2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
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.
15 #include "./vpx_dsp_rtcd.h"
16 #include "./vpx_scale_rtcd.h"
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"
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_extend.h"
35 #include "vp9/encoder/vp9_firstpass.h"
36 #include "vp9/encoder/vp9_mcomp.h"
37 #include "vp9/encoder/vp9_quantize.h"
38 #include "vp9/encoder/vp9_rd.h"
39 #include "vpx_dsp/variance.h"
42 #define ARF_STATS_OUTPUT 0
44 #define BOOST_BREAKOUT 12.5
45 #define BOOST_FACTOR 12.5
46 #define FACTOR_PT_LOW 0.70
47 #define FACTOR_PT_HIGH 0.90
48 #define FIRST_PASS_Q 10.0
49 #define GF_MAX_BOOST 96.0
50 #define INTRA_MODE_PENALTY 1024
51 #define KF_MAX_BOOST 128.0
52 #define MIN_ARF_GF_BOOST 240
53 #define MIN_DECAY_FACTOR 0.01
54 #define MIN_KF_BOOST 300
55 #define NEW_MV_MODE_PENALTY 32
56 #define SVC_FACTOR_PT_LOW 0.45
57 #define DARK_THRESH 64
58 #define DEFAULT_GRP_WEIGHT 1.0
59 #define RC_FACTOR_MIN 0.75
60 #define RC_FACTOR_MAX 1.75
61 #define SECTION_NOISE_DEF 250.0
62 #define LOW_I_THRESH 24000
64 #define NCOUNT_INTRA_THRESH 8192
65 #define NCOUNT_INTRA_FACTOR 3
67 #define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x)-0.000001 : (x) + 0.000001)
70 unsigned int arf_count = 0;
73 // Resets the first pass file to the given position using a relative seek from
74 // the current position.
75 static void reset_fpf_position(TWO_PASS *p, const FIRSTPASS_STATS *position) {
76 p->stats_in = position;
79 // Read frame stats at an offset from the current position.
80 static const FIRSTPASS_STATS *read_frame_stats(const TWO_PASS *p, int offset) {
81 if ((offset >= 0 && p->stats_in + offset >= p->stats_in_end) ||
82 (offset < 0 && p->stats_in + offset < p->stats_in_start)) {
86 return &p->stats_in[offset];
89 static int input_stats(TWO_PASS *p, FIRSTPASS_STATS *fps) {
90 if (p->stats_in >= p->stats_in_end) return EOF;
97 static void output_stats(FIRSTPASS_STATS *stats,
98 struct vpx_codec_pkt_list *pktlist) {
99 struct vpx_codec_cx_pkt pkt;
100 pkt.kind = VPX_CODEC_STATS_PKT;
101 pkt.data.twopass_stats.buf = stats;
102 pkt.data.twopass_stats.sz = sizeof(FIRSTPASS_STATS);
103 vpx_codec_pkt_list_add(pktlist, &pkt);
109 fpfile = fopen("firstpass.stt", "a");
112 "%12.0lf %12.4lf %12.0lf %12.0lf %12.0lf %12.0lf %12.4lf"
113 "%12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf"
114 "%12.4lf %12.4lf %12.4lf %12.4lf %12.0lf %12.0lf %12.0lf %12.4lf"
116 stats->frame, stats->weight, stats->intra_error, stats->coded_error,
117 stats->sr_coded_error, stats->frame_noise_energy, stats->pcnt_inter,
118 stats->pcnt_motion, stats->pcnt_second_ref, stats->pcnt_neutral,
119 stats->intra_skip_pct, stats->intra_smooth_pct,
120 stats->inactive_zone_rows, stats->inactive_zone_cols, stats->MVr,
121 stats->mvr_abs, stats->MVc, stats->mvc_abs, stats->MVrv,
122 stats->MVcv, stats->mv_in_out_count, stats->new_mv_count,
123 stats->count, stats->duration);
129 #if CONFIG_FP_MB_STATS
130 static void output_fpmb_stats(uint8_t *this_frame_mb_stats, VP9_COMMON *cm,
131 struct vpx_codec_pkt_list *pktlist) {
132 struct vpx_codec_cx_pkt pkt;
133 pkt.kind = VPX_CODEC_FPMB_STATS_PKT;
134 pkt.data.firstpass_mb_stats.buf = this_frame_mb_stats;
135 pkt.data.firstpass_mb_stats.sz = cm->initial_mbs * sizeof(uint8_t);
136 vpx_codec_pkt_list_add(pktlist, &pkt);
140 static void zero_stats(FIRSTPASS_STATS *section) {
141 section->frame = 0.0;
142 section->weight = 0.0;
143 section->intra_error = 0.0;
144 section->coded_error = 0.0;
145 section->sr_coded_error = 0.0;
146 section->frame_noise_energy = 0.0;
147 section->pcnt_inter = 0.0;
148 section->pcnt_motion = 0.0;
149 section->pcnt_second_ref = 0.0;
150 section->pcnt_neutral = 0.0;
151 section->intra_skip_pct = 0.0;
152 section->intra_smooth_pct = 0.0;
153 section->inactive_zone_rows = 0.0;
154 section->inactive_zone_cols = 0.0;
156 section->mvr_abs = 0.0;
158 section->mvc_abs = 0.0;
161 section->mv_in_out_count = 0.0;
162 section->new_mv_count = 0.0;
163 section->count = 0.0;
164 section->duration = 1.0;
165 section->spatial_layer_id = 0;
168 static void accumulate_stats(FIRSTPASS_STATS *section,
169 const FIRSTPASS_STATS *frame) {
170 section->frame += frame->frame;
171 section->weight += frame->weight;
172 section->spatial_layer_id = frame->spatial_layer_id;
173 section->intra_error += frame->intra_error;
174 section->coded_error += frame->coded_error;
175 section->sr_coded_error += frame->sr_coded_error;
176 section->frame_noise_energy += frame->frame_noise_energy;
177 section->pcnt_inter += frame->pcnt_inter;
178 section->pcnt_motion += frame->pcnt_motion;
179 section->pcnt_second_ref += frame->pcnt_second_ref;
180 section->pcnt_neutral += frame->pcnt_neutral;
181 section->intra_skip_pct += frame->intra_skip_pct;
182 section->intra_smooth_pct += frame->intra_smooth_pct;
183 section->inactive_zone_rows += frame->inactive_zone_rows;
184 section->inactive_zone_cols += frame->inactive_zone_cols;
185 section->MVr += frame->MVr;
186 section->mvr_abs += frame->mvr_abs;
187 section->MVc += frame->MVc;
188 section->mvc_abs += frame->mvc_abs;
189 section->MVrv += frame->MVrv;
190 section->MVcv += frame->MVcv;
191 section->mv_in_out_count += frame->mv_in_out_count;
192 section->new_mv_count += frame->new_mv_count;
193 section->count += frame->count;
194 section->duration += frame->duration;
197 static void subtract_stats(FIRSTPASS_STATS *section,
198 const FIRSTPASS_STATS *frame) {
199 section->frame -= frame->frame;
200 section->weight -= frame->weight;
201 section->intra_error -= frame->intra_error;
202 section->coded_error -= frame->coded_error;
203 section->sr_coded_error -= frame->sr_coded_error;
204 section->frame_noise_energy -= frame->frame_noise_energy;
205 section->pcnt_inter -= frame->pcnt_inter;
206 section->pcnt_motion -= frame->pcnt_motion;
207 section->pcnt_second_ref -= frame->pcnt_second_ref;
208 section->pcnt_neutral -= frame->pcnt_neutral;
209 section->intra_skip_pct -= frame->intra_skip_pct;
210 section->intra_smooth_pct -= frame->intra_smooth_pct;
211 section->inactive_zone_rows -= frame->inactive_zone_rows;
212 section->inactive_zone_cols -= frame->inactive_zone_cols;
213 section->MVr -= frame->MVr;
214 section->mvr_abs -= frame->mvr_abs;
215 section->MVc -= frame->MVc;
216 section->mvc_abs -= frame->mvc_abs;
217 section->MVrv -= frame->MVrv;
218 section->MVcv -= frame->MVcv;
219 section->mv_in_out_count -= frame->mv_in_out_count;
220 section->new_mv_count -= frame->new_mv_count;
221 section->count -= frame->count;
222 section->duration -= frame->duration;
225 // Calculate an active area of the image that discounts formatting
226 // bars and partially discounts other 0 energy areas.
227 #define MIN_ACTIVE_AREA 0.5
228 #define MAX_ACTIVE_AREA 1.0
229 static double calculate_active_area(const VP9_COMP *cpi,
230 const FIRSTPASS_STATS *this_frame) {
235 ((this_frame->intra_skip_pct / 2) +
236 ((this_frame->inactive_zone_rows * 2) / (double)cpi->common.mb_rows));
237 return fclamp(active_pct, MIN_ACTIVE_AREA, MAX_ACTIVE_AREA);
240 // Calculate a modified Error used in distributing bits between easier and
242 #define ACT_AREA_CORRECTION 0.5
243 static double calculate_modified_err(const VP9_COMP *cpi,
244 const TWO_PASS *twopass,
245 const VP9EncoderConfig *oxcf,
246 const FIRSTPASS_STATS *this_frame) {
247 const FIRSTPASS_STATS *const stats = &twopass->total_stats;
248 const double av_weight = stats->weight / stats->count;
249 const double av_err = (stats->coded_error * av_weight) / stats->count;
250 double modified_error =
251 av_err * pow(this_frame->coded_error * this_frame->weight /
252 DOUBLE_DIVIDE_CHECK(av_err),
253 oxcf->two_pass_vbrbias / 100.0);
255 // Correction for active area. Frames with a reduced active area
256 // (eg due to formatting bars) have a higher error per mb for the
257 // remaining active MBs. The correction here assumes that coding
258 // 0.5N blocks of complexity 2X is a little easier than coding N
259 // blocks of complexity X.
261 pow(calculate_active_area(cpi, this_frame), ACT_AREA_CORRECTION);
263 return fclamp(modified_error, twopass->modified_error_min,
264 twopass->modified_error_max);
267 // This function returns the maximum target rate per frame.
268 static int frame_max_bits(const RATE_CONTROL *rc,
269 const VP9EncoderConfig *oxcf) {
270 int64_t max_bits = ((int64_t)rc->avg_frame_bandwidth *
271 (int64_t)oxcf->two_pass_vbrmax_section) /
275 else if (max_bits > rc->max_frame_bandwidth)
276 max_bits = rc->max_frame_bandwidth;
278 return (int)max_bits;
281 void vp9_init_first_pass(VP9_COMP *cpi) {
282 zero_stats(&cpi->twopass.total_stats);
285 void vp9_end_first_pass(VP9_COMP *cpi) {
286 if (is_two_pass_svc(cpi)) {
288 for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
289 output_stats(&cpi->svc.layer_context[i].twopass.total_stats,
290 cpi->output_pkt_list);
293 output_stats(&cpi->twopass.total_stats, cpi->output_pkt_list);
297 static vpx_variance_fn_t get_block_variance_fn(BLOCK_SIZE bsize) {
299 case BLOCK_8X8: return vpx_mse8x8;
300 case BLOCK_16X8: return vpx_mse16x8;
301 case BLOCK_8X16: return vpx_mse8x16;
302 default: return vpx_mse16x16;
306 static unsigned int get_prediction_error(BLOCK_SIZE bsize,
307 const struct buf_2d *src,
308 const struct buf_2d *ref) {
310 const vpx_variance_fn_t fn = get_block_variance_fn(bsize);
311 fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
315 #if CONFIG_VP9_HIGHBITDEPTH
316 static vpx_variance_fn_t highbd_get_block_variance_fn(BLOCK_SIZE bsize,
321 case BLOCK_8X8: return vpx_highbd_8_mse8x8;
322 case BLOCK_16X8: return vpx_highbd_8_mse16x8;
323 case BLOCK_8X16: return vpx_highbd_8_mse8x16;
324 default: return vpx_highbd_8_mse16x16;
329 case BLOCK_8X8: return vpx_highbd_10_mse8x8;
330 case BLOCK_16X8: return vpx_highbd_10_mse16x8;
331 case BLOCK_8X16: return vpx_highbd_10_mse8x16;
332 default: return vpx_highbd_10_mse16x16;
337 case BLOCK_8X8: return vpx_highbd_12_mse8x8;
338 case BLOCK_16X8: return vpx_highbd_12_mse16x8;
339 case BLOCK_8X16: return vpx_highbd_12_mse8x16;
340 default: return vpx_highbd_12_mse16x16;
346 static unsigned int highbd_get_prediction_error(BLOCK_SIZE bsize,
347 const struct buf_2d *src,
348 const struct buf_2d *ref,
351 const vpx_variance_fn_t fn = highbd_get_block_variance_fn(bsize, bd);
352 fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
355 #endif // CONFIG_VP9_HIGHBITDEPTH
357 // Refine the motion search range according to the frame dimension
358 // for first pass test.
359 static int get_search_range(const VP9_COMP *cpi) {
361 const int dim = VPXMIN(cpi->initial_width, cpi->initial_height);
363 while ((dim << sr) < MAX_FULL_PEL_VAL) ++sr;
367 static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
368 const MV *ref_mv, MV *best_mv,
369 int *best_motion_err) {
370 MACROBLOCKD *const xd = &x->e_mbd;
371 MV tmp_mv = { 0, 0 };
372 MV ref_mv_full = { ref_mv->row >> 3, ref_mv->col >> 3 };
373 int num00, tmp_err, n;
374 const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
375 vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
376 const int new_mv_mode_penalty = NEW_MV_MODE_PENALTY;
379 int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
380 const int sr = get_search_range(cpi);
384 // Override the default variance function to use MSE.
385 v_fn_ptr.vf = get_block_variance_fn(bsize);
386 #if CONFIG_VP9_HIGHBITDEPTH
387 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
388 v_fn_ptr.vf = highbd_get_block_variance_fn(bsize, xd->bd);
390 #endif // CONFIG_VP9_HIGHBITDEPTH
392 // Center the initial step/diamond search on best mv.
393 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
394 step_param, x->sadperbit16, &num00,
396 if (tmp_err < INT_MAX)
397 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
398 if (tmp_err < INT_MAX - new_mv_mode_penalty) tmp_err += new_mv_mode_penalty;
400 if (tmp_err < *best_motion_err) {
401 *best_motion_err = tmp_err;
405 // Carry out further step/diamond searches as necessary.
409 while (n < further_steps) {
415 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
416 step_param + n, x->sadperbit16, &num00,
418 if (tmp_err < INT_MAX)
419 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
420 if (tmp_err < INT_MAX - new_mv_mode_penalty)
421 tmp_err += new_mv_mode_penalty;
423 if (tmp_err < *best_motion_err) {
424 *best_motion_err = tmp_err;
431 static BLOCK_SIZE get_bsize(const VP9_COMMON *cm, int mb_row, int mb_col) {
432 if (2 * mb_col + 1 < cm->mi_cols) {
433 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_16X16 : BLOCK_16X8;
435 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_8X16 : BLOCK_8X8;
439 static int find_fp_qindex(vpx_bit_depth_t bit_depth) {
442 for (i = 0; i < QINDEX_RANGE; ++i)
443 if (vp9_convert_qindex_to_q(i, bit_depth) >= FIRST_PASS_Q) break;
445 if (i == QINDEX_RANGE) i--;
450 static void set_first_pass_params(VP9_COMP *cpi) {
451 VP9_COMMON *const cm = &cpi->common;
452 if (!cpi->refresh_alt_ref_frame &&
453 (cm->current_video_frame == 0 || (cpi->frame_flags & FRAMEFLAGS_KEY))) {
454 cm->frame_type = KEY_FRAME;
456 cm->frame_type = INTER_FRAME;
458 // Do not use periodic key frames.
459 cpi->rc.frames_to_key = INT_MAX;
462 // Scale an sse threshold to account for 8/10/12 bit.
463 static int scale_sse_threshold(VP9_COMMON *cm, int thresh) {
464 int ret_val = thresh;
465 #if CONFIG_VP9_HIGHBITDEPTH
466 if (cm->use_highbitdepth) {
467 switch (cm->bit_depth) {
468 case VPX_BITS_8: ret_val = thresh; break;
469 case VPX_BITS_10: ret_val = thresh >> 4; break;
470 case VPX_BITS_12: ret_val = thresh >> 8; break;
473 "cm->bit_depth should be VPX_BITS_8, "
474 "VPX_BITS_10 or VPX_BITS_12");
479 #endif // CONFIG_VP9_HIGHBITDEPTH
483 // This threshold is used to track blocks where to all intents and purposes
484 // the intra prediction error 0. Though the metric we test against
485 // is technically a sse we are mainly interested in blocks where all the pixels
486 // in the 8 bit domain have an error of <= 1 (where error = sse) so a
487 // linear scaling for 10 and 12 bit gives similar results.
488 #define UL_INTRA_THRESH 50
489 static int get_ul_intra_threshold(VP9_COMMON *cm) {
490 int ret_val = UL_INTRA_THRESH;
491 #if CONFIG_VP9_HIGHBITDEPTH
492 if (cm->use_highbitdepth) {
493 switch (cm->bit_depth) {
494 case VPX_BITS_8: ret_val = UL_INTRA_THRESH; break;
495 case VPX_BITS_10: ret_val = UL_INTRA_THRESH << 2; break;
496 case VPX_BITS_12: ret_val = UL_INTRA_THRESH << 4; break;
499 "cm->bit_depth should be VPX_BITS_8, "
500 "VPX_BITS_10 or VPX_BITS_12");
505 #endif // CONFIG_VP9_HIGHBITDEPTH
509 #define SMOOTH_INTRA_THRESH 4000
510 static int get_smooth_intra_threshold(VP9_COMMON *cm) {
511 int ret_val = SMOOTH_INTRA_THRESH;
512 #if CONFIG_VP9_HIGHBITDEPTH
513 if (cm->use_highbitdepth) {
514 switch (cm->bit_depth) {
515 case VPX_BITS_8: ret_val = SMOOTH_INTRA_THRESH; break;
516 case VPX_BITS_10: ret_val = SMOOTH_INTRA_THRESH << 4; break;
517 case VPX_BITS_12: ret_val = SMOOTH_INTRA_THRESH << 8; break;
520 "cm->bit_depth should be VPX_BITS_8, "
521 "VPX_BITS_10 or VPX_BITS_12");
526 #endif // CONFIG_VP9_HIGHBITDEPTH
530 #define FP_DN_THRESH 8
531 #define FP_MAX_DN_THRESH 16
532 #define KERNEL_SIZE 3
534 // Baseline Kernal weights for first pass noise metric
535 static uint8_t fp_dn_kernal_3[KERNEL_SIZE * KERNEL_SIZE] = { 1, 2, 1, 2, 4,
538 // Estimate noise at a single point based on the impace of a spatial kernal
539 // on the point value
540 static int fp_estimate_point_noise(uint8_t *src_ptr, const int stride) {
550 uint8_t centre_val = *src_ptr;
552 kernal_ptr = fp_dn_kernal_3;
555 tmp_ptr = src_ptr - stride - 1;
556 for (i = 0; i < KERNEL_SIZE; ++i) {
557 for (j = 0; j < KERNEL_SIZE; ++j) {
558 diff = abs((int)centre_val - (int)tmp_ptr[j]);
559 max_diff = VPXMAX(max_diff, diff);
560 if (diff <= FP_DN_THRESH) {
561 sum_weight += *kernal_ptr;
562 sum_val += (int)tmp_ptr[j] * (int)*kernal_ptr;
569 if (max_diff < FP_MAX_DN_THRESH)
570 // Update the source value with the new filtered value
571 dn_val = (sum_val + (sum_weight >> 1)) / sum_weight;
575 // return the noise energy as the square of the difference between the
576 // denoised and raw value.
577 dn_diff = (int)*src_ptr - (int)dn_val;
578 return dn_diff * dn_diff;
580 #if CONFIG_VP9_HIGHBITDEPTH
581 static int fp_highbd_estimate_point_noise(uint8_t *src_ptr, const int stride) {
592 uint16_t centre_val = *CONVERT_TO_SHORTPTR(src_ptr);
594 kernal_ptr = fp_dn_kernal_3;
597 tmp_ptr = src_ptr - stride - 1;
598 for (i = 0; i < KERNEL_SIZE; ++i) {
599 tmp_ptr16 = CONVERT_TO_SHORTPTR(tmp_ptr);
600 for (j = 0; j < KERNEL_SIZE; ++j) {
601 diff = abs((int)centre_val - (int)tmp_ptr16[j]);
602 max_diff = VPXMAX(max_diff, diff);
603 if (diff <= FP_DN_THRESH) {
604 sum_weight += *kernal_ptr;
605 sum_val += (int)tmp_ptr16[j] * (int)*kernal_ptr;
612 if (max_diff < FP_MAX_DN_THRESH)
613 // Update the source value with the new filtered value
614 dn_val = (sum_val + (sum_weight >> 1)) / sum_weight;
616 dn_val = *CONVERT_TO_SHORTPTR(src_ptr);
618 // return the noise energy as the square of the difference between the
619 // denoised and raw value.
620 dn_diff = (int)(*CONVERT_TO_SHORTPTR(src_ptr)) - (int)dn_val;
621 return dn_diff * dn_diff;
625 // Estimate noise for a block.
626 static int fp_estimate_block_noise(MACROBLOCK *x, BLOCK_SIZE bsize) {
627 #if CONFIG_VP9_HIGHBITDEPTH
628 MACROBLOCKD *xd = &x->e_mbd;
630 uint8_t *src_ptr = &x->plane[0].src.buf[0];
631 const int width = num_4x4_blocks_wide_lookup[bsize] * 4;
632 const int height = num_4x4_blocks_high_lookup[bsize] * 4;
634 int stride = x->plane[0].src.stride;
637 // Sampled points to reduce cost overhead.
638 for (h = 0; h < height; h += 2) {
639 for (w = 0; w < width; w += 2) {
640 #if CONFIG_VP9_HIGHBITDEPTH
641 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
642 block_noise += fp_highbd_estimate_point_noise(src_ptr, stride);
644 block_noise += fp_estimate_point_noise(src_ptr, stride);
646 block_noise += fp_estimate_point_noise(src_ptr, stride);
650 src_ptr += (stride - width);
652 return block_noise << 2; // Scale << 2 to account for sampling.
655 #define INVALID_ROW -1
656 void vp9_first_pass(VP9_COMP *cpi, const struct lookahead_entry *source) {
658 MACROBLOCK *const x = &cpi->td.mb;
659 VP9_COMMON *const cm = &cpi->common;
660 MACROBLOCKD *const xd = &x->e_mbd;
662 struct macroblock_plane *const p = x->plane;
663 struct macroblockd_plane *const pd = xd->plane;
664 const PICK_MODE_CONTEXT *ctx = &cpi->td.pc_root->none;
667 int recon_yoffset, recon_uvoffset;
668 int64_t intra_error = 0;
669 int64_t coded_error = 0;
670 int64_t sr_coded_error = 0;
671 int64_t frame_noise_energy = 0;
673 int sum_mvr = 0, sum_mvc = 0;
674 int sum_mvr_abs = 0, sum_mvc_abs = 0;
675 int64_t sum_mvrs = 0, sum_mvcs = 0;
678 int second_ref_count = 0;
679 const int intrapenalty = INTRA_MODE_PENALTY;
680 double neutral_count;
681 int intra_skip_count = 0;
682 int intra_smooth_count = 0;
683 int image_data_start_row = INVALID_ROW;
684 int new_mv_count = 0;
685 int sum_in_vectors = 0;
686 MV lastmv = { 0, 0 };
687 TWO_PASS *twopass = &cpi->twopass;
688 const MV zero_mv = { 0, 0 };
689 int recon_y_stride, recon_uv_stride, uv_mb_height;
691 YV12_BUFFER_CONFIG *const lst_yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
692 YV12_BUFFER_CONFIG *gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
693 YV12_BUFFER_CONFIG *const new_yv12 = get_frame_new_buffer(cm);
694 const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;
696 LAYER_CONTEXT *const lc =
697 is_two_pass_svc(cpi) ? &cpi->svc.layer_context[cpi->svc.spatial_layer_id]
700 double brightness_factor;
701 BufferPool *const pool = cm->buffer_pool;
702 MODE_INFO mi_above, mi_left;
704 // First pass code requires valid last and new frame buffers.
705 assert(new_yv12 != NULL);
706 assert((lc != NULL) || frame_is_intra_only(cm) || (lst_yv12 != NULL));
708 #if CONFIG_FP_MB_STATS
709 if (cpi->use_fp_mb_stats) {
710 vp9_zero_array(cpi->twopass.frame_mb_stats_buf, cm->initial_mbs);
714 vpx_clear_system_state();
717 brightness_factor = 0.0;
720 set_first_pass_params(cpi);
721 vp9_set_quantizer(cm, find_fp_qindex(cm->bit_depth));
724 twopass = &lc->twopass;
726 cpi->lst_fb_idx = cpi->svc.spatial_layer_id;
727 cpi->ref_frame_flags = VP9_LAST_FLAG;
729 if (cpi->svc.number_spatial_layers + cpi->svc.spatial_layer_id <
732 cpi->svc.number_spatial_layers + cpi->svc.spatial_layer_id;
733 cpi->ref_frame_flags |= VP9_GOLD_FLAG;
734 cpi->refresh_golden_frame = (lc->current_video_frame_in_layer == 0);
736 cpi->refresh_golden_frame = 0;
739 if (lc->current_video_frame_in_layer == 0) cpi->ref_frame_flags = 0;
741 vp9_scale_references(cpi);
743 // Use either last frame or alt frame for motion search.
744 if (cpi->ref_frame_flags & VP9_LAST_FLAG) {
745 first_ref_buf = vp9_get_scaled_ref_frame(cpi, LAST_FRAME);
746 if (first_ref_buf == NULL)
747 first_ref_buf = get_ref_frame_buffer(cpi, LAST_FRAME);
750 if (cpi->ref_frame_flags & VP9_GOLD_FLAG) {
751 gld_yv12 = vp9_get_scaled_ref_frame(cpi, GOLDEN_FRAME);
752 if (gld_yv12 == NULL) {
753 gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
760 (cpi->ref_frame_flags & VP9_LAST_FLAG) ? LAST_FRAME : NONE,
761 (cpi->ref_frame_flags & VP9_GOLD_FLAG) ? GOLDEN_FRAME : NONE);
763 cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
764 &cpi->scaled_source, 0);
767 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
769 vp9_setup_src_planes(x, cpi->Source, 0, 0);
770 vp9_setup_dst_planes(xd->plane, new_yv12, 0, 0);
772 if (!frame_is_intra_only(cm)) {
773 vp9_setup_pre_planes(xd, 0, first_ref_buf, 0, 0, NULL);
776 xd->mi = cm->mi_grid_visible;
779 vp9_frame_init_quantizer(cpi);
781 for (i = 0; i < MAX_MB_PLANE; ++i) {
782 p[i].coeff = ctx->coeff_pbuf[i][1];
783 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
784 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
785 p[i].eobs = ctx->eobs_pbuf[i][1];
789 vp9_init_mv_probs(cm);
790 vp9_initialize_rd_consts(cpi);
792 // Tiling is ignored in the first pass.
793 vp9_tile_init(&tile, cm, 0, 0);
795 recon_y_stride = new_yv12->y_stride;
796 recon_uv_stride = new_yv12->uv_stride;
797 uv_mb_height = 16 >> (new_yv12->y_height > new_yv12->uv_height);
799 for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
800 MV best_ref_mv = { 0, 0 };
802 // Reset above block coeffs.
803 recon_yoffset = (mb_row * recon_y_stride * 16);
804 recon_uvoffset = (mb_row * recon_uv_stride * uv_mb_height);
806 // Set up limit values for motion vectors to prevent them extending
807 // outside the UMV borders.
808 x->mv_limits.row_min = -((mb_row * 16) + BORDER_MV_PIXELS_B16);
809 x->mv_limits.row_max =
810 ((cm->mb_rows - 1 - mb_row) * 16) + BORDER_MV_PIXELS_B16;
812 for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
814 int this_intra_error;
815 const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
816 const BLOCK_SIZE bsize = get_bsize(cm, mb_row, mb_col);
820 #if CONFIG_FP_MB_STATS
821 const int mb_index = mb_row * cm->mb_cols + mb_col;
824 vpx_clear_system_state();
826 xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
827 xd->plane[1].dst.buf = new_yv12->u_buffer + recon_uvoffset;
828 xd->plane[2].dst.buf = new_yv12->v_buffer + recon_uvoffset;
829 xd->mi[0]->sb_type = bsize;
830 xd->mi[0]->ref_frame[0] = INTRA_FRAME;
831 set_mi_row_col(xd, &tile, mb_row << 1, num_8x8_blocks_high_lookup[bsize],
832 mb_col << 1, num_8x8_blocks_wide_lookup[bsize],
833 cm->mi_rows, cm->mi_cols);
834 // Are edges available for intra prediction?
835 // Since the firstpass does not populate the mi_grid_visible,
836 // above_mi/left_mi must be overwritten with a nonzero value when edges
837 // are available. Required by vp9_predict_intra_block().
838 xd->above_mi = (mb_row != 0) ? &mi_above : NULL;
839 xd->left_mi = (mb_col > tile.mi_col_start) ? &mi_left : NULL;
841 // Do intra 16x16 prediction.
843 xd->mi[0]->mode = DC_PRED;
845 use_dc_pred ? (bsize >= BLOCK_16X16 ? TX_16X16 : TX_8X8) : TX_4X4;
846 vp9_encode_intra_block_plane(x, bsize, 0, 0);
847 this_error = vpx_get_mb_ss(x->plane[0].src_diff);
848 this_intra_error = this_error;
850 // Keep a record of blocks that have very low intra error residual
851 // (i.e. are in effect completely flat and untextured in the intra
852 // domain). In natural videos this is uncommon, but it is much more
853 // common in animations, graphics and screen content, so may be used
854 // as a signal to detect these types of content.
855 if (this_error < get_ul_intra_threshold(cm)) {
857 } else if ((mb_col > 0) && (image_data_start_row == INVALID_ROW)) {
858 image_data_start_row = mb_row;
861 // Blocks that are mainly smooth in the intra domain.
862 // Some special accounting for CQ but also these are better for testing
864 if (this_error < get_smooth_intra_threshold(cm)) {
865 ++intra_smooth_count;
868 // Special case noise measurement for first frame.
869 if (cm->current_video_frame == 0) {
870 if (this_intra_error < scale_sse_threshold(cm, LOW_I_THRESH)) {
871 frame_noise_energy += fp_estimate_block_noise(x, bsize);
873 frame_noise_energy += (int64_t)SECTION_NOISE_DEF;
877 #if CONFIG_VP9_HIGHBITDEPTH
878 if (cm->use_highbitdepth) {
879 switch (cm->bit_depth) {
880 case VPX_BITS_8: break;
881 case VPX_BITS_10: this_error >>= 4; break;
882 case VPX_BITS_12: this_error >>= 8; break;
885 "cm->bit_depth should be VPX_BITS_8, "
886 "VPX_BITS_10 or VPX_BITS_12");
890 #endif // CONFIG_VP9_HIGHBITDEPTH
892 vpx_clear_system_state();
893 log_intra = log(this_error + 1.0);
894 if (log_intra < 10.0)
895 intra_factor += 1.0 + ((10.0 - log_intra) * 0.05);
899 #if CONFIG_VP9_HIGHBITDEPTH
900 if (cm->use_highbitdepth)
901 level_sample = CONVERT_TO_SHORTPTR(x->plane[0].src.buf)[0];
903 level_sample = x->plane[0].src.buf[0];
905 level_sample = x->plane[0].src.buf[0];
907 if ((level_sample < DARK_THRESH) && (log_intra < 9.0))
908 brightness_factor += 1.0 + (0.01 * (DARK_THRESH - level_sample));
910 brightness_factor += 1.0;
912 // Intrapenalty below deals with situations where the intra and inter
913 // error scores are very low (e.g. a plain black frame).
914 // We do not have special cases in first pass for 0,0 and nearest etc so
915 // all inter modes carry an overhead cost estimate for the mv.
916 // When the error score is very low this causes us to pick all or lots of
917 // INTRA modes and throw lots of key frames.
918 // This penalty adds a cost matching that of a 0,0 mv to the intra case.
919 this_error += intrapenalty;
921 // Accumulate the intra error.
922 intra_error += (int64_t)this_error;
924 #if CONFIG_FP_MB_STATS
925 if (cpi->use_fp_mb_stats) {
927 cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
931 // Set up limit values for motion vectors to prevent them extending
932 // outside the UMV borders.
933 x->mv_limits.col_min = -((mb_col * 16) + BORDER_MV_PIXELS_B16);
934 x->mv_limits.col_max =
935 ((cm->mb_cols - 1 - mb_col) * 16) + BORDER_MV_PIXELS_B16;
937 // Other than for the first frame do a motion search.
938 if ((lc == NULL && cm->current_video_frame > 0) ||
939 (lc != NULL && lc->current_video_frame_in_layer > 0)) {
940 int tmp_err, motion_error, raw_motion_error;
941 // Assume 0,0 motion with no mv overhead.
942 MV mv = { 0, 0 }, tmp_mv = { 0, 0 };
943 struct buf_2d unscaled_last_source_buf_2d;
945 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
946 #if CONFIG_VP9_HIGHBITDEPTH
947 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
948 motion_error = highbd_get_prediction_error(
949 bsize, &x->plane[0].src, &xd->plane[0].pre[0], xd->bd);
951 motion_error = get_prediction_error(bsize, &x->plane[0].src,
952 &xd->plane[0].pre[0]);
956 get_prediction_error(bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
957 #endif // CONFIG_VP9_HIGHBITDEPTH
959 // Compute the motion error of the 0,0 motion using the last source
960 // frame as the reference. Skip the further motion search on
961 // reconstructed frame if this error is small.
962 unscaled_last_source_buf_2d.buf =
963 cpi->unscaled_last_source->y_buffer + recon_yoffset;
964 unscaled_last_source_buf_2d.stride =
965 cpi->unscaled_last_source->y_stride;
966 #if CONFIG_VP9_HIGHBITDEPTH
967 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
968 raw_motion_error = highbd_get_prediction_error(
969 bsize, &x->plane[0].src, &unscaled_last_source_buf_2d, xd->bd);
971 raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
972 &unscaled_last_source_buf_2d);
975 raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
976 &unscaled_last_source_buf_2d);
977 #endif // CONFIG_VP9_HIGHBITDEPTH
979 // TODO(pengchong): Replace the hard-coded threshold
980 if (raw_motion_error > 25 || lc != NULL) {
981 // Test last reference frame using the previous best mv as the
982 // starting point (best reference) for the search.
983 first_pass_motion_search(cpi, x, &best_ref_mv, &mv, &motion_error);
985 // If the current best reference mv is not centered on 0,0 then do a
986 // 0,0 based search as well.
987 if (!is_zero_mv(&best_ref_mv)) {
989 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv, &tmp_err);
991 if (tmp_err < motion_error) {
992 motion_error = tmp_err;
997 // Search in an older reference frame.
998 if (((lc == NULL && cm->current_video_frame > 1) ||
999 (lc != NULL && lc->current_video_frame_in_layer > 1)) &&
1001 // Assume 0,0 motion with no mv overhead.
1002 int gf_motion_error;
1004 xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
1005 #if CONFIG_VP9_HIGHBITDEPTH
1006 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1007 gf_motion_error = highbd_get_prediction_error(
1008 bsize, &x->plane[0].src, &xd->plane[0].pre[0], xd->bd);
1010 gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
1011 &xd->plane[0].pre[0]);
1014 gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
1015 &xd->plane[0].pre[0]);
1016 #endif // CONFIG_VP9_HIGHBITDEPTH
1018 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv,
1021 if (gf_motion_error < motion_error && gf_motion_error < this_error)
1024 // Reset to last frame as reference buffer.
1025 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
1026 xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
1027 xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
1029 // In accumulating a score for the older reference frame take the
1030 // best of the motion predicted score and the intra coded error
1031 // (just as will be done for) accumulation of "coded_error" for
1033 if (gf_motion_error < this_error)
1034 sr_coded_error += gf_motion_error;
1036 sr_coded_error += this_error;
1038 sr_coded_error += motion_error;
1041 sr_coded_error += motion_error;
1044 // Start by assuming that intra mode is best.
1045 best_ref_mv.row = 0;
1046 best_ref_mv.col = 0;
1048 #if CONFIG_FP_MB_STATS
1049 if (cpi->use_fp_mb_stats) {
1050 // intra prediction statistics
1051 cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
1052 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_DCINTRA_MASK;
1053 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_MOTION_ZERO_MASK;
1054 if (this_error > FPMB_ERROR_LARGE_TH) {
1055 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_LARGE_MASK;
1056 } else if (this_error < FPMB_ERROR_SMALL_TH) {
1057 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_SMALL_MASK;
1062 if (motion_error <= this_error) {
1063 vpx_clear_system_state();
1065 // Keep a count of cases where the inter and intra were very close
1066 // and very low. This helps with scene cut detection for example in
1067 // cropped clips with black bars at the sides or top and bottom.
1068 if (((this_error - intrapenalty) * 9 <= motion_error * 10) &&
1069 (this_error < (2 * intrapenalty))) {
1070 neutral_count += 1.0;
1071 // Also track cases where the intra is not much worse than the inter
1072 // and use this in limiting the GF/arf group length.
1073 } else if ((this_error > NCOUNT_INTRA_THRESH) &&
1074 (this_error < (NCOUNT_INTRA_FACTOR * motion_error))) {
1076 (double)motion_error / DOUBLE_DIVIDE_CHECK((double)this_error);
1081 this_error = motion_error;
1082 xd->mi[0]->mode = NEWMV;
1083 xd->mi[0]->mv[0].as_mv = mv;
1084 xd->mi[0]->tx_size = TX_4X4;
1085 xd->mi[0]->ref_frame[0] = LAST_FRAME;
1086 xd->mi[0]->ref_frame[1] = NONE;
1087 vp9_build_inter_predictors_sby(xd, mb_row << 1, mb_col << 1, bsize);
1088 vp9_encode_sby_pass1(x, bsize);
1090 sum_mvr_abs += abs(mv.row);
1092 sum_mvc_abs += abs(mv.col);
1093 sum_mvrs += mv.row * mv.row;
1094 sum_mvcs += mv.col * mv.col;
1099 #if CONFIG_FP_MB_STATS
1100 if (cpi->use_fp_mb_stats) {
1101 // inter prediction statistics
1102 cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
1103 cpi->twopass.frame_mb_stats_buf[mb_index] &= ~FPMB_DCINTRA_MASK;
1104 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_MOTION_ZERO_MASK;
1105 if (this_error > FPMB_ERROR_LARGE_TH) {
1106 cpi->twopass.frame_mb_stats_buf[mb_index] |=
1107 FPMB_ERROR_LARGE_MASK;
1108 } else if (this_error < FPMB_ERROR_SMALL_TH) {
1109 cpi->twopass.frame_mb_stats_buf[mb_index] |=
1110 FPMB_ERROR_SMALL_MASK;
1115 if (!is_zero_mv(&mv)) {
1118 #if CONFIG_FP_MB_STATS
1119 if (cpi->use_fp_mb_stats) {
1120 cpi->twopass.frame_mb_stats_buf[mb_index] &=
1121 ~FPMB_MOTION_ZERO_MASK;
1122 // check estimated motion direction
1123 if (mv.as_mv.col > 0 && mv.as_mv.col >= abs(mv.as_mv.row)) {
1125 cpi->twopass.frame_mb_stats_buf[mb_index] |=
1126 FPMB_MOTION_RIGHT_MASK;
1127 } else if (mv.as_mv.row < 0 &&
1128 abs(mv.as_mv.row) >= abs(mv.as_mv.col)) {
1130 cpi->twopass.frame_mb_stats_buf[mb_index] |=
1131 FPMB_MOTION_UP_MASK;
1132 } else if (mv.as_mv.col < 0 &&
1133 abs(mv.as_mv.col) >= abs(mv.as_mv.row)) {
1135 cpi->twopass.frame_mb_stats_buf[mb_index] |=
1136 FPMB_MOTION_LEFT_MASK;
1139 cpi->twopass.frame_mb_stats_buf[mb_index] |=
1140 FPMB_MOTION_DOWN_MASK;
1145 // Non-zero vector, was it different from the last non zero vector?
1146 if (!is_equal_mv(&mv, &lastmv)) ++new_mv_count;
1149 // Does the row vector point inwards or outwards?
1150 if (mb_row < cm->mb_rows / 2) {
1153 else if (mv.row < 0)
1155 } else if (mb_row > cm->mb_rows / 2) {
1158 else if (mv.row < 0)
1162 // Does the col vector point inwards or outwards?
1163 if (mb_col < cm->mb_cols / 2) {
1166 else if (mv.col < 0)
1168 } else if (mb_col > cm->mb_cols / 2) {
1171 else if (mv.col < 0)
1174 frame_noise_energy += (int64_t)SECTION_NOISE_DEF;
1175 } else if (this_intra_error < scale_sse_threshold(cm, LOW_I_THRESH)) {
1176 frame_noise_energy += fp_estimate_block_noise(x, bsize);
1177 } else { // 0,0 mv but high error
1178 frame_noise_energy += (int64_t)SECTION_NOISE_DEF;
1180 } else { // Intra < inter error
1181 if (this_intra_error < scale_sse_threshold(cm, LOW_I_THRESH))
1182 frame_noise_energy += fp_estimate_block_noise(x, bsize);
1184 frame_noise_energy += (int64_t)SECTION_NOISE_DEF;
1187 sr_coded_error += (int64_t)this_error;
1189 coded_error += (int64_t)this_error;
1191 // Adjust to the next column of MBs.
1192 x->plane[0].src.buf += 16;
1193 x->plane[1].src.buf += uv_mb_height;
1194 x->plane[2].src.buf += uv_mb_height;
1196 recon_yoffset += 16;
1197 recon_uvoffset += uv_mb_height;
1200 // Adjust to the next row of MBs.
1201 x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
1202 x->plane[1].src.buf +=
1203 uv_mb_height * x->plane[1].src.stride - uv_mb_height * cm->mb_cols;
1204 x->plane[2].src.buf +=
1205 uv_mb_height * x->plane[1].src.stride - uv_mb_height * cm->mb_cols;
1207 vpx_clear_system_state();
1210 // Clamp the image start to rows/2. This number of rows is discarded top
1211 // and bottom as dead data so rows / 2 means the frame is blank.
1212 if ((image_data_start_row > cm->mb_rows / 2) ||
1213 (image_data_start_row == INVALID_ROW)) {
1214 image_data_start_row = cm->mb_rows / 2;
1216 // Exclude any image dead zone
1217 if (image_data_start_row > 0) {
1219 VPXMAX(0, intra_skip_count - (image_data_start_row * cm->mb_cols * 2));
1223 FIRSTPASS_STATS fps;
1224 // The minimum error here insures some bit allocation to frames even
1225 // in static regions. The allocation per MB declines for larger formats
1226 // where the typical "real" energy per MB also falls.
1227 // Initial estimate here uses sqrt(mbs) to define the min_err, where the
1228 // number of mbs is proportional to the image area.
1229 const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
1232 const double min_err = 200 * sqrt(num_mbs);
1234 intra_factor = intra_factor / (double)num_mbs;
1235 brightness_factor = brightness_factor / (double)num_mbs;
1236 fps.weight = intra_factor * brightness_factor;
1238 fps.frame = cm->current_video_frame;
1239 fps.spatial_layer_id = cpi->svc.spatial_layer_id;
1240 fps.coded_error = (double)(coded_error >> 8) + min_err;
1241 fps.sr_coded_error = (double)(sr_coded_error >> 8) + min_err;
1242 fps.intra_error = (double)(intra_error >> 8) + min_err;
1243 fps.frame_noise_energy = (double)frame_noise_energy / (double)num_mbs;
1245 fps.pcnt_inter = (double)intercount / num_mbs;
1246 fps.pcnt_second_ref = (double)second_ref_count / num_mbs;
1247 fps.pcnt_neutral = (double)neutral_count / num_mbs;
1248 fps.intra_skip_pct = (double)intra_skip_count / num_mbs;
1249 fps.intra_smooth_pct = (double)intra_smooth_count / num_mbs;
1250 fps.inactive_zone_rows = (double)image_data_start_row;
1251 // Currently set to 0 as most issues relate to letter boxing.
1252 fps.inactive_zone_cols = (double)0;
1255 fps.MVr = (double)sum_mvr / mvcount;
1256 fps.mvr_abs = (double)sum_mvr_abs / mvcount;
1257 fps.MVc = (double)sum_mvc / mvcount;
1258 fps.mvc_abs = (double)sum_mvc_abs / mvcount;
1260 ((double)sum_mvrs - ((double)sum_mvr * sum_mvr / mvcount)) / mvcount;
1262 ((double)sum_mvcs - ((double)sum_mvc * sum_mvc / mvcount)) / mvcount;
1263 fps.mv_in_out_count = (double)sum_in_vectors / (mvcount * 2);
1264 fps.new_mv_count = new_mv_count;
1265 fps.pcnt_motion = (double)mvcount / num_mbs;
1273 fps.mv_in_out_count = 0.0;
1274 fps.new_mv_count = 0.0;
1275 fps.pcnt_motion = 0.0;
1278 // Dont allow a value of 0 for duration.
1279 // (Section duration is also defaulted to minimum of 1.0).
1280 fps.duration = VPXMAX(1.0, (double)(source->ts_end - source->ts_start));
1282 // Don't want to do output stats with a stack variable!
1283 twopass->this_frame_stats = fps;
1284 output_stats(&twopass->this_frame_stats, cpi->output_pkt_list);
1285 accumulate_stats(&twopass->total_stats, &fps);
1287 #if CONFIG_FP_MB_STATS
1288 if (cpi->use_fp_mb_stats) {
1289 output_fpmb_stats(twopass->frame_mb_stats_buf, cm, cpi->output_pkt_list);
1294 // Copy the previous Last Frame back into gf and and arf buffers if
1295 // the prediction is good enough... but also don't allow it to lag too far.
1296 if ((twopass->sr_update_lag > 3) ||
1297 ((cm->current_video_frame > 0) &&
1298 (twopass->this_frame_stats.pcnt_inter > 0.20) &&
1299 ((twopass->this_frame_stats.intra_error /
1300 DOUBLE_DIVIDE_CHECK(twopass->this_frame_stats.coded_error)) > 2.0))) {
1301 if (gld_yv12 != NULL) {
1302 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
1303 cm->ref_frame_map[cpi->lst_fb_idx]);
1305 twopass->sr_update_lag = 1;
1307 ++twopass->sr_update_lag;
1310 vpx_extend_frame_borders(new_yv12);
1313 vp9_update_reference_frames(cpi);
1315 // The frame we just compressed now becomes the last frame.
1316 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->lst_fb_idx],
1320 // Special case for the first frame. Copy into the GF buffer as a second
1322 if (cm->current_video_frame == 0 && cpi->gld_fb_idx != INVALID_IDX &&
1324 ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
1325 cm->ref_frame_map[cpi->lst_fb_idx]);
1328 // Use this to see what the first pass reconstruction looks like.
1332 snprintf(filename, sizeof(filename), "enc%04d.yuv",
1333 (int)cm->current_video_frame);
1335 if (cm->current_video_frame == 0)
1336 recon_file = fopen(filename, "wb");
1338 recon_file = fopen(filename, "ab");
1340 (void)fwrite(lst_yv12->buffer_alloc, lst_yv12->frame_size, 1, recon_file);
1344 ++cm->current_video_frame;
1345 if (cpi->use_svc) vp9_inc_frame_in_layer(cpi);
1348 static double calc_correction_factor(double err_per_mb, double err_divisor,
1349 double pt_low, double pt_high, int q,
1350 vpx_bit_depth_t bit_depth) {
1351 const double error_term = err_per_mb / err_divisor;
1353 // Adjustment based on actual quantizer to power term.
1354 const double power_term =
1355 VPXMIN(vp9_convert_qindex_to_q(q, bit_depth) * 0.01 + pt_low, pt_high);
1357 // Calculate correction factor.
1358 if (power_term < 1.0) assert(error_term >= 0.0);
1360 return fclamp(pow(error_term, power_term), 0.05, 5.0);
1363 #define ERR_DIVISOR 115.0
1364 #define NOISE_FACTOR_MIN 0.9
1365 #define NOISE_FACTOR_MAX 1.1
1366 static int get_twopass_worst_quality(VP9_COMP *cpi, const double section_err,
1367 double inactive_zone, double section_noise,
1368 int section_target_bandwidth) {
1369 const RATE_CONTROL *const rc = &cpi->rc;
1370 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1371 TWO_PASS *const twopass = &cpi->twopass;
1373 // Clamp the target rate to VBR min / max limts.
1374 const int target_rate =
1375 vp9_rc_clamp_pframe_target_size(cpi, section_target_bandwidth);
1376 double noise_factor = pow((section_noise / SECTION_NOISE_DEF), 0.5);
1377 noise_factor = fclamp(noise_factor, NOISE_FACTOR_MIN, NOISE_FACTOR_MAX);
1378 inactive_zone = fclamp(inactive_zone, 0.0, 1.0);
1380 if (target_rate <= 0) {
1381 return rc->worst_quality; // Highest value allowed
1383 const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
1386 const int active_mbs = VPXMAX(1, num_mbs - (int)(num_mbs * inactive_zone));
1387 const double av_err_per_mb = section_err / active_mbs;
1388 const double speed_term = 1.0 + 0.04 * oxcf->speed;
1389 double last_group_rate_err;
1390 const int target_norm_bits_per_mb =
1391 ((uint64_t)target_rate << BPER_MB_NORMBITS) / active_mbs;
1393 int is_svc_upper_layer = 0;
1395 if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0)
1396 is_svc_upper_layer = 1;
1398 // based on recent history adjust expectations of bits per macroblock.
1399 last_group_rate_err =
1400 (double)twopass->rolling_arf_group_actual_bits /
1401 DOUBLE_DIVIDE_CHECK((double)twopass->rolling_arf_group_target_bits);
1402 last_group_rate_err = VPXMAX(0.25, VPXMIN(4.0, last_group_rate_err));
1403 twopass->bpm_factor *= (3.0 + last_group_rate_err) / 4.0;
1404 twopass->bpm_factor = VPXMAX(0.25, VPXMIN(4.0, twopass->bpm_factor));
1406 // Try and pick a max Q that will be high enough to encode the
1407 // content at the given rate.
1408 for (q = rc->best_quality; q < rc->worst_quality; ++q) {
1409 const double factor = calc_correction_factor(
1410 av_err_per_mb, ERR_DIVISOR,
1411 is_svc_upper_layer ? SVC_FACTOR_PT_LOW : FACTOR_PT_LOW,
1412 FACTOR_PT_HIGH, q, cpi->common.bit_depth);
1413 const int bits_per_mb = vp9_rc_bits_per_mb(
1415 factor * speed_term * cpi->twopass.bpm_factor * noise_factor,
1416 cpi->common.bit_depth);
1417 if (bits_per_mb <= target_norm_bits_per_mb) break;
1420 // Restriction on active max q for constrained quality mode.
1421 if (cpi->oxcf.rc_mode == VPX_CQ) q = VPXMAX(q, oxcf->cq_level);
1426 static void setup_rf_level_maxq(VP9_COMP *cpi) {
1428 RATE_CONTROL *const rc = &cpi->rc;
1429 for (i = INTER_NORMAL; i < RATE_FACTOR_LEVELS; ++i) {
1430 int qdelta = vp9_frame_type_qdelta(cpi, i, rc->worst_quality);
1431 rc->rf_level_maxq[i] = VPXMAX(rc->worst_quality + qdelta, rc->best_quality);
1435 static void init_subsampling(VP9_COMP *cpi) {
1436 const VP9_COMMON *const cm = &cpi->common;
1437 RATE_CONTROL *const rc = &cpi->rc;
1438 const int w = cm->width;
1439 const int h = cm->height;
1442 for (i = 0; i < FRAME_SCALE_STEPS; ++i) {
1443 // Note: Frames with odd-sized dimensions may result from this scaling.
1444 rc->frame_width[i] = (w * 16) / frame_scale_factor[i];
1445 rc->frame_height[i] = (h * 16) / frame_scale_factor[i];
1448 setup_rf_level_maxq(cpi);
1451 void calculate_coded_size(VP9_COMP *cpi, int *scaled_frame_width,
1452 int *scaled_frame_height) {
1453 RATE_CONTROL *const rc = &cpi->rc;
1454 *scaled_frame_width = rc->frame_width[rc->frame_size_selector];
1455 *scaled_frame_height = rc->frame_height[rc->frame_size_selector];
1458 void vp9_init_second_pass(VP9_COMP *cpi) {
1459 SVC *const svc = &cpi->svc;
1460 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1461 const int is_two_pass_svc =
1462 (svc->number_spatial_layers > 1) || (svc->number_temporal_layers > 1);
1463 RATE_CONTROL *const rc = &cpi->rc;
1464 TWO_PASS *const twopass =
1465 is_two_pass_svc ? &svc->layer_context[svc->spatial_layer_id].twopass
1468 FIRSTPASS_STATS *stats;
1470 zero_stats(&twopass->total_stats);
1471 zero_stats(&twopass->total_left_stats);
1473 if (!twopass->stats_in_end) return;
1475 stats = &twopass->total_stats;
1477 *stats = *twopass->stats_in_end;
1478 twopass->total_left_stats = *stats;
1480 frame_rate = 10000000.0 * stats->count / stats->duration;
1481 // Each frame can have a different duration, as the frame rate in the source
1482 // isn't guaranteed to be constant. The frame rate prior to the first frame
1483 // encoded in the second pass is a guess. However, the sum duration is not.
1484 // It is calculated based on the actual durations of all frames from the
1487 if (is_two_pass_svc) {
1488 vp9_update_spatial_layer_framerate(cpi, frame_rate);
1489 twopass->bits_left =
1490 (int64_t)(stats->duration *
1491 svc->layer_context[svc->spatial_layer_id].target_bandwidth /
1494 vp9_new_framerate(cpi, frame_rate);
1495 twopass->bits_left =
1496 (int64_t)(stats->duration * oxcf->target_bandwidth / 10000000.0);
1499 // This variable monitors how far behind the second ref update is lagging.
1500 twopass->sr_update_lag = 1;
1502 // Scan the first pass file and calculate a modified total error based upon
1503 // the bias/power function used to allocate bits.
1505 const double avg_error =
1506 stats->coded_error / DOUBLE_DIVIDE_CHECK(stats->count);
1507 const FIRSTPASS_STATS *s = twopass->stats_in;
1508 double modified_error_total = 0.0;
1509 twopass->modified_error_min =
1510 (avg_error * oxcf->two_pass_vbrmin_section) / 100;
1511 twopass->modified_error_max =
1512 (avg_error * oxcf->two_pass_vbrmax_section) / 100;
1513 while (s < twopass->stats_in_end) {
1514 modified_error_total += calculate_modified_err(cpi, twopass, oxcf, s);
1517 twopass->modified_error_left = modified_error_total;
1520 // Reset the vbr bits off target counters
1521 rc->vbr_bits_off_target = 0;
1522 rc->vbr_bits_off_target_fast = 0;
1523 rc->rate_error_estimate = 0;
1525 // Static sequence monitor variables.
1526 twopass->kf_zeromotion_pct = 100;
1527 twopass->last_kfgroup_zeromotion_pct = 100;
1529 // Initialize bits per macro_block estimate correction factor.
1530 twopass->bpm_factor = 1.0;
1531 // Initialize actual and target bits counters for ARF groups so that
1532 // at the start we have a neutral bpm adjustment.
1533 twopass->rolling_arf_group_target_bits = 1;
1534 twopass->rolling_arf_group_actual_bits = 1;
1536 if (oxcf->resize_mode != RESIZE_NONE) {
1537 init_subsampling(cpi);
1540 // Initialize the arnr strangth adjustment to 0
1541 twopass->arnr_strength_adjustment = 0;
1544 #define SR_DIFF_PART 0.0015
1545 #define INTRA_PART 0.005
1546 #define DEFAULT_DECAY_LIMIT 0.75
1547 #define LOW_SR_DIFF_TRHESH 0.1
1548 #define SR_DIFF_MAX 128.0
1549 #define LOW_CODED_ERR_PER_MB 10.0
1550 #define NCOUNT_FRAME_II_THRESH 6.0
1552 static double get_sr_decay_rate(const VP9_COMP *cpi,
1553 const FIRSTPASS_STATS *frame) {
1554 const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE) ? cpi->initial_mbs
1556 double sr_diff = (frame->sr_coded_error - frame->coded_error) / num_mbs;
1557 double sr_decay = 1.0;
1558 double modified_pct_inter;
1559 double modified_pcnt_intra;
1560 const double motion_amplitude_part =
1561 frame->pcnt_motion * ((frame->mvc_abs + frame->mvr_abs) /
1562 (cpi->initial_height + cpi->initial_width));
1564 modified_pct_inter = frame->pcnt_inter;
1565 if (((frame->coded_error / num_mbs) > LOW_CODED_ERR_PER_MB) &&
1566 ((frame->intra_error / DOUBLE_DIVIDE_CHECK(frame->coded_error)) <
1567 (double)NCOUNT_FRAME_II_THRESH)) {
1568 modified_pct_inter = frame->pcnt_inter - frame->pcnt_neutral;
1570 modified_pcnt_intra = 100 * (1.0 - modified_pct_inter);
1572 if ((sr_diff > LOW_SR_DIFF_TRHESH)) {
1573 sr_diff = VPXMIN(sr_diff, SR_DIFF_MAX);
1574 sr_decay = 1.0 - (SR_DIFF_PART * sr_diff) - motion_amplitude_part -
1575 (INTRA_PART * modified_pcnt_intra);
1577 return VPXMAX(sr_decay, VPXMIN(DEFAULT_DECAY_LIMIT, modified_pct_inter));
1580 // This function gives an estimate of how badly we believe the prediction
1581 // quality is decaying from frame to frame.
1582 static double get_zero_motion_factor(const VP9_COMP *cpi,
1583 const FIRSTPASS_STATS *frame) {
1584 const double zero_motion_pct = frame->pcnt_inter - frame->pcnt_motion;
1585 double sr_decay = get_sr_decay_rate(cpi, frame);
1586 return VPXMIN(sr_decay, zero_motion_pct);
1589 #define ZM_POWER_FACTOR 0.75
1591 static double get_prediction_decay_rate(const VP9_COMP *cpi,
1592 const FIRSTPASS_STATS *next_frame) {
1593 const double sr_decay_rate = get_sr_decay_rate(cpi, next_frame);
1594 const double zero_motion_factor =
1595 (0.95 * pow((next_frame->pcnt_inter - next_frame->pcnt_motion),
1598 return VPXMAX(zero_motion_factor,
1599 (sr_decay_rate + ((1.0 - sr_decay_rate) * zero_motion_factor)));
1602 // Function to test for a condition where a complex transition is followed
1603 // by a static section. For example in slide shows where there is a fade
1604 // between slides. This is to help with more optimal kf and gf positioning.
1605 static int detect_transition_to_still(VP9_COMP *cpi, int frame_interval,
1607 double loop_decay_rate,
1608 double last_decay_rate) {
1609 TWO_PASS *const twopass = &cpi->twopass;
1610 RATE_CONTROL *const rc = &cpi->rc;
1612 // Break clause to detect very still sections after motion
1613 // For example a static image after a fade or other transition
1614 // instead of a clean scene cut.
1615 if (frame_interval > rc->min_gf_interval && loop_decay_rate >= 0.999 &&
1616 last_decay_rate < 0.9) {
1619 // Look ahead a few frames to see if static condition persists...
1620 for (j = 0; j < still_interval; ++j) {
1621 const FIRSTPASS_STATS *stats = &twopass->stats_in[j];
1622 if (stats >= twopass->stats_in_end) break;
1624 if (stats->pcnt_inter - stats->pcnt_motion < 0.999) break;
1627 // Only if it does do we signal a transition to still.
1628 return j == still_interval;
1634 // This function detects a flash through the high relative pcnt_second_ref
1635 // score in the frame following a flash frame. The offset passed in should
1637 static int detect_flash(const TWO_PASS *twopass, int offset) {
1638 const FIRSTPASS_STATS *const next_frame = read_frame_stats(twopass, offset);
1640 // What we are looking for here is a situation where there is a
1641 // brief break in prediction (such as a flash) but subsequent frames
1642 // are reasonably well predicted by an earlier (pre flash) frame.
1643 // The recovery after a flash is indicated by a high pcnt_second_ref
1644 // compared to pcnt_inter.
1645 return next_frame != NULL &&
1646 next_frame->pcnt_second_ref > next_frame->pcnt_inter &&
1647 next_frame->pcnt_second_ref >= 0.5;
1650 // Update the motion related elements to the GF arf boost calculation.
1651 static void accumulate_frame_motion_stats(const FIRSTPASS_STATS *stats,
1653 double *mv_in_out_accumulator,
1654 double *abs_mv_in_out_accumulator,
1655 double *mv_ratio_accumulator) {
1656 const double pct = stats->pcnt_motion;
1658 // Accumulate Motion In/Out of frame stats.
1659 *mv_in_out = stats->mv_in_out_count * pct;
1660 *mv_in_out_accumulator += *mv_in_out;
1661 *abs_mv_in_out_accumulator += fabs(*mv_in_out);
1663 // Accumulate a measure of how uniform (or conversely how random) the motion
1664 // field is (a ratio of abs(mv) / mv).
1666 const double mvr_ratio =
1667 fabs(stats->mvr_abs) / DOUBLE_DIVIDE_CHECK(fabs(stats->MVr));
1668 const double mvc_ratio =
1669 fabs(stats->mvc_abs) / DOUBLE_DIVIDE_CHECK(fabs(stats->MVc));
1671 *mv_ratio_accumulator +=
1672 pct * (mvr_ratio < stats->mvr_abs ? mvr_ratio : stats->mvr_abs);
1673 *mv_ratio_accumulator +=
1674 pct * (mvc_ratio < stats->mvc_abs ? mvc_ratio : stats->mvc_abs);
1678 #define BASELINE_ERR_PER_MB 1000.0
1679 static double calc_frame_boost(VP9_COMP *cpi, const FIRSTPASS_STATS *this_frame,
1680 double this_frame_mv_in_out, double max_boost) {
1682 const double lq = vp9_convert_qindex_to_q(
1683 cpi->rc.avg_frame_qindex[INTER_FRAME], cpi->common.bit_depth);
1684 const double boost_q_correction = VPXMIN((0.5 + (lq * 0.015)), 1.5);
1685 int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE) ? cpi->initial_mbs
1688 // Correct for any inactive region in the image
1689 num_mbs = (int)VPXMAX(1, num_mbs * calculate_active_area(cpi, this_frame));
1691 // Underlying boost factor is based on inter error ratio.
1692 frame_boost = (BASELINE_ERR_PER_MB * num_mbs) /
1693 DOUBLE_DIVIDE_CHECK(this_frame->coded_error);
1694 frame_boost = frame_boost * BOOST_FACTOR * boost_q_correction;
1696 // Increase boost for frames where new data coming into frame (e.g. zoom out).
1697 // Slightly reduce boost if there is a net balance of motion out of the frame
1698 // (zoom in). The range for this_frame_mv_in_out is -1.0 to +1.0.
1699 if (this_frame_mv_in_out > 0.0)
1700 frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
1701 // In the extreme case the boost is halved.
1703 frame_boost += frame_boost * (this_frame_mv_in_out / 2.0);
1705 return VPXMIN(frame_boost, max_boost * boost_q_correction);
1708 static int calc_arf_boost(VP9_COMP *cpi, int offset, int f_frames, int b_frames,
1709 int *f_boost, int *b_boost) {
1710 TWO_PASS *const twopass = &cpi->twopass;
1712 double boost_score = 0.0;
1713 double mv_ratio_accumulator = 0.0;
1714 double decay_accumulator = 1.0;
1715 double this_frame_mv_in_out = 0.0;
1716 double mv_in_out_accumulator = 0.0;
1717 double abs_mv_in_out_accumulator = 0.0;
1719 int flash_detected = 0;
1721 // Search forward from the proposed arf/next gf position.
1722 for (i = 0; i < f_frames; ++i) {
1723 const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
1724 if (this_frame == NULL) break;
1726 // Update the motion related elements to the boost calculation.
1727 accumulate_frame_motion_stats(
1728 this_frame, &this_frame_mv_in_out, &mv_in_out_accumulator,
1729 &abs_mv_in_out_accumulator, &mv_ratio_accumulator);
1731 // We want to discount the flash frame itself and the recovery
1732 // frame that follows as both will have poor scores.
1733 flash_detected = detect_flash(twopass, i + offset) ||
1734 detect_flash(twopass, i + offset + 1);
1736 // Accumulate the effect of prediction quality decay.
1737 if (!flash_detected) {
1738 decay_accumulator *= get_prediction_decay_rate(cpi, this_frame);
1739 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1741 : decay_accumulator;
1746 calc_frame_boost(cpi, this_frame, this_frame_mv_in_out, GF_MAX_BOOST);
1749 *f_boost = (int)boost_score;
1751 // Reset for backward looking loop.
1753 mv_ratio_accumulator = 0.0;
1754 decay_accumulator = 1.0;
1755 this_frame_mv_in_out = 0.0;
1756 mv_in_out_accumulator = 0.0;
1757 abs_mv_in_out_accumulator = 0.0;
1759 // Search backward towards last gf position.
1760 for (i = -1; i >= -b_frames; --i) {
1761 const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
1762 if (this_frame == NULL) break;
1764 // Update the motion related elements to the boost calculation.
1765 accumulate_frame_motion_stats(
1766 this_frame, &this_frame_mv_in_out, &mv_in_out_accumulator,
1767 &abs_mv_in_out_accumulator, &mv_ratio_accumulator);
1769 // We want to discount the the flash frame itself and the recovery
1770 // frame that follows as both will have poor scores.
1771 flash_detected = detect_flash(twopass, i + offset) ||
1772 detect_flash(twopass, i + offset + 1);
1774 // Cumulative effect of prediction quality decay.
1775 if (!flash_detected) {
1776 decay_accumulator *= get_prediction_decay_rate(cpi, this_frame);
1777 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1779 : decay_accumulator;
1784 calc_frame_boost(cpi, this_frame, this_frame_mv_in_out, GF_MAX_BOOST);
1786 *b_boost = (int)boost_score;
1788 arf_boost = (*f_boost + *b_boost);
1789 if (arf_boost < ((b_frames + f_frames) * 20))
1790 arf_boost = ((b_frames + f_frames) * 20);
1791 arf_boost = VPXMAX(arf_boost, MIN_ARF_GF_BOOST);
1796 // Calculate a section intra ratio used in setting max loop filter.
1797 static int calculate_section_intra_ratio(const FIRSTPASS_STATS *begin,
1798 const FIRSTPASS_STATS *end,
1799 int section_length) {
1800 const FIRSTPASS_STATS *s = begin;
1801 double intra_error = 0.0;
1802 double coded_error = 0.0;
1805 while (s < end && i < section_length) {
1806 intra_error += s->intra_error;
1807 coded_error += s->coded_error;
1812 return (int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
1815 // Calculate the total bits to allocate in this GF/ARF group.
1816 static int64_t calculate_total_gf_group_bits(VP9_COMP *cpi,
1817 double gf_group_err) {
1818 const RATE_CONTROL *const rc = &cpi->rc;
1819 const TWO_PASS *const twopass = &cpi->twopass;
1820 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
1821 int64_t total_group_bits;
1823 // Calculate the bits to be allocated to the group as a whole.
1824 if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0)) {
1825 total_group_bits = (int64_t)(twopass->kf_group_bits *
1826 (gf_group_err / twopass->kf_group_error_left));
1828 total_group_bits = 0;
1831 // Clamp odd edge cases.
1833 (total_group_bits < 0) ? 0 : (total_group_bits > twopass->kf_group_bits)
1834 ? twopass->kf_group_bits
1837 // Clip based on user supplied data rate variability limit.
1838 if (total_group_bits > (int64_t)max_bits * rc->baseline_gf_interval)
1839 total_group_bits = (int64_t)max_bits * rc->baseline_gf_interval;
1841 return total_group_bits;
1844 // Calculate the number bits extra to assign to boosted frames in a group.
1845 static int calculate_boost_bits(int frame_count, int boost,
1846 int64_t total_group_bits) {
1847 int allocation_chunks;
1849 // return 0 for invalid inputs (could arise e.g. through rounding errors)
1850 if (!boost || (total_group_bits <= 0) || (frame_count <= 0)) return 0;
1852 allocation_chunks = (frame_count * 100) + boost;
1854 // Prevent overflow.
1856 int divisor = boost >> 10;
1858 allocation_chunks /= divisor;
1861 // Calculate the number of extra bits for use in the boosted frame or frames.
1862 return VPXMAX((int)(((int64_t)boost * total_group_bits) / allocation_chunks),
1866 // Current limit on maximum number of active arfs in a GF/ARF group.
1867 #define MAX_ACTIVE_ARFS 2
1870 // This function indirects the choice of buffers for arfs.
1871 // At the moment the values are fixed but this may change as part of
1872 // the integration process with other codec features that swap buffers around.
1873 static void get_arf_buffer_indices(unsigned char *arf_buffer_indices) {
1874 arf_buffer_indices[0] = ARF_SLOT1;
1875 arf_buffer_indices[1] = ARF_SLOT2;
1878 static void allocate_gf_group_bits(VP9_COMP *cpi, int64_t gf_group_bits,
1880 RATE_CONTROL *const rc = &cpi->rc;
1881 TWO_PASS *const twopass = &cpi->twopass;
1882 GF_GROUP *const gf_group = &twopass->gf_group;
1883 FIRSTPASS_STATS frame_stats;
1885 int frame_index = 1;
1886 int target_frame_size;
1888 const int max_bits = frame_max_bits(&cpi->rc, &cpi->oxcf);
1889 int64_t total_group_bits = gf_group_bits;
1890 int mid_boost_bits = 0;
1892 unsigned char arf_buffer_indices[MAX_ACTIVE_ARFS];
1893 int alt_frame_index = frame_index;
1894 int has_temporal_layers =
1895 is_two_pass_svc(cpi) && cpi->svc.number_temporal_layers > 1;
1897 int normal_frame_bits;
1898 int last_frame_bits;
1899 int last_frame_reduction;
1901 // Only encode alt reference frame in temporal base layer.
1902 if (has_temporal_layers) alt_frame_index = cpi->svc.number_temporal_layers;
1905 cpi->common.frame_type == KEY_FRAME || vp9_is_upper_layer_key_frame(cpi);
1907 get_arf_buffer_indices(arf_buffer_indices);
1909 // For key frames the frame target rate is already set and it
1910 // is also the golden frame.
1912 if (rc->source_alt_ref_active) {
1913 gf_group->update_type[0] = OVERLAY_UPDATE;
1914 gf_group->rf_level[0] = INTER_NORMAL;
1915 gf_group->bit_allocation[0] = 0;
1917 gf_group->update_type[0] = GF_UPDATE;
1918 gf_group->rf_level[0] = GF_ARF_STD;
1919 gf_group->bit_allocation[0] = gf_arf_bits;
1921 gf_group->arf_update_idx[0] = arf_buffer_indices[0];
1922 gf_group->arf_ref_idx[0] = arf_buffer_indices[0];
1924 // Step over the golden frame / overlay frame
1925 if (EOF == input_stats(twopass, &frame_stats)) return;
1928 // Deduct the boost bits for arf (or gf if it is not a key frame)
1929 // from the group total.
1930 if (rc->source_alt_ref_pending || !key_frame) total_group_bits -= gf_arf_bits;
1932 // Store the bits to spend on the ARF if there is one.
1933 if (rc->source_alt_ref_pending) {
1934 gf_group->update_type[alt_frame_index] = ARF_UPDATE;
1935 gf_group->rf_level[alt_frame_index] = GF_ARF_STD;
1936 gf_group->bit_allocation[alt_frame_index] = gf_arf_bits;
1938 if (has_temporal_layers)
1939 gf_group->arf_src_offset[alt_frame_index] =
1940 (unsigned char)(rc->baseline_gf_interval -
1941 cpi->svc.number_temporal_layers);
1943 gf_group->arf_src_offset[alt_frame_index] =
1944 (unsigned char)(rc->baseline_gf_interval - 1);
1946 gf_group->arf_update_idx[alt_frame_index] = arf_buffer_indices[0];
1947 gf_group->arf_ref_idx[alt_frame_index] =
1948 arf_buffer_indices[cpi->multi_arf_last_grp_enabled &&
1949 rc->source_alt_ref_active];
1950 if (!has_temporal_layers) ++frame_index;
1952 if (cpi->multi_arf_enabled) {
1953 // Set aside a slot for a level 1 arf.
1954 gf_group->update_type[frame_index] = ARF_UPDATE;
1955 gf_group->rf_level[frame_index] = GF_ARF_LOW;
1956 gf_group->arf_src_offset[frame_index] =
1957 (unsigned char)((rc->baseline_gf_interval >> 1) - 1);
1958 gf_group->arf_update_idx[frame_index] = arf_buffer_indices[1];
1959 gf_group->arf_ref_idx[frame_index] = arf_buffer_indices[0];
1964 // Note index of the first normal inter frame int eh group (not gf kf arf)
1965 gf_group->first_inter_index = frame_index;
1967 // Define middle frame
1968 mid_frame_idx = frame_index + (rc->baseline_gf_interval >> 1) - 1;
1970 normal_frames = (rc->baseline_gf_interval - rc->source_alt_ref_pending);
1972 // The last frame in the group is used less as a predictor so reduce
1973 // its allocation a little.
1974 if (normal_frames > 1) {
1975 normal_frame_bits = (int)(total_group_bits / normal_frames);
1976 last_frame_reduction = normal_frame_bits / 16;
1977 last_frame_bits = normal_frame_bits - last_frame_reduction;
1979 normal_frame_bits = (int)total_group_bits;
1980 last_frame_bits = normal_frame_bits;
1981 last_frame_reduction = 0;
1984 // Allocate bits to the other frames in the group.
1985 for (i = 0; i < normal_frames; ++i) {
1987 if (EOF == input_stats(twopass, &frame_stats)) break;
1989 if (has_temporal_layers && frame_index == alt_frame_index) {
1993 target_frame_size = (i == (normal_frames - 1))
1995 : (i == mid_frame_idx)
1996 ? normal_frame_bits + last_frame_reduction
1997 : normal_frame_bits;
1999 if (rc->source_alt_ref_pending && cpi->multi_arf_enabled) {
2000 mid_boost_bits += (target_frame_size >> 4);
2001 target_frame_size -= (target_frame_size >> 4);
2003 if (frame_index <= mid_frame_idx) arf_idx = 1;
2005 gf_group->arf_update_idx[frame_index] = arf_buffer_indices[arf_idx];
2006 gf_group->arf_ref_idx[frame_index] = arf_buffer_indices[arf_idx];
2009 clamp(target_frame_size, 0, VPXMIN(max_bits, (int)total_group_bits));
2011 gf_group->update_type[frame_index] = LF_UPDATE;
2012 gf_group->rf_level[frame_index] = INTER_NORMAL;
2014 gf_group->bit_allocation[frame_index] = target_frame_size;
2019 // We need to configure the frame at the end of the sequence + 1 that will be
2020 // the start frame for the next group. Otherwise prior to the call to
2021 // vp9_rc_get_second_pass_params() the data will be undefined.
2022 gf_group->arf_update_idx[frame_index] = arf_buffer_indices[0];
2023 gf_group->arf_ref_idx[frame_index] = arf_buffer_indices[0];
2025 if (rc->source_alt_ref_pending) {
2026 gf_group->update_type[frame_index] = OVERLAY_UPDATE;
2027 gf_group->rf_level[frame_index] = INTER_NORMAL;
2029 // Final setup for second arf and its overlay.
2030 if (cpi->multi_arf_enabled) {
2031 gf_group->bit_allocation[2] =
2032 gf_group->bit_allocation[mid_frame_idx] + mid_boost_bits;
2033 gf_group->update_type[mid_frame_idx] = OVERLAY_UPDATE;
2034 gf_group->bit_allocation[mid_frame_idx] = 0;
2037 gf_group->update_type[frame_index] = GF_UPDATE;
2038 gf_group->rf_level[frame_index] = GF_ARF_STD;
2041 // Note whether multi-arf was enabled this group for next time.
2042 cpi->multi_arf_last_grp_enabled = cpi->multi_arf_enabled;
2045 // Adjusts the ARNF filter for a GF group.
2046 static void adjust_group_arnr_filter(VP9_COMP *cpi, double section_noise,
2047 double section_inter,
2048 double section_motion) {
2049 TWO_PASS *const twopass = &cpi->twopass;
2050 double section_zeromv = section_inter - section_motion;
2052 twopass->arnr_strength_adjustment = 0;
2054 if ((section_zeromv < 0.10) || (section_noise <= (SECTION_NOISE_DEF * 0.75)))
2055 twopass->arnr_strength_adjustment -= 1;
2056 if (section_zeromv > 0.50) twopass->arnr_strength_adjustment += 1;
2059 // Analyse and define a gf/arf group.
2060 static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
2061 VP9_COMMON *const cm = &cpi->common;
2062 RATE_CONTROL *const rc = &cpi->rc;
2063 VP9EncoderConfig *const oxcf = &cpi->oxcf;
2064 TWO_PASS *const twopass = &cpi->twopass;
2065 FIRSTPASS_STATS next_frame;
2066 const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
2069 double boost_score = 0.0;
2070 double old_boost_score = 0.0;
2071 double gf_group_err = 0.0;
2072 double gf_group_raw_error = 0.0;
2073 double gf_group_noise = 0.0;
2074 double gf_group_skip_pct = 0.0;
2075 double gf_group_inactive_zone_rows = 0.0;
2076 double gf_group_inter = 0.0;
2077 double gf_group_motion = 0.0;
2078 double gf_first_frame_err = 0.0;
2079 double mod_frame_err = 0.0;
2081 double mv_ratio_accumulator = 0.0;
2082 double decay_accumulator = 1.0;
2083 double zero_motion_accumulator = 1.0;
2085 double loop_decay_rate = 1.00;
2086 double last_loop_decay_rate = 1.00;
2088 double this_frame_mv_in_out = 0.0;
2089 double mv_in_out_accumulator = 0.0;
2090 double abs_mv_in_out_accumulator = 0.0;
2091 double mv_ratio_accumulator_thresh;
2092 unsigned int allow_alt_ref = is_altref_enabled(cpi);
2097 int active_max_gf_interval;
2098 int active_min_gf_interval;
2099 int64_t gf_group_bits;
2101 const int is_key_frame = frame_is_intra_only(cm);
2102 const int arf_active_or_kf = is_key_frame || rc->source_alt_ref_active;
2104 // Reset the GF group data structures unless this is a key
2105 // frame in which case it will already have been done.
2106 if (is_key_frame == 0) {
2107 vp9_zero(twopass->gf_group);
2110 vpx_clear_system_state();
2111 vp9_zero(next_frame);
2113 // Load stats for the current frame.
2114 mod_frame_err = calculate_modified_err(cpi, twopass, oxcf, this_frame);
2116 // Note the error of the frame at the start of the group. This will be
2117 // the GF frame error if we code a normal gf.
2118 gf_first_frame_err = mod_frame_err;
2120 // If this is a key frame or the overlay from a previous arf then
2121 // the error score / cost of this frame has already been accounted for.
2122 if (arf_active_or_kf) {
2123 gf_group_err -= gf_first_frame_err;
2124 gf_group_raw_error -= this_frame->coded_error;
2125 gf_group_noise -= this_frame->frame_noise_energy;
2126 gf_group_skip_pct -= this_frame->intra_skip_pct;
2127 gf_group_inactive_zone_rows -= this_frame->inactive_zone_rows;
2128 gf_group_inter -= this_frame->pcnt_inter;
2129 gf_group_motion -= this_frame->pcnt_motion;
2132 // Motion breakout threshold for loop below depends on image size.
2133 mv_ratio_accumulator_thresh =
2134 (cpi->initial_height + cpi->initial_width) / 4.0;
2136 // Set a maximum and minimum interval for the GF group.
2137 // If the image appears almost completely static we can extend beyond this.
2139 int int_max_q = (int)(vp9_convert_qindex_to_q(twopass->active_worst_quality,
2140 cpi->common.bit_depth));
2141 int int_lbq = (int)(vp9_convert_qindex_to_q(rc->last_boosted_qindex,
2142 cpi->common.bit_depth));
2143 active_min_gf_interval =
2144 rc->min_gf_interval + arf_active_or_kf + VPXMIN(2, int_max_q / 200);
2145 if (active_min_gf_interval > rc->max_gf_interval)
2146 active_min_gf_interval = rc->max_gf_interval;
2148 if (cpi->multi_arf_allowed) {
2149 active_max_gf_interval = rc->max_gf_interval;
2151 // The value chosen depends on the active Q range. At low Q we have
2152 // bits to spare and are better with a smaller interval and smaller boost.
2153 // At high Q when there are few bits to spare we are better with a longer
2154 // interval to spread the cost of the GF.
2155 active_max_gf_interval = 12 + arf_active_or_kf + VPXMIN(4, (int_lbq / 6));
2157 // We have: active_min_gf_interval <= rc->max_gf_interval
2158 if (active_max_gf_interval < active_min_gf_interval)
2159 active_max_gf_interval = active_min_gf_interval;
2160 else if (active_max_gf_interval > rc->max_gf_interval)
2161 active_max_gf_interval = rc->max_gf_interval;
2163 // Would the active max drop us out just before the near the next kf?
2164 if ((active_max_gf_interval <= rc->frames_to_key) &&
2165 (active_max_gf_interval >= (rc->frames_to_key - rc->min_gf_interval)))
2166 active_max_gf_interval = rc->frames_to_key / 2;
2171 while (i < rc->static_scene_max_gf_interval && i < rc->frames_to_key) {
2174 // Accumulate error score of frames in this gf group.
2175 mod_frame_err = calculate_modified_err(cpi, twopass, oxcf, this_frame);
2176 gf_group_err += mod_frame_err;
2177 gf_group_raw_error += this_frame->coded_error;
2178 gf_group_noise += this_frame->frame_noise_energy;
2179 gf_group_skip_pct += this_frame->intra_skip_pct;
2180 gf_group_inactive_zone_rows += this_frame->inactive_zone_rows;
2181 gf_group_inter += this_frame->pcnt_inter;
2182 gf_group_motion += this_frame->pcnt_motion;
2184 if (EOF == input_stats(twopass, &next_frame)) break;
2186 // Test for the case where there is a brief flash but the prediction
2187 // quality back to an earlier frame is then restored.
2188 flash_detected = detect_flash(twopass, 0);
2190 // Update the motion related elements to the boost calculation.
2191 accumulate_frame_motion_stats(
2192 &next_frame, &this_frame_mv_in_out, &mv_in_out_accumulator,
2193 &abs_mv_in_out_accumulator, &mv_ratio_accumulator);
2195 // Accumulate the effect of prediction quality decay.
2196 if (!flash_detected) {
2197 last_loop_decay_rate = loop_decay_rate;
2198 loop_decay_rate = get_prediction_decay_rate(cpi, &next_frame);
2200 decay_accumulator = decay_accumulator * loop_decay_rate;
2202 // Monitor for static sections.
2203 zero_motion_accumulator = VPXMIN(
2204 zero_motion_accumulator, get_zero_motion_factor(cpi, &next_frame));
2206 // Break clause to detect very still sections after motion. For example,
2207 // a static image after a fade or other transition.
2208 if (detect_transition_to_still(cpi, i, 5, loop_decay_rate,
2209 last_loop_decay_rate)) {
2215 // Calculate a boost number for this frame.
2218 calc_frame_boost(cpi, &next_frame, this_frame_mv_in_out, GF_MAX_BOOST);
2220 // Break out conditions.
2222 // Break at active_max_gf_interval unless almost totally static.
2223 ((i >= active_max_gf_interval) && (zero_motion_accumulator < 0.995)) ||
2225 // Don't break out with a very short interval.
2226 (i >= active_min_gf_interval) &&
2227 // If possible dont break very close to a kf
2228 ((rc->frames_to_key - i) >= rc->min_gf_interval) &&
2229 (!flash_detected) &&
2230 ((mv_ratio_accumulator > mv_ratio_accumulator_thresh) ||
2231 (abs_mv_in_out_accumulator > 3.0) ||
2232 (mv_in_out_accumulator < -2.0) ||
2233 ((boost_score - old_boost_score) < BOOST_BREAKOUT)))) {
2234 boost_score = old_boost_score;
2238 *this_frame = next_frame;
2239 old_boost_score = boost_score;
2242 // Was the group length constrained by the requirement for a new KF?
2243 rc->constrained_gf_group = (i >= rc->frames_to_key) ? 1 : 0;
2245 // Should we use the alternate reference frame.
2246 if (allow_alt_ref && (i < cpi->oxcf.lag_in_frames) &&
2247 (i >= rc->min_gf_interval)) {
2248 // Calculate the boost for alt ref.
2250 calc_arf_boost(cpi, 0, (i - 1), (i - 1), &f_boost, &b_boost);
2251 rc->source_alt_ref_pending = 1;
2253 // Test to see if multi arf is appropriate.
2254 cpi->multi_arf_enabled =
2255 (cpi->multi_arf_allowed && (rc->baseline_gf_interval >= 6) &&
2256 (zero_motion_accumulator < 0.995))
2260 rc->gfu_boost = VPXMAX((int)boost_score, MIN_ARF_GF_BOOST);
2261 rc->source_alt_ref_pending = 0;
2264 // Set the interval until the next gf.
2265 rc->baseline_gf_interval = i - (is_key_frame || rc->source_alt_ref_pending);
2267 // Only encode alt reference frame in temporal base layer. So
2268 // baseline_gf_interval should be multiple of a temporal layer group
2269 // (typically the frame distance between two base layer frames)
2270 if (is_two_pass_svc(cpi) && cpi->svc.number_temporal_layers > 1) {
2271 int count = (1 << (cpi->svc.number_temporal_layers - 1)) - 1;
2272 int new_gf_interval = (rc->baseline_gf_interval + count) & (~count);
2274 for (j = 0; j < new_gf_interval - rc->baseline_gf_interval; ++j) {
2275 if (EOF == input_stats(twopass, this_frame)) break;
2276 gf_group_err += calculate_modified_err(cpi, twopass, oxcf, this_frame);
2277 gf_group_raw_error += this_frame->coded_error;
2278 gf_group_noise += this_frame->frame_noise_energy;
2279 gf_group_skip_pct += this_frame->intra_skip_pct;
2280 gf_group_inactive_zone_rows += this_frame->inactive_zone_rows;
2281 gf_group_inter += this_frame->pcnt_inter;
2282 gf_group_motion += this_frame->pcnt_motion;
2284 rc->baseline_gf_interval = new_gf_interval;
2287 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
2289 // Reset the file position.
2290 reset_fpf_position(twopass, start_pos);
2292 // Calculate the bits to be allocated to the gf/arf group as a whole
2293 gf_group_bits = calculate_total_gf_group_bits(cpi, gf_group_err);
2295 // Calculate an estimate of the maxq needed for the group.
2296 // We are more agressive about correcting for sections
2297 // where there could be significant overshoot than for easier
2298 // sections where we do not wish to risk creating an overshoot
2299 // of the allocated bit budget.
2300 if ((cpi->oxcf.rc_mode != VPX_Q) && (rc->baseline_gf_interval > 1)) {
2301 const int vbr_group_bits_per_frame =
2302 (int)(gf_group_bits / rc->baseline_gf_interval);
2303 const double group_av_err = gf_group_raw_error / rc->baseline_gf_interval;
2304 const double group_av_noise = gf_group_noise / rc->baseline_gf_interval;
2305 const double group_av_skip_pct =
2306 gf_group_skip_pct / rc->baseline_gf_interval;
2307 const double group_av_inactive_zone =
2308 ((gf_group_inactive_zone_rows * 2) /
2309 (rc->baseline_gf_interval * (double)cm->mb_rows));
2310 int tmp_q = get_twopass_worst_quality(
2311 cpi, group_av_err, (group_av_skip_pct + group_av_inactive_zone),
2312 group_av_noise, vbr_group_bits_per_frame);
2313 twopass->active_worst_quality =
2314 (tmp_q + (twopass->active_worst_quality * 3)) >> 2;
2317 // Context Adjustment of ARNR filter strength
2318 if (rc->baseline_gf_interval > 1) {
2319 adjust_group_arnr_filter(cpi, (gf_group_noise / rc->baseline_gf_interval),
2320 (gf_group_inter / rc->baseline_gf_interval),
2321 (gf_group_motion / rc->baseline_gf_interval));
2323 twopass->arnr_strength_adjustment = 0;
2326 // Calculate the extra bits to be used for boosted frame(s)
2327 gf_arf_bits = calculate_boost_bits(rc->baseline_gf_interval, rc->gfu_boost,
2330 // Adjust KF group bits and error remaining.
2331 twopass->kf_group_error_left -= (int64_t)gf_group_err;
2333 // Allocate bits to each of the frames in the GF group.
2334 allocate_gf_group_bits(cpi, gf_group_bits, gf_arf_bits);
2336 // Reset the file position.
2337 reset_fpf_position(twopass, start_pos);
2339 // Calculate a section intra ratio used in setting max loop filter.
2340 if (cpi->common.frame_type != KEY_FRAME) {
2341 twopass->section_intra_rating = calculate_section_intra_ratio(
2342 start_pos, twopass->stats_in_end, rc->baseline_gf_interval);
2345 if (oxcf->resize_mode == RESIZE_DYNAMIC) {
2346 // Default to starting GF groups at normal frame size.
2347 cpi->rc.next_frame_size_selector = UNSCALED;
2350 // Reset rolling actual and target bits counters for ARF groups.
2351 twopass->rolling_arf_group_target_bits = 0;
2352 twopass->rolling_arf_group_actual_bits = 0;
2355 // Threshold for use of the lagging second reference frame. High second ref
2356 // usage may point to a transient event like a flash or occlusion rather than
2357 // a real scene cut.
2358 #define SECOND_REF_USEAGE_THRESH 0.1
2359 // Minimum % intra coding observed in first pass (1.0 = 100%)
2360 #define MIN_INTRA_LEVEL 0.25
2361 // Minimum ratio between the % of intra coding and inter coding in the first
2362 // pass after discounting neutral blocks (discounting neutral blocks in this
2363 // way helps catch scene cuts in clips with very flat areas or letter box
2364 // format clips with image padding.
2365 #define INTRA_VS_INTER_THRESH 2.0
2366 // Hard threshold where the first pass chooses intra for almost all blocks.
2367 // In such a case even if the frame is not a scene cut coding a key frame
2368 // may be a good option.
2369 #define VERY_LOW_INTER_THRESH 0.05
2370 // Maximum threshold for the relative ratio of intra error score vs best
2371 // inter error score.
2372 #define KF_II_ERR_THRESHOLD 2.5
2373 // In real scene cuts there is almost always a sharp change in the intra
2374 // or inter error score.
2375 #define ERR_CHANGE_THRESHOLD 0.4
2376 // For real scene cuts we expect an improvment in the intra inter error
2377 // ratio in the next frame.
2378 #define II_IMPROVEMENT_THRESHOLD 3.5
2379 #define KF_II_MAX 128.0
2381 static int test_candidate_kf(TWO_PASS *twopass,
2382 const FIRSTPASS_STATS *last_frame,
2383 const FIRSTPASS_STATS *this_frame,
2384 const FIRSTPASS_STATS *next_frame) {
2385 int is_viable_kf = 0;
2386 double pcnt_intra = 1.0 - this_frame->pcnt_inter;
2387 double modified_pcnt_inter =
2388 this_frame->pcnt_inter - this_frame->pcnt_neutral;
2390 // Does the frame satisfy the primary criteria of a key frame?
2391 // See above for an explanation of the test criteria.
2392 // If so, then examine how well it predicts subsequent frames.
2393 if ((this_frame->pcnt_second_ref < SECOND_REF_USEAGE_THRESH) &&
2394 (next_frame->pcnt_second_ref < SECOND_REF_USEAGE_THRESH) &&
2395 ((this_frame->pcnt_inter < VERY_LOW_INTER_THRESH) ||
2396 ((pcnt_intra > MIN_INTRA_LEVEL) &&
2397 (pcnt_intra > (INTRA_VS_INTER_THRESH * modified_pcnt_inter)) &&
2398 ((this_frame->intra_error /
2399 DOUBLE_DIVIDE_CHECK(this_frame->coded_error)) <
2400 KF_II_ERR_THRESHOLD) &&
2401 ((fabs(last_frame->coded_error - this_frame->coded_error) /
2402 DOUBLE_DIVIDE_CHECK(this_frame->coded_error) >
2403 ERR_CHANGE_THRESHOLD) ||
2404 (fabs(last_frame->intra_error - this_frame->intra_error) /
2405 DOUBLE_DIVIDE_CHECK(this_frame->intra_error) >
2406 ERR_CHANGE_THRESHOLD) ||
2407 ((next_frame->intra_error /
2408 DOUBLE_DIVIDE_CHECK(next_frame->coded_error)) >
2409 II_IMPROVEMENT_THRESHOLD))))) {
2411 const FIRSTPASS_STATS *start_pos = twopass->stats_in;
2412 FIRSTPASS_STATS local_next_frame = *next_frame;
2413 double boost_score = 0.0;
2414 double old_boost_score = 0.0;
2415 double decay_accumulator = 1.0;
2417 // Examine how well the key frame predicts subsequent frames.
2418 for (i = 0; i < 16; ++i) {
2419 double next_iiratio = (BOOST_FACTOR * local_next_frame.intra_error /
2420 DOUBLE_DIVIDE_CHECK(local_next_frame.coded_error));
2422 if (next_iiratio > KF_II_MAX) next_iiratio = KF_II_MAX;
2424 // Cumulative effect of decay in prediction quality.
2425 if (local_next_frame.pcnt_inter > 0.85)
2426 decay_accumulator *= local_next_frame.pcnt_inter;
2428 decay_accumulator *= (0.85 + local_next_frame.pcnt_inter) / 2.0;
2430 // Keep a running total.
2431 boost_score += (decay_accumulator * next_iiratio);
2433 // Test various breakout clauses.
2434 if ((local_next_frame.pcnt_inter < 0.05) || (next_iiratio < 1.5) ||
2435 (((local_next_frame.pcnt_inter - local_next_frame.pcnt_neutral) <
2437 (next_iiratio < 3.0)) ||
2438 ((boost_score - old_boost_score) < 3.0) ||
2439 (local_next_frame.intra_error < 200)) {
2443 old_boost_score = boost_score;
2445 // Get the next frame details
2446 if (EOF == input_stats(twopass, &local_next_frame)) break;
2449 // If there is tolerable prediction for at least the next 3 frames then
2450 // break out else discard this potential key frame and move on
2451 if (boost_score > 30.0 && (i > 3)) {
2454 // Reset the file position
2455 reset_fpf_position(twopass, start_pos);
2461 return is_viable_kf;
2464 #define FRAMES_TO_CHECK_DECAY 8
2466 static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
2468 RATE_CONTROL *const rc = &cpi->rc;
2469 TWO_PASS *const twopass = &cpi->twopass;
2470 GF_GROUP *const gf_group = &twopass->gf_group;
2471 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
2472 const FIRSTPASS_STATS first_frame = *this_frame;
2473 const FIRSTPASS_STATS *const start_position = twopass->stats_in;
2474 FIRSTPASS_STATS next_frame;
2475 FIRSTPASS_STATS last_frame;
2477 int loop_decay_counter = 0;
2478 double decay_accumulator = 1.0;
2479 double av_decay_accumulator = 0.0;
2480 double zero_motion_accumulator = 1.0;
2481 double boost_score = 0.0;
2482 double kf_mod_err = 0.0;
2483 double kf_group_err = 0.0;
2484 double recent_loop_decay[FRAMES_TO_CHECK_DECAY];
2486 vp9_zero(next_frame);
2488 cpi->common.frame_type = KEY_FRAME;
2490 // Reset the GF group data structures.
2491 vp9_zero(*gf_group);
2493 // Is this a forced key frame by interval.
2494 rc->this_key_frame_forced = rc->next_key_frame_forced;
2496 // Clear the alt ref active flag and last group multi arf flags as they
2497 // can never be set for a key frame.
2498 rc->source_alt_ref_active = 0;
2499 cpi->multi_arf_last_grp_enabled = 0;
2501 // KF is always a GF so clear frames till next gf counter.
2502 rc->frames_till_gf_update_due = 0;
2504 rc->frames_to_key = 1;
2506 twopass->kf_group_bits = 0; // Total bits available to kf group
2507 twopass->kf_group_error_left = 0; // Group modified error score.
2509 kf_mod_err = calculate_modified_err(cpi, twopass, oxcf, this_frame);
2511 // Initialize the decay rates for the recent frames to check
2512 for (j = 0; j < FRAMES_TO_CHECK_DECAY; ++j) recent_loop_decay[j] = 1.0;
2514 // Find the next keyframe.
2516 while (twopass->stats_in < twopass->stats_in_end &&
2517 rc->frames_to_key < cpi->oxcf.key_freq) {
2518 // Accumulate kf group error.
2519 kf_group_err += calculate_modified_err(cpi, twopass, oxcf, this_frame);
2521 // Load the next frame's stats.
2522 last_frame = *this_frame;
2523 input_stats(twopass, this_frame);
2525 // Provided that we are not at the end of the file...
2526 if (cpi->oxcf.auto_key && twopass->stats_in < twopass->stats_in_end) {
2527 double loop_decay_rate;
2529 // Check for a scene cut.
2530 if (test_candidate_kf(twopass, &last_frame, this_frame,
2534 // How fast is the prediction quality decaying?
2535 loop_decay_rate = get_prediction_decay_rate(cpi, twopass->stats_in);
2537 // We want to know something about the recent past... rather than
2538 // as used elsewhere where we are concerned with decay in prediction
2539 // quality since the last GF or KF.
2540 recent_loop_decay[i % FRAMES_TO_CHECK_DECAY] = loop_decay_rate;
2541 decay_accumulator = 1.0;
2542 for (j = 0; j < FRAMES_TO_CHECK_DECAY; ++j)
2543 decay_accumulator *= recent_loop_decay[j];
2545 // Special check for transition or high motion followed by a
2547 if (detect_transition_to_still(cpi, i, cpi->oxcf.key_freq - i,
2548 loop_decay_rate, decay_accumulator))
2551 // Step on to the next frame.
2552 ++rc->frames_to_key;
2554 // If we don't have a real key frame within the next two
2555 // key_freq intervals then break out of the loop.
2556 if (rc->frames_to_key >= 2 * cpi->oxcf.key_freq) break;
2558 ++rc->frames_to_key;
2563 // If there is a max kf interval set by the user we must obey it.
2564 // We already breakout of the loop above at 2x max.
2565 // This code centers the extra kf if the actual natural interval
2566 // is between 1x and 2x.
2567 if (cpi->oxcf.auto_key && rc->frames_to_key > cpi->oxcf.key_freq) {
2568 FIRSTPASS_STATS tmp_frame = first_frame;
2570 rc->frames_to_key /= 2;
2572 // Reset to the start of the group.
2573 reset_fpf_position(twopass, start_position);
2577 // Rescan to get the correct error data for the forced kf group.
2578 for (i = 0; i < rc->frames_to_key; ++i) {
2579 kf_group_err += calculate_modified_err(cpi, twopass, oxcf, &tmp_frame);
2580 input_stats(twopass, &tmp_frame);
2582 rc->next_key_frame_forced = 1;
2583 } else if (twopass->stats_in == twopass->stats_in_end ||
2584 rc->frames_to_key >= cpi->oxcf.key_freq) {
2585 rc->next_key_frame_forced = 1;
2587 rc->next_key_frame_forced = 0;
2590 if (is_two_pass_svc(cpi) && cpi->svc.number_temporal_layers > 1) {
2591 int count = (1 << (cpi->svc.number_temporal_layers - 1)) - 1;
2592 int new_frame_to_key = (rc->frames_to_key + count) & (~count);
2594 for (j = 0; j < new_frame_to_key - rc->frames_to_key; ++j) {
2595 if (EOF == input_stats(twopass, this_frame)) break;
2596 kf_group_err += calculate_modified_err(cpi, twopass, oxcf, this_frame);
2598 rc->frames_to_key = new_frame_to_key;
2601 // Special case for the last key frame of the file.
2602 if (twopass->stats_in >= twopass->stats_in_end) {
2603 // Accumulate kf group error.
2604 kf_group_err += calculate_modified_err(cpi, twopass, oxcf, this_frame);
2607 // Calculate the number of bits that should be assigned to the kf group.
2608 if (twopass->bits_left > 0 && twopass->modified_error_left > 0.0) {
2609 // Maximum number of bits for a single normal frame (not key frame).
2610 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
2612 // Maximum number of bits allocated to the key frame group.
2613 int64_t max_grp_bits;
2615 // Default allocation based on bits left and relative
2616 // complexity of the section.
2617 twopass->kf_group_bits = (int64_t)(
2618 twopass->bits_left * (kf_group_err / twopass->modified_error_left));
2620 // Clip based on maximum per frame rate defined by the user.
2621 max_grp_bits = (int64_t)max_bits * (int64_t)rc->frames_to_key;
2622 if (twopass->kf_group_bits > max_grp_bits)
2623 twopass->kf_group_bits = max_grp_bits;
2625 twopass->kf_group_bits = 0;
2627 twopass->kf_group_bits = VPXMAX(0, twopass->kf_group_bits);
2629 // Reset the first pass file position.
2630 reset_fpf_position(twopass, start_position);
2632 // Scan through the kf group collating various stats used to determine
2633 // how many bits to spend on it.
2634 decay_accumulator = 1.0;
2636 for (i = 0; i < (rc->frames_to_key - 1); ++i) {
2637 if (EOF == input_stats(twopass, &next_frame)) break;
2639 // Monitor for static sections.
2640 zero_motion_accumulator = VPXMIN(zero_motion_accumulator,
2641 get_zero_motion_factor(cpi, &next_frame));
2643 // Not all frames in the group are necessarily used in calculating boost.
2644 if ((i <= rc->max_gf_interval) ||
2645 ((i <= (rc->max_gf_interval * 4)) && (decay_accumulator > 0.5))) {
2646 const double frame_boost =
2647 calc_frame_boost(cpi, &next_frame, 0, KF_MAX_BOOST);
2649 // How fast is prediction quality decaying.
2650 if (!detect_flash(twopass, 0)) {
2651 const double loop_decay_rate =
2652 get_prediction_decay_rate(cpi, &next_frame);
2653 decay_accumulator *= loop_decay_rate;
2654 decay_accumulator = VPXMAX(decay_accumulator, MIN_DECAY_FACTOR);
2655 av_decay_accumulator += decay_accumulator;
2656 ++loop_decay_counter;
2658 boost_score += (decay_accumulator * frame_boost);
2661 av_decay_accumulator /= (double)loop_decay_counter;
2663 reset_fpf_position(twopass, start_position);
2665 // Store the zero motion percentage
2666 twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
2668 // Calculate a section intra ratio used in setting max loop filter.
2669 twopass->section_intra_rating = calculate_section_intra_ratio(
2670 start_position, twopass->stats_in_end, rc->frames_to_key);
2672 // Apply various clamps for min and max boost
2673 rc->kf_boost = (int)(av_decay_accumulator * boost_score);
2674 rc->kf_boost = VPXMAX(rc->kf_boost, (rc->frames_to_key * 3));
2675 rc->kf_boost = VPXMAX(rc->kf_boost, MIN_KF_BOOST);
2677 // Work out how many bits to allocate for the key frame itself.
2678 kf_bits = calculate_boost_bits((rc->frames_to_key - 1), rc->kf_boost,
2679 twopass->kf_group_bits);
2681 twopass->kf_group_bits -= kf_bits;
2683 // Save the bits to spend on the key frame.
2684 gf_group->bit_allocation[0] = kf_bits;
2685 gf_group->update_type[0] = KF_UPDATE;
2686 gf_group->rf_level[0] = KF_STD;
2688 // Note the total error score of the kf group minus the key frame itself.
2689 twopass->kf_group_error_left = (int)(kf_group_err - kf_mod_err);
2691 // Adjust the count of total modified error left.
2692 // The count of bits left is adjusted elsewhere based on real coded frame
2694 twopass->modified_error_left -= kf_group_err;
2696 if (oxcf->resize_mode == RESIZE_DYNAMIC) {
2697 // Default to normal-sized frame on keyframes.
2698 cpi->rc.next_frame_size_selector = UNSCALED;
2702 // Define the reference buffers that will be updated post encode.
2703 static void configure_buffer_updates(VP9_COMP *cpi) {
2704 TWO_PASS *const twopass = &cpi->twopass;
2706 cpi->rc.is_src_frame_alt_ref = 0;
2707 switch (twopass->gf_group.update_type[twopass->gf_group.index]) {
2709 cpi->refresh_last_frame = 1;
2710 cpi->refresh_golden_frame = 1;
2711 cpi->refresh_alt_ref_frame = 1;
2714 cpi->refresh_last_frame = 1;
2715 cpi->refresh_golden_frame = 0;
2716 cpi->refresh_alt_ref_frame = 0;
2719 cpi->refresh_last_frame = 1;
2720 cpi->refresh_golden_frame = 1;
2721 cpi->refresh_alt_ref_frame = 0;
2723 case OVERLAY_UPDATE:
2724 cpi->refresh_last_frame = 0;
2725 cpi->refresh_golden_frame = 1;
2726 cpi->refresh_alt_ref_frame = 0;
2727 cpi->rc.is_src_frame_alt_ref = 1;
2730 cpi->refresh_last_frame = 0;
2731 cpi->refresh_golden_frame = 0;
2732 cpi->refresh_alt_ref_frame = 1;
2734 default: assert(0); break;
2736 if (is_two_pass_svc(cpi)) {
2737 if (cpi->svc.temporal_layer_id > 0) {
2738 cpi->refresh_last_frame = 0;
2739 cpi->refresh_golden_frame = 0;
2741 if (cpi->svc.layer_context[cpi->svc.spatial_layer_id].gold_ref_idx < 0)
2742 cpi->refresh_golden_frame = 0;
2743 if (cpi->alt_ref_source == NULL) cpi->refresh_alt_ref_frame = 0;
2747 static int is_skippable_frame(const VP9_COMP *cpi) {
2748 // If the current frame does not have non-zero motion vector detected in the
2749 // first pass, and so do its previous and forward frames, then this frame
2750 // can be skipped for partition check, and the partition size is assigned
2751 // according to the variance
2752 const SVC *const svc = &cpi->svc;
2753 const TWO_PASS *const twopass =
2754 is_two_pass_svc(cpi) ? &svc->layer_context[svc->spatial_layer_id].twopass
2757 return (!frame_is_intra_only(&cpi->common) &&
2758 twopass->stats_in - 2 > twopass->stats_in_start &&
2759 twopass->stats_in < twopass->stats_in_end &&
2760 (twopass->stats_in - 1)->pcnt_inter -
2761 (twopass->stats_in - 1)->pcnt_motion ==
2763 (twopass->stats_in - 2)->pcnt_inter -
2764 (twopass->stats_in - 2)->pcnt_motion ==
2766 twopass->stats_in->pcnt_inter - twopass->stats_in->pcnt_motion == 1);
2769 void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
2770 VP9_COMMON *const cm = &cpi->common;
2771 RATE_CONTROL *const rc = &cpi->rc;
2772 TWO_PASS *const twopass = &cpi->twopass;
2773 GF_GROUP *const gf_group = &twopass->gf_group;
2774 FIRSTPASS_STATS this_frame;
2777 LAYER_CONTEXT *const lc =
2778 is_two_pass_svc(cpi) ? &cpi->svc.layer_context[cpi->svc.spatial_layer_id]
2781 if (!twopass->stats_in) return;
2783 // If this is an arf frame then we dont want to read the stats file or
2784 // advance the input pointer as we already have what we need.
2785 if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
2787 configure_buffer_updates(cpi);
2788 target_rate = gf_group->bit_allocation[gf_group->index];
2789 target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
2790 rc->base_frame_target = target_rate;
2792 cm->frame_type = INTER_FRAME;
2795 if (cpi->svc.spatial_layer_id == 0) {
2796 lc->is_key_frame = 0;
2798 lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
2800 if (lc->is_key_frame) cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
2804 // Do the firstpass stats indicate that this frame is skippable for the
2805 // partition search?
2806 if (cpi->sf.allow_partition_search_skip && cpi->oxcf.pass == 2 &&
2807 (!cpi->use_svc || is_two_pass_svc(cpi))) {
2808 cpi->partition_search_skippable_frame = is_skippable_frame(cpi);
2814 vpx_clear_system_state();
2816 if (cpi->oxcf.rc_mode == VPX_Q) {
2817 twopass->active_worst_quality = cpi->oxcf.cq_level;
2818 } else if (cm->current_video_frame == 0 ||
2819 (lc != NULL && lc->current_video_frame_in_layer == 0)) {
2820 const int frames_left =
2821 (int)(twopass->total_stats.count -
2822 ((lc != NULL) ? lc->current_video_frame_in_layer
2823 : cm->current_video_frame));
2824 // Special case code for first frame.
2825 const int section_target_bandwidth =
2826 (int)(twopass->bits_left / frames_left);
2827 const double section_length = twopass->total_left_stats.count;
2828 const double section_error =
2829 twopass->total_left_stats.coded_error / section_length;
2830 const double section_intra_skip =
2831 twopass->total_left_stats.intra_skip_pct / section_length;
2832 const double section_inactive_zone =
2833 (twopass->total_left_stats.inactive_zone_rows * 2) /
2834 ((double)cm->mb_rows * section_length);
2835 const double section_noise =
2836 twopass->total_left_stats.frame_noise_energy / section_length;
2839 tmp_q = get_twopass_worst_quality(
2840 cpi, section_error, section_intra_skip + section_inactive_zone,
2841 section_noise, section_target_bandwidth);
2843 twopass->active_worst_quality = tmp_q;
2844 twopass->baseline_active_worst_quality = tmp_q;
2845 rc->ni_av_qi = tmp_q;
2846 rc->last_q[INTER_FRAME] = tmp_q;
2847 rc->avg_q = vp9_convert_qindex_to_q(tmp_q, cm->bit_depth);
2848 rc->avg_frame_qindex[INTER_FRAME] = tmp_q;
2849 rc->last_q[KEY_FRAME] = (tmp_q + cpi->oxcf.best_allowed_q) / 2;
2850 rc->avg_frame_qindex[KEY_FRAME] = rc->last_q[KEY_FRAME];
2852 vp9_zero(this_frame);
2853 if (EOF == input_stats(twopass, &this_frame)) return;
2855 // Set the frame content type flag.
2856 if (this_frame.intra_skip_pct >= FC_ANIMATION_THRESH)
2857 twopass->fr_content_type = FC_GRAPHICS_ANIMATION;
2859 twopass->fr_content_type = FC_NORMAL;
2861 // Keyframe and section processing.
2862 if (rc->frames_to_key == 0 || (cpi->frame_flags & FRAMEFLAGS_KEY)) {
2863 FIRSTPASS_STATS this_frame_copy;
2864 this_frame_copy = this_frame;
2865 // Define next KF group and assign bits to it.
2866 find_next_key_frame(cpi, &this_frame);
2867 this_frame = this_frame_copy;
2869 cm->frame_type = INTER_FRAME;
2873 if (cpi->svc.spatial_layer_id == 0) {
2874 lc->is_key_frame = (cm->frame_type == KEY_FRAME);
2875 if (lc->is_key_frame) {
2876 cpi->ref_frame_flags &=
2877 (~VP9_LAST_FLAG & ~VP9_GOLD_FLAG & ~VP9_ALT_FLAG);
2878 lc->frames_from_key_frame = 0;
2879 // Encode an intra only empty frame since we have a key frame.
2880 cpi->svc.encode_intra_empty_frame = 1;
2883 cm->frame_type = INTER_FRAME;
2884 lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
2886 if (lc->is_key_frame) {
2887 cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
2888 lc->frames_from_key_frame = 0;
2893 // Define a new GF/ARF group. (Should always enter here for key frames).
2894 if (rc->frames_till_gf_update_due == 0) {
2895 define_gf_group(cpi, &this_frame);
2897 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
2898 if (lc != NULL) cpi->refresh_golden_frame = 1;
2900 #if ARF_STATS_OUTPUT
2903 fpfile = fopen("arf.stt", "a");
2905 fprintf(fpfile, "%10d %10ld %10d %10d %10ld\n", cm->current_video_frame,
2906 rc->frames_till_gf_update_due, rc->kf_boost, arf_count,
2914 configure_buffer_updates(cpi);
2916 // Do the firstpass stats indicate that this frame is skippable for the
2917 // partition search?
2918 if (cpi->sf.allow_partition_search_skip && cpi->oxcf.pass == 2 &&
2919 (!cpi->use_svc || is_two_pass_svc(cpi))) {
2920 cpi->partition_search_skippable_frame = is_skippable_frame(cpi);
2923 target_rate = gf_group->bit_allocation[gf_group->index];
2924 rc->base_frame_target = target_rate;
2927 const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
2930 // The multiplication by 256 reverses a scaling factor of (>> 8)
2931 // applied when combining MB error values for the frame.
2932 twopass->mb_av_energy =
2933 log(((this_frame.intra_error * 256.0) / num_mbs) + 1.0);
2934 twopass->mb_smooth_pct = this_frame.intra_smooth_pct;
2937 // Update the total stats remaining structure.
2938 subtract_stats(&twopass->total_left_stats, &this_frame);
2941 #define MINQ_ADJ_LIMIT 48
2942 #define MINQ_ADJ_LIMIT_CQ 20
2943 #define HIGH_UNDERSHOOT_RATIO 2
2944 void vp9_twopass_postencode_update(VP9_COMP *cpi) {
2945 TWO_PASS *const twopass = &cpi->twopass;
2946 RATE_CONTROL *const rc = &cpi->rc;
2947 VP9_COMMON *const cm = &cpi->common;
2948 const int bits_used = rc->base_frame_target;
2950 // VBR correction is done through rc->vbr_bits_off_target. Based on the
2951 // sign of this value, a limited % adjustment is made to the target rate
2952 // of subsequent frames, to try and push it back towards 0. This method
2953 // is designed to prevent extreme behaviour at the end of a clip
2954 // or group of frames.
2955 rc->vbr_bits_off_target += rc->base_frame_target - rc->projected_frame_size;
2956 twopass->bits_left = VPXMAX(twopass->bits_left - bits_used, 0);
2958 // Target vs actual bits for this arf group.
2959 twopass->rolling_arf_group_target_bits += rc->this_frame_target;
2960 twopass->rolling_arf_group_actual_bits += rc->projected_frame_size;
2962 // Calculate the pct rc error.
2963 if (rc->total_actual_bits) {
2964 rc->rate_error_estimate =
2965 (int)((rc->vbr_bits_off_target * 100) / rc->total_actual_bits);
2966 rc->rate_error_estimate = clamp(rc->rate_error_estimate, -100, 100);
2968 rc->rate_error_estimate = 0;
2971 if (cpi->common.frame_type != KEY_FRAME &&
2972 !vp9_is_upper_layer_key_frame(cpi)) {
2973 twopass->kf_group_bits -= bits_used;
2974 twopass->last_kfgroup_zeromotion_pct = twopass->kf_zeromotion_pct;
2976 twopass->kf_group_bits = VPXMAX(twopass->kf_group_bits, 0);
2978 // Increment the gf group index ready for the next frame.
2979 ++twopass->gf_group.index;
2981 // If the rate control is drifting consider adjustment to min or maxq.
2982 if ((cpi->oxcf.rc_mode != VPX_Q) && !cpi->rc.is_src_frame_alt_ref) {
2983 const int maxq_adj_limit =
2984 rc->worst_quality - twopass->active_worst_quality;
2985 const int minq_adj_limit =
2986 (cpi->oxcf.rc_mode == VPX_CQ ? MINQ_ADJ_LIMIT_CQ : MINQ_ADJ_LIMIT);
2987 int aq_extend_min = 0;
2988 int aq_extend_max = 0;
2990 // Extend min or Max Q range to account for imbalance from the base
2991 // value when using AQ.
2992 if (cpi->oxcf.aq_mode != NO_AQ) {
2993 if (cm->seg.aq_av_offset < 0) {
2994 // The balance of the AQ map tends towarda lowering the average Q.
2996 aq_extend_max = VPXMIN(maxq_adj_limit, -cm->seg.aq_av_offset);
2998 // The balance of the AQ map tends towards raising the average Q.
2999 aq_extend_min = VPXMIN(minq_adj_limit, cm->seg.aq_av_offset);
3005 if (rc->rate_error_estimate > cpi->oxcf.under_shoot_pct) {
3006 --twopass->extend_maxq;
3007 if (rc->rolling_target_bits >= rc->rolling_actual_bits)
3008 ++twopass->extend_minq;
3010 } else if (rc->rate_error_estimate < -cpi->oxcf.over_shoot_pct) {
3011 --twopass->extend_minq;
3012 if (rc->rolling_target_bits < rc->rolling_actual_bits)
3013 ++twopass->extend_maxq;
3015 // Adjustment for extreme local overshoot.
3016 if (rc->projected_frame_size > (2 * rc->base_frame_target) &&
3017 rc->projected_frame_size > (2 * rc->avg_frame_bandwidth))
3018 ++twopass->extend_maxq;
3020 // Unwind undershoot or overshoot adjustment.
3021 if (rc->rolling_target_bits < rc->rolling_actual_bits)
3022 --twopass->extend_minq;
3023 else if (rc->rolling_target_bits > rc->rolling_actual_bits)
3024 --twopass->extend_maxq;
3027 twopass->extend_minq =
3028 clamp(twopass->extend_minq, aq_extend_min, minq_adj_limit);
3029 twopass->extend_maxq =
3030 clamp(twopass->extend_maxq, aq_extend_max, maxq_adj_limit);
3032 // If there is a big and undexpected undershoot then feed the extra
3033 // bits back in quickly. One situation where this may happen is if a
3034 // frame is unexpectedly almost perfectly predicted by the ARF or GF
3035 // but not very well predcited by the previous frame.
3036 if (!frame_is_kf_gf_arf(cpi) && !cpi->rc.is_src_frame_alt_ref) {
3037 int fast_extra_thresh = rc->base_frame_target / HIGH_UNDERSHOOT_RATIO;
3038 if (rc->projected_frame_size < fast_extra_thresh) {
3039 rc->vbr_bits_off_target_fast +=
3040 fast_extra_thresh - rc->projected_frame_size;
3041 rc->vbr_bits_off_target_fast =
3042 VPXMIN(rc->vbr_bits_off_target_fast, (4 * rc->avg_frame_bandwidth));
3044 // Fast adaptation of minQ if necessary to use up the extra bits.
3045 if (rc->avg_frame_bandwidth) {
3046 twopass->extend_minq_fast =
3047 (int)(rc->vbr_bits_off_target_fast * 8 / rc->avg_frame_bandwidth);
3049 twopass->extend_minq_fast = VPXMIN(
3050 twopass->extend_minq_fast, minq_adj_limit - twopass->extend_minq);
3051 } else if (rc->vbr_bits_off_target_fast) {
3052 twopass->extend_minq_fast = VPXMIN(
3053 twopass->extend_minq_fast, minq_adj_limit - twopass->extend_minq);
3055 twopass->extend_minq_fast = 0;