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_rd.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 const FIRSTPASS_STATS *read_frame_stats(const TWO_PASS *p, int offset) {
93 if ((offset >= 0 && p->stats_in + offset >= p->stats_in_end) ||
94 (offset < 0 && p->stats_in + offset < p->stats_in_start)) {
98 return &p->stats_in[offset];
101 static int input_stats(TWO_PASS *p, FIRSTPASS_STATS *fps) {
102 if (p->stats_in >= p->stats_in_end)
110 static void output_stats(FIRSTPASS_STATS *stats,
111 struct vpx_codec_pkt_list *pktlist) {
112 struct vpx_codec_cx_pkt pkt;
113 pkt.kind = VPX_CODEC_STATS_PKT;
114 pkt.data.twopass_stats.buf = stats;
115 pkt.data.twopass_stats.sz = sizeof(FIRSTPASS_STATS);
116 vpx_codec_pkt_list_add(pktlist, &pkt);
122 fpfile = fopen("firstpass.stt", "a");
124 fprintf(fpfile, "%12.0f %12.0f %12.0f %12.0f %12.4f %12.4f"
125 "%12.4f %12.4f %12.4f %12.4f %12.4f %12.4f %12.4f"
126 "%12.0f %12.0f %12.4f %12.0f %12.0f %12.4f\n",
130 stats->sr_coded_error,
133 stats->pcnt_second_ref,
141 stats->mv_in_out_count,
150 #if CONFIG_FP_MB_STATS
151 static void output_fpmb_stats(uint8_t *this_frame_mb_stats, VP9_COMMON *cm,
152 struct vpx_codec_pkt_list *pktlist) {
153 struct vpx_codec_cx_pkt pkt;
154 pkt.kind = VPX_CODEC_FPMB_STATS_PKT;
155 pkt.data.firstpass_mb_stats.buf = this_frame_mb_stats;
156 pkt.data.firstpass_mb_stats.sz = cm->MBs * sizeof(uint8_t);
157 vpx_codec_pkt_list_add(pktlist, &pkt);
161 static void zero_stats(FIRSTPASS_STATS *section) {
162 section->frame = 0.0;
163 section->intra_error = 0.0;
164 section->coded_error = 0.0;
165 section->sr_coded_error = 0.0;
166 section->pcnt_inter = 0.0;
167 section->pcnt_motion = 0.0;
168 section->pcnt_second_ref = 0.0;
169 section->pcnt_neutral = 0.0;
171 section->mvr_abs = 0.0;
173 section->mvc_abs = 0.0;
176 section->mv_in_out_count = 0.0;
177 section->new_mv_count = 0.0;
178 section->count = 0.0;
179 section->duration = 1.0;
180 section->spatial_layer_id = 0;
183 static void accumulate_stats(FIRSTPASS_STATS *section,
184 const FIRSTPASS_STATS *frame) {
185 section->frame += frame->frame;
186 section->spatial_layer_id = frame->spatial_layer_id;
187 section->intra_error += frame->intra_error;
188 section->coded_error += frame->coded_error;
189 section->sr_coded_error += frame->sr_coded_error;
190 section->pcnt_inter += frame->pcnt_inter;
191 section->pcnt_motion += frame->pcnt_motion;
192 section->pcnt_second_ref += frame->pcnt_second_ref;
193 section->pcnt_neutral += frame->pcnt_neutral;
194 section->MVr += frame->MVr;
195 section->mvr_abs += frame->mvr_abs;
196 section->MVc += frame->MVc;
197 section->mvc_abs += frame->mvc_abs;
198 section->MVrv += frame->MVrv;
199 section->MVcv += frame->MVcv;
200 section->mv_in_out_count += frame->mv_in_out_count;
201 section->new_mv_count += frame->new_mv_count;
202 section->count += frame->count;
203 section->duration += frame->duration;
206 static void subtract_stats(FIRSTPASS_STATS *section,
207 const FIRSTPASS_STATS *frame) {
208 section->frame -= frame->frame;
209 section->intra_error -= frame->intra_error;
210 section->coded_error -= frame->coded_error;
211 section->sr_coded_error -= frame->sr_coded_error;
212 section->pcnt_inter -= frame->pcnt_inter;
213 section->pcnt_motion -= frame->pcnt_motion;
214 section->pcnt_second_ref -= frame->pcnt_second_ref;
215 section->pcnt_neutral -= frame->pcnt_neutral;
216 section->MVr -= frame->MVr;
217 section->mvr_abs -= frame->mvr_abs;
218 section->MVc -= frame->MVc;
219 section->mvc_abs -= frame->mvc_abs;
220 section->MVrv -= frame->MVrv;
221 section->MVcv -= frame->MVcv;
222 section->mv_in_out_count -= frame->mv_in_out_count;
223 section->new_mv_count -= frame->new_mv_count;
224 section->count -= frame->count;
225 section->duration -= frame->duration;
228 static void avg_stats(FIRSTPASS_STATS *section) {
229 if (section->count < 1.0)
232 section->intra_error /= section->count;
233 section->coded_error /= section->count;
234 section->sr_coded_error /= section->count;
235 section->pcnt_inter /= section->count;
236 section->pcnt_second_ref /= section->count;
237 section->pcnt_neutral /= section->count;
238 section->pcnt_motion /= section->count;
239 section->MVr /= section->count;
240 section->mvr_abs /= section->count;
241 section->MVc /= section->count;
242 section->mvc_abs /= section->count;
243 section->MVrv /= section->count;
244 section->MVcv /= section->count;
245 section->mv_in_out_count /= section->count;
246 section->duration /= section->count;
249 // Calculate a modified Error used in distributing bits between easier and
251 static double calculate_modified_err(const TWO_PASS *twopass,
252 const VP9EncoderConfig *oxcf,
253 const FIRSTPASS_STATS *this_frame) {
254 const FIRSTPASS_STATS *const stats = &twopass->total_stats;
255 const double av_err = stats->coded_error / stats->count;
256 const double modified_error = av_err *
257 pow(this_frame->coded_error / DOUBLE_DIVIDE_CHECK(av_err),
258 oxcf->two_pass_vbrbias / 100.0);
259 return fclamp(modified_error,
260 twopass->modified_error_min, twopass->modified_error_max);
263 // This function returns the maximum target rate per frame.
264 static int frame_max_bits(const RATE_CONTROL *rc,
265 const VP9EncoderConfig *oxcf) {
266 int64_t max_bits = ((int64_t)rc->avg_frame_bandwidth *
267 (int64_t)oxcf->two_pass_vbrmax_section) / 100;
270 else if (max_bits > rc->max_frame_bandwidth)
271 max_bits = rc->max_frame_bandwidth;
273 return (int)max_bits;
276 void vp9_init_first_pass(VP9_COMP *cpi) {
277 zero_stats(&cpi->twopass.total_stats);
280 void vp9_end_first_pass(VP9_COMP *cpi) {
281 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
283 for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
284 output_stats(&cpi->svc.layer_context[i].twopass.total_stats,
285 cpi->output_pkt_list);
288 output_stats(&cpi->twopass.total_stats, cpi->output_pkt_list);
292 static vp9_variance_fn_t get_block_variance_fn(BLOCK_SIZE bsize) {
305 static unsigned int get_prediction_error(BLOCK_SIZE bsize,
306 const struct buf_2d *src,
307 const struct buf_2d *ref) {
309 const vp9_variance_fn_t fn = get_block_variance_fn(bsize);
310 fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
314 // Refine the motion search range according to the frame dimension
315 // for first pass test.
316 static int get_search_range(const VP9_COMMON *cm) {
318 const int dim = MIN(cm->width, cm->height);
320 while ((dim << sr) < MAX_FULL_PEL_VAL)
325 static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
326 const MV *ref_mv, MV *best_mv,
327 int *best_motion_err) {
328 MACROBLOCKD *const xd = &x->e_mbd;
330 MV ref_mv_full = {ref_mv->row >> 3, ref_mv->col >> 3};
331 int num00, tmp_err, n;
332 const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
333 vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
334 const int new_mv_mode_penalty = 256;
337 int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
338 const int sr = get_search_range(&cpi->common);
342 // Override the default variance function to use MSE.
343 v_fn_ptr.vf = get_block_variance_fn(bsize);
345 // Center the initial step/diamond search on best mv.
346 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
348 x->sadperbit16, &num00, &v_fn_ptr, ref_mv);
349 if (tmp_err < INT_MAX)
350 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
351 if (tmp_err < INT_MAX - new_mv_mode_penalty)
352 tmp_err += new_mv_mode_penalty;
354 if (tmp_err < *best_motion_err) {
355 *best_motion_err = tmp_err;
359 // Carry out further step/diamond searches as necessary.
363 while (n < further_steps) {
369 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
370 step_param + n, x->sadperbit16,
371 &num00, &v_fn_ptr, ref_mv);
372 if (tmp_err < INT_MAX)
373 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
374 if (tmp_err < INT_MAX - new_mv_mode_penalty)
375 tmp_err += new_mv_mode_penalty;
377 if (tmp_err < *best_motion_err) {
378 *best_motion_err = tmp_err;
385 static BLOCK_SIZE get_bsize(const VP9_COMMON *cm, int mb_row, int mb_col) {
386 if (2 * mb_col + 1 < cm->mi_cols) {
387 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_16X16
390 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_8X16
395 static int find_fp_qindex() {
398 for (i = 0; i < QINDEX_RANGE; ++i)
399 if (vp9_convert_qindex_to_q(i) >= 30.0)
402 if (i == QINDEX_RANGE)
408 static void set_first_pass_params(VP9_COMP *cpi) {
409 VP9_COMMON *const cm = &cpi->common;
410 if (!cpi->refresh_alt_ref_frame &&
411 (cm->current_video_frame == 0 ||
412 (cpi->frame_flags & FRAMEFLAGS_KEY))) {
413 cm->frame_type = KEY_FRAME;
415 cm->frame_type = INTER_FRAME;
417 // Do not use periodic key frames.
418 cpi->rc.frames_to_key = INT_MAX;
421 void vp9_first_pass(VP9_COMP *cpi) {
423 MACROBLOCK *const x = &cpi->mb;
424 VP9_COMMON *const cm = &cpi->common;
425 MACROBLOCKD *const xd = &x->e_mbd;
427 struct macroblock_plane *const p = x->plane;
428 struct macroblockd_plane *const pd = xd->plane;
429 const PICK_MODE_CONTEXT *ctx = &cpi->pc_root->none;
432 int recon_yoffset, recon_uvoffset;
433 YV12_BUFFER_CONFIG *const lst_yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
434 YV12_BUFFER_CONFIG *gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
435 YV12_BUFFER_CONFIG *const new_yv12 = get_frame_new_buffer(cm);
436 int recon_y_stride = lst_yv12->y_stride;
437 int recon_uv_stride = lst_yv12->uv_stride;
438 int uv_mb_height = 16 >> (lst_yv12->y_height > lst_yv12->uv_height);
439 int64_t intra_error = 0;
440 int64_t coded_error = 0;
441 int64_t sr_coded_error = 0;
443 int sum_mvr = 0, sum_mvc = 0;
444 int sum_mvr_abs = 0, sum_mvc_abs = 0;
445 int64_t sum_mvrs = 0, sum_mvcs = 0;
448 int second_ref_count = 0;
449 int intrapenalty = 256;
450 int neutral_count = 0;
451 int new_mv_count = 0;
452 int sum_in_vectors = 0;
453 uint32_t lastmv_as_int = 0;
454 TWO_PASS *twopass = &cpi->twopass;
455 const MV zero_mv = {0, 0};
456 const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;
458 #if CONFIG_FP_MB_STATS
459 if (cpi->use_fp_mb_stats) {
460 vp9_zero_array(cpi->twopass.frame_mb_stats_buf, cm->MBs);
464 vp9_clear_system_state();
466 set_first_pass_params(cpi);
467 vp9_set_quantizer(cm, find_fp_qindex());
469 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
470 MV_REFERENCE_FRAME ref_frame = LAST_FRAME;
471 const YV12_BUFFER_CONFIG *scaled_ref_buf = NULL;
472 twopass = &cpi->svc.layer_context[cpi->svc.spatial_layer_id].twopass;
474 if (cpi->common.current_video_frame == 0) {
475 cpi->ref_frame_flags = 0;
477 LAYER_CONTEXT *lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
478 if (lc->current_video_frame_in_layer == 0)
479 cpi->ref_frame_flags = VP9_GOLD_FLAG;
481 cpi->ref_frame_flags = VP9_LAST_FLAG | VP9_GOLD_FLAG;
484 vp9_scale_references(cpi);
486 // Use either last frame or alt frame for motion search.
487 if (cpi->ref_frame_flags & VP9_LAST_FLAG) {
488 scaled_ref_buf = vp9_get_scaled_ref_frame(cpi, LAST_FRAME);
489 ref_frame = LAST_FRAME;
490 } else if (cpi->ref_frame_flags & VP9_GOLD_FLAG) {
491 scaled_ref_buf = vp9_get_scaled_ref_frame(cpi, GOLDEN_FRAME);
492 ref_frame = GOLDEN_FRAME;
495 if (scaled_ref_buf != NULL)
496 first_ref_buf = scaled_ref_buf;
498 recon_y_stride = new_yv12->y_stride;
499 recon_uv_stride = new_yv12->uv_stride;
500 uv_mb_height = 16 >> (new_yv12->y_height > new_yv12->uv_height);
502 // Disable golden frame for svc first pass for now.
504 set_ref_ptrs(cm, xd, ref_frame, NONE);
506 cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
507 &cpi->scaled_source);
510 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
512 vp9_setup_src_planes(x, cpi->Source, 0, 0);
513 vp9_setup_pre_planes(xd, 0, first_ref_buf, 0, 0, NULL);
514 vp9_setup_dst_planes(xd->plane, new_yv12, 0, 0);
516 xd->mi = cm->mi_grid_visible;
519 vp9_frame_init_quantizer(cpi);
521 for (i = 0; i < MAX_MB_PLANE; ++i) {
522 p[i].coeff = ctx->coeff_pbuf[i][1];
523 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
524 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
525 p[i].eobs = ctx->eobs_pbuf[i][1];
529 vp9_init_mv_probs(cm);
530 vp9_initialize_rd_consts(cpi);
532 // Tiling is ignored in the first pass.
533 vp9_tile_init(&tile, cm, 0, 0);
535 for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
538 best_ref_mv.as_int = 0;
540 // Reset above block coeffs.
541 xd->up_available = (mb_row != 0);
542 recon_yoffset = (mb_row * recon_y_stride * 16);
543 recon_uvoffset = (mb_row * recon_uv_stride * uv_mb_height);
545 // Set up limit values for motion vectors to prevent them extending
546 // outside the UMV borders.
547 x->mv_row_min = -((mb_row * 16) + BORDER_MV_PIXELS_B16);
548 x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16)
549 + BORDER_MV_PIXELS_B16;
551 for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
553 const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
554 double error_weight = 1.0;
555 const BLOCK_SIZE bsize = get_bsize(cm, mb_row, mb_col);
557 vp9_clear_system_state();
559 xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
560 xd->plane[1].dst.buf = new_yv12->u_buffer + recon_uvoffset;
561 xd->plane[2].dst.buf = new_yv12->v_buffer + recon_uvoffset;
562 xd->left_available = (mb_col != 0);
563 xd->mi[0]->mbmi.sb_type = bsize;
564 xd->mi[0]->mbmi.ref_frame[0] = INTRA_FRAME;
565 set_mi_row_col(xd, &tile,
566 mb_row << 1, num_8x8_blocks_high_lookup[bsize],
567 mb_col << 1, num_8x8_blocks_wide_lookup[bsize],
568 cm->mi_rows, cm->mi_cols);
570 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
571 const int energy = vp9_block_energy(cpi, x, bsize);
572 error_weight = vp9_vaq_inv_q_ratio(energy);
575 // Do intra 16x16 prediction.
577 xd->mi[0]->mbmi.mode = DC_PRED;
578 xd->mi[0]->mbmi.tx_size = use_dc_pred ?
579 (bsize >= BLOCK_16X16 ? TX_16X16 : TX_8X8) : TX_4X4;
580 vp9_encode_intra_block_plane(x, bsize, 0);
581 this_error = vp9_get_mb_ss(x->plane[0].src_diff);
583 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
584 vp9_clear_system_state();
585 this_error = (int)(this_error * error_weight);
588 // Intrapenalty below deals with situations where the intra and inter
589 // error scores are very low (e.g. a plain black frame).
590 // We do not have special cases in first pass for 0,0 and nearest etc so
591 // all inter modes carry an overhead cost estimate for the mv.
592 // When the error score is very low this causes us to pick all or lots of
593 // INTRA modes and throw lots of key frames.
594 // This penalty adds a cost matching that of a 0,0 mv to the intra case.
595 this_error += intrapenalty;
597 // Accumulate the intra error.
598 intra_error += (int64_t)this_error;
600 #if CONFIG_FP_MB_STATS
601 if (cpi->use_fp_mb_stats) {
603 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] = 0;
607 // Set up limit values for motion vectors to prevent them extending
608 // outside the UMV borders.
609 x->mv_col_min = -((mb_col * 16) + BORDER_MV_PIXELS_B16);
610 x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + BORDER_MV_PIXELS_B16;
612 // Other than for the first frame do a motion search.
613 if (cm->current_video_frame > 0) {
614 int tmp_err, motion_error, raw_motion_error;
616 struct buf_2d unscaled_last_source_buf_2d;
618 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
619 motion_error = get_prediction_error(bsize, &x->plane[0].src,
620 &xd->plane[0].pre[0]);
621 // Assume 0,0 motion with no mv overhead.
622 mv.as_int = tmp_mv.as_int = 0;
624 // Compute the motion error of the 0,0 motion using the last source
625 // frame as the reference. Skip the further motion search on
626 // reconstructed frame if this error is small.
627 unscaled_last_source_buf_2d.buf =
628 cpi->unscaled_last_source->y_buffer + recon_yoffset;
629 unscaled_last_source_buf_2d.stride =
630 cpi->unscaled_last_source->y_stride;
631 raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
632 &unscaled_last_source_buf_2d);
634 // TODO(pengchong): Replace the hard-coded threshold
635 if (raw_motion_error > 25 ||
636 (cpi->use_svc && cpi->svc.number_temporal_layers == 1)) {
637 // Test last reference frame using the previous best mv as the
638 // starting point (best reference) for the search.
639 first_pass_motion_search(cpi, x, &best_ref_mv.as_mv, &mv.as_mv,
641 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
642 vp9_clear_system_state();
643 motion_error = (int)(motion_error * error_weight);
646 // If the current best reference mv is not centered on 0,0 then do a
647 // 0,0 based search as well.
648 if (best_ref_mv.as_int) {
650 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv, &tmp_err);
651 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
652 vp9_clear_system_state();
653 tmp_err = (int)(tmp_err * error_weight);
656 if (tmp_err < motion_error) {
657 motion_error = tmp_err;
658 mv.as_int = tmp_mv.as_int;
662 // Search in an older reference frame.
663 if (cm->current_video_frame > 1 && gld_yv12 != NULL) {
664 // Assume 0,0 motion with no mv overhead.
667 xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
668 gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
669 &xd->plane[0].pre[0]);
671 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv,
673 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
674 vp9_clear_system_state();
675 gf_motion_error = (int)(gf_motion_error * error_weight);
678 if (gf_motion_error < motion_error && gf_motion_error < this_error)
681 // Reset to last frame as reference buffer.
682 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
683 xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
684 xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
686 // In accumulating a score for the older reference frame take the
687 // best of the motion predicted score and the intra coded error
688 // (just as will be done for) accumulation of "coded_error" for
690 if (gf_motion_error < this_error)
691 sr_coded_error += gf_motion_error;
693 sr_coded_error += this_error;
695 sr_coded_error += motion_error;
698 sr_coded_error += motion_error;
701 // Start by assuming that intra mode is best.
702 best_ref_mv.as_int = 0;
704 #if CONFIG_FP_MB_STATS
705 if (cpi->use_fp_mb_stats) {
706 // intra predication statistics
707 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] = 0;
708 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] |=
710 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] &=
711 (~FPMB_NONZERO_MOTION_MASK);
712 if (this_error > FPMB_ERROR_LEVEL4_TH) {
713 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] |=
714 FPMB_ERROR_LEVEL4_MASK;
715 } else if (this_error > FPMB_ERROR_LEVEL3_TH) {
716 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] |=
717 FPMB_ERROR_LEVEL3_MASK;
718 } else if (this_error > FPMB_ERROR_LEVEL2_TH) {
719 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] |=
720 FPMB_ERROR_LEVEL2_MASK;
721 } else if (this_error > FPMB_ERROR_LEVEL1_TH) {
722 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] |=
723 FPMB_ERROR_LEVEL1_MASK;
725 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] |=
726 FPMB_ERROR_LEVEL0_MASK;
731 if (motion_error <= this_error) {
732 // Keep a count of cases where the inter and intra were very close
733 // and very low. This helps with scene cut detection for example in
734 // cropped clips with black bars at the sides or top and bottom.
735 if (((this_error - intrapenalty) * 9 <= motion_error * 10) &&
736 this_error < 2 * intrapenalty)
741 this_error = motion_error;
742 xd->mi[0]->mbmi.mode = NEWMV;
743 xd->mi[0]->mbmi.mv[0] = mv;
744 xd->mi[0]->mbmi.tx_size = TX_4X4;
745 xd->mi[0]->mbmi.ref_frame[0] = LAST_FRAME;
746 xd->mi[0]->mbmi.ref_frame[1] = NONE;
747 vp9_build_inter_predictors_sby(xd, mb_row << 1, mb_col << 1, bsize);
748 vp9_encode_sby_pass1(x, bsize);
749 sum_mvr += mv.as_mv.row;
750 sum_mvr_abs += abs(mv.as_mv.row);
751 sum_mvc += mv.as_mv.col;
752 sum_mvc_abs += abs(mv.as_mv.col);
753 sum_mvrs += mv.as_mv.row * mv.as_mv.row;
754 sum_mvcs += mv.as_mv.col * mv.as_mv.col;
757 best_ref_mv.as_int = mv.as_int;
759 #if CONFIG_FP_MB_STATS
760 if (cpi->use_fp_mb_stats) {
761 // inter predication statistics
762 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] = 0;
763 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] &=
764 (~FPMB_DCINTRA_MASK);
765 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] &=
766 (~FPMB_NONZERO_MOTION_MASK);
767 if (this_error > FPMB_ERROR_LEVEL4_TH) {
768 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] |=
769 FPMB_ERROR_LEVEL4_MASK;
770 } else if (this_error > FPMB_ERROR_LEVEL3_TH) {
771 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] |=
772 FPMB_ERROR_LEVEL3_MASK;
773 } else if (this_error > FPMB_ERROR_LEVEL2_TH) {
774 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] |=
775 FPMB_ERROR_LEVEL2_MASK;
776 } else if (this_error > FPMB_ERROR_LEVEL1_TH) {
777 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] |=
778 FPMB_ERROR_LEVEL1_MASK;
780 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] |=
781 FPMB_ERROR_LEVEL0_MASK;
789 #if CONFIG_FP_MB_STATS
790 if (cpi->use_fp_mb_stats) {
791 cpi->twopass.frame_mb_stats_buf[mb_row * cm->mb_cols + mb_col] |=
792 FPMB_NONZERO_MOTION_MASK;
796 // Non-zero vector, was it different from the last non zero vector?
797 if (mv.as_int != lastmv_as_int)
799 lastmv_as_int = mv.as_int;
801 // Does the row vector point inwards or outwards?
802 if (mb_row < cm->mb_rows / 2) {
803 if (mv.as_mv.row > 0)
805 else if (mv.as_mv.row < 0)
807 } else if (mb_row > cm->mb_rows / 2) {
808 if (mv.as_mv.row > 0)
810 else if (mv.as_mv.row < 0)
814 // Does the col vector point inwards or outwards?
815 if (mb_col < cm->mb_cols / 2) {
816 if (mv.as_mv.col > 0)
818 else if (mv.as_mv.col < 0)
820 } else if (mb_col > cm->mb_cols / 2) {
821 if (mv.as_mv.col > 0)
823 else if (mv.as_mv.col < 0)
829 sr_coded_error += (int64_t)this_error;
831 coded_error += (int64_t)this_error;
833 // Adjust to the next column of MBs.
834 x->plane[0].src.buf += 16;
835 x->plane[1].src.buf += uv_mb_height;
836 x->plane[2].src.buf += uv_mb_height;
839 recon_uvoffset += uv_mb_height;
842 // Adjust to the next row of MBs.
843 x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
844 x->plane[1].src.buf += uv_mb_height * x->plane[1].src.stride -
845 uv_mb_height * cm->mb_cols;
846 x->plane[2].src.buf += uv_mb_height * x->plane[1].src.stride -
847 uv_mb_height * cm->mb_cols;
849 vp9_clear_system_state();
852 vp9_clear_system_state();
856 fps.frame = cm->current_video_frame;
857 fps.spatial_layer_id = cpi->svc.spatial_layer_id;
858 fps.intra_error = (double)(intra_error >> 8);
859 fps.coded_error = (double)(coded_error >> 8);
860 fps.sr_coded_error = (double)(sr_coded_error >> 8);
862 fps.pcnt_inter = (double)intercount / cm->MBs;
863 fps.pcnt_second_ref = (double)second_ref_count / cm->MBs;
864 fps.pcnt_neutral = (double)neutral_count / cm->MBs;
867 fps.MVr = (double)sum_mvr / mvcount;
868 fps.mvr_abs = (double)sum_mvr_abs / mvcount;
869 fps.MVc = (double)sum_mvc / mvcount;
870 fps.mvc_abs = (double)sum_mvc_abs / mvcount;
871 fps.MVrv = ((double)sum_mvrs - (fps.MVr * fps.MVr / mvcount)) / mvcount;
872 fps.MVcv = ((double)sum_mvcs - (fps.MVc * fps.MVc / mvcount)) / mvcount;
873 fps.mv_in_out_count = (double)sum_in_vectors / (mvcount * 2);
874 fps.new_mv_count = new_mv_count;
875 fps.pcnt_motion = (double)mvcount / cm->MBs;
883 fps.mv_in_out_count = 0.0;
884 fps.new_mv_count = 0.0;
885 fps.pcnt_motion = 0.0;
888 // TODO(paulwilkins): Handle the case when duration is set to 0, or
889 // something less than the full time between subsequent values of
890 // cpi->source_time_stamp.
891 fps.duration = (double)(cpi->source->ts_end - cpi->source->ts_start);
893 // Don't want to do output stats with a stack variable!
894 twopass->this_frame_stats = fps;
895 output_stats(&twopass->this_frame_stats, cpi->output_pkt_list);
896 accumulate_stats(&twopass->total_stats, &fps);
898 #if CONFIG_FP_MB_STATS
899 if (cpi->use_fp_mb_stats) {
900 output_fpmb_stats(twopass->frame_mb_stats_buf, cm, cpi->output_pkt_list);
905 // Copy the previous Last Frame back into gf and and arf buffers if
906 // the prediction is good enough... but also don't allow it to lag too far.
907 if ((twopass->sr_update_lag > 3) ||
908 ((cm->current_video_frame > 0) &&
909 (twopass->this_frame_stats.pcnt_inter > 0.20) &&
910 ((twopass->this_frame_stats.intra_error /
911 DOUBLE_DIVIDE_CHECK(twopass->this_frame_stats.coded_error)) > 2.0))) {
912 if (gld_yv12 != NULL) {
913 vp8_yv12_copy_frame(lst_yv12, gld_yv12);
915 twopass->sr_update_lag = 1;
917 ++twopass->sr_update_lag;
920 vp9_extend_frame_borders(new_yv12);
922 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
923 vp9_update_reference_frames(cpi);
925 // Swap frame pointers so last frame refers to the frame we just compressed.
926 swap_yv12(lst_yv12, new_yv12);
929 // Special case for the first frame. Copy into the GF buffer as a second
931 if (cm->current_video_frame == 0 && gld_yv12 != NULL) {
932 vp8_yv12_copy_frame(lst_yv12, gld_yv12);
935 // Use this to see what the first pass reconstruction looks like.
939 snprintf(filename, sizeof(filename), "enc%04d.yuv",
940 (int)cm->current_video_frame);
942 if (cm->current_video_frame == 0)
943 recon_file = fopen(filename, "wb");
945 recon_file = fopen(filename, "ab");
947 (void)fwrite(lst_yv12->buffer_alloc, lst_yv12->frame_size, 1, recon_file);
951 ++cm->current_video_frame;
953 vp9_inc_frame_in_layer(&cpi->svc);
956 static double calc_correction_factor(double err_per_mb,
961 const double error_term = err_per_mb / err_divisor;
963 // Adjustment based on actual quantizer to power term.
964 const double power_term = MIN(vp9_convert_qindex_to_q(q) * 0.0125 + pt_low,
967 // Calculate correction factor.
968 if (power_term < 1.0)
969 assert(error_term >= 0.0);
971 return fclamp(pow(error_term, power_term), 0.05, 5.0);
974 static int get_twopass_worst_quality(const VP9_COMP *cpi,
975 const FIRSTPASS_STATS *stats,
976 int section_target_bandwidth) {
977 const RATE_CONTROL *const rc = &cpi->rc;
978 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
980 if (section_target_bandwidth <= 0) {
981 return rc->worst_quality; // Highest value allowed
983 const int num_mbs = cpi->common.MBs;
984 const double section_err = stats->coded_error / stats->count;
985 const double err_per_mb = section_err / num_mbs;
986 const double speed_term = 1.0 + 0.04 * oxcf->speed;
987 const int target_norm_bits_per_mb = ((uint64_t)section_target_bandwidth <<
988 BPER_MB_NORMBITS) / num_mbs;
990 int is_svc_upper_layer = 0;
991 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1 &&
992 cpi->svc.spatial_layer_id > 0) {
993 is_svc_upper_layer = 1;
996 // Try and pick a max Q that will be high enough to encode the
997 // content at the given rate.
998 for (q = rc->best_quality; q < rc->worst_quality; ++q) {
999 const double factor =
1000 calc_correction_factor(err_per_mb, ERR_DIVISOR,
1001 is_svc_upper_layer ? SVC_FACTOR_PT_LOW :
1002 FACTOR_PT_LOW, FACTOR_PT_HIGH, q);
1003 const int bits_per_mb = vp9_rc_bits_per_mb(INTER_FRAME, q,
1004 factor * speed_term);
1005 if (bits_per_mb <= target_norm_bits_per_mb)
1009 // Restriction on active max q for constrained quality mode.
1010 if (cpi->oxcf.rc_mode == VPX_CQ)
1011 q = MAX(q, oxcf->cq_level);
1016 extern void vp9_new_framerate(VP9_COMP *cpi, double framerate);
1018 void vp9_init_second_pass(VP9_COMP *cpi) {
1019 SVC *const svc = &cpi->svc;
1020 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1021 const int is_spatial_svc = (svc->number_spatial_layers > 1) &&
1022 (svc->number_temporal_layers == 1);
1023 TWO_PASS *const twopass = is_spatial_svc ?
1024 &svc->layer_context[svc->spatial_layer_id].twopass : &cpi->twopass;
1026 FIRSTPASS_STATS *stats;
1028 zero_stats(&twopass->total_stats);
1029 zero_stats(&twopass->total_left_stats);
1031 if (!twopass->stats_in_end)
1034 stats = &twopass->total_stats;
1036 *stats = *twopass->stats_in_end;
1037 twopass->total_left_stats = *stats;
1039 frame_rate = 10000000.0 * stats->count / stats->duration;
1040 // Each frame can have a different duration, as the frame rate in the source
1041 // isn't guaranteed to be constant. The frame rate prior to the first frame
1042 // encoded in the second pass is a guess. However, the sum duration is not.
1043 // It is calculated based on the actual durations of all frames from the
1046 if (is_spatial_svc) {
1047 vp9_update_spatial_layer_framerate(cpi, frame_rate);
1048 twopass->bits_left = (int64_t)(stats->duration *
1049 svc->layer_context[svc->spatial_layer_id].target_bandwidth /
1052 vp9_new_framerate(cpi, frame_rate);
1053 twopass->bits_left = (int64_t)(stats->duration * oxcf->target_bandwidth /
1057 // Calculate a minimum intra value to be used in determining the IIratio
1058 // scores used in the second pass. We have this minimum to make sure
1059 // that clips that are static but "low complexity" in the intra domain
1060 // are still boosted appropriately for KF/GF/ARF.
1061 if (!is_spatial_svc) {
1062 // We don't know the number of MBs for each layer at this point.
1063 // So we will do it later.
1064 twopass->kf_intra_err_min = KF_MB_INTRA_MIN * cpi->common.MBs;
1065 twopass->gf_intra_err_min = GF_MB_INTRA_MIN * cpi->common.MBs;
1068 // This variable monitors how far behind the second ref update is lagging.
1069 twopass->sr_update_lag = 1;
1071 // Scan the first pass file and calculate a modified total error based upon
1072 // the bias/power function used to allocate bits.
1074 const double avg_error = stats->coded_error /
1075 DOUBLE_DIVIDE_CHECK(stats->count);
1076 const FIRSTPASS_STATS *s = twopass->stats_in;
1077 double modified_error_total = 0.0;
1078 twopass->modified_error_min = (avg_error *
1079 oxcf->two_pass_vbrmin_section) / 100;
1080 twopass->modified_error_max = (avg_error *
1081 oxcf->two_pass_vbrmax_section) / 100;
1082 while (s < twopass->stats_in_end) {
1083 modified_error_total += calculate_modified_err(twopass, oxcf, s);
1086 twopass->modified_error_left = modified_error_total;
1089 // Reset the vbr bits off target counter
1090 cpi->rc.vbr_bits_off_target = 0;
1093 // This function gives an estimate of how badly we believe the prediction
1094 // quality is decaying from frame to frame.
1095 static double get_prediction_decay_rate(const VP9_COMMON *cm,
1096 const FIRSTPASS_STATS *next_frame) {
1097 // Look at the observed drop in prediction quality between the last frame
1098 // and the GF buffer (which contains an older frame).
1099 const double mb_sr_err_diff = (next_frame->sr_coded_error -
1100 next_frame->coded_error) / cm->MBs;
1101 const double second_ref_decay = mb_sr_err_diff <= 512.0
1102 ? fclamp(pow(1.0 - (mb_sr_err_diff / 512.0), 0.5), 0.85, 1.0)
1105 return MIN(second_ref_decay, next_frame->pcnt_inter);
1108 // Function to test for a condition where a complex transition is followed
1109 // by a static section. For example in slide shows where there is a fade
1110 // between slides. This is to help with more optimal kf and gf positioning.
1111 static int detect_transition_to_still(TWO_PASS *twopass,
1112 int frame_interval, int still_interval,
1113 double loop_decay_rate,
1114 double last_decay_rate) {
1115 int trans_to_still = 0;
1117 // Break clause to detect very still sections after motion
1118 // For example a static image after a fade or other transition
1119 // instead of a clean scene cut.
1120 if (frame_interval > MIN_GF_INTERVAL &&
1121 loop_decay_rate >= 0.999 &&
1122 last_decay_rate < 0.9) {
1124 const FIRSTPASS_STATS *position = twopass->stats_in;
1125 FIRSTPASS_STATS tmp_next_frame;
1127 // Look ahead a few frames to see if static condition persists...
1128 for (j = 0; j < still_interval; ++j) {
1129 if (EOF == input_stats(twopass, &tmp_next_frame))
1132 if (tmp_next_frame.pcnt_inter - tmp_next_frame.pcnt_motion < 0.999)
1136 reset_fpf_position(twopass, position);
1138 // Only if it does do we signal a transition to still.
1139 if (j == still_interval)
1143 return trans_to_still;
1146 // This function detects a flash through the high relative pcnt_second_ref
1147 // score in the frame following a flash frame. The offset passed in should
1149 static int detect_flash(const TWO_PASS *twopass, int offset) {
1150 const FIRSTPASS_STATS *const next_frame = read_frame_stats(twopass, offset);
1152 // What we are looking for here is a situation where there is a
1153 // brief break in prediction (such as a flash) but subsequent frames
1154 // are reasonably well predicted by an earlier (pre flash) frame.
1155 // The recovery after a flash is indicated by a high pcnt_second_ref
1156 // compared to pcnt_inter.
1157 return next_frame != NULL &&
1158 next_frame->pcnt_second_ref > next_frame->pcnt_inter &&
1159 next_frame->pcnt_second_ref >= 0.5;
1162 // Update the motion related elements to the GF arf boost calculation.
1163 static void accumulate_frame_motion_stats(const FIRSTPASS_STATS *stats,
1165 double *mv_in_out_accumulator,
1166 double *abs_mv_in_out_accumulator,
1167 double *mv_ratio_accumulator) {
1168 const double pct = stats->pcnt_motion;
1170 // Accumulate Motion In/Out of frame stats.
1171 *mv_in_out = stats->mv_in_out_count * pct;
1172 *mv_in_out_accumulator += *mv_in_out;
1173 *abs_mv_in_out_accumulator += fabs(*mv_in_out);
1175 // Accumulate a measure of how uniform (or conversely how random) the motion
1176 // field is (a ratio of abs(mv) / mv).
1178 const double mvr_ratio = fabs(stats->mvr_abs) /
1179 DOUBLE_DIVIDE_CHECK(fabs(stats->MVr));
1180 const double mvc_ratio = fabs(stats->mvc_abs) /
1181 DOUBLE_DIVIDE_CHECK(fabs(stats->MVc));
1183 *mv_ratio_accumulator += pct * (mvr_ratio < stats->mvr_abs ?
1184 mvr_ratio : stats->mvr_abs);
1185 *mv_ratio_accumulator += pct * (mvc_ratio < stats->mvc_abs ?
1186 mvc_ratio : stats->mvc_abs);
1190 // Calculate a baseline boost number for the current frame.
1191 static double calc_frame_boost(const TWO_PASS *twopass,
1192 const FIRSTPASS_STATS *this_frame,
1193 double this_frame_mv_in_out) {
1196 // Underlying boost factor is based on inter intra error ratio.
1197 if (this_frame->intra_error > twopass->gf_intra_err_min)
1198 frame_boost = (IIFACTOR * this_frame->intra_error /
1199 DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
1201 frame_boost = (IIFACTOR * twopass->gf_intra_err_min /
1202 DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
1204 // Increase boost for frames where new data coming into frame (e.g. zoom out).
1205 // Slightly reduce boost if there is a net balance of motion out of the frame
1206 // (zoom in). The range for this_frame_mv_in_out is -1.0 to +1.0.
1207 if (this_frame_mv_in_out > 0.0)
1208 frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
1209 // In the extreme case the boost is halved.
1211 frame_boost += frame_boost * (this_frame_mv_in_out / 2.0);
1213 return MIN(frame_boost, GF_RMAX);
1216 static int calc_arf_boost(VP9_COMP *cpi, int offset,
1217 int f_frames, int b_frames,
1218 int *f_boost, int *b_boost) {
1219 TWO_PASS *const twopass = &cpi->twopass;
1221 double boost_score = 0.0;
1222 double mv_ratio_accumulator = 0.0;
1223 double decay_accumulator = 1.0;
1224 double this_frame_mv_in_out = 0.0;
1225 double mv_in_out_accumulator = 0.0;
1226 double abs_mv_in_out_accumulator = 0.0;
1228 int flash_detected = 0;
1230 // Search forward from the proposed arf/next gf position.
1231 for (i = 0; i < f_frames; ++i) {
1232 const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
1233 if (this_frame == NULL)
1236 // Update the motion related elements to the boost calculation.
1237 accumulate_frame_motion_stats(this_frame,
1238 &this_frame_mv_in_out, &mv_in_out_accumulator,
1239 &abs_mv_in_out_accumulator,
1240 &mv_ratio_accumulator);
1242 // We want to discount the flash frame itself and the recovery
1243 // frame that follows as both will have poor scores.
1244 flash_detected = detect_flash(twopass, i + offset) ||
1245 detect_flash(twopass, i + offset + 1);
1247 // Accumulate the effect of prediction quality decay.
1248 if (!flash_detected) {
1249 decay_accumulator *= get_prediction_decay_rate(&cpi->common, this_frame);
1250 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1251 ? MIN_DECAY_FACTOR : decay_accumulator;
1254 boost_score += decay_accumulator * calc_frame_boost(twopass, this_frame,
1255 this_frame_mv_in_out);
1258 *f_boost = (int)boost_score;
1260 // Reset for backward looking loop.
1262 mv_ratio_accumulator = 0.0;
1263 decay_accumulator = 1.0;
1264 this_frame_mv_in_out = 0.0;
1265 mv_in_out_accumulator = 0.0;
1266 abs_mv_in_out_accumulator = 0.0;
1268 // Search backward towards last gf position.
1269 for (i = -1; i >= -b_frames; --i) {
1270 const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
1271 if (this_frame == NULL)
1274 // Update the motion related elements to the boost calculation.
1275 accumulate_frame_motion_stats(this_frame,
1276 &this_frame_mv_in_out, &mv_in_out_accumulator,
1277 &abs_mv_in_out_accumulator,
1278 &mv_ratio_accumulator);
1280 // We want to discount the the flash frame itself and the recovery
1281 // frame that follows as both will have poor scores.
1282 flash_detected = detect_flash(twopass, i + offset) ||
1283 detect_flash(twopass, i + offset + 1);
1285 // Cumulative effect of prediction quality decay.
1286 if (!flash_detected) {
1287 decay_accumulator *= get_prediction_decay_rate(&cpi->common, this_frame);
1288 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1289 ? MIN_DECAY_FACTOR : decay_accumulator;
1292 boost_score += decay_accumulator * calc_frame_boost(twopass, this_frame,
1293 this_frame_mv_in_out);
1295 *b_boost = (int)boost_score;
1297 arf_boost = (*f_boost + *b_boost);
1298 if (arf_boost < ((b_frames + f_frames) * 20))
1299 arf_boost = ((b_frames + f_frames) * 20);
1304 // Calculate a section intra ratio used in setting max loop filter.
1305 static int calculate_section_intra_ratio(const FIRSTPASS_STATS *begin,
1306 const FIRSTPASS_STATS *end,
1307 int section_length) {
1308 const FIRSTPASS_STATS *s = begin;
1309 double intra_error = 0.0;
1310 double coded_error = 0.0;
1313 while (s < end && i < section_length) {
1314 intra_error += s->intra_error;
1315 coded_error += s->coded_error;
1320 return (int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
1323 // Calculate the total bits to allocate in this GF/ARF group.
1324 static int64_t calculate_total_gf_group_bits(VP9_COMP *cpi,
1325 double gf_group_err) {
1326 const RATE_CONTROL *const rc = &cpi->rc;
1327 const TWO_PASS *const twopass = &cpi->twopass;
1328 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
1329 int64_t total_group_bits;
1331 // Calculate the bits to be allocated to the group as a whole.
1332 if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0)) {
1333 total_group_bits = (int64_t)(twopass->kf_group_bits *
1334 (gf_group_err / twopass->kf_group_error_left));
1336 total_group_bits = 0;
1339 // Clamp odd edge cases.
1340 total_group_bits = (total_group_bits < 0) ?
1341 0 : (total_group_bits > twopass->kf_group_bits) ?
1342 twopass->kf_group_bits : total_group_bits;
1344 // Clip based on user supplied data rate variability limit.
1345 if (total_group_bits > (int64_t)max_bits * rc->baseline_gf_interval)
1346 total_group_bits = (int64_t)max_bits * rc->baseline_gf_interval;
1348 return total_group_bits;
1351 // Calculate the number bits extra to assign to boosted frames in a group.
1352 static int calculate_boost_bits(int frame_count,
1353 int boost, int64_t total_group_bits) {
1354 int allocation_chunks;
1356 // return 0 for invalid inputs (could arise e.g. through rounding errors)
1357 if (!boost || (total_group_bits <= 0) || (frame_count <= 0) )
1360 allocation_chunks = (frame_count * 100) + boost;
1362 // Prevent overflow.
1364 int divisor = boost >> 10;
1366 allocation_chunks /= divisor;
1369 // Calculate the number of extra bits for use in the boosted frame or frames.
1370 return MAX((int)(((int64_t)boost * total_group_bits) / allocation_chunks), 0);
1373 // Current limit on maximum number of active arfs in a GF/ARF group.
1374 #define MAX_ACTIVE_ARFS 2
1377 // This function indirects the choice of buffers for arfs.
1378 // At the moment the values are fixed but this may change as part of
1379 // the integration process with other codec features that swap buffers around.
1380 static void get_arf_buffer_indices(unsigned char *arf_buffer_indices) {
1381 arf_buffer_indices[0] = ARF_SLOT1;
1382 arf_buffer_indices[1] = ARF_SLOT2;
1385 static void allocate_gf_group_bits(VP9_COMP *cpi, int64_t gf_group_bits,
1386 double group_error, int gf_arf_bits) {
1387 RATE_CONTROL *const rc = &cpi->rc;
1388 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1389 TWO_PASS *twopass = &cpi->twopass;
1390 FIRSTPASS_STATS frame_stats;
1392 int frame_index = 1;
1393 int target_frame_size;
1395 const int max_bits = frame_max_bits(&cpi->rc, &cpi->oxcf);
1396 int64_t total_group_bits = gf_group_bits;
1397 double modified_err = 0.0;
1398 double err_fraction;
1399 int mid_boost_bits = 0;
1401 unsigned char arf_buffer_indices[MAX_ACTIVE_ARFS];
1403 key_frame = cpi->common.frame_type == KEY_FRAME ||
1404 vp9_is_upper_layer_key_frame(cpi);
1406 get_arf_buffer_indices(arf_buffer_indices);
1408 // For key frames the frame target rate is already set and it
1409 // is also the golden frame.
1411 if (rc->source_alt_ref_active) {
1412 twopass->gf_group.update_type[0] = OVERLAY_UPDATE;
1413 twopass->gf_group.rf_level[0] = INTER_NORMAL;
1414 twopass->gf_group.bit_allocation[0] = 0;
1415 twopass->gf_group.arf_update_idx[0] = arf_buffer_indices[0];
1416 twopass->gf_group.arf_ref_idx[0] = arf_buffer_indices[0];
1418 twopass->gf_group.update_type[0] = GF_UPDATE;
1419 twopass->gf_group.rf_level[0] = GF_ARF_STD;
1420 twopass->gf_group.bit_allocation[0] = gf_arf_bits;
1421 twopass->gf_group.arf_update_idx[0] = arf_buffer_indices[0];
1422 twopass->gf_group.arf_ref_idx[0] = arf_buffer_indices[0];
1425 // Step over the golden frame / overlay frame
1426 if (EOF == input_stats(twopass, &frame_stats))
1430 // Deduct the boost bits for arf (or gf if it is not a key frame)
1431 // from the group total.
1432 if (rc->source_alt_ref_pending || !key_frame)
1433 total_group_bits -= gf_arf_bits;
1435 // Store the bits to spend on the ARF if there is one.
1436 if (rc->source_alt_ref_pending) {
1437 twopass->gf_group.update_type[frame_index] = ARF_UPDATE;
1438 twopass->gf_group.rf_level[frame_index] = GF_ARF_STD;
1439 twopass->gf_group.bit_allocation[frame_index] = gf_arf_bits;
1440 twopass->gf_group.arf_src_offset[frame_index] =
1441 (unsigned char)(rc->baseline_gf_interval - 1);
1442 twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[0];
1443 twopass->gf_group.arf_ref_idx[frame_index] =
1444 arf_buffer_indices[cpi->multi_arf_last_grp_enabled &&
1445 rc->source_alt_ref_active];
1448 if (cpi->multi_arf_enabled) {
1449 // Set aside a slot for a level 1 arf.
1450 twopass->gf_group.update_type[frame_index] = ARF_UPDATE;
1451 twopass->gf_group.rf_level[frame_index] = GF_ARF_LOW;
1452 twopass->gf_group.arf_src_offset[frame_index] =
1453 (unsigned char)((rc->baseline_gf_interval >> 1) - 1);
1454 twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[1];
1455 twopass->gf_group.arf_ref_idx[frame_index] = arf_buffer_indices[0];
1460 // Define middle frame
1461 mid_frame_idx = frame_index + (rc->baseline_gf_interval >> 1) - 1;
1463 // Allocate bits to the other frames in the group.
1464 for (i = 0; i < rc->baseline_gf_interval - 1; ++i) {
1466 if (EOF == input_stats(twopass, &frame_stats))
1469 modified_err = calculate_modified_err(twopass, oxcf, &frame_stats);
1471 if (group_error > 0)
1472 err_fraction = modified_err / DOUBLE_DIVIDE_CHECK(group_error);
1476 target_frame_size = (int)((double)total_group_bits * err_fraction);
1478 if (rc->source_alt_ref_pending && cpi->multi_arf_enabled) {
1479 mid_boost_bits += (target_frame_size >> 4);
1480 target_frame_size -= (target_frame_size >> 4);
1482 if (frame_index <= mid_frame_idx)
1485 twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[arf_idx];
1486 twopass->gf_group.arf_ref_idx[frame_index] = arf_buffer_indices[arf_idx];
1488 target_frame_size = clamp(target_frame_size, 0,
1489 MIN(max_bits, (int)total_group_bits));
1491 twopass->gf_group.update_type[frame_index] = LF_UPDATE;
1492 twopass->gf_group.rf_level[frame_index] = INTER_NORMAL;
1494 twopass->gf_group.bit_allocation[frame_index] = target_frame_size;
1499 // We need to configure the frame at the end of the sequence + 1 that will be
1500 // the start frame for the next group. Otherwise prior to the call to
1501 // vp9_rc_get_second_pass_params() the data will be undefined.
1502 twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[0];
1503 twopass->gf_group.arf_ref_idx[frame_index] = arf_buffer_indices[0];
1505 if (rc->source_alt_ref_pending) {
1506 twopass->gf_group.update_type[frame_index] = OVERLAY_UPDATE;
1507 twopass->gf_group.rf_level[frame_index] = INTER_NORMAL;
1509 // Final setup for second arf and its overlay.
1510 if (cpi->multi_arf_enabled) {
1511 twopass->gf_group.bit_allocation[2] =
1512 twopass->gf_group.bit_allocation[mid_frame_idx] + mid_boost_bits;
1513 twopass->gf_group.update_type[mid_frame_idx] = OVERLAY_UPDATE;
1514 twopass->gf_group.bit_allocation[mid_frame_idx] = 0;
1517 twopass->gf_group.update_type[frame_index] = GF_UPDATE;
1518 twopass->gf_group.rf_level[frame_index] = GF_ARF_STD;
1521 // Note whether multi-arf was enabled this group for next time.
1522 cpi->multi_arf_last_grp_enabled = cpi->multi_arf_enabled;
1525 // Analyse and define a gf/arf group.
1526 static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
1527 RATE_CONTROL *const rc = &cpi->rc;
1528 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1529 TWO_PASS *const twopass = &cpi->twopass;
1530 FIRSTPASS_STATS next_frame;
1531 const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
1534 double boost_score = 0.0;
1535 double old_boost_score = 0.0;
1536 double gf_group_err = 0.0;
1537 double gf_first_frame_err = 0.0;
1538 double mod_frame_err = 0.0;
1540 double mv_ratio_accumulator = 0.0;
1541 double decay_accumulator = 1.0;
1542 double zero_motion_accumulator = 1.0;
1544 double loop_decay_rate = 1.00;
1545 double last_loop_decay_rate = 1.00;
1547 double this_frame_mv_in_out = 0.0;
1548 double mv_in_out_accumulator = 0.0;
1549 double abs_mv_in_out_accumulator = 0.0;
1550 double mv_ratio_accumulator_thresh;
1551 unsigned int allow_alt_ref = is_altref_enabled(cpi);
1556 int active_max_gf_interval;
1557 int64_t gf_group_bits;
1558 double gf_group_error_left;
1561 // Reset the GF group data structures unless this is a key
1562 // frame in which case it will already have been done.
1563 if (cpi->common.frame_type != KEY_FRAME) {
1564 vp9_zero(twopass->gf_group);
1567 vp9_clear_system_state();
1568 vp9_zero(next_frame);
1572 // Load stats for the current frame.
1573 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1575 // Note the error of the frame at the start of the group. This will be
1576 // the GF frame error if we code a normal gf.
1577 gf_first_frame_err = mod_frame_err;
1579 // If this is a key frame or the overlay from a previous arf then
1580 // the error score / cost of this frame has already been accounted for.
1581 if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
1582 gf_group_err -= gf_first_frame_err;
1584 // Motion breakout threshold for loop below depends on image size.
1585 mv_ratio_accumulator_thresh = (cpi->common.width + cpi->common.height) / 10.0;
1587 // Work out a maximum interval for the GF group.
1588 // If the image appears almost completely static we can extend beyond this.
1589 if (cpi->multi_arf_allowed) {
1590 active_max_gf_interval = rc->max_gf_interval;
1592 // The value chosen depends on the active Q range. At low Q we have
1593 // bits to spare and are better with a smaller interval and smaller boost.
1594 // At high Q when there are few bits to spare we are better with a longer
1595 // interval to spread the cost of the GF.
1596 active_max_gf_interval =
1597 12 + ((int)vp9_convert_qindex_to_q(rc->last_q[INTER_FRAME]) >> 5);
1599 if (active_max_gf_interval > rc->max_gf_interval)
1600 active_max_gf_interval = rc->max_gf_interval;
1604 while (i < rc->static_scene_max_gf_interval && i < rc->frames_to_key) {
1607 // Accumulate error score of frames in this gf group.
1608 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1609 gf_group_err += mod_frame_err;
1611 if (EOF == input_stats(twopass, &next_frame))
1614 // Test for the case where there is a brief flash but the prediction
1615 // quality back to an earlier frame is then restored.
1616 flash_detected = detect_flash(twopass, 0);
1618 // Update the motion related elements to the boost calculation.
1619 accumulate_frame_motion_stats(&next_frame,
1620 &this_frame_mv_in_out, &mv_in_out_accumulator,
1621 &abs_mv_in_out_accumulator,
1622 &mv_ratio_accumulator);
1624 // Accumulate the effect of prediction quality decay.
1625 if (!flash_detected) {
1626 last_loop_decay_rate = loop_decay_rate;
1627 loop_decay_rate = get_prediction_decay_rate(&cpi->common, &next_frame);
1628 decay_accumulator = decay_accumulator * loop_decay_rate;
1630 // Monitor for static sections.
1631 if ((next_frame.pcnt_inter - next_frame.pcnt_motion) <
1632 zero_motion_accumulator) {
1633 zero_motion_accumulator = next_frame.pcnt_inter -
1634 next_frame.pcnt_motion;
1637 // Break clause to detect very still sections after motion. For example,
1638 // a static image after a fade or other transition.
1639 if (detect_transition_to_still(twopass, i, 5, loop_decay_rate,
1640 last_loop_decay_rate)) {
1646 // Calculate a boost number for this frame.
1647 boost_score += decay_accumulator * calc_frame_boost(twopass, &next_frame,
1648 this_frame_mv_in_out);
1650 // Break out conditions.
1652 // Break at active_max_gf_interval unless almost totally static.
1653 (i >= active_max_gf_interval && (zero_motion_accumulator < 0.995)) ||
1655 // Don't break out with a very short interval.
1656 (i > MIN_GF_INTERVAL) &&
1657 ((boost_score > 125.0) || (next_frame.pcnt_inter < 0.75)) &&
1658 (!flash_detected) &&
1659 ((mv_ratio_accumulator > mv_ratio_accumulator_thresh) ||
1660 (abs_mv_in_out_accumulator > 3.0) ||
1661 (mv_in_out_accumulator < -2.0) ||
1662 ((boost_score - old_boost_score) < IIFACTOR)))) {
1663 boost_score = old_boost_score;
1667 *this_frame = next_frame;
1669 old_boost_score = boost_score;
1672 twopass->gf_zeromotion_pct = (int)(zero_motion_accumulator * 1000.0);
1674 // Don't allow a gf too near the next kf.
1675 if ((rc->frames_to_key - i) < MIN_GF_INTERVAL) {
1676 while (i < (rc->frames_to_key + !rc->next_key_frame_forced)) {
1679 if (EOF == input_stats(twopass, this_frame))
1682 if (i < rc->frames_to_key) {
1683 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1684 gf_group_err += mod_frame_err;
1689 // Set the interval until the next gf.
1690 if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
1691 rc->baseline_gf_interval = i - 1;
1693 rc->baseline_gf_interval = i;
1695 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
1697 // Should we use the alternate reference frame.
1698 if (allow_alt_ref &&
1699 (i < cpi->oxcf.lag_in_frames) &&
1700 (i >= MIN_GF_INTERVAL) &&
1701 // For real scene cuts (not forced kfs) don't allow arf very near kf.
1702 (rc->next_key_frame_forced ||
1703 (i <= (rc->frames_to_key - MIN_GF_INTERVAL)))) {
1704 // Calculate the boost for alt ref.
1705 rc->gfu_boost = calc_arf_boost(cpi, 0, (i - 1), (i - 1), &f_boost,
1707 rc->source_alt_ref_pending = 1;
1709 // Test to see if multi arf is appropriate.
1710 cpi->multi_arf_enabled =
1711 (cpi->multi_arf_allowed && (rc->baseline_gf_interval >= 6) &&
1712 (zero_motion_accumulator < 0.995)) ? 1 : 0;
1714 rc->gfu_boost = (int)boost_score;
1715 rc->source_alt_ref_pending = 0;
1718 // Reset the file position.
1719 reset_fpf_position(twopass, start_pos);
1721 // Calculate the bits to be allocated to the gf/arf group as a whole
1722 gf_group_bits = calculate_total_gf_group_bits(cpi, gf_group_err);
1724 // Calculate the extra bits to be used for boosted frame(s)
1726 int q = rc->last_q[INTER_FRAME];
1727 int boost = (rc->gfu_boost * gfboost_qadjust(q)) / 100;
1729 // Set max and minimum boost and hence minimum allocation.
1730 boost = clamp(boost, 125, (rc->baseline_gf_interval + 1) * 200);
1732 // Calculate the extra bits to be used for boosted frame(s)
1733 gf_arf_bits = calculate_boost_bits(rc->baseline_gf_interval,
1734 boost, gf_group_bits);
1737 // Adjust KF group bits and error remaining.
1738 twopass->kf_group_error_left -= (int64_t)gf_group_err;
1740 // If this is an arf update we want to remove the score for the overlay
1741 // frame at the end which will usually be very cheap to code.
1742 // The overlay frame has already, in effect, been coded so we want to spread
1743 // the remaining bits among the other frames.
1744 // For normal GFs remove the score for the GF itself unless this is
1745 // also a key frame in which case it has already been accounted for.
1746 if (rc->source_alt_ref_pending) {
1747 gf_group_error_left = gf_group_err - mod_frame_err;
1748 } else if (cpi->common.frame_type != KEY_FRAME) {
1749 gf_group_error_left = gf_group_err - gf_first_frame_err;
1751 gf_group_error_left = gf_group_err;
1754 // Allocate bits to each of the frames in the GF group.
1755 allocate_gf_group_bits(cpi, gf_group_bits, gf_group_error_left, gf_arf_bits);
1757 // Reset the file position.
1758 reset_fpf_position(twopass, start_pos);
1760 // Calculate a section intra ratio used in setting max loop filter.
1761 if (cpi->common.frame_type != KEY_FRAME) {
1762 twopass->section_intra_rating =
1763 calculate_section_intra_ratio(start_pos, twopass->stats_in_end,
1764 rc->baseline_gf_interval);
1768 static int test_candidate_kf(TWO_PASS *twopass,
1769 const FIRSTPASS_STATS *last_frame,
1770 const FIRSTPASS_STATS *this_frame,
1771 const FIRSTPASS_STATS *next_frame) {
1772 int is_viable_kf = 0;
1774 // Does the frame satisfy the primary criteria of a key frame?
1775 // If so, then examine how well it predicts subsequent frames.
1776 if ((this_frame->pcnt_second_ref < 0.10) &&
1777 (next_frame->pcnt_second_ref < 0.10) &&
1778 ((this_frame->pcnt_inter < 0.05) ||
1779 (((this_frame->pcnt_inter - this_frame->pcnt_neutral) < 0.35) &&
1780 ((this_frame->intra_error /
1781 DOUBLE_DIVIDE_CHECK(this_frame->coded_error)) < 2.5) &&
1782 ((fabs(last_frame->coded_error - this_frame->coded_error) /
1783 DOUBLE_DIVIDE_CHECK(this_frame->coded_error) > 0.40) ||
1784 (fabs(last_frame->intra_error - this_frame->intra_error) /
1785 DOUBLE_DIVIDE_CHECK(this_frame->intra_error) > 0.40) ||
1786 ((next_frame->intra_error /
1787 DOUBLE_DIVIDE_CHECK(next_frame->coded_error)) > 3.5))))) {
1789 const FIRSTPASS_STATS *start_pos = twopass->stats_in;
1790 FIRSTPASS_STATS local_next_frame = *next_frame;
1791 double boost_score = 0.0;
1792 double old_boost_score = 0.0;
1793 double decay_accumulator = 1.0;
1795 // Examine how well the key frame predicts subsequent frames.
1796 for (i = 0; i < 16; ++i) {
1797 double next_iiratio = (IIKFACTOR1 * local_next_frame.intra_error /
1798 DOUBLE_DIVIDE_CHECK(local_next_frame.coded_error));
1800 if (next_iiratio > RMAX)
1801 next_iiratio = RMAX;
1803 // Cumulative effect of decay in prediction quality.
1804 if (local_next_frame.pcnt_inter > 0.85)
1805 decay_accumulator *= local_next_frame.pcnt_inter;
1807 decay_accumulator *= (0.85 + local_next_frame.pcnt_inter) / 2.0;
1809 // Keep a running total.
1810 boost_score += (decay_accumulator * next_iiratio);
1812 // Test various breakout clauses.
1813 if ((local_next_frame.pcnt_inter < 0.05) ||
1814 (next_iiratio < 1.5) ||
1815 (((local_next_frame.pcnt_inter -
1816 local_next_frame.pcnt_neutral) < 0.20) &&
1817 (next_iiratio < 3.0)) ||
1818 ((boost_score - old_boost_score) < 3.0) ||
1819 (local_next_frame.intra_error < 200)) {
1823 old_boost_score = boost_score;
1825 // Get the next frame details
1826 if (EOF == input_stats(twopass, &local_next_frame))
1830 // If there is tolerable prediction for at least the next 3 frames then
1831 // break out else discard this potential key frame and move on
1832 if (boost_score > 30.0 && (i > 3)) {
1835 // Reset the file position
1836 reset_fpf_position(twopass, start_pos);
1842 return is_viable_kf;
1845 static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
1847 RATE_CONTROL *const rc = &cpi->rc;
1848 TWO_PASS *const twopass = &cpi->twopass;
1849 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1850 const FIRSTPASS_STATS first_frame = *this_frame;
1851 const FIRSTPASS_STATS *const start_position = twopass->stats_in;
1852 FIRSTPASS_STATS next_frame;
1853 FIRSTPASS_STATS last_frame;
1855 double decay_accumulator = 1.0;
1856 double zero_motion_accumulator = 1.0;
1857 double boost_score = 0.0;
1858 double kf_mod_err = 0.0;
1859 double kf_group_err = 0.0;
1860 double recent_loop_decay[8] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
1862 vp9_zero(next_frame);
1864 cpi->common.frame_type = KEY_FRAME;
1866 // Reset the GF group data structures.
1867 vp9_zero(twopass->gf_group);
1869 // Is this a forced key frame by interval.
1870 rc->this_key_frame_forced = rc->next_key_frame_forced;
1872 // Clear the alt ref active flag and last group multi arf flags as they
1873 // can never be set for a key frame.
1874 rc->source_alt_ref_active = 0;
1875 cpi->multi_arf_last_grp_enabled = 0;
1877 // KF is always a GF so clear frames till next gf counter.
1878 rc->frames_till_gf_update_due = 0;
1880 rc->frames_to_key = 1;
1882 twopass->kf_group_bits = 0; // Total bits available to kf group
1883 twopass->kf_group_error_left = 0; // Group modified error score.
1885 kf_mod_err = calculate_modified_err(twopass, oxcf, this_frame);
1887 // Find the next keyframe.
1889 while (twopass->stats_in < twopass->stats_in_end &&
1890 rc->frames_to_key < cpi->oxcf.key_freq) {
1891 // Accumulate kf group error.
1892 kf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
1894 // Load the next frame's stats.
1895 last_frame = *this_frame;
1896 input_stats(twopass, this_frame);
1898 // Provided that we are not at the end of the file...
1899 if (cpi->oxcf.auto_key &&
1900 lookup_next_frame_stats(twopass, &next_frame) != EOF) {
1901 double loop_decay_rate;
1903 // Check for a scene cut.
1904 if (test_candidate_kf(twopass, &last_frame, this_frame, &next_frame))
1907 // How fast is the prediction quality decaying?
1908 loop_decay_rate = get_prediction_decay_rate(&cpi->common, &next_frame);
1910 // We want to know something about the recent past... rather than
1911 // as used elsewhere where we are concerned with decay in prediction
1912 // quality since the last GF or KF.
1913 recent_loop_decay[i % 8] = loop_decay_rate;
1914 decay_accumulator = 1.0;
1915 for (j = 0; j < 8; ++j)
1916 decay_accumulator *= recent_loop_decay[j];
1918 // Special check for transition or high motion followed by a
1920 if (detect_transition_to_still(twopass, i, cpi->oxcf.key_freq - i,
1921 loop_decay_rate, decay_accumulator))
1924 // Step on to the next frame.
1925 ++rc->frames_to_key;
1927 // If we don't have a real key frame within the next two
1928 // key_freq intervals then break out of the loop.
1929 if (rc->frames_to_key >= 2 * cpi->oxcf.key_freq)
1932 ++rc->frames_to_key;
1937 // If there is a max kf interval set by the user we must obey it.
1938 // We already breakout of the loop above at 2x max.
1939 // This code centers the extra kf if the actual natural interval
1940 // is between 1x and 2x.
1941 if (cpi->oxcf.auto_key &&
1942 rc->frames_to_key > cpi->oxcf.key_freq) {
1943 FIRSTPASS_STATS tmp_frame = first_frame;
1945 rc->frames_to_key /= 2;
1947 // Reset to the start of the group.
1948 reset_fpf_position(twopass, start_position);
1952 // Rescan to get the correct error data for the forced kf group.
1953 for (i = 0; i < rc->frames_to_key; ++i) {
1954 kf_group_err += calculate_modified_err(twopass, oxcf, &tmp_frame);
1955 input_stats(twopass, &tmp_frame);
1957 rc->next_key_frame_forced = 1;
1958 } else if (twopass->stats_in == twopass->stats_in_end ||
1959 rc->frames_to_key >= cpi->oxcf.key_freq) {
1960 rc->next_key_frame_forced = 1;
1962 rc->next_key_frame_forced = 0;
1965 // Special case for the last key frame of the file.
1966 if (twopass->stats_in >= twopass->stats_in_end) {
1967 // Accumulate kf group error.
1968 kf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
1971 // Calculate the number of bits that should be assigned to the kf group.
1972 if (twopass->bits_left > 0 && twopass->modified_error_left > 0.0) {
1973 // Maximum number of bits for a single normal frame (not key frame).
1974 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
1976 // Maximum number of bits allocated to the key frame group.
1977 int64_t max_grp_bits;
1979 // Default allocation based on bits left and relative
1980 // complexity of the section.
1981 twopass->kf_group_bits = (int64_t)(twopass->bits_left *
1982 (kf_group_err / twopass->modified_error_left));
1984 // Clip based on maximum per frame rate defined by the user.
1985 max_grp_bits = (int64_t)max_bits * (int64_t)rc->frames_to_key;
1986 if (twopass->kf_group_bits > max_grp_bits)
1987 twopass->kf_group_bits = max_grp_bits;
1989 twopass->kf_group_bits = 0;
1991 twopass->kf_group_bits = MAX(0, twopass->kf_group_bits);
1993 // Reset the first pass file position.
1994 reset_fpf_position(twopass, start_position);
1996 // Scan through the kf group collating various stats used to deteermine
1997 // how many bits to spend on it.
1998 decay_accumulator = 1.0;
2000 for (i = 0; i < rc->frames_to_key; ++i) {
2001 if (EOF == input_stats(twopass, &next_frame))
2004 // Monitor for static sections.
2005 if ((next_frame.pcnt_inter - next_frame.pcnt_motion) <
2006 zero_motion_accumulator) {
2007 zero_motion_accumulator = (next_frame.pcnt_inter -
2008 next_frame.pcnt_motion);
2011 // For the first few frames collect data to decide kf boost.
2012 if (i <= (rc->max_gf_interval * 2)) {
2014 if (next_frame.intra_error > twopass->kf_intra_err_min)
2015 r = (IIKFACTOR2 * next_frame.intra_error /
2016 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
2018 r = (IIKFACTOR2 * twopass->kf_intra_err_min /
2019 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
2024 // How fast is prediction quality decaying.
2025 if (!detect_flash(twopass, 0)) {
2026 const double loop_decay_rate = get_prediction_decay_rate(&cpi->common,
2028 decay_accumulator *= loop_decay_rate;
2029 decay_accumulator = MAX(decay_accumulator, MIN_DECAY_FACTOR);
2032 boost_score += (decay_accumulator * r);
2036 reset_fpf_position(twopass, start_position);
2038 // Store the zero motion percentage
2039 twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
2041 // Calculate a section intra ratio used in setting max loop filter.
2042 twopass->section_intra_rating =
2043 calculate_section_intra_ratio(start_position, twopass->stats_in_end,
2046 // Work out how many bits to allocate for the key frame itself.
2047 rc->kf_boost = (int)boost_score;
2049 if (rc->kf_boost < (rc->frames_to_key * 3))
2050 rc->kf_boost = (rc->frames_to_key * 3);
2051 if (rc->kf_boost < MIN_KF_BOOST)
2052 rc->kf_boost = MIN_KF_BOOST;
2054 kf_bits = calculate_boost_bits((rc->frames_to_key - 1),
2055 rc->kf_boost, twopass->kf_group_bits);
2057 twopass->kf_group_bits -= kf_bits;
2059 // Save the bits to spend on the key frame.
2060 twopass->gf_group.bit_allocation[0] = kf_bits;
2061 twopass->gf_group.update_type[0] = KF_UPDATE;
2062 twopass->gf_group.rf_level[0] = KF_STD;
2064 // Note the total error score of the kf group minus the key frame itself.
2065 twopass->kf_group_error_left = (int)(kf_group_err - kf_mod_err);
2067 // Adjust the count of total modified error left.
2068 // The count of bits left is adjusted elsewhere based on real coded frame
2070 twopass->modified_error_left -= kf_group_err;
2073 // For VBR...adjustment to the frame target based on error from previous frames
2074 void vbr_rate_correction(int * this_frame_target,
2075 const int64_t vbr_bits_off_target) {
2076 int max_delta = (*this_frame_target * 15) / 100;
2078 // vbr_bits_off_target > 0 means we have extra bits to spend
2079 if (vbr_bits_off_target > 0) {
2080 *this_frame_target +=
2081 (vbr_bits_off_target > max_delta) ? max_delta
2082 : (int)vbr_bits_off_target;
2084 *this_frame_target -=
2085 (vbr_bits_off_target < -max_delta) ? max_delta
2086 : (int)-vbr_bits_off_target;
2090 // Define the reference buffers that will be updated post encode.
2091 void configure_buffer_updates(VP9_COMP *cpi) {
2092 TWO_PASS *const twopass = &cpi->twopass;
2094 cpi->rc.is_src_frame_alt_ref = 0;
2095 switch (twopass->gf_group.update_type[twopass->gf_group.index]) {
2097 cpi->refresh_last_frame = 1;
2098 cpi->refresh_golden_frame = 1;
2099 cpi->refresh_alt_ref_frame = 1;
2102 cpi->refresh_last_frame = 1;
2103 cpi->refresh_golden_frame = 0;
2104 cpi->refresh_alt_ref_frame = 0;
2107 cpi->refresh_last_frame = 1;
2108 cpi->refresh_golden_frame = 1;
2109 cpi->refresh_alt_ref_frame = 0;
2111 case OVERLAY_UPDATE:
2112 cpi->refresh_last_frame = 0;
2113 cpi->refresh_golden_frame = 1;
2114 cpi->refresh_alt_ref_frame = 0;
2115 cpi->rc.is_src_frame_alt_ref = 1;
2118 cpi->refresh_last_frame = 0;
2119 cpi->refresh_golden_frame = 0;
2120 cpi->refresh_alt_ref_frame = 1;
2125 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
2126 cpi->refresh_golden_frame = 0;
2127 if (cpi->alt_ref_source == NULL)
2128 cpi->refresh_alt_ref_frame = 0;
2133 void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
2134 VP9_COMMON *const cm = &cpi->common;
2135 RATE_CONTROL *const rc = &cpi->rc;
2136 TWO_PASS *const twopass = &cpi->twopass;
2138 FIRSTPASS_STATS this_frame;
2139 FIRSTPASS_STATS this_frame_copy;
2142 LAYER_CONTEXT *lc = NULL;
2143 const int is_spatial_svc = (cpi->use_svc &&
2144 cpi->svc.number_temporal_layers == 1);
2145 if (is_spatial_svc) {
2146 lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
2147 frames_left = (int)(twopass->total_stats.count -
2148 lc->current_video_frame_in_layer);
2150 frames_left = (int)(twopass->total_stats.count -
2151 cm->current_video_frame);
2154 if (!twopass->stats_in)
2157 // If this is an arf frame then we dont want to read the stats file or
2158 // advance the input pointer as we already have what we need.
2159 if (twopass->gf_group.update_type[twopass->gf_group.index] == ARF_UPDATE) {
2161 configure_buffer_updates(cpi);
2162 target_rate = twopass->gf_group.bit_allocation[twopass->gf_group.index];
2163 target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
2164 rc->base_frame_target = target_rate;
2165 #ifdef LONG_TERM_VBR_CORRECTION
2166 // Correction to rate target based on prior over or under shoot.
2167 if (cpi->oxcf.rc_mode == VPX_VBR)
2168 vbr_rate_correction(&target_rate, rc->vbr_bits_off_target);
2170 vp9_rc_set_frame_target(cpi, target_rate);
2171 cm->frame_type = INTER_FRAME;
2173 if (is_spatial_svc) {
2174 if (cpi->svc.spatial_layer_id == 0) {
2175 lc->is_key_frame = 0;
2177 lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
2179 if (lc->is_key_frame)
2180 cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
2187 vp9_clear_system_state();
2189 if (is_spatial_svc && twopass->kf_intra_err_min == 0) {
2190 twopass->kf_intra_err_min = KF_MB_INTRA_MIN * cpi->common.MBs;
2191 twopass->gf_intra_err_min = GF_MB_INTRA_MIN * cpi->common.MBs;
2194 if (cpi->oxcf.rc_mode == VPX_Q) {
2195 twopass->active_worst_quality = cpi->oxcf.cq_level;
2196 } else if (cm->current_video_frame == 0 ||
2197 (is_spatial_svc && lc->current_video_frame_in_layer == 0)) {
2198 // Special case code for first frame.
2199 const int section_target_bandwidth = (int)(twopass->bits_left /
2201 const int tmp_q = get_twopass_worst_quality(cpi, &twopass->total_left_stats,
2202 section_target_bandwidth);
2203 twopass->active_worst_quality = tmp_q;
2204 rc->ni_av_qi = tmp_q;
2205 rc->avg_q = vp9_convert_qindex_to_q(tmp_q);
2207 vp9_zero(this_frame);
2208 if (EOF == input_stats(twopass, &this_frame))
2211 // Local copy of the current frame's first pass stats.
2212 this_frame_copy = this_frame;
2214 // Keyframe and section processing.
2215 if (rc->frames_to_key == 0 ||
2216 (cpi->frame_flags & FRAMEFLAGS_KEY)) {
2217 // Define next KF group and assign bits to it.
2218 find_next_key_frame(cpi, &this_frame_copy);
2220 cm->frame_type = INTER_FRAME;
2223 if (is_spatial_svc) {
2224 if (cpi->svc.spatial_layer_id == 0) {
2225 lc->is_key_frame = (cm->frame_type == KEY_FRAME);
2227 cm->frame_type = INTER_FRAME;
2228 lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
2230 if (lc->is_key_frame) {
2231 cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
2236 // Define a new GF/ARF group. (Should always enter here for key frames).
2237 if (rc->frames_till_gf_update_due == 0) {
2238 define_gf_group(cpi, &this_frame_copy);
2240 if (twopass->gf_zeromotion_pct > 995) {
2241 // As long as max_thresh for encode breakout is small enough, it is ok
2242 // to enable it for show frame, i.e. set allow_encode_breakout to
2243 // ENCODE_BREAKOUT_LIMITED.
2244 if (!cm->show_frame)
2245 cpi->allow_encode_breakout = ENCODE_BREAKOUT_DISABLED;
2247 cpi->allow_encode_breakout = ENCODE_BREAKOUT_LIMITED;
2250 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
2251 if (!is_spatial_svc)
2252 cpi->refresh_golden_frame = 1;
2255 configure_buffer_updates(cpi);
2257 target_rate = twopass->gf_group.bit_allocation[twopass->gf_group.index];
2258 if (cpi->common.frame_type == KEY_FRAME)
2259 target_rate = vp9_rc_clamp_iframe_target_size(cpi, target_rate);
2261 target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
2263 rc->base_frame_target = target_rate;
2264 #ifdef LONG_TERM_VBR_CORRECTION
2265 // Correction to rate target based on prior over or under shoot.
2266 if (cpi->oxcf.rc_mode == VPX_VBR)
2267 vbr_rate_correction(&target_rate, rc->vbr_bits_off_target);
2269 vp9_rc_set_frame_target(cpi, target_rate);
2271 // Update the total stats remaining structure.
2272 subtract_stats(&twopass->total_left_stats, &this_frame);
2275 void vp9_twopass_postencode_update(VP9_COMP *cpi) {
2276 TWO_PASS *const twopass = &cpi->twopass;
2277 RATE_CONTROL *const rc = &cpi->rc;
2278 #ifdef LONG_TERM_VBR_CORRECTION
2279 // In this experimental mode, the VBR correction is done exclusively through
2280 // rc->vbr_bits_off_target. Based on the sign of this value, a limited %
2281 // adjustment is made to the target rate of subsequent frames, to try and
2282 // push it back towards 0. This mode is less likely to suffer from
2283 // extreme behaviour at the end of a clip or group of frames.
2284 const int bits_used = rc->base_frame_target;
2285 rc->vbr_bits_off_target += rc->base_frame_target - rc->projected_frame_size;
2287 // In this mode, VBR correction is acheived by altering bits_left,
2288 // kf_group_bits & gf_group_bits to reflect any deviation from the target
2289 // rate in this frame. This alters the allocation of bits to the
2290 // remaning frames in the group / clip.
2292 // This method can give rise to unstable behaviour near the end of a clip
2293 // or kf/gf group of frames where any accumulated error is corrected over an
2294 // ever decreasing number of frames. Hence we change the balance of target
2295 // vs. actual bitrate gradually as we progress towards the end of the
2296 // sequence in order to mitigate this effect.
2297 const double progress =
2298 (double)(twopass->stats_in - twopass->stats_in_start) /
2299 (twopass->stats_in_end - twopass->stats_in_start);
2300 const int bits_used = (int)(progress * rc->this_frame_target +
2301 (1.0 - progress) * rc->projected_frame_size);
2304 twopass->bits_left = MAX(twopass->bits_left - bits_used, 0);
2306 #ifdef LONG_TERM_VBR_CORRECTION
2307 if (cpi->common.frame_type != KEY_FRAME &&
2308 !vp9_is_upper_layer_key_frame(cpi)) {
2310 if (cpi->common.frame_type == KEY_FRAME ||
2311 vp9_is_upper_layer_key_frame(cpi)) {
2312 // For key frames kf_group_bits already had the target bits subtracted out.
2313 // So now update to the correct value based on the actual bits used.
2314 twopass->kf_group_bits += rc->this_frame_target - bits_used;
2317 twopass->kf_group_bits -= bits_used;
2319 twopass->kf_group_bits = MAX(twopass->kf_group_bits, 0);
2321 // Increment the gf group index ready for the next frame.
2322 ++twopass->gf_group.index;