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_scale_rtcd.h"
17 #include "vpx_mem/vpx_mem.h"
18 #include "vpx_scale/vpx_scale.h"
19 #include "vpx_scale/yv12config.h"
21 #include "vp9/common/vp9_entropymv.h"
22 #include "vp9/common/vp9_quant_common.h"
23 #include "vp9/common/vp9_reconinter.h" // vp9_setup_dst_planes()
24 #include "vp9/common/vp9_systemdependent.h"
26 #include "vp9/encoder/vp9_aq_variance.h"
27 #include "vp9/encoder/vp9_block.h"
28 #include "vp9/encoder/vp9_encodeframe.h"
29 #include "vp9/encoder/vp9_encodemb.h"
30 #include "vp9/encoder/vp9_encodemv.h"
31 #include "vp9/encoder/vp9_encoder.h"
32 #include "vp9/encoder/vp9_extend.h"
33 #include "vp9/encoder/vp9_firstpass.h"
34 #include "vp9/encoder/vp9_mcomp.h"
35 #include "vp9/encoder/vp9_quantize.h"
36 #include "vp9/encoder/vp9_rdopt.h"
37 #include "vp9/encoder/vp9_variance.h"
42 #define IIKFACTOR1 12.5
43 #define IIKFACTOR2 15.0
46 #define ERR_DIVISOR 150.0
47 #define MIN_DECAY_FACTOR 0.1
48 #define SVC_FACTOR_PT_LOW 0.45
49 #define FACTOR_PT_LOW 0.5
50 #define FACTOR_PT_HIGH 0.9
52 #define KF_MB_INTRA_MIN 150
53 #define GF_MB_INTRA_MIN 100
55 #define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x) - 0.000001 : (x) + 0.000001)
57 #define MIN_KF_BOOST 300
58 #define MIN_GF_INTERVAL 4
59 #define LONG_TERM_VBR_CORRECTION
61 static void swap_yv12(YV12_BUFFER_CONFIG *a, YV12_BUFFER_CONFIG *b) {
62 YV12_BUFFER_CONFIG temp = *a;
67 static int gfboost_qadjust(int qindex) {
68 const double q = vp9_convert_qindex_to_q(qindex);
69 return (int)((0.00000828 * q * q * q) +
74 // Resets the first pass file to the given position using a relative seek from
75 // the current position.
76 static void reset_fpf_position(TWO_PASS *p,
77 const FIRSTPASS_STATS *position) {
78 p->stats_in = position;
81 static int lookup_next_frame_stats(const TWO_PASS *p,
82 FIRSTPASS_STATS *next_frame) {
83 if (p->stats_in >= p->stats_in_end)
86 *next_frame = *p->stats_in;
91 // Read frame stats at an offset from the current position.
92 static int read_frame_stats(const TWO_PASS *p,
93 FIRSTPASS_STATS *frame_stats, int offset) {
94 const FIRSTPASS_STATS *fps_ptr = p->stats_in;
96 // Check legality of offset.
98 if (&fps_ptr[offset] >= p->stats_in_end)
100 } else if (offset < 0) {
101 if (&fps_ptr[offset] < p->stats_in_start)
105 *frame_stats = fps_ptr[offset];
109 #if CONFIG_FP_MB_STATS
110 static int input_mb_stats(FIRSTPASS_FRAME_MB_STATS *fp_frame_stats,
111 const VP9_COMMON *const cm) {
115 fpfile = fopen("firstpass_mb.stt", "r");
116 fseek(fpfile, cm->current_video_frame * cm->MBs * sizeof(FIRSTPASS_MB_STATS),
118 ret = fread(fp_frame_stats->mb_stats, sizeof(FIRSTPASS_MB_STATS), cm->MBs,
127 static void output_mb_stats(FIRSTPASS_FRAME_MB_STATS *fp_frame_stats,
128 const VP9_COMMON *const cm) {
131 fpfile = fopen("firstpass_mb.stt", "a");
132 fwrite(fp_frame_stats->mb_stats, sizeof(FIRSTPASS_MB_STATS), cm->MBs, fpfile);
137 static int input_stats(TWO_PASS *p, FIRSTPASS_STATS *fps) {
138 if (p->stats_in >= p->stats_in_end)
146 static void output_stats(FIRSTPASS_STATS *stats,
147 struct vpx_codec_pkt_list *pktlist) {
148 struct vpx_codec_cx_pkt pkt;
149 pkt.kind = VPX_CODEC_STATS_PKT;
150 pkt.data.twopass_stats.buf = stats;
151 pkt.data.twopass_stats.sz = sizeof(FIRSTPASS_STATS);
152 vpx_codec_pkt_list_add(pktlist, &pkt);
158 fpfile = fopen("firstpass.stt", "a");
160 fprintf(fpfile, "%12.0f %12.0f %12.0f %12.0f %12.4f %12.4f"
161 "%12.4f %12.4f %12.4f %12.4f %12.4f %12.4f %12.4f"
162 "%12.0f %12.0f %12.4f %12.0f %12.0f %12.4f\n",
166 stats->sr_coded_error,
169 stats->pcnt_second_ref,
177 stats->mv_in_out_count,
186 static void zero_stats(FIRSTPASS_STATS *section) {
187 section->frame = 0.0;
188 section->intra_error = 0.0;
189 section->coded_error = 0.0;
190 section->sr_coded_error = 0.0;
191 section->pcnt_inter = 0.0;
192 section->pcnt_motion = 0.0;
193 section->pcnt_second_ref = 0.0;
194 section->pcnt_neutral = 0.0;
196 section->mvr_abs = 0.0;
198 section->mvc_abs = 0.0;
201 section->mv_in_out_count = 0.0;
202 section->new_mv_count = 0.0;
203 section->count = 0.0;
204 section->duration = 1.0;
205 section->spatial_layer_id = 0;
208 static void accumulate_stats(FIRSTPASS_STATS *section,
209 const FIRSTPASS_STATS *frame) {
210 section->frame += frame->frame;
211 section->spatial_layer_id = frame->spatial_layer_id;
212 section->intra_error += frame->intra_error;
213 section->coded_error += frame->coded_error;
214 section->sr_coded_error += frame->sr_coded_error;
215 section->pcnt_inter += frame->pcnt_inter;
216 section->pcnt_motion += frame->pcnt_motion;
217 section->pcnt_second_ref += frame->pcnt_second_ref;
218 section->pcnt_neutral += frame->pcnt_neutral;
219 section->MVr += frame->MVr;
220 section->mvr_abs += frame->mvr_abs;
221 section->MVc += frame->MVc;
222 section->mvc_abs += frame->mvc_abs;
223 section->MVrv += frame->MVrv;
224 section->MVcv += frame->MVcv;
225 section->mv_in_out_count += frame->mv_in_out_count;
226 section->new_mv_count += frame->new_mv_count;
227 section->count += frame->count;
228 section->duration += frame->duration;
231 static void subtract_stats(FIRSTPASS_STATS *section,
232 const FIRSTPASS_STATS *frame) {
233 section->frame -= frame->frame;
234 section->intra_error -= frame->intra_error;
235 section->coded_error -= frame->coded_error;
236 section->sr_coded_error -= frame->sr_coded_error;
237 section->pcnt_inter -= frame->pcnt_inter;
238 section->pcnt_motion -= frame->pcnt_motion;
239 section->pcnt_second_ref -= frame->pcnt_second_ref;
240 section->pcnt_neutral -= frame->pcnt_neutral;
241 section->MVr -= frame->MVr;
242 section->mvr_abs -= frame->mvr_abs;
243 section->MVc -= frame->MVc;
244 section->mvc_abs -= frame->mvc_abs;
245 section->MVrv -= frame->MVrv;
246 section->MVcv -= frame->MVcv;
247 section->mv_in_out_count -= frame->mv_in_out_count;
248 section->new_mv_count -= frame->new_mv_count;
249 section->count -= frame->count;
250 section->duration -= frame->duration;
253 static void avg_stats(FIRSTPASS_STATS *section) {
254 if (section->count < 1.0)
257 section->intra_error /= section->count;
258 section->coded_error /= section->count;
259 section->sr_coded_error /= section->count;
260 section->pcnt_inter /= section->count;
261 section->pcnt_second_ref /= section->count;
262 section->pcnt_neutral /= section->count;
263 section->pcnt_motion /= section->count;
264 section->MVr /= section->count;
265 section->mvr_abs /= section->count;
266 section->MVc /= section->count;
267 section->mvc_abs /= section->count;
268 section->MVrv /= section->count;
269 section->MVcv /= section->count;
270 section->mv_in_out_count /= section->count;
271 section->duration /= section->count;
274 // Calculate a modified Error used in distributing bits between easier and
276 static double calculate_modified_err(const TWO_PASS *twopass,
277 const VP9EncoderConfig *oxcf,
278 const FIRSTPASS_STATS *this_frame) {
279 const FIRSTPASS_STATS *const stats = &twopass->total_stats;
280 const double av_err = stats->coded_error / stats->count;
281 const double modified_error = av_err *
282 pow(this_frame->coded_error / DOUBLE_DIVIDE_CHECK(av_err),
283 oxcf->two_pass_vbrbias / 100.0);
284 return fclamp(modified_error,
285 twopass->modified_error_min, twopass->modified_error_max);
288 // This function returns the maximum target rate per frame.
289 static int frame_max_bits(const RATE_CONTROL *rc,
290 const VP9EncoderConfig *oxcf) {
291 int64_t max_bits = ((int64_t)rc->avg_frame_bandwidth *
292 (int64_t)oxcf->two_pass_vbrmax_section) / 100;
295 else if (max_bits > rc->max_frame_bandwidth)
296 max_bits = rc->max_frame_bandwidth;
298 return (int)max_bits;
301 void vp9_init_first_pass(VP9_COMP *cpi) {
302 zero_stats(&cpi->twopass.total_stats);
305 void vp9_end_first_pass(VP9_COMP *cpi) {
306 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
308 for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
309 output_stats(&cpi->svc.layer_context[i].twopass.total_stats,
310 cpi->output_pkt_list);
313 output_stats(&cpi->twopass.total_stats, cpi->output_pkt_list);
317 static vp9_variance_fn_t get_block_variance_fn(BLOCK_SIZE bsize) {
330 static unsigned int get_prediction_error(BLOCK_SIZE bsize,
331 const struct buf_2d *src,
332 const struct buf_2d *ref) {
334 const vp9_variance_fn_t fn = get_block_variance_fn(bsize);
335 fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
339 // Refine the motion search range according to the frame dimension
340 // for first pass test.
341 static int get_search_range(const VP9_COMMON *cm) {
343 const int dim = MIN(cm->width, cm->height);
345 while ((dim << sr) < MAX_FULL_PEL_VAL)
350 static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
351 const MV *ref_mv, MV *best_mv,
352 int *best_motion_err) {
353 MACROBLOCKD *const xd = &x->e_mbd;
355 MV ref_mv_full = {ref_mv->row >> 3, ref_mv->col >> 3};
356 int num00, tmp_err, n;
357 const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
358 vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
359 const int new_mv_mode_penalty = 256;
362 int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
363 const int sr = get_search_range(&cpi->common);
367 // Override the default variance function to use MSE.
368 v_fn_ptr.vf = get_block_variance_fn(bsize);
370 // Center the initial step/diamond search on best mv.
371 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
373 x->sadperbit16, &num00, &v_fn_ptr, ref_mv);
374 if (tmp_err < INT_MAX)
375 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
376 if (tmp_err < INT_MAX - new_mv_mode_penalty)
377 tmp_err += new_mv_mode_penalty;
379 if (tmp_err < *best_motion_err) {
380 *best_motion_err = tmp_err;
384 // Carry out further step/diamond searches as necessary.
388 while (n < further_steps) {
394 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
395 step_param + n, x->sadperbit16,
396 &num00, &v_fn_ptr, ref_mv);
397 if (tmp_err < INT_MAX)
398 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
399 if (tmp_err < INT_MAX - new_mv_mode_penalty)
400 tmp_err += new_mv_mode_penalty;
402 if (tmp_err < *best_motion_err) {
403 *best_motion_err = tmp_err;
410 static BLOCK_SIZE get_bsize(const VP9_COMMON *cm, int mb_row, int mb_col) {
411 if (2 * mb_col + 1 < cm->mi_cols) {
412 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_16X16
415 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_8X16
420 static int find_fp_qindex() {
423 for (i = 0; i < QINDEX_RANGE; ++i)
424 if (vp9_convert_qindex_to_q(i) >= 30.0)
427 if (i == QINDEX_RANGE)
433 static void set_first_pass_params(VP9_COMP *cpi) {
434 VP9_COMMON *const cm = &cpi->common;
435 if (!cpi->refresh_alt_ref_frame &&
436 (cm->current_video_frame == 0 ||
437 (cpi->frame_flags & FRAMEFLAGS_KEY))) {
438 cm->frame_type = KEY_FRAME;
440 cm->frame_type = INTER_FRAME;
442 // Do not use periodic key frames.
443 cpi->rc.frames_to_key = INT_MAX;
446 void vp9_first_pass(VP9_COMP *cpi) {
448 MACROBLOCK *const x = &cpi->mb;
449 VP9_COMMON *const cm = &cpi->common;
450 MACROBLOCKD *const xd = &x->e_mbd;
452 struct macroblock_plane *const p = x->plane;
453 struct macroblockd_plane *const pd = xd->plane;
454 const PICK_MODE_CONTEXT *ctx = &cpi->pc_root->none;
457 int recon_yoffset, recon_uvoffset;
458 YV12_BUFFER_CONFIG *const lst_yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
459 YV12_BUFFER_CONFIG *gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
460 YV12_BUFFER_CONFIG *const new_yv12 = get_frame_new_buffer(cm);
461 int recon_y_stride = lst_yv12->y_stride;
462 int recon_uv_stride = lst_yv12->uv_stride;
463 int uv_mb_height = 16 >> (lst_yv12->y_height > lst_yv12->uv_height);
464 int64_t intra_error = 0;
465 int64_t coded_error = 0;
466 int64_t sr_coded_error = 0;
468 int sum_mvr = 0, sum_mvc = 0;
469 int sum_mvr_abs = 0, sum_mvc_abs = 0;
470 int64_t sum_mvrs = 0, sum_mvcs = 0;
473 int second_ref_count = 0;
474 int intrapenalty = 256;
475 int neutral_count = 0;
476 int new_mv_count = 0;
477 int sum_in_vectors = 0;
478 uint32_t lastmv_as_int = 0;
479 TWO_PASS *twopass = &cpi->twopass;
480 const MV zero_mv = {0, 0};
481 const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;
483 #if CONFIG_FP_MB_STATS
484 FIRSTPASS_FRAME_MB_STATS *this_frame_mb_stats = &twopass->this_frame_mb_stats;
487 vp9_clear_system_state();
489 set_first_pass_params(cpi);
490 vp9_set_quantizer(cm, find_fp_qindex());
492 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
493 MV_REFERENCE_FRAME ref_frame = LAST_FRAME;
494 const YV12_BUFFER_CONFIG *scaled_ref_buf = NULL;
495 twopass = &cpi->svc.layer_context[cpi->svc.spatial_layer_id].twopass;
497 vp9_scale_references(cpi);
499 // Use either last frame or alt frame for motion search.
500 if (cpi->ref_frame_flags & VP9_LAST_FLAG) {
501 scaled_ref_buf = vp9_get_scaled_ref_frame(cpi, LAST_FRAME);
502 ref_frame = LAST_FRAME;
503 } else if (cpi->ref_frame_flags & VP9_ALT_FLAG) {
504 scaled_ref_buf = vp9_get_scaled_ref_frame(cpi, ALTREF_FRAME);
505 ref_frame = ALTREF_FRAME;
508 if (scaled_ref_buf != NULL) {
509 // Update the stride since we are using scaled reference buffer
510 first_ref_buf = scaled_ref_buf;
511 recon_y_stride = first_ref_buf->y_stride;
512 recon_uv_stride = first_ref_buf->uv_stride;
513 uv_mb_height = 16 >> (first_ref_buf->y_height > first_ref_buf->uv_height);
516 // Disable golden frame for svc first pass for now.
518 set_ref_ptrs(cm, xd, ref_frame, NONE);
520 cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
521 &cpi->scaled_source);
524 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
526 vp9_setup_src_planes(x, cpi->Source, 0, 0);
527 vp9_setup_pre_planes(xd, 0, first_ref_buf, 0, 0, NULL);
528 vp9_setup_dst_planes(xd->plane, new_yv12, 0, 0);
530 xd->mi = cm->mi_grid_visible;
533 vp9_frame_init_quantizer(cpi);
535 for (i = 0; i < MAX_MB_PLANE; ++i) {
536 p[i].coeff = ctx->coeff_pbuf[i][1];
537 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
538 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
539 p[i].eobs = ctx->eobs_pbuf[i][1];
543 vp9_init_mv_probs(cm);
544 vp9_initialize_rd_consts(cpi);
546 // Tiling is ignored in the first pass.
547 vp9_tile_init(&tile, cm, 0, 0);
549 for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
552 best_ref_mv.as_int = 0;
554 // Reset above block coeffs.
555 xd->up_available = (mb_row != 0);
556 recon_yoffset = (mb_row * recon_y_stride * 16);
557 recon_uvoffset = (mb_row * recon_uv_stride * uv_mb_height);
559 // Set up limit values for motion vectors to prevent them extending
560 // outside the UMV borders.
561 x->mv_row_min = -((mb_row * 16) + BORDER_MV_PIXELS_B16);
562 x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16)
563 + BORDER_MV_PIXELS_B16;
565 for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
567 const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
568 double error_weight = 1.0;
569 const BLOCK_SIZE bsize = get_bsize(cm, mb_row, mb_col);
571 vp9_clear_system_state();
573 xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
574 xd->plane[1].dst.buf = new_yv12->u_buffer + recon_uvoffset;
575 xd->plane[2].dst.buf = new_yv12->v_buffer + recon_uvoffset;
576 xd->left_available = (mb_col != 0);
577 xd->mi[0]->mbmi.sb_type = bsize;
578 xd->mi[0]->mbmi.ref_frame[0] = INTRA_FRAME;
579 set_mi_row_col(xd, &tile,
580 mb_row << 1, num_8x8_blocks_high_lookup[bsize],
581 mb_col << 1, num_8x8_blocks_wide_lookup[bsize],
582 cm->mi_rows, cm->mi_cols);
584 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
585 const int energy = vp9_block_energy(cpi, x, bsize);
586 error_weight = vp9_vaq_inv_q_ratio(energy);
589 // Do intra 16x16 prediction.
591 xd->mi[0]->mbmi.mode = DC_PRED;
592 xd->mi[0]->mbmi.tx_size = use_dc_pred ?
593 (bsize >= BLOCK_16X16 ? TX_16X16 : TX_8X8) : TX_4X4;
594 vp9_encode_intra_block_plane(x, bsize, 0);
595 this_error = vp9_get_mb_ss(x->plane[0].src_diff);
597 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
598 vp9_clear_system_state();
599 this_error = (int)(this_error * error_weight);
602 // Intrapenalty below deals with situations where the intra and inter
603 // error scores are very low (e.g. a plain black frame).
604 // We do not have special cases in first pass for 0,0 and nearest etc so
605 // all inter modes carry an overhead cost estimate for the mv.
606 // When the error score is very low this causes us to pick all or lots of
607 // INTRA modes and throw lots of key frames.
608 // This penalty adds a cost matching that of a 0,0 mv to the intra case.
609 this_error += intrapenalty;
611 // Accumulate the intra error.
612 intra_error += (int64_t)this_error;
614 #if CONFIG_FP_MB_STATS
615 if (cpi->use_fp_mb_stats) {
616 this_frame_mb_stats->mb_stats[mb_row * cm->mb_cols + mb_col].mode =
618 this_frame_mb_stats->mb_stats[mb_row * cm->mb_cols + mb_col].err =
620 this_frame_mb_stats->mb_stats[mb_row * cm->mb_cols + mb_col].mv.as_int
625 // Set up limit values for motion vectors to prevent them extending
626 // outside the UMV borders.
627 x->mv_col_min = -((mb_col * 16) + BORDER_MV_PIXELS_B16);
628 x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + BORDER_MV_PIXELS_B16;
630 // Other than for the first frame do a motion search.
631 if (cm->current_video_frame > 0) {
632 int tmp_err, motion_error, raw_motion_error;
634 struct buf_2d unscaled_last_source_buf_2d;
636 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
637 motion_error = get_prediction_error(bsize, &x->plane[0].src,
638 &xd->plane[0].pre[0]);
639 // Assume 0,0 motion with no mv overhead.
640 mv.as_int = tmp_mv.as_int = 0;
642 // Compute the motion error of the 0,0 motion using the last source
643 // frame as the reference. Skip the further motion search on
644 // reconstructed frame if this error is small.
645 unscaled_last_source_buf_2d.buf =
646 cpi->unscaled_last_source->y_buffer + recon_yoffset;
647 unscaled_last_source_buf_2d.stride =
648 cpi->unscaled_last_source->y_stride;
649 raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
650 &unscaled_last_source_buf_2d);
652 // TODO(pengchong): Replace the hard-coded threshold
653 if (raw_motion_error > 25 ||
654 (cpi->use_svc && cpi->svc.number_temporal_layers == 1)) {
655 // Test last reference frame using the previous best mv as the
656 // starting point (best reference) for the search.
657 first_pass_motion_search(cpi, x, &best_ref_mv.as_mv, &mv.as_mv,
659 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
660 vp9_clear_system_state();
661 motion_error = (int)(motion_error * error_weight);
664 // If the current best reference mv is not centered on 0,0 then do a
665 // 0,0 based search as well.
666 if (best_ref_mv.as_int) {
668 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv, &tmp_err);
669 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
670 vp9_clear_system_state();
671 tmp_err = (int)(tmp_err * error_weight);
674 if (tmp_err < motion_error) {
675 motion_error = tmp_err;
676 mv.as_int = tmp_mv.as_int;
680 // Search in an older reference frame.
681 if (cm->current_video_frame > 1 && gld_yv12 != NULL) {
682 // Assume 0,0 motion with no mv overhead.
685 xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
686 gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
687 &xd->plane[0].pre[0]);
689 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv,
691 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
692 vp9_clear_system_state();
693 gf_motion_error = (int)(gf_motion_error * error_weight);
696 if (gf_motion_error < motion_error && gf_motion_error < this_error)
699 // Reset to last frame as reference buffer.
700 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
701 xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
702 xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
704 // In accumulating a score for the older reference frame take the
705 // best of the motion predicted score and the intra coded error
706 // (just as will be done for) accumulation of "coded_error" for
708 if (gf_motion_error < this_error)
709 sr_coded_error += gf_motion_error;
711 sr_coded_error += this_error;
713 sr_coded_error += motion_error;
716 sr_coded_error += motion_error;
719 // Start by assuming that intra mode is best.
720 best_ref_mv.as_int = 0;
722 if (motion_error <= this_error) {
723 // Keep a count of cases where the inter and intra were very close
724 // and very low. This helps with scene cut detection for example in
725 // cropped clips with black bars at the sides or top and bottom.
726 if (((this_error - intrapenalty) * 9 <= motion_error * 10) &&
727 this_error < 2 * intrapenalty)
732 this_error = motion_error;
733 xd->mi[0]->mbmi.mode = NEWMV;
734 xd->mi[0]->mbmi.mv[0] = mv;
735 xd->mi[0]->mbmi.tx_size = TX_4X4;
736 xd->mi[0]->mbmi.ref_frame[0] = LAST_FRAME;
737 xd->mi[0]->mbmi.ref_frame[1] = NONE;
738 vp9_build_inter_predictors_sby(xd, mb_row << 1, mb_col << 1, bsize);
739 vp9_encode_sby_pass1(x, bsize);
740 sum_mvr += mv.as_mv.row;
741 sum_mvr_abs += abs(mv.as_mv.row);
742 sum_mvc += mv.as_mv.col;
743 sum_mvc_abs += abs(mv.as_mv.col);
744 sum_mvrs += mv.as_mv.row * mv.as_mv.row;
745 sum_mvcs += mv.as_mv.col * mv.as_mv.col;
748 best_ref_mv.as_int = mv.as_int;
750 #if CONFIG_FP_MB_STATS
751 if (cpi->use_fp_mb_stats) {
752 this_frame_mb_stats->mb_stats[mb_row * cm->mb_cols + mb_col].mode =
754 this_frame_mb_stats->mb_stats[mb_row * cm->mb_cols + mb_col].err =
756 this_frame_mb_stats->mb_stats[mb_row * cm->mb_cols + mb_col].mv.
764 // Non-zero vector, was it different from the last non zero vector?
765 if (mv.as_int != lastmv_as_int)
767 lastmv_as_int = mv.as_int;
769 // Does the row vector point inwards or outwards?
770 if (mb_row < cm->mb_rows / 2) {
771 if (mv.as_mv.row > 0)
773 else if (mv.as_mv.row < 0)
775 } else if (mb_row > cm->mb_rows / 2) {
776 if (mv.as_mv.row > 0)
778 else if (mv.as_mv.row < 0)
782 // Does the col vector point inwards or outwards?
783 if (mb_col < cm->mb_cols / 2) {
784 if (mv.as_mv.col > 0)
786 else if (mv.as_mv.col < 0)
788 } else if (mb_col > cm->mb_cols / 2) {
789 if (mv.as_mv.col > 0)
791 else if (mv.as_mv.col < 0)
797 sr_coded_error += (int64_t)this_error;
799 coded_error += (int64_t)this_error;
801 // Adjust to the next column of MBs.
802 x->plane[0].src.buf += 16;
803 x->plane[1].src.buf += uv_mb_height;
804 x->plane[2].src.buf += uv_mb_height;
807 recon_uvoffset += uv_mb_height;
810 // Adjust to the next row of MBs.
811 x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
812 x->plane[1].src.buf += uv_mb_height * x->plane[1].src.stride -
813 uv_mb_height * cm->mb_cols;
814 x->plane[2].src.buf += uv_mb_height * x->plane[1].src.stride -
815 uv_mb_height * cm->mb_cols;
817 vp9_clear_system_state();
820 vp9_clear_system_state();
824 fps.frame = cm->current_video_frame;
825 fps.spatial_layer_id = cpi->svc.spatial_layer_id;
826 fps.intra_error = (double)(intra_error >> 8);
827 fps.coded_error = (double)(coded_error >> 8);
828 fps.sr_coded_error = (double)(sr_coded_error >> 8);
830 fps.pcnt_inter = (double)intercount / cm->MBs;
831 fps.pcnt_second_ref = (double)second_ref_count / cm->MBs;
832 fps.pcnt_neutral = (double)neutral_count / cm->MBs;
835 fps.MVr = (double)sum_mvr / mvcount;
836 fps.mvr_abs = (double)sum_mvr_abs / mvcount;
837 fps.MVc = (double)sum_mvc / mvcount;
838 fps.mvc_abs = (double)sum_mvc_abs / mvcount;
839 fps.MVrv = ((double)sum_mvrs - (fps.MVr * fps.MVr / mvcount)) / mvcount;
840 fps.MVcv = ((double)sum_mvcs - (fps.MVc * fps.MVc / mvcount)) / mvcount;
841 fps.mv_in_out_count = (double)sum_in_vectors / (mvcount * 2);
842 fps.new_mv_count = new_mv_count;
843 fps.pcnt_motion = (double)mvcount / cm->MBs;
851 fps.mv_in_out_count = 0.0;
852 fps.new_mv_count = 0.0;
853 fps.pcnt_motion = 0.0;
856 // TODO(paulwilkins): Handle the case when duration is set to 0, or
857 // something less than the full time between subsequent values of
858 // cpi->source_time_stamp.
859 fps.duration = (double)(cpi->source->ts_end - cpi->source->ts_start);
861 // Don't want to do output stats with a stack variable!
862 twopass->this_frame_stats = fps;
863 output_stats(&twopass->this_frame_stats, cpi->output_pkt_list);
864 accumulate_stats(&twopass->total_stats, &fps);
866 #if CONFIG_FP_MB_STATS
867 if (cpi->use_fp_mb_stats) {
868 output_mb_stats(this_frame_mb_stats, cm);
873 // Copy the previous Last Frame back into gf and and arf buffers if
874 // the prediction is good enough... but also don't allow it to lag too far.
875 if ((twopass->sr_update_lag > 3) ||
876 ((cm->current_video_frame > 0) &&
877 (twopass->this_frame_stats.pcnt_inter > 0.20) &&
878 ((twopass->this_frame_stats.intra_error /
879 DOUBLE_DIVIDE_CHECK(twopass->this_frame_stats.coded_error)) > 2.0))) {
880 if (gld_yv12 != NULL) {
881 vp8_yv12_copy_frame(lst_yv12, gld_yv12);
883 twopass->sr_update_lag = 1;
885 ++twopass->sr_update_lag;
888 vp9_extend_frame_borders(new_yv12);
890 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
891 vp9_update_reference_frames(cpi);
893 // Swap frame pointers so last frame refers to the frame we just compressed.
894 swap_yv12(lst_yv12, new_yv12);
897 // Special case for the first frame. Copy into the GF buffer as a second
899 if (cm->current_video_frame == 0 && gld_yv12 != NULL) {
900 vp8_yv12_copy_frame(lst_yv12, gld_yv12);
903 // Use this to see what the first pass reconstruction looks like.
907 snprintf(filename, sizeof(filename), "enc%04d.yuv",
908 (int)cm->current_video_frame);
910 if (cm->current_video_frame == 0)
911 recon_file = fopen(filename, "wb");
913 recon_file = fopen(filename, "ab");
915 (void)fwrite(lst_yv12->buffer_alloc, lst_yv12->frame_size, 1, recon_file);
919 ++cm->current_video_frame;
922 static double calc_correction_factor(double err_per_mb,
927 const double error_term = err_per_mb / err_divisor;
929 // Adjustment based on actual quantizer to power term.
930 const double power_term = MIN(vp9_convert_qindex_to_q(q) * 0.0125 + pt_low,
933 // Calculate correction factor.
934 if (power_term < 1.0)
935 assert(error_term >= 0.0);
937 return fclamp(pow(error_term, power_term), 0.05, 5.0);
940 static int get_twopass_worst_quality(const VP9_COMP *cpi,
941 const FIRSTPASS_STATS *stats,
942 int section_target_bandwidth) {
943 const RATE_CONTROL *const rc = &cpi->rc;
944 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
946 if (section_target_bandwidth <= 0) {
947 return rc->worst_quality; // Highest value allowed
949 const int num_mbs = cpi->common.MBs;
950 const double section_err = stats->coded_error / stats->count;
951 const double err_per_mb = section_err / num_mbs;
952 const double speed_term = 1.0 + 0.04 * oxcf->speed;
953 const int target_norm_bits_per_mb = ((uint64_t)section_target_bandwidth <<
954 BPER_MB_NORMBITS) / num_mbs;
956 int is_svc_upper_layer = 0;
957 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1 &&
958 cpi->svc.spatial_layer_id > 0) {
959 is_svc_upper_layer = 1;
962 // Try and pick a max Q that will be high enough to encode the
963 // content at the given rate.
964 for (q = rc->best_quality; q < rc->worst_quality; ++q) {
965 const double factor =
966 calc_correction_factor(err_per_mb, ERR_DIVISOR,
967 is_svc_upper_layer ? SVC_FACTOR_PT_LOW :
968 FACTOR_PT_LOW, FACTOR_PT_HIGH, q);
969 const int bits_per_mb = vp9_rc_bits_per_mb(INTER_FRAME, q,
970 factor * speed_term);
971 if (bits_per_mb <= target_norm_bits_per_mb)
975 // Restriction on active max q for constrained quality mode.
976 if (cpi->oxcf.rc_mode == VPX_CQ)
977 q = MAX(q, oxcf->cq_level);
982 extern void vp9_new_framerate(VP9_COMP *cpi, double framerate);
984 void vp9_init_second_pass(VP9_COMP *cpi) {
985 SVC *const svc = &cpi->svc;
986 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
987 const int is_spatial_svc = (svc->number_spatial_layers > 1) &&
988 (svc->number_temporal_layers == 1);
989 TWO_PASS *const twopass = is_spatial_svc ?
990 &svc->layer_context[svc->spatial_layer_id].twopass : &cpi->twopass;
992 FIRSTPASS_STATS *stats;
994 zero_stats(&twopass->total_stats);
995 zero_stats(&twopass->total_left_stats);
997 if (!twopass->stats_in_end)
1000 stats = &twopass->total_stats;
1002 *stats = *twopass->stats_in_end;
1003 twopass->total_left_stats = *stats;
1005 frame_rate = 10000000.0 * stats->count / stats->duration;
1006 // Each frame can have a different duration, as the frame rate in the source
1007 // isn't guaranteed to be constant. The frame rate prior to the first frame
1008 // encoded in the second pass is a guess. However, the sum duration is not.
1009 // It is calculated based on the actual durations of all frames from the
1012 if (is_spatial_svc) {
1013 vp9_update_spatial_layer_framerate(cpi, frame_rate);
1014 twopass->bits_left = (int64_t)(stats->duration *
1015 svc->layer_context[svc->spatial_layer_id].target_bandwidth /
1018 vp9_new_framerate(cpi, frame_rate);
1019 twopass->bits_left = (int64_t)(stats->duration * oxcf->target_bandwidth /
1023 // Calculate a minimum intra value to be used in determining the IIratio
1024 // scores used in the second pass. We have this minimum to make sure
1025 // that clips that are static but "low complexity" in the intra domain
1026 // are still boosted appropriately for KF/GF/ARF.
1027 if (!is_spatial_svc) {
1028 // We don't know the number of MBs for each layer at this point.
1029 // So we will do it later.
1030 twopass->kf_intra_err_min = KF_MB_INTRA_MIN * cpi->common.MBs;
1031 twopass->gf_intra_err_min = GF_MB_INTRA_MIN * cpi->common.MBs;
1034 // This variable monitors how far behind the second ref update is lagging.
1035 twopass->sr_update_lag = 1;
1037 // Scan the first pass file and calculate a modified total error based upon
1038 // the bias/power function used to allocate bits.
1040 const double avg_error = stats->coded_error /
1041 DOUBLE_DIVIDE_CHECK(stats->count);
1042 const FIRSTPASS_STATS *s = twopass->stats_in;
1043 double modified_error_total = 0.0;
1044 twopass->modified_error_min = (avg_error *
1045 oxcf->two_pass_vbrmin_section) / 100;
1046 twopass->modified_error_max = (avg_error *
1047 oxcf->two_pass_vbrmax_section) / 100;
1048 while (s < twopass->stats_in_end) {
1049 modified_error_total += calculate_modified_err(twopass, oxcf, s);
1052 twopass->modified_error_left = modified_error_total;
1055 // Reset the vbr bits off target counter
1056 cpi->rc.vbr_bits_off_target = 0;
1059 // This function gives an estimate of how badly we believe the prediction
1060 // quality is decaying from frame to frame.
1061 static double get_prediction_decay_rate(const VP9_COMMON *cm,
1062 const FIRSTPASS_STATS *next_frame) {
1063 // Look at the observed drop in prediction quality between the last frame
1064 // and the GF buffer (which contains an older frame).
1065 const double mb_sr_err_diff = (next_frame->sr_coded_error -
1066 next_frame->coded_error) / cm->MBs;
1067 const double second_ref_decay = mb_sr_err_diff <= 512.0
1068 ? fclamp(pow(1.0 - (mb_sr_err_diff / 512.0), 0.5), 0.85, 1.0)
1071 return MIN(second_ref_decay, next_frame->pcnt_inter);
1074 // Function to test for a condition where a complex transition is followed
1075 // by a static section. For example in slide shows where there is a fade
1076 // between slides. This is to help with more optimal kf and gf positioning.
1077 static int detect_transition_to_still(TWO_PASS *twopass,
1078 int frame_interval, int still_interval,
1079 double loop_decay_rate,
1080 double last_decay_rate) {
1081 int trans_to_still = 0;
1083 // Break clause to detect very still sections after motion
1084 // For example a static image after a fade or other transition
1085 // instead of a clean scene cut.
1086 if (frame_interval > MIN_GF_INTERVAL &&
1087 loop_decay_rate >= 0.999 &&
1088 last_decay_rate < 0.9) {
1090 const FIRSTPASS_STATS *position = twopass->stats_in;
1091 FIRSTPASS_STATS tmp_next_frame;
1093 // Look ahead a few frames to see if static condition persists...
1094 for (j = 0; j < still_interval; ++j) {
1095 if (EOF == input_stats(twopass, &tmp_next_frame))
1098 if (tmp_next_frame.pcnt_inter - tmp_next_frame.pcnt_motion < 0.999)
1102 reset_fpf_position(twopass, position);
1104 // Only if it does do we signal a transition to still.
1105 if (j == still_interval)
1109 return trans_to_still;
1112 // This function detects a flash through the high relative pcnt_second_ref
1113 // score in the frame following a flash frame. The offset passed in should
1115 static int detect_flash(const TWO_PASS *twopass, int offset) {
1116 FIRSTPASS_STATS next_frame;
1118 int flash_detected = 0;
1120 // Read the frame data.
1121 // The return is FALSE (no flash detected) if not a valid frame
1122 if (read_frame_stats(twopass, &next_frame, offset) != EOF) {
1123 // What we are looking for here is a situation where there is a
1124 // brief break in prediction (such as a flash) but subsequent frames
1125 // are reasonably well predicted by an earlier (pre flash) frame.
1126 // The recovery after a flash is indicated by a high pcnt_second_ref
1127 // compared to pcnt_inter.
1128 if (next_frame.pcnt_second_ref > next_frame.pcnt_inter &&
1129 next_frame.pcnt_second_ref >= 0.5)
1133 return flash_detected;
1136 // Update the motion related elements to the GF arf boost calculation.
1137 static void accumulate_frame_motion_stats(const FIRSTPASS_STATS *stats,
1139 double *mv_in_out_accumulator,
1140 double *abs_mv_in_out_accumulator,
1141 double *mv_ratio_accumulator) {
1142 const double pct = stats->pcnt_motion;
1144 // Accumulate Motion In/Out of frame stats.
1145 *mv_in_out = stats->mv_in_out_count * pct;
1146 *mv_in_out_accumulator += *mv_in_out;
1147 *abs_mv_in_out_accumulator += fabs(*mv_in_out);
1149 // Accumulate a measure of how uniform (or conversely how random) the motion
1150 // field is (a ratio of abs(mv) / mv).
1152 const double mvr_ratio = fabs(stats->mvr_abs) /
1153 DOUBLE_DIVIDE_CHECK(fabs(stats->MVr));
1154 const double mvc_ratio = fabs(stats->mvc_abs) /
1155 DOUBLE_DIVIDE_CHECK(fabs(stats->MVc));
1157 *mv_ratio_accumulator += pct * (mvr_ratio < stats->mvr_abs ?
1158 mvr_ratio : stats->mvr_abs);
1159 *mv_ratio_accumulator += pct * (mvc_ratio < stats->mvc_abs ?
1160 mvc_ratio : stats->mvc_abs);
1164 // Calculate a baseline boost number for the current frame.
1165 static double calc_frame_boost(const TWO_PASS *twopass,
1166 const FIRSTPASS_STATS *this_frame,
1167 double this_frame_mv_in_out) {
1170 // Underlying boost factor is based on inter intra error ratio.
1171 if (this_frame->intra_error > twopass->gf_intra_err_min)
1172 frame_boost = (IIFACTOR * this_frame->intra_error /
1173 DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
1175 frame_boost = (IIFACTOR * twopass->gf_intra_err_min /
1176 DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
1178 // Increase boost for frames where new data coming into frame (e.g. zoom out).
1179 // Slightly reduce boost if there is a net balance of motion out of the frame
1180 // (zoom in). The range for this_frame_mv_in_out is -1.0 to +1.0.
1181 if (this_frame_mv_in_out > 0.0)
1182 frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
1183 // In the extreme case the boost is halved.
1185 frame_boost += frame_boost * (this_frame_mv_in_out / 2.0);
1187 return MIN(frame_boost, GF_RMAX);
1190 static int calc_arf_boost(VP9_COMP *cpi, int offset,
1191 int f_frames, int b_frames,
1192 int *f_boost, int *b_boost) {
1193 FIRSTPASS_STATS this_frame;
1194 TWO_PASS *const twopass = &cpi->twopass;
1196 double boost_score = 0.0;
1197 double mv_ratio_accumulator = 0.0;
1198 double decay_accumulator = 1.0;
1199 double this_frame_mv_in_out = 0.0;
1200 double mv_in_out_accumulator = 0.0;
1201 double abs_mv_in_out_accumulator = 0.0;
1203 int flash_detected = 0;
1205 // Search forward from the proposed arf/next gf position.
1206 for (i = 0; i < f_frames; ++i) {
1207 if (read_frame_stats(twopass, &this_frame, (i + offset)) == EOF)
1210 // Update the motion related elements to the boost calculation.
1211 accumulate_frame_motion_stats(&this_frame,
1212 &this_frame_mv_in_out, &mv_in_out_accumulator,
1213 &abs_mv_in_out_accumulator,
1214 &mv_ratio_accumulator);
1216 // We want to discount the flash frame itself and the recovery
1217 // frame that follows as both will have poor scores.
1218 flash_detected = detect_flash(twopass, i + offset) ||
1219 detect_flash(twopass, i + offset + 1);
1221 // Accumulate the effect of prediction quality decay.
1222 if (!flash_detected) {
1223 decay_accumulator *= get_prediction_decay_rate(&cpi->common, &this_frame);
1224 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1225 ? MIN_DECAY_FACTOR : decay_accumulator;
1228 boost_score += decay_accumulator * calc_frame_boost(twopass, &this_frame,
1229 this_frame_mv_in_out);
1232 *f_boost = (int)boost_score;
1234 // Reset for backward looking loop.
1236 mv_ratio_accumulator = 0.0;
1237 decay_accumulator = 1.0;
1238 this_frame_mv_in_out = 0.0;
1239 mv_in_out_accumulator = 0.0;
1240 abs_mv_in_out_accumulator = 0.0;
1242 // Search backward towards last gf position.
1243 for (i = -1; i >= -b_frames; --i) {
1244 if (read_frame_stats(twopass, &this_frame, (i + offset)) == EOF)
1247 // Update the motion related elements to the boost calculation.
1248 accumulate_frame_motion_stats(&this_frame,
1249 &this_frame_mv_in_out, &mv_in_out_accumulator,
1250 &abs_mv_in_out_accumulator,
1251 &mv_ratio_accumulator);
1253 // We want to discount the the flash frame itself and the recovery
1254 // frame that follows as both will have poor scores.
1255 flash_detected = detect_flash(twopass, i + offset) ||
1256 detect_flash(twopass, i + offset + 1);
1258 // Cumulative effect of prediction quality decay.
1259 if (!flash_detected) {
1260 decay_accumulator *= get_prediction_decay_rate(&cpi->common, &this_frame);
1261 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1262 ? MIN_DECAY_FACTOR : decay_accumulator;
1265 boost_score += decay_accumulator * calc_frame_boost(twopass, &this_frame,
1266 this_frame_mv_in_out);
1268 *b_boost = (int)boost_score;
1270 arf_boost = (*f_boost + *b_boost);
1271 if (arf_boost < ((b_frames + f_frames) * 20))
1272 arf_boost = ((b_frames + f_frames) * 20);
1277 // Calculate a section intra ratio used in setting max loop filter.
1278 static int calculate_section_intra_ratio(const FIRSTPASS_STATS *begin,
1279 const FIRSTPASS_STATS *end,
1280 int section_length) {
1281 const FIRSTPASS_STATS *s = begin;
1282 double intra_error = 0.0;
1283 double coded_error = 0.0;
1286 while (s < end && i < section_length) {
1287 intra_error += s->intra_error;
1288 coded_error += s->coded_error;
1293 return (int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
1296 // Calculate the total bits to allocate in this GF/ARF group.
1297 static int64_t calculate_total_gf_group_bits(VP9_COMP *cpi,
1298 double gf_group_err) {
1299 const RATE_CONTROL *const rc = &cpi->rc;
1300 const TWO_PASS *const twopass = &cpi->twopass;
1301 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
1302 int64_t total_group_bits;
1304 // Calculate the bits to be allocated to the group as a whole.
1305 if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0)) {
1306 total_group_bits = (int64_t)(twopass->kf_group_bits *
1307 (gf_group_err / twopass->kf_group_error_left));
1309 total_group_bits = 0;
1312 // Clamp odd edge cases.
1313 total_group_bits = (total_group_bits < 0) ?
1314 0 : (total_group_bits > twopass->kf_group_bits) ?
1315 twopass->kf_group_bits : total_group_bits;
1317 // Clip based on user supplied data rate variability limit.
1318 if (total_group_bits > (int64_t)max_bits * rc->baseline_gf_interval)
1319 total_group_bits = (int64_t)max_bits * rc->baseline_gf_interval;
1321 return total_group_bits;
1324 // Calculate the number bits extra to assign to boosted frames in a group.
1325 static int calculate_boost_bits(int frame_count,
1326 int boost, int64_t total_group_bits) {
1327 int allocation_chunks;
1329 // return 0 for invalid inputs (could arise e.g. through rounding errors)
1330 if (!boost || (total_group_bits <= 0) || (frame_count <= 0) )
1333 allocation_chunks = (frame_count * 100) + boost;
1335 // Prevent overflow.
1337 int divisor = boost >> 10;
1339 allocation_chunks /= divisor;
1342 // Calculate the number of extra bits for use in the boosted frame or frames.
1343 return MAX((int)(((int64_t)boost * total_group_bits) / allocation_chunks), 0);
1346 // Current limit on maximum number of active arfs in a GF/ARF group.
1347 #define MAX_ACTIVE_ARFS 2
1350 // This function indirects the choice of buffers for arfs.
1351 // At the moment the values are fixed but this may change as part of
1352 // the integration process with other codec features that swap buffers around.
1353 static void get_arf_buffer_indices(unsigned char *arf_buffer_indices) {
1354 arf_buffer_indices[0] = ARF_SLOT1;
1355 arf_buffer_indices[1] = ARF_SLOT2;
1358 static void allocate_gf_group_bits(VP9_COMP *cpi, int64_t gf_group_bits,
1359 double group_error, int gf_arf_bits) {
1360 RATE_CONTROL *const rc = &cpi->rc;
1361 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1362 TWO_PASS *twopass = &cpi->twopass;
1363 FIRSTPASS_STATS frame_stats;
1365 int frame_index = 1;
1366 int target_frame_size;
1368 const int max_bits = frame_max_bits(&cpi->rc, &cpi->oxcf);
1369 int64_t total_group_bits = gf_group_bits;
1370 double modified_err = 0.0;
1371 double err_fraction;
1372 int mid_boost_bits = 0;
1374 unsigned char arf_buffer_indices[MAX_ACTIVE_ARFS];
1376 key_frame = cpi->common.frame_type == KEY_FRAME ||
1377 vp9_is_upper_layer_key_frame(cpi);
1379 get_arf_buffer_indices(arf_buffer_indices);
1381 // For key frames the frame target rate is already set and it
1382 // is also the golden frame.
1384 if (rc->source_alt_ref_active) {
1385 twopass->gf_group.update_type[0] = OVERLAY_UPDATE;
1386 twopass->gf_group.rf_level[0] = INTER_NORMAL;
1387 twopass->gf_group.bit_allocation[0] = 0;
1388 twopass->gf_group.arf_update_idx[0] = arf_buffer_indices[0];
1389 twopass->gf_group.arf_ref_idx[0] = arf_buffer_indices[0];
1391 twopass->gf_group.update_type[0] = GF_UPDATE;
1392 twopass->gf_group.rf_level[0] = GF_ARF_STD;
1393 twopass->gf_group.bit_allocation[0] = gf_arf_bits;
1394 twopass->gf_group.arf_update_idx[0] = arf_buffer_indices[0];
1395 twopass->gf_group.arf_ref_idx[0] = arf_buffer_indices[0];
1398 // Step over the golden frame / overlay frame
1399 if (EOF == input_stats(twopass, &frame_stats))
1403 // Deduct the boost bits for arf (or gf if it is not a key frame)
1404 // from the group total.
1405 if (rc->source_alt_ref_pending || !key_frame)
1406 total_group_bits -= gf_arf_bits;
1408 // Store the bits to spend on the ARF if there is one.
1409 if (rc->source_alt_ref_pending) {
1410 if (cpi->multi_arf_enabled) {
1411 // A portion of the gf / arf extra bits are set asside for lower level
1412 // boosted frames in the middle of the group.
1413 mid_boost_bits += gf_arf_bits >> 5;
1414 gf_arf_bits -= (gf_arf_bits >> 5);
1417 twopass->gf_group.update_type[frame_index] = ARF_UPDATE;
1418 twopass->gf_group.rf_level[frame_index] = GF_ARF_STD;
1419 twopass->gf_group.bit_allocation[frame_index] = gf_arf_bits;
1420 twopass->gf_group.arf_src_offset[frame_index] =
1421 (unsigned char)(rc->baseline_gf_interval - 1);
1422 twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[0];
1423 twopass->gf_group.arf_ref_idx[frame_index] =
1424 arf_buffer_indices[cpi->multi_arf_enabled && rc->source_alt_ref_active];
1427 if (cpi->multi_arf_enabled) {
1428 // Set aside a slot for a level 1 arf.
1429 twopass->gf_group.update_type[frame_index] = ARF_UPDATE;
1430 twopass->gf_group.rf_level[frame_index] = GF_ARF_LOW;
1431 twopass->gf_group.arf_src_offset[frame_index] =
1432 (unsigned char)((rc->baseline_gf_interval >> 1) - 1);
1433 twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[1];
1434 twopass->gf_group.arf_ref_idx[frame_index] = arf_buffer_indices[0];
1439 // Define middle frame
1440 mid_frame_idx = frame_index + (rc->baseline_gf_interval >> 1) - 1;
1442 // Allocate bits to the other frames in the group.
1443 for (i = 0; i < rc->baseline_gf_interval - 1; ++i) {
1445 if (EOF == input_stats(twopass, &frame_stats))
1448 modified_err = calculate_modified_err(twopass, oxcf, &frame_stats);
1450 if (group_error > 0)
1451 err_fraction = modified_err / DOUBLE_DIVIDE_CHECK(group_error);
1455 target_frame_size = (int)((double)total_group_bits * err_fraction);
1457 if (rc->source_alt_ref_pending && cpi->multi_arf_enabled) {
1458 mid_boost_bits += (target_frame_size >> 4);
1459 target_frame_size -= (target_frame_size >> 4);
1461 if (frame_index <= mid_frame_idx)
1464 twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[arf_idx];
1465 twopass->gf_group.arf_ref_idx[frame_index] = arf_buffer_indices[arf_idx];
1467 target_frame_size = clamp(target_frame_size, 0,
1468 MIN(max_bits, (int)total_group_bits));
1470 twopass->gf_group.update_type[frame_index] = LF_UPDATE;
1471 twopass->gf_group.rf_level[frame_index] = INTER_NORMAL;
1473 twopass->gf_group.bit_allocation[frame_index] = target_frame_size;
1478 // We need to configure the frame at the end of the sequence + 1 that will be
1479 // the start frame for the next group. Otherwise prior to the call to
1480 // vp9_rc_get_second_pass_params() the data will be undefined.
1481 twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[0];
1482 twopass->gf_group.arf_ref_idx[frame_index] = arf_buffer_indices[0];
1484 if (rc->source_alt_ref_pending) {
1485 twopass->gf_group.update_type[frame_index] = OVERLAY_UPDATE;
1486 twopass->gf_group.rf_level[frame_index] = INTER_NORMAL;
1488 // Final setup for second arf and its overlay.
1489 if (cpi->multi_arf_enabled) {
1490 twopass->gf_group.bit_allocation[2] =
1491 twopass->gf_group.bit_allocation[mid_frame_idx] + mid_boost_bits;
1492 twopass->gf_group.update_type[mid_frame_idx] = OVERLAY_UPDATE;
1493 twopass->gf_group.bit_allocation[mid_frame_idx] = 0;
1496 twopass->gf_group.update_type[frame_index] = GF_UPDATE;
1497 twopass->gf_group.rf_level[frame_index] = GF_ARF_STD;
1501 // Analyse and define a gf/arf group.
1502 static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
1503 RATE_CONTROL *const rc = &cpi->rc;
1504 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1505 TWO_PASS *const twopass = &cpi->twopass;
1506 FIRSTPASS_STATS next_frame;
1507 const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
1510 double boost_score = 0.0;
1511 double old_boost_score = 0.0;
1512 double gf_group_err = 0.0;
1513 double gf_first_frame_err = 0.0;
1514 double mod_frame_err = 0.0;
1516 double mv_ratio_accumulator = 0.0;
1517 double decay_accumulator = 1.0;
1518 double zero_motion_accumulator = 1.0;
1520 double loop_decay_rate = 1.00;
1521 double last_loop_decay_rate = 1.00;
1523 double this_frame_mv_in_out = 0.0;
1524 double mv_in_out_accumulator = 0.0;
1525 double abs_mv_in_out_accumulator = 0.0;
1526 double mv_ratio_accumulator_thresh;
1527 unsigned int allow_alt_ref = is_altref_enabled(oxcf);
1532 int active_max_gf_interval;
1533 int64_t gf_group_bits;
1534 double gf_group_error_left;
1537 // Reset the GF group data structures unless this is a key
1538 // frame in which case it will already have been done.
1539 if (cpi->common.frame_type != KEY_FRAME) {
1540 vp9_zero(twopass->gf_group);
1543 vp9_clear_system_state();
1544 vp9_zero(next_frame);
1548 // Load stats for the current frame.
1549 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1551 // Note the error of the frame at the start of the group. This will be
1552 // the GF frame error if we code a normal gf.
1553 gf_first_frame_err = mod_frame_err;
1555 // If this is a key frame or the overlay from a previous arf then
1556 // the error score / cost of this frame has already been accounted for.
1557 if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
1558 gf_group_err -= gf_first_frame_err;
1560 // Motion breakout threshold for loop below depends on image size.
1561 mv_ratio_accumulator_thresh = (cpi->common.width + cpi->common.height) / 10.0;
1563 // Work out a maximum interval for the GF.
1564 // If the image appears completely static we can extend beyond this.
1565 // The value chosen depends on the active Q range. At low Q we have
1566 // bits to spare and are better with a smaller interval and smaller boost.
1567 // At high Q when there are few bits to spare we are better with a longer
1568 // interval to spread the cost of the GF.
1570 active_max_gf_interval =
1571 12 + ((int)vp9_convert_qindex_to_q(rc->last_q[INTER_FRAME]) >> 5);
1573 if (active_max_gf_interval > rc->max_gf_interval)
1574 active_max_gf_interval = rc->max_gf_interval;
1577 while (i < rc->static_scene_max_gf_interval && i < rc->frames_to_key) {
1580 // Accumulate error score of frames in this gf group.
1581 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1582 gf_group_err += mod_frame_err;
1584 if (EOF == input_stats(twopass, &next_frame))
1587 // Test for the case where there is a brief flash but the prediction
1588 // quality back to an earlier frame is then restored.
1589 flash_detected = detect_flash(twopass, 0);
1591 // Update the motion related elements to the boost calculation.
1592 accumulate_frame_motion_stats(&next_frame,
1593 &this_frame_mv_in_out, &mv_in_out_accumulator,
1594 &abs_mv_in_out_accumulator,
1595 &mv_ratio_accumulator);
1597 // Accumulate the effect of prediction quality decay.
1598 if (!flash_detected) {
1599 last_loop_decay_rate = loop_decay_rate;
1600 loop_decay_rate = get_prediction_decay_rate(&cpi->common, &next_frame);
1601 decay_accumulator = decay_accumulator * loop_decay_rate;
1603 // Monitor for static sections.
1604 if ((next_frame.pcnt_inter - next_frame.pcnt_motion) <
1605 zero_motion_accumulator) {
1606 zero_motion_accumulator = next_frame.pcnt_inter -
1607 next_frame.pcnt_motion;
1610 // Break clause to detect very still sections after motion. For example,
1611 // a static image after a fade or other transition.
1612 if (detect_transition_to_still(twopass, i, 5, loop_decay_rate,
1613 last_loop_decay_rate)) {
1619 // Calculate a boost number for this frame.
1620 boost_score += decay_accumulator * calc_frame_boost(twopass, &next_frame,
1621 this_frame_mv_in_out);
1623 // Break out conditions.
1625 // Break at active_max_gf_interval unless almost totally static.
1626 (i >= active_max_gf_interval && (zero_motion_accumulator < 0.995)) ||
1628 // Don't break out with a very short interval.
1629 (i > MIN_GF_INTERVAL) &&
1630 ((boost_score > 125.0) || (next_frame.pcnt_inter < 0.75)) &&
1631 (!flash_detected) &&
1632 ((mv_ratio_accumulator > mv_ratio_accumulator_thresh) ||
1633 (abs_mv_in_out_accumulator > 3.0) ||
1634 (mv_in_out_accumulator < -2.0) ||
1635 ((boost_score - old_boost_score) < IIFACTOR)))) {
1636 boost_score = old_boost_score;
1640 *this_frame = next_frame;
1642 old_boost_score = boost_score;
1645 twopass->gf_zeromotion_pct = (int)(zero_motion_accumulator * 1000.0);
1647 // Don't allow a gf too near the next kf.
1648 if ((rc->frames_to_key - i) < MIN_GF_INTERVAL) {
1649 while (i < (rc->frames_to_key + !rc->next_key_frame_forced)) {
1652 if (EOF == input_stats(twopass, this_frame))
1655 if (i < rc->frames_to_key) {
1656 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1657 gf_group_err += mod_frame_err;
1662 // Set the interval until the next gf.
1663 if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
1664 rc->baseline_gf_interval = i - 1;
1666 rc->baseline_gf_interval = i;
1668 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
1670 // Should we use the alternate reference frame.
1671 if (allow_alt_ref &&
1672 (i < cpi->oxcf.lag_in_frames) &&
1673 (i >= MIN_GF_INTERVAL) &&
1674 // For real scene cuts (not forced kfs) don't allow arf very near kf.
1675 (rc->next_key_frame_forced ||
1676 (i <= (rc->frames_to_key - MIN_GF_INTERVAL)))) {
1677 // Calculate the boost for alt ref.
1678 rc->gfu_boost = calc_arf_boost(cpi, 0, (i - 1), (i - 1), &f_boost,
1680 rc->source_alt_ref_pending = 1;
1683 rc->gfu_boost = (int)boost_score;
1684 rc->source_alt_ref_pending = 0;
1687 // Reset the file position.
1688 reset_fpf_position(twopass, start_pos);
1690 // Calculate the bits to be allocated to the gf/arf group as a whole
1691 gf_group_bits = calculate_total_gf_group_bits(cpi, gf_group_err);
1693 // Calculate the extra bits to be used for boosted frame(s)
1695 int q = rc->last_q[INTER_FRAME];
1696 int boost = (rc->gfu_boost * gfboost_qadjust(q)) / 100;
1698 // Set max and minimum boost and hence minimum allocation.
1699 boost = clamp(boost, 125, (rc->baseline_gf_interval + 1) * 200);
1701 // Calculate the extra bits to be used for boosted frame(s)
1702 gf_arf_bits = calculate_boost_bits(rc->baseline_gf_interval,
1703 boost, gf_group_bits);
1706 // Adjust KF group bits and error remaining.
1707 twopass->kf_group_error_left -= (int64_t)gf_group_err;
1709 // If this is an arf update we want to remove the score for the overlay
1710 // frame at the end which will usually be very cheap to code.
1711 // The overlay frame has already, in effect, been coded so we want to spread
1712 // the remaining bits among the other frames.
1713 // For normal GFs remove the score for the GF itself unless this is
1714 // also a key frame in which case it has already been accounted for.
1715 if (rc->source_alt_ref_pending) {
1716 gf_group_error_left = gf_group_err - mod_frame_err;
1717 } else if (cpi->common.frame_type != KEY_FRAME) {
1718 gf_group_error_left = gf_group_err - gf_first_frame_err;
1720 gf_group_error_left = gf_group_err;
1723 // Allocate bits to each of the frames in the GF group.
1724 allocate_gf_group_bits(cpi, gf_group_bits, gf_group_error_left, gf_arf_bits);
1726 // Reset the file position.
1727 reset_fpf_position(twopass, start_pos);
1729 // Calculate a section intra ratio used in setting max loop filter.
1730 if (cpi->common.frame_type != KEY_FRAME) {
1731 twopass->section_intra_rating =
1732 calculate_section_intra_ratio(start_pos, twopass->stats_in_end,
1733 rc->baseline_gf_interval);
1737 static int test_candidate_kf(TWO_PASS *twopass,
1738 const FIRSTPASS_STATS *last_frame,
1739 const FIRSTPASS_STATS *this_frame,
1740 const FIRSTPASS_STATS *next_frame) {
1741 int is_viable_kf = 0;
1743 // Does the frame satisfy the primary criteria of a key frame?
1744 // If so, then examine how well it predicts subsequent frames.
1745 if ((this_frame->pcnt_second_ref < 0.10) &&
1746 (next_frame->pcnt_second_ref < 0.10) &&
1747 ((this_frame->pcnt_inter < 0.05) ||
1748 (((this_frame->pcnt_inter - this_frame->pcnt_neutral) < 0.35) &&
1749 ((this_frame->intra_error /
1750 DOUBLE_DIVIDE_CHECK(this_frame->coded_error)) < 2.5) &&
1751 ((fabs(last_frame->coded_error - this_frame->coded_error) /
1752 DOUBLE_DIVIDE_CHECK(this_frame->coded_error) > 0.40) ||
1753 (fabs(last_frame->intra_error - this_frame->intra_error) /
1754 DOUBLE_DIVIDE_CHECK(this_frame->intra_error) > 0.40) ||
1755 ((next_frame->intra_error /
1756 DOUBLE_DIVIDE_CHECK(next_frame->coded_error)) > 3.5))))) {
1758 const FIRSTPASS_STATS *start_pos = twopass->stats_in;
1759 FIRSTPASS_STATS local_next_frame = *next_frame;
1760 double boost_score = 0.0;
1761 double old_boost_score = 0.0;
1762 double decay_accumulator = 1.0;
1764 // Examine how well the key frame predicts subsequent frames.
1765 for (i = 0; i < 16; ++i) {
1766 double next_iiratio = (IIKFACTOR1 * local_next_frame.intra_error /
1767 DOUBLE_DIVIDE_CHECK(local_next_frame.coded_error));
1769 if (next_iiratio > RMAX)
1770 next_iiratio = RMAX;
1772 // Cumulative effect of decay in prediction quality.
1773 if (local_next_frame.pcnt_inter > 0.85)
1774 decay_accumulator *= local_next_frame.pcnt_inter;
1776 decay_accumulator *= (0.85 + local_next_frame.pcnt_inter) / 2.0;
1778 // Keep a running total.
1779 boost_score += (decay_accumulator * next_iiratio);
1781 // Test various breakout clauses.
1782 if ((local_next_frame.pcnt_inter < 0.05) ||
1783 (next_iiratio < 1.5) ||
1784 (((local_next_frame.pcnt_inter -
1785 local_next_frame.pcnt_neutral) < 0.20) &&
1786 (next_iiratio < 3.0)) ||
1787 ((boost_score - old_boost_score) < 3.0) ||
1788 (local_next_frame.intra_error < 200)) {
1792 old_boost_score = boost_score;
1794 // Get the next frame details
1795 if (EOF == input_stats(twopass, &local_next_frame))
1799 // If there is tolerable prediction for at least the next 3 frames then
1800 // break out else discard this potential key frame and move on
1801 if (boost_score > 30.0 && (i > 3)) {
1804 // Reset the file position
1805 reset_fpf_position(twopass, start_pos);
1811 return is_viable_kf;
1814 static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
1816 RATE_CONTROL *const rc = &cpi->rc;
1817 TWO_PASS *const twopass = &cpi->twopass;
1818 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1819 const FIRSTPASS_STATS first_frame = *this_frame;
1820 const FIRSTPASS_STATS *const start_position = twopass->stats_in;
1821 FIRSTPASS_STATS next_frame;
1822 FIRSTPASS_STATS last_frame;
1824 double decay_accumulator = 1.0;
1825 double zero_motion_accumulator = 1.0;
1826 double boost_score = 0.0;
1827 double kf_mod_err = 0.0;
1828 double kf_group_err = 0.0;
1829 double recent_loop_decay[8] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
1831 vp9_zero(next_frame);
1833 cpi->common.frame_type = KEY_FRAME;
1835 // Reset the GF group data structures.
1836 vp9_zero(twopass->gf_group);
1838 // Is this a forced key frame by interval.
1839 rc->this_key_frame_forced = rc->next_key_frame_forced;
1841 // Clear the alt ref active flag as this can never be active on a key frame.
1842 rc->source_alt_ref_active = 0;
1844 // KF is always a GF so clear frames till next gf counter.
1845 rc->frames_till_gf_update_due = 0;
1847 rc->frames_to_key = 1;
1849 twopass->kf_group_bits = 0; // Total bits available to kf group
1850 twopass->kf_group_error_left = 0; // Group modified error score.
1852 kf_mod_err = calculate_modified_err(twopass, oxcf, this_frame);
1854 // Find the next keyframe.
1856 while (twopass->stats_in < twopass->stats_in_end &&
1857 rc->frames_to_key < cpi->oxcf.key_freq) {
1858 // Accumulate kf group error.
1859 kf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
1861 // Load the next frame's stats.
1862 last_frame = *this_frame;
1863 input_stats(twopass, this_frame);
1865 // Provided that we are not at the end of the file...
1866 if (cpi->oxcf.auto_key &&
1867 lookup_next_frame_stats(twopass, &next_frame) != EOF) {
1868 double loop_decay_rate;
1870 // Check for a scene cut.
1871 if (test_candidate_kf(twopass, &last_frame, this_frame, &next_frame))
1874 // How fast is the prediction quality decaying?
1875 loop_decay_rate = get_prediction_decay_rate(&cpi->common, &next_frame);
1877 // We want to know something about the recent past... rather than
1878 // as used elsewhere where we are concerned with decay in prediction
1879 // quality since the last GF or KF.
1880 recent_loop_decay[i % 8] = loop_decay_rate;
1881 decay_accumulator = 1.0;
1882 for (j = 0; j < 8; ++j)
1883 decay_accumulator *= recent_loop_decay[j];
1885 // Special check for transition or high motion followed by a
1887 if (detect_transition_to_still(twopass, i, cpi->oxcf.key_freq - i,
1888 loop_decay_rate, decay_accumulator))
1891 // Step on to the next frame.
1892 ++rc->frames_to_key;
1894 // If we don't have a real key frame within the next two
1895 // key_freq intervals then break out of the loop.
1896 if (rc->frames_to_key >= 2 * cpi->oxcf.key_freq)
1899 ++rc->frames_to_key;
1904 // If there is a max kf interval set by the user we must obey it.
1905 // We already breakout of the loop above at 2x max.
1906 // This code centers the extra kf if the actual natural interval
1907 // is between 1x and 2x.
1908 if (cpi->oxcf.auto_key &&
1909 rc->frames_to_key > cpi->oxcf.key_freq) {
1910 FIRSTPASS_STATS tmp_frame = first_frame;
1912 rc->frames_to_key /= 2;
1914 // Reset to the start of the group.
1915 reset_fpf_position(twopass, start_position);
1919 // Rescan to get the correct error data for the forced kf group.
1920 for (i = 0; i < rc->frames_to_key; ++i) {
1921 kf_group_err += calculate_modified_err(twopass, oxcf, &tmp_frame);
1922 input_stats(twopass, &tmp_frame);
1924 rc->next_key_frame_forced = 1;
1925 } else if (twopass->stats_in == twopass->stats_in_end ||
1926 rc->frames_to_key >= cpi->oxcf.key_freq) {
1927 rc->next_key_frame_forced = 1;
1929 rc->next_key_frame_forced = 0;
1932 // Special case for the last key frame of the file.
1933 if (twopass->stats_in >= twopass->stats_in_end) {
1934 // Accumulate kf group error.
1935 kf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
1938 // Calculate the number of bits that should be assigned to the kf group.
1939 if (twopass->bits_left > 0 && twopass->modified_error_left > 0.0) {
1940 // Maximum number of bits for a single normal frame (not key frame).
1941 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
1943 // Maximum number of bits allocated to the key frame group.
1944 int64_t max_grp_bits;
1946 // Default allocation based on bits left and relative
1947 // complexity of the section.
1948 twopass->kf_group_bits = (int64_t)(twopass->bits_left *
1949 (kf_group_err / twopass->modified_error_left));
1951 // Clip based on maximum per frame rate defined by the user.
1952 max_grp_bits = (int64_t)max_bits * (int64_t)rc->frames_to_key;
1953 if (twopass->kf_group_bits > max_grp_bits)
1954 twopass->kf_group_bits = max_grp_bits;
1956 twopass->kf_group_bits = 0;
1958 twopass->kf_group_bits = MAX(0, twopass->kf_group_bits);
1960 // Reset the first pass file position.
1961 reset_fpf_position(twopass, start_position);
1963 // Scan through the kf group collating various stats used to deteermine
1964 // how many bits to spend on it.
1965 decay_accumulator = 1.0;
1967 for (i = 0; i < rc->frames_to_key; ++i) {
1968 if (EOF == input_stats(twopass, &next_frame))
1971 // Monitor for static sections.
1972 if ((next_frame.pcnt_inter - next_frame.pcnt_motion) <
1973 zero_motion_accumulator) {
1974 zero_motion_accumulator = (next_frame.pcnt_inter -
1975 next_frame.pcnt_motion);
1978 // For the first few frames collect data to decide kf boost.
1979 if (i <= (rc->max_gf_interval * 2)) {
1981 if (next_frame.intra_error > twopass->kf_intra_err_min)
1982 r = (IIKFACTOR2 * next_frame.intra_error /
1983 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
1985 r = (IIKFACTOR2 * twopass->kf_intra_err_min /
1986 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
1991 // How fast is prediction quality decaying.
1992 if (!detect_flash(twopass, 0)) {
1993 const double loop_decay_rate = get_prediction_decay_rate(&cpi->common,
1995 decay_accumulator *= loop_decay_rate;
1996 decay_accumulator = MAX(decay_accumulator, MIN_DECAY_FACTOR);
1999 boost_score += (decay_accumulator * r);
2003 reset_fpf_position(twopass, start_position);
2005 // Store the zero motion percentage
2006 twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
2008 // Calculate a section intra ratio used in setting max loop filter.
2009 twopass->section_intra_rating =
2010 calculate_section_intra_ratio(start_position, twopass->stats_in_end,
2013 // Work out how many bits to allocate for the key frame itself.
2014 rc->kf_boost = (int)boost_score;
2016 if (rc->kf_boost < (rc->frames_to_key * 3))
2017 rc->kf_boost = (rc->frames_to_key * 3);
2018 if (rc->kf_boost < MIN_KF_BOOST)
2019 rc->kf_boost = MIN_KF_BOOST;
2021 kf_bits = calculate_boost_bits((rc->frames_to_key - 1),
2022 rc->kf_boost, twopass->kf_group_bits);
2024 twopass->kf_group_bits -= kf_bits;
2026 // Save the bits to spend on the key frame.
2027 twopass->gf_group.bit_allocation[0] = kf_bits;
2028 twopass->gf_group.update_type[0] = KF_UPDATE;
2029 twopass->gf_group.rf_level[0] = KF_STD;
2031 // Note the total error score of the kf group minus the key frame itself.
2032 twopass->kf_group_error_left = (int)(kf_group_err - kf_mod_err);
2034 // Adjust the count of total modified error left.
2035 // The count of bits left is adjusted elsewhere based on real coded frame
2037 twopass->modified_error_left -= kf_group_err;
2040 // For VBR...adjustment to the frame target based on error from previous frames
2041 void vbr_rate_correction(int * this_frame_target,
2042 const int64_t vbr_bits_off_target) {
2043 int max_delta = (*this_frame_target * 15) / 100;
2045 // vbr_bits_off_target > 0 means we have extra bits to spend
2046 if (vbr_bits_off_target > 0) {
2047 *this_frame_target +=
2048 (vbr_bits_off_target > max_delta) ? max_delta
2049 : (int)vbr_bits_off_target;
2051 *this_frame_target -=
2052 (vbr_bits_off_target < -max_delta) ? max_delta
2053 : (int)-vbr_bits_off_target;
2057 // Define the reference buffers that will be updated post encode.
2058 void configure_buffer_updates(VP9_COMP *cpi) {
2059 TWO_PASS *const twopass = &cpi->twopass;
2061 cpi->rc.is_src_frame_alt_ref = 0;
2062 switch (twopass->gf_group.update_type[twopass->gf_group.index]) {
2064 cpi->refresh_last_frame = 1;
2065 cpi->refresh_golden_frame = 1;
2066 cpi->refresh_alt_ref_frame = 1;
2069 cpi->refresh_last_frame = 1;
2070 cpi->refresh_golden_frame = 0;
2071 cpi->refresh_alt_ref_frame = 0;
2074 cpi->refresh_last_frame = 1;
2075 cpi->refresh_golden_frame = 1;
2076 cpi->refresh_alt_ref_frame = 0;
2078 case OVERLAY_UPDATE:
2079 cpi->refresh_last_frame = 0;
2080 cpi->refresh_golden_frame = 1;
2081 cpi->refresh_alt_ref_frame = 0;
2082 cpi->rc.is_src_frame_alt_ref = 1;
2085 cpi->refresh_last_frame = 0;
2086 cpi->refresh_golden_frame = 0;
2087 cpi->refresh_alt_ref_frame = 1;
2095 void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
2096 VP9_COMMON *const cm = &cpi->common;
2097 RATE_CONTROL *const rc = &cpi->rc;
2098 TWO_PASS *const twopass = &cpi->twopass;
2100 FIRSTPASS_STATS this_frame;
2101 FIRSTPASS_STATS this_frame_copy;
2104 LAYER_CONTEXT *lc = NULL;
2105 const int is_spatial_svc = (cpi->use_svc &&
2106 cpi->svc.number_temporal_layers == 1);
2107 if (is_spatial_svc) {
2108 lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
2109 frames_left = (int)(twopass->total_stats.count -
2110 lc->current_video_frame_in_layer);
2112 frames_left = (int)(twopass->total_stats.count -
2113 cm->current_video_frame);
2116 if (!twopass->stats_in)
2119 // If this is an arf frame then we dont want to read the stats file or
2120 // advance the input pointer as we already have what we need.
2121 if (twopass->gf_group.update_type[twopass->gf_group.index] == ARF_UPDATE) {
2123 configure_buffer_updates(cpi);
2124 target_rate = twopass->gf_group.bit_allocation[twopass->gf_group.index];
2125 target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
2126 rc->base_frame_target = target_rate;
2127 #ifdef LONG_TERM_VBR_CORRECTION
2128 // Correction to rate target based on prior over or under shoot.
2129 if (cpi->oxcf.rc_mode == VPX_VBR)
2130 vbr_rate_correction(&target_rate, rc->vbr_bits_off_target);
2132 vp9_rc_set_frame_target(cpi, target_rate);
2133 cm->frame_type = INTER_FRAME;
2137 vp9_clear_system_state();
2139 if (is_spatial_svc && twopass->kf_intra_err_min == 0) {
2140 twopass->kf_intra_err_min = KF_MB_INTRA_MIN * cpi->common.MBs;
2141 twopass->gf_intra_err_min = GF_MB_INTRA_MIN * cpi->common.MBs;
2144 if (cpi->oxcf.rc_mode == VPX_Q) {
2145 twopass->active_worst_quality = cpi->oxcf.cq_level;
2146 } else if (cm->current_video_frame == 0 ||
2147 (is_spatial_svc && lc->current_video_frame_in_layer == 0)) {
2148 // Special case code for first frame.
2149 const int section_target_bandwidth = (int)(twopass->bits_left /
2151 const int tmp_q = get_twopass_worst_quality(cpi, &twopass->total_left_stats,
2152 section_target_bandwidth);
2153 twopass->active_worst_quality = tmp_q;
2154 rc->ni_av_qi = tmp_q;
2155 rc->avg_q = vp9_convert_qindex_to_q(tmp_q);
2157 vp9_zero(this_frame);
2158 if (EOF == input_stats(twopass, &this_frame))
2161 // Local copy of the current frame's first pass stats.
2162 this_frame_copy = this_frame;
2164 // Keyframe and section processing.
2165 if (rc->frames_to_key == 0 ||
2166 (cpi->frame_flags & FRAMEFLAGS_KEY)) {
2167 // Define next KF group and assign bits to it.
2168 find_next_key_frame(cpi, &this_frame_copy);
2170 cm->frame_type = INTER_FRAME;
2173 if (is_spatial_svc) {
2174 if (cpi->svc.spatial_layer_id == 0) {
2175 lc->is_key_frame = (cm->frame_type == KEY_FRAME);
2177 cm->frame_type = INTER_FRAME;
2178 lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
2180 if (lc->is_key_frame) {
2181 cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
2186 // Define a new GF/ARF group. (Should always enter here for key frames).
2187 if (rc->frames_till_gf_update_due == 0) {
2188 define_gf_group(cpi, &this_frame_copy);
2190 if (twopass->gf_zeromotion_pct > 995) {
2191 // As long as max_thresh for encode breakout is small enough, it is ok
2192 // to enable it for show frame, i.e. set allow_encode_breakout to
2193 // ENCODE_BREAKOUT_LIMITED.
2194 if (!cm->show_frame)
2195 cpi->allow_encode_breakout = ENCODE_BREAKOUT_DISABLED;
2197 cpi->allow_encode_breakout = ENCODE_BREAKOUT_LIMITED;
2200 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
2201 cpi->refresh_golden_frame = 1;
2205 FIRSTPASS_STATS next_frame;
2206 if (lookup_next_frame_stats(twopass, &next_frame) != EOF) {
2207 twopass->next_iiratio = (int)(next_frame.intra_error /
2208 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
2212 configure_buffer_updates(cpi);
2214 target_rate = twopass->gf_group.bit_allocation[twopass->gf_group.index];
2215 if (cpi->common.frame_type == KEY_FRAME)
2216 target_rate = vp9_rc_clamp_iframe_target_size(cpi, target_rate);
2218 target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
2220 rc->base_frame_target = target_rate;
2221 #ifdef LONG_TERM_VBR_CORRECTION
2222 // Correction to rate target based on prior over or under shoot.
2223 if (cpi->oxcf.rc_mode == VPX_VBR)
2224 vbr_rate_correction(&target_rate, rc->vbr_bits_off_target);
2226 vp9_rc_set_frame_target(cpi, target_rate);
2228 // Update the total stats remaining structure.
2229 subtract_stats(&twopass->total_left_stats, &this_frame);
2231 #if CONFIG_FP_MB_STATS
2232 if (cpi->use_fp_mb_stats) {
2233 input_mb_stats(&twopass->this_frame_mb_stats, cm);
2238 void vp9_twopass_postencode_update(VP9_COMP *cpi) {
2239 TWO_PASS *const twopass = &cpi->twopass;
2240 RATE_CONTROL *const rc = &cpi->rc;
2241 #ifdef LONG_TERM_VBR_CORRECTION
2242 // In this experimental mode, the VBR correction is done exclusively through
2243 // rc->vbr_bits_off_target. Based on the sign of this value, a limited %
2244 // adjustment is made to the target rate of subsequent frames, to try and
2245 // push it back towards 0. This mode is less likely to suffer from
2246 // extreme behaviour at the end of a clip or group of frames.
2247 const int bits_used = rc->base_frame_target;
2248 rc->vbr_bits_off_target += rc->base_frame_target - rc->projected_frame_size;
2250 // In this mode, VBR correction is acheived by altering bits_left,
2251 // kf_group_bits & gf_group_bits to reflect any deviation from the target
2252 // rate in this frame. This alters the allocation of bits to the
2253 // remaning frames in the group / clip.
2255 // This method can give rise to unstable behaviour near the end of a clip
2256 // or kf/gf group of frames where any accumulated error is corrected over an
2257 // ever decreasing number of frames. Hence we change the balance of target
2258 // vs. actual bitrate gradually as we progress towards the end of the
2259 // sequence in order to mitigate this effect.
2260 const double progress =
2261 (double)(twopass->stats_in - twopass->stats_in_start) /
2262 (twopass->stats_in_end - twopass->stats_in_start);
2263 const int bits_used = (int)(progress * rc->this_frame_target +
2264 (1.0 - progress) * rc->projected_frame_size);
2267 twopass->bits_left = MAX(twopass->bits_left - bits_used, 0);
2269 #ifdef LONG_TERM_VBR_CORRECTION
2270 if (cpi->common.frame_type != KEY_FRAME &&
2271 !vp9_is_upper_layer_key_frame(cpi)) {
2273 if (cpi->common.frame_type == KEY_FRAME ||
2274 vp9_is_upper_layer_key_frame(cpi)) {
2275 // For key frames kf_group_bits already had the target bits subtracted out.
2276 // So now update to the correct value based on the actual bits used.
2277 twopass->kf_group_bits += rc->this_frame_target - bits_used;
2280 twopass->kf_group_bits -= bits_used;
2282 twopass->kf_group_bits = MAX(twopass->kf_group_bits, 0);
2284 // Increment the gf group index ready for the next frame.
2285 ++twopass->gf_group.index;