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 #include <stdint.h>
13
14 #include "config/aom_config.h"
15 #include "config/aom_scale_rtcd.h"
16
17 #include "aom/aom_codec.h"
18 #include "aom/aom_encoder.h"
19
20 #include "aom_ports/system_state.h"
21
22 #include "av1/common/av1_common_int.h"
23
24 #include "av1/encoder/encoder.h"
25 #include "av1/encoder/firstpass.h"
26 #include "av1/encoder/gop_structure.h"
27 #include "av1/encoder/pass2_strategy.h"
28 #include "av1/encoder/ratectrl.h"
29 #include "av1/encoder/tpl_model.h"
30 #include "av1/encoder/use_flat_gop_model_params.h"
31 #include "av1/encoder/encode_strategy.h"
32
33 #define DEFAULT_KF_BOOST 2300
34 #define DEFAULT_GF_BOOST 2000
35 #define GROUP_ADAPTIVE_MAXQ 1
36 static void init_gf_stats(GF_GROUP_STATS *gf_stats);
37
38 // Calculate an active area of the image that discounts formatting
39 // bars and partially discounts other 0 energy areas.
40 #define MIN_ACTIVE_AREA 0.5
41 #define MAX_ACTIVE_AREA 1.0
calculate_active_area(const FRAME_INFO * frame_info,const FIRSTPASS_STATS * this_frame)42 static double calculate_active_area(const FRAME_INFO *frame_info,
43 const FIRSTPASS_STATS *this_frame) {
44 const double active_pct =
45 1.0 -
46 ((this_frame->intra_skip_pct / 2) +
47 ((this_frame->inactive_zone_rows * 2) / (double)frame_info->mb_rows));
48 return fclamp(active_pct, MIN_ACTIVE_AREA, MAX_ACTIVE_AREA);
49 }
50
51 // Calculate a modified Error used in distributing bits between easier and
52 // harder frames.
53 #define ACT_AREA_CORRECTION 0.5
calculate_modified_err(const FRAME_INFO * frame_info,const TWO_PASS * twopass,const AV1EncoderConfig * oxcf,const FIRSTPASS_STATS * this_frame)54 static double calculate_modified_err(const FRAME_INFO *frame_info,
55 const TWO_PASS *twopass,
56 const AV1EncoderConfig *oxcf,
57 const FIRSTPASS_STATS *this_frame) {
58 const FIRSTPASS_STATS *const stats = twopass->stats_buf_ctx->total_stats;
59 if (stats == NULL) {
60 return 0;
61 }
62 const double av_weight = stats->weight / stats->count;
63 const double av_err = (stats->coded_error * av_weight) / stats->count;
64 double modified_error =
65 av_err * pow(this_frame->coded_error * this_frame->weight /
66 DOUBLE_DIVIDE_CHECK(av_err),
67 oxcf->two_pass_vbrbias / 100.0);
68
69 // Correction for active area. Frames with a reduced active area
70 // (eg due to formatting bars) have a higher error per mb for the
71 // remaining active MBs. The correction here assumes that coding
72 // 0.5N blocks of complexity 2X is a little easier than coding N
73 // blocks of complexity X.
74 modified_error *=
75 pow(calculate_active_area(frame_info, this_frame), ACT_AREA_CORRECTION);
76
77 return fclamp(modified_error, twopass->modified_error_min,
78 twopass->modified_error_max);
79 }
80
81 // Resets the first pass file to the given position using a relative seek from
82 // the current position.
reset_fpf_position(TWO_PASS * p,const FIRSTPASS_STATS * position)83 static void reset_fpf_position(TWO_PASS *p, const FIRSTPASS_STATS *position) {
84 p->stats_in = position;
85 }
86
input_stats(TWO_PASS * p,FIRSTPASS_STATS * fps)87 static int input_stats(TWO_PASS *p, FIRSTPASS_STATS *fps) {
88 if (p->stats_in >= p->stats_buf_ctx->stats_in_end) return EOF;
89
90 *fps = *p->stats_in;
91 ++p->stats_in;
92 return 1;
93 }
94
input_stats_lap(TWO_PASS * p,FIRSTPASS_STATS * fps)95 static int input_stats_lap(TWO_PASS *p, FIRSTPASS_STATS *fps) {
96 if (p->stats_in >= p->stats_buf_ctx->stats_in_end) return EOF;
97
98 *fps = *p->stats_in;
99 /* Move old stats[0] out to accommodate for next frame stats */
100 memmove(p->frame_stats_arr[0], p->frame_stats_arr[1],
101 (p->stats_buf_ctx->stats_in_end - p->stats_in - 1) *
102 sizeof(FIRSTPASS_STATS));
103 p->stats_buf_ctx->stats_in_end--;
104 return 1;
105 }
106
107 // Read frame stats at an offset from the current position.
read_frame_stats(const TWO_PASS * p,int offset)108 static const FIRSTPASS_STATS *read_frame_stats(const TWO_PASS *p, int offset) {
109 if ((offset >= 0 && p->stats_in + offset >= p->stats_buf_ctx->stats_in_end) ||
110 (offset < 0 && p->stats_in + offset < p->stats_buf_ctx->stats_in_start)) {
111 return NULL;
112 }
113
114 return &p->stats_in[offset];
115 }
116
subtract_stats(FIRSTPASS_STATS * section,const FIRSTPASS_STATS * frame)117 static void subtract_stats(FIRSTPASS_STATS *section,
118 const FIRSTPASS_STATS *frame) {
119 section->frame -= frame->frame;
120 section->weight -= frame->weight;
121 section->intra_error -= frame->intra_error;
122 section->frame_avg_wavelet_energy -= frame->frame_avg_wavelet_energy;
123 section->coded_error -= frame->coded_error;
124 section->sr_coded_error -= frame->sr_coded_error;
125 section->pcnt_inter -= frame->pcnt_inter;
126 section->pcnt_motion -= frame->pcnt_motion;
127 section->pcnt_second_ref -= frame->pcnt_second_ref;
128 section->pcnt_neutral -= frame->pcnt_neutral;
129 section->intra_skip_pct -= frame->intra_skip_pct;
130 section->inactive_zone_rows -= frame->inactive_zone_rows;
131 section->inactive_zone_cols -= frame->inactive_zone_cols;
132 section->MVr -= frame->MVr;
133 section->mvr_abs -= frame->mvr_abs;
134 section->MVc -= frame->MVc;
135 section->mvc_abs -= frame->mvc_abs;
136 section->MVrv -= frame->MVrv;
137 section->MVcv -= frame->MVcv;
138 section->mv_in_out_count -= frame->mv_in_out_count;
139 section->new_mv_count -= frame->new_mv_count;
140 section->count -= frame->count;
141 section->duration -= frame->duration;
142 }
143
144 // This function returns the maximum target rate per frame.
frame_max_bits(const RATE_CONTROL * rc,const AV1EncoderConfig * oxcf)145 static int frame_max_bits(const RATE_CONTROL *rc,
146 const AV1EncoderConfig *oxcf) {
147 int64_t max_bits = ((int64_t)rc->avg_frame_bandwidth *
148 (int64_t)oxcf->two_pass_vbrmax_section) /
149 100;
150 if (max_bits < 0)
151 max_bits = 0;
152 else if (max_bits > rc->max_frame_bandwidth)
153 max_bits = rc->max_frame_bandwidth;
154
155 return (int)max_bits;
156 }
157
158 static const double q_pow_term[(QINDEX_RANGE >> 5) + 1] = { 0.65, 0.70, 0.75,
159 0.80, 0.85, 0.90,
160 0.95, 0.95, 0.95 };
161 #define ERR_DIVISOR 96.0
calc_correction_factor(double err_per_mb,int q)162 static double calc_correction_factor(double err_per_mb, int q) {
163 const double error_term = err_per_mb / ERR_DIVISOR;
164 const int index = q >> 5;
165 // Adjustment to power term based on qindex
166 const double power_term =
167 q_pow_term[index] +
168 (((q_pow_term[index + 1] - q_pow_term[index]) * (q % 32)) / 32.0);
169 assert(error_term >= 0.0);
170 return fclamp(pow(error_term, power_term), 0.05, 5.0);
171 }
172
twopass_update_bpm_factor(TWO_PASS * twopass)173 static void twopass_update_bpm_factor(TWO_PASS *twopass) {
174 // Based on recent history adjust expectations of bits per macroblock.
175 double last_group_rate_err =
176 (double)twopass->rolling_arf_group_actual_bits /
177 DOUBLE_DIVIDE_CHECK((double)twopass->rolling_arf_group_target_bits);
178 last_group_rate_err = AOMMAX(0.25, AOMMIN(4.0, last_group_rate_err));
179 twopass->bpm_factor *= (3.0 + last_group_rate_err) / 4.0;
180 twopass->bpm_factor = AOMMAX(0.25, AOMMIN(4.0, twopass->bpm_factor));
181 }
182
qbpm_enumerator(int rate_err_tol)183 static int qbpm_enumerator(int rate_err_tol) {
184 return 1350000 + ((300000 * AOMMIN(75, AOMMAX(rate_err_tol - 25, 0))) / 75);
185 }
186
187 // Similar to find_qindex_by_rate() function in ratectrl.c, but includes
188 // calculation of a correction_factor.
find_qindex_by_rate_with_correction(int desired_bits_per_mb,aom_bit_depth_t bit_depth,double error_per_mb,double group_weight_factor,int rate_err_tol,int best_qindex,int worst_qindex)189 static int find_qindex_by_rate_with_correction(
190 int desired_bits_per_mb, aom_bit_depth_t bit_depth, double error_per_mb,
191 double group_weight_factor, int rate_err_tol, int best_qindex,
192 int worst_qindex) {
193 assert(best_qindex <= worst_qindex);
194 int low = best_qindex;
195 int high = worst_qindex;
196
197 while (low < high) {
198 const int mid = (low + high) >> 1;
199 const double mid_factor = calc_correction_factor(error_per_mb, mid);
200 const double q = av1_convert_qindex_to_q(mid, bit_depth);
201 const int enumerator = qbpm_enumerator(rate_err_tol);
202 const int mid_bits_per_mb =
203 (int)((enumerator * mid_factor * group_weight_factor) / q);
204
205 if (mid_bits_per_mb > desired_bits_per_mb) {
206 low = mid + 1;
207 } else {
208 high = mid;
209 }
210 }
211 return low;
212 }
213
get_twopass_worst_quality(AV1_COMP * cpi,const double section_err,double inactive_zone,int section_target_bandwidth,double group_weight_factor)214 static int get_twopass_worst_quality(AV1_COMP *cpi, const double section_err,
215 double inactive_zone,
216 int section_target_bandwidth,
217 double group_weight_factor) {
218 const RATE_CONTROL *const rc = &cpi->rc;
219 const AV1EncoderConfig *const oxcf = &cpi->oxcf;
220
221 inactive_zone = fclamp(inactive_zone, 0.0, 1.0);
222
223 if (section_target_bandwidth <= 0) {
224 return rc->worst_quality; // Highest value allowed
225 } else {
226 const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
227 ? cpi->initial_mbs
228 : cpi->common.mi_params.MBs;
229 const int active_mbs = AOMMAX(1, num_mbs - (int)(num_mbs * inactive_zone));
230 const double av_err_per_mb = section_err / active_mbs;
231 const int target_norm_bits_per_mb =
232 (int)((uint64_t)section_target_bandwidth << BPER_MB_NORMBITS) /
233 active_mbs;
234 int rate_err_tol =
235 AOMMIN(cpi->oxcf.under_shoot_pct, cpi->oxcf.over_shoot_pct);
236
237 twopass_update_bpm_factor(&cpi->twopass);
238 // Try and pick a max Q that will be high enough to encode the
239 // content at the given rate.
240 int q = find_qindex_by_rate_with_correction(
241 target_norm_bits_per_mb, cpi->common.seq_params.bit_depth,
242 av_err_per_mb, group_weight_factor, rate_err_tol, rc->best_quality,
243 rc->worst_quality);
244
245 // Restriction on active max q for constrained quality mode.
246 if (cpi->oxcf.rc_mode == AOM_CQ) q = AOMMAX(q, oxcf->cq_level);
247 return q;
248 }
249 }
250
251 #define SR_DIFF_PART 0.0015
252 #define MOTION_AMP_PART 0.003
253 #define INTRA_PART 0.005
254 #define DEFAULT_DECAY_LIMIT 0.75
255 #define LOW_SR_DIFF_TRHESH 0.1
256 #define SR_DIFF_MAX 128.0
257 #define NCOUNT_FRAME_II_THRESH 5.0
258
get_sr_decay_rate(const FRAME_INFO * frame_info,const FIRSTPASS_STATS * frame)259 static double get_sr_decay_rate(const FRAME_INFO *frame_info,
260 const FIRSTPASS_STATS *frame) {
261 const int num_mbs = frame_info->num_mbs;
262 double sr_diff = (frame->sr_coded_error - frame->coded_error) / num_mbs;
263 double sr_decay = 1.0;
264 double modified_pct_inter;
265 double modified_pcnt_intra;
266 const double motion_amplitude_factor =
267 frame->pcnt_motion * ((frame->mvc_abs + frame->mvr_abs) / 2);
268
269 modified_pct_inter = frame->pcnt_inter;
270 if ((frame->intra_error / DOUBLE_DIVIDE_CHECK(frame->coded_error)) <
271 (double)NCOUNT_FRAME_II_THRESH) {
272 modified_pct_inter = frame->pcnt_inter - frame->pcnt_neutral;
273 }
274 modified_pcnt_intra = 100 * (1.0 - modified_pct_inter);
275
276 if ((sr_diff > LOW_SR_DIFF_TRHESH)) {
277 sr_diff = AOMMIN(sr_diff, SR_DIFF_MAX);
278 sr_decay = 1.0 - (SR_DIFF_PART * sr_diff) -
279 (MOTION_AMP_PART * motion_amplitude_factor) -
280 (INTRA_PART * modified_pcnt_intra);
281 }
282 return AOMMAX(sr_decay, AOMMIN(DEFAULT_DECAY_LIMIT, modified_pct_inter));
283 }
284
285 // This function gives an estimate of how badly we believe the prediction
286 // quality is decaying from frame to frame.
get_zero_motion_factor(const FRAME_INFO * frame_info,const FIRSTPASS_STATS * frame)287 static double get_zero_motion_factor(const FRAME_INFO *frame_info,
288 const FIRSTPASS_STATS *frame) {
289 const double zero_motion_pct = frame->pcnt_inter - frame->pcnt_motion;
290 double sr_decay = get_sr_decay_rate(frame_info, frame);
291 return AOMMIN(sr_decay, zero_motion_pct);
292 }
293
294 #define ZM_POWER_FACTOR 0.75
295
get_prediction_decay_rate(const FRAME_INFO * frame_info,const FIRSTPASS_STATS * next_frame)296 static double get_prediction_decay_rate(const FRAME_INFO *frame_info,
297 const FIRSTPASS_STATS *next_frame) {
298 const double sr_decay_rate = get_sr_decay_rate(frame_info, next_frame);
299 const double zero_motion_factor =
300 (0.95 * pow((next_frame->pcnt_inter - next_frame->pcnt_motion),
301 ZM_POWER_FACTOR));
302
303 return AOMMAX(zero_motion_factor,
304 (sr_decay_rate + ((1.0 - sr_decay_rate) * zero_motion_factor)));
305 }
306
307 // Function to test for a condition where a complex transition is followed
308 // by a static section. For example in slide shows where there is a fade
309 // between slides. This is to help with more optimal kf and gf positioning.
detect_transition_to_still(TWO_PASS * const twopass,const int min_gf_interval,const int frame_interval,const int still_interval,const double loop_decay_rate,const double last_decay_rate)310 static int detect_transition_to_still(TWO_PASS *const twopass,
311 const int min_gf_interval,
312 const int frame_interval,
313 const int still_interval,
314 const double loop_decay_rate,
315 const double last_decay_rate) {
316 // Break clause to detect very still sections after motion
317 // For example a static image after a fade or other transition
318 // instead of a clean scene cut.
319 if (frame_interval > min_gf_interval && loop_decay_rate >= 0.999 &&
320 last_decay_rate < 0.9) {
321 int j;
322 // Look ahead a few frames to see if static condition persists...
323 for (j = 0; j < still_interval; ++j) {
324 const FIRSTPASS_STATS *stats = &twopass->stats_in[j];
325 if (stats >= twopass->stats_buf_ctx->stats_in_end) break;
326
327 if (stats->pcnt_inter - stats->pcnt_motion < 0.999) break;
328 }
329 // Only if it does do we signal a transition to still.
330 return j == still_interval;
331 }
332 return 0;
333 }
334
335 // This function detects a flash through the high relative pcnt_second_ref
336 // score in the frame following a flash frame. The offset passed in should
337 // reflect this.
detect_flash(const TWO_PASS * twopass,const int offset)338 static int detect_flash(const TWO_PASS *twopass, const int offset) {
339 const FIRSTPASS_STATS *const next_frame = read_frame_stats(twopass, offset);
340
341 // What we are looking for here is a situation where there is a
342 // brief break in prediction (such as a flash) but subsequent frames
343 // are reasonably well predicted by an earlier (pre flash) frame.
344 // The recovery after a flash is indicated by a high pcnt_second_ref
345 // compared to pcnt_inter.
346 return next_frame != NULL &&
347 next_frame->pcnt_second_ref > next_frame->pcnt_inter &&
348 next_frame->pcnt_second_ref >= 0.5;
349 }
350
351 // Update the motion related elements to the GF arf boost calculation.
accumulate_frame_motion_stats(const FIRSTPASS_STATS * stats,GF_GROUP_STATS * gf_stats)352 static void accumulate_frame_motion_stats(const FIRSTPASS_STATS *stats,
353 GF_GROUP_STATS *gf_stats) {
354 const double pct = stats->pcnt_motion;
355
356 // Accumulate Motion In/Out of frame stats.
357 gf_stats->this_frame_mv_in_out = stats->mv_in_out_count * pct;
358 gf_stats->mv_in_out_accumulator += gf_stats->this_frame_mv_in_out;
359 gf_stats->abs_mv_in_out_accumulator += fabs(gf_stats->this_frame_mv_in_out);
360
361 // Accumulate a measure of how uniform (or conversely how random) the motion
362 // field is (a ratio of abs(mv) / mv).
363 if (pct > 0.05) {
364 const double mvr_ratio =
365 fabs(stats->mvr_abs) / DOUBLE_DIVIDE_CHECK(fabs(stats->MVr));
366 const double mvc_ratio =
367 fabs(stats->mvc_abs) / DOUBLE_DIVIDE_CHECK(fabs(stats->MVc));
368
369 gf_stats->mv_ratio_accumulator +=
370 pct * (mvr_ratio < stats->mvr_abs ? mvr_ratio : stats->mvr_abs);
371 gf_stats->mv_ratio_accumulator +=
372 pct * (mvc_ratio < stats->mvc_abs ? mvc_ratio : stats->mvc_abs);
373 }
374 }
375
accumulate_this_frame_stats(const FIRSTPASS_STATS * stats,const double mod_frame_err,GF_GROUP_STATS * gf_stats)376 static void accumulate_this_frame_stats(const FIRSTPASS_STATS *stats,
377 const double mod_frame_err,
378 GF_GROUP_STATS *gf_stats) {
379 gf_stats->gf_group_err += mod_frame_err;
380 #if GROUP_ADAPTIVE_MAXQ
381 gf_stats->gf_group_raw_error += stats->coded_error;
382 #endif
383 gf_stats->gf_group_skip_pct += stats->intra_skip_pct;
384 gf_stats->gf_group_inactive_zone_rows += stats->inactive_zone_rows;
385 }
386
accumulate_next_frame_stats(const FIRSTPASS_STATS * stats,const FRAME_INFO * frame_info,TWO_PASS * const twopass,const int flash_detected,const int frames_since_key,const int cur_idx,const int can_disable_arf,const int min_gf_interval,GF_GROUP_STATS * gf_stats)387 static void accumulate_next_frame_stats(
388 const FIRSTPASS_STATS *stats, const FRAME_INFO *frame_info,
389 TWO_PASS *const twopass, const int flash_detected,
390 const int frames_since_key, const int cur_idx, const int can_disable_arf,
391 const int min_gf_interval, GF_GROUP_STATS *gf_stats) {
392 accumulate_frame_motion_stats(stats, gf_stats);
393 // sum up the metric values of current gf group
394 gf_stats->avg_sr_coded_error += stats->sr_coded_error;
395 gf_stats->avg_tr_coded_error += stats->tr_coded_error;
396 gf_stats->avg_pcnt_second_ref += stats->pcnt_second_ref;
397 gf_stats->avg_pcnt_third_ref += stats->pcnt_third_ref;
398 gf_stats->avg_new_mv_count += stats->new_mv_count;
399 gf_stats->avg_wavelet_energy += stats->frame_avg_wavelet_energy;
400 if (fabs(stats->raw_error_stdev) > 0.000001) {
401 gf_stats->non_zero_stdev_count++;
402 gf_stats->avg_raw_err_stdev += stats->raw_error_stdev;
403 }
404
405 // Accumulate the effect of prediction quality decay
406 if (!flash_detected) {
407 gf_stats->last_loop_decay_rate = gf_stats->loop_decay_rate;
408 gf_stats->loop_decay_rate = get_prediction_decay_rate(frame_info, stats);
409
410 gf_stats->decay_accumulator =
411 gf_stats->decay_accumulator * gf_stats->loop_decay_rate;
412
413 // Monitor for static sections.
414 if ((frames_since_key + cur_idx - 1) > 1) {
415 gf_stats->zero_motion_accumulator =
416 AOMMIN(gf_stats->zero_motion_accumulator,
417 get_zero_motion_factor(frame_info, stats));
418 }
419
420 // Break clause to detect very still sections after motion. For example,
421 // a static image after a fade or other transition.
422 if (can_disable_arf &&
423 detect_transition_to_still(twopass, min_gf_interval, cur_idx, 5,
424 gf_stats->loop_decay_rate,
425 gf_stats->last_loop_decay_rate)) {
426 gf_stats->allow_alt_ref = 0;
427 }
428 }
429 }
430
average_gf_stats(const int total_frame,const FIRSTPASS_STATS * last_stat,GF_GROUP_STATS * gf_stats)431 static void average_gf_stats(const int total_frame,
432 const FIRSTPASS_STATS *last_stat,
433 GF_GROUP_STATS *gf_stats) {
434 if (total_frame) {
435 gf_stats->avg_sr_coded_error /= total_frame;
436 gf_stats->avg_tr_coded_error /= total_frame;
437 gf_stats->avg_pcnt_second_ref /= total_frame;
438 if (total_frame - 1) {
439 gf_stats->avg_pcnt_third_ref_nolast =
440 (gf_stats->avg_pcnt_third_ref - last_stat->pcnt_third_ref) /
441 (total_frame - 1);
442 } else {
443 gf_stats->avg_pcnt_third_ref_nolast =
444 gf_stats->avg_pcnt_third_ref / total_frame;
445 }
446 gf_stats->avg_pcnt_third_ref /= total_frame;
447 gf_stats->avg_new_mv_count /= total_frame;
448 gf_stats->avg_wavelet_energy /= total_frame;
449 }
450
451 if (gf_stats->non_zero_stdev_count)
452 gf_stats->avg_raw_err_stdev /= gf_stats->non_zero_stdev_count;
453 }
454
get_features_from_gf_stats(const GF_GROUP_STATS * gf_stats,const GF_FRAME_STATS * first_frame,const GF_FRAME_STATS * last_frame,const int num_mbs,const int constrained_gf_group,const int kf_zeromotion_pct,const int num_frames,float * features)455 static void get_features_from_gf_stats(const GF_GROUP_STATS *gf_stats,
456 const GF_FRAME_STATS *first_frame,
457 const GF_FRAME_STATS *last_frame,
458 const int num_mbs,
459 const int constrained_gf_group,
460 const int kf_zeromotion_pct,
461 const int num_frames, float *features) {
462 *features++ = (float)gf_stats->abs_mv_in_out_accumulator;
463 *features++ = (float)(gf_stats->avg_new_mv_count / num_mbs);
464 *features++ = (float)gf_stats->avg_pcnt_second_ref;
465 *features++ = (float)gf_stats->avg_pcnt_third_ref;
466 *features++ = (float)gf_stats->avg_pcnt_third_ref_nolast;
467 *features++ = (float)(gf_stats->avg_sr_coded_error / num_mbs);
468 *features++ = (float)(gf_stats->avg_tr_coded_error / num_mbs);
469 *features++ = (float)(gf_stats->avg_wavelet_energy / num_mbs);
470 *features++ = (float)(constrained_gf_group);
471 *features++ = (float)gf_stats->decay_accumulator;
472 *features++ = (float)(first_frame->frame_coded_error / num_mbs);
473 *features++ = (float)(first_frame->frame_sr_coded_error / num_mbs);
474 *features++ = (float)(first_frame->frame_tr_coded_error / num_mbs);
475 *features++ = (float)(first_frame->frame_err / num_mbs);
476 *features++ = (float)(kf_zeromotion_pct);
477 *features++ = (float)(last_frame->frame_coded_error / num_mbs);
478 *features++ = (float)(last_frame->frame_sr_coded_error / num_mbs);
479 *features++ = (float)(last_frame->frame_tr_coded_error / num_mbs);
480 *features++ = (float)num_frames;
481 *features++ = (float)gf_stats->mv_ratio_accumulator;
482 *features++ = (float)gf_stats->non_zero_stdev_count;
483 }
484
485 #define BOOST_FACTOR 12.5
baseline_err_per_mb(const FRAME_INFO * frame_info)486 static double baseline_err_per_mb(const FRAME_INFO *frame_info) {
487 unsigned int screen_area = frame_info->frame_height * frame_info->frame_width;
488
489 // Use a different error per mb factor for calculating boost for
490 // different formats.
491 if (screen_area <= 640 * 360) {
492 return 500.0;
493 } else {
494 return 1000.0;
495 }
496 }
497
calc_frame_boost(const RATE_CONTROL * rc,const FRAME_INFO * frame_info,const FIRSTPASS_STATS * this_frame,double this_frame_mv_in_out,double max_boost)498 static double calc_frame_boost(const RATE_CONTROL *rc,
499 const FRAME_INFO *frame_info,
500 const FIRSTPASS_STATS *this_frame,
501 double this_frame_mv_in_out, double max_boost) {
502 double frame_boost;
503 const double lq = av1_convert_qindex_to_q(rc->avg_frame_qindex[INTER_FRAME],
504 frame_info->bit_depth);
505 const double boost_q_correction = AOMMIN((0.5 + (lq * 0.015)), 1.5);
506 const double active_area = calculate_active_area(frame_info, this_frame);
507 int num_mbs = frame_info->num_mbs;
508
509 // Correct for any inactive region in the image
510 num_mbs = (int)AOMMAX(1, num_mbs * active_area);
511
512 // Underlying boost factor is based on inter error ratio.
513 frame_boost = AOMMAX(baseline_err_per_mb(frame_info) * num_mbs,
514 this_frame->intra_error * active_area) /
515 DOUBLE_DIVIDE_CHECK(this_frame->coded_error);
516 frame_boost = frame_boost * BOOST_FACTOR * boost_q_correction;
517
518 // Increase boost for frames where new data coming into frame (e.g. zoom out).
519 // Slightly reduce boost if there is a net balance of motion out of the frame
520 // (zoom in). The range for this_frame_mv_in_out is -1.0 to +1.0.
521 if (this_frame_mv_in_out > 0.0)
522 frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
523 // In the extreme case the boost is halved.
524 else
525 frame_boost += frame_boost * (this_frame_mv_in_out / 2.0);
526
527 return AOMMIN(frame_boost, max_boost * boost_q_correction);
528 }
529
calc_kf_frame_boost(const RATE_CONTROL * rc,const FRAME_INFO * frame_info,const FIRSTPASS_STATS * this_frame,double * sr_accumulator,double max_boost)530 static double calc_kf_frame_boost(const RATE_CONTROL *rc,
531 const FRAME_INFO *frame_info,
532 const FIRSTPASS_STATS *this_frame,
533 double *sr_accumulator, double max_boost) {
534 double frame_boost;
535 const double lq = av1_convert_qindex_to_q(rc->avg_frame_qindex[INTER_FRAME],
536 frame_info->bit_depth);
537 const double boost_q_correction = AOMMIN((0.50 + (lq * 0.015)), 2.00);
538 const double active_area = calculate_active_area(frame_info, this_frame);
539 int num_mbs = frame_info->num_mbs;
540
541 // Correct for any inactive region in the image
542 num_mbs = (int)AOMMAX(1, num_mbs * active_area);
543
544 // Underlying boost factor is based on inter error ratio.
545 frame_boost = AOMMAX(baseline_err_per_mb(frame_info) * num_mbs,
546 this_frame->intra_error * active_area) /
547 DOUBLE_DIVIDE_CHECK(
548 (this_frame->coded_error + *sr_accumulator) * active_area);
549
550 // Update the accumulator for second ref error difference.
551 // This is intended to give an indication of how much the coded error is
552 // increasing over time.
553 *sr_accumulator += (this_frame->sr_coded_error - this_frame->coded_error);
554 *sr_accumulator = AOMMAX(0.0, *sr_accumulator);
555
556 // Q correction and scaling
557 // The 40.0 value here is an experimentally derived baseline minimum.
558 // This value is in line with the minimum per frame boost in the alt_ref
559 // boost calculation.
560 frame_boost = ((frame_boost + 40.0) * boost_q_correction);
561
562 return AOMMIN(frame_boost, max_boost * boost_q_correction);
563 }
564
get_projected_gfu_boost(const RATE_CONTROL * rc,int gfu_boost,int frames_to_project,int num_stats_used_for_gfu_boost)565 static int get_projected_gfu_boost(const RATE_CONTROL *rc, int gfu_boost,
566 int frames_to_project,
567 int num_stats_used_for_gfu_boost) {
568 /*
569 * If frames_to_project is equal to num_stats_used_for_gfu_boost,
570 * it means that gfu_boost was calculated over frames_to_project to
571 * begin with(ie; all stats required were available), hence return
572 * the original boost.
573 */
574 if (num_stats_used_for_gfu_boost >= frames_to_project) return gfu_boost;
575
576 double min_boost_factor = sqrt(rc->baseline_gf_interval);
577 // Get the current tpl factor (number of frames = frames_to_project).
578 double tpl_factor = av1_get_gfu_boost_projection_factor(
579 min_boost_factor, MAX_GFUBOOST_FACTOR, frames_to_project);
580 // Get the tpl factor when number of frames = num_stats_used_for_prior_boost.
581 double tpl_factor_num_stats = av1_get_gfu_boost_projection_factor(
582 min_boost_factor, MAX_GFUBOOST_FACTOR, num_stats_used_for_gfu_boost);
583 int projected_gfu_boost =
584 (int)rint((tpl_factor * gfu_boost) / tpl_factor_num_stats);
585 return projected_gfu_boost;
586 }
587
588 #define GF_MAX_BOOST 90.0
589 #define MIN_DECAY_FACTOR 0.01
av1_calc_arf_boost(const TWO_PASS * twopass,const RATE_CONTROL * rc,FRAME_INFO * frame_info,int offset,int f_frames,int b_frames,int * num_fpstats_used,int * num_fpstats_required)590 int av1_calc_arf_boost(const TWO_PASS *twopass, const RATE_CONTROL *rc,
591 FRAME_INFO *frame_info, int offset, int f_frames,
592 int b_frames, int *num_fpstats_used,
593 int *num_fpstats_required) {
594 int i;
595 GF_GROUP_STATS gf_stats;
596 init_gf_stats(&gf_stats);
597 double boost_score = (double)NORMAL_BOOST;
598 int arf_boost;
599 int flash_detected = 0;
600 if (num_fpstats_used) *num_fpstats_used = 0;
601
602 // Search forward from the proposed arf/next gf position.
603 for (i = 0; i < f_frames; ++i) {
604 const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
605 if (this_frame == NULL) break;
606
607 // Update the motion related elements to the boost calculation.
608 accumulate_frame_motion_stats(this_frame, &gf_stats);
609
610 // We want to discount the flash frame itself and the recovery
611 // frame that follows as both will have poor scores.
612 flash_detected = detect_flash(twopass, i + offset) ||
613 detect_flash(twopass, i + offset + 1);
614
615 // Accumulate the effect of prediction quality decay.
616 if (!flash_detected) {
617 gf_stats.decay_accumulator *=
618 get_prediction_decay_rate(frame_info, this_frame);
619 gf_stats.decay_accumulator = gf_stats.decay_accumulator < MIN_DECAY_FACTOR
620 ? MIN_DECAY_FACTOR
621 : gf_stats.decay_accumulator;
622 }
623
624 boost_score +=
625 gf_stats.decay_accumulator *
626 calc_frame_boost(rc, frame_info, this_frame,
627 gf_stats.this_frame_mv_in_out, GF_MAX_BOOST);
628 if (num_fpstats_used) (*num_fpstats_used)++;
629 }
630
631 arf_boost = (int)boost_score;
632
633 // Reset for backward looking loop.
634 boost_score = 0.0;
635 init_gf_stats(&gf_stats);
636 // Search backward towards last gf position.
637 for (i = -1; i >= -b_frames; --i) {
638 const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
639 if (this_frame == NULL) break;
640
641 // Update the motion related elements to the boost calculation.
642 accumulate_frame_motion_stats(this_frame, &gf_stats);
643
644 // We want to discount the the flash frame itself and the recovery
645 // frame that follows as both will have poor scores.
646 flash_detected = detect_flash(twopass, i + offset) ||
647 detect_flash(twopass, i + offset + 1);
648
649 // Cumulative effect of prediction quality decay.
650 if (!flash_detected) {
651 gf_stats.decay_accumulator *=
652 get_prediction_decay_rate(frame_info, this_frame);
653 gf_stats.decay_accumulator = gf_stats.decay_accumulator < MIN_DECAY_FACTOR
654 ? MIN_DECAY_FACTOR
655 : gf_stats.decay_accumulator;
656 }
657
658 boost_score +=
659 gf_stats.decay_accumulator *
660 calc_frame_boost(rc, frame_info, this_frame,
661 gf_stats.this_frame_mv_in_out, GF_MAX_BOOST);
662 if (num_fpstats_used) (*num_fpstats_used)++;
663 }
664 arf_boost += (int)boost_score;
665
666 if (num_fpstats_required) {
667 *num_fpstats_required = f_frames + b_frames;
668 if (num_fpstats_used) {
669 arf_boost = get_projected_gfu_boost(rc, arf_boost, *num_fpstats_required,
670 *num_fpstats_used);
671 }
672 }
673
674 if (arf_boost < ((b_frames + f_frames) * 50))
675 arf_boost = ((b_frames + f_frames) * 50);
676
677 return arf_boost;
678 }
679
680 // Calculate a section intra ratio used in setting max loop filter.
calculate_section_intra_ratio(const FIRSTPASS_STATS * begin,const FIRSTPASS_STATS * end,int section_length)681 static int calculate_section_intra_ratio(const FIRSTPASS_STATS *begin,
682 const FIRSTPASS_STATS *end,
683 int section_length) {
684 const FIRSTPASS_STATS *s = begin;
685 double intra_error = 0.0;
686 double coded_error = 0.0;
687 int i = 0;
688
689 while (s < end && i < section_length) {
690 intra_error += s->intra_error;
691 coded_error += s->coded_error;
692 ++s;
693 ++i;
694 }
695
696 return (int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
697 }
698
699 // Calculate the total bits to allocate in this GF/ARF group.
calculate_total_gf_group_bits(AV1_COMP * cpi,double gf_group_err)700 static int64_t calculate_total_gf_group_bits(AV1_COMP *cpi,
701 double gf_group_err) {
702 const RATE_CONTROL *const rc = &cpi->rc;
703 const TWO_PASS *const twopass = &cpi->twopass;
704 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
705 int64_t total_group_bits;
706
707 // Calculate the bits to be allocated to the group as a whole.
708 if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0)) {
709 total_group_bits = (int64_t)(twopass->kf_group_bits *
710 (gf_group_err / twopass->kf_group_error_left));
711 } else {
712 total_group_bits = 0;
713 }
714
715 // Clamp odd edge cases.
716 total_group_bits = (total_group_bits < 0)
717 ? 0
718 : (total_group_bits > twopass->kf_group_bits)
719 ? twopass->kf_group_bits
720 : total_group_bits;
721
722 // Clip based on user supplied data rate variability limit.
723 if (total_group_bits > (int64_t)max_bits * rc->baseline_gf_interval)
724 total_group_bits = (int64_t)max_bits * rc->baseline_gf_interval;
725
726 return total_group_bits;
727 }
728
729 // Calculate the number of bits to assign to boosted frames in a group.
calculate_boost_bits(int frame_count,int boost,int64_t total_group_bits)730 static int calculate_boost_bits(int frame_count, int boost,
731 int64_t total_group_bits) {
732 int allocation_chunks;
733
734 // return 0 for invalid inputs (could arise e.g. through rounding errors)
735 if (!boost || (total_group_bits <= 0)) return 0;
736
737 if (frame_count <= 0) return (int)(AOMMIN(total_group_bits, INT_MAX));
738
739 allocation_chunks = (frame_count * 100) + boost;
740
741 // Prevent overflow.
742 if (boost > 1023) {
743 int divisor = boost >> 10;
744 boost /= divisor;
745 allocation_chunks /= divisor;
746 }
747
748 // Calculate the number of extra bits for use in the boosted frame or frames.
749 return AOMMAX((int)(((int64_t)boost * total_group_bits) / allocation_chunks),
750 0);
751 }
752
753 // Calculate the boost factor based on the number of bits assigned, i.e. the
754 // inverse of calculate_boost_bits().
calculate_boost_factor(int frame_count,int bits,int64_t total_group_bits)755 static int calculate_boost_factor(int frame_count, int bits,
756 int64_t total_group_bits) {
757 aom_clear_system_state();
758 return (int)(100.0 * frame_count * bits / (total_group_bits - bits));
759 }
760
761 // Reduce the number of bits assigned to keyframe or arf if necessary, to
762 // prevent bitrate spikes that may break level constraints.
763 // frame_type: 0: keyframe; 1: arf.
adjust_boost_bits_for_target_level(const AV1_COMP * const cpi,RATE_CONTROL * const rc,int bits_assigned,int64_t group_bits,int frame_type)764 static int adjust_boost_bits_for_target_level(const AV1_COMP *const cpi,
765 RATE_CONTROL *const rc,
766 int bits_assigned,
767 int64_t group_bits,
768 int frame_type) {
769 const AV1_COMMON *const cm = &cpi->common;
770 const SequenceHeader *const seq_params = &cm->seq_params;
771 const int temporal_layer_id = cm->temporal_layer_id;
772 const int spatial_layer_id = cm->spatial_layer_id;
773 for (int index = 0; index < seq_params->operating_points_cnt_minus_1 + 1;
774 ++index) {
775 if (!is_in_operating_point(seq_params->operating_point_idc[index],
776 temporal_layer_id, spatial_layer_id)) {
777 continue;
778 }
779
780 const AV1_LEVEL target_level =
781 cpi->level_params.target_seq_level_idx[index];
782 if (target_level >= SEQ_LEVELS) continue;
783
784 assert(is_valid_seq_level_idx(target_level));
785
786 const double level_bitrate_limit = av1_get_max_bitrate_for_level(
787 target_level, seq_params->tier[0], seq_params->profile);
788 const int target_bits_per_frame =
789 (int)(level_bitrate_limit / cpi->framerate);
790 if (frame_type == 0) {
791 // Maximum bits for keyframe is 8 times the target_bits_per_frame.
792 const int level_enforced_max_kf_bits = target_bits_per_frame * 8;
793 if (bits_assigned > level_enforced_max_kf_bits) {
794 const int frames = rc->frames_to_key - 1;
795 rc->kf_boost = calculate_boost_factor(
796 frames, level_enforced_max_kf_bits, group_bits);
797 bits_assigned = calculate_boost_bits(frames, rc->kf_boost, group_bits);
798 }
799 } else if (frame_type == 1) {
800 // Maximum bits for arf is 4 times the target_bits_per_frame.
801 const int level_enforced_max_arf_bits = target_bits_per_frame * 4;
802 if (bits_assigned > level_enforced_max_arf_bits) {
803 rc->gfu_boost = calculate_boost_factor(
804 rc->baseline_gf_interval, level_enforced_max_arf_bits, group_bits);
805 bits_assigned = calculate_boost_bits(rc->baseline_gf_interval,
806 rc->gfu_boost, group_bits);
807 }
808 } else {
809 assert(0);
810 }
811 }
812
813 return bits_assigned;
814 }
815
816 // Compile time switch on alternate algorithm to allocate bits in ARF groups
817 // #define ALT_ARF_ALLOCATION
818 #ifdef ALT_ARF_ALLOCATION
819 double layer_fraction[MAX_ARF_LAYERS + 1] = { 1.0, 0.70, 0.55, 0.60,
820 0.60, 1.0, 1.0 };
allocate_gf_group_bits(GF_GROUP * gf_group,RATE_CONTROL * const rc,int64_t gf_group_bits,int gf_arf_bits,int key_frame,int use_arf)821 static void allocate_gf_group_bits(GF_GROUP *gf_group, RATE_CONTROL *const rc,
822 int64_t gf_group_bits, int gf_arf_bits,
823 int key_frame, int use_arf) {
824 int64_t total_group_bits = gf_group_bits;
825 int base_frame_bits;
826 const int gf_group_size = gf_group->size;
827 int layer_frames[MAX_ARF_LAYERS + 1] = { 0 };
828
829 // Subtract the extra bits set aside for ARF frames from the Group Total
830 if (use_arf || !key_frame) total_group_bits -= gf_arf_bits;
831
832 if (rc->baseline_gf_interval)
833 base_frame_bits = (int)(total_group_bits / rc->baseline_gf_interval);
834 else
835 base_frame_bits = (int)1;
836
837 // For key frames the frame target rate is already set and it
838 // is also the golden frame.
839 // === [frame_index == 0] ===
840 int frame_index = 0;
841 if (!key_frame) {
842 if (rc->source_alt_ref_active)
843 gf_group->bit_allocation[frame_index] = 0;
844 else
845 gf_group->bit_allocation[frame_index] =
846 base_frame_bits + (int)(gf_arf_bits * layer_fraction[1]);
847 }
848 frame_index++;
849
850 // Check the number of frames in each layer in case we have a
851 // non standard group length.
852 int max_arf_layer = gf_group->max_layer_depth - 1;
853 for (int idx = frame_index; idx < gf_group_size; ++idx) {
854 if ((gf_group->update_type[idx] == ARF_UPDATE) ||
855 (gf_group->update_type[idx] == INTNL_ARF_UPDATE)) {
856 // max_arf_layer = AOMMAX(max_arf_layer, gf_group->layer_depth[idx]);
857 layer_frames[gf_group->layer_depth[idx]]++;
858 }
859 }
860
861 // Allocate extra bits to each ARF layer
862 int i;
863 int layer_extra_bits[MAX_ARF_LAYERS + 1] = { 0 };
864 for (i = 1; i <= max_arf_layer; ++i) {
865 double fraction = (i == max_arf_layer) ? 1.0 : layer_fraction[i];
866 layer_extra_bits[i] =
867 (int)((gf_arf_bits * fraction) / AOMMAX(1, layer_frames[i]));
868 gf_arf_bits -= (int)(gf_arf_bits * fraction);
869 }
870
871 // Now combine ARF layer and baseline bits to give total bits for each frame.
872 int arf_extra_bits;
873 for (int idx = frame_index; idx < gf_group_size; ++idx) {
874 switch (gf_group->update_type[idx]) {
875 case ARF_UPDATE:
876 case INTNL_ARF_UPDATE:
877 arf_extra_bits = layer_extra_bits[gf_group->layer_depth[idx]];
878 gf_group->bit_allocation[idx] = base_frame_bits + arf_extra_bits;
879 break;
880 case INTNL_OVERLAY_UPDATE:
881 case OVERLAY_UPDATE: gf_group->bit_allocation[idx] = 0; break;
882 default: gf_group->bit_allocation[idx] = base_frame_bits; break;
883 }
884 }
885
886 // Set the frame following the current GOP to 0 bit allocation. For ARF
887 // groups, this next frame will be overlay frame, which is the first frame
888 // in the next GOP. For GF group, next GOP will overwrite the rate allocation.
889 // Setting this frame to use 0 bit (of out the current GOP budget) will
890 // simplify logics in reference frame management.
891 gf_group->bit_allocation[gf_group_size] = 0;
892 }
893 #else
allocate_gf_group_bits(GF_GROUP * gf_group,RATE_CONTROL * const rc,int64_t gf_group_bits,int gf_arf_bits,int key_frame,int use_arf)894 static void allocate_gf_group_bits(GF_GROUP *gf_group, RATE_CONTROL *const rc,
895 int64_t gf_group_bits, int gf_arf_bits,
896 int key_frame, int use_arf) {
897 int64_t total_group_bits = gf_group_bits;
898
899 // For key frames the frame target rate is already set and it
900 // is also the golden frame.
901 // === [frame_index == 0] ===
902 int frame_index = 0;
903 if (!key_frame) {
904 if (rc->source_alt_ref_active)
905 gf_group->bit_allocation[frame_index] = 0;
906 else
907 gf_group->bit_allocation[frame_index] = gf_arf_bits;
908 }
909
910 // Deduct the boost bits for arf (or gf if it is not a key frame)
911 // from the group total.
912 if (use_arf || !key_frame) total_group_bits -= gf_arf_bits;
913
914 frame_index++;
915
916 // Store the bits to spend on the ARF if there is one.
917 // === [frame_index == 1] ===
918 if (use_arf) {
919 gf_group->bit_allocation[frame_index] = gf_arf_bits;
920 ++frame_index;
921 }
922
923 const int gf_group_size = gf_group->size;
924 int arf_depth_bits[MAX_ARF_LAYERS + 1] = { 0 };
925 int arf_depth_count[MAX_ARF_LAYERS + 1] = { 0 };
926 int arf_depth_boost[MAX_ARF_LAYERS + 1] = { 0 };
927 int total_arfs = 0;
928 int total_overlays = rc->source_alt_ref_active;
929
930 for (int idx = 0; idx < gf_group_size; ++idx) {
931 if (gf_group->update_type[idx] == ARF_UPDATE ||
932 gf_group->update_type[idx] == INTNL_ARF_UPDATE ||
933 gf_group->update_type[idx] == LF_UPDATE) {
934 arf_depth_boost[gf_group->layer_depth[idx]] += gf_group->arf_boost[idx];
935 ++arf_depth_count[gf_group->layer_depth[idx]];
936 }
937 }
938
939 for (int idx = 2; idx <= MAX_ARF_LAYERS; ++idx) {
940 arf_depth_bits[idx] =
941 calculate_boost_bits(rc->baseline_gf_interval - total_arfs -
942 total_overlays - arf_depth_count[idx],
943 arf_depth_boost[idx], total_group_bits);
944 total_group_bits -= arf_depth_bits[idx];
945 total_arfs += arf_depth_count[idx];
946 }
947
948 for (int idx = frame_index; idx < gf_group_size; ++idx) {
949 switch (gf_group->update_type[idx]) {
950 case ARF_UPDATE:
951 case INTNL_ARF_UPDATE:
952 case LF_UPDATE:
953 gf_group->bit_allocation[idx] =
954 (int)(((int64_t)arf_depth_bits[gf_group->layer_depth[idx]] *
955 gf_group->arf_boost[idx]) /
956 arf_depth_boost[gf_group->layer_depth[idx]]);
957 break;
958 case INTNL_OVERLAY_UPDATE:
959 case OVERLAY_UPDATE:
960 default: gf_group->bit_allocation[idx] = 0; break;
961 }
962 }
963
964 // Set the frame following the current GOP to 0 bit allocation. For ARF
965 // groups, this next frame will be overlay frame, which is the first frame
966 // in the next GOP. For GF group, next GOP will overwrite the rate allocation.
967 // Setting this frame to use 0 bit (of out the current GOP budget) will
968 // simplify logics in reference frame management.
969 gf_group->bit_allocation[gf_group_size] = 0;
970 }
971 #endif
972
973 // Returns true if KF group and GF group both are almost completely static.
is_almost_static(double gf_zero_motion,int kf_zero_motion)974 static INLINE int is_almost_static(double gf_zero_motion, int kf_zero_motion) {
975 return (gf_zero_motion >= 0.995) &&
976 (kf_zero_motion >= STATIC_KF_GROUP_THRESH);
977 }
978
979 #define ARF_ABS_ZOOM_THRESH 4.4
detect_gf_cut(AV1_COMP * cpi,int frame_index,int cur_start,int flash_detected,int active_max_gf_interval,int active_min_gf_interval,GF_GROUP_STATS * gf_stats)980 static INLINE int detect_gf_cut(AV1_COMP *cpi, int frame_index, int cur_start,
981 int flash_detected, int active_max_gf_interval,
982 int active_min_gf_interval,
983 GF_GROUP_STATS *gf_stats) {
984 RATE_CONTROL *const rc = &cpi->rc;
985 TWO_PASS *const twopass = &cpi->twopass;
986 // Motion breakout threshold for loop below depends on image size.
987 const double mv_ratio_accumulator_thresh =
988 (cpi->initial_height + cpi->initial_width) / 4.0;
989
990 if (!flash_detected) {
991 // Break clause to detect very still sections after motion. For example,
992 // a static image after a fade or other transition.
993 if (detect_transition_to_still(
994 twopass, rc->min_gf_interval, frame_index - cur_start, 5,
995 gf_stats->loop_decay_rate, gf_stats->last_loop_decay_rate)) {
996 return 1;
997 }
998 }
999
1000 // Some conditions to breakout after min interval.
1001 if (frame_index - cur_start >= active_min_gf_interval &&
1002 // If possible don't break very close to a kf
1003 (rc->frames_to_key - frame_index >= rc->min_gf_interval) &&
1004 ((frame_index - cur_start) & 0x01) && !flash_detected &&
1005 (gf_stats->mv_ratio_accumulator > mv_ratio_accumulator_thresh ||
1006 gf_stats->abs_mv_in_out_accumulator > ARF_ABS_ZOOM_THRESH)) {
1007 return 1;
1008 }
1009
1010 // If almost totally static, we will not use the the max GF length later,
1011 // so we can continue for more frames.
1012 if (((frame_index - cur_start) >= active_max_gf_interval + 1) &&
1013 !is_almost_static(gf_stats->zero_motion_accumulator,
1014 twopass->kf_zeromotion_pct)) {
1015 return 1;
1016 }
1017 return 0;
1018 }
1019
1020 #define MAX_PAD_GF_CHECK 6 // padding length to check for gf length
1021 #define AVG_SI_THRES 0.6 // thres for average silouette
1022 #define GF_SHRINK_OUTPUT 0 // print output for gf length decision
determine_high_err_gf(double * errs,int * is_high,double * si,int len,double * ratio,int gf_start,int gf_end,int before_pad)1023 int determine_high_err_gf(double *errs, int *is_high, double *si, int len,
1024 double *ratio, int gf_start, int gf_end,
1025 int before_pad) {
1026 (void)gf_start;
1027 (void)gf_end;
1028 (void)before_pad;
1029 // alpha and beta controls the threshold placement
1030 // e.g. a smaller alpha makes the lower group more rigid
1031 const double alpha = 0.5;
1032 const double beta = 1 - alpha;
1033 double mean = 0;
1034 double mean_low = 0;
1035 double mean_high = 0;
1036 double prev_mean_low = 0;
1037 double prev_mean_high = 0;
1038 int count_low = 0;
1039 int count_high = 0;
1040 // calculate mean of errs
1041 for (int i = 0; i < len; i++) {
1042 mean += errs[i];
1043 }
1044 mean /= len;
1045 // separate into two initial groups with greater / lower than mean
1046 for (int i = 0; i < len; i++) {
1047 if (errs[i] <= mean) {
1048 is_high[i] = 0;
1049 count_low++;
1050 prev_mean_low += errs[i];
1051 } else {
1052 is_high[i] = 1;
1053 count_high++;
1054 prev_mean_high += errs[i];
1055 }
1056 }
1057 prev_mean_low /= count_low;
1058 prev_mean_high /= count_high;
1059 // kmeans to refine
1060 int count = 0;
1061 while (count < 10) {
1062 // re-group
1063 mean_low = 0;
1064 mean_high = 0;
1065 count_low = 0;
1066 count_high = 0;
1067 double thres = prev_mean_low * alpha + prev_mean_high * beta;
1068 for (int i = 0; i < len; i++) {
1069 if (errs[i] <= thres) {
1070 is_high[i] = 0;
1071 count_low++;
1072 mean_low += errs[i];
1073 } else {
1074 is_high[i] = 1;
1075 count_high++;
1076 mean_high += errs[i];
1077 }
1078 }
1079 mean_low /= count_low;
1080 mean_high /= count_high;
1081
1082 // break if not changed much
1083 if (fabs((mean_low - prev_mean_low) / (prev_mean_low + 0.00001)) <
1084 0.00001 &&
1085 fabs((mean_high - prev_mean_high) / (prev_mean_high + 0.00001)) <
1086 0.00001)
1087 break;
1088
1089 // update means
1090 prev_mean_high = mean_high;
1091 prev_mean_low = mean_low;
1092
1093 count++;
1094 }
1095
1096 // count how many jumps of group changes
1097 int num_change = 0;
1098 for (int i = 0; i < len - 1; i++) {
1099 if (is_high[i] != is_high[i + 1]) num_change++;
1100 }
1101
1102 // get silhouette as a measure of the classification quality
1103 double avg_si = 0;
1104 // ai: avg dist of its own class, bi: avg dist to the other class
1105 double ai, bi;
1106 if (count_low > 1 && count_high > 1) {
1107 for (int i = 0; i < len; i++) {
1108 ai = 0;
1109 bi = 0;
1110 // calculate average distance to everyone in the same group
1111 // and in the other group
1112 for (int j = 0; j < len; j++) {
1113 if (i == j) continue;
1114 if (is_high[i] == is_high[j]) {
1115 ai += fabs(errs[i] - errs[j]);
1116 } else {
1117 bi += fabs(errs[i] - errs[j]);
1118 }
1119 }
1120 if (is_high[i] == 0) {
1121 ai = ai / (count_low - 1);
1122 bi = bi / count_high;
1123 } else {
1124 ai = ai / (count_high - 1);
1125 bi = bi / count_low;
1126 }
1127 if (ai <= bi) {
1128 si[i] = 1 - ai / (bi + 0.00001);
1129 } else {
1130 si[i] = bi / (ai + 0.00001) - 1;
1131 }
1132 avg_si += si[i];
1133 }
1134 avg_si /= len;
1135 }
1136
1137 int reset = 0;
1138 *ratio = mean_high / (mean_low + 0.00001);
1139 // if the two groups too similar, or
1140 // if too many numbers of changes, or
1141 // silhouette is too small, not confident
1142 // reset everything to 0 later so we fallback to the original decision
1143 if (*ratio < 1.3 || num_change > AOMMAX(len / 3, 6) ||
1144 avg_si < AVG_SI_THRES) {
1145 reset = 1;
1146 }
1147
1148 #if GF_SHRINK_OUTPUT
1149 printf("\n");
1150 for (int i = 0; i < len; i++) {
1151 printf("%d: err %.1f, ishigh %d, si %.2f, (i=%d)\n",
1152 gf_start + i - before_pad, errs[i], is_high[i], si[i], gf_end);
1153 }
1154 printf(
1155 "count: %d, mean_high: %.1f, mean_low: %.1f, avg_si: %.2f, num_change: "
1156 "%d, ratio %.2f, reset: %d\n",
1157 count, mean_high, mean_low, avg_si, num_change,
1158 mean_high / (mean_low + 0.000001), reset);
1159 #endif
1160
1161 if (reset) {
1162 memset(is_high, 0, sizeof(is_high[0]) * len);
1163 memset(si, 0, sizeof(si[0]) * len);
1164 }
1165 return reset;
1166 }
1167
1168 #if GROUP_ADAPTIVE_MAXQ
1169 #define RC_FACTOR_MIN 0.75
1170 #define RC_FACTOR_MAX 1.25
1171 #endif // GROUP_ADAPTIVE_MAXQ
1172 #define MIN_FWD_KF_INTERVAL 8
1173 #define MIN_SHRINK_LEN 6 // the minimum length of gf if we are shrinking
1174 #define SI_HIGH AVG_SI_THRES // high quality classification
1175 #define SI_LOW 0.3 // very unsure classification
1176 // this function finds an low error frame previously to the current last frame
1177 // in the gf group, and set the last frame to it.
1178 // The resulting last frame is then returned by *cur_last_ptr
1179 // *cur_start_ptr and cut_pos[n] could also change due to shrinking
1180 // previous gf groups
set_last_prev_low_err(int * cur_start_ptr,int * cur_last_ptr,int * cut_pos,int count_cuts,int before_pad,double ratio,int * is_high,double * si,int prev_lows)1181 void set_last_prev_low_err(int *cur_start_ptr, int *cur_last_ptr, int *cut_pos,
1182 int count_cuts, int before_pad, double ratio,
1183 int *is_high, double *si, int prev_lows) {
1184 int n;
1185 int cur_start = *cur_start_ptr;
1186 int cur_last = *cur_last_ptr;
1187 for (n = cur_last; n >= cur_start + MIN_SHRINK_LEN; n--) {
1188 // try to find a point that is very probable to be good
1189 if (is_high[n - cur_start + before_pad] == 0 &&
1190 si[n - cur_start + before_pad] > SI_HIGH) {
1191 *cur_last_ptr = n;
1192 return;
1193 }
1194 }
1195 // could not find a low-err point, then let's try find an "unsure"
1196 // point at least
1197 for (n = cur_last; n >= cur_start + MIN_SHRINK_LEN; n--) {
1198 if ((is_high[n - cur_start + before_pad] == 0) ||
1199 (is_high[n - cur_start + before_pad] &&
1200 si[n - cur_start + before_pad] < SI_LOW)) {
1201 *cur_last_ptr = n;
1202 return;
1203 }
1204 }
1205 if (prev_lows) {
1206 // try with shrinking previous all_zero interval
1207 for (n = cur_start + MIN_SHRINK_LEN - 1; n > cur_start; n--) {
1208 if (is_high[n - cur_start + before_pad] == 0 &&
1209 si[n - cur_start + before_pad] > SI_HIGH) {
1210 int tentative_start = n - MIN_SHRINK_LEN;
1211 // check if the previous interval can shrink this much
1212 int available =
1213 tentative_start - cut_pos[count_cuts - 2] > MIN_SHRINK_LEN &&
1214 cur_start - tentative_start < prev_lows;
1215 // shrinking too agressively may worsen performance
1216 // set stricter thres for shorter length
1217 double ratio_thres =
1218 1.0 * (cur_start - tentative_start) / (double)(MIN_SHRINK_LEN) +
1219 1.0;
1220
1221 if (available && (ratio > ratio_thres)) {
1222 cut_pos[count_cuts - 1] = tentative_start;
1223 *cur_start_ptr = tentative_start;
1224 *cur_last_ptr = n;
1225 return;
1226 }
1227 }
1228 }
1229 }
1230 if (prev_lows) {
1231 // try with shrinking previous all_zero interval with unsure points
1232 for (n = cur_start + MIN_SHRINK_LEN - 1; n > cur_start; n--) {
1233 if ((is_high[n - cur_start + before_pad] == 0) ||
1234 (is_high[n - cur_start + before_pad] &&
1235 si[n - cur_start + before_pad] < SI_LOW)) {
1236 int tentative_start = n - MIN_SHRINK_LEN;
1237 // check if the previous interval can shrink this much
1238 int available =
1239 tentative_start - cut_pos[count_cuts - 2] > MIN_SHRINK_LEN &&
1240 cur_start - tentative_start < prev_lows;
1241 // shrinking too agressively may worsen performance
1242 double ratio_thres =
1243 1.0 * (cur_start - tentative_start) / (double)(MIN_SHRINK_LEN) +
1244 1.0;
1245
1246 if (available && (ratio > ratio_thres)) {
1247 cut_pos[count_cuts - 1] = tentative_start;
1248 *cur_start_ptr = tentative_start;
1249 *cur_last_ptr = n;
1250 return;
1251 }
1252 }
1253 }
1254 } // prev_lows
1255 return;
1256 }
1257
1258 // This function decides the gf group length of future frames in batch
1259 // rc->gf_intervals is modified to store the group lengths
calculate_gf_length(AV1_COMP * cpi,int max_gop_length,int max_intervals)1260 static void calculate_gf_length(AV1_COMP *cpi, int max_gop_length,
1261 int max_intervals) {
1262 RATE_CONTROL *const rc = &cpi->rc;
1263 TWO_PASS *const twopass = &cpi->twopass;
1264 FIRSTPASS_STATS next_frame;
1265 const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
1266 FRAME_INFO *frame_info = &cpi->frame_info;
1267 int i;
1268
1269 int flash_detected;
1270
1271 aom_clear_system_state();
1272 av1_zero(next_frame);
1273
1274 if (has_no_stats_stage(cpi)) {
1275 for (i = 0; i < MAX_NUM_GF_INTERVALS; i++) {
1276 rc->gf_intervals[i] = AOMMIN(rc->max_gf_interval, max_gop_length);
1277 }
1278 rc->cur_gf_index = 0;
1279 rc->intervals_till_gf_calculate_due = MAX_NUM_GF_INTERVALS;
1280 return;
1281 }
1282
1283 // TODO(urvang): Try logic to vary min and max interval based on q.
1284 const int active_min_gf_interval = rc->min_gf_interval;
1285 const int active_max_gf_interval =
1286 AOMMIN(rc->max_gf_interval, max_gop_length);
1287
1288 i = 0;
1289 max_intervals = cpi->lap_enabled ? 1 : max_intervals;
1290 int cut_pos[MAX_NUM_GF_INTERVALS + 1] = { 0 };
1291 int count_cuts = 1;
1292 int cur_start = 0, cur_last;
1293 int cut_here;
1294 int prev_lows = 0;
1295 GF_GROUP_STATS gf_stats;
1296 init_gf_stats(&gf_stats);
1297 while (count_cuts < max_intervals + 1) {
1298 ++i;
1299
1300 // reaches next key frame, break here
1301 if (i >= rc->frames_to_key) {
1302 cut_pos[count_cuts] = i - 1;
1303 count_cuts++;
1304 break;
1305 }
1306
1307 // reached maximum len, but nothing special yet (almost static)
1308 // let's look at the next interval
1309 if (i - cur_start >= rc->static_scene_max_gf_interval) {
1310 cut_here = 1;
1311 } else {
1312 // reaches last frame, break
1313 if (EOF == input_stats(twopass, &next_frame)) {
1314 cut_pos[count_cuts] = i - 1;
1315 count_cuts++;
1316 break;
1317 }
1318 // Test for the case where there is a brief flash but the prediction
1319 // quality back to an earlier frame is then restored.
1320 flash_detected = detect_flash(twopass, 0);
1321 // TODO(bohanli): remove redundant accumulations here, or unify
1322 // this and the ones in define_gf_group
1323 accumulate_next_frame_stats(&next_frame, frame_info, twopass,
1324 flash_detected, rc->frames_since_key, i, 0,
1325 rc->min_gf_interval, &gf_stats);
1326
1327 cut_here = detect_gf_cut(cpi, i, cur_start, flash_detected,
1328 active_max_gf_interval, active_min_gf_interval,
1329 &gf_stats);
1330 }
1331 if (cut_here) {
1332 cur_last = i - 1; // the current last frame in the gf group
1333 // only try shrinking if interval smaller than active_max_gf_interval
1334 if (cur_last - cur_start <= active_max_gf_interval) {
1335 // determine in the current decided gop the higher and lower errs
1336 int n;
1337 double ratio;
1338
1339 // load neighboring coded errs
1340 int is_high[MAX_GF_INTERVAL + 1 + MAX_PAD_GF_CHECK * 2] = { 0 };
1341 double errs[MAX_GF_INTERVAL + 1 + MAX_PAD_GF_CHECK * 2] = { 0 };
1342 double si[MAX_GF_INTERVAL + 1 + MAX_PAD_GF_CHECK * 2] = { 0 };
1343 int before_pad =
1344 AOMMIN(MAX_PAD_GF_CHECK, rc->frames_since_key - 1 + cur_start);
1345 int after_pad =
1346 AOMMIN(MAX_PAD_GF_CHECK, rc->frames_to_key - cur_last - 1);
1347 for (n = cur_start - before_pad; n <= cur_last + after_pad; n++) {
1348 if (start_pos + n - 1 > twopass->stats_buf_ctx->stats_in_end) {
1349 after_pad = n - cur_last - 1;
1350 assert(after_pad >= 0);
1351 break;
1352 } else if (start_pos + n - 1 <
1353 twopass->stats_buf_ctx->stats_in_start) {
1354 before_pad = cur_start - n - 1;
1355 continue;
1356 }
1357 errs[n + before_pad - cur_start] = (start_pos + n - 1)->coded_error;
1358 }
1359 const int len = before_pad + after_pad + cur_last - cur_start + 1;
1360 const int reset = determine_high_err_gf(
1361 errs, is_high, si, len, &ratio, cur_start, cur_last, before_pad);
1362
1363 // if the current frame may have high error, try shrinking
1364 if (is_high[cur_last - cur_start + before_pad] == 1 ||
1365 (!reset && si[cur_last - cur_start + before_pad] < SI_LOW)) {
1366 // try not to cut in high err area
1367 set_last_prev_low_err(&cur_start, &cur_last, cut_pos, count_cuts,
1368 before_pad, ratio, is_high, si, prev_lows);
1369 } // if current frame high error
1370 // count how many trailing lower error frames we have in this decided
1371 // gf group
1372 prev_lows = 0;
1373 for (n = cur_last - 1; n > cur_start + MIN_SHRINK_LEN; n--) {
1374 if (is_high[n - cur_start + before_pad] == 0 &&
1375 (si[n - cur_start + before_pad] > SI_HIGH || reset)) {
1376 prev_lows++;
1377 } else {
1378 break;
1379 }
1380 }
1381 }
1382 cut_pos[count_cuts] = cur_last;
1383 count_cuts++;
1384
1385 // reset pointers to the shrinked location
1386 twopass->stats_in = start_pos + cur_last;
1387 cur_start = cur_last;
1388 i = cur_last;
1389
1390 // reset accumulators
1391 init_gf_stats(&gf_stats);
1392 }
1393 }
1394
1395 // save intervals
1396 rc->intervals_till_gf_calculate_due = count_cuts - 1;
1397 for (int n = 1; n < count_cuts; n++) {
1398 rc->gf_intervals[n - 1] = cut_pos[n] + 1 - cut_pos[n - 1];
1399 }
1400 rc->cur_gf_index = 0;
1401 twopass->stats_in = start_pos;
1402
1403 #if GF_SHRINK_OUTPUT
1404 printf("\nf_to_key: %d, count_cut: %d. ", rc->frames_to_key, count_cuts);
1405 for (int n = 0; n < count_cuts; n++) {
1406 printf("%d ", cut_pos[n]);
1407 }
1408 printf("\n");
1409
1410 for (int n = 0; n < rc->intervals_till_gf_calculate_due; n++) {
1411 printf("%d ", rc->gf_intervals[n]);
1412 }
1413 printf("\n\n");
1414 #endif
1415 }
1416
correct_frames_to_key(AV1_COMP * cpi)1417 static void correct_frames_to_key(AV1_COMP *cpi) {
1418 int lookahead_size =
1419 (int)av1_lookahead_depth(cpi->lookahead, cpi->compressor_stage) + 1;
1420 if (lookahead_size <
1421 av1_lookahead_pop_sz(cpi->lookahead, cpi->compressor_stage)) {
1422 cpi->rc.frames_to_key = AOMMIN(cpi->rc.frames_to_key, lookahead_size);
1423 }
1424 }
1425
define_gf_group_pass0(AV1_COMP * cpi,const EncodeFrameParams * const frame_params)1426 static void define_gf_group_pass0(AV1_COMP *cpi,
1427 const EncodeFrameParams *const frame_params) {
1428 RATE_CONTROL *const rc = &cpi->rc;
1429 GF_GROUP *const gf_group = &cpi->gf_group;
1430 int target;
1431
1432 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
1433 av1_cyclic_refresh_set_golden_update(cpi);
1434 } else {
1435 rc->baseline_gf_interval = rc->gf_intervals[rc->cur_gf_index];
1436 rc->intervals_till_gf_calculate_due--;
1437 rc->cur_gf_index++;
1438 }
1439
1440 // correct frames_to_key when lookahead queue is flushing
1441 correct_frames_to_key(cpi);
1442
1443 if (rc->baseline_gf_interval > rc->frames_to_key)
1444 rc->baseline_gf_interval = rc->frames_to_key;
1445
1446 rc->gfu_boost = DEFAULT_GF_BOOST;
1447 rc->constrained_gf_group =
1448 (rc->baseline_gf_interval >= rc->frames_to_key) ? 1 : 0;
1449
1450 gf_group->max_layer_depth_allowed = cpi->oxcf.gf_max_pyr_height;
1451
1452 // Rare case when the look-ahead is less than the target GOP length, can't
1453 // generate ARF frame.
1454 if (rc->baseline_gf_interval > cpi->oxcf.lag_in_frames ||
1455 !is_altref_enabled(cpi) || rc->baseline_gf_interval < rc->min_gf_interval)
1456 gf_group->max_layer_depth_allowed = 0;
1457
1458 // Set up the structure of this Group-Of-Pictures (same as GF_GROUP)
1459 av1_gop_setup_structure(cpi, frame_params);
1460
1461 // Allocate bits to each of the frames in the GF group.
1462 // TODO(sarahparker) Extend this to work with pyramid structure.
1463 for (int cur_index = 0; cur_index < gf_group->size; ++cur_index) {
1464 const FRAME_UPDATE_TYPE cur_update_type = gf_group->update_type[cur_index];
1465 if (cpi->oxcf.rc_mode == AOM_CBR) {
1466 if (cur_update_type == KEY_FRAME) {
1467 target = av1_calc_iframe_target_size_one_pass_cbr(cpi);
1468 } else {
1469 target = av1_calc_pframe_target_size_one_pass_cbr(cpi, cur_update_type);
1470 }
1471 } else {
1472 if (cur_update_type == KEY_FRAME) {
1473 target = av1_calc_iframe_target_size_one_pass_vbr(cpi);
1474 } else {
1475 target = av1_calc_pframe_target_size_one_pass_vbr(cpi, cur_update_type);
1476 }
1477 }
1478 gf_group->bit_allocation[cur_index] = target;
1479 }
1480 }
1481
set_baseline_gf_interval(AV1_COMP * cpi,int arf_position,int active_max_gf_interval,int use_alt_ref,int is_final_pass)1482 static INLINE void set_baseline_gf_interval(AV1_COMP *cpi, int arf_position,
1483 int active_max_gf_interval,
1484 int use_alt_ref,
1485 int is_final_pass) {
1486 RATE_CONTROL *const rc = &cpi->rc;
1487 TWO_PASS *const twopass = &cpi->twopass;
1488 // Set the interval until the next gf.
1489 // If forward keyframes are enabled, ensure the final gf group obeys the
1490 // MIN_FWD_KF_INTERVAL.
1491 if (cpi->oxcf.fwd_kf_enabled && use_alt_ref &&
1492 ((twopass->stats_in - arf_position + rc->frames_to_key) <
1493 twopass->stats_buf_ctx->stats_in_end) &&
1494 cpi->rc.next_is_fwd_key) {
1495 if (arf_position == rc->frames_to_key) {
1496 rc->baseline_gf_interval = arf_position;
1497 // if the last gf group will be smaller than MIN_FWD_KF_INTERVAL
1498 } else if ((rc->frames_to_key - arf_position <
1499 AOMMAX(MIN_FWD_KF_INTERVAL, rc->min_gf_interval)) &&
1500 (rc->frames_to_key != arf_position)) {
1501 // if possible, merge the last two gf groups
1502 if (rc->frames_to_key <= active_max_gf_interval) {
1503 rc->baseline_gf_interval = rc->frames_to_key;
1504 if (is_final_pass) rc->intervals_till_gf_calculate_due = 0;
1505 // if merging the last two gf groups creates a group that is too long,
1506 // split them and force the last gf group to be the MIN_FWD_KF_INTERVAL
1507 } else {
1508 rc->baseline_gf_interval = rc->frames_to_key - MIN_FWD_KF_INTERVAL;
1509 if (is_final_pass) rc->intervals_till_gf_calculate_due = 0;
1510 }
1511 } else {
1512 rc->baseline_gf_interval = arf_position - rc->source_alt_ref_pending;
1513 }
1514 } else {
1515 rc->baseline_gf_interval = arf_position - rc->source_alt_ref_pending;
1516 }
1517 }
1518
1519 // initialize GF_GROUP_STATS
init_gf_stats(GF_GROUP_STATS * gf_stats)1520 static void init_gf_stats(GF_GROUP_STATS *gf_stats) {
1521 gf_stats->gf_group_err = 0.0;
1522 gf_stats->gf_group_raw_error = 0.0;
1523 gf_stats->gf_group_skip_pct = 0.0;
1524 gf_stats->gf_group_inactive_zone_rows = 0.0;
1525
1526 gf_stats->mv_ratio_accumulator = 0.0;
1527 gf_stats->decay_accumulator = 1.0;
1528 gf_stats->zero_motion_accumulator = 1.0;
1529 gf_stats->loop_decay_rate = 1.0;
1530 gf_stats->last_loop_decay_rate = 1.0;
1531 gf_stats->this_frame_mv_in_out = 0.0;
1532 gf_stats->mv_in_out_accumulator = 0.0;
1533 gf_stats->abs_mv_in_out_accumulator = 0.0;
1534
1535 gf_stats->avg_sr_coded_error = 0.0;
1536 gf_stats->avg_tr_coded_error = 0.0;
1537 gf_stats->avg_pcnt_second_ref = 0.0;
1538 gf_stats->avg_pcnt_third_ref = 0.0;
1539 gf_stats->avg_pcnt_third_ref_nolast = 0.0;
1540 gf_stats->avg_new_mv_count = 0.0;
1541 gf_stats->avg_wavelet_energy = 0.0;
1542 gf_stats->avg_raw_err_stdev = 0.0;
1543 gf_stats->non_zero_stdev_count = 0;
1544
1545 gf_stats->allow_alt_ref = 0;
1546 }
1547
1548 // Analyse and define a gf/arf group.
1549 #define MAX_GF_BOOST 5400
define_gf_group(AV1_COMP * cpi,FIRSTPASS_STATS * this_frame,const EncodeFrameParams * const frame_params,int max_gop_length,int is_final_pass)1550 static void define_gf_group(AV1_COMP *cpi, FIRSTPASS_STATS *this_frame,
1551 const EncodeFrameParams *const frame_params,
1552 int max_gop_length, int is_final_pass) {
1553 AV1_COMMON *const cm = &cpi->common;
1554 RATE_CONTROL *const rc = &cpi->rc;
1555 AV1EncoderConfig *const oxcf = &cpi->oxcf;
1556 TWO_PASS *const twopass = &cpi->twopass;
1557 FIRSTPASS_STATS next_frame;
1558 const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
1559 GF_GROUP *gf_group = &cpi->gf_group;
1560 FRAME_INFO *frame_info = &cpi->frame_info;
1561 int i;
1562
1563 int flash_detected;
1564 int64_t gf_group_bits;
1565 const int is_intra_only = frame_params->frame_type == KEY_FRAME ||
1566 frame_params->frame_type == INTRA_ONLY_FRAME;
1567 const int arf_active_or_kf = is_intra_only || rc->source_alt_ref_active;
1568
1569 cpi->internal_altref_allowed = (oxcf->gf_max_pyr_height > 1);
1570
1571 // Reset the GF group data structures unless this is a key
1572 // frame in which case it will already have been done.
1573 if (!is_intra_only) {
1574 av1_zero(cpi->gf_group);
1575 }
1576
1577 aom_clear_system_state();
1578 av1_zero(next_frame);
1579
1580 if (has_no_stats_stage(cpi)) {
1581 define_gf_group_pass0(cpi, frame_params);
1582 return;
1583 }
1584
1585 // correct frames_to_key when lookahead queue is emptying
1586 if (cpi->lap_enabled) {
1587 correct_frames_to_key(cpi);
1588 }
1589
1590 GF_GROUP_STATS gf_stats;
1591 init_gf_stats(&gf_stats);
1592 GF_FRAME_STATS first_frame_stats, last_frame_stats;
1593
1594 gf_stats.allow_alt_ref = is_altref_enabled(cpi);
1595 const int can_disable_arf = (oxcf->gf_min_pyr_height == MIN_PYRAMID_LVL);
1596
1597 // Load stats for the current frame.
1598 double mod_frame_err =
1599 calculate_modified_err(frame_info, twopass, oxcf, this_frame);
1600
1601 // Note the error of the frame at the start of the group. This will be
1602 // the GF frame error if we code a normal gf.
1603 first_frame_stats.frame_err = mod_frame_err;
1604 first_frame_stats.frame_coded_error = this_frame->coded_error;
1605 first_frame_stats.frame_sr_coded_error = this_frame->sr_coded_error;
1606 first_frame_stats.frame_tr_coded_error = this_frame->tr_coded_error;
1607
1608 // If this is a key frame or the overlay from a previous arf then
1609 // the error score / cost of this frame has already been accounted for.
1610 if (arf_active_or_kf) {
1611 gf_stats.gf_group_err -= first_frame_stats.frame_err;
1612 #if GROUP_ADAPTIVE_MAXQ
1613 gf_stats.gf_group_raw_error -= this_frame->coded_error;
1614 #endif
1615 gf_stats.gf_group_skip_pct -= this_frame->intra_skip_pct;
1616 gf_stats.gf_group_inactive_zone_rows -= this_frame->inactive_zone_rows;
1617 }
1618
1619 // TODO(urvang): Try logic to vary min and max interval based on q.
1620 const int active_min_gf_interval = rc->min_gf_interval;
1621 const int active_max_gf_interval =
1622 AOMMIN(rc->max_gf_interval, max_gop_length);
1623
1624 i = 0;
1625 // get the determined gf group length from rc->gf_intervals
1626 while (i < rc->gf_intervals[rc->cur_gf_index]) {
1627 ++i;
1628 // Accumulate error score of frames in this gf group.
1629 mod_frame_err =
1630 calculate_modified_err(frame_info, twopass, oxcf, this_frame);
1631 // accumulate stats for this frame
1632 accumulate_this_frame_stats(this_frame, mod_frame_err, &gf_stats);
1633
1634 // read in the next frame
1635 if (EOF == input_stats(twopass, &next_frame)) break;
1636
1637 // Test for the case where there is a brief flash but the prediction
1638 // quality back to an earlier frame is then restored.
1639 flash_detected = detect_flash(twopass, 0);
1640
1641 // accumulate stats for next frame
1642 accumulate_next_frame_stats(
1643 &next_frame, frame_info, twopass, flash_detected, rc->frames_since_key,
1644 i, can_disable_arf, rc->min_gf_interval, &gf_stats);
1645
1646 *this_frame = next_frame;
1647 }
1648 // save the errs for the last frame
1649 last_frame_stats.frame_coded_error = next_frame.coded_error;
1650 last_frame_stats.frame_sr_coded_error = next_frame.sr_coded_error;
1651 last_frame_stats.frame_tr_coded_error = next_frame.tr_coded_error;
1652
1653 if (is_final_pass) {
1654 rc->intervals_till_gf_calculate_due--;
1655 rc->cur_gf_index++;
1656 }
1657
1658 // Was the group length constrained by the requirement for a new KF?
1659 rc->constrained_gf_group = (i >= rc->frames_to_key) ? 1 : 0;
1660
1661 const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
1662 ? cpi->initial_mbs
1663 : cm->mi_params.MBs;
1664 assert(num_mbs > 0);
1665
1666 average_gf_stats(i, &next_frame, &gf_stats);
1667
1668 // Disable internal ARFs for "still" gf groups.
1669 // zero_motion_accumulator: minimum percentage of (0,0) motion;
1670 // avg_sr_coded_error: average of the SSE per pixel of each frame;
1671 // avg_raw_err_stdev: average of the standard deviation of (0,0)
1672 // motion error per block of each frame.
1673 const int can_disable_internal_arfs =
1674 (oxcf->gf_min_pyr_height <= MIN_PYRAMID_LVL + 1);
1675 if (can_disable_internal_arfs &&
1676 gf_stats.zero_motion_accumulator > MIN_ZERO_MOTION &&
1677 gf_stats.avg_sr_coded_error / num_mbs < MAX_SR_CODED_ERROR &&
1678 gf_stats.avg_raw_err_stdev < MAX_RAW_ERR_VAR) {
1679 cpi->internal_altref_allowed = 0;
1680 }
1681
1682 int use_alt_ref;
1683 if (can_disable_arf) {
1684 use_alt_ref = !is_almost_static(gf_stats.zero_motion_accumulator,
1685 twopass->kf_zeromotion_pct) &&
1686 gf_stats.allow_alt_ref && (i < cpi->oxcf.lag_in_frames) &&
1687 (i >= MIN_GF_INTERVAL) &&
1688 (cpi->oxcf.gf_max_pyr_height > MIN_PYRAMID_LVL);
1689
1690 // TODO(urvang): Improve and use model for VBR, CQ etc as well.
1691 if (use_alt_ref && cpi->oxcf.rc_mode == AOM_Q &&
1692 cpi->oxcf.cq_level <= 200) {
1693 aom_clear_system_state();
1694 float features[21];
1695 get_features_from_gf_stats(
1696 &gf_stats, &first_frame_stats, &last_frame_stats, num_mbs,
1697 rc->constrained_gf_group, twopass->kf_zeromotion_pct, i, features);
1698 // Infer using ML model.
1699 float score;
1700 av1_nn_predict(features, &av1_use_flat_gop_nn_config, 1, &score);
1701 use_alt_ref = (score <= 0.0);
1702 }
1703 } else {
1704 assert(cpi->oxcf.gf_max_pyr_height > MIN_PYRAMID_LVL);
1705 use_alt_ref =
1706 gf_stats.allow_alt_ref && (i < cpi->oxcf.lag_in_frames) && (i > 2);
1707 }
1708
1709 #define REDUCE_GF_LENGTH_THRESH 4
1710 #define REDUCE_GF_LENGTH_TO_KEY_THRESH 9
1711 #define REDUCE_GF_LENGTH_BY 1
1712 int alt_offset = 0;
1713 // The length reduction strategy is tweaked for certain cases, and doesn't
1714 // work well for certain other cases.
1715 const int allow_gf_length_reduction =
1716 ((cpi->oxcf.rc_mode == AOM_Q && cpi->oxcf.cq_level <= 128) ||
1717 !cpi->internal_altref_allowed) &&
1718 !is_lossless_requested(&cpi->oxcf);
1719
1720 if (allow_gf_length_reduction && use_alt_ref) {
1721 // adjust length of this gf group if one of the following condition met
1722 // 1: only one overlay frame left and this gf is too long
1723 // 2: next gf group is too short to have arf compared to the current gf
1724
1725 // maximum length of next gf group
1726 const int next_gf_len = rc->frames_to_key - i;
1727 const int single_overlay_left =
1728 next_gf_len == 0 && i > REDUCE_GF_LENGTH_THRESH;
1729 // the next gf is probably going to have a ARF but it will be shorter than
1730 // this gf
1731 const int unbalanced_gf =
1732 i > REDUCE_GF_LENGTH_TO_KEY_THRESH &&
1733 next_gf_len + 1 < REDUCE_GF_LENGTH_TO_KEY_THRESH &&
1734 next_gf_len + 1 >= rc->min_gf_interval;
1735
1736 if (single_overlay_left || unbalanced_gf) {
1737 const int roll_back = REDUCE_GF_LENGTH_BY;
1738 // Reduce length only if active_min_gf_interval will be respected later.
1739 if (i - roll_back >= active_min_gf_interval + 1) {
1740 alt_offset = -roll_back;
1741 i -= roll_back;
1742 if (is_final_pass) rc->intervals_till_gf_calculate_due = 0;
1743 }
1744 }
1745 }
1746
1747 // Should we use the alternate reference frame.
1748 if (use_alt_ref) {
1749 rc->source_alt_ref_pending = 1;
1750 gf_group->max_layer_depth_allowed = cpi->oxcf.gf_max_pyr_height;
1751 set_baseline_gf_interval(cpi, i, active_max_gf_interval, use_alt_ref,
1752 is_final_pass);
1753
1754 const int forward_frames = (rc->frames_to_key - i >= i - 1)
1755 ? i - 1
1756 : AOMMAX(0, rc->frames_to_key - i);
1757
1758 // Calculate the boost for alt ref.
1759 rc->gfu_boost = av1_calc_arf_boost(
1760 twopass, rc, frame_info, alt_offset, forward_frames, (i - 1),
1761 cpi->lap_enabled ? &rc->num_stats_used_for_gfu_boost : NULL,
1762 cpi->lap_enabled ? &rc->num_stats_required_for_gfu_boost : NULL);
1763 } else {
1764 reset_fpf_position(twopass, start_pos);
1765 rc->source_alt_ref_pending = 0;
1766 gf_group->max_layer_depth_allowed = 0;
1767 set_baseline_gf_interval(cpi, i, active_max_gf_interval, use_alt_ref,
1768 is_final_pass);
1769
1770 rc->gfu_boost = AOMMIN(
1771 MAX_GF_BOOST,
1772 av1_calc_arf_boost(
1773 twopass, rc, frame_info, alt_offset, (i - 1), 0,
1774 cpi->lap_enabled ? &rc->num_stats_used_for_gfu_boost : NULL,
1775 cpi->lap_enabled ? &rc->num_stats_required_for_gfu_boost : NULL));
1776 }
1777
1778 // rc->gf_intervals assumes the usage of alt_ref, therefore adding one overlay
1779 // frame to the next gf. If no alt_ref is used, should substract 1 frame from
1780 // the next gf group.
1781 // TODO(bohanli): should incorporate the usage of alt_ref into
1782 // calculate_gf_length
1783 if (is_final_pass && rc->source_alt_ref_pending == 0 &&
1784 rc->intervals_till_gf_calculate_due > 0) {
1785 rc->gf_intervals[rc->cur_gf_index]--;
1786 }
1787
1788 #define LAST_ALR_BOOST_FACTOR 0.2f
1789 rc->arf_boost_factor = 1.0;
1790 if (rc->source_alt_ref_pending && !is_lossless_requested(&cpi->oxcf)) {
1791 // Reduce the boost of altref in the last gf group
1792 if (rc->frames_to_key - i == REDUCE_GF_LENGTH_BY ||
1793 rc->frames_to_key - i == 0) {
1794 rc->arf_boost_factor = LAST_ALR_BOOST_FACTOR;
1795 }
1796 }
1797
1798 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
1799
1800 // Reset the file position.
1801 reset_fpf_position(twopass, start_pos);
1802
1803 // Calculate the bits to be allocated to the gf/arf group as a whole
1804 gf_group_bits = calculate_total_gf_group_bits(cpi, gf_stats.gf_group_err);
1805 rc->gf_group_bits = gf_group_bits;
1806
1807 #if GROUP_ADAPTIVE_MAXQ
1808 // Calculate an estimate of the maxq needed for the group.
1809 // We are more agressive about correcting for sections
1810 // where there could be significant overshoot than for easier
1811 // sections where we do not wish to risk creating an overshoot
1812 // of the allocated bit budget.
1813 if ((cpi->oxcf.rc_mode != AOM_Q) && (rc->baseline_gf_interval > 1)) {
1814 const int vbr_group_bits_per_frame =
1815 (int)(gf_group_bits / rc->baseline_gf_interval);
1816 const double group_av_err =
1817 gf_stats.gf_group_raw_error / rc->baseline_gf_interval;
1818 const double group_av_skip_pct =
1819 gf_stats.gf_group_skip_pct / rc->baseline_gf_interval;
1820 const double group_av_inactive_zone =
1821 ((gf_stats.gf_group_inactive_zone_rows * 2) /
1822 (rc->baseline_gf_interval * (double)cm->mi_params.mb_rows));
1823
1824 int tmp_q;
1825 // rc factor is a weight factor that corrects for local rate control drift.
1826 double rc_factor = 1.0;
1827 int64_t bits = cpi->oxcf.target_bandwidth;
1828
1829 if (bits > 0) {
1830 int rate_error;
1831
1832 rate_error = (int)((rc->vbr_bits_off_target * 100) / bits);
1833 rate_error = clamp(rate_error, -100, 100);
1834 if (rate_error > 0) {
1835 rc_factor = AOMMAX(RC_FACTOR_MIN, (double)(100 - rate_error) / 100.0);
1836 } else {
1837 rc_factor = AOMMIN(RC_FACTOR_MAX, (double)(100 - rate_error) / 100.0);
1838 }
1839 }
1840
1841 tmp_q = get_twopass_worst_quality(
1842 cpi, group_av_err, (group_av_skip_pct + group_av_inactive_zone),
1843 vbr_group_bits_per_frame, rc_factor);
1844 rc->active_worst_quality = AOMMAX(tmp_q, rc->active_worst_quality >> 1);
1845 }
1846 #endif
1847
1848 // Adjust KF group bits and error remaining.
1849 if (is_final_pass)
1850 twopass->kf_group_error_left -= (int64_t)gf_stats.gf_group_err;
1851
1852 // Set up the structure of this Group-Of-Pictures (same as GF_GROUP)
1853 av1_gop_setup_structure(cpi, frame_params);
1854
1855 // Reset the file position.
1856 reset_fpf_position(twopass, start_pos);
1857
1858 // Calculate a section intra ratio used in setting max loop filter.
1859 if (frame_params->frame_type != KEY_FRAME) {
1860 twopass->section_intra_rating = calculate_section_intra_ratio(
1861 start_pos, twopass->stats_buf_ctx->stats_in_end,
1862 rc->baseline_gf_interval);
1863 }
1864
1865 // Reset rolling actual and target bits counters for ARF groups.
1866 twopass->rolling_arf_group_target_bits = 1;
1867 twopass->rolling_arf_group_actual_bits = 1;
1868
1869 av1_gop_bit_allocation(cpi, rc, gf_group,
1870 frame_params->frame_type == KEY_FRAME, use_alt_ref,
1871 gf_group_bits);
1872 }
1873
1874 // #define FIXED_ARF_BITS
1875 #ifdef FIXED_ARF_BITS
1876 #define ARF_BITS_FRACTION 0.75
1877 #endif
av1_gop_bit_allocation(const AV1_COMP * cpi,RATE_CONTROL * const rc,GF_GROUP * gf_group,int is_key_frame,int use_arf,int64_t gf_group_bits)1878 void av1_gop_bit_allocation(const AV1_COMP *cpi, RATE_CONTROL *const rc,
1879 GF_GROUP *gf_group, int is_key_frame, int use_arf,
1880 int64_t gf_group_bits) {
1881 // Calculate the extra bits to be used for boosted frame(s)
1882 #ifdef FIXED_ARF_BITS
1883 int gf_arf_bits = (int)(ARF_BITS_FRACTION * gf_group_bits);
1884 #else
1885 int gf_arf_bits = calculate_boost_bits(rc->baseline_gf_interval,
1886 rc->gfu_boost, gf_group_bits);
1887 #endif
1888
1889 gf_arf_bits = adjust_boost_bits_for_target_level(cpi, rc, gf_arf_bits,
1890 gf_group_bits, 1);
1891
1892 // Allocate bits to each of the frames in the GF group.
1893 allocate_gf_group_bits(gf_group, rc, gf_group_bits, gf_arf_bits, is_key_frame,
1894 use_arf);
1895 }
1896
1897 // Minimum % intra coding observed in first pass (1.0 = 100%)
1898 #define MIN_INTRA_LEVEL 0.25
1899 // Minimum ratio between the % of intra coding and inter coding in the first
1900 // pass after discounting neutral blocks (discounting neutral blocks in this
1901 // way helps catch scene cuts in clips with very flat areas or letter box
1902 // format clips with image padding.
1903 #define INTRA_VS_INTER_THRESH 2.0
1904 // Hard threshold where the first pass chooses intra for almost all blocks.
1905 // In such a case even if the frame is not a scene cut coding a key frame
1906 // may be a good option.
1907 #define VERY_LOW_INTER_THRESH 0.05
1908 // Maximum threshold for the relative ratio of intra error score vs best
1909 // inter error score.
1910 #define KF_II_ERR_THRESHOLD 2.5
1911 // In real scene cuts there is almost always a sharp change in the intra
1912 // or inter error score.
1913 #define ERR_CHANGE_THRESHOLD 0.4
1914 // For real scene cuts we expect an improvment in the intra inter error
1915 // ratio in the next frame.
1916 #define II_IMPROVEMENT_THRESHOLD 3.5
1917 #define KF_II_MAX 128.0
1918
1919 // Threshold for use of the lagging second reference frame. High second ref
1920 // usage may point to a transient event like a flash or occlusion rather than
1921 // a real scene cut.
1922 // We adapt the threshold based on number of frames in this key-frame group so
1923 // far.
get_second_ref_usage_thresh(int frame_count_so_far)1924 static double get_second_ref_usage_thresh(int frame_count_so_far) {
1925 const int adapt_upto = 32;
1926 const double min_second_ref_usage_thresh = 0.085;
1927 const double second_ref_usage_thresh_max_delta = 0.035;
1928 if (frame_count_so_far >= adapt_upto) {
1929 return min_second_ref_usage_thresh + second_ref_usage_thresh_max_delta;
1930 }
1931 return min_second_ref_usage_thresh +
1932 ((double)frame_count_so_far / (adapt_upto - 1)) *
1933 second_ref_usage_thresh_max_delta;
1934 }
1935
test_candidate_kf(TWO_PASS * twopass,const FIRSTPASS_STATS * last_frame,const FIRSTPASS_STATS * this_frame,const FIRSTPASS_STATS * next_frame,int frame_count_so_far,enum aom_rc_mode rc_mode)1936 static int test_candidate_kf(TWO_PASS *twopass,
1937 const FIRSTPASS_STATS *last_frame,
1938 const FIRSTPASS_STATS *this_frame,
1939 const FIRSTPASS_STATS *next_frame,
1940 int frame_count_so_far, enum aom_rc_mode rc_mode) {
1941 int is_viable_kf = 0;
1942 double pcnt_intra = 1.0 - this_frame->pcnt_inter;
1943 double modified_pcnt_inter =
1944 this_frame->pcnt_inter - this_frame->pcnt_neutral;
1945 const double second_ref_usage_thresh =
1946 get_second_ref_usage_thresh(frame_count_so_far);
1947
1948 // Does the frame satisfy the primary criteria of a key frame?
1949 // See above for an explanation of the test criteria.
1950 // If so, then examine how well it predicts subsequent frames.
1951 if (IMPLIES(rc_mode == AOM_Q, frame_count_so_far >= 3) &&
1952 (this_frame->pcnt_second_ref < second_ref_usage_thresh) &&
1953 (next_frame->pcnt_second_ref < second_ref_usage_thresh) &&
1954 ((this_frame->pcnt_inter < VERY_LOW_INTER_THRESH) ||
1955 ((pcnt_intra > MIN_INTRA_LEVEL) &&
1956 (pcnt_intra > (INTRA_VS_INTER_THRESH * modified_pcnt_inter)) &&
1957 ((this_frame->intra_error /
1958 DOUBLE_DIVIDE_CHECK(this_frame->coded_error)) <
1959 KF_II_ERR_THRESHOLD) &&
1960 ((fabs(last_frame->coded_error - this_frame->coded_error) /
1961 DOUBLE_DIVIDE_CHECK(this_frame->coded_error) >
1962 ERR_CHANGE_THRESHOLD) ||
1963 (fabs(last_frame->intra_error - this_frame->intra_error) /
1964 DOUBLE_DIVIDE_CHECK(this_frame->intra_error) >
1965 ERR_CHANGE_THRESHOLD) ||
1966 ((next_frame->intra_error /
1967 DOUBLE_DIVIDE_CHECK(next_frame->coded_error)) >
1968 II_IMPROVEMENT_THRESHOLD))))) {
1969 int i;
1970 const FIRSTPASS_STATS *start_pos = twopass->stats_in;
1971 FIRSTPASS_STATS local_next_frame = *next_frame;
1972 double boost_score = 0.0;
1973 double old_boost_score = 0.0;
1974 double decay_accumulator = 1.0;
1975
1976 // Examine how well the key frame predicts subsequent frames.
1977 for (i = 0; i < SCENE_CUT_KEY_TEST_INTERVAL; ++i) {
1978 double next_iiratio = (BOOST_FACTOR * local_next_frame.intra_error /
1979 DOUBLE_DIVIDE_CHECK(local_next_frame.coded_error));
1980
1981 if (next_iiratio > KF_II_MAX) next_iiratio = KF_II_MAX;
1982
1983 // Cumulative effect of decay in prediction quality.
1984 if (local_next_frame.pcnt_inter > 0.85)
1985 decay_accumulator *= local_next_frame.pcnt_inter;
1986 else
1987 decay_accumulator *= (0.85 + local_next_frame.pcnt_inter) / 2.0;
1988
1989 // Keep a running total.
1990 boost_score += (decay_accumulator * next_iiratio);
1991
1992 // Test various breakout clauses.
1993 if ((local_next_frame.pcnt_inter < 0.05) || (next_iiratio < 1.5) ||
1994 (((local_next_frame.pcnt_inter - local_next_frame.pcnt_neutral) <
1995 0.20) &&
1996 (next_iiratio < 3.0)) ||
1997 ((boost_score - old_boost_score) < 3.0) ||
1998 (local_next_frame.intra_error < 200)) {
1999 break;
2000 }
2001
2002 old_boost_score = boost_score;
2003
2004 // Get the next frame details
2005 if (EOF == input_stats(twopass, &local_next_frame)) break;
2006 }
2007
2008 // If there is tolerable prediction for at least the next 3 frames then
2009 // break out else discard this potential key frame and move on
2010 if (boost_score > 30.0 && (i > 3)) {
2011 is_viable_kf = 1;
2012 } else {
2013 // Reset the file position
2014 reset_fpf_position(twopass, start_pos);
2015
2016 is_viable_kf = 0;
2017 }
2018 }
2019
2020 return is_viable_kf;
2021 }
2022
2023 #define FRAMES_TO_CHECK_DECAY 8
2024 #define KF_MIN_FRAME_BOOST 80.0
2025 #define KF_MAX_FRAME_BOOST 128.0
2026 #define MIN_KF_BOOST 600 // Minimum boost for non-static KF interval
2027 #define MAX_KF_BOOST 3200
2028 #define MIN_STATIC_KF_BOOST 5400 // Minimum boost for static KF interval
2029
detect_app_forced_key(AV1_COMP * cpi)2030 static int detect_app_forced_key(AV1_COMP *cpi) {
2031 if (cpi->oxcf.fwd_kf_enabled) cpi->rc.next_is_fwd_key = 1;
2032 int num_frames_to_app_forced_key = is_forced_keyframe_pending(
2033 cpi->lookahead, cpi->lookahead->max_sz, cpi->compressor_stage);
2034 if (num_frames_to_app_forced_key != -1) cpi->rc.next_is_fwd_key = 0;
2035 return num_frames_to_app_forced_key;
2036 }
2037
get_projected_kf_boost(AV1_COMP * cpi)2038 static int get_projected_kf_boost(AV1_COMP *cpi) {
2039 /*
2040 * If num_stats_used_for_kf_boost >= frames_to_key, then
2041 * all stats needed for prior boost calculation are available.
2042 * Hence projecting the prior boost is not needed in this cases.
2043 */
2044 if (cpi->rc.num_stats_used_for_kf_boost >= cpi->rc.frames_to_key)
2045 return cpi->rc.kf_boost;
2046
2047 // Get the current tpl factor (number of frames = frames_to_key).
2048 double tpl_factor = av1_get_kf_boost_projection_factor(cpi->rc.frames_to_key);
2049 // Get the tpl factor when number of frames = num_stats_used_for_kf_boost.
2050 double tpl_factor_num_stats =
2051 av1_get_kf_boost_projection_factor(cpi->rc.num_stats_used_for_kf_boost);
2052 int projected_kf_boost =
2053 (int)rint((tpl_factor * cpi->rc.kf_boost) / tpl_factor_num_stats);
2054 return projected_kf_boost;
2055 }
2056
define_kf_interval(AV1_COMP * cpi,FIRSTPASS_STATS * this_frame,double * kf_group_err,int num_frames_to_detect_scenecut)2057 static int define_kf_interval(AV1_COMP *cpi, FIRSTPASS_STATS *this_frame,
2058 double *kf_group_err,
2059 int num_frames_to_detect_scenecut) {
2060 TWO_PASS *const twopass = &cpi->twopass;
2061 RATE_CONTROL *const rc = &cpi->rc;
2062 const AV1EncoderConfig *const oxcf = &cpi->oxcf;
2063 double recent_loop_decay[FRAMES_TO_CHECK_DECAY];
2064 FIRSTPASS_STATS last_frame;
2065 double decay_accumulator = 1.0;
2066 int i = 0, j;
2067 int frames_to_key = 1;
2068 int frames_since_key = rc->frames_since_key + 1;
2069 FRAME_INFO *const frame_info = &cpi->frame_info;
2070 int num_stats_used_for_kf_boost = 1;
2071 int scenecut_detected = 0;
2072
2073 int num_frames_to_next_key = detect_app_forced_key(cpi);
2074
2075 if (num_frames_to_detect_scenecut == 0) {
2076 if (num_frames_to_next_key != -1)
2077 return num_frames_to_next_key;
2078 else
2079 return rc->frames_to_key;
2080 }
2081
2082 if (num_frames_to_next_key != -1)
2083 num_frames_to_detect_scenecut =
2084 AOMMIN(num_frames_to_detect_scenecut, num_frames_to_next_key);
2085
2086 // Initialize the decay rates for the recent frames to check
2087 for (j = 0; j < FRAMES_TO_CHECK_DECAY; ++j) recent_loop_decay[j] = 1.0;
2088
2089 i = 0;
2090 while (twopass->stats_in < twopass->stats_buf_ctx->stats_in_end &&
2091 frames_to_key < num_frames_to_detect_scenecut) {
2092 // Accumulate total number of stats available till next key frame
2093 num_stats_used_for_kf_boost++;
2094
2095 // Accumulate kf group error.
2096 if (kf_group_err != NULL)
2097 *kf_group_err +=
2098 calculate_modified_err(frame_info, twopass, oxcf, this_frame);
2099
2100 // Load the next frame's stats.
2101 last_frame = *this_frame;
2102 input_stats(twopass, this_frame);
2103
2104 // Provided that we are not at the end of the file...
2105 if (cpi->rc.enable_scenecut_detection && cpi->oxcf.auto_key &&
2106 twopass->stats_in < twopass->stats_buf_ctx->stats_in_end) {
2107 double loop_decay_rate;
2108
2109 // Check for a scene cut.
2110 if (test_candidate_kf(twopass, &last_frame, this_frame, twopass->stats_in,
2111 frames_since_key, oxcf->rc_mode)) {
2112 scenecut_detected = 1;
2113 break;
2114 }
2115
2116 // How fast is the prediction quality decaying?
2117 loop_decay_rate =
2118 get_prediction_decay_rate(frame_info, twopass->stats_in);
2119
2120 // We want to know something about the recent past... rather than
2121 // as used elsewhere where we are concerned with decay in prediction
2122 // quality since the last GF or KF.
2123 recent_loop_decay[i % FRAMES_TO_CHECK_DECAY] = loop_decay_rate;
2124 decay_accumulator = 1.0;
2125 for (j = 0; j < FRAMES_TO_CHECK_DECAY; ++j)
2126 decay_accumulator *= recent_loop_decay[j];
2127
2128 // Special check for transition or high motion followed by a
2129 // static scene.
2130 if (detect_transition_to_still(twopass, rc->min_gf_interval, i,
2131 cpi->oxcf.key_freq - i, loop_decay_rate,
2132 decay_accumulator)) {
2133 scenecut_detected = 1;
2134 break;
2135 }
2136
2137 // Step on to the next frame.
2138 ++frames_to_key;
2139 ++frames_since_key;
2140
2141 // If we don't have a real key frame within the next two
2142 // key_freq intervals then break out of the loop.
2143 if (frames_to_key >= 2 * cpi->oxcf.key_freq) break;
2144 } else {
2145 ++frames_to_key;
2146 ++frames_since_key;
2147 }
2148 ++i;
2149 }
2150
2151 if (kf_group_err != NULL)
2152 rc->num_stats_used_for_kf_boost = num_stats_used_for_kf_boost;
2153
2154 if (cpi->lap_enabled && !scenecut_detected)
2155 frames_to_key = num_frames_to_next_key;
2156
2157 return frames_to_key;
2158 }
2159
find_next_key_frame(AV1_COMP * cpi,FIRSTPASS_STATS * this_frame)2160 static void find_next_key_frame(AV1_COMP *cpi, FIRSTPASS_STATS *this_frame) {
2161 RATE_CONTROL *const rc = &cpi->rc;
2162 TWO_PASS *const twopass = &cpi->twopass;
2163 GF_GROUP *const gf_group = &cpi->gf_group;
2164 FRAME_INFO *const frame_info = &cpi->frame_info;
2165 AV1_COMMON *const cm = &cpi->common;
2166 CurrentFrame *const current_frame = &cm->current_frame;
2167 const AV1EncoderConfig *const oxcf = &cpi->oxcf;
2168 const FIRSTPASS_STATS first_frame = *this_frame;
2169 FIRSTPASS_STATS next_frame;
2170 av1_zero(next_frame);
2171
2172 rc->frames_since_key = 0;
2173
2174 // Reset the GF group data structures.
2175 av1_zero(*gf_group);
2176
2177 // Clear the alt ref active flag and last group multi arf flags as they
2178 // can never be set for a key frame.
2179 rc->source_alt_ref_active = 0;
2180
2181 // KF is always a GF so clear frames till next gf counter.
2182 rc->frames_till_gf_update_due = 0;
2183
2184 rc->frames_to_key = 1;
2185
2186 if (has_no_stats_stage(cpi)) {
2187 int num_frames_to_app_forced_key = detect_app_forced_key(cpi);
2188 rc->this_key_frame_forced =
2189 current_frame->frame_number != 0 && rc->frames_to_key == 0;
2190 if (num_frames_to_app_forced_key != -1)
2191 rc->frames_to_key = num_frames_to_app_forced_key;
2192 else
2193 rc->frames_to_key = AOMMAX(1, cpi->oxcf.key_freq);
2194 correct_frames_to_key(cpi);
2195 rc->kf_boost = DEFAULT_KF_BOOST;
2196 rc->source_alt_ref_active = 0;
2197 gf_group->update_type[0] = KF_UPDATE;
2198 return;
2199 }
2200 int i;
2201 const FIRSTPASS_STATS *const start_position = twopass->stats_in;
2202 int kf_bits = 0;
2203 double zero_motion_accumulator = 1.0;
2204 double boost_score = 0.0;
2205 double kf_raw_err = 0.0;
2206 double kf_mod_err = 0.0;
2207 double kf_group_err = 0.0;
2208 double sr_accumulator = 0.0;
2209 int frames_to_key;
2210 // Is this a forced key frame by interval.
2211 rc->this_key_frame_forced = rc->next_key_frame_forced;
2212
2213 twopass->kf_group_bits = 0; // Total bits available to kf group
2214 twopass->kf_group_error_left = 0; // Group modified error score.
2215
2216 kf_raw_err = this_frame->intra_error;
2217 kf_mod_err = calculate_modified_err(frame_info, twopass, oxcf, this_frame);
2218
2219 frames_to_key =
2220 define_kf_interval(cpi, this_frame, &kf_group_err, oxcf->key_freq);
2221
2222 if (frames_to_key != -1)
2223 rc->frames_to_key = AOMMIN(oxcf->key_freq, frames_to_key);
2224 else
2225 rc->frames_to_key = oxcf->key_freq;
2226
2227 if (cpi->lap_enabled) correct_frames_to_key(cpi);
2228
2229 // If there is a max kf interval set by the user we must obey it.
2230 // We already breakout of the loop above at 2x max.
2231 // This code centers the extra kf if the actual natural interval
2232 // is between 1x and 2x.
2233 if (cpi->oxcf.auto_key && rc->frames_to_key > cpi->oxcf.key_freq) {
2234 FIRSTPASS_STATS tmp_frame = first_frame;
2235
2236 rc->frames_to_key /= 2;
2237
2238 // Reset to the start of the group.
2239 reset_fpf_position(twopass, start_position);
2240
2241 kf_group_err = 0.0;
2242
2243 // Rescan to get the correct error data for the forced kf group.
2244 for (i = 0; i < rc->frames_to_key; ++i) {
2245 kf_group_err +=
2246 calculate_modified_err(frame_info, twopass, oxcf, &tmp_frame);
2247 if (EOF == input_stats(twopass, &tmp_frame)) break;
2248 }
2249 rc->next_key_frame_forced = 1;
2250 } else if ((twopass->stats_in == twopass->stats_buf_ctx->stats_in_end &&
2251 is_stat_consumption_stage_twopass(cpi)) ||
2252 rc->frames_to_key >= cpi->oxcf.key_freq) {
2253 rc->next_key_frame_forced = 1;
2254 } else {
2255 rc->next_key_frame_forced = 0;
2256 }
2257
2258 // Special case for the last key frame of the file.
2259 if (twopass->stats_in >= twopass->stats_buf_ctx->stats_in_end) {
2260 // Accumulate kf group error.
2261 kf_group_err +=
2262 calculate_modified_err(frame_info, twopass, oxcf, this_frame);
2263 }
2264
2265 // Calculate the number of bits that should be assigned to the kf group.
2266 if (twopass->bits_left > 0 && twopass->modified_error_left > 0.0) {
2267 // Maximum number of bits for a single normal frame (not key frame).
2268 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
2269
2270 // Maximum number of bits allocated to the key frame group.
2271 int64_t max_grp_bits;
2272
2273 // Default allocation based on bits left and relative
2274 // complexity of the section.
2275 twopass->kf_group_bits = (int64_t)(
2276 twopass->bits_left * (kf_group_err / twopass->modified_error_left));
2277
2278 // Clip based on maximum per frame rate defined by the user.
2279 max_grp_bits = (int64_t)max_bits * (int64_t)rc->frames_to_key;
2280 if (twopass->kf_group_bits > max_grp_bits)
2281 twopass->kf_group_bits = max_grp_bits;
2282 } else {
2283 twopass->kf_group_bits = 0;
2284 }
2285 twopass->kf_group_bits = AOMMAX(0, twopass->kf_group_bits);
2286
2287 // Reset the first pass file position.
2288 reset_fpf_position(twopass, start_position);
2289
2290 // Scan through the kf group collating various stats used to determine
2291 // how many bits to spend on it.
2292 boost_score = 0.0;
2293 const double kf_max_boost =
2294 cpi->oxcf.rc_mode == AOM_Q
2295 ? AOMMIN(AOMMAX(rc->frames_to_key * 2.0, KF_MIN_FRAME_BOOST),
2296 KF_MAX_FRAME_BOOST)
2297 : KF_MAX_FRAME_BOOST;
2298 for (i = 0; i < (rc->frames_to_key - 1); ++i) {
2299 if (EOF == input_stats(twopass, &next_frame)) break;
2300
2301 // Monitor for static sections.
2302 // For the first frame in kf group, the second ref indicator is invalid.
2303 if (i > 0) {
2304 zero_motion_accumulator =
2305 AOMMIN(zero_motion_accumulator,
2306 get_zero_motion_factor(frame_info, &next_frame));
2307 } else {
2308 zero_motion_accumulator = next_frame.pcnt_inter - next_frame.pcnt_motion;
2309 }
2310
2311 // Not all frames in the group are necessarily used in calculating boost.
2312 if ((sr_accumulator < (kf_raw_err * 1.50)) &&
2313 (i <= rc->max_gf_interval * 2)) {
2314 double frame_boost;
2315 double zm_factor;
2316
2317 // Factor 0.75-1.25 based on how much of frame is static.
2318 zm_factor = (0.75 + (zero_motion_accumulator / 2.0));
2319
2320 if (i < 2) sr_accumulator = 0.0;
2321 frame_boost = calc_kf_frame_boost(rc, frame_info, &next_frame,
2322 &sr_accumulator, kf_max_boost);
2323 boost_score += frame_boost * zm_factor;
2324 }
2325 }
2326
2327 reset_fpf_position(twopass, start_position);
2328
2329 // Store the zero motion percentage
2330 twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
2331
2332 // Calculate a section intra ratio used in setting max loop filter.
2333 twopass->section_intra_rating = calculate_section_intra_ratio(
2334 start_position, twopass->stats_buf_ctx->stats_in_end, rc->frames_to_key);
2335
2336 rc->kf_boost = (int)boost_score;
2337
2338 if (cpi->lap_enabled) {
2339 rc->kf_boost = get_projected_kf_boost(cpi);
2340 }
2341
2342 // Special case for static / slide show content but don't apply
2343 // if the kf group is very short.
2344 if ((zero_motion_accumulator > STATIC_KF_GROUP_FLOAT_THRESH) &&
2345 (rc->frames_to_key > 8)) {
2346 rc->kf_boost = AOMMAX(rc->kf_boost, MIN_STATIC_KF_BOOST);
2347 } else {
2348 // Apply various clamps for min and max boost
2349 rc->kf_boost = AOMMAX(rc->kf_boost, (rc->frames_to_key * 3));
2350 rc->kf_boost = AOMMAX(rc->kf_boost, MIN_KF_BOOST);
2351 #ifdef STRICT_RC
2352 rc->kf_boost = AOMMIN(rc->kf_boost, MAX_KF_BOOST);
2353 #endif
2354 }
2355
2356 // Work out how many bits to allocate for the key frame itself.
2357 kf_bits = calculate_boost_bits((rc->frames_to_key - 1), rc->kf_boost,
2358 twopass->kf_group_bits);
2359 // printf("kf boost = %d kf_bits = %d kf_zeromotion_pct = %d\n", rc->kf_boost,
2360 // kf_bits, twopass->kf_zeromotion_pct);
2361 kf_bits = adjust_boost_bits_for_target_level(cpi, rc, kf_bits,
2362 twopass->kf_group_bits, 0);
2363
2364 twopass->kf_group_bits -= kf_bits;
2365
2366 // Save the bits to spend on the key frame.
2367 gf_group->bit_allocation[0] = kf_bits;
2368 gf_group->update_type[0] = KF_UPDATE;
2369
2370 // Note the total error score of the kf group minus the key frame itself.
2371 twopass->kf_group_error_left = (int)(kf_group_err - kf_mod_err);
2372
2373 // Adjust the count of total modified error left.
2374 // The count of bits left is adjusted elsewhere based on real coded frame
2375 // sizes.
2376 twopass->modified_error_left -= kf_group_err;
2377 }
2378
is_skippable_frame(const AV1_COMP * cpi)2379 static int is_skippable_frame(const AV1_COMP *cpi) {
2380 if (has_no_stats_stage(cpi)) return 0;
2381 // If the current frame does not have non-zero motion vector detected in the
2382 // first pass, and so do its previous and forward frames, then this frame
2383 // can be skipped for partition check, and the partition size is assigned
2384 // according to the variance
2385 const TWO_PASS *const twopass = &cpi->twopass;
2386
2387 return (!frame_is_intra_only(&cpi->common) &&
2388 twopass->stats_in - 2 > twopass->stats_buf_ctx->stats_in_start &&
2389 twopass->stats_in < twopass->stats_buf_ctx->stats_in_end &&
2390 (twopass->stats_in - 1)->pcnt_inter -
2391 (twopass->stats_in - 1)->pcnt_motion ==
2392 1 &&
2393 (twopass->stats_in - 2)->pcnt_inter -
2394 (twopass->stats_in - 2)->pcnt_motion ==
2395 1 &&
2396 twopass->stats_in->pcnt_inter - twopass->stats_in->pcnt_motion == 1);
2397 }
2398
2399 #define ARF_STATS_OUTPUT 0
2400 #if ARF_STATS_OUTPUT
2401 unsigned int arf_count = 0;
2402 #endif
2403 #define DEFAULT_GRP_WEIGHT 1.0
2404
process_first_pass_stats(AV1_COMP * cpi,FIRSTPASS_STATS * this_frame)2405 static void process_first_pass_stats(AV1_COMP *cpi,
2406 FIRSTPASS_STATS *this_frame) {
2407 AV1_COMMON *const cm = &cpi->common;
2408 CurrentFrame *const current_frame = &cm->current_frame;
2409 RATE_CONTROL *const rc = &cpi->rc;
2410 TWO_PASS *const twopass = &cpi->twopass;
2411
2412 if (cpi->oxcf.rc_mode != AOM_Q && current_frame->frame_number == 0 &&
2413 cpi->twopass.stats_buf_ctx->total_stats &&
2414 cpi->twopass.stats_buf_ctx->total_left_stats) {
2415 if (cpi->lap_enabled) {
2416 /*
2417 * Accumulate total_stats using available limited number of stats,
2418 * and assign it to total_left_stats.
2419 */
2420 *cpi->twopass.stats_buf_ctx->total_left_stats =
2421 *cpi->twopass.stats_buf_ctx->total_stats;
2422 }
2423 const int frames_left = (int)(twopass->stats_buf_ctx->total_stats->count -
2424 current_frame->frame_number);
2425
2426 // Special case code for first frame.
2427 const int section_target_bandwidth =
2428 (int)(twopass->bits_left / frames_left);
2429 const double section_length =
2430 twopass->stats_buf_ctx->total_left_stats->count;
2431 const double section_error =
2432 twopass->stats_buf_ctx->total_left_stats->coded_error / section_length;
2433 const double section_intra_skip =
2434 twopass->stats_buf_ctx->total_left_stats->intra_skip_pct /
2435 section_length;
2436 const double section_inactive_zone =
2437 (twopass->stats_buf_ctx->total_left_stats->inactive_zone_rows * 2) /
2438 ((double)cm->mi_params.mb_rows * section_length);
2439 const int tmp_q = get_twopass_worst_quality(
2440 cpi, section_error, section_intra_skip + section_inactive_zone,
2441 section_target_bandwidth, DEFAULT_GRP_WEIGHT);
2442
2443 rc->active_worst_quality = tmp_q;
2444 rc->ni_av_qi = tmp_q;
2445 rc->last_q[INTER_FRAME] = tmp_q;
2446 rc->avg_q = av1_convert_qindex_to_q(tmp_q, cm->seq_params.bit_depth);
2447 rc->avg_frame_qindex[INTER_FRAME] = tmp_q;
2448 rc->last_q[KEY_FRAME] = (tmp_q + cpi->oxcf.best_allowed_q) / 2;
2449 rc->avg_frame_qindex[KEY_FRAME] = rc->last_q[KEY_FRAME];
2450 }
2451
2452 int err = 0;
2453 if (cpi->lap_enabled) {
2454 err = input_stats_lap(twopass, this_frame);
2455 } else {
2456 err = input_stats(twopass, this_frame);
2457 }
2458 if (err == EOF) return;
2459
2460 {
2461 const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
2462 ? cpi->initial_mbs
2463 : cm->mi_params.MBs;
2464 // The multiplication by 256 reverses a scaling factor of (>> 8)
2465 // applied when combining MB error values for the frame.
2466 twopass->mb_av_energy = log((this_frame->intra_error / num_mbs) + 1.0);
2467 twopass->frame_avg_haar_energy =
2468 log((this_frame->frame_avg_wavelet_energy / num_mbs) + 1.0);
2469 }
2470
2471 // Update the total stats remaining structure.
2472 if (twopass->stats_buf_ctx->total_left_stats)
2473 subtract_stats(twopass->stats_buf_ctx->total_left_stats, this_frame);
2474
2475 // Set the frame content type flag.
2476 if (this_frame->intra_skip_pct >= FC_ANIMATION_THRESH)
2477 twopass->fr_content_type = FC_GRAPHICS_ANIMATION;
2478 else
2479 twopass->fr_content_type = FC_NORMAL;
2480 }
2481
setup_target_rate(AV1_COMP * cpi)2482 static void setup_target_rate(AV1_COMP *cpi) {
2483 RATE_CONTROL *const rc = &cpi->rc;
2484 GF_GROUP *const gf_group = &cpi->gf_group;
2485
2486 int target_rate = gf_group->bit_allocation[gf_group->index];
2487
2488 if (has_no_stats_stage(cpi)) {
2489 av1_rc_set_frame_target(cpi, target_rate, cpi->common.width,
2490 cpi->common.height);
2491 }
2492
2493 rc->base_frame_target = target_rate;
2494 }
2495
av1_get_second_pass_params(AV1_COMP * cpi,EncodeFrameParams * const frame_params,const EncodeFrameInput * const frame_input,unsigned int frame_flags)2496 void av1_get_second_pass_params(AV1_COMP *cpi,
2497 EncodeFrameParams *const frame_params,
2498 const EncodeFrameInput *const frame_input,
2499 unsigned int frame_flags) {
2500 RATE_CONTROL *const rc = &cpi->rc;
2501 TWO_PASS *const twopass = &cpi->twopass;
2502 GF_GROUP *const gf_group = &cpi->gf_group;
2503 AV1_COMMON *cm = &cpi->common;
2504
2505 if (frame_is_intra_only(cm)) {
2506 FeatureFlags *const features = &cm->features;
2507 av1_set_screen_content_options(cpi, features);
2508 cpi->is_screen_content_type = features->allow_screen_content_tools;
2509 }
2510
2511 if (is_stat_consumption_stage(cpi) && !twopass->stats_in) return;
2512
2513 if (rc->frames_till_gf_update_due > 0 && !(frame_flags & FRAMEFLAGS_KEY)) {
2514 assert(gf_group->index < gf_group->size);
2515 const int update_type = gf_group->update_type[gf_group->index];
2516
2517 setup_target_rate(cpi);
2518
2519 // If this is an arf frame then we dont want to read the stats file or
2520 // advance the input pointer as we already have what we need.
2521 if (update_type == ARF_UPDATE || update_type == INTNL_ARF_UPDATE) {
2522 if (cpi->no_show_kf) {
2523 assert(update_type == ARF_UPDATE);
2524 frame_params->frame_type = KEY_FRAME;
2525 } else {
2526 frame_params->frame_type = INTER_FRAME;
2527 }
2528
2529 // Do the firstpass stats indicate that this frame is skippable for the
2530 // partition search?
2531 if (cpi->sf.part_sf.allow_partition_search_skip && cpi->oxcf.pass == 2) {
2532 cpi->partition_search_skippable_frame = is_skippable_frame(cpi);
2533 }
2534
2535 return;
2536 }
2537 }
2538
2539 aom_clear_system_state();
2540
2541 if (cpi->oxcf.rc_mode == AOM_Q) rc->active_worst_quality = cpi->oxcf.cq_level;
2542 FIRSTPASS_STATS this_frame;
2543 av1_zero(this_frame);
2544 // call above fn
2545 if (is_stat_consumption_stage(cpi)) {
2546 process_first_pass_stats(cpi, &this_frame);
2547 } else {
2548 rc->active_worst_quality = cpi->oxcf.cq_level;
2549 }
2550
2551 // Keyframe and section processing.
2552 if (rc->frames_to_key == 0 || (frame_flags & FRAMEFLAGS_KEY)) {
2553 FIRSTPASS_STATS this_frame_copy;
2554 this_frame_copy = this_frame;
2555 frame_params->frame_type = KEY_FRAME;
2556 // Define next KF group and assign bits to it.
2557 find_next_key_frame(cpi, &this_frame);
2558 this_frame = this_frame_copy;
2559 } else {
2560 frame_params->frame_type = INTER_FRAME;
2561 const int altref_enabled = is_altref_enabled(cpi);
2562 const int sframe_dist = cpi->oxcf.sframe_dist;
2563 const int sframe_mode = cpi->oxcf.sframe_mode;
2564 const int sframe_enabled = cpi->oxcf.sframe_enabled;
2565 const int update_type = gf_group->update_type[gf_group->index];
2566 CurrentFrame *const current_frame = &cpi->common.current_frame;
2567 if (sframe_enabled) {
2568 if (altref_enabled) {
2569 if (sframe_mode == 1) {
2570 // sframe_mode == 1: insert sframe if it matches altref frame.
2571 if (current_frame->frame_number % sframe_dist == 0 &&
2572 current_frame->frame_number != 0 && update_type == ARF_UPDATE) {
2573 frame_params->frame_type = S_FRAME;
2574 }
2575 } else {
2576 // sframe_mode != 1: if sframe will be inserted at the next available
2577 // altref frame
2578 if (current_frame->frame_number % sframe_dist == 0 &&
2579 current_frame->frame_number != 0) {
2580 rc->sframe_due = 1;
2581 }
2582 if (rc->sframe_due && update_type == ARF_UPDATE) {
2583 frame_params->frame_type = S_FRAME;
2584 rc->sframe_due = 0;
2585 }
2586 }
2587 } else {
2588 if (current_frame->frame_number % sframe_dist == 0 &&
2589 current_frame->frame_number != 0) {
2590 frame_params->frame_type = S_FRAME;
2591 }
2592 }
2593 }
2594 }
2595
2596 // Define a new GF/ARF group. (Should always enter here for key frames).
2597 if (rc->frames_till_gf_update_due == 0) {
2598 assert(cpi->common.current_frame.frame_number == 0 ||
2599 gf_group->index == gf_group->size);
2600 const FIRSTPASS_STATS *const start_position = twopass->stats_in;
2601 int num_frames_to_detect_scenecut, frames_to_key;
2602 if (cpi->lap_enabled && cpi->rc.enable_scenecut_detection)
2603 num_frames_to_detect_scenecut = MAX_GF_LENGTH_LAP + 1;
2604 else
2605 num_frames_to_detect_scenecut = 0;
2606 frames_to_key = define_kf_interval(cpi, &this_frame, NULL,
2607 num_frames_to_detect_scenecut);
2608 reset_fpf_position(twopass, start_position);
2609 if (frames_to_key != -1)
2610 rc->frames_to_key = AOMMIN(rc->frames_to_key, frames_to_key);
2611
2612 int max_gop_length = (cpi->oxcf.lag_in_frames >= 32 &&
2613 is_stat_consumption_stage_twopass(cpi))
2614 ? MAX_GF_INTERVAL
2615 : MAX_GF_LENGTH_LAP;
2616 if (rc->intervals_till_gf_calculate_due == 0) {
2617 calculate_gf_length(cpi, max_gop_length, MAX_NUM_GF_INTERVALS);
2618 }
2619
2620 if (max_gop_length > 16) {
2621 if (rc->gf_intervals[rc->cur_gf_index] - 1 > 16) {
2622 // The calculate_gf_length function is previously used with
2623 // max_gop_length = 32 with look-ahead gf intervals.
2624 define_gf_group(cpi, &this_frame, frame_params, max_gop_length, 0);
2625 if (!av1_tpl_setup_stats(cpi, 1, frame_params, frame_input)) {
2626 // Tpl decides that a shorter gf interval is better.
2627 // TODO(jingning): Remove redundant computations here.
2628 max_gop_length = 16;
2629 calculate_gf_length(cpi, max_gop_length, 1);
2630 }
2631 } else {
2632 // Even based on 32 we still decide to use a short gf interval.
2633 // Better to re-decide based on 16 then
2634 max_gop_length = 16;
2635 calculate_gf_length(cpi, max_gop_length, 1);
2636 }
2637 }
2638 define_gf_group(cpi, &this_frame, frame_params, max_gop_length, 1);
2639 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
2640 cpi->num_gf_group_show_frames = 0;
2641 assert(gf_group->index == 0);
2642
2643 #if ARF_STATS_OUTPUT
2644 {
2645 FILE *fpfile;
2646 fpfile = fopen("arf.stt", "a");
2647 ++arf_count;
2648 fprintf(fpfile, "%10d %10d %10d %10d %10d\n",
2649 cpi->common.current_frame.frame_number,
2650 rc->frames_till_gf_update_due, rc->kf_boost, arf_count,
2651 rc->gfu_boost);
2652
2653 fclose(fpfile);
2654 }
2655 #endif
2656 }
2657 assert(gf_group->index < gf_group->size);
2658
2659 // Do the firstpass stats indicate that this frame is skippable for the
2660 // partition search?
2661 if (cpi->sf.part_sf.allow_partition_search_skip && cpi->oxcf.pass == 2) {
2662 cpi->partition_search_skippable_frame = is_skippable_frame(cpi);
2663 }
2664
2665 setup_target_rate(cpi);
2666 }
2667
av1_init_second_pass(AV1_COMP * cpi)2668 void av1_init_second_pass(AV1_COMP *cpi) {
2669 const AV1EncoderConfig *const oxcf = &cpi->oxcf;
2670 TWO_PASS *const twopass = &cpi->twopass;
2671 FRAME_INFO *const frame_info = &cpi->frame_info;
2672 double frame_rate;
2673 FIRSTPASS_STATS *stats;
2674
2675 if (!twopass->stats_buf_ctx->stats_in_end) return;
2676
2677 stats = twopass->stats_buf_ctx->total_stats;
2678
2679 *stats = *twopass->stats_buf_ctx->stats_in_end;
2680 *twopass->stats_buf_ctx->total_left_stats = *stats;
2681
2682 frame_rate = 10000000.0 * stats->count / stats->duration;
2683 // Each frame can have a different duration, as the frame rate in the source
2684 // isn't guaranteed to be constant. The frame rate prior to the first frame
2685 // encoded in the second pass is a guess. However, the sum duration is not.
2686 // It is calculated based on the actual durations of all frames from the
2687 // first pass.
2688 av1_new_framerate(cpi, frame_rate);
2689 twopass->bits_left =
2690 (int64_t)(stats->duration * oxcf->target_bandwidth / 10000000.0);
2691
2692 // This variable monitors how far behind the second ref update is lagging.
2693 twopass->sr_update_lag = 1;
2694
2695 // Scan the first pass file and calculate a modified total error based upon
2696 // the bias/power function used to allocate bits.
2697 {
2698 const double avg_error =
2699 stats->coded_error / DOUBLE_DIVIDE_CHECK(stats->count);
2700 const FIRSTPASS_STATS *s = twopass->stats_in;
2701 double modified_error_total = 0.0;
2702 twopass->modified_error_min =
2703 (avg_error * oxcf->two_pass_vbrmin_section) / 100;
2704 twopass->modified_error_max =
2705 (avg_error * oxcf->two_pass_vbrmax_section) / 100;
2706 while (s < twopass->stats_buf_ctx->stats_in_end) {
2707 modified_error_total +=
2708 calculate_modified_err(frame_info, twopass, oxcf, s);
2709 ++s;
2710 }
2711 twopass->modified_error_left = modified_error_total;
2712 }
2713
2714 // Reset the vbr bits off target counters
2715 cpi->rc.vbr_bits_off_target = 0;
2716 cpi->rc.vbr_bits_off_target_fast = 0;
2717
2718 cpi->rc.rate_error_estimate = 0;
2719
2720 // Static sequence monitor variables.
2721 twopass->kf_zeromotion_pct = 100;
2722 twopass->last_kfgroup_zeromotion_pct = 100;
2723
2724 // Initialize bits per macro_block estimate correction factor.
2725 twopass->bpm_factor = 1.0;
2726 // Initialize actual and target bits counters for ARF groups so that
2727 // at the start we have a neutral bpm adjustment.
2728 twopass->rolling_arf_group_target_bits = 1;
2729 twopass->rolling_arf_group_actual_bits = 1;
2730 }
2731
av1_init_single_pass_lap(AV1_COMP * cpi)2732 void av1_init_single_pass_lap(AV1_COMP *cpi) {
2733 TWO_PASS *const twopass = &cpi->twopass;
2734
2735 if (!twopass->stats_buf_ctx->stats_in_end) return;
2736
2737 // This variable monitors how far behind the second ref update is lagging.
2738 twopass->sr_update_lag = 1;
2739
2740 twopass->bits_left = 0;
2741 twopass->modified_error_min = 0.0;
2742 twopass->modified_error_max = 0.0;
2743 twopass->modified_error_left = 0.0;
2744
2745 // Reset the vbr bits off target counters
2746 cpi->rc.vbr_bits_off_target = 0;
2747 cpi->rc.vbr_bits_off_target_fast = 0;
2748
2749 cpi->rc.rate_error_estimate = 0;
2750
2751 // Static sequence monitor variables.
2752 twopass->kf_zeromotion_pct = 100;
2753 twopass->last_kfgroup_zeromotion_pct = 100;
2754
2755 // Initialize bits per macro_block estimate correction factor.
2756 twopass->bpm_factor = 1.0;
2757 // Initialize actual and target bits counters for ARF groups so that
2758 // at the start we have a neutral bpm adjustment.
2759 twopass->rolling_arf_group_target_bits = 1;
2760 twopass->rolling_arf_group_actual_bits = 1;
2761 }
2762
2763 #define MINQ_ADJ_LIMIT 48
2764 #define MINQ_ADJ_LIMIT_CQ 20
2765 #define HIGH_UNDERSHOOT_RATIO 2
av1_twopass_postencode_update(AV1_COMP * cpi)2766 void av1_twopass_postencode_update(AV1_COMP *cpi) {
2767 TWO_PASS *const twopass = &cpi->twopass;
2768 RATE_CONTROL *const rc = &cpi->rc;
2769 const int bits_used = rc->base_frame_target;
2770
2771 // VBR correction is done through rc->vbr_bits_off_target. Based on the
2772 // sign of this value, a limited % adjustment is made to the target rate
2773 // of subsequent frames, to try and push it back towards 0. This method
2774 // is designed to prevent extreme behaviour at the end of a clip
2775 // or group of frames.
2776 rc->vbr_bits_off_target += rc->base_frame_target - rc->projected_frame_size;
2777 twopass->bits_left = AOMMAX(twopass->bits_left - bits_used, 0);
2778
2779 // Target vs actual bits for this arf group.
2780 twopass->rolling_arf_group_target_bits += rc->this_frame_target;
2781 twopass->rolling_arf_group_actual_bits += rc->projected_frame_size;
2782
2783 // Calculate the pct rc error.
2784 if (rc->total_actual_bits) {
2785 rc->rate_error_estimate =
2786 (int)((rc->vbr_bits_off_target * 100) / rc->total_actual_bits);
2787 rc->rate_error_estimate = clamp(rc->rate_error_estimate, -100, 100);
2788 } else {
2789 rc->rate_error_estimate = 0;
2790 }
2791
2792 // Update the active best quality pyramid.
2793 if (!rc->is_src_frame_alt_ref) {
2794 const int pyramid_level = cpi->gf_group.layer_depth[cpi->gf_group.index];
2795 int i;
2796 for (i = pyramid_level; i <= MAX_ARF_LAYERS; ++i) {
2797 rc->active_best_quality[i] = cpi->common.quant_params.base_qindex;
2798 // if (pyramid_level >= 2) {
2799 // rc->active_best_quality[pyramid_level] =
2800 // AOMMAX(rc->active_best_quality[pyramid_level],
2801 // cpi->common.base_qindex);
2802 // }
2803 }
2804 }
2805
2806 #if 0
2807 {
2808 AV1_COMMON *cm = &cpi->common;
2809 FILE *fpfile;
2810 fpfile = fopen("details.stt", "a");
2811 fprintf(fpfile,
2812 "%10d %10d %10d %10" PRId64 " %10" PRId64
2813 " %10d %10d %10d %10.4lf %10.4lf %10.4lf %10.4lf\n",
2814 cm->current_frame.frame_number, rc->base_frame_target,
2815 rc->projected_frame_size, rc->total_actual_bits,
2816 rc->vbr_bits_off_target, rc->rate_error_estimate,
2817 twopass->rolling_arf_group_target_bits,
2818 twopass->rolling_arf_group_actual_bits,
2819 (double)twopass->rolling_arf_group_actual_bits /
2820 (double)twopass->rolling_arf_group_target_bits,
2821 twopass->bpm_factor,
2822 av1_convert_qindex_to_q(quant_params->base_qindex,
2823 cm->seq_params.bit_depth),
2824 av1_convert_qindex_to_q(rc->active_worst_quality,
2825 cm->seq_params.bit_depth));
2826 fclose(fpfile);
2827 }
2828 #endif
2829
2830 if (cpi->common.current_frame.frame_type != KEY_FRAME) {
2831 twopass->kf_group_bits -= bits_used;
2832 twopass->last_kfgroup_zeromotion_pct = twopass->kf_zeromotion_pct;
2833 }
2834 twopass->kf_group_bits = AOMMAX(twopass->kf_group_bits, 0);
2835
2836 // If the rate control is drifting consider adjustment to min or maxq.
2837 if ((cpi->oxcf.rc_mode != AOM_Q) && !cpi->rc.is_src_frame_alt_ref) {
2838 const int maxq_adj_limit = rc->worst_quality - rc->active_worst_quality;
2839 const int minq_adj_limit =
2840 (cpi->oxcf.rc_mode == AOM_CQ ? MINQ_ADJ_LIMIT_CQ : MINQ_ADJ_LIMIT);
2841
2842 // Undershoot.
2843 if (rc->rate_error_estimate > cpi->oxcf.under_shoot_pct) {
2844 --twopass->extend_maxq;
2845 if (rc->rolling_target_bits >= rc->rolling_actual_bits)
2846 ++twopass->extend_minq;
2847 // Overshoot.
2848 } else if (rc->rate_error_estimate < -cpi->oxcf.over_shoot_pct) {
2849 --twopass->extend_minq;
2850 if (rc->rolling_target_bits < rc->rolling_actual_bits)
2851 ++twopass->extend_maxq;
2852 } else {
2853 // Adjustment for extreme local overshoot.
2854 if (rc->projected_frame_size > (2 * rc->base_frame_target) &&
2855 rc->projected_frame_size > (2 * rc->avg_frame_bandwidth))
2856 ++twopass->extend_maxq;
2857
2858 // Unwind undershoot or overshoot adjustment.
2859 if (rc->rolling_target_bits < rc->rolling_actual_bits)
2860 --twopass->extend_minq;
2861 else if (rc->rolling_target_bits > rc->rolling_actual_bits)
2862 --twopass->extend_maxq;
2863 }
2864
2865 twopass->extend_minq = clamp(twopass->extend_minq, 0, minq_adj_limit);
2866 twopass->extend_maxq = clamp(twopass->extend_maxq, 0, maxq_adj_limit);
2867
2868 // If there is a big and undexpected undershoot then feed the extra
2869 // bits back in quickly. One situation where this may happen is if a
2870 // frame is unexpectedly almost perfectly predicted by the ARF or GF
2871 // but not very well predcited by the previous frame.
2872 if (!frame_is_kf_gf_arf(cpi) && !cpi->rc.is_src_frame_alt_ref) {
2873 int fast_extra_thresh = rc->base_frame_target / HIGH_UNDERSHOOT_RATIO;
2874 if (rc->projected_frame_size < fast_extra_thresh) {
2875 rc->vbr_bits_off_target_fast +=
2876 fast_extra_thresh - rc->projected_frame_size;
2877 rc->vbr_bits_off_target_fast =
2878 AOMMIN(rc->vbr_bits_off_target_fast, (4 * rc->avg_frame_bandwidth));
2879
2880 // Fast adaptation of minQ if necessary to use up the extra bits.
2881 if (rc->avg_frame_bandwidth) {
2882 twopass->extend_minq_fast =
2883 (int)(rc->vbr_bits_off_target_fast * 8 / rc->avg_frame_bandwidth);
2884 }
2885 twopass->extend_minq_fast = AOMMIN(
2886 twopass->extend_minq_fast, minq_adj_limit - twopass->extend_minq);
2887 } else if (rc->vbr_bits_off_target_fast) {
2888 twopass->extend_minq_fast = AOMMIN(
2889 twopass->extend_minq_fast, minq_adj_limit - twopass->extend_minq);
2890 } else {
2891 twopass->extend_minq_fast = 0;
2892 }
2893 }
2894 }
2895 }
2896