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);
556 #if CONFIG_FP_MB_STATS
557 const int mb_index = mb_row * cm->mb_cols + mb_col;
560 vp9_clear_system_state();
562 xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
563 xd->plane[1].dst.buf = new_yv12->u_buffer + recon_uvoffset;
564 xd->plane[2].dst.buf = new_yv12->v_buffer + recon_uvoffset;
565 xd->left_available = (mb_col != 0);
566 xd->mi[0]->mbmi.sb_type = bsize;
567 xd->mi[0]->mbmi.ref_frame[0] = INTRA_FRAME;
568 set_mi_row_col(xd, &tile,
569 mb_row << 1, num_8x8_blocks_high_lookup[bsize],
570 mb_col << 1, num_8x8_blocks_wide_lookup[bsize],
571 cm->mi_rows, cm->mi_cols);
573 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
574 const int energy = vp9_block_energy(cpi, x, bsize);
575 error_weight = vp9_vaq_inv_q_ratio(energy);
578 // Do intra 16x16 prediction.
580 xd->mi[0]->mbmi.mode = DC_PRED;
581 xd->mi[0]->mbmi.tx_size = use_dc_pred ?
582 (bsize >= BLOCK_16X16 ? TX_16X16 : TX_8X8) : TX_4X4;
583 vp9_encode_intra_block_plane(x, bsize, 0);
584 this_error = vp9_get_mb_ss(x->plane[0].src_diff);
586 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
587 vp9_clear_system_state();
588 this_error = (int)(this_error * error_weight);
591 // Intrapenalty below deals with situations where the intra and inter
592 // error scores are very low (e.g. a plain black frame).
593 // We do not have special cases in first pass for 0,0 and nearest etc so
594 // all inter modes carry an overhead cost estimate for the mv.
595 // When the error score is very low this causes us to pick all or lots of
596 // INTRA modes and throw lots of key frames.
597 // This penalty adds a cost matching that of a 0,0 mv to the intra case.
598 this_error += intrapenalty;
600 // Accumulate the intra error.
601 intra_error += (int64_t)this_error;
603 #if CONFIG_FP_MB_STATS
604 if (cpi->use_fp_mb_stats) {
606 cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
610 // Set up limit values for motion vectors to prevent them extending
611 // outside the UMV borders.
612 x->mv_col_min = -((mb_col * 16) + BORDER_MV_PIXELS_B16);
613 x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + BORDER_MV_PIXELS_B16;
615 // Other than for the first frame do a motion search.
616 if (cm->current_video_frame > 0) {
617 int tmp_err, motion_error, raw_motion_error;
619 struct buf_2d unscaled_last_source_buf_2d;
621 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
622 motion_error = get_prediction_error(bsize, &x->plane[0].src,
623 &xd->plane[0].pre[0]);
624 // Assume 0,0 motion with no mv overhead.
625 mv.as_int = tmp_mv.as_int = 0;
627 // Compute the motion error of the 0,0 motion using the last source
628 // frame as the reference. Skip the further motion search on
629 // reconstructed frame if this error is small.
630 unscaled_last_source_buf_2d.buf =
631 cpi->unscaled_last_source->y_buffer + recon_yoffset;
632 unscaled_last_source_buf_2d.stride =
633 cpi->unscaled_last_source->y_stride;
634 raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
635 &unscaled_last_source_buf_2d);
637 // TODO(pengchong): Replace the hard-coded threshold
638 if (raw_motion_error > 25 ||
639 (cpi->use_svc && cpi->svc.number_temporal_layers == 1)) {
640 // Test last reference frame using the previous best mv as the
641 // starting point (best reference) for the search.
642 first_pass_motion_search(cpi, x, &best_ref_mv.as_mv, &mv.as_mv,
644 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
645 vp9_clear_system_state();
646 motion_error = (int)(motion_error * error_weight);
649 // If the current best reference mv is not centered on 0,0 then do a
650 // 0,0 based search as well.
651 if (best_ref_mv.as_int) {
653 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv, &tmp_err);
654 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
655 vp9_clear_system_state();
656 tmp_err = (int)(tmp_err * error_weight);
659 if (tmp_err < motion_error) {
660 motion_error = tmp_err;
661 mv.as_int = tmp_mv.as_int;
665 // Search in an older reference frame.
666 if (cm->current_video_frame > 1 && gld_yv12 != NULL) {
667 // Assume 0,0 motion with no mv overhead.
670 xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
671 gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
672 &xd->plane[0].pre[0]);
674 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv,
676 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
677 vp9_clear_system_state();
678 gf_motion_error = (int)(gf_motion_error * error_weight);
681 if (gf_motion_error < motion_error && gf_motion_error < this_error)
684 // Reset to last frame as reference buffer.
685 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
686 xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
687 xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
689 // In accumulating a score for the older reference frame take the
690 // best of the motion predicted score and the intra coded error
691 // (just as will be done for) accumulation of "coded_error" for
693 if (gf_motion_error < this_error)
694 sr_coded_error += gf_motion_error;
696 sr_coded_error += this_error;
698 sr_coded_error += motion_error;
701 sr_coded_error += motion_error;
704 // Start by assuming that intra mode is best.
705 best_ref_mv.as_int = 0;
707 #if CONFIG_FP_MB_STATS
708 if (cpi->use_fp_mb_stats) {
709 // intra predication statistics
710 cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
711 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_DCINTRA_MASK;
712 cpi->twopass.frame_mb_stats_buf[mb_index] &=
713 ~FPMB_NONZERO_MOTION_MASK;
714 if (this_error > FPMB_ERROR_LEVEL4_TH) {
715 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_LEVEL4_MASK;
716 } else if (this_error > FPMB_ERROR_LEVEL3_TH) {
717 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_LEVEL3_MASK;
718 } else if (this_error > FPMB_ERROR_LEVEL2_TH) {
719 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_LEVEL2_MASK;
720 } else if (this_error > FPMB_ERROR_LEVEL1_TH) {
721 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_LEVEL1_MASK;
723 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_LEVEL0_MASK;
728 if (motion_error <= this_error) {
729 // Keep a count of cases where the inter and intra were very close
730 // and very low. This helps with scene cut detection for example in
731 // cropped clips with black bars at the sides or top and bottom.
732 if (((this_error - intrapenalty) * 9 <= motion_error * 10) &&
733 this_error < 2 * intrapenalty)
738 this_error = motion_error;
739 xd->mi[0]->mbmi.mode = NEWMV;
740 xd->mi[0]->mbmi.mv[0] = mv;
741 xd->mi[0]->mbmi.tx_size = TX_4X4;
742 xd->mi[0]->mbmi.ref_frame[0] = LAST_FRAME;
743 xd->mi[0]->mbmi.ref_frame[1] = NONE;
744 vp9_build_inter_predictors_sby(xd, mb_row << 1, mb_col << 1, bsize);
745 vp9_encode_sby_pass1(x, bsize);
746 sum_mvr += mv.as_mv.row;
747 sum_mvr_abs += abs(mv.as_mv.row);
748 sum_mvc += mv.as_mv.col;
749 sum_mvc_abs += abs(mv.as_mv.col);
750 sum_mvrs += mv.as_mv.row * mv.as_mv.row;
751 sum_mvcs += mv.as_mv.col * mv.as_mv.col;
754 best_ref_mv.as_int = mv.as_int;
756 #if CONFIG_FP_MB_STATS
757 if (cpi->use_fp_mb_stats) {
758 // inter predication statistics
759 cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
760 cpi->twopass.frame_mb_stats_buf[mb_index] &= ~FPMB_DCINTRA_MASK;
761 cpi->twopass.frame_mb_stats_buf[mb_index] &=
762 ~FPMB_NONZERO_MOTION_MASK;
763 if (this_error > FPMB_ERROR_LEVEL4_TH) {
764 cpi->twopass.frame_mb_stats_buf[mb_index] |=
765 FPMB_ERROR_LEVEL4_MASK;
766 } else if (this_error > FPMB_ERROR_LEVEL3_TH) {
767 cpi->twopass.frame_mb_stats_buf[mb_index] |=
768 FPMB_ERROR_LEVEL3_MASK;
769 } else if (this_error > FPMB_ERROR_LEVEL2_TH) {
770 cpi->twopass.frame_mb_stats_buf[mb_index] |=
771 FPMB_ERROR_LEVEL2_MASK;
772 } else if (this_error > FPMB_ERROR_LEVEL1_TH) {
773 cpi->twopass.frame_mb_stats_buf[mb_index] |=
774 FPMB_ERROR_LEVEL1_MASK;
776 cpi->twopass.frame_mb_stats_buf[mb_index] |=
777 FPMB_ERROR_LEVEL0_MASK;
785 #if CONFIG_FP_MB_STATS
786 if (cpi->use_fp_mb_stats) {
787 cpi->twopass.frame_mb_stats_buf[mb_index] |=
788 FPMB_NONZERO_MOTION_MASK;
792 // Non-zero vector, was it different from the last non zero vector?
793 if (mv.as_int != lastmv_as_int)
795 lastmv_as_int = mv.as_int;
797 // Does the row vector point inwards or outwards?
798 if (mb_row < cm->mb_rows / 2) {
799 if (mv.as_mv.row > 0)
801 else if (mv.as_mv.row < 0)
803 } else if (mb_row > cm->mb_rows / 2) {
804 if (mv.as_mv.row > 0)
806 else if (mv.as_mv.row < 0)
810 // Does the col vector point inwards or outwards?
811 if (mb_col < cm->mb_cols / 2) {
812 if (mv.as_mv.col > 0)
814 else if (mv.as_mv.col < 0)
816 } else if (mb_col > cm->mb_cols / 2) {
817 if (mv.as_mv.col > 0)
819 else if (mv.as_mv.col < 0)
825 sr_coded_error += (int64_t)this_error;
827 coded_error += (int64_t)this_error;
829 // Adjust to the next column of MBs.
830 x->plane[0].src.buf += 16;
831 x->plane[1].src.buf += uv_mb_height;
832 x->plane[2].src.buf += uv_mb_height;
835 recon_uvoffset += uv_mb_height;
838 // Adjust to the next row of MBs.
839 x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
840 x->plane[1].src.buf += uv_mb_height * x->plane[1].src.stride -
841 uv_mb_height * cm->mb_cols;
842 x->plane[2].src.buf += uv_mb_height * x->plane[1].src.stride -
843 uv_mb_height * cm->mb_cols;
845 vp9_clear_system_state();
848 vp9_clear_system_state();
852 fps.frame = cm->current_video_frame;
853 fps.spatial_layer_id = cpi->svc.spatial_layer_id;
854 fps.intra_error = (double)(intra_error >> 8);
855 fps.coded_error = (double)(coded_error >> 8);
856 fps.sr_coded_error = (double)(sr_coded_error >> 8);
858 fps.pcnt_inter = (double)intercount / cm->MBs;
859 fps.pcnt_second_ref = (double)second_ref_count / cm->MBs;
860 fps.pcnt_neutral = (double)neutral_count / cm->MBs;
863 fps.MVr = (double)sum_mvr / mvcount;
864 fps.mvr_abs = (double)sum_mvr_abs / mvcount;
865 fps.MVc = (double)sum_mvc / mvcount;
866 fps.mvc_abs = (double)sum_mvc_abs / mvcount;
867 fps.MVrv = ((double)sum_mvrs - (fps.MVr * fps.MVr / mvcount)) / mvcount;
868 fps.MVcv = ((double)sum_mvcs - (fps.MVc * fps.MVc / mvcount)) / mvcount;
869 fps.mv_in_out_count = (double)sum_in_vectors / (mvcount * 2);
870 fps.new_mv_count = new_mv_count;
871 fps.pcnt_motion = (double)mvcount / cm->MBs;
879 fps.mv_in_out_count = 0.0;
880 fps.new_mv_count = 0.0;
881 fps.pcnt_motion = 0.0;
884 // TODO(paulwilkins): Handle the case when duration is set to 0, or
885 // something less than the full time between subsequent values of
886 // cpi->source_time_stamp.
887 fps.duration = (double)(cpi->source->ts_end - cpi->source->ts_start);
889 // Don't want to do output stats with a stack variable!
890 twopass->this_frame_stats = fps;
891 output_stats(&twopass->this_frame_stats, cpi->output_pkt_list);
892 accumulate_stats(&twopass->total_stats, &fps);
894 #if CONFIG_FP_MB_STATS
895 if (cpi->use_fp_mb_stats) {
896 output_fpmb_stats(twopass->frame_mb_stats_buf, cm, cpi->output_pkt_list);
901 // Copy the previous Last Frame back into gf and and arf buffers if
902 // the prediction is good enough... but also don't allow it to lag too far.
903 if ((twopass->sr_update_lag > 3) ||
904 ((cm->current_video_frame > 0) &&
905 (twopass->this_frame_stats.pcnt_inter > 0.20) &&
906 ((twopass->this_frame_stats.intra_error /
907 DOUBLE_DIVIDE_CHECK(twopass->this_frame_stats.coded_error)) > 2.0))) {
908 if (gld_yv12 != NULL) {
909 vp8_yv12_copy_frame(lst_yv12, gld_yv12);
911 twopass->sr_update_lag = 1;
913 ++twopass->sr_update_lag;
916 vp9_extend_frame_borders(new_yv12);
918 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
919 vp9_update_reference_frames(cpi);
921 // Swap frame pointers so last frame refers to the frame we just compressed.
922 swap_yv12(lst_yv12, new_yv12);
925 // Special case for the first frame. Copy into the GF buffer as a second
927 if (cm->current_video_frame == 0 && gld_yv12 != NULL) {
928 vp8_yv12_copy_frame(lst_yv12, gld_yv12);
931 // Use this to see what the first pass reconstruction looks like.
935 snprintf(filename, sizeof(filename), "enc%04d.yuv",
936 (int)cm->current_video_frame);
938 if (cm->current_video_frame == 0)
939 recon_file = fopen(filename, "wb");
941 recon_file = fopen(filename, "ab");
943 (void)fwrite(lst_yv12->buffer_alloc, lst_yv12->frame_size, 1, recon_file);
947 ++cm->current_video_frame;
949 vp9_inc_frame_in_layer(&cpi->svc);
952 static double calc_correction_factor(double err_per_mb,
957 const double error_term = err_per_mb / err_divisor;
959 // Adjustment based on actual quantizer to power term.
960 const double power_term = MIN(vp9_convert_qindex_to_q(q) * 0.0125 + pt_low,
963 // Calculate correction factor.
964 if (power_term < 1.0)
965 assert(error_term >= 0.0);
967 return fclamp(pow(error_term, power_term), 0.05, 5.0);
970 static int get_twopass_worst_quality(const VP9_COMP *cpi,
971 const FIRSTPASS_STATS *stats,
972 int section_target_bandwidth) {
973 const RATE_CONTROL *const rc = &cpi->rc;
974 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
976 if (section_target_bandwidth <= 0) {
977 return rc->worst_quality; // Highest value allowed
979 const int num_mbs = cpi->common.MBs;
980 const double section_err = stats->coded_error / stats->count;
981 const double err_per_mb = section_err / num_mbs;
982 const double speed_term = 1.0 + 0.04 * oxcf->speed;
983 const int target_norm_bits_per_mb = ((uint64_t)section_target_bandwidth <<
984 BPER_MB_NORMBITS) / num_mbs;
986 int is_svc_upper_layer = 0;
987 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1 &&
988 cpi->svc.spatial_layer_id > 0) {
989 is_svc_upper_layer = 1;
992 // Try and pick a max Q that will be high enough to encode the
993 // content at the given rate.
994 for (q = rc->best_quality; q < rc->worst_quality; ++q) {
995 const double factor =
996 calc_correction_factor(err_per_mb, ERR_DIVISOR,
997 is_svc_upper_layer ? SVC_FACTOR_PT_LOW :
998 FACTOR_PT_LOW, FACTOR_PT_HIGH, q);
999 const int bits_per_mb = vp9_rc_bits_per_mb(INTER_FRAME, q,
1000 factor * speed_term);
1001 if (bits_per_mb <= target_norm_bits_per_mb)
1005 // Restriction on active max q for constrained quality mode.
1006 if (cpi->oxcf.rc_mode == VPX_CQ)
1007 q = MAX(q, oxcf->cq_level);
1012 extern void vp9_new_framerate(VP9_COMP *cpi, double framerate);
1014 void vp9_init_second_pass(VP9_COMP *cpi) {
1015 SVC *const svc = &cpi->svc;
1016 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1017 const int is_spatial_svc = (svc->number_spatial_layers > 1) &&
1018 (svc->number_temporal_layers == 1);
1019 TWO_PASS *const twopass = is_spatial_svc ?
1020 &svc->layer_context[svc->spatial_layer_id].twopass : &cpi->twopass;
1022 FIRSTPASS_STATS *stats;
1024 zero_stats(&twopass->total_stats);
1025 zero_stats(&twopass->total_left_stats);
1027 if (!twopass->stats_in_end)
1030 stats = &twopass->total_stats;
1032 *stats = *twopass->stats_in_end;
1033 twopass->total_left_stats = *stats;
1035 frame_rate = 10000000.0 * stats->count / stats->duration;
1036 // Each frame can have a different duration, as the frame rate in the source
1037 // isn't guaranteed to be constant. The frame rate prior to the first frame
1038 // encoded in the second pass is a guess. However, the sum duration is not.
1039 // It is calculated based on the actual durations of all frames from the
1042 if (is_spatial_svc) {
1043 vp9_update_spatial_layer_framerate(cpi, frame_rate);
1044 twopass->bits_left = (int64_t)(stats->duration *
1045 svc->layer_context[svc->spatial_layer_id].target_bandwidth /
1048 vp9_new_framerate(cpi, frame_rate);
1049 twopass->bits_left = (int64_t)(stats->duration * oxcf->target_bandwidth /
1053 // Calculate a minimum intra value to be used in determining the IIratio
1054 // scores used in the second pass. We have this minimum to make sure
1055 // that clips that are static but "low complexity" in the intra domain
1056 // are still boosted appropriately for KF/GF/ARF.
1057 if (!is_spatial_svc) {
1058 // We don't know the number of MBs for each layer at this point.
1059 // So we will do it later.
1060 twopass->kf_intra_err_min = KF_MB_INTRA_MIN * cpi->common.MBs;
1061 twopass->gf_intra_err_min = GF_MB_INTRA_MIN * cpi->common.MBs;
1064 // This variable monitors how far behind the second ref update is lagging.
1065 twopass->sr_update_lag = 1;
1067 // Scan the first pass file and calculate a modified total error based upon
1068 // the bias/power function used to allocate bits.
1070 const double avg_error = stats->coded_error /
1071 DOUBLE_DIVIDE_CHECK(stats->count);
1072 const FIRSTPASS_STATS *s = twopass->stats_in;
1073 double modified_error_total = 0.0;
1074 twopass->modified_error_min = (avg_error *
1075 oxcf->two_pass_vbrmin_section) / 100;
1076 twopass->modified_error_max = (avg_error *
1077 oxcf->two_pass_vbrmax_section) / 100;
1078 while (s < twopass->stats_in_end) {
1079 modified_error_total += calculate_modified_err(twopass, oxcf, s);
1082 twopass->modified_error_left = modified_error_total;
1085 // Reset the vbr bits off target counter
1086 cpi->rc.vbr_bits_off_target = 0;
1089 // This function gives an estimate of how badly we believe the prediction
1090 // quality is decaying from frame to frame.
1091 static double get_prediction_decay_rate(const VP9_COMMON *cm,
1092 const FIRSTPASS_STATS *next_frame) {
1093 // Look at the observed drop in prediction quality between the last frame
1094 // and the GF buffer (which contains an older frame).
1095 const double mb_sr_err_diff = (next_frame->sr_coded_error -
1096 next_frame->coded_error) / cm->MBs;
1097 const double second_ref_decay = mb_sr_err_diff <= 512.0
1098 ? fclamp(pow(1.0 - (mb_sr_err_diff / 512.0), 0.5), 0.85, 1.0)
1101 return MIN(second_ref_decay, next_frame->pcnt_inter);
1104 // Function to test for a condition where a complex transition is followed
1105 // by a static section. For example in slide shows where there is a fade
1106 // between slides. This is to help with more optimal kf and gf positioning.
1107 static int detect_transition_to_still(TWO_PASS *twopass,
1108 int frame_interval, int still_interval,
1109 double loop_decay_rate,
1110 double last_decay_rate) {
1111 int trans_to_still = 0;
1113 // Break clause to detect very still sections after motion
1114 // For example a static image after a fade or other transition
1115 // instead of a clean scene cut.
1116 if (frame_interval > MIN_GF_INTERVAL &&
1117 loop_decay_rate >= 0.999 &&
1118 last_decay_rate < 0.9) {
1120 const FIRSTPASS_STATS *position = twopass->stats_in;
1121 FIRSTPASS_STATS tmp_next_frame;
1123 // Look ahead a few frames to see if static condition persists...
1124 for (j = 0; j < still_interval; ++j) {
1125 if (EOF == input_stats(twopass, &tmp_next_frame))
1128 if (tmp_next_frame.pcnt_inter - tmp_next_frame.pcnt_motion < 0.999)
1132 reset_fpf_position(twopass, position);
1134 // Only if it does do we signal a transition to still.
1135 if (j == still_interval)
1139 return trans_to_still;
1142 // This function detects a flash through the high relative pcnt_second_ref
1143 // score in the frame following a flash frame. The offset passed in should
1145 static int detect_flash(const TWO_PASS *twopass, int offset) {
1146 const FIRSTPASS_STATS *const next_frame = read_frame_stats(twopass, offset);
1148 // What we are looking for here is a situation where there is a
1149 // brief break in prediction (such as a flash) but subsequent frames
1150 // are reasonably well predicted by an earlier (pre flash) frame.
1151 // The recovery after a flash is indicated by a high pcnt_second_ref
1152 // compared to pcnt_inter.
1153 return next_frame != NULL &&
1154 next_frame->pcnt_second_ref > next_frame->pcnt_inter &&
1155 next_frame->pcnt_second_ref >= 0.5;
1158 // Update the motion related elements to the GF arf boost calculation.
1159 static void accumulate_frame_motion_stats(const FIRSTPASS_STATS *stats,
1161 double *mv_in_out_accumulator,
1162 double *abs_mv_in_out_accumulator,
1163 double *mv_ratio_accumulator) {
1164 const double pct = stats->pcnt_motion;
1166 // Accumulate Motion In/Out of frame stats.
1167 *mv_in_out = stats->mv_in_out_count * pct;
1168 *mv_in_out_accumulator += *mv_in_out;
1169 *abs_mv_in_out_accumulator += fabs(*mv_in_out);
1171 // Accumulate a measure of how uniform (or conversely how random) the motion
1172 // field is (a ratio of abs(mv) / mv).
1174 const double mvr_ratio = fabs(stats->mvr_abs) /
1175 DOUBLE_DIVIDE_CHECK(fabs(stats->MVr));
1176 const double mvc_ratio = fabs(stats->mvc_abs) /
1177 DOUBLE_DIVIDE_CHECK(fabs(stats->MVc));
1179 *mv_ratio_accumulator += pct * (mvr_ratio < stats->mvr_abs ?
1180 mvr_ratio : stats->mvr_abs);
1181 *mv_ratio_accumulator += pct * (mvc_ratio < stats->mvc_abs ?
1182 mvc_ratio : stats->mvc_abs);
1186 // Calculate a baseline boost number for the current frame.
1187 static double calc_frame_boost(const TWO_PASS *twopass,
1188 const FIRSTPASS_STATS *this_frame,
1189 double this_frame_mv_in_out) {
1192 // Underlying boost factor is based on inter intra error ratio.
1193 if (this_frame->intra_error > twopass->gf_intra_err_min)
1194 frame_boost = (IIFACTOR * this_frame->intra_error /
1195 DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
1197 frame_boost = (IIFACTOR * twopass->gf_intra_err_min /
1198 DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
1200 // Increase boost for frames where new data coming into frame (e.g. zoom out).
1201 // Slightly reduce boost if there is a net balance of motion out of the frame
1202 // (zoom in). The range for this_frame_mv_in_out is -1.0 to +1.0.
1203 if (this_frame_mv_in_out > 0.0)
1204 frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
1205 // In the extreme case the boost is halved.
1207 frame_boost += frame_boost * (this_frame_mv_in_out / 2.0);
1209 return MIN(frame_boost, GF_RMAX);
1212 static int calc_arf_boost(VP9_COMP *cpi, int offset,
1213 int f_frames, int b_frames,
1214 int *f_boost, int *b_boost) {
1215 TWO_PASS *const twopass = &cpi->twopass;
1217 double boost_score = 0.0;
1218 double mv_ratio_accumulator = 0.0;
1219 double decay_accumulator = 1.0;
1220 double this_frame_mv_in_out = 0.0;
1221 double mv_in_out_accumulator = 0.0;
1222 double abs_mv_in_out_accumulator = 0.0;
1224 int flash_detected = 0;
1226 // Search forward from the proposed arf/next gf position.
1227 for (i = 0; i < f_frames; ++i) {
1228 const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
1229 if (this_frame == NULL)
1232 // Update the motion related elements to the boost calculation.
1233 accumulate_frame_motion_stats(this_frame,
1234 &this_frame_mv_in_out, &mv_in_out_accumulator,
1235 &abs_mv_in_out_accumulator,
1236 &mv_ratio_accumulator);
1238 // We want to discount the flash frame itself and the recovery
1239 // frame that follows as both will have poor scores.
1240 flash_detected = detect_flash(twopass, i + offset) ||
1241 detect_flash(twopass, i + offset + 1);
1243 // Accumulate the effect of prediction quality decay.
1244 if (!flash_detected) {
1245 decay_accumulator *= get_prediction_decay_rate(&cpi->common, this_frame);
1246 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1247 ? MIN_DECAY_FACTOR : decay_accumulator;
1250 boost_score += decay_accumulator * calc_frame_boost(twopass, this_frame,
1251 this_frame_mv_in_out);
1254 *f_boost = (int)boost_score;
1256 // Reset for backward looking loop.
1258 mv_ratio_accumulator = 0.0;
1259 decay_accumulator = 1.0;
1260 this_frame_mv_in_out = 0.0;
1261 mv_in_out_accumulator = 0.0;
1262 abs_mv_in_out_accumulator = 0.0;
1264 // Search backward towards last gf position.
1265 for (i = -1; i >= -b_frames; --i) {
1266 const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
1267 if (this_frame == NULL)
1270 // Update the motion related elements to the boost calculation.
1271 accumulate_frame_motion_stats(this_frame,
1272 &this_frame_mv_in_out, &mv_in_out_accumulator,
1273 &abs_mv_in_out_accumulator,
1274 &mv_ratio_accumulator);
1276 // We want to discount the the flash frame itself and the recovery
1277 // frame that follows as both will have poor scores.
1278 flash_detected = detect_flash(twopass, i + offset) ||
1279 detect_flash(twopass, i + offset + 1);
1281 // Cumulative effect of prediction quality decay.
1282 if (!flash_detected) {
1283 decay_accumulator *= get_prediction_decay_rate(&cpi->common, this_frame);
1284 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1285 ? MIN_DECAY_FACTOR : decay_accumulator;
1288 boost_score += decay_accumulator * calc_frame_boost(twopass, this_frame,
1289 this_frame_mv_in_out);
1291 *b_boost = (int)boost_score;
1293 arf_boost = (*f_boost + *b_boost);
1294 if (arf_boost < ((b_frames + f_frames) * 20))
1295 arf_boost = ((b_frames + f_frames) * 20);
1300 // Calculate a section intra ratio used in setting max loop filter.
1301 static int calculate_section_intra_ratio(const FIRSTPASS_STATS *begin,
1302 const FIRSTPASS_STATS *end,
1303 int section_length) {
1304 const FIRSTPASS_STATS *s = begin;
1305 double intra_error = 0.0;
1306 double coded_error = 0.0;
1309 while (s < end && i < section_length) {
1310 intra_error += s->intra_error;
1311 coded_error += s->coded_error;
1316 return (int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
1319 // Calculate the total bits to allocate in this GF/ARF group.
1320 static int64_t calculate_total_gf_group_bits(VP9_COMP *cpi,
1321 double gf_group_err) {
1322 const RATE_CONTROL *const rc = &cpi->rc;
1323 const TWO_PASS *const twopass = &cpi->twopass;
1324 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
1325 int64_t total_group_bits;
1327 // Calculate the bits to be allocated to the group as a whole.
1328 if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0)) {
1329 total_group_bits = (int64_t)(twopass->kf_group_bits *
1330 (gf_group_err / twopass->kf_group_error_left));
1332 total_group_bits = 0;
1335 // Clamp odd edge cases.
1336 total_group_bits = (total_group_bits < 0) ?
1337 0 : (total_group_bits > twopass->kf_group_bits) ?
1338 twopass->kf_group_bits : total_group_bits;
1340 // Clip based on user supplied data rate variability limit.
1341 if (total_group_bits > (int64_t)max_bits * rc->baseline_gf_interval)
1342 total_group_bits = (int64_t)max_bits * rc->baseline_gf_interval;
1344 return total_group_bits;
1347 // Calculate the number bits extra to assign to boosted frames in a group.
1348 static int calculate_boost_bits(int frame_count,
1349 int boost, int64_t total_group_bits) {
1350 int allocation_chunks;
1352 // return 0 for invalid inputs (could arise e.g. through rounding errors)
1353 if (!boost || (total_group_bits <= 0) || (frame_count <= 0) )
1356 allocation_chunks = (frame_count * 100) + boost;
1358 // Prevent overflow.
1360 int divisor = boost >> 10;
1362 allocation_chunks /= divisor;
1365 // Calculate the number of extra bits for use in the boosted frame or frames.
1366 return MAX((int)(((int64_t)boost * total_group_bits) / allocation_chunks), 0);
1369 // Current limit on maximum number of active arfs in a GF/ARF group.
1370 #define MAX_ACTIVE_ARFS 2
1373 // This function indirects the choice of buffers for arfs.
1374 // At the moment the values are fixed but this may change as part of
1375 // the integration process with other codec features that swap buffers around.
1376 static void get_arf_buffer_indices(unsigned char *arf_buffer_indices) {
1377 arf_buffer_indices[0] = ARF_SLOT1;
1378 arf_buffer_indices[1] = ARF_SLOT2;
1381 static void allocate_gf_group_bits(VP9_COMP *cpi, int64_t gf_group_bits,
1382 double group_error, int gf_arf_bits) {
1383 RATE_CONTROL *const rc = &cpi->rc;
1384 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1385 TWO_PASS *twopass = &cpi->twopass;
1386 FIRSTPASS_STATS frame_stats;
1388 int frame_index = 1;
1389 int target_frame_size;
1391 const int max_bits = frame_max_bits(&cpi->rc, &cpi->oxcf);
1392 int64_t total_group_bits = gf_group_bits;
1393 double modified_err = 0.0;
1394 double err_fraction;
1395 int mid_boost_bits = 0;
1397 unsigned char arf_buffer_indices[MAX_ACTIVE_ARFS];
1399 key_frame = cpi->common.frame_type == KEY_FRAME ||
1400 vp9_is_upper_layer_key_frame(cpi);
1402 get_arf_buffer_indices(arf_buffer_indices);
1404 // For key frames the frame target rate is already set and it
1405 // is also the golden frame.
1407 if (rc->source_alt_ref_active) {
1408 twopass->gf_group.update_type[0] = OVERLAY_UPDATE;
1409 twopass->gf_group.rf_level[0] = INTER_NORMAL;
1410 twopass->gf_group.bit_allocation[0] = 0;
1411 twopass->gf_group.arf_update_idx[0] = arf_buffer_indices[0];
1412 twopass->gf_group.arf_ref_idx[0] = arf_buffer_indices[0];
1414 twopass->gf_group.update_type[0] = GF_UPDATE;
1415 twopass->gf_group.rf_level[0] = GF_ARF_STD;
1416 twopass->gf_group.bit_allocation[0] = gf_arf_bits;
1417 twopass->gf_group.arf_update_idx[0] = arf_buffer_indices[0];
1418 twopass->gf_group.arf_ref_idx[0] = arf_buffer_indices[0];
1421 // Step over the golden frame / overlay frame
1422 if (EOF == input_stats(twopass, &frame_stats))
1426 // Deduct the boost bits for arf (or gf if it is not a key frame)
1427 // from the group total.
1428 if (rc->source_alt_ref_pending || !key_frame)
1429 total_group_bits -= gf_arf_bits;
1431 // Store the bits to spend on the ARF if there is one.
1432 if (rc->source_alt_ref_pending) {
1433 twopass->gf_group.update_type[frame_index] = ARF_UPDATE;
1434 twopass->gf_group.rf_level[frame_index] = GF_ARF_STD;
1435 twopass->gf_group.bit_allocation[frame_index] = gf_arf_bits;
1436 twopass->gf_group.arf_src_offset[frame_index] =
1437 (unsigned char)(rc->baseline_gf_interval - 1);
1438 twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[0];
1439 twopass->gf_group.arf_ref_idx[frame_index] =
1440 arf_buffer_indices[cpi->multi_arf_last_grp_enabled &&
1441 rc->source_alt_ref_active];
1444 if (cpi->multi_arf_enabled) {
1445 // Set aside a slot for a level 1 arf.
1446 twopass->gf_group.update_type[frame_index] = ARF_UPDATE;
1447 twopass->gf_group.rf_level[frame_index] = GF_ARF_LOW;
1448 twopass->gf_group.arf_src_offset[frame_index] =
1449 (unsigned char)((rc->baseline_gf_interval >> 1) - 1);
1450 twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[1];
1451 twopass->gf_group.arf_ref_idx[frame_index] = arf_buffer_indices[0];
1456 // Define middle frame
1457 mid_frame_idx = frame_index + (rc->baseline_gf_interval >> 1) - 1;
1459 // Allocate bits to the other frames in the group.
1460 for (i = 0; i < rc->baseline_gf_interval - 1; ++i) {
1462 if (EOF == input_stats(twopass, &frame_stats))
1465 modified_err = calculate_modified_err(twopass, oxcf, &frame_stats);
1467 if (group_error > 0)
1468 err_fraction = modified_err / DOUBLE_DIVIDE_CHECK(group_error);
1472 target_frame_size = (int)((double)total_group_bits * err_fraction);
1474 if (rc->source_alt_ref_pending && cpi->multi_arf_enabled) {
1475 mid_boost_bits += (target_frame_size >> 4);
1476 target_frame_size -= (target_frame_size >> 4);
1478 if (frame_index <= mid_frame_idx)
1481 twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[arf_idx];
1482 twopass->gf_group.arf_ref_idx[frame_index] = arf_buffer_indices[arf_idx];
1484 target_frame_size = clamp(target_frame_size, 0,
1485 MIN(max_bits, (int)total_group_bits));
1487 twopass->gf_group.update_type[frame_index] = LF_UPDATE;
1488 twopass->gf_group.rf_level[frame_index] = INTER_NORMAL;
1490 twopass->gf_group.bit_allocation[frame_index] = target_frame_size;
1495 // We need to configure the frame at the end of the sequence + 1 that will be
1496 // the start frame for the next group. Otherwise prior to the call to
1497 // vp9_rc_get_second_pass_params() the data will be undefined.
1498 twopass->gf_group.arf_update_idx[frame_index] = arf_buffer_indices[0];
1499 twopass->gf_group.arf_ref_idx[frame_index] = arf_buffer_indices[0];
1501 if (rc->source_alt_ref_pending) {
1502 twopass->gf_group.update_type[frame_index] = OVERLAY_UPDATE;
1503 twopass->gf_group.rf_level[frame_index] = INTER_NORMAL;
1505 // Final setup for second arf and its overlay.
1506 if (cpi->multi_arf_enabled) {
1507 twopass->gf_group.bit_allocation[2] =
1508 twopass->gf_group.bit_allocation[mid_frame_idx] + mid_boost_bits;
1509 twopass->gf_group.update_type[mid_frame_idx] = OVERLAY_UPDATE;
1510 twopass->gf_group.bit_allocation[mid_frame_idx] = 0;
1513 twopass->gf_group.update_type[frame_index] = GF_UPDATE;
1514 twopass->gf_group.rf_level[frame_index] = GF_ARF_STD;
1517 // Note whether multi-arf was enabled this group for next time.
1518 cpi->multi_arf_last_grp_enabled = cpi->multi_arf_enabled;
1521 // Analyse and define a gf/arf group.
1522 static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
1523 RATE_CONTROL *const rc = &cpi->rc;
1524 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1525 TWO_PASS *const twopass = &cpi->twopass;
1526 FIRSTPASS_STATS next_frame;
1527 const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
1530 double boost_score = 0.0;
1531 double old_boost_score = 0.0;
1532 double gf_group_err = 0.0;
1533 double gf_first_frame_err = 0.0;
1534 double mod_frame_err = 0.0;
1536 double mv_ratio_accumulator = 0.0;
1537 double decay_accumulator = 1.0;
1538 double zero_motion_accumulator = 1.0;
1540 double loop_decay_rate = 1.00;
1541 double last_loop_decay_rate = 1.00;
1543 double this_frame_mv_in_out = 0.0;
1544 double mv_in_out_accumulator = 0.0;
1545 double abs_mv_in_out_accumulator = 0.0;
1546 double mv_ratio_accumulator_thresh;
1547 unsigned int allow_alt_ref = is_altref_enabled(cpi);
1552 int active_max_gf_interval;
1553 int64_t gf_group_bits;
1554 double gf_group_error_left;
1557 // Reset the GF group data structures unless this is a key
1558 // frame in which case it will already have been done.
1559 if (cpi->common.frame_type != KEY_FRAME) {
1560 vp9_zero(twopass->gf_group);
1563 vp9_clear_system_state();
1564 vp9_zero(next_frame);
1568 // Load stats for the current frame.
1569 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1571 // Note the error of the frame at the start of the group. This will be
1572 // the GF frame error if we code a normal gf.
1573 gf_first_frame_err = mod_frame_err;
1575 // If this is a key frame or the overlay from a previous arf then
1576 // the error score / cost of this frame has already been accounted for.
1577 if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
1578 gf_group_err -= gf_first_frame_err;
1580 // Motion breakout threshold for loop below depends on image size.
1581 mv_ratio_accumulator_thresh = (cpi->common.width + cpi->common.height) / 10.0;
1583 // Work out a maximum interval for the GF group.
1584 // If the image appears almost completely static we can extend beyond this.
1585 if (cpi->multi_arf_allowed) {
1586 active_max_gf_interval = rc->max_gf_interval;
1588 // The value chosen depends on the active Q range. At low Q we have
1589 // bits to spare and are better with a smaller interval and smaller boost.
1590 // At high Q when there are few bits to spare we are better with a longer
1591 // interval to spread the cost of the GF.
1592 active_max_gf_interval =
1593 12 + ((int)vp9_convert_qindex_to_q(rc->last_q[INTER_FRAME]) >> 5);
1595 if (active_max_gf_interval > rc->max_gf_interval)
1596 active_max_gf_interval = rc->max_gf_interval;
1600 while (i < rc->static_scene_max_gf_interval && i < rc->frames_to_key) {
1603 // Accumulate error score of frames in this gf group.
1604 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1605 gf_group_err += mod_frame_err;
1607 if (EOF == input_stats(twopass, &next_frame))
1610 // Test for the case where there is a brief flash but the prediction
1611 // quality back to an earlier frame is then restored.
1612 flash_detected = detect_flash(twopass, 0);
1614 // Update the motion related elements to the boost calculation.
1615 accumulate_frame_motion_stats(&next_frame,
1616 &this_frame_mv_in_out, &mv_in_out_accumulator,
1617 &abs_mv_in_out_accumulator,
1618 &mv_ratio_accumulator);
1620 // Accumulate the effect of prediction quality decay.
1621 if (!flash_detected) {
1622 last_loop_decay_rate = loop_decay_rate;
1623 loop_decay_rate = get_prediction_decay_rate(&cpi->common, &next_frame);
1624 decay_accumulator = decay_accumulator * loop_decay_rate;
1626 // Monitor for static sections.
1627 if ((next_frame.pcnt_inter - next_frame.pcnt_motion) <
1628 zero_motion_accumulator) {
1629 zero_motion_accumulator = next_frame.pcnt_inter -
1630 next_frame.pcnt_motion;
1633 // Break clause to detect very still sections after motion. For example,
1634 // a static image after a fade or other transition.
1635 if (detect_transition_to_still(twopass, i, 5, loop_decay_rate,
1636 last_loop_decay_rate)) {
1642 // Calculate a boost number for this frame.
1643 boost_score += decay_accumulator * calc_frame_boost(twopass, &next_frame,
1644 this_frame_mv_in_out);
1646 // Break out conditions.
1648 // Break at active_max_gf_interval unless almost totally static.
1649 (i >= active_max_gf_interval && (zero_motion_accumulator < 0.995)) ||
1651 // Don't break out with a very short interval.
1652 (i > MIN_GF_INTERVAL) &&
1653 ((boost_score > 125.0) || (next_frame.pcnt_inter < 0.75)) &&
1654 (!flash_detected) &&
1655 ((mv_ratio_accumulator > mv_ratio_accumulator_thresh) ||
1656 (abs_mv_in_out_accumulator > 3.0) ||
1657 (mv_in_out_accumulator < -2.0) ||
1658 ((boost_score - old_boost_score) < IIFACTOR)))) {
1659 boost_score = old_boost_score;
1663 *this_frame = next_frame;
1665 old_boost_score = boost_score;
1668 twopass->gf_zeromotion_pct = (int)(zero_motion_accumulator * 1000.0);
1670 // Don't allow a gf too near the next kf.
1671 if ((rc->frames_to_key - i) < MIN_GF_INTERVAL) {
1672 while (i < (rc->frames_to_key + !rc->next_key_frame_forced)) {
1675 if (EOF == input_stats(twopass, this_frame))
1678 if (i < rc->frames_to_key) {
1679 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1680 gf_group_err += mod_frame_err;
1685 // Set the interval until the next gf.
1686 if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
1687 rc->baseline_gf_interval = i - 1;
1689 rc->baseline_gf_interval = i;
1691 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
1693 // Should we use the alternate reference frame.
1694 if (allow_alt_ref &&
1695 (i < cpi->oxcf.lag_in_frames) &&
1696 (i >= MIN_GF_INTERVAL) &&
1697 // For real scene cuts (not forced kfs) don't allow arf very near kf.
1698 (rc->next_key_frame_forced ||
1699 (i <= (rc->frames_to_key - MIN_GF_INTERVAL)))) {
1700 // Calculate the boost for alt ref.
1701 rc->gfu_boost = calc_arf_boost(cpi, 0, (i - 1), (i - 1), &f_boost,
1703 rc->source_alt_ref_pending = 1;
1705 // Test to see if multi arf is appropriate.
1706 cpi->multi_arf_enabled =
1707 (cpi->multi_arf_allowed && (rc->baseline_gf_interval >= 6) &&
1708 (zero_motion_accumulator < 0.995)) ? 1 : 0;
1710 rc->gfu_boost = (int)boost_score;
1711 rc->source_alt_ref_pending = 0;
1714 // Reset the file position.
1715 reset_fpf_position(twopass, start_pos);
1717 // Calculate the bits to be allocated to the gf/arf group as a whole
1718 gf_group_bits = calculate_total_gf_group_bits(cpi, gf_group_err);
1720 // Calculate the extra bits to be used for boosted frame(s)
1722 int q = rc->last_q[INTER_FRAME];
1723 int boost = (rc->gfu_boost * gfboost_qadjust(q)) / 100;
1725 // Set max and minimum boost and hence minimum allocation.
1726 boost = clamp(boost, 125, (rc->baseline_gf_interval + 1) * 200);
1728 // Calculate the extra bits to be used for boosted frame(s)
1729 gf_arf_bits = calculate_boost_bits(rc->baseline_gf_interval,
1730 boost, gf_group_bits);
1733 // Adjust KF group bits and error remaining.
1734 twopass->kf_group_error_left -= (int64_t)gf_group_err;
1736 // If this is an arf update we want to remove the score for the overlay
1737 // frame at the end which will usually be very cheap to code.
1738 // The overlay frame has already, in effect, been coded so we want to spread
1739 // the remaining bits among the other frames.
1740 // For normal GFs remove the score for the GF itself unless this is
1741 // also a key frame in which case it has already been accounted for.
1742 if (rc->source_alt_ref_pending) {
1743 gf_group_error_left = gf_group_err - mod_frame_err;
1744 } else if (cpi->common.frame_type != KEY_FRAME) {
1745 gf_group_error_left = gf_group_err - gf_first_frame_err;
1747 gf_group_error_left = gf_group_err;
1750 // Allocate bits to each of the frames in the GF group.
1751 allocate_gf_group_bits(cpi, gf_group_bits, gf_group_error_left, gf_arf_bits);
1753 // Reset the file position.
1754 reset_fpf_position(twopass, start_pos);
1756 // Calculate a section intra ratio used in setting max loop filter.
1757 if (cpi->common.frame_type != KEY_FRAME) {
1758 twopass->section_intra_rating =
1759 calculate_section_intra_ratio(start_pos, twopass->stats_in_end,
1760 rc->baseline_gf_interval);
1764 static int test_candidate_kf(TWO_PASS *twopass,
1765 const FIRSTPASS_STATS *last_frame,
1766 const FIRSTPASS_STATS *this_frame,
1767 const FIRSTPASS_STATS *next_frame) {
1768 int is_viable_kf = 0;
1770 // Does the frame satisfy the primary criteria of a key frame?
1771 // If so, then examine how well it predicts subsequent frames.
1772 if ((this_frame->pcnt_second_ref < 0.10) &&
1773 (next_frame->pcnt_second_ref < 0.10) &&
1774 ((this_frame->pcnt_inter < 0.05) ||
1775 (((this_frame->pcnt_inter - this_frame->pcnt_neutral) < 0.35) &&
1776 ((this_frame->intra_error /
1777 DOUBLE_DIVIDE_CHECK(this_frame->coded_error)) < 2.5) &&
1778 ((fabs(last_frame->coded_error - this_frame->coded_error) /
1779 DOUBLE_DIVIDE_CHECK(this_frame->coded_error) > 0.40) ||
1780 (fabs(last_frame->intra_error - this_frame->intra_error) /
1781 DOUBLE_DIVIDE_CHECK(this_frame->intra_error) > 0.40) ||
1782 ((next_frame->intra_error /
1783 DOUBLE_DIVIDE_CHECK(next_frame->coded_error)) > 3.5))))) {
1785 const FIRSTPASS_STATS *start_pos = twopass->stats_in;
1786 FIRSTPASS_STATS local_next_frame = *next_frame;
1787 double boost_score = 0.0;
1788 double old_boost_score = 0.0;
1789 double decay_accumulator = 1.0;
1791 // Examine how well the key frame predicts subsequent frames.
1792 for (i = 0; i < 16; ++i) {
1793 double next_iiratio = (IIKFACTOR1 * local_next_frame.intra_error /
1794 DOUBLE_DIVIDE_CHECK(local_next_frame.coded_error));
1796 if (next_iiratio > RMAX)
1797 next_iiratio = RMAX;
1799 // Cumulative effect of decay in prediction quality.
1800 if (local_next_frame.pcnt_inter > 0.85)
1801 decay_accumulator *= local_next_frame.pcnt_inter;
1803 decay_accumulator *= (0.85 + local_next_frame.pcnt_inter) / 2.0;
1805 // Keep a running total.
1806 boost_score += (decay_accumulator * next_iiratio);
1808 // Test various breakout clauses.
1809 if ((local_next_frame.pcnt_inter < 0.05) ||
1810 (next_iiratio < 1.5) ||
1811 (((local_next_frame.pcnt_inter -
1812 local_next_frame.pcnt_neutral) < 0.20) &&
1813 (next_iiratio < 3.0)) ||
1814 ((boost_score - old_boost_score) < 3.0) ||
1815 (local_next_frame.intra_error < 200)) {
1819 old_boost_score = boost_score;
1821 // Get the next frame details
1822 if (EOF == input_stats(twopass, &local_next_frame))
1826 // If there is tolerable prediction for at least the next 3 frames then
1827 // break out else discard this potential key frame and move on
1828 if (boost_score > 30.0 && (i > 3)) {
1831 // Reset the file position
1832 reset_fpf_position(twopass, start_pos);
1838 return is_viable_kf;
1841 static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
1843 RATE_CONTROL *const rc = &cpi->rc;
1844 TWO_PASS *const twopass = &cpi->twopass;
1845 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1846 const FIRSTPASS_STATS first_frame = *this_frame;
1847 const FIRSTPASS_STATS *const start_position = twopass->stats_in;
1848 FIRSTPASS_STATS next_frame;
1849 FIRSTPASS_STATS last_frame;
1851 double decay_accumulator = 1.0;
1852 double zero_motion_accumulator = 1.0;
1853 double boost_score = 0.0;
1854 double kf_mod_err = 0.0;
1855 double kf_group_err = 0.0;
1856 double recent_loop_decay[8] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
1858 vp9_zero(next_frame);
1860 cpi->common.frame_type = KEY_FRAME;
1862 // Reset the GF group data structures.
1863 vp9_zero(twopass->gf_group);
1865 // Is this a forced key frame by interval.
1866 rc->this_key_frame_forced = rc->next_key_frame_forced;
1868 // Clear the alt ref active flag and last group multi arf flags as they
1869 // can never be set for a key frame.
1870 rc->source_alt_ref_active = 0;
1871 cpi->multi_arf_last_grp_enabled = 0;
1873 // KF is always a GF so clear frames till next gf counter.
1874 rc->frames_till_gf_update_due = 0;
1876 rc->frames_to_key = 1;
1878 twopass->kf_group_bits = 0; // Total bits available to kf group
1879 twopass->kf_group_error_left = 0; // Group modified error score.
1881 kf_mod_err = calculate_modified_err(twopass, oxcf, this_frame);
1883 // Find the next keyframe.
1885 while (twopass->stats_in < twopass->stats_in_end &&
1886 rc->frames_to_key < cpi->oxcf.key_freq) {
1887 // Accumulate kf group error.
1888 kf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
1890 // Load the next frame's stats.
1891 last_frame = *this_frame;
1892 input_stats(twopass, this_frame);
1894 // Provided that we are not at the end of the file...
1895 if (cpi->oxcf.auto_key &&
1896 lookup_next_frame_stats(twopass, &next_frame) != EOF) {
1897 double loop_decay_rate;
1899 // Check for a scene cut.
1900 if (test_candidate_kf(twopass, &last_frame, this_frame, &next_frame))
1903 // How fast is the prediction quality decaying?
1904 loop_decay_rate = get_prediction_decay_rate(&cpi->common, &next_frame);
1906 // We want to know something about the recent past... rather than
1907 // as used elsewhere where we are concerned with decay in prediction
1908 // quality since the last GF or KF.
1909 recent_loop_decay[i % 8] = loop_decay_rate;
1910 decay_accumulator = 1.0;
1911 for (j = 0; j < 8; ++j)
1912 decay_accumulator *= recent_loop_decay[j];
1914 // Special check for transition or high motion followed by a
1916 if (detect_transition_to_still(twopass, i, cpi->oxcf.key_freq - i,
1917 loop_decay_rate, decay_accumulator))
1920 // Step on to the next frame.
1921 ++rc->frames_to_key;
1923 // If we don't have a real key frame within the next two
1924 // key_freq intervals then break out of the loop.
1925 if (rc->frames_to_key >= 2 * cpi->oxcf.key_freq)
1928 ++rc->frames_to_key;
1933 // If there is a max kf interval set by the user we must obey it.
1934 // We already breakout of the loop above at 2x max.
1935 // This code centers the extra kf if the actual natural interval
1936 // is between 1x and 2x.
1937 if (cpi->oxcf.auto_key &&
1938 rc->frames_to_key > cpi->oxcf.key_freq) {
1939 FIRSTPASS_STATS tmp_frame = first_frame;
1941 rc->frames_to_key /= 2;
1943 // Reset to the start of the group.
1944 reset_fpf_position(twopass, start_position);
1948 // Rescan to get the correct error data for the forced kf group.
1949 for (i = 0; i < rc->frames_to_key; ++i) {
1950 kf_group_err += calculate_modified_err(twopass, oxcf, &tmp_frame);
1951 input_stats(twopass, &tmp_frame);
1953 rc->next_key_frame_forced = 1;
1954 } else if (twopass->stats_in == twopass->stats_in_end ||
1955 rc->frames_to_key >= cpi->oxcf.key_freq) {
1956 rc->next_key_frame_forced = 1;
1958 rc->next_key_frame_forced = 0;
1961 // Special case for the last key frame of the file.
1962 if (twopass->stats_in >= twopass->stats_in_end) {
1963 // Accumulate kf group error.
1964 kf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
1967 // Calculate the number of bits that should be assigned to the kf group.
1968 if (twopass->bits_left > 0 && twopass->modified_error_left > 0.0) {
1969 // Maximum number of bits for a single normal frame (not key frame).
1970 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
1972 // Maximum number of bits allocated to the key frame group.
1973 int64_t max_grp_bits;
1975 // Default allocation based on bits left and relative
1976 // complexity of the section.
1977 twopass->kf_group_bits = (int64_t)(twopass->bits_left *
1978 (kf_group_err / twopass->modified_error_left));
1980 // Clip based on maximum per frame rate defined by the user.
1981 max_grp_bits = (int64_t)max_bits * (int64_t)rc->frames_to_key;
1982 if (twopass->kf_group_bits > max_grp_bits)
1983 twopass->kf_group_bits = max_grp_bits;
1985 twopass->kf_group_bits = 0;
1987 twopass->kf_group_bits = MAX(0, twopass->kf_group_bits);
1989 // Reset the first pass file position.
1990 reset_fpf_position(twopass, start_position);
1992 // Scan through the kf group collating various stats used to deteermine
1993 // how many bits to spend on it.
1994 decay_accumulator = 1.0;
1996 for (i = 0; i < rc->frames_to_key; ++i) {
1997 if (EOF == input_stats(twopass, &next_frame))
2000 // Monitor for static sections.
2001 if ((next_frame.pcnt_inter - next_frame.pcnt_motion) <
2002 zero_motion_accumulator) {
2003 zero_motion_accumulator = (next_frame.pcnt_inter -
2004 next_frame.pcnt_motion);
2007 // For the first few frames collect data to decide kf boost.
2008 if (i <= (rc->max_gf_interval * 2)) {
2010 if (next_frame.intra_error > twopass->kf_intra_err_min)
2011 r = (IIKFACTOR2 * next_frame.intra_error /
2012 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
2014 r = (IIKFACTOR2 * twopass->kf_intra_err_min /
2015 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
2020 // How fast is prediction quality decaying.
2021 if (!detect_flash(twopass, 0)) {
2022 const double loop_decay_rate = get_prediction_decay_rate(&cpi->common,
2024 decay_accumulator *= loop_decay_rate;
2025 decay_accumulator = MAX(decay_accumulator, MIN_DECAY_FACTOR);
2028 boost_score += (decay_accumulator * r);
2032 reset_fpf_position(twopass, start_position);
2034 // Store the zero motion percentage
2035 twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
2037 // Calculate a section intra ratio used in setting max loop filter.
2038 twopass->section_intra_rating =
2039 calculate_section_intra_ratio(start_position, twopass->stats_in_end,
2042 // Work out how many bits to allocate for the key frame itself.
2043 rc->kf_boost = (int)boost_score;
2045 if (rc->kf_boost < (rc->frames_to_key * 3))
2046 rc->kf_boost = (rc->frames_to_key * 3);
2047 if (rc->kf_boost < MIN_KF_BOOST)
2048 rc->kf_boost = MIN_KF_BOOST;
2050 kf_bits = calculate_boost_bits((rc->frames_to_key - 1),
2051 rc->kf_boost, twopass->kf_group_bits);
2053 twopass->kf_group_bits -= kf_bits;
2055 // Save the bits to spend on the key frame.
2056 twopass->gf_group.bit_allocation[0] = kf_bits;
2057 twopass->gf_group.update_type[0] = KF_UPDATE;
2058 twopass->gf_group.rf_level[0] = KF_STD;
2060 // Note the total error score of the kf group minus the key frame itself.
2061 twopass->kf_group_error_left = (int)(kf_group_err - kf_mod_err);
2063 // Adjust the count of total modified error left.
2064 // The count of bits left is adjusted elsewhere based on real coded frame
2066 twopass->modified_error_left -= kf_group_err;
2069 // For VBR...adjustment to the frame target based on error from previous frames
2070 void vbr_rate_correction(int * this_frame_target,
2071 const int64_t vbr_bits_off_target) {
2072 int max_delta = (*this_frame_target * 15) / 100;
2074 // vbr_bits_off_target > 0 means we have extra bits to spend
2075 if (vbr_bits_off_target > 0) {
2076 *this_frame_target +=
2077 (vbr_bits_off_target > max_delta) ? max_delta
2078 : (int)vbr_bits_off_target;
2080 *this_frame_target -=
2081 (vbr_bits_off_target < -max_delta) ? max_delta
2082 : (int)-vbr_bits_off_target;
2086 // Define the reference buffers that will be updated post encode.
2087 void configure_buffer_updates(VP9_COMP *cpi) {
2088 TWO_PASS *const twopass = &cpi->twopass;
2090 cpi->rc.is_src_frame_alt_ref = 0;
2091 switch (twopass->gf_group.update_type[twopass->gf_group.index]) {
2093 cpi->refresh_last_frame = 1;
2094 cpi->refresh_golden_frame = 1;
2095 cpi->refresh_alt_ref_frame = 1;
2098 cpi->refresh_last_frame = 1;
2099 cpi->refresh_golden_frame = 0;
2100 cpi->refresh_alt_ref_frame = 0;
2103 cpi->refresh_last_frame = 1;
2104 cpi->refresh_golden_frame = 1;
2105 cpi->refresh_alt_ref_frame = 0;
2107 case OVERLAY_UPDATE:
2108 cpi->refresh_last_frame = 0;
2109 cpi->refresh_golden_frame = 1;
2110 cpi->refresh_alt_ref_frame = 0;
2111 cpi->rc.is_src_frame_alt_ref = 1;
2114 cpi->refresh_last_frame = 0;
2115 cpi->refresh_golden_frame = 0;
2116 cpi->refresh_alt_ref_frame = 1;
2121 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
2122 cpi->refresh_golden_frame = 0;
2123 if (cpi->alt_ref_source == NULL)
2124 cpi->refresh_alt_ref_frame = 0;
2129 void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
2130 VP9_COMMON *const cm = &cpi->common;
2131 RATE_CONTROL *const rc = &cpi->rc;
2132 TWO_PASS *const twopass = &cpi->twopass;
2134 FIRSTPASS_STATS this_frame;
2135 FIRSTPASS_STATS this_frame_copy;
2138 LAYER_CONTEXT *lc = NULL;
2139 const int is_spatial_svc = (cpi->use_svc &&
2140 cpi->svc.number_temporal_layers == 1);
2141 if (is_spatial_svc) {
2142 lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
2143 frames_left = (int)(twopass->total_stats.count -
2144 lc->current_video_frame_in_layer);
2146 frames_left = (int)(twopass->total_stats.count -
2147 cm->current_video_frame);
2150 if (!twopass->stats_in)
2153 // If this is an arf frame then we dont want to read the stats file or
2154 // advance the input pointer as we already have what we need.
2155 if (twopass->gf_group.update_type[twopass->gf_group.index] == ARF_UPDATE) {
2157 configure_buffer_updates(cpi);
2158 target_rate = twopass->gf_group.bit_allocation[twopass->gf_group.index];
2159 target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
2160 rc->base_frame_target = target_rate;
2161 #ifdef LONG_TERM_VBR_CORRECTION
2162 // Correction to rate target based on prior over or under shoot.
2163 if (cpi->oxcf.rc_mode == VPX_VBR)
2164 vbr_rate_correction(&target_rate, rc->vbr_bits_off_target);
2166 vp9_rc_set_frame_target(cpi, target_rate);
2167 cm->frame_type = INTER_FRAME;
2169 if (is_spatial_svc) {
2170 if (cpi->svc.spatial_layer_id == 0) {
2171 lc->is_key_frame = 0;
2173 lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
2175 if (lc->is_key_frame)
2176 cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
2183 vp9_clear_system_state();
2185 if (is_spatial_svc && twopass->kf_intra_err_min == 0) {
2186 twopass->kf_intra_err_min = KF_MB_INTRA_MIN * cpi->common.MBs;
2187 twopass->gf_intra_err_min = GF_MB_INTRA_MIN * cpi->common.MBs;
2190 if (cpi->oxcf.rc_mode == VPX_Q) {
2191 twopass->active_worst_quality = cpi->oxcf.cq_level;
2192 } else if (cm->current_video_frame == 0 ||
2193 (is_spatial_svc && lc->current_video_frame_in_layer == 0)) {
2194 // Special case code for first frame.
2195 const int section_target_bandwidth = (int)(twopass->bits_left /
2197 const int tmp_q = get_twopass_worst_quality(cpi, &twopass->total_left_stats,
2198 section_target_bandwidth);
2199 twopass->active_worst_quality = tmp_q;
2200 rc->ni_av_qi = tmp_q;
2201 rc->avg_q = vp9_convert_qindex_to_q(tmp_q);
2203 vp9_zero(this_frame);
2204 if (EOF == input_stats(twopass, &this_frame))
2207 // Local copy of the current frame's first pass stats.
2208 this_frame_copy = this_frame;
2210 // Keyframe and section processing.
2211 if (rc->frames_to_key == 0 ||
2212 (cpi->frame_flags & FRAMEFLAGS_KEY)) {
2213 // Define next KF group and assign bits to it.
2214 find_next_key_frame(cpi, &this_frame_copy);
2216 cm->frame_type = INTER_FRAME;
2219 if (is_spatial_svc) {
2220 if (cpi->svc.spatial_layer_id == 0) {
2221 lc->is_key_frame = (cm->frame_type == KEY_FRAME);
2223 cm->frame_type = INTER_FRAME;
2224 lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
2226 if (lc->is_key_frame) {
2227 cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
2232 // Define a new GF/ARF group. (Should always enter here for key frames).
2233 if (rc->frames_till_gf_update_due == 0) {
2234 define_gf_group(cpi, &this_frame_copy);
2236 if (twopass->gf_zeromotion_pct > 995) {
2237 // As long as max_thresh for encode breakout is small enough, it is ok
2238 // to enable it for show frame, i.e. set allow_encode_breakout to
2239 // ENCODE_BREAKOUT_LIMITED.
2240 if (!cm->show_frame)
2241 cpi->allow_encode_breakout = ENCODE_BREAKOUT_DISABLED;
2243 cpi->allow_encode_breakout = ENCODE_BREAKOUT_LIMITED;
2246 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
2247 if (!is_spatial_svc)
2248 cpi->refresh_golden_frame = 1;
2251 configure_buffer_updates(cpi);
2253 target_rate = twopass->gf_group.bit_allocation[twopass->gf_group.index];
2254 if (cpi->common.frame_type == KEY_FRAME)
2255 target_rate = vp9_rc_clamp_iframe_target_size(cpi, target_rate);
2257 target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
2259 rc->base_frame_target = target_rate;
2260 #ifdef LONG_TERM_VBR_CORRECTION
2261 // Correction to rate target based on prior over or under shoot.
2262 if (cpi->oxcf.rc_mode == VPX_VBR)
2263 vbr_rate_correction(&target_rate, rc->vbr_bits_off_target);
2265 vp9_rc_set_frame_target(cpi, target_rate);
2267 // Update the total stats remaining structure.
2268 subtract_stats(&twopass->total_left_stats, &this_frame);
2271 void vp9_twopass_postencode_update(VP9_COMP *cpi) {
2272 TWO_PASS *const twopass = &cpi->twopass;
2273 RATE_CONTROL *const rc = &cpi->rc;
2274 #ifdef LONG_TERM_VBR_CORRECTION
2275 // In this experimental mode, the VBR correction is done exclusively through
2276 // rc->vbr_bits_off_target. Based on the sign of this value, a limited %
2277 // adjustment is made to the target rate of subsequent frames, to try and
2278 // push it back towards 0. This mode is less likely to suffer from
2279 // extreme behaviour at the end of a clip or group of frames.
2280 const int bits_used = rc->base_frame_target;
2281 rc->vbr_bits_off_target += rc->base_frame_target - rc->projected_frame_size;
2283 // In this mode, VBR correction is acheived by altering bits_left,
2284 // kf_group_bits & gf_group_bits to reflect any deviation from the target
2285 // rate in this frame. This alters the allocation of bits to the
2286 // remaning frames in the group / clip.
2288 // This method can give rise to unstable behaviour near the end of a clip
2289 // or kf/gf group of frames where any accumulated error is corrected over an
2290 // ever decreasing number of frames. Hence we change the balance of target
2291 // vs. actual bitrate gradually as we progress towards the end of the
2292 // sequence in order to mitigate this effect.
2293 const double progress =
2294 (double)(twopass->stats_in - twopass->stats_in_start) /
2295 (twopass->stats_in_end - twopass->stats_in_start);
2296 const int bits_used = (int)(progress * rc->this_frame_target +
2297 (1.0 - progress) * rc->projected_frame_size);
2300 twopass->bits_left = MAX(twopass->bits_left - bits_used, 0);
2302 #ifdef LONG_TERM_VBR_CORRECTION
2303 if (cpi->common.frame_type != KEY_FRAME &&
2304 !vp9_is_upper_layer_key_frame(cpi)) {
2306 if (cpi->common.frame_type == KEY_FRAME ||
2307 vp9_is_upper_layer_key_frame(cpi)) {
2308 // For key frames kf_group_bits already had the target bits subtracted out.
2309 // So now update to the correct value based on the actual bits used.
2310 twopass->kf_group_bits += rc->this_frame_target - bits_used;
2313 twopass->kf_group_bits -= bits_used;
2315 twopass->kf_group_bits = MAX(twopass->kf_group_bits, 0);
2317 // Increment the gf group index ready for the next frame.
2318 ++twopass->gf_group.index;