1 /*
2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include <limits.h>
12 #include <math.h>
13 #include <stdio.h>
14
15 #include "./vpx_dsp_rtcd.h"
16 #include "./vpx_scale_rtcd.h"
17
18 #include "vpx_dsp/vpx_dsp_common.h"
19 #include "vpx_mem/vpx_mem.h"
20 #include "vpx_ports/mem.h"
21 #include "vpx_ports/system_state.h"
22 #include "vpx_scale/vpx_scale.h"
23 #include "vpx_scale/yv12config.h"
24
25 #include "vp9/common/vp9_entropymv.h"
26 #include "vp9/common/vp9_quant_common.h"
27 #include "vp9/common/vp9_reconinter.h" // vp9_setup_dst_planes()
28 #include "vp9/encoder/vp9_aq_variance.h"
29 #include "vp9/encoder/vp9_block.h"
30 #include "vp9/encoder/vp9_encodeframe.h"
31 #include "vp9/encoder/vp9_encodemb.h"
32 #include "vp9/encoder/vp9_encodemv.h"
33 #include "vp9/encoder/vp9_encoder.h"
34 #include "vp9/encoder/vp9_ethread.h"
35 #include "vp9/encoder/vp9_extend.h"
36 #include "vp9/encoder/vp9_firstpass.h"
37 #include "vp9/encoder/vp9_mcomp.h"
38 #include "vp9/encoder/vp9_quantize.h"
39 #include "vp9/encoder/vp9_rd.h"
40 #include "vpx_dsp/variance.h"
41
42 #define OUTPUT_FPF 0
43 #define ARF_STATS_OUTPUT 0
44 #define COMPLEXITY_STATS_OUTPUT 0
45
46 #define FIRST_PASS_Q 10.0
47 #define NORMAL_BOOST 100
48 #define MIN_ARF_GF_BOOST 250
49 #define MIN_DECAY_FACTOR 0.01
50 #define NEW_MV_MODE_PENALTY 32
51 #define DARK_THRESH 64
52 #define LOW_I_THRESH 24000
53
54 #define NCOUNT_INTRA_THRESH 8192
55 #define NCOUNT_INTRA_FACTOR 3
56
57 #define INTRA_PART 0.005
58 #define DEFAULT_DECAY_LIMIT 0.75
59 #define LOW_SR_DIFF_TRHESH 0.1
60 #define LOW_CODED_ERR_PER_MB 10.0
61 #define NCOUNT_FRAME_II_THRESH 6.0
62 #define BASELINE_ERR_PER_MB 12500.0
63 #define GF_MAX_FRAME_BOOST 96.0
64
65 #ifdef AGGRESSIVE_VBR
66 #define KF_MIN_FRAME_BOOST 40.0
67 #define KF_MAX_FRAME_BOOST 80.0
68 #define MAX_KF_TOT_BOOST 4800
69 #else
70 #define KF_MIN_FRAME_BOOST 40.0
71 #define KF_MAX_FRAME_BOOST 96.0
72 #define MAX_KF_TOT_BOOST 5400
73 #endif
74
75 #define DEFAULT_ZM_FACTOR 0.5
76 #define MINQ_ADJ_LIMIT 48
77 #define MINQ_ADJ_LIMIT_CQ 20
78 #define HIGH_UNDERSHOOT_RATIO 2
79 #define AV_WQ_FACTOR 4.0
80
81 #define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x)-0.000001 : (x) + 0.000001)
82
83 #if ARF_STATS_OUTPUT
84 unsigned int arf_count = 0;
85 #endif
86
87 // Resets the first pass file to the given position using a relative seek from
88 // the current position.
reset_fpf_position(TWO_PASS * p,const FIRSTPASS_STATS * position)89 static void reset_fpf_position(TWO_PASS *p, const FIRSTPASS_STATS *position) {
90 p->stats_in = position;
91 }
92
93 // Read frame stats at an offset from the current position.
read_frame_stats(const TWO_PASS * p,int offset)94 static const FIRSTPASS_STATS *read_frame_stats(const TWO_PASS *p, int offset) {
95 if ((offset >= 0 && p->stats_in + offset >= p->stats_in_end) ||
96 (offset < 0 && p->stats_in + offset < p->stats_in_start)) {
97 return NULL;
98 }
99
100 return &p->stats_in[offset];
101 }
102
input_stats(TWO_PASS * p,FIRSTPASS_STATS * fps)103 static int input_stats(TWO_PASS *p, FIRSTPASS_STATS *fps) {
104 if (p->stats_in >= p->stats_in_end) return EOF;
105
106 *fps = *p->stats_in;
107 ++p->stats_in;
108 return 1;
109 }
110
output_stats(FIRSTPASS_STATS * stats)111 static void output_stats(FIRSTPASS_STATS *stats) {
112 (void)stats;
113 // TEMP debug code
114 #if OUTPUT_FPF
115 {
116 FILE *fpfile;
117 fpfile = fopen("firstpass.stt", "a");
118
119 fprintf(fpfile,
120 "%12.0lf %12.4lf %12.2lf %12.2lf %12.2lf %12.0lf %12.4lf %12.4lf"
121 "%12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf"
122 "%12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.0lf %12.4lf %12.0lf"
123 "%12.4lf"
124 "\n",
125 stats->frame, stats->weight, stats->intra_error, stats->coded_error,
126 stats->sr_coded_error, stats->frame_noise_energy, stats->pcnt_inter,
127 stats->pcnt_motion, stats->pcnt_second_ref, stats->pcnt_neutral,
128 stats->pcnt_intra_low, stats->pcnt_intra_high,
129 stats->intra_skip_pct, stats->intra_smooth_pct,
130 stats->inactive_zone_rows, stats->inactive_zone_cols, stats->MVr,
131 stats->mvr_abs, stats->MVc, stats->mvc_abs, stats->MVrv,
132 stats->MVcv, stats->mv_in_out_count, stats->count, stats->duration);
133 fclose(fpfile);
134 }
135 #endif
136 }
137
zero_stats(FIRSTPASS_STATS * section)138 static void zero_stats(FIRSTPASS_STATS *section) {
139 section->frame = 0.0;
140 section->weight = 0.0;
141 section->intra_error = 0.0;
142 section->coded_error = 0.0;
143 section->sr_coded_error = 0.0;
144 section->frame_noise_energy = 0.0;
145 section->pcnt_inter = 0.0;
146 section->pcnt_motion = 0.0;
147 section->pcnt_second_ref = 0.0;
148 section->pcnt_neutral = 0.0;
149 section->intra_skip_pct = 0.0;
150 section->intra_smooth_pct = 0.0;
151 section->pcnt_intra_low = 0.0;
152 section->pcnt_intra_high = 0.0;
153 section->inactive_zone_rows = 0.0;
154 section->inactive_zone_cols = 0.0;
155 section->MVr = 0.0;
156 section->mvr_abs = 0.0;
157 section->MVc = 0.0;
158 section->mvc_abs = 0.0;
159 section->MVrv = 0.0;
160 section->MVcv = 0.0;
161 section->mv_in_out_count = 0.0;
162 section->count = 0.0;
163 section->duration = 1.0;
164 section->spatial_layer_id = 0;
165 }
166
accumulate_stats(FIRSTPASS_STATS * section,const FIRSTPASS_STATS * frame)167 static void accumulate_stats(FIRSTPASS_STATS *section,
168 const FIRSTPASS_STATS *frame) {
169 section->frame += frame->frame;
170 section->weight += frame->weight;
171 section->spatial_layer_id = frame->spatial_layer_id;
172 section->intra_error += frame->intra_error;
173 section->coded_error += frame->coded_error;
174 section->sr_coded_error += frame->sr_coded_error;
175 section->frame_noise_energy += frame->frame_noise_energy;
176 section->pcnt_inter += frame->pcnt_inter;
177 section->pcnt_motion += frame->pcnt_motion;
178 section->pcnt_second_ref += frame->pcnt_second_ref;
179 section->pcnt_neutral += frame->pcnt_neutral;
180 section->intra_skip_pct += frame->intra_skip_pct;
181 section->intra_smooth_pct += frame->intra_smooth_pct;
182 section->pcnt_intra_low += frame->pcnt_intra_low;
183 section->pcnt_intra_high += frame->pcnt_intra_high;
184 section->inactive_zone_rows += frame->inactive_zone_rows;
185 section->inactive_zone_cols += frame->inactive_zone_cols;
186 section->MVr += frame->MVr;
187 section->mvr_abs += frame->mvr_abs;
188 section->MVc += frame->MVc;
189 section->mvc_abs += frame->mvc_abs;
190 section->MVrv += frame->MVrv;
191 section->MVcv += frame->MVcv;
192 section->mv_in_out_count += frame->mv_in_out_count;
193 section->count += frame->count;
194 section->duration += frame->duration;
195 }
196
subtract_stats(FIRSTPASS_STATS * section,const FIRSTPASS_STATS * frame)197 static void subtract_stats(FIRSTPASS_STATS *section,
198 const FIRSTPASS_STATS *frame) {
199 section->frame -= frame->frame;
200 section->weight -= frame->weight;
201 section->intra_error -= frame->intra_error;
202 section->coded_error -= frame->coded_error;
203 section->sr_coded_error -= frame->sr_coded_error;
204 section->frame_noise_energy -= frame->frame_noise_energy;
205 section->pcnt_inter -= frame->pcnt_inter;
206 section->pcnt_motion -= frame->pcnt_motion;
207 section->pcnt_second_ref -= frame->pcnt_second_ref;
208 section->pcnt_neutral -= frame->pcnt_neutral;
209 section->intra_skip_pct -= frame->intra_skip_pct;
210 section->intra_smooth_pct -= frame->intra_smooth_pct;
211 section->pcnt_intra_low -= frame->pcnt_intra_low;
212 section->pcnt_intra_high -= frame->pcnt_intra_high;
213 section->inactive_zone_rows -= frame->inactive_zone_rows;
214 section->inactive_zone_cols -= frame->inactive_zone_cols;
215 section->MVr -= frame->MVr;
216 section->mvr_abs -= frame->mvr_abs;
217 section->MVc -= frame->MVc;
218 section->mvc_abs -= frame->mvc_abs;
219 section->MVrv -= frame->MVrv;
220 section->MVcv -= frame->MVcv;
221 section->mv_in_out_count -= frame->mv_in_out_count;
222 section->count -= frame->count;
223 section->duration -= frame->duration;
224 }
225
226 // Calculate an active area of the image that discounts formatting
227 // bars and partially discounts other 0 energy areas.
228 #define MIN_ACTIVE_AREA 0.5
229 #define MAX_ACTIVE_AREA 1.0
calculate_active_area(const FRAME_INFO * frame_info,const FIRSTPASS_STATS * this_frame)230 static double calculate_active_area(const FRAME_INFO *frame_info,
231 const FIRSTPASS_STATS *this_frame) {
232 double active_pct;
233
234 active_pct =
235 1.0 -
236 ((this_frame->intra_skip_pct / 2) +
237 ((this_frame->inactive_zone_rows * 2) / (double)frame_info->mb_rows));
238 return fclamp(active_pct, MIN_ACTIVE_AREA, MAX_ACTIVE_AREA);
239 }
240
241 // Get the average weighted error for the clip (or corpus)
get_distribution_av_err(VP9_COMP * cpi,TWO_PASS * const twopass)242 static double get_distribution_av_err(VP9_COMP *cpi, TWO_PASS *const twopass) {
243 const double av_weight =
244 twopass->total_stats.weight / twopass->total_stats.count;
245
246 if (cpi->oxcf.vbr_corpus_complexity)
247 return av_weight * twopass->mean_mod_score;
248 else
249 return (twopass->total_stats.coded_error * av_weight) /
250 twopass->total_stats.count;
251 }
252
253 #define ACT_AREA_CORRECTION 0.5
254 // Calculate a modified Error used in distributing bits between easier and
255 // harder frames.
calculate_mod_frame_score(const VP9_COMP * cpi,const VP9EncoderConfig * oxcf,const FIRSTPASS_STATS * this_frame,const double av_err)256 static double calculate_mod_frame_score(const VP9_COMP *cpi,
257 const VP9EncoderConfig *oxcf,
258 const FIRSTPASS_STATS *this_frame,
259 const double av_err) {
260 double modified_score =
261 av_err * pow(this_frame->coded_error * this_frame->weight /
262 DOUBLE_DIVIDE_CHECK(av_err),
263 oxcf->two_pass_vbrbias / 100.0);
264
265 // Correction for active area. Frames with a reduced active area
266 // (eg due to formatting bars) have a higher error per mb for the
267 // remaining active MBs. The correction here assumes that coding
268 // 0.5N blocks of complexity 2X is a little easier than coding N
269 // blocks of complexity X.
270 modified_score *= pow(calculate_active_area(&cpi->frame_info, this_frame),
271 ACT_AREA_CORRECTION);
272
273 return modified_score;
274 }
275
calc_norm_frame_score(const VP9EncoderConfig * oxcf,const FRAME_INFO * frame_info,const FIRSTPASS_STATS * this_frame,double mean_mod_score,double av_err)276 static double calc_norm_frame_score(const VP9EncoderConfig *oxcf,
277 const FRAME_INFO *frame_info,
278 const FIRSTPASS_STATS *this_frame,
279 double mean_mod_score, double av_err) {
280 double modified_score =
281 av_err * pow(this_frame->coded_error * this_frame->weight /
282 DOUBLE_DIVIDE_CHECK(av_err),
283 oxcf->two_pass_vbrbias / 100.0);
284
285 const double min_score = (double)(oxcf->two_pass_vbrmin_section) / 100.0;
286 const double max_score = (double)(oxcf->two_pass_vbrmax_section) / 100.0;
287
288 // Correction for active area. Frames with a reduced active area
289 // (eg due to formatting bars) have a higher error per mb for the
290 // remaining active MBs. The correction here assumes that coding
291 // 0.5N blocks of complexity 2X is a little easier than coding N
292 // blocks of complexity X.
293 modified_score *=
294 pow(calculate_active_area(frame_info, this_frame), ACT_AREA_CORRECTION);
295
296 // Normalize to a midpoint score.
297 modified_score /= DOUBLE_DIVIDE_CHECK(mean_mod_score);
298 return fclamp(modified_score, min_score, max_score);
299 }
300
calculate_norm_frame_score(const VP9_COMP * cpi,const TWO_PASS * twopass,const VP9EncoderConfig * oxcf,const FIRSTPASS_STATS * this_frame,const double av_err)301 static double calculate_norm_frame_score(const VP9_COMP *cpi,
302 const TWO_PASS *twopass,
303 const VP9EncoderConfig *oxcf,
304 const FIRSTPASS_STATS *this_frame,
305 const double av_err) {
306 return calc_norm_frame_score(oxcf, &cpi->frame_info, this_frame,
307 twopass->mean_mod_score, av_err);
308 }
309
310 // This function returns the maximum target rate per frame.
frame_max_bits(const RATE_CONTROL * rc,const VP9EncoderConfig * oxcf)311 static int frame_max_bits(const RATE_CONTROL *rc,
312 const VP9EncoderConfig *oxcf) {
313 int64_t max_bits = ((int64_t)rc->avg_frame_bandwidth *
314 (int64_t)oxcf->two_pass_vbrmax_section) /
315 100;
316 if (max_bits < 0)
317 max_bits = 0;
318 else if (max_bits > rc->max_frame_bandwidth)
319 max_bits = rc->max_frame_bandwidth;
320
321 return (int)max_bits;
322 }
323
vp9_init_first_pass(VP9_COMP * cpi)324 void vp9_init_first_pass(VP9_COMP *cpi) {
325 zero_stats(&cpi->twopass.total_stats);
326 }
327
vp9_end_first_pass(VP9_COMP * cpi)328 void vp9_end_first_pass(VP9_COMP *cpi) {
329 output_stats(&cpi->twopass.total_stats);
330 cpi->twopass.first_pass_done = 1;
331 vpx_free(cpi->twopass.fp_mb_float_stats);
332 cpi->twopass.fp_mb_float_stats = NULL;
333 }
334
get_block_variance_fn(BLOCK_SIZE bsize)335 static vpx_variance_fn_t get_block_variance_fn(BLOCK_SIZE bsize) {
336 switch (bsize) {
337 case BLOCK_8X8: return vpx_mse8x8;
338 case BLOCK_16X8: return vpx_mse16x8;
339 case BLOCK_8X16: return vpx_mse8x16;
340 default: return vpx_mse16x16;
341 }
342 }
343
get_prediction_error(BLOCK_SIZE bsize,const struct buf_2d * src,const struct buf_2d * ref)344 static unsigned int get_prediction_error(BLOCK_SIZE bsize,
345 const struct buf_2d *src,
346 const struct buf_2d *ref) {
347 unsigned int sse;
348 const vpx_variance_fn_t fn = get_block_variance_fn(bsize);
349 fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
350 return sse;
351 }
352
353 #if CONFIG_VP9_HIGHBITDEPTH
highbd_get_block_variance_fn(BLOCK_SIZE bsize,int bd)354 static vpx_variance_fn_t highbd_get_block_variance_fn(BLOCK_SIZE bsize,
355 int bd) {
356 switch (bd) {
357 default:
358 switch (bsize) {
359 case BLOCK_8X8: return vpx_highbd_8_mse8x8;
360 case BLOCK_16X8: return vpx_highbd_8_mse16x8;
361 case BLOCK_8X16: return vpx_highbd_8_mse8x16;
362 default: return vpx_highbd_8_mse16x16;
363 }
364 break;
365 case 10:
366 switch (bsize) {
367 case BLOCK_8X8: return vpx_highbd_10_mse8x8;
368 case BLOCK_16X8: return vpx_highbd_10_mse16x8;
369 case BLOCK_8X16: return vpx_highbd_10_mse8x16;
370 default: return vpx_highbd_10_mse16x16;
371 }
372 break;
373 case 12:
374 switch (bsize) {
375 case BLOCK_8X8: return vpx_highbd_12_mse8x8;
376 case BLOCK_16X8: return vpx_highbd_12_mse16x8;
377 case BLOCK_8X16: return vpx_highbd_12_mse8x16;
378 default: return vpx_highbd_12_mse16x16;
379 }
380 break;
381 }
382 }
383
highbd_get_prediction_error(BLOCK_SIZE bsize,const struct buf_2d * src,const struct buf_2d * ref,int bd)384 static unsigned int highbd_get_prediction_error(BLOCK_SIZE bsize,
385 const struct buf_2d *src,
386 const struct buf_2d *ref,
387 int bd) {
388 unsigned int sse;
389 const vpx_variance_fn_t fn = highbd_get_block_variance_fn(bsize, bd);
390 fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
391 return sse;
392 }
393 #endif // CONFIG_VP9_HIGHBITDEPTH
394
395 // Refine the motion search range according to the frame dimension
396 // for first pass test.
get_search_range(const VP9_COMP * cpi)397 static int get_search_range(const VP9_COMP *cpi) {
398 int sr = 0;
399 const int dim = VPXMIN(cpi->initial_width, cpi->initial_height);
400
401 while ((dim << sr) < MAX_FULL_PEL_VAL) ++sr;
402 return sr;
403 }
404
405 // Reduce limits to keep the motion search within MV_MAX of ref_mv. Not doing
406 // this can be problematic for big videos (8K) and may cause assert failure
407 // (or memory violation) in mv_cost. Limits are only modified if they would
408 // be non-empty. Returns 1 if limits are non-empty.
intersect_limits_with_mv_max(MvLimits * mv_limits,const MV * ref_mv)409 static int intersect_limits_with_mv_max(MvLimits *mv_limits, const MV *ref_mv) {
410 const int row_min =
411 VPXMAX(mv_limits->row_min, (ref_mv->row + 7 - MV_MAX) >> 3);
412 const int row_max =
413 VPXMIN(mv_limits->row_max, (ref_mv->row - 1 + MV_MAX) >> 3);
414 const int col_min =
415 VPXMAX(mv_limits->col_min, (ref_mv->col + 7 - MV_MAX) >> 3);
416 const int col_max =
417 VPXMIN(mv_limits->col_max, (ref_mv->col - 1 + MV_MAX) >> 3);
418 if (row_min > row_max || col_min > col_max) {
419 return 0;
420 }
421 mv_limits->row_min = row_min;
422 mv_limits->row_max = row_max;
423 mv_limits->col_min = col_min;
424 mv_limits->col_max = col_max;
425 return 1;
426 }
427
first_pass_motion_search(VP9_COMP * cpi,MACROBLOCK * x,const MV * ref_mv,MV * best_mv,int * best_motion_err)428 static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
429 const MV *ref_mv, MV *best_mv,
430 int *best_motion_err) {
431 MACROBLOCKD *const xd = &x->e_mbd;
432 MV tmp_mv = { 0, 0 };
433 MV ref_mv_full = { ref_mv->row >> 3, ref_mv->col >> 3 };
434 int num00, tmp_err, n;
435 const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
436 vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
437 const int new_mv_mode_penalty = NEW_MV_MODE_PENALTY;
438
439 int step_param = 3;
440 int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
441 const int sr = get_search_range(cpi);
442 const MvLimits tmp_mv_limits = x->mv_limits;
443 step_param += sr;
444 further_steps -= sr;
445
446 if (!intersect_limits_with_mv_max(&x->mv_limits, ref_mv)) {
447 return;
448 }
449
450 // Override the default variance function to use MSE.
451 v_fn_ptr.vf = get_block_variance_fn(bsize);
452 #if CONFIG_VP9_HIGHBITDEPTH
453 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
454 v_fn_ptr.vf = highbd_get_block_variance_fn(bsize, xd->bd);
455 }
456 #endif // CONFIG_VP9_HIGHBITDEPTH
457
458 // Center the initial step/diamond search on best mv.
459 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
460 step_param, x->sadperbit16, &num00,
461 &v_fn_ptr, ref_mv);
462 if (tmp_err < INT_MAX)
463 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
464 if (tmp_err < INT_MAX - new_mv_mode_penalty) tmp_err += new_mv_mode_penalty;
465
466 if (tmp_err < *best_motion_err) {
467 *best_motion_err = tmp_err;
468 *best_mv = tmp_mv;
469 }
470
471 // Carry out further step/diamond searches as necessary.
472 n = num00;
473 num00 = 0;
474
475 while (n < further_steps) {
476 ++n;
477
478 if (num00) {
479 --num00;
480 } else {
481 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
482 step_param + n, x->sadperbit16, &num00,
483 &v_fn_ptr, ref_mv);
484 if (tmp_err < INT_MAX)
485 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
486 if (tmp_err < INT_MAX - new_mv_mode_penalty)
487 tmp_err += new_mv_mode_penalty;
488
489 if (tmp_err < *best_motion_err) {
490 *best_motion_err = tmp_err;
491 *best_mv = tmp_mv;
492 }
493 }
494 }
495 x->mv_limits = tmp_mv_limits;
496 }
497
get_bsize(const VP9_COMMON * cm,int mb_row,int mb_col)498 static BLOCK_SIZE get_bsize(const VP9_COMMON *cm, int mb_row, int mb_col) {
499 if (2 * mb_col + 1 < cm->mi_cols) {
500 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_16X16 : BLOCK_16X8;
501 } else {
502 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_8X16 : BLOCK_8X8;
503 }
504 }
505
find_fp_qindex(vpx_bit_depth_t bit_depth)506 static int find_fp_qindex(vpx_bit_depth_t bit_depth) {
507 int i;
508
509 for (i = 0; i < QINDEX_RANGE; ++i)
510 if (vp9_convert_qindex_to_q(i, bit_depth) >= FIRST_PASS_Q) break;
511
512 if (i == QINDEX_RANGE) i--;
513
514 return i;
515 }
516
set_first_pass_params(VP9_COMP * cpi)517 static void set_first_pass_params(VP9_COMP *cpi) {
518 VP9_COMMON *const cm = &cpi->common;
519 if (!cpi->refresh_alt_ref_frame &&
520 (cm->current_video_frame == 0 || (cpi->frame_flags & FRAMEFLAGS_KEY))) {
521 cm->frame_type = KEY_FRAME;
522 } else {
523 cm->frame_type = INTER_FRAME;
524 }
525 // Do not use periodic key frames.
526 cpi->rc.frames_to_key = INT_MAX;
527 }
528
529 // Scale an sse threshold to account for 8/10/12 bit.
scale_sse_threshold(VP9_COMMON * cm,int thresh)530 static int scale_sse_threshold(VP9_COMMON *cm, int thresh) {
531 int ret_val = thresh;
532 #if CONFIG_VP9_HIGHBITDEPTH
533 if (cm->use_highbitdepth) {
534 switch (cm->bit_depth) {
535 case VPX_BITS_8: ret_val = thresh; break;
536 case VPX_BITS_10: ret_val = thresh << 4; break;
537 default:
538 assert(cm->bit_depth == VPX_BITS_12);
539 ret_val = thresh << 8;
540 break;
541 }
542 }
543 #else
544 (void)cm;
545 #endif // CONFIG_VP9_HIGHBITDEPTH
546 return ret_val;
547 }
548
549 // This threshold is used to track blocks where to all intents and purposes
550 // the intra prediction error 0. Though the metric we test against
551 // is technically a sse we are mainly interested in blocks where all the pixels
552 // in the 8 bit domain have an error of <= 1 (where error = sse) so a
553 // linear scaling for 10 and 12 bit gives similar results.
554 #define UL_INTRA_THRESH 50
get_ul_intra_threshold(VP9_COMMON * cm)555 static int get_ul_intra_threshold(VP9_COMMON *cm) {
556 int ret_val = UL_INTRA_THRESH;
557 #if CONFIG_VP9_HIGHBITDEPTH
558 if (cm->use_highbitdepth) {
559 switch (cm->bit_depth) {
560 case VPX_BITS_8: ret_val = UL_INTRA_THRESH; break;
561 case VPX_BITS_10: ret_val = UL_INTRA_THRESH << 2; break;
562 default:
563 assert(cm->bit_depth == VPX_BITS_12);
564 ret_val = UL_INTRA_THRESH << 4;
565 break;
566 }
567 }
568 #else
569 (void)cm;
570 #endif // CONFIG_VP9_HIGHBITDEPTH
571 return ret_val;
572 }
573
574 #define SMOOTH_INTRA_THRESH 4000
get_smooth_intra_threshold(VP9_COMMON * cm)575 static int get_smooth_intra_threshold(VP9_COMMON *cm) {
576 int ret_val = SMOOTH_INTRA_THRESH;
577 #if CONFIG_VP9_HIGHBITDEPTH
578 if (cm->use_highbitdepth) {
579 switch (cm->bit_depth) {
580 case VPX_BITS_8: ret_val = SMOOTH_INTRA_THRESH; break;
581 case VPX_BITS_10: ret_val = SMOOTH_INTRA_THRESH << 4; break;
582 default:
583 assert(cm->bit_depth == VPX_BITS_12);
584 ret_val = SMOOTH_INTRA_THRESH << 8;
585 break;
586 }
587 }
588 #else
589 (void)cm;
590 #endif // CONFIG_VP9_HIGHBITDEPTH
591 return ret_val;
592 }
593
594 #define FP_DN_THRESH 8
595 #define FP_MAX_DN_THRESH 24
596 #define KERNEL_SIZE 3
597
598 // Baseline Kernal weights for first pass noise metric
599 static uint8_t fp_dn_kernal_3[KERNEL_SIZE * KERNEL_SIZE] = { 1, 2, 1, 2, 4,
600 2, 1, 2, 1 };
601
602 // Estimate noise at a single point based on the impace of a spatial kernal
603 // on the point value
fp_estimate_point_noise(uint8_t * src_ptr,const int stride)604 static int fp_estimate_point_noise(uint8_t *src_ptr, const int stride) {
605 int sum_weight = 0;
606 int sum_val = 0;
607 int i, j;
608 int max_diff = 0;
609 int diff;
610 int dn_diff;
611 uint8_t *tmp_ptr;
612 uint8_t *kernal_ptr;
613 uint8_t dn_val;
614 uint8_t centre_val = *src_ptr;
615
616 kernal_ptr = fp_dn_kernal_3;
617
618 // Apply the kernal
619 tmp_ptr = src_ptr - stride - 1;
620 for (i = 0; i < KERNEL_SIZE; ++i) {
621 for (j = 0; j < KERNEL_SIZE; ++j) {
622 diff = abs((int)centre_val - (int)tmp_ptr[j]);
623 max_diff = VPXMAX(max_diff, diff);
624 if (diff <= FP_DN_THRESH) {
625 sum_weight += *kernal_ptr;
626 sum_val += (int)tmp_ptr[j] * (int)*kernal_ptr;
627 }
628 ++kernal_ptr;
629 }
630 tmp_ptr += stride;
631 }
632
633 if (max_diff < FP_MAX_DN_THRESH)
634 // Update the source value with the new filtered value
635 dn_val = (sum_val + (sum_weight >> 1)) / sum_weight;
636 else
637 dn_val = *src_ptr;
638
639 // return the noise energy as the square of the difference between the
640 // denoised and raw value.
641 dn_diff = (int)*src_ptr - (int)dn_val;
642 return dn_diff * dn_diff;
643 }
644 #if CONFIG_VP9_HIGHBITDEPTH
fp_highbd_estimate_point_noise(uint8_t * src_ptr,const int stride)645 static int fp_highbd_estimate_point_noise(uint8_t *src_ptr, const int stride) {
646 int sum_weight = 0;
647 int sum_val = 0;
648 int i, j;
649 int max_diff = 0;
650 int diff;
651 int dn_diff;
652 uint8_t *tmp_ptr;
653 uint16_t *tmp_ptr16;
654 uint8_t *kernal_ptr;
655 uint16_t dn_val;
656 uint16_t centre_val = *CONVERT_TO_SHORTPTR(src_ptr);
657
658 kernal_ptr = fp_dn_kernal_3;
659
660 // Apply the kernal
661 tmp_ptr = src_ptr - stride - 1;
662 for (i = 0; i < KERNEL_SIZE; ++i) {
663 tmp_ptr16 = CONVERT_TO_SHORTPTR(tmp_ptr);
664 for (j = 0; j < KERNEL_SIZE; ++j) {
665 diff = abs((int)centre_val - (int)tmp_ptr16[j]);
666 max_diff = VPXMAX(max_diff, diff);
667 if (diff <= FP_DN_THRESH) {
668 sum_weight += *kernal_ptr;
669 sum_val += (int)tmp_ptr16[j] * (int)*kernal_ptr;
670 }
671 ++kernal_ptr;
672 }
673 tmp_ptr += stride;
674 }
675
676 if (max_diff < FP_MAX_DN_THRESH)
677 // Update the source value with the new filtered value
678 dn_val = (sum_val + (sum_weight >> 1)) / sum_weight;
679 else
680 dn_val = *CONVERT_TO_SHORTPTR(src_ptr);
681
682 // return the noise energy as the square of the difference between the
683 // denoised and raw value.
684 dn_diff = (int)(*CONVERT_TO_SHORTPTR(src_ptr)) - (int)dn_val;
685 return dn_diff * dn_diff;
686 }
687 #endif
688
689 // Estimate noise for a block.
fp_estimate_block_noise(MACROBLOCK * x,BLOCK_SIZE bsize)690 static int fp_estimate_block_noise(MACROBLOCK *x, BLOCK_SIZE bsize) {
691 #if CONFIG_VP9_HIGHBITDEPTH
692 MACROBLOCKD *xd = &x->e_mbd;
693 #endif
694 uint8_t *src_ptr = &x->plane[0].src.buf[0];
695 const int width = num_4x4_blocks_wide_lookup[bsize] * 4;
696 const int height = num_4x4_blocks_high_lookup[bsize] * 4;
697 int w, h;
698 int stride = x->plane[0].src.stride;
699 int block_noise = 0;
700
701 // Sampled points to reduce cost overhead.
702 for (h = 0; h < height; h += 2) {
703 for (w = 0; w < width; w += 2) {
704 #if CONFIG_VP9_HIGHBITDEPTH
705 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
706 block_noise += fp_highbd_estimate_point_noise(src_ptr, stride);
707 else
708 block_noise += fp_estimate_point_noise(src_ptr, stride);
709 #else
710 block_noise += fp_estimate_point_noise(src_ptr, stride);
711 #endif
712 ++src_ptr;
713 }
714 src_ptr += (stride - width);
715 }
716 return block_noise << 2; // Scale << 2 to account for sampling.
717 }
718
719 // This function is called to test the functionality of row based
720 // multi-threading in unit tests for bit-exactness
accumulate_floating_point_stats(VP9_COMP * cpi,TileDataEnc * first_tile_col)721 static void accumulate_floating_point_stats(VP9_COMP *cpi,
722 TileDataEnc *first_tile_col) {
723 VP9_COMMON *const cm = &cpi->common;
724 int mb_row, mb_col;
725 first_tile_col->fp_data.intra_factor = 0;
726 first_tile_col->fp_data.brightness_factor = 0;
727 first_tile_col->fp_data.neutral_count = 0;
728 for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
729 for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
730 const int mb_index = mb_row * cm->mb_cols + mb_col;
731 first_tile_col->fp_data.intra_factor +=
732 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_intra_factor;
733 first_tile_col->fp_data.brightness_factor +=
734 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_brightness_factor;
735 first_tile_col->fp_data.neutral_count +=
736 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_neutral_count;
737 }
738 }
739 }
740
first_pass_stat_calc(VP9_COMP * cpi,FIRSTPASS_STATS * fps,FIRSTPASS_DATA * fp_acc_data)741 static void first_pass_stat_calc(VP9_COMP *cpi, FIRSTPASS_STATS *fps,
742 FIRSTPASS_DATA *fp_acc_data) {
743 VP9_COMMON *const cm = &cpi->common;
744 // The minimum error here insures some bit allocation to frames even
745 // in static regions. The allocation per MB declines for larger formats
746 // where the typical "real" energy per MB also falls.
747 // Initial estimate here uses sqrt(mbs) to define the min_err, where the
748 // number of mbs is proportional to the image area.
749 const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE) ? cpi->initial_mbs
750 : cpi->common.MBs;
751 const double min_err = 200 * sqrt(num_mbs);
752
753 // Clamp the image start to rows/2. This number of rows is discarded top
754 // and bottom as dead data so rows / 2 means the frame is blank.
755 if ((fp_acc_data->image_data_start_row > cm->mb_rows / 2) ||
756 (fp_acc_data->image_data_start_row == INVALID_ROW)) {
757 fp_acc_data->image_data_start_row = cm->mb_rows / 2;
758 }
759 // Exclude any image dead zone
760 if (fp_acc_data->image_data_start_row > 0) {
761 fp_acc_data->intra_skip_count =
762 VPXMAX(0, fp_acc_data->intra_skip_count -
763 (fp_acc_data->image_data_start_row * cm->mb_cols * 2));
764 }
765
766 fp_acc_data->intra_factor = fp_acc_data->intra_factor / (double)num_mbs;
767 fp_acc_data->brightness_factor =
768 fp_acc_data->brightness_factor / (double)num_mbs;
769 fps->weight = fp_acc_data->intra_factor * fp_acc_data->brightness_factor;
770
771 fps->frame = cm->current_video_frame;
772 fps->spatial_layer_id = cpi->svc.spatial_layer_id;
773
774 fps->coded_error =
775 ((double)(fp_acc_data->coded_error >> 8) + min_err) / num_mbs;
776 fps->sr_coded_error =
777 ((double)(fp_acc_data->sr_coded_error >> 8) + min_err) / num_mbs;
778 fps->intra_error =
779 ((double)(fp_acc_data->intra_error >> 8) + min_err) / num_mbs;
780
781 fps->frame_noise_energy =
782 (double)(fp_acc_data->frame_noise_energy) / (double)num_mbs;
783 fps->count = 1.0;
784 fps->pcnt_inter = (double)(fp_acc_data->intercount) / num_mbs;
785 fps->pcnt_second_ref = (double)(fp_acc_data->second_ref_count) / num_mbs;
786 fps->pcnt_neutral = (double)(fp_acc_data->neutral_count) / num_mbs;
787 fps->pcnt_intra_low = (double)(fp_acc_data->intra_count_low) / num_mbs;
788 fps->pcnt_intra_high = (double)(fp_acc_data->intra_count_high) / num_mbs;
789 fps->intra_skip_pct = (double)(fp_acc_data->intra_skip_count) / num_mbs;
790 fps->intra_smooth_pct = (double)(fp_acc_data->intra_smooth_count) / num_mbs;
791 fps->inactive_zone_rows = (double)(fp_acc_data->image_data_start_row);
792 // Currently set to 0 as most issues relate to letter boxing.
793 fps->inactive_zone_cols = (double)0;
794
795 if (fp_acc_data->mvcount > 0) {
796 fps->MVr = (double)(fp_acc_data->sum_mvr) / fp_acc_data->mvcount;
797 fps->mvr_abs = (double)(fp_acc_data->sum_mvr_abs) / fp_acc_data->mvcount;
798 fps->MVc = (double)(fp_acc_data->sum_mvc) / fp_acc_data->mvcount;
799 fps->mvc_abs = (double)(fp_acc_data->sum_mvc_abs) / fp_acc_data->mvcount;
800 fps->MVrv = ((double)(fp_acc_data->sum_mvrs) -
801 ((double)(fp_acc_data->sum_mvr) * (fp_acc_data->sum_mvr) /
802 fp_acc_data->mvcount)) /
803 fp_acc_data->mvcount;
804 fps->MVcv = ((double)(fp_acc_data->sum_mvcs) -
805 ((double)(fp_acc_data->sum_mvc) * (fp_acc_data->sum_mvc) /
806 fp_acc_data->mvcount)) /
807 fp_acc_data->mvcount;
808 fps->mv_in_out_count =
809 (double)(fp_acc_data->sum_in_vectors) / (fp_acc_data->mvcount * 2);
810 fps->pcnt_motion = (double)(fp_acc_data->mvcount) / num_mbs;
811 } else {
812 fps->MVr = 0.0;
813 fps->mvr_abs = 0.0;
814 fps->MVc = 0.0;
815 fps->mvc_abs = 0.0;
816 fps->MVrv = 0.0;
817 fps->MVcv = 0.0;
818 fps->mv_in_out_count = 0.0;
819 fps->pcnt_motion = 0.0;
820 }
821 }
822
accumulate_fp_mb_row_stat(TileDataEnc * this_tile,FIRSTPASS_DATA * fp_acc_data)823 static void accumulate_fp_mb_row_stat(TileDataEnc *this_tile,
824 FIRSTPASS_DATA *fp_acc_data) {
825 this_tile->fp_data.intra_factor += fp_acc_data->intra_factor;
826 this_tile->fp_data.brightness_factor += fp_acc_data->brightness_factor;
827 this_tile->fp_data.coded_error += fp_acc_data->coded_error;
828 this_tile->fp_data.sr_coded_error += fp_acc_data->sr_coded_error;
829 this_tile->fp_data.frame_noise_energy += fp_acc_data->frame_noise_energy;
830 this_tile->fp_data.intra_error += fp_acc_data->intra_error;
831 this_tile->fp_data.intercount += fp_acc_data->intercount;
832 this_tile->fp_data.second_ref_count += fp_acc_data->second_ref_count;
833 this_tile->fp_data.neutral_count += fp_acc_data->neutral_count;
834 this_tile->fp_data.intra_count_low += fp_acc_data->intra_count_low;
835 this_tile->fp_data.intra_count_high += fp_acc_data->intra_count_high;
836 this_tile->fp_data.intra_skip_count += fp_acc_data->intra_skip_count;
837 this_tile->fp_data.mvcount += fp_acc_data->mvcount;
838 this_tile->fp_data.sum_mvr += fp_acc_data->sum_mvr;
839 this_tile->fp_data.sum_mvr_abs += fp_acc_data->sum_mvr_abs;
840 this_tile->fp_data.sum_mvc += fp_acc_data->sum_mvc;
841 this_tile->fp_data.sum_mvc_abs += fp_acc_data->sum_mvc_abs;
842 this_tile->fp_data.sum_mvrs += fp_acc_data->sum_mvrs;
843 this_tile->fp_data.sum_mvcs += fp_acc_data->sum_mvcs;
844 this_tile->fp_data.sum_in_vectors += fp_acc_data->sum_in_vectors;
845 this_tile->fp_data.intra_smooth_count += fp_acc_data->intra_smooth_count;
846 this_tile->fp_data.image_data_start_row =
847 VPXMIN(this_tile->fp_data.image_data_start_row,
848 fp_acc_data->image_data_start_row) == INVALID_ROW
849 ? VPXMAX(this_tile->fp_data.image_data_start_row,
850 fp_acc_data->image_data_start_row)
851 : VPXMIN(this_tile->fp_data.image_data_start_row,
852 fp_acc_data->image_data_start_row);
853 }
854
855 #if CONFIG_RATE_CTRL
store_fp_motion_vector(VP9_COMP * cpi,const MV * mv,const int mb_row,const int mb_col,MV_REFERENCE_FRAME frame_type,const int mv_idx)856 static void store_fp_motion_vector(VP9_COMP *cpi, const MV *mv,
857 const int mb_row, const int mb_col,
858 MV_REFERENCE_FRAME frame_type,
859 const int mv_idx) {
860 VP9_COMMON *const cm = &cpi->common;
861 const int mb_index = mb_row * cm->mb_cols + mb_col;
862 MOTION_VECTOR_INFO *this_motion_vector_info =
863 &cpi->fp_motion_vector_info[mb_index];
864 this_motion_vector_info->ref_frame[mv_idx] = frame_type;
865 if (frame_type != INTRA_FRAME) {
866 this_motion_vector_info->mv[mv_idx].as_mv = *mv;
867 }
868 }
869 #endif // CONFIG_RATE_CTRL
870
871 #define NZ_MOTION_PENALTY 128
872 #define INTRA_MODE_PENALTY 1024
vp9_first_pass_encode_tile_mb_row(VP9_COMP * cpi,ThreadData * td,FIRSTPASS_DATA * fp_acc_data,TileDataEnc * tile_data,MV * best_ref_mv,int mb_row)873 void vp9_first_pass_encode_tile_mb_row(VP9_COMP *cpi, ThreadData *td,
874 FIRSTPASS_DATA *fp_acc_data,
875 TileDataEnc *tile_data, MV *best_ref_mv,
876 int mb_row) {
877 int mb_col;
878 MACROBLOCK *const x = &td->mb;
879 VP9_COMMON *const cm = &cpi->common;
880 MACROBLOCKD *const xd = &x->e_mbd;
881 TileInfo tile = tile_data->tile_info;
882 const int mb_col_start = ROUND_POWER_OF_TWO(tile.mi_col_start, 1);
883 const int mb_col_end = ROUND_POWER_OF_TWO(tile.mi_col_end, 1);
884 struct macroblock_plane *const p = x->plane;
885 struct macroblockd_plane *const pd = xd->plane;
886 const PICK_MODE_CONTEXT *ctx = &td->pc_root->none;
887 int i, c;
888 int num_mb_cols = get_num_cols(tile_data->tile_info, 1);
889
890 int recon_yoffset, recon_uvoffset;
891 const int intrapenalty = INTRA_MODE_PENALTY;
892 const MV zero_mv = { 0, 0 };
893 int recon_y_stride, recon_uv_stride, uv_mb_height;
894
895 YV12_BUFFER_CONFIG *const lst_yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
896 YV12_BUFFER_CONFIG *gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
897 YV12_BUFFER_CONFIG *const new_yv12 = get_frame_new_buffer(cm);
898 const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;
899
900 MODE_INFO mi_above, mi_left;
901
902 double mb_intra_factor;
903 double mb_brightness_factor;
904 double mb_neutral_count;
905 int scaled_low_intra_thresh = scale_sse_threshold(cm, LOW_I_THRESH);
906
907 // First pass code requires valid last and new frame buffers.
908 assert(new_yv12 != NULL);
909 assert(frame_is_intra_only(cm) || (lst_yv12 != NULL));
910
911 xd->mi = cm->mi_grid_visible + xd->mi_stride * (mb_row << 1) + mb_col_start;
912 xd->mi[0] = cm->mi + xd->mi_stride * (mb_row << 1) + mb_col_start;
913
914 for (i = 0; i < MAX_MB_PLANE; ++i) {
915 p[i].coeff = ctx->coeff_pbuf[i][1];
916 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
917 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
918 p[i].eobs = ctx->eobs_pbuf[i][1];
919 }
920
921 recon_y_stride = new_yv12->y_stride;
922 recon_uv_stride = new_yv12->uv_stride;
923 uv_mb_height = 16 >> (new_yv12->y_height > new_yv12->uv_height);
924
925 // Reset above block coeffs.
926 recon_yoffset = (mb_row * recon_y_stride * 16) + mb_col_start * 16;
927 recon_uvoffset =
928 (mb_row * recon_uv_stride * uv_mb_height) + mb_col_start * uv_mb_height;
929
930 // Set up limit values for motion vectors to prevent them extending
931 // outside the UMV borders.
932 x->mv_limits.row_min = -((mb_row * 16) + BORDER_MV_PIXELS_B16);
933 x->mv_limits.row_max =
934 ((cm->mb_rows - 1 - mb_row) * 16) + BORDER_MV_PIXELS_B16;
935
936 for (mb_col = mb_col_start, c = 0; mb_col < mb_col_end; ++mb_col, c++) {
937 int this_error;
938 int this_intra_error;
939 const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
940 const BLOCK_SIZE bsize = get_bsize(cm, mb_row, mb_col);
941 double log_intra;
942 int level_sample;
943 const int mb_index = mb_row * cm->mb_cols + mb_col;
944
945 (*(cpi->row_mt_sync_read_ptr))(&tile_data->row_mt_sync, mb_row, c);
946
947 // Adjust to the next column of MBs.
948 x->plane[0].src.buf = cpi->Source->y_buffer +
949 mb_row * 16 * x->plane[0].src.stride + mb_col * 16;
950 x->plane[1].src.buf = cpi->Source->u_buffer +
951 mb_row * uv_mb_height * x->plane[1].src.stride +
952 mb_col * uv_mb_height;
953 x->plane[2].src.buf = cpi->Source->v_buffer +
954 mb_row * uv_mb_height * x->plane[1].src.stride +
955 mb_col * uv_mb_height;
956
957 vpx_clear_system_state();
958
959 xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
960 xd->plane[1].dst.buf = new_yv12->u_buffer + recon_uvoffset;
961 xd->plane[2].dst.buf = new_yv12->v_buffer + recon_uvoffset;
962 xd->mi[0]->sb_type = bsize;
963 xd->mi[0]->ref_frame[0] = INTRA_FRAME;
964 set_mi_row_col(xd, &tile, mb_row << 1, num_8x8_blocks_high_lookup[bsize],
965 mb_col << 1, num_8x8_blocks_wide_lookup[bsize], cm->mi_rows,
966 cm->mi_cols);
967 // Are edges available for intra prediction?
968 // Since the firstpass does not populate the mi_grid_visible,
969 // above_mi/left_mi must be overwritten with a nonzero value when edges
970 // are available. Required by vp9_predict_intra_block().
971 xd->above_mi = (mb_row != 0) ? &mi_above : NULL;
972 xd->left_mi = ((mb_col << 1) > tile.mi_col_start) ? &mi_left : NULL;
973
974 // Do intra 16x16 prediction.
975 x->skip_encode = 0;
976 x->fp_src_pred = 0;
977 // Do intra prediction based on source pixels for tile boundaries
978 if (mb_col == mb_col_start && mb_col != 0) {
979 xd->left_mi = &mi_left;
980 x->fp_src_pred = 1;
981 }
982 xd->mi[0]->mode = DC_PRED;
983 xd->mi[0]->tx_size =
984 use_dc_pred ? (bsize >= BLOCK_16X16 ? TX_16X16 : TX_8X8) : TX_4X4;
985 // Fix - zero the 16x16 block first. This ensures correct this_error for
986 // block sizes smaller than 16x16.
987 vp9_zero_array(x->plane[0].src_diff, 256);
988 vp9_encode_intra_block_plane(x, bsize, 0, 0);
989 this_error = vpx_get_mb_ss(x->plane[0].src_diff);
990 this_intra_error = this_error;
991
992 // Keep a record of blocks that have very low intra error residual
993 // (i.e. are in effect completely flat and untextured in the intra
994 // domain). In natural videos this is uncommon, but it is much more
995 // common in animations, graphics and screen content, so may be used
996 // as a signal to detect these types of content.
997 if (this_error < get_ul_intra_threshold(cm)) {
998 ++(fp_acc_data->intra_skip_count);
999 } else if ((mb_col > 0) &&
1000 (fp_acc_data->image_data_start_row == INVALID_ROW)) {
1001 fp_acc_data->image_data_start_row = mb_row;
1002 }
1003
1004 // Blocks that are mainly smooth in the intra domain.
1005 // Some special accounting for CQ but also these are better for testing
1006 // noise levels.
1007 if (this_error < get_smooth_intra_threshold(cm)) {
1008 ++(fp_acc_data->intra_smooth_count);
1009 }
1010
1011 // Special case noise measurement for first frame.
1012 if (cm->current_video_frame == 0) {
1013 if (this_intra_error < scale_sse_threshold(cm, LOW_I_THRESH)) {
1014 fp_acc_data->frame_noise_energy += fp_estimate_block_noise(x, bsize);
1015 } else {
1016 fp_acc_data->frame_noise_energy += (int64_t)SECTION_NOISE_DEF;
1017 }
1018 }
1019
1020 #if CONFIG_VP9_HIGHBITDEPTH
1021 if (cm->use_highbitdepth) {
1022 switch (cm->bit_depth) {
1023 case VPX_BITS_8: break;
1024 case VPX_BITS_10: this_error >>= 4; break;
1025 default:
1026 assert(cm->bit_depth == VPX_BITS_12);
1027 this_error >>= 8;
1028 break;
1029 }
1030 }
1031 #endif // CONFIG_VP9_HIGHBITDEPTH
1032
1033 vpx_clear_system_state();
1034 log_intra = log(this_error + 1.0);
1035 if (log_intra < 10.0) {
1036 mb_intra_factor = 1.0 + ((10.0 - log_intra) * 0.05);
1037 fp_acc_data->intra_factor += mb_intra_factor;
1038 if (cpi->row_mt_bit_exact)
1039 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_intra_factor =
1040 mb_intra_factor;
1041 } else {
1042 fp_acc_data->intra_factor += 1.0;
1043 if (cpi->row_mt_bit_exact)
1044 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_intra_factor = 1.0;
1045 }
1046
1047 #if CONFIG_VP9_HIGHBITDEPTH
1048 if (cm->use_highbitdepth)
1049 level_sample = CONVERT_TO_SHORTPTR(x->plane[0].src.buf)[0];
1050 else
1051 level_sample = x->plane[0].src.buf[0];
1052 #else
1053 level_sample = x->plane[0].src.buf[0];
1054 #endif
1055 if ((level_sample < DARK_THRESH) && (log_intra < 9.0)) {
1056 mb_brightness_factor = 1.0 + (0.01 * (DARK_THRESH - level_sample));
1057 fp_acc_data->brightness_factor += mb_brightness_factor;
1058 if (cpi->row_mt_bit_exact)
1059 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_brightness_factor =
1060 mb_brightness_factor;
1061 } else {
1062 fp_acc_data->brightness_factor += 1.0;
1063 if (cpi->row_mt_bit_exact)
1064 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_brightness_factor =
1065 1.0;
1066 }
1067
1068 // Intrapenalty below deals with situations where the intra and inter
1069 // error scores are very low (e.g. a plain black frame).
1070 // We do not have special cases in first pass for 0,0 and nearest etc so
1071 // all inter modes carry an overhead cost estimate for the mv.
1072 // When the error score is very low this causes us to pick all or lots of
1073 // INTRA modes and throw lots of key frames.
1074 // This penalty adds a cost matching that of a 0,0 mv to the intra case.
1075 this_error += intrapenalty;
1076
1077 // Accumulate the intra error.
1078 fp_acc_data->intra_error += (int64_t)this_error;
1079
1080 // Set up limit values for motion vectors to prevent them extending
1081 // outside the UMV borders.
1082 x->mv_limits.col_min = -((mb_col * 16) + BORDER_MV_PIXELS_B16);
1083 x->mv_limits.col_max =
1084 ((cm->mb_cols - 1 - mb_col) * 16) + BORDER_MV_PIXELS_B16;
1085
1086 // Other than for intra-only frame do a motion search.
1087 if (!frame_is_intra_only(cm)) {
1088 int tmp_err, motion_error, this_motion_error, raw_motion_error;
1089 // Assume 0,0 motion with no mv overhead.
1090 MV mv = { 0, 0 }, tmp_mv = { 0, 0 };
1091 struct buf_2d unscaled_last_source_buf_2d;
1092 vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
1093
1094 #if CONFIG_RATE_CTRL
1095 if (cpi->oxcf.use_simple_encode_api) {
1096 // Store zero mv as default
1097 store_fp_motion_vector(cpi, &mv, mb_row, mb_col, LAST_FRAME, 0);
1098 }
1099 #endif // CONFIG_RAGE_CTRL
1100
1101 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
1102 #if CONFIG_VP9_HIGHBITDEPTH
1103 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1104 motion_error = highbd_get_prediction_error(
1105 bsize, &x->plane[0].src, &xd->plane[0].pre[0], xd->bd);
1106 this_motion_error = highbd_get_prediction_error(
1107 bsize, &x->plane[0].src, &xd->plane[0].pre[0], 8);
1108 } else {
1109 motion_error =
1110 get_prediction_error(bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
1111 this_motion_error = motion_error;
1112 }
1113 #else
1114 motion_error =
1115 get_prediction_error(bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
1116 this_motion_error = motion_error;
1117 #endif // CONFIG_VP9_HIGHBITDEPTH
1118
1119 // Compute the motion error of the 0,0 motion using the last source
1120 // frame as the reference. Skip the further motion search on
1121 // reconstructed frame if this error is very small.
1122 unscaled_last_source_buf_2d.buf =
1123 cpi->unscaled_last_source->y_buffer + recon_yoffset;
1124 unscaled_last_source_buf_2d.stride = cpi->unscaled_last_source->y_stride;
1125 #if CONFIG_VP9_HIGHBITDEPTH
1126 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1127 raw_motion_error = highbd_get_prediction_error(
1128 bsize, &x->plane[0].src, &unscaled_last_source_buf_2d, xd->bd);
1129 } else {
1130 raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
1131 &unscaled_last_source_buf_2d);
1132 }
1133 #else
1134 raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
1135 &unscaled_last_source_buf_2d);
1136 #endif // CONFIG_VP9_HIGHBITDEPTH
1137
1138 if (raw_motion_error > NZ_MOTION_PENALTY) {
1139 // Test last reference frame using the previous best mv as the
1140 // starting point (best reference) for the search.
1141 first_pass_motion_search(cpi, x, best_ref_mv, &mv, &motion_error);
1142
1143 v_fn_ptr.vf = get_block_variance_fn(bsize);
1144 #if CONFIG_VP9_HIGHBITDEPTH
1145 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1146 v_fn_ptr.vf = highbd_get_block_variance_fn(bsize, 8);
1147 }
1148 #endif // CONFIG_VP9_HIGHBITDEPTH
1149 this_motion_error =
1150 vp9_get_mvpred_var(x, &mv, best_ref_mv, &v_fn_ptr, 0);
1151
1152 // If the current best reference mv is not centered on 0,0 then do a
1153 // 0,0 based search as well.
1154 if (!is_zero_mv(best_ref_mv)) {
1155 tmp_err = INT_MAX;
1156 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv, &tmp_err);
1157
1158 if (tmp_err < motion_error) {
1159 motion_error = tmp_err;
1160 mv = tmp_mv;
1161 this_motion_error =
1162 vp9_get_mvpred_var(x, &tmp_mv, &zero_mv, &v_fn_ptr, 0);
1163 }
1164 }
1165 #if CONFIG_RATE_CTRL
1166 if (cpi->oxcf.use_simple_encode_api) {
1167 store_fp_motion_vector(cpi, &mv, mb_row, mb_col, LAST_FRAME, 0);
1168 }
1169 #endif // CONFIG_RAGE_CTRL
1170
1171 // Search in an older reference frame.
1172 if ((cm->current_video_frame > 1) && gld_yv12 != NULL) {
1173 // Assume 0,0 motion with no mv overhead.
1174 int gf_motion_error;
1175
1176 xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
1177 #if CONFIG_VP9_HIGHBITDEPTH
1178 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1179 gf_motion_error = highbd_get_prediction_error(
1180 bsize, &x->plane[0].src, &xd->plane[0].pre[0], xd->bd);
1181 } else {
1182 gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
1183 &xd->plane[0].pre[0]);
1184 }
1185 #else
1186 gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
1187 &xd->plane[0].pre[0]);
1188 #endif // CONFIG_VP9_HIGHBITDEPTH
1189
1190 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv, &gf_motion_error);
1191 #if CONFIG_RATE_CTRL
1192 if (cpi->oxcf.use_simple_encode_api) {
1193 store_fp_motion_vector(cpi, &tmp_mv, mb_row, mb_col, GOLDEN_FRAME,
1194 1);
1195 }
1196 #endif // CONFIG_RAGE_CTRL
1197
1198 if (gf_motion_error < motion_error && gf_motion_error < this_error)
1199 ++(fp_acc_data->second_ref_count);
1200
1201 // Reset to last frame as reference buffer.
1202 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
1203 xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
1204 xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
1205
1206 // In accumulating a score for the older reference frame take the
1207 // best of the motion predicted score and the intra coded error
1208 // (just as will be done for) accumulation of "coded_error" for
1209 // the last frame.
1210 if (gf_motion_error < this_error)
1211 fp_acc_data->sr_coded_error += gf_motion_error;
1212 else
1213 fp_acc_data->sr_coded_error += this_error;
1214 } else {
1215 fp_acc_data->sr_coded_error += motion_error;
1216 }
1217 } else {
1218 fp_acc_data->sr_coded_error += motion_error;
1219 }
1220
1221 // Start by assuming that intra mode is best.
1222 best_ref_mv->row = 0;
1223 best_ref_mv->col = 0;
1224
1225 if (motion_error <= this_error) {
1226 vpx_clear_system_state();
1227
1228 // Keep a count of cases where the inter and intra were very close
1229 // and very low. This helps with scene cut detection for example in
1230 // cropped clips with black bars at the sides or top and bottom.
1231 if (((this_error - intrapenalty) * 9 <= motion_error * 10) &&
1232 (this_error < (2 * intrapenalty))) {
1233 fp_acc_data->neutral_count += 1.0;
1234 if (cpi->row_mt_bit_exact)
1235 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_neutral_count =
1236 1.0;
1237 // Also track cases where the intra is not much worse than the inter
1238 // and use this in limiting the GF/arf group length.
1239 } else if ((this_error > NCOUNT_INTRA_THRESH) &&
1240 (this_error < (NCOUNT_INTRA_FACTOR * motion_error))) {
1241 mb_neutral_count =
1242 (double)motion_error / DOUBLE_DIVIDE_CHECK((double)this_error);
1243 fp_acc_data->neutral_count += mb_neutral_count;
1244 if (cpi->row_mt_bit_exact)
1245 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_neutral_count =
1246 mb_neutral_count;
1247 }
1248
1249 mv.row *= 8;
1250 mv.col *= 8;
1251 this_error = motion_error;
1252 xd->mi[0]->mode = NEWMV;
1253 xd->mi[0]->mv[0].as_mv = mv;
1254 xd->mi[0]->tx_size = TX_4X4;
1255 xd->mi[0]->ref_frame[0] = LAST_FRAME;
1256 xd->mi[0]->ref_frame[1] = NONE;
1257 vp9_build_inter_predictors_sby(xd, mb_row << 1, mb_col << 1, bsize);
1258 vp9_encode_sby_pass1(x, bsize);
1259 fp_acc_data->sum_mvr += mv.row;
1260 fp_acc_data->sum_mvr_abs += abs(mv.row);
1261 fp_acc_data->sum_mvc += mv.col;
1262 fp_acc_data->sum_mvc_abs += abs(mv.col);
1263 fp_acc_data->sum_mvrs += mv.row * mv.row;
1264 fp_acc_data->sum_mvcs += mv.col * mv.col;
1265 ++(fp_acc_data->intercount);
1266
1267 *best_ref_mv = mv;
1268
1269 if (!is_zero_mv(&mv)) {
1270 ++(fp_acc_data->mvcount);
1271
1272 // Does the row vector point inwards or outwards?
1273 if (mb_row < cm->mb_rows / 2) {
1274 if (mv.row > 0)
1275 --(fp_acc_data->sum_in_vectors);
1276 else if (mv.row < 0)
1277 ++(fp_acc_data->sum_in_vectors);
1278 } else if (mb_row > cm->mb_rows / 2) {
1279 if (mv.row > 0)
1280 ++(fp_acc_data->sum_in_vectors);
1281 else if (mv.row < 0)
1282 --(fp_acc_data->sum_in_vectors);
1283 }
1284
1285 // Does the col vector point inwards or outwards?
1286 if (mb_col < cm->mb_cols / 2) {
1287 if (mv.col > 0)
1288 --(fp_acc_data->sum_in_vectors);
1289 else if (mv.col < 0)
1290 ++(fp_acc_data->sum_in_vectors);
1291 } else if (mb_col > cm->mb_cols / 2) {
1292 if (mv.col > 0)
1293 ++(fp_acc_data->sum_in_vectors);
1294 else if (mv.col < 0)
1295 --(fp_acc_data->sum_in_vectors);
1296 }
1297 }
1298 if (this_intra_error < scaled_low_intra_thresh) {
1299 fp_acc_data->frame_noise_energy += fp_estimate_block_noise(x, bsize);
1300 } else {
1301 fp_acc_data->frame_noise_energy += (int64_t)SECTION_NOISE_DEF;
1302 }
1303 } else { // Intra < inter error
1304 if (this_intra_error < scaled_low_intra_thresh) {
1305 fp_acc_data->frame_noise_energy += fp_estimate_block_noise(x, bsize);
1306 if (this_motion_error < scaled_low_intra_thresh) {
1307 fp_acc_data->intra_count_low += 1.0;
1308 } else {
1309 fp_acc_data->intra_count_high += 1.0;
1310 }
1311 } else {
1312 fp_acc_data->frame_noise_energy += (int64_t)SECTION_NOISE_DEF;
1313 fp_acc_data->intra_count_high += 1.0;
1314 }
1315 }
1316 } else {
1317 fp_acc_data->sr_coded_error += (int64_t)this_error;
1318 #if CONFIG_RATE_CTRL
1319 if (cpi->oxcf.use_simple_encode_api) {
1320 store_fp_motion_vector(cpi, NULL, mb_row, mb_col, INTRA_FRAME, 0);
1321 }
1322 #endif // CONFIG_RAGE_CTRL
1323 }
1324 fp_acc_data->coded_error += (int64_t)this_error;
1325
1326 recon_yoffset += 16;
1327 recon_uvoffset += uv_mb_height;
1328
1329 // Accumulate row level stats to the corresponding tile stats
1330 if (cpi->row_mt && mb_col == mb_col_end - 1)
1331 accumulate_fp_mb_row_stat(tile_data, fp_acc_data);
1332
1333 (*(cpi->row_mt_sync_write_ptr))(&tile_data->row_mt_sync, mb_row, c,
1334 num_mb_cols);
1335 }
1336 vpx_clear_system_state();
1337 }
1338
first_pass_encode(VP9_COMP * cpi,FIRSTPASS_DATA * fp_acc_data)1339 static void first_pass_encode(VP9_COMP *cpi, FIRSTPASS_DATA *fp_acc_data) {
1340 VP9_COMMON *const cm = &cpi->common;
1341 int mb_row;
1342 TileDataEnc tile_data;
1343 TileInfo *tile = &tile_data.tile_info;
1344 MV zero_mv = { 0, 0 };
1345 MV best_ref_mv;
1346 // Tiling is ignored in the first pass.
1347 vp9_tile_init(tile, cm, 0, 0);
1348
1349 #if CONFIG_RATE_CTRL
1350 if (cpi->oxcf.use_simple_encode_api) {
1351 fp_motion_vector_info_reset(cpi->frame_info.frame_width,
1352 cpi->frame_info.frame_height,
1353 cpi->fp_motion_vector_info);
1354 }
1355 #endif
1356
1357 for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
1358 best_ref_mv = zero_mv;
1359 vp9_first_pass_encode_tile_mb_row(cpi, &cpi->td, fp_acc_data, &tile_data,
1360 &best_ref_mv, mb_row);
1361 }
1362 }
1363
vp9_first_pass(VP9_COMP * cpi,const struct lookahead_entry * source)1364 void vp9_first_pass(VP9_COMP *cpi, const struct lookahead_entry *source) {
1365 MACROBLOCK *const x = &cpi->td.mb;
1366 VP9_COMMON *const cm = &cpi->common;
1367 MACROBLOCKD *const xd = &x->e_mbd;
1368 TWO_PASS *twopass = &cpi->twopass;
1369
1370 YV12_BUFFER_CONFIG *const lst_yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
1371 YV12_BUFFER_CONFIG *gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
1372 YV12_BUFFER_CONFIG *const new_yv12 = get_frame_new_buffer(cm);
1373 const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;
1374
1375 BufferPool *const pool = cm->buffer_pool;
1376
1377 FIRSTPASS_DATA fp_temp_data;
1378 FIRSTPASS_DATA *fp_acc_data = &fp_temp_data;
1379
1380 vpx_clear_system_state();
1381 vp9_zero(fp_temp_data);
1382 fp_acc_data->image_data_start_row = INVALID_ROW;
1383
1384 // First pass code requires valid last and new frame buffers.
1385 assert(new_yv12 != NULL);
1386 assert(frame_is_intra_only(cm) || (lst_yv12 != NULL));
1387
1388 set_first_pass_params(cpi);
1389 vp9_set_quantizer(cpi, find_fp_qindex(cm->bit_depth));
1390
1391 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
1392
1393 vp9_setup_src_planes(x, cpi->Source, 0, 0);
1394 vp9_setup_dst_planes(xd->plane, new_yv12, 0, 0);
1395
1396 if (!frame_is_intra_only(cm)) {
1397 vp9_setup_pre_planes(xd, 0, first_ref_buf, 0, 0, NULL);
1398 }
1399
1400 xd->mi = cm->mi_grid_visible;
1401 xd->mi[0] = cm->mi;
1402
1403 vp9_frame_init_quantizer(cpi);
1404
1405 x->skip_recode = 0;
1406
1407 vp9_init_mv_probs(cm);
1408 vp9_initialize_rd_consts(cpi);
1409
1410 cm->log2_tile_rows = 0;
1411
1412 if (cpi->row_mt_bit_exact && cpi->twopass.fp_mb_float_stats == NULL)
1413 CHECK_MEM_ERROR(
1414 cm, cpi->twopass.fp_mb_float_stats,
1415 vpx_calloc(cm->MBs * sizeof(*cpi->twopass.fp_mb_float_stats), 1));
1416
1417 {
1418 FIRSTPASS_STATS fps;
1419 TileDataEnc *first_tile_col;
1420 if (!cpi->row_mt) {
1421 cm->log2_tile_cols = 0;
1422 cpi->row_mt_sync_read_ptr = vp9_row_mt_sync_read_dummy;
1423 cpi->row_mt_sync_write_ptr = vp9_row_mt_sync_write_dummy;
1424 first_pass_encode(cpi, fp_acc_data);
1425 first_pass_stat_calc(cpi, &fps, fp_acc_data);
1426 } else {
1427 cpi->row_mt_sync_read_ptr = vp9_row_mt_sync_read;
1428 cpi->row_mt_sync_write_ptr = vp9_row_mt_sync_write;
1429 if (cpi->row_mt_bit_exact) {
1430 cm->log2_tile_cols = 0;
1431 vp9_zero_array(cpi->twopass.fp_mb_float_stats, cm->MBs);
1432 }
1433 vp9_encode_fp_row_mt(cpi);
1434 first_tile_col = &cpi->tile_data[0];
1435 if (cpi->row_mt_bit_exact)
1436 accumulate_floating_point_stats(cpi, first_tile_col);
1437 first_pass_stat_calc(cpi, &fps, &(first_tile_col->fp_data));
1438 }
1439
1440 // Dont allow a value of 0 for duration.
1441 // (Section duration is also defaulted to minimum of 1.0).
1442 fps.duration = VPXMAX(1.0, (double)(source->ts_end - source->ts_start));
1443
1444 // Don't want to do output stats with a stack variable!
1445 twopass->this_frame_stats = fps;
1446 output_stats(&twopass->this_frame_stats);
1447 accumulate_stats(&twopass->total_stats, &fps);
1448 }
1449
1450 // Copy the previous Last Frame back into gf and and arf buffers if
1451 // the prediction is good enough... but also don't allow it to lag too far.
1452 if ((twopass->sr_update_lag > 3) ||
1453 ((cm->current_video_frame > 0) &&
1454 (twopass->this_frame_stats.pcnt_inter > 0.20) &&
1455 ((twopass->this_frame_stats.intra_error /
1456 DOUBLE_DIVIDE_CHECK(twopass->this_frame_stats.coded_error)) > 2.0))) {
1457 if (gld_yv12 != NULL) {
1458 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
1459 cm->ref_frame_map[cpi->lst_fb_idx]);
1460 }
1461 twopass->sr_update_lag = 1;
1462 } else {
1463 ++twopass->sr_update_lag;
1464 }
1465
1466 vpx_extend_frame_borders(new_yv12);
1467
1468 // The frame we just compressed now becomes the last frame.
1469 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->lst_fb_idx],
1470 cm->new_fb_idx);
1471
1472 // Special case for the first frame. Copy into the GF buffer as a second
1473 // reference.
1474 if (cm->current_video_frame == 0 && cpi->gld_fb_idx != INVALID_IDX) {
1475 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
1476 cm->ref_frame_map[cpi->lst_fb_idx]);
1477 }
1478
1479 // Use this to see what the first pass reconstruction looks like.
1480 if (0) {
1481 char filename[512];
1482 FILE *recon_file;
1483 snprintf(filename, sizeof(filename), "enc%04d.yuv",
1484 (int)cm->current_video_frame);
1485
1486 if (cm->current_video_frame == 0)
1487 recon_file = fopen(filename, "wb");
1488 else
1489 recon_file = fopen(filename, "ab");
1490
1491 (void)fwrite(lst_yv12->buffer_alloc, lst_yv12->frame_size, 1, recon_file);
1492 fclose(recon_file);
1493 }
1494
1495 // In the first pass, every frame is considered as a show frame.
1496 update_frame_indexes(cm, /*show_frame=*/1);
1497 if (cpi->use_svc) vp9_inc_frame_in_layer(cpi);
1498 }
1499
1500 static const double q_pow_term[(QINDEX_RANGE >> 5) + 1] = { 0.65, 0.70, 0.75,
1501 0.85, 0.90, 0.90,
1502 0.90, 1.00, 1.25 };
1503
calc_correction_factor(double err_per_mb,double err_divisor,int q)1504 static double calc_correction_factor(double err_per_mb, double err_divisor,
1505 int q) {
1506 const double error_term = err_per_mb / DOUBLE_DIVIDE_CHECK(err_divisor);
1507 const int index = q >> 5;
1508 double power_term;
1509
1510 assert((index >= 0) && (index < (QINDEX_RANGE >> 5)));
1511
1512 // Adjustment based on quantizer to the power term.
1513 power_term =
1514 q_pow_term[index] +
1515 (((q_pow_term[index + 1] - q_pow_term[index]) * (q % 32)) / 32.0);
1516
1517 // Calculate correction factor.
1518 if (power_term < 1.0) assert(error_term >= 0.0);
1519
1520 return fclamp(pow(error_term, power_term), 0.05, 5.0);
1521 }
1522
wq_err_divisor(VP9_COMP * cpi)1523 static double wq_err_divisor(VP9_COMP *cpi) {
1524 const VP9_COMMON *const cm = &cpi->common;
1525 unsigned int screen_area = (cm->width * cm->height);
1526
1527 // Use a different error per mb factor for calculating boost for
1528 // different formats.
1529 if (screen_area <= 640 * 360) {
1530 return 115.0;
1531 } else if (screen_area < 1280 * 720) {
1532 return 125.0;
1533 } else if (screen_area <= 1920 * 1080) {
1534 return 130.0;
1535 } else if (screen_area < 3840 * 2160) {
1536 return 150.0;
1537 }
1538
1539 // Fall through to here only for 4K and above.
1540 return 200.0;
1541 }
1542
1543 #define NOISE_FACTOR_MIN 0.9
1544 #define NOISE_FACTOR_MAX 1.1
get_twopass_worst_quality(VP9_COMP * cpi,const double section_err,double inactive_zone,double section_noise,int section_target_bandwidth)1545 static int get_twopass_worst_quality(VP9_COMP *cpi, const double section_err,
1546 double inactive_zone, double section_noise,
1547 int section_target_bandwidth) {
1548 const RATE_CONTROL *const rc = &cpi->rc;
1549 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1550 TWO_PASS *const twopass = &cpi->twopass;
1551 double last_group_rate_err;
1552
1553 // Clamp the target rate to VBR min / max limts.
1554 const int target_rate =
1555 vp9_rc_clamp_pframe_target_size(cpi, section_target_bandwidth);
1556 double noise_factor = pow((section_noise / SECTION_NOISE_DEF), 0.5);
1557 noise_factor = fclamp(noise_factor, NOISE_FACTOR_MIN, NOISE_FACTOR_MAX);
1558 inactive_zone = fclamp(inactive_zone, 0.0, 1.0);
1559
1560 // TODO(jimbankoski): remove #if here or below when this has been
1561 // well tested.
1562 #if CONFIG_ALWAYS_ADJUST_BPM
1563 // based on recent history adjust expectations of bits per macroblock.
1564 last_group_rate_err =
1565 (double)twopass->rolling_arf_group_actual_bits /
1566 DOUBLE_DIVIDE_CHECK((double)twopass->rolling_arf_group_target_bits);
1567 last_group_rate_err = VPXMAX(0.25, VPXMIN(4.0, last_group_rate_err));
1568 twopass->bpm_factor *= (3.0 + last_group_rate_err) / 4.0;
1569 twopass->bpm_factor = VPXMAX(0.25, VPXMIN(4.0, twopass->bpm_factor));
1570 #endif
1571
1572 if (target_rate <= 0) {
1573 return rc->worst_quality; // Highest value allowed
1574 } else {
1575 const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
1576 ? cpi->initial_mbs
1577 : cpi->common.MBs;
1578 const double active_pct = VPXMAX(0.01, 1.0 - inactive_zone);
1579 const int active_mbs = (int)VPXMAX(1, (double)num_mbs * active_pct);
1580 const double av_err_per_mb = section_err / active_pct;
1581 const double speed_term = 1.0 + 0.04 * oxcf->speed;
1582 const int target_norm_bits_per_mb =
1583 (int)(((uint64_t)target_rate << BPER_MB_NORMBITS) / active_mbs);
1584 int q;
1585
1586 // TODO(jimbankoski): remove #if here or above when this has been
1587 // well tested.
1588 #if !CONFIG_ALWAYS_ADJUST_BPM
1589 // based on recent history adjust expectations of bits per macroblock.
1590 last_group_rate_err =
1591 (double)twopass->rolling_arf_group_actual_bits /
1592 DOUBLE_DIVIDE_CHECK((double)twopass->rolling_arf_group_target_bits);
1593 last_group_rate_err = VPXMAX(0.25, VPXMIN(4.0, last_group_rate_err));
1594 twopass->bpm_factor *= (3.0 + last_group_rate_err) / 4.0;
1595 twopass->bpm_factor = VPXMAX(0.25, VPXMIN(4.0, twopass->bpm_factor));
1596 #endif
1597
1598 // Try and pick a max Q that will be high enough to encode the
1599 // content at the given rate.
1600 for (q = rc->best_quality; q < rc->worst_quality; ++q) {
1601 const double factor =
1602 calc_correction_factor(av_err_per_mb, wq_err_divisor(cpi), q);
1603 const int bits_per_mb = vp9_rc_bits_per_mb(
1604 INTER_FRAME, q,
1605 factor * speed_term * cpi->twopass.bpm_factor * noise_factor,
1606 cpi->common.bit_depth);
1607 if (bits_per_mb <= target_norm_bits_per_mb) break;
1608 }
1609
1610 // Restriction on active max q for constrained quality mode.
1611 if (cpi->oxcf.rc_mode == VPX_CQ) q = VPXMAX(q, oxcf->cq_level);
1612 return q;
1613 }
1614 }
1615
setup_rf_level_maxq(VP9_COMP * cpi)1616 static void setup_rf_level_maxq(VP9_COMP *cpi) {
1617 int i;
1618 RATE_CONTROL *const rc = &cpi->rc;
1619 for (i = INTER_NORMAL; i < RATE_FACTOR_LEVELS; ++i) {
1620 int qdelta = vp9_frame_type_qdelta(cpi, i, rc->worst_quality);
1621 rc->rf_level_maxq[i] = VPXMAX(rc->worst_quality + qdelta, rc->best_quality);
1622 }
1623 }
1624
init_subsampling(VP9_COMP * cpi)1625 static void init_subsampling(VP9_COMP *cpi) {
1626 const VP9_COMMON *const cm = &cpi->common;
1627 RATE_CONTROL *const rc = &cpi->rc;
1628 const int w = cm->width;
1629 const int h = cm->height;
1630 int i;
1631
1632 for (i = 0; i < FRAME_SCALE_STEPS; ++i) {
1633 // Note: Frames with odd-sized dimensions may result from this scaling.
1634 rc->frame_width[i] = (w * 16) / frame_scale_factor[i];
1635 rc->frame_height[i] = (h * 16) / frame_scale_factor[i];
1636 }
1637
1638 setup_rf_level_maxq(cpi);
1639 }
1640
calculate_coded_size(VP9_COMP * cpi,int * scaled_frame_width,int * scaled_frame_height)1641 void calculate_coded_size(VP9_COMP *cpi, int *scaled_frame_width,
1642 int *scaled_frame_height) {
1643 RATE_CONTROL *const rc = &cpi->rc;
1644 *scaled_frame_width = rc->frame_width[rc->frame_size_selector];
1645 *scaled_frame_height = rc->frame_height[rc->frame_size_selector];
1646 }
1647
vp9_init_second_pass(VP9_COMP * cpi)1648 void vp9_init_second_pass(VP9_COMP *cpi) {
1649 VP9EncoderConfig *const oxcf = &cpi->oxcf;
1650 RATE_CONTROL *const rc = &cpi->rc;
1651 TWO_PASS *const twopass = &cpi->twopass;
1652 double frame_rate;
1653 FIRSTPASS_STATS *stats;
1654
1655 zero_stats(&twopass->total_stats);
1656 zero_stats(&twopass->total_left_stats);
1657
1658 if (!twopass->stats_in_end) return;
1659
1660 stats = &twopass->total_stats;
1661
1662 *stats = *twopass->stats_in_end;
1663 twopass->total_left_stats = *stats;
1664
1665 // Scan the first pass file and calculate a modified score for each
1666 // frame that is used to distribute bits. The modified score is assumed
1667 // to provide a linear basis for bit allocation. I.e a frame A with a score
1668 // that is double that of frame B will be allocated 2x as many bits.
1669 {
1670 double modified_score_total = 0.0;
1671 const FIRSTPASS_STATS *s = twopass->stats_in;
1672 double av_err;
1673
1674 if (oxcf->vbr_corpus_complexity) {
1675 twopass->mean_mod_score = (double)oxcf->vbr_corpus_complexity / 10.0;
1676 av_err = get_distribution_av_err(cpi, twopass);
1677 } else {
1678 av_err = get_distribution_av_err(cpi, twopass);
1679 // The first scan is unclamped and gives a raw average.
1680 while (s < twopass->stats_in_end) {
1681 modified_score_total += calculate_mod_frame_score(cpi, oxcf, s, av_err);
1682 ++s;
1683 }
1684
1685 // The average error from this first scan is used to define the midpoint
1686 // error for the rate distribution function.
1687 twopass->mean_mod_score =
1688 modified_score_total / DOUBLE_DIVIDE_CHECK(stats->count);
1689 }
1690
1691 // Second scan using clamps based on the previous cycle average.
1692 // This may modify the total and average somewhat but we dont bother with
1693 // further itterations.
1694 modified_score_total = 0.0;
1695 s = twopass->stats_in;
1696 while (s < twopass->stats_in_end) {
1697 modified_score_total +=
1698 calculate_norm_frame_score(cpi, twopass, oxcf, s, av_err);
1699 ++s;
1700 }
1701 twopass->normalized_score_left = modified_score_total;
1702
1703 // If using Corpus wide VBR mode then update the clip target bandwidth to
1704 // reflect how the clip compares to the rest of the corpus.
1705 if (oxcf->vbr_corpus_complexity) {
1706 oxcf->target_bandwidth =
1707 (int64_t)((double)oxcf->target_bandwidth *
1708 (twopass->normalized_score_left / stats->count));
1709 }
1710
1711 #if COMPLEXITY_STATS_OUTPUT
1712 {
1713 FILE *compstats;
1714 compstats = fopen("complexity_stats.stt", "a");
1715 fprintf(compstats, "%10.3lf\n",
1716 twopass->normalized_score_left / stats->count);
1717 fclose(compstats);
1718 }
1719 #endif
1720 }
1721
1722 frame_rate = 10000000.0 * stats->count / stats->duration;
1723 // Each frame can have a different duration, as the frame rate in the source
1724 // isn't guaranteed to be constant. The frame rate prior to the first frame
1725 // encoded in the second pass is a guess. However, the sum duration is not.
1726 // It is calculated based on the actual durations of all frames from the
1727 // first pass.
1728 vp9_new_framerate(cpi, frame_rate);
1729 twopass->bits_left =
1730 (int64_t)(stats->duration * oxcf->target_bandwidth / 10000000.0);
1731
1732 // This variable monitors how far behind the second ref update is lagging.
1733 twopass->sr_update_lag = 1;
1734
1735 // Reset the vbr bits off target counters
1736 rc->vbr_bits_off_target = 0;
1737 rc->vbr_bits_off_target_fast = 0;
1738 rc->rate_error_estimate = 0;
1739
1740 // Static sequence monitor variables.
1741 twopass->kf_zeromotion_pct = 100;
1742 twopass->last_kfgroup_zeromotion_pct = 100;
1743
1744 // Initialize bits per macro_block estimate correction factor.
1745 twopass->bpm_factor = 1.0;
1746 // Initialize actual and target bits counters for ARF groups so that
1747 // at the start we have a neutral bpm adjustment.
1748 twopass->rolling_arf_group_target_bits = 1;
1749 twopass->rolling_arf_group_actual_bits = 1;
1750
1751 if (oxcf->resize_mode != RESIZE_NONE) {
1752 init_subsampling(cpi);
1753 }
1754
1755 // Initialize the arnr strangth adjustment to 0
1756 twopass->arnr_strength_adjustment = 0;
1757 }
1758
1759 /* This function considers how the quality of prediction may be deteriorating
1760 * with distance. It compares the coded error for the last frame and the
1761 * second reference frame (usually two frames old) and also applies a factor
1762 * based on the extent of INTRA coding.
1763 *
1764 * The decay factor is then used to reduce the contribution of frames further
1765 * from the alt-ref or golden frame, to the bitrate boost calculation for that
1766 * alt-ref or golden frame.
1767 */
get_sr_decay_rate(const TWO_PASS * const twopass,const FIRSTPASS_STATS * frame)1768 static double get_sr_decay_rate(const TWO_PASS *const twopass,
1769 const FIRSTPASS_STATS *frame) {
1770 double sr_diff = (frame->sr_coded_error - frame->coded_error);
1771 double sr_decay = 1.0;
1772
1773 // Do nothing if the second ref to last frame error difference is
1774 // very small or even negative.
1775 if ((sr_diff > LOW_SR_DIFF_TRHESH)) {
1776 const double sr_diff_part =
1777 twopass->sr_diff_factor * ((sr_diff * 0.25) / frame->intra_error);
1778 double modified_pct_inter = frame->pcnt_inter;
1779 double modified_pcnt_intra;
1780
1781 if ((frame->coded_error > LOW_CODED_ERR_PER_MB) &&
1782 ((frame->intra_error / DOUBLE_DIVIDE_CHECK(frame->coded_error)) <
1783 (double)NCOUNT_FRAME_II_THRESH)) {
1784 modified_pct_inter =
1785 frame->pcnt_inter + frame->pcnt_intra_low - frame->pcnt_neutral;
1786 }
1787 modified_pcnt_intra = 100 * (1.0 - modified_pct_inter);
1788
1789 sr_decay = 1.0 - sr_diff_part - (INTRA_PART * modified_pcnt_intra);
1790 }
1791 return VPXMAX(sr_decay, twopass->sr_default_decay_limit);
1792 }
1793
1794 // This function gives an estimate of how badly we believe the prediction
1795 // quality is decaying from frame to frame.
get_zero_motion_factor(const TWO_PASS * const twopass,const FIRSTPASS_STATS * frame_stats)1796 static double get_zero_motion_factor(const TWO_PASS *const twopass,
1797 const FIRSTPASS_STATS *frame_stats) {
1798 const double zero_motion_pct =
1799 frame_stats->pcnt_inter - frame_stats->pcnt_motion;
1800 double sr_decay = get_sr_decay_rate(twopass, frame_stats);
1801 return VPXMIN(sr_decay, zero_motion_pct);
1802 }
1803
get_prediction_decay_rate(const TWO_PASS * const twopass,const FIRSTPASS_STATS * frame_stats)1804 static double get_prediction_decay_rate(const TWO_PASS *const twopass,
1805 const FIRSTPASS_STATS *frame_stats) {
1806 const double sr_decay_rate = get_sr_decay_rate(twopass, frame_stats);
1807 double zero_motion_factor =
1808 twopass->zm_factor * (frame_stats->pcnt_inter - frame_stats->pcnt_motion);
1809
1810 // Check that the zero motion factor is valid
1811 assert(zero_motion_factor >= 0.0 && zero_motion_factor <= 1.0);
1812
1813 return VPXMAX(zero_motion_factor,
1814 (sr_decay_rate + ((1.0 - sr_decay_rate) * zero_motion_factor)));
1815 }
1816
get_show_idx(const TWO_PASS * twopass)1817 static int get_show_idx(const TWO_PASS *twopass) {
1818 return (int)(twopass->stats_in - twopass->stats_in_start);
1819 }
1820 // Function to test for a condition where a complex transition is followed
1821 // by a static section. For example in slide shows where there is a fade
1822 // between slides. This is to help with more optimal kf and gf positioning.
check_transition_to_still(const FIRST_PASS_INFO * first_pass_info,int show_idx,int still_interval)1823 static int check_transition_to_still(const FIRST_PASS_INFO *first_pass_info,
1824 int show_idx, int still_interval) {
1825 int j;
1826 int num_frames = fps_get_num_frames(first_pass_info);
1827 if (show_idx + still_interval > num_frames) {
1828 return 0;
1829 }
1830
1831 // Look ahead a few frames to see if static condition persists...
1832 for (j = 0; j < still_interval; ++j) {
1833 const FIRSTPASS_STATS *stats =
1834 fps_get_frame_stats(first_pass_info, show_idx + j);
1835 if (stats->pcnt_inter - stats->pcnt_motion < 0.999) break;
1836 }
1837
1838 // Only if it does do we signal a transition to still.
1839 return j == still_interval;
1840 }
1841
1842 // This function detects a flash through the high relative pcnt_second_ref
1843 // score in the frame following a flash frame. The offset passed in should
1844 // reflect this.
detect_flash_from_frame_stats(const FIRSTPASS_STATS * frame_stats)1845 static int detect_flash_from_frame_stats(const FIRSTPASS_STATS *frame_stats) {
1846 // What we are looking for here is a situation where there is a
1847 // brief break in prediction (such as a flash) but subsequent frames
1848 // are reasonably well predicted by an earlier (pre flash) frame.
1849 // The recovery after a flash is indicated by a high pcnt_second_ref
1850 // useage or a second ref coded error notabley lower than the last
1851 // frame coded error.
1852 if (frame_stats == NULL) {
1853 return 0;
1854 }
1855 return (frame_stats->sr_coded_error < frame_stats->coded_error) ||
1856 ((frame_stats->pcnt_second_ref > frame_stats->pcnt_inter) &&
1857 (frame_stats->pcnt_second_ref >= 0.5));
1858 }
1859
detect_flash(const TWO_PASS * twopass,int offset)1860 static int detect_flash(const TWO_PASS *twopass, int offset) {
1861 const FIRSTPASS_STATS *const next_frame = read_frame_stats(twopass, offset);
1862 return detect_flash_from_frame_stats(next_frame);
1863 }
1864
1865 // Update the motion related elements to the GF arf boost calculation.
accumulate_frame_motion_stats(const FIRSTPASS_STATS * stats,double * mv_in_out,double * mv_in_out_accumulator,double * abs_mv_in_out_accumulator,double * mv_ratio_accumulator)1866 static void accumulate_frame_motion_stats(const FIRSTPASS_STATS *stats,
1867 double *mv_in_out,
1868 double *mv_in_out_accumulator,
1869 double *abs_mv_in_out_accumulator,
1870 double *mv_ratio_accumulator) {
1871 const double pct = stats->pcnt_motion;
1872
1873 // Accumulate Motion In/Out of frame stats.
1874 *mv_in_out = stats->mv_in_out_count * pct;
1875 *mv_in_out_accumulator += *mv_in_out;
1876 *abs_mv_in_out_accumulator += fabs(*mv_in_out);
1877
1878 // Accumulate a measure of how uniform (or conversely how random) the motion
1879 // field is (a ratio of abs(mv) / mv).
1880 if (pct > 0.05) {
1881 const double mvr_ratio =
1882 fabs(stats->mvr_abs) / DOUBLE_DIVIDE_CHECK(fabs(stats->MVr));
1883 const double mvc_ratio =
1884 fabs(stats->mvc_abs) / DOUBLE_DIVIDE_CHECK(fabs(stats->MVc));
1885
1886 *mv_ratio_accumulator +=
1887 pct * (mvr_ratio < stats->mvr_abs ? mvr_ratio : stats->mvr_abs);
1888 *mv_ratio_accumulator +=
1889 pct * (mvc_ratio < stats->mvc_abs ? mvc_ratio : stats->mvc_abs);
1890 }
1891 }
1892
calc_frame_boost(const FRAME_INFO * frame_info,const FIRSTPASS_STATS * this_frame,const TWO_PASS * const twopass,int avg_frame_qindex,double this_frame_mv_in_out)1893 static double calc_frame_boost(const FRAME_INFO *frame_info,
1894 const FIRSTPASS_STATS *this_frame,
1895 const TWO_PASS *const twopass,
1896 int avg_frame_qindex,
1897 double this_frame_mv_in_out) {
1898 double frame_boost;
1899 const double lq =
1900 vp9_convert_qindex_to_q(avg_frame_qindex, frame_info->bit_depth);
1901 const double boost_q_correction = VPXMIN((0.5 + (lq * 0.015)), 1.5);
1902 const double active_area = calculate_active_area(frame_info, this_frame);
1903
1904 // Frame booost is based on inter error.
1905 frame_boost = (twopass->err_per_mb * active_area) /
1906 DOUBLE_DIVIDE_CHECK(this_frame->coded_error);
1907
1908 // Small adjustment for cases where there is a zoom out
1909 if (this_frame_mv_in_out > 0.0)
1910 frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
1911
1912 // Q correction and scalling
1913 frame_boost = frame_boost * boost_q_correction;
1914
1915 return VPXMIN(frame_boost, twopass->gf_frame_max_boost * boost_q_correction);
1916 }
1917
calc_kf_frame_boost(VP9_COMP * cpi,const FIRSTPASS_STATS * this_frame,double * sr_accumulator,double this_frame_mv_in_out,double zm_factor)1918 static double calc_kf_frame_boost(VP9_COMP *cpi,
1919 const FIRSTPASS_STATS *this_frame,
1920 double *sr_accumulator,
1921 double this_frame_mv_in_out,
1922 double zm_factor) {
1923 TWO_PASS *const twopass = &cpi->twopass;
1924 double frame_boost;
1925 const double lq = vp9_convert_qindex_to_q(
1926 cpi->rc.avg_frame_qindex[INTER_FRAME], cpi->common.bit_depth);
1927 const double boost_q_correction = VPXMIN((0.50 + (lq * 0.015)), 2.00);
1928 const double active_area =
1929 calculate_active_area(&cpi->frame_info, this_frame);
1930 double max_boost;
1931
1932 // Frame booost is based on inter error.
1933 frame_boost = (twopass->kf_err_per_mb * active_area) /
1934 DOUBLE_DIVIDE_CHECK(this_frame->coded_error + *sr_accumulator);
1935
1936 // Update the accumulator for second ref error difference.
1937 // This is intended to give an indication of how much the coded error is
1938 // increasing over time.
1939 *sr_accumulator += (this_frame->sr_coded_error - this_frame->coded_error);
1940 *sr_accumulator = VPXMAX(0.0, *sr_accumulator);
1941
1942 // Small adjustment for cases where there is a zoom out
1943 if (this_frame_mv_in_out > 0.0)
1944 frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
1945
1946 // Q correction and scaling
1947 // The 40.0 value here is an experimentally derived baseline minimum.
1948 // This value is in line with the minimum per frame boost in the alt_ref
1949 // boost calculation.
1950 frame_boost =
1951 (frame_boost + twopass->kf_frame_min_boost) * boost_q_correction;
1952
1953 // Maximum allowed boost this frame. May be different for first vs subsequent
1954 // key frames.
1955 max_boost = (cpi->common.current_video_frame == 0)
1956 ? twopass->kf_frame_max_boost_first
1957 : twopass->kf_frame_max_boost_subs;
1958 max_boost *= zm_factor * boost_q_correction;
1959
1960 return VPXMIN(frame_boost, max_boost);
1961 }
1962
compute_arf_boost(const FRAME_INFO * frame_info,TWO_PASS * const twopass,int arf_show_idx,int f_frames,int b_frames,int avg_frame_qindex)1963 static int compute_arf_boost(const FRAME_INFO *frame_info,
1964 TWO_PASS *const twopass, int arf_show_idx,
1965 int f_frames, int b_frames, int avg_frame_qindex) {
1966 const FIRST_PASS_INFO *first_pass_info = &twopass->first_pass_info;
1967 int i;
1968 double boost_score = 0.0;
1969 double mv_ratio_accumulator = 0.0;
1970 double decay_accumulator = 1.0;
1971 double this_frame_mv_in_out = 0.0;
1972 double mv_in_out_accumulator = 0.0;
1973 double abs_mv_in_out_accumulator = 0.0;
1974 int arf_boost;
1975 int flash_detected = 0;
1976
1977 // Search forward from the proposed arf/next gf position.
1978 for (i = 0; i < f_frames; ++i) {
1979 const FIRSTPASS_STATS *this_frame =
1980 fps_get_frame_stats(first_pass_info, arf_show_idx + i);
1981 const FIRSTPASS_STATS *next_frame =
1982 fps_get_frame_stats(first_pass_info, arf_show_idx + i + 1);
1983 if (this_frame == NULL) break;
1984
1985 // Update the motion related elements to the boost calculation.
1986 accumulate_frame_motion_stats(
1987 this_frame, &this_frame_mv_in_out, &mv_in_out_accumulator,
1988 &abs_mv_in_out_accumulator, &mv_ratio_accumulator);
1989
1990 // We want to discount the flash frame itself and the recovery
1991 // frame that follows as both will have poor scores.
1992 flash_detected = detect_flash_from_frame_stats(this_frame) ||
1993 detect_flash_from_frame_stats(next_frame);
1994
1995 // Accumulate the effect of prediction quality decay.
1996 if (!flash_detected) {
1997 decay_accumulator *= get_prediction_decay_rate(twopass, this_frame);
1998 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1999 ? MIN_DECAY_FACTOR
2000 : decay_accumulator;
2001 }
2002 boost_score += decay_accumulator *
2003 calc_frame_boost(frame_info, this_frame, twopass,
2004 avg_frame_qindex, this_frame_mv_in_out);
2005 }
2006
2007 arf_boost = (int)boost_score;
2008
2009 // Reset for backward looking loop.
2010 boost_score = 0.0;
2011 mv_ratio_accumulator = 0.0;
2012 decay_accumulator = 1.0;
2013 this_frame_mv_in_out = 0.0;
2014 mv_in_out_accumulator = 0.0;
2015 abs_mv_in_out_accumulator = 0.0;
2016
2017 // Search backward towards last gf position.
2018 for (i = -1; i >= -b_frames; --i) {
2019 const FIRSTPASS_STATS *this_frame =
2020 fps_get_frame_stats(first_pass_info, arf_show_idx + i);
2021 const FIRSTPASS_STATS *next_frame =
2022 fps_get_frame_stats(first_pass_info, arf_show_idx + i + 1);
2023 if (this_frame == NULL) break;
2024
2025 // Update the motion related elements to the boost calculation.
2026 accumulate_frame_motion_stats(
2027 this_frame, &this_frame_mv_in_out, &mv_in_out_accumulator,
2028 &abs_mv_in_out_accumulator, &mv_ratio_accumulator);
2029
2030 // We want to discount the the flash frame itself and the recovery
2031 // frame that follows as both will have poor scores.
2032 flash_detected = detect_flash_from_frame_stats(this_frame) ||
2033 detect_flash_from_frame_stats(next_frame);
2034
2035 // Cumulative effect of prediction quality decay.
2036 if (!flash_detected) {
2037 decay_accumulator *= get_prediction_decay_rate(twopass, this_frame);
2038 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
2039 ? MIN_DECAY_FACTOR
2040 : decay_accumulator;
2041 }
2042 boost_score += decay_accumulator *
2043 calc_frame_boost(frame_info, this_frame, twopass,
2044 avg_frame_qindex, this_frame_mv_in_out);
2045 }
2046 arf_boost += (int)boost_score;
2047
2048 if (arf_boost < ((b_frames + f_frames) * 40))
2049 arf_boost = ((b_frames + f_frames) * 40);
2050 arf_boost = VPXMAX(arf_boost, MIN_ARF_GF_BOOST);
2051
2052 return arf_boost;
2053 }
2054
calc_arf_boost(VP9_COMP * cpi,int f_frames,int b_frames)2055 static int calc_arf_boost(VP9_COMP *cpi, int f_frames, int b_frames) {
2056 const FRAME_INFO *frame_info = &cpi->frame_info;
2057 TWO_PASS *const twopass = &cpi->twopass;
2058 const int avg_inter_frame_qindex = cpi->rc.avg_frame_qindex[INTER_FRAME];
2059 int arf_show_idx = get_show_idx(twopass);
2060 return compute_arf_boost(frame_info, twopass, arf_show_idx, f_frames,
2061 b_frames, avg_inter_frame_qindex);
2062 }
2063
2064 // 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)2065 static int calculate_section_intra_ratio(const FIRSTPASS_STATS *begin,
2066 const FIRSTPASS_STATS *end,
2067 int section_length) {
2068 const FIRSTPASS_STATS *s = begin;
2069 double intra_error = 0.0;
2070 double coded_error = 0.0;
2071 int i = 0;
2072
2073 while (s < end && i < section_length) {
2074 intra_error += s->intra_error;
2075 coded_error += s->coded_error;
2076 ++s;
2077 ++i;
2078 }
2079
2080 return (int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
2081 }
2082
2083 // Calculate the total bits to allocate in this GF/ARF group.
calculate_total_gf_group_bits(VP9_COMP * cpi,double gf_group_err)2084 static int64_t calculate_total_gf_group_bits(VP9_COMP *cpi,
2085 double gf_group_err) {
2086 VP9_COMMON *const cm = &cpi->common;
2087 const RATE_CONTROL *const rc = &cpi->rc;
2088 const TWO_PASS *const twopass = &cpi->twopass;
2089 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
2090 int64_t total_group_bits;
2091 const int is_key_frame = frame_is_intra_only(cm);
2092 const int arf_active_or_kf = is_key_frame || rc->source_alt_ref_active;
2093 int gop_frames =
2094 rc->baseline_gf_interval + rc->source_alt_ref_pending - arf_active_or_kf;
2095
2096 // Calculate the bits to be allocated to the group as a whole.
2097 if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0.0)) {
2098 int key_frame_interval = rc->frames_since_key + rc->frames_to_key;
2099 int distance_from_next_key_frame =
2100 rc->frames_to_key -
2101 (rc->baseline_gf_interval + rc->source_alt_ref_pending);
2102 int max_gf_bits_bias = rc->avg_frame_bandwidth;
2103 double gf_interval_bias_bits_normalize_factor =
2104 (double)rc->baseline_gf_interval / 16;
2105 total_group_bits = (int64_t)(twopass->kf_group_bits *
2106 (gf_group_err / twopass->kf_group_error_left));
2107 // TODO(ravi): Experiment with different values of max_gf_bits_bias
2108 total_group_bits +=
2109 (int64_t)((double)distance_from_next_key_frame / key_frame_interval *
2110 max_gf_bits_bias * gf_interval_bias_bits_normalize_factor);
2111 } else {
2112 total_group_bits = 0;
2113 }
2114
2115 // Clamp odd edge cases.
2116 total_group_bits = (total_group_bits < 0)
2117 ? 0
2118 : (total_group_bits > twopass->kf_group_bits)
2119 ? twopass->kf_group_bits
2120 : total_group_bits;
2121
2122 // Clip based on user supplied data rate variability limit.
2123 if (total_group_bits > (int64_t)max_bits * gop_frames)
2124 total_group_bits = (int64_t)max_bits * gop_frames;
2125
2126 return total_group_bits;
2127 }
2128
2129 // Calculate the number bits extra to assign to boosted frames in a group.
calculate_boost_bits(int frame_count,int boost,int64_t total_group_bits)2130 static int calculate_boost_bits(int frame_count, int boost,
2131 int64_t total_group_bits) {
2132 int allocation_chunks;
2133
2134 // return 0 for invalid inputs (could arise e.g. through rounding errors)
2135 if (!boost || (total_group_bits <= 0) || (frame_count < 0)) return 0;
2136
2137 allocation_chunks = (frame_count * NORMAL_BOOST) + boost;
2138
2139 // Prevent overflow.
2140 if (boost > 1023) {
2141 int divisor = boost >> 10;
2142 boost /= divisor;
2143 allocation_chunks /= divisor;
2144 }
2145
2146 // Calculate the number of extra bits for use in the boosted frame or frames.
2147 return VPXMAX((int)(((int64_t)boost * total_group_bits) / allocation_chunks),
2148 0);
2149 }
2150
2151 // Used in corpus vbr: Calculates the total normalized group complexity score
2152 // for a given number of frames starting at the current position in the stats
2153 // file.
calculate_group_score(VP9_COMP * cpi,double av_score,int frame_count)2154 static double calculate_group_score(VP9_COMP *cpi, double av_score,
2155 int frame_count) {
2156 VP9EncoderConfig *const oxcf = &cpi->oxcf;
2157 TWO_PASS *const twopass = &cpi->twopass;
2158 const FIRSTPASS_STATS *s = twopass->stats_in;
2159 double score_total = 0.0;
2160 int i = 0;
2161
2162 // We dont ever want to return a 0 score here.
2163 if (frame_count == 0) return 1.0;
2164
2165 while ((i < frame_count) && (s < twopass->stats_in_end)) {
2166 score_total += calculate_norm_frame_score(cpi, twopass, oxcf, s, av_score);
2167 ++s;
2168 ++i;
2169 }
2170
2171 return score_total;
2172 }
2173
find_arf_order(VP9_COMP * cpi,GF_GROUP * gf_group,int * index_counter,int depth,int start,int end)2174 static void find_arf_order(VP9_COMP *cpi, GF_GROUP *gf_group,
2175 int *index_counter, int depth, int start, int end) {
2176 TWO_PASS *twopass = &cpi->twopass;
2177 const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
2178 FIRSTPASS_STATS fpf_frame;
2179 const int mid = (start + end + 1) >> 1;
2180 const int min_frame_interval = 2;
2181 int idx;
2182
2183 // Process regular P frames
2184 if ((end - start < min_frame_interval) ||
2185 (depth > gf_group->allowed_max_layer_depth)) {
2186 for (idx = start; idx <= end; ++idx) {
2187 gf_group->update_type[*index_counter] = LF_UPDATE;
2188 gf_group->arf_src_offset[*index_counter] = 0;
2189 gf_group->frame_gop_index[*index_counter] = idx;
2190 gf_group->rf_level[*index_counter] = INTER_NORMAL;
2191 gf_group->layer_depth[*index_counter] = depth;
2192 gf_group->gfu_boost[*index_counter] = NORMAL_BOOST;
2193 ++(*index_counter);
2194 }
2195 gf_group->max_layer_depth = VPXMAX(gf_group->max_layer_depth, depth);
2196 return;
2197 }
2198
2199 assert(abs(mid - start) >= 1 && abs(mid - end) >= 1);
2200
2201 // Process ARF frame
2202 gf_group->layer_depth[*index_counter] = depth;
2203 gf_group->update_type[*index_counter] = ARF_UPDATE;
2204 gf_group->arf_src_offset[*index_counter] = mid - start;
2205 gf_group->frame_gop_index[*index_counter] = mid;
2206 gf_group->rf_level[*index_counter] = GF_ARF_LOW;
2207
2208 for (idx = 0; idx <= mid; ++idx)
2209 if (EOF == input_stats(twopass, &fpf_frame)) break;
2210
2211 gf_group->gfu_boost[*index_counter] =
2212 VPXMAX(MIN_ARF_GF_BOOST,
2213 calc_arf_boost(cpi, end - mid + 1, mid - start) >> depth);
2214
2215 reset_fpf_position(twopass, start_pos);
2216
2217 ++(*index_counter);
2218
2219 find_arf_order(cpi, gf_group, index_counter, depth + 1, start, mid - 1);
2220
2221 gf_group->update_type[*index_counter] = USE_BUF_FRAME;
2222 gf_group->arf_src_offset[*index_counter] = 0;
2223 gf_group->frame_gop_index[*index_counter] = mid;
2224 gf_group->rf_level[*index_counter] = INTER_NORMAL;
2225 gf_group->layer_depth[*index_counter] = depth;
2226 ++(*index_counter);
2227
2228 find_arf_order(cpi, gf_group, index_counter, depth + 1, mid + 1, end);
2229 }
2230
set_gf_overlay_frame_type(GF_GROUP * gf_group,int frame_index,int source_alt_ref_active)2231 static INLINE void set_gf_overlay_frame_type(GF_GROUP *gf_group,
2232 int frame_index,
2233 int source_alt_ref_active) {
2234 if (source_alt_ref_active) {
2235 gf_group->update_type[frame_index] = OVERLAY_UPDATE;
2236 gf_group->rf_level[frame_index] = INTER_NORMAL;
2237 gf_group->layer_depth[frame_index] = MAX_ARF_LAYERS - 1;
2238 gf_group->gfu_boost[frame_index] = NORMAL_BOOST;
2239 } else {
2240 gf_group->update_type[frame_index] = GF_UPDATE;
2241 gf_group->rf_level[frame_index] = GF_ARF_STD;
2242 gf_group->layer_depth[frame_index] = 0;
2243 }
2244 }
2245
define_gf_group_structure(VP9_COMP * cpi)2246 static void define_gf_group_structure(VP9_COMP *cpi) {
2247 RATE_CONTROL *const rc = &cpi->rc;
2248 TWO_PASS *const twopass = &cpi->twopass;
2249 GF_GROUP *const gf_group = &twopass->gf_group;
2250 int frame_index = 0;
2251 int key_frame = cpi->common.frame_type == KEY_FRAME;
2252 int layer_depth = 1;
2253 int gop_frames =
2254 rc->baseline_gf_interval - (key_frame || rc->source_alt_ref_pending);
2255
2256 gf_group->frame_start = cpi->common.current_video_frame;
2257 gf_group->frame_end = gf_group->frame_start + rc->baseline_gf_interval;
2258 gf_group->max_layer_depth = 0;
2259 gf_group->allowed_max_layer_depth = 0;
2260
2261 // For key frames the frame target rate is already set and it
2262 // is also the golden frame.
2263 // === [frame_index == 0] ===
2264 if (!key_frame)
2265 set_gf_overlay_frame_type(gf_group, frame_index, rc->source_alt_ref_active);
2266
2267 ++frame_index;
2268
2269 // === [frame_index == 1] ===
2270 if (rc->source_alt_ref_pending) {
2271 gf_group->update_type[frame_index] = ARF_UPDATE;
2272 gf_group->rf_level[frame_index] = GF_ARF_STD;
2273 gf_group->layer_depth[frame_index] = layer_depth;
2274 gf_group->arf_src_offset[frame_index] =
2275 (unsigned char)(rc->baseline_gf_interval - 1);
2276 gf_group->frame_gop_index[frame_index] = rc->baseline_gf_interval;
2277 gf_group->max_layer_depth = 1;
2278 ++frame_index;
2279 ++layer_depth;
2280 gf_group->allowed_max_layer_depth = cpi->oxcf.enable_auto_arf;
2281 }
2282
2283 find_arf_order(cpi, gf_group, &frame_index, layer_depth, 1, gop_frames);
2284
2285 set_gf_overlay_frame_type(gf_group, frame_index, rc->source_alt_ref_pending);
2286 gf_group->arf_src_offset[frame_index] = 0;
2287 gf_group->frame_gop_index[frame_index] = rc->baseline_gf_interval;
2288
2289 // Set the frame ops number.
2290 gf_group->gf_group_size = frame_index;
2291 }
2292
allocate_gf_group_bits(VP9_COMP * cpi,int64_t gf_group_bits,int gf_arf_bits)2293 static void allocate_gf_group_bits(VP9_COMP *cpi, int64_t gf_group_bits,
2294 int gf_arf_bits) {
2295 VP9EncoderConfig *const oxcf = &cpi->oxcf;
2296 RATE_CONTROL *const rc = &cpi->rc;
2297 TWO_PASS *const twopass = &cpi->twopass;
2298 GF_GROUP *const gf_group = &twopass->gf_group;
2299 FIRSTPASS_STATS frame_stats;
2300 int i;
2301 int frame_index = 0;
2302 int target_frame_size;
2303 int key_frame;
2304 const int max_bits = frame_max_bits(&cpi->rc, oxcf);
2305 int64_t total_group_bits = gf_group_bits;
2306 int mid_frame_idx;
2307 int normal_frames;
2308 int normal_frame_bits;
2309 int last_frame_reduction = 0;
2310 double av_score = 1.0;
2311 double tot_norm_frame_score = 1.0;
2312 double this_frame_score = 1.0;
2313
2314 // Define the GF structure and specify
2315 int gop_frames = gf_group->gf_group_size;
2316
2317 key_frame = cpi->common.frame_type == KEY_FRAME;
2318
2319 // For key frames the frame target rate is already set and it
2320 // is also the golden frame.
2321 // === [frame_index == 0] ===
2322 if (!key_frame) {
2323 gf_group->bit_allocation[frame_index] =
2324 rc->source_alt_ref_active ? 0 : gf_arf_bits;
2325 }
2326
2327 // Deduct the boost bits for arf (or gf if it is not a key frame)
2328 // from the group total.
2329 if (rc->source_alt_ref_pending || !key_frame) total_group_bits -= gf_arf_bits;
2330
2331 ++frame_index;
2332
2333 // === [frame_index == 1] ===
2334 // Store the bits to spend on the ARF if there is one.
2335 if (rc->source_alt_ref_pending) {
2336 gf_group->bit_allocation[frame_index] = gf_arf_bits;
2337
2338 ++frame_index;
2339 }
2340
2341 // Define middle frame
2342 mid_frame_idx = frame_index + (rc->baseline_gf_interval >> 1) - 1;
2343
2344 normal_frames = (rc->baseline_gf_interval - 1);
2345 if (normal_frames > 1)
2346 normal_frame_bits = (int)(total_group_bits / normal_frames);
2347 else
2348 normal_frame_bits = (int)total_group_bits;
2349
2350 gf_group->gfu_boost[1] = rc->gfu_boost;
2351
2352 if (cpi->multi_layer_arf) {
2353 int idx;
2354 int arf_depth_bits[MAX_ARF_LAYERS] = { 0 };
2355 int arf_depth_count[MAX_ARF_LAYERS] = { 0 };
2356 int arf_depth_boost[MAX_ARF_LAYERS] = { 0 };
2357 int total_arfs = 1; // Account for the base layer ARF.
2358
2359 for (idx = 0; idx < gop_frames; ++idx) {
2360 if (gf_group->update_type[idx] == ARF_UPDATE) {
2361 arf_depth_boost[gf_group->layer_depth[idx]] += gf_group->gfu_boost[idx];
2362 ++arf_depth_count[gf_group->layer_depth[idx]];
2363 }
2364 }
2365
2366 for (idx = 2; idx < MAX_ARF_LAYERS; ++idx) {
2367 if (arf_depth_boost[idx] == 0) break;
2368 arf_depth_bits[idx] = calculate_boost_bits(
2369 rc->baseline_gf_interval - total_arfs - arf_depth_count[idx],
2370 arf_depth_boost[idx], total_group_bits);
2371
2372 total_group_bits -= arf_depth_bits[idx];
2373 total_arfs += arf_depth_count[idx];
2374 }
2375
2376 // offset the base layer arf
2377 normal_frames -= (total_arfs - 1);
2378 if (normal_frames > 1)
2379 normal_frame_bits = (int)(total_group_bits / normal_frames);
2380 else
2381 normal_frame_bits = (int)total_group_bits;
2382
2383 target_frame_size = normal_frame_bits;
2384 target_frame_size =
2385 clamp(target_frame_size, 0, VPXMIN(max_bits, (int)total_group_bits));
2386
2387 // The first layer ARF has its bit allocation assigned.
2388 for (idx = frame_index; idx < gop_frames; ++idx) {
2389 switch (gf_group->update_type[idx]) {
2390 case ARF_UPDATE:
2391 gf_group->bit_allocation[idx] =
2392 (int)(((int64_t)arf_depth_bits[gf_group->layer_depth[idx]] *
2393 gf_group->gfu_boost[idx]) /
2394 arf_depth_boost[gf_group->layer_depth[idx]]);
2395 break;
2396 case USE_BUF_FRAME: gf_group->bit_allocation[idx] = 0; break;
2397 default: gf_group->bit_allocation[idx] = target_frame_size; break;
2398 }
2399 }
2400 gf_group->bit_allocation[idx] = 0;
2401
2402 return;
2403 }
2404
2405 if (oxcf->vbr_corpus_complexity) {
2406 av_score = get_distribution_av_err(cpi, twopass);
2407 tot_norm_frame_score = calculate_group_score(cpi, av_score, normal_frames);
2408 }
2409
2410 // Allocate bits to the other frames in the group.
2411 for (i = 0; i < normal_frames; ++i) {
2412 if (EOF == input_stats(twopass, &frame_stats)) break;
2413 if (oxcf->vbr_corpus_complexity) {
2414 this_frame_score = calculate_norm_frame_score(cpi, twopass, oxcf,
2415 &frame_stats, av_score);
2416 normal_frame_bits = (int)((double)total_group_bits *
2417 (this_frame_score / tot_norm_frame_score));
2418 }
2419
2420 target_frame_size = normal_frame_bits;
2421 if ((i == (normal_frames - 1)) && (i >= 1)) {
2422 last_frame_reduction = normal_frame_bits / 16;
2423 target_frame_size -= last_frame_reduction;
2424 }
2425
2426 target_frame_size =
2427 clamp(target_frame_size, 0, VPXMIN(max_bits, (int)total_group_bits));
2428
2429 gf_group->bit_allocation[frame_index] = target_frame_size;
2430 ++frame_index;
2431 }
2432
2433 // Add in some extra bits for the middle frame in the group.
2434 gf_group->bit_allocation[mid_frame_idx] += last_frame_reduction;
2435
2436 // Note:
2437 // We need to configure the frame at the end of the sequence + 1 that will be
2438 // the start frame for the next group. Otherwise prior to the call to
2439 // vp9_rc_get_second_pass_params() the data will be undefined.
2440 }
2441
2442 // Adjusts the ARNF filter for a GF group.
adjust_group_arnr_filter(VP9_COMP * cpi,double section_noise,double section_inter,double section_motion)2443 static void adjust_group_arnr_filter(VP9_COMP *cpi, double section_noise,
2444 double section_inter,
2445 double section_motion) {
2446 TWO_PASS *const twopass = &cpi->twopass;
2447 double section_zeromv = section_inter - section_motion;
2448
2449 twopass->arnr_strength_adjustment = 0;
2450
2451 if (section_noise < 150) {
2452 twopass->arnr_strength_adjustment -= 1;
2453 if (section_noise < 75) twopass->arnr_strength_adjustment -= 1;
2454 } else if (section_noise > 250)
2455 twopass->arnr_strength_adjustment += 1;
2456
2457 if (section_zeromv > 0.50) twopass->arnr_strength_adjustment += 1;
2458 }
2459
2460 // Analyse and define a gf/arf group.
2461 #define ARF_ABS_ZOOM_THRESH 4.0
2462
2463 #define MAX_GF_BOOST 5400
2464
2465 typedef struct RANGE {
2466 int min;
2467 int max;
2468 } RANGE;
2469
2470 /* get_gop_coding_frame_num() depends on several fields in RATE_CONTROL *rc as
2471 * follows.
2472 * Static fields:
2473 * (The following fields will remain unchanged after initialization of encoder.)
2474 * rc->static_scene_max_gf_interval
2475 * rc->min_gf_interval
2476 * twopass->sr_diff_factor
2477 * twopass->sr_default_decay_limit
2478 * twopass->zm_factor
2479 *
2480 * Dynamic fields:
2481 * (The following fields will be updated before or after coding each frame.)
2482 * rc->frames_to_key
2483 * rc->frames_since_key
2484 * rc->source_alt_ref_active
2485 *
2486 * Special case: if CONFIG_RATE_CTRL is true, the external arf indexes will
2487 * determine the arf position.
2488 *
2489 * TODO(angiebird): Separate the dynamic fields and static fields into two
2490 * structs.
2491 */
get_gop_coding_frame_num(int * use_alt_ref,const FRAME_INFO * frame_info,const TWO_PASS * const twopass,const RATE_CONTROL * rc,int gf_start_show_idx,const RANGE * active_gf_interval,double gop_intra_factor,int lag_in_frames)2492 static int get_gop_coding_frame_num(
2493 int *use_alt_ref, const FRAME_INFO *frame_info,
2494 const TWO_PASS *const twopass, const RATE_CONTROL *rc,
2495 int gf_start_show_idx, const RANGE *active_gf_interval,
2496 double gop_intra_factor, int lag_in_frames) {
2497 const FIRST_PASS_INFO *first_pass_info = &twopass->first_pass_info;
2498 double loop_decay_rate = 1.00;
2499 double mv_ratio_accumulator = 0.0;
2500 double this_frame_mv_in_out = 0.0;
2501 double mv_in_out_accumulator = 0.0;
2502 double abs_mv_in_out_accumulator = 0.0;
2503 double sr_accumulator = 0.0;
2504 // Motion breakout threshold for loop below depends on image size.
2505 double mv_ratio_accumulator_thresh =
2506 (frame_info->frame_height + frame_info->frame_width) / 4.0;
2507 double zero_motion_accumulator = 1.0;
2508 int gop_coding_frames;
2509
2510 *use_alt_ref = 1;
2511 gop_coding_frames = 0;
2512 while (gop_coding_frames < rc->static_scene_max_gf_interval &&
2513 gop_coding_frames < rc->frames_to_key) {
2514 const FIRSTPASS_STATS *next_next_frame;
2515 const FIRSTPASS_STATS *next_frame;
2516 int flash_detected;
2517 ++gop_coding_frames;
2518
2519 next_frame = fps_get_frame_stats(first_pass_info,
2520 gf_start_show_idx + gop_coding_frames);
2521 if (next_frame == NULL) {
2522 break;
2523 }
2524
2525 // Test for the case where there is a brief flash but the prediction
2526 // quality back to an earlier frame is then restored.
2527 next_next_frame = fps_get_frame_stats(
2528 first_pass_info, gf_start_show_idx + gop_coding_frames + 1);
2529 flash_detected = detect_flash_from_frame_stats(next_next_frame);
2530
2531 // Update the motion related elements to the boost calculation.
2532 accumulate_frame_motion_stats(
2533 next_frame, &this_frame_mv_in_out, &mv_in_out_accumulator,
2534 &abs_mv_in_out_accumulator, &mv_ratio_accumulator);
2535
2536 // Monitor for static sections.
2537 if ((rc->frames_since_key + gop_coding_frames - 1) > 1) {
2538 zero_motion_accumulator = VPXMIN(
2539 zero_motion_accumulator, get_zero_motion_factor(twopass, next_frame));
2540 }
2541
2542 // Accumulate the effect of prediction quality decay.
2543 if (!flash_detected) {
2544 double last_loop_decay_rate = loop_decay_rate;
2545 loop_decay_rate = get_prediction_decay_rate(twopass, next_frame);
2546
2547 // Break clause to detect very still sections after motion. For example,
2548 // a static image after a fade or other transition.
2549 if (gop_coding_frames > rc->min_gf_interval && loop_decay_rate >= 0.999 &&
2550 last_loop_decay_rate < 0.9) {
2551 int still_interval = 5;
2552 if (check_transition_to_still(first_pass_info,
2553 gf_start_show_idx + gop_coding_frames,
2554 still_interval)) {
2555 *use_alt_ref = 0;
2556 break;
2557 }
2558 }
2559
2560 // Update the accumulator for second ref error difference.
2561 // This is intended to give an indication of how much the coded error is
2562 // increasing over time.
2563 if (gop_coding_frames == 1) {
2564 sr_accumulator += next_frame->coded_error;
2565 } else {
2566 sr_accumulator +=
2567 (next_frame->sr_coded_error - next_frame->coded_error);
2568 }
2569 }
2570
2571 // Break out conditions.
2572 // Break at maximum of active_gf_interval->max unless almost totally
2573 // static.
2574 //
2575 // Note that the addition of a test of rc->source_alt_ref_active is
2576 // deliberate. The effect of this is that after a normal altref group even
2577 // if the material is static there will be one normal length GF group
2578 // before allowing longer GF groups. The reason for this is that in cases
2579 // such as slide shows where slides are separated by a complex transition
2580 // such as a fade, the arf group spanning the transition may not be coded
2581 // at a very high quality and hence this frame (with its overlay) is a
2582 // poor golden frame to use for an extended group.
2583 if ((gop_coding_frames >= active_gf_interval->max) &&
2584 ((zero_motion_accumulator < 0.995) || (rc->source_alt_ref_active))) {
2585 break;
2586 }
2587 if (
2588 // Don't break out with a very short interval.
2589 (gop_coding_frames >= active_gf_interval->min) &&
2590 // If possible dont break very close to a kf
2591 ((rc->frames_to_key - gop_coding_frames) >= rc->min_gf_interval) &&
2592 (gop_coding_frames & 0x01) && (!flash_detected) &&
2593 ((mv_ratio_accumulator > mv_ratio_accumulator_thresh) ||
2594 (abs_mv_in_out_accumulator > ARF_ABS_ZOOM_THRESH) ||
2595 (sr_accumulator > gop_intra_factor * next_frame->intra_error))) {
2596 break;
2597 }
2598 }
2599 *use_alt_ref &= zero_motion_accumulator < 0.995;
2600 *use_alt_ref &= gop_coding_frames < lag_in_frames;
2601 *use_alt_ref &= gop_coding_frames >= rc->min_gf_interval;
2602 return gop_coding_frames;
2603 }
2604
get_active_gf_inverval_range_simple(int min_gf_interval,int arf_active_or_kf,int frames_to_key)2605 static RANGE get_active_gf_inverval_range_simple(int min_gf_interval,
2606 int arf_active_or_kf,
2607 int frames_to_key) {
2608 RANGE active_gf_interval;
2609 active_gf_interval.min = min_gf_interval + arf_active_or_kf + 2;
2610 active_gf_interval.max = 16 + arf_active_or_kf;
2611
2612 if ((active_gf_interval.max <= frames_to_key) &&
2613 (active_gf_interval.max >= (frames_to_key - min_gf_interval))) {
2614 active_gf_interval.min = frames_to_key / 2;
2615 active_gf_interval.max = frames_to_key / 2;
2616 }
2617 return active_gf_interval;
2618 }
2619
get_active_gf_inverval_range(const FRAME_INFO * frame_info,const RATE_CONTROL * rc,int arf_active_or_kf,int gf_start_show_idx,int active_worst_quality,int last_boosted_qindex)2620 static RANGE get_active_gf_inverval_range(
2621 const FRAME_INFO *frame_info, const RATE_CONTROL *rc, int arf_active_or_kf,
2622 int gf_start_show_idx, int active_worst_quality, int last_boosted_qindex) {
2623 RANGE active_gf_interval;
2624 int int_max_q = (int)(vp9_convert_qindex_to_q(active_worst_quality,
2625 frame_info->bit_depth));
2626 int q_term = (gf_start_show_idx == 0)
2627 ? int_max_q / 32
2628 : (int)(vp9_convert_qindex_to_q(last_boosted_qindex,
2629 frame_info->bit_depth) /
2630 6);
2631 active_gf_interval.min =
2632 rc->min_gf_interval + arf_active_or_kf + VPXMIN(2, int_max_q / 200);
2633 active_gf_interval.min =
2634 VPXMIN(active_gf_interval.min, rc->max_gf_interval + arf_active_or_kf);
2635
2636 // The value chosen depends on the active Q range. At low Q we have
2637 // bits to spare and are better with a smaller interval and smaller boost.
2638 // At high Q when there are few bits to spare we are better with a longer
2639 // interval to spread the cost of the GF.
2640 active_gf_interval.max = 11 + arf_active_or_kf + VPXMIN(5, q_term);
2641
2642 // Force max GF interval to be odd.
2643 active_gf_interval.max = active_gf_interval.max | 0x01;
2644
2645 // We have: active_gf_interval.min <=
2646 // rc->max_gf_interval + arf_active_or_kf.
2647 if (active_gf_interval.max < active_gf_interval.min) {
2648 active_gf_interval.max = active_gf_interval.min;
2649 } else {
2650 active_gf_interval.max =
2651 VPXMIN(active_gf_interval.max, rc->max_gf_interval + arf_active_or_kf);
2652 }
2653
2654 // Would the active max drop us out just before the near the next kf?
2655 if ((active_gf_interval.max <= rc->frames_to_key) &&
2656 (active_gf_interval.max >= (rc->frames_to_key - rc->min_gf_interval))) {
2657 active_gf_interval.max = rc->frames_to_key / 2;
2658 }
2659 active_gf_interval.max =
2660 VPXMAX(active_gf_interval.max, active_gf_interval.min);
2661 return active_gf_interval;
2662 }
2663
get_arf_layers(int multi_layer_arf,int max_layers,int coding_frame_num)2664 static int get_arf_layers(int multi_layer_arf, int max_layers,
2665 int coding_frame_num) {
2666 assert(max_layers <= MAX_ARF_LAYERS);
2667 if (multi_layer_arf) {
2668 int layers = 0;
2669 int i;
2670 for (i = coding_frame_num; i > 0; i >>= 1) {
2671 ++layers;
2672 }
2673 layers = VPXMIN(max_layers, layers);
2674 return layers;
2675 } else {
2676 return 1;
2677 }
2678 }
2679
define_gf_group(VP9_COMP * cpi,int gf_start_show_idx)2680 static void define_gf_group(VP9_COMP *cpi, int gf_start_show_idx) {
2681 VP9_COMMON *const cm = &cpi->common;
2682 RATE_CONTROL *const rc = &cpi->rc;
2683 VP9EncoderConfig *const oxcf = &cpi->oxcf;
2684 TWO_PASS *const twopass = &cpi->twopass;
2685 const FRAME_INFO *frame_info = &cpi->frame_info;
2686 const FIRST_PASS_INFO *first_pass_info = &twopass->first_pass_info;
2687 const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
2688 int gop_coding_frames;
2689
2690 double gf_group_err = 0.0;
2691 double gf_group_raw_error = 0.0;
2692 double gf_group_noise = 0.0;
2693 double gf_group_skip_pct = 0.0;
2694 double gf_group_inactive_zone_rows = 0.0;
2695 double gf_group_inter = 0.0;
2696 double gf_group_motion = 0.0;
2697
2698 int allow_alt_ref = is_altref_enabled(cpi);
2699 int use_alt_ref;
2700
2701 int64_t gf_group_bits;
2702 int gf_arf_bits;
2703 const int is_key_frame = frame_is_intra_only(cm);
2704 // If this is a key frame or the overlay from a previous arf then
2705 // the error score / cost of this frame has already been accounted for.
2706 const int arf_active_or_kf = is_key_frame || rc->source_alt_ref_active;
2707 int is_alt_ref_flash = 0;
2708
2709 double gop_intra_factor;
2710 int gop_frames;
2711 RANGE active_gf_interval;
2712
2713 // Reset the GF group data structures unless this is a key
2714 // frame in which case it will already have been done.
2715 if (is_key_frame == 0) {
2716 vp9_zero(twopass->gf_group);
2717 }
2718
2719 vpx_clear_system_state();
2720
2721 if (oxcf->use_simple_encode_api) {
2722 active_gf_interval = get_active_gf_inverval_range_simple(
2723 rc->min_gf_interval, arf_active_or_kf, rc->frames_to_key);
2724 } else {
2725 active_gf_interval = get_active_gf_inverval_range(
2726 frame_info, rc, arf_active_or_kf, gf_start_show_idx,
2727 twopass->active_worst_quality, rc->last_boosted_qindex);
2728 }
2729
2730 if (cpi->multi_layer_arf) {
2731 int arf_layers = get_arf_layers(cpi->multi_layer_arf, oxcf->enable_auto_arf,
2732 active_gf_interval.max);
2733 gop_intra_factor = 1.0 + 0.25 * arf_layers;
2734 } else {
2735 gop_intra_factor = 1.0;
2736 }
2737
2738 gop_coding_frames = get_gop_coding_frame_num(
2739 &use_alt_ref, frame_info, twopass, rc, gf_start_show_idx,
2740 &active_gf_interval, gop_intra_factor, cpi->oxcf.lag_in_frames);
2741 use_alt_ref &= allow_alt_ref;
2742 #if CONFIG_RATE_CTRL
2743 // If the external gop_command is on, we will override the decisions
2744 // of gop_coding_frames and use_alt_ref.
2745 if (cpi->oxcf.use_simple_encode_api) {
2746 const GOP_COMMAND *gop_command = &cpi->encode_command.gop_command;
2747 assert(allow_alt_ref == 1);
2748 if (gop_command->use) {
2749 gop_coding_frames = gop_command_coding_frame_count(gop_command);
2750 use_alt_ref = gop_command->use_alt_ref;
2751 }
2752 }
2753 #endif
2754
2755 // Was the group length constrained by the requirement for a new KF?
2756 rc->constrained_gf_group = (gop_coding_frames >= rc->frames_to_key) ? 1 : 0;
2757
2758 // Should we use the alternate reference frame.
2759 if (use_alt_ref) {
2760 const int f_frames =
2761 (rc->frames_to_key - gop_coding_frames >= gop_coding_frames - 1)
2762 ? gop_coding_frames - 1
2763 : VPXMAX(0, rc->frames_to_key - gop_coding_frames);
2764 const int b_frames = gop_coding_frames - 1;
2765 const int avg_inter_frame_qindex = rc->avg_frame_qindex[INTER_FRAME];
2766 // TODO(angiebird): figure out why arf's location is assigned this way
2767 const int arf_show_idx = VPXMIN(gf_start_show_idx + gop_coding_frames + 1,
2768 fps_get_num_frames(first_pass_info));
2769
2770 // Calculate the boost for alt ref.
2771 rc->gfu_boost =
2772 compute_arf_boost(frame_info, twopass, arf_show_idx, f_frames, b_frames,
2773 avg_inter_frame_qindex);
2774 rc->source_alt_ref_pending = 1;
2775 } else {
2776 const int f_frames = gop_coding_frames - 1;
2777 const int b_frames = 0;
2778 const int avg_inter_frame_qindex = rc->avg_frame_qindex[INTER_FRAME];
2779 // TODO(angiebird): figure out why arf's location is assigned this way
2780 const int gld_show_idx =
2781 VPXMIN(gf_start_show_idx + 1, fps_get_num_frames(first_pass_info));
2782 const int arf_boost =
2783 compute_arf_boost(frame_info, twopass, gld_show_idx, f_frames, b_frames,
2784 avg_inter_frame_qindex);
2785 rc->gfu_boost = VPXMIN((int)twopass->gf_max_total_boost, arf_boost);
2786 rc->source_alt_ref_pending = 0;
2787 }
2788
2789 #define LAST_ALR_ACTIVE_BEST_QUALITY_ADJUSTMENT_FACTOR 0.2
2790 rc->arf_active_best_quality_adjustment_factor = 1.0;
2791 rc->arf_increase_active_best_quality = 0;
2792
2793 if (!is_lossless_requested(&cpi->oxcf)) {
2794 if (rc->frames_since_key >= rc->frames_to_key) {
2795 // Increase the active best quality in the second half of key frame
2796 // interval.
2797 rc->arf_active_best_quality_adjustment_factor =
2798 LAST_ALR_ACTIVE_BEST_QUALITY_ADJUSTMENT_FACTOR +
2799 (1.0 - LAST_ALR_ACTIVE_BEST_QUALITY_ADJUSTMENT_FACTOR) *
2800 (rc->frames_to_key - gop_coding_frames) /
2801 (VPXMAX(1, ((rc->frames_to_key + rc->frames_since_key) / 2 -
2802 gop_coding_frames)));
2803 rc->arf_increase_active_best_quality = 1;
2804 } else if ((rc->frames_to_key - gop_coding_frames) > 0) {
2805 // Reduce the active best quality in the first half of key frame interval.
2806 rc->arf_active_best_quality_adjustment_factor =
2807 LAST_ALR_ACTIVE_BEST_QUALITY_ADJUSTMENT_FACTOR +
2808 (1.0 - LAST_ALR_ACTIVE_BEST_QUALITY_ADJUSTMENT_FACTOR) *
2809 (rc->frames_since_key + gop_coding_frames) /
2810 (VPXMAX(1, (rc->frames_to_key + rc->frames_since_key) / 2 +
2811 gop_coding_frames));
2812 rc->arf_increase_active_best_quality = -1;
2813 }
2814 }
2815
2816 #ifdef AGGRESSIVE_VBR
2817 // Limit maximum boost based on interval length.
2818 rc->gfu_boost = VPXMIN((int)rc->gfu_boost, gop_coding_frames * 140);
2819 #else
2820 rc->gfu_boost = VPXMIN((int)rc->gfu_boost, gop_coding_frames * 200);
2821 #endif
2822
2823 // Cap the ARF boost when perceptual quality AQ mode is enabled. This is
2824 // designed to improve the perceptual quality of high value content and to
2825 // make consistent quality across consecutive frames. It will hurt objective
2826 // quality.
2827 if (oxcf->aq_mode == PERCEPTUAL_AQ)
2828 rc->gfu_boost = VPXMIN(rc->gfu_boost, MIN_ARF_GF_BOOST);
2829
2830 rc->baseline_gf_interval = gop_coding_frames - rc->source_alt_ref_pending;
2831
2832 if (rc->source_alt_ref_pending)
2833 is_alt_ref_flash = detect_flash(twopass, rc->baseline_gf_interval);
2834
2835 {
2836 const double av_err = get_distribution_av_err(cpi, twopass);
2837 const double mean_mod_score = twopass->mean_mod_score;
2838 // If the first frame is a key frame or the overlay from a previous arf then
2839 // the error score / cost of this frame has already been accounted for.
2840 int start_idx = arf_active_or_kf ? 1 : 0;
2841 int j;
2842 for (j = start_idx; j < gop_coding_frames; ++j) {
2843 int show_idx = gf_start_show_idx + j;
2844 const FIRSTPASS_STATS *frame_stats =
2845 fps_get_frame_stats(first_pass_info, show_idx);
2846 // Accumulate error score of frames in this gf group.
2847 gf_group_err += calc_norm_frame_score(oxcf, frame_info, frame_stats,
2848 mean_mod_score, av_err);
2849 gf_group_raw_error += frame_stats->coded_error;
2850 gf_group_noise += frame_stats->frame_noise_energy;
2851 gf_group_skip_pct += frame_stats->intra_skip_pct;
2852 gf_group_inactive_zone_rows += frame_stats->inactive_zone_rows;
2853 gf_group_inter += frame_stats->pcnt_inter;
2854 gf_group_motion += frame_stats->pcnt_motion;
2855 }
2856 }
2857
2858 // Calculate the bits to be allocated to the gf/arf group as a whole
2859 gf_group_bits = calculate_total_gf_group_bits(cpi, gf_group_err);
2860
2861 gop_frames =
2862 rc->baseline_gf_interval + rc->source_alt_ref_pending - arf_active_or_kf;
2863
2864 // Store the average moise level measured for the group
2865 // TODO(any): Experiment with removal of else condition (gop_frames = 0) so
2866 // that consumption of group noise energy is based on previous gf group
2867 if (gop_frames > 0)
2868 twopass->gf_group.group_noise_energy = (int)(gf_group_noise / gop_frames);
2869 else
2870 twopass->gf_group.group_noise_energy = 0;
2871
2872 // Calculate an estimate of the maxq needed for the group.
2873 // We are more aggressive about correcting for sections
2874 // where there could be significant overshoot than for easier
2875 // sections where we do not wish to risk creating an overshoot
2876 // of the allocated bit budget.
2877 if ((cpi->oxcf.rc_mode != VPX_Q) && (rc->baseline_gf_interval > 1)) {
2878 const int vbr_group_bits_per_frame = (int)(gf_group_bits / gop_frames);
2879 const double group_av_err = gf_group_raw_error / gop_frames;
2880 const double group_av_noise = gf_group_noise / gop_frames;
2881 const double group_av_skip_pct = gf_group_skip_pct / gop_frames;
2882 const double group_av_inactive_zone = ((gf_group_inactive_zone_rows * 2) /
2883 (gop_frames * (double)cm->mb_rows));
2884 int tmp_q = get_twopass_worst_quality(
2885 cpi, group_av_err, (group_av_skip_pct + group_av_inactive_zone),
2886 group_av_noise, vbr_group_bits_per_frame);
2887 twopass->active_worst_quality =
2888 (int)((tmp_q + (twopass->active_worst_quality *
2889 (twopass->active_wq_factor - 1))) /
2890 twopass->active_wq_factor);
2891
2892 #if CONFIG_ALWAYS_ADJUST_BPM
2893 // Reset rolling actual and target bits counters for ARF groups.
2894 twopass->rolling_arf_group_target_bits = 0;
2895 twopass->rolling_arf_group_actual_bits = 0;
2896 #endif
2897 }
2898
2899 // Context Adjustment of ARNR filter strength
2900 if (rc->baseline_gf_interval > 1) {
2901 adjust_group_arnr_filter(cpi, (gf_group_noise / gop_frames),
2902 (gf_group_inter / gop_frames),
2903 (gf_group_motion / gop_frames));
2904 } else {
2905 twopass->arnr_strength_adjustment = 0;
2906 }
2907
2908 // Calculate the extra bits to be used for boosted frame(s)
2909 gf_arf_bits = calculate_boost_bits((rc->baseline_gf_interval - 1),
2910 rc->gfu_boost, gf_group_bits);
2911
2912 // Adjust KF group bits and error remaining.
2913 twopass->kf_group_error_left -= gf_group_err;
2914
2915 // Decide GOP structure.
2916 define_gf_group_structure(cpi);
2917
2918 // Allocate bits to each of the frames in the GF group.
2919 allocate_gf_group_bits(cpi, gf_group_bits, gf_arf_bits);
2920
2921 // Reset the file position.
2922 reset_fpf_position(twopass, start_pos);
2923
2924 // Calculate a section intra ratio used in setting max loop filter.
2925 twopass->section_intra_rating = calculate_section_intra_ratio(
2926 start_pos, twopass->stats_in_end, rc->baseline_gf_interval);
2927
2928 if (oxcf->resize_mode == RESIZE_DYNAMIC) {
2929 // Default to starting GF groups at normal frame size.
2930 cpi->rc.next_frame_size_selector = UNSCALED;
2931 }
2932 #if !CONFIG_ALWAYS_ADJUST_BPM
2933 // Reset rolling actual and target bits counters for ARF groups.
2934 twopass->rolling_arf_group_target_bits = 0;
2935 twopass->rolling_arf_group_actual_bits = 0;
2936 #endif
2937 rc->preserve_arf_as_gld = rc->preserve_next_arf_as_gld;
2938 rc->preserve_next_arf_as_gld = 0;
2939 // If alt ref frame is flash do not set preserve_arf_as_gld
2940 if (!is_lossless_requested(&cpi->oxcf) && !cpi->use_svc &&
2941 cpi->oxcf.aq_mode == NO_AQ && cpi->multi_layer_arf && !is_alt_ref_flash)
2942 rc->preserve_next_arf_as_gld = 1;
2943 }
2944
2945 // Intra / Inter threshold very low
2946 #define VERY_LOW_II 1.5
2947 // Clean slide transitions we expect a sharp single frame spike in error.
2948 #define ERROR_SPIKE 5.0
2949
2950 // Slide show transition detection.
2951 // Tests for case where there is very low error either side of the current frame
2952 // but much higher just for this frame. This can help detect key frames in
2953 // slide shows even where the slides are pictures of different sizes.
2954 // Also requires that intra and inter errors are very similar to help eliminate
2955 // harmful false positives.
2956 // It will not help if the transition is a fade or other multi-frame effect.
slide_transition(const FIRSTPASS_STATS * this_frame,const FIRSTPASS_STATS * last_frame,const FIRSTPASS_STATS * next_frame)2957 static int slide_transition(const FIRSTPASS_STATS *this_frame,
2958 const FIRSTPASS_STATS *last_frame,
2959 const FIRSTPASS_STATS *next_frame) {
2960 return (this_frame->intra_error < (this_frame->coded_error * VERY_LOW_II)) &&
2961 (this_frame->coded_error > (last_frame->coded_error * ERROR_SPIKE)) &&
2962 (this_frame->coded_error > (next_frame->coded_error * ERROR_SPIKE));
2963 }
2964
2965 // This test looks for anomalous changes in the nature of the intra signal
2966 // related to the previous and next frame as an indicator for coding a key
2967 // frame. This test serves to detect some additional scene cuts,
2968 // especially in lowish motion and low contrast sections, that are missed
2969 // by the other tests.
intra_step_transition(const FIRSTPASS_STATS * this_frame,const FIRSTPASS_STATS * last_frame,const FIRSTPASS_STATS * next_frame)2970 static int intra_step_transition(const FIRSTPASS_STATS *this_frame,
2971 const FIRSTPASS_STATS *last_frame,
2972 const FIRSTPASS_STATS *next_frame) {
2973 double last_ii_ratio;
2974 double this_ii_ratio;
2975 double next_ii_ratio;
2976 double last_pcnt_intra = 1.0 - last_frame->pcnt_inter;
2977 double this_pcnt_intra = 1.0 - this_frame->pcnt_inter;
2978 double next_pcnt_intra = 1.0 - next_frame->pcnt_inter;
2979 double mod_this_intra = this_pcnt_intra + this_frame->pcnt_neutral;
2980
2981 // Calculate ii ratio for this frame last frame and next frame.
2982 last_ii_ratio =
2983 last_frame->intra_error / DOUBLE_DIVIDE_CHECK(last_frame->coded_error);
2984 this_ii_ratio =
2985 this_frame->intra_error / DOUBLE_DIVIDE_CHECK(this_frame->coded_error);
2986 next_ii_ratio =
2987 next_frame->intra_error / DOUBLE_DIVIDE_CHECK(next_frame->coded_error);
2988
2989 // Return true the intra/inter ratio for the current frame is
2990 // low but better in the next and previous frame and the relative useage of
2991 // intra in the current frame is markedly higher than the last and next frame.
2992 if ((this_ii_ratio < 2.0) && (last_ii_ratio > 2.25) &&
2993 (next_ii_ratio > 2.25) && (this_pcnt_intra > (3 * last_pcnt_intra)) &&
2994 (this_pcnt_intra > (3 * next_pcnt_intra)) &&
2995 ((this_pcnt_intra > 0.075) || (mod_this_intra > 0.85))) {
2996 return 1;
2997 // Very low inter intra ratio (i.e. not much gain from inter coding), most
2998 // blocks neutral on coding method and better inter prediction either side
2999 } else if ((this_ii_ratio < 1.25) && (mod_this_intra > 0.85) &&
3000 (this_ii_ratio < last_ii_ratio * 0.9) &&
3001 (this_ii_ratio < next_ii_ratio * 0.9)) {
3002 return 1;
3003 } else {
3004 return 0;
3005 }
3006 }
3007
3008 // Minimum % intra coding observed in first pass (1.0 = 100%)
3009 #define MIN_INTRA_LEVEL 0.25
3010 // Threshold for use of the lagging second reference frame. Scene cuts do not
3011 // usually have a high second ref useage.
3012 #define SECOND_REF_USEAGE_THRESH 0.2
3013 // Hard threshold where the first pass chooses intra for almost all blocks.
3014 // In such a case even if the frame is not a scene cut coding a key frame
3015 // may be a good option.
3016 #define VERY_LOW_INTER_THRESH 0.05
3017 // Maximum threshold for the relative ratio of intra error score vs best
3018 // inter error score.
3019 #define KF_II_ERR_THRESHOLD 2.5
3020 #define KF_II_MAX 128.0
3021 #define II_FACTOR 12.5
3022 // Test for very low intra complexity which could cause false key frames
3023 #define V_LOW_INTRA 0.5
3024
test_candidate_kf(const FIRST_PASS_INFO * first_pass_info,int show_idx)3025 static int test_candidate_kf(const FIRST_PASS_INFO *first_pass_info,
3026 int show_idx) {
3027 const FIRSTPASS_STATS *last_frame =
3028 fps_get_frame_stats(first_pass_info, show_idx - 1);
3029 const FIRSTPASS_STATS *this_frame =
3030 fps_get_frame_stats(first_pass_info, show_idx);
3031 const FIRSTPASS_STATS *next_frame =
3032 fps_get_frame_stats(first_pass_info, show_idx + 1);
3033 int is_viable_kf = 0;
3034 double pcnt_intra = 1.0 - this_frame->pcnt_inter;
3035
3036 // Does the frame satisfy the primary criteria of a key frame?
3037 // See above for an explanation of the test criteria.
3038 // If so, then examine how well it predicts subsequent frames.
3039 detect_flash_from_frame_stats(next_frame);
3040 if (!detect_flash_from_frame_stats(this_frame) &&
3041 !detect_flash_from_frame_stats(next_frame) &&
3042 (this_frame->pcnt_second_ref < SECOND_REF_USEAGE_THRESH) &&
3043 ((this_frame->pcnt_inter < VERY_LOW_INTER_THRESH) ||
3044 (slide_transition(this_frame, last_frame, next_frame)) ||
3045 (intra_step_transition(this_frame, last_frame, next_frame)) ||
3046 (((this_frame->coded_error > (next_frame->coded_error * 1.2)) &&
3047 (this_frame->coded_error > (last_frame->coded_error * 1.2))) &&
3048 (pcnt_intra > MIN_INTRA_LEVEL) &&
3049 ((pcnt_intra + this_frame->pcnt_neutral) > 0.5) &&
3050 ((this_frame->intra_error /
3051 DOUBLE_DIVIDE_CHECK(this_frame->coded_error)) <
3052 KF_II_ERR_THRESHOLD)))) {
3053 int i;
3054 double boost_score = 0.0;
3055 double old_boost_score = 0.0;
3056 double decay_accumulator = 1.0;
3057
3058 // Examine how well the key frame predicts subsequent frames.
3059 for (i = 0; i < 16; ++i) {
3060 const FIRSTPASS_STATS *frame_stats =
3061 fps_get_frame_stats(first_pass_info, show_idx + 1 + i);
3062 double next_iiratio = (II_FACTOR * frame_stats->intra_error /
3063 DOUBLE_DIVIDE_CHECK(frame_stats->coded_error));
3064
3065 if (next_iiratio > KF_II_MAX) next_iiratio = KF_II_MAX;
3066
3067 // Cumulative effect of decay in prediction quality.
3068 if (frame_stats->pcnt_inter > 0.85)
3069 decay_accumulator *= frame_stats->pcnt_inter;
3070 else
3071 decay_accumulator *= (0.85 + frame_stats->pcnt_inter) / 2.0;
3072
3073 // Keep a running total.
3074 boost_score += (decay_accumulator * next_iiratio);
3075
3076 // Test various breakout clauses.
3077 if ((frame_stats->pcnt_inter < 0.05) || (next_iiratio < 1.5) ||
3078 (((frame_stats->pcnt_inter - frame_stats->pcnt_neutral) < 0.20) &&
3079 (next_iiratio < 3.0)) ||
3080 ((boost_score - old_boost_score) < 3.0) ||
3081 (frame_stats->intra_error < V_LOW_INTRA)) {
3082 break;
3083 }
3084
3085 old_boost_score = boost_score;
3086
3087 // Get the next frame details
3088 if (show_idx + 1 + i == fps_get_num_frames(first_pass_info) - 1) break;
3089 }
3090
3091 // If there is tolerable prediction for at least the next 3 frames then
3092 // break out else discard this potential key frame and move on
3093 if (boost_score > 30.0 && (i > 3)) {
3094 is_viable_kf = 1;
3095 } else {
3096 is_viable_kf = 0;
3097 }
3098 }
3099
3100 return is_viable_kf;
3101 }
3102
3103 #define FRAMES_TO_CHECK_DECAY 8
3104 #define MIN_KF_TOT_BOOST 300
3105 #define DEFAULT_SCAN_FRAMES_FOR_KF_BOOST 32
3106 #define MAX_SCAN_FRAMES_FOR_KF_BOOST 48
3107 #define MIN_SCAN_FRAMES_FOR_KF_BOOST 32
3108 #define KF_ABS_ZOOM_THRESH 6.0
3109
vp9_get_frames_to_next_key(const VP9EncoderConfig * oxcf,const TWO_PASS * const twopass,int kf_show_idx,int min_gf_interval)3110 int vp9_get_frames_to_next_key(const VP9EncoderConfig *oxcf,
3111 const TWO_PASS *const twopass, int kf_show_idx,
3112 int min_gf_interval) {
3113 const FIRST_PASS_INFO *first_pass_info = &twopass->first_pass_info;
3114 double recent_loop_decay[FRAMES_TO_CHECK_DECAY];
3115 int j;
3116 int frames_to_key;
3117 int max_frames_to_key = first_pass_info->num_frames - kf_show_idx;
3118 max_frames_to_key = VPXMIN(max_frames_to_key, oxcf->key_freq);
3119
3120 // Initialize the decay rates for the recent frames to check
3121 for (j = 0; j < FRAMES_TO_CHECK_DECAY; ++j) recent_loop_decay[j] = 1.0;
3122 // Find the next keyframe.
3123 if (!oxcf->auto_key) {
3124 frames_to_key = max_frames_to_key;
3125 } else {
3126 frames_to_key = 1;
3127 while (frames_to_key < max_frames_to_key) {
3128 // Provided that we are not at the end of the file...
3129 if (kf_show_idx + frames_to_key + 1 < first_pass_info->num_frames) {
3130 double loop_decay_rate;
3131 double decay_accumulator;
3132 const FIRSTPASS_STATS *next_frame = fps_get_frame_stats(
3133 first_pass_info, kf_show_idx + frames_to_key + 1);
3134
3135 // Check for a scene cut.
3136 if (test_candidate_kf(first_pass_info, kf_show_idx + frames_to_key))
3137 break;
3138
3139 // How fast is the prediction quality decaying?
3140 loop_decay_rate = get_prediction_decay_rate(twopass, next_frame);
3141
3142 // We want to know something about the recent past... rather than
3143 // as used elsewhere where we are concerned with decay in prediction
3144 // quality since the last GF or KF.
3145 recent_loop_decay[(frames_to_key - 1) % FRAMES_TO_CHECK_DECAY] =
3146 loop_decay_rate;
3147 decay_accumulator = 1.0;
3148 for (j = 0; j < FRAMES_TO_CHECK_DECAY; ++j)
3149 decay_accumulator *= recent_loop_decay[j];
3150
3151 // Special check for transition or high motion followed by a
3152 // static scene.
3153 if ((frames_to_key - 1) > min_gf_interval && loop_decay_rate >= 0.999 &&
3154 decay_accumulator < 0.9) {
3155 int still_interval = oxcf->key_freq - (frames_to_key - 1);
3156 // TODO(angiebird): Figure out why we use "+1" here
3157 int show_idx = kf_show_idx + frames_to_key;
3158 if (check_transition_to_still(first_pass_info, show_idx,
3159 still_interval)) {
3160 break;
3161 }
3162 }
3163 }
3164 ++frames_to_key;
3165 }
3166 }
3167 return frames_to_key;
3168 }
3169
find_next_key_frame(VP9_COMP * cpi,int kf_show_idx)3170 static void find_next_key_frame(VP9_COMP *cpi, int kf_show_idx) {
3171 int i;
3172 RATE_CONTROL *const rc = &cpi->rc;
3173 TWO_PASS *const twopass = &cpi->twopass;
3174 GF_GROUP *const gf_group = &twopass->gf_group;
3175 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
3176 const FIRST_PASS_INFO *first_pass_info = &twopass->first_pass_info;
3177 const FRAME_INFO *frame_info = &cpi->frame_info;
3178 const FIRSTPASS_STATS *const start_position = twopass->stats_in;
3179 const FIRSTPASS_STATS *keyframe_stats =
3180 fps_get_frame_stats(first_pass_info, kf_show_idx);
3181 FIRSTPASS_STATS next_frame;
3182 int kf_bits = 0;
3183 int64_t max_kf_bits;
3184 double zero_motion_accumulator = 1.0;
3185 double zero_motion_sum = 0.0;
3186 double zero_motion_avg;
3187 double motion_compensable_sum = 0.0;
3188 double motion_compensable_avg;
3189 int num_frames = 0;
3190 int kf_boost_scan_frames = DEFAULT_SCAN_FRAMES_FOR_KF_BOOST;
3191 double boost_score = 0.0;
3192 double kf_mod_err = 0.0;
3193 double kf_raw_err = 0.0;
3194 double kf_group_err = 0.0;
3195 double sr_accumulator = 0.0;
3196 double abs_mv_in_out_accumulator = 0.0;
3197 const double av_err = get_distribution_av_err(cpi, twopass);
3198 const double mean_mod_score = twopass->mean_mod_score;
3199 vp9_zero(next_frame);
3200
3201 cpi->common.frame_type = KEY_FRAME;
3202 rc->frames_since_key = 0;
3203
3204 // Reset the GF group data structures.
3205 vp9_zero(*gf_group);
3206
3207 // Is this a forced key frame by interval.
3208 rc->this_key_frame_forced = rc->next_key_frame_forced;
3209
3210 // Clear the alt ref active flag and last group multi arf flags as they
3211 // can never be set for a key frame.
3212 rc->source_alt_ref_active = 0;
3213
3214 // KF is always a GF so clear frames till next gf counter.
3215 rc->frames_till_gf_update_due = 0;
3216
3217 rc->frames_to_key = 1;
3218
3219 twopass->kf_group_bits = 0; // Total bits available to kf group
3220 twopass->kf_group_error_left = 0.0; // Group modified error score.
3221
3222 kf_raw_err = keyframe_stats->intra_error;
3223 kf_mod_err = calc_norm_frame_score(oxcf, frame_info, keyframe_stats,
3224 mean_mod_score, av_err);
3225
3226 rc->frames_to_key = vp9_get_frames_to_next_key(oxcf, twopass, kf_show_idx,
3227 rc->min_gf_interval);
3228
3229 // If there is a max kf interval set by the user we must obey it.
3230 // We already breakout of the loop above at 2x max.
3231 // This code centers the extra kf if the actual natural interval
3232 // is between 1x and 2x.
3233 if (rc->frames_to_key >= cpi->oxcf.key_freq) {
3234 rc->next_key_frame_forced = 1;
3235 } else {
3236 rc->next_key_frame_forced = 0;
3237 }
3238
3239 for (i = 0; i < rc->frames_to_key; ++i) {
3240 const FIRSTPASS_STATS *frame_stats =
3241 fps_get_frame_stats(first_pass_info, kf_show_idx + i);
3242 // Accumulate kf group error.
3243 kf_group_err += calc_norm_frame_score(oxcf, frame_info, frame_stats,
3244 mean_mod_score, av_err);
3245 }
3246
3247 // Calculate the number of bits that should be assigned to the kf group.
3248 if (twopass->bits_left > 0 && twopass->normalized_score_left > 0.0) {
3249 // Maximum number of bits for a single normal frame (not key frame).
3250 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
3251
3252 // Maximum number of bits allocated to the key frame group.
3253 int64_t max_grp_bits;
3254
3255 // Default allocation based on bits left and relative
3256 // complexity of the section.
3257 twopass->kf_group_bits = (int64_t)(
3258 twopass->bits_left * (kf_group_err / twopass->normalized_score_left));
3259
3260 // Clip based on maximum per frame rate defined by the user.
3261 max_grp_bits = (int64_t)max_bits * (int64_t)rc->frames_to_key;
3262 if (twopass->kf_group_bits > max_grp_bits)
3263 twopass->kf_group_bits = max_grp_bits;
3264 } else {
3265 twopass->kf_group_bits = 0;
3266 }
3267 twopass->kf_group_bits = VPXMAX(0, twopass->kf_group_bits);
3268
3269 // Scan through the kf group collating various stats used to determine
3270 // how many bits to spend on it.
3271 boost_score = 0.0;
3272
3273 for (i = 0; i < VPXMIN(MAX_SCAN_FRAMES_FOR_KF_BOOST, (rc->frames_to_key - 1));
3274 ++i) {
3275 if (EOF == input_stats(twopass, &next_frame)) break;
3276
3277 zero_motion_sum += next_frame.pcnt_inter - next_frame.pcnt_motion;
3278 motion_compensable_sum +=
3279 1 - (double)next_frame.coded_error / next_frame.intra_error;
3280 num_frames++;
3281 }
3282
3283 if (num_frames >= MIN_SCAN_FRAMES_FOR_KF_BOOST) {
3284 zero_motion_avg = zero_motion_sum / num_frames;
3285 motion_compensable_avg = motion_compensable_sum / num_frames;
3286 kf_boost_scan_frames = (int)(VPXMAX(64 * zero_motion_avg - 16,
3287 160 * motion_compensable_avg - 112));
3288 kf_boost_scan_frames =
3289 VPXMAX(VPXMIN(kf_boost_scan_frames, MAX_SCAN_FRAMES_FOR_KF_BOOST),
3290 MIN_SCAN_FRAMES_FOR_KF_BOOST);
3291 }
3292 reset_fpf_position(twopass, start_position);
3293
3294 for (i = 0; i < (rc->frames_to_key - 1); ++i) {
3295 if (EOF == input_stats(twopass, &next_frame)) break;
3296
3297 // The zero motion test here insures that if we mark a kf group as static
3298 // it is static throughout not just the first KF_BOOST_SCAN_MAX_FRAMES.
3299 // It also allows for a larger boost on long static groups.
3300 if ((i <= kf_boost_scan_frames) || (zero_motion_accumulator >= 0.99)) {
3301 double frame_boost;
3302 double zm_factor;
3303
3304 // Monitor for static sections.
3305 // First frame in kf group the second ref indicator is invalid.
3306 if (i > 0) {
3307 zero_motion_accumulator =
3308 VPXMIN(zero_motion_accumulator,
3309 get_zero_motion_factor(twopass, &next_frame));
3310 } else {
3311 zero_motion_accumulator =
3312 next_frame.pcnt_inter - next_frame.pcnt_motion;
3313 }
3314
3315 // Factor 0.75-1.25 based on how much of frame is static.
3316 zm_factor = (0.75 + (zero_motion_accumulator / 2.0));
3317
3318 // The second (lagging) ref error is not valid immediately after
3319 // a key frame because either the lag has not built up (in the case of
3320 // the first key frame or it points to a refernce before the new key
3321 // frame.
3322 if (i < 2) sr_accumulator = 0.0;
3323 frame_boost =
3324 calc_kf_frame_boost(cpi, &next_frame, &sr_accumulator, 0, zm_factor);
3325
3326 boost_score += frame_boost;
3327
3328 // Measure of zoom. Large zoom tends to indicate reduced boost.
3329 abs_mv_in_out_accumulator +=
3330 fabs(next_frame.mv_in_out_count * next_frame.pcnt_motion);
3331
3332 if ((frame_boost < 25.00) ||
3333 (abs_mv_in_out_accumulator > KF_ABS_ZOOM_THRESH) ||
3334 (sr_accumulator > (kf_raw_err * 1.50)))
3335 break;
3336 } else {
3337 break;
3338 }
3339 }
3340
3341 reset_fpf_position(twopass, start_position);
3342
3343 // Store the zero motion percentage
3344 twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
3345
3346 // Calculate a section intra ratio used in setting max loop filter.
3347 twopass->key_frame_section_intra_rating = calculate_section_intra_ratio(
3348 start_position, twopass->stats_in_end, rc->frames_to_key);
3349
3350 // Special case for static / slide show content but dont apply
3351 // if the kf group is very short.
3352 if ((zero_motion_accumulator > 0.99) && (rc->frames_to_key > 8)) {
3353 rc->kf_boost = (int)(twopass->kf_max_total_boost);
3354 } else {
3355 // Apply various clamps for min and max oost
3356 rc->kf_boost = VPXMAX((int)boost_score, (rc->frames_to_key * 3));
3357 rc->kf_boost = VPXMAX(rc->kf_boost, MIN_KF_TOT_BOOST);
3358 rc->kf_boost = VPXMIN(rc->kf_boost, (int)(twopass->kf_max_total_boost));
3359 }
3360
3361 // Work out how many bits to allocate for the key frame itself.
3362 kf_bits = calculate_boost_bits((rc->frames_to_key - 1), rc->kf_boost,
3363 twopass->kf_group_bits);
3364 // Based on the spatial complexity, increase the bits allocated to key frame.
3365 kf_bits +=
3366 (int)((twopass->kf_group_bits - kf_bits) * (kf_mod_err / kf_group_err));
3367 max_kf_bits =
3368 twopass->kf_group_bits - (rc->frames_to_key - 1) * FRAME_OVERHEAD_BITS;
3369 max_kf_bits = lclamp(max_kf_bits, 0, INT_MAX);
3370 kf_bits = VPXMIN(kf_bits, (int)max_kf_bits);
3371
3372 twopass->kf_group_bits -= kf_bits;
3373
3374 // Save the bits to spend on the key frame.
3375 gf_group->bit_allocation[0] = kf_bits;
3376 gf_group->update_type[0] = KF_UPDATE;
3377 gf_group->rf_level[0] = KF_STD;
3378 gf_group->layer_depth[0] = 0;
3379
3380 // Note the total error score of the kf group minus the key frame itself.
3381 twopass->kf_group_error_left = (kf_group_err - kf_mod_err);
3382
3383 // Adjust the count of total modified error left.
3384 // The count of bits left is adjusted elsewhere based on real coded frame
3385 // sizes.
3386 twopass->normalized_score_left -= kf_group_err;
3387
3388 if (oxcf->resize_mode == RESIZE_DYNAMIC) {
3389 // Default to normal-sized frame on keyframes.
3390 cpi->rc.next_frame_size_selector = UNSCALED;
3391 }
3392 }
3393
3394 // Configure image size specific vizier parameters.
3395 // Later these will be set via additional command line options
vp9_init_vizier_params(TWO_PASS * const twopass,int screen_area)3396 void vp9_init_vizier_params(TWO_PASS *const twopass, int screen_area) {
3397 // When |use_vizier_rc_params| is 1, we expect the rc parameters below to
3398 // have been initialised on the command line as adjustment factors such
3399 // that a factor of 1.0 will match the default behavior when
3400 // |use_vizier_rc_params| is 0
3401 if (twopass->use_vizier_rc_params) {
3402 twopass->active_wq_factor *= AV_WQ_FACTOR;
3403 twopass->err_per_mb *= BASELINE_ERR_PER_MB;
3404 twopass->sr_default_decay_limit *= DEFAULT_DECAY_LIMIT;
3405 if (twopass->sr_default_decay_limit > 1.0) // > 1.0 here makes no sense
3406 twopass->sr_default_decay_limit = 1.0;
3407 twopass->sr_diff_factor *= 1.0;
3408 twopass->gf_frame_max_boost *= GF_MAX_FRAME_BOOST;
3409 twopass->gf_max_total_boost *= MAX_GF_BOOST;
3410 // NOTE: In use max boost has precedence over min boost. So even if min is
3411 // somehow set higher than max the final boost value will be clamped to the
3412 // appropriate maximum.
3413 twopass->kf_frame_min_boost *= KF_MIN_FRAME_BOOST;
3414 twopass->kf_frame_max_boost_first *= KF_MAX_FRAME_BOOST;
3415 twopass->kf_frame_max_boost_subs *= KF_MAX_FRAME_BOOST;
3416 twopass->kf_max_total_boost *= MAX_KF_TOT_BOOST;
3417 twopass->zm_factor *= DEFAULT_ZM_FACTOR;
3418 if (twopass->zm_factor > 1.0) // > 1.0 here makes no sense
3419 twopass->zm_factor = 1.0;
3420
3421 // Correction for the fact that the kf_err_per_mb_factor default is
3422 // already different for different video formats and ensures that a passed
3423 // in value of 1.0 on the vizier command line will still match the current
3424 // default.
3425 if (screen_area < 1280 * 720) {
3426 twopass->kf_err_per_mb *= 2000.0;
3427 } else if (screen_area < 1920 * 1080) {
3428 twopass->kf_err_per_mb *= 500.0;
3429 } else {
3430 twopass->kf_err_per_mb *= 250.0;
3431 }
3432 } else {
3433 // When |use_vizier_rc_params| is 0, use defaults.
3434 twopass->active_wq_factor = AV_WQ_FACTOR;
3435 twopass->err_per_mb = BASELINE_ERR_PER_MB;
3436 twopass->sr_default_decay_limit = DEFAULT_DECAY_LIMIT;
3437 twopass->sr_diff_factor = 1.0;
3438 twopass->gf_frame_max_boost = GF_MAX_FRAME_BOOST;
3439 twopass->gf_max_total_boost = MAX_GF_BOOST;
3440 twopass->kf_frame_min_boost = KF_MIN_FRAME_BOOST;
3441 twopass->kf_frame_max_boost_first = KF_MAX_FRAME_BOOST;
3442 twopass->kf_frame_max_boost_subs = KF_MAX_FRAME_BOOST;
3443 twopass->kf_max_total_boost = MAX_KF_TOT_BOOST;
3444 twopass->zm_factor = DEFAULT_ZM_FACTOR;
3445
3446 if (screen_area < 1280 * 720) {
3447 twopass->kf_err_per_mb = 2000.0;
3448 } else if (screen_area < 1920 * 1080) {
3449 twopass->kf_err_per_mb = 500.0;
3450 } else {
3451 twopass->kf_err_per_mb = 250.0;
3452 }
3453 }
3454 }
3455
vp9_rc_get_second_pass_params(VP9_COMP * cpi)3456 void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
3457 VP9_COMMON *const cm = &cpi->common;
3458 RATE_CONTROL *const rc = &cpi->rc;
3459 TWO_PASS *const twopass = &cpi->twopass;
3460 GF_GROUP *const gf_group = &twopass->gf_group;
3461 FIRSTPASS_STATS this_frame;
3462 const int show_idx = cm->current_video_frame;
3463
3464 if (!twopass->stats_in) return;
3465
3466 // Configure image size specific vizier parameters
3467 if (cm->current_video_frame == 0) {
3468 unsigned int screen_area = (cm->width * cm->height);
3469
3470 vp9_init_vizier_params(twopass, screen_area);
3471 }
3472
3473 // If this is an arf frame then we dont want to read the stats file or
3474 // advance the input pointer as we already have what we need.
3475 if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
3476 int target_rate;
3477
3478 vp9_zero(this_frame);
3479 this_frame =
3480 cpi->twopass.stats_in_start[cm->current_video_frame +
3481 gf_group->arf_src_offset[gf_group->index]];
3482
3483 vp9_configure_buffer_updates(cpi, gf_group->index);
3484
3485 target_rate = gf_group->bit_allocation[gf_group->index];
3486 target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
3487 rc->base_frame_target = target_rate;
3488
3489 cm->frame_type = INTER_FRAME;
3490
3491 // The multiplication by 256 reverses a scaling factor of (>> 8)
3492 // applied when combining MB error values for the frame.
3493 twopass->mb_av_energy = log((this_frame.intra_error * 256.0) + 1.0);
3494 twopass->mb_smooth_pct = this_frame.intra_smooth_pct;
3495
3496 return;
3497 }
3498
3499 vpx_clear_system_state();
3500
3501 if (cpi->oxcf.rc_mode == VPX_Q) {
3502 twopass->active_worst_quality = cpi->oxcf.cq_level;
3503 } else if (cm->current_video_frame == 0) {
3504 const int frames_left =
3505 (int)(twopass->total_stats.count - cm->current_video_frame);
3506 // Special case code for first frame.
3507 const int section_target_bandwidth =
3508 (int)(twopass->bits_left / frames_left);
3509 const double section_length = twopass->total_left_stats.count;
3510 const double section_error =
3511 twopass->total_left_stats.coded_error / section_length;
3512 const double section_intra_skip =
3513 twopass->total_left_stats.intra_skip_pct / section_length;
3514 const double section_inactive_zone =
3515 (twopass->total_left_stats.inactive_zone_rows * 2) /
3516 ((double)cm->mb_rows * section_length);
3517 const double section_noise =
3518 twopass->total_left_stats.frame_noise_energy / section_length;
3519 int tmp_q;
3520
3521 tmp_q = get_twopass_worst_quality(
3522 cpi, section_error, section_intra_skip + section_inactive_zone,
3523 section_noise, section_target_bandwidth);
3524
3525 twopass->active_worst_quality = tmp_q;
3526 twopass->baseline_active_worst_quality = tmp_q;
3527 rc->ni_av_qi = tmp_q;
3528 rc->last_q[INTER_FRAME] = tmp_q;
3529 rc->avg_q = vp9_convert_qindex_to_q(tmp_q, cm->bit_depth);
3530 rc->avg_frame_qindex[INTER_FRAME] = tmp_q;
3531 rc->last_q[KEY_FRAME] = (tmp_q + cpi->oxcf.best_allowed_q) / 2;
3532 rc->avg_frame_qindex[KEY_FRAME] = rc->last_q[KEY_FRAME];
3533 }
3534 vp9_zero(this_frame);
3535 if (EOF == input_stats(twopass, &this_frame)) return;
3536
3537 // Set the frame content type flag.
3538 if (this_frame.intra_skip_pct >= FC_ANIMATION_THRESH)
3539 twopass->fr_content_type = FC_GRAPHICS_ANIMATION;
3540 else
3541 twopass->fr_content_type = FC_NORMAL;
3542
3543 // Keyframe and section processing.
3544 if (rc->frames_to_key == 0 || (cpi->frame_flags & FRAMEFLAGS_KEY)) {
3545 // Define next KF group and assign bits to it.
3546 find_next_key_frame(cpi, show_idx);
3547 } else {
3548 cm->frame_type = INTER_FRAME;
3549 }
3550
3551 // Define a new GF/ARF group. (Should always enter here for key frames).
3552 if (rc->frames_till_gf_update_due == 0) {
3553 define_gf_group(cpi, show_idx);
3554
3555 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
3556
3557 #if ARF_STATS_OUTPUT
3558 {
3559 FILE *fpfile;
3560 fpfile = fopen("arf.stt", "a");
3561 ++arf_count;
3562 fprintf(fpfile, "%10d %10ld %10d %10d %10ld %10ld\n",
3563 cm->current_video_frame, rc->frames_till_gf_update_due,
3564 rc->kf_boost, arf_count, rc->gfu_boost, cm->frame_type);
3565
3566 fclose(fpfile);
3567 }
3568 #endif
3569 }
3570
3571 vp9_configure_buffer_updates(cpi, gf_group->index);
3572
3573 rc->base_frame_target = gf_group->bit_allocation[gf_group->index];
3574
3575 // The multiplication by 256 reverses a scaling factor of (>> 8)
3576 // applied when combining MB error values for the frame.
3577 twopass->mb_av_energy = log((this_frame.intra_error * 256.0) + 1.0);
3578 twopass->mb_smooth_pct = this_frame.intra_smooth_pct;
3579
3580 // Update the total stats remaining structure.
3581 subtract_stats(&twopass->total_left_stats, &this_frame);
3582 }
3583
vp9_twopass_postencode_update(VP9_COMP * cpi)3584 void vp9_twopass_postencode_update(VP9_COMP *cpi) {
3585 TWO_PASS *const twopass = &cpi->twopass;
3586 RATE_CONTROL *const rc = &cpi->rc;
3587 VP9_COMMON *const cm = &cpi->common;
3588 const int bits_used = rc->base_frame_target;
3589
3590 // VBR correction is done through rc->vbr_bits_off_target. Based on the
3591 // sign of this value, a limited % adjustment is made to the target rate
3592 // of subsequent frames, to try and push it back towards 0. This method
3593 // is designed to prevent extreme behaviour at the end of a clip
3594 // or group of frames.
3595 rc->vbr_bits_off_target += rc->base_frame_target - rc->projected_frame_size;
3596 twopass->bits_left = VPXMAX(twopass->bits_left - bits_used, 0);
3597
3598 // Target vs actual bits for this arf group.
3599 twopass->rolling_arf_group_target_bits += rc->this_frame_target;
3600 twopass->rolling_arf_group_actual_bits += rc->projected_frame_size;
3601
3602 // Calculate the pct rc error.
3603 if (rc->total_actual_bits) {
3604 rc->rate_error_estimate =
3605 (int)((rc->vbr_bits_off_target * 100) / rc->total_actual_bits);
3606 rc->rate_error_estimate = clamp(rc->rate_error_estimate, -100, 100);
3607 } else {
3608 rc->rate_error_estimate = 0;
3609 }
3610
3611 if (cpi->common.frame_type != KEY_FRAME) {
3612 twopass->kf_group_bits -= bits_used;
3613 twopass->last_kfgroup_zeromotion_pct = twopass->kf_zeromotion_pct;
3614 }
3615 twopass->kf_group_bits = VPXMAX(twopass->kf_group_bits, 0);
3616
3617 // Increment the gf group index ready for the next frame.
3618 ++twopass->gf_group.index;
3619
3620 // If the rate control is drifting consider adjustment to min or maxq.
3621 if ((cpi->oxcf.rc_mode != VPX_Q) && !cpi->rc.is_src_frame_alt_ref) {
3622 const int maxq_adj_limit =
3623 rc->worst_quality - twopass->active_worst_quality;
3624 const int minq_adj_limit =
3625 (cpi->oxcf.rc_mode == VPX_CQ ? MINQ_ADJ_LIMIT_CQ : MINQ_ADJ_LIMIT);
3626 int aq_extend_min = 0;
3627 int aq_extend_max = 0;
3628
3629 // Extend min or Max Q range to account for imbalance from the base
3630 // value when using AQ.
3631 if (cpi->oxcf.aq_mode != NO_AQ && cpi->oxcf.aq_mode != PSNR_AQ &&
3632 cpi->oxcf.aq_mode != PERCEPTUAL_AQ) {
3633 if (cm->seg.aq_av_offset < 0) {
3634 // The balance of the AQ map tends towarda lowering the average Q.
3635 aq_extend_min = 0;
3636 aq_extend_max = VPXMIN(maxq_adj_limit, -cm->seg.aq_av_offset);
3637 } else {
3638 // The balance of the AQ map tends towards raising the average Q.
3639 aq_extend_min = VPXMIN(minq_adj_limit, cm->seg.aq_av_offset);
3640 aq_extend_max = 0;
3641 }
3642 }
3643
3644 // Undershoot.
3645 if (rc->rate_error_estimate > cpi->oxcf.under_shoot_pct) {
3646 --twopass->extend_maxq;
3647 if (rc->rolling_target_bits >= rc->rolling_actual_bits)
3648 ++twopass->extend_minq;
3649 // Overshoot.
3650 } else if (rc->rate_error_estimate < -cpi->oxcf.over_shoot_pct) {
3651 --twopass->extend_minq;
3652 if (rc->rolling_target_bits < rc->rolling_actual_bits)
3653 ++twopass->extend_maxq;
3654 } else {
3655 // Adjustment for extreme local overshoot.
3656 if (rc->projected_frame_size > (2 * rc->base_frame_target) &&
3657 rc->projected_frame_size > (2 * rc->avg_frame_bandwidth))
3658 ++twopass->extend_maxq;
3659
3660 // Unwind undershoot or overshoot adjustment.
3661 if (rc->rolling_target_bits < rc->rolling_actual_bits)
3662 --twopass->extend_minq;
3663 else if (rc->rolling_target_bits > rc->rolling_actual_bits)
3664 --twopass->extend_maxq;
3665 }
3666
3667 twopass->extend_minq =
3668 clamp(twopass->extend_minq, aq_extend_min, minq_adj_limit);
3669 twopass->extend_maxq =
3670 clamp(twopass->extend_maxq, aq_extend_max, maxq_adj_limit);
3671
3672 // If there is a big and undexpected undershoot then feed the extra
3673 // bits back in quickly. One situation where this may happen is if a
3674 // frame is unexpectedly almost perfectly predicted by the ARF or GF
3675 // but not very well predcited by the previous frame.
3676 if (!frame_is_kf_gf_arf(cpi) && !cpi->rc.is_src_frame_alt_ref) {
3677 int fast_extra_thresh = rc->base_frame_target / HIGH_UNDERSHOOT_RATIO;
3678 if (rc->projected_frame_size < fast_extra_thresh) {
3679 rc->vbr_bits_off_target_fast +=
3680 fast_extra_thresh - rc->projected_frame_size;
3681 rc->vbr_bits_off_target_fast =
3682 VPXMIN(rc->vbr_bits_off_target_fast, (4 * rc->avg_frame_bandwidth));
3683
3684 // Fast adaptation of minQ if necessary to use up the extra bits.
3685 if (rc->avg_frame_bandwidth) {
3686 twopass->extend_minq_fast =
3687 (int)(rc->vbr_bits_off_target_fast * 8 / rc->avg_frame_bandwidth);
3688 }
3689 twopass->extend_minq_fast = VPXMIN(
3690 twopass->extend_minq_fast, minq_adj_limit - twopass->extend_minq);
3691 } else if (rc->vbr_bits_off_target_fast) {
3692 twopass->extend_minq_fast = VPXMIN(
3693 twopass->extend_minq_fast, minq_adj_limit - twopass->extend_minq);
3694 } else {
3695 twopass->extend_minq_fast = 0;
3696 }
3697 }
3698 }
3699 }
3700
3701 #if CONFIG_RATE_CTRL
vp9_get_next_group_of_picture(const VP9_COMP * cpi,int * first_is_key_frame,int * use_alt_ref,int * coding_frame_count,int * first_show_idx,int * last_gop_use_alt_ref)3702 void vp9_get_next_group_of_picture(const VP9_COMP *cpi, int *first_is_key_frame,
3703 int *use_alt_ref, int *coding_frame_count,
3704 int *first_show_idx,
3705 int *last_gop_use_alt_ref) {
3706 const GOP_COMMAND *gop_command = &cpi->encode_command.gop_command;
3707 // We make a copy of rc here because we want to get information from the
3708 // encoder without changing its state.
3709 // TODO(angiebird): Avoid copying rc here.
3710 RATE_CONTROL rc = cpi->rc;
3711 const int multi_layer_arf = 0;
3712 const int allow_alt_ref = 1;
3713 // We assume that current_video_frame is updated to the show index of the
3714 // frame we are about to called. Note that current_video_frame is updated at
3715 // the end of encode_frame_to_data_rate().
3716 // TODO(angiebird): Avoid this kind of fragile style.
3717 *first_show_idx = cpi->common.current_video_frame;
3718 *last_gop_use_alt_ref = rc.source_alt_ref_active;
3719
3720 *first_is_key_frame = 0;
3721 if (rc.frames_to_key == 0) {
3722 rc.frames_to_key = vp9_get_frames_to_next_key(
3723 &cpi->oxcf, &cpi->twopass, *first_show_idx, rc.min_gf_interval);
3724 rc.frames_since_key = 0;
3725 *first_is_key_frame = 1;
3726 }
3727
3728 if (gop_command->use) {
3729 *coding_frame_count = gop_command_coding_frame_count(gop_command);
3730 *use_alt_ref = gop_command->use_alt_ref;
3731 assert(gop_command->show_frame_count <= rc.frames_to_key);
3732 } else {
3733 *coding_frame_count = vp9_get_gop_coding_frame_count(
3734 &cpi->oxcf, &cpi->twopass, &cpi->frame_info, &rc, *first_show_idx,
3735 multi_layer_arf, allow_alt_ref, *first_is_key_frame,
3736 *last_gop_use_alt_ref, use_alt_ref);
3737 }
3738 }
3739
vp9_get_gop_coding_frame_count(const VP9EncoderConfig * oxcf,const TWO_PASS * const twopass,const FRAME_INFO * frame_info,const RATE_CONTROL * rc,int show_idx,int multi_layer_arf,int allow_alt_ref,int first_is_key_frame,int last_gop_use_alt_ref,int * use_alt_ref)3740 int vp9_get_gop_coding_frame_count(const VP9EncoderConfig *oxcf,
3741 const TWO_PASS *const twopass,
3742 const FRAME_INFO *frame_info,
3743 const RATE_CONTROL *rc, int show_idx,
3744 int multi_layer_arf, int allow_alt_ref,
3745 int first_is_key_frame,
3746 int last_gop_use_alt_ref, int *use_alt_ref) {
3747 int frame_count;
3748 double gop_intra_factor;
3749 const int arf_active_or_kf = last_gop_use_alt_ref || first_is_key_frame;
3750 RANGE active_gf_interval;
3751 int arf_layers;
3752 if (oxcf->use_simple_encode_api) {
3753 active_gf_interval = get_active_gf_inverval_range_simple(
3754 rc->min_gf_interval, arf_active_or_kf, rc->frames_to_key);
3755 } else {
3756 active_gf_interval = get_active_gf_inverval_range(
3757 frame_info, rc, arf_active_or_kf, show_idx, /*active_worst_quality=*/0,
3758 /*last_boosted_qindex=*/0);
3759 }
3760
3761 arf_layers = get_arf_layers(multi_layer_arf, oxcf->enable_auto_arf,
3762 active_gf_interval.max);
3763 if (multi_layer_arf) {
3764 gop_intra_factor = 1.0 + 0.25 * arf_layers;
3765 } else {
3766 gop_intra_factor = 1.0;
3767 }
3768
3769 frame_count = get_gop_coding_frame_num(use_alt_ref, frame_info, twopass, rc,
3770 show_idx, &active_gf_interval,
3771 gop_intra_factor, oxcf->lag_in_frames);
3772 *use_alt_ref &= allow_alt_ref;
3773 return frame_count;
3774 }
3775
3776 // Under CONFIG_RATE_CTRL, once the first_pass_info is ready, the number of
3777 // coding frames (including show frame and alt ref) can be determined.
vp9_get_coding_frame_num(const VP9EncoderConfig * oxcf,const TWO_PASS * const twopass,const FRAME_INFO * frame_info,int multi_layer_arf,int allow_alt_ref)3778 int vp9_get_coding_frame_num(const VP9EncoderConfig *oxcf,
3779 const TWO_PASS *const twopass,
3780 const FRAME_INFO *frame_info, int multi_layer_arf,
3781 int allow_alt_ref) {
3782 const FIRST_PASS_INFO *first_pass_info = &twopass->first_pass_info;
3783 int coding_frame_num = 0;
3784 RATE_CONTROL rc;
3785 int gop_coding_frame_count;
3786 int gop_show_frames;
3787 int show_idx = 0;
3788 int last_gop_use_alt_ref = 0;
3789 vp9_rc_init(oxcf, 1, &rc);
3790
3791 while (show_idx < first_pass_info->num_frames) {
3792 int use_alt_ref;
3793 int first_is_key_frame = 0;
3794 if (rc.frames_to_key == 0) {
3795 rc.frames_to_key = vp9_get_frames_to_next_key(oxcf, twopass, show_idx,
3796 rc.min_gf_interval);
3797 rc.frames_since_key = 0;
3798 first_is_key_frame = 1;
3799 }
3800
3801 gop_coding_frame_count = vp9_get_gop_coding_frame_count(
3802 oxcf, twopass, frame_info, &rc, show_idx, multi_layer_arf,
3803 allow_alt_ref, first_is_key_frame, last_gop_use_alt_ref, &use_alt_ref);
3804
3805 rc.source_alt_ref_active = use_alt_ref;
3806 last_gop_use_alt_ref = use_alt_ref;
3807 gop_show_frames = gop_coding_frame_count - use_alt_ref;
3808 rc.frames_to_key -= gop_show_frames;
3809 rc.frames_since_key += gop_show_frames;
3810 show_idx += gop_show_frames;
3811 coding_frame_num += gop_show_frames + use_alt_ref;
3812 }
3813 return coding_frame_num;
3814 }
3815
vp9_get_key_frame_map(const VP9EncoderConfig * oxcf,const TWO_PASS * const twopass,int * key_frame_map)3816 void vp9_get_key_frame_map(const VP9EncoderConfig *oxcf,
3817 const TWO_PASS *const twopass, int *key_frame_map) {
3818 const FIRST_PASS_INFO *first_pass_info = &twopass->first_pass_info;
3819 int show_idx = 0;
3820 RATE_CONTROL rc;
3821 vp9_rc_init(oxcf, 1, &rc);
3822
3823 // key_frame_map points to an int array with size equal to
3824 // first_pass_info->num_frames, which is also the number of show frames in the
3825 // video.
3826 memset(key_frame_map, 0,
3827 sizeof(*key_frame_map) * first_pass_info->num_frames);
3828 while (show_idx < first_pass_info->num_frames) {
3829 int key_frame_group_size;
3830 key_frame_map[show_idx] = 1;
3831 key_frame_group_size =
3832 vp9_get_frames_to_next_key(oxcf, twopass, show_idx, rc.min_gf_interval);
3833 assert(key_frame_group_size > 0);
3834 show_idx += key_frame_group_size;
3835 }
3836 assert(show_idx == first_pass_info->num_frames);
3837 }
3838 #endif // CONFIG_RATE_CTRL
3839
vp9_get_frame_stats(const TWO_PASS * twopass)3840 FIRSTPASS_STATS vp9_get_frame_stats(const TWO_PASS *twopass) {
3841 return twopass->this_frame_stats;
3842 }
vp9_get_total_stats(const TWO_PASS * twopass)3843 FIRSTPASS_STATS vp9_get_total_stats(const TWO_PASS *twopass) {
3844 return twopass->total_stats;
3845 }
3846