2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
15 #include "./vpx_scale_rtcd.h"
17 #include "vpx_mem/vpx_mem.h"
18 #include "vpx_scale/vpx_scale.h"
19 #include "vpx_scale/yv12config.h"
21 #include "vp9/common/vp9_entropymv.h"
22 #include "vp9/common/vp9_quant_common.h"
23 #include "vp9/common/vp9_reconinter.h" // vp9_setup_dst_planes()
24 #include "vp9/common/vp9_systemdependent.h"
26 #include "vp9/encoder/vp9_aq_variance.h"
27 #include "vp9/encoder/vp9_block.h"
28 #include "vp9/encoder/vp9_encodeframe.h"
29 #include "vp9/encoder/vp9_encodemb.h"
30 #include "vp9/encoder/vp9_encodemv.h"
31 #include "vp9/encoder/vp9_encoder.h"
32 #include "vp9/encoder/vp9_extend.h"
33 #include "vp9/encoder/vp9_firstpass.h"
34 #include "vp9/encoder/vp9_mcomp.h"
35 #include "vp9/encoder/vp9_quantize.h"
36 #include "vp9/encoder/vp9_rdopt.h"
37 #include "vp9/encoder/vp9_variance.h"
42 #define IIKFACTOR1 12.5
43 #define IIKFACTOR2 15.0
46 #define ERR_DIVISOR 150.0
47 #define MIN_DECAY_FACTOR 0.1
48 #define SVC_FACTOR_PT_LOW 0.45
49 #define FACTOR_PT_LOW 0.5
50 #define FACTOR_PT_HIGH 0.9
52 #define KF_MB_INTRA_MIN 150
53 #define GF_MB_INTRA_MIN 100
55 #define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x) - 0.000001 : (x) + 0.000001)
57 #define MIN_KF_BOOST 300
58 #define MIN_GF_INTERVAL 4
59 #define LONG_TERM_VBR_CORRECTION
61 static void swap_yv12(YV12_BUFFER_CONFIG *a, YV12_BUFFER_CONFIG *b) {
62 YV12_BUFFER_CONFIG temp = *a;
67 static int gfboost_qadjust(int qindex) {
68 const double q = vp9_convert_qindex_to_q(qindex);
69 return (int)((0.00000828 * q * q * q) +
74 // Resets the first pass file to the given position using a relative seek from
75 // the current position.
76 static void reset_fpf_position(TWO_PASS *p,
77 const FIRSTPASS_STATS *position) {
78 p->stats_in = position;
81 static int lookup_next_frame_stats(const TWO_PASS *p,
82 FIRSTPASS_STATS *next_frame) {
83 if (p->stats_in >= p->stats_in_end)
86 *next_frame = *p->stats_in;
91 // Read frame stats at an offset from the current position.
92 static int read_frame_stats(const TWO_PASS *p,
93 FIRSTPASS_STATS *frame_stats, int offset) {
94 const FIRSTPASS_STATS *fps_ptr = p->stats_in;
96 // Check legality of offset.
98 if (&fps_ptr[offset] >= p->stats_in_end)
100 } else if (offset < 0) {
101 if (&fps_ptr[offset] < p->stats_in_start)
105 *frame_stats = fps_ptr[offset];
109 static int input_stats(TWO_PASS *p, FIRSTPASS_STATS *fps) {
110 if (p->stats_in >= p->stats_in_end)
118 static void output_stats(FIRSTPASS_STATS *stats,
119 struct vpx_codec_pkt_list *pktlist) {
120 struct vpx_codec_cx_pkt pkt;
121 pkt.kind = VPX_CODEC_STATS_PKT;
122 pkt.data.twopass_stats.buf = stats;
123 pkt.data.twopass_stats.sz = sizeof(FIRSTPASS_STATS);
124 vpx_codec_pkt_list_add(pktlist, &pkt);
130 fpfile = fopen("firstpass.stt", "a");
132 fprintf(fpfile, "%12.0f %12.0f %12.0f %12.0f %12.4f %12.4f"
133 "%12.4f %12.4f %12.4f %12.4f %12.4f %12.4f %12.4f"
134 "%12.0f %12.0f %12.4f %12.0f %12.0f %12.4f\n",
138 stats->sr_coded_error,
141 stats->pcnt_second_ref,
149 stats->mv_in_out_count,
158 static void zero_stats(FIRSTPASS_STATS *section) {
159 section->frame = 0.0;
160 section->intra_error = 0.0;
161 section->coded_error = 0.0;
162 section->sr_coded_error = 0.0;
163 section->pcnt_inter = 0.0;
164 section->pcnt_motion = 0.0;
165 section->pcnt_second_ref = 0.0;
166 section->pcnt_neutral = 0.0;
168 section->mvr_abs = 0.0;
170 section->mvc_abs = 0.0;
173 section->mv_in_out_count = 0.0;
174 section->new_mv_count = 0.0;
175 section->count = 0.0;
176 section->duration = 1.0;
177 section->spatial_layer_id = 0;
180 static void accumulate_stats(FIRSTPASS_STATS *section,
181 const FIRSTPASS_STATS *frame) {
182 section->frame += frame->frame;
183 section->spatial_layer_id = frame->spatial_layer_id;
184 section->intra_error += frame->intra_error;
185 section->coded_error += frame->coded_error;
186 section->sr_coded_error += frame->sr_coded_error;
187 section->pcnt_inter += frame->pcnt_inter;
188 section->pcnt_motion += frame->pcnt_motion;
189 section->pcnt_second_ref += frame->pcnt_second_ref;
190 section->pcnt_neutral += frame->pcnt_neutral;
191 section->MVr += frame->MVr;
192 section->mvr_abs += frame->mvr_abs;
193 section->MVc += frame->MVc;
194 section->mvc_abs += frame->mvc_abs;
195 section->MVrv += frame->MVrv;
196 section->MVcv += frame->MVcv;
197 section->mv_in_out_count += frame->mv_in_out_count;
198 section->new_mv_count += frame->new_mv_count;
199 section->count += frame->count;
200 section->duration += frame->duration;
203 static void subtract_stats(FIRSTPASS_STATS *section,
204 const FIRSTPASS_STATS *frame) {
205 section->frame -= frame->frame;
206 section->intra_error -= frame->intra_error;
207 section->coded_error -= frame->coded_error;
208 section->sr_coded_error -= frame->sr_coded_error;
209 section->pcnt_inter -= frame->pcnt_inter;
210 section->pcnt_motion -= frame->pcnt_motion;
211 section->pcnt_second_ref -= frame->pcnt_second_ref;
212 section->pcnt_neutral -= frame->pcnt_neutral;
213 section->MVr -= frame->MVr;
214 section->mvr_abs -= frame->mvr_abs;
215 section->MVc -= frame->MVc;
216 section->mvc_abs -= frame->mvc_abs;
217 section->MVrv -= frame->MVrv;
218 section->MVcv -= frame->MVcv;
219 section->mv_in_out_count -= frame->mv_in_out_count;
220 section->new_mv_count -= frame->new_mv_count;
221 section->count -= frame->count;
222 section->duration -= frame->duration;
225 static void avg_stats(FIRSTPASS_STATS *section) {
226 if (section->count < 1.0)
229 section->intra_error /= section->count;
230 section->coded_error /= section->count;
231 section->sr_coded_error /= section->count;
232 section->pcnt_inter /= section->count;
233 section->pcnt_second_ref /= section->count;
234 section->pcnt_neutral /= section->count;
235 section->pcnt_motion /= section->count;
236 section->MVr /= section->count;
237 section->mvr_abs /= section->count;
238 section->MVc /= section->count;
239 section->mvc_abs /= section->count;
240 section->MVrv /= section->count;
241 section->MVcv /= section->count;
242 section->mv_in_out_count /= section->count;
243 section->duration /= section->count;
246 // Calculate a modified Error used in distributing bits between easier and
248 static double calculate_modified_err(const TWO_PASS *twopass,
249 const VP9EncoderConfig *oxcf,
250 const FIRSTPASS_STATS *this_frame) {
251 const FIRSTPASS_STATS *const stats = &twopass->total_stats;
252 const double av_err = stats->coded_error / stats->count;
253 const double modified_error = av_err *
254 pow(this_frame->coded_error / DOUBLE_DIVIDE_CHECK(av_err),
255 oxcf->two_pass_vbrbias / 100.0);
256 return fclamp(modified_error,
257 twopass->modified_error_min, twopass->modified_error_max);
260 // This function returns the maximum target rate per frame.
261 static int frame_max_bits(const RATE_CONTROL *rc,
262 const VP9EncoderConfig *oxcf) {
263 int64_t max_bits = ((int64_t)rc->avg_frame_bandwidth *
264 (int64_t)oxcf->two_pass_vbrmax_section) / 100;
267 else if (max_bits > rc->max_frame_bandwidth)
268 max_bits = rc->max_frame_bandwidth;
270 return (int)max_bits;
273 void vp9_init_first_pass(VP9_COMP *cpi) {
274 zero_stats(&cpi->twopass.total_stats);
277 void vp9_end_first_pass(VP9_COMP *cpi) {
278 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
280 for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
281 output_stats(&cpi->svc.layer_context[i].twopass.total_stats,
282 cpi->output_pkt_list);
285 output_stats(&cpi->twopass.total_stats, cpi->output_pkt_list);
289 static vp9_variance_fn_t get_block_variance_fn(BLOCK_SIZE bsize) {
302 static unsigned int get_prediction_error(BLOCK_SIZE bsize,
303 const struct buf_2d *src,
304 const struct buf_2d *ref) {
306 const vp9_variance_fn_t fn = get_block_variance_fn(bsize);
307 fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
311 // Refine the motion search range according to the frame dimension
312 // for first pass test.
313 static int get_search_range(const VP9_COMMON *cm) {
315 const int dim = MIN(cm->width, cm->height);
317 while ((dim << sr) < MAX_FULL_PEL_VAL)
322 static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
323 const MV *ref_mv, MV *best_mv,
324 int *best_motion_err) {
325 MACROBLOCKD *const xd = &x->e_mbd;
327 MV ref_mv_full = {ref_mv->row >> 3, ref_mv->col >> 3};
328 int num00, tmp_err, n;
329 const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
330 vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
331 const int new_mv_mode_penalty = 256;
334 int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
335 const int sr = get_search_range(&cpi->common);
339 // Override the default variance function to use MSE.
340 v_fn_ptr.vf = get_block_variance_fn(bsize);
342 // Center the initial step/diamond search on best mv.
343 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
345 x->sadperbit16, &num00, &v_fn_ptr, ref_mv);
346 if (tmp_err < INT_MAX)
347 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
348 if (tmp_err < INT_MAX - new_mv_mode_penalty)
349 tmp_err += new_mv_mode_penalty;
351 if (tmp_err < *best_motion_err) {
352 *best_motion_err = tmp_err;
356 // Carry out further step/diamond searches as necessary.
360 while (n < further_steps) {
366 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
367 step_param + n, x->sadperbit16,
368 &num00, &v_fn_ptr, ref_mv);
369 if (tmp_err < INT_MAX)
370 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
371 if (tmp_err < INT_MAX - new_mv_mode_penalty)
372 tmp_err += new_mv_mode_penalty;
374 if (tmp_err < *best_motion_err) {
375 *best_motion_err = tmp_err;
382 static BLOCK_SIZE get_bsize(const VP9_COMMON *cm, int mb_row, int mb_col) {
383 if (2 * mb_col + 1 < cm->mi_cols) {
384 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_16X16
387 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_8X16
392 static int find_fp_qindex() {
395 for (i = 0; i < QINDEX_RANGE; ++i)
396 if (vp9_convert_qindex_to_q(i) >= 30.0)
399 if (i == QINDEX_RANGE)
405 static void set_first_pass_params(VP9_COMP *cpi) {
406 VP9_COMMON *const cm = &cpi->common;
407 if (!cpi->refresh_alt_ref_frame &&
408 (cm->current_video_frame == 0 ||
409 (cpi->frame_flags & FRAMEFLAGS_KEY))) {
410 cm->frame_type = KEY_FRAME;
412 cm->frame_type = INTER_FRAME;
414 // Do not use periodic key frames.
415 cpi->rc.frames_to_key = INT_MAX;
418 void vp9_first_pass(VP9_COMP *cpi) {
420 MACROBLOCK *const x = &cpi->mb;
421 VP9_COMMON *const cm = &cpi->common;
422 MACROBLOCKD *const xd = &x->e_mbd;
424 struct macroblock_plane *const p = x->plane;
425 struct macroblockd_plane *const pd = xd->plane;
426 const PICK_MODE_CONTEXT *ctx = &cpi->pc_root->none;
429 int recon_yoffset, recon_uvoffset;
430 YV12_BUFFER_CONFIG *const lst_yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
431 YV12_BUFFER_CONFIG *gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
432 YV12_BUFFER_CONFIG *const new_yv12 = get_frame_new_buffer(cm);
433 int recon_y_stride = lst_yv12->y_stride;
434 int recon_uv_stride = lst_yv12->uv_stride;
435 int uv_mb_height = 16 >> (lst_yv12->y_height > lst_yv12->uv_height);
436 int64_t intra_error = 0;
437 int64_t coded_error = 0;
438 int64_t sr_coded_error = 0;
440 int sum_mvr = 0, sum_mvc = 0;
441 int sum_mvr_abs = 0, sum_mvc_abs = 0;
442 int64_t sum_mvrs = 0, sum_mvcs = 0;
445 int second_ref_count = 0;
446 int intrapenalty = 256;
447 int neutral_count = 0;
448 int new_mv_count = 0;
449 int sum_in_vectors = 0;
450 uint32_t lastmv_as_int = 0;
451 TWO_PASS *twopass = &cpi->twopass;
452 const MV zero_mv = {0, 0};
453 const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;
455 vp9_clear_system_state();
457 set_first_pass_params(cpi);
458 vp9_set_quantizer(cm, find_fp_qindex());
460 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
461 MV_REFERENCE_FRAME ref_frame = LAST_FRAME;
462 const YV12_BUFFER_CONFIG *scaled_ref_buf = NULL;
463 twopass = &cpi->svc.layer_context[cpi->svc.spatial_layer_id].twopass;
465 vp9_scale_references(cpi);
467 // Use either last frame or alt frame for motion search.
468 if (cpi->ref_frame_flags & VP9_LAST_FLAG) {
469 scaled_ref_buf = vp9_get_scaled_ref_frame(cpi, LAST_FRAME);
470 ref_frame = LAST_FRAME;
471 } else if (cpi->ref_frame_flags & VP9_ALT_FLAG) {
472 scaled_ref_buf = vp9_get_scaled_ref_frame(cpi, ALTREF_FRAME);
473 ref_frame = ALTREF_FRAME;
476 if (scaled_ref_buf != NULL) {
477 // Update the stride since we are using scaled reference buffer
478 first_ref_buf = scaled_ref_buf;
479 recon_y_stride = first_ref_buf->y_stride;
480 recon_uv_stride = first_ref_buf->uv_stride;
481 uv_mb_height = 16 >> (first_ref_buf->y_height > first_ref_buf->uv_height);
484 // Disable golden frame for svc first pass for now.
486 set_ref_ptrs(cm, xd, ref_frame, NONE);
488 cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
489 &cpi->scaled_source);
492 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
494 vp9_setup_src_planes(x, cpi->Source, 0, 0);
495 vp9_setup_pre_planes(xd, 0, first_ref_buf, 0, 0, NULL);
496 vp9_setup_dst_planes(xd->plane, new_yv12, 0, 0);
498 xd->mi = cm->mi_grid_visible;
501 vp9_frame_init_quantizer(cpi);
503 for (i = 0; i < MAX_MB_PLANE; ++i) {
504 p[i].coeff = ctx->coeff_pbuf[i][1];
505 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
506 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
507 p[i].eobs = ctx->eobs_pbuf[i][1];
511 vp9_init_mv_probs(cm);
512 vp9_initialize_rd_consts(cpi);
514 // Tiling is ignored in the first pass.
515 vp9_tile_init(&tile, cm, 0, 0);
517 for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
520 best_ref_mv.as_int = 0;
522 // Reset above block coeffs.
523 xd->up_available = (mb_row != 0);
524 recon_yoffset = (mb_row * recon_y_stride * 16);
525 recon_uvoffset = (mb_row * recon_uv_stride * uv_mb_height);
527 // Set up limit values for motion vectors to prevent them extending
528 // outside the UMV borders.
529 x->mv_row_min = -((mb_row * 16) + BORDER_MV_PIXELS_B16);
530 x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16)
531 + BORDER_MV_PIXELS_B16;
533 for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
535 const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
536 double error_weight = 1.0;
537 const BLOCK_SIZE bsize = get_bsize(cm, mb_row, mb_col);
539 vp9_clear_system_state();
541 xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
542 xd->plane[1].dst.buf = new_yv12->u_buffer + recon_uvoffset;
543 xd->plane[2].dst.buf = new_yv12->v_buffer + recon_uvoffset;
544 xd->left_available = (mb_col != 0);
545 xd->mi[0]->mbmi.sb_type = bsize;
546 xd->mi[0]->mbmi.ref_frame[0] = INTRA_FRAME;
547 set_mi_row_col(xd, &tile,
548 mb_row << 1, num_8x8_blocks_high_lookup[bsize],
549 mb_col << 1, num_8x8_blocks_wide_lookup[bsize],
550 cm->mi_rows, cm->mi_cols);
552 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
553 const int energy = vp9_block_energy(cpi, x, bsize);
554 error_weight = vp9_vaq_inv_q_ratio(energy);
557 // Do intra 16x16 prediction.
559 xd->mi[0]->mbmi.mode = DC_PRED;
560 xd->mi[0]->mbmi.tx_size = use_dc_pred ?
561 (bsize >= BLOCK_16X16 ? TX_16X16 : TX_8X8) : TX_4X4;
562 vp9_encode_intra_block_plane(x, bsize, 0);
563 this_error = vp9_get_mb_ss(x->plane[0].src_diff);
565 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
566 vp9_clear_system_state();
567 this_error = (int)(this_error * error_weight);
570 // Intrapenalty below deals with situations where the intra and inter
571 // error scores are very low (e.g. a plain black frame).
572 // We do not have special cases in first pass for 0,0 and nearest etc so
573 // all inter modes carry an overhead cost estimate for the mv.
574 // When the error score is very low this causes us to pick all or lots of
575 // INTRA modes and throw lots of key frames.
576 // This penalty adds a cost matching that of a 0,0 mv to the intra case.
577 this_error += intrapenalty;
579 // Accumulate the intra error.
580 intra_error += (int64_t)this_error;
582 // Set up limit values for motion vectors to prevent them extending
583 // outside the UMV borders.
584 x->mv_col_min = -((mb_col * 16) + BORDER_MV_PIXELS_B16);
585 x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + BORDER_MV_PIXELS_B16;
587 // Other than for the first frame do a motion search.
588 if (cm->current_video_frame > 0) {
589 int tmp_err, motion_error, raw_motion_error;
591 struct buf_2d unscaled_last_source_buf_2d;
593 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
594 motion_error = get_prediction_error(bsize, &x->plane[0].src,
595 &xd->plane[0].pre[0]);
596 // Assume 0,0 motion with no mv overhead.
597 mv.as_int = tmp_mv.as_int = 0;
599 // Compute the motion error of the 0,0 motion using the last source
600 // frame as the reference. Skip the further motion search on
601 // reconstructed frame if this error is small.
602 unscaled_last_source_buf_2d.buf =
603 cpi->unscaled_last_source->y_buffer + recon_yoffset;
604 unscaled_last_source_buf_2d.stride =
605 cpi->unscaled_last_source->y_stride;
606 raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
607 &unscaled_last_source_buf_2d);
609 // TODO(pengchong): Replace the hard-coded threshold
610 if (raw_motion_error > 25 ||
611 (cpi->use_svc && cpi->svc.number_temporal_layers == 1)) {
612 // Test last reference frame using the previous best mv as the
613 // starting point (best reference) for the search.
614 first_pass_motion_search(cpi, x, &best_ref_mv.as_mv, &mv.as_mv,
616 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
617 vp9_clear_system_state();
618 motion_error = (int)(motion_error * error_weight);
621 // If the current best reference mv is not centered on 0,0 then do a
622 // 0,0 based search as well.
623 if (best_ref_mv.as_int) {
625 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv, &tmp_err);
626 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
627 vp9_clear_system_state();
628 tmp_err = (int)(tmp_err * error_weight);
631 if (tmp_err < motion_error) {
632 motion_error = tmp_err;
633 mv.as_int = tmp_mv.as_int;
637 // Search in an older reference frame.
638 if (cm->current_video_frame > 1 && gld_yv12 != NULL) {
639 // Assume 0,0 motion with no mv overhead.
642 xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
643 gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
644 &xd->plane[0].pre[0]);
646 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv,
648 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
649 vp9_clear_system_state();
650 gf_motion_error = (int)(gf_motion_error * error_weight);
653 if (gf_motion_error < motion_error && gf_motion_error < this_error)
656 // Reset to last frame as reference buffer.
657 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
658 xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
659 xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
661 // In accumulating a score for the older reference frame take the
662 // best of the motion predicted score and the intra coded error
663 // (just as will be done for) accumulation of "coded_error" for
665 if (gf_motion_error < this_error)
666 sr_coded_error += gf_motion_error;
668 sr_coded_error += this_error;
670 sr_coded_error += motion_error;
673 sr_coded_error += motion_error;
676 // Start by assuming that intra mode is best.
677 best_ref_mv.as_int = 0;
679 if (motion_error <= this_error) {
680 // Keep a count of cases where the inter and intra were very close
681 // and very low. This helps with scene cut detection for example in
682 // cropped clips with black bars at the sides or top and bottom.
683 if (((this_error - intrapenalty) * 9 <= motion_error * 10) &&
684 this_error < 2 * intrapenalty)
689 this_error = motion_error;
690 xd->mi[0]->mbmi.mode = NEWMV;
691 xd->mi[0]->mbmi.mv[0] = mv;
692 xd->mi[0]->mbmi.tx_size = TX_4X4;
693 xd->mi[0]->mbmi.ref_frame[0] = LAST_FRAME;
694 xd->mi[0]->mbmi.ref_frame[1] = NONE;
695 vp9_build_inter_predictors_sby(xd, mb_row << 1, mb_col << 1, bsize);
696 vp9_encode_sby_pass1(x, bsize);
697 sum_mvr += mv.as_mv.row;
698 sum_mvr_abs += abs(mv.as_mv.row);
699 sum_mvc += mv.as_mv.col;
700 sum_mvc_abs += abs(mv.as_mv.col);
701 sum_mvrs += mv.as_mv.row * mv.as_mv.row;
702 sum_mvcs += mv.as_mv.col * mv.as_mv.col;
705 best_ref_mv.as_int = mv.as_int;
710 // Non-zero vector, was it different from the last non zero vector?
711 if (mv.as_int != lastmv_as_int)
713 lastmv_as_int = mv.as_int;
715 // Does the row vector point inwards or outwards?
716 if (mb_row < cm->mb_rows / 2) {
717 if (mv.as_mv.row > 0)
719 else if (mv.as_mv.row < 0)
721 } else if (mb_row > cm->mb_rows / 2) {
722 if (mv.as_mv.row > 0)
724 else if (mv.as_mv.row < 0)
728 // Does the col vector point inwards or outwards?
729 if (mb_col < cm->mb_cols / 2) {
730 if (mv.as_mv.col > 0)
732 else if (mv.as_mv.col < 0)
734 } else if (mb_col > cm->mb_cols / 2) {
735 if (mv.as_mv.col > 0)
737 else if (mv.as_mv.col < 0)
743 sr_coded_error += (int64_t)this_error;
745 coded_error += (int64_t)this_error;
747 // Adjust to the next column of MBs.
748 x->plane[0].src.buf += 16;
749 x->plane[1].src.buf += uv_mb_height;
750 x->plane[2].src.buf += uv_mb_height;
753 recon_uvoffset += uv_mb_height;
756 // Adjust to the next row of MBs.
757 x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
758 x->plane[1].src.buf += uv_mb_height * x->plane[1].src.stride -
759 uv_mb_height * cm->mb_cols;
760 x->plane[2].src.buf += uv_mb_height * x->plane[1].src.stride -
761 uv_mb_height * cm->mb_cols;
763 vp9_clear_system_state();
766 vp9_clear_system_state();
770 fps.frame = cm->current_video_frame;
771 fps.spatial_layer_id = cpi->svc.spatial_layer_id;
772 fps.intra_error = (double)(intra_error >> 8);
773 fps.coded_error = (double)(coded_error >> 8);
774 fps.sr_coded_error = (double)(sr_coded_error >> 8);
776 fps.pcnt_inter = (double)intercount / cm->MBs;
777 fps.pcnt_second_ref = (double)second_ref_count / cm->MBs;
778 fps.pcnt_neutral = (double)neutral_count / cm->MBs;
781 fps.MVr = (double)sum_mvr / mvcount;
782 fps.mvr_abs = (double)sum_mvr_abs / mvcount;
783 fps.MVc = (double)sum_mvc / mvcount;
784 fps.mvc_abs = (double)sum_mvc_abs / mvcount;
785 fps.MVrv = ((double)sum_mvrs - (fps.MVr * fps.MVr / mvcount)) / mvcount;
786 fps.MVcv = ((double)sum_mvcs - (fps.MVc * fps.MVc / mvcount)) / mvcount;
787 fps.mv_in_out_count = (double)sum_in_vectors / (mvcount * 2);
788 fps.new_mv_count = new_mv_count;
789 fps.pcnt_motion = (double)mvcount / cm->MBs;
797 fps.mv_in_out_count = 0.0;
798 fps.new_mv_count = 0.0;
799 fps.pcnt_motion = 0.0;
802 // TODO(paulwilkins): Handle the case when duration is set to 0, or
803 // something less than the full time between subsequent values of
804 // cpi->source_time_stamp.
805 fps.duration = (double)(cpi->source->ts_end - cpi->source->ts_start);
807 // Don't want to do output stats with a stack variable!
808 twopass->this_frame_stats = fps;
809 output_stats(&twopass->this_frame_stats, cpi->output_pkt_list);
810 accumulate_stats(&twopass->total_stats, &fps);
813 // Copy the previous Last Frame back into gf and and arf buffers if
814 // the prediction is good enough... but also don't allow it to lag too far.
815 if ((twopass->sr_update_lag > 3) ||
816 ((cm->current_video_frame > 0) &&
817 (twopass->this_frame_stats.pcnt_inter > 0.20) &&
818 ((twopass->this_frame_stats.intra_error /
819 DOUBLE_DIVIDE_CHECK(twopass->this_frame_stats.coded_error)) > 2.0))) {
820 if (gld_yv12 != NULL) {
821 vp8_yv12_copy_frame(lst_yv12, gld_yv12);
823 twopass->sr_update_lag = 1;
825 ++twopass->sr_update_lag;
828 vp9_extend_frame_borders(new_yv12);
830 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1) {
831 vp9_update_reference_frames(cpi);
833 // Swap frame pointers so last frame refers to the frame we just compressed.
834 swap_yv12(lst_yv12, new_yv12);
837 // Special case for the first frame. Copy into the GF buffer as a second
839 if (cm->current_video_frame == 0 && gld_yv12 != NULL) {
840 vp8_yv12_copy_frame(lst_yv12, gld_yv12);
843 // Use this to see what the first pass reconstruction looks like.
847 snprintf(filename, sizeof(filename), "enc%04d.yuv",
848 (int)cm->current_video_frame);
850 if (cm->current_video_frame == 0)
851 recon_file = fopen(filename, "wb");
853 recon_file = fopen(filename, "ab");
855 (void)fwrite(lst_yv12->buffer_alloc, lst_yv12->frame_size, 1, recon_file);
859 ++cm->current_video_frame;
862 static double calc_correction_factor(double err_per_mb,
867 const double error_term = err_per_mb / err_divisor;
869 // Adjustment based on actual quantizer to power term.
870 const double power_term = MIN(vp9_convert_qindex_to_q(q) * 0.0125 + pt_low,
873 // Calculate correction factor.
874 if (power_term < 1.0)
875 assert(error_term >= 0.0);
877 return fclamp(pow(error_term, power_term), 0.05, 5.0);
880 static int get_twopass_worst_quality(const VP9_COMP *cpi,
881 const FIRSTPASS_STATS *stats,
882 int section_target_bandwidth) {
883 const RATE_CONTROL *const rc = &cpi->rc;
884 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
886 if (section_target_bandwidth <= 0) {
887 return rc->worst_quality; // Highest value allowed
889 const int num_mbs = cpi->common.MBs;
890 const double section_err = stats->coded_error / stats->count;
891 const double err_per_mb = section_err / num_mbs;
892 const double speed_term = 1.0 + 0.04 * oxcf->speed;
893 const int target_norm_bits_per_mb = ((uint64_t)section_target_bandwidth <<
894 BPER_MB_NORMBITS) / num_mbs;
896 int is_svc_upper_layer = 0;
897 if (cpi->use_svc && cpi->svc.number_temporal_layers == 1 &&
898 cpi->svc.spatial_layer_id > 0) {
899 is_svc_upper_layer = 1;
902 // Try and pick a max Q that will be high enough to encode the
903 // content at the given rate.
904 for (q = rc->best_quality; q < rc->worst_quality; ++q) {
905 const double factor =
906 calc_correction_factor(err_per_mb, ERR_DIVISOR,
907 is_svc_upper_layer ? SVC_FACTOR_PT_LOW :
908 FACTOR_PT_LOW, FACTOR_PT_HIGH, q);
909 const int bits_per_mb = vp9_rc_bits_per_mb(INTER_FRAME, q,
910 factor * speed_term);
911 if (bits_per_mb <= target_norm_bits_per_mb)
915 // Restriction on active max q for constrained quality mode.
916 if (cpi->oxcf.rc_mode == VPX_CQ)
917 q = MAX(q, oxcf->cq_level);
922 extern void vp9_new_framerate(VP9_COMP *cpi, double framerate);
924 void vp9_init_second_pass(VP9_COMP *cpi) {
925 SVC *const svc = &cpi->svc;
926 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
927 const int is_spatial_svc = (svc->number_spatial_layers > 1) &&
928 (svc->number_temporal_layers == 1);
929 TWO_PASS *const twopass = is_spatial_svc ?
930 &svc->layer_context[svc->spatial_layer_id].twopass : &cpi->twopass;
932 FIRSTPASS_STATS *stats;
934 zero_stats(&twopass->total_stats);
935 zero_stats(&twopass->total_left_stats);
937 if (!twopass->stats_in_end)
940 stats = &twopass->total_stats;
942 *stats = *twopass->stats_in_end;
943 twopass->total_left_stats = *stats;
945 frame_rate = 10000000.0 * stats->count / stats->duration;
946 // Each frame can have a different duration, as the frame rate in the source
947 // isn't guaranteed to be constant. The frame rate prior to the first frame
948 // encoded in the second pass is a guess. However, the sum duration is not.
949 // It is calculated based on the actual durations of all frames from the
952 if (is_spatial_svc) {
953 vp9_update_spatial_layer_framerate(cpi, frame_rate);
954 twopass->bits_left = (int64_t)(stats->duration *
955 svc->layer_context[svc->spatial_layer_id].target_bandwidth /
958 vp9_new_framerate(cpi, frame_rate);
959 twopass->bits_left = (int64_t)(stats->duration * oxcf->target_bandwidth /
963 // Calculate a minimum intra value to be used in determining the IIratio
964 // scores used in the second pass. We have this minimum to make sure
965 // that clips that are static but "low complexity" in the intra domain
966 // are still boosted appropriately for KF/GF/ARF.
967 if (!is_spatial_svc) {
968 // We don't know the number of MBs for each layer at this point.
969 // So we will do it later.
970 twopass->kf_intra_err_min = KF_MB_INTRA_MIN * cpi->common.MBs;
971 twopass->gf_intra_err_min = GF_MB_INTRA_MIN * cpi->common.MBs;
974 // This variable monitors how far behind the second ref update is lagging.
975 twopass->sr_update_lag = 1;
977 // Scan the first pass file and calculate a modified total error based upon
978 // the bias/power function used to allocate bits.
980 const double avg_error = stats->coded_error /
981 DOUBLE_DIVIDE_CHECK(stats->count);
982 const FIRSTPASS_STATS *s = twopass->stats_in;
983 double modified_error_total = 0.0;
984 twopass->modified_error_min = (avg_error *
985 oxcf->two_pass_vbrmin_section) / 100;
986 twopass->modified_error_max = (avg_error *
987 oxcf->two_pass_vbrmax_section) / 100;
988 while (s < twopass->stats_in_end) {
989 modified_error_total += calculate_modified_err(twopass, oxcf, s);
992 twopass->modified_error_left = modified_error_total;
995 // Reset the vbr bits off target counter
996 cpi->rc.vbr_bits_off_target = 0;
999 // This function gives an estimate of how badly we believe the prediction
1000 // quality is decaying from frame to frame.
1001 static double get_prediction_decay_rate(const VP9_COMMON *cm,
1002 const FIRSTPASS_STATS *next_frame) {
1003 // Look at the observed drop in prediction quality between the last frame
1004 // and the GF buffer (which contains an older frame).
1005 const double mb_sr_err_diff = (next_frame->sr_coded_error -
1006 next_frame->coded_error) / cm->MBs;
1007 const double second_ref_decay = mb_sr_err_diff <= 512.0
1008 ? fclamp(pow(1.0 - (mb_sr_err_diff / 512.0), 0.5), 0.85, 1.0)
1011 return MIN(second_ref_decay, next_frame->pcnt_inter);
1014 // Function to test for a condition where a complex transition is followed
1015 // by a static section. For example in slide shows where there is a fade
1016 // between slides. This is to help with more optimal kf and gf positioning.
1017 static int detect_transition_to_still(TWO_PASS *twopass,
1018 int frame_interval, int still_interval,
1019 double loop_decay_rate,
1020 double last_decay_rate) {
1021 int trans_to_still = 0;
1023 // Break clause to detect very still sections after motion
1024 // For example a static image after a fade or other transition
1025 // instead of a clean scene cut.
1026 if (frame_interval > MIN_GF_INTERVAL &&
1027 loop_decay_rate >= 0.999 &&
1028 last_decay_rate < 0.9) {
1030 const FIRSTPASS_STATS *position = twopass->stats_in;
1031 FIRSTPASS_STATS tmp_next_frame;
1033 // Look ahead a few frames to see if static condition persists...
1034 for (j = 0; j < still_interval; ++j) {
1035 if (EOF == input_stats(twopass, &tmp_next_frame))
1038 if (tmp_next_frame.pcnt_inter - tmp_next_frame.pcnt_motion < 0.999)
1042 reset_fpf_position(twopass, position);
1044 // Only if it does do we signal a transition to still.
1045 if (j == still_interval)
1049 return trans_to_still;
1052 // This function detects a flash through the high relative pcnt_second_ref
1053 // score in the frame following a flash frame. The offset passed in should
1055 static int detect_flash(const TWO_PASS *twopass, int offset) {
1056 FIRSTPASS_STATS next_frame;
1058 int flash_detected = 0;
1060 // Read the frame data.
1061 // The return is FALSE (no flash detected) if not a valid frame
1062 if (read_frame_stats(twopass, &next_frame, offset) != EOF) {
1063 // What we are looking for here is a situation where there is a
1064 // brief break in prediction (such as a flash) but subsequent frames
1065 // are reasonably well predicted by an earlier (pre flash) frame.
1066 // The recovery after a flash is indicated by a high pcnt_second_ref
1067 // compared to pcnt_inter.
1068 if (next_frame.pcnt_second_ref > next_frame.pcnt_inter &&
1069 next_frame.pcnt_second_ref >= 0.5)
1073 return flash_detected;
1076 // Update the motion related elements to the GF arf boost calculation.
1077 static void accumulate_frame_motion_stats(const FIRSTPASS_STATS *stats,
1079 double *mv_in_out_accumulator,
1080 double *abs_mv_in_out_accumulator,
1081 double *mv_ratio_accumulator) {
1082 const double pct = stats->pcnt_motion;
1084 // Accumulate Motion In/Out of frame stats.
1085 *mv_in_out = stats->mv_in_out_count * pct;
1086 *mv_in_out_accumulator += *mv_in_out;
1087 *abs_mv_in_out_accumulator += fabs(*mv_in_out);
1089 // Accumulate a measure of how uniform (or conversely how random) the motion
1090 // field is (a ratio of abs(mv) / mv).
1092 const double mvr_ratio = fabs(stats->mvr_abs) /
1093 DOUBLE_DIVIDE_CHECK(fabs(stats->MVr));
1094 const double mvc_ratio = fabs(stats->mvc_abs) /
1095 DOUBLE_DIVIDE_CHECK(fabs(stats->MVc));
1097 *mv_ratio_accumulator += pct * (mvr_ratio < stats->mvr_abs ?
1098 mvr_ratio : stats->mvr_abs);
1099 *mv_ratio_accumulator += pct * (mvc_ratio < stats->mvc_abs ?
1100 mvc_ratio : stats->mvc_abs);
1104 // Calculate a baseline boost number for the current frame.
1105 static double calc_frame_boost(const TWO_PASS *twopass,
1106 const FIRSTPASS_STATS *this_frame,
1107 double this_frame_mv_in_out) {
1110 // Underlying boost factor is based on inter intra error ratio.
1111 if (this_frame->intra_error > twopass->gf_intra_err_min)
1112 frame_boost = (IIFACTOR * this_frame->intra_error /
1113 DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
1115 frame_boost = (IIFACTOR * twopass->gf_intra_err_min /
1116 DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
1118 // Increase boost for frames where new data coming into frame (e.g. zoom out).
1119 // Slightly reduce boost if there is a net balance of motion out of the frame
1120 // (zoom in). The range for this_frame_mv_in_out is -1.0 to +1.0.
1121 if (this_frame_mv_in_out > 0.0)
1122 frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
1123 // In the extreme case the boost is halved.
1125 frame_boost += frame_boost * (this_frame_mv_in_out / 2.0);
1127 return MIN(frame_boost, GF_RMAX);
1130 static int calc_arf_boost(VP9_COMP *cpi, int offset,
1131 int f_frames, int b_frames,
1132 int *f_boost, int *b_boost) {
1133 FIRSTPASS_STATS this_frame;
1134 TWO_PASS *const twopass = &cpi->twopass;
1136 double boost_score = 0.0;
1137 double mv_ratio_accumulator = 0.0;
1138 double decay_accumulator = 1.0;
1139 double this_frame_mv_in_out = 0.0;
1140 double mv_in_out_accumulator = 0.0;
1141 double abs_mv_in_out_accumulator = 0.0;
1143 int flash_detected = 0;
1145 // Search forward from the proposed arf/next gf position.
1146 for (i = 0; i < f_frames; ++i) {
1147 if (read_frame_stats(twopass, &this_frame, (i + offset)) == EOF)
1150 // Update the motion related elements to the boost calculation.
1151 accumulate_frame_motion_stats(&this_frame,
1152 &this_frame_mv_in_out, &mv_in_out_accumulator,
1153 &abs_mv_in_out_accumulator,
1154 &mv_ratio_accumulator);
1156 // We want to discount the flash frame itself and the recovery
1157 // frame that follows as both will have poor scores.
1158 flash_detected = detect_flash(twopass, i + offset) ||
1159 detect_flash(twopass, i + offset + 1);
1161 // Accumulate the effect of prediction quality decay.
1162 if (!flash_detected) {
1163 decay_accumulator *= get_prediction_decay_rate(&cpi->common, &this_frame);
1164 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1165 ? MIN_DECAY_FACTOR : decay_accumulator;
1168 boost_score += decay_accumulator * calc_frame_boost(twopass, &this_frame,
1169 this_frame_mv_in_out);
1172 *f_boost = (int)boost_score;
1174 // Reset for backward looking loop.
1176 mv_ratio_accumulator = 0.0;
1177 decay_accumulator = 1.0;
1178 this_frame_mv_in_out = 0.0;
1179 mv_in_out_accumulator = 0.0;
1180 abs_mv_in_out_accumulator = 0.0;
1182 // Search backward towards last gf position.
1183 for (i = -1; i >= -b_frames; --i) {
1184 if (read_frame_stats(twopass, &this_frame, (i + offset)) == EOF)
1187 // Update the motion related elements to the boost calculation.
1188 accumulate_frame_motion_stats(&this_frame,
1189 &this_frame_mv_in_out, &mv_in_out_accumulator,
1190 &abs_mv_in_out_accumulator,
1191 &mv_ratio_accumulator);
1193 // We want to discount the the flash frame itself and the recovery
1194 // frame that follows as both will have poor scores.
1195 flash_detected = detect_flash(twopass, i + offset) ||
1196 detect_flash(twopass, i + offset + 1);
1198 // Cumulative effect of prediction quality decay.
1199 if (!flash_detected) {
1200 decay_accumulator *= get_prediction_decay_rate(&cpi->common, &this_frame);
1201 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1202 ? MIN_DECAY_FACTOR : decay_accumulator;
1205 boost_score += decay_accumulator * calc_frame_boost(twopass, &this_frame,
1206 this_frame_mv_in_out);
1208 *b_boost = (int)boost_score;
1210 arf_boost = (*f_boost + *b_boost);
1211 if (arf_boost < ((b_frames + f_frames) * 20))
1212 arf_boost = ((b_frames + f_frames) * 20);
1217 // Calculate a section intra ratio used in setting max loop filter.
1218 static int calculate_section_intra_ratio(const FIRSTPASS_STATS *begin,
1219 const FIRSTPASS_STATS *end,
1220 int section_length) {
1221 const FIRSTPASS_STATS *s = begin;
1222 double intra_error = 0.0;
1223 double coded_error = 0.0;
1226 while (s < end && i < section_length) {
1227 intra_error += s->intra_error;
1228 coded_error += s->coded_error;
1233 return (int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
1236 // Calculate the total bits to allocate in this GF/ARF group.
1237 static int64_t calculate_total_gf_group_bits(VP9_COMP *cpi,
1238 double gf_group_err) {
1239 const RATE_CONTROL *const rc = &cpi->rc;
1240 const TWO_PASS *const twopass = &cpi->twopass;
1241 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
1242 int64_t total_group_bits;
1244 // Calculate the bits to be allocated to the group as a whole.
1245 if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0)) {
1246 total_group_bits = (int64_t)(twopass->kf_group_bits *
1247 (gf_group_err / twopass->kf_group_error_left));
1249 total_group_bits = 0;
1252 // Clamp odd edge cases.
1253 total_group_bits = (total_group_bits < 0) ?
1254 0 : (total_group_bits > twopass->kf_group_bits) ?
1255 twopass->kf_group_bits : total_group_bits;
1257 // Clip based on user supplied data rate variability limit.
1258 if (total_group_bits > (int64_t)max_bits * rc->baseline_gf_interval)
1259 total_group_bits = (int64_t)max_bits * rc->baseline_gf_interval;
1261 return total_group_bits;
1264 // Calculate the number bits extra to assign to boosted frames in a group.
1265 static int calculate_boost_bits(int frame_count,
1266 int boost, int64_t total_group_bits) {
1267 int allocation_chunks;
1269 // return 0 for invalid inputs (could arise e.g. through rounding errors)
1270 if (!boost || (total_group_bits <= 0) || (frame_count <= 0) )
1273 allocation_chunks = (frame_count * 100) + boost;
1275 // Prevent overflow.
1277 int divisor = boost >> 10;
1279 allocation_chunks /= divisor;
1282 // Calculate the number of extra bits for use in the boosted frame or frames.
1283 return MAX((int)(((int64_t)boost * total_group_bits) / allocation_chunks), 0);
1286 static void allocate_gf_group_bits(VP9_COMP *cpi, int64_t gf_group_bits,
1287 double group_error, int gf_arf_bits) {
1288 RATE_CONTROL *const rc = &cpi->rc;
1289 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1290 TWO_PASS *twopass = &cpi->twopass;
1291 FIRSTPASS_STATS frame_stats;
1293 int frame_index = 1;
1294 int target_frame_size;
1296 const int max_bits = frame_max_bits(&cpi->rc, &cpi->oxcf);
1297 int64_t total_group_bits = gf_group_bits;
1298 double modified_err = 0.0;
1299 double err_fraction;
1300 int mid_boost_bits = 0;
1301 int middle_frame_idx;
1303 key_frame = cpi->common.frame_type == KEY_FRAME ||
1304 vp9_is_upper_layer_key_frame(cpi);
1306 // For key frames the frame target rate is already set and it
1307 // is also the golden frame.
1309 if (rc->source_alt_ref_active) {
1310 twopass->gf_group.update_type[0] = OVERLAY_UPDATE;
1311 twopass->gf_group.rf_level[0] = INTER_NORMAL;
1312 twopass->gf_group.bit_allocation[0] = 0;
1313 twopass->gf_group.arf_update_idx[0] = 2;
1314 twopass->gf_group.arf_ref_idx[0] = 2;
1316 twopass->gf_group.update_type[0] = GF_UPDATE;
1317 twopass->gf_group.rf_level[0] = GF_ARF_STD;
1318 twopass->gf_group.bit_allocation[0] = gf_arf_bits;
1319 twopass->gf_group.arf_update_idx[0] = 2;
1320 twopass->gf_group.arf_ref_idx[0] = 2;
1323 // Step over the golden frame / overlay frame
1324 if (EOF == input_stats(twopass, &frame_stats))
1328 // Deduct the boost bits for arf (or gf if it is not a key frame)
1329 // from the group total.
1330 if (rc->source_alt_ref_pending || !key_frame)
1331 total_group_bits -= gf_arf_bits;
1333 // Store the bits to spend on the ARF if there is one.
1334 if (rc->source_alt_ref_pending) {
1335 // A portion of the gf / arf extra bits are set asside for lower level
1336 // boosted frames in the middle of the group.
1337 mid_boost_bits = gf_arf_bits >> 5;
1338 gf_arf_bits -= (gf_arf_bits >> 5);
1340 twopass->gf_group.update_type[frame_index] = ARF_UPDATE;
1341 twopass->gf_group.rf_level[frame_index] = GF_ARF_STD;
1342 twopass->gf_group.bit_allocation[frame_index] = gf_arf_bits;
1343 twopass->gf_group.arf_src_offset[frame_index] =
1344 (unsigned char)(rc->baseline_gf_interval - 1);
1345 twopass->gf_group.arf_update_idx[frame_index] = 2;
1346 twopass->gf_group.arf_ref_idx[frame_index] = 2;
1349 if (cpi->multi_arf_enabled) {
1350 // Set aside a slot for a level 1 arf.
1351 twopass->gf_group.update_type[frame_index] = ARF_UPDATE;
1352 twopass->gf_group.rf_level[frame_index] = GF_ARF_LOW;
1353 twopass->gf_group.arf_src_offset[frame_index] =
1354 (unsigned char)((rc->baseline_gf_interval >> 1) - 1);
1355 twopass->gf_group.arf_update_idx[frame_index] = 3;
1356 twopass->gf_group.arf_ref_idx[frame_index] = 2;
1361 // Define middle frame
1362 middle_frame_idx = frame_index + (rc->baseline_gf_interval >> 1) - 1;
1364 // Allocate bits to the other frames in the group.
1365 for (i = 0; i < rc->baseline_gf_interval - 1; ++i) {
1366 if (EOF == input_stats(twopass, &frame_stats))
1369 modified_err = calculate_modified_err(twopass, oxcf, &frame_stats);
1371 if (group_error > 0)
1372 err_fraction = modified_err / DOUBLE_DIVIDE_CHECK(group_error);
1376 target_frame_size = (int)((double)total_group_bits * err_fraction);
1378 if (rc->source_alt_ref_pending && cpi->multi_arf_enabled) {
1379 mid_boost_bits += (target_frame_size >> 4);
1380 target_frame_size -= (target_frame_size >> 4);
1382 if (frame_index <= middle_frame_idx) {
1383 twopass->gf_group.arf_update_idx[frame_index] = 3;
1384 twopass->gf_group.arf_ref_idx[frame_index] = 3;
1386 twopass->gf_group.arf_update_idx[frame_index] = 2;
1387 twopass->gf_group.arf_ref_idx[frame_index] = 2;
1390 twopass->gf_group.arf_update_idx[frame_index] = 2;
1391 twopass->gf_group.arf_ref_idx[frame_index] = 2;
1394 target_frame_size = clamp(target_frame_size, 0,
1395 MIN(max_bits, (int)total_group_bits));
1397 twopass->gf_group.update_type[frame_index] = LF_UPDATE;
1398 twopass->gf_group.rf_level[frame_index] = INTER_NORMAL;
1400 twopass->gf_group.bit_allocation[frame_index] = target_frame_size;
1404 if (rc->source_alt_ref_pending) {
1405 if (cpi->multi_arf_enabled) {
1406 twopass->gf_group.bit_allocation[2] =
1407 twopass->gf_group.bit_allocation[middle_frame_idx] + mid_boost_bits;
1408 twopass->gf_group.update_type[middle_frame_idx] = OVERLAY_UPDATE;
1409 twopass->gf_group.bit_allocation[middle_frame_idx] = 0;
1412 // Configure the overlay frame at the end of the sequence that will also
1413 // be the start frame of the next group. The reason for doing this here
1414 // is that on entry to vp9_get_compressed_data() for the overlay
1415 // frame, but before the call to vp9_rc_get_second_pass_params() the
1416 // data will otherwise be undefined.
1417 twopass->gf_group.update_type[frame_index] = OVERLAY_UPDATE;
1418 twopass->gf_group.rf_level[frame_index] = INTER_NORMAL;
1419 twopass->gf_group.arf_update_idx[frame_index] = 2;
1420 twopass->gf_group.arf_ref_idx[frame_index] = 2;
1424 // Analyse and define a gf/arf group.
1425 static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
1426 RATE_CONTROL *const rc = &cpi->rc;
1427 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1428 TWO_PASS *const twopass = &cpi->twopass;
1429 FIRSTPASS_STATS next_frame;
1430 const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
1433 double boost_score = 0.0;
1434 double old_boost_score = 0.0;
1435 double gf_group_err = 0.0;
1436 double gf_first_frame_err = 0.0;
1437 double mod_frame_err = 0.0;
1439 double mv_ratio_accumulator = 0.0;
1440 double decay_accumulator = 1.0;
1441 double zero_motion_accumulator = 1.0;
1443 double loop_decay_rate = 1.00;
1444 double last_loop_decay_rate = 1.00;
1446 double this_frame_mv_in_out = 0.0;
1447 double mv_in_out_accumulator = 0.0;
1448 double abs_mv_in_out_accumulator = 0.0;
1449 double mv_ratio_accumulator_thresh;
1450 unsigned int allow_alt_ref = is_altref_enabled(oxcf);
1455 int active_max_gf_interval;
1456 int64_t gf_group_bits;
1457 double gf_group_error_left;
1460 // Reset the GF group data structures unless this is a key
1461 // frame in which case it will already have been done.
1462 if (cpi->common.frame_type != KEY_FRAME) {
1463 vp9_zero(twopass->gf_group);
1466 vp9_clear_system_state();
1467 vp9_zero(next_frame);
1471 // Load stats for the current frame.
1472 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1474 // Note the error of the frame at the start of the group. This will be
1475 // the GF frame error if we code a normal gf.
1476 gf_first_frame_err = mod_frame_err;
1478 // If this is a key frame or the overlay from a previous arf then
1479 // the error score / cost of this frame has already been accounted for.
1480 if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
1481 gf_group_err -= gf_first_frame_err;
1483 // Motion breakout threshold for loop below depends on image size.
1484 mv_ratio_accumulator_thresh = (cpi->common.width + cpi->common.height) / 10.0;
1486 // Work out a maximum interval for the GF.
1487 // If the image appears completely static we can extend beyond this.
1488 // The value chosen depends on the active Q range. At low Q we have
1489 // bits to spare and are better with a smaller interval and smaller boost.
1490 // At high Q when there are few bits to spare we are better with a longer
1491 // interval to spread the cost of the GF.
1493 active_max_gf_interval =
1494 12 + ((int)vp9_convert_qindex_to_q(rc->last_q[INTER_FRAME]) >> 5);
1496 if (active_max_gf_interval > rc->max_gf_interval)
1497 active_max_gf_interval = rc->max_gf_interval;
1500 while (i < rc->static_scene_max_gf_interval && i < rc->frames_to_key) {
1503 // Accumulate error score of frames in this gf group.
1504 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1505 gf_group_err += mod_frame_err;
1507 if (EOF == input_stats(twopass, &next_frame))
1510 // Test for the case where there is a brief flash but the prediction
1511 // quality back to an earlier frame is then restored.
1512 flash_detected = detect_flash(twopass, 0);
1514 // Update the motion related elements to the boost calculation.
1515 accumulate_frame_motion_stats(&next_frame,
1516 &this_frame_mv_in_out, &mv_in_out_accumulator,
1517 &abs_mv_in_out_accumulator,
1518 &mv_ratio_accumulator);
1520 // Accumulate the effect of prediction quality decay.
1521 if (!flash_detected) {
1522 last_loop_decay_rate = loop_decay_rate;
1523 loop_decay_rate = get_prediction_decay_rate(&cpi->common, &next_frame);
1524 decay_accumulator = decay_accumulator * loop_decay_rate;
1526 // Monitor for static sections.
1527 if ((next_frame.pcnt_inter - next_frame.pcnt_motion) <
1528 zero_motion_accumulator) {
1529 zero_motion_accumulator = next_frame.pcnt_inter -
1530 next_frame.pcnt_motion;
1533 // Break clause to detect very still sections after motion. For example,
1534 // a static image after a fade or other transition.
1535 if (detect_transition_to_still(twopass, i, 5, loop_decay_rate,
1536 last_loop_decay_rate)) {
1542 // Calculate a boost number for this frame.
1543 boost_score += decay_accumulator * calc_frame_boost(twopass, &next_frame,
1544 this_frame_mv_in_out);
1546 // Break out conditions.
1548 // Break at active_max_gf_interval unless almost totally static.
1549 (i >= active_max_gf_interval && (zero_motion_accumulator < 0.995)) ||
1551 // Don't break out with a very short interval.
1552 (i > MIN_GF_INTERVAL) &&
1553 ((boost_score > 125.0) || (next_frame.pcnt_inter < 0.75)) &&
1554 (!flash_detected) &&
1555 ((mv_ratio_accumulator > mv_ratio_accumulator_thresh) ||
1556 (abs_mv_in_out_accumulator > 3.0) ||
1557 (mv_in_out_accumulator < -2.0) ||
1558 ((boost_score - old_boost_score) < IIFACTOR)))) {
1559 boost_score = old_boost_score;
1563 *this_frame = next_frame;
1565 old_boost_score = boost_score;
1568 twopass->gf_zeromotion_pct = (int)(zero_motion_accumulator * 1000.0);
1570 // Don't allow a gf too near the next kf.
1571 if ((rc->frames_to_key - i) < MIN_GF_INTERVAL) {
1572 while (i < (rc->frames_to_key + !rc->next_key_frame_forced)) {
1575 if (EOF == input_stats(twopass, this_frame))
1578 if (i < rc->frames_to_key) {
1579 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1580 gf_group_err += mod_frame_err;
1585 // Set the interval until the next gf.
1586 if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
1587 rc->baseline_gf_interval = i - 1;
1589 rc->baseline_gf_interval = i;
1591 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
1593 // Should we use the alternate reference frame.
1594 if (allow_alt_ref &&
1595 (i < cpi->oxcf.lag_in_frames) &&
1596 (i >= MIN_GF_INTERVAL) &&
1597 // For real scene cuts (not forced kfs) don't allow arf very near kf.
1598 (rc->next_key_frame_forced ||
1599 (i <= (rc->frames_to_key - MIN_GF_INTERVAL)))) {
1600 // Calculate the boost for alt ref.
1601 rc->gfu_boost = calc_arf_boost(cpi, 0, (i - 1), (i - 1), &f_boost,
1603 rc->source_alt_ref_pending = 1;
1606 rc->gfu_boost = (int)boost_score;
1607 rc->source_alt_ref_pending = 0;
1610 // Reset the file position.
1611 reset_fpf_position(twopass, start_pos);
1613 // Calculate the bits to be allocated to the gf/arf group as a whole
1614 gf_group_bits = calculate_total_gf_group_bits(cpi, gf_group_err);
1616 // Calculate the extra bits to be used for boosted frame(s)
1618 int q = rc->last_q[INTER_FRAME];
1619 int boost = (rc->gfu_boost * gfboost_qadjust(q)) / 100;
1621 // Set max and minimum boost and hence minimum allocation.
1622 boost = clamp(boost, 125, (rc->baseline_gf_interval + 1) * 200);
1624 // Calculate the extra bits to be used for boosted frame(s)
1625 gf_arf_bits = calculate_boost_bits(rc->baseline_gf_interval,
1626 boost, gf_group_bits);
1629 // Adjust KF group bits and error remaining.
1630 twopass->kf_group_error_left -= (int64_t)gf_group_err;
1632 // If this is an arf update we want to remove the score for the overlay
1633 // frame at the end which will usually be very cheap to code.
1634 // The overlay frame has already, in effect, been coded so we want to spread
1635 // the remaining bits among the other frames.
1636 // For normal GFs remove the score for the GF itself unless this is
1637 // also a key frame in which case it has already been accounted for.
1638 if (rc->source_alt_ref_pending) {
1639 gf_group_error_left = gf_group_err - mod_frame_err;
1640 } else if (cpi->common.frame_type != KEY_FRAME) {
1641 gf_group_error_left = gf_group_err - gf_first_frame_err;
1643 gf_group_error_left = gf_group_err;
1646 // Allocate bits to each of the frames in the GF group.
1647 allocate_gf_group_bits(cpi, gf_group_bits, gf_group_error_left, gf_arf_bits);
1649 // Reset the file position.
1650 reset_fpf_position(twopass, start_pos);
1652 // Calculate a section intra ratio used in setting max loop filter.
1653 if (cpi->common.frame_type != KEY_FRAME) {
1654 twopass->section_intra_rating =
1655 calculate_section_intra_ratio(start_pos, twopass->stats_in_end,
1656 rc->baseline_gf_interval);
1660 static int test_candidate_kf(TWO_PASS *twopass,
1661 const FIRSTPASS_STATS *last_frame,
1662 const FIRSTPASS_STATS *this_frame,
1663 const FIRSTPASS_STATS *next_frame) {
1664 int is_viable_kf = 0;
1666 // Does the frame satisfy the primary criteria of a key frame?
1667 // If so, then examine how well it predicts subsequent frames.
1668 if ((this_frame->pcnt_second_ref < 0.10) &&
1669 (next_frame->pcnt_second_ref < 0.10) &&
1670 ((this_frame->pcnt_inter < 0.05) ||
1671 (((this_frame->pcnt_inter - this_frame->pcnt_neutral) < 0.35) &&
1672 ((this_frame->intra_error /
1673 DOUBLE_DIVIDE_CHECK(this_frame->coded_error)) < 2.5) &&
1674 ((fabs(last_frame->coded_error - this_frame->coded_error) /
1675 DOUBLE_DIVIDE_CHECK(this_frame->coded_error) > 0.40) ||
1676 (fabs(last_frame->intra_error - this_frame->intra_error) /
1677 DOUBLE_DIVIDE_CHECK(this_frame->intra_error) > 0.40) ||
1678 ((next_frame->intra_error /
1679 DOUBLE_DIVIDE_CHECK(next_frame->coded_error)) > 3.5))))) {
1681 const FIRSTPASS_STATS *start_pos = twopass->stats_in;
1682 FIRSTPASS_STATS local_next_frame = *next_frame;
1683 double boost_score = 0.0;
1684 double old_boost_score = 0.0;
1685 double decay_accumulator = 1.0;
1687 // Examine how well the key frame predicts subsequent frames.
1688 for (i = 0; i < 16; ++i) {
1689 double next_iiratio = (IIKFACTOR1 * local_next_frame.intra_error /
1690 DOUBLE_DIVIDE_CHECK(local_next_frame.coded_error));
1692 if (next_iiratio > RMAX)
1693 next_iiratio = RMAX;
1695 // Cumulative effect of decay in prediction quality.
1696 if (local_next_frame.pcnt_inter > 0.85)
1697 decay_accumulator *= local_next_frame.pcnt_inter;
1699 decay_accumulator *= (0.85 + local_next_frame.pcnt_inter) / 2.0;
1701 // Keep a running total.
1702 boost_score += (decay_accumulator * next_iiratio);
1704 // Test various breakout clauses.
1705 if ((local_next_frame.pcnt_inter < 0.05) ||
1706 (next_iiratio < 1.5) ||
1707 (((local_next_frame.pcnt_inter -
1708 local_next_frame.pcnt_neutral) < 0.20) &&
1709 (next_iiratio < 3.0)) ||
1710 ((boost_score - old_boost_score) < 3.0) ||
1711 (local_next_frame.intra_error < 200)) {
1715 old_boost_score = boost_score;
1717 // Get the next frame details
1718 if (EOF == input_stats(twopass, &local_next_frame))
1722 // If there is tolerable prediction for at least the next 3 frames then
1723 // break out else discard this potential key frame and move on
1724 if (boost_score > 30.0 && (i > 3)) {
1727 // Reset the file position
1728 reset_fpf_position(twopass, start_pos);
1734 return is_viable_kf;
1737 static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
1739 RATE_CONTROL *const rc = &cpi->rc;
1740 TWO_PASS *const twopass = &cpi->twopass;
1741 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1742 const FIRSTPASS_STATS first_frame = *this_frame;
1743 const FIRSTPASS_STATS *const start_position = twopass->stats_in;
1744 FIRSTPASS_STATS next_frame;
1745 FIRSTPASS_STATS last_frame;
1747 double decay_accumulator = 1.0;
1748 double zero_motion_accumulator = 1.0;
1749 double boost_score = 0.0;
1750 double kf_mod_err = 0.0;
1751 double kf_group_err = 0.0;
1752 double recent_loop_decay[8] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
1754 vp9_zero(next_frame);
1756 cpi->common.frame_type = KEY_FRAME;
1758 // Reset the GF group data structures.
1759 vp9_zero(twopass->gf_group);
1761 // Is this a forced key frame by interval.
1762 rc->this_key_frame_forced = rc->next_key_frame_forced;
1764 // Clear the alt ref active flag as this can never be active on a key frame.
1765 rc->source_alt_ref_active = 0;
1767 // KF is always a GF so clear frames till next gf counter.
1768 rc->frames_till_gf_update_due = 0;
1770 rc->frames_to_key = 1;
1772 twopass->kf_group_bits = 0; // Total bits available to kf group
1773 twopass->kf_group_error_left = 0; // Group modified error score.
1775 kf_mod_err = calculate_modified_err(twopass, oxcf, this_frame);
1777 // Find the next keyframe.
1779 while (twopass->stats_in < twopass->stats_in_end &&
1780 rc->frames_to_key < cpi->oxcf.key_freq) {
1781 // Accumulate kf group error.
1782 kf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
1784 // Load the next frame's stats.
1785 last_frame = *this_frame;
1786 input_stats(twopass, this_frame);
1788 // Provided that we are not at the end of the file...
1789 if (cpi->oxcf.auto_key &&
1790 lookup_next_frame_stats(twopass, &next_frame) != EOF) {
1791 double loop_decay_rate;
1793 // Check for a scene cut.
1794 if (test_candidate_kf(twopass, &last_frame, this_frame, &next_frame))
1797 // How fast is the prediction quality decaying?
1798 loop_decay_rate = get_prediction_decay_rate(&cpi->common, &next_frame);
1800 // We want to know something about the recent past... rather than
1801 // as used elsewhere where we are concerned with decay in prediction
1802 // quality since the last GF or KF.
1803 recent_loop_decay[i % 8] = loop_decay_rate;
1804 decay_accumulator = 1.0;
1805 for (j = 0; j < 8; ++j)
1806 decay_accumulator *= recent_loop_decay[j];
1808 // Special check for transition or high motion followed by a
1810 if (detect_transition_to_still(twopass, i, cpi->oxcf.key_freq - i,
1811 loop_decay_rate, decay_accumulator))
1814 // Step on to the next frame.
1815 ++rc->frames_to_key;
1817 // If we don't have a real key frame within the next two
1818 // key_freq intervals then break out of the loop.
1819 if (rc->frames_to_key >= 2 * cpi->oxcf.key_freq)
1822 ++rc->frames_to_key;
1827 // If there is a max kf interval set by the user we must obey it.
1828 // We already breakout of the loop above at 2x max.
1829 // This code centers the extra kf if the actual natural interval
1830 // is between 1x and 2x.
1831 if (cpi->oxcf.auto_key &&
1832 rc->frames_to_key > cpi->oxcf.key_freq) {
1833 FIRSTPASS_STATS tmp_frame = first_frame;
1835 rc->frames_to_key /= 2;
1837 // Reset to the start of the group.
1838 reset_fpf_position(twopass, start_position);
1842 // Rescan to get the correct error data for the forced kf group.
1843 for (i = 0; i < rc->frames_to_key; ++i) {
1844 kf_group_err += calculate_modified_err(twopass, oxcf, &tmp_frame);
1845 input_stats(twopass, &tmp_frame);
1847 rc->next_key_frame_forced = 1;
1848 } else if (twopass->stats_in == twopass->stats_in_end ||
1849 rc->frames_to_key >= cpi->oxcf.key_freq) {
1850 rc->next_key_frame_forced = 1;
1852 rc->next_key_frame_forced = 0;
1855 // Special case for the last key frame of the file.
1856 if (twopass->stats_in >= twopass->stats_in_end) {
1857 // Accumulate kf group error.
1858 kf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
1861 // Calculate the number of bits that should be assigned to the kf group.
1862 if (twopass->bits_left > 0 && twopass->modified_error_left > 0.0) {
1863 // Maximum number of bits for a single normal frame (not key frame).
1864 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
1866 // Maximum number of bits allocated to the key frame group.
1867 int64_t max_grp_bits;
1869 // Default allocation based on bits left and relative
1870 // complexity of the section.
1871 twopass->kf_group_bits = (int64_t)(twopass->bits_left *
1872 (kf_group_err / twopass->modified_error_left));
1874 // Clip based on maximum per frame rate defined by the user.
1875 max_grp_bits = (int64_t)max_bits * (int64_t)rc->frames_to_key;
1876 if (twopass->kf_group_bits > max_grp_bits)
1877 twopass->kf_group_bits = max_grp_bits;
1879 twopass->kf_group_bits = 0;
1881 twopass->kf_group_bits = MAX(0, twopass->kf_group_bits);
1883 // Reset the first pass file position.
1884 reset_fpf_position(twopass, start_position);
1886 // Scan through the kf group collating various stats used to deteermine
1887 // how many bits to spend on it.
1888 decay_accumulator = 1.0;
1890 for (i = 0; i < rc->frames_to_key; ++i) {
1891 if (EOF == input_stats(twopass, &next_frame))
1894 // Monitor for static sections.
1895 if ((next_frame.pcnt_inter - next_frame.pcnt_motion) <
1896 zero_motion_accumulator) {
1897 zero_motion_accumulator = (next_frame.pcnt_inter -
1898 next_frame.pcnt_motion);
1901 // For the first few frames collect data to decide kf boost.
1902 if (i <= (rc->max_gf_interval * 2)) {
1904 if (next_frame.intra_error > twopass->kf_intra_err_min)
1905 r = (IIKFACTOR2 * next_frame.intra_error /
1906 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
1908 r = (IIKFACTOR2 * twopass->kf_intra_err_min /
1909 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
1914 // How fast is prediction quality decaying.
1915 if (!detect_flash(twopass, 0)) {
1916 const double loop_decay_rate = get_prediction_decay_rate(&cpi->common,
1918 decay_accumulator *= loop_decay_rate;
1919 decay_accumulator = MAX(decay_accumulator, MIN_DECAY_FACTOR);
1922 boost_score += (decay_accumulator * r);
1926 reset_fpf_position(twopass, start_position);
1928 // Store the zero motion percentage
1929 twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
1931 // Calculate a section intra ratio used in setting max loop filter.
1932 twopass->section_intra_rating =
1933 calculate_section_intra_ratio(start_position, twopass->stats_in_end,
1936 // Work out how many bits to allocate for the key frame itself.
1937 rc->kf_boost = (int)boost_score;
1939 if (rc->kf_boost < (rc->frames_to_key * 3))
1940 rc->kf_boost = (rc->frames_to_key * 3);
1941 if (rc->kf_boost < MIN_KF_BOOST)
1942 rc->kf_boost = MIN_KF_BOOST;
1944 kf_bits = calculate_boost_bits((rc->frames_to_key - 1),
1945 rc->kf_boost, twopass->kf_group_bits);
1947 twopass->kf_group_bits -= kf_bits;
1949 // Save the bits to spend on the key frame.
1950 twopass->gf_group.bit_allocation[0] = kf_bits;
1951 twopass->gf_group.update_type[0] = KF_UPDATE;
1952 twopass->gf_group.rf_level[0] = KF_STD;
1954 // Note the total error score of the kf group minus the key frame itself.
1955 twopass->kf_group_error_left = (int)(kf_group_err - kf_mod_err);
1957 // Adjust the count of total modified error left.
1958 // The count of bits left is adjusted elsewhere based on real coded frame
1960 twopass->modified_error_left -= kf_group_err;
1963 // For VBR...adjustment to the frame target based on error from previous frames
1964 void vbr_rate_correction(int * this_frame_target,
1965 const int64_t vbr_bits_off_target) {
1966 int max_delta = (*this_frame_target * 15) / 100;
1968 // vbr_bits_off_target > 0 means we have extra bits to spend
1969 if (vbr_bits_off_target > 0) {
1970 *this_frame_target +=
1971 (vbr_bits_off_target > max_delta) ? max_delta
1972 : (int)vbr_bits_off_target;
1974 *this_frame_target -=
1975 (vbr_bits_off_target < -max_delta) ? max_delta
1976 : (int)-vbr_bits_off_target;
1980 // Define the reference buffers that will be updated post encode.
1981 void configure_buffer_updates(VP9_COMP *cpi) {
1982 TWO_PASS *const twopass = &cpi->twopass;
1984 cpi->rc.is_src_frame_alt_ref = 0;
1985 switch (twopass->gf_group.update_type[twopass->gf_group.index]) {
1987 cpi->refresh_last_frame = 1;
1988 cpi->refresh_golden_frame = 1;
1989 cpi->refresh_alt_ref_frame = 1;
1992 cpi->refresh_last_frame = 1;
1993 cpi->refresh_golden_frame = 0;
1994 cpi->refresh_alt_ref_frame = 0;
1997 cpi->refresh_last_frame = 1;
1998 cpi->refresh_golden_frame = 1;
1999 cpi->refresh_alt_ref_frame = 0;
2001 case OVERLAY_UPDATE:
2002 cpi->refresh_last_frame = 0;
2003 cpi->refresh_golden_frame = 1;
2004 cpi->refresh_alt_ref_frame = 0;
2005 cpi->rc.is_src_frame_alt_ref = 1;
2008 cpi->refresh_last_frame = 0;
2009 cpi->refresh_golden_frame = 0;
2010 cpi->refresh_alt_ref_frame = 1;
2018 void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
2019 VP9_COMMON *const cm = &cpi->common;
2020 RATE_CONTROL *const rc = &cpi->rc;
2021 TWO_PASS *const twopass = &cpi->twopass;
2023 FIRSTPASS_STATS this_frame;
2024 FIRSTPASS_STATS this_frame_copy;
2027 LAYER_CONTEXT *lc = NULL;
2028 const int is_spatial_svc = (cpi->use_svc &&
2029 cpi->svc.number_temporal_layers == 1);
2030 if (is_spatial_svc) {
2031 lc = &cpi->svc.layer_context[cpi->svc.spatial_layer_id];
2032 frames_left = (int)(twopass->total_stats.count -
2033 lc->current_video_frame_in_layer);
2035 frames_left = (int)(twopass->total_stats.count -
2036 cm->current_video_frame);
2039 if (!twopass->stats_in)
2042 // If this is an arf frame then we dont want to read the stats file or
2043 // advance the input pointer as we already have what we need.
2044 if (twopass->gf_group.update_type[twopass->gf_group.index] == ARF_UPDATE) {
2046 configure_buffer_updates(cpi);
2047 target_rate = twopass->gf_group.bit_allocation[twopass->gf_group.index];
2048 target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
2049 rc->base_frame_target = target_rate;
2050 #ifdef LONG_TERM_VBR_CORRECTION
2051 // Correction to rate target based on prior over or under shoot.
2052 if (cpi->oxcf.rc_mode == VPX_VBR)
2053 vbr_rate_correction(&target_rate, rc->vbr_bits_off_target);
2055 vp9_rc_set_frame_target(cpi, target_rate);
2056 cm->frame_type = INTER_FRAME;
2060 vp9_clear_system_state();
2062 if (is_spatial_svc && twopass->kf_intra_err_min == 0) {
2063 twopass->kf_intra_err_min = KF_MB_INTRA_MIN * cpi->common.MBs;
2064 twopass->gf_intra_err_min = GF_MB_INTRA_MIN * cpi->common.MBs;
2067 if (cpi->oxcf.rc_mode == VPX_Q) {
2068 twopass->active_worst_quality = cpi->oxcf.cq_level;
2069 } else if (cm->current_video_frame == 0 ||
2070 (is_spatial_svc && lc->current_video_frame_in_layer == 0)) {
2071 // Special case code for first frame.
2072 const int section_target_bandwidth = (int)(twopass->bits_left /
2074 const int tmp_q = get_twopass_worst_quality(cpi, &twopass->total_left_stats,
2075 section_target_bandwidth);
2076 twopass->active_worst_quality = tmp_q;
2077 rc->ni_av_qi = tmp_q;
2078 rc->avg_q = vp9_convert_qindex_to_q(tmp_q);
2080 vp9_zero(this_frame);
2081 if (EOF == input_stats(twopass, &this_frame))
2084 // Local copy of the current frame's first pass stats.
2085 this_frame_copy = this_frame;
2087 // Keyframe and section processing.
2088 if (rc->frames_to_key == 0 ||
2089 (cpi->frame_flags & FRAMEFLAGS_KEY)) {
2090 // Define next KF group and assign bits to it.
2091 find_next_key_frame(cpi, &this_frame_copy);
2093 cm->frame_type = INTER_FRAME;
2096 if (is_spatial_svc) {
2097 if (cpi->svc.spatial_layer_id == 0) {
2098 lc->is_key_frame = (cm->frame_type == KEY_FRAME);
2100 cm->frame_type = INTER_FRAME;
2101 lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
2103 if (lc->is_key_frame) {
2104 cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
2109 // Define a new GF/ARF group. (Should always enter here for key frames).
2110 if (rc->frames_till_gf_update_due == 0) {
2111 define_gf_group(cpi, &this_frame_copy);
2113 if (twopass->gf_zeromotion_pct > 995) {
2114 // As long as max_thresh for encode breakout is small enough, it is ok
2115 // to enable it for show frame, i.e. set allow_encode_breakout to
2116 // ENCODE_BREAKOUT_LIMITED.
2117 if (!cm->show_frame)
2118 cpi->allow_encode_breakout = ENCODE_BREAKOUT_DISABLED;
2120 cpi->allow_encode_breakout = ENCODE_BREAKOUT_LIMITED;
2123 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
2124 cpi->refresh_golden_frame = 1;
2128 FIRSTPASS_STATS next_frame;
2129 if (lookup_next_frame_stats(twopass, &next_frame) != EOF) {
2130 twopass->next_iiratio = (int)(next_frame.intra_error /
2131 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
2135 configure_buffer_updates(cpi);
2137 target_rate = twopass->gf_group.bit_allocation[twopass->gf_group.index];
2138 if (cpi->common.frame_type == KEY_FRAME)
2139 target_rate = vp9_rc_clamp_iframe_target_size(cpi, target_rate);
2141 target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
2143 rc->base_frame_target = target_rate;
2144 #ifdef LONG_TERM_VBR_CORRECTION
2145 // Correction to rate target based on prior over or under shoot.
2146 if (cpi->oxcf.rc_mode == VPX_VBR)
2147 vbr_rate_correction(&target_rate, rc->vbr_bits_off_target);
2149 vp9_rc_set_frame_target(cpi, target_rate);
2151 // Update the total stats remaining structure.
2152 subtract_stats(&twopass->total_left_stats, &this_frame);
2155 void vp9_twopass_postencode_update(VP9_COMP *cpi) {
2156 TWO_PASS *const twopass = &cpi->twopass;
2157 RATE_CONTROL *const rc = &cpi->rc;
2158 #ifdef LONG_TERM_VBR_CORRECTION
2159 // In this experimental mode, the VBR correction is done exclusively through
2160 // rc->vbr_bits_off_target. Based on the sign of this value, a limited %
2161 // adjustment is made to the target rate of subsequent frames, to try and
2162 // push it back towards 0. This mode is less likely to suffer from
2163 // extreme behaviour at the end of a clip or group of frames.
2164 const int bits_used = rc->base_frame_target;
2165 rc->vbr_bits_off_target += rc->base_frame_target - rc->projected_frame_size;
2167 // In this mode, VBR correction is acheived by altering bits_left,
2168 // kf_group_bits & gf_group_bits to reflect any deviation from the target
2169 // rate in this frame. This alters the allocation of bits to the
2170 // remaning frames in the group / clip.
2172 // This method can give rise to unstable behaviour near the end of a clip
2173 // or kf/gf group of frames where any accumulated error is corrected over an
2174 // ever decreasing number of frames. Hence we change the balance of target
2175 // vs. actual bitrate gradually as we progress towards the end of the
2176 // sequence in order to mitigate this effect.
2177 const double progress =
2178 (double)(twopass->stats_in - twopass->stats_in_start) /
2179 (twopass->stats_in_end - twopass->stats_in_start);
2180 const int bits_used = (int)(progress * rc->this_frame_target +
2181 (1.0 - progress) * rc->projected_frame_size);
2184 twopass->bits_left = MAX(twopass->bits_left - bits_used, 0);
2186 #ifdef LONG_TERM_VBR_CORRECTION
2187 if (cpi->common.frame_type != KEY_FRAME &&
2188 !vp9_is_upper_layer_key_frame(cpi)) {
2190 if (cpi->common.frame_type == KEY_FRAME ||
2191 vp9_is_upper_layer_key_frame(cpi)) {
2192 // For key frames kf_group_bits already had the target bits subtracted out.
2193 // So now update to the correct value based on the actual bits used.
2194 twopass->kf_group_bits += rc->this_frame_target - bits_used;
2197 twopass->kf_group_bits -= bits_used;
2199 twopass->kf_group_bits = MAX(twopass->kf_group_bits, 0);
2201 // Increment the gf group index ready for the next frame.
2202 ++twopass->gf_group.index;