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.
16 #include "encodeintra.h"
17 #include "vp8/common/setupintrarecon.h"
19 #include "vpx_scale/vpxscale.h"
21 #include "vp8/common/extend.h"
22 #include "vp8/common/systemdependent.h"
23 #include "vpx_scale/yv12extend.h"
24 #include "vpx_mem/vpx_mem.h"
25 #include "vp8/common/swapyv12buffer.h"
28 #include "vp8/common/quant_common.h"
31 //#define OUTPUT_FPF 1
33 #if CONFIG_RUNTIME_CPU_DETECT
34 #define IF_RTCD(x) (x)
36 #define IF_RTCD(x) NULL
39 extern void vp8_build_block_offsets(MACROBLOCK *x);
40 extern void vp8_setup_block_ptrs(MACROBLOCK *x);
41 extern void vp8cx_frame_init_quantizer(VP8_COMP *cpi);
42 extern void vp8_set_mbmode_and_mvs(MACROBLOCK *x, MB_PREDICTION_MODE mb, int_mv *mv);
43 extern void vp8_alloc_compressor_data(VP8_COMP *cpi);
45 //#define GFQ_ADJUSTMENT (40 + ((15*Q)/10))
46 //#define GFQ_ADJUSTMENT (80 + ((15*Q)/10))
47 #define GFQ_ADJUSTMENT vp8_gf_boost_qadjustment[Q]
48 extern int vp8_kf_boost_qadjustment[QINDEX_RANGE];
50 extern const int vp8_gf_boost_qadjustment[QINDEX_RANGE];
53 #define IIKFACTOR1 1.40
54 #define IIKFACTOR2 1.5
58 #define KF_MB_INTRA_MIN 300
59 #define GF_MB_INTRA_MIN 200
61 #define DOUBLE_DIVIDE_CHECK(X) ((X)<0?(X)-.000001:(X)+.000001)
63 #define POW1 (double)cpi->oxcf.two_pass_vbrbias/100.0
64 #define POW2 (double)cpi->oxcf.two_pass_vbrbias/100.0
66 static int vscale_lookup[7] = {0, 1, 1, 2, 2, 3, 3};
67 static int hscale_lookup[7] = {0, 0, 1, 1, 2, 2, 3};
70 static const int cq_level[QINDEX_RANGE] =
72 0,0,1,1,2,3,3,4,4,5,6,6,7,8,8,9,
73 9,10,11,11,12,13,13,14,15,15,16,17,17,18,19,20,
74 20,21,22,22,23,24,24,25,26,27,27,28,29,30,30,31,
75 32,33,33,34,35,36,36,37,38,39,39,40,41,42,42,43,
76 44,45,46,46,47,48,49,50,50,51,52,53,54,55,55,56,
77 57,58,59,60,60,61,62,63,64,65,66,67,67,68,69,70,
78 71,72,73,74,75,75,76,77,78,79,80,81,82,83,84,85,
79 86,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100
82 static void find_next_key_frame(VP8_COMP *cpi, FIRSTPASS_STATS *this_frame);
84 static int encode_intra(VP8_COMP *cpi, MACROBLOCK *x, int use_dc_pred)
88 int intra_pred_var = 0;
93 x->e_mbd.mode_info_context->mbmi.mode = DC_PRED;
94 x->e_mbd.mode_info_context->mbmi.uv_mode = DC_PRED;
95 x->e_mbd.mode_info_context->mbmi.ref_frame = INTRA_FRAME;
97 vp8_encode_intra16x16mby(IF_RTCD(&cpi->rtcd), x);
101 for (i = 0; i < 16; i++)
103 x->e_mbd.block[i].bmi.as_mode = B_DC_PRED;
104 vp8_encode_intra4x4block(IF_RTCD(&cpi->rtcd), x, i);
108 intra_pred_var = VARIANCE_INVOKE(&cpi->rtcd.variance, getmbss)(x->src_diff);
110 return intra_pred_var;
113 // Resets the first pass file to the given position using a relative seek from the current position
114 static void reset_fpf_position(VP8_COMP *cpi, FIRSTPASS_STATS *Position)
116 cpi->twopass.stats_in = Position;
119 static int lookup_next_frame_stats(VP8_COMP *cpi, FIRSTPASS_STATS *next_frame)
121 if (cpi->twopass.stats_in >= cpi->twopass.stats_in_end)
124 *next_frame = *cpi->twopass.stats_in;
128 // Calculate a modified Error used in distributing bits between easier and harder frames
129 static double calculate_modified_err(VP8_COMP *cpi, FIRSTPASS_STATS *this_frame)
131 double av_err = cpi->twopass.total_stats->ssim_weighted_pred_err;
132 double this_err = this_frame->ssim_weighted_pred_err;
135 //double relative_next_iiratio;
136 //double next_iiratio;
137 //double sum_iiratio;
140 //FIRSTPASS_STATS next_frame;
141 //FIRSTPASS_STATS *start_pos;
143 /*start_pos = cpi->twopass.stats_in;
146 while ( (i < 1) && input_stats(cpi,&next_frame) != EOF )
149 next_iiratio = next_frame.intra_error / DOUBLE_DIVIDE_CHECK(next_frame.coded_error);
150 next_iiratio = ( next_iiratio < 1.0 ) ? 1.0 : (next_iiratio > 20.0) ? 20.0 : next_iiratio;
151 sum_iiratio += next_iiratio;
156 relative_next_iiratio = sum_iiratio / DOUBLE_DIVIDE_CHECK(cpi->twopass.avg_iiratio * (double)i);
160 relative_next_iiratio = 1.0;
162 reset_fpf_position(cpi, start_pos);*/
164 if (this_err > av_err)
165 modified_err = av_err * pow((this_err / DOUBLE_DIVIDE_CHECK(av_err)), POW1);
167 modified_err = av_err * pow((this_err / DOUBLE_DIVIDE_CHECK(av_err)), POW2);
170 relative_next_iiratio = pow(relative_next_iiratio,0.25);
171 modified_err = modified_err * relative_next_iiratio;
177 static const double weight_table[256] = {
178 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
179 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
180 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
181 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
182 0.020000, 0.031250, 0.062500, 0.093750, 0.125000, 0.156250, 0.187500, 0.218750,
183 0.250000, 0.281250, 0.312500, 0.343750, 0.375000, 0.406250, 0.437500, 0.468750,
184 0.500000, 0.531250, 0.562500, 0.593750, 0.625000, 0.656250, 0.687500, 0.718750,
185 0.750000, 0.781250, 0.812500, 0.843750, 0.875000, 0.906250, 0.937500, 0.968750,
186 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
187 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
188 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
189 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
190 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
191 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
192 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
193 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
194 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
195 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
196 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
197 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
198 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
199 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
200 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
201 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
202 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
203 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
204 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
205 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
206 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
207 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
208 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
209 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000
212 static double simple_weight(YV12_BUFFER_CONFIG *source)
216 unsigned char *src = source->y_buffer;
217 double sum_weights = 0.0;
219 // Loop throught the Y plane raw examining levels and creating a weight for the image
220 i = source->y_height;
226 sum_weights += weight_table[ *src];
229 src -= source->y_width;
230 src += source->y_stride;
233 sum_weights /= (source->y_height * source->y_width);
239 // This function returns the current per frame maximum bitrate target
240 static int frame_max_bits(VP8_COMP *cpi)
242 // Max allocation for a single frame based on the max section guidelines passed in and how many bits are left
245 // For CBR we need to also consider buffer fullness.
246 // If we are running below the optimal level then we need to gradually tighten up on max_bits.
247 if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
249 double buffer_fullness_ratio = (double)cpi->buffer_level / DOUBLE_DIVIDE_CHECK((double)cpi->oxcf.optimal_buffer_level);
251 // For CBR base this on the target average bits per frame plus the maximum sedction rate passed in by the user
252 max_bits = (int)(cpi->av_per_frame_bandwidth * ((double)cpi->oxcf.two_pass_vbrmax_section / 100.0));
254 // If our buffer is below the optimum level
255 if (buffer_fullness_ratio < 1.0)
257 // The lower of max_bits / 4 or cpi->av_per_frame_bandwidth / 4.
258 int min_max_bits = ((cpi->av_per_frame_bandwidth >> 2) < (max_bits >> 2)) ? cpi->av_per_frame_bandwidth >> 2 : max_bits >> 2;
260 max_bits = (int)(max_bits * buffer_fullness_ratio);
262 if (max_bits < min_max_bits)
263 max_bits = min_max_bits; // Lowest value we will set ... which should allow the buffer to refil.
269 // For VBR base this on the bits and frames left plus the two_pass_vbrmax_section rate passed in by the user
270 max_bits = (int)(((double)cpi->twopass.bits_left / (cpi->twopass.total_stats->count - (double)cpi->common.current_video_frame)) * ((double)cpi->oxcf.two_pass_vbrmax_section / 100.0));
273 // Trap case where we are out of bits
281 static void output_stats(const VP8_COMP *cpi,
282 struct vpx_codec_pkt_list *pktlist,
283 FIRSTPASS_STATS *stats)
285 struct vpx_codec_cx_pkt pkt;
286 pkt.kind = VPX_CODEC_STATS_PKT;
287 pkt.data.twopass_stats.buf = stats;
288 pkt.data.twopass_stats.sz = sizeof(FIRSTPASS_STATS);
289 vpx_codec_pkt_list_add(pktlist, &pkt);
296 fpfile = fopen("firstpass.stt", "a");
298 fprintf(fpfile, "%12.0f %12.0f %12.0f %12.4f %12.4f %12.4f %12.4f"
299 " %12.4f %12.4f %12.4f %12.4f %12.4f %12.4f %12.4f %12.4f"
304 stats->ssim_weighted_pred_err,
307 stats->pcnt_second_ref,
315 stats->mv_in_out_count,
323 static int input_stats(VP8_COMP *cpi, FIRSTPASS_STATS *fps)
325 if (cpi->twopass.stats_in >= cpi->twopass.stats_in_end)
328 *fps = *cpi->twopass.stats_in;
329 cpi->twopass.stats_in =
330 (void*)((char *)cpi->twopass.stats_in + sizeof(FIRSTPASS_STATS));
334 static void zero_stats(FIRSTPASS_STATS *section)
336 section->frame = 0.0;
337 section->intra_error = 0.0;
338 section->coded_error = 0.0;
339 section->ssim_weighted_pred_err = 0.0;
340 section->pcnt_inter = 0.0;
341 section->pcnt_motion = 0.0;
342 section->pcnt_second_ref = 0.0;
343 section->pcnt_neutral = 0.0;
345 section->mvr_abs = 0.0;
347 section->mvc_abs = 0.0;
350 section->mv_in_out_count = 0.0;
351 section->count = 0.0;
352 section->duration = 1.0;
354 static void accumulate_stats(FIRSTPASS_STATS *section, FIRSTPASS_STATS *frame)
356 section->frame += frame->frame;
357 section->intra_error += frame->intra_error;
358 section->coded_error += frame->coded_error;
359 section->ssim_weighted_pred_err += frame->ssim_weighted_pred_err;
360 section->pcnt_inter += frame->pcnt_inter;
361 section->pcnt_motion += frame->pcnt_motion;
362 section->pcnt_second_ref += frame->pcnt_second_ref;
363 section->pcnt_neutral += frame->pcnt_neutral;
364 section->MVr += frame->MVr;
365 section->mvr_abs += frame->mvr_abs;
366 section->MVc += frame->MVc;
367 section->mvc_abs += frame->mvc_abs;
368 section->MVrv += frame->MVrv;
369 section->MVcv += frame->MVcv;
370 section->mv_in_out_count += frame->mv_in_out_count;
371 section->count += frame->count;
372 section->duration += frame->duration;
374 static void avg_stats(FIRSTPASS_STATS *section)
376 if (section->count < 1.0)
379 section->intra_error /= section->count;
380 section->coded_error /= section->count;
381 section->ssim_weighted_pred_err /= section->count;
382 section->pcnt_inter /= section->count;
383 section->pcnt_second_ref /= section->count;
384 section->pcnt_neutral /= section->count;
385 section->pcnt_motion /= section->count;
386 section->MVr /= section->count;
387 section->mvr_abs /= section->count;
388 section->MVc /= section->count;
389 section->mvc_abs /= section->count;
390 section->MVrv /= section->count;
391 section->MVcv /= section->count;
392 section->mv_in_out_count /= section->count;
393 section->duration /= section->count;
396 void vp8_init_first_pass(VP8_COMP *cpi)
398 zero_stats(cpi->twopass.total_stats);
401 void vp8_end_first_pass(VP8_COMP *cpi)
403 output_stats(cpi, cpi->output_pkt_list, cpi->twopass.total_stats);
406 static void zz_motion_search( VP8_COMP *cpi, MACROBLOCK * x, YV12_BUFFER_CONFIG * recon_buffer, int * best_motion_err, int recon_yoffset )
408 MACROBLOCKD * const xd = & x->e_mbd;
409 BLOCK *b = &x->block[0];
410 BLOCKD *d = &x->e_mbd.block[0];
412 unsigned char *src_ptr = (*(b->base_src) + b->src);
413 int src_stride = b->src_stride;
414 unsigned char *ref_ptr;
415 int ref_stride=d->pre_stride;
417 // Set up pointers for this macro block recon buffer
418 xd->pre.y_buffer = recon_buffer->y_buffer + recon_yoffset;
420 ref_ptr = (unsigned char *)(*(d->base_pre) + d->pre );
422 VARIANCE_INVOKE(IF_RTCD(&cpi->rtcd.variance), mse16x16) ( src_ptr, src_stride, ref_ptr, ref_stride, (unsigned int *)(best_motion_err));
425 static void first_pass_motion_search(VP8_COMP *cpi, MACROBLOCK *x,
426 int_mv *ref_mv, MV *best_mv,
427 YV12_BUFFER_CONFIG *recon_buffer,
428 int *best_motion_err, int recon_yoffset )
430 MACROBLOCKD *const xd = & x->e_mbd;
431 BLOCK *b = &x->block[0];
432 BLOCKD *d = &x->e_mbd.block[0];
438 int step_param = 3; //3; // Dont search over full range for first pass
439 int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param; //3;
441 vp8_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
442 int new_mv_mode_penalty = 256;
444 // override the default variance function to use MSE
445 v_fn_ptr.vf = VARIANCE_INVOKE(IF_RTCD(&cpi->rtcd.variance), mse16x16);
447 // Set up pointers for this macro block recon buffer
448 xd->pre.y_buffer = recon_buffer->y_buffer + recon_yoffset;
450 // Initial step/diamond search centred on best mv
452 tmp_err = cpi->diamond_search_sad(x, b, d, ref_mv, &tmp_mv, step_param,
453 x->sadperbit16, &num00, &v_fn_ptr,
455 if ( tmp_err < INT_MAX-new_mv_mode_penalty )
456 tmp_err += new_mv_mode_penalty;
458 if (tmp_err < *best_motion_err)
460 *best_motion_err = tmp_err;
461 best_mv->row = tmp_mv.as_mv.row;
462 best_mv->col = tmp_mv.as_mv.col;
465 // Further step/diamond searches as necessary
469 while (n < further_steps)
477 tmp_err = cpi->diamond_search_sad(x, b, d, ref_mv, &tmp_mv,
478 step_param + n, x->sadperbit16,
479 &num00, &v_fn_ptr, x->mvcost,
481 if ( tmp_err < INT_MAX-new_mv_mode_penalty )
482 tmp_err += new_mv_mode_penalty;
484 if (tmp_err < *best_motion_err)
486 *best_motion_err = tmp_err;
487 best_mv->row = tmp_mv.as_mv.row;
488 best_mv->col = tmp_mv.as_mv.col;
494 void vp8_first_pass(VP8_COMP *cpi)
497 MACROBLOCK *const x = & cpi->mb;
498 VP8_COMMON *const cm = & cpi->common;
499 MACROBLOCKD *const xd = & x->e_mbd;
501 int recon_yoffset, recon_uvoffset;
502 YV12_BUFFER_CONFIG *lst_yv12 = &cm->yv12_fb[cm->lst_fb_idx];
503 YV12_BUFFER_CONFIG *new_yv12 = &cm->yv12_fb[cm->new_fb_idx];
504 YV12_BUFFER_CONFIG *gld_yv12 = &cm->yv12_fb[cm->gld_fb_idx];
505 int recon_y_stride = lst_yv12->y_stride;
506 int recon_uv_stride = lst_yv12->uv_stride;
507 long long intra_error = 0;
508 long long coded_error = 0;
510 int sum_mvr = 0, sum_mvc = 0;
511 int sum_mvr_abs = 0, sum_mvc_abs = 0;
512 int sum_mvrs = 0, sum_mvcs = 0;
515 int second_ref_count = 0;
516 int intrapenalty = 256;
517 int neutral_count = 0;
519 int sum_in_vectors = 0;
523 zero_ref_mv.as_int = 0;
525 vp8_clear_system_state(); //__asm emms;
527 x->src = * cpi->Source;
531 x->partition_info = x->pi;
533 xd->mode_info_context = cm->mi;
535 vp8_build_block_offsets(x);
537 vp8_setup_block_dptrs(&x->e_mbd);
539 vp8_setup_block_ptrs(x);
541 // set up frame new frame for intra coded blocks
542 vp8_setup_intra_recon(new_yv12);
543 vp8cx_frame_init_quantizer(cpi);
545 // Initialise the MV cost table to the defaults
546 //if( cm->current_video_frame == 0)
549 int flag[2] = {1, 1};
550 vp8_initialize_rd_consts(cpi, vp8_dc_quant(cm->base_qindex, cm->y1dc_delta_q));
551 vpx_memcpy(cm->fc.mvc, vp8_default_mv_context, sizeof(vp8_default_mv_context));
552 vp8_build_component_cost_table(cpi->mb.mvcost, (const MV_CONTEXT *) cm->fc.mvc, flag);
555 // for each macroblock row in image
556 for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
560 best_ref_mv.as_int = 0;
562 // reset above block coeffs
563 xd->up_available = (mb_row != 0);
564 recon_yoffset = (mb_row * recon_y_stride * 16);
565 recon_uvoffset = (mb_row * recon_uv_stride * 8);
567 // Set up limit values for motion vectors to prevent them extending outside the UMV borders
568 x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
569 x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
572 // for each macroblock col in image
573 for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
576 int gf_motion_error = INT_MAX;
577 int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
579 xd->dst.y_buffer = new_yv12->y_buffer + recon_yoffset;
580 xd->dst.u_buffer = new_yv12->u_buffer + recon_uvoffset;
581 xd->dst.v_buffer = new_yv12->v_buffer + recon_uvoffset;
582 xd->left_available = (mb_col != 0);
584 // do intra 16x16 prediction
585 this_error = encode_intra(cpi, x, use_dc_pred);
587 // "intrapenalty" below deals with situations where the intra and inter error scores are very low (eg a plain black frame)
588 // We do not have special cases in first pass for 0,0 and nearest etc so all inter modes carry an overhead cost estimate fot the mv.
589 // When the error score is very low this causes us to pick all or lots of INTRA modes and throw lots of key frames.
590 // This penalty adds a cost matching that of a 0,0 mv to the intra case.
591 this_error += intrapenalty;
593 // Cumulative intra error total
594 intra_error += (long long)this_error;
596 // Set up limit values for motion vectors to prevent them extending outside the UMV borders
597 x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
598 x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
600 // Other than for the first frame do a motion search
601 if (cm->current_video_frame > 0)
603 BLOCKD *d = &x->e_mbd.block[0];
606 int motion_error = INT_MAX;
608 // Simple 0,0 motion with no mv overhead
609 zz_motion_search( cpi, x, lst_yv12, &motion_error, recon_yoffset );
610 d->bmi.mv.as_mv.row = 0;
611 d->bmi.mv.as_mv.col = 0;
613 // Test last reference frame using the previous best mv as the
614 // starting point (best reference) for the search
615 first_pass_motion_search(cpi, x, &best_ref_mv,
616 &d->bmi.mv.as_mv, lst_yv12,
617 &motion_error, recon_yoffset);
619 // If the current best reference mv is not centred on 0,0 then do a 0,0 based search as well
620 if (best_ref_mv.as_int)
623 first_pass_motion_search(cpi, x, &zero_ref_mv, &tmp_mv,
624 lst_yv12, &tmp_err, recon_yoffset);
626 if ( tmp_err < motion_error )
628 motion_error = tmp_err;
629 d->bmi.mv.as_mv.row = tmp_mv.row;
630 d->bmi.mv.as_mv.col = tmp_mv.col;
634 // Experimental search in a second reference frame ((0,0) based only)
635 if (cm->current_video_frame > 1)
637 first_pass_motion_search(cpi, x, &zero_ref_mv, &tmp_mv, gld_yv12, &gf_motion_error, recon_yoffset);
639 if ((gf_motion_error < motion_error) && (gf_motion_error < this_error))
642 //motion_error = gf_motion_error;
643 //d->bmi.mv.as_mv.row = tmp_mv.row;
644 //d->bmi.mv.as_mv.col = tmp_mv.col;
648 xd->pre.y_buffer = cm->last_frame.y_buffer + recon_yoffset;
649 xd->pre.u_buffer = cm->last_frame.u_buffer + recon_uvoffset;
650 xd->pre.v_buffer = cm->last_frame.v_buffer + recon_uvoffset;
654 // Reset to last frame as reference buffer
655 xd->pre.y_buffer = lst_yv12->y_buffer + recon_yoffset;
656 xd->pre.u_buffer = lst_yv12->u_buffer + recon_uvoffset;
657 xd->pre.v_buffer = lst_yv12->v_buffer + recon_uvoffset;
660 /* Intra assumed best */
661 best_ref_mv.as_int = 0;
663 if (motion_error <= this_error)
665 // Keep a count of cases where the inter and intra were
666 // very close and very low. This helps with scene cut
667 // detection for example in cropped clips with black bars
668 // at the sides or top and bottom.
669 if( (((this_error-intrapenalty) * 9) <=
670 (motion_error*10)) &&
671 (this_error < (2*intrapenalty)) )
676 d->bmi.mv.as_mv.row <<= 3;
677 d->bmi.mv.as_mv.col <<= 3;
678 this_error = motion_error;
679 vp8_set_mbmode_and_mvs(x, NEWMV, &d->bmi.mv);
680 vp8_encode_inter16x16y(IF_RTCD(&cpi->rtcd), x);
681 sum_mvr += d->bmi.mv.as_mv.row;
682 sum_mvr_abs += abs(d->bmi.mv.as_mv.row);
683 sum_mvc += d->bmi.mv.as_mv.col;
684 sum_mvc_abs += abs(d->bmi.mv.as_mv.col);
685 sum_mvrs += d->bmi.mv.as_mv.row * d->bmi.mv.as_mv.row;
686 sum_mvcs += d->bmi.mv.as_mv.col * d->bmi.mv.as_mv.col;
689 best_ref_mv.as_int = d->bmi.mv.as_int;
691 // Was the vector non-zero
692 if (d->bmi.mv.as_int)
696 // Does the Row vector point inwards or outwards
697 if (mb_row < cm->mb_rows / 2)
699 if (d->bmi.mv.as_mv.row > 0)
701 else if (d->bmi.mv.as_mv.row < 0)
704 else if (mb_row > cm->mb_rows / 2)
706 if (d->bmi.mv.as_mv.row > 0)
708 else if (d->bmi.mv.as_mv.row < 0)
712 // Does the Row vector point inwards or outwards
713 if (mb_col < cm->mb_cols / 2)
715 if (d->bmi.mv.as_mv.col > 0)
717 else if (d->bmi.mv.as_mv.col < 0)
720 else if (mb_col > cm->mb_cols / 2)
722 if (d->bmi.mv.as_mv.col > 0)
724 else if (d->bmi.mv.as_mv.col < 0)
731 coded_error += (long long)this_error;
733 // adjust to the next column of macroblocks
734 x->src.y_buffer += 16;
735 x->src.u_buffer += 8;
736 x->src.v_buffer += 8;
742 // adjust to the next row of mbs
743 x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
744 x->src.u_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
745 x->src.v_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
747 //extend the recon for intra prediction
748 vp8_extend_mb_row(new_yv12, xd->dst.y_buffer + 16, xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
749 vp8_clear_system_state(); //__asm emms;
752 vp8_clear_system_state(); //__asm emms;
758 fps.frame = cm->current_video_frame ;
759 fps.intra_error = intra_error >> 8;
760 fps.coded_error = coded_error >> 8;
761 weight = simple_weight(cpi->Source);
767 fps.ssim_weighted_pred_err = fps.coded_error * weight;
769 fps.pcnt_inter = 0.0;
770 fps.pcnt_motion = 0.0;
777 fps.mv_in_out_count = 0.0;
780 fps.pcnt_inter = 1.0 * (double)intercount / cm->MBs;
781 fps.pcnt_second_ref = 1.0 * (double)second_ref_count / cm->MBs;
782 fps.pcnt_neutral = 1.0 * (double)neutral_count / cm->MBs;
786 fps.MVr = (double)sum_mvr / (double)mvcount;
787 fps.mvr_abs = (double)sum_mvr_abs / (double)mvcount;
788 fps.MVc = (double)sum_mvc / (double)mvcount;
789 fps.mvc_abs = (double)sum_mvc_abs / (double)mvcount;
790 fps.MVrv = ((double)sum_mvrs - (fps.MVr * fps.MVr / (double)mvcount)) / (double)mvcount;
791 fps.MVcv = ((double)sum_mvcs - (fps.MVc * fps.MVc / (double)mvcount)) / (double)mvcount;
792 fps.mv_in_out_count = (double)sum_in_vectors / (double)(mvcount * 2);
794 fps.pcnt_motion = 1.0 * (double)mvcount / cpi->common.MBs;
797 // TODO: handle the case when duration is set to 0, or something less
798 // than the full time between subsequent cpi->source_time_stamp s .
799 fps.duration = cpi->source->ts_end
800 - cpi->source->ts_start;
802 // don't want to do output stats with a stack variable!
803 memcpy(cpi->twopass.this_frame_stats,
805 sizeof(FIRSTPASS_STATS));
806 output_stats(cpi, cpi->output_pkt_list, cpi->twopass.this_frame_stats);
807 accumulate_stats(cpi->twopass.total_stats, &fps);
810 // Copy the previous Last Frame into the GF buffer if specific conditions for doing so are met
811 if ((cm->current_video_frame > 0) &&
812 (cpi->twopass.this_frame_stats->pcnt_inter > 0.20) &&
813 ((cpi->twopass.this_frame_stats->intra_error / cpi->twopass.this_frame_stats->coded_error) > 2.0))
815 vp8_yv12_copy_frame_ptr(lst_yv12, gld_yv12);
818 // swap frame pointers so last frame refers to the frame we just compressed
819 vp8_swap_yv12_buffer(lst_yv12, new_yv12);
820 vp8_yv12_extend_frame_borders(lst_yv12);
822 // Special case for the first frame. Copy into the GF buffer as a second reference.
823 if (cm->current_video_frame == 0)
825 vp8_yv12_copy_frame_ptr(lst_yv12, gld_yv12);
829 // use this to see what the first pass reconstruction looks like
834 sprintf(filename, "enc%04d.yuv", (int) cm->current_video_frame);
836 if (cm->current_video_frame == 0)
837 recon_file = fopen(filename, "wb");
839 recon_file = fopen(filename, "ab");
841 if(fwrite(lst_yv12->buffer_alloc, lst_yv12->frame_size, 1, recon_file));
845 cm->current_video_frame++;
848 extern const int vp8_bits_per_mb[2][QINDEX_RANGE];
850 #define BASE_ERRPERMB 150
851 static int estimate_max_q(VP8_COMP *cpi, double section_err, int section_target_bandwitdh)
854 int num_mbs = cpi->common.MBs;
855 int target_norm_bits_per_mb;
857 double err_per_mb = section_err / num_mbs;
858 double correction_factor;
860 double speed_correction = 1.0;
861 double rolling_ratio;
863 double pow_highq = 0.90;
864 double pow_lowq = 0.40;
866 if (section_target_bandwitdh <= 0)
867 return cpi->twopass.maxq_max_limit; // Highest value allowed
869 target_norm_bits_per_mb = (section_target_bandwitdh < (1 << 20)) ? (512 * section_target_bandwitdh) / num_mbs : 512 * (section_target_bandwitdh / num_mbs);
871 // Calculate a corrective factor based on a rolling ratio of bits spent vs target bits
872 if ((cpi->rolling_target_bits > 0.0) && (cpi->active_worst_quality < cpi->worst_quality))
874 rolling_ratio = (double)cpi->rolling_actual_bits / (double)cpi->rolling_target_bits;
876 //if ( cpi->twopass.est_max_qcorrection_factor > rolling_ratio )
877 if (rolling_ratio < 0.95)
878 //cpi->twopass.est_max_qcorrection_factor *= adjustment_rate;
879 cpi->twopass.est_max_qcorrection_factor -= 0.005;
880 //else if ( cpi->twopass.est_max_qcorrection_factor < rolling_ratio )
881 else if (rolling_ratio > 1.05)
882 cpi->twopass.est_max_qcorrection_factor += 0.005;
884 //cpi->twopass.est_max_qcorrection_factor /= adjustment_rate;
886 cpi->twopass.est_max_qcorrection_factor = (cpi->twopass.est_max_qcorrection_factor < 0.1) ? 0.1 : (cpi->twopass.est_max_qcorrection_factor > 10.0) ? 10.0 : cpi->twopass.est_max_qcorrection_factor;
889 // Corrections for higher compression speed settings (reduced compression expected)
890 if ((cpi->compressor_speed == 3) || (cpi->compressor_speed == 1))
892 if (cpi->oxcf.cpu_used <= 5)
893 speed_correction = 1.04 + (cpi->oxcf.cpu_used * 0.04);
895 speed_correction = 1.25;
898 // Correction factor used for Q values >= 20
899 corr_high = pow(err_per_mb / BASE_ERRPERMB, pow_highq);
900 corr_high = (corr_high < 0.05)
901 ? 0.05 : (corr_high > 5.0) ? 5.0 : corr_high;
903 // Try and pick a max Q that will be high enough to encode the
904 // content at the given rate.
905 for (Q = cpi->twopass.maxq_min_limit; Q < cpi->twopass.maxq_max_limit; Q++)
907 int bits_per_mb_at_this_q;
911 correction_factor = pow(err_per_mb / BASE_ERRPERMB, (pow_lowq + Q * 0.01));
912 correction_factor = (correction_factor < 0.05) ? 0.05 : (correction_factor > 5.0) ? 5.0 : correction_factor;
915 correction_factor = corr_high;
917 bits_per_mb_at_this_q = (int)(.5 + correction_factor
918 * speed_correction * cpi->twopass.est_max_qcorrection_factor
919 * cpi->twopass.section_max_qfactor
920 * (double)vp8_bits_per_mb[INTER_FRAME][Q] / 1.0);
921 //bits_per_mb_at_this_q = (int)(.5 + correction_factor * speed_correction * cpi->twopass.est_max_qcorrection_factor * (double)vp8_bits_per_mb[INTER_FRAME][Q] / 1.0);
923 if (bits_per_mb_at_this_q <= target_norm_bits_per_mb)
927 // Restriction on active max q for constrained quality mode.
928 if ( (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY) &&
929 (Q < cpi->cq_target_quality) )
930 //(Q < cpi->oxcf.cq_level;) )
932 Q = cpi->cq_target_quality;
933 //Q = cpi->oxcf.cq_level;
936 // Adjust maxq_min_limit and maxq_max_limit limits based on
937 // averaga q observed in clip for non kf/gf.arf frames
938 // Give average a chance to settle though.
939 if ( (cpi->ni_frames >
940 ((unsigned int)cpi->twopass.total_stats->count >> 8)) &&
941 (cpi->ni_frames > 150) )
943 cpi->twopass.maxq_max_limit = ((cpi->ni_av_qi + 32) < cpi->worst_quality)
944 ? (cpi->ni_av_qi + 32) : cpi->worst_quality;
945 cpi->twopass.maxq_min_limit = ((cpi->ni_av_qi - 32) > cpi->best_quality)
946 ? (cpi->ni_av_qi - 32) : cpi->best_quality;
951 static int estimate_q(VP8_COMP *cpi, double section_err, int section_target_bandwitdh)
954 int num_mbs = cpi->common.MBs;
955 int target_norm_bits_per_mb;
957 double err_per_mb = section_err / num_mbs;
958 double correction_factor;
960 double speed_correction = 1.0;
961 double pow_highq = 0.90;
962 double pow_lowq = 0.40;
964 target_norm_bits_per_mb = (section_target_bandwitdh < (1 << 20)) ? (512 * section_target_bandwitdh) / num_mbs : 512 * (section_target_bandwitdh / num_mbs);
966 // Corrections for higher compression speed settings (reduced compression expected)
967 if ((cpi->compressor_speed == 3) || (cpi->compressor_speed == 1))
969 if (cpi->oxcf.cpu_used <= 5)
970 speed_correction = 1.04 + (cpi->oxcf.cpu_used * 0.04);
972 speed_correction = 1.25;
975 // Correction factor used for Q values >= 20
976 corr_high = pow(err_per_mb / BASE_ERRPERMB, pow_highq);
977 corr_high = (corr_high < 0.05) ? 0.05 : (corr_high > 5.0) ? 5.0 : corr_high;
979 // Try and pick a Q that can encode the content at the given rate.
980 for (Q = 0; Q < MAXQ; Q++)
982 int bits_per_mb_at_this_q;
986 correction_factor = pow(err_per_mb / BASE_ERRPERMB, (pow_lowq + Q * 0.01));
987 correction_factor = (correction_factor < 0.05) ? 0.05 : (correction_factor > 5.0) ? 5.0 : correction_factor;
990 correction_factor = corr_high;
992 bits_per_mb_at_this_q = (int)(.5 + correction_factor * speed_correction * cpi->twopass.est_max_qcorrection_factor * (double)vp8_bits_per_mb[INTER_FRAME][Q] / 1.0);
994 if (bits_per_mb_at_this_q <= target_norm_bits_per_mb)
1001 // Estimate a worst case Q for a KF group
1002 static int estimate_kf_group_q(VP8_COMP *cpi, double section_err, int section_target_bandwitdh, double group_iiratio)
1005 int num_mbs = cpi->common.MBs;
1006 int target_norm_bits_per_mb = (512 * section_target_bandwitdh) / num_mbs;
1007 int bits_per_mb_at_this_q;
1009 double err_per_mb = section_err / num_mbs;
1010 double err_correction_factor;
1012 double speed_correction = 1.0;
1013 double current_spend_ratio = 1.0;
1015 double pow_highq = (POW1 < 0.6) ? POW1 + 0.3 : 0.90;
1016 double pow_lowq = (POW1 < 0.7) ? POW1 + 0.1 : 0.80;
1018 double iiratio_correction_factor = 1.0;
1020 double combined_correction_factor;
1022 // Trap special case where the target is <= 0
1023 if (target_norm_bits_per_mb <= 0)
1026 // Calculate a corrective factor based on a rolling ratio of bits spent vs target bits
1027 // This is clamped to the range 0.1 to 10.0
1028 if (cpi->long_rolling_target_bits <= 0)
1029 current_spend_ratio = 10.0;
1032 current_spend_ratio = (double)cpi->long_rolling_actual_bits / (double)cpi->long_rolling_target_bits;
1033 current_spend_ratio = (current_spend_ratio > 10.0) ? 10.0 : (current_spend_ratio < 0.1) ? 0.1 : current_spend_ratio;
1036 // Calculate a correction factor based on the quality of prediction in the sequence as indicated by intra_inter error score ratio (IIRatio)
1037 // The idea here is to favour subsampling in the hardest sections vs the easyest.
1038 iiratio_correction_factor = 1.0 - ((group_iiratio - 6.0) * 0.1);
1040 if (iiratio_correction_factor < 0.5)
1041 iiratio_correction_factor = 0.5;
1043 // Corrections for higher compression speed settings (reduced compression expected)
1044 if ((cpi->compressor_speed == 3) || (cpi->compressor_speed == 1))
1046 if (cpi->oxcf.cpu_used <= 5)
1047 speed_correction = 1.04 + (cpi->oxcf.cpu_used * 0.04);
1049 speed_correction = 1.25;
1052 // Combine the various factors calculated above
1053 combined_correction_factor = speed_correction * iiratio_correction_factor * current_spend_ratio;
1055 // Correction factor used for Q values >= 20
1056 corr_high = pow(err_per_mb / BASE_ERRPERMB, pow_highq);
1057 corr_high = (corr_high < 0.05) ? 0.05 : (corr_high > 5.0) ? 5.0 : corr_high;
1059 // Try and pick a Q that should be high enough to encode the content at the given rate.
1060 for (Q = 0; Q < MAXQ; Q++)
1062 // Q values < 20 treated as a special case
1065 err_correction_factor = pow(err_per_mb / BASE_ERRPERMB, (pow_lowq + Q * 0.01));
1066 err_correction_factor = (err_correction_factor < 0.05) ? 0.05 : (err_correction_factor > 5.0) ? 5.0 : err_correction_factor;
1069 err_correction_factor = corr_high;
1071 bits_per_mb_at_this_q = (int)(.5 + err_correction_factor * combined_correction_factor * (double)vp8_bits_per_mb[INTER_FRAME][Q]);
1073 if (bits_per_mb_at_this_q <= target_norm_bits_per_mb)
1077 // If we could not hit the target even at Max Q then estimate what Q would have bee required
1078 while ((bits_per_mb_at_this_q > target_norm_bits_per_mb) && (Q < (MAXQ * 2)))
1081 bits_per_mb_at_this_q = (int)(0.96 * bits_per_mb_at_this_q);
1087 FILE *f = fopen("estkf_q.stt", "a");
1088 fprintf(f, "%8d %8d %8d %8.2f %8.3f %8.2f %8.3f %8.3f %8.3f %8d\n", cpi->common.current_video_frame, bits_per_mb_at_this_q,
1089 target_norm_bits_per_mb, err_per_mb, err_correction_factor,
1090 current_spend_ratio, group_iiratio, iiratio_correction_factor,
1091 (double)cpi->buffer_level / (double)cpi->oxcf.optimal_buffer_level, Q);
1098 // For cq mode estimate a cq level that matches the observed
1099 // complexity and data rate.
1100 static int estimate_cq(VP8_COMP *cpi, double section_err, int section_target_bandwitdh)
1103 int num_mbs = cpi->common.MBs;
1104 int target_norm_bits_per_mb;
1106 double err_per_mb = section_err / num_mbs;
1107 double correction_factor;
1109 double speed_correction = 1.0;
1110 double pow_highq = 0.90;
1111 double pow_lowq = 0.40;
1112 double clip_iiratio;
1113 double clip_iifactor;
1115 target_norm_bits_per_mb = (section_target_bandwitdh < (1 << 20))
1116 ? (512 * section_target_bandwitdh) / num_mbs
1117 : 512 * (section_target_bandwitdh / num_mbs);
1119 // Corrections for higher compression speed settings
1120 // (reduced compression expected)
1121 if ((cpi->compressor_speed == 3) || (cpi->compressor_speed == 1))
1123 if (cpi->oxcf.cpu_used <= 5)
1124 speed_correction = 1.04 + (cpi->oxcf.cpu_used * 0.04);
1126 speed_correction = 1.25;
1128 // II ratio correction factor for clip as a whole
1129 clip_iiratio = cpi->twopass.total_stats->intra_error /
1130 DOUBLE_DIVIDE_CHECK(cpi->twopass.total_stats->coded_error);
1131 clip_iifactor = 1.0 - ((clip_iiratio - 10.0) * 0.025);
1132 if (clip_iifactor < 0.80)
1133 clip_iifactor = 0.80;
1135 // Correction factor used for Q values >= 20
1136 corr_high = pow(err_per_mb / BASE_ERRPERMB, pow_highq);
1137 corr_high = (corr_high < 0.05) ? 0.05 : (corr_high > 5.0) ? 5.0 : corr_high;
1139 // Try and pick a Q that can encode the content at the given rate.
1140 for (Q = 0; Q < MAXQ; Q++)
1142 int bits_per_mb_at_this_q;
1147 pow( err_per_mb / BASE_ERRPERMB, (pow_lowq + Q * 0.01));
1149 correction_factor = (correction_factor < 0.05) ? 0.05
1150 : (correction_factor > 5.0) ? 5.0
1151 : correction_factor;
1154 correction_factor = corr_high;
1156 bits_per_mb_at_this_q =
1157 (int)( .5 + correction_factor *
1160 (double)vp8_bits_per_mb[INTER_FRAME][Q] / 1.0);
1162 if (bits_per_mb_at_this_q <= target_norm_bits_per_mb)
1169 extern void vp8_new_frame_rate(VP8_COMP *cpi, double framerate);
1171 void vp8_init_second_pass(VP8_COMP *cpi)
1173 FIRSTPASS_STATS this_frame;
1174 FIRSTPASS_STATS *start_pos;
1176 double two_pass_min_rate = (double)(cpi->oxcf.target_bandwidth * cpi->oxcf.two_pass_vbrmin_section / 100);
1178 zero_stats(cpi->twopass.total_stats);
1180 if (!cpi->twopass.stats_in_end)
1183 *cpi->twopass.total_stats = *cpi->twopass.stats_in_end;
1185 cpi->twopass.total_error_left = cpi->twopass.total_stats->ssim_weighted_pred_err;
1186 cpi->twopass.total_intra_error_left = cpi->twopass.total_stats->intra_error;
1187 cpi->twopass.total_coded_error_left = cpi->twopass.total_stats->coded_error;
1188 cpi->twopass.start_tot_err_left = cpi->twopass.total_error_left;
1190 //cpi->twopass.bits_left = (long long)(cpi->twopass.total_stats->count * cpi->oxcf.target_bandwidth / DOUBLE_DIVIDE_CHECK((double)cpi->oxcf.frame_rate));
1191 //cpi->twopass.bits_left -= (long long)(cpi->twopass.total_stats->count * two_pass_min_rate / DOUBLE_DIVIDE_CHECK((double)cpi->oxcf.frame_rate));
1193 // each frame can have a different duration, as the frame rate in the source
1194 // isn't guaranteed to be constant. The frame rate prior to the first frame
1195 // encoded in the second pass is a guess. However the sum duration is not.
1196 // Its calculated based on the actual durations of all frames from the first
1198 vp8_new_frame_rate(cpi, 10000000.0 * cpi->twopass.total_stats->count / cpi->twopass.total_stats->duration);
1200 cpi->output_frame_rate = cpi->oxcf.frame_rate;
1201 cpi->twopass.bits_left = (long long)(cpi->twopass.total_stats->duration * cpi->oxcf.target_bandwidth / 10000000.0) ;
1202 cpi->twopass.bits_left -= (long long)(cpi->twopass.total_stats->duration * two_pass_min_rate / 10000000.0);
1203 cpi->twopass.clip_bits_total = cpi->twopass.bits_left;
1205 // Calculate a minimum intra value to be used in determining the IIratio
1206 // scores used in the second pass. We have this minimum to make sure
1207 // that clips that are static but "low complexity" in the intra domain
1208 // are still boosted appropriately for KF/GF/ARF
1209 cpi->twopass.kf_intra_err_min = KF_MB_INTRA_MIN * cpi->common.MBs;
1210 cpi->twopass.gf_intra_err_min = GF_MB_INTRA_MIN * cpi->common.MBs;
1212 avg_stats(cpi->twopass.total_stats);
1214 // Scan the first pass file and calculate an average Intra / Inter error score ratio for the sequence
1216 double sum_iiratio = 0.0;
1219 start_pos = cpi->twopass.stats_in; // Note starting "file" position
1221 while (input_stats(cpi, &this_frame) != EOF)
1223 IIRatio = this_frame.intra_error / DOUBLE_DIVIDE_CHECK(this_frame.coded_error);
1224 IIRatio = (IIRatio < 1.0) ? 1.0 : (IIRatio > 20.0) ? 20.0 : IIRatio;
1225 sum_iiratio += IIRatio;
1228 cpi->twopass.avg_iiratio = sum_iiratio / DOUBLE_DIVIDE_CHECK((double)cpi->twopass.total_stats->count);
1230 // Reset file position
1231 reset_fpf_position(cpi, start_pos);
1234 // Scan the first pass file and calculate a modified total error based upon the bias/power function
1235 // used to allocate bits
1237 start_pos = cpi->twopass.stats_in; // Note starting "file" position
1239 cpi->twopass.modified_error_total = 0.0;
1240 cpi->twopass.modified_error_used = 0.0;
1242 while (input_stats(cpi, &this_frame) != EOF)
1244 cpi->twopass.modified_error_total += calculate_modified_err(cpi, &this_frame);
1246 cpi->twopass.modified_error_left = cpi->twopass.modified_error_total;
1248 reset_fpf_position(cpi, start_pos); // Reset file position
1253 void vp8_end_second_pass(VP8_COMP *cpi)
1257 // This function gives and estimate of how badly we believe
1258 // the prediction quality is decaying from frame to frame.
1259 static double get_prediction_decay_rate(VP8_COMP *cpi, FIRSTPASS_STATS *next_frame)
1261 double prediction_decay_rate;
1262 double motion_decay;
1263 double motion_pct = next_frame->pcnt_motion;
1266 // Initial basis is the % mbs inter coded
1267 prediction_decay_rate = next_frame->pcnt_inter;
1269 // High % motion -> somewhat higher decay rate
1270 motion_decay = (1.0 - (motion_pct / 20.0));
1271 if (motion_decay < prediction_decay_rate)
1272 prediction_decay_rate = motion_decay;
1274 // Adjustment to decay rate based on speed of motion
1276 double this_mv_rabs;
1277 double this_mv_cabs;
1278 double distance_factor;
1280 this_mv_rabs = fabs(next_frame->mvr_abs * motion_pct);
1281 this_mv_cabs = fabs(next_frame->mvc_abs * motion_pct);
1283 distance_factor = sqrt((this_mv_rabs * this_mv_rabs) +
1284 (this_mv_cabs * this_mv_cabs)) / 250.0;
1285 distance_factor = ((distance_factor > 1.0)
1286 ? 0.0 : (1.0 - distance_factor));
1287 if (distance_factor < prediction_decay_rate)
1288 prediction_decay_rate = distance_factor;
1291 return prediction_decay_rate;
1294 // Function to test for a condition where a complex transition is followed
1295 // by a static section. For example in slide shows where there is a fade
1296 // between slides. This is to help with more optimal kf and gf positioning.
1297 static int detect_transition_to_still(
1301 double loop_decay_rate,
1302 double decay_accumulator )
1304 BOOL trans_to_still = FALSE;
1306 // Break clause to detect very still sections after motion
1307 // For example a static image after a fade or other transition
1308 // instead of a clean scene cut.
1309 if ( (frame_interval > MIN_GF_INTERVAL) &&
1310 (loop_decay_rate >= 0.999) &&
1311 (decay_accumulator < 0.9) )
1314 FIRSTPASS_STATS * position = cpi->twopass.stats_in;
1315 FIRSTPASS_STATS tmp_next_frame;
1318 // Look ahead a few frames to see if static condition
1320 for ( j = 0; j < still_interval; j++ )
1322 if (EOF == input_stats(cpi, &tmp_next_frame))
1325 decay_rate = get_prediction_decay_rate(cpi, &tmp_next_frame);
1326 if ( decay_rate < 0.999 )
1329 // Reset file position
1330 reset_fpf_position(cpi, position);
1332 // Only if it does do we signal a transition to still
1333 if ( j == still_interval )
1334 trans_to_still = TRUE;
1337 return trans_to_still;
1340 // Analyse and define a gf/arf group .
1341 static void define_gf_group(VP8_COMP *cpi, FIRSTPASS_STATS *this_frame)
1343 FIRSTPASS_STATS next_frame;
1344 FIRSTPASS_STATS *start_pos;
1346 double boost_score = 0.0;
1347 double old_boost_score = 0.0;
1348 double gf_group_err = 0.0;
1349 double gf_first_frame_err = 0.0;
1350 double mod_frame_err = 0.0;
1352 double mv_accumulator_rabs = 0.0;
1353 double mv_accumulator_cabs = 0.0;
1354 double mv_ratio_accumulator = 0.0;
1355 double decay_accumulator = 1.0;
1357 double boost_factor = IIFACTOR;
1358 double loop_decay_rate = 1.00; // Starting decay rate
1360 double this_frame_mv_in_out = 0.0;
1361 double mv_in_out_accumulator = 0.0;
1362 double abs_mv_in_out_accumulator = 0.0;
1363 double mod_err_per_mb_accumulator = 0.0;
1365 int max_bits = frame_max_bits(cpi); // Max for a single frame
1367 unsigned int allow_alt_ref =
1368 cpi->oxcf.play_alternate && cpi->oxcf.lag_in_frames;
1370 cpi->twopass.gf_group_bits = 0;
1371 cpi->twopass.gf_decay_rate = 0;
1373 vp8_clear_system_state(); //__asm emms;
1375 start_pos = cpi->twopass.stats_in;
1377 vpx_memset(&next_frame, 0, sizeof(next_frame)); // assure clean
1379 // Preload the stats for the next frame.
1380 mod_frame_err = calculate_modified_err(cpi, this_frame);
1382 // Note the error of the frame at the start of the group (this will be
1383 // the GF frame error if we code a normal gf
1384 gf_first_frame_err = mod_frame_err;
1386 // Special treatment if the current frame is a key frame (which is also
1387 // a gf). If it is then its error score (and hence bit allocation) need
1388 // to be subtracted out from the calculation for the GF group
1389 if (cpi->common.frame_type == KEY_FRAME)
1390 gf_group_err -= gf_first_frame_err;
1392 // Scan forward to try and work out how many frames the next gf group
1393 // should contain and what level of boost is appropriate for the GF
1394 // or ARF that will be coded with the group
1397 while (((i < cpi->twopass.static_scene_max_gf_interval) ||
1398 ((cpi->twopass.frames_to_key - i) < MIN_GF_INTERVAL)) &&
1399 (i < cpi->twopass.frames_to_key))
1402 double this_frame_mvr_ratio;
1403 double this_frame_mvc_ratio;
1404 //double motion_pct = next_frame.pcnt_motion;
1407 i++; // Increment the loop counter
1409 // Accumulate error score of frames in this gf group
1410 mod_frame_err = calculate_modified_err(cpi, this_frame);
1412 gf_group_err += mod_frame_err;
1414 mod_err_per_mb_accumulator +=
1415 mod_frame_err / DOUBLE_DIVIDE_CHECK((double)cpi->common.MBs);
1417 if (EOF == input_stats(cpi, &next_frame))
1420 // Accumulate motion stats.
1421 motion_pct = next_frame.pcnt_motion;
1422 mv_accumulator_rabs += fabs(next_frame.mvr_abs * motion_pct);
1423 mv_accumulator_cabs += fabs(next_frame.mvc_abs * motion_pct);
1425 //Accumulate Motion In/Out of frame stats
1426 this_frame_mv_in_out =
1427 next_frame.mv_in_out_count * motion_pct;
1428 mv_in_out_accumulator +=
1429 next_frame.mv_in_out_count * motion_pct;
1430 abs_mv_in_out_accumulator +=
1431 fabs(next_frame.mv_in_out_count * motion_pct);
1433 // If there is a significant amount of motion
1434 if (motion_pct > 0.05)
1436 this_frame_mvr_ratio = fabs(next_frame.mvr_abs) /
1437 DOUBLE_DIVIDE_CHECK(fabs(next_frame.MVr));
1439 this_frame_mvc_ratio = fabs(next_frame.mvc_abs) /
1440 DOUBLE_DIVIDE_CHECK(fabs(next_frame.MVc));
1442 mv_ratio_accumulator +=
1443 (this_frame_mvr_ratio < next_frame.mvr_abs)
1444 ? (this_frame_mvr_ratio * motion_pct)
1445 : next_frame.mvr_abs * motion_pct;
1447 mv_ratio_accumulator +=
1448 (this_frame_mvc_ratio < next_frame.mvc_abs)
1449 ? (this_frame_mvc_ratio * motion_pct)
1450 : next_frame.mvc_abs * motion_pct;
1454 mv_ratio_accumulator += 0.0;
1455 this_frame_mvr_ratio = 1.0;
1456 this_frame_mvc_ratio = 1.0;
1459 // Underlying boost factor is based on inter intra error ratio
1460 r = ( boost_factor *
1461 ( next_frame.intra_error /
1462 DOUBLE_DIVIDE_CHECK(next_frame.coded_error)));
1464 if (next_frame.intra_error > cpi->twopass.gf_intra_err_min)
1465 r = (IIKFACTOR2 * next_frame.intra_error /
1466 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
1468 r = (IIKFACTOR2 * cpi->twopass.gf_intra_err_min /
1469 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
1471 // Increase boost for frames where new data coming into frame
1472 // (eg zoom out). Slightly reduce boost if there is a net balance
1473 // of motion out of the frame (zoom in).
1474 // The range for this_frame_mv_in_out is -1.0 to +1.0
1475 if (this_frame_mv_in_out > 0.0)
1476 r += r * (this_frame_mv_in_out * 2.0);
1477 // In extreme case boost is halved
1479 r += r * (this_frame_mv_in_out / 2.0);
1484 loop_decay_rate = get_prediction_decay_rate(cpi, &next_frame);
1486 // Cumulative effect of decay
1487 decay_accumulator = decay_accumulator * loop_decay_rate;
1488 decay_accumulator = decay_accumulator < 0.1 ? 0.1 : decay_accumulator;
1490 boost_score += (decay_accumulator * r);
1492 // Break clause to detect very still sections after motion
1493 // For example a staic image after a fade or other transition.
1494 if ( detect_transition_to_still( cpi, i, 5,
1495 loop_decay_rate, decay_accumulator ) )
1497 allow_alt_ref = FALSE;
1498 boost_score = old_boost_score;
1502 // Break out conditions.
1504 // Break at cpi->max_gf_interval unless almost totally static
1505 (i >= cpi->max_gf_interval && (decay_accumulator < 0.995)) ||
1507 // Dont break out with a very short interval
1508 (i > MIN_GF_INTERVAL) &&
1509 // Dont break out very close to a key frame
1510 ((cpi->twopass.frames_to_key - i) >= MIN_GF_INTERVAL) &&
1511 ((boost_score > 20.0) || (next_frame.pcnt_inter < 0.75)) &&
1512 ((mv_ratio_accumulator > 100.0) ||
1513 (abs_mv_in_out_accumulator > 3.0) ||
1514 (mv_in_out_accumulator < -2.0) ||
1515 ((boost_score - old_boost_score) < 2.0))
1518 boost_score = old_boost_score;
1522 vpx_memcpy(this_frame, &next_frame, sizeof(*this_frame));
1524 old_boost_score = boost_score;
1527 cpi->twopass.gf_decay_rate =
1528 (i > 0) ? (int)(100.0 * (1.0 - decay_accumulator)) / i : 0;
1530 // When using CBR apply additional buffer related upper limits
1531 if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
1535 // For cbr apply buffer related limits
1536 if (cpi->drop_frames_allowed)
1538 int df_buffer_level = cpi->oxcf.drop_frames_water_mark *
1539 (cpi->oxcf.optimal_buffer_level / 100);
1541 if (cpi->buffer_level > df_buffer_level)
1542 max_boost = ((double)((cpi->buffer_level - df_buffer_level) * 2 / 3) * 16.0) / DOUBLE_DIVIDE_CHECK((double)cpi->av_per_frame_bandwidth);
1546 else if (cpi->buffer_level > 0)
1548 max_boost = ((double)(cpi->buffer_level * 2 / 3) * 16.0) / DOUBLE_DIVIDE_CHECK((double)cpi->av_per_frame_bandwidth);
1555 if (boost_score > max_boost)
1556 boost_score = max_boost;
1559 cpi->gfu_boost = (int)(boost_score * 100.0) >> 4;
1561 // Dont allow conventional gf too near the next kf
1562 if ((cpi->twopass.frames_to_key - i) < MIN_GF_INTERVAL)
1564 while (i < cpi->twopass.frames_to_key)
1568 if (EOF == input_stats(cpi, this_frame))
1571 if (i < cpi->twopass.frames_to_key)
1573 mod_frame_err = calculate_modified_err(cpi, this_frame);
1574 gf_group_err += mod_frame_err;
1579 // Should we use the alternate refernce frame
1580 if (allow_alt_ref &&
1581 (i >= MIN_GF_INTERVAL) &&
1582 // dont use ARF very near next kf
1583 (i <= (cpi->twopass.frames_to_key - MIN_GF_INTERVAL)) &&
1584 (((next_frame.pcnt_inter > 0.75) &&
1585 ((mv_in_out_accumulator / (double)i > -0.2) || (mv_in_out_accumulator > -2.0)) &&
1586 //(cpi->gfu_boost>150) &&
1587 (cpi->gfu_boost > 100) &&
1588 //(cpi->gfu_boost>AF_THRESH2) &&
1589 //((cpi->gfu_boost/i)>AF_THRESH) &&
1590 //(decay_accumulator > 0.5) &&
1591 (cpi->twopass.gf_decay_rate <= (ARF_DECAY_THRESH + (cpi->gfu_boost / 200)))
1597 int allocation_chunks;
1598 int Q = (cpi->oxcf.fixed_q < 0) ? cpi->last_q[INTER_FRAME] : cpi->oxcf.fixed_q;
1600 int arf_frame_bits = 0;
1603 // Estimate the bits to be allocated to the group as a whole
1604 if ((cpi->twopass.kf_group_bits > 0) && (cpi->twopass.kf_group_error_left > 0))
1605 group_bits = (int)((double)cpi->twopass.kf_group_bits * (gf_group_err / (double)cpi->twopass.kf_group_error_left));
1609 // Boost for arf frame
1610 Boost = (cpi->gfu_boost * 3 * GFQ_ADJUSTMENT) / (2 * 100);
1612 allocation_chunks = (i * 100) + Boost;
1614 // Normalize Altboost and allocations chunck down to prevent overflow
1615 while (Boost > 1000)
1618 allocation_chunks /= 2;
1621 // Calculate the number of bits to be spent on the arf based on the boost number
1622 arf_frame_bits = (int)((double)Boost * (group_bits / (double)allocation_chunks));
1624 // Estimate if there are enough bits available to make worthwhile use of an arf.
1625 tmp_q = estimate_q(cpi, mod_frame_err, (int)arf_frame_bits);
1627 // Only use an arf if it is likely we will be able to code it at a lower Q than the surrounding frames.
1628 if (tmp_q < cpi->worst_quality)
1631 int frames_after_arf;
1632 int frames_bwd = cpi->oxcf.arnr_max_frames - 1;
1633 int frames_fwd = cpi->oxcf.arnr_max_frames - 1;
1635 cpi->source_alt_ref_pending = TRUE;
1637 // For alt ref frames the error score for the end frame of the group (the alt ref frame) should not contribute to the group total and hence
1638 // the number of bit allocated to the group. Rather it forms part of the next group (it is the GF at the start of the next group)
1639 gf_group_err -= mod_frame_err;
1641 // Set the interval till the next gf or arf. For ARFs this is the number of frames to be coded before the future frame that is coded as an ARF.
1642 // The future frame itself is part of the next group
1643 cpi->baseline_gf_interval = i - 1;
1645 // Define the arnr filter width for this group of frames:
1646 // We only filter frames that lie within a distance of half
1647 // the GF interval from the ARF frame. We also have to trap
1648 // cases where the filter extends beyond the end of clip.
1649 // Note: this_frame->frame has been updated in the loop
1650 // so it now points at the ARF frame.
1651 half_gf_int = cpi->baseline_gf_interval >> 1;
1652 frames_after_arf = cpi->twopass.total_stats->count - this_frame->frame - 1;
1654 switch (cpi->oxcf.arnr_type)
1656 case 1: // Backward filter
1658 if (frames_bwd > half_gf_int)
1659 frames_bwd = half_gf_int;
1662 case 2: // Forward filter
1663 if (frames_fwd > half_gf_int)
1664 frames_fwd = half_gf_int;
1665 if (frames_fwd > frames_after_arf)
1666 frames_fwd = frames_after_arf;
1670 case 3: // Centered filter
1673 if (frames_fwd > frames_after_arf)
1674 frames_fwd = frames_after_arf;
1675 if (frames_fwd > half_gf_int)
1676 frames_fwd = half_gf_int;
1678 frames_bwd = frames_fwd;
1680 // For even length filter there is one more frame backward
1681 // than forward: e.g. len=6 ==> bbbAff, len=7 ==> bbbAfff.
1682 if (frames_bwd < half_gf_int)
1683 frames_bwd += (cpi->oxcf.arnr_max_frames+1) & 0x1;
1687 cpi->active_arnr_frames = frames_bwd + 1 + frames_fwd;
1691 cpi->source_alt_ref_pending = FALSE;
1692 cpi->baseline_gf_interval = i;
1697 cpi->source_alt_ref_pending = FALSE;
1698 cpi->baseline_gf_interval = i;
1701 // Now decide how many bits should be allocated to the GF group as a proportion of those remaining in the kf group.
1702 // The final key frame group in the clip is treated as a special case where cpi->twopass.kf_group_bits is tied to cpi->twopass.bits_left.
1703 // This is also important for short clips where there may only be one key frame.
1704 if (cpi->twopass.frames_to_key >= (int)(cpi->twopass.total_stats->count - cpi->common.current_video_frame))
1706 cpi->twopass.kf_group_bits = (cpi->twopass.bits_left > 0) ? cpi->twopass.bits_left : 0;
1709 // Calculate the bits to be allocated to the group as a whole
1710 if ((cpi->twopass.kf_group_bits > 0) && (cpi->twopass.kf_group_error_left > 0))
1711 cpi->twopass.gf_group_bits = (int)((double)cpi->twopass.kf_group_bits * (gf_group_err / (double)cpi->twopass.kf_group_error_left));
1713 cpi->twopass.gf_group_bits = 0;
1715 cpi->twopass.gf_group_bits = (cpi->twopass.gf_group_bits < 0) ? 0 : (cpi->twopass.gf_group_bits > cpi->twopass.kf_group_bits) ? cpi->twopass.kf_group_bits : cpi->twopass.gf_group_bits;
1717 // Clip cpi->twopass.gf_group_bits based on user supplied data rate variability limit (cpi->oxcf.two_pass_vbrmax_section)
1718 if (cpi->twopass.gf_group_bits > max_bits * cpi->baseline_gf_interval)
1719 cpi->twopass.gf_group_bits = max_bits * cpi->baseline_gf_interval;
1721 // Reset the file position
1722 reset_fpf_position(cpi, start_pos);
1724 // Update the record of error used so far (only done once per gf group)
1725 cpi->twopass.modified_error_used += gf_group_err;
1727 // Assign bits to the arf or gf.
1730 int frames_in_section;
1731 int allocation_chunks;
1732 int Q = (cpi->oxcf.fixed_q < 0) ? cpi->last_q[INTER_FRAME] : cpi->oxcf.fixed_q;
1735 if (cpi->source_alt_ref_pending)
1737 Boost = (cpi->gfu_boost * 3 * GFQ_ADJUSTMENT) / (2 * 100);
1738 //Boost += (cpi->baseline_gf_interval * 25);
1739 Boost += (cpi->baseline_gf_interval * 50);
1741 // Set max and minimum boost and hence minimum allocation
1742 if (Boost > ((cpi->baseline_gf_interval + 1) * 200))
1743 Boost = ((cpi->baseline_gf_interval + 1) * 200);
1744 else if (Boost < 125)
1747 frames_in_section = cpi->baseline_gf_interval + 1;
1748 allocation_chunks = (frames_in_section * 100) + Boost;
1750 // Else for standard golden frames
1753 // boost based on inter / intra ratio of subsequent frames
1754 Boost = (cpi->gfu_boost * GFQ_ADJUSTMENT) / 100;
1756 // Set max and minimum boost and hence minimum allocation
1757 if (Boost > (cpi->baseline_gf_interval * 150))
1758 Boost = (cpi->baseline_gf_interval * 150);
1759 else if (Boost < 125)
1762 frames_in_section = cpi->baseline_gf_interval;
1763 allocation_chunks = (frames_in_section * 100) + (Boost - 100);
1766 // Normalize Altboost and allocations chunck down to prevent overflow
1767 while (Boost > 1000)
1770 allocation_chunks /= 2;
1773 // Calculate the number of bits to be spent on the gf or arf based on the boost number
1774 cpi->twopass.gf_bits = (int)((double)Boost * (cpi->twopass.gf_group_bits / (double)allocation_chunks));
1776 // If the frame that is to be boosted is simpler than the average for
1777 // the gf/arf group then use an alternative calculation
1778 // based on the error score of the frame itself
1779 if (mod_frame_err < gf_group_err / (double)cpi->baseline_gf_interval)
1781 double alt_gf_grp_bits;
1785 (double)cpi->twopass.kf_group_bits *
1786 (mod_frame_err * (double)cpi->baseline_gf_interval) /
1787 DOUBLE_DIVIDE_CHECK((double)cpi->twopass.kf_group_error_left);
1789 alt_gf_bits = (int)((double)Boost * (alt_gf_grp_bits /
1790 (double)allocation_chunks));
1792 if (cpi->twopass.gf_bits > alt_gf_bits)
1794 cpi->twopass.gf_bits = alt_gf_bits;
1797 // Else if it is harder than other frames in the group make sure it at
1798 // least receives an allocation in keeping with its relative error
1799 // score, otherwise it may be worse off than an "un-boosted" frame
1803 (int)((double)cpi->twopass.kf_group_bits *
1805 DOUBLE_DIVIDE_CHECK((double)cpi->twopass.kf_group_error_left));
1807 if (alt_gf_bits > cpi->twopass.gf_bits)
1809 cpi->twopass.gf_bits = alt_gf_bits;
1813 // Apply an additional limit for CBR
1814 if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
1816 if (cpi->twopass.gf_bits > (cpi->buffer_level >> 1))
1817 cpi->twopass.gf_bits = cpi->buffer_level >> 1;
1820 // Dont allow a negative value for gf_bits
1821 if (cpi->twopass.gf_bits < 0)
1822 cpi->twopass.gf_bits = 0;
1824 // Adjust KF group bits and error remainin
1825 cpi->twopass.kf_group_error_left -= gf_group_err;
1826 cpi->twopass.kf_group_bits -= cpi->twopass.gf_group_bits;
1828 if (cpi->twopass.kf_group_bits < 0)
1829 cpi->twopass.kf_group_bits = 0;
1831 // Note the error score left in the remaining frames of the group.
1832 // For normal GFs we want to remove the error score for the first frame of the group (except in Key frame case where this has already happened)
1833 if (!cpi->source_alt_ref_pending && cpi->common.frame_type != KEY_FRAME)
1834 cpi->twopass.gf_group_error_left = gf_group_err - gf_first_frame_err;
1836 cpi->twopass.gf_group_error_left = gf_group_err;
1838 cpi->twopass.gf_group_bits -= cpi->twopass.gf_bits;
1840 if (cpi->twopass.gf_group_bits < 0)
1841 cpi->twopass.gf_group_bits = 0;
1843 // Set aside some bits for a mid gf sequence boost
1844 if ((cpi->gfu_boost > 150) && (cpi->baseline_gf_interval > 5))
1846 int pct_extra = (cpi->gfu_boost - 100) / 50;
1847 pct_extra = (pct_extra > 10) ? 10 : pct_extra;
1849 cpi->twopass.mid_gf_extra_bits = (cpi->twopass.gf_group_bits * pct_extra) / 100;
1850 cpi->twopass.gf_group_bits -= cpi->twopass.mid_gf_extra_bits;
1853 cpi->twopass.mid_gf_extra_bits = 0;
1855 cpi->twopass.gf_bits += cpi->min_frame_bandwidth; // Add in minimum for a frame
1858 if (!cpi->source_alt_ref_pending && (cpi->common.frame_type != KEY_FRAME)) // Normal GF and not a KF
1860 cpi->per_frame_bandwidth = cpi->twopass.gf_bits; // Per frame bit target for this frame
1863 // Adjustment to estimate_max_q based on a measure of complexity of the section
1864 if (cpi->common.frame_type != KEY_FRAME)
1866 FIRSTPASS_STATS sectionstats;
1869 zero_stats(§ionstats);
1870 reset_fpf_position(cpi, start_pos);
1872 for (i = 0 ; i < cpi->baseline_gf_interval ; i++)
1874 input_stats(cpi, &next_frame);
1875 accumulate_stats(§ionstats, &next_frame);
1878 avg_stats(§ionstats);
1880 cpi->twopass.section_intra_rating =
1881 sectionstats.intra_error /
1882 DOUBLE_DIVIDE_CHECK(sectionstats.coded_error);
1884 Ratio = sectionstats.intra_error / DOUBLE_DIVIDE_CHECK(sectionstats.coded_error);
1885 //if( (Ratio > 11) ) //&& (sectionstats.pcnt_second_ref < .20) )
1887 cpi->twopass.section_max_qfactor = 1.0 - ((Ratio - 10.0) * 0.025);
1889 if (cpi->twopass.section_max_qfactor < 0.80)
1890 cpi->twopass.section_max_qfactor = 0.80;
1894 // cpi->twopass.section_max_qfactor = 1.0;
1896 reset_fpf_position(cpi, start_pos);
1900 // Allocate bits to a normal frame that is neither a gf an arf or a key frame.
1901 static void assign_std_frame_bits(VP8_COMP *cpi, FIRSTPASS_STATS *this_frame)
1903 int target_frame_size; // gf_group_error_left
1905 double modified_err;
1906 double err_fraction; // What portion of the remaining GF group error is used by this frame
1908 int max_bits = frame_max_bits(cpi); // Max for a single frame
1910 // Calculate modified prediction error used in bit allocation
1911 modified_err = calculate_modified_err(cpi, this_frame);
1913 if (cpi->twopass.gf_group_error_left > 0)
1914 err_fraction = modified_err / cpi->twopass.gf_group_error_left; // What portion of the remaining GF group error is used by this frame
1918 target_frame_size = (int)((double)cpi->twopass.gf_group_bits * err_fraction); // How many of those bits available for allocation should we give it?
1920 // Clip to target size to 0 - max_bits (or cpi->twopass.gf_group_bits) at the top end.
1921 if (target_frame_size < 0)
1922 target_frame_size = 0;
1925 if (target_frame_size > max_bits)
1926 target_frame_size = max_bits;
1928 if (target_frame_size > cpi->twopass.gf_group_bits)
1929 target_frame_size = cpi->twopass.gf_group_bits;
1932 cpi->twopass.gf_group_error_left -= modified_err; // Adjust error remaining
1933 cpi->twopass.gf_group_bits -= target_frame_size; // Adjust bits remaining
1935 if (cpi->twopass.gf_group_bits < 0)
1936 cpi->twopass.gf_group_bits = 0;
1938 target_frame_size += cpi->min_frame_bandwidth; // Add in the minimum number of bits that is set aside for every frame.
1940 // Special case for the frame that lies half way between two gfs
1941 if (cpi->common.frames_since_golden == cpi->baseline_gf_interval / 2)
1942 target_frame_size += cpi->twopass.mid_gf_extra_bits;
1944 cpi->per_frame_bandwidth = target_frame_size; // Per frame bit target for this frame
1947 void vp8_second_pass(VP8_COMP *cpi)
1950 int frames_left = (int)(cpi->twopass.total_stats->count - cpi->common.current_video_frame);
1952 FIRSTPASS_STATS this_frame;
1953 FIRSTPASS_STATS this_frame_copy;
1955 double this_frame_error;
1956 double this_frame_intra_error;
1957 double this_frame_coded_error;
1959 FIRSTPASS_STATS *start_pos;
1961 if (!cpi->twopass.stats_in)
1966 vp8_clear_system_state();
1968 if (EOF == input_stats(cpi, &this_frame))
1971 this_frame_error = this_frame.ssim_weighted_pred_err;
1972 this_frame_intra_error = this_frame.intra_error;
1973 this_frame_coded_error = this_frame.coded_error;
1975 start_pos = cpi->twopass.stats_in;
1977 // keyframe and section processing !
1978 if (cpi->twopass.frames_to_key == 0)
1980 // Define next KF group and assign bits to it
1981 vpx_memcpy(&this_frame_copy, &this_frame, sizeof(this_frame));
1982 find_next_key_frame(cpi, &this_frame_copy);
1984 // Special case: Error error_resilient_mode mode does not make much sense for two pass but with its current meaning but this code is designed to stop
1985 // outlandish behaviour if someone does set it when using two pass. It effectively disables GF groups.
1986 // This is temporary code till we decide what should really happen in this case.
1987 if (cpi->oxcf.error_resilient_mode)
1989 cpi->twopass.gf_group_bits = cpi->twopass.kf_group_bits;
1990 cpi->twopass.gf_group_error_left = cpi->twopass.kf_group_error_left;
1991 cpi->baseline_gf_interval = cpi->twopass.frames_to_key;
1992 cpi->frames_till_gf_update_due = cpi->baseline_gf_interval;
1993 cpi->source_alt_ref_pending = FALSE;
1998 // Is this a GF / ARF (Note that a KF is always also a GF)
1999 if (cpi->frames_till_gf_update_due == 0)
2001 // Define next gf group and assign bits to it
2002 vpx_memcpy(&this_frame_copy, &this_frame, sizeof(this_frame));
2003 define_gf_group(cpi, &this_frame_copy);
2005 // If we are going to code an altref frame at the end of the group and the current frame is not a key frame....
2006 // If the previous group used an arf this frame has already benefited from that arf boost and it should not be given extra bits
2007 // If the previous group was NOT coded using arf we may want to apply some boost to this GF as well
2008 if (cpi->source_alt_ref_pending && (cpi->common.frame_type != KEY_FRAME))
2010 // Assign a standard frames worth of bits from those allocated to the GF group
2011 vpx_memcpy(&this_frame_copy, &this_frame, sizeof(this_frame));
2012 assign_std_frame_bits(cpi, &this_frame_copy);
2014 // If appropriate (we are switching into ARF active but it was not previously active) apply a boost for the gf at the start of the group.
2015 //if ( !cpi->source_alt_ref_active && (cpi->gfu_boost > 150) )
2019 int pct_extra = (cpi->gfu_boost - 100) / 50;
2021 pct_extra = (pct_extra > 20) ? 20 : pct_extra;
2023 extra_bits = (cpi->twopass.gf_group_bits * pct_extra) / 100;
2024 cpi->twopass.gf_group_bits -= extra_bits;
2025 cpi->per_frame_bandwidth += extra_bits;
2030 // Otherwise this is an ordinary frame
2033 // Special case: Error error_resilient_mode mode does not make much sense for two pass but with its current meaning but this code is designed to stop
2034 // outlandish behaviour if someone does set it when using two pass. It effectively disables GF groups.
2035 // This is temporary code till we decide what should really happen in this case.
2036 if (cpi->oxcf.error_resilient_mode)
2038 cpi->frames_till_gf_update_due = cpi->twopass.frames_to_key;
2040 if (cpi->common.frame_type != KEY_FRAME)
2042 // Assign bits from those allocated to the GF group
2043 vpx_memcpy(&this_frame_copy, &this_frame, sizeof(this_frame));
2044 assign_std_frame_bits(cpi, &this_frame_copy);
2049 // Assign bits from those allocated to the GF group
2050 vpx_memcpy(&this_frame_copy, &this_frame, sizeof(this_frame));
2051 assign_std_frame_bits(cpi, &this_frame_copy);
2055 // Keep a globally available copy of this and the next frame's iiratio.
2056 cpi->twopass.this_iiratio = this_frame_intra_error /
2057 DOUBLE_DIVIDE_CHECK(this_frame_coded_error);
2059 FIRSTPASS_STATS next_frame;
2060 if ( lookup_next_frame_stats(cpi, &next_frame) != EOF )
2062 cpi->twopass.next_iiratio = next_frame.intra_error /
2063 DOUBLE_DIVIDE_CHECK(next_frame.coded_error);
2067 // Set nominal per second bandwidth for this frame
2068 cpi->target_bandwidth = cpi->per_frame_bandwidth * cpi->output_frame_rate;
2069 if (cpi->target_bandwidth < 0)
2070 cpi->target_bandwidth = 0;
2072 if (cpi->common.current_video_frame == 0)
2074 cpi->twopass.est_max_qcorrection_factor = 1.0;
2076 // Experimental code to try and set a cq_level in constrained
2078 if ( cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY )
2084 (cpi->twopass.total_coded_error_left / frames_left),
2085 (int)(cpi->twopass.bits_left / frames_left));
2087 cpi->cq_target_quality = cpi->oxcf.cq_level;
2088 if ( est_cq > cpi->cq_target_quality )
2089 cpi->cq_target_quality = est_cq;
2092 // guess at maxq needed in 2nd pass
2093 cpi->twopass.maxq_max_limit = cpi->worst_quality;
2094 cpi->twopass.maxq_min_limit = cpi->best_quality;
2095 tmp_q = estimate_max_q( cpi,
2096 (cpi->twopass.total_coded_error_left / frames_left),
2097 (int)(cpi->twopass.bits_left / frames_left));
2099 // Limit the maxq value returned subsequently.
2100 // This increases the risk of overspend or underspend if the initial
2101 // estimate for the clip is bad, but helps prevent excessive
2102 // variation in Q, especially near the end of a clip
2103 // where for example a small overspend may cause Q to crash
2104 cpi->twopass.maxq_max_limit = ((tmp_q + 32) < cpi->worst_quality)
2105 ? (tmp_q + 32) : cpi->worst_quality;
2106 cpi->twopass.maxq_min_limit = ((tmp_q - 32) > cpi->best_quality)
2107 ? (tmp_q - 32) : cpi->best_quality;
2109 cpi->active_worst_quality = tmp_q;
2110 cpi->ni_av_qi = tmp_q;
2113 // The last few frames of a clip almost always have to few or too many
2114 // bits and for the sake of over exact rate control we dont want to make
2115 // radical adjustments to the allowed quantizer range just to use up a
2116 // few surplus bits or get beneath the target rate.
2117 else if ( (cpi->common.current_video_frame <
2118 (((unsigned int)cpi->twopass.total_stats->count * 255)>>8)) &&
2119 ((cpi->common.current_video_frame + cpi->baseline_gf_interval) <
2120 (unsigned int)cpi->twopass.total_stats->count) )
2122 if (frames_left < 1)
2125 tmp_q = estimate_max_q(cpi, (cpi->twopass.total_coded_error_left / frames_left), (int)(cpi->twopass.bits_left / frames_left));
2127 // Move active_worst_quality but in a damped way
2128 if (tmp_q > cpi->active_worst_quality)
2129 cpi->active_worst_quality ++;
2130 else if (tmp_q < cpi->active_worst_quality)
2131 cpi->active_worst_quality --;
2133 cpi->active_worst_quality = ((cpi->active_worst_quality * 3) + tmp_q + 2) / 4;
2136 cpi->twopass.frames_to_key --;
2137 cpi->twopass.total_error_left -= this_frame_error;
2138 cpi->twopass.total_intra_error_left -= this_frame_intra_error;
2139 cpi->twopass.total_coded_error_left -= this_frame_coded_error;
2143 static BOOL test_candidate_kf(VP8_COMP *cpi, FIRSTPASS_STATS *last_frame, FIRSTPASS_STATS *this_frame, FIRSTPASS_STATS *next_frame)
2145 BOOL is_viable_kf = FALSE;
2147 // Does the frame satisfy the primary criteria of a key frame
2148 // If so, then examine how well it predicts subsequent frames
2149 if ((this_frame->pcnt_second_ref < 0.10) &&
2150 (next_frame->pcnt_second_ref < 0.10) &&
2151 ((this_frame->pcnt_inter < 0.05) ||
2153 ((this_frame->pcnt_inter - this_frame->pcnt_neutral) < .25) &&
2154 ((this_frame->intra_error / DOUBLE_DIVIDE_CHECK(this_frame->coded_error)) < 2.5) &&
2155 ((fabs(last_frame->coded_error - this_frame->coded_error) / DOUBLE_DIVIDE_CHECK(this_frame->coded_error) > .40) ||
2156 (fabs(last_frame->intra_error - this_frame->intra_error) / DOUBLE_DIVIDE_CHECK(this_frame->intra_error) > .40) ||
2157 ((next_frame->intra_error / DOUBLE_DIVIDE_CHECK(next_frame->coded_error)) > 3.5)
2164 FIRSTPASS_STATS *start_pos;
2166 FIRSTPASS_STATS local_next_frame;
2168 double boost_score = 0.0;
2169 double old_boost_score = 0.0;
2170 double decay_accumulator = 1.0;
2171 double next_iiratio;
2173 vpx_memcpy(&local_next_frame, next_frame, sizeof(*next_frame));
2175 // Note the starting file position so we can reset to it
2176 start_pos = cpi->twopass.stats_in;
2178 // Examine how well the key frame predicts subsequent frames
2179 for (i = 0 ; i < 16; i++)
2181 next_iiratio = (IIKFACTOR1 * local_next_frame.intra_error / DOUBLE_DIVIDE_CHECK(local_next_frame.coded_error)) ;
2183 if (next_iiratio > RMAX)
2184 next_iiratio = RMAX;
2186 // Cumulative effect of decay in prediction quality
2187 if (local_next_frame.pcnt_inter > 0.85)
2188 decay_accumulator = decay_accumulator * local_next_frame.pcnt_inter;
2190 decay_accumulator = decay_accumulator * ((0.85 + local_next_frame.pcnt_inter) / 2.0);
2192 //decay_accumulator = decay_accumulator * local_next_frame.pcnt_inter;
2194 // Keep a running total
2195 boost_score += (decay_accumulator * next_iiratio);
2197 // Test various breakout clauses
2198 if ((local_next_frame.pcnt_inter < 0.05) ||
2199 (next_iiratio < 1.5) ||
2200 (((local_next_frame.pcnt_inter -
2201 local_next_frame.pcnt_neutral) < 0.20) &&
2202 (next_iiratio < 3.0)) ||
2203 ((boost_score - old_boost_score) < 0.5) ||
2204 (local_next_frame.intra_error < 200)
2210 old_boost_score = boost_score;
2212 // Get the next frame details
2213 if (EOF == input_stats(cpi, &local_next_frame))
2217 // If there is tolerable prediction for at least the next 3 frames then break out else discard this pottential key frame and move on
2218 if (boost_score > 5.0 && (i > 3))
2219 is_viable_kf = TRUE;
2222 // Reset the file position
2223 reset_fpf_position(cpi, start_pos);
2225 is_viable_kf = FALSE;
2229 return is_viable_kf;
2231 static void find_next_key_frame(VP8_COMP *cpi, FIRSTPASS_STATS *this_frame)
2234 FIRSTPASS_STATS last_frame;
2235 FIRSTPASS_STATS first_frame;
2236 FIRSTPASS_STATS next_frame;
2237 FIRSTPASS_STATS *start_position;
2239 double decay_accumulator = 1.0;
2240 double boost_score = 0;
2241 double old_boost_score = 0.0;
2242 double loop_decay_rate;
2244 double kf_mod_err = 0.0;
2245 double kf_group_err = 0.0;
2246 double kf_group_intra_err = 0.0;
2247 double kf_group_coded_err = 0.0;
2248 double recent_loop_decay[8] = {1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0};
2250 vpx_memset(&next_frame, 0, sizeof(next_frame)); // assure clean
2252 vp8_clear_system_state(); //__asm emms;
2253 start_position = cpi->twopass.stats_in;
2255 cpi->common.frame_type = KEY_FRAME;
2257 // is this a forced key frame by interval
2258 cpi->this_key_frame_forced = cpi->next_key_frame_forced;
2260 // Clear the alt ref active flag as this can never be active on a key frame
2261 cpi->source_alt_ref_active = FALSE;
2263 // Kf is always a gf so clear frames till next gf counter
2264 cpi->frames_till_gf_update_due = 0;
2266 cpi->twopass.frames_to_key = 1;
2268 // Take a copy of the initial frame details
2269 vpx_memcpy(&first_frame, this_frame, sizeof(*this_frame));
2271 cpi->twopass.kf_group_bits = 0; // Total bits avaialable to kf group
2272 cpi->twopass.kf_group_error_left = 0; // Group modified error score.
2274 kf_mod_err = calculate_modified_err(cpi, this_frame);
2276 // find the next keyframe
2278 while (cpi->twopass.stats_in < cpi->twopass.stats_in_end)
2280 // Accumulate kf group error
2281 kf_group_err += calculate_modified_err(cpi, this_frame);
2283 // These figures keep intra and coded error counts for all frames including key frames in the group.
2284 // The effect of the key frame itself can be subtracted out using the first_frame data collected above
2285 kf_group_intra_err += this_frame->intra_error;
2286 kf_group_coded_err += this_frame->coded_error;
2288 // load a the next frame's stats
2289 vpx_memcpy(&last_frame, this_frame, sizeof(*this_frame));
2290 input_stats(cpi, this_frame);
2292 // Provided that we are not at the end of the file...
2293 if (cpi->oxcf.auto_key
2294 && lookup_next_frame_stats(cpi, &next_frame) != EOF)
2296 // Normal scene cut check
2297 if (test_candidate_kf(cpi, &last_frame, this_frame, &next_frame))
2300 // How fast is prediction quality decaying
2301 loop_decay_rate = get_prediction_decay_rate(cpi, &next_frame);
2303 // We want to know something about the recent past... rather than
2304 // as used elsewhere where we are concened with decay in prediction
2305 // quality since the last GF or KF.
2306 recent_loop_decay[i%8] = loop_decay_rate;
2307 decay_accumulator = 1.0;
2308 for (j = 0; j < 8; j++)
2310 decay_accumulator = decay_accumulator * recent_loop_decay[j];
2313 // Special check for transition or high motion followed by a
2314 // to a static scene.
2315 if ( detect_transition_to_still( cpi, i,
2316 (cpi->key_frame_frequency-i),
2318 decay_accumulator ) )
2324 // Step on to the next frame
2325 cpi->twopass.frames_to_key ++;
2327 // If we don't have a real key frame within the next two
2328 // forcekeyframeevery intervals then break out of the loop.
2329 if (cpi->twopass.frames_to_key >= 2 *(int)cpi->key_frame_frequency)
2332 cpi->twopass.frames_to_key ++;
2337 // If there is a max kf interval set by the user we must obey it.
2338 // We already breakout of the loop above at 2x max.
2339 // This code centers the extra kf if the actual natural
2340 // interval is between 1x and 2x
2341 if (cpi->oxcf.auto_key
2342 && cpi->twopass.frames_to_key > (int)cpi->key_frame_frequency )
2344 FIRSTPASS_STATS *current_pos = cpi->twopass.stats_in;
2345 FIRSTPASS_STATS tmp_frame;
2347 cpi->twopass.frames_to_key /= 2;
2349 // Copy first frame details
2350 vpx_memcpy(&tmp_frame, &first_frame, sizeof(first_frame));
2352 // Reset to the start of the group
2353 reset_fpf_position(cpi, start_position);
2356 kf_group_intra_err = 0;
2357 kf_group_coded_err = 0;
2359 // Rescan to get the correct error data for the forced kf group
2360 for( i = 0; i < cpi->twopass.frames_to_key; i++ )
2362 // Accumulate kf group errors
2363 kf_group_err += calculate_modified_err(cpi, &tmp_frame);
2364 kf_group_intra_err += tmp_frame.intra_error;
2365 kf_group_coded_err += tmp_frame.coded_error;
2367 // Load a the next frame's stats
2368 input_stats(cpi, &tmp_frame);
2371 // Reset to the start of the group
2372 reset_fpf_position(cpi, current_pos);
2374 cpi->next_key_frame_forced = TRUE;
2377 cpi->next_key_frame_forced = FALSE;
2379 // Special case for the last frame of the file
2380 if (cpi->twopass.stats_in >= cpi->twopass.stats_in_end)
2382 // Accumulate kf group error
2383 kf_group_err += calculate_modified_err(cpi, this_frame);
2385 // These figures keep intra and coded error counts for all frames including key frames in the group.
2386 // The effect of the key frame itself can be subtracted out using the first_frame data collected above
2387 kf_group_intra_err += this_frame->intra_error;
2388 kf_group_coded_err += this_frame->coded_error;
2391 // Calculate the number of bits that should be assigned to the kf group.
2392 if ((cpi->twopass.bits_left > 0) && (cpi->twopass.modified_error_left > 0.0))
2394 // Max for a single normal frame (not key frame)
2395 int max_bits = frame_max_bits(cpi);
2397 // Maximum bits for the kf group
2398 long long max_grp_bits;
2400 // Default allocation based on bits left and relative
2401 // complexity of the section
2402 cpi->twopass.kf_group_bits = (long long)( cpi->twopass.bits_left *
2404 cpi->twopass.modified_error_left ));
2406 // Clip based on maximum per frame rate defined by the user.
2407 max_grp_bits = (long long)max_bits * (long long)cpi->twopass.frames_to_key;
2408 if (cpi->twopass.kf_group_bits > max_grp_bits)
2409 cpi->twopass.kf_group_bits = max_grp_bits;
2411 // Additional special case for CBR if buffer is getting full.
2412 if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
2414 int opt_buffer_lvl = cpi->oxcf.optimal_buffer_level;
2415 int buffer_lvl = cpi->buffer_level;
2417 // If the buffer is near or above the optimal and this kf group is
2418 // not being allocated much then increase the allocation a bit.
2419 if (buffer_lvl >= opt_buffer_lvl)
2421 int high_water_mark = (opt_buffer_lvl +
2422 cpi->oxcf.maximum_buffer_size) >> 1;
2424 long long av_group_bits;
2426 // Av bits per frame * number of frames
2427 av_group_bits = (long long)cpi->av_per_frame_bandwidth *
2428 (long long)cpi->twopass.frames_to_key;
2430 // We are at or above the maximum.
2431 if (cpi->buffer_level >= high_water_mark)
2433 long long min_group_bits;
2435 min_group_bits = av_group_bits +
2436 (long long)(buffer_lvl -
2439 if (cpi->twopass.kf_group_bits < min_group_bits)
2440 cpi->twopass.kf_group_bits = min_group_bits;
2442 // We are above optimal but below the maximum
2443 else if (cpi->twopass.kf_group_bits < av_group_bits)
2445 long long bits_below_av = av_group_bits -
2446 cpi->twopass.kf_group_bits;
2448 cpi->twopass.kf_group_bits +=
2449 (long long)((double)bits_below_av *
2450 (double)(buffer_lvl - opt_buffer_lvl) /
2451 (double)(high_water_mark - opt_buffer_lvl));
2457 cpi->twopass.kf_group_bits = 0;
2459 // Reset the first pass file position
2460 reset_fpf_position(cpi, start_position);
2462 // determine how big to make this keyframe based on how well the subsequent frames use inter blocks
2463 decay_accumulator = 1.0;
2465 loop_decay_rate = 1.00; // Starting decay rate
2467 for (i = 0 ; i < cpi->twopass.frames_to_key ; i++)
2471 if (EOF == input_stats(cpi, &next_frame))
2474 if (next_frame.intra_error > cpi->twopass.kf_intra_err_min)
2475 r = (IIKFACTOR2 * next_frame.intra_error /
2476 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
2478 r = (IIKFACTOR2 * cpi->twopass.kf_intra_err_min /
2479 DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
2484 // How fast is prediction quality decaying
2485 loop_decay_rate = get_prediction_decay_rate(cpi, &next_frame);
2487 decay_accumulator = decay_accumulator * loop_decay_rate;
2488 decay_accumulator = decay_accumulator < 0.1 ? 0.1 : decay_accumulator;
2490 boost_score += (decay_accumulator * r);
2492 if ((i > MIN_GF_INTERVAL) &&
2493 ((boost_score - old_boost_score) < 1.0))
2498 old_boost_score = boost_score;
2503 FIRSTPASS_STATS sectionstats;
2506 zero_stats(§ionstats);
2507 reset_fpf_position(cpi, start_position);
2509 for (i = 0 ; i < cpi->twopass.frames_to_key ; i++)
2511 input_stats(cpi, &next_frame);
2512 accumulate_stats(§ionstats, &next_frame);
2515 avg_stats(§ionstats);
2517 cpi->twopass.section_intra_rating =
2518 sectionstats.intra_error
2519 / DOUBLE_DIVIDE_CHECK(sectionstats.coded_error);
2521 Ratio = sectionstats.intra_error / DOUBLE_DIVIDE_CHECK(sectionstats.coded_error);
2522 // if( (Ratio > 11) ) //&& (sectionstats.pcnt_second_ref < .20) )
2524 cpi->twopass.section_max_qfactor = 1.0 - ((Ratio - 10.0) * 0.025);
2526 if (cpi->twopass.section_max_qfactor < 0.80)
2527 cpi->twopass.section_max_qfactor = 0.80;
2531 // cpi->twopass.section_max_qfactor = 1.0;
2534 // When using CBR apply additional buffer fullness related upper limits
2535 if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
2539 if (cpi->drop_frames_allowed)
2541 int df_buffer_level = cpi->oxcf.drop_frames_water_mark * (cpi->oxcf.optimal_buffer_level / 100);
2543 if (cpi->buffer_level > df_buffer_level)
2544 max_boost = ((double)((cpi->buffer_level - df_buffer_level) * 2 / 3) * 16.0) / DOUBLE_DIVIDE_CHECK((double)cpi->av_per_frame_bandwidth);
2548 else if (cpi->buffer_level > 0)
2550 max_boost = ((double)(cpi->buffer_level * 2 / 3) * 16.0) / DOUBLE_DIVIDE_CHECK((double)cpi->av_per_frame_bandwidth);
2557 if (boost_score > max_boost)
2558 boost_score = max_boost;
2561 // Reset the first pass file position
2562 reset_fpf_position(cpi, start_position);
2564 // Work out how many bits to allocate for the key frame itself
2567 int kf_boost = boost_score;
2568 int allocation_chunks;
2569 int Counter = cpi->twopass.frames_to_key;
2571 YV12_BUFFER_CONFIG *lst_yv12 = &cpi->common.yv12_fb[cpi->common.lst_fb_idx];
2572 // Min boost based on kf interval
2575 while ((kf_boost < 48) && (Counter > 0))
2585 kf_boost += ((Counter + 1) >> 1);
2587 if (kf_boost > 48) kf_boost = 48;
2590 // bigger frame sizes need larger kf boosts, smaller frames smaller boosts...
2591 if ((lst_yv12->y_width * lst_yv12->y_height) > (320 * 240))
2592 kf_boost += 2 * (lst_yv12->y_width * lst_yv12->y_height) / (320 * 240);
2593 else if ((lst_yv12->y_width * lst_yv12->y_height) < (320 * 240))
2594 kf_boost -= 4 * (320 * 240) / (lst_yv12->y_width * lst_yv12->y_height);
2596 kf_boost = (int)((double)kf_boost * 100.0) >> 4; // Scale 16 to 100
2598 // Adjustment to boost based on recent average q
2599 //kf_boost = kf_boost * vp8_kf_boost_qadjustment[cpi->ni_av_qi] / 100;
2601 if (kf_boost < 250) // Min KF boost
2604 // We do three calculations for kf size.
2605 // The first is based on the error score for the whole kf group.
2606 // The second (optionaly) on the key frames own error if this is smaller than the average for the group.
2607 // The final one insures that the frame receives at least the allocation it would have received based on its own error score vs the error score remaining
2609 allocation_chunks = ((cpi->twopass.frames_to_key - 1) * 100) + kf_boost; // cpi->twopass.frames_to_key-1 because key frame itself is taken care of by kf_boost
2611 // Normalize Altboost and allocations chunck down to prevent overflow
2612 while (kf_boost > 1000)
2615 allocation_chunks /= 2;
2618 cpi->twopass.kf_group_bits = (cpi->twopass.kf_group_bits < 0) ? 0 : cpi->twopass.kf_group_bits;
2620 // Calculate the number of bits to be spent on the key frame
2621 cpi->twopass.kf_bits = (int)((double)kf_boost * ((double)cpi->twopass.kf_group_bits / (double)allocation_chunks));
2623 // Apply an additional limit for CBR
2624 if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
2626 if (cpi->twopass.kf_bits > ((3 * cpi->buffer_level) >> 2))
2627 cpi->twopass.kf_bits = (3 * cpi->buffer_level) >> 2;
2630 // If the key frame is actually easier than the average for the
2631 // kf group (which does sometimes happen... eg a blank intro frame)
2632 // Then use an alternate calculation based on the kf error score
2633 // which should give a smaller key frame.
2634 if (kf_mod_err < kf_group_err / cpi->twopass.frames_to_key)
2636 double alt_kf_grp_bits =
2637 ((double)cpi->twopass.bits_left *
2638 (kf_mod_err * (double)cpi->twopass.frames_to_key) /
2639 DOUBLE_DIVIDE_CHECK(cpi->twopass.modified_error_left));
2641 alt_kf_bits = (int)((double)kf_boost *
2642 (alt_kf_grp_bits / (double)allocation_chunks));
2644 if (cpi->twopass.kf_bits > alt_kf_bits)
2646 cpi->twopass.kf_bits = alt_kf_bits;
2649 // Else if it is much harder than other frames in the group make sure
2650 // it at least receives an allocation in keeping with its relative
2655 (int)((double)cpi->twopass.bits_left *
2657 DOUBLE_DIVIDE_CHECK(cpi->twopass.modified_error_left)));
2659 if (alt_kf_bits > cpi->twopass.kf_bits)
2661 cpi->twopass.kf_bits = alt_kf_bits;
2665 cpi->twopass.kf_group_bits -= cpi->twopass.kf_bits;
2666 cpi->twopass.kf_bits += cpi->min_frame_bandwidth; // Add in the minimum frame allowance
2668 cpi->per_frame_bandwidth = cpi->twopass.kf_bits; // Peer frame bit target for this frame
2669 cpi->target_bandwidth = cpi->twopass.kf_bits * cpi->output_frame_rate; // Convert to a per second bitrate
2672 // Note the total error score of the kf group minus the key frame itself
2673 cpi->twopass.kf_group_error_left = (int)(kf_group_err - kf_mod_err);
2675 // Adjust the count of total modified error left.
2676 // The count of bits left is adjusted elsewhere based on real coded frame sizes
2677 cpi->twopass.modified_error_left -= kf_group_err;
2679 if (cpi->oxcf.allow_spatial_resampling)
2681 int resample_trigger = FALSE;
2682 int last_kf_resampled = FALSE;
2686 int new_width = cpi->oxcf.Width;
2687 int new_height = cpi->oxcf.Height;
2689 int projected_buffer_level = cpi->buffer_level;
2692 double projected_bits_perframe;
2693 double group_iiratio = (kf_group_intra_err - first_frame.intra_error) / (kf_group_coded_err - first_frame.coded_error);
2694 double err_per_frame = kf_group_err / cpi->twopass.frames_to_key;
2695 double bits_per_frame;
2696 double av_bits_per_frame;
2697 double effective_size_ratio;
2699 if ((cpi->common.Width != cpi->oxcf.Width) || (cpi->common.Height != cpi->oxcf.Height))
2700 last_kf_resampled = TRUE;
2702 // Set back to unscaled by defaults
2703 cpi->common.horiz_scale = NORMAL;
2704 cpi->common.vert_scale = NORMAL;
2706 // Calculate Average bits per frame.
2707 //av_bits_per_frame = cpi->twopass.bits_left/(double)(cpi->twopass.total_stats->count - cpi->common.current_video_frame);
2708 av_bits_per_frame = cpi->oxcf.target_bandwidth / DOUBLE_DIVIDE_CHECK((double)cpi->oxcf.frame_rate);
2709 //if ( av_bits_per_frame < 0.0 )
2710 // av_bits_per_frame = 0.0
2712 // CBR... Use the clip average as the target for deciding resample
2713 if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
2715 bits_per_frame = av_bits_per_frame;
2718 // In VBR we want to avoid downsampling in easy section unless we are under extreme pressure
2719 // So use the larger of target bitrate for this sectoion or average bitrate for sequence
2722 bits_per_frame = cpi->twopass.kf_group_bits / cpi->twopass.frames_to_key; // This accounts for how hard the section is...
2724 if (bits_per_frame < av_bits_per_frame) // Dont turn to resampling in easy sections just because they have been assigned a small number of bits
2725 bits_per_frame = av_bits_per_frame;
2728 // bits_per_frame should comply with our minimum
2729 if (bits_per_frame < (cpi->oxcf.target_bandwidth * cpi->oxcf.two_pass_vbrmin_section / 100))
2730 bits_per_frame = (cpi->oxcf.target_bandwidth * cpi->oxcf.two_pass_vbrmin_section / 100);
2732 // Work out if spatial resampling is necessary
2733 kf_q = estimate_kf_group_q(cpi, err_per_frame, bits_per_frame, group_iiratio);
2735 // If we project a required Q higher than the maximum allowed Q then make a guess at the actual size of frames in this section
2736 projected_bits_perframe = bits_per_frame;
2739 while (tmp_q > cpi->worst_quality)
2741 projected_bits_perframe *= 1.04;
2745 // Guess at buffer level at the end of the section
2746 projected_buffer_level = cpi->buffer_level - (int)((projected_bits_perframe - av_bits_per_frame) * cpi->twopass.frames_to_key);
2750 FILE *f = fopen("Subsamle.stt", "a");
2751 fprintf(f, " %8d %8d %8d %8d %12.0f %8d %8d %8d\n", cpi->common.current_video_frame, kf_q, cpi->common.horiz_scale, cpi->common.vert_scale, kf_group_err / cpi->twopass.frames_to_key, (int)(cpi->twopass.kf_group_bits / cpi->twopass.frames_to_key), new_height, new_width);
2755 // The trigger for spatial resampling depends on the various parameters such as whether we are streaming (CBR) or VBR.
2756 if (cpi->oxcf.end_usage == USAGE_STREAM_FROM_SERVER)
2758 // Trigger resample if we are projected to fall below down sample level or
2759 // resampled last time and are projected to remain below the up sample level
2760 if ((projected_buffer_level < (cpi->oxcf.resample_down_water_mark * cpi->oxcf.optimal_buffer_level / 100)) ||
2761 (last_kf_resampled && (projected_buffer_level < (cpi->oxcf.resample_up_water_mark * cpi->oxcf.optimal_buffer_level / 100))))
2762 //( ((cpi->buffer_level < (cpi->oxcf.resample_down_water_mark * cpi->oxcf.optimal_buffer_level / 100))) &&
2763 // ((projected_buffer_level < (cpi->oxcf.resample_up_water_mark * cpi->oxcf.optimal_buffer_level / 100))) ))
2764 resample_trigger = TRUE;
2766 resample_trigger = FALSE;
2770 long long clip_bits = (long long)(cpi->twopass.total_stats->count * cpi->oxcf.target_bandwidth / DOUBLE_DIVIDE_CHECK((double)cpi->oxcf.frame_rate));
2771 long long over_spend = cpi->oxcf.starting_buffer_level - cpi->buffer_level;
2773 if ((last_kf_resampled && (kf_q > cpi->worst_quality)) || // If triggered last time the threshold for triggering again is reduced
2774 ((kf_q > cpi->worst_quality) && // Projected Q higher than allowed and ...
2775 (over_spend > clip_bits / 20))) // ... Overspend > 5% of total bits
2776 resample_trigger = TRUE;
2778 resample_trigger = FALSE;
2782 if (resample_trigger)
2784 while ((kf_q >= cpi->worst_quality) && (scale_val < 6))
2788 cpi->common.vert_scale = vscale_lookup[scale_val];
2789 cpi->common.horiz_scale = hscale_lookup[scale_val];
2791 Scale2Ratio(cpi->common.horiz_scale, &hr, &hs);
2792 Scale2Ratio(cpi->common.vert_scale, &vr, &vs);
2794 new_width = ((hs - 1) + (cpi->oxcf.Width * hr)) / hs;
2795 new_height = ((vs - 1) + (cpi->oxcf.Height * vr)) / vs;
2797 // Reducing the area to 1/4 does not reduce the complexity (err_per_frame) to 1/4...
2798 // effective_sizeratio attempts to provide a crude correction for this
2799 effective_size_ratio = (double)(new_width * new_height) / (double)(cpi->oxcf.Width * cpi->oxcf.Height);
2800 effective_size_ratio = (1.0 + (3.0 * effective_size_ratio)) / 4.0;
2802 // Now try again and see what Q we get with the smaller image size
2803 kf_q = estimate_kf_group_q(cpi, err_per_frame * effective_size_ratio, bits_per_frame, group_iiratio);
2807 FILE *f = fopen("Subsamle.stt", "a");
2808 fprintf(f, "******** %8d %8d %8d %12.0f %8d %8d %8d\n", kf_q, cpi->common.horiz_scale, cpi->common.vert_scale, kf_group_err / cpi->twopass.frames_to_key, (int)(cpi->twopass.kf_group_bits / cpi->twopass.frames_to_key), new_height, new_width);
2814 if ((cpi->common.Width != new_width) || (cpi->common.Height != new_height))
2816 cpi->common.Width = new_width;
2817 cpi->common.Height = new_height;
2818 vp8_alloc_compressor_data(cpi);