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"
25 #include "vp9/encoder/vp9_aq_variance.h"
26 #include "vp9/encoder/vp9_block.h"
27 #include "vp9/encoder/vp9_encodeframe.h"
28 #include "vp9/encoder/vp9_encodemb.h"
29 #include "vp9/encoder/vp9_encodemv.h"
30 #include "vp9/encoder/vp9_encoder.h"
31 #include "vp9/encoder/vp9_extend.h"
32 #include "vp9/encoder/vp9_firstpass.h"
33 #include "vp9/encoder/vp9_mcomp.h"
34 #include "vp9/encoder/vp9_quantize.h"
35 #include "vp9/encoder/vp9_rd.h"
36 #include "vp9/encoder/vp9_variance.h"
39 #define ARF_STATS_OUTPUT 0
41 #define BOOST_BREAKOUT 12.5
42 #define BOOST_FACTOR 12.5
43 #define ERR_DIVISOR 128.0
44 #define FACTOR_PT_LOW 0.70
45 #define FACTOR_PT_HIGH 0.90
46 #define FIRST_PASS_Q 10.0
47 #define GF_MAX_BOOST 96.0
48 #define INTRA_MODE_PENALTY 1024
49 #define KF_MAX_BOOST 128.0
50 #define MIN_ARF_GF_BOOST 240
51 #define MIN_DECAY_FACTOR 0.01
52 #define MIN_GF_INTERVAL 4
53 #define MIN_KF_BOOST 300
54 #define NEW_MV_MODE_PENALTY 32
55 #define SVC_FACTOR_PT_LOW 0.45
57 #define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x) - 0.000001 : (x) + 0.000001)
60 unsigned int arf_count = 0;
63 static void swap_yv12(YV12_BUFFER_CONFIG *a, YV12_BUFFER_CONFIG *b) {
64 YV12_BUFFER_CONFIG temp = *a;
69 // Resets the first pass file to the given position using a relative seek from
70 // the current position.
71 static void reset_fpf_position(TWO_PASS *p,
72 const FIRSTPASS_STATS *position) {
73 p->stats_in = position;
76 // Read frame stats at an offset from the current position.
77 static const FIRSTPASS_STATS *read_frame_stats(const TWO_PASS *p, int offset) {
78 if ((offset >= 0 && p->stats_in + offset >= p->stats_in_end) ||
79 (offset < 0 && p->stats_in + offset < p->stats_in_start)) {
83 return &p->stats_in[offset];
86 static int input_stats(TWO_PASS *p, FIRSTPASS_STATS *fps) {
87 if (p->stats_in >= p->stats_in_end)
95 static void output_stats(FIRSTPASS_STATS *stats,
96 struct vpx_codec_pkt_list *pktlist) {
97 struct vpx_codec_cx_pkt pkt;
98 pkt.kind = VPX_CODEC_STATS_PKT;
99 pkt.data.twopass_stats.buf = stats;
100 pkt.data.twopass_stats.sz = sizeof(FIRSTPASS_STATS);
101 vpx_codec_pkt_list_add(pktlist, &pkt);
107 fpfile = fopen("firstpass.stt", "a");
109 fprintf(fpfile, "%12.0f %12.0f %12.0f %12.0f %12.4f %12.4f"
110 "%12.4f %12.4f %12.4f %12.4f %12.4f %12.4f %12.4f"
111 "%12.0f %12.0f %12.4f %12.0f %12.0f %12.4f\n",
115 stats->sr_coded_error,
118 stats->pcnt_second_ref,
126 stats->mv_in_out_count,
135 #if CONFIG_FP_MB_STATS
136 static void output_fpmb_stats(uint8_t *this_frame_mb_stats, VP9_COMMON *cm,
137 struct vpx_codec_pkt_list *pktlist) {
138 struct vpx_codec_cx_pkt pkt;
139 pkt.kind = VPX_CODEC_FPMB_STATS_PKT;
140 pkt.data.firstpass_mb_stats.buf = this_frame_mb_stats;
141 pkt.data.firstpass_mb_stats.sz = cm->MBs * sizeof(uint8_t);
142 vpx_codec_pkt_list_add(pktlist, &pkt);
146 static void zero_stats(FIRSTPASS_STATS *section) {
147 section->frame = 0.0;
148 section->intra_error = 0.0;
149 section->coded_error = 0.0;
150 section->sr_coded_error = 0.0;
151 section->pcnt_inter = 0.0;
152 section->pcnt_motion = 0.0;
153 section->pcnt_second_ref = 0.0;
154 section->pcnt_neutral = 0.0;
156 section->mvr_abs = 0.0;
158 section->mvc_abs = 0.0;
161 section->mv_in_out_count = 0.0;
162 section->new_mv_count = 0.0;
163 section->count = 0.0;
164 section->duration = 1.0;
165 section->spatial_layer_id = 0;
168 static void accumulate_stats(FIRSTPASS_STATS *section,
169 const FIRSTPASS_STATS *frame) {
170 section->frame += frame->frame;
171 section->spatial_layer_id = frame->spatial_layer_id;
172 section->intra_error += frame->intra_error;
173 section->coded_error += frame->coded_error;
174 section->sr_coded_error += frame->sr_coded_error;
175 section->pcnt_inter += frame->pcnt_inter;
176 section->pcnt_motion += frame->pcnt_motion;
177 section->pcnt_second_ref += frame->pcnt_second_ref;
178 section->pcnt_neutral += frame->pcnt_neutral;
179 section->MVr += frame->MVr;
180 section->mvr_abs += frame->mvr_abs;
181 section->MVc += frame->MVc;
182 section->mvc_abs += frame->mvc_abs;
183 section->MVrv += frame->MVrv;
184 section->MVcv += frame->MVcv;
185 section->mv_in_out_count += frame->mv_in_out_count;
186 section->new_mv_count += frame->new_mv_count;
187 section->count += frame->count;
188 section->duration += frame->duration;
191 static void subtract_stats(FIRSTPASS_STATS *section,
192 const FIRSTPASS_STATS *frame) {
193 section->frame -= frame->frame;
194 section->intra_error -= frame->intra_error;
195 section->coded_error -= frame->coded_error;
196 section->sr_coded_error -= frame->sr_coded_error;
197 section->pcnt_inter -= frame->pcnt_inter;
198 section->pcnt_motion -= frame->pcnt_motion;
199 section->pcnt_second_ref -= frame->pcnt_second_ref;
200 section->pcnt_neutral -= frame->pcnt_neutral;
201 section->MVr -= frame->MVr;
202 section->mvr_abs -= frame->mvr_abs;
203 section->MVc -= frame->MVc;
204 section->mvc_abs -= frame->mvc_abs;
205 section->MVrv -= frame->MVrv;
206 section->MVcv -= frame->MVcv;
207 section->mv_in_out_count -= frame->mv_in_out_count;
208 section->new_mv_count -= frame->new_mv_count;
209 section->count -= frame->count;
210 section->duration -= frame->duration;
214 // Calculate a modified Error used in distributing bits between easier and
216 static double calculate_modified_err(const TWO_PASS *twopass,
217 const VP9EncoderConfig *oxcf,
218 const FIRSTPASS_STATS *this_frame) {
219 const FIRSTPASS_STATS *const stats = &twopass->total_stats;
220 const double av_err = stats->coded_error / stats->count;
221 const double modified_error = av_err *
222 pow(this_frame->coded_error / DOUBLE_DIVIDE_CHECK(av_err),
223 oxcf->two_pass_vbrbias / 100.0);
224 return fclamp(modified_error,
225 twopass->modified_error_min, twopass->modified_error_max);
228 // This function returns the maximum target rate per frame.
229 static int frame_max_bits(const RATE_CONTROL *rc,
230 const VP9EncoderConfig *oxcf) {
231 int64_t max_bits = ((int64_t)rc->avg_frame_bandwidth *
232 (int64_t)oxcf->two_pass_vbrmax_section) / 100;
235 else if (max_bits > rc->max_frame_bandwidth)
236 max_bits = rc->max_frame_bandwidth;
238 return (int)max_bits;
241 void vp9_init_first_pass(VP9_COMP *cpi) {
242 zero_stats(&cpi->twopass.total_stats);
245 void vp9_end_first_pass(VP9_COMP *cpi) {
246 if (is_two_pass_svc(cpi)) {
248 for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
249 output_stats(&cpi->svc.layer_context[i].twopass.total_stats,
250 cpi->output_pkt_list);
253 output_stats(&cpi->twopass.total_stats, cpi->output_pkt_list);
257 static vp9_variance_fn_t get_block_variance_fn(BLOCK_SIZE bsize) {
270 static unsigned int get_prediction_error(BLOCK_SIZE bsize,
271 const struct buf_2d *src,
272 const struct buf_2d *ref) {
274 const vp9_variance_fn_t fn = get_block_variance_fn(bsize);
275 fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
279 #if CONFIG_VP9_HIGHBITDEPTH
280 static vp9_variance_fn_t highbd_get_block_variance_fn(BLOCK_SIZE bsize,
286 return vp9_highbd_mse8x8;
288 return vp9_highbd_mse16x8;
290 return vp9_highbd_mse8x16;
292 return vp9_highbd_mse16x16;
298 return vp9_highbd_10_mse8x8;
300 return vp9_highbd_10_mse16x8;
302 return vp9_highbd_10_mse8x16;
304 return vp9_highbd_10_mse16x16;
310 return vp9_highbd_12_mse8x8;
312 return vp9_highbd_12_mse16x8;
314 return vp9_highbd_12_mse8x16;
316 return vp9_highbd_12_mse16x16;
322 static unsigned int highbd_get_prediction_error(BLOCK_SIZE bsize,
323 const struct buf_2d *src,
324 const struct buf_2d *ref,
327 const vp9_variance_fn_t fn = highbd_get_block_variance_fn(bsize, bd);
328 fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
331 #endif // CONFIG_VP9_HIGHBITDEPTH
333 // Refine the motion search range according to the frame dimension
334 // for first pass test.
335 static int get_search_range(const VP9_COMMON *cm) {
337 const int dim = MIN(cm->width, cm->height);
339 while ((dim << sr) < MAX_FULL_PEL_VAL)
344 static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
345 const MV *ref_mv, MV *best_mv,
346 int *best_motion_err) {
347 MACROBLOCKD *const xd = &x->e_mbd;
349 MV ref_mv_full = {ref_mv->row >> 3, ref_mv->col >> 3};
350 int num00, tmp_err, n;
351 const BLOCK_SIZE bsize = xd->mi[0].src_mi->mbmi.sb_type;
352 vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
353 const int new_mv_mode_penalty = NEW_MV_MODE_PENALTY;
356 int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
357 const int sr = get_search_range(&cpi->common);
361 // Override the default variance function to use MSE.
362 v_fn_ptr.vf = get_block_variance_fn(bsize);
363 #if CONFIG_VP9_HIGHBITDEPTH
364 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
365 v_fn_ptr.vf = highbd_get_block_variance_fn(bsize, xd->bd);
367 #endif // CONFIG_VP9_HIGHBITDEPTH
369 // Center the initial step/diamond search on best mv.
370 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
372 x->sadperbit16, &num00, &v_fn_ptr, ref_mv);
373 if (tmp_err < INT_MAX)
374 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
375 if (tmp_err < INT_MAX - new_mv_mode_penalty)
376 tmp_err += new_mv_mode_penalty;
378 if (tmp_err < *best_motion_err) {
379 *best_motion_err = tmp_err;
383 // Carry out further step/diamond searches as necessary.
387 while (n < further_steps) {
393 tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
394 step_param + n, x->sadperbit16,
395 &num00, &v_fn_ptr, ref_mv);
396 if (tmp_err < INT_MAX)
397 tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
398 if (tmp_err < INT_MAX - new_mv_mode_penalty)
399 tmp_err += new_mv_mode_penalty;
401 if (tmp_err < *best_motion_err) {
402 *best_motion_err = tmp_err;
409 static BLOCK_SIZE get_bsize(const VP9_COMMON *cm, int mb_row, int mb_col) {
410 if (2 * mb_col + 1 < cm->mi_cols) {
411 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_16X16
414 return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_8X16
419 static int find_fp_qindex(vpx_bit_depth_t bit_depth) {
422 for (i = 0; i < QINDEX_RANGE; ++i)
423 if (vp9_convert_qindex_to_q(i, bit_depth) >= FIRST_PASS_Q)
426 if (i == QINDEX_RANGE)
432 static void set_first_pass_params(VP9_COMP *cpi) {
433 VP9_COMMON *const cm = &cpi->common;
434 if (!cpi->refresh_alt_ref_frame &&
435 (cm->current_video_frame == 0 ||
436 (cpi->frame_flags & FRAMEFLAGS_KEY))) {
437 cm->frame_type = KEY_FRAME;
439 cm->frame_type = INTER_FRAME;
441 // Do not use periodic key frames.
442 cpi->rc.frames_to_key = INT_MAX;
445 void vp9_first_pass(VP9_COMP *cpi, const struct lookahead_entry *source) {
447 MACROBLOCK *const x = &cpi->mb;
448 VP9_COMMON *const cm = &cpi->common;
449 MACROBLOCKD *const xd = &x->e_mbd;
451 struct macroblock_plane *const p = x->plane;
452 struct macroblockd_plane *const pd = xd->plane;
453 const PICK_MODE_CONTEXT *ctx = &cpi->pc_root->none;
456 int recon_yoffset, recon_uvoffset;
457 YV12_BUFFER_CONFIG *const lst_yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
458 YV12_BUFFER_CONFIG *gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
459 YV12_BUFFER_CONFIG *const new_yv12 = get_frame_new_buffer(cm);
460 int recon_y_stride = lst_yv12->y_stride;
461 int recon_uv_stride = lst_yv12->uv_stride;
462 int uv_mb_height = 16 >> (lst_yv12->y_height > lst_yv12->uv_height);
463 int64_t intra_error = 0;
464 int64_t coded_error = 0;
465 int64_t sr_coded_error = 0;
467 int sum_mvr = 0, sum_mvc = 0;
468 int sum_mvr_abs = 0, sum_mvc_abs = 0;
469 int64_t sum_mvrs = 0, sum_mvcs = 0;
472 int second_ref_count = 0;
473 const int intrapenalty = INTRA_MODE_PENALTY;
474 int neutral_count = 0;
475 int new_mv_count = 0;
476 int sum_in_vectors = 0;
478 TWO_PASS *twopass = &cpi->twopass;
479 const MV zero_mv = {0, 0};
480 const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;
481 LAYER_CONTEXT *const lc = is_two_pass_svc(cpi) ?
482 &cpi->svc.layer_context[cpi->svc.spatial_layer_id] : NULL;
484 #if CONFIG_FP_MB_STATS
485 if (cpi->use_fp_mb_stats) {
486 vp9_zero_array(cpi->twopass.frame_mb_stats_buf, cm->MBs);
490 vp9_clear_system_state();
492 set_first_pass_params(cpi);
493 vp9_set_quantizer(cm, find_fp_qindex(cm->bit_depth));
496 twopass = &lc->twopass;
498 cpi->lst_fb_idx = cpi->svc.spatial_layer_id;
499 cpi->ref_frame_flags = VP9_LAST_FLAG;
501 if (cpi->svc.number_spatial_layers + cpi->svc.spatial_layer_id <
504 cpi->svc.number_spatial_layers + cpi->svc.spatial_layer_id;
505 cpi->ref_frame_flags |= VP9_GOLD_FLAG;
506 cpi->refresh_golden_frame = (lc->current_video_frame_in_layer == 0);
508 cpi->refresh_golden_frame = 0;
511 if (lc->current_video_frame_in_layer == 0)
512 cpi->ref_frame_flags = 0;
514 vp9_scale_references(cpi);
516 // Use either last frame or alt frame for motion search.
517 if (cpi->ref_frame_flags & VP9_LAST_FLAG) {
518 first_ref_buf = vp9_get_scaled_ref_frame(cpi, LAST_FRAME);
519 if (first_ref_buf == NULL)
520 first_ref_buf = get_ref_frame_buffer(cpi, LAST_FRAME);
523 if (cpi->ref_frame_flags & VP9_GOLD_FLAG) {
525 cm->ref_frame_map[get_ref_frame_idx(cpi, GOLDEN_FRAME)];
526 const int scaled_idx = cpi->scaled_ref_idx[GOLDEN_FRAME - 1];
528 gld_yv12 = (scaled_idx != ref_idx) ? &cm->frame_bufs[scaled_idx].buf :
529 get_ref_frame_buffer(cpi, GOLDEN_FRAME);
534 recon_y_stride = new_yv12->y_stride;
535 recon_uv_stride = new_yv12->uv_stride;
536 uv_mb_height = 16 >> (new_yv12->y_height > new_yv12->uv_height);
539 (cpi->ref_frame_flags & VP9_LAST_FLAG) ? LAST_FRAME: NONE,
540 (cpi->ref_frame_flags & VP9_GOLD_FLAG) ? GOLDEN_FRAME : NONE);
542 cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
543 &cpi->scaled_source);
546 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
548 vp9_setup_src_planes(x, cpi->Source, 0, 0);
549 vp9_setup_pre_planes(xd, 0, first_ref_buf, 0, 0, NULL);
550 vp9_setup_dst_planes(xd->plane, new_yv12, 0, 0);
553 xd->mi[0].src_mi = &xd->mi[0];
555 vp9_frame_init_quantizer(cpi);
557 for (i = 0; i < MAX_MB_PLANE; ++i) {
558 p[i].coeff = ctx->coeff_pbuf[i][1];
559 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
560 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
561 p[i].eobs = ctx->eobs_pbuf[i][1];
565 vp9_init_mv_probs(cm);
566 vp9_initialize_rd_consts(cpi);
568 // Tiling is ignored in the first pass.
569 vp9_tile_init(&tile, cm, 0, 0);
571 for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
572 MV best_ref_mv = {0, 0};
574 // Reset above block coeffs.
575 xd->up_available = (mb_row != 0);
576 recon_yoffset = (mb_row * recon_y_stride * 16);
577 recon_uvoffset = (mb_row * recon_uv_stride * uv_mb_height);
579 // Set up limit values for motion vectors to prevent them extending
580 // outside the UMV borders.
581 x->mv_row_min = -((mb_row * 16) + BORDER_MV_PIXELS_B16);
582 x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16)
583 + BORDER_MV_PIXELS_B16;
585 for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
587 const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
588 double error_weight = 1.0;
589 const BLOCK_SIZE bsize = get_bsize(cm, mb_row, mb_col);
590 #if CONFIG_FP_MB_STATS
591 const int mb_index = mb_row * cm->mb_cols + mb_col;
594 vp9_clear_system_state();
596 xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
597 xd->plane[1].dst.buf = new_yv12->u_buffer + recon_uvoffset;
598 xd->plane[2].dst.buf = new_yv12->v_buffer + recon_uvoffset;
599 xd->left_available = (mb_col != 0);
600 xd->mi[0].src_mi->mbmi.sb_type = bsize;
601 xd->mi[0].src_mi->mbmi.ref_frame[0] = INTRA_FRAME;
602 set_mi_row_col(xd, &tile,
603 mb_row << 1, num_8x8_blocks_high_lookup[bsize],
604 mb_col << 1, num_8x8_blocks_wide_lookup[bsize],
605 cm->mi_rows, cm->mi_cols);
607 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
608 const int energy = vp9_block_energy(cpi, x, bsize);
609 error_weight = vp9_vaq_inv_q_ratio(energy);
612 // Do intra 16x16 prediction.
614 xd->mi[0].src_mi->mbmi.mode = DC_PRED;
615 xd->mi[0].src_mi->mbmi.tx_size = use_dc_pred ?
616 (bsize >= BLOCK_16X16 ? TX_16X16 : TX_8X8) : TX_4X4;
617 vp9_encode_intra_block_plane(x, bsize, 0);
618 this_error = vp9_get_mb_ss(x->plane[0].src_diff);
619 #if CONFIG_VP9_HIGHBITDEPTH
620 if (cm->use_highbitdepth) {
621 switch (cm->bit_depth) {
631 assert(0 && "cm->bit_depth should be VPX_BITS_8, "
632 "VPX_BITS_10 or VPX_BITS_12");
636 #endif // CONFIG_VP9_HIGHBITDEPTH
638 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
639 vp9_clear_system_state();
640 this_error = (int)(this_error * error_weight);
643 // Intrapenalty below deals with situations where the intra and inter
644 // error scores are very low (e.g. a plain black frame).
645 // We do not have special cases in first pass for 0,0 and nearest etc so
646 // all inter modes carry an overhead cost estimate for the mv.
647 // When the error score is very low this causes us to pick all or lots of
648 // INTRA modes and throw lots of key frames.
649 // This penalty adds a cost matching that of a 0,0 mv to the intra case.
650 this_error += intrapenalty;
652 // Accumulate the intra error.
653 intra_error += (int64_t)this_error;
655 #if CONFIG_FP_MB_STATS
656 if (cpi->use_fp_mb_stats) {
658 cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
662 // Set up limit values for motion vectors to prevent them extending
663 // outside the UMV borders.
664 x->mv_col_min = -((mb_col * 16) + BORDER_MV_PIXELS_B16);
665 x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + BORDER_MV_PIXELS_B16;
667 // Other than for the first frame do a motion search.
668 if ((lc == NULL && cm->current_video_frame > 0) ||
669 (lc != NULL && lc->current_video_frame_in_layer > 0)) {
670 int tmp_err, motion_error, raw_motion_error;
671 // Assume 0,0 motion with no mv overhead.
672 MV mv = {0, 0} , tmp_mv = {0, 0};
673 struct buf_2d unscaled_last_source_buf_2d;
675 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
676 #if CONFIG_VP9_HIGHBITDEPTH
677 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
678 motion_error = highbd_get_prediction_error(
679 bsize, &x->plane[0].src, &xd->plane[0].pre[0], xd->bd);
681 motion_error = get_prediction_error(
682 bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
685 motion_error = get_prediction_error(
686 bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
687 #endif // CONFIG_VP9_HIGHBITDEPTH
689 // Compute the motion error of the 0,0 motion using the last source
690 // frame as the reference. Skip the further motion search on
691 // reconstructed frame if this error is small.
692 unscaled_last_source_buf_2d.buf =
693 cpi->unscaled_last_source->y_buffer + recon_yoffset;
694 unscaled_last_source_buf_2d.stride =
695 cpi->unscaled_last_source->y_stride;
696 #if CONFIG_VP9_HIGHBITDEPTH
697 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
698 raw_motion_error = highbd_get_prediction_error(
699 bsize, &x->plane[0].src, &unscaled_last_source_buf_2d, xd->bd);
701 raw_motion_error = get_prediction_error(
702 bsize, &x->plane[0].src, &unscaled_last_source_buf_2d);
705 raw_motion_error = get_prediction_error(
706 bsize, &x->plane[0].src, &unscaled_last_source_buf_2d);
707 #endif // CONFIG_VP9_HIGHBITDEPTH
709 // TODO(pengchong): Replace the hard-coded threshold
710 if (raw_motion_error > 25 || lc != NULL) {
711 // Test last reference frame using the previous best mv as the
712 // starting point (best reference) for the search.
713 first_pass_motion_search(cpi, x, &best_ref_mv, &mv, &motion_error);
714 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
715 vp9_clear_system_state();
716 motion_error = (int)(motion_error * error_weight);
719 // If the current best reference mv is not centered on 0,0 then do a
720 // 0,0 based search as well.
721 if (!is_zero_mv(&best_ref_mv)) {
723 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv, &tmp_err);
724 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
725 vp9_clear_system_state();
726 tmp_err = (int)(tmp_err * error_weight);
729 if (tmp_err < motion_error) {
730 motion_error = tmp_err;
735 // Search in an older reference frame.
736 if (((lc == NULL && cm->current_video_frame > 1) ||
737 (lc != NULL && lc->current_video_frame_in_layer > 1))
738 && gld_yv12 != NULL) {
739 // Assume 0,0 motion with no mv overhead.
742 xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
743 #if CONFIG_VP9_HIGHBITDEPTH
744 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
745 gf_motion_error = highbd_get_prediction_error(
746 bsize, &x->plane[0].src, &xd->plane[0].pre[0], xd->bd);
748 gf_motion_error = get_prediction_error(
749 bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
752 gf_motion_error = get_prediction_error(
753 bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
754 #endif // CONFIG_VP9_HIGHBITDEPTH
756 first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv,
758 if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
759 vp9_clear_system_state();
760 gf_motion_error = (int)(gf_motion_error * error_weight);
763 if (gf_motion_error < motion_error && gf_motion_error < this_error)
766 // Reset to last frame as reference buffer.
767 xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
768 xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
769 xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
771 // In accumulating a score for the older reference frame take the
772 // best of the motion predicted score and the intra coded error
773 // (just as will be done for) accumulation of "coded_error" for
775 if (gf_motion_error < this_error)
776 sr_coded_error += gf_motion_error;
778 sr_coded_error += this_error;
780 sr_coded_error += motion_error;
783 sr_coded_error += motion_error;
786 // Start by assuming that intra mode is best.
790 #if CONFIG_FP_MB_STATS
791 if (cpi->use_fp_mb_stats) {
792 // intra predication statistics
793 cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
794 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_DCINTRA_MASK;
795 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_MOTION_ZERO_MASK;
796 if (this_error > FPMB_ERROR_LARGE_TH) {
797 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_LARGE_MASK;
798 } else if (this_error < FPMB_ERROR_SMALL_TH) {
799 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_SMALL_MASK;
804 if (motion_error <= this_error) {
805 // Keep a count of cases where the inter and intra were very close
806 // and very low. This helps with scene cut detection for example in
807 // cropped clips with black bars at the sides or top and bottom.
808 if (((this_error - intrapenalty) * 9 <= motion_error * 10) &&
809 this_error < 2 * intrapenalty)
814 this_error = motion_error;
815 xd->mi[0].src_mi->mbmi.mode = NEWMV;
816 xd->mi[0].src_mi->mbmi.mv[0].as_mv = mv;
817 xd->mi[0].src_mi->mbmi.tx_size = TX_4X4;
818 xd->mi[0].src_mi->mbmi.ref_frame[0] = LAST_FRAME;
819 xd->mi[0].src_mi->mbmi.ref_frame[1] = NONE;
820 vp9_build_inter_predictors_sby(xd, mb_row << 1, mb_col << 1, bsize);
821 vp9_encode_sby_pass1(x, bsize);
823 sum_mvr_abs += abs(mv.row);
825 sum_mvc_abs += abs(mv.col);
826 sum_mvrs += mv.row * mv.row;
827 sum_mvcs += mv.col * mv.col;
832 #if CONFIG_FP_MB_STATS
833 if (cpi->use_fp_mb_stats) {
834 // inter predication statistics
835 cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
836 cpi->twopass.frame_mb_stats_buf[mb_index] &= ~FPMB_DCINTRA_MASK;
837 cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_MOTION_ZERO_MASK;
838 if (this_error > FPMB_ERROR_LARGE_TH) {
839 cpi->twopass.frame_mb_stats_buf[mb_index] |=
840 FPMB_ERROR_LARGE_MASK;
841 } else if (this_error < FPMB_ERROR_SMALL_TH) {
842 cpi->twopass.frame_mb_stats_buf[mb_index] |=
843 FPMB_ERROR_SMALL_MASK;
848 if (!is_zero_mv(&mv)) {
851 #if CONFIG_FP_MB_STATS
852 if (cpi->use_fp_mb_stats) {
853 cpi->twopass.frame_mb_stats_buf[mb_index] &=
854 ~FPMB_MOTION_ZERO_MASK;
855 // check estimated motion direction
856 if (mv.as_mv.col > 0 && mv.as_mv.col >= abs(mv.as_mv.row)) {
858 cpi->twopass.frame_mb_stats_buf[mb_index] |=
859 FPMB_MOTION_RIGHT_MASK;
860 } else if (mv.as_mv.row < 0 &&
861 abs(mv.as_mv.row) >= abs(mv.as_mv.col)) {
863 cpi->twopass.frame_mb_stats_buf[mb_index] |=
865 } else if (mv.as_mv.col < 0 &&
866 abs(mv.as_mv.col) >= abs(mv.as_mv.row)) {
868 cpi->twopass.frame_mb_stats_buf[mb_index] |=
869 FPMB_MOTION_LEFT_MASK;
872 cpi->twopass.frame_mb_stats_buf[mb_index] |=
873 FPMB_MOTION_DOWN_MASK;
878 // Non-zero vector, was it different from the last non zero vector?
879 if (!is_equal_mv(&mv, &lastmv))
883 // Does the row vector point inwards or outwards?
884 if (mb_row < cm->mb_rows / 2) {
889 } else if (mb_row > cm->mb_rows / 2) {
896 // Does the col vector point inwards or outwards?
897 if (mb_col < cm->mb_cols / 2) {
902 } else if (mb_col > cm->mb_cols / 2) {
911 sr_coded_error += (int64_t)this_error;
913 coded_error += (int64_t)this_error;
915 // Adjust to the next column of MBs.
916 x->plane[0].src.buf += 16;
917 x->plane[1].src.buf += uv_mb_height;
918 x->plane[2].src.buf += uv_mb_height;
921 recon_uvoffset += uv_mb_height;
924 // Adjust to the next row of MBs.
925 x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
926 x->plane[1].src.buf += uv_mb_height * x->plane[1].src.stride -
927 uv_mb_height * cm->mb_cols;
928 x->plane[2].src.buf += uv_mb_height * x->plane[1].src.stride -
929 uv_mb_height * cm->mb_cols;
931 vp9_clear_system_state();
934 vp9_clear_system_state();
937 // The minimum error here insures some bit alocation to frames even
938 // in static regions. The allocation per MB declines for larger formats
939 // where the typical "real" energy per MB also falls.
940 // Initial estimate here uses sqrt(mbs) to define the min_err, where the
941 // number of mbs is propotional to image area.
942 const double min_err = 200 * sqrt(cm->MBs);
944 fps.frame = cm->current_video_frame;
945 fps.spatial_layer_id = cpi->svc.spatial_layer_id;
946 fps.coded_error = (double)(coded_error >> 8) + min_err;
947 fps.sr_coded_error = (double)(sr_coded_error >> 8) + min_err;
948 fps.intra_error = (double)(intra_error >> 8) + min_err;
950 fps.pcnt_inter = (double)intercount / cm->MBs;
951 fps.pcnt_second_ref = (double)second_ref_count / cm->MBs;
952 fps.pcnt_neutral = (double)neutral_count / cm->MBs;
955 fps.MVr = (double)sum_mvr / mvcount;
956 fps.mvr_abs = (double)sum_mvr_abs / mvcount;
957 fps.MVc = (double)sum_mvc / mvcount;
958 fps.mvc_abs = (double)sum_mvc_abs / mvcount;
959 fps.MVrv = ((double)sum_mvrs - (fps.MVr * fps.MVr / mvcount)) / mvcount;
960 fps.MVcv = ((double)sum_mvcs - (fps.MVc * fps.MVc / mvcount)) / mvcount;
961 fps.mv_in_out_count = (double)sum_in_vectors / (mvcount * 2);
962 fps.new_mv_count = new_mv_count;
963 fps.pcnt_motion = (double)mvcount / cm->MBs;
971 fps.mv_in_out_count = 0.0;
972 fps.new_mv_count = 0.0;
973 fps.pcnt_motion = 0.0;
976 // TODO(paulwilkins): Handle the case when duration is set to 0, or
977 // something less than the full time between subsequent values of
978 // cpi->source_time_stamp.
979 fps.duration = (double)(source->ts_end - source->ts_start);
981 // Don't want to do output stats with a stack variable!
982 twopass->this_frame_stats = fps;
983 output_stats(&twopass->this_frame_stats, cpi->output_pkt_list);
984 accumulate_stats(&twopass->total_stats, &fps);
986 #if CONFIG_FP_MB_STATS
987 if (cpi->use_fp_mb_stats) {
988 output_fpmb_stats(twopass->frame_mb_stats_buf, cm, cpi->output_pkt_list);
993 // Copy the previous Last Frame back into gf and and arf buffers if
994 // the prediction is good enough... but also don't allow it to lag too far.
995 if ((twopass->sr_update_lag > 3) ||
996 ((cm->current_video_frame > 0) &&
997 (twopass->this_frame_stats.pcnt_inter > 0.20) &&
998 ((twopass->this_frame_stats.intra_error /
999 DOUBLE_DIVIDE_CHECK(twopass->this_frame_stats.coded_error)) > 2.0))) {
1000 if (gld_yv12 != NULL) {
1001 vp8_yv12_copy_frame(lst_yv12, gld_yv12);
1003 twopass->sr_update_lag = 1;
1005 ++twopass->sr_update_lag;
1008 vp9_extend_frame_borders(new_yv12);
1011 vp9_update_reference_frames(cpi);
1013 // Swap frame pointers so last frame refers to the frame we just compressed.
1014 swap_yv12(lst_yv12, new_yv12);
1017 // Special case for the first frame. Copy into the GF buffer as a second
1019 if (cm->current_video_frame == 0 && gld_yv12 != NULL && lc == NULL) {
1020 vp8_yv12_copy_frame(lst_yv12, gld_yv12);
1023 // Use this to see what the first pass reconstruction looks like.
1027 snprintf(filename, sizeof(filename), "enc%04d.yuv",
1028 (int)cm->current_video_frame);
1030 if (cm->current_video_frame == 0)
1031 recon_file = fopen(filename, "wb");
1033 recon_file = fopen(filename, "ab");
1035 (void)fwrite(lst_yv12->buffer_alloc, lst_yv12->frame_size, 1, recon_file);
1039 ++cm->current_video_frame;
1041 vp9_inc_frame_in_layer(cpi);
1044 static double calc_correction_factor(double err_per_mb,
1049 vpx_bit_depth_t bit_depth) {
1050 const double error_term = err_per_mb / err_divisor;
1052 // Adjustment based on actual quantizer to power term.
1053 const double power_term =
1054 MIN(vp9_convert_qindex_to_q(q, bit_depth) * 0.01 + pt_low, pt_high);
1056 // Calculate correction factor.
1057 if (power_term < 1.0)
1058 assert(error_term >= 0.0);
1060 return fclamp(pow(error_term, power_term), 0.05, 5.0);
1063 // Larger image formats are expected to be a little harder to code relatively
1064 // given the same prediction error score. This in part at least relates to the
1065 // increased size and hence coding cost of motion vectors.
1066 #define EDIV_SIZE_FACTOR 800
1068 static int get_twopass_worst_quality(const VP9_COMP *cpi,
1069 const FIRSTPASS_STATS *stats,
1070 int section_target_bandwidth) {
1071 const RATE_CONTROL *const rc = &cpi->rc;
1072 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1074 if (section_target_bandwidth <= 0) {
1075 return rc->worst_quality; // Highest value allowed
1077 const int num_mbs = cpi->common.MBs;
1078 const double section_err = stats->coded_error / stats->count;
1079 const double err_per_mb = section_err / num_mbs;
1080 const double speed_term = 1.0 + 0.04 * oxcf->speed;
1081 const double ediv_size_correction = num_mbs / EDIV_SIZE_FACTOR;
1082 const int target_norm_bits_per_mb = ((uint64_t)section_target_bandwidth <<
1083 BPER_MB_NORMBITS) / num_mbs;
1086 int is_svc_upper_layer = 0;
1087 if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0)
1088 is_svc_upper_layer = 1;
1090 // Try and pick a max Q that will be high enough to encode the
1091 // content at the given rate.
1092 for (q = rc->best_quality; q < rc->worst_quality; ++q) {
1093 const double factor =
1094 calc_correction_factor(err_per_mb, ERR_DIVISOR - ediv_size_correction,
1095 is_svc_upper_layer ? SVC_FACTOR_PT_LOW :
1096 FACTOR_PT_LOW, FACTOR_PT_HIGH, q,
1097 cpi->common.bit_depth);
1098 const int bits_per_mb = vp9_rc_bits_per_mb(INTER_FRAME, q,
1099 factor * speed_term,
1100 cpi->common.bit_depth);
1101 if (bits_per_mb <= target_norm_bits_per_mb)
1105 // Restriction on active max q for constrained quality mode.
1106 if (cpi->oxcf.rc_mode == VPX_CQ)
1107 q = MAX(q, oxcf->cq_level);
1112 extern void vp9_new_framerate(VP9_COMP *cpi, double framerate);
1114 void vp9_init_second_pass(VP9_COMP *cpi) {
1115 SVC *const svc = &cpi->svc;
1116 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1117 const int is_two_pass_svc = (svc->number_spatial_layers > 1) ||
1118 (svc->number_temporal_layers > 1);
1119 TWO_PASS *const twopass = is_two_pass_svc ?
1120 &svc->layer_context[svc->spatial_layer_id].twopass : &cpi->twopass;
1122 FIRSTPASS_STATS *stats;
1124 zero_stats(&twopass->total_stats);
1125 zero_stats(&twopass->total_left_stats);
1127 if (!twopass->stats_in_end)
1130 stats = &twopass->total_stats;
1132 *stats = *twopass->stats_in_end;
1133 twopass->total_left_stats = *stats;
1135 frame_rate = 10000000.0 * stats->count / stats->duration;
1136 // Each frame can have a different duration, as the frame rate in the source
1137 // isn't guaranteed to be constant. The frame rate prior to the first frame
1138 // encoded in the second pass is a guess. However, the sum duration is not.
1139 // It is calculated based on the actual durations of all frames from the
1142 if (is_two_pass_svc) {
1143 vp9_update_spatial_layer_framerate(cpi, frame_rate);
1144 twopass->bits_left = (int64_t)(stats->duration *
1145 svc->layer_context[svc->spatial_layer_id].target_bandwidth /
1148 vp9_new_framerate(cpi, frame_rate);
1149 twopass->bits_left = (int64_t)(stats->duration * oxcf->target_bandwidth /
1153 // This variable monitors how far behind the second ref update is lagging.
1154 twopass->sr_update_lag = 1;
1156 // Scan the first pass file and calculate a modified total error based upon
1157 // the bias/power function used to allocate bits.
1159 const double avg_error = stats->coded_error /
1160 DOUBLE_DIVIDE_CHECK(stats->count);
1161 const FIRSTPASS_STATS *s = twopass->stats_in;
1162 double modified_error_total = 0.0;
1163 twopass->modified_error_min = (avg_error *
1164 oxcf->two_pass_vbrmin_section) / 100;
1165 twopass->modified_error_max = (avg_error *
1166 oxcf->two_pass_vbrmax_section) / 100;
1167 while (s < twopass->stats_in_end) {
1168 modified_error_total += calculate_modified_err(twopass, oxcf, s);
1171 twopass->modified_error_left = modified_error_total;
1174 // Reset the vbr bits off target counter
1175 cpi->rc.vbr_bits_off_target = 0;
1177 cpi->rc.rate_error_estimate = 0;
1179 // Static sequence monitor variables.
1180 twopass->kf_zeromotion_pct = 100;
1181 twopass->last_kfgroup_zeromotion_pct = 100;
1184 #define SR_DIFF_PART 0.0015
1185 #define MOTION_AMP_PART 0.003
1186 #define INTRA_PART 0.005
1187 #define DEFAULT_DECAY_LIMIT 0.75
1188 #define LOW_SR_DIFF_TRHESH 0.1
1189 #define SR_DIFF_MAX 128.0
1191 static double get_sr_decay_rate(const VP9_COMMON *cm,
1192 const FIRSTPASS_STATS *frame) {
1193 double sr_diff = (frame->sr_coded_error - frame->coded_error) / cm->MBs;
1194 double sr_decay = 1.0;
1195 const double motion_amplitude_factor =
1196 frame->pcnt_motion * ((frame->mvc_abs + frame->mvr_abs) / 2);
1197 const double pcnt_intra = 100 * (1.0 - frame->pcnt_inter);
1199 if ((sr_diff > LOW_SR_DIFF_TRHESH)) {
1200 sr_diff = MIN(sr_diff, SR_DIFF_MAX);
1201 sr_decay = 1.0 - (SR_DIFF_PART * sr_diff) -
1202 (MOTION_AMP_PART * motion_amplitude_factor) -
1203 (INTRA_PART * pcnt_intra);
1205 return MAX(sr_decay, MIN(DEFAULT_DECAY_LIMIT, frame->pcnt_inter));
1208 // This function gives an estimate of how badly we believe the prediction
1209 // quality is decaying from frame to frame.
1210 static double get_zero_motion_factor(const VP9_COMMON *cm,
1211 const FIRSTPASS_STATS *frame) {
1212 const double zero_motion_pct = frame->pcnt_inter -
1214 double sr_decay = get_sr_decay_rate(cm, frame);
1215 return MIN(sr_decay, zero_motion_pct);
1218 #define ZM_POWER_FACTOR 0.75
1220 static double get_prediction_decay_rate(const VP9_COMMON *cm,
1221 const FIRSTPASS_STATS *next_frame) {
1222 const double sr_decay_rate = get_sr_decay_rate(cm, next_frame);
1223 const double zero_motion_factor =
1224 (0.95 * pow((next_frame->pcnt_inter - next_frame->pcnt_motion),
1227 return MAX(zero_motion_factor,
1228 (sr_decay_rate + ((1.0 - sr_decay_rate) * zero_motion_factor)));
1231 // Function to test for a condition where a complex transition is followed
1232 // by a static section. For example in slide shows where there is a fade
1233 // between slides. This is to help with more optimal kf and gf positioning.
1234 static int detect_transition_to_still(const TWO_PASS *twopass,
1235 int frame_interval, int still_interval,
1236 double loop_decay_rate,
1237 double last_decay_rate) {
1238 // Break clause to detect very still sections after motion
1239 // For example a static image after a fade or other transition
1240 // instead of a clean scene cut.
1241 if (frame_interval > MIN_GF_INTERVAL &&
1242 loop_decay_rate >= 0.999 &&
1243 last_decay_rate < 0.9) {
1246 // Look ahead a few frames to see if static condition persists...
1247 for (j = 0; j < still_interval; ++j) {
1248 const FIRSTPASS_STATS *stats = &twopass->stats_in[j];
1249 if (stats >= twopass->stats_in_end)
1252 if (stats->pcnt_inter - stats->pcnt_motion < 0.999)
1256 // Only if it does do we signal a transition to still.
1257 return j == still_interval;
1263 // This function detects a flash through the high relative pcnt_second_ref
1264 // score in the frame following a flash frame. The offset passed in should
1266 static int detect_flash(const TWO_PASS *twopass, int offset) {
1267 const FIRSTPASS_STATS *const next_frame = read_frame_stats(twopass, offset);
1269 // What we are looking for here is a situation where there is a
1270 // brief break in prediction (such as a flash) but subsequent frames
1271 // are reasonably well predicted by an earlier (pre flash) frame.
1272 // The recovery after a flash is indicated by a high pcnt_second_ref
1273 // compared to pcnt_inter.
1274 return next_frame != NULL &&
1275 next_frame->pcnt_second_ref > next_frame->pcnt_inter &&
1276 next_frame->pcnt_second_ref >= 0.5;
1279 // Update the motion related elements to the GF arf boost calculation.
1280 static void accumulate_frame_motion_stats(const FIRSTPASS_STATS *stats,
1282 double *mv_in_out_accumulator,
1283 double *abs_mv_in_out_accumulator,
1284 double *mv_ratio_accumulator) {
1285 const double pct = stats->pcnt_motion;
1287 // Accumulate Motion In/Out of frame stats.
1288 *mv_in_out = stats->mv_in_out_count * pct;
1289 *mv_in_out_accumulator += *mv_in_out;
1290 *abs_mv_in_out_accumulator += fabs(*mv_in_out);
1292 // Accumulate a measure of how uniform (or conversely how random) the motion
1293 // field is (a ratio of abs(mv) / mv).
1295 const double mvr_ratio = fabs(stats->mvr_abs) /
1296 DOUBLE_DIVIDE_CHECK(fabs(stats->MVr));
1297 const double mvc_ratio = fabs(stats->mvc_abs) /
1298 DOUBLE_DIVIDE_CHECK(fabs(stats->MVc));
1300 *mv_ratio_accumulator += pct * (mvr_ratio < stats->mvr_abs ?
1301 mvr_ratio : stats->mvr_abs);
1302 *mv_ratio_accumulator += pct * (mvc_ratio < stats->mvc_abs ?
1303 mvc_ratio : stats->mvc_abs);
1307 #define BASELINE_ERR_PER_MB 1000.0
1308 static double calc_frame_boost(VP9_COMP *cpi,
1309 const FIRSTPASS_STATS *this_frame,
1310 double this_frame_mv_in_out,
1314 vp9_convert_qindex_to_q(cpi->rc.avg_frame_qindex[INTER_FRAME],
1315 cpi->common.bit_depth);
1316 const double boost_correction = MIN((0.5 + (lq * 0.015)), 1.5);
1318 // Underlying boost factor is based on inter error ratio.
1319 frame_boost = (BASELINE_ERR_PER_MB * cpi->common.MBs) /
1320 DOUBLE_DIVIDE_CHECK(this_frame->coded_error);
1321 frame_boost = frame_boost * BOOST_FACTOR * boost_correction;
1323 // Increase boost for frames where new data coming into frame (e.g. zoom out).
1324 // Slightly reduce boost if there is a net balance of motion out of the frame
1325 // (zoom in). The range for this_frame_mv_in_out is -1.0 to +1.0.
1326 if (this_frame_mv_in_out > 0.0)
1327 frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
1328 // In the extreme case the boost is halved.
1330 frame_boost += frame_boost * (this_frame_mv_in_out / 2.0);
1332 return MIN(frame_boost, max_boost * boost_correction);
1335 static int calc_arf_boost(VP9_COMP *cpi, int offset,
1336 int f_frames, int b_frames,
1337 int *f_boost, int *b_boost) {
1338 TWO_PASS *const twopass = &cpi->twopass;
1340 double boost_score = 0.0;
1341 double mv_ratio_accumulator = 0.0;
1342 double decay_accumulator = 1.0;
1343 double this_frame_mv_in_out = 0.0;
1344 double mv_in_out_accumulator = 0.0;
1345 double abs_mv_in_out_accumulator = 0.0;
1347 int flash_detected = 0;
1349 // Search forward from the proposed arf/next gf position.
1350 for (i = 0; i < f_frames; ++i) {
1351 const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
1352 if (this_frame == NULL)
1355 // Update the motion related elements to the boost calculation.
1356 accumulate_frame_motion_stats(this_frame,
1357 &this_frame_mv_in_out, &mv_in_out_accumulator,
1358 &abs_mv_in_out_accumulator,
1359 &mv_ratio_accumulator);
1361 // We want to discount the flash frame itself and the recovery
1362 // frame that follows as both will have poor scores.
1363 flash_detected = detect_flash(twopass, i + offset) ||
1364 detect_flash(twopass, i + offset + 1);
1366 // Accumulate the effect of prediction quality decay.
1367 if (!flash_detected) {
1368 decay_accumulator *= get_prediction_decay_rate(&cpi->common, this_frame);
1369 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1370 ? MIN_DECAY_FACTOR : decay_accumulator;
1373 boost_score += decay_accumulator * calc_frame_boost(cpi, this_frame,
1374 this_frame_mv_in_out,
1378 *f_boost = (int)boost_score;
1380 // Reset for backward looking loop.
1382 mv_ratio_accumulator = 0.0;
1383 decay_accumulator = 1.0;
1384 this_frame_mv_in_out = 0.0;
1385 mv_in_out_accumulator = 0.0;
1386 abs_mv_in_out_accumulator = 0.0;
1388 // Search backward towards last gf position.
1389 for (i = -1; i >= -b_frames; --i) {
1390 const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
1391 if (this_frame == NULL)
1394 // Update the motion related elements to the boost calculation.
1395 accumulate_frame_motion_stats(this_frame,
1396 &this_frame_mv_in_out, &mv_in_out_accumulator,
1397 &abs_mv_in_out_accumulator,
1398 &mv_ratio_accumulator);
1400 // We want to discount the the flash frame itself and the recovery
1401 // frame that follows as both will have poor scores.
1402 flash_detected = detect_flash(twopass, i + offset) ||
1403 detect_flash(twopass, i + offset + 1);
1405 // Cumulative effect of prediction quality decay.
1406 if (!flash_detected) {
1407 decay_accumulator *= get_prediction_decay_rate(&cpi->common, this_frame);
1408 decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
1409 ? MIN_DECAY_FACTOR : decay_accumulator;
1412 boost_score += decay_accumulator * calc_frame_boost(cpi, this_frame,
1413 this_frame_mv_in_out,
1416 *b_boost = (int)boost_score;
1418 arf_boost = (*f_boost + *b_boost);
1419 if (arf_boost < ((b_frames + f_frames) * 20))
1420 arf_boost = ((b_frames + f_frames) * 20);
1421 arf_boost = MAX(arf_boost, MIN_ARF_GF_BOOST);
1426 // Calculate a section intra ratio used in setting max loop filter.
1427 static int calculate_section_intra_ratio(const FIRSTPASS_STATS *begin,
1428 const FIRSTPASS_STATS *end,
1429 int section_length) {
1430 const FIRSTPASS_STATS *s = begin;
1431 double intra_error = 0.0;
1432 double coded_error = 0.0;
1435 while (s < end && i < section_length) {
1436 intra_error += s->intra_error;
1437 coded_error += s->coded_error;
1442 return (int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
1445 // Calculate the total bits to allocate in this GF/ARF group.
1446 static int64_t calculate_total_gf_group_bits(VP9_COMP *cpi,
1447 double gf_group_err) {
1448 const RATE_CONTROL *const rc = &cpi->rc;
1449 const TWO_PASS *const twopass = &cpi->twopass;
1450 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
1451 int64_t total_group_bits;
1453 // Calculate the bits to be allocated to the group as a whole.
1454 if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0)) {
1455 total_group_bits = (int64_t)(twopass->kf_group_bits *
1456 (gf_group_err / twopass->kf_group_error_left));
1458 total_group_bits = 0;
1461 // Clamp odd edge cases.
1462 total_group_bits = (total_group_bits < 0) ?
1463 0 : (total_group_bits > twopass->kf_group_bits) ?
1464 twopass->kf_group_bits : total_group_bits;
1466 // Clip based on user supplied data rate variability limit.
1467 if (total_group_bits > (int64_t)max_bits * rc->baseline_gf_interval)
1468 total_group_bits = (int64_t)max_bits * rc->baseline_gf_interval;
1470 return total_group_bits;
1473 // Calculate the number bits extra to assign to boosted frames in a group.
1474 static int calculate_boost_bits(int frame_count,
1475 int boost, int64_t total_group_bits) {
1476 int allocation_chunks;
1478 // return 0 for invalid inputs (could arise e.g. through rounding errors)
1479 if (!boost || (total_group_bits <= 0) || (frame_count <= 0) )
1482 allocation_chunks = (frame_count * 100) + boost;
1484 // Prevent overflow.
1486 int divisor = boost >> 10;
1488 allocation_chunks /= divisor;
1491 // Calculate the number of extra bits for use in the boosted frame or frames.
1492 return MAX((int)(((int64_t)boost * total_group_bits) / allocation_chunks), 0);
1495 // Current limit on maximum number of active arfs in a GF/ARF group.
1496 #define MAX_ACTIVE_ARFS 2
1499 // This function indirects the choice of buffers for arfs.
1500 // At the moment the values are fixed but this may change as part of
1501 // the integration process with other codec features that swap buffers around.
1502 static void get_arf_buffer_indices(unsigned char *arf_buffer_indices) {
1503 arf_buffer_indices[0] = ARF_SLOT1;
1504 arf_buffer_indices[1] = ARF_SLOT2;
1507 static void allocate_gf_group_bits(VP9_COMP *cpi, int64_t gf_group_bits,
1508 double group_error, int gf_arf_bits) {
1509 RATE_CONTROL *const rc = &cpi->rc;
1510 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1511 TWO_PASS *const twopass = &cpi->twopass;
1512 GF_GROUP *const gf_group = &twopass->gf_group;
1513 FIRSTPASS_STATS frame_stats;
1515 int frame_index = 1;
1516 int target_frame_size;
1518 const int max_bits = frame_max_bits(&cpi->rc, &cpi->oxcf);
1519 int64_t total_group_bits = gf_group_bits;
1520 double modified_err = 0.0;
1521 double err_fraction;
1522 int mid_boost_bits = 0;
1524 unsigned char arf_buffer_indices[MAX_ACTIVE_ARFS];
1525 int alt_frame_index = frame_index;
1526 int has_temporal_layers = is_two_pass_svc(cpi) &&
1527 cpi->svc.number_temporal_layers > 1;
1529 // Only encode alt reference frame in temporal base layer.
1530 if (has_temporal_layers)
1531 alt_frame_index = cpi->svc.number_temporal_layers;
1533 key_frame = cpi->common.frame_type == KEY_FRAME ||
1534 vp9_is_upper_layer_key_frame(cpi);
1536 get_arf_buffer_indices(arf_buffer_indices);
1538 // For key frames the frame target rate is already set and it
1539 // is also the golden frame.
1541 if (rc->source_alt_ref_active) {
1542 gf_group->update_type[0] = OVERLAY_UPDATE;
1543 gf_group->rf_level[0] = INTER_NORMAL;
1544 gf_group->bit_allocation[0] = 0;
1545 gf_group->arf_update_idx[0] = arf_buffer_indices[0];
1546 gf_group->arf_ref_idx[0] = arf_buffer_indices[0];
1548 gf_group->update_type[0] = GF_UPDATE;
1549 gf_group->rf_level[0] = GF_ARF_STD;
1550 gf_group->bit_allocation[0] = gf_arf_bits;
1551 gf_group->arf_update_idx[0] = arf_buffer_indices[0];
1552 gf_group->arf_ref_idx[0] = arf_buffer_indices[0];
1555 // Step over the golden frame / overlay frame
1556 if (EOF == input_stats(twopass, &frame_stats))
1560 // Deduct the boost bits for arf (or gf if it is not a key frame)
1561 // from the group total.
1562 if (rc->source_alt_ref_pending || !key_frame)
1563 total_group_bits -= gf_arf_bits;
1565 // Store the bits to spend on the ARF if there is one.
1566 if (rc->source_alt_ref_pending) {
1567 gf_group->update_type[alt_frame_index] = ARF_UPDATE;
1568 gf_group->rf_level[alt_frame_index] = GF_ARF_STD;
1569 gf_group->bit_allocation[alt_frame_index] = gf_arf_bits;
1571 if (has_temporal_layers)
1572 gf_group->arf_src_offset[alt_frame_index] =
1573 (unsigned char)(rc->baseline_gf_interval -
1574 cpi->svc.number_temporal_layers);
1576 gf_group->arf_src_offset[alt_frame_index] =
1577 (unsigned char)(rc->baseline_gf_interval - 1);
1579 gf_group->arf_update_idx[alt_frame_index] = arf_buffer_indices[0];
1580 gf_group->arf_ref_idx[alt_frame_index] =
1581 arf_buffer_indices[cpi->multi_arf_last_grp_enabled &&
1582 rc->source_alt_ref_active];
1583 if (!has_temporal_layers)
1586 if (cpi->multi_arf_enabled) {
1587 // Set aside a slot for a level 1 arf.
1588 gf_group->update_type[frame_index] = ARF_UPDATE;
1589 gf_group->rf_level[frame_index] = GF_ARF_LOW;
1590 gf_group->arf_src_offset[frame_index] =
1591 (unsigned char)((rc->baseline_gf_interval >> 1) - 1);
1592 gf_group->arf_update_idx[frame_index] = arf_buffer_indices[1];
1593 gf_group->arf_ref_idx[frame_index] = arf_buffer_indices[0];
1598 // Define middle frame
1599 mid_frame_idx = frame_index + (rc->baseline_gf_interval >> 1) - 1;
1601 // Allocate bits to the other frames in the group.
1602 for (i = 0; i < rc->baseline_gf_interval - 1; ++i) {
1604 if (EOF == input_stats(twopass, &frame_stats))
1607 if (has_temporal_layers && frame_index == alt_frame_index) {
1611 modified_err = calculate_modified_err(twopass, oxcf, &frame_stats);
1613 if (group_error > 0)
1614 err_fraction = modified_err / DOUBLE_DIVIDE_CHECK(group_error);
1618 target_frame_size = (int)((double)total_group_bits * err_fraction);
1620 if (rc->source_alt_ref_pending && cpi->multi_arf_enabled) {
1621 mid_boost_bits += (target_frame_size >> 4);
1622 target_frame_size -= (target_frame_size >> 4);
1624 if (frame_index <= mid_frame_idx)
1627 gf_group->arf_update_idx[frame_index] = arf_buffer_indices[arf_idx];
1628 gf_group->arf_ref_idx[frame_index] = arf_buffer_indices[arf_idx];
1630 target_frame_size = clamp(target_frame_size, 0,
1631 MIN(max_bits, (int)total_group_bits));
1633 gf_group->update_type[frame_index] = LF_UPDATE;
1634 gf_group->rf_level[frame_index] = INTER_NORMAL;
1636 gf_group->bit_allocation[frame_index] = target_frame_size;
1641 // We need to configure the frame at the end of the sequence + 1 that will be
1642 // the start frame for the next group. Otherwise prior to the call to
1643 // vp9_rc_get_second_pass_params() the data will be undefined.
1644 gf_group->arf_update_idx[frame_index] = arf_buffer_indices[0];
1645 gf_group->arf_ref_idx[frame_index] = arf_buffer_indices[0];
1647 if (rc->source_alt_ref_pending) {
1648 gf_group->update_type[frame_index] = OVERLAY_UPDATE;
1649 gf_group->rf_level[frame_index] = INTER_NORMAL;
1651 // Final setup for second arf and its overlay.
1652 if (cpi->multi_arf_enabled) {
1653 gf_group->bit_allocation[2] =
1654 gf_group->bit_allocation[mid_frame_idx] + mid_boost_bits;
1655 gf_group->update_type[mid_frame_idx] = OVERLAY_UPDATE;
1656 gf_group->bit_allocation[mid_frame_idx] = 0;
1659 gf_group->update_type[frame_index] = GF_UPDATE;
1660 gf_group->rf_level[frame_index] = GF_ARF_STD;
1663 // Note whether multi-arf was enabled this group for next time.
1664 cpi->multi_arf_last_grp_enabled = cpi->multi_arf_enabled;
1667 // Analyse and define a gf/arf group.
1668 static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
1669 RATE_CONTROL *const rc = &cpi->rc;
1670 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1671 TWO_PASS *const twopass = &cpi->twopass;
1672 FIRSTPASS_STATS next_frame;
1673 const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
1676 double boost_score = 0.0;
1677 double old_boost_score = 0.0;
1678 double gf_group_err = 0.0;
1679 double gf_first_frame_err = 0.0;
1680 double mod_frame_err = 0.0;
1682 double mv_ratio_accumulator = 0.0;
1683 double decay_accumulator = 1.0;
1684 double zero_motion_accumulator = 1.0;
1686 double loop_decay_rate = 1.00;
1687 double last_loop_decay_rate = 1.00;
1689 double this_frame_mv_in_out = 0.0;
1690 double mv_in_out_accumulator = 0.0;
1691 double abs_mv_in_out_accumulator = 0.0;
1692 double mv_ratio_accumulator_thresh;
1693 unsigned int allow_alt_ref = is_altref_enabled(cpi);
1698 int active_max_gf_interval;
1699 int active_min_gf_interval;
1700 int64_t gf_group_bits;
1701 double gf_group_error_left;
1704 // Reset the GF group data structures unless this is a key
1705 // frame in which case it will already have been done.
1706 if (cpi->common.frame_type != KEY_FRAME) {
1707 vp9_zero(twopass->gf_group);
1710 vp9_clear_system_state();
1711 vp9_zero(next_frame);
1713 // Load stats for the current frame.
1714 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1716 // Note the error of the frame at the start of the group. This will be
1717 // the GF frame error if we code a normal gf.
1718 gf_first_frame_err = mod_frame_err;
1720 // If this is a key frame or the overlay from a previous arf then
1721 // the error score / cost of this frame has already been accounted for.
1722 if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
1723 gf_group_err -= gf_first_frame_err;
1725 // Motion breakout threshold for loop below depends on image size.
1726 mv_ratio_accumulator_thresh = (cpi->common.width + cpi->common.height) / 4.0;
1728 // Set a maximum and minimum interval for the GF group.
1729 // If the image appears almost completely static we can extend beyond this.
1732 (int)(vp9_convert_qindex_to_q(twopass->active_worst_quality,
1733 cpi->common.bit_depth));
1735 (int)(vp9_convert_qindex_to_q(rc->last_boosted_qindex,
1736 cpi->common.bit_depth));
1737 active_min_gf_interval = MIN_GF_INTERVAL + MIN(2, int_max_q / 200);
1738 if (active_min_gf_interval > rc->max_gf_interval)
1739 active_min_gf_interval = rc->max_gf_interval;
1741 if (cpi->multi_arf_allowed) {
1742 active_max_gf_interval = rc->max_gf_interval;
1744 // The value chosen depends on the active Q range. At low Q we have
1745 // bits to spare and are better with a smaller interval and smaller boost.
1746 // At high Q when there are few bits to spare we are better with a longer
1747 // interval to spread the cost of the GF.
1748 active_max_gf_interval = 12 + MIN(4, (int_lbq / 6));
1749 if (active_max_gf_interval > rc->max_gf_interval)
1750 active_max_gf_interval = rc->max_gf_interval;
1755 while (i < rc->static_scene_max_gf_interval && i < rc->frames_to_key) {
1758 // Accumulate error score of frames in this gf group.
1759 mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
1760 gf_group_err += mod_frame_err;
1762 if (EOF == input_stats(twopass, &next_frame))
1765 // Test for the case where there is a brief flash but the prediction
1766 // quality back to an earlier frame is then restored.
1767 flash_detected = detect_flash(twopass, 0);
1769 // Update the motion related elements to the boost calculation.
1770 accumulate_frame_motion_stats(&next_frame,
1771 &this_frame_mv_in_out, &mv_in_out_accumulator,
1772 &abs_mv_in_out_accumulator,
1773 &mv_ratio_accumulator);
1775 // Accumulate the effect of prediction quality decay.
1776 if (!flash_detected) {
1777 last_loop_decay_rate = loop_decay_rate;
1778 loop_decay_rate = get_prediction_decay_rate(&cpi->common, &next_frame);
1780 decay_accumulator = decay_accumulator * loop_decay_rate;
1782 // Monitor for static sections.
1783 zero_motion_accumulator =
1784 MIN(zero_motion_accumulator,
1785 get_zero_motion_factor(&cpi->common, &next_frame));
1787 // Break clause to detect very still sections after motion. For example,
1788 // a static image after a fade or other transition.
1789 if (detect_transition_to_still(twopass, i, 5, loop_decay_rate,
1790 last_loop_decay_rate)) {
1796 // Calculate a boost number for this frame.
1797 boost_score += decay_accumulator * calc_frame_boost(cpi, &next_frame,
1798 this_frame_mv_in_out,
1801 // Break out conditions.
1803 // Break at active_max_gf_interval unless almost totally static.
1804 (i >= active_max_gf_interval && (zero_motion_accumulator < 0.995)) ||
1806 // Don't break out with a very short interval.
1807 (i > active_min_gf_interval) &&
1808 (!flash_detected) &&
1809 ((mv_ratio_accumulator > mv_ratio_accumulator_thresh) ||
1810 (abs_mv_in_out_accumulator > 3.0) ||
1811 (mv_in_out_accumulator < -2.0) ||
1812 ((boost_score - old_boost_score) < BOOST_BREAKOUT)))) {
1813 boost_score = old_boost_score;
1817 *this_frame = next_frame;
1818 old_boost_score = boost_score;
1821 twopass->gf_zeromotion_pct = (int)(zero_motion_accumulator * 1000.0);
1823 // Set the interval until the next gf.
1824 if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
1825 rc->baseline_gf_interval = i - 1;
1827 rc->baseline_gf_interval = i;
1829 // Only encode alt reference frame in temporal base layer. So
1830 // baseline_gf_interval should be multiple of a temporal layer group
1831 // (typically the frame distance between two base layer frames)
1832 if (is_two_pass_svc(cpi) && cpi->svc.number_temporal_layers > 1) {
1833 int count = (1 << (cpi->svc.number_temporal_layers - 1)) - 1;
1834 int new_gf_interval = (rc->baseline_gf_interval + count) & (~count);
1836 for (j = 0; j < new_gf_interval - rc->baseline_gf_interval; ++j) {
1837 if (EOF == input_stats(twopass, this_frame))
1839 gf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
1841 rc->baseline_gf_interval = new_gf_interval;
1844 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
1846 // Should we use the alternate reference frame.
1847 if (allow_alt_ref &&
1848 (i < cpi->oxcf.lag_in_frames) &&
1849 (i >= MIN_GF_INTERVAL)) {
1850 // Calculate the boost for alt ref.
1851 rc->gfu_boost = calc_arf_boost(cpi, 0, (i - 1), (i - 1), &f_boost,
1853 rc->source_alt_ref_pending = 1;
1855 // Test to see if multi arf is appropriate.
1856 cpi->multi_arf_enabled =
1857 (cpi->multi_arf_allowed && (rc->baseline_gf_interval >= 6) &&
1858 (zero_motion_accumulator < 0.995)) ? 1 : 0;
1860 rc->gfu_boost = MAX((int)boost_score, MIN_ARF_GF_BOOST);
1861 rc->source_alt_ref_pending = 0;
1864 // Reset the file position.
1865 reset_fpf_position(twopass, start_pos);
1867 // Calculate the bits to be allocated to the gf/arf group as a whole
1868 gf_group_bits = calculate_total_gf_group_bits(cpi, gf_group_err);
1870 // Calculate the extra bits to be used for boosted frame(s)
1871 gf_arf_bits = calculate_boost_bits(rc->baseline_gf_interval,
1872 rc->gfu_boost, gf_group_bits);
1874 // Adjust KF group bits and error remaining.
1875 twopass->kf_group_error_left -= (int64_t)gf_group_err;
1877 // If this is an arf update we want to remove the score for the overlay
1878 // frame at the end which will usually be very cheap to code.
1879 // The overlay frame has already, in effect, been coded so we want to spread
1880 // the remaining bits among the other frames.
1881 // For normal GFs remove the score for the GF itself unless this is
1882 // also a key frame in which case it has already been accounted for.
1883 if (rc->source_alt_ref_pending) {
1884 gf_group_error_left = gf_group_err - mod_frame_err;
1885 } else if (cpi->common.frame_type != KEY_FRAME) {
1886 gf_group_error_left = gf_group_err - gf_first_frame_err;
1888 gf_group_error_left = gf_group_err;
1891 // Allocate bits to each of the frames in the GF group.
1892 allocate_gf_group_bits(cpi, gf_group_bits, gf_group_error_left, gf_arf_bits);
1894 // Reset the file position.
1895 reset_fpf_position(twopass, start_pos);
1897 // Calculate a section intra ratio used in setting max loop filter.
1898 if (cpi->common.frame_type != KEY_FRAME) {
1899 twopass->section_intra_rating =
1900 calculate_section_intra_ratio(start_pos, twopass->stats_in_end,
1901 rc->baseline_gf_interval);
1905 // TODO(PGW) Re-examine the use of II ration in this code in the light of#
1906 // changes elsewhere
1907 #define KF_II_MAX 128.0
1908 static int test_candidate_kf(TWO_PASS *twopass,
1909 const FIRSTPASS_STATS *last_frame,
1910 const FIRSTPASS_STATS *this_frame,
1911 const FIRSTPASS_STATS *next_frame) {
1912 int is_viable_kf = 0;
1914 // Does the frame satisfy the primary criteria of a key frame?
1915 // If so, then examine how well it predicts subsequent frames.
1916 if ((this_frame->pcnt_second_ref < 0.10) &&
1917 (next_frame->pcnt_second_ref < 0.10) &&
1918 ((this_frame->pcnt_inter < 0.05) ||
1919 (((this_frame->pcnt_inter - this_frame->pcnt_neutral) < 0.35) &&
1920 ((this_frame->intra_error /
1921 DOUBLE_DIVIDE_CHECK(this_frame->coded_error)) < 2.5) &&
1922 ((fabs(last_frame->coded_error - this_frame->coded_error) /
1923 DOUBLE_DIVIDE_CHECK(this_frame->coded_error) > 0.40) ||
1924 (fabs(last_frame->intra_error - this_frame->intra_error) /
1925 DOUBLE_DIVIDE_CHECK(this_frame->intra_error) > 0.40) ||
1926 ((next_frame->intra_error /
1927 DOUBLE_DIVIDE_CHECK(next_frame->coded_error)) > 3.5))))) {
1929 const FIRSTPASS_STATS *start_pos = twopass->stats_in;
1930 FIRSTPASS_STATS local_next_frame = *next_frame;
1931 double boost_score = 0.0;
1932 double old_boost_score = 0.0;
1933 double decay_accumulator = 1.0;
1935 // Examine how well the key frame predicts subsequent frames.
1936 for (i = 0; i < 16; ++i) {
1937 double next_iiratio = (BOOST_FACTOR * local_next_frame.intra_error /
1938 DOUBLE_DIVIDE_CHECK(local_next_frame.coded_error));
1940 if (next_iiratio > KF_II_MAX)
1941 next_iiratio = KF_II_MAX;
1943 // Cumulative effect of decay in prediction quality.
1944 if (local_next_frame.pcnt_inter > 0.85)
1945 decay_accumulator *= local_next_frame.pcnt_inter;
1947 decay_accumulator *= (0.85 + local_next_frame.pcnt_inter) / 2.0;
1949 // Keep a running total.
1950 boost_score += (decay_accumulator * next_iiratio);
1952 // Test various breakout clauses.
1953 if ((local_next_frame.pcnt_inter < 0.05) ||
1954 (next_iiratio < 1.5) ||
1955 (((local_next_frame.pcnt_inter -
1956 local_next_frame.pcnt_neutral) < 0.20) &&
1957 (next_iiratio < 3.0)) ||
1958 ((boost_score - old_boost_score) < 3.0) ||
1959 (local_next_frame.intra_error < 200)) {
1963 old_boost_score = boost_score;
1965 // Get the next frame details
1966 if (EOF == input_stats(twopass, &local_next_frame))
1970 // If there is tolerable prediction for at least the next 3 frames then
1971 // break out else discard this potential key frame and move on
1972 if (boost_score > 30.0 && (i > 3)) {
1975 // Reset the file position
1976 reset_fpf_position(twopass, start_pos);
1982 return is_viable_kf;
1985 static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
1987 RATE_CONTROL *const rc = &cpi->rc;
1988 TWO_PASS *const twopass = &cpi->twopass;
1989 GF_GROUP *const gf_group = &twopass->gf_group;
1990 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
1991 const FIRSTPASS_STATS first_frame = *this_frame;
1992 const FIRSTPASS_STATS *const start_position = twopass->stats_in;
1993 FIRSTPASS_STATS next_frame;
1994 FIRSTPASS_STATS last_frame;
1996 int loop_decay_counter = 0;
1997 double decay_accumulator = 1.0;
1998 double av_decay_accumulator = 0.0;
1999 double zero_motion_accumulator = 1.0;
2000 double boost_score = 0.0;
2001 double kf_mod_err = 0.0;
2002 double kf_group_err = 0.0;
2003 double recent_loop_decay[8] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
2005 vp9_zero(next_frame);
2007 cpi->common.frame_type = KEY_FRAME;
2009 // Reset the GF group data structures.
2010 vp9_zero(*gf_group);
2012 // Is this a forced key frame by interval.
2013 rc->this_key_frame_forced = rc->next_key_frame_forced;
2015 // Clear the alt ref active flag and last group multi arf flags as they
2016 // can never be set for a key frame.
2017 rc->source_alt_ref_active = 0;
2018 cpi->multi_arf_last_grp_enabled = 0;
2020 // KF is always a GF so clear frames till next gf counter.
2021 rc->frames_till_gf_update_due = 0;
2023 rc->frames_to_key = 1;
2025 twopass->kf_group_bits = 0; // Total bits available to kf group
2026 twopass->kf_group_error_left = 0; // Group modified error score.
2028 kf_mod_err = calculate_modified_err(twopass, oxcf, this_frame);
2030 // Find the next keyframe.
2032 while (twopass->stats_in < twopass->stats_in_end &&
2033 rc->frames_to_key < cpi->oxcf.key_freq) {
2034 // Accumulate kf group error.
2035 kf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
2037 // Load the next frame's stats.
2038 last_frame = *this_frame;
2039 input_stats(twopass, this_frame);
2041 // Provided that we are not at the end of the file...
2042 if (cpi->oxcf.auto_key && twopass->stats_in < twopass->stats_in_end) {
2043 double loop_decay_rate;
2045 // Check for a scene cut.
2046 if (test_candidate_kf(twopass, &last_frame, this_frame,
2050 // How fast is the prediction quality decaying?
2051 loop_decay_rate = get_prediction_decay_rate(&cpi->common,
2054 // We want to know something about the recent past... rather than
2055 // as used elsewhere where we are concerned with decay in prediction
2056 // quality since the last GF or KF.
2057 recent_loop_decay[i % 8] = loop_decay_rate;
2058 decay_accumulator = 1.0;
2059 for (j = 0; j < 8; ++j)
2060 decay_accumulator *= recent_loop_decay[j];
2062 // Special check for transition or high motion followed by a
2064 if (detect_transition_to_still(twopass, i, cpi->oxcf.key_freq - i,
2065 loop_decay_rate, decay_accumulator))
2068 // Step on to the next frame.
2069 ++rc->frames_to_key;
2071 // If we don't have a real key frame within the next two
2072 // key_freq intervals then break out of the loop.
2073 if (rc->frames_to_key >= 2 * cpi->oxcf.key_freq)
2076 ++rc->frames_to_key;
2081 // If there is a max kf interval set by the user we must obey it.
2082 // We already breakout of the loop above at 2x max.
2083 // This code centers the extra kf if the actual natural interval
2084 // is between 1x and 2x.
2085 if (cpi->oxcf.auto_key &&
2086 rc->frames_to_key > cpi->oxcf.key_freq) {
2087 FIRSTPASS_STATS tmp_frame = first_frame;
2089 rc->frames_to_key /= 2;
2091 // Reset to the start of the group.
2092 reset_fpf_position(twopass, start_position);
2096 // Rescan to get the correct error data for the forced kf group.
2097 for (i = 0; i < rc->frames_to_key; ++i) {
2098 kf_group_err += calculate_modified_err(twopass, oxcf, &tmp_frame);
2099 input_stats(twopass, &tmp_frame);
2101 rc->next_key_frame_forced = 1;
2102 } else if (twopass->stats_in == twopass->stats_in_end ||
2103 rc->frames_to_key >= cpi->oxcf.key_freq) {
2104 rc->next_key_frame_forced = 1;
2106 rc->next_key_frame_forced = 0;
2109 if (is_two_pass_svc(cpi) && cpi->svc.number_temporal_layers > 1) {
2110 int count = (1 << (cpi->svc.number_temporal_layers - 1)) - 1;
2111 int new_frame_to_key = (rc->frames_to_key + count) & (~count);
2113 for (j = 0; j < new_frame_to_key - rc->frames_to_key; ++j) {
2114 if (EOF == input_stats(twopass, this_frame))
2116 kf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
2118 rc->frames_to_key = new_frame_to_key;
2121 // Special case for the last key frame of the file.
2122 if (twopass->stats_in >= twopass->stats_in_end) {
2123 // Accumulate kf group error.
2124 kf_group_err += calculate_modified_err(twopass, oxcf, this_frame);
2127 // Calculate the number of bits that should be assigned to the kf group.
2128 if (twopass->bits_left > 0 && twopass->modified_error_left > 0.0) {
2129 // Maximum number of bits for a single normal frame (not key frame).
2130 const int max_bits = frame_max_bits(rc, &cpi->oxcf);
2132 // Maximum number of bits allocated to the key frame group.
2133 int64_t max_grp_bits;
2135 // Default allocation based on bits left and relative
2136 // complexity of the section.
2137 twopass->kf_group_bits = (int64_t)(twopass->bits_left *
2138 (kf_group_err / twopass->modified_error_left));
2140 // Clip based on maximum per frame rate defined by the user.
2141 max_grp_bits = (int64_t)max_bits * (int64_t)rc->frames_to_key;
2142 if (twopass->kf_group_bits > max_grp_bits)
2143 twopass->kf_group_bits = max_grp_bits;
2145 twopass->kf_group_bits = 0;
2147 twopass->kf_group_bits = MAX(0, twopass->kf_group_bits);
2149 // Reset the first pass file position.
2150 reset_fpf_position(twopass, start_position);
2152 // Scan through the kf group collating various stats used to determine
2153 // how many bits to spend on it.
2154 decay_accumulator = 1.0;
2156 for (i = 0; i < (rc->frames_to_key - 1); ++i) {
2157 if (EOF == input_stats(twopass, &next_frame))
2160 // Monitor for static sections.
2161 zero_motion_accumulator =
2162 MIN(zero_motion_accumulator,
2163 get_zero_motion_factor(&cpi->common, &next_frame));
2165 // Not all frames in the group are necessarily used in calculating boost.
2166 if ((i <= rc->max_gf_interval) ||
2167 ((i <= (rc->max_gf_interval * 4)) && (decay_accumulator > 0.5))) {
2168 const double frame_boost =
2169 calc_frame_boost(cpi, this_frame, 0, KF_MAX_BOOST);
2171 // How fast is prediction quality decaying.
2172 if (!detect_flash(twopass, 0)) {
2173 const double loop_decay_rate =
2174 get_prediction_decay_rate(&cpi->common, &next_frame);
2175 decay_accumulator *= loop_decay_rate;
2176 decay_accumulator = MAX(decay_accumulator, MIN_DECAY_FACTOR);
2177 av_decay_accumulator += decay_accumulator;
2178 ++loop_decay_counter;
2180 boost_score += (decay_accumulator * frame_boost);
2183 av_decay_accumulator /= (double)loop_decay_counter;
2185 reset_fpf_position(twopass, start_position);
2187 // Store the zero motion percentage
2188 twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
2190 // Calculate a section intra ratio used in setting max loop filter.
2191 twopass->section_intra_rating =
2192 calculate_section_intra_ratio(start_position, twopass->stats_in_end,
2195 // Apply various clamps for min and max boost
2196 rc->kf_boost = (int)(av_decay_accumulator * boost_score);
2197 rc->kf_boost = MAX(rc->kf_boost, (rc->frames_to_key * 3));
2198 rc->kf_boost = MAX(rc->kf_boost, MIN_KF_BOOST);
2200 // Work out how many bits to allocate for the key frame itself.
2201 kf_bits = calculate_boost_bits((rc->frames_to_key - 1),
2202 rc->kf_boost, twopass->kf_group_bits);
2204 twopass->kf_group_bits -= kf_bits;
2206 // Save the bits to spend on the key frame.
2207 gf_group->bit_allocation[0] = kf_bits;
2208 gf_group->update_type[0] = KF_UPDATE;
2209 gf_group->rf_level[0] = KF_STD;
2211 // Note the total error score of the kf group minus the key frame itself.
2212 twopass->kf_group_error_left = (int)(kf_group_err - kf_mod_err);
2214 // Adjust the count of total modified error left.
2215 // The count of bits left is adjusted elsewhere based on real coded frame
2217 twopass->modified_error_left -= kf_group_err;
2220 #define VBR_PCT_ADJUSTMENT_LIMIT 50
2221 // For VBR...adjustment to the frame target based on error from previous frames
2222 void vbr_rate_correction(VP9_COMP *cpi,
2223 int * this_frame_target,
2224 const int64_t vbr_bits_off_target) {
2226 double position_factor = 1.0;
2228 // How far through the clip are we.
2229 // This number is used to damp the per frame rate correction.
2231 if (cpi->twopass.total_stats.count) {
2232 position_factor = sqrt((double)cpi->common.current_video_frame /
2233 cpi->twopass.total_stats.count);
2235 max_delta = (int)(position_factor *
2236 ((*this_frame_target * VBR_PCT_ADJUSTMENT_LIMIT) / 100));
2238 // vbr_bits_off_target > 0 means we have extra bits to spend
2239 if (vbr_bits_off_target > 0) {
2240 *this_frame_target +=
2241 (vbr_bits_off_target > max_delta) ? max_delta
2242 : (int)vbr_bits_off_target;
2244 *this_frame_target -=
2245 (vbr_bits_off_target < -max_delta) ? max_delta
2246 : (int)-vbr_bits_off_target;
2250 // Define the reference buffers that will be updated post encode.
2251 void configure_buffer_updates(VP9_COMP *cpi) {
2252 TWO_PASS *const twopass = &cpi->twopass;
2254 cpi->rc.is_src_frame_alt_ref = 0;
2255 switch (twopass->gf_group.update_type[twopass->gf_group.index]) {
2257 cpi->refresh_last_frame = 1;
2258 cpi->refresh_golden_frame = 1;
2259 cpi->refresh_alt_ref_frame = 1;
2262 cpi->refresh_last_frame = 1;
2263 cpi->refresh_golden_frame = 0;
2264 cpi->refresh_alt_ref_frame = 0;
2267 cpi->refresh_last_frame = 1;
2268 cpi->refresh_golden_frame = 1;
2269 cpi->refresh_alt_ref_frame = 0;
2271 case OVERLAY_UPDATE:
2272 cpi->refresh_last_frame = 0;
2273 cpi->refresh_golden_frame = 1;
2274 cpi->refresh_alt_ref_frame = 0;
2275 cpi->rc.is_src_frame_alt_ref = 1;
2278 cpi->refresh_last_frame = 0;
2279 cpi->refresh_golden_frame = 0;
2280 cpi->refresh_alt_ref_frame = 1;
2286 if (is_two_pass_svc(cpi)) {
2287 if (cpi->svc.temporal_layer_id > 0) {
2288 cpi->refresh_last_frame = 0;
2289 cpi->refresh_golden_frame = 0;
2291 if (cpi->svc.layer_context[cpi->svc.spatial_layer_id].gold_ref_idx < 0)
2292 cpi->refresh_golden_frame = 0;
2293 if (cpi->alt_ref_source == NULL)
2294 cpi->refresh_alt_ref_frame = 0;
2299 void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
2300 VP9_COMMON *const cm = &cpi->common;
2301 RATE_CONTROL *const rc = &cpi->rc;
2302 TWO_PASS *const twopass = &cpi->twopass;
2303 GF_GROUP *const gf_group = &twopass->gf_group;
2305 FIRSTPASS_STATS this_frame;
2306 FIRSTPASS_STATS this_frame_copy;
2309 LAYER_CONTEXT *const lc = is_two_pass_svc(cpi) ?
2310 &cpi->svc.layer_context[cpi->svc.spatial_layer_id] : 0;
2313 frames_left = (int)(twopass->total_stats.count -
2314 lc->current_video_frame_in_layer);
2316 frames_left = (int)(twopass->total_stats.count -
2317 cm->current_video_frame);
2320 if (!twopass->stats_in)
2323 // If this is an arf frame then we dont want to read the stats file or
2324 // advance the input pointer as we already have what we need.
2325 if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
2327 configure_buffer_updates(cpi);
2328 target_rate = gf_group->bit_allocation[gf_group->index];
2329 target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
2330 rc->base_frame_target = target_rate;
2332 // Correction to rate target based on prior over or under shoot.
2333 if (cpi->oxcf.rc_mode == VPX_VBR)
2334 vbr_rate_correction(cpi, &target_rate, rc->vbr_bits_off_target);
2336 vp9_rc_set_frame_target(cpi, target_rate);
2337 cm->frame_type = INTER_FRAME;
2340 if (cpi->svc.spatial_layer_id == 0) {
2341 lc->is_key_frame = 0;
2343 lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
2345 if (lc->is_key_frame)
2346 cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
2353 vp9_clear_system_state();
2355 if (cpi->oxcf.rc_mode == VPX_Q) {
2356 twopass->active_worst_quality = cpi->oxcf.cq_level;
2357 } else if (cm->current_video_frame == 0 ||
2358 (lc != NULL && lc->current_video_frame_in_layer == 0)) {
2359 // Special case code for first frame.
2360 const int section_target_bandwidth = (int)(twopass->bits_left /
2362 const int tmp_q = get_twopass_worst_quality(cpi, &twopass->total_left_stats,
2363 section_target_bandwidth);
2364 twopass->active_worst_quality = tmp_q;
2365 rc->ni_av_qi = tmp_q;
2366 rc->last_q[INTER_FRAME] = tmp_q;
2367 rc->avg_q = vp9_convert_qindex_to_q(tmp_q, cm->bit_depth);
2368 rc->avg_frame_qindex[INTER_FRAME] = tmp_q;
2369 rc->last_q[KEY_FRAME] = (tmp_q + cpi->oxcf.best_allowed_q) / 2;
2370 rc->avg_frame_qindex[KEY_FRAME] = rc->last_q[KEY_FRAME];
2372 vp9_zero(this_frame);
2373 if (EOF == input_stats(twopass, &this_frame))
2376 // Local copy of the current frame's first pass stats.
2377 this_frame_copy = this_frame;
2379 // Keyframe and section processing.
2380 if (rc->frames_to_key == 0 ||
2381 (cpi->frame_flags & FRAMEFLAGS_KEY)) {
2382 // Define next KF group and assign bits to it.
2383 find_next_key_frame(cpi, &this_frame_copy);
2385 cm->frame_type = INTER_FRAME;
2389 if (cpi->svc.spatial_layer_id == 0) {
2390 lc->is_key_frame = (cm->frame_type == KEY_FRAME);
2391 if (lc->is_key_frame) {
2392 cpi->ref_frame_flags &=
2393 (~VP9_LAST_FLAG & ~VP9_GOLD_FLAG & ~VP9_ALT_FLAG);
2394 lc->frames_from_key_frame = 0;
2395 // Reset the empty frame resolution since we have a key frame.
2396 cpi->svc.empty_frame_width = cm->width;
2397 cpi->svc.empty_frame_height = cm->height;
2400 cm->frame_type = INTER_FRAME;
2401 lc->is_key_frame = cpi->svc.layer_context[0].is_key_frame;
2403 if (lc->is_key_frame) {
2404 cpi->ref_frame_flags &= (~VP9_LAST_FLAG);
2405 lc->frames_from_key_frame = 0;
2410 // Define a new GF/ARF group. (Should always enter here for key frames).
2411 if (rc->frames_till_gf_update_due == 0) {
2412 define_gf_group(cpi, &this_frame_copy);
2414 rc->frames_till_gf_update_due = rc->baseline_gf_interval;
2416 cpi->refresh_golden_frame = 1;
2418 #if ARF_STATS_OUTPUT
2421 fpfile = fopen("arf.stt", "a");
2423 fprintf(fpfile, "%10d %10ld %10d %10d %10ld\n",
2424 cm->current_video_frame, rc->frames_till_gf_update_due,
2425 rc->kf_boost, arf_count, rc->gfu_boost);
2432 configure_buffer_updates(cpi);
2434 target_rate = gf_group->bit_allocation[gf_group->index];
2435 if (cpi->common.frame_type == KEY_FRAME)
2436 target_rate = vp9_rc_clamp_iframe_target_size(cpi, target_rate);
2438 target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
2440 rc->base_frame_target = target_rate;
2442 // Correction to rate target based on prior over or under shoot.
2443 if (cpi->oxcf.rc_mode == VPX_VBR)
2444 vbr_rate_correction(cpi, &target_rate, rc->vbr_bits_off_target);
2446 vp9_rc_set_frame_target(cpi, target_rate);
2448 // Update the total stats remaining structure.
2449 subtract_stats(&twopass->total_left_stats, &this_frame);
2452 #define MINQ_ADJ_LIMIT 32
2453 #define Q_LIMIT_STEP 1
2454 void vp9_twopass_postencode_update(VP9_COMP *cpi) {
2455 TWO_PASS *const twopass = &cpi->twopass;
2456 RATE_CONTROL *const rc = &cpi->rc;
2457 const int bits_used = rc->base_frame_target;
2459 // VBR correction is done through rc->vbr_bits_off_target. Based on the
2460 // sign of this value, a limited % adjustment is made to the target rate
2461 // of subsequent frames, to try and push it back towards 0. This method
2462 // is designed to prevent extreme behaviour at the end of a clip
2463 // or group of frames.
2464 rc->vbr_bits_off_target += rc->base_frame_target - rc->projected_frame_size;
2465 twopass->bits_left = MAX(twopass->bits_left - bits_used, 0);
2467 // Calculate the pct rc error.
2468 if (rc->total_actual_bits) {
2469 rc->rate_error_estimate =
2470 (int)((rc->vbr_bits_off_target * 100) / rc->total_actual_bits);
2471 rc->rate_error_estimate = clamp(rc->rate_error_estimate, -100, 100);
2473 rc->rate_error_estimate = 0;
2476 if (cpi->common.frame_type != KEY_FRAME &&
2477 !vp9_is_upper_layer_key_frame(cpi)) {
2478 twopass->kf_group_bits -= bits_used;
2479 twopass->last_kfgroup_zeromotion_pct = twopass->kf_zeromotion_pct;
2481 twopass->kf_group_bits = MAX(twopass->kf_group_bits, 0);
2483 // Increment the gf group index ready for the next frame.
2484 ++twopass->gf_group.index;
2486 // If the rate control is drifting consider adjustment ot min or maxq.
2487 // Only make adjustments on gf/arf
2488 if ((cpi->oxcf.rc_mode == VPX_VBR) &&
2489 (cpi->twopass.gf_zeromotion_pct < VLOW_MOTION_THRESHOLD) &&
2490 !cpi->rc.is_src_frame_alt_ref) {
2491 const int maxq_adj_limit =
2492 rc->worst_quality - twopass->active_worst_quality;
2495 if (rc->rate_error_estimate > cpi->oxcf.under_shoot_pct) {
2496 --twopass->extend_maxq;
2497 if (rc->rolling_target_bits >= rc->rolling_actual_bits)
2498 twopass->extend_minq += Q_LIMIT_STEP;
2500 } else if (rc->rate_error_estimate < -cpi->oxcf.over_shoot_pct) {
2501 --twopass->extend_minq;
2502 if (rc->rolling_target_bits < rc->rolling_actual_bits)
2503 twopass->extend_maxq += Q_LIMIT_STEP;
2505 if (rc->rolling_target_bits < rc->rolling_actual_bits)
2506 --twopass->extend_minq;
2507 if (rc->rolling_target_bits > rc->rolling_actual_bits)
2508 --twopass->extend_maxq;
2510 twopass->extend_minq = clamp(twopass->extend_minq, 0, MINQ_ADJ_LIMIT);
2511 twopass->extend_maxq = clamp(twopass->extend_maxq, 0, maxq_adj_limit);