1 /*
2 * Copyright (c) 2020, 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 #include "av1/common/common_data.h"
13 #include "av1/common/quant_common.h"
14 #include "av1/common/reconintra.h"
15
16 #include "av1/encoder/encoder.h"
17 #include "av1/encoder/encodeframe_utils.h"
18 #include "av1/encoder/encoder_utils.h"
19 #include "av1/encoder/rdopt.h"
20
av1_set_ssim_rdmult(const AV1_COMP * const cpi,int * errorperbit,const BLOCK_SIZE bsize,const int mi_row,const int mi_col,int * const rdmult)21 void av1_set_ssim_rdmult(const AV1_COMP *const cpi, int *errorperbit,
22 const BLOCK_SIZE bsize, const int mi_row,
23 const int mi_col, int *const rdmult) {
24 const AV1_COMMON *const cm = &cpi->common;
25
26 const BLOCK_SIZE bsize_base = BLOCK_16X16;
27 const int num_mi_w = mi_size_wide[bsize_base];
28 const int num_mi_h = mi_size_high[bsize_base];
29 const int num_cols = (cm->mi_params.mi_cols + num_mi_w - 1) / num_mi_w;
30 const int num_rows = (cm->mi_params.mi_rows + num_mi_h - 1) / num_mi_h;
31 const int num_bcols = (mi_size_wide[bsize] + num_mi_w - 1) / num_mi_w;
32 const int num_brows = (mi_size_high[bsize] + num_mi_h - 1) / num_mi_h;
33 int row, col;
34 double num_of_mi = 0.0;
35 double geom_mean_of_scale = 1.0;
36
37 // To avoid overflow of 'geom_mean_of_scale', bsize_base must be at least
38 // BLOCK_8X8.
39 //
40 // For bsize=BLOCK_128X128 and bsize_base=BLOCK_8X8, the loop below would
41 // iterate 256 times. Considering the maximum value of
42 // cpi->ssim_rdmult_scaling_factors (see av1_set_mb_ssim_rdmult_scaling()),
43 // geom_mean_of_scale can go up to 4.8323^256, which is within DBL_MAX
44 // (maximum value a double data type can hold). If bsize_base is modified to
45 // BLOCK_4X4 (minimum possible block size), geom_mean_of_scale can go up
46 // to 4.8323^1024 and exceed DBL_MAX, resulting in data overflow.
47 assert(bsize_base >= BLOCK_8X8);
48 assert(cpi->oxcf.tune_cfg.tuning == AOM_TUNE_SSIM ||
49 cpi->oxcf.tune_cfg.tuning == AOM_TUNE_IQ);
50
51 for (row = mi_row / num_mi_w;
52 row < num_rows && row < mi_row / num_mi_w + num_brows; ++row) {
53 for (col = mi_col / num_mi_h;
54 col < num_cols && col < mi_col / num_mi_h + num_bcols; ++col) {
55 const int index = row * num_cols + col;
56 assert(cpi->ssim_rdmult_scaling_factors[index] != 0.0);
57 geom_mean_of_scale *= cpi->ssim_rdmult_scaling_factors[index];
58 num_of_mi += 1.0;
59 }
60 }
61 geom_mean_of_scale = pow(geom_mean_of_scale, (1.0 / num_of_mi));
62
63 *rdmult = (int)((double)(*rdmult) * geom_mean_of_scale + 0.5);
64 *rdmult = AOMMAX(*rdmult, 0);
65 av1_set_error_per_bit(errorperbit, *rdmult);
66 }
67
68 #if CONFIG_SALIENCY_MAP
av1_set_saliency_map_vmaf_rdmult(const AV1_COMP * const cpi,int * errorperbit,const BLOCK_SIZE bsize,const int mi_row,const int mi_col,int * const rdmult)69 void av1_set_saliency_map_vmaf_rdmult(const AV1_COMP *const cpi,
70 int *errorperbit, const BLOCK_SIZE bsize,
71 const int mi_row, const int mi_col,
72 int *const rdmult) {
73 const AV1_COMMON *const cm = &cpi->common;
74 const int num_mi_w = mi_size_wide[bsize];
75 const int num_mi_h = mi_size_high[bsize];
76 const int num_cols = (cm->mi_params.mi_cols + num_mi_w - 1) / num_mi_w;
77
78 *rdmult =
79 (int)(*rdmult * cpi->sm_scaling_factor[(mi_row / num_mi_h) * num_cols +
80 (mi_col / num_mi_w)]);
81
82 *rdmult = AOMMAX(*rdmult, 0);
83 av1_set_error_per_bit(errorperbit, *rdmult);
84 }
85 #endif
86
87 // TODO(angiebird): Move this function to tpl_model.c
88 #if !CONFIG_REALTIME_ONLY
av1_get_cb_rdmult(const AV1_COMP * const cpi,MACROBLOCK * const x,const BLOCK_SIZE bsize,const int mi_row,const int mi_col)89 int av1_get_cb_rdmult(const AV1_COMP *const cpi, MACROBLOCK *const x,
90 const BLOCK_SIZE bsize, const int mi_row,
91 const int mi_col) {
92 const AV1_COMMON *const cm = &cpi->common;
93 assert(IMPLIES(cpi->ppi->gf_group.size > 0,
94 cpi->gf_frame_index < cpi->ppi->gf_group.size));
95 const int tpl_idx = cpi->gf_frame_index;
96 int deltaq_rdmult = set_rdmult(cpi, x, -1);
97 if (!av1_tpl_stats_ready(&cpi->ppi->tpl_data, tpl_idx)) return deltaq_rdmult;
98 if (cm->superres_scale_denominator != SCALE_NUMERATOR) return deltaq_rdmult;
99 if (cpi->oxcf.q_cfg.aq_mode != NO_AQ) return deltaq_rdmult;
100 if (x->rb == 0) return deltaq_rdmult;
101
102 TplParams *const tpl_data = &cpi->ppi->tpl_data;
103 TplDepFrame *tpl_frame = &tpl_data->tpl_frame[tpl_idx];
104 TplDepStats *tpl_stats = tpl_frame->tpl_stats_ptr;
105
106 const int mi_wide = mi_size_wide[bsize];
107 const int mi_high = mi_size_high[bsize];
108
109 int tpl_stride = tpl_frame->stride;
110 double intra_cost_base = 0;
111 double mc_dep_cost_base = 0;
112 double cbcmp_base = 0;
113 const int step = 1 << tpl_data->tpl_stats_block_mis_log2;
114
115 for (int row = mi_row; row < mi_row + mi_high; row += step) {
116 for (int col = mi_col; col < mi_col + mi_wide; col += step) {
117 if (row >= cm->mi_params.mi_rows || col >= cm->mi_params.mi_cols)
118 continue;
119
120 TplDepStats *this_stats = &tpl_stats[av1_tpl_ptr_pos(
121 row, col, tpl_stride, tpl_data->tpl_stats_block_mis_log2)];
122
123 double cbcmp = (double)this_stats->srcrf_dist;
124 int64_t mc_dep_delta =
125 RDCOST(tpl_frame->base_rdmult, this_stats->mc_dep_rate,
126 this_stats->mc_dep_dist);
127 double dist_scaled = (double)(this_stats->recrf_dist << RDDIV_BITS);
128 intra_cost_base += log(dist_scaled) * cbcmp;
129 mc_dep_cost_base += log(3 * dist_scaled + mc_dep_delta) * cbcmp;
130 cbcmp_base += cbcmp;
131 }
132 }
133
134 if (cbcmp_base == 0) return deltaq_rdmult;
135
136 double rk = exp((intra_cost_base - mc_dep_cost_base) / cbcmp_base);
137 deltaq_rdmult = (int)(deltaq_rdmult * (rk / x->rb));
138
139 return AOMMAX(deltaq_rdmult, 1);
140 }
141 #endif // !CONFIG_REALTIME_ONLY
142
update_filter_type_count(FRAME_COUNTS * counts,const MACROBLOCKD * xd,const MB_MODE_INFO * mbmi)143 static inline void update_filter_type_count(FRAME_COUNTS *counts,
144 const MACROBLOCKD *xd,
145 const MB_MODE_INFO *mbmi) {
146 int dir;
147 for (dir = 0; dir < 2; ++dir) {
148 const int ctx = av1_get_pred_context_switchable_interp(xd, dir);
149 InterpFilter filter = av1_extract_interp_filter(mbmi->interp_filters, dir);
150
151 // Only allow the 3 valid SWITCHABLE_FILTERS.
152 assert(filter < SWITCHABLE_FILTERS);
153 ++counts->switchable_interp[ctx][filter];
154 }
155 }
156
157 // This function will copy the best reference mode information from
158 // MB_MODE_INFO_EXT_FRAME to MB_MODE_INFO_EXT.
copy_mbmi_ext_frame_to_mbmi_ext(MB_MODE_INFO_EXT * mbmi_ext,const MB_MODE_INFO_EXT_FRAME * const mbmi_ext_best,uint8_t ref_frame_type)159 static inline void copy_mbmi_ext_frame_to_mbmi_ext(
160 MB_MODE_INFO_EXT *mbmi_ext,
161 const MB_MODE_INFO_EXT_FRAME *const mbmi_ext_best, uint8_t ref_frame_type) {
162 memcpy(mbmi_ext->ref_mv_stack[ref_frame_type], mbmi_ext_best->ref_mv_stack,
163 sizeof(mbmi_ext->ref_mv_stack[USABLE_REF_MV_STACK_SIZE]));
164 memcpy(mbmi_ext->weight[ref_frame_type], mbmi_ext_best->weight,
165 sizeof(mbmi_ext->weight[USABLE_REF_MV_STACK_SIZE]));
166 mbmi_ext->mode_context[ref_frame_type] = mbmi_ext_best->mode_context;
167 mbmi_ext->ref_mv_count[ref_frame_type] = mbmi_ext_best->ref_mv_count;
168 memcpy(mbmi_ext->global_mvs, mbmi_ext_best->global_mvs,
169 sizeof(mbmi_ext->global_mvs));
170 }
171
av1_update_state(const AV1_COMP * const cpi,ThreadData * td,const PICK_MODE_CONTEXT * const ctx,int mi_row,int mi_col,BLOCK_SIZE bsize,RUN_TYPE dry_run)172 void av1_update_state(const AV1_COMP *const cpi, ThreadData *td,
173 const PICK_MODE_CONTEXT *const ctx, int mi_row,
174 int mi_col, BLOCK_SIZE bsize, RUN_TYPE dry_run) {
175 int i, x_idx, y;
176 const AV1_COMMON *const cm = &cpi->common;
177 const CommonModeInfoParams *const mi_params = &cm->mi_params;
178 const int num_planes = av1_num_planes(cm);
179 MACROBLOCK *const x = &td->mb;
180 MACROBLOCKD *const xd = &x->e_mbd;
181 struct macroblock_plane *const p = x->plane;
182 struct macroblockd_plane *const pd = xd->plane;
183 const MB_MODE_INFO *const mi = &ctx->mic;
184 MB_MODE_INFO *const mi_addr = xd->mi[0];
185 const struct segmentation *const seg = &cm->seg;
186 assert(bsize < BLOCK_SIZES_ALL);
187 const int bw = mi_size_wide[mi->bsize];
188 const int bh = mi_size_high[mi->bsize];
189 const int mis = mi_params->mi_stride;
190 const int mi_width = mi_size_wide[bsize];
191 const int mi_height = mi_size_high[bsize];
192 TxfmSearchInfo *txfm_info = &x->txfm_search_info;
193
194 assert(mi->bsize == bsize);
195
196 *mi_addr = *mi;
197 copy_mbmi_ext_frame_to_mbmi_ext(&x->mbmi_ext, &ctx->mbmi_ext_best,
198 av1_ref_frame_type(ctx->mic.ref_frame));
199
200 memcpy(txfm_info->blk_skip, ctx->blk_skip,
201 sizeof(txfm_info->blk_skip[0]) * ctx->num_4x4_blk);
202
203 txfm_info->skip_txfm = ctx->rd_stats.skip_txfm;
204
205 xd->tx_type_map = ctx->tx_type_map;
206 xd->tx_type_map_stride = mi_size_wide[bsize];
207 // If not dry_run, copy the transform type data into the frame level buffer.
208 // Encoder will fetch tx types when writing bitstream.
209 if (!dry_run) {
210 const int grid_idx = get_mi_grid_idx(mi_params, mi_row, mi_col);
211 uint8_t *const tx_type_map = mi_params->tx_type_map + grid_idx;
212 const int mi_stride = mi_params->mi_stride;
213 for (int blk_row = 0; blk_row < bh; ++blk_row) {
214 av1_copy_array(tx_type_map + blk_row * mi_stride,
215 xd->tx_type_map + blk_row * xd->tx_type_map_stride, bw);
216 }
217 xd->tx_type_map = tx_type_map;
218 xd->tx_type_map_stride = mi_stride;
219 }
220
221 // If segmentation in use
222 if (seg->enabled) {
223 // For in frame complexity AQ copy the segment id from the segment map.
224 if (cpi->oxcf.q_cfg.aq_mode == COMPLEXITY_AQ) {
225 const uint8_t *const map =
226 seg->update_map ? cpi->enc_seg.map : cm->last_frame_seg_map;
227 mi_addr->segment_id =
228 map ? get_segment_id(mi_params, map, bsize, mi_row, mi_col) : 0;
229 }
230 // Else for cyclic refresh mode update the segment map, set the segment id
231 // and then update the quantizer.
232 if (cpi->oxcf.q_cfg.aq_mode == CYCLIC_REFRESH_AQ &&
233 mi_addr->segment_id != AM_SEGMENT_ID_INACTIVE &&
234 !cpi->rc.rtc_external_ratectrl) {
235 av1_cyclic_refresh_update_segment(cpi, x, mi_row, mi_col, bsize,
236 ctx->rd_stats.rate, ctx->rd_stats.dist,
237 txfm_info->skip_txfm, dry_run);
238 }
239 if (mi_addr->uv_mode == UV_CFL_PRED && !is_cfl_allowed(xd))
240 mi_addr->uv_mode = UV_DC_PRED;
241
242 if (!dry_run && !mi_addr->skip_txfm) {
243 int cdf_num;
244 const uint8_t spatial_pred = av1_get_spatial_seg_pred(
245 cm, xd, &cdf_num, cpi->cyclic_refresh->skip_over4x4);
246 const uint8_t coded_id = av1_neg_interleave(
247 mi_addr->segment_id, spatial_pred, seg->last_active_segid + 1);
248 int64_t spatial_cost = x->mode_costs.spatial_pred_cost[cdf_num][coded_id];
249 td->rd_counts.seg_tmp_pred_cost[0] += spatial_cost;
250
251 const int pred_segment_id =
252 cm->last_frame_seg_map
253 ? get_segment_id(mi_params, cm->last_frame_seg_map, bsize, mi_row,
254 mi_col)
255 : 0;
256 const int use_tmp_pred = pred_segment_id == mi_addr->segment_id;
257 const uint8_t tmp_pred_ctx = av1_get_pred_context_seg_id(xd);
258 td->rd_counts.seg_tmp_pred_cost[1] +=
259 x->mode_costs.tmp_pred_cost[tmp_pred_ctx][use_tmp_pred];
260 if (!use_tmp_pred) {
261 td->rd_counts.seg_tmp_pred_cost[1] += spatial_cost;
262 }
263 }
264 }
265
266 // Count zero motion vector.
267 if (!dry_run && !frame_is_intra_only(cm)) {
268 const MV mv = mi->mv[0].as_mv;
269 if (is_inter_block(mi) && mi->ref_frame[0] == LAST_FRAME &&
270 abs(mv.row) < 8 && abs(mv.col) < 8) {
271 const int ymis = AOMMIN(cm->mi_params.mi_rows - mi_row, bh);
272 // Accumulate low_content_frame.
273 for (int mi_y = 0; mi_y < ymis; mi_y += 2) x->cnt_zeromv += bw << 1;
274 }
275 }
276
277 for (i = 0; i < num_planes; ++i) {
278 p[i].coeff = ctx->coeff[i];
279 p[i].qcoeff = ctx->qcoeff[i];
280 p[i].dqcoeff = ctx->dqcoeff[i];
281 p[i].eobs = ctx->eobs[i];
282 p[i].txb_entropy_ctx = ctx->txb_entropy_ctx[i];
283 }
284 for (i = 0; i < 2; ++i) pd[i].color_index_map = ctx->color_index_map[i];
285 // Restore the coding context of the MB to that that was in place
286 // when the mode was picked for it
287
288 const int cols =
289 AOMMIN((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width, mi_width);
290 const int rows = AOMMIN(
291 (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height, mi_height);
292 for (y = 0; y < rows; y++) {
293 for (x_idx = 0; x_idx < cols; x_idx++) xd->mi[x_idx + y * mis] = mi_addr;
294 }
295
296 if (cpi->oxcf.q_cfg.aq_mode)
297 av1_init_plane_quantizers(cpi, x, mi_addr->segment_id, 0);
298
299 if (dry_run) return;
300
301 #if CONFIG_INTERNAL_STATS
302 {
303 unsigned int *const mode_chosen_counts =
304 (unsigned int *)cpi->mode_chosen_counts; // Cast const away.
305 if (frame_is_intra_only(cm)) {
306 static const int kf_mode_index[] = {
307 THR_DC /*DC_PRED*/,
308 THR_V_PRED /*V_PRED*/,
309 THR_H_PRED /*H_PRED*/,
310 THR_D45_PRED /*D45_PRED*/,
311 THR_D135_PRED /*D135_PRED*/,
312 THR_D113_PRED /*D113_PRED*/,
313 THR_D157_PRED /*D157_PRED*/,
314 THR_D203_PRED /*D203_PRED*/,
315 THR_D67_PRED /*D67_PRED*/,
316 THR_SMOOTH, /*SMOOTH_PRED*/
317 THR_SMOOTH_V, /*SMOOTH_V_PRED*/
318 THR_SMOOTH_H, /*SMOOTH_H_PRED*/
319 THR_PAETH /*PAETH_PRED*/,
320 };
321 ++mode_chosen_counts[kf_mode_index[mi_addr->mode]];
322 } else {
323 // Note how often each mode chosen as best
324 ++mode_chosen_counts[ctx->best_mode_index];
325 }
326 }
327 #endif
328 if (!frame_is_intra_only(cm)) {
329 if (is_inter_block(mi) && cm->features.interp_filter == SWITCHABLE) {
330 // When the frame interp filter is SWITCHABLE, several cases that always
331 // use the default type (EIGHTTAP_REGULAR) are described in
332 // av1_is_interp_needed(). Here, we should keep the counts for all
333 // applicable blocks, so the frame filter resetting decision in
334 // fix_interp_filter() is made correctly.
335 update_filter_type_count(td->counts, xd, mi_addr);
336 }
337 }
338
339 const int x_mis = AOMMIN(bw, mi_params->mi_cols - mi_col);
340 const int y_mis = AOMMIN(bh, mi_params->mi_rows - mi_row);
341 if (cm->seq_params->order_hint_info.enable_ref_frame_mvs)
342 av1_copy_frame_mvs(cm, mi, mi_row, mi_col, x_mis, y_mis);
343 }
344
av1_update_inter_mode_stats(FRAME_CONTEXT * fc,FRAME_COUNTS * counts,PREDICTION_MODE mode,int16_t mode_context)345 void av1_update_inter_mode_stats(FRAME_CONTEXT *fc, FRAME_COUNTS *counts,
346 PREDICTION_MODE mode, int16_t mode_context) {
347 (void)counts;
348
349 int16_t mode_ctx = mode_context & NEWMV_CTX_MASK;
350 if (mode == NEWMV) {
351 #if CONFIG_ENTROPY_STATS
352 ++counts->newmv_mode[mode_ctx][0];
353 #endif
354 update_cdf(fc->newmv_cdf[mode_ctx], 0, 2);
355 return;
356 }
357
358 #if CONFIG_ENTROPY_STATS
359 ++counts->newmv_mode[mode_ctx][1];
360 #endif
361 update_cdf(fc->newmv_cdf[mode_ctx], 1, 2);
362
363 mode_ctx = (mode_context >> GLOBALMV_OFFSET) & GLOBALMV_CTX_MASK;
364 if (mode == GLOBALMV) {
365 #if CONFIG_ENTROPY_STATS
366 ++counts->zeromv_mode[mode_ctx][0];
367 #endif
368 update_cdf(fc->zeromv_cdf[mode_ctx], 0, 2);
369 return;
370 }
371
372 #if CONFIG_ENTROPY_STATS
373 ++counts->zeromv_mode[mode_ctx][1];
374 #endif
375 update_cdf(fc->zeromv_cdf[mode_ctx], 1, 2);
376
377 mode_ctx = (mode_context >> REFMV_OFFSET) & REFMV_CTX_MASK;
378 #if CONFIG_ENTROPY_STATS
379 ++counts->refmv_mode[mode_ctx][mode != NEARESTMV];
380 #endif
381 update_cdf(fc->refmv_cdf[mode_ctx], mode != NEARESTMV, 2);
382 }
383
update_palette_cdf(MACROBLOCKD * xd,const MB_MODE_INFO * const mbmi,FRAME_COUNTS * counts)384 static void update_palette_cdf(MACROBLOCKD *xd, const MB_MODE_INFO *const mbmi,
385 FRAME_COUNTS *counts) {
386 FRAME_CONTEXT *fc = xd->tile_ctx;
387 const BLOCK_SIZE bsize = mbmi->bsize;
388 const PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
389 const int palette_bsize_ctx = av1_get_palette_bsize_ctx(bsize);
390
391 (void)counts;
392
393 if (mbmi->mode == DC_PRED) {
394 const int n = pmi->palette_size[0];
395 const int palette_mode_ctx = av1_get_palette_mode_ctx(xd);
396
397 #if CONFIG_ENTROPY_STATS
398 ++counts->palette_y_mode[palette_bsize_ctx][palette_mode_ctx][n > 0];
399 #endif
400 update_cdf(fc->palette_y_mode_cdf[palette_bsize_ctx][palette_mode_ctx],
401 n > 0, 2);
402 if (n > 0) {
403 #if CONFIG_ENTROPY_STATS
404 ++counts->palette_y_size[palette_bsize_ctx][n - PALETTE_MIN_SIZE];
405 #endif
406 update_cdf(fc->palette_y_size_cdf[palette_bsize_ctx],
407 n - PALETTE_MIN_SIZE, PALETTE_SIZES);
408 }
409 }
410
411 if (mbmi->uv_mode == UV_DC_PRED) {
412 const int n = pmi->palette_size[1];
413 const int palette_uv_mode_ctx = (pmi->palette_size[0] > 0);
414
415 #if CONFIG_ENTROPY_STATS
416 ++counts->palette_uv_mode[palette_uv_mode_ctx][n > 0];
417 #endif
418 update_cdf(fc->palette_uv_mode_cdf[palette_uv_mode_ctx], n > 0, 2);
419
420 if (n > 0) {
421 #if CONFIG_ENTROPY_STATS
422 ++counts->palette_uv_size[palette_bsize_ctx][n - PALETTE_MIN_SIZE];
423 #endif
424 update_cdf(fc->palette_uv_size_cdf[palette_bsize_ctx],
425 n - PALETTE_MIN_SIZE, PALETTE_SIZES);
426 }
427 }
428 }
429
av1_sum_intra_stats(const AV1_COMMON * const cm,FRAME_COUNTS * counts,MACROBLOCKD * xd,const MB_MODE_INFO * const mbmi,const MB_MODE_INFO * above_mi,const MB_MODE_INFO * left_mi,const int intraonly)430 void av1_sum_intra_stats(const AV1_COMMON *const cm, FRAME_COUNTS *counts,
431 MACROBLOCKD *xd, const MB_MODE_INFO *const mbmi,
432 const MB_MODE_INFO *above_mi,
433 const MB_MODE_INFO *left_mi, const int intraonly) {
434 FRAME_CONTEXT *fc = xd->tile_ctx;
435 const PREDICTION_MODE y_mode = mbmi->mode;
436 (void)counts;
437 const BLOCK_SIZE bsize = mbmi->bsize;
438
439 if (intraonly) {
440 #if CONFIG_ENTROPY_STATS
441 const PREDICTION_MODE above = av1_above_block_mode(above_mi);
442 const PREDICTION_MODE left = av1_left_block_mode(left_mi);
443 const int above_ctx = intra_mode_context[above];
444 const int left_ctx = intra_mode_context[left];
445 ++counts->kf_y_mode[above_ctx][left_ctx][y_mode];
446 #endif // CONFIG_ENTROPY_STATS
447 update_cdf(get_y_mode_cdf(fc, above_mi, left_mi), y_mode, INTRA_MODES);
448 } else {
449 #if CONFIG_ENTROPY_STATS
450 ++counts->y_mode[size_group_lookup[bsize]][y_mode];
451 #endif // CONFIG_ENTROPY_STATS
452 update_cdf(fc->y_mode_cdf[size_group_lookup[bsize]], y_mode, INTRA_MODES);
453 }
454
455 if (av1_filter_intra_allowed(cm, mbmi)) {
456 const int use_filter_intra_mode =
457 mbmi->filter_intra_mode_info.use_filter_intra;
458 #if CONFIG_ENTROPY_STATS
459 ++counts->filter_intra[mbmi->bsize][use_filter_intra_mode];
460 if (use_filter_intra_mode) {
461 ++counts
462 ->filter_intra_mode[mbmi->filter_intra_mode_info.filter_intra_mode];
463 }
464 #endif // CONFIG_ENTROPY_STATS
465 update_cdf(fc->filter_intra_cdfs[mbmi->bsize], use_filter_intra_mode, 2);
466 if (use_filter_intra_mode) {
467 update_cdf(fc->filter_intra_mode_cdf,
468 mbmi->filter_intra_mode_info.filter_intra_mode,
469 FILTER_INTRA_MODES);
470 }
471 }
472 if (av1_is_directional_mode(mbmi->mode) && av1_use_angle_delta(bsize)) {
473 #if CONFIG_ENTROPY_STATS
474 ++counts->angle_delta[mbmi->mode - V_PRED]
475 [mbmi->angle_delta[PLANE_TYPE_Y] + MAX_ANGLE_DELTA];
476 #endif
477 update_cdf(fc->angle_delta_cdf[mbmi->mode - V_PRED],
478 mbmi->angle_delta[PLANE_TYPE_Y] + MAX_ANGLE_DELTA,
479 2 * MAX_ANGLE_DELTA + 1);
480 }
481
482 if (!xd->is_chroma_ref) return;
483
484 const UV_PREDICTION_MODE uv_mode = mbmi->uv_mode;
485 const CFL_ALLOWED_TYPE cfl_allowed = is_cfl_allowed(xd);
486 #if CONFIG_ENTROPY_STATS
487 ++counts->uv_mode[cfl_allowed][y_mode][uv_mode];
488 #endif // CONFIG_ENTROPY_STATS
489 update_cdf(fc->uv_mode_cdf[cfl_allowed][y_mode], uv_mode,
490 UV_INTRA_MODES - !cfl_allowed);
491 if (uv_mode == UV_CFL_PRED) {
492 const int8_t joint_sign = mbmi->cfl_alpha_signs;
493 const uint8_t idx = mbmi->cfl_alpha_idx;
494
495 #if CONFIG_ENTROPY_STATS
496 ++counts->cfl_sign[joint_sign];
497 #endif
498 update_cdf(fc->cfl_sign_cdf, joint_sign, CFL_JOINT_SIGNS);
499 if (CFL_SIGN_U(joint_sign) != CFL_SIGN_ZERO) {
500 aom_cdf_prob *cdf_u = fc->cfl_alpha_cdf[CFL_CONTEXT_U(joint_sign)];
501
502 #if CONFIG_ENTROPY_STATS
503 ++counts->cfl_alpha[CFL_CONTEXT_U(joint_sign)][CFL_IDX_U(idx)];
504 #endif
505 update_cdf(cdf_u, CFL_IDX_U(idx), CFL_ALPHABET_SIZE);
506 }
507 if (CFL_SIGN_V(joint_sign) != CFL_SIGN_ZERO) {
508 aom_cdf_prob *cdf_v = fc->cfl_alpha_cdf[CFL_CONTEXT_V(joint_sign)];
509
510 #if CONFIG_ENTROPY_STATS
511 ++counts->cfl_alpha[CFL_CONTEXT_V(joint_sign)][CFL_IDX_V(idx)];
512 #endif
513 update_cdf(cdf_v, CFL_IDX_V(idx), CFL_ALPHABET_SIZE);
514 }
515 }
516 const PREDICTION_MODE intra_mode = get_uv_mode(uv_mode);
517 if (av1_is_directional_mode(intra_mode) && av1_use_angle_delta(bsize)) {
518 #if CONFIG_ENTROPY_STATS
519 ++counts->angle_delta[intra_mode - V_PRED]
520 [mbmi->angle_delta[PLANE_TYPE_UV] + MAX_ANGLE_DELTA];
521 #endif
522 update_cdf(fc->angle_delta_cdf[intra_mode - V_PRED],
523 mbmi->angle_delta[PLANE_TYPE_UV] + MAX_ANGLE_DELTA,
524 2 * MAX_ANGLE_DELTA + 1);
525 }
526 if (av1_allow_palette(cm->features.allow_screen_content_tools, bsize)) {
527 update_palette_cdf(xd, mbmi, counts);
528 }
529 }
530
av1_restore_context(MACROBLOCK * x,const RD_SEARCH_MACROBLOCK_CONTEXT * ctx,int mi_row,int mi_col,BLOCK_SIZE bsize,const int num_planes)531 void av1_restore_context(MACROBLOCK *x, const RD_SEARCH_MACROBLOCK_CONTEXT *ctx,
532 int mi_row, int mi_col, BLOCK_SIZE bsize,
533 const int num_planes) {
534 MACROBLOCKD *xd = &x->e_mbd;
535 int p;
536 const int num_4x4_blocks_wide = mi_size_wide[bsize];
537 const int num_4x4_blocks_high = mi_size_high[bsize];
538 int mi_width = mi_size_wide[bsize];
539 int mi_height = mi_size_high[bsize];
540 for (p = 0; p < num_planes; p++) {
541 int tx_col = mi_col;
542 int tx_row = mi_row & MAX_MIB_MASK;
543 memcpy(
544 xd->above_entropy_context[p] + (tx_col >> xd->plane[p].subsampling_x),
545 ctx->a + num_4x4_blocks_wide * p,
546 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
547 xd->plane[p].subsampling_x);
548 memcpy(xd->left_entropy_context[p] + (tx_row >> xd->plane[p].subsampling_y),
549 ctx->l + num_4x4_blocks_high * p,
550 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
551 xd->plane[p].subsampling_y);
552 }
553 memcpy(xd->above_partition_context + mi_col, ctx->sa,
554 sizeof(*xd->above_partition_context) * mi_width);
555 memcpy(xd->left_partition_context + (mi_row & MAX_MIB_MASK), ctx->sl,
556 sizeof(xd->left_partition_context[0]) * mi_height);
557 xd->above_txfm_context = ctx->p_ta;
558 xd->left_txfm_context = ctx->p_tl;
559 memcpy(xd->above_txfm_context, ctx->ta,
560 sizeof(*xd->above_txfm_context) * mi_width);
561 memcpy(xd->left_txfm_context, ctx->tl,
562 sizeof(*xd->left_txfm_context) * mi_height);
563 }
564
av1_save_context(const MACROBLOCK * x,RD_SEARCH_MACROBLOCK_CONTEXT * ctx,int mi_row,int mi_col,BLOCK_SIZE bsize,const int num_planes)565 void av1_save_context(const MACROBLOCK *x, RD_SEARCH_MACROBLOCK_CONTEXT *ctx,
566 int mi_row, int mi_col, BLOCK_SIZE bsize,
567 const int num_planes) {
568 const MACROBLOCKD *xd = &x->e_mbd;
569 int p;
570 int mi_width = mi_size_wide[bsize];
571 int mi_height = mi_size_high[bsize];
572
573 // buffer the above/left context information of the block in search.
574 for (p = 0; p < num_planes; ++p) {
575 int tx_col = mi_col;
576 int tx_row = mi_row & MAX_MIB_MASK;
577 memcpy(
578 ctx->a + mi_width * p,
579 xd->above_entropy_context[p] + (tx_col >> xd->plane[p].subsampling_x),
580 (sizeof(ENTROPY_CONTEXT) * mi_width) >> xd->plane[p].subsampling_x);
581 memcpy(ctx->l + mi_height * p,
582 xd->left_entropy_context[p] + (tx_row >> xd->plane[p].subsampling_y),
583 (sizeof(ENTROPY_CONTEXT) * mi_height) >> xd->plane[p].subsampling_y);
584 }
585 memcpy(ctx->sa, xd->above_partition_context + mi_col,
586 sizeof(*xd->above_partition_context) * mi_width);
587 memcpy(ctx->sl, xd->left_partition_context + (mi_row & MAX_MIB_MASK),
588 sizeof(xd->left_partition_context[0]) * mi_height);
589 memcpy(ctx->ta, xd->above_txfm_context,
590 sizeof(*xd->above_txfm_context) * mi_width);
591 memcpy(ctx->tl, xd->left_txfm_context,
592 sizeof(*xd->left_txfm_context) * mi_height);
593 ctx->p_ta = xd->above_txfm_context;
594 ctx->p_tl = xd->left_txfm_context;
595 }
596
set_partial_sb_partition(const AV1_COMMON * const cm,MB_MODE_INFO * mi,int bh_in,int bw_in,int mi_rows_remaining,int mi_cols_remaining,BLOCK_SIZE bsize,MB_MODE_INFO ** mib)597 static void set_partial_sb_partition(const AV1_COMMON *const cm,
598 MB_MODE_INFO *mi, int bh_in, int bw_in,
599 int mi_rows_remaining,
600 int mi_cols_remaining, BLOCK_SIZE bsize,
601 MB_MODE_INFO **mib) {
602 int bh = bh_in;
603 int r, c;
604 for (r = 0; r < cm->seq_params->mib_size; r += bh) {
605 int bw = bw_in;
606 for (c = 0; c < cm->seq_params->mib_size; c += bw) {
607 const int grid_index = get_mi_grid_idx(&cm->mi_params, r, c);
608 const int mi_index = get_alloc_mi_idx(&cm->mi_params, r, c);
609 mib[grid_index] = mi + mi_index;
610 mib[grid_index]->bsize = find_partition_size(
611 bsize, mi_rows_remaining - r, mi_cols_remaining - c, &bh, &bw);
612 }
613 }
614 }
615
616 // This function attempts to set all mode info entries in a given superblock
617 // to the same block partition size.
618 // However, at the bottom and right borders of the image the requested size
619 // may not be allowed in which case this code attempts to choose the largest
620 // allowable partition.
av1_set_fixed_partitioning(AV1_COMP * cpi,const TileInfo * const tile,MB_MODE_INFO ** mib,int mi_row,int mi_col,BLOCK_SIZE bsize)621 void av1_set_fixed_partitioning(AV1_COMP *cpi, const TileInfo *const tile,
622 MB_MODE_INFO **mib, int mi_row, int mi_col,
623 BLOCK_SIZE bsize) {
624 AV1_COMMON *const cm = &cpi->common;
625 const CommonModeInfoParams *const mi_params = &cm->mi_params;
626 const int mi_rows_remaining = tile->mi_row_end - mi_row;
627 const int mi_cols_remaining = tile->mi_col_end - mi_col;
628 MB_MODE_INFO *const mi_upper_left =
629 mi_params->mi_alloc + get_alloc_mi_idx(mi_params, mi_row, mi_col);
630 int bh = mi_size_high[bsize];
631 int bw = mi_size_wide[bsize];
632
633 assert(bsize >= mi_params->mi_alloc_bsize &&
634 "Attempted to use bsize < mi_params->mi_alloc_bsize");
635 assert((mi_rows_remaining > 0) && (mi_cols_remaining > 0));
636
637 // Apply the requested partition size to the SB if it is all "in image"
638 if ((mi_cols_remaining >= cm->seq_params->mib_size) &&
639 (mi_rows_remaining >= cm->seq_params->mib_size)) {
640 for (int block_row = 0; block_row < cm->seq_params->mib_size;
641 block_row += bh) {
642 for (int block_col = 0; block_col < cm->seq_params->mib_size;
643 block_col += bw) {
644 const int grid_index = get_mi_grid_idx(mi_params, block_row, block_col);
645 const int mi_index = get_alloc_mi_idx(mi_params, block_row, block_col);
646 mib[grid_index] = mi_upper_left + mi_index;
647 mib[grid_index]->bsize = bsize;
648 }
649 }
650 } else {
651 // Else this is a partial SB.
652 set_partial_sb_partition(cm, mi_upper_left, bh, bw, mi_rows_remaining,
653 mi_cols_remaining, bsize, mib);
654 }
655 }
656
av1_is_leaf_split_partition(AV1_COMMON * cm,int mi_row,int mi_col,BLOCK_SIZE bsize)657 int av1_is_leaf_split_partition(AV1_COMMON *cm, int mi_row, int mi_col,
658 BLOCK_SIZE bsize) {
659 const int bs = mi_size_wide[bsize];
660 const int hbs = bs / 2;
661 assert(bsize >= BLOCK_8X8);
662 const BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_SPLIT);
663
664 for (int i = 0; i < 4; i++) {
665 int x_idx = (i & 1) * hbs;
666 int y_idx = (i >> 1) * hbs;
667 if ((mi_row + y_idx >= cm->mi_params.mi_rows) ||
668 (mi_col + x_idx >= cm->mi_params.mi_cols))
669 return 0;
670 if (get_partition(cm, mi_row + y_idx, mi_col + x_idx, subsize) !=
671 PARTITION_NONE &&
672 subsize != BLOCK_8X8)
673 return 0;
674 }
675 return 1;
676 }
677
678 #if !CONFIG_REALTIME_ONLY
av1_get_rdmult_delta(AV1_COMP * cpi,BLOCK_SIZE bsize,int mi_row,int mi_col,int orig_rdmult)679 int av1_get_rdmult_delta(AV1_COMP *cpi, BLOCK_SIZE bsize, int mi_row,
680 int mi_col, int orig_rdmult) {
681 AV1_COMMON *const cm = &cpi->common;
682 const GF_GROUP *const gf_group = &cpi->ppi->gf_group;
683 assert(IMPLIES(cpi->ppi->gf_group.size > 0,
684 cpi->gf_frame_index < cpi->ppi->gf_group.size));
685 const int tpl_idx = cpi->gf_frame_index;
686 TplParams *const tpl_data = &cpi->ppi->tpl_data;
687 const uint8_t block_mis_log2 = tpl_data->tpl_stats_block_mis_log2;
688 int64_t intra_cost = 0;
689 int64_t mc_dep_cost = 0;
690 const int mi_wide = mi_size_wide[bsize];
691 const int mi_high = mi_size_high[bsize];
692
693 TplDepFrame *tpl_frame = &tpl_data->tpl_frame[tpl_idx];
694 TplDepStats *tpl_stats = tpl_frame->tpl_stats_ptr;
695 int tpl_stride = tpl_frame->stride;
696
697 if (!av1_tpl_stats_ready(&cpi->ppi->tpl_data, cpi->gf_frame_index)) {
698 return orig_rdmult;
699 }
700 if (!is_frame_tpl_eligible(gf_group, cpi->gf_frame_index)) {
701 return orig_rdmult;
702 }
703
704 #ifndef NDEBUG
705 int mi_count = 0;
706 #endif
707 const int mi_col_sr =
708 coded_to_superres_mi(mi_col, cm->superres_scale_denominator);
709 const int mi_col_end_sr =
710 coded_to_superres_mi(mi_col + mi_wide, cm->superres_scale_denominator);
711 const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width);
712 const int step = 1 << block_mis_log2;
713 const int row_step = step;
714 const int col_step_sr =
715 coded_to_superres_mi(step, cm->superres_scale_denominator);
716 for (int row = mi_row; row < mi_row + mi_high; row += row_step) {
717 for (int col = mi_col_sr; col < mi_col_end_sr; col += col_step_sr) {
718 if (row >= cm->mi_params.mi_rows || col >= mi_cols_sr) continue;
719 TplDepStats *this_stats =
720 &tpl_stats[av1_tpl_ptr_pos(row, col, tpl_stride, block_mis_log2)];
721 int64_t mc_dep_delta =
722 RDCOST(tpl_frame->base_rdmult, this_stats->mc_dep_rate,
723 this_stats->mc_dep_dist);
724 intra_cost += this_stats->recrf_dist << RDDIV_BITS;
725 mc_dep_cost += (this_stats->recrf_dist << RDDIV_BITS) + mc_dep_delta;
726 #ifndef NDEBUG
727 mi_count++;
728 #endif
729 }
730 }
731 assert(mi_count <= MAX_TPL_BLK_IN_SB * MAX_TPL_BLK_IN_SB);
732
733 double beta = 1.0;
734 if (mc_dep_cost > 0 && intra_cost > 0) {
735 const double r0 = cpi->rd.r0;
736 const double rk = (double)intra_cost / mc_dep_cost;
737 beta = (r0 / rk);
738 }
739
740 int rdmult = av1_get_adaptive_rdmult(cpi, beta);
741
742 rdmult = AOMMIN(rdmult, orig_rdmult * 3 / 2);
743 rdmult = AOMMAX(rdmult, orig_rdmult * 1 / 2);
744
745 rdmult = AOMMAX(1, rdmult);
746
747 return rdmult;
748 }
749
750 // Checks to see if a super block is on a horizontal image edge.
751 // In most cases this is the "real" edge unless there are formatting
752 // bars embedded in the stream.
av1_active_h_edge(const AV1_COMP * cpi,int mi_row,int mi_step)753 int av1_active_h_edge(const AV1_COMP *cpi, int mi_row, int mi_step) {
754 int top_edge = 0;
755 int bottom_edge = cpi->common.mi_params.mi_rows;
756 int is_active_h_edge = 0;
757
758 // For two pass account for any formatting bars detected.
759 if (is_stat_consumption_stage_twopass(cpi)) {
760 const AV1_COMMON *const cm = &cpi->common;
761 const FIRSTPASS_STATS *const this_frame_stats = read_one_frame_stats(
762 &cpi->ppi->twopass, cm->current_frame.display_order_hint);
763 if (this_frame_stats == NULL) return AOM_CODEC_ERROR;
764
765 // The inactive region is specified in MBs not mi units.
766 // The image edge is in the following MB row.
767 top_edge += (int)(this_frame_stats->inactive_zone_rows * 4);
768
769 bottom_edge -= (int)(this_frame_stats->inactive_zone_rows * 4);
770 bottom_edge = AOMMAX(top_edge, bottom_edge);
771 }
772
773 if (((top_edge >= mi_row) && (top_edge < (mi_row + mi_step))) ||
774 ((bottom_edge >= mi_row) && (bottom_edge < (mi_row + mi_step)))) {
775 is_active_h_edge = 1;
776 }
777 return is_active_h_edge;
778 }
779
780 // Checks to see if a super block is on a vertical image edge.
781 // In most cases this is the "real" edge unless there are formatting
782 // bars embedded in the stream.
av1_active_v_edge(const AV1_COMP * cpi,int mi_col,int mi_step)783 int av1_active_v_edge(const AV1_COMP *cpi, int mi_col, int mi_step) {
784 int left_edge = 0;
785 int right_edge = cpi->common.mi_params.mi_cols;
786 int is_active_v_edge = 0;
787
788 // For two pass account for any formatting bars detected.
789 if (is_stat_consumption_stage_twopass(cpi)) {
790 const AV1_COMMON *const cm = &cpi->common;
791 const FIRSTPASS_STATS *const this_frame_stats = read_one_frame_stats(
792 &cpi->ppi->twopass, cm->current_frame.display_order_hint);
793 if (this_frame_stats == NULL) return AOM_CODEC_ERROR;
794
795 // The inactive region is specified in MBs not mi units.
796 // The image edge is in the following MB row.
797 left_edge += (int)(this_frame_stats->inactive_zone_cols * 4);
798
799 right_edge -= (int)(this_frame_stats->inactive_zone_cols * 4);
800 right_edge = AOMMAX(left_edge, right_edge);
801 }
802
803 if (((left_edge >= mi_col) && (left_edge < (mi_col + mi_step))) ||
804 ((right_edge >= mi_col) && (right_edge < (mi_col + mi_step)))) {
805 is_active_v_edge = 1;
806 }
807 return is_active_v_edge;
808 }
809
av1_get_tpl_stats_sb(AV1_COMP * cpi,BLOCK_SIZE bsize,int mi_row,int mi_col,SuperBlockEnc * sb_enc)810 void av1_get_tpl_stats_sb(AV1_COMP *cpi, BLOCK_SIZE bsize, int mi_row,
811 int mi_col, SuperBlockEnc *sb_enc) {
812 sb_enc->tpl_data_count = 0;
813
814 if (!cpi->oxcf.algo_cfg.enable_tpl_model) return;
815 if (cpi->common.current_frame.frame_type == KEY_FRAME) return;
816 const FRAME_UPDATE_TYPE update_type =
817 get_frame_update_type(&cpi->ppi->gf_group, cpi->gf_frame_index);
818 if (update_type == INTNL_OVERLAY_UPDATE || update_type == OVERLAY_UPDATE)
819 return;
820 assert(IMPLIES(cpi->ppi->gf_group.size > 0,
821 cpi->gf_frame_index < cpi->ppi->gf_group.size));
822
823 AV1_COMMON *const cm = &cpi->common;
824 const int gf_group_index = cpi->gf_frame_index;
825 TplParams *const tpl_data = &cpi->ppi->tpl_data;
826 if (!av1_tpl_stats_ready(tpl_data, gf_group_index)) return;
827 const int mi_wide = mi_size_wide[bsize];
828 const int mi_high = mi_size_high[bsize];
829
830 TplDepFrame *tpl_frame = &tpl_data->tpl_frame[gf_group_index];
831 TplDepStats *tpl_stats = tpl_frame->tpl_stats_ptr;
832 int tpl_stride = tpl_frame->stride;
833
834 int mi_count = 0;
835 int count = 0;
836 const int mi_col_sr =
837 coded_to_superres_mi(mi_col, cm->superres_scale_denominator);
838 const int mi_col_end_sr =
839 coded_to_superres_mi(mi_col + mi_wide, cm->superres_scale_denominator);
840 // mi_cols_sr is mi_cols at superres case.
841 const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width);
842
843 // TPL store unit size is not the same as the motion estimation unit size.
844 // Here always use motion estimation size to avoid getting repetitive inter/
845 // intra cost.
846 const BLOCK_SIZE tpl_bsize = convert_length_to_bsize(tpl_data->tpl_bsize_1d);
847 assert(mi_size_wide[tpl_bsize] == mi_size_high[tpl_bsize]);
848 const int row_step = mi_size_high[tpl_bsize];
849 const int col_step_sr = coded_to_superres_mi(mi_size_wide[tpl_bsize],
850 cm->superres_scale_denominator);
851
852 // Stride is only based on SB size, and we fill in values for every 16x16
853 // block in a SB.
854 sb_enc->tpl_stride = (mi_col_end_sr - mi_col_sr) / col_step_sr;
855
856 for (int row = mi_row; row < mi_row + mi_high; row += row_step) {
857 for (int col = mi_col_sr; col < mi_col_end_sr; col += col_step_sr) {
858 assert(count < MAX_TPL_BLK_IN_SB * MAX_TPL_BLK_IN_SB);
859 // Handle partial SB, so that no invalid values are used later.
860 if (row >= cm->mi_params.mi_rows || col >= mi_cols_sr) {
861 sb_enc->tpl_inter_cost[count] = INT64_MAX;
862 sb_enc->tpl_intra_cost[count] = INT64_MAX;
863 for (int i = 0; i < INTER_REFS_PER_FRAME; ++i) {
864 sb_enc->tpl_mv[count][i].as_int = INVALID_MV;
865 }
866 count++;
867 continue;
868 }
869
870 TplDepStats *this_stats = &tpl_stats[av1_tpl_ptr_pos(
871 row, col, tpl_stride, tpl_data->tpl_stats_block_mis_log2)];
872 sb_enc->tpl_inter_cost[count] = this_stats->inter_cost
873 << TPL_DEP_COST_SCALE_LOG2;
874 sb_enc->tpl_intra_cost[count] = this_stats->intra_cost
875 << TPL_DEP_COST_SCALE_LOG2;
876 memcpy(sb_enc->tpl_mv[count], this_stats->mv, sizeof(this_stats->mv));
877 mi_count++;
878 count++;
879 }
880 }
881
882 assert(mi_count <= MAX_TPL_BLK_IN_SB * MAX_TPL_BLK_IN_SB);
883 sb_enc->tpl_data_count = mi_count;
884 }
885
886 // analysis_type 0: Use mc_dep_cost and intra_cost
887 // analysis_type 1: Use count of best inter predictor chosen
888 // analysis_type 2: Use cost reduction from intra to inter for best inter
889 // predictor chosen
av1_get_q_for_deltaq_objective(AV1_COMP * const cpi,ThreadData * td,int64_t * delta_dist,BLOCK_SIZE bsize,int mi_row,int mi_col)890 int av1_get_q_for_deltaq_objective(AV1_COMP *const cpi, ThreadData *td,
891 int64_t *delta_dist, BLOCK_SIZE bsize,
892 int mi_row, int mi_col) {
893 AV1_COMMON *const cm = &cpi->common;
894 assert(IMPLIES(cpi->ppi->gf_group.size > 0,
895 cpi->gf_frame_index < cpi->ppi->gf_group.size));
896 const int tpl_idx = cpi->gf_frame_index;
897 TplParams *const tpl_data = &cpi->ppi->tpl_data;
898 const uint8_t block_mis_log2 = tpl_data->tpl_stats_block_mis_log2;
899 double intra_cost = 0;
900 double mc_dep_reg = 0;
901 double mc_dep_cost = 0;
902 double cbcmp_base = 1;
903 double srcrf_dist = 0;
904 double srcrf_sse = 0;
905 double srcrf_rate = 0;
906 const int mi_wide = mi_size_wide[bsize];
907 const int mi_high = mi_size_high[bsize];
908 const int base_qindex = cm->quant_params.base_qindex;
909
910 if (tpl_idx >= MAX_TPL_FRAME_IDX) return base_qindex;
911
912 TplDepFrame *tpl_frame = &tpl_data->tpl_frame[tpl_idx];
913 TplDepStats *tpl_stats = tpl_frame->tpl_stats_ptr;
914 int tpl_stride = tpl_frame->stride;
915 if (!tpl_frame->is_valid) return base_qindex;
916
917 #ifndef NDEBUG
918 int mi_count = 0;
919 #endif
920 const int mi_col_sr =
921 coded_to_superres_mi(mi_col, cm->superres_scale_denominator);
922 const int mi_col_end_sr =
923 coded_to_superres_mi(mi_col + mi_wide, cm->superres_scale_denominator);
924 const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width);
925 const int step = 1 << block_mis_log2;
926 const int row_step = step;
927 const int col_step_sr =
928 coded_to_superres_mi(step, cm->superres_scale_denominator);
929 for (int row = mi_row; row < mi_row + mi_high; row += row_step) {
930 for (int col = mi_col_sr; col < mi_col_end_sr; col += col_step_sr) {
931 if (row >= cm->mi_params.mi_rows || col >= mi_cols_sr) continue;
932 TplDepStats *this_stats =
933 &tpl_stats[av1_tpl_ptr_pos(row, col, tpl_stride, block_mis_log2)];
934 double cbcmp = (double)this_stats->srcrf_dist;
935 int64_t mc_dep_delta =
936 RDCOST(tpl_frame->base_rdmult, this_stats->mc_dep_rate,
937 this_stats->mc_dep_dist);
938 double dist_scaled = (double)(this_stats->recrf_dist << RDDIV_BITS);
939 intra_cost += log(dist_scaled) * cbcmp;
940 mc_dep_cost += log(dist_scaled + mc_dep_delta) * cbcmp;
941 mc_dep_reg += log(3 * dist_scaled + mc_dep_delta) * cbcmp;
942 srcrf_dist += (double)(this_stats->srcrf_dist << RDDIV_BITS);
943 srcrf_sse += (double)(this_stats->srcrf_sse << RDDIV_BITS);
944 srcrf_rate += (double)(this_stats->srcrf_rate << TPL_DEP_COST_SCALE_LOG2);
945 #ifndef NDEBUG
946 mi_count++;
947 #endif
948 cbcmp_base += cbcmp;
949 }
950 }
951 assert(mi_count <= MAX_TPL_BLK_IN_SB * MAX_TPL_BLK_IN_SB);
952
953 int offset = 0;
954 double beta = 1.0;
955 double rk;
956 if (mc_dep_cost > 0 && intra_cost > 0) {
957 const double r0 = cpi->rd.r0;
958 rk = exp((intra_cost - mc_dep_cost) / cbcmp_base);
959 td->mb.rb = exp((intra_cost - mc_dep_reg) / cbcmp_base);
960 beta = (r0 / rk);
961 assert(beta > 0.0);
962 } else {
963 return base_qindex;
964 }
965 offset = av1_get_deltaq_offset(cm->seq_params->bit_depth, base_qindex, beta);
966
967 const DeltaQInfo *const delta_q_info = &cm->delta_q_info;
968 offset = AOMMIN(offset, delta_q_info->delta_q_res * 9 - 1);
969 offset = AOMMAX(offset, -delta_q_info->delta_q_res * 9 + 1);
970 int qindex = cm->quant_params.base_qindex + offset;
971 qindex = AOMMIN(qindex, MAXQ);
972 qindex = AOMMAX(qindex, MINQ);
973
974 int frm_qstep = av1_dc_quant_QTX(base_qindex, 0, cm->seq_params->bit_depth);
975 int sbs_qstep =
976 av1_dc_quant_QTX(base_qindex, offset, cm->seq_params->bit_depth);
977
978 if (delta_dist) {
979 double sbs_dist = srcrf_dist * pow((double)sbs_qstep / frm_qstep, 2.0);
980 double sbs_rate = srcrf_rate * ((double)frm_qstep / sbs_qstep);
981 sbs_dist = AOMMIN(sbs_dist, srcrf_sse);
982 *delta_dist = (int64_t)((sbs_dist - srcrf_dist) / rk);
983 *delta_dist += RDCOST(tpl_frame->base_rdmult, 4 * 256, 0);
984 *delta_dist += RDCOST(tpl_frame->base_rdmult, sbs_rate - srcrf_rate, 0);
985 }
986 return qindex;
987 }
988
989 #if !DISABLE_HDR_LUMA_DELTAQ
990 // offset table defined in Table3 of T-REC-H.Sup15 document.
991 static const int hdr_thres[HDR_QP_LEVELS + 1] = { 0, 301, 367, 434, 501, 567,
992 634, 701, 767, 834, 1024 };
993
994 static const int hdr10_qp_offset[HDR_QP_LEVELS] = { 3, 2, 1, 0, -1,
995 -2, -3, -4, -5, -6 };
996 #endif
997
av1_get_q_for_hdr(AV1_COMP * const cpi,MACROBLOCK * const x,BLOCK_SIZE bsize,int mi_row,int mi_col)998 int av1_get_q_for_hdr(AV1_COMP *const cpi, MACROBLOCK *const x,
999 BLOCK_SIZE bsize, int mi_row, int mi_col) {
1000 AV1_COMMON *const cm = &cpi->common;
1001 assert(cm->seq_params->bit_depth == AOM_BITS_10);
1002
1003 #if DISABLE_HDR_LUMA_DELTAQ
1004 (void)x;
1005 (void)bsize;
1006 (void)mi_row;
1007 (void)mi_col;
1008 return cm->quant_params.base_qindex;
1009 #else
1010 // calculate pixel average
1011 const int block_luma_avg = av1_log_block_avg(cpi, x, bsize, mi_row, mi_col);
1012 // adjust offset based on average of the pixel block
1013 int offset = 0;
1014 for (int i = 0; i < HDR_QP_LEVELS; i++) {
1015 if (block_luma_avg >= hdr_thres[i] && block_luma_avg < hdr_thres[i + 1]) {
1016 offset = (int)(hdr10_qp_offset[i] * QP_SCALE_FACTOR);
1017 break;
1018 }
1019 }
1020
1021 const DeltaQInfo *const delta_q_info = &cm->delta_q_info;
1022 offset = AOMMIN(offset, delta_q_info->delta_q_res * 9 - 1);
1023 offset = AOMMAX(offset, -delta_q_info->delta_q_res * 9 + 1);
1024 int qindex = cm->quant_params.base_qindex + offset;
1025 qindex = AOMMIN(qindex, MAXQ);
1026 qindex = AOMMAX(qindex, MINQ);
1027
1028 return qindex;
1029 #endif
1030 }
1031 #endif // !CONFIG_REALTIME_ONLY
1032
av1_reset_simple_motion_tree_partition(SIMPLE_MOTION_DATA_TREE * sms_tree,BLOCK_SIZE bsize)1033 void av1_reset_simple_motion_tree_partition(SIMPLE_MOTION_DATA_TREE *sms_tree,
1034 BLOCK_SIZE bsize) {
1035 if (sms_tree == NULL) return;
1036 sms_tree->partitioning = PARTITION_NONE;
1037
1038 if (bsize >= BLOCK_8X8) {
1039 BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_SPLIT);
1040 for (int idx = 0; idx < 4; ++idx)
1041 av1_reset_simple_motion_tree_partition(sms_tree->split[idx], subsize);
1042 }
1043 }
1044
1045 // Record the ref frames that have been selected by square partition blocks.
av1_update_picked_ref_frames_mask(MACROBLOCK * const x,int ref_type,BLOCK_SIZE bsize,int mib_size,int mi_row,int mi_col)1046 void av1_update_picked_ref_frames_mask(MACROBLOCK *const x, int ref_type,
1047 BLOCK_SIZE bsize, int mib_size,
1048 int mi_row, int mi_col) {
1049 assert(mi_size_wide[bsize] == mi_size_high[bsize]);
1050 const int sb_size_mask = mib_size - 1;
1051 const int mi_row_in_sb = mi_row & sb_size_mask;
1052 const int mi_col_in_sb = mi_col & sb_size_mask;
1053 const int mi_size = mi_size_wide[bsize];
1054 for (int i = mi_row_in_sb; i < mi_row_in_sb + mi_size; ++i) {
1055 for (int j = mi_col_in_sb; j < mi_col_in_sb + mi_size; ++j) {
1056 x->picked_ref_frames_mask[i * 32 + j] |= 1 << ref_type;
1057 }
1058 }
1059 }
1060
avg_cdf_symbol(aom_cdf_prob * cdf_ptr_left,aom_cdf_prob * cdf_ptr_tr,int num_cdfs,int cdf_stride,int nsymbs,int wt_left,int wt_tr)1061 static void avg_cdf_symbol(aom_cdf_prob *cdf_ptr_left, aom_cdf_prob *cdf_ptr_tr,
1062 int num_cdfs, int cdf_stride, int nsymbs,
1063 int wt_left, int wt_tr) {
1064 for (int i = 0; i < num_cdfs; i++) {
1065 for (int j = 0; j <= nsymbs; j++) {
1066 cdf_ptr_left[i * cdf_stride + j] =
1067 (aom_cdf_prob)(((int)cdf_ptr_left[i * cdf_stride + j] * wt_left +
1068 (int)cdf_ptr_tr[i * cdf_stride + j] * wt_tr +
1069 ((wt_left + wt_tr) / 2)) /
1070 (wt_left + wt_tr));
1071 assert(cdf_ptr_left[i * cdf_stride + j] >= 0 &&
1072 cdf_ptr_left[i * cdf_stride + j] < CDF_PROB_TOP);
1073 }
1074 }
1075 }
1076
1077 #define AVERAGE_CDF(cname_left, cname_tr, nsymbs) \
1078 AVG_CDF_STRIDE(cname_left, cname_tr, nsymbs, CDF_SIZE(nsymbs))
1079
1080 #define AVG_CDF_STRIDE(cname_left, cname_tr, nsymbs, cdf_stride) \
1081 do { \
1082 aom_cdf_prob *cdf_ptr_left = (aom_cdf_prob *)cname_left; \
1083 aom_cdf_prob *cdf_ptr_tr = (aom_cdf_prob *)cname_tr; \
1084 int array_size = (int)sizeof(cname_left) / sizeof(aom_cdf_prob); \
1085 int num_cdfs = array_size / cdf_stride; \
1086 avg_cdf_symbol(cdf_ptr_left, cdf_ptr_tr, num_cdfs, cdf_stride, nsymbs, \
1087 wt_left, wt_tr); \
1088 } while (0)
1089
avg_nmv(nmv_context * nmv_left,nmv_context * nmv_tr,int wt_left,int wt_tr)1090 static void avg_nmv(nmv_context *nmv_left, nmv_context *nmv_tr, int wt_left,
1091 int wt_tr) {
1092 AVERAGE_CDF(nmv_left->joints_cdf, nmv_tr->joints_cdf, 4);
1093 for (int i = 0; i < 2; i++) {
1094 AVERAGE_CDF(nmv_left->comps[i].classes_cdf, nmv_tr->comps[i].classes_cdf,
1095 MV_CLASSES);
1096 AVERAGE_CDF(nmv_left->comps[i].class0_fp_cdf,
1097 nmv_tr->comps[i].class0_fp_cdf, MV_FP_SIZE);
1098 AVERAGE_CDF(nmv_left->comps[i].fp_cdf, nmv_tr->comps[i].fp_cdf, MV_FP_SIZE);
1099 AVERAGE_CDF(nmv_left->comps[i].sign_cdf, nmv_tr->comps[i].sign_cdf, 2);
1100 AVERAGE_CDF(nmv_left->comps[i].class0_hp_cdf,
1101 nmv_tr->comps[i].class0_hp_cdf, 2);
1102 AVERAGE_CDF(nmv_left->comps[i].hp_cdf, nmv_tr->comps[i].hp_cdf, 2);
1103 AVERAGE_CDF(nmv_left->comps[i].class0_cdf, nmv_tr->comps[i].class0_cdf,
1104 CLASS0_SIZE);
1105 AVERAGE_CDF(nmv_left->comps[i].bits_cdf, nmv_tr->comps[i].bits_cdf, 2);
1106 }
1107 }
1108
1109 // In case of row-based multi-threading of encoder, since we always
1110 // keep a top - right sync, we can average the top - right SB's CDFs and
1111 // the left SB's CDFs and use the same for current SB's encoding to
1112 // improve the performance. This function facilitates the averaging
1113 // of CDF and used only when row-mt is enabled in encoder.
av1_avg_cdf_symbols(FRAME_CONTEXT * ctx_left,FRAME_CONTEXT * ctx_tr,int wt_left,int wt_tr)1114 void av1_avg_cdf_symbols(FRAME_CONTEXT *ctx_left, FRAME_CONTEXT *ctx_tr,
1115 int wt_left, int wt_tr) {
1116 AVERAGE_CDF(ctx_left->txb_skip_cdf, ctx_tr->txb_skip_cdf, 2);
1117 AVERAGE_CDF(ctx_left->eob_extra_cdf, ctx_tr->eob_extra_cdf, 2);
1118 AVERAGE_CDF(ctx_left->dc_sign_cdf, ctx_tr->dc_sign_cdf, 2);
1119 AVERAGE_CDF(ctx_left->eob_flag_cdf16, ctx_tr->eob_flag_cdf16, 5);
1120 AVERAGE_CDF(ctx_left->eob_flag_cdf32, ctx_tr->eob_flag_cdf32, 6);
1121 AVERAGE_CDF(ctx_left->eob_flag_cdf64, ctx_tr->eob_flag_cdf64, 7);
1122 AVERAGE_CDF(ctx_left->eob_flag_cdf128, ctx_tr->eob_flag_cdf128, 8);
1123 AVERAGE_CDF(ctx_left->eob_flag_cdf256, ctx_tr->eob_flag_cdf256, 9);
1124 AVERAGE_CDF(ctx_left->eob_flag_cdf512, ctx_tr->eob_flag_cdf512, 10);
1125 AVERAGE_CDF(ctx_left->eob_flag_cdf1024, ctx_tr->eob_flag_cdf1024, 11);
1126 AVERAGE_CDF(ctx_left->coeff_base_eob_cdf, ctx_tr->coeff_base_eob_cdf, 3);
1127 AVERAGE_CDF(ctx_left->coeff_base_cdf, ctx_tr->coeff_base_cdf, 4);
1128 AVERAGE_CDF(ctx_left->coeff_br_cdf, ctx_tr->coeff_br_cdf, BR_CDF_SIZE);
1129 AVERAGE_CDF(ctx_left->newmv_cdf, ctx_tr->newmv_cdf, 2);
1130 AVERAGE_CDF(ctx_left->zeromv_cdf, ctx_tr->zeromv_cdf, 2);
1131 AVERAGE_CDF(ctx_left->refmv_cdf, ctx_tr->refmv_cdf, 2);
1132 AVERAGE_CDF(ctx_left->drl_cdf, ctx_tr->drl_cdf, 2);
1133 AVERAGE_CDF(ctx_left->inter_compound_mode_cdf,
1134 ctx_tr->inter_compound_mode_cdf, INTER_COMPOUND_MODES);
1135 AVERAGE_CDF(ctx_left->compound_type_cdf, ctx_tr->compound_type_cdf,
1136 MASKED_COMPOUND_TYPES);
1137 AVERAGE_CDF(ctx_left->wedge_idx_cdf, ctx_tr->wedge_idx_cdf, 16);
1138 AVERAGE_CDF(ctx_left->interintra_cdf, ctx_tr->interintra_cdf, 2);
1139 AVERAGE_CDF(ctx_left->wedge_interintra_cdf, ctx_tr->wedge_interintra_cdf, 2);
1140 AVERAGE_CDF(ctx_left->interintra_mode_cdf, ctx_tr->interintra_mode_cdf,
1141 INTERINTRA_MODES);
1142 AVERAGE_CDF(ctx_left->motion_mode_cdf, ctx_tr->motion_mode_cdf, MOTION_MODES);
1143 AVERAGE_CDF(ctx_left->obmc_cdf, ctx_tr->obmc_cdf, 2);
1144 AVERAGE_CDF(ctx_left->palette_y_size_cdf, ctx_tr->palette_y_size_cdf,
1145 PALETTE_SIZES);
1146 AVERAGE_CDF(ctx_left->palette_uv_size_cdf, ctx_tr->palette_uv_size_cdf,
1147 PALETTE_SIZES);
1148 for (int j = 0; j < PALETTE_SIZES; j++) {
1149 int nsymbs = j + PALETTE_MIN_SIZE;
1150 AVG_CDF_STRIDE(ctx_left->palette_y_color_index_cdf[j],
1151 ctx_tr->palette_y_color_index_cdf[j], nsymbs,
1152 CDF_SIZE(PALETTE_COLORS));
1153 AVG_CDF_STRIDE(ctx_left->palette_uv_color_index_cdf[j],
1154 ctx_tr->palette_uv_color_index_cdf[j], nsymbs,
1155 CDF_SIZE(PALETTE_COLORS));
1156 }
1157 AVERAGE_CDF(ctx_left->palette_y_mode_cdf, ctx_tr->palette_y_mode_cdf, 2);
1158 AVERAGE_CDF(ctx_left->palette_uv_mode_cdf, ctx_tr->palette_uv_mode_cdf, 2);
1159 AVERAGE_CDF(ctx_left->comp_inter_cdf, ctx_tr->comp_inter_cdf, 2);
1160 AVERAGE_CDF(ctx_left->single_ref_cdf, ctx_tr->single_ref_cdf, 2);
1161 AVERAGE_CDF(ctx_left->comp_ref_type_cdf, ctx_tr->comp_ref_type_cdf, 2);
1162 AVERAGE_CDF(ctx_left->uni_comp_ref_cdf, ctx_tr->uni_comp_ref_cdf, 2);
1163 AVERAGE_CDF(ctx_left->comp_ref_cdf, ctx_tr->comp_ref_cdf, 2);
1164 AVERAGE_CDF(ctx_left->comp_bwdref_cdf, ctx_tr->comp_bwdref_cdf, 2);
1165 AVERAGE_CDF(ctx_left->txfm_partition_cdf, ctx_tr->txfm_partition_cdf, 2);
1166 AVERAGE_CDF(ctx_left->compound_index_cdf, ctx_tr->compound_index_cdf, 2);
1167 AVERAGE_CDF(ctx_left->comp_group_idx_cdf, ctx_tr->comp_group_idx_cdf, 2);
1168 AVERAGE_CDF(ctx_left->skip_mode_cdfs, ctx_tr->skip_mode_cdfs, 2);
1169 AVERAGE_CDF(ctx_left->skip_txfm_cdfs, ctx_tr->skip_txfm_cdfs, 2);
1170 AVERAGE_CDF(ctx_left->intra_inter_cdf, ctx_tr->intra_inter_cdf, 2);
1171 avg_nmv(&ctx_left->nmvc, &ctx_tr->nmvc, wt_left, wt_tr);
1172 avg_nmv(&ctx_left->ndvc, &ctx_tr->ndvc, wt_left, wt_tr);
1173 AVERAGE_CDF(ctx_left->intrabc_cdf, ctx_tr->intrabc_cdf, 2);
1174 AVERAGE_CDF(ctx_left->seg.pred_cdf, ctx_tr->seg.pred_cdf, 2);
1175 AVERAGE_CDF(ctx_left->seg.spatial_pred_seg_cdf,
1176 ctx_tr->seg.spatial_pred_seg_cdf, MAX_SEGMENTS);
1177 AVERAGE_CDF(ctx_left->filter_intra_cdfs, ctx_tr->filter_intra_cdfs, 2);
1178 AVERAGE_CDF(ctx_left->filter_intra_mode_cdf, ctx_tr->filter_intra_mode_cdf,
1179 FILTER_INTRA_MODES);
1180 AVERAGE_CDF(ctx_left->switchable_restore_cdf, ctx_tr->switchable_restore_cdf,
1181 RESTORE_SWITCHABLE_TYPES);
1182 AVERAGE_CDF(ctx_left->wiener_restore_cdf, ctx_tr->wiener_restore_cdf, 2);
1183 AVERAGE_CDF(ctx_left->sgrproj_restore_cdf, ctx_tr->sgrproj_restore_cdf, 2);
1184 AVERAGE_CDF(ctx_left->y_mode_cdf, ctx_tr->y_mode_cdf, INTRA_MODES);
1185 AVG_CDF_STRIDE(ctx_left->uv_mode_cdf[0], ctx_tr->uv_mode_cdf[0],
1186 UV_INTRA_MODES - 1, CDF_SIZE(UV_INTRA_MODES));
1187 AVERAGE_CDF(ctx_left->uv_mode_cdf[1], ctx_tr->uv_mode_cdf[1], UV_INTRA_MODES);
1188 for (int i = 0; i < PARTITION_CONTEXTS; i++) {
1189 if (i < 4) {
1190 AVG_CDF_STRIDE(ctx_left->partition_cdf[i], ctx_tr->partition_cdf[i], 4,
1191 CDF_SIZE(10));
1192 } else if (i < 16) {
1193 AVERAGE_CDF(ctx_left->partition_cdf[i], ctx_tr->partition_cdf[i], 10);
1194 } else {
1195 AVG_CDF_STRIDE(ctx_left->partition_cdf[i], ctx_tr->partition_cdf[i], 8,
1196 CDF_SIZE(10));
1197 }
1198 }
1199 AVERAGE_CDF(ctx_left->switchable_interp_cdf, ctx_tr->switchable_interp_cdf,
1200 SWITCHABLE_FILTERS);
1201 AVERAGE_CDF(ctx_left->kf_y_cdf, ctx_tr->kf_y_cdf, INTRA_MODES);
1202 AVERAGE_CDF(ctx_left->angle_delta_cdf, ctx_tr->angle_delta_cdf,
1203 2 * MAX_ANGLE_DELTA + 1);
1204 AVG_CDF_STRIDE(ctx_left->tx_size_cdf[0], ctx_tr->tx_size_cdf[0], MAX_TX_DEPTH,
1205 CDF_SIZE(MAX_TX_DEPTH + 1));
1206 AVERAGE_CDF(ctx_left->tx_size_cdf[1], ctx_tr->tx_size_cdf[1],
1207 MAX_TX_DEPTH + 1);
1208 AVERAGE_CDF(ctx_left->tx_size_cdf[2], ctx_tr->tx_size_cdf[2],
1209 MAX_TX_DEPTH + 1);
1210 AVERAGE_CDF(ctx_left->tx_size_cdf[3], ctx_tr->tx_size_cdf[3],
1211 MAX_TX_DEPTH + 1);
1212 AVERAGE_CDF(ctx_left->delta_q_cdf, ctx_tr->delta_q_cdf, DELTA_Q_PROBS + 1);
1213 AVERAGE_CDF(ctx_left->delta_lf_cdf, ctx_tr->delta_lf_cdf, DELTA_LF_PROBS + 1);
1214 for (int i = 0; i < FRAME_LF_COUNT; i++) {
1215 AVERAGE_CDF(ctx_left->delta_lf_multi_cdf[i], ctx_tr->delta_lf_multi_cdf[i],
1216 DELTA_LF_PROBS + 1);
1217 }
1218 AVG_CDF_STRIDE(ctx_left->intra_ext_tx_cdf[1], ctx_tr->intra_ext_tx_cdf[1], 7,
1219 CDF_SIZE(TX_TYPES));
1220 AVG_CDF_STRIDE(ctx_left->intra_ext_tx_cdf[2], ctx_tr->intra_ext_tx_cdf[2], 5,
1221 CDF_SIZE(TX_TYPES));
1222 AVG_CDF_STRIDE(ctx_left->inter_ext_tx_cdf[1], ctx_tr->inter_ext_tx_cdf[1], 16,
1223 CDF_SIZE(TX_TYPES));
1224 AVG_CDF_STRIDE(ctx_left->inter_ext_tx_cdf[2], ctx_tr->inter_ext_tx_cdf[2], 12,
1225 CDF_SIZE(TX_TYPES));
1226 AVG_CDF_STRIDE(ctx_left->inter_ext_tx_cdf[3], ctx_tr->inter_ext_tx_cdf[3], 2,
1227 CDF_SIZE(TX_TYPES));
1228 AVERAGE_CDF(ctx_left->cfl_sign_cdf, ctx_tr->cfl_sign_cdf, CFL_JOINT_SIGNS);
1229 AVERAGE_CDF(ctx_left->cfl_alpha_cdf, ctx_tr->cfl_alpha_cdf,
1230 CFL_ALPHABET_SIZE);
1231 }
1232
1233 // Check neighbor blocks' motion information.
check_neighbor_blocks(MB_MODE_INFO ** mi,int mi_stride,const TileInfo * const tile_info,int mi_row,int mi_col)1234 static int check_neighbor_blocks(MB_MODE_INFO **mi, int mi_stride,
1235 const TileInfo *const tile_info, int mi_row,
1236 int mi_col) {
1237 int is_above_low_motion = 1;
1238 int is_left_low_motion = 1;
1239 const int thr = 24;
1240
1241 // Check above block.
1242 if (mi_row > tile_info->mi_row_start) {
1243 const MB_MODE_INFO *above_mbmi = mi[-mi_stride];
1244 const int_mv above_mv = above_mbmi->mv[0];
1245 if (above_mbmi->mode >= INTRA_MODE_END &&
1246 (abs(above_mv.as_mv.row) > thr || abs(above_mv.as_mv.col) > thr))
1247 is_above_low_motion = 0;
1248 }
1249
1250 // Check left block.
1251 if (mi_col > tile_info->mi_col_start) {
1252 const MB_MODE_INFO *left_mbmi = mi[-1];
1253 const int_mv left_mv = left_mbmi->mv[0];
1254 if (left_mbmi->mode >= INTRA_MODE_END &&
1255 (abs(left_mv.as_mv.row) > thr || abs(left_mv.as_mv.col) > thr))
1256 is_left_low_motion = 0;
1257 }
1258
1259 return (is_above_low_motion && is_left_low_motion);
1260 }
1261
1262 // Check this block's motion in a fast way.
fast_detect_non_zero_motion(AV1_COMP * cpi,const uint8_t * src_y,int src_ystride,const uint8_t * last_src_y,int last_src_ystride,int mi_row,int mi_col)1263 static int fast_detect_non_zero_motion(AV1_COMP *cpi, const uint8_t *src_y,
1264 int src_ystride,
1265 const uint8_t *last_src_y,
1266 int last_src_ystride, int mi_row,
1267 int mi_col) {
1268 AV1_COMMON *const cm = &cpi->common;
1269 const BLOCK_SIZE bsize = cm->seq_params->sb_size;
1270 unsigned int blk_sad = INT_MAX;
1271 if (cpi->src_sad_blk_64x64 != NULL) {
1272 const int sb_size_by_mb = (bsize == BLOCK_128X128)
1273 ? (cm->seq_params->mib_size >> 1)
1274 : cm->seq_params->mib_size;
1275 const int sb_cols =
1276 (cm->mi_params.mi_cols + sb_size_by_mb - 1) / sb_size_by_mb;
1277 const int sbi_col = mi_col / sb_size_by_mb;
1278 const int sbi_row = mi_row / sb_size_by_mb;
1279 blk_sad = (unsigned int)cpi->src_sad_blk_64x64[sbi_col + sbi_row * sb_cols];
1280 } else {
1281 blk_sad = cpi->ppi->fn_ptr[bsize].sdf(src_y, src_ystride, last_src_y,
1282 last_src_ystride);
1283 }
1284
1285 // Search 4 1-away points.
1286 const uint8_t *const search_pos[4] = {
1287 last_src_y - last_src_ystride,
1288 last_src_y - 1,
1289 last_src_y + 1,
1290 last_src_y + last_src_ystride,
1291 };
1292 unsigned int sad_arr[4];
1293 cpi->ppi->fn_ptr[bsize].sdx4df(src_y, src_ystride, search_pos,
1294 last_src_ystride, sad_arr);
1295
1296 blk_sad = (blk_sad * 5) >> 3;
1297 return (blk_sad < sad_arr[0] && blk_sad < sad_arr[1] &&
1298 blk_sad < sad_arr[2] && blk_sad < sad_arr[3]);
1299 }
1300
1301 // Grade the temporal variation of the source by comparing the current sb and
1302 // its collocated block in the last frame.
av1_source_content_sb(AV1_COMP * cpi,MACROBLOCK * x,TileDataEnc * tile_data,int mi_row,int mi_col)1303 void av1_source_content_sb(AV1_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data,
1304 int mi_row, int mi_col) {
1305 if (cpi->last_source->y_width != cpi->source->y_width ||
1306 cpi->last_source->y_height != cpi->source->y_height)
1307 return;
1308 #if CONFIG_AV1_HIGHBITDEPTH
1309 if (x->e_mbd.cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) return;
1310 #endif
1311
1312 unsigned int tmp_sse;
1313 unsigned int tmp_variance;
1314 const BLOCK_SIZE bsize = cpi->common.seq_params->sb_size;
1315 uint8_t *src_y = cpi->source->y_buffer;
1316 const int src_ystride = cpi->source->y_stride;
1317 const int src_offset = src_ystride * (mi_row << 2) + (mi_col << 2);
1318 uint8_t *last_src_y = cpi->last_source->y_buffer;
1319 const int last_src_ystride = cpi->last_source->y_stride;
1320 const int last_src_offset = last_src_ystride * (mi_row << 2) + (mi_col << 2);
1321 uint64_t avg_source_sse_threshold_verylow = 10000; // ~1.5*1.5*(64*64)
1322 uint64_t avg_source_sse_threshold_low[2] = { 100000, // ~5*5*(64*64)
1323 36000 }; // ~3*3*(64*64)
1324
1325 uint64_t avg_source_sse_threshold_high = 1000000; // ~15*15*(64*64)
1326 if (cpi->sf.rt_sf.increase_source_sad_thresh) {
1327 avg_source_sse_threshold_high = avg_source_sse_threshold_high << 1;
1328 avg_source_sse_threshold_low[0] = avg_source_sse_threshold_low[0] << 1;
1329 avg_source_sse_threshold_verylow = avg_source_sse_threshold_verylow << 1;
1330 }
1331 uint64_t sum_sq_thresh = 10000; // sum = sqrt(thresh / 64*64)) ~1.5
1332 src_y += src_offset;
1333 last_src_y += last_src_offset;
1334 tmp_variance = cpi->ppi->fn_ptr[bsize].vf(src_y, src_ystride, last_src_y,
1335 last_src_ystride, &tmp_sse);
1336 // rd thresholds
1337 if (tmp_sse < avg_source_sse_threshold_low[1])
1338 x->content_state_sb.source_sad_rd = kLowSad;
1339
1340 // nonrd thresholds
1341 if (tmp_sse == 0) {
1342 x->content_state_sb.source_sad_nonrd = kZeroSad;
1343 return;
1344 }
1345 if (tmp_sse < avg_source_sse_threshold_verylow)
1346 x->content_state_sb.source_sad_nonrd = kVeryLowSad;
1347 else if (tmp_sse < avg_source_sse_threshold_low[0])
1348 x->content_state_sb.source_sad_nonrd = kLowSad;
1349 else if (tmp_sse > avg_source_sse_threshold_high)
1350 x->content_state_sb.source_sad_nonrd = kHighSad;
1351
1352 // Detect large lighting change.
1353 // Note: tmp_sse - tmp_variance = ((sum * sum) >> 12)
1354 if (tmp_variance < (tmp_sse >> 1) && (tmp_sse - tmp_variance) > sum_sq_thresh)
1355 x->content_state_sb.lighting_change = 1;
1356 if ((tmp_sse - tmp_variance) < (sum_sq_thresh >> 1))
1357 x->content_state_sb.low_sumdiff = 1;
1358
1359 if (tmp_sse > ((avg_source_sse_threshold_high * 7) >> 3) &&
1360 !x->content_state_sb.lighting_change && !x->content_state_sb.low_sumdiff)
1361 x->sb_force_fixed_part = 0;
1362
1363 if (!cpi->sf.rt_sf.use_rtc_tf || cpi->rc.high_source_sad ||
1364 cpi->rc.frame_source_sad > 20000 || cpi->svc.number_spatial_layers > 1)
1365 return;
1366
1367 // In-place temporal filter. If psnr calculation is enabled, we store the
1368 // source for that.
1369 AV1_COMMON *const cm = &cpi->common;
1370 // Calculate n*mean^2
1371 const unsigned int nmean2 = tmp_sse - tmp_variance;
1372 const int ac_q_step = av1_ac_quant_QTX(cm->quant_params.base_qindex, 0,
1373 cm->seq_params->bit_depth);
1374 const PRIMARY_RATE_CONTROL *const p_rc = &cpi->ppi->p_rc;
1375 const int avg_q_step = av1_ac_quant_QTX(p_rc->avg_frame_qindex[INTER_FRAME],
1376 0, cm->seq_params->bit_depth);
1377
1378 const unsigned int threshold =
1379 (cpi->sf.rt_sf.use_rtc_tf == 1)
1380 ? (clamp(avg_q_step, 250, 1000)) * ac_q_step
1381 : 250 * ac_q_step;
1382
1383 // TODO(yunqing): use a weighted sum instead of averaging in filtering.
1384 if (tmp_variance <= threshold && nmean2 <= 15) {
1385 // Check neighbor blocks. If neighbor blocks aren't low-motion blocks,
1386 // skip temporal filtering for this block.
1387 MB_MODE_INFO **mi = cm->mi_params.mi_grid_base +
1388 get_mi_grid_idx(&cm->mi_params, mi_row, mi_col);
1389 const TileInfo *const tile_info = &tile_data->tile_info;
1390 const int is_neighbor_blocks_low_motion = check_neighbor_blocks(
1391 mi, cm->mi_params.mi_stride, tile_info, mi_row, mi_col);
1392 if (!is_neighbor_blocks_low_motion) return;
1393
1394 // Only consider 64x64 SB for now. Need to extend to 128x128 for large SB
1395 // size.
1396 // Test several nearby points. If non-zero mv exists, don't do temporal
1397 // filtering.
1398 const int is_this_blk_low_motion = fast_detect_non_zero_motion(
1399 cpi, src_y, src_ystride, last_src_y, last_src_ystride, mi_row, mi_col);
1400
1401 if (!is_this_blk_low_motion) return;
1402
1403 const int shift_x[2] = { 0, cpi->source->subsampling_x };
1404 const int shift_y[2] = { 0, cpi->source->subsampling_y };
1405 const uint8_t h = block_size_high[bsize];
1406 const uint8_t w = block_size_wide[bsize];
1407
1408 for (int plane = 0; plane < av1_num_planes(cm); ++plane) {
1409 uint8_t *src = cpi->source->buffers[plane];
1410 const int src_stride = cpi->source->strides[plane != 0];
1411 uint8_t *last_src = cpi->last_source->buffers[plane];
1412 const int last_src_stride = cpi->last_source->strides[plane != 0];
1413 src += src_stride * (mi_row << (2 - shift_y[plane != 0])) +
1414 (mi_col << (2 - shift_x[plane != 0]));
1415 last_src += last_src_stride * (mi_row << (2 - shift_y[plane != 0])) +
1416 (mi_col << (2 - shift_x[plane != 0]));
1417
1418 for (int i = 0; i < (h >> shift_y[plane != 0]); ++i) {
1419 for (int j = 0; j < (w >> shift_x[plane != 0]); ++j) {
1420 src[j] = (last_src[j] + src[j]) >> 1;
1421 }
1422 src += src_stride;
1423 last_src += last_src_stride;
1424 }
1425 }
1426 }
1427 }
1428
1429 // Memset the mbmis at the current superblock to 0
av1_reset_mbmi(CommonModeInfoParams * const mi_params,BLOCK_SIZE sb_size,int mi_row,int mi_col)1430 void av1_reset_mbmi(CommonModeInfoParams *const mi_params, BLOCK_SIZE sb_size,
1431 int mi_row, int mi_col) {
1432 // size of sb in unit of mi (BLOCK_4X4)
1433 const int sb_size_mi = mi_size_wide[sb_size];
1434 const int mi_alloc_size_1d = mi_size_wide[mi_params->mi_alloc_bsize];
1435 // size of sb in unit of allocated mi size
1436 const int sb_size_alloc_mi = mi_size_wide[sb_size] / mi_alloc_size_1d;
1437 assert(mi_params->mi_alloc_stride % sb_size_alloc_mi == 0 &&
1438 "mi is not allocated as a multiple of sb!");
1439 assert(mi_params->mi_stride % sb_size_mi == 0 &&
1440 "mi_grid_base is not allocated as a multiple of sb!");
1441
1442 const int mi_rows = mi_size_high[sb_size];
1443 for (int cur_mi_row = 0; cur_mi_row < mi_rows; cur_mi_row++) {
1444 assert(get_mi_grid_idx(mi_params, 0, mi_col + mi_alloc_size_1d) <
1445 mi_params->mi_stride);
1446 const int mi_grid_idx =
1447 get_mi_grid_idx(mi_params, mi_row + cur_mi_row, mi_col);
1448 const int alloc_mi_idx =
1449 get_alloc_mi_idx(mi_params, mi_row + cur_mi_row, mi_col);
1450 memset(&mi_params->mi_grid_base[mi_grid_idx], 0,
1451 sb_size_mi * sizeof(*mi_params->mi_grid_base));
1452 memset(&mi_params->tx_type_map[mi_grid_idx], 0,
1453 sb_size_mi * sizeof(*mi_params->tx_type_map));
1454 if (cur_mi_row % mi_alloc_size_1d == 0) {
1455 memset(&mi_params->mi_alloc[alloc_mi_idx], 0,
1456 sb_size_alloc_mi * sizeof(*mi_params->mi_alloc));
1457 }
1458 }
1459 }
1460
av1_backup_sb_state(SB_FIRST_PASS_STATS * sb_fp_stats,const AV1_COMP * cpi,ThreadData * td,const TileDataEnc * tile_data,int mi_row,int mi_col)1461 void av1_backup_sb_state(SB_FIRST_PASS_STATS *sb_fp_stats, const AV1_COMP *cpi,
1462 ThreadData *td, const TileDataEnc *tile_data,
1463 int mi_row, int mi_col) {
1464 MACROBLOCK *x = &td->mb;
1465 MACROBLOCKD *xd = &x->e_mbd;
1466 const TileInfo *tile_info = &tile_data->tile_info;
1467
1468 const AV1_COMMON *cm = &cpi->common;
1469 const int num_planes = av1_num_planes(cm);
1470 const BLOCK_SIZE sb_size = cm->seq_params->sb_size;
1471
1472 xd->above_txfm_context =
1473 cm->above_contexts.txfm[tile_info->tile_row] + mi_col;
1474 xd->left_txfm_context =
1475 xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
1476 av1_save_context(x, &sb_fp_stats->x_ctx, mi_row, mi_col, sb_size, num_planes);
1477
1478 sb_fp_stats->rd_count = td->rd_counts;
1479 sb_fp_stats->split_count = x->txfm_search_info.txb_split_count;
1480
1481 sb_fp_stats->fc = *td->counts;
1482
1483 // Don't copy in row_mt case, otherwise run into data race. No behavior change
1484 // in row_mt case.
1485 if (cpi->sf.inter_sf.inter_mode_rd_model_estimation == 1) {
1486 memcpy(sb_fp_stats->inter_mode_rd_models, tile_data->inter_mode_rd_models,
1487 sizeof(sb_fp_stats->inter_mode_rd_models));
1488 }
1489
1490 memcpy(sb_fp_stats->thresh_freq_fact, x->thresh_freq_fact,
1491 sizeof(sb_fp_stats->thresh_freq_fact));
1492
1493 const int alloc_mi_idx = get_alloc_mi_idx(&cm->mi_params, mi_row, mi_col);
1494 sb_fp_stats->current_qindex =
1495 cm->mi_params.mi_alloc[alloc_mi_idx].current_qindex;
1496
1497 #if CONFIG_INTERNAL_STATS
1498 memcpy(sb_fp_stats->mode_chosen_counts, cpi->mode_chosen_counts,
1499 sizeof(sb_fp_stats->mode_chosen_counts));
1500 #endif // CONFIG_INTERNAL_STATS
1501 }
1502
av1_restore_sb_state(const SB_FIRST_PASS_STATS * sb_fp_stats,AV1_COMP * cpi,ThreadData * td,TileDataEnc * tile_data,int mi_row,int mi_col)1503 void av1_restore_sb_state(const SB_FIRST_PASS_STATS *sb_fp_stats, AV1_COMP *cpi,
1504 ThreadData *td, TileDataEnc *tile_data, int mi_row,
1505 int mi_col) {
1506 MACROBLOCK *x = &td->mb;
1507
1508 const AV1_COMMON *cm = &cpi->common;
1509 const int num_planes = av1_num_planes(cm);
1510 const BLOCK_SIZE sb_size = cm->seq_params->sb_size;
1511
1512 av1_restore_context(x, &sb_fp_stats->x_ctx, mi_row, mi_col, sb_size,
1513 num_planes);
1514
1515 td->rd_counts = sb_fp_stats->rd_count;
1516 x->txfm_search_info.txb_split_count = sb_fp_stats->split_count;
1517
1518 *td->counts = sb_fp_stats->fc;
1519
1520 if (cpi->sf.inter_sf.inter_mode_rd_model_estimation == 1) {
1521 memcpy(tile_data->inter_mode_rd_models, sb_fp_stats->inter_mode_rd_models,
1522 sizeof(sb_fp_stats->inter_mode_rd_models));
1523 }
1524
1525 memcpy(x->thresh_freq_fact, sb_fp_stats->thresh_freq_fact,
1526 sizeof(sb_fp_stats->thresh_freq_fact));
1527
1528 const int alloc_mi_idx = get_alloc_mi_idx(&cm->mi_params, mi_row, mi_col);
1529 cm->mi_params.mi_alloc[alloc_mi_idx].current_qindex =
1530 sb_fp_stats->current_qindex;
1531
1532 #if CONFIG_INTERNAL_STATS
1533 memcpy(cpi->mode_chosen_counts, sb_fp_stats->mode_chosen_counts,
1534 sizeof(sb_fp_stats->mode_chosen_counts));
1535 #endif // CONFIG_INTERNAL_STATS
1536 }
1537
1538 /*! Checks whether to skip updating the entropy cost based on tile info.
1539 *
1540 * This function contains the common code used to skip the cost update of coeff,
1541 * mode, mv and dv symbols.
1542 */
skip_cost_update(const SequenceHeader * seq_params,const TileInfo * const tile_info,const int mi_row,const int mi_col,INTERNAL_COST_UPDATE_TYPE upd_level)1543 static int skip_cost_update(const SequenceHeader *seq_params,
1544 const TileInfo *const tile_info, const int mi_row,
1545 const int mi_col,
1546 INTERNAL_COST_UPDATE_TYPE upd_level) {
1547 if (upd_level == INTERNAL_COST_UPD_SB) return 0;
1548 if (upd_level == INTERNAL_COST_UPD_OFF) return 1;
1549
1550 // upd_level is at most as frequent as each sb_row in a tile.
1551 if (mi_col != tile_info->mi_col_start) return 1;
1552
1553 if (upd_level == INTERNAL_COST_UPD_SBROW_SET) {
1554 const int mib_size_log2 = seq_params->mib_size_log2;
1555 const int sb_row = (mi_row - tile_info->mi_row_start) >> mib_size_log2;
1556 const int sb_size = seq_params->mib_size * MI_SIZE;
1557 const int tile_height =
1558 (tile_info->mi_row_end - tile_info->mi_row_start) * MI_SIZE;
1559 // When upd_level = INTERNAL_COST_UPD_SBROW_SET, the cost update happens
1560 // once for 2, 4 sb rows for sb size 128, sb size 64 respectively. However,
1561 // as the update will not be equally spaced in smaller resolutions making
1562 // it equally spaced by calculating (mv_num_rows_cost_update) the number of
1563 // rows after which the cost update should happen.
1564 const int sb_size_update_freq_map[2] = { 2, 4 };
1565 const int update_freq_sb_rows =
1566 sb_size_update_freq_map[sb_size != MAX_SB_SIZE];
1567 const int update_freq_num_rows = sb_size * update_freq_sb_rows;
1568 // Round-up the division result to next integer.
1569 const int num_updates_per_tile =
1570 (tile_height + update_freq_num_rows - 1) / update_freq_num_rows;
1571 const int num_rows_update_per_tile = num_updates_per_tile * sb_size;
1572 // Round-up the division result to next integer.
1573 const int num_sb_rows_per_update =
1574 (tile_height + num_rows_update_per_tile - 1) / num_rows_update_per_tile;
1575 if ((sb_row % num_sb_rows_per_update) != 0) return 1;
1576 }
1577 return 0;
1578 }
1579
1580 // Checks for skip status of mv cost update.
skip_mv_cost_update(AV1_COMP * cpi,const TileInfo * const tile_info,const int mi_row,const int mi_col)1581 static int skip_mv_cost_update(AV1_COMP *cpi, const TileInfo *const tile_info,
1582 const int mi_row, const int mi_col) {
1583 const AV1_COMMON *cm = &cpi->common;
1584 // For intra frames, mv cdfs are not updated during the encode. Hence, the mv
1585 // cost calculation is skipped in this case.
1586 if (frame_is_intra_only(cm)) return 1;
1587
1588 return skip_cost_update(cm->seq_params, tile_info, mi_row, mi_col,
1589 cpi->sf.inter_sf.mv_cost_upd_level);
1590 }
1591
1592 // Checks for skip status of dv cost update.
skip_dv_cost_update(AV1_COMP * cpi,const TileInfo * const tile_info,const int mi_row,const int mi_col)1593 static int skip_dv_cost_update(AV1_COMP *cpi, const TileInfo *const tile_info,
1594 const int mi_row, const int mi_col) {
1595 const AV1_COMMON *cm = &cpi->common;
1596 // Intrabc is only applicable to intra frames. So skip if intrabc is not
1597 // allowed.
1598 if (!av1_allow_intrabc(cm) || is_stat_generation_stage(cpi)) {
1599 return 1;
1600 }
1601
1602 return skip_cost_update(cm->seq_params, tile_info, mi_row, mi_col,
1603 cpi->sf.intra_sf.dv_cost_upd_level);
1604 }
1605
1606 // Update the rate costs of some symbols according to the frequency directed
1607 // by speed features
av1_set_cost_upd_freq(AV1_COMP * cpi,ThreadData * td,const TileInfo * const tile_info,const int mi_row,const int mi_col)1608 void av1_set_cost_upd_freq(AV1_COMP *cpi, ThreadData *td,
1609 const TileInfo *const tile_info, const int mi_row,
1610 const int mi_col) {
1611 AV1_COMMON *const cm = &cpi->common;
1612 const int num_planes = av1_num_planes(cm);
1613 MACROBLOCK *const x = &td->mb;
1614 MACROBLOCKD *const xd = &x->e_mbd;
1615
1616 if (cm->features.disable_cdf_update) {
1617 return;
1618 }
1619
1620 switch (cpi->sf.inter_sf.coeff_cost_upd_level) {
1621 case INTERNAL_COST_UPD_OFF:
1622 case INTERNAL_COST_UPD_TILE: // Tile level
1623 break;
1624 case INTERNAL_COST_UPD_SBROW_SET: // SB row set level in tile
1625 case INTERNAL_COST_UPD_SBROW: // SB row level in tile
1626 case INTERNAL_COST_UPD_SB: // SB level
1627 if (skip_cost_update(cm->seq_params, tile_info, mi_row, mi_col,
1628 cpi->sf.inter_sf.coeff_cost_upd_level))
1629 break;
1630 av1_fill_coeff_costs(&x->coeff_costs, xd->tile_ctx, num_planes);
1631 break;
1632 default: assert(0);
1633 }
1634
1635 switch (cpi->sf.inter_sf.mode_cost_upd_level) {
1636 case INTERNAL_COST_UPD_OFF:
1637 case INTERNAL_COST_UPD_TILE: // Tile level
1638 break;
1639 case INTERNAL_COST_UPD_SBROW_SET: // SB row set level in tile
1640 case INTERNAL_COST_UPD_SBROW: // SB row level in tile
1641 case INTERNAL_COST_UPD_SB: // SB level
1642 if (skip_cost_update(cm->seq_params, tile_info, mi_row, mi_col,
1643 cpi->sf.inter_sf.mode_cost_upd_level))
1644 break;
1645 av1_fill_mode_rates(cm, &x->mode_costs, xd->tile_ctx);
1646 break;
1647 default: assert(0);
1648 }
1649
1650 switch (cpi->sf.inter_sf.mv_cost_upd_level) {
1651 case INTERNAL_COST_UPD_OFF:
1652 case INTERNAL_COST_UPD_TILE: // Tile level
1653 break;
1654 case INTERNAL_COST_UPD_SBROW_SET: // SB row set level in tile
1655 case INTERNAL_COST_UPD_SBROW: // SB row level in tile
1656 case INTERNAL_COST_UPD_SB: // SB level
1657 // Checks for skip status of mv cost update.
1658 if (skip_mv_cost_update(cpi, tile_info, mi_row, mi_col)) break;
1659 av1_fill_mv_costs(&xd->tile_ctx->nmvc,
1660 cm->features.cur_frame_force_integer_mv,
1661 cm->features.allow_high_precision_mv, x->mv_costs);
1662 break;
1663 default: assert(0);
1664 }
1665
1666 switch (cpi->sf.intra_sf.dv_cost_upd_level) {
1667 case INTERNAL_COST_UPD_OFF:
1668 case INTERNAL_COST_UPD_TILE: // Tile level
1669 break;
1670 case INTERNAL_COST_UPD_SBROW_SET: // SB row set level in tile
1671 case INTERNAL_COST_UPD_SBROW: // SB row level in tile
1672 case INTERNAL_COST_UPD_SB: // SB level
1673 // Checks for skip status of dv cost update.
1674 if (skip_dv_cost_update(cpi, tile_info, mi_row, mi_col)) break;
1675 av1_fill_dv_costs(&xd->tile_ctx->ndvc, x->dv_costs);
1676 break;
1677 default: assert(0);
1678 }
1679 }
1680
av1_dealloc_src_diff_buf(struct macroblock * mb,int num_planes)1681 void av1_dealloc_src_diff_buf(struct macroblock *mb, int num_planes) {
1682 for (int plane = 0; plane < num_planes; ++plane) {
1683 aom_free(mb->plane[plane].src_diff);
1684 mb->plane[plane].src_diff = NULL;
1685 }
1686 }
1687
av1_alloc_src_diff_buf(const struct AV1Common * cm,struct macroblock * mb)1688 void av1_alloc_src_diff_buf(const struct AV1Common *cm, struct macroblock *mb) {
1689 const int num_planes = av1_num_planes(cm);
1690 #ifndef NDEBUG
1691 for (int plane = 0; plane < num_planes; ++plane) {
1692 assert(!mb->plane[plane].src_diff);
1693 }
1694 #endif
1695 for (int plane = 0; plane < num_planes; ++plane) {
1696 const int subsampling_xy =
1697 plane ? cm->seq_params->subsampling_x + cm->seq_params->subsampling_y
1698 : 0;
1699 const int sb_size = MAX_SB_SQUARE >> subsampling_xy;
1700 CHECK_MEM_ERROR(cm, mb->plane[plane].src_diff,
1701 (int16_t *)aom_memalign(
1702 32, sizeof(*mb->plane[plane].src_diff) * sb_size));
1703 }
1704 }
1705