1 /*
2 * Copyright (c) 2019, 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 #ifndef AOM_AV1_ENCODER_RDOPT_UTILS_H_
13 #define AOM_AV1_ENCODER_RDOPT_UTILS_H_
14
15 #include "aom/aom_integer.h"
16 #include "av1/encoder/block.h"
17 #include "av1/common/cfl.h"
18 #include "av1/common/pred_common.h"
19 #include "av1/encoder/rdopt_data_defs.h"
20
21 #ifdef __cplusplus
22 extern "C" {
23 #endif
24
25 #define MAX_REF_MV_SEARCH 3
26 #define INTER_INTRA_RD_THRESH_SCALE 9
27 #define INTER_INTRA_RD_THRESH_SHIFT 4
28
29 typedef struct {
30 PREDICTION_MODE mode;
31 MV_REFERENCE_FRAME ref_frame[2];
32 } MODE_DEFINITION;
33
34 // This array defines the mapping from the enums in THR_MODES to the actual
35 // prediction modes and refrence frames
36 static const MODE_DEFINITION av1_mode_defs[MAX_MODES] = {
37 { NEARESTMV, { LAST_FRAME, NONE_FRAME } },
38 { NEARESTMV, { LAST2_FRAME, NONE_FRAME } },
39 { NEARESTMV, { LAST3_FRAME, NONE_FRAME } },
40 { NEARESTMV, { BWDREF_FRAME, NONE_FRAME } },
41 { NEARESTMV, { ALTREF2_FRAME, NONE_FRAME } },
42 { NEARESTMV, { ALTREF_FRAME, NONE_FRAME } },
43 { NEARESTMV, { GOLDEN_FRAME, NONE_FRAME } },
44
45 { NEWMV, { LAST_FRAME, NONE_FRAME } },
46 { NEWMV, { LAST2_FRAME, NONE_FRAME } },
47 { NEWMV, { LAST3_FRAME, NONE_FRAME } },
48 { NEWMV, { BWDREF_FRAME, NONE_FRAME } },
49 { NEWMV, { ALTREF2_FRAME, NONE_FRAME } },
50 { NEWMV, { ALTREF_FRAME, NONE_FRAME } },
51 { NEWMV, { GOLDEN_FRAME, NONE_FRAME } },
52
53 { NEARMV, { LAST_FRAME, NONE_FRAME } },
54 { NEARMV, { LAST2_FRAME, NONE_FRAME } },
55 { NEARMV, { LAST3_FRAME, NONE_FRAME } },
56 { NEARMV, { BWDREF_FRAME, NONE_FRAME } },
57 { NEARMV, { ALTREF2_FRAME, NONE_FRAME } },
58 { NEARMV, { ALTREF_FRAME, NONE_FRAME } },
59 { NEARMV, { GOLDEN_FRAME, NONE_FRAME } },
60
61 { GLOBALMV, { LAST_FRAME, NONE_FRAME } },
62 { GLOBALMV, { LAST2_FRAME, NONE_FRAME } },
63 { GLOBALMV, { LAST3_FRAME, NONE_FRAME } },
64 { GLOBALMV, { BWDREF_FRAME, NONE_FRAME } },
65 { GLOBALMV, { ALTREF2_FRAME, NONE_FRAME } },
66 { GLOBALMV, { ALTREF_FRAME, NONE_FRAME } },
67 { GLOBALMV, { GOLDEN_FRAME, NONE_FRAME } },
68
69 // TODO(zoeliu): May need to reconsider the order on the modes to check
70
71 { NEAREST_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
72 { NEAREST_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
73 { NEAREST_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
74 { NEAREST_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
75 { NEAREST_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
76 { NEAREST_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
77 { NEAREST_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
78 { NEAREST_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
79 { NEAREST_NEARESTMV, { LAST_FRAME, ALTREF2_FRAME } },
80 { NEAREST_NEARESTMV, { LAST2_FRAME, ALTREF2_FRAME } },
81 { NEAREST_NEARESTMV, { LAST3_FRAME, ALTREF2_FRAME } },
82 { NEAREST_NEARESTMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
83
84 { NEAREST_NEARESTMV, { LAST_FRAME, LAST2_FRAME } },
85 { NEAREST_NEARESTMV, { LAST_FRAME, LAST3_FRAME } },
86 { NEAREST_NEARESTMV, { LAST_FRAME, GOLDEN_FRAME } },
87 { NEAREST_NEARESTMV, { BWDREF_FRAME, ALTREF_FRAME } },
88
89 { NEAR_NEARMV, { LAST_FRAME, BWDREF_FRAME } },
90 { NEW_NEWMV, { LAST_FRAME, BWDREF_FRAME } },
91 { NEW_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
92 { NEAREST_NEWMV, { LAST_FRAME, BWDREF_FRAME } },
93 { NEW_NEARMV, { LAST_FRAME, BWDREF_FRAME } },
94 { NEAR_NEWMV, { LAST_FRAME, BWDREF_FRAME } },
95 { GLOBAL_GLOBALMV, { LAST_FRAME, BWDREF_FRAME } },
96
97 { NEAR_NEARMV, { LAST_FRAME, ALTREF_FRAME } },
98 { NEW_NEWMV, { LAST_FRAME, ALTREF_FRAME } },
99 { NEW_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
100 { NEAREST_NEWMV, { LAST_FRAME, ALTREF_FRAME } },
101 { NEW_NEARMV, { LAST_FRAME, ALTREF_FRAME } },
102 { NEAR_NEWMV, { LAST_FRAME, ALTREF_FRAME } },
103 { GLOBAL_GLOBALMV, { LAST_FRAME, ALTREF_FRAME } },
104
105 { NEAR_NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
106 { NEW_NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
107 { NEW_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
108 { NEAREST_NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
109 { NEW_NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
110 { NEAR_NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
111 { GLOBAL_GLOBALMV, { LAST2_FRAME, ALTREF_FRAME } },
112
113 { NEAR_NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
114 { NEW_NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
115 { NEW_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
116 { NEAREST_NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
117 { NEW_NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
118 { NEAR_NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
119 { GLOBAL_GLOBALMV, { LAST3_FRAME, ALTREF_FRAME } },
120
121 { NEAR_NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
122 { NEW_NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
123 { NEW_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
124 { NEAREST_NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
125 { NEW_NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
126 { NEAR_NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
127 { GLOBAL_GLOBALMV, { GOLDEN_FRAME, ALTREF_FRAME } },
128
129 { NEAR_NEARMV, { LAST2_FRAME, BWDREF_FRAME } },
130 { NEW_NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
131 { NEW_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
132 { NEAREST_NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
133 { NEW_NEARMV, { LAST2_FRAME, BWDREF_FRAME } },
134 { NEAR_NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
135 { GLOBAL_GLOBALMV, { LAST2_FRAME, BWDREF_FRAME } },
136
137 { NEAR_NEARMV, { LAST3_FRAME, BWDREF_FRAME } },
138 { NEW_NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
139 { NEW_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
140 { NEAREST_NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
141 { NEW_NEARMV, { LAST3_FRAME, BWDREF_FRAME } },
142 { NEAR_NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
143 { GLOBAL_GLOBALMV, { LAST3_FRAME, BWDREF_FRAME } },
144
145 { NEAR_NEARMV, { GOLDEN_FRAME, BWDREF_FRAME } },
146 { NEW_NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
147 { NEW_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
148 { NEAREST_NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
149 { NEW_NEARMV, { GOLDEN_FRAME, BWDREF_FRAME } },
150 { NEAR_NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
151 { GLOBAL_GLOBALMV, { GOLDEN_FRAME, BWDREF_FRAME } },
152
153 { NEAR_NEARMV, { LAST_FRAME, ALTREF2_FRAME } },
154 { NEW_NEWMV, { LAST_FRAME, ALTREF2_FRAME } },
155 { NEW_NEARESTMV, { LAST_FRAME, ALTREF2_FRAME } },
156 { NEAREST_NEWMV, { LAST_FRAME, ALTREF2_FRAME } },
157 { NEW_NEARMV, { LAST_FRAME, ALTREF2_FRAME } },
158 { NEAR_NEWMV, { LAST_FRAME, ALTREF2_FRAME } },
159 { GLOBAL_GLOBALMV, { LAST_FRAME, ALTREF2_FRAME } },
160
161 { NEAR_NEARMV, { LAST2_FRAME, ALTREF2_FRAME } },
162 { NEW_NEWMV, { LAST2_FRAME, ALTREF2_FRAME } },
163 { NEW_NEARESTMV, { LAST2_FRAME, ALTREF2_FRAME } },
164 { NEAREST_NEWMV, { LAST2_FRAME, ALTREF2_FRAME } },
165 { NEW_NEARMV, { LAST2_FRAME, ALTREF2_FRAME } },
166 { NEAR_NEWMV, { LAST2_FRAME, ALTREF2_FRAME } },
167 { GLOBAL_GLOBALMV, { LAST2_FRAME, ALTREF2_FRAME } },
168
169 { NEAR_NEARMV, { LAST3_FRAME, ALTREF2_FRAME } },
170 { NEW_NEWMV, { LAST3_FRAME, ALTREF2_FRAME } },
171 { NEW_NEARESTMV, { LAST3_FRAME, ALTREF2_FRAME } },
172 { NEAREST_NEWMV, { LAST3_FRAME, ALTREF2_FRAME } },
173 { NEW_NEARMV, { LAST3_FRAME, ALTREF2_FRAME } },
174 { NEAR_NEWMV, { LAST3_FRAME, ALTREF2_FRAME } },
175 { GLOBAL_GLOBALMV, { LAST3_FRAME, ALTREF2_FRAME } },
176
177 { NEAR_NEARMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
178 { NEW_NEWMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
179 { NEW_NEARESTMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
180 { NEAREST_NEWMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
181 { NEW_NEARMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
182 { NEAR_NEWMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
183 { GLOBAL_GLOBALMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
184
185 { NEAR_NEARMV, { LAST_FRAME, LAST2_FRAME } },
186 { NEW_NEWMV, { LAST_FRAME, LAST2_FRAME } },
187 { NEW_NEARESTMV, { LAST_FRAME, LAST2_FRAME } },
188 { NEAREST_NEWMV, { LAST_FRAME, LAST2_FRAME } },
189 { NEW_NEARMV, { LAST_FRAME, LAST2_FRAME } },
190 { NEAR_NEWMV, { LAST_FRAME, LAST2_FRAME } },
191 { GLOBAL_GLOBALMV, { LAST_FRAME, LAST2_FRAME } },
192
193 { NEAR_NEARMV, { LAST_FRAME, LAST3_FRAME } },
194 { NEW_NEWMV, { LAST_FRAME, LAST3_FRAME } },
195 { NEW_NEARESTMV, { LAST_FRAME, LAST3_FRAME } },
196 { NEAREST_NEWMV, { LAST_FRAME, LAST3_FRAME } },
197 { NEW_NEARMV, { LAST_FRAME, LAST3_FRAME } },
198 { NEAR_NEWMV, { LAST_FRAME, LAST3_FRAME } },
199 { GLOBAL_GLOBALMV, { LAST_FRAME, LAST3_FRAME } },
200
201 { NEAR_NEARMV, { LAST_FRAME, GOLDEN_FRAME } },
202 { NEW_NEWMV, { LAST_FRAME, GOLDEN_FRAME } },
203 { NEW_NEARESTMV, { LAST_FRAME, GOLDEN_FRAME } },
204 { NEAREST_NEWMV, { LAST_FRAME, GOLDEN_FRAME } },
205 { NEW_NEARMV, { LAST_FRAME, GOLDEN_FRAME } },
206 { NEAR_NEWMV, { LAST_FRAME, GOLDEN_FRAME } },
207 { GLOBAL_GLOBALMV, { LAST_FRAME, GOLDEN_FRAME } },
208
209 { NEAR_NEARMV, { BWDREF_FRAME, ALTREF_FRAME } },
210 { NEW_NEWMV, { BWDREF_FRAME, ALTREF_FRAME } },
211 { NEW_NEARESTMV, { BWDREF_FRAME, ALTREF_FRAME } },
212 { NEAREST_NEWMV, { BWDREF_FRAME, ALTREF_FRAME } },
213 { NEW_NEARMV, { BWDREF_FRAME, ALTREF_FRAME } },
214 { NEAR_NEWMV, { BWDREF_FRAME, ALTREF_FRAME } },
215 { GLOBAL_GLOBALMV, { BWDREF_FRAME, ALTREF_FRAME } },
216
217 // intra modes
218 { DC_PRED, { INTRA_FRAME, NONE_FRAME } },
219 { PAETH_PRED, { INTRA_FRAME, NONE_FRAME } },
220 { SMOOTH_PRED, { INTRA_FRAME, NONE_FRAME } },
221 { SMOOTH_V_PRED, { INTRA_FRAME, NONE_FRAME } },
222 { SMOOTH_H_PRED, { INTRA_FRAME, NONE_FRAME } },
223 { H_PRED, { INTRA_FRAME, NONE_FRAME } },
224 { V_PRED, { INTRA_FRAME, NONE_FRAME } },
225 { D135_PRED, { INTRA_FRAME, NONE_FRAME } },
226 { D203_PRED, { INTRA_FRAME, NONE_FRAME } },
227 { D157_PRED, { INTRA_FRAME, NONE_FRAME } },
228 { D67_PRED, { INTRA_FRAME, NONE_FRAME } },
229 { D113_PRED, { INTRA_FRAME, NONE_FRAME } },
230 { D45_PRED, { INTRA_FRAME, NONE_FRAME } },
231 };
232
233 // Number of winner modes allowed for different values of the speed feature
234 // multi_winner_mode_type.
235 static const int winner_mode_count_allowed[MULTI_WINNER_MODE_LEVELS] = {
236 1, // MULTI_WINNER_MODE_OFF
237 2, // MULTI_WINNER_MODE_FAST
238 3 // MULTI_WINNER_MODE_DEFAULT
239 };
240
restore_dst_buf(MACROBLOCKD * xd,const BUFFER_SET dst,const int num_planes)241 static AOM_INLINE void restore_dst_buf(MACROBLOCKD *xd, const BUFFER_SET dst,
242 const int num_planes) {
243 for (int i = 0; i < num_planes; i++) {
244 xd->plane[i].dst.buf = dst.plane[i];
245 xd->plane[i].dst.stride = dst.stride[i];
246 }
247 }
248
swap_dst_buf(MACROBLOCKD * xd,const BUFFER_SET * dst_bufs[2],int num_planes)249 static AOM_INLINE void swap_dst_buf(MACROBLOCKD *xd,
250 const BUFFER_SET *dst_bufs[2],
251 int num_planes) {
252 const BUFFER_SET *buf0 = dst_bufs[0];
253 dst_bufs[0] = dst_bufs[1];
254 dst_bufs[1] = buf0;
255 restore_dst_buf(xd, *dst_bufs[0], num_planes);
256 }
257
258 /* clang-format on */
259 // Calculate rd threshold based on ref best rd and relevant scaling factors
get_rd_thresh_from_best_rd(int64_t ref_best_rd,int mul_factor,int div_factor)260 static AOM_INLINE int64_t get_rd_thresh_from_best_rd(int64_t ref_best_rd,
261 int mul_factor,
262 int div_factor) {
263 int64_t rd_thresh = ref_best_rd;
264 if (div_factor != 0) {
265 rd_thresh = ref_best_rd < (div_factor * (INT64_MAX / mul_factor))
266 ? ((ref_best_rd / div_factor) * mul_factor)
267 : INT64_MAX;
268 }
269 return rd_thresh;
270 }
271
272 static AOM_INLINE THR_MODES
get_prediction_mode_idx(PREDICTION_MODE this_mode,MV_REFERENCE_FRAME ref_frame,MV_REFERENCE_FRAME second_ref_frame)273 get_prediction_mode_idx(PREDICTION_MODE this_mode, MV_REFERENCE_FRAME ref_frame,
274 MV_REFERENCE_FRAME second_ref_frame) {
275 if (this_mode < INTRA_MODE_END) {
276 assert(ref_frame == INTRA_FRAME);
277 assert(second_ref_frame == NONE_FRAME);
278 return intra_to_mode_idx[this_mode - INTRA_MODE_START];
279 }
280 if (this_mode >= SINGLE_INTER_MODE_START &&
281 this_mode < SINGLE_INTER_MODE_END) {
282 assert((ref_frame > INTRA_FRAME) && (ref_frame <= ALTREF_FRAME));
283 return single_inter_to_mode_idx[this_mode - SINGLE_INTER_MODE_START]
284 [ref_frame];
285 }
286 if (this_mode >= COMP_INTER_MODE_START && this_mode < COMP_INTER_MODE_END &&
287 second_ref_frame != NONE_FRAME) {
288 assert((ref_frame > INTRA_FRAME) && (ref_frame <= ALTREF_FRAME));
289 assert((second_ref_frame > INTRA_FRAME) &&
290 (second_ref_frame <= ALTREF_FRAME));
291 return comp_inter_to_mode_idx[this_mode - COMP_INTER_MODE_START][ref_frame]
292 [second_ref_frame];
293 }
294 assert(0);
295 return THR_INVALID;
296 }
297
inter_mode_data_block_idx(BLOCK_SIZE bsize)298 static AOM_INLINE int inter_mode_data_block_idx(BLOCK_SIZE bsize) {
299 if (bsize == BLOCK_4X4 || bsize == BLOCK_4X8 || bsize == BLOCK_8X4 ||
300 bsize == BLOCK_4X16 || bsize == BLOCK_16X4) {
301 return -1;
302 }
303 return 1;
304 }
305
306 // Get transform block visible dimensions cropped to the MI units.
get_txb_dimensions(const MACROBLOCKD * xd,int plane,BLOCK_SIZE plane_bsize,int blk_row,int blk_col,BLOCK_SIZE tx_bsize,int * width,int * height,int * visible_width,int * visible_height)307 static AOM_INLINE void get_txb_dimensions(const MACROBLOCKD *xd, int plane,
308 BLOCK_SIZE plane_bsize, int blk_row,
309 int blk_col, BLOCK_SIZE tx_bsize,
310 int *width, int *height,
311 int *visible_width,
312 int *visible_height) {
313 assert(tx_bsize <= plane_bsize);
314 const int txb_height = block_size_high[tx_bsize];
315 const int txb_width = block_size_wide[tx_bsize];
316 const struct macroblockd_plane *const pd = &xd->plane[plane];
317
318 // TODO(aconverse@google.com): Investigate using crop_width/height here rather
319 // than the MI size
320 if (xd->mb_to_bottom_edge >= 0) {
321 *visible_height = txb_height;
322 } else {
323 const int block_height = block_size_high[plane_bsize];
324 const int block_rows =
325 (xd->mb_to_bottom_edge >> (3 + pd->subsampling_y)) + block_height;
326 *visible_height =
327 clamp(block_rows - (blk_row << MI_SIZE_LOG2), 0, txb_height);
328 }
329 if (height) *height = txb_height;
330
331 if (xd->mb_to_right_edge >= 0) {
332 *visible_width = txb_width;
333 } else {
334 const int block_width = block_size_wide[plane_bsize];
335 const int block_cols =
336 (xd->mb_to_right_edge >> (3 + pd->subsampling_x)) + block_width;
337 *visible_width =
338 clamp(block_cols - (blk_col << MI_SIZE_LOG2), 0, txb_width);
339 }
340 if (width) *width = txb_width;
341 }
342
bsize_to_num_blk(BLOCK_SIZE bsize)343 static AOM_INLINE int bsize_to_num_blk(BLOCK_SIZE bsize) {
344 int num_blk = 1 << (num_pels_log2_lookup[bsize] - 2 * MI_SIZE_LOG2);
345 return num_blk;
346 }
347
check_txfm_eval(MACROBLOCK * const x,BLOCK_SIZE bsize,int64_t best_skip_rd,int64_t skip_rd,int level,int is_luma_only)348 static INLINE int check_txfm_eval(MACROBLOCK *const x, BLOCK_SIZE bsize,
349 int64_t best_skip_rd, int64_t skip_rd,
350 int level, int is_luma_only) {
351 int eval_txfm = 1;
352 // Derive aggressiveness factor for gating the transform search
353 // Lower value indicates more aggressiveness. Be more conservative (high
354 // value) for (i) low quantizers (ii) regions where prediction is poor
355 const int scale[5] = { INT_MAX, 4, 3, 2, 2 };
356 const int qslope = 2 * (!is_luma_only);
357 const int level_to_qindex_map[5] = { 0, 0, 0, 80, 100 };
358 int aggr_factor = 4;
359 const int pred_qindex_thresh = level_to_qindex_map[level];
360 if (!is_luma_only && level <= 2) {
361 aggr_factor = 4 * AOMMAX(1, ROUND_POWER_OF_TWO((MAXQ - x->qindex) * qslope,
362 QINDEX_BITS));
363 }
364 if ((best_skip_rd >
365 (x->source_variance << (num_pels_log2_lookup[bsize] + RDDIV_BITS))) &&
366 (x->qindex >= pred_qindex_thresh))
367 aggr_factor *= scale[level];
368 // For level setting 1, be more conservative for non-luma-only case even when
369 // prediction is good.
370 else if ((level <= 1) && !is_luma_only)
371 aggr_factor = (aggr_factor >> 2) * 6;
372
373 // Be more conservative for luma only cases (called from compound type rd)
374 // since best_skip_rd is computed after and skip_rd is computed (with 8-bit
375 // prediction signals blended for WEDGE/DIFFWTD rather than 16-bit) before
376 // interpolation filter search
377 const int luma_mul[5] = { INT_MAX, 32, 29, 17, 17 };
378 int mul_factor = is_luma_only ? luma_mul[level] : 16;
379 int64_t rd_thresh =
380 (best_skip_rd == INT64_MAX)
381 ? best_skip_rd
382 : (int64_t)(best_skip_rd * aggr_factor * mul_factor >> 6);
383 if (skip_rd > rd_thresh) eval_txfm = 0;
384 return eval_txfm;
385 }
386
select_tx_mode(const AV1_COMMON * cm,const TX_SIZE_SEARCH_METHOD tx_size_search_method)387 static TX_MODE select_tx_mode(
388 const AV1_COMMON *cm, const TX_SIZE_SEARCH_METHOD tx_size_search_method) {
389 if (cm->features.coded_lossless) return ONLY_4X4;
390 if (tx_size_search_method == USE_LARGESTALL) {
391 return TX_MODE_LARGEST;
392 } else {
393 assert(tx_size_search_method == USE_FULL_RD ||
394 tx_size_search_method == USE_FAST_RD);
395 return TX_MODE_SELECT;
396 }
397 }
398
399 // Checks the conditions to disable winner mode processing
bypass_winner_mode_processing(const MACROBLOCK * const x,const SPEED_FEATURES * sf,int use_txfm_skip,int actual_txfm_skip,PREDICTION_MODE best_mode)400 static INLINE int bypass_winner_mode_processing(const MACROBLOCK *const x,
401 const SPEED_FEATURES *sf,
402 int use_txfm_skip,
403 int actual_txfm_skip,
404 PREDICTION_MODE best_mode) {
405 const int prune_winner_mode_eval_level =
406 sf->winner_mode_sf.prune_winner_mode_eval_level;
407
408 // Disable winner mode processing for blocks with low source variance.
409 // The aggressiveness of this pruning logic reduces as qindex increases.
410 // The threshold decreases linearly from 64 as qindex varies from 0 to 255.
411 if (prune_winner_mode_eval_level == 1) {
412 const unsigned int src_var_thresh = 64 - 48 * x->qindex / (MAXQ + 1);
413 if (x->source_variance < src_var_thresh) return 1;
414 } else if (prune_winner_mode_eval_level == 2) {
415 // Skip winner mode processing of blocks for which transform turns out to be
416 // skip due to nature of eob alone except NEWMV mode.
417 if (!have_newmv_in_inter_mode(best_mode) && actual_txfm_skip) return 1;
418 } else if (prune_winner_mode_eval_level == 3) {
419 // Skip winner mode processing of blocks for which transform turns out to be
420 // skip except NEWMV mode and considered based on the quantizer.
421 // At high quantizers: Take conservative approach by considering transform
422 // skip based on eob alone.
423 // At low quantizers: Consider transform skip based on eob nature or RD cost
424 // evaluation.
425 const int is_txfm_skip =
426 x->qindex > 127 ? actual_txfm_skip : actual_txfm_skip || use_txfm_skip;
427
428 if (!have_newmv_in_inter_mode(best_mode) && is_txfm_skip) return 1;
429 } else if (prune_winner_mode_eval_level >= 4) {
430 // Do not skip winner mode evaluation at low quantizers if normal mode's
431 // transform search was too aggressive.
432 if (sf->rd_sf.perform_coeff_opt >= 5 && x->qindex <= 70) return 0;
433
434 if (use_txfm_skip || actual_txfm_skip) return 1;
435 }
436
437 return 0;
438 }
439
440 // Checks the conditions to enable winner mode processing
is_winner_mode_processing_enabled(const struct AV1_COMP * cpi,const MACROBLOCK * const x,MB_MODE_INFO * const mbmi,int actual_txfm_skip)441 static INLINE int is_winner_mode_processing_enabled(const struct AV1_COMP *cpi,
442 const MACROBLOCK *const x,
443 MB_MODE_INFO *const mbmi,
444 int actual_txfm_skip) {
445 const SPEED_FEATURES *sf = &cpi->sf;
446 const PREDICTION_MODE best_mode = mbmi->mode;
447
448 if (bypass_winner_mode_processing(x, sf, mbmi->skip_txfm, actual_txfm_skip,
449 best_mode))
450 return 0;
451
452 // TODO(any): Move block independent condition checks to frame level
453 if (is_inter_block(mbmi)) {
454 if (is_inter_mode(best_mode) &&
455 (sf->tx_sf.tx_type_search.fast_inter_tx_type_prob_thresh != INT_MAX) &&
456 !cpi->oxcf.txfm_cfg.use_inter_dct_only)
457 return 1;
458 } else {
459 if (sf->tx_sf.tx_type_search.fast_intra_tx_type_search &&
460 !cpi->oxcf.txfm_cfg.use_intra_default_tx_only &&
461 !cpi->oxcf.txfm_cfg.use_intra_dct_only)
462 return 1;
463 }
464
465 // Check speed feature related to winner mode processing
466 if (sf->winner_mode_sf.enable_winner_mode_for_coeff_opt &&
467 cpi->optimize_seg_arr[mbmi->segment_id] != NO_TRELLIS_OPT &&
468 cpi->optimize_seg_arr[mbmi->segment_id] != FINAL_PASS_TRELLIS_OPT)
469 return 1;
470 if (sf->winner_mode_sf.enable_winner_mode_for_tx_size_srch) return 1;
471
472 return 0;
473 }
474
set_tx_size_search_method(const AV1_COMMON * cm,const WinnerModeParams * winner_mode_params,TxfmSearchParams * txfm_params,int enable_winner_mode_for_tx_size_srch,int is_winner_mode)475 static INLINE void set_tx_size_search_method(
476 const AV1_COMMON *cm, const WinnerModeParams *winner_mode_params,
477 TxfmSearchParams *txfm_params, int enable_winner_mode_for_tx_size_srch,
478 int is_winner_mode) {
479 // Populate transform size search method/transform mode appropriately
480 txfm_params->tx_size_search_method =
481 winner_mode_params->tx_size_search_methods[DEFAULT_EVAL];
482 if (enable_winner_mode_for_tx_size_srch) {
483 if (is_winner_mode)
484 txfm_params->tx_size_search_method =
485 winner_mode_params->tx_size_search_methods[WINNER_MODE_EVAL];
486 else
487 txfm_params->tx_size_search_method =
488 winner_mode_params->tx_size_search_methods[MODE_EVAL];
489 }
490 txfm_params->tx_mode_search_type =
491 select_tx_mode(cm, txfm_params->tx_size_search_method);
492 }
493
set_tx_type_prune(const SPEED_FEATURES * sf,TxfmSearchParams * txfm_params,int winner_mode_tx_type_pruning,int is_winner_mode)494 static INLINE void set_tx_type_prune(const SPEED_FEATURES *sf,
495 TxfmSearchParams *txfm_params,
496 int winner_mode_tx_type_pruning,
497 int is_winner_mode) {
498 // Populate prune transform mode appropriately
499 txfm_params->prune_2d_txfm_mode = sf->tx_sf.tx_type_search.prune_2d_txfm_mode;
500 if (!winner_mode_tx_type_pruning) return;
501
502 const int prune_mode[4][2] = { { TX_TYPE_PRUNE_3, TX_TYPE_PRUNE_0 },
503 { TX_TYPE_PRUNE_4, TX_TYPE_PRUNE_0 },
504 { TX_TYPE_PRUNE_5, TX_TYPE_PRUNE_2 },
505 { TX_TYPE_PRUNE_5, TX_TYPE_PRUNE_3 } };
506 txfm_params->prune_2d_txfm_mode =
507 prune_mode[winner_mode_tx_type_pruning - 1][is_winner_mode];
508 }
509
set_tx_domain_dist_params(const WinnerModeParams * winner_mode_params,TxfmSearchParams * txfm_params,int enable_winner_mode_for_tx_domain_dist,int is_winner_mode)510 static INLINE void set_tx_domain_dist_params(
511 const WinnerModeParams *winner_mode_params, TxfmSearchParams *txfm_params,
512 int enable_winner_mode_for_tx_domain_dist, int is_winner_mode) {
513 if (txfm_params->use_qm_dist_metric) {
514 // QM-weighted PSNR is computed in transform space, so we need to forcibly
515 // enable the use of tx domain distortion.
516 txfm_params->use_transform_domain_distortion = 1;
517 txfm_params->tx_domain_dist_threshold = 0;
518 return;
519 }
520
521 if (!enable_winner_mode_for_tx_domain_dist) {
522 txfm_params->use_transform_domain_distortion =
523 winner_mode_params->use_transform_domain_distortion[DEFAULT_EVAL];
524 txfm_params->tx_domain_dist_threshold =
525 winner_mode_params->tx_domain_dist_threshold[DEFAULT_EVAL];
526 return;
527 }
528
529 if (is_winner_mode) {
530 txfm_params->use_transform_domain_distortion =
531 winner_mode_params->use_transform_domain_distortion[WINNER_MODE_EVAL];
532 txfm_params->tx_domain_dist_threshold =
533 winner_mode_params->tx_domain_dist_threshold[WINNER_MODE_EVAL];
534 } else {
535 txfm_params->use_transform_domain_distortion =
536 winner_mode_params->use_transform_domain_distortion[MODE_EVAL];
537 txfm_params->tx_domain_dist_threshold =
538 winner_mode_params->tx_domain_dist_threshold[MODE_EVAL];
539 }
540 }
541
542 // This function sets mode parameters for different mode evaluation stages
set_mode_eval_params(const struct AV1_COMP * cpi,MACROBLOCK * x,MODE_EVAL_TYPE mode_eval_type)543 static INLINE void set_mode_eval_params(const struct AV1_COMP *cpi,
544 MACROBLOCK *x,
545 MODE_EVAL_TYPE mode_eval_type) {
546 const AV1_COMMON *cm = &cpi->common;
547 const SPEED_FEATURES *sf = &cpi->sf;
548 const WinnerModeParams *winner_mode_params = &cpi->winner_mode_params;
549 TxfmSearchParams *txfm_params = &x->txfm_search_params;
550
551 txfm_params->use_qm_dist_metric =
552 cpi->oxcf.tune_cfg.dist_metric == AOM_DIST_METRIC_QM_PSNR;
553
554 switch (mode_eval_type) {
555 case DEFAULT_EVAL:
556 txfm_params->default_inter_tx_type_prob_thresh = INT_MAX;
557 txfm_params->use_default_intra_tx_type = 0;
558 txfm_params->skip_txfm_level =
559 winner_mode_params->skip_txfm_level[DEFAULT_EVAL];
560 txfm_params->predict_dc_level =
561 winner_mode_params->predict_dc_level[DEFAULT_EVAL];
562 // Set default transform domain distortion type
563 set_tx_domain_dist_params(winner_mode_params, txfm_params, 0, 0);
564
565 // Get default threshold for R-D optimization of coefficients
566 get_rd_opt_coeff_thresh(winner_mode_params->coeff_opt_thresholds,
567 txfm_params, 0, 0);
568
569 // Set default transform size search method
570 set_tx_size_search_method(cm, winner_mode_params, txfm_params, 0, 0);
571 // Set default transform type prune
572 set_tx_type_prune(sf, txfm_params, 0, 0);
573 break;
574 case MODE_EVAL:
575 txfm_params->use_default_intra_tx_type =
576 (cpi->sf.tx_sf.tx_type_search.fast_intra_tx_type_search ||
577 cpi->oxcf.txfm_cfg.use_intra_default_tx_only);
578 txfm_params->default_inter_tx_type_prob_thresh =
579 cpi->sf.tx_sf.tx_type_search.fast_inter_tx_type_prob_thresh;
580 txfm_params->skip_txfm_level =
581 winner_mode_params->skip_txfm_level[MODE_EVAL];
582 txfm_params->predict_dc_level =
583 winner_mode_params->predict_dc_level[MODE_EVAL];
584 // Set transform domain distortion type for mode evaluation
585 set_tx_domain_dist_params(
586 winner_mode_params, txfm_params,
587 sf->winner_mode_sf.enable_winner_mode_for_use_tx_domain_dist, 0);
588
589 // Get threshold for R-D optimization of coefficients during mode
590 // evaluation
591 get_rd_opt_coeff_thresh(
592 winner_mode_params->coeff_opt_thresholds, txfm_params,
593 sf->winner_mode_sf.enable_winner_mode_for_coeff_opt, 0);
594
595 // Set the transform size search method for mode evaluation
596 set_tx_size_search_method(
597 cm, winner_mode_params, txfm_params,
598 sf->winner_mode_sf.enable_winner_mode_for_tx_size_srch, 0);
599 // Set transform type prune for mode evaluation
600 set_tx_type_prune(sf, txfm_params,
601 sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning,
602 0);
603 break;
604 case WINNER_MODE_EVAL:
605 txfm_params->default_inter_tx_type_prob_thresh = INT_MAX;
606 txfm_params->use_default_intra_tx_type = 0;
607 txfm_params->skip_txfm_level =
608 winner_mode_params->skip_txfm_level[WINNER_MODE_EVAL];
609 txfm_params->predict_dc_level =
610 winner_mode_params->predict_dc_level[WINNER_MODE_EVAL];
611
612 // Set transform domain distortion type for winner mode evaluation
613 set_tx_domain_dist_params(
614 winner_mode_params, txfm_params,
615 sf->winner_mode_sf.enable_winner_mode_for_use_tx_domain_dist, 1);
616
617 // Get threshold for R-D optimization of coefficients for winner mode
618 // evaluation
619 get_rd_opt_coeff_thresh(
620 winner_mode_params->coeff_opt_thresholds, txfm_params,
621 sf->winner_mode_sf.enable_winner_mode_for_coeff_opt, 1);
622
623 // Set the transform size search method for winner mode evaluation
624 set_tx_size_search_method(
625 cm, winner_mode_params, txfm_params,
626 sf->winner_mode_sf.enable_winner_mode_for_tx_size_srch, 1);
627 // Set default transform type prune mode for winner mode evaluation
628 set_tx_type_prune(sf, txfm_params,
629 sf->tx_sf.tx_type_search.winner_mode_tx_type_pruning,
630 1);
631 break;
632 default: assert(0);
633 }
634
635 // Rd record collected at a specific mode evaluation stage can not be used
636 // across other evaluation stages as the transform parameters are different.
637 // Hence, reset mb rd record whenever mode evaluation stage type changes.
638 if (txfm_params->mode_eval_type != mode_eval_type)
639 reset_mb_rd_record(x->txfm_search_info.mb_rd_record);
640
641 txfm_params->mode_eval_type = mode_eval_type;
642 }
643
644 // Similar to store_cfl_required(), but for use during the RDO process,
645 // where we haven't yet determined whether this block uses CfL.
store_cfl_required_rdo(const AV1_COMMON * cm,const MACROBLOCK * x)646 static INLINE CFL_ALLOWED_TYPE store_cfl_required_rdo(const AV1_COMMON *cm,
647 const MACROBLOCK *x) {
648 const MACROBLOCKD *xd = &x->e_mbd;
649
650 if (cm->seq_params->monochrome || !xd->is_chroma_ref) return CFL_DISALLOWED;
651
652 if (!xd->is_chroma_ref) {
653 // For non-chroma-reference blocks, we should always store the luma pixels,
654 // in case the corresponding chroma-reference block uses CfL.
655 // Note that this can only happen for block sizes which are <8 on
656 // their shortest side, as otherwise they would be chroma reference
657 // blocks.
658 return CFL_ALLOWED;
659 }
660
661 // For chroma reference blocks, we should store data in the encoder iff we're
662 // allowed to try out CfL.
663 return is_cfl_allowed(xd);
664 }
665
init_sbuv_mode(MB_MODE_INFO * const mbmi)666 static AOM_INLINE void init_sbuv_mode(MB_MODE_INFO *const mbmi) {
667 mbmi->uv_mode = UV_DC_PRED;
668 mbmi->palette_mode_info.palette_size[1] = 0;
669 }
670
671 // Store best mode stats for winner mode processing
store_winner_mode_stats(const AV1_COMMON * const cm,MACROBLOCK * x,const MB_MODE_INFO * mbmi,RD_STATS * rd_cost,RD_STATS * rd_cost_y,RD_STATS * rd_cost_uv,THR_MODES mode_index,uint8_t * color_map,BLOCK_SIZE bsize,int64_t this_rd,int multi_winner_mode_type,int txfm_search_done)672 static INLINE void store_winner_mode_stats(
673 const AV1_COMMON *const cm, MACROBLOCK *x, const MB_MODE_INFO *mbmi,
674 RD_STATS *rd_cost, RD_STATS *rd_cost_y, RD_STATS *rd_cost_uv,
675 THR_MODES mode_index, uint8_t *color_map, BLOCK_SIZE bsize, int64_t this_rd,
676 int multi_winner_mode_type, int txfm_search_done) {
677 WinnerModeStats *winner_mode_stats = x->winner_mode_stats;
678 int mode_idx = 0;
679 int is_palette_mode = mbmi->palette_mode_info.palette_size[PLANE_TYPE_Y] > 0;
680 // Mode stat is not required when multiwinner mode processing is disabled
681 if (multi_winner_mode_type == MULTI_WINNER_MODE_OFF) return;
682 // Ignore mode with maximum rd
683 if (this_rd == INT64_MAX) return;
684 // TODO(any): Winner mode processing is currently not applicable for palette
685 // mode in Inter frames. Clean-up the following code, once support is added
686 if (!frame_is_intra_only(cm) && is_palette_mode) return;
687
688 int max_winner_mode_count = winner_mode_count_allowed[multi_winner_mode_type];
689 assert(x->winner_mode_count >= 0 &&
690 x->winner_mode_count <= max_winner_mode_count);
691
692 if (x->winner_mode_count) {
693 // Find the mode which has higher rd cost than this_rd
694 for (mode_idx = 0; mode_idx < x->winner_mode_count; mode_idx++)
695 if (winner_mode_stats[mode_idx].rd > this_rd) break;
696
697 if (mode_idx == max_winner_mode_count) {
698 // No mode has higher rd cost than this_rd
699 return;
700 } else if (mode_idx < max_winner_mode_count - 1) {
701 // Create a slot for current mode and move others to the next slot
702 memmove(
703 &winner_mode_stats[mode_idx + 1], &winner_mode_stats[mode_idx],
704 (max_winner_mode_count - mode_idx - 1) * sizeof(*winner_mode_stats));
705 }
706 }
707 // Add a mode stat for winner mode processing
708 winner_mode_stats[mode_idx].mbmi = *mbmi;
709 winner_mode_stats[mode_idx].rd = this_rd;
710 winner_mode_stats[mode_idx].mode_index = mode_index;
711
712 // Update rd stats required for inter frame
713 if (!frame_is_intra_only(cm) && rd_cost && rd_cost_y && rd_cost_uv) {
714 const MACROBLOCKD *xd = &x->e_mbd;
715 const int skip_ctx = av1_get_skip_txfm_context(xd);
716 const int is_intra_mode = av1_mode_defs[mode_index].mode < INTRA_MODE_END;
717 const int skip_txfm = mbmi->skip_txfm && !is_intra_mode;
718
719 winner_mode_stats[mode_idx].rd_cost = *rd_cost;
720 if (txfm_search_done) {
721 winner_mode_stats[mode_idx].rate_y =
722 rd_cost_y->rate +
723 x->mode_costs
724 .skip_txfm_cost[skip_ctx][rd_cost->skip_txfm || skip_txfm];
725 winner_mode_stats[mode_idx].rate_uv = rd_cost_uv->rate;
726 }
727 }
728
729 if (color_map) {
730 // Store color_index_map for palette mode
731 const MACROBLOCKD *const xd = &x->e_mbd;
732 int block_width, block_height;
733 av1_get_block_dimensions(bsize, AOM_PLANE_Y, xd, &block_width,
734 &block_height, NULL, NULL);
735 memcpy(winner_mode_stats[mode_idx].color_index_map, color_map,
736 block_width * block_height * sizeof(color_map[0]));
737 }
738
739 x->winner_mode_count =
740 AOMMIN(x->winner_mode_count + 1, max_winner_mode_count);
741 }
742
743 unsigned int av1_get_perpixel_variance(const AV1_COMP *cpi,
744 const MACROBLOCKD *xd,
745 const struct buf_2d *ref,
746 BLOCK_SIZE bsize, int plane,
747 int use_hbd);
748
749 unsigned int av1_get_perpixel_variance_facade(const struct AV1_COMP *cpi,
750 const MACROBLOCKD *xd,
751 const struct buf_2d *ref,
752 BLOCK_SIZE bsize, int plane);
753
is_mode_intra(PREDICTION_MODE mode)754 static INLINE int is_mode_intra(PREDICTION_MODE mode) {
755 return mode < INTRA_MODE_END;
756 }
757
758 // This function will copy usable ref_mv_stack[ref_frame][4] and
759 // weight[ref_frame][4] information from ref_mv_stack[ref_frame][8] and
760 // weight[ref_frame][8].
av1_copy_usable_ref_mv_stack_and_weight(const MACROBLOCKD * xd,MB_MODE_INFO_EXT * const mbmi_ext,MV_REFERENCE_FRAME ref_frame)761 static INLINE void av1_copy_usable_ref_mv_stack_and_weight(
762 const MACROBLOCKD *xd, MB_MODE_INFO_EXT *const mbmi_ext,
763 MV_REFERENCE_FRAME ref_frame) {
764 memcpy(mbmi_ext->weight[ref_frame], xd->weight[ref_frame],
765 USABLE_REF_MV_STACK_SIZE * sizeof(xd->weight[0][0]));
766 memcpy(mbmi_ext->ref_mv_stack[ref_frame], xd->ref_mv_stack[ref_frame],
767 USABLE_REF_MV_STACK_SIZE * sizeof(xd->ref_mv_stack[0][0]));
768 }
769
770 #ifdef __cplusplus
771 } // extern "C"
772 #endif
773
774 #endif // AOM_AV1_ENCODER_RDOPT_UTILS_H_
775