_mv,const ModeCosts * mode_costs,const int (* single_inter_mode_costs)[REF_FRAMES],const MB_MODE_INFO_EXT * mbmi_ext)2102 static AOM_INLINE bool is_globalmv_better(
2103 PREDICTION_MODE this_mode, MV_REFERENCE_FRAME ref_frame, int rate_mv,
2104 const ModeCosts *mode_costs,
2105 const int (*single_inter_mode_costs)[REF_FRAMES],
2106 const MB_MODE_INFO_EXT *mbmi_ext) {
2107 const int globalmv_mode_cost =
2108 single_inter_mode_costs[INTER_OFFSET(GLOBALMV)][ref_frame];
2109 int this_mode_cost =
2110 rate_mv + single_inter_mode_costs[INTER_OFFSET(this_mode)][ref_frame];
2111 if (this_mode == NEWMV || this_mode == NEARMV) {
2112 const MV_REFERENCE_FRAME rf[2] = { ref_frame, NONE_FRAME };
2113 this_mode_cost += get_drl_cost(
2114 NEWMV, 0, mbmi_ext, mode_costs->drl_mode_cost0, av1_ref_frame_type(rf));
2115 }
2116 return this_mode_cost > globalmv_mode_cost;
2117 }
2118
2119 // Set up the mv/ref_frames etc based on the comp_index. Returns 1 if it
2120 // succeeds, 0 if it fails.
setup_compound_params_from_comp_idx(const AV1_COMP * cpi,MACROBLOCK * x,struct buf_2d yv12_mb[8][MAX_MB_PLANE],PREDICTION_MODE * this_mode,MV_REFERENCE_FRAME * ref_frame,MV_REFERENCE_FRAME * ref_frame2,int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES],const int * use_ref_frame_mask,int comp_index,bool comp_use_zero_zeromv_only,MV_REFERENCE_FRAME * last_comp_ref_frame,BLOCK_SIZE bsize)2121 static AOM_INLINE int setup_compound_params_from_comp_idx(
2122 const AV1_COMP *cpi, MACROBLOCK *x, struct buf_2d yv12_mb[8][MAX_MB_PLANE],
2123 PREDICTION_MODE *this_mode, MV_REFERENCE_FRAME *ref_frame,
2124 MV_REFERENCE_FRAME *ref_frame2, int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES],
2125 const int *use_ref_frame_mask, int comp_index,
2126 bool comp_use_zero_zeromv_only, MV_REFERENCE_FRAME *last_comp_ref_frame,
2127 BLOCK_SIZE bsize) {
2128 const MV_REFERENCE_FRAME *rf = comp_ref_mode_set[comp_index].ref_frame;
2129 int skip_gf = 0;
2130 int skip_alt = 0;
2131 *this_mode = comp_ref_mode_set[comp_index].pred_mode;
2132 *ref_frame = rf[0];
2133 *ref_frame2 = rf[1];
2134 assert(*ref_frame == LAST_FRAME);
2135 assert(*this_mode == GLOBAL_GLOBALMV || *this_mode == NEAREST_NEARESTMV);
2136 if (x->source_variance < 50 && bsize > BLOCK_16X16) {
2137 if (x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_U)] == 1 ||
2138 x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_V)] == 1)
2139 skip_gf = 1;
2140 if (x->color_sensitivity_sb_alt[COLOR_SENS_IDX(AOM_PLANE_U)] == 1 ||
2141 x->color_sensitivity_sb_alt[COLOR_SENS_IDX(AOM_PLANE_V)] == 1)
2142 skip_alt = 1;
2143 }
2144 if (comp_use_zero_zeromv_only && *this_mode != GLOBAL_GLOBALMV) {
2145 return 0;
2146 }
2147 if (*ref_frame2 == GOLDEN_FRAME &&
2148 (cpi->sf.rt_sf.ref_frame_comp_nonrd[0] == 0 || skip_gf ||
2149 !(cpi->ref_frame_flags & AOM_GOLD_FLAG))) {
2150 return 0;
2151 } else if (*ref_frame2 == LAST2_FRAME &&
2152 (cpi->sf.rt_sf.ref_frame_comp_nonrd[1] == 0 ||
2153 !(cpi->ref_frame_flags & AOM_LAST2_FLAG))) {
2154 return 0;
2155 } else if (*ref_frame2 == ALTREF_FRAME &&
2156 (cpi->sf.rt_sf.ref_frame_comp_nonrd[2] == 0 || skip_alt ||
2157 !(cpi->ref_frame_flags & AOM_ALT_FLAG))) {
2158 return 0;
2159 }
2160 int ref_mv_idx = 0;
2161 if (*last_comp_ref_frame != rf[1]) {
2162 // Only needs to be done once per reference pair.
2163 setup_compound_prediction(&cpi->common, x, yv12_mb, use_ref_frame_mask, rf,
2164 &ref_mv_idx);
2165 *last_comp_ref_frame = rf[1];
2166 }
2167 set_compound_mode(x, *ref_frame, *ref_frame2, ref_mv_idx, frame_mv,
2168 *this_mode);
2169 if (*this_mode != GLOBAL_GLOBALMV &&
2170 frame_mv[*this_mode][*ref_frame].as_int == 0 &&
2171 frame_mv[*this_mode][*ref_frame2].as_int == 0) {
2172 return 0;
2173 }
2174
2175 return 1;
2176 }
2177
previous_mode_performed_poorly(PREDICTION_MODE mode,MV_REFERENCE_FRAME ref_frame,const unsigned int (* vars)[REF_FRAMES],const int64_t (* uv_dist)[REF_FRAMES])2178 static AOM_INLINE bool previous_mode_performed_poorly(
2179 PREDICTION_MODE mode, MV_REFERENCE_FRAME ref_frame,
2180 const unsigned int (*vars)[REF_FRAMES],
2181 const int64_t (*uv_dist)[REF_FRAMES]) {
2182 unsigned int best_var = UINT_MAX;
2183 int64_t best_uv_dist = INT64_MAX;
2184 for (int midx = 0; midx < RTC_INTER_MODES; midx++) {
2185 best_var = AOMMIN(best_var, vars[midx][ref_frame]);
2186 best_uv_dist = AOMMIN(best_uv_dist, uv_dist[midx][ref_frame]);
2187 }
2188 assert(best_var != UINT_MAX && "Invalid variance data.");
2189 const float mult = 1.125f;
2190 bool var_bad = mult * best_var < vars[INTER_OFFSET(mode)][ref_frame];
2191 if (uv_dist[INTER_OFFSET(mode)][ref_frame] < INT64_MAX &&
2192 best_uv_dist != uv_dist[INTER_OFFSET(mode)][ref_frame]) {
2193 // If we have chroma info, then take it into account
2194 var_bad &= mult * best_uv_dist < uv_dist[INTER_OFFSET(mode)][ref_frame];
2195 }
2196 return var_bad;
2197 }
2198
prune_compoundmode_with_singlemode_var(PREDICTION_MODE compound_mode,MV_REFERENCE_FRAME ref_frame,MV_REFERENCE_FRAME ref_frame2,const int_mv (* frame_mv)[REF_FRAMES],const uint8_t (* mode_checked)[REF_FRAMES],const unsigned int (* vars)[REF_FRAMES],const int64_t (* uv_dist)[REF_FRAMES])2199 static AOM_INLINE bool prune_compoundmode_with_singlemode_var(
2200 PREDICTION_MODE compound_mode, MV_REFERENCE_FRAME ref_frame,
2201 MV_REFERENCE_FRAME ref_frame2, const int_mv (*frame_mv)[REF_FRAMES],
2202 const uint8_t (*mode_checked)[REF_FRAMES],
2203 const unsigned int (*vars)[REF_FRAMES],
2204 const int64_t (*uv_dist)[REF_FRAMES]) {
2205 const PREDICTION_MODE single_mode0 = compound_ref0_mode(compound_mode);
2206 const PREDICTION_MODE single_mode1 = compound_ref1_mode(compound_mode);
2207
2208 bool first_ref_valid = false, second_ref_valid = false;
2209 bool first_ref_bad = false, second_ref_bad = false;
2210 if (mode_checked[single_mode0][ref_frame] &&
2211 frame_mv[single_mode0][ref_frame].as_int ==
2212 frame_mv[compound_mode][ref_frame].as_int &&
2213 vars[INTER_OFFSET(single_mode0)][ref_frame] < UINT_MAX) {
2214 first_ref_valid = true;
2215 first_ref_bad =
2216 previous_mode_performed_poorly(single_mode0, ref_frame, vars, uv_dist);
2217 }
2218 if (mode_checked[single_mode1][ref_frame2] &&
2219 frame_mv[single_mode1][ref_frame2].as_int ==
2220 frame_mv[compound_mode][ref_frame2].as_int &&
2221 vars[INTER_OFFSET(single_mode1)][ref_frame2] < UINT_MAX) {
2222 second_ref_valid = true;
2223 second_ref_bad =
2224 previous_mode_performed_poorly(single_mode1, ref_frame2, vars, uv_dist);
2225 }
2226 if (first_ref_valid && second_ref_valid) {
2227 return first_ref_bad && second_ref_bad;
2228 } else if (first_ref_valid || second_ref_valid) {
2229 return first_ref_bad || second_ref_bad;
2230 }
2231 return false;
2232 }
2233
2234 // Function to setup parameters used for inter mode evaluation in non-rd.
set_params_nonrd_pick_inter_mode(AV1_COMP * cpi,MACROBLOCK * x,InterModeSearchStateNonrd * search_state,RD_STATS * rd_cost,int * force_skip_low_temp_var,int mi_row,int mi_col,int gf_temporal_ref,unsigned char segment_id,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx,int denoise_svc_pickmode)2235 static AOM_FORCE_INLINE void set_params_nonrd_pick_inter_mode(
2236 AV1_COMP *cpi, MACROBLOCK *x, InterModeSearchStateNonrd *search_state,
2237 RD_STATS *rd_cost, int *force_skip_low_temp_var, int mi_row, int mi_col,
2238 int gf_temporal_ref, unsigned char segment_id, BLOCK_SIZE bsize
2239 #if CONFIG_AV1_TEMPORAL_DENOISING
2240 ,
2241 PICK_MODE_CONTEXT *ctx, int denoise_svc_pickmode
2242 #endif
2243 ) {
2244 AV1_COMMON *const cm = &cpi->common;
2245 MACROBLOCKD *const xd = &x->e_mbd;
2246 TxfmSearchInfo *txfm_info = &x->txfm_search_info;
2247 MB_MODE_INFO *const mi = xd->mi[0];
2248 const ModeCosts *mode_costs = &x->mode_costs;
2249 int skip_pred_mv = 0;
2250
2251 // Initialize variance and distortion (chroma) for all modes and reference
2252 // frames
2253 for (int idx = 0; idx < RTC_INTER_MODES; idx++) {
2254 for (int ref = 0; ref < REF_FRAMES; ref++) {
2255 search_state->vars[idx][ref] = UINT_MAX;
2256 search_state->uv_dist[idx][ref] = INT64_MAX;
2257 }
2258 }
2259
2260 // Initialize values of color sensitivity with sb level color sensitivity
2261 av1_copy(x->color_sensitivity, x->color_sensitivity_sb);
2262
2263 init_best_pickmode(&search_state->best_pickmode);
2264
2265 // Estimate cost for single reference frames
2266 estimate_single_ref_frame_costs(cm, xd, mode_costs, segment_id, bsize,
2267 search_state->ref_costs_single);
2268
2269 // Reset flag to indicate modes evaluated
2270 av1_zero(search_state->mode_checked);
2271
2272 txfm_info->skip_txfm = 0;
2273
2274 // Initialize mode decisions
2275 av1_invalid_rd_stats(&search_state->best_rdc);
2276 av1_invalid_rd_stats(&search_state->this_rdc);
2277 av1_invalid_rd_stats(rd_cost);
2278 for (int ref_idx = 0; ref_idx < REF_FRAMES; ++ref_idx) {
2279 x->warp_sample_info[ref_idx].num = -1;
2280 }
2281
2282 mi->bsize = bsize;
2283 mi->ref_frame[0] = NONE_FRAME;
2284 mi->ref_frame[1] = NONE_FRAME;
2285
2286 #if CONFIG_AV1_TEMPORAL_DENOISING
2287 if (cpi->oxcf.noise_sensitivity > 0) {
2288 // if (cpi->ppi->use_svc) denoise_svc_pickmode =
2289 // av1_denoise_svc_non_key(cpi);
2290 if (cpi->denoiser.denoising_level > kDenLowLow && denoise_svc_pickmode)
2291 av1_denoiser_reset_frame_stats(ctx);
2292 }
2293 #endif
2294
2295 // Populate predicated motion vectors for LAST_FRAME
2296 if (cpi->ref_frame_flags & AOM_LAST_FLAG) {
2297 find_predictors(cpi, x, LAST_FRAME, search_state->frame_mv,
2298 search_state->yv12_mb, bsize, *force_skip_low_temp_var,
2299 x->force_zeromv_skip_for_blk,
2300 &search_state->use_scaled_ref_frame[LAST_FRAME]);
2301 }
2302 // Update mask to use all reference frame
2303 get_ref_frame_use_mask(cpi, x, mi, mi_row, mi_col, bsize, gf_temporal_ref,
2304 search_state->use_ref_frame_mask,
2305 force_skip_low_temp_var);
2306
2307 skip_pred_mv = x->force_zeromv_skip_for_blk ||
2308 (x->nonrd_prune_ref_frame_search > 2 &&
2309 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] != 2 &&
2310 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] != 2);
2311
2312 // Populate predicated motion vectors for other single reference frame
2313 // Start at LAST_FRAME + 1.
2314 for (MV_REFERENCE_FRAME ref_frame_iter = LAST_FRAME + 1;
2315 ref_frame_iter <= ALTREF_FRAME; ++ref_frame_iter) {
2316 if (search_state->use_ref_frame_mask[ref_frame_iter]) {
2317 find_predictors(cpi, x, ref_frame_iter, search_state->frame_mv,
2318 search_state->yv12_mb, bsize, *force_skip_low_temp_var,
2319 skip_pred_mv,
2320 &search_state->use_scaled_ref_frame[ref_frame_iter]);
2321 }
2322 }
2323 }
2324
2325 // Function to check the inter mode can be skipped based on mode statistics and
2326 // speed features settings.
skip_inter_mode_nonrd(AV1_COMP * cpi,MACROBLOCK * x,InterModeSearchStateNonrd * search_state,int64_t * thresh_sad_pred,int * force_mv_inter_layer,int * is_single_pred,PREDICTION_MODE * this_mode,MV_REFERENCE_FRAME * last_comp_ref_frame,MV_REFERENCE_FRAME * ref_frame,MV_REFERENCE_FRAME * ref_frame2,int idx,int_mv svc_mv,int force_skip_low_temp_var,unsigned int sse_zeromv_norm,int num_inter_modes,unsigned char segment_id,BLOCK_SIZE bsize,bool comp_use_zero_zeromv_only,bool check_globalmv)2327 static AOM_FORCE_INLINE bool skip_inter_mode_nonrd(
2328 AV1_COMP *cpi, MACROBLOCK *x, InterModeSearchStateNonrd *search_state,
2329 int64_t *thresh_sad_pred, int *force_mv_inter_layer, int *is_single_pred,
2330 PREDICTION_MODE *this_mode, MV_REFERENCE_FRAME *last_comp_ref_frame,
2331 MV_REFERENCE_FRAME *ref_frame, MV_REFERENCE_FRAME *ref_frame2, int idx,
2332 int_mv svc_mv, int force_skip_low_temp_var, unsigned int sse_zeromv_norm,
2333 int num_inter_modes, unsigned char segment_id, BLOCK_SIZE bsize,
2334 bool comp_use_zero_zeromv_only, bool check_globalmv) {
2335 AV1_COMMON *const cm = &cpi->common;
2336 const struct segmentation *const seg = &cm->seg;
2337 const SVC *const svc = &cpi->svc;
2338 MACROBLOCKD *const xd = &x->e_mbd;
2339 MB_MODE_INFO *const mi = xd->mi[0];
2340 const REAL_TIME_SPEED_FEATURES *const rt_sf = &cpi->sf.rt_sf;
2341
2342 // Skip compound mode based on reference frame mask and type of the mode and
2343 // for allowed compound modes, setup ref mv stack and reference frame.
2344 if (idx >= num_inter_modes) {
2345 const int comp_index = idx - num_inter_modes;
2346 if (!setup_compound_params_from_comp_idx(
2347 cpi, x, search_state->yv12_mb, this_mode, ref_frame, ref_frame2,
2348 search_state->frame_mv, search_state->use_ref_frame_mask,
2349 comp_index, comp_use_zero_zeromv_only, last_comp_ref_frame,
2350 bsize)) {
2351 return true;
2352 }
2353 *is_single_pred = 0;
2354 } else {
2355 *this_mode = ref_mode_set[idx].pred_mode;
2356 *ref_frame = ref_mode_set[idx].ref_frame;
2357 *ref_frame2 = NONE_FRAME;
2358 }
2359
2360 if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP) &&
2361 (*this_mode != GLOBALMV || *ref_frame != LAST_FRAME))
2362 return true;
2363
2364 if (x->sb_me_block && *ref_frame == LAST_FRAME) {
2365 // We want to make sure to test the superblock MV:
2366 // so don't skip (return false) for NEAREST_LAST or NEAR_LAST if they
2367 // have this sb MV. And don't skip NEWMV_LAST: this will be set to
2368 // sb MV in handle_inter_mode_nonrd(), in case NEAREST or NEAR don't
2369 // have it.
2370 if (*this_mode == NEARESTMV &&
2371 search_state->frame_mv[NEARESTMV][LAST_FRAME].as_int ==
2372 x->sb_me_mv.as_int) {
2373 return false;
2374 }
2375 if (*this_mode == NEARMV &&
2376 search_state->frame_mv[NEARMV][LAST_FRAME].as_int ==
2377 x->sb_me_mv.as_int) {
2378 return false;
2379 }
2380 if (*this_mode == NEWMV) {
2381 return false;
2382 }
2383 }
2384
2385 // Skip the single reference mode for which mode check flag is set.
2386 if (*is_single_pred && search_state->mode_checked[*this_mode][*ref_frame]) {
2387 return true;
2388 }
2389
2390 // Skip GLOBALMV mode if check_globalmv flag is not enabled.
2391 if (!check_globalmv && *this_mode == GLOBALMV) {
2392 return true;
2393 }
2394
2395 #if COLLECT_NONRD_PICK_MODE_STAT
2396 aom_usec_timer_start(&x->ms_stat_nonrd.timer1);
2397 x->ms_stat_nonrd.num_searches[bsize][*this_mode]++;
2398 #endif
2399 mi->mode = *this_mode;
2400 mi->ref_frame[0] = *ref_frame;
2401 mi->ref_frame[1] = *ref_frame2;
2402
2403 // Skip the mode if use reference frame mask flag is not set.
2404 if (!search_state->use_ref_frame_mask[*ref_frame]) return true;
2405
2406 // Skip mode for some modes and reference frames when
2407 // force_zeromv_skip_for_blk flag is true.
2408 if (x->force_zeromv_skip_for_blk &&
2409 ((!(*this_mode == NEARESTMV &&
2410 search_state->frame_mv[*this_mode][*ref_frame].as_int == 0) &&
2411 *this_mode != GLOBALMV) ||
2412 *ref_frame != LAST_FRAME))
2413 return true;
2414
2415 // Skip compound mode based on variance of previously evaluated single
2416 // reference modes.
2417 if (rt_sf->prune_compoundmode_with_singlemode_var && !*is_single_pred &&
2418 prune_compoundmode_with_singlemode_var(
2419 *this_mode, *ref_frame, *ref_frame2, search_state->frame_mv,
2420 search_state->mode_checked, search_state->vars,
2421 search_state->uv_dist)) {
2422 return true;
2423 }
2424
2425 *force_mv_inter_layer = 0;
2426 if (cpi->ppi->use_svc && svc->spatial_layer_id > 0 &&
2427 ((*ref_frame == LAST_FRAME && svc->skip_mvsearch_last) ||
2428 (*ref_frame == GOLDEN_FRAME && svc->skip_mvsearch_gf) ||
2429 (*ref_frame == ALTREF_FRAME && svc->skip_mvsearch_altref))) {
2430 // Only test mode if NEARESTMV/NEARMV is (svc_mv.mv.col, svc_mv.mv.row),
2431 // otherwise set NEWMV to (svc_mv.mv.col, svc_mv.mv.row).
2432 // Skip newmv and filter search.
2433 *force_mv_inter_layer = 1;
2434 if (*this_mode == NEWMV) {
2435 search_state->frame_mv[*this_mode][*ref_frame] = svc_mv;
2436 } else if (search_state->frame_mv[*this_mode][*ref_frame].as_int !=
2437 svc_mv.as_int) {
2438 return true;
2439 }
2440 }
2441
2442 // If the segment reference frame feature is enabled then do nothing if the
2443 // current ref frame is not allowed.
2444 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
2445 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)(*ref_frame))
2446 return true;
2447
2448 // For screen content: skip mode testing based on source_sad.
2449 if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN) {
2450 // If source_sad is computed: skip non-zero motion
2451 // check for stationary (super)blocks. Otherwise if superblock
2452 // has motion skip the modes with zero motion on last reference
2453 // for flat blocks, and color is not set.
2454 // For the latter condition: the same condition should apply
2455 // to newmv if (0, 0), so this latter condition is repeated
2456 // below after search_new_mv.
2457 if (rt_sf->source_metrics_sb_nonrd) {
2458 if ((search_state->frame_mv[*this_mode][*ref_frame].as_int != 0 &&
2459 x->content_state_sb.source_sad_nonrd == kZeroSad) ||
2460 (search_state->frame_mv[*this_mode][*ref_frame].as_int == 0 &&
2461 x->block_is_zero_sad == 0 && *ref_frame == LAST_FRAME &&
2462 ((x->color_sensitivity_sb[COLOR_SENS_IDX(AOM_PLANE_U)] == 0 &&
2463 x->color_sensitivity_sb[COLOR_SENS_IDX(AOM_PLANE_V)] == 0) ||
2464 cpi->rc.high_source_sad) &&
2465 x->source_variance == 0))
2466 return true;
2467 }
2468 // Skip NEWMV search for flat blocks.
2469 if (*this_mode == NEWMV && x->source_variance < 100) return true;
2470 // Skip non-LAST for color on flat blocks.
2471 if (*ref_frame > LAST_FRAME && x->source_variance == 0 &&
2472 (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] == 1 ||
2473 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] == 1))
2474 return true;
2475 }
2476
2477 // Skip mode based on block size, reference frame mode and other block
2478 // properties.
2479 if (skip_mode_by_bsize_and_ref_frame(
2480 *this_mode, *ref_frame, bsize, x->nonrd_prune_ref_frame_search,
2481 sse_zeromv_norm, rt_sf->nonrd_aggressive_skip))
2482 return true;
2483
2484 // Skip mode based on low temporal variance and souce sad.
2485 if (skip_mode_by_low_temp(*this_mode, *ref_frame, bsize, x->content_state_sb,
2486 search_state->frame_mv[*this_mode][*ref_frame],
2487 force_skip_low_temp_var))
2488 return true;
2489
2490 // Disable this drop out case if the ref frame segment level feature is
2491 // enabled for this segment. This is to prevent the possibility that we
2492 // end up unable to pick any mode.
2493 if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
2494 // Check for skipping GOLDEN and ALTREF based pred_mv_sad.
2495 if (rt_sf->nonrd_prune_ref_frame_search > 0 &&
2496 x->pred_mv_sad[*ref_frame] != INT_MAX && *ref_frame != LAST_FRAME) {
2497 if ((int64_t)(x->pred_mv_sad[*ref_frame]) > *thresh_sad_pred) return true;
2498 }
2499 }
2500
2501 // Check for skipping NEARMV based on pred_mv_sad.
2502 if (*this_mode == NEARMV && x->pred_mv1_sad[*ref_frame] != INT_MAX &&
2503 x->pred_mv1_sad[*ref_frame] > (x->pred_mv0_sad[*ref_frame] << 1))
2504 return true;
2505
2506 // Skip single reference mode based on rd threshold.
2507 if (*is_single_pred) {
2508 if (skip_mode_by_threshold(
2509 *this_mode, *ref_frame,
2510 search_state->frame_mv[*this_mode][*ref_frame],
2511 cpi->rc.frames_since_golden, cpi->rd.threshes[segment_id][bsize],
2512 x->thresh_freq_fact[bsize], search_state->best_rdc.rdcost,
2513 search_state->best_pickmode.best_mode_skip_txfm,
2514 (rt_sf->nonrd_aggressive_skip ? 1 : 0)))
2515 return true;
2516 }
2517 return false;
2518 }
2519
2520 // Function to perform inter mode evaluation for non-rd
handle_inter_mode_nonrd(AV1_COMP * cpi,MACROBLOCK * x,InterModeSearchStateNonrd * search_state,PICK_MODE_CONTEXT * ctx,PRED_BUFFER ** this_mode_pred,PRED_BUFFER * tmp_buffer,InterPredParams inter_pred_params_sr,int * best_early_term,unsigned int * sse_zeromv_norm,bool * check_globalmv,int64_t * zero_last_cost_orig,int denoise_svc_pickmode,int idx,int force_mv_inter_layer,int is_single_pred,int gf_temporal_ref,int use_model_yrd_large,int filter_search_enabled_blk,BLOCK_SIZE bsize,PREDICTION_MODE this_mode,InterpFilter filt_select,int cb_pred_filter_search,int reuse_inter_pred,int * sb_me_has_been_tested)2521 static AOM_FORCE_INLINE bool handle_inter_mode_nonrd(
2522 AV1_COMP *cpi, MACROBLOCK *x, InterModeSearchStateNonrd *search_state,
2523 PICK_MODE_CONTEXT *ctx, PRED_BUFFER **this_mode_pred,
2524 PRED_BUFFER *tmp_buffer, InterPredParams inter_pred_params_sr,
2525 int *best_early_term, unsigned int *sse_zeromv_norm, bool *check_globalmv,
2526 #if CONFIG_AV1_TEMPORAL_DENOISING
2527 int64_t *zero_last_cost_orig, int denoise_svc_pickmode,
2528 #endif
2529 int idx, int force_mv_inter_layer, int is_single_pred, int gf_temporal_ref,
2530 int use_model_yrd_large, int filter_search_enabled_blk, BLOCK_SIZE bsize,
2531 PREDICTION_MODE this_mode, InterpFilter filt_select,
2532 int cb_pred_filter_search, int reuse_inter_pred,
2533 int *sb_me_has_been_tested) {
2534 AV1_COMMON *const cm = &cpi->common;
2535 MACROBLOCKD *const xd = &x->e_mbd;
2536 MB_MODE_INFO *const mi = xd->mi[0];
2537 const MB_MODE_INFO_EXT *const mbmi_ext = &x->mbmi_ext;
2538 const int mi_row = xd->mi_row;
2539 const int mi_col = xd->mi_col;
2540 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
2541 const int bw = block_size_wide[bsize];
2542 const InterpFilter filter_ref = cm->features.interp_filter;
2543 const InterpFilter default_interp_filter = EIGHTTAP_REGULAR;
2544 TxfmSearchInfo *txfm_info = &x->txfm_search_info;
2545 const ModeCosts *mode_costs = &x->mode_costs;
2546 const REAL_TIME_SPEED_FEATURES *const rt_sf = &cpi->sf.rt_sf;
2547 BEST_PICKMODE *const best_pickmode = &search_state->best_pickmode;
2548
2549 MV_REFERENCE_FRAME ref_frame = mi->ref_frame[0];
2550 MV_REFERENCE_FRAME ref_frame2 = mi->ref_frame[1];
2551 int_mv *const this_mv = &search_state->frame_mv[this_mode][ref_frame];
2552 unsigned int var = UINT_MAX;
2553 int this_early_term = 0;
2554 int rate_mv = 0;
2555 int is_skippable;
2556 int skip_this_mv = 0;
2557 unsigned int var_threshold = UINT_MAX;
2558 PREDICTION_MODE this_best_mode;
2559 RD_STATS nonskip_rdc;
2560 av1_invalid_rd_stats(&nonskip_rdc);
2561
2562 if (x->sb_me_block && this_mode == NEWMV && ref_frame == LAST_FRAME) {
2563 // Set the NEWMV_LAST to the sb MV.
2564 search_state->frame_mv[NEWMV][LAST_FRAME].as_int = x->sb_me_mv.as_int;
2565 } else if (this_mode == NEWMV && !force_mv_inter_layer) {
2566 #if COLLECT_NONRD_PICK_MODE_STAT
2567 aom_usec_timer_start(&x->ms_stat_nonrd.timer2);
2568 #endif
2569 // Find the best motion vector for single/compound mode.
2570 const bool skip_newmv = search_new_mv(
2571 cpi, x, search_state->frame_mv, ref_frame, gf_temporal_ref, bsize,
2572 mi_row, mi_col, &rate_mv, &search_state->best_rdc);
2573 #if COLLECT_NONRD_PICK_MODE_STAT
2574 aom_usec_timer_mark(&x->ms_stat_nonrd.timer2);
2575 x->ms_stat_nonrd.ms_time[bsize][this_mode] +=
2576 aom_usec_timer_elapsed(&x->ms_stat_nonrd.timer2);
2577 #endif
2578 // Skip NEWMV mode,
2579 // (i). For bsize smaller than 16X16
2580 // (ii). Based on sad of the predicted mv w.r.t LAST_FRAME
2581 // (iii). When motion vector is same as that of reference mv
2582 if (skip_newmv) {
2583 return true;
2584 }
2585 }
2586
2587 // Check the current motion vector is same as that of previously evaluated
2588 // motion vectors.
2589 for (PREDICTION_MODE inter_mv_mode = NEARESTMV; inter_mv_mode <= NEWMV;
2590 inter_mv_mode++) {
2591 if (inter_mv_mode == this_mode) continue;
2592 if (is_single_pred &&
2593 search_state->mode_checked[inter_mv_mode][ref_frame] &&
2594 this_mv->as_int ==
2595 search_state->frame_mv[inter_mv_mode][ref_frame].as_int) {
2596 skip_this_mv = 1;
2597 break;
2598 }
2599 }
2600
2601 // Skip single mode if current motion vector is same that of previously
2602 // evaluated motion vectors.
2603 if (skip_this_mv && is_single_pred) return true;
2604
2605 // For screen: for spatially flat blocks with non-zero motion,
2606 // skip newmv if the motion vector is (0, 0)-LAST, and color is not set.
2607 if (this_mode == NEWMV && cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
2608 cpi->svc.spatial_layer_id == 0 && rt_sf->source_metrics_sb_nonrd) {
2609 if (this_mv->as_int == 0 && ref_frame == LAST_FRAME &&
2610 x->block_is_zero_sad == 0 &&
2611 ((x->color_sensitivity_sb[COLOR_SENS_IDX(AOM_PLANE_U)] == 0 &&
2612 x->color_sensitivity_sb[COLOR_SENS_IDX(AOM_PLANE_V)] == 0) ||
2613 cpi->rc.high_source_sad) &&
2614 x->source_variance == 0)
2615 return true;
2616 }
2617
2618 mi->mode = this_mode;
2619 mi->mv[0].as_int = this_mv->as_int;
2620 mi->mv[1].as_int = 0;
2621 if (!is_single_pred)
2622 mi->mv[1].as_int = search_state->frame_mv[this_mode][ref_frame2].as_int;
2623
2624 // Set buffers to store predicted samples for reuse
2625 if (reuse_inter_pred) {
2626 if (!*this_mode_pred) {
2627 *this_mode_pred = &tmp_buffer[3];
2628 } else {
2629 *this_mode_pred = &tmp_buffer[get_pred_buffer(tmp_buffer, 3)];
2630 pd->dst.buf = (*this_mode_pred)->data;
2631 pd->dst.stride = bw;
2632 }
2633 }
2634
2635 mi->motion_mode = SIMPLE_TRANSLATION;
2636 #if !CONFIG_REALTIME_ONLY
2637 if (cpi->oxcf.motion_mode_cfg.allow_warped_motion) {
2638 calc_num_proj_ref(cpi, x, mi);
2639 }
2640 #endif
2641 // set variance threshold for compound mode pruning
2642 if (rt_sf->prune_compoundmode_with_singlecompound_var && !is_single_pred &&
2643 use_model_yrd_large) {
2644 const PREDICTION_MODE single_mode0 = compound_ref0_mode(this_mode);
2645 const PREDICTION_MODE single_mode1 = compound_ref1_mode(this_mode);
2646 var_threshold =
2647 AOMMIN(var_threshold,
2648 search_state->vars[INTER_OFFSET(single_mode0)][ref_frame]);
2649 var_threshold =
2650 AOMMIN(var_threshold,
2651 search_state->vars[INTER_OFFSET(single_mode1)][ref_frame2]);
2652 }
2653
2654 // decide interpolation filter, build prediction signal, get sse
2655 const bool is_mv_subpel =
2656 (mi->mv[0].as_mv.row & 0x07) || (mi->mv[0].as_mv.col & 0x07);
2657 const bool enable_filt_search_this_mode =
2658 (filter_search_enabled_blk == 2)
2659 ? true
2660 : (filter_search_enabled_blk && !force_mv_inter_layer &&
2661 is_single_pred &&
2662 (ref_frame == LAST_FRAME || !x->nonrd_prune_ref_frame_search));
2663 if (is_mv_subpel && enable_filt_search_this_mode) {
2664 #if COLLECT_NONRD_PICK_MODE_STAT
2665 aom_usec_timer_start(&x->ms_stat_nonrd.timer2);
2666 #endif
2667 search_filter_ref(
2668 cpi, x, &search_state->this_rdc, &inter_pred_params_sr, mi_row, mi_col,
2669 tmp_buffer, bsize, reuse_inter_pred, this_mode_pred, &this_early_term,
2670 &var, use_model_yrd_large, best_pickmode->best_sse, is_single_pred);
2671 #if COLLECT_NONRD_PICK_MODE_STAT
2672 aom_usec_timer_mark(&x->ms_stat_nonrd.timer2);
2673 x->ms_stat_nonrd.ifs_time[bsize][this_mode] +=
2674 aom_usec_timer_elapsed(&x->ms_stat_nonrd.timer2);
2675 #endif
2676 #if !CONFIG_REALTIME_ONLY
2677 } else if (cpi->oxcf.motion_mode_cfg.allow_warped_motion &&
2678 this_mode == NEWMV) {
2679 // Find the best motion mode when current mode is NEWMV
2680 search_motion_mode(cpi, x, &search_state->this_rdc, mi_row, mi_col, bsize,
2681 &this_early_term, use_model_yrd_large, &rate_mv,
2682 best_pickmode->best_sse);
2683 if (this_mode == NEWMV) {
2684 this_mv[0] = mi->mv[0];
2685 }
2686 #endif
2687 } else {
2688 mi->interp_filters =
2689 (filter_ref == SWITCHABLE)
2690 ? av1_broadcast_interp_filter(default_interp_filter)
2691 : av1_broadcast_interp_filter(filter_ref);
2692 if (force_mv_inter_layer)
2693 mi->interp_filters = av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
2694
2695 // If it is sub-pel motion and cb_pred_filter_search is enabled, select
2696 // the pre-decided filter
2697 if (is_mv_subpel && cb_pred_filter_search)
2698 mi->interp_filters = av1_broadcast_interp_filter(filt_select);
2699
2700 #if COLLECT_NONRD_PICK_MODE_STAT
2701 aom_usec_timer_start(&x->ms_stat_nonrd.timer2);
2702 #endif
2703 if (is_single_pred) {
2704 SubpelParams subpel_params;
2705 // Initialize inter mode level params for single reference mode.
2706 init_inter_mode_params(&mi->mv[0].as_mv, &inter_pred_params_sr,
2707 &subpel_params, xd->block_ref_scale_factors[0],
2708 pd->pre->width, pd->pre->height);
2709 av1_enc_build_inter_predictor_y_nonrd(xd, &inter_pred_params_sr,
2710 &subpel_params);
2711 } else {
2712 av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
2713 AOM_PLANE_Y, AOM_PLANE_Y);
2714 }
2715
2716 if (use_model_yrd_large) {
2717 model_skip_for_sb_y_large(cpi, bsize, mi_row, mi_col, x, xd,
2718 &search_state->this_rdc, &this_early_term, 0,
2719 best_pickmode->best_sse, &var, var_threshold);
2720 } else {
2721 model_rd_for_sb_y(cpi, bsize, x, xd, &search_state->this_rdc, &var, 0,
2722 &this_early_term);
2723 }
2724 #if COLLECT_NONRD_PICK_MODE_STAT
2725 aom_usec_timer_mark(&x->ms_stat_nonrd.timer2);
2726 x->ms_stat_nonrd.model_rd_time[bsize][this_mode] +=
2727 aom_usec_timer_elapsed(&x->ms_stat_nonrd.timer2);
2728 #endif
2729 }
2730
2731 // update variance for single mode
2732 if (is_single_pred) {
2733 search_state->vars[INTER_OFFSET(this_mode)][ref_frame] = var;
2734 if (this_mv->as_int == 0) {
2735 search_state->vars[INTER_OFFSET(GLOBALMV)][ref_frame] = var;
2736 }
2737 }
2738 // prune compound mode based on single mode var threshold
2739 if (!is_single_pred && var > var_threshold) {
2740 if (reuse_inter_pred) free_pred_buffer(*this_mode_pred);
2741 return true;
2742 }
2743
2744 if (ref_frame == LAST_FRAME && this_mv->as_int == 0) {
2745 *sse_zeromv_norm = (unsigned int)(search_state->this_rdc.sse >>
2746 (b_width_log2_lookup[bsize] +
2747 b_height_log2_lookup[bsize]));
2748 }
2749
2750 // Perform early termination based on sse.
2751 if (rt_sf->sse_early_term_inter_search &&
2752 early_term_inter_search_with_sse(rt_sf->sse_early_term_inter_search,
2753 bsize, search_state->this_rdc.sse,
2754 best_pickmode->best_sse, this_mode)) {
2755 if (reuse_inter_pred) free_pred_buffer(*this_mode_pred);
2756 return true;
2757 }
2758
2759 #if COLLECT_NONRD_PICK_MODE_STAT
2760 x->ms_stat_nonrd.num_nonskipped_searches[bsize][this_mode]++;
2761 #endif
2762
2763 const int skip_ctx = av1_get_skip_txfm_context(xd);
2764 const int skip_txfm_cost = mode_costs->skip_txfm_cost[skip_ctx][1];
2765 const int no_skip_txfm_cost = mode_costs->skip_txfm_cost[skip_ctx][0];
2766 const int64_t sse_y = search_state->this_rdc.sse;
2767
2768 if (this_early_term) {
2769 search_state->this_rdc.skip_txfm = 1;
2770 search_state->this_rdc.rate = skip_txfm_cost;
2771 search_state->this_rdc.dist = search_state->this_rdc.sse << 4;
2772 } else {
2773 #if COLLECT_NONRD_PICK_MODE_STAT
2774 aom_usec_timer_start(&x->ms_stat_nonrd.timer2);
2775 #endif
2776 // Calculates RD Cost using Hadamard transform.
2777 av1_block_yrd(x, &search_state->this_rdc, &is_skippable, bsize,
2778 mi->tx_size);
2779 if (search_state->this_rdc.skip_txfm ||
2780 RDCOST(x->rdmult, search_state->this_rdc.rate,
2781 search_state->this_rdc.dist) >=
2782 RDCOST(x->rdmult, 0, search_state->this_rdc.sse)) {
2783 if (!search_state->this_rdc.skip_txfm) {
2784 // Need to store "real" rdc for possible future use if UV rdc
2785 // disallows tx skip
2786 nonskip_rdc = search_state->this_rdc;
2787 nonskip_rdc.rate += no_skip_txfm_cost;
2788 }
2789 search_state->this_rdc.rate = skip_txfm_cost;
2790 search_state->this_rdc.skip_txfm = 1;
2791 search_state->this_rdc.dist = search_state->this_rdc.sse;
2792 } else {
2793 search_state->this_rdc.rate += no_skip_txfm_cost;
2794 }
2795
2796 // Populate predicted sample for chroma planes based on color sensitivity.
2797 if ((x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] ||
2798 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)])) {
2799 RD_STATS rdc_uv;
2800 const BLOCK_SIZE uv_bsize =
2801 get_plane_block_size(bsize, xd->plane[AOM_PLANE_U].subsampling_x,
2802 xd->plane[AOM_PLANE_U].subsampling_y);
2803 if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)]) {
2804 av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
2805 AOM_PLANE_U, AOM_PLANE_U);
2806 }
2807 if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)]) {
2808 av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
2809 AOM_PLANE_V, AOM_PLANE_V);
2810 }
2811 // Compute sse for chroma planes.
2812 const int64_t sse_uv = av1_model_rd_for_sb_uv(
2813 cpi, uv_bsize, x, xd, &rdc_uv, AOM_PLANE_U, AOM_PLANE_V);
2814 if (rdc_uv.dist < x->min_dist_inter_uv)
2815 x->min_dist_inter_uv = rdc_uv.dist;
2816 search_state->this_rdc.sse += sse_uv;
2817 // Restore Y rdc if UV rdc disallows txfm skip
2818 if (search_state->this_rdc.skip_txfm && !rdc_uv.skip_txfm &&
2819 nonskip_rdc.rate != INT_MAX)
2820 search_state->this_rdc = nonskip_rdc;
2821 if (is_single_pred) {
2822 search_state->uv_dist[INTER_OFFSET(this_mode)][ref_frame] = rdc_uv.dist;
2823 }
2824 search_state->this_rdc.rate += rdc_uv.rate;
2825 search_state->this_rdc.dist += rdc_uv.dist;
2826 search_state->this_rdc.skip_txfm =
2827 search_state->this_rdc.skip_txfm && rdc_uv.skip_txfm;
2828 }
2829 #if COLLECT_NONRD_PICK_MODE_STAT
2830 aom_usec_timer_mark(&x->ms_stat_nonrd.timer2);
2831 x->ms_stat_nonrd.txfm_time[bsize][this_mode] +=
2832 aom_usec_timer_elapsed(&x->ms_stat_nonrd.timer2);
2833 #endif
2834 }
2835
2836 this_best_mode = this_mode;
2837 // TODO(kyslov) account for UV prediction cost
2838 search_state->this_rdc.rate += rate_mv;
2839 if (!is_single_pred) {
2840 const int16_t mode_ctx =
2841 av1_mode_context_analyzer(mbmi_ext->mode_context, mi->ref_frame);
2842 search_state->this_rdc.rate += cost_mv_ref(mode_costs, this_mode, mode_ctx);
2843 } else {
2844 // If the current mode has zeromv but is not GLOBALMV, compare the rate
2845 // cost. If GLOBALMV is cheaper, use GLOBALMV instead.
2846 if (this_mode != GLOBALMV &&
2847 this_mv->as_int == search_state->frame_mv[GLOBALMV][ref_frame].as_int) {
2848 if (is_globalmv_better(this_mode, ref_frame, rate_mv, mode_costs,
2849 search_state->single_inter_mode_costs, mbmi_ext)) {
2850 this_best_mode = GLOBALMV;
2851 }
2852 }
2853
2854 search_state->this_rdc.rate +=
2855 search_state
2856 ->single_inter_mode_costs[INTER_OFFSET(this_best_mode)][ref_frame];
2857 }
2858
2859 if (is_single_pred && this_mv->as_int == 0 && var < UINT_MAX) {
2860 search_state->vars[INTER_OFFSET(GLOBALMV)][ref_frame] = var;
2861 }
2862
2863 search_state->this_rdc.rate += search_state->ref_costs_single[ref_frame];
2864
2865 search_state->this_rdc.rdcost = RDCOST(x->rdmult, search_state->this_rdc.rate,
2866 search_state->this_rdc.dist);
2867 if (cpi->oxcf.rc_cfg.mode == AOM_CBR && is_single_pred) {
2868 newmv_diff_bias(xd, this_best_mode, &search_state->this_rdc, bsize,
2869 search_state->frame_mv[this_best_mode][ref_frame].as_mv.row,
2870 search_state->frame_mv[this_best_mode][ref_frame].as_mv.col,
2871 cpi->speed, x->source_variance, x->content_state_sb);
2872 }
2873
2874 #if CONFIG_AV1_TEMPORAL_DENOISING
2875 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc_pickmode &&
2876 cpi->denoiser.denoising_level > kDenLowLow) {
2877 av1_denoiser_update_frame_stats(mi, sse_y, this_mode, ctx);
2878 // Keep track of zero_last cost.
2879 if (ref_frame == LAST_FRAME && this_mv->as_int == 0)
2880 *zero_last_cost_orig = search_state->this_rdc.rdcost;
2881 }
2882 #else
2883 (void)(sse_y);
2884 #endif
2885
2886 search_state->mode_checked[this_mode][ref_frame] = 1;
2887 search_state->mode_checked[this_best_mode][ref_frame] = 1;
2888
2889 if (*check_globalmv) {
2890 int32_t abs_mv =
2891 abs(search_state->frame_mv[this_best_mode][ref_frame].as_mv.row) +
2892 abs(search_state->frame_mv[this_best_mode][ref_frame].as_mv.col);
2893 // Early exit check: if the magnitude of this_best_mode's mv is small
2894 // enough, we skip GLOBALMV check in the next loop iteration.
2895 if (abs_mv < 2) {
2896 *check_globalmv = false;
2897 }
2898 }
2899 #if COLLECT_NONRD_PICK_MODE_STAT
2900 aom_usec_timer_mark(&x->ms_stat_nonrd.timer1);
2901 x->ms_stat_nonrd.nonskipped_search_times[bsize][this_mode] +=
2902 aom_usec_timer_elapsed(&x->ms_stat_nonrd.timer1);
2903 #endif
2904
2905 if (x->sb_me_block && ref_frame == LAST_FRAME &&
2906 search_state->frame_mv[this_best_mode][ref_frame].as_int ==
2907 x->sb_me_mv.as_int)
2908 *sb_me_has_been_tested = 1;
2909
2910 // Copy best mode params to search state
2911 if (search_state->this_rdc.rdcost < search_state->best_rdc.rdcost) {
2912 search_state->best_rdc = search_state->this_rdc;
2913 *best_early_term = this_early_term;
2914 update_search_state_nonrd(search_state, mi, txfm_info, &nonskip_rdc, ctx,
2915 this_best_mode, sse_y);
2916
2917 // This is needed for the compound modes.
2918 search_state->frame_mv_best[this_best_mode][ref_frame].as_int =
2919 search_state->frame_mv[this_best_mode][ref_frame].as_int;
2920 if (ref_frame2 > NONE_FRAME) {
2921 search_state->frame_mv_best[this_best_mode][ref_frame2].as_int =
2922 search_state->frame_mv[this_best_mode][ref_frame2].as_int;
2923 }
2924
2925 if (reuse_inter_pred) {
2926 free_pred_buffer(best_pickmode->best_pred);
2927 best_pickmode->best_pred = *this_mode_pred;
2928 }
2929 } else {
2930 if (reuse_inter_pred) free_pred_buffer(*this_mode_pred);
2931 }
2932
2933 if (*best_early_term && (idx > 0 || rt_sf->nonrd_aggressive_skip)) {
2934 txfm_info->skip_txfm = 1;
2935 if (!x->sb_me_block || *sb_me_has_been_tested) return false;
2936 }
2937 return true;
2938 }
2939
2940 // Function to perform screen content mode evaluation for non-rd
handle_screen_content_mode_nonrd(AV1_COMP * cpi,MACROBLOCK * x,InterModeSearchStateNonrd * search_state,PRED_BUFFER * this_mode_pred,PICK_MODE_CONTEXT * ctx,PRED_BUFFER * tmp_buffer,struct buf_2d * orig_dst,int skip_idtx_palette,int try_palette,BLOCK_SIZE bsize,int reuse_inter_pred,int mi_col,int mi_row)2941 static AOM_FORCE_INLINE void handle_screen_content_mode_nonrd(
2942 AV1_COMP *cpi, MACROBLOCK *x, InterModeSearchStateNonrd *search_state,
2943 PRED_BUFFER *this_mode_pred, PICK_MODE_CONTEXT *ctx,
2944 PRED_BUFFER *tmp_buffer, struct buf_2d *orig_dst, int skip_idtx_palette,
2945 int try_palette, BLOCK_SIZE bsize, int reuse_inter_pred, int mi_col,
2946 int mi_row) {
2947 AV1_COMMON *const cm = &cpi->common;
2948 const REAL_TIME_SPEED_FEATURES *const rt_sf = &cpi->sf.rt_sf;
2949 MACROBLOCKD *const xd = &x->e_mbd;
2950 MB_MODE_INFO *const mi = xd->mi[0];
2951 struct macroblockd_plane *const pd = &xd->plane[0];
2952 const int bw = block_size_wide[bsize];
2953 const int bh = block_size_high[bsize];
2954 TxfmSearchInfo *txfm_info = &x->txfm_search_info;
2955 BEST_PICKMODE *const best_pickmode = &search_state->best_pickmode;
2956
2957 // TODO(marpan): Only allow for 8 bit-depth for now, re-enable for 10/12 bit
2958 // when issue 3359 is fixed.
2959 if (cm->seq_params->bit_depth == 8 &&
2960 cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN && !skip_idtx_palette &&
2961 !cpi->oxcf.txfm_cfg.use_inter_dct_only && !x->force_zeromv_skip_for_blk &&
2962 is_inter_mode(best_pickmode->best_mode) &&
2963 best_pickmode->best_pred != NULL &&
2964 (!rt_sf->prune_idtx_nonrd ||
2965 (rt_sf->prune_idtx_nonrd && bsize <= BLOCK_32X32 &&
2966 best_pickmode->best_mode_skip_txfm != 1 && x->source_variance > 200))) {
2967 RD_STATS idtx_rdc;
2968 av1_init_rd_stats(&idtx_rdc);
2969 int is_skippable;
2970 this_mode_pred = &tmp_buffer[get_pred_buffer(tmp_buffer, 3)];
2971 pd->dst.buf = this_mode_pred->data;
2972 pd->dst.stride = bw;
2973 const PRED_BUFFER *const best_pred = best_pickmode->best_pred;
2974 av1_block_yrd_idtx(x, best_pred->data, best_pred->stride, &idtx_rdc,
2975 &is_skippable, bsize, mi->tx_size);
2976 int64_t idx_rdcost_y = RDCOST(x->rdmult, idtx_rdc.rate, idtx_rdc.dist);
2977 int allow_idtx = 1;
2978 // Incorporate color into rd cost.
2979 if ((x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] ||
2980 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)])) {
2981 RD_STATS rdc_uv;
2982 const BLOCK_SIZE uv_bsize =
2983 get_plane_block_size(bsize, xd->plane[AOM_PLANE_U].subsampling_x,
2984 xd->plane[AOM_PLANE_U].subsampling_y);
2985 if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)]) {
2986 av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
2987 AOM_PLANE_U, AOM_PLANE_U);
2988 }
2989 if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)]) {
2990 av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
2991 AOM_PLANE_V, AOM_PLANE_V);
2992 }
2993 av1_model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &rdc_uv, AOM_PLANE_U,
2994 AOM_PLANE_V);
2995 if (rdc_uv.dist < x->min_dist_inter_uv)
2996 x->min_dist_inter_uv = rdc_uv.dist;
2997 idtx_rdc.rate += rdc_uv.rate;
2998 idtx_rdc.dist += rdc_uv.dist;
2999 idtx_rdc.skip_txfm = idtx_rdc.skip_txfm && rdc_uv.skip_txfm;
3000 if (idx_rdcost_y == 0 && rdc_uv.dist > 0 && x->source_variance < 3000 &&
3001 x->content_state_sb.source_sad_nonrd > kMedSad)
3002 allow_idtx = 0;
3003 }
3004 int64_t idx_rdcost = RDCOST(x->rdmult, idtx_rdc.rate, idtx_rdc.dist);
3005 if (allow_idtx && idx_rdcost < search_state->best_rdc.rdcost) {
3006 best_pickmode->tx_type = IDTX;
3007 search_state->best_rdc.rdcost = idx_rdcost;
3008 best_pickmode->best_mode_skip_txfm = idtx_rdc.skip_txfm;
3009 if (!idtx_rdc.skip_txfm) {
3010 memcpy(ctx->blk_skip, txfm_info->blk_skip,
3011 sizeof(txfm_info->blk_skip[0]) * ctx->num_4x4_blk);
3012 }
3013 xd->tx_type_map[0] = best_pickmode->tx_type;
3014 memset(ctx->tx_type_map, best_pickmode->tx_type, ctx->num_4x4_blk);
3015 memset(xd->tx_type_map, best_pickmode->tx_type, ctx->num_4x4_blk);
3016 }
3017 pd->dst = *orig_dst;
3018 }
3019
3020 if (!try_palette) return;
3021 const unsigned int intra_ref_frame_cost =
3022 search_state->ref_costs_single[INTRA_FRAME];
3023
3024 if (!is_mode_intra(best_pickmode->best_mode)) {
3025 PRED_BUFFER *const best_pred = best_pickmode->best_pred;
3026 if (reuse_inter_pred && best_pred != NULL) {
3027 if (best_pred->data == orig_dst->buf) {
3028 this_mode_pred = &tmp_buffer[get_pred_buffer(tmp_buffer, 3)];
3029 aom_convolve_copy(best_pred->data, best_pred->stride,
3030 this_mode_pred->data, this_mode_pred->stride, bw, bh);
3031 best_pickmode->best_pred = this_mode_pred;
3032 }
3033 }
3034 pd->dst = *orig_dst;
3035 }
3036 // Search palette mode for Luma plane in inter frame.
3037 av1_search_palette_mode_luma(cpi, x, bsize, intra_ref_frame_cost, ctx,
3038 &search_state->this_rdc,
3039 search_state->best_rdc.rdcost);
3040 // Update best mode data in search_state
3041 if (search_state->this_rdc.rdcost < search_state->best_rdc.rdcost) {
3042 best_pickmode->pmi = mi->palette_mode_info;
3043 best_pickmode->best_mode = DC_PRED;
3044 mi->mv[0].as_int = INVALID_MV;
3045 mi->mv[1].as_int = INVALID_MV;
3046 best_pickmode->best_ref_frame = INTRA_FRAME;
3047 best_pickmode->best_second_ref_frame = NONE;
3048 search_state->best_rdc.rate = search_state->this_rdc.rate;
3049 search_state->best_rdc.dist = search_state->this_rdc.dist;
3050 search_state->best_rdc.rdcost = search_state->this_rdc.rdcost;
3051 best_pickmode->best_mode_skip_txfm = search_state->this_rdc.skip_txfm;
3052 // Keep the skip_txfm off if the color_sensitivity is set.
3053 if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] ||
3054 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)])
3055 search_state->this_rdc.skip_txfm = 0;
3056 if (!search_state->this_rdc.skip_txfm) {
3057 memcpy(ctx->blk_skip, txfm_info->blk_skip,
3058 sizeof(txfm_info->blk_skip[0]) * ctx->num_4x4_blk);
3059 }
3060 if (xd->tx_type_map[0] != DCT_DCT)
3061 av1_copy_array(ctx->tx_type_map, xd->tx_type_map, ctx->num_4x4_blk);
3062 }
3063 }
3064
3065 /*!\brief AV1 inter mode selection based on Non-RD optimized model.
3066 *
3067 * \ingroup nonrd_mode_search
3068 * \callgraph
3069 * Top level function for Non-RD optimized inter mode selection.
3070 * This finction will loop over subset of inter modes and select the best one
3071 * based on calculated modelled RD cost. While making decisions which modes to
3072 * check, this function applies heuristics based on previously checked modes,
3073 * block residual variance, block size, and other factors to prune certain
3074 * modes and reference frames. Currently only single reference frame modes
3075 * are checked. Additional heuristics are applied to decide if intra modes
3076 * need to be checked.
3077 * *
3078 * \param[in] cpi Top-level encoder structure
3079 * \param[in] tile_data Pointer to struct holding adaptive
3080 data/contexts/models for the tile during
3081 encoding
3082 * \param[in] x Pointer to structure holding all the data for
3083 the current macroblock
3084 * \param[in] rd_cost Struct to keep track of the RD information
3085 * \param[in] bsize Current block size
3086 * \param[in] ctx Structure to hold snapshot of coding context
3087 during the mode picking process
3088 *
3089 * \remark Nothing is returned. Instead, the MB_MODE_INFO struct inside x
3090 * is modified to store information about the best mode computed
3091 * in this function. The rd_cost struct is also updated with the RD stats
3092 * corresponding to the best mode found.
3093 */
av1_nonrd_pick_inter_mode_sb(AV1_COMP * cpi,TileDataEnc * tile_data,MACROBLOCK * x,RD_STATS * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)3094 void av1_nonrd_pick_inter_mode_sb(AV1_COMP *cpi, TileDataEnc *tile_data,
3095 MACROBLOCK *x, RD_STATS *rd_cost,
3096 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
3097 AV1_COMMON *const cm = &cpi->common;
3098 SVC *const svc = &cpi->svc;
3099 MACROBLOCKD *const xd = &x->e_mbd;
3100 MB_MODE_INFO *const mi = xd->mi[0];
3101 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
3102 const MB_MODE_INFO_EXT *const mbmi_ext = &x->mbmi_ext;
3103 MV_REFERENCE_FRAME ref_frame, ref_frame2;
3104 const unsigned char segment_id = mi->segment_id;
3105 int best_early_term = 0;
3106 int force_skip_low_temp_var = 0;
3107 unsigned int sse_zeromv_norm = UINT_MAX;
3108 const int num_inter_modes = NUM_INTER_MODES;
3109 const REAL_TIME_SPEED_FEATURES *const rt_sf = &cpi->sf.rt_sf;
3110 bool check_globalmv = rt_sf->check_globalmv_on_single_ref;
3111 PRED_BUFFER tmp_buffer[4];
3112 DECLARE_ALIGNED(16, uint8_t, pred_buf[MAX_MB_PLANE * MAX_SB_SQUARE]);
3113 PRED_BUFFER *this_mode_pred = NULL;
3114 const int reuse_inter_pred =
3115 rt_sf->reuse_inter_pred_nonrd && cm->seq_params->bit_depth == AOM_BITS_8;
3116 InterModeSearchStateNonrd search_state;
3117 av1_zero(search_state.use_ref_frame_mask);
3118 av1_zero(search_state.use_scaled_ref_frame);
3119 BEST_PICKMODE *const best_pickmode = &search_state.best_pickmode;
3120 (void)tile_data;
3121
3122 const int bh = block_size_high[bsize];
3123 const int bw = block_size_wide[bsize];
3124 const int pixels_in_block = bh * bw;
3125 struct buf_2d orig_dst = pd->dst;
3126 const TxfmSearchParams *txfm_params = &x->txfm_search_params;
3127 TxfmSearchInfo *txfm_info = &x->txfm_search_info;
3128 #if COLLECT_NONRD_PICK_MODE_STAT
3129 // Mode statistics can be collected only when num_workers is 1
3130 assert(cpi->mt_info.num_workers <= 1);
3131 aom_usec_timer_start(&x->ms_stat_nonrd.bsize_timer);
3132 #endif
3133 int64_t thresh_sad_pred = INT64_MAX;
3134 const int mi_row = xd->mi_row;
3135 const int mi_col = xd->mi_col;
3136 int_mv svc_mv = { .as_int = 0 };
3137 int force_mv_inter_layer = 0;
3138 bool comp_use_zero_zeromv_only = 0;
3139 int tot_num_comp_modes = NUM_COMP_INTER_MODES_RT;
3140 #if CONFIG_AV1_TEMPORAL_DENOISING
3141 const int denoise_recheck_zeromv = 1;
3142 AV1_PICKMODE_CTX_DEN ctx_den;
3143 int64_t zero_last_cost_orig = INT64_MAX;
3144 int denoise_svc_pickmode = 1;
3145 const int resize_pending = is_frame_resize_pending(cpi);
3146 #endif
3147 const ModeCosts *mode_costs = &x->mode_costs;
3148 struct scale_factors sf_no_scale;
3149 av1_setup_scale_factors_for_frame(&sf_no_scale, cm->width, cm->height,
3150 cm->width, cm->height);
3151 if (reuse_inter_pred) {
3152 for (int buf_idx = 0; buf_idx < 3; buf_idx++) {
3153 tmp_buffer[buf_idx].data = &pred_buf[pixels_in_block * buf_idx];
3154 tmp_buffer[buf_idx].stride = bw;
3155 tmp_buffer[buf_idx].in_use = 0;
3156 }
3157 tmp_buffer[3].data = pd->dst.buf;
3158 tmp_buffer[3].stride = pd->dst.stride;
3159 tmp_buffer[3].in_use = 0;
3160 }
3161
3162 const int gf_temporal_ref = is_same_gf_and_last_scale(cm);
3163
3164 // If the lower spatial layer uses an averaging filter for downsampling
3165 // (phase = 8), the target decimated pixel is shifted by (1/2, 1/2) relative
3166 // to source, so use subpel motion vector to compensate. The nonzero motion
3167 // is half pixel shifted to left and top, so (-4, -4). This has more effect
3168 // on higher resolutions, so condition it on that for now.
3169 // Exclude quality layers, which have the same resolution and hence no shift.
3170 if (cpi->ppi->use_svc && svc->spatial_layer_id > 0 &&
3171 !svc->has_lower_quality_layer &&
3172 svc->downsample_filter_phase[svc->spatial_layer_id - 1] == 8 &&
3173 cm->width * cm->height > 640 * 480) {
3174 svc_mv.as_mv.row = -4;
3175 svc_mv.as_mv.col = -4;
3176 }
3177
3178 // Setup parameters used for inter mode evaluation.
3179 set_params_nonrd_pick_inter_mode(cpi, x, &search_state, rd_cost,
3180 &force_skip_low_temp_var, mi_row, mi_col,
3181 gf_temporal_ref, segment_id, bsize
3182 #if CONFIG_AV1_TEMPORAL_DENOISING
3183 ,
3184 ctx, denoise_svc_pickmode
3185 #endif
3186 );
3187
3188 if (rt_sf->use_comp_ref_nonrd && is_comp_ref_allowed(bsize)) {
3189 // Only search compound if bsize \gt BLOCK_16X16.
3190 if (bsize > BLOCK_16X16) {
3191 comp_use_zero_zeromv_only = rt_sf->check_only_zero_zeromv_on_large_blocks;
3192 } else {
3193 tot_num_comp_modes = 0;
3194 }
3195 } else {
3196 tot_num_comp_modes = 0;
3197 }
3198
3199 if (x->pred_mv_sad[LAST_FRAME] != INT_MAX) {
3200 thresh_sad_pred = ((int64_t)x->pred_mv_sad[LAST_FRAME]) << 1;
3201 // Increase threshold for less aggressive pruning.
3202 if (rt_sf->nonrd_prune_ref_frame_search == 1)
3203 thresh_sad_pred += (x->pred_mv_sad[LAST_FRAME] >> 2);
3204 }
3205
3206 const int use_model_yrd_large = get_model_rd_flag(cpi, xd, bsize);
3207
3208 // decide block-level interp filter search flags:
3209 // filter_search_enabled_blk:
3210 // 0: disabled
3211 // 1: filter search depends on mode properties
3212 // 2: filter search forced since prediction is unreliable
3213 // cb_pred_filter_search 0: disabled cb prediction
3214 InterpFilter filt_select = EIGHTTAP_REGULAR;
3215 const int cb_pred_filter_search =
3216 x->content_state_sb.source_sad_nonrd > kVeryLowSad
3217 ? cpi->sf.interp_sf.cb_pred_filter_search
3218 : 0;
3219 const int filter_search_enabled_blk =
3220 is_filter_search_enabled_blk(cpi, x, mi_row, mi_col, bsize, segment_id,
3221 cb_pred_filter_search, &filt_select);
3222
3223 #if COLLECT_NONRD_PICK_MODE_STAT
3224 x->ms_stat_nonrd.num_blocks[bsize]++;
3225 #endif
3226 init_mbmi_nonrd(mi, DC_PRED, NONE_FRAME, NONE_FRAME, cm);
3227 mi->tx_size = AOMMIN(
3228 AOMMIN(max_txsize_lookup[bsize],
3229 tx_mode_to_biggest_tx_size[txfm_params->tx_mode_search_type]),
3230 TX_16X16);
3231
3232 fill_single_inter_mode_costs(search_state.single_inter_mode_costs,
3233 num_inter_modes, ref_mode_set, mode_costs,
3234 mbmi_ext->mode_context);
3235
3236 MV_REFERENCE_FRAME last_comp_ref_frame = NONE_FRAME;
3237
3238 // Initialize inter prediction params at block level for single reference
3239 // mode.
3240 InterPredParams inter_pred_params_sr;
3241 init_inter_block_params(&inter_pred_params_sr, pd->width, pd->height,
3242 mi_row * MI_SIZE, mi_col * MI_SIZE, pd->subsampling_x,
3243 pd->subsampling_y, xd->bd, is_cur_buf_hbd(xd),
3244 /*is_intrabc=*/0);
3245 inter_pred_params_sr.conv_params =
3246 get_conv_params(/*do_average=*/0, AOM_PLANE_Y, xd->bd);
3247
3248 x->block_is_zero_sad = x->content_state_sb.source_sad_nonrd == kZeroSad ||
3249 segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP);
3250 if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
3251 !x->force_zeromv_skip_for_blk &&
3252 x->content_state_sb.source_sad_nonrd != kZeroSad &&
3253 x->source_variance == 0 && bsize < cm->seq_params->sb_size &&
3254 search_state.yv12_mb[LAST_FRAME][0].width == cm->width &&
3255 search_state.yv12_mb[LAST_FRAME][0].height == cm->height) {
3256 set_block_source_sad(cpi, x, bsize, &search_state.yv12_mb[LAST_FRAME][0]);
3257 }
3258
3259 int sb_me_has_been_tested = 0;
3260 x->sb_me_block = x->sb_me_partition;
3261 // Only use this feature (force testing of superblock motion) if coding
3262 // block size is large.
3263 if (x->sb_me_block) {
3264 if (cm->seq_params->sb_size == BLOCK_128X128 && bsize < BLOCK_64X64)
3265 x->sb_me_block = 0;
3266 else if (cm->seq_params->sb_size == BLOCK_64X64 && bsize < BLOCK_32X32)
3267 x->sb_me_block = 0;
3268 }
3269
3270 x->min_dist_inter_uv = INT64_MAX;
3271 for (int idx = 0; idx < num_inter_modes + tot_num_comp_modes; ++idx) {
3272 // If we are at the first compound mode, and the single modes already
3273 // perform well, then end the search.
3274 if (rt_sf->skip_compound_based_on_var && idx == num_inter_modes &&
3275 skip_comp_based_on_var(search_state.vars, bsize)) {
3276 break;
3277 }
3278
3279 int is_single_pred = 1;
3280 PREDICTION_MODE this_mode;
3281
3282 if (idx == 0 && !x->force_zeromv_skip_for_blk) {
3283 // Set color sensitivity on first tested mode only.
3284 // Use y-sad already computed in find_predictors: take the sad with motion
3285 // vector closest to 0; the uv-sad computed below in set_color_sensitivity
3286 // is for zeromv.
3287 // For screen: first check if golden reference is being used, if so,
3288 // force color_sensitivity on (=1) if the color sensitivity for sb_g is 1.
3289 // The check in set_color_sensitivity() will then follow and check for
3290 // setting the flag if the level is still 2 or 0.
3291 if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
3292 search_state.use_ref_frame_mask[GOLDEN_FRAME]) {
3293 if (x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_U)] == 1)
3294 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] = 1;
3295 if (x->color_sensitivity_sb_g[COLOR_SENS_IDX(AOM_PLANE_V)] == 1)
3296 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] = 1;
3297 }
3298 if (search_state.use_ref_frame_mask[LAST_FRAME] &&
3299 x->pred_mv0_sad[LAST_FRAME] != INT_MAX) {
3300 int y_sad = x->pred_mv0_sad[LAST_FRAME];
3301 if (x->pred_mv1_sad[LAST_FRAME] != INT_MAX &&
3302 (abs(search_state.frame_mv[NEARMV][LAST_FRAME].as_mv.col) +
3303 abs(search_state.frame_mv[NEARMV][LAST_FRAME].as_mv.row)) <
3304 (abs(search_state.frame_mv[NEARESTMV][LAST_FRAME].as_mv.col) +
3305 abs(search_state.frame_mv[NEARESTMV][LAST_FRAME].as_mv.row)))
3306 y_sad = x->pred_mv1_sad[LAST_FRAME];
3307 set_color_sensitivity(cpi, x, bsize, y_sad, x->source_variance,
3308 search_state.yv12_mb[LAST_FRAME]);
3309 }
3310 }
3311
3312 // Check the inter mode can be skipped based on mode statistics and speed
3313 // features settings.
3314 if (skip_inter_mode_nonrd(cpi, x, &search_state, &thresh_sad_pred,
3315 &force_mv_inter_layer, &is_single_pred,
3316 &this_mode, &last_comp_ref_frame, &ref_frame,
3317 &ref_frame2, idx, svc_mv, force_skip_low_temp_var,
3318 sse_zeromv_norm, num_inter_modes, segment_id,
3319 bsize, comp_use_zero_zeromv_only, check_globalmv))
3320 continue;
3321
3322 // Select prediction reference frames.
3323 for (int plane = 0; plane < MAX_MB_PLANE; plane++) {
3324 xd->plane[plane].pre[0] = search_state.yv12_mb[ref_frame][plane];
3325 if (!is_single_pred)
3326 xd->plane[plane].pre[1] = search_state.yv12_mb[ref_frame2][plane];
3327 }
3328
3329 mi->ref_frame[0] = ref_frame;
3330 mi->ref_frame[1] = ref_frame2;
3331 set_ref_ptrs(cm, xd, ref_frame, ref_frame2);
3332
3333 // Check if the scaled reference frame should be used. This is set in the
3334 // find_predictors() for each usable reference. If so, set the
3335 // block_ref_scale_factors[] to no reference scaling.
3336 if (search_state.use_scaled_ref_frame[ref_frame]) {
3337 xd->block_ref_scale_factors[0] = &sf_no_scale;
3338 }
3339 if (!is_single_pred && search_state.use_scaled_ref_frame[ref_frame2]) {
3340 xd->block_ref_scale_factors[1] = &sf_no_scale;
3341 }
3342
3343 // Perform inter mode evaluation for non-rd
3344 if (!handle_inter_mode_nonrd(
3345 cpi, x, &search_state, ctx, &this_mode_pred, tmp_buffer,
3346 inter_pred_params_sr, &best_early_term, &sse_zeromv_norm,
3347 &check_globalmv,
3348 #if CONFIG_AV1_TEMPORAL_DENOISING
3349 &zero_last_cost_orig, denoise_svc_pickmode,
3350 #endif
3351 idx, force_mv_inter_layer, is_single_pred, gf_temporal_ref,
3352 use_model_yrd_large, filter_search_enabled_blk, bsize, this_mode,
3353 filt_select, cb_pred_filter_search, reuse_inter_pred,
3354 &sb_me_has_been_tested)) {
3355 break;
3356 }
3357 }
3358
3359 // Restore mode data of best inter mode
3360 mi->mode = best_pickmode->best_mode;
3361 mi->motion_mode = best_pickmode->best_motion_mode;
3362 mi->wm_params = best_pickmode->wm_params;
3363 mi->num_proj_ref = best_pickmode->num_proj_ref;
3364 mi->interp_filters = best_pickmode->best_pred_filter;
3365 mi->tx_size = best_pickmode->best_tx_size;
3366 memset(mi->inter_tx_size, mi->tx_size, sizeof(mi->inter_tx_size));
3367 mi->ref_frame[0] = best_pickmode->best_ref_frame;
3368 mi->mv[0].as_int = search_state
3369 .frame_mv_best[best_pickmode->best_mode]
3370 [best_pickmode->best_ref_frame]
3371 .as_int;
3372 mi->mv[1].as_int = 0;
3373 if (best_pickmode->best_second_ref_frame > INTRA_FRAME) {
3374 mi->ref_frame[1] = best_pickmode->best_second_ref_frame;
3375 mi->mv[1].as_int = search_state
3376 .frame_mv_best[best_pickmode->best_mode]
3377 [best_pickmode->best_second_ref_frame]
3378 .as_int;
3379 }
3380 // Perform intra prediction search, if the best SAD is above a certain
3381 // threshold.
3382 mi->angle_delta[PLANE_TYPE_Y] = 0;
3383 mi->angle_delta[PLANE_TYPE_UV] = 0;
3384 mi->filter_intra_mode_info.use_filter_intra = 0;
3385
3386 #if COLLECT_NONRD_PICK_MODE_STAT
3387 aom_usec_timer_start(&x->ms_stat_nonrd.timer1);
3388 x->ms_stat_nonrd.num_searches[bsize][DC_PRED]++;
3389 x->ms_stat_nonrd.num_nonskipped_searches[bsize][DC_PRED]++;
3390 #endif
3391
3392 int force_palette_test = 0;
3393 if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
3394 x->content_state_sb.source_sad_nonrd != kZeroSad &&
3395 bsize <= BLOCK_16X16) {
3396 unsigned int thresh_sse = cpi->rc.high_source_sad ? 15000 : 200000;
3397 unsigned int thresh_source_var = cpi->rc.high_source_sad ? 50 : 200;
3398 unsigned int best_sse_inter_motion =
3399 (unsigned int)(search_state.best_rdc.sse >>
3400 (b_width_log2_lookup[bsize] +
3401 b_height_log2_lookup[bsize]));
3402 if (best_sse_inter_motion > thresh_sse &&
3403 x->source_variance > thresh_source_var)
3404 force_palette_test = 1;
3405 }
3406
3407 // Evaluate Intra modes in inter frame
3408 if (!x->force_zeromv_skip_for_blk)
3409 av1_estimate_intra_mode(cpi, x, bsize, best_early_term,
3410 search_state.ref_costs_single[INTRA_FRAME],
3411 reuse_inter_pred, &orig_dst, tmp_buffer,
3412 &this_mode_pred, &search_state.best_rdc,
3413 best_pickmode, ctx);
3414
3415 int skip_idtx_palette = (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] ||
3416 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)]) &&
3417 x->content_state_sb.source_sad_nonrd != kZeroSad &&
3418 !cpi->rc.high_source_sad;
3419
3420 int try_palette =
3421 !skip_idtx_palette && cpi->oxcf.tool_cfg.enable_palette &&
3422 av1_allow_palette(cpi->common.features.allow_screen_content_tools,
3423 mi->bsize);
3424 try_palette =
3425 try_palette &&
3426 (is_mode_intra(best_pickmode->best_mode) || force_palette_test) &&
3427 x->source_variance > 0 && !x->force_zeromv_skip_for_blk &&
3428 (cpi->rc.high_source_sad || x->source_variance > 300);
3429
3430 if (rt_sf->prune_palette_nonrd && bsize > BLOCK_16X16) try_palette = 0;
3431
3432 // Perform screen content mode evaluation for non-rd
3433 handle_screen_content_mode_nonrd(
3434 cpi, x, &search_state, this_mode_pred, ctx, tmp_buffer, &orig_dst,
3435 skip_idtx_palette, try_palette, bsize, reuse_inter_pred, mi_col, mi_row);
3436
3437 #if COLLECT_NONRD_PICK_MODE_STAT
3438 aom_usec_timer_mark(&x->ms_stat_nonrd.timer1);
3439 x->ms_stat_nonrd.nonskipped_search_times[bsize][DC_PRED] +=
3440 aom_usec_timer_elapsed(&x->ms_stat_nonrd.timer1);
3441 #endif
3442
3443 pd->dst = orig_dst;
3444 // Best mode is finalized. Restore the mode data to mbmi
3445 if (try_palette) mi->palette_mode_info = best_pickmode->pmi;
3446 mi->mode = best_pickmode->best_mode;
3447 mi->ref_frame[0] = best_pickmode->best_ref_frame;
3448 mi->ref_frame[1] = best_pickmode->best_second_ref_frame;
3449 // For lossless: always force the skip flags off.
3450 if (is_lossless_requested(&cpi->oxcf.rc_cfg)) {
3451 txfm_info->skip_txfm = 0;
3452 memset(ctx->blk_skip, 0, sizeof(ctx->blk_skip[0]) * ctx->num_4x4_blk);
3453 } else {
3454 txfm_info->skip_txfm = best_pickmode->best_mode_skip_txfm;
3455 }
3456 if (has_second_ref(mi)) {
3457 mi->comp_group_idx = 0;
3458 mi->compound_idx = 1;
3459 mi->interinter_comp.type = COMPOUND_AVERAGE;
3460 }
3461
3462 if (!is_inter_block(mi)) {
3463 mi->interp_filters = av1_broadcast_interp_filter(SWITCHABLE_FILTERS);
3464 } else {
3465 // If inter mode is selected and ref_frame was one that uses the
3466 // scaled reference frame, then we can't use reuse_inter_pred.
3467 if (search_state.use_scaled_ref_frame[best_pickmode->best_ref_frame] ||
3468 (has_second_ref(mi) &&
3469 search_state
3470 .use_scaled_ref_frame[best_pickmode->best_second_ref_frame]))
3471 x->reuse_inter_pred = 0;
3472 }
3473
3474 // Restore the predicted samples of best mode to final buffer
3475 if (reuse_inter_pred && best_pickmode->best_pred != NULL) {
3476 PRED_BUFFER *const best_pred = best_pickmode->best_pred;
3477 if (best_pred->data != orig_dst.buf && is_inter_mode(mi->mode)) {
3478 aom_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
3479 pd->dst.stride, bw, bh);
3480 }
3481 }
3482
3483 #if CONFIG_AV1_TEMPORAL_DENOISING
3484 if (cpi->oxcf.noise_sensitivity > 0 && resize_pending == 0 &&
3485 denoise_svc_pickmode && cpi->denoiser.denoising_level > kDenLowLow &&
3486 cpi->denoiser.reset == 0) {
3487 AV1_DENOISER_DECISION decision = COPY_BLOCK;
3488 ctx->sb_skip_denoising = 0;
3489 av1_pickmode_ctx_den_update(
3490 &ctx_den, zero_last_cost_orig, search_state.ref_costs_single,
3491 search_state.frame_mv, reuse_inter_pred, best_pickmode);
3492 av1_denoiser_denoise(cpi, x, mi_row, mi_col, bsize, ctx, &decision,
3493 gf_temporal_ref);
3494 if (denoise_recheck_zeromv)
3495 recheck_zeromv_after_denoising(
3496 cpi, mi, x, xd, decision, &ctx_den, search_state.yv12_mb,
3497 &search_state.best_rdc, best_pickmode, bsize, mi_row, mi_col);
3498 best_pickmode->best_ref_frame = ctx_den.best_ref_frame;
3499 }
3500 #endif
3501
3502 // Update the factors used for RD thresholding for all modes.
3503 if (cpi->sf.inter_sf.adaptive_rd_thresh && !has_second_ref(mi)) {
3504 THR_MODES best_mode_idx =
3505 mode_idx[best_pickmode->best_ref_frame][mode_offset(mi->mode)];
3506 if (best_pickmode->best_ref_frame == INTRA_FRAME) {
3507 // Only consider the modes that are included in the intra_mode_list.
3508 int intra_modes = sizeof(intra_mode_list) / sizeof(PREDICTION_MODE);
3509 for (int mode_index = 0; mode_index < intra_modes; mode_index++) {
3510 update_thresh_freq_fact(cpi, x, bsize, INTRA_FRAME, best_mode_idx,
3511 intra_mode_list[mode_index]);
3512 }
3513 } else {
3514 PREDICTION_MODE this_mode;
3515 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
3516 update_thresh_freq_fact(cpi, x, bsize, best_pickmode->best_ref_frame,
3517 best_mode_idx, this_mode);
3518 }
3519 }
3520 }
3521
3522 #if CONFIG_INTERNAL_STATS
3523 store_coding_context_nonrd(x, ctx, mi->mode);
3524 #else
3525 store_coding_context_nonrd(x, ctx);
3526 #endif // CONFIG_INTERNAL_STATS
3527
3528 #if COLLECT_NONRD_PICK_MODE_STAT
3529 aom_usec_timer_mark(&x->ms_stat_nonrd.bsize_timer);
3530 x->ms_stat_nonrd.total_block_times[bsize] +=
3531 aom_usec_timer_elapsed(&x->ms_stat_nonrd.bsize_timer);
3532 print_time(&x->ms_stat_nonrd, bsize, cm->mi_params.mi_rows,
3533 cm->mi_params.mi_cols, mi_row, mi_col);
3534 #endif // COLLECT_NONRD_PICK_MODE_STAT
3535
3536 *rd_cost = search_state.best_rdc;
3537
3538 // Reset the xd->block_ref_scale_factors[i], as they may have
3539 // been set to pointer &sf_no_scale, which becomes invalid afer
3540 // this function.
3541 set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
3542 }
3543