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 240
49 #define MIN_DECAY_FACTOR 0.01
50 #define NEW_MV_MODE_PENALTY 32
51 #define DARK_THRESH 64
52 #define SECTION_NOISE_DEF 250.0
53 #define LOW_I_THRESH 24000
54
55 #define NCOUNT_INTRA_THRESH 8192
56 #define NCOUNT_INTRA_FACTOR 3
57
58 #define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x)-0.000001 : (x) + 0.000001)
59
60 #if ARF_STATS_OUTPUT
61 unsigned int arf_count = 0;
62 #endif
63
64 // Resets the first pass file to the given position using a relative seek from
65 // the current position.
reset_fpf_position(TWO_PASS * p,const FIRSTPASS_STATS * position)66 static void reset_fpf_position(TWO_PASS *p, const FIRSTPASS_STATS *position) {
67 p->stats_in = position;
68 }
69
70 // Read frame stats at an offset from the current position.
read_frame_stats(const TWO_PASS * p,int offset)71 static const FIRSTPASS_STATS *read_frame_stats(const TWO_PASS *p, int offset) {
72 if ((offset >= 0 && p->stats_in + offset >= p->stats_in_end) ||
73 (offset < 0 && p->stats_in + offset < p->stats_in_start)) {
74 return NULL;
75 }
76
77 return &p->stats_in[offset];
78 }
79
input_stats(TWO_PASS * p,FIRSTPASS_STATS * fps)80 static int input_stats(TWO_PASS *p, FIRSTPASS_STATS *fps) {
81 if (p->stats_in >= p->stats_in_end) return EOF;
82
83 *fps = *p->stats_in;
84 ++p->stats_in;
85 return 1;
86 }
87
output_stats(FIRSTPASS_STATS * stats,struct vpx_codec_pkt_list * pktlist)88 static void output_stats(FIRSTPASS_STATS *stats,
89 struct vpx_codec_pkt_list *pktlist) {
90 struct vpx_codec_cx_pkt pkt;
91 pkt.kind = VPX_CODEC_STATS_PKT;
92 pkt.data.twopass_stats.buf = stats;
93 pkt.data.twopass_stats.sz = sizeof(FIRSTPASS_STATS);
94 vpx_codec_pkt_list_add(pktlist, &pkt);
95
96 // TEMP debug code
97 #if OUTPUT_FPF
98 {
99 FILE *fpfile;
100 fpfile = fopen("firstpass.stt", "a");
101
102 fprintf(fpfile,
103 "%12.0lf %12.4lf %12.2lf %12.2lf %12.2lf %12.0lf %12.4lf %12.4lf"
104 "%12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf"
105 "%12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.0lf %12.4lf %12.0lf"
106 "%12.4lf"
107 "\n",
108 stats->frame, stats->weight, stats->intra_error, stats->coded_error,
109 stats->sr_coded_error, stats->frame_noise_energy, stats->pcnt_inter,
110 stats->pcnt_motion, stats->pcnt_second_ref, stats->pcnt_neutral,
111 stats->pcnt_intra_low, stats->pcnt_intra_high,
112 stats->intra_skip_pct, stats->intra_smooth_pct,
113 stats->inactive_zone_rows, stats->inactive_zone_cols, stats->MVr,
114 stats->mvr_abs, stats->MVc, stats->mvc_abs, stats->MVrv,
115 stats->MVcv, stats->mv_in_out_count, stats->count, stats->duration);
116 fclose(fpfile);
117 }
118 #endif
119 }
120
121 #if CONFIG_FP_MB_STATS
output_fpmb_stats(uint8_t * this_frame_mb_stats,VP9_COMMON * cm,struct vpx_codec_pkt_list * pktlist)122 static void output_fpmb_stats(uint8_t *this_frame_mb_stats, VP9_COMMON *cm,
123 struct vpx_codec_pkt_list *pktlist) {
124 struct vpx_codec_cx_pkt pkt;
125 pkt.kind = VPX_CODEC_FPMB_STATS_PKT;
126 pkt.data.firstpass_mb_stats.buf = this_frame_mb_stats;
127 pkt.data.firstpass_mb_stats.sz = cm->initial_mbs * sizeof(uint8_t);
128 vpx_codec_pkt_list_add(pktlist, &pkt);
129 }
130 #endif
131
zero_stats(FIRSTPASS_STATS * section)132 static void zero_stats(FIRSTPASS_STATS *section) {
133 section->frame = 0.0;
134 section->weight = 0.0;
135 section->intra_error = 0.0;
136 section->coded_error = 0.0;
137 section->sr_coded_error = 0.0;
138 section->frame_noise_energy = 0.0;
139 section->pcnt_inter = 0.0;
140 section->pcnt_motion = 0.0;
141 section->pcnt_second_ref = 0.0;
142 section->pcnt_neutral = 0.0;
143 section->intra_skip_pct = 0.0;
144 section->intra_smooth_pct = 0.0;
145 section->pcnt_intra_low = 0.0;
146 section->pcnt_intra_high = 0.0;
147 section->inactive_zone_rows = 0.0;
148 section->inactive_zone_cols = 0.0;
149 section->MVr = 0.0;
150 section->mvr_abs = 0.0;
151 section->MVc = 0.0;
152 section->mvc_abs = 0.0;
153 section->MVrv = 0.0;
154 section->MVcv = 0.0;
155 section->mv_in_out_count = 0.0;
156 section->count = 0.0;
157 section->duration = 1.0;
158 section->spatial_layer_id = 0;
159 }
160
accumulate_stats(FIRSTPASS_STATS * section,const FIRSTPASS_STATS * frame)161 static void accumulate_stats(FIRSTPASS_STATS *section,
162 const FIRSTPASS_STATS *frame) {
163 section->frame += frame->frame;
164 section->weight += frame->weight;
165 section->spatial_layer_id = frame->spatial_layer_id;
166 section->intra_error += frame->intra_error;
167 section->coded_error += frame->coded_error;
168 section->sr_coded_error += frame->sr_coded_error;
169 section->frame_noise_energy += frame->frame_noise_energy;
170 section->pcnt_inter += frame->pcnt_inter;
171 section->pcnt_motion += frame->pcnt_motion;
172 section->pcnt_second_ref += frame->pcnt_second_ref;
173 section->pcnt_neutral += frame->pcnt_neutral;
174 section->intra_skip_pct += frame->intra_skip_pct;
175 section->intra_smooth_pct += frame->intra_smooth_pct;
176 section->pcnt_intra_low += frame->pcnt_intra_low;
177 section->pcnt_intra_high += frame->pcnt_intra_high;
178 section->inactive_zone_rows += frame->inactive_zone_rows;
179 section->inactive_zone_cols += frame->inactive_zone_cols;
180 section->MVr += frame->MVr;
181 section->mvr_abs += frame->mvr_abs;
182 section->MVc += frame->MVc;
183 section->mvc_abs += frame->mvc_abs;
184 section->MVrv += frame->MVrv;
185 section->MVcv += frame->MVcv;
186 section->mv_in_out_count += frame->mv_in_out_count;
187 section->count += frame->count;
188 section->duration += frame->duration;
189 }
190
subtract_stats(FIRSTPASS_STATS * section,const FIRSTPASS_STATS * frame)191 static void subtract_stats(FIRSTPASS_STATS *section,
192 const FIRSTPASS_STATS *frame) {
193 section->frame -= frame->frame;
194 section->weight -= frame->weight;
195 section->intra_error -= frame->intra_error;
196 section->coded_error -= frame->coded_error;
197 section->sr_coded_error -= frame->sr_coded_error;
198 section->frame_noise_energy -= frame->frame_noise_energy;
199 section->pcnt_inter -= frame->pcnt_inter;
200 section->pcnt_motion -= frame->pcnt_motion;
201 section->pcnt_second_ref -= frame->pcnt_second_ref;
202 section->pcnt_neutral -= frame->pcnt_neutral;
203 section->intra_skip_pct -= frame->intra_skip_pct;
204 section->intra_smooth_pct -= frame->intra_smooth_pct;
205 section->pcnt_intra_low -= frame->pcnt_intra_low;
206 section->pcnt_intra_high -= frame->pcnt_intra_high;
207 section->inactive_zone_rows -= frame->inactive_zone_rows;
208 section->inactive_zone_cols -= frame->inactive_zone_cols;
209 section->MVr -= frame->MVr;
210 section->mvr_abs -= frame->mvr_abs;
211 section->MVc -= frame->MVc;
212 section->mvc_abs -= frame->mvc_abs;
213 section->MVrv -= frame->MVrv;
214 section->MVcv -= frame->MVcv;
215 section->mv_in_out_count -= frame->mv_in_out_count;
216 section->count -= frame->count;
217 section->duration -= frame->duration;
218 }
219
220 // Calculate an active area of the image that discounts formatting
221 // bars and partially discounts other 0 energy areas.
222 #define MIN_ACTIVE_AREA 0.5
223 #define MAX_ACTIVE_AREA 1.0
calculate_active_area(const VP9_COMP * cpi,const FIRSTPASS_STATS * this_frame)224 static double calculate_active_area(const VP9_COMP *cpi,
225 const FIRSTPASS_STATS *this_frame) {
226 double active_pct;
227
228 active_pct =
229 1.0 -
230 ((this_frame->intra_skip_pct / 2) +
231 ((this_frame->inactive_zone_rows * 2) / (double)cpi->common.mb_rows));
232 return fclamp(active_pct, MIN_ACTIVE_AREA, MAX_ACTIVE_AREA);
233 }
234
235 // Get the average weighted error for the clip (or corpus)
get_distribution_av_err(VP9_COMP * cpi,TWO_PASS * const twopass)236 static double get_distribution_av_err(VP9_COMP *cpi, TWO_PASS *const twopass) {
237 const double av_weight =
238 twopass->total_stats.weight / twopass->total_stats.count;
239
240 if (cpi->oxcf.vbr_corpus_complexity)
241 return av_weight * twopass->mean_mod_score;
242 else
243 return (twopass->total_stats.coded_error * av_weight) /
244 twopass->total_stats.count;
245 }
246
247 #define ACT_AREA_CORRECTION 0.5
248 // Calculate a modified Error used in distributing bits between easier and
249 // harder frames.
calculate_mod_frame_score(const VP9_COMP * cpi,const VP9EncoderConfig * oxcf,const FIRSTPASS_STATS * this_frame,const double av_err)250 static double calculate_mod_frame_score(const VP9_COMP *cpi,
251 const VP9EncoderConfig *oxcf,
252 const FIRSTPASS_STATS *this_frame,
253 const double av_err) {
254 double modified_score =
255 av_err * pow(this_frame->coded_error * this_frame->weight /
256 DOUBLE_DIVIDE_CHECK(av_err),
257 oxcf->two_pass_vbrbias / 100.0);
258
259 // Correction for active area. Frames with a reduced active area
260 // (eg due to formatting bars) have a higher error per mb for the
261 // remaining active MBs. The correction here assumes that coding
262 // 0.5N blocks of complexity 2X is a little easier than coding N
263 // blocks of complexity X.
264 modified_score *=
265 pow(calculate_active_area(cpi, this_frame), ACT_AREA_CORRECTION);
266
267 return modified_score;
268 }
269
calculate_norm_frame_score(const VP9_COMP * cpi,const TWO_PASS * twopass,const VP9EncoderConfig * oxcf,const FIRSTPASS_STATS * this_frame,const double av_err)270 static double calculate_norm_frame_score(const VP9_COMP *cpi,
271 const TWO_PASS *twopass,
272 const VP9EncoderConfig *oxcf,
273 const FIRSTPASS_STATS *this_frame,
274 const double av_err) {
275 double modified_score =
276 av_err * pow(this_frame->coded_error * this_frame->weight /
277 DOUBLE_DIVIDE_CHECK(av_err),
278 oxcf->two_pass_vbrbias / 100.0);
279
280 const double min_score = (double)(oxcf->two_pass_vbrmin_section) / 100.0;
281 const double max_score = (double)(oxcf->two_pass_vbrmax_section) / 100.0;
282
283 // Correction for active area. Frames with a reduced active area
284 // (eg due to formatting bars) have a higher error per mb for the
285 // remaining active MBs. The correction here assumes that coding
286 // 0.5N blocks of complexity 2X is a little easier than coding N
287 // blocks of complexity X.
288 modified_score *=
289 pow(calculate_active_area(cpi, this_frame), ACT_AREA_CORRECTION);
290
291 // Normalize to a midpoint score.
292 modified_score /= DOUBLE_DIVIDE_CHECK(twopass->mean_mod_score);
293
294 return fclamp(modified_score, min_score, max_score);
295 }
296
297 // This function returns the maximum target rate per frame.
frame_max_bits(const RATE_CONTROL * rc,const VP9EncoderConfig * oxcf)298 static int frame_max_bits(const RATE_CONTROL *rc,
299 const VP9EncoderConfig *oxcf) {
300 int64_t max_bits = ((int64_t)rc->avg_frame_bandwidth *
301 (int64_t)oxcf->two_pass_vbrmax_section) /
302 100;
303 if (max_bits < 0)
304 max_bits = 0;
305 else if (max_bits > rc->max_frame_bandwidth)
306 max_bits = rc->max_frame_bandwidth;
307
308 return (int)max_bits;
309 }
310
vp9_init_first_pass(VP9_COMP * cpi)311 void vp9_init_first_pass(VP9_COMP *cpi) {
312 zero_stats(&cpi->twopass.total_stats);
313 }
314
vp9_end_first_pass(VP9_COMP * cpi)315 void vp9_end_first_pass(VP9_COMP *cpi) {
316 output_stats(&cpi->twopass.total_stats, cpi->output_pkt_list);
317 vpx_free(cpi->twopass.fp_mb_float_stats);
318 cpi->twopass.fp_mb_float_stats = NULL;
319 }
320
get_block_variance_fn(BLOCK_SIZE bsize)321 static vpx_variance_fn_t get_block_variance_fn(BLOCK_SIZE bsize) {
322 switch (bsize) {
323 case BLOCK_8X8: return vpx_mse8x8;
324 case BLOCK_16X8: return vpx_mse16x8;
325 case BLOCK_8X16: return vpx_mse8x16;
326 default: return vpx_mse16x16;
327 }
328 }
329
get_prediction_error(BLOCK_SIZE bsize,const struct buf_2d * src,const struct buf_2d * ref)330 static unsigned int get_prediction_error(BLOCK_SIZE bsize,
331 const struct buf_2d *src,
332 const struct buf_2d *ref) {
333 unsigned int sse;
334 const vpx_variance_fn_t fn = get_block_variance_fn(bsize);
335 fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
336 return sse;
337 }
338
339 #if CONFIG_VP9_HIGHBITDEPTH
highbd_get_block_variance_fn(BLOCK_SIZE bsize,int bd)340 static vpx_variance_fn_t highbd_get_block_variance_fn(BLOCK_SIZE bsize,
341 int bd) {
342 switch (bd) {
343 default:
344 switch (bsize) {
345 case BLOCK_8X8: return vpx_highbd_8_mse8x8;
346 case BLOCK_16X8: return vpx_highbd_8_mse16x8;
347 case BLOCK_8X16: return vpx_highbd_8_mse8x16;
348 default: return vpx_highbd_8_mse16x16;
349 }
350 break;
351 case 10:
352 switch (bsize) {
353 case BLOCK_8X8: return vpx_highbd_10_mse8x8;
354 case BLOCK_16X8: return vpx_highbd_10_mse16x8;
355 case BLOCK_8X16: return vpx_highbd_10_mse8x16;
356 default: return vpx_highbd_10_mse16x16;
357 }
358 break;
359 case 12:
360 switch (bsize) {
361 case BLOCK_8X8: return vpx_highbd_12_mse8x8;
362 case BLOCK_16X8: return vpx_highbd_12_mse16x8;
363 case BLOCK_8X16: return vpx_highbd_12_mse8x16;
364 default: return vpx_highbd_12_mse16x16;
365 }
366 break;
367 }
368 }
369
highbd_get_prediction_error(BLOCK_SIZE bsize,const struct buf_2d * src,const struct buf_2d * ref,int bd)370 static unsigned int highbd_get_prediction_error(BLOCK_SIZE bsize,
371 const struct buf_2d *src,
372 const struct buf_2d *ref,
373 int bd) {
374 unsigned int sse;
375 const vpx_variance_fn_t fn = highbd_get_block_variance_fn(bsize, bd);
376 fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
377 return sse;
378 }
379 #endif // CONFIG_VP9_HIGHBITDEPTH
380
381 // Refine the motion search range according to the frame dimension
382 // for first pass test.
get_search_range(const VP9_COMP * cpi)383 static int get_search_range(const VP9_COMP *cpi) {
384 int sr = 0;
385 const int dim = VPXMIN(cpi->initial_width, cpi->initial_height);
386
387 while ((dim << sr) < MAX_FULL_PEL_VAL) ++sr;
388 return sr;
389 }
390
first_pass_motion_search(VP9_COMP * cpi,MACROBLOCK * x,const MV * ref_mv,MV * best_mv,int * best_motion_err)391 static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
392 const MV *ref_mv, MV *best_mv,
393 int *best_motion_err) {
394 MACROBLOCKD *const xd = &x->e_mbd;
395 MV tmp_mv = { 0, 0 };
396 MV ref_mv_full = { ref_mv->row >> 3, ref_mv->col >> 3 };
397 int num00, tmp_err, n;
398 const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
399 vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
400 const int new_mv_mode_penalty = NEW_MV_MODE_PENALTY;
401
402 int step_param = 3;
403 int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
404 const int sr = get_search_range(cpi);
405 step_param += sr;
406 further_steps -= sr;
407
408 // Override the default variance function to use MSE.
409 v_fn_ptr.vf = get_block_variance_fn(bsize);
410 #if CONFIG_VP9_HIGHBITDEPTH
411 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
412 v_fn_ptr.vf = highbd_get_block_variance_fn(bsize, xd->bd);
413 }
414 #endif // CONFIG_VP9_HIGHBITDEPTH
415
416 // Center the initial step/diamond search on best mv.
417 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
418 step_param, x->sadperbit16, &num00,
419 &v_fn_ptr, ref_mv);
420 if (tmp_err < INT_MAX)
421 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
422 if (tmp_err < INT_MAX - new_mv_mode_penalty) tmp_err += new_mv_mode_penalty;
423
424 if (tmp_err < *best_motion_err) {
425 *best_motion_err = tmp_err;
426 *best_mv = tmp_mv;
427 }
428
429 // Carry out further step/diamond searches as necessary.
430 n = num00;
431 num00 = 0;
432
433 while (n < further_steps) {
434 ++n;
435
436 if (num00) {
437 --num00;
438 } else {
439 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
440 step_param + n, x->sadperbit16, &num00,
441 &v_fn_ptr, ref_mv);
442 if (tmp_err < INT_MAX)
443 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
444 if (tmp_err < INT_MAX - new_mv_mode_penalty)
445 tmp_err += new_mv_mode_penalty;
446
447 if (tmp_err < *best_motion_err) {
448 *best_motion_err = tmp_err;
449 *best_mv = tmp_mv;
450 }
451 }
452 }
453 }
454
get_bsize(const VP9_COMMON * cm,int mb_row,int mb_col)455 static BLOCK_SIZE get_bsize(const VP9_COMMON *cm, int mb_row, int mb_col) {
456 if (2 * mb_col + 1 < cm->mi_cols) {
457 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_16X16 : BLOCK_16X8;
458 } else {
459 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_8X16 : BLOCK_8X8;
460 }
461 }
462
find_fp_qindex(vpx_bit_depth_t bit_depth)463 static int find_fp_qindex(vpx_bit_depth_t bit_depth) {
464 int i;
465
466 for (i = 0; i < QINDEX_RANGE; ++i)
467 if (vp9_convert_qindex_to_q(i, bit_depth) >= FIRST_PASS_Q) break;
468
469 if (i == QINDEX_RANGE) i--;
470
471 return i;
472 }
473
set_first_pass_params(VP9_COMP * cpi)474 static void set_first_pass_params(VP9_COMP *cpi) {
475 VP9_COMMON *const cm = &cpi->common;
476 if (!cpi->refresh_alt_ref_frame &&
477 (cm->current_video_frame == 0 || (cpi->frame_flags & FRAMEFLAGS_KEY))) {
478 cm->frame_type = KEY_FRAME;
479 } else {
480 cm->frame_type = INTER_FRAME;
481 }
482 // Do not use periodic key frames.
483 cpi->rc.frames_to_key = INT_MAX;
484 }
485
486 // Scale an sse threshold to account for 8/10/12 bit.
scale_sse_threshold(VP9_COMMON * cm,int thresh)487 static int scale_sse_threshold(VP9_COMMON *cm, int thresh) {
488 int ret_val = thresh;
489 #if CONFIG_VP9_HIGHBITDEPTH
490 if (cm->use_highbitdepth) {
491 switch (cm->bit_depth) {
492 case VPX_BITS_8: ret_val = thresh; break;
493 case VPX_BITS_10: ret_val = thresh << 4; break;
494 default:
495 assert(cm->bit_depth == VPX_BITS_12);
496 ret_val = thresh << 8;
497 break;
498 }
499 }
500 #else
501 (void)cm;
502 #endif // CONFIG_VP9_HIGHBITDEPTH
503 return ret_val;
504 }
505
506 // This threshold is used to track blocks where to all intents and purposes
507 // the intra prediction error 0. Though the metric we test against
508 // is technically a sse we are mainly interested in blocks where all the pixels
509 // in the 8 bit domain have an error of <= 1 (where error = sse) so a
510 // linear scaling for 10 and 12 bit gives similar results.
511 #define UL_INTRA_THRESH 50
get_ul_intra_threshold(VP9_COMMON * cm)512 static int get_ul_intra_threshold(VP9_COMMON *cm) {
513 int ret_val = UL_INTRA_THRESH;
514 #if CONFIG_VP9_HIGHBITDEPTH
515 if (cm->use_highbitdepth) {
516 switch (cm->bit_depth) {
517 case VPX_BITS_8: ret_val = UL_INTRA_THRESH; break;
518 case VPX_BITS_10: ret_val = UL_INTRA_THRESH << 2; break;
519 default:
520 assert(cm->bit_depth == VPX_BITS_12);
521 ret_val = UL_INTRA_THRESH << 4;
522 break;
523 }
524 }
525 #else
526 (void)cm;
527 #endif // CONFIG_VP9_HIGHBITDEPTH
528 return ret_val;
529 }
530
531 #define SMOOTH_INTRA_THRESH 4000
get_smooth_intra_threshold(VP9_COMMON * cm)532 static int get_smooth_intra_threshold(VP9_COMMON *cm) {
533 int ret_val = SMOOTH_INTRA_THRESH;
534 #if CONFIG_VP9_HIGHBITDEPTH
535 if (cm->use_highbitdepth) {
536 switch (cm->bit_depth) {
537 case VPX_BITS_8: ret_val = SMOOTH_INTRA_THRESH; break;
538 case VPX_BITS_10: ret_val = SMOOTH_INTRA_THRESH << 4; break;
539 default:
540 assert(cm->bit_depth == VPX_BITS_12);
541 ret_val = SMOOTH_INTRA_THRESH << 8;
542 break;
543 }
544 }
545 #else
546 (void)cm;
547 #endif // CONFIG_VP9_HIGHBITDEPTH
548 return ret_val;
549 }
550
551 #define FP_DN_THRESH 8
552 #define FP_MAX_DN_THRESH 16
553 #define KERNEL_SIZE 3
554
555 // Baseline Kernal weights for first pass noise metric
556 static uint8_t fp_dn_kernal_3[KERNEL_SIZE * KERNEL_SIZE] = { 1, 2, 1, 2, 4,
557 2, 1, 2, 1 };
558
559 // Estimate noise at a single point based on the impace of a spatial kernal
560 // on the point value
fp_estimate_point_noise(uint8_t * src_ptr,const int stride)561 static int fp_estimate_point_noise(uint8_t *src_ptr, const int stride) {
562 int sum_weight = 0;
563 int sum_val = 0;
564 int i, j;
565 int max_diff = 0;
566 int diff;
567 int dn_diff;
568 uint8_t *tmp_ptr;
569 uint8_t *kernal_ptr;
570 uint8_t dn_val;
571 uint8_t centre_val = *src_ptr;
572
573 kernal_ptr = fp_dn_kernal_3;
574
575 // Apply the kernal
576 tmp_ptr = src_ptr - stride - 1;
577 for (i = 0; i < KERNEL_SIZE; ++i) {
578 for (j = 0; j < KERNEL_SIZE; ++j) {
579 diff = abs((int)centre_val - (int)tmp_ptr[j]);
580 max_diff = VPXMAX(max_diff, diff);
581 if (diff <= FP_DN_THRESH) {
582 sum_weight += *kernal_ptr;
583 sum_val += (int)tmp_ptr[j] * (int)*kernal_ptr;
584 }
585 ++kernal_ptr;
586 }
587 tmp_ptr += stride;
588 }
589
590 if (max_diff < FP_MAX_DN_THRESH)
591 // Update the source value with the new filtered value
592 dn_val = (sum_val + (sum_weight >> 1)) / sum_weight;
593 else
594 dn_val = *src_ptr;
595
596 // return the noise energy as the square of the difference between the
597 // denoised and raw value.
598 dn_diff = (int)*src_ptr - (int)dn_val;
599 return dn_diff * dn_diff;
600 }
601 #if CONFIG_VP9_HIGHBITDEPTH
fp_highbd_estimate_point_noise(uint8_t * src_ptr,const int stride)602 static int fp_highbd_estimate_point_noise(uint8_t *src_ptr, const int stride) {
603 int sum_weight = 0;
604 int sum_val = 0;
605 int i, j;
606 int max_diff = 0;
607 int diff;
608 int dn_diff;
609 uint8_t *tmp_ptr;
610 uint16_t *tmp_ptr16;
611 uint8_t *kernal_ptr;
612 uint16_t dn_val;
613 uint16_t centre_val = *CONVERT_TO_SHORTPTR(src_ptr);
614
615 kernal_ptr = fp_dn_kernal_3;
616
617 // Apply the kernal
618 tmp_ptr = src_ptr - stride - 1;
619 for (i = 0; i < KERNEL_SIZE; ++i) {
620 tmp_ptr16 = CONVERT_TO_SHORTPTR(tmp_ptr);
621 for (j = 0; j < KERNEL_SIZE; ++j) {
622 diff = abs((int)centre_val - (int)tmp_ptr16[j]);
623 max_diff = VPXMAX(max_diff, diff);
624 if (diff <= FP_DN_THRESH) {
625 sum_weight += *kernal_ptr;
626 sum_val += (int)tmp_ptr16[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 = *CONVERT_TO_SHORTPTR(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)(*CONVERT_TO_SHORTPTR(src_ptr)) - (int)dn_val;
642 return dn_diff * dn_diff;
643 }
644 #endif
645
646 // Estimate noise for a block.
fp_estimate_block_noise(MACROBLOCK * x,BLOCK_SIZE bsize)647 static int fp_estimate_block_noise(MACROBLOCK *x, BLOCK_SIZE bsize) {
648 #if CONFIG_VP9_HIGHBITDEPTH
649 MACROBLOCKD *xd = &x->e_mbd;
650 #endif
651 uint8_t *src_ptr = &x->plane[0].src.buf[0];
652 const int width = num_4x4_blocks_wide_lookup[bsize] * 4;
653 const int height = num_4x4_blocks_high_lookup[bsize] * 4;
654 int w, h;
655 int stride = x->plane[0].src.stride;
656 int block_noise = 0;
657
658 // Sampled points to reduce cost overhead.
659 for (h = 0; h < height; h += 2) {
660 for (w = 0; w < width; w += 2) {
661 #if CONFIG_VP9_HIGHBITDEPTH
662 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
663 block_noise += fp_highbd_estimate_point_noise(src_ptr, stride);
664 else
665 block_noise += fp_estimate_point_noise(src_ptr, stride);
666 #else
667 block_noise += fp_estimate_point_noise(src_ptr, stride);
668 #endif
669 ++src_ptr;
670 }
671 src_ptr += (stride - width);
672 }
673 return block_noise << 2; // Scale << 2 to account for sampling.
674 }
675
676 // This function is called to test the functionality of row based
677 // multi-threading in unit tests for bit-exactness
accumulate_floating_point_stats(VP9_COMP * cpi,TileDataEnc * first_tile_col)678 static void accumulate_floating_point_stats(VP9_COMP *cpi,
679 TileDataEnc *first_tile_col) {
680 VP9_COMMON *const cm = &cpi->common;
681 int mb_row, mb_col;
682 first_tile_col->fp_data.intra_factor = 0;
683 first_tile_col->fp_data.brightness_factor = 0;
684 first_tile_col->fp_data.neutral_count = 0;
685 for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
686 for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
687 const int mb_index = mb_row * cm->mb_cols + mb_col;
688 first_tile_col->fp_data.intra_factor +=
689 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_intra_factor;
690 first_tile_col->fp_data.brightness_factor +=
691 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_brightness_factor;
692 first_tile_col->fp_data.neutral_count +=
693 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_neutral_count;
694 }
695 }
696 }
697
first_pass_stat_calc(VP9_COMP * cpi,FIRSTPASS_STATS * fps,FIRSTPASS_DATA * fp_acc_data)698 static void first_pass_stat_calc(VP9_COMP *cpi, FIRSTPASS_STATS *fps,
699 FIRSTPASS_DATA *fp_acc_data) {
700 VP9_COMMON *const cm = &cpi->common;
701 // The minimum error here insures some bit allocation to frames even
702 // in static regions. The allocation per MB declines for larger formats
703 // where the typical "real" energy per MB also falls.
704 // Initial estimate here uses sqrt(mbs) to define the min_err, where the
705 // number of mbs is proportional to the image area.
706 const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE) ? cpi->initial_mbs
707 : cpi->common.MBs;
708 const double min_err = 200 * sqrt(num_mbs);
709
710 // Clamp the image start to rows/2. This number of rows is discarded top
711 // and bottom as dead data so rows / 2 means the frame is blank.
712 if ((fp_acc_data->image_data_start_row > cm->mb_rows / 2) ||
713 (fp_acc_data->image_data_start_row == INVALID_ROW)) {
714 fp_acc_data->image_data_start_row = cm->mb_rows / 2;
715 }
716 // Exclude any image dead zone
717 if (fp_acc_data->image_data_start_row > 0) {
718 fp_acc_data->intra_skip_count =
719 VPXMAX(0, fp_acc_data->intra_skip_count -
720 (fp_acc_data->image_data_start_row * cm->mb_cols * 2));
721 }
722
723 fp_acc_data->intra_factor = fp_acc_data->intra_factor / (double)num_mbs;
724 fp_acc_data->brightness_factor =
725 fp_acc_data->brightness_factor / (double)num_mbs;
726 fps->weight = fp_acc_data->intra_factor * fp_acc_data->brightness_factor;
727
728 fps->frame = cm->current_video_frame;
729 fps->spatial_layer_id = cpi->svc.spatial_layer_id;
730
731 fps->coded_error =
732 ((double)(fp_acc_data->coded_error >> 8) + min_err) / num_mbs;
733 fps->sr_coded_error =
734 ((double)(fp_acc_data->sr_coded_error >> 8) + min_err) / num_mbs;
735 fps->intra_error =
736 ((double)(fp_acc_data->intra_error >> 8) + min_err) / num_mbs;
737
738 fps->frame_noise_energy =
739 (double)(fp_acc_data->frame_noise_energy) / (double)num_mbs;
740 fps->count = 1.0;
741 fps->pcnt_inter = (double)(fp_acc_data->intercount) / num_mbs;
742 fps->pcnt_second_ref = (double)(fp_acc_data->second_ref_count) / num_mbs;
743 fps->pcnt_neutral = (double)(fp_acc_data->neutral_count) / num_mbs;
744 fps->pcnt_intra_low = (double)(fp_acc_data->intra_count_low) / num_mbs;
745 fps->pcnt_intra_high = (double)(fp_acc_data->intra_count_high) / num_mbs;
746 fps->intra_skip_pct = (double)(fp_acc_data->intra_skip_count) / num_mbs;
747 fps->intra_smooth_pct = (double)(fp_acc_data->intra_smooth_count) / num_mbs;
748 fps->inactive_zone_rows = (double)(fp_acc_data->image_data_start_row);
749 // Currently set to 0 as most issues relate to letter boxing.
750 fps->inactive_zone_cols = (double)0;
751
752 if (fp_acc_data->mvcount > 0) {
753 fps->MVr = (double)(fp_acc_data->sum_mvr) / fp_acc_data->mvcount;
754 fps->mvr_abs = (double)(fp_acc_data->sum_mvr_abs) / fp_acc_data->mvcount;
755 fps->MVc = (double)(fp_acc_data->sum_mvc) / fp_acc_data->mvcount;
756 fps->mvc_abs = (double)(fp_acc_data->sum_mvc_abs) / fp_acc_data->mvcount;
757 fps->MVrv = ((double)(fp_acc_data->sum_mvrs) -
758 ((double)(fp_acc_data->sum_mvr) * (fp_acc_data->sum_mvr) /
759 fp_acc_data->mvcount)) /
760 fp_acc_data->mvcount;
761 fps->MVcv = ((double)(fp_acc_data->sum_mvcs) -
762 ((double)(fp_acc_data->sum_mvc) * (fp_acc_data->sum_mvc) /
763 fp_acc_data->mvcount)) /
764 fp_acc_data->mvcount;
765 fps->mv_in_out_count =
766 (double)(fp_acc_data->sum_in_vectors) / (fp_acc_data->mvcount * 2);
767 fps->pcnt_motion = (double)(fp_acc_data->mvcount) / num_mbs;
768 } else {
769 fps->MVr = 0.0;
770 fps->mvr_abs = 0.0;
771 fps->MVc = 0.0;
772 fps->mvc_abs = 0.0;
773 fps->MVrv = 0.0;
774 fps->MVcv = 0.0;
775 fps->mv_in_out_count = 0.0;
776 fps->pcnt_motion = 0.0;
777 }
778 }
779
accumulate_fp_mb_row_stat(TileDataEnc * this_tile,FIRSTPASS_DATA * fp_acc_data)780 static void accumulate_fp_mb_row_stat(TileDataEnc *this_tile,
781 FIRSTPASS_DATA *fp_acc_data) {
782 this_tile->fp_data.intra_factor += fp_acc_data->intra_factor;
783 this_tile->fp_data.brightness_factor += fp_acc_data->brightness_factor;
784 this_tile->fp_data.coded_error += fp_acc_data->coded_error;
785 this_tile->fp_data.sr_coded_error += fp_acc_data->sr_coded_error;
786 this_tile->fp_data.frame_noise_energy += fp_acc_data->frame_noise_energy;
787 this_tile->fp_data.intra_error += fp_acc_data->intra_error;
788 this_tile->fp_data.intercount += fp_acc_data->intercount;
789 this_tile->fp_data.second_ref_count += fp_acc_data->second_ref_count;
790 this_tile->fp_data.neutral_count += fp_acc_data->neutral_count;
791 this_tile->fp_data.intra_count_low += fp_acc_data->intra_count_low;
792 this_tile->fp_data.intra_count_high += fp_acc_data->intra_count_high;
793 this_tile->fp_data.intra_skip_count += fp_acc_data->intra_skip_count;
794 this_tile->fp_data.mvcount += fp_acc_data->mvcount;
795 this_tile->fp_data.sum_mvr += fp_acc_data->sum_mvr;
796 this_tile->fp_data.sum_mvr_abs += fp_acc_data->sum_mvr_abs;
797 this_tile->fp_data.sum_mvc += fp_acc_data->sum_mvc;
798 this_tile->fp_data.sum_mvc_abs += fp_acc_data->sum_mvc_abs;
799 this_tile->fp_data.sum_mvrs += fp_acc_data->sum_mvrs;
800 this_tile->fp_data.sum_mvcs += fp_acc_data->sum_mvcs;
801 this_tile->fp_data.sum_in_vectors += fp_acc_data->sum_in_vectors;
802 this_tile->fp_data.intra_smooth_count += fp_acc_data->intra_smooth_count;
803 this_tile->fp_data.image_data_start_row =
804 VPXMIN(this_tile->fp_data.image_data_start_row,
805 fp_acc_data->image_data_start_row) == INVALID_ROW
806 ? VPXMAX(this_tile->fp_data.image_data_start_row,
807 fp_acc_data->image_data_start_row)
808 : VPXMIN(this_tile->fp_data.image_data_start_row,
809 fp_acc_data->image_data_start_row);
810 }
811
812 #define NZ_MOTION_PENALTY 128
813 #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)814 void vp9_first_pass_encode_tile_mb_row(VP9_COMP *cpi, ThreadData *td,
815 FIRSTPASS_DATA *fp_acc_data,
816 TileDataEnc *tile_data, MV *best_ref_mv,
817 int mb_row) {
818 int mb_col;
819 MACROBLOCK *const x = &td->mb;
820 VP9_COMMON *const cm = &cpi->common;
821 MACROBLOCKD *const xd = &x->e_mbd;
822 TileInfo tile = tile_data->tile_info;
823 const int mb_col_start = ROUND_POWER_OF_TWO(tile.mi_col_start, 1);
824 const int mb_col_end = ROUND_POWER_OF_TWO(tile.mi_col_end, 1);
825 struct macroblock_plane *const p = x->plane;
826 struct macroblockd_plane *const pd = xd->plane;
827 const PICK_MODE_CONTEXT *ctx = &td->pc_root->none;
828 int i, c;
829 int num_mb_cols = get_num_cols(tile_data->tile_info, 1);
830
831 int recon_yoffset, recon_uvoffset;
832 const int intrapenalty = INTRA_MODE_PENALTY;
833 const MV zero_mv = { 0, 0 };
834 int recon_y_stride, recon_uv_stride, uv_mb_height;
835
836 YV12_BUFFER_CONFIG *const lst_yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
837 YV12_BUFFER_CONFIG *gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
838 YV12_BUFFER_CONFIG *const new_yv12 = get_frame_new_buffer(cm);
839 const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;
840
841 MODE_INFO mi_above, mi_left;
842
843 double mb_intra_factor;
844 double mb_brightness_factor;
845 double mb_neutral_count;
846
847 // First pass code requires valid last and new frame buffers.
848 assert(new_yv12 != NULL);
849 assert(frame_is_intra_only(cm) || (lst_yv12 != NULL));
850
851 xd->mi = cm->mi_grid_visible + xd->mi_stride * (mb_row << 1) + mb_col_start;
852 xd->mi[0] = cm->mi + xd->mi_stride * (mb_row << 1) + mb_col_start;
853
854 for (i = 0; i < MAX_MB_PLANE; ++i) {
855 p[i].coeff = ctx->coeff_pbuf[i][1];
856 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
857 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
858 p[i].eobs = ctx->eobs_pbuf[i][1];
859 }
860
861 recon_y_stride = new_yv12->y_stride;
862 recon_uv_stride = new_yv12->uv_stride;
863 uv_mb_height = 16 >> (new_yv12->y_height > new_yv12->uv_height);
864
865 // Reset above block coeffs.
866 recon_yoffset = (mb_row * recon_y_stride * 16) + mb_col_start * 16;
867 recon_uvoffset =
868 (mb_row * recon_uv_stride * uv_mb_height) + mb_col_start * uv_mb_height;
869
870 // Set up limit values for motion vectors to prevent them extending
871 // outside the UMV borders.
872 x->mv_limits.row_min = -((mb_row * 16) + BORDER_MV_PIXELS_B16);
873 x->mv_limits.row_max =
874 ((cm->mb_rows - 1 - mb_row) * 16) + BORDER_MV_PIXELS_B16;
875
876 for (mb_col = mb_col_start, c = 0; mb_col < mb_col_end; ++mb_col, c++) {
877 int this_error;
878 int this_intra_error;
879 const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
880 const BLOCK_SIZE bsize = get_bsize(cm, mb_row, mb_col);
881 double log_intra;
882 int level_sample;
883 const int mb_index = mb_row * cm->mb_cols + mb_col;
884
885 #if CONFIG_FP_MB_STATS
886 const int mb_index = mb_row * cm->mb_cols + mb_col;
887 #endif
888
889 (*(cpi->row_mt_sync_read_ptr))(&tile_data->row_mt_sync, mb_row, c);
890
891 // Adjust to the next column of MBs.
892 x->plane[0].src.buf = cpi->Source->y_buffer +
893 mb_row * 16 * x->plane[0].src.stride + mb_col * 16;
894 x->plane[1].src.buf = cpi->Source->u_buffer +
895 mb_row * uv_mb_height * x->plane[1].src.stride +
896 mb_col * uv_mb_height;
897 x->plane[2].src.buf = cpi->Source->v_buffer +
898 mb_row * uv_mb_height * x->plane[1].src.stride +
899 mb_col * uv_mb_height;
900
901 vpx_clear_system_state();
902
903 xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
904 xd->plane[1].dst.buf = new_yv12->u_buffer + recon_uvoffset;
905 xd->plane[2].dst.buf = new_yv12->v_buffer + recon_uvoffset;
906 xd->mi[0]->sb_type = bsize;
907 xd->mi[0]->ref_frame[0] = INTRA_FRAME;
908 set_mi_row_col(xd, &tile, mb_row << 1, num_8x8_blocks_high_lookup[bsize],
909 mb_col << 1, num_8x8_blocks_wide_lookup[bsize], cm->mi_rows,
910 cm->mi_cols);
911 // Are edges available for intra prediction?
912 // Since the firstpass does not populate the mi_grid_visible,
913 // above_mi/left_mi must be overwritten with a nonzero value when edges
914 // are available. Required by vp9_predict_intra_block().
915 xd->above_mi = (mb_row != 0) ? &mi_above : NULL;
916 xd->left_mi = ((mb_col << 1) > tile.mi_col_start) ? &mi_left : NULL;
917
918 // Do intra 16x16 prediction.
919 x->skip_encode = 0;
920 x->fp_src_pred = 0;
921 // Do intra prediction based on source pixels for tile boundaries
922 if (mb_col == mb_col_start && mb_col != 0) {
923 xd->left_mi = &mi_left;
924 x->fp_src_pred = 1;
925 }
926 xd->mi[0]->mode = DC_PRED;
927 xd->mi[0]->tx_size =
928 use_dc_pred ? (bsize >= BLOCK_16X16 ? TX_16X16 : TX_8X8) : TX_4X4;
929 // Fix - zero the 16x16 block first. This ensures correct this_error for
930 // block sizes smaller than 16x16.
931 vp9_zero_array(x->plane[0].src_diff, 256);
932 vp9_encode_intra_block_plane(x, bsize, 0, 0);
933 this_error = vpx_get_mb_ss(x->plane[0].src_diff);
934 this_intra_error = this_error;
935
936 // Keep a record of blocks that have very low intra error residual
937 // (i.e. are in effect completely flat and untextured in the intra
938 // domain). In natural videos this is uncommon, but it is much more
939 // common in animations, graphics and screen content, so may be used
940 // as a signal to detect these types of content.
941 if (this_error < get_ul_intra_threshold(cm)) {
942 ++(fp_acc_data->intra_skip_count);
943 } else if ((mb_col > 0) &&
944 (fp_acc_data->image_data_start_row == INVALID_ROW)) {
945 fp_acc_data->image_data_start_row = mb_row;
946 }
947
948 // Blocks that are mainly smooth in the intra domain.
949 // Some special accounting for CQ but also these are better for testing
950 // noise levels.
951 if (this_error < get_smooth_intra_threshold(cm)) {
952 ++(fp_acc_data->intra_smooth_count);
953 }
954
955 // Special case noise measurement for first frame.
956 if (cm->current_video_frame == 0) {
957 if (this_intra_error < scale_sse_threshold(cm, LOW_I_THRESH)) {
958 fp_acc_data->frame_noise_energy += fp_estimate_block_noise(x, bsize);
959 } else {
960 fp_acc_data->frame_noise_energy += (int64_t)SECTION_NOISE_DEF;
961 }
962 }
963
964 #if CONFIG_VP9_HIGHBITDEPTH
965 if (cm->use_highbitdepth) {
966 switch (cm->bit_depth) {
967 case VPX_BITS_8: break;
968 case VPX_BITS_10: this_error >>= 4; break;
969 default:
970 assert(cm->bit_depth == VPX_BITS_12);
971 this_error >>= 8;
972 break;
973 }
974 }
975 #endif // CONFIG_VP9_HIGHBITDEPTH
976
977 vpx_clear_system_state();
978 log_intra = log(this_error + 1.0);
979 if (log_intra < 10.0) {
980 mb_intra_factor = 1.0 + ((10.0 - log_intra) * 0.05);
981 fp_acc_data->intra_factor += mb_intra_factor;
982 if (cpi->row_mt_bit_exact)
983 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_intra_factor =
984 mb_intra_factor;
985 } else {
986 fp_acc_data->intra_factor += 1.0;
987 if (cpi->row_mt_bit_exact)
988 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_intra_factor = 1.0;
989 }
990
991 #if CONFIG_VP9_HIGHBITDEPTH
992 if (cm->use_highbitdepth)
993 level_sample = CONVERT_TO_SHORTPTR(x->plane[0].src.buf)[0];
994 else
995 level_sample = x->plane[0].src.buf[0];
996 #else
997 level_sample = x->plane[0].src.buf[0];
998 #endif
999 if ((level_sample < DARK_THRESH) && (log_intra < 9.0)) {
1000 mb_brightness_factor = 1.0 + (0.01 * (DARK_THRESH - level_sample));
1001 fp_acc_data->brightness_factor += mb_brightness_factor;
1002 if (cpi->row_mt_bit_exact)
1003 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_brightness_factor =
1004 mb_brightness_factor;
1005 } else {
1006 fp_acc_data->brightness_factor += 1.0;
1007 if (cpi->row_mt_bit_exact)
1008 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_brightness_factor =
1009 1.0;
1010 }
1011
1012 // Intrapenalty below deals with situations where the intra and inter
1013 // error scores are very low (e.g. a plain black frame).
1014 // We do not have special cases in first pass for 0,0 and nearest etc so
1015 // all inter modes carry an overhead cost estimate for the mv.
1016 // When the error score is very low this causes us to pick all or lots of
1017 // INTRA modes and throw lots of key frames.
1018 // This penalty adds a cost matching that of a 0,0 mv to the intra case.
1019 this_error += intrapenalty;
1020
1021 // Accumulate the intra error.
1022 fp_acc_data->intra_error += (int64_t)this_error;
1023
1024 #if CONFIG_FP_MB_STATS
1025 if (cpi->use_fp_mb_stats) {
1026 // initialization
1027 cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
1028 }
1029 #endif
1030
1031 // Set up limit values for motion vectors to prevent them extending
1032 // outside the UMV borders.
1033 x->mv_limits.col_min = -((mb_col * 16) + BORDER_MV_PIXELS_B16);
1034 x->mv_limits.col_max =
1035 ((cm->mb_cols - 1 - mb_col) * 16) + BORDER_MV_PIXELS_B16;
1036
1037 // Other than for the first frame do a motion search.
1038 if (cm->current_video_frame > 0) {
1039 int tmp_err, motion_error, this_motion_error, raw_motion_error;
1040 // Assume 0,0 motion with no mv overhead.
1041 MV mv = { 0, 0 }, tmp_mv = { 0, 0 };
1042 struct buf_2d unscaled_last_source_buf_2d;
1043 vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
1044
1045 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
1046 #if CONFIG_VP9_HIGHBITDEPTH
1047 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1048 motion_error = highbd_get_prediction_error(
1049 bsize, &x->plane[0].src, &xd->plane[0].pre[0], xd->bd);
1050 this_motion_error = highbd_get_prediction_error(
1051 bsize, &x->plane[0].src, &xd->plane[0].pre[0], 8);
1052 } else {
1053 motion_error =
1054 get_prediction_error(bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
1055 this_motion_error = motion_error;
1056 }
1057 #else
1058 motion_error =
1059 get_prediction_error(bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
1060 this_motion_error = motion_error;
1061 #endif // CONFIG_VP9_HIGHBITDEPTH
1062
1063 // Compute the motion error of the 0,0 motion using the last source
1064 // frame as the reference. Skip the further motion search on
1065 // reconstructed frame if this error is very small.
1066 unscaled_last_source_buf_2d.buf =
1067 cpi->unscaled_last_source->y_buffer + recon_yoffset;
1068 unscaled_last_source_buf_2d.stride = cpi->unscaled_last_source->y_stride;
1069 #if CONFIG_VP9_HIGHBITDEPTH
1070 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1071 raw_motion_error = highbd_get_prediction_error(
1072 bsize, &x->plane[0].src, &unscaled_last_source_buf_2d, xd->bd);
1073 } else {
1074 raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
1075 &unscaled_last_source_buf_2d);
1076 }
1077 #else
1078 raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
1079 &unscaled_last_source_buf_2d);
1080 #endif // CONFIG_VP9_HIGHBITDEPTH
1081
1082 if (raw_motion_error > NZ_MOTION_PENALTY) {
1083 // Test last reference frame using the previous best mv as the
1084 // starting point (best reference) for the search.
1085 first_pass_motion_search(cpi, x, best_ref_mv, &mv, &motion_error);
1086
1087 v_fn_ptr.vf = get_block_variance_fn(bsize);
1088 #if CONFIG_VP9_HIGHBITDEPTH
1089 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1090 v_fn_ptr.vf = highbd_get_block_variance_fn(bsize, 8);
1091 }
1092 #endif // CONFIG_VP9_HIGHBITDEPTH
1093 this_motion_error =
1094 vp9_get_mvpred_var(x, &mv, best_ref_mv, &v_fn_ptr, 0);
1095
1096 // If the current best reference mv is not centered on 0,0 then do a
1097 // 0,0 based search as well.
1098 if (!is_zero_mv(best_ref_mv)) {
1099 tmp_err = INT_MAX;
1100 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv, &tmp_err);
1101
1102 if (tmp_err < motion_error) {
1103 motion_error = tmp_err;
1104 mv = tmp_mv;
1105 this_motion_error =
1106 vp9_get_mvpred_var(x, &tmp_mv, &zero_mv, &v_fn_ptr, 0);
1107 }
1108 }
1109
1110 // Search in an older reference frame.
1111 if ((cm->current_video_frame > 1) && gld_yv12 != NULL) {
1112 // Assume 0,0 motion with no mv overhead.
1113 int gf_motion_error;
1114
1115 xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
1116 #if CONFIG_VP9_HIGHBITDEPTH
1117 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1118 gf_motion_error = highbd_get_prediction_error(
1119 bsize, &x->plane[0].src, &xd->plane[0].pre[0], xd->bd);
1120 } else {
1121 gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
1122 &xd->plane[0].pre[0]);
1123 }
1124 #else
1125 gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
1126 &xd->plane[0].pre[0]);
1127 #endif // CONFIG_VP9_HIGHBITDEPTH
1128
1129 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv, &gf_motion_error);
1130
1131 if (gf_motion_error < motion_error && gf_motion_error < this_error)
1132 ++(fp_acc_data->second_ref_count);
1133
1134 // Reset to last frame as reference buffer.
1135 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
1136 xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
1137 xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
1138
1139 // In accumulating a score for the older reference frame take the
1140 // best of the motion predicted score and the intra coded error
1141 // (just as will be done for) accumulation of "coded_error" for
1142 // the last frame.
1143 if (gf_motion_error < this_error)
1144 fp_acc_data->sr_coded_error += gf_motion_error;
1145 else
1146 fp_acc_data->sr_coded_error += this_error;
1147 } else {
1148 fp_acc_data->sr_coded_error += motion_error;
1149 }
1150 } else {
1151 fp_acc_data->sr_coded_error += motion_error;
1152 }
1153
1154 // Start by assuming that intra mode is best.
1155 best_ref_mv->row = 0;
1156 best_ref_mv->col = 0;
1157
1158 #if CONFIG_FP_MB_STATS
1159 if (cpi->use_fp_mb_stats) {
1160 // intra prediction statistics
1161 cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
1162 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_DCINTRA_MASK;
1163 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_MOTION_ZERO_MASK;
1164 if (this_error > FPMB_ERROR_LARGE_TH) {
1165 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_LARGE_MASK;
1166 } else if (this_error < FPMB_ERROR_SMALL_TH) {
1167 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_SMALL_MASK;
1168 }
1169 }
1170 #endif
1171
1172 if (motion_error <= this_error) {
1173 vpx_clear_system_state();
1174
1175 // Keep a count of cases where the inter and intra were very close
1176 // and very low. This helps with scene cut detection for example in
1177 // cropped clips with black bars at the sides or top and bottom.
1178 if (((this_error - intrapenalty) * 9 <= motion_error * 10) &&
1179 (this_error < (2 * intrapenalty))) {
1180 fp_acc_data->neutral_count += 1.0;
1181 if (cpi->row_mt_bit_exact)
1182 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_neutral_count =
1183 1.0;
1184 // Also track cases where the intra is not much worse than the inter
1185 // and use this in limiting the GF/arf group length.
1186 } else if ((this_error > NCOUNT_INTRA_THRESH) &&
1187 (this_error < (NCOUNT_INTRA_FACTOR * motion_error))) {
1188 mb_neutral_count =
1189 (double)motion_error / DOUBLE_DIVIDE_CHECK((double)this_error);
1190 fp_acc_data->neutral_count += mb_neutral_count;
1191 if (cpi->row_mt_bit_exact)
1192 cpi->twopass.fp_mb_float_stats[mb_index].frame_mb_neutral_count =
1193 mb_neutral_count;
1194 }
1195
1196 mv.row *= 8;
1197 mv.col *= 8;
1198 this_error = motion_error;
1199 xd->mi[0]->mode = NEWMV;
1200 xd->mi[0]->mv[0].as_mv = mv;
1201 xd->mi[0]->tx_size = TX_4X4;
1202 xd->mi[0]->ref_frame[0] = LAST_FRAME;
1203 xd->mi[0]->ref_frame[1] = NONE;
1204 vp9_build_inter_predictors_sby(xd, mb_row << 1, mb_col << 1, bsize);
1205 vp9_encode_sby_pass1(x, bsize);
1206 fp_acc_data->sum_mvr += mv.row;
1207 fp_acc_data->sum_mvr_abs += abs(mv.row);
1208 fp_acc_data->sum_mvc += mv.col;
1209 fp_acc_data->sum_mvc_abs += abs(mv.col);
1210 fp_acc_data->sum_mvrs += mv.row * mv.row;
1211 fp_acc_data->sum_mvcs += mv.col * mv.col;
1212 ++(fp_acc_data->intercount);
1213
1214 *best_ref_mv = mv;
1215
1216 #if CONFIG_FP_MB_STATS
1217 if (cpi->use_fp_mb_stats) {
1218 // inter prediction statistics
1219 cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
1220 cpi->twopass.frame_mb_stats_buf[mb_index] &= ~FPMB_DCINTRA_MASK;
1221 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_MOTION_ZERO_MASK;
1222 if (this_error > FPMB_ERROR_LARGE_TH) {
1223 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_LARGE_MASK;
1224 } else if (this_error < FPMB_ERROR_SMALL_TH) {
1225 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_SMALL_MASK;
1226 }
1227 }
1228 #endif
1229
1230 if (!is_zero_mv(&mv)) {
1231 ++(fp_acc_data->mvcount);
1232
1233 #if CONFIG_FP_MB_STATS
1234 if (cpi->use_fp_mb_stats) {
1235 cpi->twopass.frame_mb_stats_buf[mb_index] &= ~FPMB_MOTION_ZERO_MASK;
1236 // check estimated motion direction
1237 if (mv.as_mv.col > 0 && mv.as_mv.col >= abs(mv.as_mv.row)) {
1238 // right direction
1239 cpi->twopass.frame_mb_stats_buf[mb_index] |=
1240 FPMB_MOTION_RIGHT_MASK;
1241 } else if (mv.as_mv.row < 0 &&
1242 abs(mv.as_mv.row) >= abs(mv.as_mv.col)) {
1243 // up direction
1244 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_MOTION_UP_MASK;
1245 } else if (mv.as_mv.col < 0 &&
1246 abs(mv.as_mv.col) >= abs(mv.as_mv.row)) {
1247 // left direction
1248 cpi->twopass.frame_mb_stats_buf[mb_index] |=
1249 FPMB_MOTION_LEFT_MASK;
1250 } else {
1251 // down direction
1252 cpi->twopass.frame_mb_stats_buf[mb_index] |=
1253 FPMB_MOTION_DOWN_MASK;
1254 }
1255 }
1256 #endif
1257
1258 // Does the row vector point inwards or outwards?
1259 if (mb_row < cm->mb_rows / 2) {
1260 if (mv.row > 0)
1261 --(fp_acc_data->sum_in_vectors);
1262 else if (mv.row < 0)
1263 ++(fp_acc_data->sum_in_vectors);
1264 } else if (mb_row > cm->mb_rows / 2) {
1265 if (mv.row > 0)
1266 ++(fp_acc_data->sum_in_vectors);
1267 else if (mv.row < 0)
1268 --(fp_acc_data->sum_in_vectors);
1269 }
1270
1271 // Does the col vector point inwards or outwards?
1272 if (mb_col < cm->mb_cols / 2) {
1273 if (mv.col > 0)
1274 --(fp_acc_data->sum_in_vectors);
1275 else if (mv.col < 0)
1276 ++(fp_acc_data->sum_in_vectors);
1277 } else if (mb_col > cm->mb_cols / 2) {
1278 if (mv.col > 0)
1279 ++(fp_acc_data->sum_in_vectors);
1280 else if (mv.col < 0)
1281 --(fp_acc_data->sum_in_vectors);
1282 }
1283 fp_acc_data->frame_noise_energy += (int64_t)SECTION_NOISE_DEF;
1284 } else if (this_intra_error < scale_sse_threshold(cm, LOW_I_THRESH)) {
1285 fp_acc_data->frame_noise_energy += fp_estimate_block_noise(x, bsize);
1286 } else { // 0,0 mv but high error
1287 fp_acc_data->frame_noise_energy += (int64_t)SECTION_NOISE_DEF;
1288 }
1289 } else { // Intra < inter error
1290 int scaled_low_intra_thresh = scale_sse_threshold(cm, LOW_I_THRESH);
1291 if (this_intra_error < scaled_low_intra_thresh) {
1292 fp_acc_data->frame_noise_energy += fp_estimate_block_noise(x, bsize);
1293 if (this_motion_error < scaled_low_intra_thresh) {
1294 fp_acc_data->intra_count_low += 1.0;
1295 } else {
1296 fp_acc_data->intra_count_high += 1.0;
1297 }
1298 } else {
1299 fp_acc_data->frame_noise_energy += (int64_t)SECTION_NOISE_DEF;
1300 fp_acc_data->intra_count_high += 1.0;
1301 }
1302 }
1303 } else {
1304 fp_acc_data->sr_coded_error += (int64_t)this_error;
1305 }
1306 fp_acc_data->coded_error += (int64_t)this_error;
1307
1308 recon_yoffset += 16;
1309 recon_uvoffset += uv_mb_height;
1310
1311 // Accumulate row level stats to the corresponding tile stats
1312 if (cpi->row_mt && mb_col == mb_col_end - 1)
1313 accumulate_fp_mb_row_stat(tile_data, fp_acc_data);
1314
1315 (*(cpi->row_mt_sync_write_ptr))(&tile_data->row_mt_sync, mb_row, c,
1316 num_mb_cols);
1317 }
1318 vpx_clear_system_state();
1319 }
1320
first_pass_encode(VP9_COMP * cpi,FIRSTPASS_DATA * fp_acc_data)1321 static void first_pass_encode(VP9_COMP *cpi, FIRSTPASS_DATA *fp_acc_data) {
1322 VP9_COMMON *const cm = &cpi->common;
1323 int mb_row;
1324 TileDataEnc tile_data;
1325 TileInfo *tile = &tile_data.tile_info;
1326 MV zero_mv = { 0, 0 };
1327 MV best_ref_mv;
1328 // Tiling is ignored in the first pass.
1329 vp9_tile_init(tile, cm, 0, 0);
1330
1331 for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
1332 best_ref_mv = zero_mv;
1333 vp9_first_pass_encode_tile_mb_row(cpi, &cpi->td, fp_acc_data, &tile_data,
1334 &best_ref_mv, mb_row);
1335 }
1336 }
1337
vp9_first_pass(VP9_COMP * cpi,const struct lookahead_entry * source)1338 void vp9_first_pass(VP9_COMP *cpi, const struct lookahead_entry *source) {
1339 MACROBLOCK *const x = &cpi->td.mb;
1340 VP9_COMMON *const cm = &cpi->common;
1341 MACROBLOCKD *const xd = &x->e_mbd;
1342 TWO_PASS *twopass = &cpi->twopass;
1343
1344 YV12_BUFFER_CONFIG *const lst_yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
1345 YV12_BUFFER_CONFIG *gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
1346 YV12_BUFFER_CONFIG *const new_yv12 = get_frame_new_buffer(cm);
1347 const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;
1348
1349 BufferPool *const pool = cm->buffer_pool;
1350
1351 FIRSTPASS_DATA fp_temp_data;
1352 FIRSTPASS_DATA *fp_acc_data = &fp_temp_data;
1353
1354 vpx_clear_system_state();
1355 vp9_zero(fp_temp_data);
1356 fp_acc_data->image_data_start_row = INVALID_ROW;
1357
1358 // First pass code requires valid last and new frame buffers.
1359 assert(new_yv12 != NULL);
1360 assert(frame_is_intra_only(cm) || (lst_yv12 != NULL));
1361
1362 #if CONFIG_FP_MB_STATS
1363 if (cpi->use_fp_mb_stats) {
1364 vp9_zero_array(cpi->twopass.frame_mb_stats_buf, cm->initial_mbs);
1365 }
1366 #endif
1367
1368 set_first_pass_params(cpi);
1369 vp9_set_quantizer(cm, find_fp_qindex(cm->bit_depth));
1370
1371 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
1372
1373 vp9_setup_src_planes(x, cpi->Source, 0, 0);
1374 vp9_setup_dst_planes(xd->plane, new_yv12, 0, 0);
1375
1376 if (!frame_is_intra_only(cm)) {
1377 vp9_setup_pre_planes(xd, 0, first_ref_buf, 0, 0, NULL);
1378 }
1379
1380 xd->mi = cm->mi_grid_visible;
1381 xd->mi[0] = cm->mi;
1382
1383 vp9_frame_init_quantizer(cpi);
1384
1385 x->skip_recode = 0;
1386
1387 vp9_init_mv_probs(cm);
1388 vp9_initialize_rd_consts(cpi);
1389
1390 cm->log2_tile_rows = 0;
1391
1392 if (cpi->row_mt_bit_exact && cpi->twopass.fp_mb_float_stats == NULL)
1393 CHECK_MEM_ERROR(
1394 cm, cpi->twopass.fp_mb_float_stats,
1395 vpx_calloc(cm->MBs * sizeof(*cpi->twopass.fp_mb_float_stats), 1));
1396
1397 {
1398 FIRSTPASS_STATS fps;
1399 TileDataEnc *first_tile_col;
1400 if (!cpi->row_mt) {
1401 cm->log2_tile_cols = 0;
1402 cpi->row_mt_sync_read_ptr = vp9_row_mt_sync_read_dummy;
1403 cpi->row_mt_sync_write_ptr = vp9_row_mt_sync_write_dummy;
1404 first_pass_encode(cpi, fp_acc_data);
1405 first_pass_stat_calc(cpi, &fps, fp_acc_data);
1406 } else {
1407 cpi->row_mt_sync_read_ptr = vp9_row_mt_sync_read;
1408 cpi->row_mt_sync_write_ptr = vp9_row_mt_sync_write;
1409 if (cpi->row_mt_bit_exact) {
1410 cm->log2_tile_cols = 0;
1411 vp9_zero_array(cpi->twopass.fp_mb_float_stats, cm->MBs);
1412 }
1413 vp9_encode_fp_row_mt(cpi);
1414 first_tile_col = &cpi->tile_data[0];
1415 if (cpi->row_mt_bit_exact)
1416 accumulate_floating_point_stats(cpi, first_tile_col);
1417 first_pass_stat_calc(cpi, &fps, &(first_tile_col->fp_data));
1418 }
1419
1420 // Dont allow a value of 0 for duration.
1421 // (Section duration is also defaulted to minimum of 1.0).
1422 fps.duration = VPXMAX(1.0, (double)(source->ts_end - source->ts_start));
1423
1424 // Don't want to do output stats with a stack variable!
1425 twopass->this_frame_stats = fps;
1426 output_stats(&twopass->this_frame_stats, cpi->output_pkt_list);
1427 accumulate_stats(&twopass->total_stats, &fps);
1428
1429 #if CONFIG_FP_MB_STATS
1430 if (cpi->use_fp_mb_stats) {
1431 output_fpmb_stats(twopass->frame_mb_stats_buf, cm, cpi->output_pkt_list);
1432 }
1433 #endif
1434 }
1435
1436 // Copy the previous Last Frame back into gf and and arf buffers if
1437 // the prediction is good enough... but also don't allow it to lag too far.
1438 if ((twopass->sr_update_lag > 3) ||
1439 ((cm->current_video_frame > 0) &&
1440 (twopass->this_frame_stats.pcnt_inter > 0.20) &&
1441 ((twopass->this_frame_stats.intra_error /
1442 DOUBLE_DIVIDE_CHECK(twopass->this_frame_stats.coded_error)) > 2.0))) {
1443 if (gld_yv12 != NULL) {
1444 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
1445 cm->ref_frame_map[cpi->lst_fb_idx]);
1446 }
1447 twopass->sr_update_lag = 1;
1448 } else {
1449 ++twopass->sr_update_lag;
1450 }
1451
1452 vpx_extend_frame_borders(new_yv12);
1453
1454 // The frame we just compressed now becomes the last frame.
1455 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->lst_fb_idx],
1456 cm->new_fb_idx);
1457
1458 // Special case for the first frame. Copy into the GF buffer as a second
1459 // reference.
1460 if (cm->current_video_frame == 0 && cpi->gld_fb_idx != INVALID_IDX) {
1461 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
1462 cm->ref_frame_map[cpi->lst_fb_idx]);
1463 }
1464
1465 // Use this to see what the first pass reconstruction looks like.
1466 if (0) {
1467 char filename[512];
1468 FILE *recon_file;
1469 snprintf(filename, sizeof(filename), "enc%04d.yuv",
1470 (int)cm->current_video_frame);
1471
1472 if (cm->current_video_frame == 0)
1473 recon_file = fopen(filename, "wb");
1474 else
1475 recon_file = fopen(filename, "ab");
1476
1477 (void)fwrite(lst_yv12->buffer_alloc, lst_yv12->frame_size, 1, recon_file);
1478 fclose(recon_file);
1479 }
1480
1481 ++cm->current_video_frame;
1482 if (cpi->use_svc) vp9_inc_frame_in_layer(cpi);
1483 }
1484
1485 static const double q_pow_term[(QINDEX_RANGE >> 5) + 1] = {
1486 0.65, 0.70, 0.75, 0.85, 0.90, 0.90, 0.90, 1.00, 1.25
1487 };
1488
calc_correction_factor(double err_per_mb,double err_divisor,int q)1489 static double calc_correction_factor(double err_per_mb, double err_divisor,
1490 int q) {
1491 const double error_term = err_per_mb / DOUBLE_DIVIDE_CHECK(err_divisor);
1492 const int index = q >> 5;
1493 double power_term;
1494
1495 assert((index >= 0) && (index < (QINDEX_RANGE >> 5)));
1496
1497 // Adjustment based on quantizer to the power term.
1498 power_term =
1499 q_pow_term[index] +
1500 (((q_pow_term[index + 1] - q_pow_term[index]) * (q % 32)) / 32.0);
1501
1502 // Calculate correction factor.
1503 if (power_term < 1.0) assert(error_term >= 0.0);
1504
1505 return fclamp(pow(error_term, power_term), 0.05, 5.0);
1506 }
1507
wq_err_divisor(VP9_COMP * cpi)1508 static double wq_err_divisor(VP9_COMP *cpi) {
1509 const VP9_COMMON *const cm = &cpi->common;
1510 unsigned int screen_area = (cm->width * cm->height);
1511
1512 // Use a different error per mb factor for calculating boost for
1513 // different formats.
1514 if (screen_area <= 640 * 360) {
1515 return 115.0;
1516 } else if (screen_area < 1280 * 720) {
1517 return 125.0;
1518 } else if (screen_area <= 1920 * 1080) {
1519 return 130.0;
1520 } else if (screen_area < 3840 * 2160) {
1521 return 150.0;
1522 }
1523
1524 // Fall through to here only for 4K and above.
1525 return 200.0;
1526 }
1527
1528 #define NOISE_FACTOR_MIN 0.9
1529 #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)1530 static int get_twopass_worst_quality(VP9_COMP *cpi, const double section_err,
1531 double inactive_zone, double section_noise,
1532 int section_target_bandwidth) {
1533 const RATE_CONTROL *const rc = &cpi->rc;
1534 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1535 TWO_PASS *const twopass = &cpi->twopass;
1536 double last_group_rate_err;
1537
1538 // Clamp the target rate to VBR min / max limts.
1539 const int target_rate =
1540 vp9_rc_clamp_pframe_target_size(cpi, section_target_bandwidth);
1541 double noise_factor = pow((section_noise / SECTION_NOISE_DEF), 0.5);
1542 noise_factor = fclamp(noise_factor, NOISE_FACTOR_MIN, NOISE_FACTOR_MAX);
1543 inactive_zone = fclamp(inactive_zone, 0.0, 1.0);
1544
1545 // TODO(jimbankoski): remove #if here or below when this has been
1546 // well tested.
1547 #if CONFIG_ALWAYS_ADJUST_BPM
1548 // based on recent history adjust expectations of bits per macroblock.
1549 last_group_rate_err =
1550 (double)twopass->rolling_arf_group_actual_bits /
1551 DOUBLE_DIVIDE_CHECK((double)twopass->rolling_arf_group_target_bits);
1552 last_group_rate_err = VPXMAX(0.25, VPXMIN(4.0, last_group_rate_err));
1553 twopass->bpm_factor *= (3.0 + last_group_rate_err) / 4.0;
1554 twopass->bpm_factor = VPXMAX(0.25, VPXMIN(4.0, twopass->bpm_factor));
1555 #endif
1556
1557 if (target_rate <= 0) {
1558 return rc->worst_quality; // Highest value allowed
1559 } else {
1560 const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
1561 ? cpi->initial_mbs
1562 : cpi->common.MBs;
1563 const double active_pct = VPXMAX(0.01, 1.0 - inactive_zone);
1564 const int active_mbs = (int)VPXMAX(1, (double)num_mbs * active_pct);
1565 const double av_err_per_mb = section_err / active_pct;
1566 const double speed_term = 1.0 + 0.04 * oxcf->speed;
1567 const int target_norm_bits_per_mb =
1568 (int)(((uint64_t)target_rate << BPER_MB_NORMBITS) / active_mbs);
1569 int q;
1570
1571 // TODO(jimbankoski): remove #if here or above when this has been
1572 // well tested.
1573 #if !CONFIG_ALWAYS_ADJUST_BPM
1574 // based on recent history adjust expectations of bits per macroblock.
1575 last_group_rate_err =
1576 (double)twopass->rolling_arf_group_actual_bits /
1577 DOUBLE_DIVIDE_CHECK((double)twopass->rolling_arf_group_target_bits);
1578 last_group_rate_err = VPXMAX(0.25, VPXMIN(4.0, last_group_rate_err));
1579 twopass->bpm_factor *= (3.0 + last_group_rate_err) / 4.0;
1580 twopass->bpm_factor = VPXMAX(0.25, VPXMIN(4.0, twopass->bpm_factor));
1581 #endif
1582
1583 // Try and pick a max Q that will be high enough to encode the
1584 // content at the given rate.
1585 for (q = rc->best_quality; q < rc->worst_quality; ++q) {
1586 const double factor =
1587 calc_correction_factor(av_err_per_mb, wq_err_divisor(cpi), q);
1588 const int bits_per_mb = vp9_rc_bits_per_mb(
1589 INTER_FRAME, q,
1590 factor * speed_term * cpi->twopass.bpm_factor * noise_factor,
1591 cpi->common.bit_depth);
1592 if (bits_per_mb <= target_norm_bits_per_mb) break;
1593 }
1594
1595 // Restriction on active max q for constrained quality mode.
1596 if (cpi->oxcf.rc_mode == VPX_CQ) q = VPXMAX(q, oxcf->cq_level);
1597 return q;
1598 }
1599 }
1600
setup_rf_level_maxq(VP9_COMP * cpi)1601 static void setup_rf_level_maxq(VP9_COMP *cpi) {
1602 int i;
1603 RATE_CONTROL *const rc = &cpi->rc;
1604 for (i = INTER_NORMAL; i < RATE_FACTOR_LEVELS; ++i) {
1605 int qdelta = vp9_frame_type_qdelta(cpi, i, rc->worst_quality);
1606 rc->rf_level_maxq[i] = VPXMAX(rc->worst_quality + qdelta, rc->best_quality);
1607 }
1608 }
1609
init_subsampling(VP9_COMP * cpi)1610 static void init_subsampling(VP9_COMP *cpi) {
1611 const VP9_COMMON *const cm = &cpi->common;
1612 RATE_CONTROL *const rc = &cpi->rc;
1613 const int w = cm->width;
1614 const int h = cm->height;
1615 int i;
1616
1617 for (i = 0; i < FRAME_SCALE_STEPS; ++i) {
1618 // Note: Frames with odd-sized dimensions may result from this scaling.
1619 rc->frame_width[i] = (w * 16) / frame_scale_factor[i];
1620 rc->frame_height[i] = (h * 16) / frame_scale_factor[i];
1621 }
1622
1623 setup_rf_level_maxq(cpi);
1624 }
1625
calculate_coded_size(VP9_COMP * cpi,int * scaled_frame_width,int * scaled_frame_height)1626 void calculate_coded_size(VP9_COMP *cpi, int *scaled_frame_width,
1627 int *scaled_frame_height) {
1628 RATE_CONTROL *const rc = &cpi->rc;
1629 *scaled_frame_width = rc->frame_width[rc->frame_size_selector];
1630 *scaled_frame_height = rc->frame_height[rc->frame_size_selector];
1631 }
1632
vp9_init_second_pass(VP9_COMP * cpi)1633 void vp9_init_second_pass(VP9_COMP *cpi) {
1634 VP9EncoderConfig *const oxcf = &cpi->oxcf;
1635 RATE_CONTROL *const rc = &cpi->rc;
1636 TWO_PASS *const twopass = &cpi->twopass;
1637 double frame_rate;
1638 FIRSTPASS_STATS *stats;
1639
1640 zero_stats(&twopass->total_stats);
1641 zero_stats(&twopass->total_left_stats);
1642
1643 if (!twopass->stats_in_end) return;
1644
1645 stats = &twopass->total_stats;
1646
1647 *stats = *twopass->stats_in_end;
1648 twopass->total_left_stats = *stats;
1649
1650 // Scan the first pass file and calculate a modified score for each
1651 // frame that is used to distribute bits. The modified score is assumed
1652 // to provide a linear basis for bit allocation. I.e a frame A with a score
1653 // that is double that of frame B will be allocated 2x as many bits.
1654 {
1655 double modified_score_total = 0.0;
1656 const FIRSTPASS_STATS *s = twopass->stats_in;
1657 double av_err;
1658
1659 if (oxcf->vbr_corpus_complexity) {
1660 twopass->mean_mod_score = (double)oxcf->vbr_corpus_complexity / 10.0;
1661 av_err = get_distribution_av_err(cpi, twopass);
1662 } else {
1663 av_err = get_distribution_av_err(cpi, twopass);
1664 // The first scan is unclamped and gives a raw average.
1665 while (s < twopass->stats_in_end) {
1666 modified_score_total += calculate_mod_frame_score(cpi, oxcf, s, av_err);
1667 ++s;
1668 }
1669
1670 // The average error from this first scan is used to define the midpoint
1671 // error for the rate distribution function.
1672 twopass->mean_mod_score =
1673 modified_score_total / DOUBLE_DIVIDE_CHECK(stats->count);
1674 }
1675
1676 // Second scan using clamps based on the previous cycle average.
1677 // This may modify the total and average somewhat but we dont bother with
1678 // further itterations.
1679 modified_score_total = 0.0;
1680 s = twopass->stats_in;
1681 while (s < twopass->stats_in_end) {
1682 modified_score_total +=
1683 calculate_norm_frame_score(cpi, twopass, oxcf, s, av_err);
1684 ++s;
1685 }
1686 twopass->normalized_score_left = modified_score_total;
1687
1688 // If using Corpus wide VBR mode then update the clip target bandwidth to
1689 // reflect how the clip compares to the rest of the corpus.
1690 if (oxcf->vbr_corpus_complexity) {
1691 oxcf->target_bandwidth =
1692 (int64_t)((double)oxcf->target_bandwidth *
1693 (twopass->normalized_score_left / stats->count));
1694 }
1695
1696 #if COMPLEXITY_STATS_OUTPUT
1697 {
1698 FILE *compstats;
1699 compstats = fopen("complexity_stats.stt", "a");
1700 fprintf(compstats, "%10.3lf\n",
1701 twopass->normalized_score_left / stats->count);
1702 fclose(compstats);
1703 }
1704 #endif
1705 }
1706
1707 frame_rate = 10000000.0 * stats->count / stats->duration;
1708 // Each frame can have a different duration, as the frame rate in the source
1709 // isn't guaranteed to be constant. The frame rate prior to the first frame
1710 // encoded in the second pass is a guess. However, the sum duration is not.
1711 // It is calculated based on the actual durations of all frames from the
1712 // first pass.
1713 vp9_new_framerate(cpi, frame_rate);
1714 twopass->bits_left =
1715 (int64_t)(stats->duration * oxcf->target_bandwidth / 10000000.0);
1716
1717 // This variable monitors how far behind the second ref update is lagging.
1718 twopass->sr_update_lag = 1;
1719
1720 // Reset the vbr bits off target counters
1721 rc->vbr_bits_off_target = 0;
1722 rc->vbr_bits_off_target_fast = 0;
1723 rc->rate_error_estimate = 0;
1724
1725 // Static sequence monitor variables.
1726 twopass->kf_zeromotion_pct = 100;
1727 twopass->last_kfgroup_zeromotion_pct = 100;
1728
1729 // Initialize bits per macro_block estimate correction factor.
1730 twopass->bpm_factor = 1.0;
1731 // Initialize actual and target bits counters for ARF groups so that
1732 // at the start we have a neutral bpm adjustment.
1733 twopass->rolling_arf_group_target_bits = 1;
1734 twopass->rolling_arf_group_actual_bits = 1;
1735
1736 if (oxcf->resize_mode != RESIZE_NONE) {
1737 init_subsampling(cpi);
1738 }
1739
1740 // Initialize the arnr strangth adjustment to 0
1741 twopass->arnr_strength_adjustment = 0;
1742 }
1743
1744 #define SR_DIFF_PART 0.0015
1745 #define INTRA_PART 0.005
1746 #define DEFAULT_DECAY_LIMIT 0.75
1747 #define LOW_SR_DIFF_TRHESH 0.1
1748 #define SR_DIFF_MAX 128.0
1749 #define LOW_CODED_ERR_PER_MB 10.0
1750 #define NCOUNT_FRAME_II_THRESH 6.0
1751
get_sr_decay_rate(const VP9_COMP * cpi,const FIRSTPASS_STATS * frame)1752 static double get_sr_decay_rate(const VP9_COMP *cpi,
1753 const FIRSTPASS_STATS *frame) {
1754 double sr_diff = (frame->sr_coded_error - frame->coded_error);
1755 double sr_decay = 1.0;
1756 double modified_pct_inter;
1757 double modified_pcnt_intra;
1758 const double motion_amplitude_part =
1759 frame->pcnt_motion * ((frame->mvc_abs + frame->mvr_abs) /
1760 (cpi->initial_height + cpi->initial_width));
1761
1762 modified_pct_inter = frame->pcnt_inter;
1763 if ((frame->coded_error > LOW_CODED_ERR_PER_MB) &&
1764 ((frame->intra_error / DOUBLE_DIVIDE_CHECK(frame->coded_error)) <
1765 (double)NCOUNT_FRAME_II_THRESH)) {
1766 modified_pct_inter =
1767 frame->pcnt_inter + frame->pcnt_intra_low - frame->pcnt_neutral;
1768 }
1769 modified_pcnt_intra = 100 * (1.0 - modified_pct_inter);
1770
1771 if ((sr_diff > LOW_SR_DIFF_TRHESH)) {
1772 sr_diff = VPXMIN(sr_diff, SR_DIFF_MAX);
1773 sr_decay = 1.0 - (SR_DIFF_PART * sr_diff) - motion_amplitude_part -
1774 (INTRA_PART * modified_pcnt_intra);
1775 }
1776 return VPXMAX(sr_decay, DEFAULT_DECAY_LIMIT);
1777 }
1778
1779 // This function gives an estimate of how badly we believe the prediction
1780 // quality is decaying from frame to frame.
get_zero_motion_factor(const VP9_COMP * cpi,const FIRSTPASS_STATS * frame)1781 static double get_zero_motion_factor(const VP9_COMP *cpi,
1782 const FIRSTPASS_STATS *frame) {
1783 const double zero_motion_pct = frame->pcnt_inter - frame->pcnt_motion;
1784 double sr_decay = get_sr_decay_rate(cpi, frame);
1785 return VPXMIN(sr_decay, zero_motion_pct);
1786 }
1787
1788 #define ZM_POWER_FACTOR 0.75
1789
get_prediction_decay_rate(const VP9_COMP * cpi,const FIRSTPASS_STATS * next_frame)1790 static double get_prediction_decay_rate(const VP9_COMP *cpi,
1791 const FIRSTPASS_STATS *next_frame) {
1792 const double sr_decay_rate = get_sr_decay_rate(cpi, next_frame);
1793 const double zero_motion_factor =
1794 (0.95 * pow((next_frame->pcnt_inter - next_frame->pcnt_motion),
1795 ZM_POWER_FACTOR));
1796
1797 return VPXMAX(zero_motion_factor,
1798 (sr_decay_rate + ((1.0 - sr_decay_rate) * zero_motion_factor)));
1799 }
1800
1801 // Function to test for a condition where a complex transition is followed
1802 // by a static section. For example in slide shows where there is a fade
1803 // between slides. This is to help with more optimal kf and gf positioning.
detect_transition_to_still(VP9_COMP * cpi,int frame_interval,int still_interval,double loop_decay_rate,double last_decay_rate)1804 static int detect_transition_to_still(VP9_COMP *cpi, int frame_interval,
1805 int still_interval,
1806 double loop_decay_rate,
1807 double last_decay_rate) {
1808 TWO_PASS *const twopass = &cpi->twopass;
1809 RATE_CONTROL *const rc = &cpi->rc;
1810
1811 // Break clause to detect very still sections after motion
1812 // For example a static image after a fade or other transition
1813 // instead of a clean scene cut.
1814 if (frame_interval > rc->min_gf_interval && loop_decay_rate >= 0.999 &&
1815 last_decay_rate < 0.9) {
1816 int j;
1817
1818 // Look ahead a few frames to see if static condition persists...
1819 for (j = 0; j < still_interval; ++j) {
1820 const FIRSTPASS_STATS *stats = &twopass->stats_in[j];
1821 if (stats >= twopass->stats_in_end) break;
1822
1823 if (stats->pcnt_inter - stats->pcnt_motion < 0.999) break;
1824 }
1825
1826 // Only if it does do we signal a transition to still.
1827 return j == still_interval;
1828 }
1829
1830 return 0;
1831 }
1832
1833 // This function detects a flash through the high relative pcnt_second_ref
1834 // score in the frame following a flash frame. The offset passed in should
1835 // reflect this.
detect_flash(const TWO_PASS * twopass,int offset)1836 static int detect_flash(const TWO_PASS *twopass, int offset) {
1837 const FIRSTPASS_STATS *const next_frame = read_frame_stats(twopass, offset);
1838
1839 // What we are looking for here is a situation where there is a
1840 // brief break in prediction (such as a flash) but subsequent frames
1841 // are reasonably well predicted by an earlier (pre flash) frame.
1842 // The recovery after a flash is indicated by a high pcnt_second_ref
1843 // useage or a second ref coded error notabley lower than the last
1844 // frame coded error.
1845 return next_frame != NULL &&
1846 ((next_frame->sr_coded_error < next_frame->coded_error) ||
1847 ((next_frame->pcnt_second_ref > next_frame->pcnt_inter) &&
1848 (next_frame->pcnt_second_ref >= 0.5)));
1849 }
1850
1851 // 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)1852 static void accumulate_frame_motion_stats(const FIRSTPASS_STATS *stats,
1853 double *mv_in_out,
1854 double *mv_in_out_accumulator,
1855 double *abs_mv_in_out_accumulator,
1856 double *mv_ratio_accumulator) {
1857 const double pct = stats->pcnt_motion;
1858
1859 // Accumulate Motion In/Out of frame stats.
1860 *mv_in_out = stats->mv_in_out_count * pct;
1861 *mv_in_out_accumulator += *mv_in_out;
1862 *abs_mv_in_out_accumulator += fabs(*mv_in_out);
1863
1864 // Accumulate a measure of how uniform (or conversely how random) the motion
1865 // field is (a ratio of abs(mv) / mv).
1866 if (pct > 0.05) {
1867 const double mvr_ratio =
1868 fabs(stats->mvr_abs) / DOUBLE_DIVIDE_CHECK(fabs(stats->MVr));
1869 const double mvc_ratio =
1870 fabs(stats->mvc_abs) / DOUBLE_DIVIDE_CHECK(fabs(stats->MVc));
1871
1872 *mv_ratio_accumulator +=
1873 pct * (mvr_ratio < stats->mvr_abs ? mvr_ratio : stats->mvr_abs);
1874 *mv_ratio_accumulator +=
1875 pct * (mvc_ratio < stats->mvc_abs ? mvc_ratio : stats->mvc_abs);
1876 }
1877 }
1878
1879 #define BASELINE_ERR_PER_MB 12500.0
1880 #define GF_MAX_BOOST 96.0
calc_frame_boost(VP9_COMP * cpi,const FIRSTPASS_STATS * this_frame,double this_frame_mv_in_out)1881 static double calc_frame_boost(VP9_COMP *cpi, const FIRSTPASS_STATS *this_frame,
1882 double this_frame_mv_in_out) {
1883 double frame_boost;
1884 const double lq = vp9_convert_qindex_to_q(
1885 cpi->rc.avg_frame_qindex[INTER_FRAME], cpi->common.bit_depth);
1886 const double boost_q_correction = VPXMIN((0.5 + (lq * 0.015)), 1.5);
1887 const double active_area = calculate_active_area(cpi, this_frame);
1888
1889 // Underlying boost factor is based on inter error ratio.
1890 frame_boost = (BASELINE_ERR_PER_MB * active_area) /
1891 DOUBLE_DIVIDE_CHECK(this_frame->coded_error);
1892
1893 // Small adjustment for cases where there is a zoom out
1894 if (this_frame_mv_in_out > 0.0)
1895 frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
1896
1897 // Q correction and scalling
1898 frame_boost = frame_boost * boost_q_correction;
1899
1900 return VPXMIN(frame_boost, GF_MAX_BOOST * boost_q_correction);
1901 }
1902
kf_err_per_mb(VP9_COMP * cpi)1903 static double kf_err_per_mb(VP9_COMP *cpi) {
1904 const VP9_COMMON *const cm = &cpi->common;
1905 unsigned int screen_area = (cm->width * cm->height);
1906
1907 // Use a different error per mb factor for calculating boost for
1908 // different formats.
1909 if (screen_area < 1280 * 720) {
1910 return 2000.0;
1911 } else if (screen_area < 1920 * 1080) {
1912 return 500.0;
1913 }
1914 return 250.0;
1915 }
1916
calc_kf_frame_boost(VP9_COMP * cpi,const FIRSTPASS_STATS * this_frame,double * sr_accumulator,double this_frame_mv_in_out,double max_boost)1917 static double calc_kf_frame_boost(VP9_COMP *cpi,
1918 const FIRSTPASS_STATS *this_frame,
1919 double *sr_accumulator,
1920 double this_frame_mv_in_out,
1921 double max_boost) {
1922 double frame_boost;
1923 const double lq = vp9_convert_qindex_to_q(
1924 cpi->rc.avg_frame_qindex[INTER_FRAME], cpi->common.bit_depth);
1925 const double boost_q_correction = VPXMIN((0.50 + (lq * 0.015)), 2.00);
1926 const double active_area = calculate_active_area(cpi, this_frame);
1927
1928 // Underlying boost factor is based on inter error ratio.
1929 frame_boost = (kf_err_per_mb(cpi) * active_area) /
1930 DOUBLE_DIVIDE_CHECK(this_frame->coded_error + *sr_accumulator);
1931
1932 // Update the accumulator for second ref error difference.
1933 // This is intended to give an indication of how much the coded error is
1934 // increasing over time.
1935 *sr_accumulator += (this_frame->sr_coded_error - this_frame->coded_error);
1936 *sr_accumulator = VPXMAX(0.0, *sr_accumulator);
1937
1938 // Small adjustment for cases where there is a zoom out
1939 if (this_frame_mv_in_out > 0.0)
1940 frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
1941
1942 // Q correction and scaling
1943 // The 40.0 value here is an experimentally derived baseline minimum.
1944 // This value is in line with the minimum per frame boost in the alt_ref
1945 // boost calculation.
1946 frame_boost = ((frame_boost + 40.0) * boost_q_correction);
1947
1948 return VPXMIN(frame_boost, max_boost * boost_q_correction);
1949 }
1950
calc_arf_boost(VP9_COMP * cpi,int f_frames,int b_frames)1951 static int calc_arf_boost(VP9_COMP *cpi, int f_frames, int b_frames) {
1952 TWO_PASS *const twopass = &cpi->twopass;
1953 int i;
1954 double boost_score = 0.0;
1955 double mv_ratio_accumulator = 0.0;
1956 double decay_accumulator = 1.0;
1957 double this_frame_mv_in_out = 0.0;
1958 double mv_in_out_accumulator = 0.0;
1959 double abs_mv_in_out_accumulator = 0.0;
1960 int arf_boost;
1961 int flash_detected = 0;
1962
1963 // Search forward from the proposed arf/next gf position.
1964 for (i = 0; i < f_frames; ++i) {
1965 const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i);
1966 if (this_frame == NULL) break;
1967
1968 // Update the motion related elements to the boost calculation.
1969 accumulate_frame_motion_stats(
1970 this_frame, &this_frame_mv_in_out, &mv_in_out_accumulator,
1971 &abs_mv_in_out_accumulator, &mv_ratio_accumulator);
1972
1973 // We want to discount the flash frame itself and the recovery
1974 // frame that follows as both will have poor scores.
1975 flash_detected = detect_flash(twopass, i) || detect_flash(twopass, i + 1);
1976
1977 // Accumulate the effect of prediction quality decay.
1978 if (!flash_detected) {
1979 decay_accumulator *= get_prediction_decay_rate(cpi, this_frame);
1980 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1981 ? MIN_DECAY_FACTOR
1982 : decay_accumulator;
1983 }
1984 boost_score += decay_accumulator *
1985 calc_frame_boost(cpi, this_frame, this_frame_mv_in_out);
1986 }
1987
1988 arf_boost = (int)boost_score;
1989
1990 // Reset for backward looking loop.
1991 boost_score = 0.0;
1992 mv_ratio_accumulator = 0.0;
1993 decay_accumulator = 1.0;
1994 this_frame_mv_in_out = 0.0;
1995 mv_in_out_accumulator = 0.0;
1996 abs_mv_in_out_accumulator = 0.0;
1997
1998 // Search backward towards last gf position.
1999 for (i = -1; i >= -b_frames; --i) {
2000 const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i);
2001 if (this_frame == NULL) break;
2002
2003 // Update the motion related elements to the boost calculation.
2004 accumulate_frame_motion_stats(
2005 this_frame, &this_frame_mv_in_out, &mv_in_out_accumulator,
2006 &abs_mv_in_out_accumulator, &mv_ratio_accumulator);
2007
2008 // We want to discount the the flash frame itself and the recovery
2009 // frame that follows as both will have poor scores.
2010 flash_detected = detect_flash(twopass, i) || detect_flash(twopass, i + 1);
2011
2012 // Cumulative effect of prediction quality decay.
2013 if (!flash_detected) {
2014 decay_accumulator *= get_prediction_decay_rate(cpi, this_frame);
2015 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
2016 ? MIN_DECAY_FACTOR
2017 : decay_accumulator;
2018 }
2019 boost_score += decay_accumulator *
2020 calc_frame_boost(cpi, this_frame, this_frame_mv_in_out);
2021 }
2022 arf_boost += (int)boost_score;
2023
2024 if (arf_boost < ((b_frames + f_frames) * 40))
2025 arf_boost = ((b_frames + f_frames) * 40);
2026 arf_boost = VPXMAX(arf_boost, MIN_ARF_GF_BOOST);
2027
2028 return arf_boost;
2029 }
2030
2031 // 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)2032 static int calculate_section_intra_ratio(const FIRSTPASS_STATS *begin,
2033 const FIRSTPASS_STATS *end,
2034 int section_length) {
2035 const FIRSTPASS_STATS *s = begin;
2036 double intra_error = 0.0;
2037 double coded_error = 0.0;
2038 int i = 0;
2039
2040 while (s < end && i < section_length) {
2041 intra_error += s->intra_error;
2042 coded_error += s->coded_error;
2043 ++s;
2044 ++i;
2045 }
2046
2047 return (int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
2048 }
2049
2050 // Calculate the total bits to allocate in this GF/ARF group.
calculate_total_gf_group_bits(VP9_COMP * cpi,double gf_group_err)2051 static int64_t calculate_total_gf_group_bits(VP9_COMP *cpi,
2052 double gf_group_err) {
2053 const RATE_CONTROL *const rc = &cpi->rc;
2054 const TWO_PASS *const twopass = &cpi->twopass;
2055 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
2056 int64_t total_group_bits;
2057
2058 // Calculate the bits to be allocated to the group as a whole.
2059 if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0.0)) {
2060 total_group_bits = (int64_t)(twopass->kf_group_bits *
2061 (gf_group_err / twopass->kf_group_error_left));
2062 } else {
2063 total_group_bits = 0;
2064 }
2065
2066 // Clamp odd edge cases.
2067 total_group_bits = (total_group_bits < 0)
2068 ? 0
2069 : (total_group_bits > twopass->kf_group_bits)
2070 ? twopass->kf_group_bits
2071 : total_group_bits;
2072
2073 // Clip based on user supplied data rate variability limit.
2074 if (total_group_bits > (int64_t)max_bits * rc->baseline_gf_interval)
2075 total_group_bits = (int64_t)max_bits * rc->baseline_gf_interval;
2076
2077 return total_group_bits;
2078 }
2079
2080 // 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)2081 static int calculate_boost_bits(int frame_count, int boost,
2082 int64_t total_group_bits) {
2083 int allocation_chunks;
2084
2085 // return 0 for invalid inputs (could arise e.g. through rounding errors)
2086 if (!boost || (total_group_bits <= 0) || (frame_count < 0)) return 0;
2087
2088 allocation_chunks = (frame_count * NORMAL_BOOST) + boost;
2089
2090 // Prevent overflow.
2091 if (boost > 1023) {
2092 int divisor = boost >> 10;
2093 boost /= divisor;
2094 allocation_chunks /= divisor;
2095 }
2096
2097 // Calculate the number of extra bits for use in the boosted frame or frames.
2098 return VPXMAX((int)(((int64_t)boost * total_group_bits) / allocation_chunks),
2099 0);
2100 }
2101
2102 // Used in corpus vbr: Calculates the total normalized group complexity score
2103 // for a given number of frames starting at the current position in the stats
2104 // file.
calculate_group_score(VP9_COMP * cpi,double av_score,int frame_count)2105 static double calculate_group_score(VP9_COMP *cpi, double av_score,
2106 int frame_count) {
2107 VP9EncoderConfig *const oxcf = &cpi->oxcf;
2108 TWO_PASS *const twopass = &cpi->twopass;
2109 const FIRSTPASS_STATS *s = twopass->stats_in;
2110 double score_total = 0.0;
2111 int i = 0;
2112
2113 // We dont ever want to return a 0 score here.
2114 if (frame_count == 0) return 1.0;
2115
2116 while ((i < frame_count) && (s < twopass->stats_in_end)) {
2117 score_total += calculate_norm_frame_score(cpi, twopass, oxcf, s, av_score);
2118 ++s;
2119 ++i;
2120 }
2121
2122 return score_total;
2123 }
2124
find_arf_order(VP9_COMP * cpi,GF_GROUP * gf_group,int * index_counter,int depth,int start,int end)2125 static void find_arf_order(VP9_COMP *cpi, GF_GROUP *gf_group,
2126 int *index_counter, int depth, int start, int end) {
2127 TWO_PASS *twopass = &cpi->twopass;
2128 const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
2129 FIRSTPASS_STATS fpf_frame;
2130 const int mid = (start + end + 1) >> 1;
2131 const int min_frame_interval = 2;
2132 int idx;
2133
2134 // Process regular P frames
2135 if ((end - start < min_frame_interval) ||
2136 (depth > gf_group->allowed_max_layer_depth)) {
2137 for (idx = start; idx <= end; ++idx) {
2138 gf_group->update_type[*index_counter] = LF_UPDATE;
2139 gf_group->arf_src_offset[*index_counter] = 0;
2140 gf_group->frame_gop_index[*index_counter] = idx;
2141 gf_group->rf_level[*index_counter] = INTER_NORMAL;
2142 gf_group->layer_depth[*index_counter] = depth;
2143 gf_group->gfu_boost[*index_counter] = NORMAL_BOOST;
2144 ++(*index_counter);
2145 }
2146 gf_group->max_layer_depth = VPXMAX(gf_group->max_layer_depth, depth);
2147 return;
2148 }
2149
2150 assert(abs(mid - start) >= 1 && abs(mid - end) >= 1);
2151
2152 // Process ARF frame
2153 gf_group->layer_depth[*index_counter] = depth;
2154 gf_group->update_type[*index_counter] = ARF_UPDATE;
2155 gf_group->arf_src_offset[*index_counter] = mid - start;
2156 gf_group->frame_gop_index[*index_counter] = mid;
2157 gf_group->rf_level[*index_counter] = GF_ARF_LOW;
2158
2159 for (idx = 0; idx <= mid; ++idx)
2160 if (EOF == input_stats(twopass, &fpf_frame)) break;
2161
2162 gf_group->gfu_boost[*index_counter] =
2163 VPXMAX(MIN_ARF_GF_BOOST,
2164 calc_arf_boost(cpi, end - mid + 1, mid - start) >> depth);
2165
2166 reset_fpf_position(twopass, start_pos);
2167
2168 ++(*index_counter);
2169
2170 find_arf_order(cpi, gf_group, index_counter, depth + 1, start, mid - 1);
2171
2172 gf_group->update_type[*index_counter] = USE_BUF_FRAME;
2173 gf_group->arf_src_offset[*index_counter] = 0;
2174 gf_group->frame_gop_index[*index_counter] = mid;
2175 gf_group->rf_level[*index_counter] = INTER_NORMAL;
2176 gf_group->layer_depth[*index_counter] = depth;
2177 ++(*index_counter);
2178
2179 find_arf_order(cpi, gf_group, index_counter, depth + 1, mid + 1, end);
2180 }
2181
set_gf_overlay_frame_type(GF_GROUP * gf_group,int frame_index,int source_alt_ref_active)2182 static INLINE void set_gf_overlay_frame_type(GF_GROUP *gf_group,
2183 int frame_index,
2184 int source_alt_ref_active) {
2185 if (source_alt_ref_active) {
2186 gf_group->update_type[frame_index] = OVERLAY_UPDATE;
2187 gf_group->rf_level[frame_index] = INTER_NORMAL;
2188 gf_group->layer_depth[frame_index] = MAX_ARF_LAYERS - 1;
2189 gf_group->gfu_boost[frame_index] = NORMAL_BOOST;
2190 } else {
2191 gf_group->update_type[frame_index] = GF_UPDATE;
2192 gf_group->rf_level[frame_index] = GF_ARF_STD;
2193 gf_group->layer_depth[frame_index] = 0;
2194 }
2195 }
2196
define_gf_group_structure(VP9_COMP * cpi)2197 static void define_gf_group_structure(VP9_COMP *cpi) {
2198 RATE_CONTROL *const rc = &cpi->rc;
2199 TWO_PASS *const twopass = &cpi->twopass;
2200 GF_GROUP *const gf_group = &twopass->gf_group;
2201 int frame_index = 0;
2202 int key_frame = cpi->common.frame_type == KEY_FRAME;
2203 int layer_depth = 1;
2204 int gop_frames =
2205 rc->baseline_gf_interval - (key_frame || rc->source_alt_ref_pending);
2206
2207 gf_group->frame_start = cpi->common.current_video_frame;
2208 gf_group->frame_end = gf_group->frame_start + rc->baseline_gf_interval;
2209 gf_group->max_layer_depth = 0;
2210 gf_group->allowed_max_layer_depth = 0;
2211
2212 // For key frames the frame target rate is already set and it
2213 // is also the golden frame.
2214 // === [frame_index == 0] ===
2215 if (!key_frame)
2216 set_gf_overlay_frame_type(gf_group, frame_index, rc->source_alt_ref_active);
2217
2218 ++frame_index;
2219
2220 // === [frame_index == 1] ===
2221 if (rc->source_alt_ref_pending) {
2222 gf_group->update_type[frame_index] = ARF_UPDATE;
2223 gf_group->rf_level[frame_index] = GF_ARF_STD;
2224 gf_group->layer_depth[frame_index] = layer_depth;
2225 gf_group->arf_src_offset[frame_index] =
2226 (unsigned char)(rc->baseline_gf_interval - 1);
2227 gf_group->frame_gop_index[frame_index] = rc->baseline_gf_interval;
2228 gf_group->max_layer_depth = 1;
2229 ++frame_index;
2230 ++layer_depth;
2231 gf_group->allowed_max_layer_depth = cpi->oxcf.enable_auto_arf;
2232 }
2233
2234 find_arf_order(cpi, gf_group, &frame_index, layer_depth, 1, gop_frames);
2235
2236 set_gf_overlay_frame_type(gf_group, frame_index, rc->source_alt_ref_pending);
2237 gf_group->arf_src_offset[frame_index] = 0;
2238 gf_group->frame_gop_index[frame_index] = rc->baseline_gf_interval;
2239
2240 // Set the frame ops number.
2241 gf_group->gf_group_size = frame_index;
2242 }
2243
allocate_gf_group_bits(VP9_COMP * cpi,int64_t gf_group_bits,int gf_arf_bits)2244 static void allocate_gf_group_bits(VP9_COMP *cpi, int64_t gf_group_bits,
2245 int gf_arf_bits) {
2246 VP9EncoderConfig *const oxcf = &cpi->oxcf;
2247 RATE_CONTROL *const rc = &cpi->rc;
2248 TWO_PASS *const twopass = &cpi->twopass;
2249 GF_GROUP *const gf_group = &twopass->gf_group;
2250 FIRSTPASS_STATS frame_stats;
2251 int i;
2252 int frame_index = 0;
2253 int target_frame_size;
2254 int key_frame;
2255 const int max_bits = frame_max_bits(&cpi->rc, oxcf);
2256 int64_t total_group_bits = gf_group_bits;
2257 int mid_frame_idx;
2258 int normal_frames;
2259 int normal_frame_bits;
2260 int last_frame_reduction = 0;
2261 double av_score = 1.0;
2262 double tot_norm_frame_score = 1.0;
2263 double this_frame_score = 1.0;
2264
2265 // Define the GF structure and specify
2266 int gop_frames = gf_group->gf_group_size;
2267
2268 key_frame = cpi->common.frame_type == KEY_FRAME;
2269
2270 // For key frames the frame target rate is already set and it
2271 // is also the golden frame.
2272 // === [frame_index == 0] ===
2273 if (!key_frame) {
2274 gf_group->bit_allocation[frame_index] =
2275 rc->source_alt_ref_active ? 0 : gf_arf_bits;
2276 }
2277
2278 // Deduct the boost bits for arf (or gf if it is not a key frame)
2279 // from the group total.
2280 if (rc->source_alt_ref_pending || !key_frame) total_group_bits -= gf_arf_bits;
2281
2282 ++frame_index;
2283
2284 // === [frame_index == 1] ===
2285 // Store the bits to spend on the ARF if there is one.
2286 if (rc->source_alt_ref_pending) {
2287 gf_group->bit_allocation[frame_index] = gf_arf_bits;
2288
2289 ++frame_index;
2290 }
2291
2292 // Define middle frame
2293 mid_frame_idx = frame_index + (rc->baseline_gf_interval >> 1) - 1;
2294
2295 normal_frames = (rc->baseline_gf_interval - rc->source_alt_ref_pending);
2296 if (normal_frames > 1)
2297 normal_frame_bits = (int)(total_group_bits / normal_frames);
2298 else
2299 normal_frame_bits = (int)total_group_bits;
2300
2301 gf_group->gfu_boost[1] = rc->gfu_boost;
2302
2303 if (cpi->multi_layer_arf) {
2304 int idx;
2305 int arf_depth_bits[MAX_ARF_LAYERS] = { 0 };
2306 int arf_depth_count[MAX_ARF_LAYERS] = { 0 };
2307 int arf_depth_boost[MAX_ARF_LAYERS] = { 0 };
2308 int total_arfs = 1; // Account for the base layer ARF.
2309
2310 for (idx = 0; idx < gop_frames; ++idx) {
2311 if (gf_group->update_type[idx] == ARF_UPDATE) {
2312 arf_depth_boost[gf_group->layer_depth[idx]] += gf_group->gfu_boost[idx];
2313 ++arf_depth_count[gf_group->layer_depth[idx]];
2314 }
2315 }
2316
2317 for (idx = 2; idx < MAX_ARF_LAYERS; ++idx) {
2318 if (arf_depth_boost[idx] == 0) break;
2319 arf_depth_bits[idx] = calculate_boost_bits(
2320 rc->baseline_gf_interval - total_arfs - arf_depth_count[idx],
2321 arf_depth_boost[idx], total_group_bits);
2322
2323 total_group_bits -= arf_depth_bits[idx];
2324 total_arfs += arf_depth_count[idx];
2325 }
2326
2327 // offset the base layer arf
2328 normal_frames -= (total_arfs - 1);
2329 if (normal_frames > 1)
2330 normal_frame_bits = (int)(total_group_bits / normal_frames);
2331 else
2332 normal_frame_bits = (int)total_group_bits;
2333
2334 target_frame_size = normal_frame_bits;
2335 target_frame_size =
2336 clamp(target_frame_size, 0, VPXMIN(max_bits, (int)total_group_bits));
2337
2338 // The first layer ARF has its bit allocation assigned.
2339 for (idx = frame_index; idx < gop_frames; ++idx) {
2340 switch (gf_group->update_type[idx]) {
2341 case ARF_UPDATE:
2342 gf_group->bit_allocation[idx] =
2343 (int)((arf_depth_bits[gf_group->layer_depth[idx]] *
2344 gf_group->gfu_boost[idx]) /
2345 arf_depth_boost[gf_group->layer_depth[idx]]);
2346 break;
2347 case USE_BUF_FRAME: gf_group->bit_allocation[idx] = 0; break;
2348 default: gf_group->bit_allocation[idx] = target_frame_size; break;
2349 }
2350 }
2351 gf_group->bit_allocation[idx] = 0;
2352
2353 return;
2354 }
2355
2356 if (oxcf->vbr_corpus_complexity) {
2357 av_score = get_distribution_av_err(cpi, twopass);
2358 tot_norm_frame_score = calculate_group_score(cpi, av_score, normal_frames);
2359 }
2360
2361 // Allocate bits to the other frames in the group.
2362 for (i = 0; i < normal_frames; ++i) {
2363 if (EOF == input_stats(twopass, &frame_stats)) break;
2364 if (oxcf->vbr_corpus_complexity) {
2365 this_frame_score = calculate_norm_frame_score(cpi, twopass, oxcf,
2366 &frame_stats, av_score);
2367 normal_frame_bits = (int)((double)total_group_bits *
2368 (this_frame_score / tot_norm_frame_score));
2369 }
2370
2371 target_frame_size = normal_frame_bits;
2372 if ((i == (normal_frames - 1)) && (i >= 1)) {
2373 last_frame_reduction = normal_frame_bits / 16;
2374 target_frame_size -= last_frame_reduction;
2375 }
2376
2377 target_frame_size =
2378 clamp(target_frame_size, 0, VPXMIN(max_bits, (int)total_group_bits));
2379
2380 gf_group->bit_allocation[frame_index] = target_frame_size;
2381 ++frame_index;
2382 }
2383
2384 // Add in some extra bits for the middle frame in the group.
2385 gf_group->bit_allocation[mid_frame_idx] += last_frame_reduction;
2386
2387 // Note:
2388 // We need to configure the frame at the end of the sequence + 1 that will be
2389 // the start frame for the next group. Otherwise prior to the call to
2390 // vp9_rc_get_second_pass_params() the data will be undefined.
2391 }
2392
2393 // Adjusts the ARNF filter for a GF group.
adjust_group_arnr_filter(VP9_COMP * cpi,double section_noise,double section_inter,double section_motion)2394 static void adjust_group_arnr_filter(VP9_COMP *cpi, double section_noise,
2395 double section_inter,
2396 double section_motion) {
2397 TWO_PASS *const twopass = &cpi->twopass;
2398 double section_zeromv = section_inter - section_motion;
2399
2400 twopass->arnr_strength_adjustment = 0;
2401
2402 if ((section_zeromv < 0.10) || (section_noise <= (SECTION_NOISE_DEF * 0.75)))
2403 twopass->arnr_strength_adjustment -= 1;
2404 if (section_zeromv > 0.50) twopass->arnr_strength_adjustment += 1;
2405 }
2406
2407 // Analyse and define a gf/arf group.
2408 #define ARF_ABS_ZOOM_THRESH 4.0
2409
2410 #define MAX_GF_BOOST 5400
define_gf_group(VP9_COMP * cpi,FIRSTPASS_STATS * this_frame)2411 static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
2412 VP9_COMMON *const cm = &cpi->common;
2413 RATE_CONTROL *const rc = &cpi->rc;
2414 VP9EncoderConfig *const oxcf = &cpi->oxcf;
2415 TWO_PASS *const twopass = &cpi->twopass;
2416 FIRSTPASS_STATS next_frame;
2417 const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
2418 int i;
2419
2420 double gf_group_err = 0.0;
2421 double gf_group_raw_error = 0.0;
2422 double gf_group_noise = 0.0;
2423 double gf_group_skip_pct = 0.0;
2424 double gf_group_inactive_zone_rows = 0.0;
2425 double gf_group_inter = 0.0;
2426 double gf_group_motion = 0.0;
2427 double gf_first_frame_err = 0.0;
2428 double mod_frame_err = 0.0;
2429
2430 double mv_ratio_accumulator = 0.0;
2431 double zero_motion_accumulator = 1.0;
2432 double loop_decay_rate = 1.00;
2433 double last_loop_decay_rate = 1.00;
2434
2435 double this_frame_mv_in_out = 0.0;
2436 double mv_in_out_accumulator = 0.0;
2437 double abs_mv_in_out_accumulator = 0.0;
2438 double mv_ratio_accumulator_thresh;
2439 double abs_mv_in_out_thresh;
2440 double sr_accumulator = 0.0;
2441 const double av_err = get_distribution_av_err(cpi, twopass);
2442 unsigned int allow_alt_ref = is_altref_enabled(cpi);
2443
2444 int flash_detected;
2445 int active_max_gf_interval;
2446 int active_min_gf_interval;
2447 int64_t gf_group_bits;
2448 int gf_arf_bits;
2449 const int is_key_frame = frame_is_intra_only(cm);
2450 const int arf_active_or_kf = is_key_frame || rc->source_alt_ref_active;
2451
2452 double gop_intra_factor = 1.0;
2453
2454 // Reset the GF group data structures unless this is a key
2455 // frame in which case it will already have been done.
2456 if (is_key_frame == 0) {
2457 vp9_zero(twopass->gf_group);
2458 }
2459
2460 vpx_clear_system_state();
2461 vp9_zero(next_frame);
2462
2463 // Load stats for the current frame.
2464 mod_frame_err =
2465 calculate_norm_frame_score(cpi, twopass, oxcf, this_frame, av_err);
2466
2467 // Note the error of the frame at the start of the group. This will be
2468 // the GF frame error if we code a normal gf.
2469 gf_first_frame_err = mod_frame_err;
2470
2471 // If this is a key frame or the overlay from a previous arf then
2472 // the error score / cost of this frame has already been accounted for.
2473 if (arf_active_or_kf) {
2474 gf_group_err -= gf_first_frame_err;
2475 gf_group_raw_error -= this_frame->coded_error;
2476 gf_group_noise -= this_frame->frame_noise_energy;
2477 gf_group_skip_pct -= this_frame->intra_skip_pct;
2478 gf_group_inactive_zone_rows -= this_frame->inactive_zone_rows;
2479 gf_group_inter -= this_frame->pcnt_inter;
2480 gf_group_motion -= this_frame->pcnt_motion;
2481 }
2482
2483 // Motion breakout threshold for loop below depends on image size.
2484 mv_ratio_accumulator_thresh =
2485 (cpi->initial_height + cpi->initial_width) / 4.0;
2486 abs_mv_in_out_thresh = ARF_ABS_ZOOM_THRESH;
2487
2488 // Set a maximum and minimum interval for the GF group.
2489 // If the image appears almost completely static we can extend beyond this.
2490 {
2491 int int_max_q = (int)(vp9_convert_qindex_to_q(twopass->active_worst_quality,
2492 cpi->common.bit_depth));
2493 int q_term = (cm->current_video_frame == 0)
2494 ? int_max_q / 32
2495 : (int)(vp9_convert_qindex_to_q(rc->last_boosted_qindex,
2496 cpi->common.bit_depth) /
2497 6);
2498 active_min_gf_interval =
2499 rc->min_gf_interval + arf_active_or_kf + VPXMIN(2, int_max_q / 200);
2500 active_min_gf_interval =
2501 VPXMIN(active_min_gf_interval, rc->max_gf_interval + arf_active_or_kf);
2502
2503 // The value chosen depends on the active Q range. At low Q we have
2504 // bits to spare and are better with a smaller interval and smaller boost.
2505 // At high Q when there are few bits to spare we are better with a longer
2506 // interval to spread the cost of the GF.
2507 active_max_gf_interval = 11 + arf_active_or_kf + VPXMIN(5, q_term);
2508
2509 // Force max GF interval to be odd.
2510 active_max_gf_interval = active_max_gf_interval | 0x01;
2511
2512 // We have: active_min_gf_interval <=
2513 // rc->max_gf_interval + arf_active_or_kf.
2514 if (active_max_gf_interval < active_min_gf_interval) {
2515 active_max_gf_interval = active_min_gf_interval;
2516 } else {
2517 active_max_gf_interval = VPXMIN(active_max_gf_interval,
2518 rc->max_gf_interval + arf_active_or_kf);
2519 }
2520
2521 // Would the active max drop us out just before the near the next kf?
2522 if ((active_max_gf_interval <= rc->frames_to_key) &&
2523 (active_max_gf_interval >= (rc->frames_to_key - rc->min_gf_interval)))
2524 active_max_gf_interval = rc->frames_to_key / 2;
2525 }
2526
2527 if (cpi->multi_layer_arf) {
2528 int layers = 0;
2529 int max_layers = VPXMIN(MAX_ARF_LAYERS, cpi->oxcf.enable_auto_arf);
2530
2531 // Adapt the intra_error factor to active_max_gf_interval limit.
2532 for (i = active_max_gf_interval; i > 0; i >>= 1) ++layers;
2533
2534 layers = VPXMIN(max_layers, layers);
2535 gop_intra_factor += (layers * 0.25);
2536 }
2537
2538 i = 0;
2539 while (i < rc->static_scene_max_gf_interval && i < rc->frames_to_key) {
2540 ++i;
2541
2542 // Accumulate error score of frames in this gf group.
2543 mod_frame_err =
2544 calculate_norm_frame_score(cpi, twopass, oxcf, this_frame, av_err);
2545 gf_group_err += mod_frame_err;
2546 gf_group_raw_error += this_frame->coded_error;
2547 gf_group_noise += this_frame->frame_noise_energy;
2548 gf_group_skip_pct += this_frame->intra_skip_pct;
2549 gf_group_inactive_zone_rows += this_frame->inactive_zone_rows;
2550 gf_group_inter += this_frame->pcnt_inter;
2551 gf_group_motion += this_frame->pcnt_motion;
2552
2553 if (EOF == input_stats(twopass, &next_frame)) break;
2554
2555 // Test for the case where there is a brief flash but the prediction
2556 // quality back to an earlier frame is then restored.
2557 flash_detected = detect_flash(twopass, 0);
2558
2559 // Update the motion related elements to the boost calculation.
2560 accumulate_frame_motion_stats(
2561 &next_frame, &this_frame_mv_in_out, &mv_in_out_accumulator,
2562 &abs_mv_in_out_accumulator, &mv_ratio_accumulator);
2563
2564 // Monitor for static sections.
2565 if ((rc->frames_since_key + i - 1) > 1) {
2566 zero_motion_accumulator = VPXMIN(
2567 zero_motion_accumulator, get_zero_motion_factor(cpi, &next_frame));
2568 }
2569
2570 // Accumulate the effect of prediction quality decay.
2571 if (!flash_detected) {
2572 last_loop_decay_rate = loop_decay_rate;
2573 loop_decay_rate = get_prediction_decay_rate(cpi, &next_frame);
2574
2575 // Break clause to detect very still sections after motion. For example,
2576 // a static image after a fade or other transition.
2577 if (detect_transition_to_still(cpi, i, 5, loop_decay_rate,
2578 last_loop_decay_rate)) {
2579 allow_alt_ref = 0;
2580 break;
2581 }
2582
2583 // Update the accumulator for second ref error difference.
2584 // This is intended to give an indication of how much the coded error is
2585 // increasing over time.
2586 if (i == 1) {
2587 sr_accumulator += next_frame.coded_error;
2588 } else {
2589 sr_accumulator += (next_frame.sr_coded_error - next_frame.coded_error);
2590 }
2591 }
2592
2593 // Break out conditions.
2594 // Break at maximum of active_max_gf_interval unless almost totally static.
2595 //
2596 // Note that the addition of a test of rc->source_alt_ref_active is
2597 // deliberate. The effect of this is that after a normal altref group even
2598 // if the material is static there will be one normal length GF group
2599 // before allowing longer GF groups. The reason for this is that in cases
2600 // such as slide shows where slides are separated by a complex transition
2601 // such as a fade, the arf group spanning the transition may not be coded
2602 // at a very high quality and hence this frame (with its overlay) is a
2603 // poor golden frame to use for an extended group.
2604 if (((i >= active_max_gf_interval) &&
2605 ((zero_motion_accumulator < 0.995) || (rc->source_alt_ref_active))) ||
2606 (
2607 // Don't break out with a very short interval.
2608 (i >= active_min_gf_interval) &&
2609 // If possible dont break very close to a kf
2610 ((rc->frames_to_key - i) >= rc->min_gf_interval) && (i & 0x01) &&
2611 (!flash_detected) &&
2612 ((mv_ratio_accumulator > mv_ratio_accumulator_thresh) ||
2613 (abs_mv_in_out_accumulator > abs_mv_in_out_thresh) ||
2614 (sr_accumulator > gop_intra_factor * next_frame.intra_error)))) {
2615 break;
2616 }
2617
2618 *this_frame = next_frame;
2619 }
2620
2621 // Was the group length constrained by the requirement for a new KF?
2622 rc->constrained_gf_group = (i >= rc->frames_to_key) ? 1 : 0;
2623
2624 // Should we use the alternate reference frame.
2625 if ((zero_motion_accumulator < 0.995) && allow_alt_ref &&
2626 (twopass->kf_zeromotion_pct < STATIC_KF_GROUP_THRESH) &&
2627 (i < cpi->oxcf.lag_in_frames) && (i >= rc->min_gf_interval)) {
2628 const int forward_frames = (rc->frames_to_key - i >= i - 1)
2629 ? i - 1
2630 : VPXMAX(0, rc->frames_to_key - i);
2631
2632 // Calculate the boost for alt ref.
2633 rc->gfu_boost = calc_arf_boost(cpi, forward_frames, (i - 1));
2634 rc->source_alt_ref_pending = 1;
2635 } else {
2636 rc->gfu_boost = VPXMIN(MAX_GF_BOOST, calc_arf_boost(cpi, 0, (i - 1)));
2637 rc->source_alt_ref_pending = 0;
2638 }
2639
2640 #ifdef AGGRESSIVE_VBR
2641 // Limit maximum boost based on interval length.
2642 rc->gfu_boost = VPXMIN((int)rc->gfu_boost, i * 140);
2643 #else
2644 rc->gfu_boost = VPXMIN((int)rc->gfu_boost, i * 200);
2645 #endif
2646
2647 rc->baseline_gf_interval = i - rc->source_alt_ref_pending;
2648
2649 // Reset the file position.
2650 reset_fpf_position(twopass, start_pos);
2651
2652 // Calculate the bits to be allocated to the gf/arf group as a whole
2653 gf_group_bits = calculate_total_gf_group_bits(cpi, gf_group_err);
2654
2655 // Calculate an estimate of the maxq needed for the group.
2656 // We are more aggressive about correcting for sections
2657 // where there could be significant overshoot than for easier
2658 // sections where we do not wish to risk creating an overshoot
2659 // of the allocated bit budget.
2660 if ((cpi->oxcf.rc_mode != VPX_Q) && (rc->baseline_gf_interval > 1)) {
2661 const int vbr_group_bits_per_frame =
2662 (int)(gf_group_bits / rc->baseline_gf_interval);
2663 const double group_av_err = gf_group_raw_error / rc->baseline_gf_interval;
2664 const double group_av_noise = gf_group_noise / rc->baseline_gf_interval;
2665 const double group_av_skip_pct =
2666 gf_group_skip_pct / rc->baseline_gf_interval;
2667 const double group_av_inactive_zone =
2668 ((gf_group_inactive_zone_rows * 2) /
2669 (rc->baseline_gf_interval * (double)cm->mb_rows));
2670 int tmp_q = get_twopass_worst_quality(
2671 cpi, group_av_err, (group_av_skip_pct + group_av_inactive_zone),
2672 group_av_noise, vbr_group_bits_per_frame);
2673 twopass->active_worst_quality =
2674 (tmp_q + (twopass->active_worst_quality * 3)) >> 2;
2675
2676 #if CONFIG_ALWAYS_ADJUST_BPM
2677 // Reset rolling actual and target bits counters for ARF groups.
2678 twopass->rolling_arf_group_target_bits = 0;
2679 twopass->rolling_arf_group_actual_bits = 0;
2680 #endif
2681 }
2682
2683 // Context Adjustment of ARNR filter strength
2684 if (rc->baseline_gf_interval > 1) {
2685 adjust_group_arnr_filter(cpi, (gf_group_noise / rc->baseline_gf_interval),
2686 (gf_group_inter / rc->baseline_gf_interval),
2687 (gf_group_motion / rc->baseline_gf_interval));
2688 } else {
2689 twopass->arnr_strength_adjustment = 0;
2690 }
2691
2692 // Calculate the extra bits to be used for boosted frame(s)
2693 gf_arf_bits = calculate_boost_bits((rc->baseline_gf_interval - 1),
2694 rc->gfu_boost, gf_group_bits);
2695
2696 // Adjust KF group bits and error remaining.
2697 twopass->kf_group_error_left -= gf_group_err;
2698
2699 // Decide GOP structure.
2700 define_gf_group_structure(cpi);
2701
2702 // Allocate bits to each of the frames in the GF group.
2703 allocate_gf_group_bits(cpi, gf_group_bits, gf_arf_bits);
2704
2705 // Reset the file position.
2706 reset_fpf_position(twopass, start_pos);
2707
2708 // Calculate a section intra ratio used in setting max loop filter.
2709 if (cpi->common.frame_type != KEY_FRAME) {
2710 twopass->section_intra_rating = calculate_section_intra_ratio(
2711 start_pos, twopass->stats_in_end, rc->baseline_gf_interval);
2712 }
2713
2714 if (oxcf->resize_mode == RESIZE_DYNAMIC) {
2715 // Default to starting GF groups at normal frame size.
2716 cpi->rc.next_frame_size_selector = UNSCALED;
2717 }
2718 #if !CONFIG_ALWAYS_ADJUST_BPM
2719 // Reset rolling actual and target bits counters for ARF groups.
2720 twopass->rolling_arf_group_target_bits = 0;
2721 twopass->rolling_arf_group_actual_bits = 0;
2722 #endif
2723 }
2724
2725 // Intra / Inter threshold very low
2726 #define VERY_LOW_II 1.5
2727 // Clean slide transitions we expect a sharp single frame spike in error.
2728 #define ERROR_SPIKE 5.0
2729
2730 // Slide show transition detection.
2731 // Tests for case where there is very low error either side of the current frame
2732 // but much higher just for this frame. This can help detect key frames in
2733 // slide shows even where the slides are pictures of different sizes.
2734 // Also requires that intra and inter errors are very similar to help eliminate
2735 // harmful false positives.
2736 // 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)2737 static int slide_transition(const FIRSTPASS_STATS *this_frame,
2738 const FIRSTPASS_STATS *last_frame,
2739 const FIRSTPASS_STATS *next_frame) {
2740 return (this_frame->intra_error < (this_frame->coded_error * VERY_LOW_II)) &&
2741 (this_frame->coded_error > (last_frame->coded_error * ERROR_SPIKE)) &&
2742 (this_frame->coded_error > (next_frame->coded_error * ERROR_SPIKE));
2743 }
2744
2745 // Minimum % intra coding observed in first pass (1.0 = 100%)
2746 #define MIN_INTRA_LEVEL 0.25
2747 // Threshold for use of the lagging second reference frame. Scene cuts do not
2748 // usually have a high second ref useage.
2749 #define SECOND_REF_USEAGE_THRESH 0.125
2750 // Hard threshold where the first pass chooses intra for almost all blocks.
2751 // In such a case even if the frame is not a scene cut coding a key frame
2752 // may be a good option.
2753 #define VERY_LOW_INTER_THRESH 0.05
2754 // Maximum threshold for the relative ratio of intra error score vs best
2755 // inter error score.
2756 #define KF_II_ERR_THRESHOLD 2.5
2757 #define KF_II_MAX 128.0
2758 #define II_FACTOR 12.5
2759 // Test for very low intra complexity which could cause false key frames
2760 #define V_LOW_INTRA 0.5
2761
test_candidate_kf(TWO_PASS * twopass,const FIRSTPASS_STATS * last_frame,const FIRSTPASS_STATS * this_frame,const FIRSTPASS_STATS * next_frame)2762 static int test_candidate_kf(TWO_PASS *twopass,
2763 const FIRSTPASS_STATS *last_frame,
2764 const FIRSTPASS_STATS *this_frame,
2765 const FIRSTPASS_STATS *next_frame) {
2766 int is_viable_kf = 0;
2767 double pcnt_intra = 1.0 - this_frame->pcnt_inter;
2768
2769 // Does the frame satisfy the primary criteria of a key frame?
2770 // See above for an explanation of the test criteria.
2771 // If so, then examine how well it predicts subsequent frames.
2772 if (!detect_flash(twopass, -1) && !detect_flash(twopass, 0) &&
2773 (this_frame->pcnt_second_ref < SECOND_REF_USEAGE_THRESH) &&
2774 ((this_frame->pcnt_inter < VERY_LOW_INTER_THRESH) ||
2775 (slide_transition(this_frame, last_frame, next_frame)) ||
2776 (((this_frame->coded_error > (next_frame->coded_error * 1.1)) &&
2777 (this_frame->coded_error > (last_frame->coded_error * 1.1))) &&
2778 (pcnt_intra > MIN_INTRA_LEVEL) &&
2779 ((pcnt_intra + this_frame->pcnt_neutral) > 0.5) &&
2780 ((this_frame->intra_error /
2781 DOUBLE_DIVIDE_CHECK(this_frame->coded_error)) <
2782 KF_II_ERR_THRESHOLD)))) {
2783 int i;
2784 const FIRSTPASS_STATS *start_pos = twopass->stats_in;
2785 FIRSTPASS_STATS local_next_frame = *next_frame;
2786 double boost_score = 0.0;
2787 double old_boost_score = 0.0;
2788 double decay_accumulator = 1.0;
2789
2790 // Examine how well the key frame predicts subsequent frames.
2791 for (i = 0; i < 16; ++i) {
2792 double next_iiratio = (II_FACTOR * local_next_frame.intra_error /
2793 DOUBLE_DIVIDE_CHECK(local_next_frame.coded_error));
2794
2795 if (next_iiratio > KF_II_MAX) next_iiratio = KF_II_MAX;
2796
2797 // Cumulative effect of decay in prediction quality.
2798 if (local_next_frame.pcnt_inter > 0.85)
2799 decay_accumulator *= local_next_frame.pcnt_inter;
2800 else
2801 decay_accumulator *= (0.85 + local_next_frame.pcnt_inter) / 2.0;
2802
2803 // Keep a running total.
2804 boost_score += (decay_accumulator * next_iiratio);
2805
2806 // Test various breakout clauses.
2807 if ((local_next_frame.pcnt_inter < 0.05) || (next_iiratio < 1.5) ||
2808 (((local_next_frame.pcnt_inter - local_next_frame.pcnt_neutral) <
2809 0.20) &&
2810 (next_iiratio < 3.0)) ||
2811 ((boost_score - old_boost_score) < 3.0) ||
2812 (local_next_frame.intra_error < V_LOW_INTRA)) {
2813 break;
2814 }
2815
2816 old_boost_score = boost_score;
2817
2818 // Get the next frame details
2819 if (EOF == input_stats(twopass, &local_next_frame)) break;
2820 }
2821
2822 // If there is tolerable prediction for at least the next 3 frames then
2823 // break out else discard this potential key frame and move on
2824 if (boost_score > 30.0 && (i > 3)) {
2825 is_viable_kf = 1;
2826 } else {
2827 // Reset the file position
2828 reset_fpf_position(twopass, start_pos);
2829
2830 is_viable_kf = 0;
2831 }
2832 }
2833
2834 return is_viable_kf;
2835 }
2836
2837 #define FRAMES_TO_CHECK_DECAY 8
2838 #define MIN_KF_TOT_BOOST 300
2839 #define KF_BOOST_SCAN_MAX_FRAMES 32
2840 #define KF_ABS_ZOOM_THRESH 6.0
2841
2842 #ifdef AGGRESSIVE_VBR
2843 #define KF_MAX_FRAME_BOOST 80.0
2844 #define MAX_KF_TOT_BOOST 4800
2845 #else
2846 #define KF_MAX_FRAME_BOOST 96.0
2847 #define MAX_KF_TOT_BOOST 5400
2848 #endif
2849
find_next_key_frame(VP9_COMP * cpi,FIRSTPASS_STATS * this_frame)2850 static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
2851 int i, j;
2852 RATE_CONTROL *const rc = &cpi->rc;
2853 TWO_PASS *const twopass = &cpi->twopass;
2854 GF_GROUP *const gf_group = &twopass->gf_group;
2855 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
2856 const FIRSTPASS_STATS first_frame = *this_frame;
2857 const FIRSTPASS_STATS *const start_position = twopass->stats_in;
2858 FIRSTPASS_STATS next_frame;
2859 FIRSTPASS_STATS last_frame;
2860 int kf_bits = 0;
2861 double decay_accumulator = 1.0;
2862 double zero_motion_accumulator = 1.0;
2863 double boost_score = 0.0;
2864 double kf_mod_err = 0.0;
2865 double kf_raw_err = 0.0;
2866 double kf_group_err = 0.0;
2867 double recent_loop_decay[FRAMES_TO_CHECK_DECAY];
2868 double sr_accumulator = 0.0;
2869 double abs_mv_in_out_accumulator = 0.0;
2870 const double av_err = get_distribution_av_err(cpi, twopass);
2871 vp9_zero(next_frame);
2872
2873 cpi->common.frame_type = KEY_FRAME;
2874 rc->frames_since_key = 0;
2875
2876 // Reset the GF group data structures.
2877 vp9_zero(*gf_group);
2878
2879 // Is this a forced key frame by interval.
2880 rc->this_key_frame_forced = rc->next_key_frame_forced;
2881
2882 // Clear the alt ref active flag and last group multi arf flags as they
2883 // can never be set for a key frame.
2884 rc->source_alt_ref_active = 0;
2885
2886 // KF is always a GF so clear frames till next gf counter.
2887 rc->frames_till_gf_update_due = 0;
2888
2889 rc->frames_to_key = 1;
2890
2891 twopass->kf_group_bits = 0; // Total bits available to kf group
2892 twopass->kf_group_error_left = 0.0; // Group modified error score.
2893
2894 kf_raw_err = this_frame->intra_error;
2895 kf_mod_err =
2896 calculate_norm_frame_score(cpi, twopass, oxcf, this_frame, av_err);
2897
2898 // Initialize the decay rates for the recent frames to check
2899 for (j = 0; j < FRAMES_TO_CHECK_DECAY; ++j) recent_loop_decay[j] = 1.0;
2900
2901 // Find the next keyframe.
2902 i = 0;
2903 while (twopass->stats_in < twopass->stats_in_end &&
2904 rc->frames_to_key < cpi->oxcf.key_freq) {
2905 // Accumulate kf group error.
2906 kf_group_err +=
2907 calculate_norm_frame_score(cpi, twopass, oxcf, this_frame, av_err);
2908
2909 // Load the next frame's stats.
2910 last_frame = *this_frame;
2911 input_stats(twopass, this_frame);
2912
2913 // Provided that we are not at the end of the file...
2914 if (cpi->oxcf.auto_key && twopass->stats_in < twopass->stats_in_end) {
2915 double loop_decay_rate;
2916
2917 // Check for a scene cut.
2918 if (test_candidate_kf(twopass, &last_frame, this_frame,
2919 twopass->stats_in))
2920 break;
2921
2922 // How fast is the prediction quality decaying?
2923 loop_decay_rate = get_prediction_decay_rate(cpi, twopass->stats_in);
2924
2925 // We want to know something about the recent past... rather than
2926 // as used elsewhere where we are concerned with decay in prediction
2927 // quality since the last GF or KF.
2928 recent_loop_decay[i % FRAMES_TO_CHECK_DECAY] = loop_decay_rate;
2929 decay_accumulator = 1.0;
2930 for (j = 0; j < FRAMES_TO_CHECK_DECAY; ++j)
2931 decay_accumulator *= recent_loop_decay[j];
2932
2933 // Special check for transition or high motion followed by a
2934 // static scene.
2935 if (detect_transition_to_still(cpi, i, cpi->oxcf.key_freq - i,
2936 loop_decay_rate, decay_accumulator))
2937 break;
2938
2939 // Step on to the next frame.
2940 ++rc->frames_to_key;
2941
2942 // If we don't have a real key frame within the next two
2943 // key_freq intervals then break out of the loop.
2944 if (rc->frames_to_key >= 2 * cpi->oxcf.key_freq) break;
2945 } else {
2946 ++rc->frames_to_key;
2947 }
2948 ++i;
2949 }
2950
2951 // If there is a max kf interval set by the user we must obey it.
2952 // We already breakout of the loop above at 2x max.
2953 // This code centers the extra kf if the actual natural interval
2954 // is between 1x and 2x.
2955 if (cpi->oxcf.auto_key && rc->frames_to_key > cpi->oxcf.key_freq) {
2956 FIRSTPASS_STATS tmp_frame = first_frame;
2957
2958 rc->frames_to_key /= 2;
2959
2960 // Reset to the start of the group.
2961 reset_fpf_position(twopass, start_position);
2962
2963 kf_group_err = 0.0;
2964
2965 // Rescan to get the correct error data for the forced kf group.
2966 for (i = 0; i < rc->frames_to_key; ++i) {
2967 kf_group_err +=
2968 calculate_norm_frame_score(cpi, twopass, oxcf, &tmp_frame, av_err);
2969 input_stats(twopass, &tmp_frame);
2970 }
2971 rc->next_key_frame_forced = 1;
2972 } else if (twopass->stats_in == twopass->stats_in_end ||
2973 rc->frames_to_key >= cpi->oxcf.key_freq) {
2974 rc->next_key_frame_forced = 1;
2975 } else {
2976 rc->next_key_frame_forced = 0;
2977 }
2978
2979 // Special case for the last key frame of the file.
2980 if (twopass->stats_in >= twopass->stats_in_end) {
2981 // Accumulate kf group error.
2982 kf_group_err +=
2983 calculate_norm_frame_score(cpi, twopass, oxcf, this_frame, av_err);
2984 }
2985
2986 // Calculate the number of bits that should be assigned to the kf group.
2987 if (twopass->bits_left > 0 && twopass->normalized_score_left > 0.0) {
2988 // Maximum number of bits for a single normal frame (not key frame).
2989 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
2990
2991 // Maximum number of bits allocated to the key frame group.
2992 int64_t max_grp_bits;
2993
2994 // Default allocation based on bits left and relative
2995 // complexity of the section.
2996 twopass->kf_group_bits = (int64_t)(
2997 twopass->bits_left * (kf_group_err / twopass->normalized_score_left));
2998
2999 // Clip based on maximum per frame rate defined by the user.
3000 max_grp_bits = (int64_t)max_bits * (int64_t)rc->frames_to_key;
3001 if (twopass->kf_group_bits > max_grp_bits)
3002 twopass->kf_group_bits = max_grp_bits;
3003 } else {
3004 twopass->kf_group_bits = 0;
3005 }
3006 twopass->kf_group_bits = VPXMAX(0, twopass->kf_group_bits);
3007
3008 // Reset the first pass file position.
3009 reset_fpf_position(twopass, start_position);
3010
3011 // Scan through the kf group collating various stats used to determine
3012 // how many bits to spend on it.
3013 boost_score = 0.0;
3014
3015 for (i = 0; i < (rc->frames_to_key - 1); ++i) {
3016 if (EOF == input_stats(twopass, &next_frame)) break;
3017
3018 // The zero motion test here insures that if we mark a kf group as static
3019 // it is static throughout not just the first KF_BOOST_SCAN_MAX_FRAMES.
3020 // It also allows for a larger boost on long static groups.
3021 if ((i <= KF_BOOST_SCAN_MAX_FRAMES) || (zero_motion_accumulator >= 0.99)) {
3022 double frame_boost;
3023 double zm_factor;
3024
3025 // Monitor for static sections.
3026 // First frame in kf group the second ref indicator is invalid.
3027 if (i > 0) {
3028 zero_motion_accumulator = VPXMIN(
3029 zero_motion_accumulator, get_zero_motion_factor(cpi, &next_frame));
3030 } else {
3031 zero_motion_accumulator =
3032 next_frame.pcnt_inter - next_frame.pcnt_motion;
3033 }
3034
3035 // Factor 0.75-1.25 based on how much of frame is static.
3036 zm_factor = (0.75 + (zero_motion_accumulator / 2.0));
3037
3038 // The second (lagging) ref error is not valid immediately after
3039 // a key frame because either the lag has not built up (in the case of
3040 // the first key frame or it points to a refernce before the new key
3041 // frame.
3042 if (i < 2) sr_accumulator = 0.0;
3043 frame_boost = calc_kf_frame_boost(cpi, &next_frame, &sr_accumulator, 0,
3044 KF_MAX_FRAME_BOOST * zm_factor);
3045
3046 boost_score += frame_boost;
3047
3048 // Measure of zoom. Large zoom tends to indicate reduced boost.
3049 abs_mv_in_out_accumulator +=
3050 fabs(next_frame.mv_in_out_count * next_frame.pcnt_motion);
3051
3052 if ((frame_boost < 25.00) ||
3053 (abs_mv_in_out_accumulator > KF_ABS_ZOOM_THRESH) ||
3054 (sr_accumulator > (kf_raw_err * 1.50)))
3055 break;
3056 } else {
3057 break;
3058 }
3059 }
3060
3061 reset_fpf_position(twopass, start_position);
3062
3063 // Store the zero motion percentage
3064 twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
3065
3066 // Calculate a section intra ratio used in setting max loop filter.
3067 twopass->section_intra_rating = calculate_section_intra_ratio(
3068 start_position, twopass->stats_in_end, rc->frames_to_key);
3069
3070 // Special case for static / slide show content but dont apply
3071 // if the kf group is very short.
3072 if ((zero_motion_accumulator > 0.99) && (rc->frames_to_key > 8)) {
3073 rc->kf_boost = MAX_KF_TOT_BOOST;
3074 } else {
3075 // Apply various clamps for min and max boost
3076 rc->kf_boost = VPXMAX((int)boost_score, (rc->frames_to_key * 3));
3077 rc->kf_boost = VPXMAX(rc->kf_boost, MIN_KF_TOT_BOOST);
3078 rc->kf_boost = VPXMIN(rc->kf_boost, MAX_KF_TOT_BOOST);
3079 }
3080
3081 // Work out how many bits to allocate for the key frame itself.
3082 kf_bits = calculate_boost_bits((rc->frames_to_key - 1), rc->kf_boost,
3083 twopass->kf_group_bits);
3084
3085 twopass->kf_group_bits -= kf_bits;
3086
3087 // Save the bits to spend on the key frame.
3088 gf_group->bit_allocation[0] = kf_bits;
3089 gf_group->update_type[0] = KF_UPDATE;
3090 gf_group->rf_level[0] = KF_STD;
3091
3092 // Note the total error score of the kf group minus the key frame itself.
3093 twopass->kf_group_error_left = (kf_group_err - kf_mod_err);
3094
3095 // Adjust the count of total modified error left.
3096 // The count of bits left is adjusted elsewhere based on real coded frame
3097 // sizes.
3098 twopass->normalized_score_left -= kf_group_err;
3099
3100 if (oxcf->resize_mode == RESIZE_DYNAMIC) {
3101 // Default to normal-sized frame on keyframes.
3102 cpi->rc.next_frame_size_selector = UNSCALED;
3103 }
3104 }
3105
is_skippable_frame(const VP9_COMP * cpi)3106 static int is_skippable_frame(const VP9_COMP *cpi) {
3107 // If the current frame does not have non-zero motion vector detected in the
3108 // first pass, and so do its previous and forward frames, then this frame
3109 // can be skipped for partition check, and the partition size is assigned
3110 // according to the variance
3111 const TWO_PASS *const twopass = &cpi->twopass;
3112
3113 return (!frame_is_intra_only(&cpi->common) &&
3114 twopass->stats_in - 2 > twopass->stats_in_start &&
3115 twopass->stats_in < twopass->stats_in_end &&
3116 (twopass->stats_in - 1)->pcnt_inter -
3117 (twopass->stats_in - 1)->pcnt_motion ==
3118 1 &&
3119 (twopass->stats_in - 2)->pcnt_inter -
3120 (twopass->stats_in - 2)->pcnt_motion ==
3121 1 &&
3122 twopass->stats_in->pcnt_inter - twopass->stats_in->pcnt_motion == 1);
3123 }
3124
vp9_rc_get_second_pass_params(VP9_COMP * cpi)3125 void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
3126 VP9_COMMON *const cm = &cpi->common;
3127 RATE_CONTROL *const rc = &cpi->rc;
3128 TWO_PASS *const twopass = &cpi->twopass;
3129 GF_GROUP *const gf_group = &twopass->gf_group;
3130 FIRSTPASS_STATS this_frame;
3131
3132 if (!twopass->stats_in) return;
3133
3134 // If this is an arf frame then we dont want to read the stats file or
3135 // advance the input pointer as we already have what we need.
3136 if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
3137 int target_rate;
3138
3139 vp9_configure_buffer_updates(cpi, gf_group->index);
3140
3141 target_rate = gf_group->bit_allocation[gf_group->index];
3142 target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
3143 rc->base_frame_target = target_rate;
3144
3145 cm->frame_type = INTER_FRAME;
3146
3147 // Do the firstpass stats indicate that this frame is skippable for the
3148 // partition search?
3149 if (cpi->sf.allow_partition_search_skip && cpi->oxcf.pass == 2 &&
3150 !cpi->use_svc) {
3151 cpi->partition_search_skippable_frame = is_skippable_frame(cpi);
3152 }
3153
3154 return;
3155 }
3156
3157 vpx_clear_system_state();
3158
3159 if (cpi->oxcf.rc_mode == VPX_Q) {
3160 twopass->active_worst_quality = cpi->oxcf.cq_level;
3161 } else if (cm->current_video_frame == 0) {
3162 const int frames_left =
3163 (int)(twopass->total_stats.count - cm->current_video_frame);
3164 // Special case code for first frame.
3165 const int section_target_bandwidth =
3166 (int)(twopass->bits_left / frames_left);
3167 const double section_length = twopass->total_left_stats.count;
3168 const double section_error =
3169 twopass->total_left_stats.coded_error / section_length;
3170 const double section_intra_skip =
3171 twopass->total_left_stats.intra_skip_pct / section_length;
3172 const double section_inactive_zone =
3173 (twopass->total_left_stats.inactive_zone_rows * 2) /
3174 ((double)cm->mb_rows * section_length);
3175 const double section_noise =
3176 twopass->total_left_stats.frame_noise_energy / section_length;
3177 int tmp_q;
3178
3179 tmp_q = get_twopass_worst_quality(
3180 cpi, section_error, section_intra_skip + section_inactive_zone,
3181 section_noise, section_target_bandwidth);
3182
3183 twopass->active_worst_quality = tmp_q;
3184 twopass->baseline_active_worst_quality = tmp_q;
3185 rc->ni_av_qi = tmp_q;
3186 rc->last_q[INTER_FRAME] = tmp_q;
3187 rc->avg_q = vp9_convert_qindex_to_q(tmp_q, cm->bit_depth);
3188 rc->avg_frame_qindex[INTER_FRAME] = tmp_q;
3189 rc->last_q[KEY_FRAME] = (tmp_q + cpi->oxcf.best_allowed_q) / 2;
3190 rc->avg_frame_qindex[KEY_FRAME] = rc->last_q[KEY_FRAME];
3191 }
3192 vp9_zero(this_frame);
3193 if (EOF == input_stats(twopass, &this_frame)) return;
3194
3195 // Set the frame content type flag.
3196 if (this_frame.intra_skip_pct >= FC_ANIMATION_THRESH)
3197 twopass->fr_content_type = FC_GRAPHICS_ANIMATION;
3198 else
3199 twopass->fr_content_type = FC_NORMAL;
3200
3201 // Keyframe and section processing.
3202 if (rc->frames_to_key == 0 || (cpi->frame_flags & FRAMEFLAGS_KEY)) {
3203 FIRSTPASS_STATS this_frame_copy;
3204 this_frame_copy = this_frame;
3205 // Define next KF group and assign bits to it.
3206 find_next_key_frame(cpi, &this_frame);
3207 this_frame = this_frame_copy;
3208 } else {
3209 cm->frame_type = INTER_FRAME;
3210 }
3211
3212 // Define a new GF/ARF group. (Should always enter here for key frames).
3213 if (rc->frames_till_gf_update_due == 0) {
3214 define_gf_group(cpi, &this_frame);
3215
3216 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
3217
3218 #if ARF_STATS_OUTPUT
3219 {
3220 FILE *fpfile;
3221 fpfile = fopen("arf.stt", "a");
3222 ++arf_count;
3223 fprintf(fpfile, "%10d %10ld %10d %10d %10ld %10ld\n",
3224 cm->current_video_frame, rc->frames_till_gf_update_due,
3225 rc->kf_boost, arf_count, rc->gfu_boost, cm->frame_type);
3226
3227 fclose(fpfile);
3228 }
3229 #endif
3230 }
3231
3232 vp9_configure_buffer_updates(cpi, gf_group->index);
3233
3234 // Do the firstpass stats indicate that this frame is skippable for the
3235 // partition search?
3236 if (cpi->sf.allow_partition_search_skip && cpi->oxcf.pass == 2 &&
3237 !cpi->use_svc) {
3238 cpi->partition_search_skippable_frame = is_skippable_frame(cpi);
3239 }
3240
3241 rc->base_frame_target = gf_group->bit_allocation[gf_group->index];
3242
3243 // The multiplication by 256 reverses a scaling factor of (>> 8)
3244 // applied when combining MB error values for the frame.
3245 twopass->mb_av_energy = log((this_frame.intra_error * 256.0) + 1.0);
3246 twopass->mb_smooth_pct = this_frame.intra_smooth_pct;
3247
3248 // Update the total stats remaining structure.
3249 subtract_stats(&twopass->total_left_stats, &this_frame);
3250 }
3251
3252 #define MINQ_ADJ_LIMIT 48
3253 #define MINQ_ADJ_LIMIT_CQ 20
3254 #define HIGH_UNDERSHOOT_RATIO 2
vp9_twopass_postencode_update(VP9_COMP * cpi)3255 void vp9_twopass_postencode_update(VP9_COMP *cpi) {
3256 TWO_PASS *const twopass = &cpi->twopass;
3257 RATE_CONTROL *const rc = &cpi->rc;
3258 VP9_COMMON *const cm = &cpi->common;
3259 const int bits_used = rc->base_frame_target;
3260
3261 // VBR correction is done through rc->vbr_bits_off_target. Based on the
3262 // sign of this value, a limited % adjustment is made to the target rate
3263 // of subsequent frames, to try and push it back towards 0. This method
3264 // is designed to prevent extreme behaviour at the end of a clip
3265 // or group of frames.
3266 rc->vbr_bits_off_target += rc->base_frame_target - rc->projected_frame_size;
3267 twopass->bits_left = VPXMAX(twopass->bits_left - bits_used, 0);
3268
3269 // Target vs actual bits for this arf group.
3270 twopass->rolling_arf_group_target_bits += rc->this_frame_target;
3271 twopass->rolling_arf_group_actual_bits += rc->projected_frame_size;
3272
3273 // Calculate the pct rc error.
3274 if (rc->total_actual_bits) {
3275 rc->rate_error_estimate =
3276 (int)((rc->vbr_bits_off_target * 100) / rc->total_actual_bits);
3277 rc->rate_error_estimate = clamp(rc->rate_error_estimate, -100, 100);
3278 } else {
3279 rc->rate_error_estimate = 0;
3280 }
3281
3282 if (cpi->common.frame_type != KEY_FRAME) {
3283 twopass->kf_group_bits -= bits_used;
3284 twopass->last_kfgroup_zeromotion_pct = twopass->kf_zeromotion_pct;
3285 }
3286 twopass->kf_group_bits = VPXMAX(twopass->kf_group_bits, 0);
3287
3288 // Increment the gf group index ready for the next frame.
3289 ++twopass->gf_group.index;
3290
3291 // If the rate control is drifting consider adjustment to min or maxq.
3292 if ((cpi->oxcf.rc_mode != VPX_Q) && !cpi->rc.is_src_frame_alt_ref) {
3293 const int maxq_adj_limit =
3294 rc->worst_quality - twopass->active_worst_quality;
3295 const int minq_adj_limit =
3296 (cpi->oxcf.rc_mode == VPX_CQ ? MINQ_ADJ_LIMIT_CQ : MINQ_ADJ_LIMIT);
3297 int aq_extend_min = 0;
3298 int aq_extend_max = 0;
3299
3300 // Extend min or Max Q range to account for imbalance from the base
3301 // value when using AQ.
3302 if (cpi->oxcf.aq_mode != NO_AQ) {
3303 if (cm->seg.aq_av_offset < 0) {
3304 // The balance of the AQ map tends towarda lowering the average Q.
3305 aq_extend_min = 0;
3306 aq_extend_max = VPXMIN(maxq_adj_limit, -cm->seg.aq_av_offset);
3307 } else {
3308 // The balance of the AQ map tends towards raising the average Q.
3309 aq_extend_min = VPXMIN(minq_adj_limit, cm->seg.aq_av_offset);
3310 aq_extend_max = 0;
3311 }
3312 }
3313
3314 // Undershoot.
3315 if (rc->rate_error_estimate > cpi->oxcf.under_shoot_pct) {
3316 --twopass->extend_maxq;
3317 if (rc->rolling_target_bits >= rc->rolling_actual_bits)
3318 ++twopass->extend_minq;
3319 // Overshoot.
3320 } else if (rc->rate_error_estimate < -cpi->oxcf.over_shoot_pct) {
3321 --twopass->extend_minq;
3322 if (rc->rolling_target_bits < rc->rolling_actual_bits)
3323 ++twopass->extend_maxq;
3324 } else {
3325 // Adjustment for extreme local overshoot.
3326 if (rc->projected_frame_size > (2 * rc->base_frame_target) &&
3327 rc->projected_frame_size > (2 * rc->avg_frame_bandwidth))
3328 ++twopass->extend_maxq;
3329
3330 // Unwind undershoot or overshoot adjustment.
3331 if (rc->rolling_target_bits < rc->rolling_actual_bits)
3332 --twopass->extend_minq;
3333 else if (rc->rolling_target_bits > rc->rolling_actual_bits)
3334 --twopass->extend_maxq;
3335 }
3336
3337 twopass->extend_minq =
3338 clamp(twopass->extend_minq, aq_extend_min, minq_adj_limit);
3339 twopass->extend_maxq =
3340 clamp(twopass->extend_maxq, aq_extend_max, maxq_adj_limit);
3341
3342 // If there is a big and undexpected undershoot then feed the extra
3343 // bits back in quickly. One situation where this may happen is if a
3344 // frame is unexpectedly almost perfectly predicted by the ARF or GF
3345 // but not very well predcited by the previous frame.
3346 if (!frame_is_kf_gf_arf(cpi) && !cpi->rc.is_src_frame_alt_ref) {
3347 int fast_extra_thresh = rc->base_frame_target / HIGH_UNDERSHOOT_RATIO;
3348 if (rc->projected_frame_size < fast_extra_thresh) {
3349 rc->vbr_bits_off_target_fast +=
3350 fast_extra_thresh - rc->projected_frame_size;
3351 rc->vbr_bits_off_target_fast =
3352 VPXMIN(rc->vbr_bits_off_target_fast, (4 * rc->avg_frame_bandwidth));
3353
3354 // Fast adaptation of minQ if necessary to use up the extra bits.
3355 if (rc->avg_frame_bandwidth) {
3356 twopass->extend_minq_fast =
3357 (int)(rc->vbr_bits_off_target_fast * 8 / rc->avg_frame_bandwidth);
3358 }
3359 twopass->extend_minq_fast = VPXMIN(
3360 twopass->extend_minq_fast, minq_adj_limit - twopass->extend_minq);
3361 } else if (rc->vbr_bits_off_target_fast) {
3362 twopass->extend_minq_fast = VPXMIN(
3363 twopass->extend_minq_fast, minq_adj_limit - twopass->extend_minq);
3364 } else {
3365 twopass->extend_minq_fast = 0;
3366 }
3367 }
3368 }
3369 }
3370