X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=encoder%2Fratecontrol.c;h=4a7cc5c5edb5104a18bc01731d01915c2b0af849;hb=87e5994706c76bd628c7e23f0dca95f05e922a7c;hp=e73c9e3521d7e764b7d44c0db77f5e95c7704584;hpb=53295729673187df0ab1143c08fae578b5447376;p=x264 diff --git a/encoder/ratecontrol.c b/encoder/ratecontrol.c index e73c9e35..4a7cc5c5 100644 --- a/encoder/ratecontrol.c +++ b/encoder/ratecontrol.c @@ -1,12 +1,12 @@ /***************************************************-*- coding: iso-8859-1 -*- * ratecontrol.c: h264 encoder library (Rate Control) ***************************************************************************** - * Copyright (C) 2003 Laurent Aimar + * Copyright (C) 2005 x264 project * $Id: ratecontrol.c,v 1.1 2004/06/03 19:27:08 fenrir Exp $ * - * Authors: Måns Rullgård - * 2 pass code: Michael Niedermayer - * Loren Merritt + * Authors: Loren Merritt + * Michael Niedermayer + * Måns Rullgård * * This program is free software; you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -34,27 +34,29 @@ #include "common/common.h" #include "common/cpu.h" -#include "common/macroblock.h" #include "ratecontrol.h" -#ifdef SYS_MACOSX -#define exp2f(x) ( (float) exp2( (x) ) ) -#endif -#ifdef SYS_FREEBSD +#if defined(SYS_FREEBSD) || defined(SYS_BEOS) || defined(SYS_NETBSD) #define exp2f(x) powf( 2, (x) ) #endif - +#if defined(_MSC_VER) || defined(SYS_SunOS) +#define exp2f(x) pow( 2, (x) ) +#define sqrtf sqrt +#endif +#ifdef WIN32 // POSIX says that rename() removes the destination, but win32 doesn't. +#define rename(src,dst) (unlink(dst), rename(src,dst)) +#endif typedef struct { int pict_type; + int kept_as_ref; float qscale; int mv_bits; int i_tex_bits; int p_tex_bits; int misc_bits; uint64_t expected_bits; - int new_pict_type; float new_qscale; int new_qp; int i_count; @@ -63,47 +65,60 @@ typedef struct float blurred_complexity; } ratecontrol_entry_t; +typedef struct +{ + double coeff; + double count; + double decay; +} predictor_t; + struct x264_ratecontrol_t { /* constants */ + int b_abr; + int b_2pass; double fps; - int gop_size; - int bitrate; + double bitrate; + double rate_tolerance; int nmb; /* number of macroblocks in a frame */ - int buffer_size; - int rcbufrate; - int init_qp; int qp_constant[5]; - /* 1st pass stuff */ - int gop_qp; - int buffer_fullness; - int frames; /* frames in current gop */ - int pframes; - int slice_type; - int mb; /* MBs processed in current frame */ - int bits_gop; /* allocated bits current gop */ - int bits_last_gop; /* bits consumed in gop */ + /* current frame */ + ratecontrol_entry_t *rce; int qp; /* qp for current frame */ - int qpm; /* qp for next MB */ - float qpa; /* average qp for last frame */ - int qps; - float qp_avg_p; /* average QP for P frames */ - float qp_last_p; - int fbits; /* bits allocated for current frame */ - int ufbits; /* bits used for current frame */ - int nzcoeffs; /* # of 0-quantized coefficients */ - int ncoeffs; /* total # of coefficients */ - int overhead; + float qpa; /* average of macroblocks' qp (same as qp if no adaptive quant) */ + int slice_type; + int qp_force; + + /* VBV stuff */ + double buffer_size; + double buffer_fill; + double buffer_rate; /* # of bits added to buffer_fill after each frame */ + predictor_t pred[5]; /* predict frame size from satd */ + + /* ABR stuff */ + int last_satd; + double last_rceq; + double cplxr_sum; /* sum of bits*qscale/rceq */ + double expected_bits_sum; /* sum of qscale2bits after rceq, ratefactor, and overflow */ + double wanted_bits_window; /* target bitrate * window */ + double cbr_decay; + double short_term_cplxsum; + double short_term_cplxcount; + double rate_factor_constant; /* 2pass stuff */ FILE *p_stat_file_out; + char *psz_stat_file_tmpname; int num_entries; /* number of ratecontrol_entry_ts */ ratecontrol_entry_t *entry; /* FIXME: copy needed data and free this once init is done */ double last_qscale; double last_qscale_for[5]; /* last qscale for a specific pict type, used for max_diff & ipb factor stuff */ int last_non_b_pict_type; + double accum_p_qp; /* for determining I-frame quant */ + double accum_p_norm; + double last_accum_p_norm; double lmin[5]; /* min qscale by frame type */ double lmax[5]; double lstep; /* max change (multiply) in qscale per frame */ @@ -111,11 +126,17 @@ struct x264_ratecontrol_t double p_cplx_sum[5]; double mv_bits_sum[5]; int frame_count[5]; /* number of frames of each type */ + + int i_zones; + x264_zone_t *zones; }; +static int parse_zones( x264_t *h ); static int init_pass2(x264_t *); static float rate_estimate_qscale( x264_t *h, int pict_type ); +static void update_vbv( x264_t *h, int bits ); +int x264_rc_analyse_slice( x264_t *h ); /* Terminology: * qp = h.264's quantizer @@ -130,147 +151,158 @@ static inline double qscale2qp(double qscale) return 12.0 + 6.0 * log(qscale/0.85) / log(2.0); } +/* Texture bitrate is not quite inversely proportional to qscale, + * probably due the the changing number of SKIP blocks. + * MV bits level off at about qp<=12, because the lambda used + * for motion estimation is constant there. */ static inline double qscale2bits(ratecontrol_entry_t *rce, double qscale) { if(qscale<0.1) - { - fprintf(stderr, "qscale<0.1\n"); qscale = 0.1; - } - return (double)(rce->i_tex_bits + rce->p_tex_bits + .1) * rce->qscale / qscale; -} - -static inline double bits2qscale(ratecontrol_entry_t *rce, double bits) -{ - if(bits<0.9) - { -// fprintf(stderr, "bits<0.9\n"); - bits = 1.0; - } - return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + .1) / bits; + return (rce->i_tex_bits + rce->p_tex_bits + .1) * pow( rce->qscale / qscale, 1.1 ) + + rce->mv_bits * pow( X264_MAX(rce->qscale, 12) / X264_MAX(qscale, 12), 0.5 ); } int x264_ratecontrol_new( x264_t *h ) { x264_ratecontrol_t *rc; - float bpp; int i; - /* Needed(?) for 2 pass */ x264_cpu_restore( h->param.cpu ); h->rc = rc = x264_malloc( sizeof( x264_ratecontrol_t ) ); memset(rc, 0, sizeof(*rc)); + rc->b_abr = ( h->param.rc.b_cbr || h->param.rc.i_rf_constant ) && !h->param.rc.b_stat_read; + rc->b_2pass = h->param.rc.b_cbr && h->param.rc.b_stat_read; + h->mb.b_variable_qp = 0; + /* FIXME: use integers */ if(h->param.i_fps_num > 0 && h->param.i_fps_den > 0) rc->fps = (float) h->param.i_fps_num / h->param.i_fps_den; else rc->fps = 25.0; - rc->gop_size = h->param.i_iframe; rc->bitrate = h->param.rc.i_bitrate * 1000; + rc->rate_tolerance = h->param.rc.f_rate_tolerance; rc->nmb = h->mb.i_mb_count; - - rc->qp = h->param.rc.i_qp_constant; - rc->qpa = rc->qp; - rc->qpm = rc->qp; - - rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant; - rc->qp_constant[SLICE_TYPE_I] = x264_clip3( (int)( qscale2qp( qp2qscale( h->param.rc.i_qp_constant ) / fabs( h->param.rc.f_ip_factor )) + 0.5 ), 0, 51 ); - rc->qp_constant[SLICE_TYPE_B] = x264_clip3( (int)( qscale2qp( qp2qscale( h->param.rc.i_qp_constant ) * fabs( h->param.rc.f_pb_factor )) + 0.5 ), 0, 51 ); - - /* Init 1pass CBR algo */ - if( h->param.rc.b_cbr ){ - rc->buffer_size = h->param.rc.i_rc_buffer_size * 1000; - rc->buffer_fullness = h->param.rc.i_rc_init_buffer; - rc->rcbufrate = rc->bitrate / rc->fps; - - if(rc->buffer_size < rc->rcbufrate){ - x264_log(h, X264_LOG_WARNING, "rc buffer size %i too small\n", - rc->buffer_size); - rc->buffer_size = 0; - } - - if(rc->buffer_size <= 0) - rc->buffer_size = rc->bitrate / 2; - - if(rc->buffer_fullness > rc->buffer_size || rc->buffer_fullness < 0){ - x264_log(h, X264_LOG_WARNING, "invalid initial buffer fullness %i\n", - rc->buffer_fullness); - rc->buffer_fullness = 0; + rc->last_non_b_pict_type = -1; + rc->cbr_decay = 1.0; + + if( rc->b_2pass && h->param.rc.i_rf_constant ) + x264_log(h, X264_LOG_ERROR, "constant rate-factor is incompatible with 2pass.\n"); + if( h->param.rc.i_vbv_max_bitrate && !h->param.rc.b_cbr && !h->param.rc.i_rf_constant ) + x264_log(h, X264_LOG_ERROR, "VBV is incompatible with constant QP.\n"); + if( h->param.rc.i_vbv_max_bitrate < h->param.rc.i_bitrate && + h->param.rc.i_vbv_max_bitrate > 0) + x264_log(h, X264_LOG_ERROR, "max bitrate less than average bitrate, ignored.\n"); + else if( h->param.rc.i_vbv_max_bitrate > 0 && + h->param.rc.i_vbv_buffer_size > 0 ) + { + if( h->param.rc.i_vbv_buffer_size < 10 * h->param.rc.i_vbv_max_bitrate / rc->fps ) { + h->param.rc.i_vbv_buffer_size = 10 * h->param.rc.i_vbv_max_bitrate / rc->fps; + x264_log( h, X264_LOG_ERROR, "VBV buffer size too small, using %d kbit\n", + h->param.rc.i_vbv_buffer_size ); } + rc->buffer_rate = h->param.rc.i_vbv_max_bitrate * 1000 / rc->fps; + rc->buffer_size = h->param.rc.i_vbv_buffer_size * 1000; + rc->buffer_fill = rc->buffer_size * h->param.rc.f_vbv_buffer_init; + rc->cbr_decay = 1.0 - rc->buffer_rate / rc->buffer_size + * 0.5 * X264_MAX(0, 1.5 - rc->buffer_rate * rc->fps / rc->bitrate); + } + else if( h->param.rc.i_vbv_max_bitrate || h->param.rc.i_vbv_buffer_size ) + x264_log(h, X264_LOG_ERROR, "VBV maxrate or buffer size specified, but not both.\n"); + if(rc->rate_tolerance < 0.01) { + x264_log(h, X264_LOG_ERROR, "bitrate tolerance too small, using .01\n"); + rc->rate_tolerance = 0.01; + } - bpp = rc->bitrate / (rc->fps * h->param.i_width * h->param.i_height); - if(bpp <= 0.6) - rc->init_qp = 31; - else if(bpp <= 1.4) - rc->init_qp = 25; - else if(bpp <= 2.4) - rc->init_qp = 20; - else - rc->init_qp = 10; - rc->gop_qp = rc->init_qp; - - rc->bits_last_gop = 0; + if( rc->b_abr ) + { + /* FIXME shouldn't need to arbitrarily specify a QP, + * but this is more robust than BPP measures */ +#define ABR_INIT_QP ( h->param.rc.i_rf_constant > 0 ? h->param.rc.i_rf_constant : 24 ) + rc->accum_p_norm = .01; + rc->accum_p_qp = ABR_INIT_QP * rc->accum_p_norm; + rc->cplxr_sum = .01; + rc->wanted_bits_window = .01; + } - x264_log(h, X264_LOG_DEBUG, "%f fps, %i bps, bufsize %i\n", - rc->fps, rc->bitrate, rc->buffer_size); + if( h->param.rc.i_rf_constant ) + { + /* arbitrary rescaling to make CRF somewhat similar to QP */ + double base_cplx = h->mb.i_mb_count * (h->param.i_bframe ? 120 : 80); + rc->rate_factor_constant = pow( base_cplx, 1 - h->param.rc.f_qcompress ) + / qp2qscale( h->param.rc.i_rf_constant ); } + rc->qp_constant[SLICE_TYPE_P] = h->param.rc.i_qp_constant; + rc->qp_constant[SLICE_TYPE_I] = x264_clip3( (int)( qscale2qp( qp2qscale( h->param.rc.i_qp_constant ) / fabs( h->param.rc.f_ip_factor )) + 0.5 ), 0, 51 ); + rc->qp_constant[SLICE_TYPE_B] = x264_clip3( (int)( qscale2qp( qp2qscale( h->param.rc.i_qp_constant ) * fabs( h->param.rc.f_pb_factor )) + 0.5 ), 0, 51 ); rc->lstep = exp2f(h->param.rc.i_qp_step / 6.0); + rc->last_qscale = qp2qscale(26); for( i = 0; i < 5; i++ ) { rc->last_qscale_for[i] = qp2qscale(26); rc->lmin[i] = qp2qscale( h->param.rc.i_qp_min ); rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max ); + rc->pred[i].coeff= 2.0; + rc->pred[i].count= 1.0; + rc->pred[i].decay= 0.5; } -#if 0 // FIXME: do we want to assign lmin/lmax based on ip_factor, or leave them all the same? - rc->lmin[SLICE_TYPE_I] /= fabs(h->param.f_ip_factor); - rc->lmax[SLICE_TYPE_I] /= fabs(h->param.f_ip_factor); - rc->lmin[SLICE_TYPE_B] *= fabs(h->param.f_pb_factor); - rc->lmax[SLICE_TYPE_B] *= fabs(h->param.f_pb_factor); -#endif + + if( parse_zones( h ) < 0 ) + return -1; /* Load stat file and init 2pass algo */ if( h->param.rc.b_stat_read ) { - int stats_size; char *p, *stats_in; - FILE *stats_file; /* read 1st pass stats */ assert( h->param.rc.psz_stat_in ); - stats_file = fopen( h->param.rc.psz_stat_in, "rb"); - if(!stats_file) + stats_in = x264_slurp_file( h->param.rc.psz_stat_in ); + if( !stats_in ) { x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n"); return -1; } - // FIXME: error checking - fseek(stats_file, 0, SEEK_END); - stats_size = ftell(stats_file); - fseek(stats_file, 0, SEEK_SET); - stats_in = x264_malloc(stats_size+10); - fread(stats_in, 1, stats_size, stats_file); - fclose(stats_file); /* find number of pics */ p = stats_in; - for(i=-1; p; i++){ + for(i=-1; p; i++) p = strchr(p+1, ';'); + if(i==0) + { + x264_log(h, X264_LOG_ERROR, "empty stats file\n"); + return -1; + } + rc->num_entries = i; + + if( h->param.i_frame_total < rc->num_entries && h->param.i_frame_total > 0 ) + { + x264_log( h, X264_LOG_WARNING, "2nd pass has fewer frames than 1st pass (%d vs %d)\n", + h->param.i_frame_total, rc->num_entries ); + } + if( h->param.i_frame_total > rc->num_entries + h->param.i_bframe ) + { + x264_log( h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d vs %d)\n", + h->param.i_frame_total, rc->num_entries ); + return -1; } - i += h->param.i_bframe; - rc->entry = (ratecontrol_entry_t*) x264_malloc(i*sizeof(ratecontrol_entry_t)); - memset(rc->entry, 0, i*sizeof(ratecontrol_entry_t)); - rc->num_entries= i; + + /* FIXME: ugly padding because VfW drops delayed B-frames */ + rc->num_entries += h->param.i_bframe; + + rc->entry = (ratecontrol_entry_t*) x264_malloc(rc->num_entries * sizeof(ratecontrol_entry_t)); + memset(rc->entry, 0, rc->num_entries * sizeof(ratecontrol_entry_t)); /* init all to skipped p frames */ for(i=0; inum_entries; i++){ ratecontrol_entry_t *rce = &rc->entry[i]; - rce->pict_type = rce->new_pict_type = SLICE_TYPE_P; + rce->pict_type = SLICE_TYPE_P; rce->qscale = rce->new_qscale = qp2qscale(20); rce->misc_bits = rc->nmb + 10; rce->new_qp = 0; @@ -280,7 +312,8 @@ int x264_ratecontrol_new( x264_t *h ) p = stats_in; for(i=0; i < rc->num_entries - h->param.i_bframe; i++){ ratecontrol_entry_t *rce; - int picture_number; + int frame_number; + char pict_type; int e; char *next; float qp; @@ -290,15 +323,27 @@ int x264_ratecontrol_new( x264_t *h ) (*next)=0; //sscanf is unbelievably slow on looong strings next++; } - e = sscanf(p, " in:%d ", &picture_number); + e = sscanf(p, " in:%d ", &frame_number); - assert(picture_number >= 0); - assert(picture_number < rc->num_entries); - rce = &rc->entry[picture_number]; + if(frame_number < 0 || frame_number >= rc->num_entries) + { + x264_log(h, X264_LOG_ERROR, "bad frame number (%d) at stats line %d\n", frame_number, i); + return -1; + } + rce = &rc->entry[frame_number]; - e += sscanf(p, " in:%*d out:%*d type:%d q:%f itex:%d ptex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d", - &rce->pict_type, &qp, &rce->i_tex_bits, &rce->p_tex_bits, + e += sscanf(p, " in:%*d out:%*d type:%c q:%f itex:%d ptex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d", + &pict_type, &qp, &rce->i_tex_bits, &rce->p_tex_bits, &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count, &rce->s_count); + + switch(pict_type){ + case 'I': rce->kept_as_ref = 1; + case 'i': rce->pict_type = SLICE_TYPE_I; break; + case 'P': rce->pict_type = SLICE_TYPE_P; break; + case 'B': rce->kept_as_ref = 1; + case 'b': rce->pict_type = SLICE_TYPE_B; break; + default: e = -1; break; + } if(e != 10){ x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e); return -1; @@ -309,13 +354,22 @@ int x264_ratecontrol_new( x264_t *h ) x264_free(stats_in); - if(init_pass2(h) < 0) return -1; + if(h->param.rc.b_cbr) + { + if(init_pass2(h) < 0) return -1; + } /* else we're using constant quant, so no need to run the bitrate allocation */ } /* Open output file */ + /* If input and output files are the same, output to a temp file + * and move it to the real name only when it's complete */ if( h->param.rc.b_stat_write ) { - rc->p_stat_file_out = fopen( h->param.rc.psz_stat_out, "wb" ); + rc->psz_stat_file_tmpname = x264_malloc( strlen(h->param.rc.psz_stat_out) + 6 ); + strcpy( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ); + strcat( rc->psz_stat_file_tmpname, ".temp" ); + + rc->p_stat_file_out = fopen( rc->psz_stat_file_tmpname, "wb" ); if( rc->p_stat_file_out == NULL ) { x264_log(h, X264_LOG_ERROR, "ratecontrol_init: can't open stats file\n"); @@ -326,223 +380,173 @@ int x264_ratecontrol_new( x264_t *h ) return 0; } +static int parse_zones( x264_t *h ) +{ + x264_ratecontrol_t *rc = h->rc; + int i; + if( h->param.rc.psz_zones && !h->param.rc.i_zones ) + { + char *p; + h->param.rc.i_zones = 1; + for( p = h->param.rc.psz_zones; *p; p++ ) + h->param.rc.i_zones += (*p == '/'); + h->param.rc.zones = x264_malloc( h->param.rc.i_zones * sizeof(x264_zone_t) ); + p = h->param.rc.psz_zones; + for( i = 0; i < h->param.rc.i_zones; i++) + { + x264_zone_t *z = &h->param.rc.zones[i]; + if( 3 == sscanf(p, "%u,%u,q=%u", &z->i_start, &z->i_end, &z->i_qp) ) + z->b_force_qp = 1; + else if( 3 == sscanf(p, "%u,%u,b=%f", &z->i_start, &z->i_end, &z->f_bitrate_factor) ) + z->b_force_qp = 0; + else + { + char *slash = strchr(p, '/'); + if(slash) *slash = '\0'; + x264_log( h, X264_LOG_ERROR, "invalid zone: \"%s\"\n", p ); + return -1; + } + p = strchr(p, '/') + 1; + } + } + + if( h->param.rc.i_zones > 0 ) + { + for( i = 0; i < h->param.rc.i_zones; i++ ) + { + x264_zone_t z = h->param.rc.zones[i]; + if( z.i_start < 0 || z.i_start > z.i_end ) + { + x264_log( h, X264_LOG_ERROR, "invalid zone: start=%d end=%d\n", + z.i_start, z.i_end ); + return -1; + } + else if( !z.b_force_qp && z.f_bitrate_factor <= 0 ) + { + x264_log( h, X264_LOG_ERROR, "invalid zone: bitrate_factor=%f\n", + z.f_bitrate_factor ); + return -1; + } + } + + rc->i_zones = h->param.rc.i_zones; + rc->zones = x264_malloc( rc->i_zones * sizeof(x264_zone_t) ); + memcpy( rc->zones, h->param.rc.zones, rc->i_zones * sizeof(x264_zone_t) ); + } + + return 0; +} + void x264_ratecontrol_delete( x264_t *h ) { x264_ratecontrol_t *rc = h->rc; if( rc->p_stat_file_out ) + { fclose( rc->p_stat_file_out ); - if( rc->entry ) - x264_free(rc->entry); + if( h->i_frame >= rc->num_entries - h->param.i_bframe ) + if( rename( rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ) != 0 ) + { + x264_log( h, X264_LOG_ERROR, "failed to rename \"%s\" to \"%s\"\n", + rc->psz_stat_file_tmpname, h->param.rc.psz_stat_out ); + } + x264_free( rc->psz_stat_file_tmpname ); + } + x264_free( rc->entry ); + x264_free( rc->zones ); x264_free( rc ); } -void x264_ratecontrol_start( x264_t *h, int i_slice_type ) +/* Before encoding a frame, choose a QP for it */ +void x264_ratecontrol_start( x264_t *h, int i_slice_type, int i_force_qp ) { x264_ratecontrol_t *rc = h->rc; - int gframes, iframes, pframes, bframes; - int minbits, maxbits; - int gbits, fbits; - int zn = 0; - float kp; - int gbuf; - - rc->slice_type = i_slice_type; x264_cpu_restore( h->param.cpu ); - if( h->param.rc.b_stat_read ) + rc->qp_force = i_force_qp; + rc->slice_type = i_slice_type; + + if( i_force_qp ) + { + rc->qpa = rc->qp = i_force_qp - 1; + } + else if( rc->b_abr ) + { + rc->qpa = rc->qp = + x264_clip3( (int)(qscale2qp( rate_estimate_qscale( h, i_slice_type ) ) + .5), 0, 51 ); + } + else if( rc->b_2pass ) { int frame = h->fenc->i_frame; - ratecontrol_entry_t *rce = &h->rc->entry[frame]; - + ratecontrol_entry_t *rce; assert( frame >= 0 && frame < rc->num_entries ); + rce = h->rc->rce = &h->rc->entry[frame]; rce->new_qscale = rate_estimate_qscale( h, i_slice_type ); - rc->qpm = rc->qpa = rc->qp = rce->new_qp = - (int)(qscale2qp(rce->new_qscale) + 0.5); - return; + rc->qpa = rc->qp = rce->new_qp = + x264_clip3( (int)(qscale2qp(rce->new_qscale) + 0.5), 0, 51 ); } - else if( !h->param.rc.b_cbr ) + else /* CQP */ { - rc->qpm = rc->qpa = rc->qp = - rc->qp_constant[ i_slice_type ]; - return; - } - - switch(i_slice_type){ - case SLICE_TYPE_I: - gbuf = rc->buffer_fullness + (rc->gop_size-1) * rc->rcbufrate; - rc->bits_gop = gbuf - rc->buffer_size / 2; - - if(!rc->mb && rc->pframes){ - int qp = rc->qp_avg_p / rc->pframes + 0.5; -#if 0 /* JM does this without explaining why */ - int gdq = (float) rc->gop_size / 15 + 0.5; - if(gdq > 2) - gdq = 2; - qp -= gdq; - if(qp > rc->qp_last_p - 2) - qp--; -#endif - qp = x264_clip3(qp, rc->gop_qp - 4, rc->gop_qp + 4); - qp = x264_clip3(qp, h->param.rc.i_qp_min, h->param.rc.i_qp_max); - rc->gop_qp = qp; - } else if(rc->frames > 4){ - rc->gop_qp = rc->init_qp; - } - - kp = h->param.rc.f_ip_factor * h->param.rc.f_pb_factor; - - x264_log(h, X264_LOG_DEBUG,"gbuf=%i bits_gop=%i frames=%i gop_qp=%i\n", - gbuf, rc->bits_gop, rc->frames, rc->gop_qp); - - rc->bits_last_gop = 0; - rc->frames = 0; - rc->pframes = 0; - rc->qp_avg_p = 0; - break; - - case SLICE_TYPE_P: - kp = h->param.rc.f_pb_factor; - break; - - case SLICE_TYPE_B: - kp = 1.0; - break; - - default: - x264_log(h, X264_LOG_WARNING, "ratecontrol: unknown slice type %i\n", - i_slice_type); - kp = 1.0; - break; - } - - gframes = rc->gop_size - rc->frames; - iframes = gframes / rc->gop_size; - pframes = gframes / (h->param.i_bframe + 1) - iframes; - bframes = gframes - pframes - iframes; - - gbits = rc->bits_gop - rc->bits_last_gop; - fbits = kp * gbits / - (h->param.rc.f_ip_factor * h->param.rc.f_pb_factor * iframes + - h->param.rc.f_pb_factor * pframes + bframes); - - minbits = rc->buffer_fullness + rc->rcbufrate - rc->buffer_size; - if(minbits < 0) - minbits = 0; - maxbits = rc->buffer_fullness; - rc->fbits = x264_clip3(fbits, minbits, maxbits); - - if(i_slice_type == SLICE_TYPE_I){ - rc->qp = rc->gop_qp; - } else if(rc->ncoeffs && rc->ufbits){ - int dqp, nonzc; - - nonzc = (rc->ncoeffs - rc->nzcoeffs); - if(nonzc == 0) - zn = rc->ncoeffs; - else if(rc->fbits < INT_MAX / nonzc) - zn = rc->ncoeffs - rc->fbits * nonzc / rc->ufbits; + int q; + if( i_slice_type == SLICE_TYPE_B && h->fdec->b_kept_as_ref ) + q = ( rc->qp_constant[ SLICE_TYPE_B ] + rc->qp_constant[ SLICE_TYPE_P ] ) / 2; else - zn = 0; - zn = x264_clip3(zn, 0, rc->ncoeffs); - dqp = h->param.rc.i_rc_sens * exp2f(rc->qpa / 6) * - (zn - rc->nzcoeffs) / rc->nzcoeffs; - dqp = x264_clip3(dqp, -h->param.rc.i_qp_step, h->param.rc.i_qp_step); - rc->qp = (int)(rc->qpa + dqp + .5); - } - - if(rc->fbits > 0.9 * maxbits) - rc->qp += 2; - else if(rc->fbits > 0.8 * maxbits) - rc->qp += 1; - else if(rc->fbits < 1.1 * minbits) - rc->qp -= 2; - else if(rc->fbits < 1.2 * minbits) - rc->qp -= 1; - - rc->qp = x264_clip3(rc->qp, h->param.rc.i_qp_min, h->param.rc.i_qp_max); - rc->qpm = rc->qp; - - x264_log(h, X264_LOG_DEBUG, "fbits=%i, qp=%i, z=%i, min=%i, max=%i\n", - rc->fbits, rc->qpm, zn, minbits, maxbits); - - rc->fbits -= rc->overhead; - rc->ufbits = 0; - rc->ncoeffs = 0; - rc->nzcoeffs = 0; - rc->mb = 0; - rc->qps = 0; + q = rc->qp_constant[ i_slice_type ]; + rc->qpa = rc->qp = q; + } } void x264_ratecontrol_mb( x264_t *h, int bits ) { - x264_ratecontrol_t *rc = h->rc; - int rbits; - int zn, enz, nonz; - int rcoeffs; - int dqp; - int i; - - if( !h->param.rc.b_cbr || h->param.rc.b_stat_read ) - return; - - x264_cpu_restore( h->param.cpu ); - - rc->qps += rc->qpm; - rc->ufbits += bits; - rc->mb++; - - for(i = 0; i < 16 + 8; i++) - rc->nzcoeffs += 16 - h->mb.cache.non_zero_count[x264_scan8[i]]; - rc->ncoeffs += 16 * (16 + 8); - - if(rc->mb < rc->nmb / 16) - return; - else if(rc->mb == rc->nmb) - return; - - rcoeffs = (rc->nmb - rc->mb) * 16 * 24; - rbits = rc->fbits - rc->ufbits; -/* if(rbits < 0) */ -/* rbits = 0; */ - -/* zn = (rc->nmb - rc->mb) * 16 * 24; */ - nonz = (rc->ncoeffs - rc->nzcoeffs); - if(nonz == 0) - zn = rcoeffs; - else if(rc->ufbits && rbits < INT_MAX / nonz) - zn = rcoeffs - rbits * nonz / rc->ufbits; - else - zn = 0; - zn = x264_clip3(zn, 0, rcoeffs); - enz = rc->nzcoeffs * (rc->nmb - rc->mb) / rc->mb; - dqp = (float) 2*h->param.rc.i_rc_sens * exp2f((float) rc->qps / rc->mb / 6) * - (zn - enz) / enz; - rc->qpm = x264_clip3(rc->qpm + dqp, rc->qp - 3, rc->qp + 3); - if(rbits <= 0) - rc->qpm++; - rc->qpm = x264_clip3(rc->qpm, h->param.rc.i_qp_min, h->param.rc.i_qp_max); + /* currently no adaptive quant */ } -int x264_ratecontrol_qp( x264_t *h ) +int x264_ratecontrol_qp( x264_t *h ) { - return h->rc->qpm; + return h->rc->qp; } +/* In 2pass, force the same frame types as in the 1st pass */ int x264_ratecontrol_slice_type( x264_t *h, int frame_num ) { + x264_ratecontrol_t *rc = h->rc; if( h->param.rc.b_stat_read ) { - assert(frame_num < h->rc->num_entries); - switch( h->rc->entry[frame_num].new_pict_type ) + if( frame_num >= rc->num_entries ) + { + /* We could try to initialize everything required for ABR and + * adaptive B-frames, but that would be complicated. + * So just calculate the average QP used so far. */ + + h->param.rc.i_qp_constant = (h->stat.i_slice_count[SLICE_TYPE_P] == 0) ? 24 + : 1 + h->stat.i_slice_qp[SLICE_TYPE_P] / h->stat.i_slice_count[SLICE_TYPE_P]; + rc->qp_constant[SLICE_TYPE_P] = x264_clip3( h->param.rc.i_qp_constant, 0, 51 ); + rc->qp_constant[SLICE_TYPE_I] = x264_clip3( (int)( qscale2qp( qp2qscale( h->param.rc.i_qp_constant ) / fabs( h->param.rc.f_ip_factor )) + 0.5 ), 0, 51 ); + rc->qp_constant[SLICE_TYPE_B] = x264_clip3( (int)( qscale2qp( qp2qscale( h->param.rc.i_qp_constant ) * fabs( h->param.rc.f_pb_factor )) + 0.5 ), 0, 51 ); + + x264_log(h, X264_LOG_ERROR, "2nd pass has more frames than 1st pass (%d)\n", rc->num_entries); + x264_log(h, X264_LOG_ERROR, "continuing anyway, at constant QP=%d\n", h->param.rc.i_qp_constant); + if( h->param.b_bframe_adaptive ) + x264_log(h, X264_LOG_ERROR, "disabling adaptive B-frames\n"); + + rc->b_abr = 0; + rc->b_2pass = 0; + h->param.rc.b_cbr = 0; + h->param.rc.b_stat_read = 0; + h->param.b_bframe_adaptive = 0; + if( h->param.i_bframe > 1 ) + h->param.i_bframe = 1; + return X264_TYPE_P; + } + switch( rc->entry[frame_num].pict_type ) { case SLICE_TYPE_I: - return X264_TYPE_I; + return rc->entry[frame_num].kept_as_ref ? X264_TYPE_IDR : X264_TYPE_I; case SLICE_TYPE_B: - return X264_TYPE_B; + return rc->entry[frame_num].kept_as_ref ? X264_TYPE_BREF : X264_TYPE_B; case SLICE_TYPE_P: default: @@ -555,87 +559,95 @@ int x264_ratecontrol_slice_type( x264_t *h, int frame_num ) } } +/* After encoding one frame, save stats and update ratecontrol state */ void x264_ratecontrol_end( x264_t *h, int bits ) { x264_ratecontrol_t *rc = h->rc; + const int *mbs = h->stat.frame.i_mb_count; int i; x264_cpu_restore( h->param.cpu ); - h->stat.frame.i_mb_count_skip = h->stat.frame.i_mb_count[P_SKIP] + h->stat.frame.i_mb_count[B_SKIP]; - h->stat.frame.i_mb_count_p = h->stat.frame.i_mb_count[P_L0] + h->stat.frame.i_mb_count[P_8x8]; + h->stat.frame.i_mb_count_skip = mbs[P_SKIP] + mbs[B_SKIP]; + h->stat.frame.i_mb_count_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4]; + h->stat.frame.i_mb_count_p = mbs[P_L0] + mbs[P_8x8]; for( i = B_DIRECT; i < B_8x8; i++ ) - h->stat.frame.i_mb_count_p += h->stat.frame.i_mb_count[i]; + h->stat.frame.i_mb_count_p += mbs[i]; if( h->param.rc.b_stat_write ) { + char c_type = rc->slice_type==SLICE_TYPE_I ? (h->fenc->i_poc==0 ? 'I' : 'i') + : rc->slice_type==SLICE_TYPE_P ? 'P' + : h->fenc->b_kept_as_ref ? 'B' : 'b'; fprintf( rc->p_stat_file_out, - "in:%d out:%d type:%d q:%.3f itex:%d ptex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d;\n", + "in:%d out:%d type:%c q:%.2f itex:%d ptex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d;\n", h->fenc->i_frame, h->i_frame-1, - rc->slice_type, rc->qpa, + c_type, rc->qpa, h->stat.frame.i_itex_bits, h->stat.frame.i_ptex_bits, h->stat.frame.i_hdr_bits, h->stat.frame.i_misc_bits, - h->stat.frame.i_mb_count[I_4x4] + h->stat.frame.i_mb_count[I_16x16], + h->stat.frame.i_mb_count_i, h->stat.frame.i_mb_count_p, h->stat.frame.i_mb_count_skip); } - if( !h->param.rc.b_cbr || h->param.rc.b_stat_read ) - return; - - rc->buffer_fullness += rc->rcbufrate - bits; - if(rc->buffer_fullness < 0){ - x264_log(h, X264_LOG_WARNING, "buffer underflow %i\n", - rc->buffer_fullness); - rc->buffer_fullness = 0; + if( rc->b_abr ) + { + if( rc->slice_type != SLICE_TYPE_B ) + rc->cplxr_sum += bits * qp2qscale(rc->qpa) / rc->last_rceq; + else + { + /* Depends on the fact that B-frame's QP is an offset from the following P-frame's. + * Not perfectly accurate with B-refs, but good enough. */ + rc->cplxr_sum += bits * qp2qscale(rc->qpa) / (rc->last_rceq * fabs(h->param.rc.f_pb_factor)); + } + rc->cplxr_sum *= rc->cbr_decay; + rc->wanted_bits_window += rc->bitrate / rc->fps; + rc->wanted_bits_window *= rc->cbr_decay; + + rc->accum_p_qp *= .95; + rc->accum_p_norm *= .95; + rc->accum_p_norm += 1; + if( rc->slice_type == SLICE_TYPE_I ) + rc->accum_p_qp += rc->qpa * fabs(h->param.rc.f_ip_factor); + else + rc->accum_p_qp += rc->qpa; } - rc->qpa = (float)rc->qps / rc->mb; - if(rc->slice_type == SLICE_TYPE_P){ - rc->qp_avg_p += rc->qpa; - rc->qp_last_p = rc->qpa; - rc->pframes++; - } else if(rc->slice_type == SLICE_TYPE_I){ - float err = (float) rc->ufbits / rc->fbits; - if(err > 1.1) - rc->gop_qp++; - else if(err < 0.9) - rc->gop_qp--; + if( rc->b_2pass ) + { + rc->expected_bits_sum += qscale2bits( rc->rce, qp2qscale(rc->rce->new_qp) ); } - rc->overhead = bits - rc->ufbits; - - x264_log(h, X264_LOG_DEBUG, "bits=%i, qp=%.1f, z=%i, zr=%6.3f, buf=%i\n", - bits, rc->qpa, rc->nzcoeffs, (float) rc->nzcoeffs / rc->ncoeffs, - rc->buffer_fullness); + update_vbv( h, bits ); - rc->bits_last_gop += bits; - rc->frames++; - rc->mb = 0; + if( rc->slice_type != SLICE_TYPE_B ) + rc->last_non_b_pict_type = rc->slice_type; } /**************************************************************************** * 2 pass functions ***************************************************************************/ + double x264_eval( char *s, double *const_value, const char **const_name, double (**func1)(void *, double), const char **func1_name, double (**func2)(void *, double, double), char **func2_name, void *opaque ); /** - * modifies the bitrate curve from pass1 for one frame + * modify the bitrate curve from pass1 for one frame */ -static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor) +static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor, int frame_num) { x264_ratecontrol_t *rcc= h->rc; - double bits; - const int pict_type = rce->new_pict_type; + const int pict_type = rce->pict_type; + double q; + int i; double const_values[]={ rce->i_tex_bits * rce->qscale, rce->p_tex_bits * rce->qscale, (rce->i_tex_bits + rce->p_tex_bits) * rce->qscale, - rce->mv_bits / rcc->nmb, + rce->mv_bits * rce->qscale, (double)rce->i_count / rcc->nmb, (double)rce->p_count / rcc->nmb, (double)rce->s_count / rcc->nmb, @@ -668,57 +680,87 @@ static double get_qscale(x264_t *h, ratecontrol_entry_t *rce, double rate_factor "avgPPTex", "avgBPTex", "avgTex", - "blurTex", + "blurCplx", NULL }; static double (*func1[])(void *, double)={ - (void *)bits2qscale, +// (void *)bits2qscale, (void *)qscale2bits, NULL }; static const char *func1_names[]={ - "bits2qp", +// "bits2qp", "qp2bits", NULL }; - bits = x264_eval((char*)h->param.rc.psz_rc_eq, const_values, const_names, func1, func1_names, NULL, NULL, rce); + q = x264_eval((char*)h->param.rc.psz_rc_eq, const_values, const_names, func1, func1_names, NULL, NULL, rce); // avoid NaN's in the rc_eq - if(bits != bits || rce->i_tex_bits + rce->p_tex_bits == 0) - bits = 0; - - bits *= rate_factor; - if(bits<0.0) bits=0.0; - bits += 1.0; //avoid 1/0 issues + if(q != q || rce->i_tex_bits + rce->p_tex_bits + rce->mv_bits == 0) + q = rcc->last_qscale; + else { + rcc->last_rceq = q; + q /= rate_factor; + rcc->last_qscale = q; + } - /* I/B difference */ - if( pict_type==SLICE_TYPE_I && h->param.rc.f_ip_factor > 0 ) - bits *= h->param.rc.f_ip_factor; - else if( pict_type==SLICE_TYPE_B && h->param.rc.f_pb_factor > 0 ) - bits /= h->param.rc.f_pb_factor; + for( i = rcc->i_zones-1; i >= 0; i-- ) + { + x264_zone_t *z = &rcc->zones[i]; + if( frame_num >= z->i_start && frame_num <= z->i_end ) + { + if( z->b_force_qp ) + q = qp2qscale(z->i_qp); + else + q /= z->f_bitrate_factor; + break; + } + } - return bits2qscale(rce, bits); + return q; } static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q) { x264_ratecontrol_t *rcc = h->rc; - const int pict_type = rce->new_pict_type; + const int pict_type = rce->pict_type; // force I/B quants as a function of P quants const double last_p_q = rcc->last_qscale_for[SLICE_TYPE_P]; const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type]; - if( pict_type == SLICE_TYPE_I && h->param.rc.f_ip_factor < 0 ) - q = last_p_q / fabs( h->param.rc.f_ip_factor ); - else if( pict_type == SLICE_TYPE_B - && ( h->param.rc.f_pb_factor < 0 || rce->i_tex_bits + rce->p_tex_bits == 0 ) ) - q = last_non_b_q * fabs( h->param.rc.f_pb_factor ); - else if( pict_type == SLICE_TYPE_P && rce->i_tex_bits + rce->p_tex_bits == 0 ) + if( pict_type == SLICE_TYPE_I ) + { + double iq = q; + double pq = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm ); + double ip_factor = fabs( h->param.rc.f_ip_factor ); + /* don't apply ip_factor if the following frame is also I */ + if( rcc->accum_p_norm <= 0 ) + q = iq; + else if( h->param.rc.f_ip_factor < 0 ) + q = iq / ip_factor; + else if( rcc->accum_p_norm >= 1 ) + q = pq / ip_factor; + else + q = rcc->accum_p_norm * pq / ip_factor + (1 - rcc->accum_p_norm) * iq; + } + else if( pict_type == SLICE_TYPE_B ) + { + if( h->param.rc.f_pb_factor > 0 ) + q = last_non_b_q; + if( !rce->kept_as_ref ) + q *= fabs( h->param.rc.f_pb_factor ); + } + else if( pict_type == SLICE_TYPE_P + && rcc->last_non_b_pict_type == SLICE_TYPE_P + && rce->i_tex_bits + rce->p_tex_bits == 0 ) + { q = last_p_q; + } /* last qscale / qdiff stuff */ - if(rcc->last_non_b_pict_type==pict_type || pict_type!=SLICE_TYPE_I) + if(rcc->last_non_b_pict_type==pict_type + && (pict_type!=SLICE_TYPE_I || rcc->last_accum_p_norm < 1)) { double last_q = rcc->last_qscale_for[pict_type]; double max_qscale = last_q * rcc->lstep; @@ -731,18 +773,87 @@ static double get_diff_limited_q(x264_t *h, ratecontrol_entry_t *rce, double q) rcc->last_qscale_for[pict_type] = q; if(pict_type!=SLICE_TYPE_B) rcc->last_non_b_pict_type = pict_type; + if(pict_type==SLICE_TYPE_I) + { + rcc->last_accum_p_norm = rcc->accum_p_norm; + rcc->accum_p_norm = 0; + rcc->accum_p_qp = 0; + } + if(pict_type==SLICE_TYPE_P) + { + float mask = 1 - pow( (float)rce->i_count / rcc->nmb, 2 ); + rcc->accum_p_qp = mask * (qscale2qp(q) + rcc->accum_p_qp); + rcc->accum_p_norm = mask * (1 + rcc->accum_p_norm); + } return q; } -// clip a qscale to between lmin and lmax -static double clip_qscale( x264_t *h, ratecontrol_entry_t *rce, double q ) +static double predict_size( predictor_t *p, double q, double var ) +{ + return p->coeff*var / (q*p->count); +} + +static void update_predictor( predictor_t *p, double q, double var, double bits ) { - double lmin = h->rc->lmin[rce->new_pict_type]; - double lmax = h->rc->lmax[rce->new_pict_type]; + p->count *= p->decay; + p->coeff *= p->decay; + p->count ++; + p->coeff += bits*q / var; +} - if(lmin==lmax){ +static void update_vbv( x264_t *h, int bits ) +{ + x264_ratecontrol_t *rcc = h->rc; + if( !rcc->buffer_size ) + return; + + rcc->buffer_fill += rcc->buffer_rate - bits; + if( rcc->buffer_fill < 0 && !rcc->b_2pass ) + x264_log( h, X264_LOG_WARNING, "VBV underflow (%.0f bits)\n", rcc->buffer_fill ); + rcc->buffer_fill = x264_clip3( rcc->buffer_fill, 0, rcc->buffer_size ); + + if(rcc->last_satd > 100) + update_predictor( &rcc->pred[rcc->slice_type], qp2qscale(rcc->qpa), rcc->last_satd, bits ); +} + +// apply VBV constraints and clip qscale to between lmin and lmax +static double clip_qscale( x264_t *h, int pict_type, double q ) +{ + x264_ratecontrol_t *rcc = h->rc; + double lmin = rcc->lmin[pict_type]; + double lmax = rcc->lmax[pict_type]; + double q0 = q; + + /* B-frames are not directly subject to VBV, + * since they are controlled by the P-frames' QPs. + * FIXME: in 2pass we could modify previous frames' QP too, + * instead of waiting for the buffer to fill */ + if( rcc->buffer_size && + ( pict_type == SLICE_TYPE_P || + ( pict_type == SLICE_TYPE_I && rcc->last_non_b_pict_type == SLICE_TYPE_I ) ) ) + { + if( rcc->buffer_fill/rcc->buffer_size < 0.5 ) + q /= x264_clip3f( 2.0*rcc->buffer_fill/rcc->buffer_size, 0.5, 1.0 ); + } + /* Now a hard threshold to make sure the frame fits in VBV. + * This one is mostly for I-frames. */ + if( rcc->buffer_size && rcc->last_satd > 0 ) + { + double bits = predict_size( &rcc->pred[rcc->slice_type], q, rcc->last_satd ); + double qf = 1.0; + if( bits > rcc->buffer_fill/2 ) + qf = x264_clip3f( rcc->buffer_fill/(2*bits), 0.2, 1.0 ); + q /= qf; + bits *= qf; + if( bits < rcc->buffer_rate/2 ) + q *= bits*2/rcc->buffer_rate; + q = X264_MAX( q0, q ); + } + + if(lmin==lmax) return lmin; - }else{ + else if(rcc->b_2pass) + { double min2 = log(lmin); double max2 = log(lmax); q = (log(q) - min2)/(max2-min2) - 0.5; @@ -750,38 +861,142 @@ static double clip_qscale( x264_t *h, ratecontrol_entry_t *rce, double q ) q = q*(max2-min2) + min2; return exp(q); } + else + return x264_clip3f(q, lmin, lmax); } // update qscale for 1 frame based on actual bits used so far static float rate_estimate_qscale(x264_t *h, int pict_type) { float q; - float br_compensation; - double diff; - int picture_number = h->fenc->i_frame; x264_ratecontrol_t *rcc = h->rc; - ratecontrol_entry_t *rce; + ratecontrol_entry_t rce; double lmin = rcc->lmin[pict_type]; double lmax = rcc->lmax[pict_type]; int64_t total_bits = 8*(h->stat.i_slice_size[SLICE_TYPE_I] + h->stat.i_slice_size[SLICE_TYPE_P] + h->stat.i_slice_size[SLICE_TYPE_B]); -//printf("input_pic_num:%d pic_num:%d frame_rate:%d\n", s->input_picture_number, s->picture_number, s->frame_rate); + if( rcc->b_2pass ) + { + rce = *rcc->rce; + if(pict_type != rce.pict_type) + { + x264_log(h, X264_LOG_ERROR, "slice=%c but 2pass stats say %c\n", + slice_type_to_char[pict_type], slice_type_to_char[rce.pict_type]); + } + } + + if( pict_type == SLICE_TYPE_B ) + { + rcc->last_satd = 0; + if(h->fenc->b_kept_as_ref) + q = rcc->last_qscale * sqrtf(h->param.rc.f_pb_factor); + else + q = rcc->last_qscale * h->param.rc.f_pb_factor; + return x264_clip3f(q, lmin, lmax); + } + else + { + double abr_buffer = 2 * rcc->rate_tolerance * rcc->bitrate; + if( rcc->b_2pass ) + { + //FIXME adjust abr_buffer based on distance to the end of the video + int64_t diff = total_bits - (int64_t)rce.expected_bits; + q = rce.new_qscale; + q /= x264_clip3f((double)(abr_buffer - diff) / abr_buffer, .5, 2); + if( h->fenc->i_frame > 30 ) + { + /* Adjust quant based on the difference between + * achieved and expected bitrate so far */ + double time = (double)h->fenc->i_frame / rcc->num_entries; + double w = x264_clip3f( time*100, 0.0, 1.0 ); + q *= pow( (double)total_bits / rcc->expected_bits_sum, w ); + } + q = x264_clip3f( q, lmin, lmax ); + } + else /* 1pass ABR */ + { + /* Calculate the quantizer which would have produced the desired + * average bitrate if it had been applied to all frames so far. + * Then modulate that quant based on the current frame's complexity + * relative to the average complexity so far (using the 2pass RCEQ). + * Then bias the quant up or down if total size so far was far from + * the target. + * Result: Depending on the value of rate_tolerance, there is a + * tradeoff between quality and bitrate precision. But at large + * tolerances, the bit distribution approaches that of 2pass. */ + + double wanted_bits, overflow, lmin, lmax; + + rcc->last_satd = x264_rc_analyse_slice( h ); + rcc->short_term_cplxsum *= 0.5; + rcc->short_term_cplxcount *= 0.5; + rcc->short_term_cplxsum += rcc->last_satd; + rcc->short_term_cplxcount ++; + + rce.p_tex_bits = rcc->last_satd; + rce.blurred_complexity = rcc->short_term_cplxsum / rcc->short_term_cplxcount; + rce.i_tex_bits = 0; + rce.mv_bits = 0; + rce.p_count = rcc->nmb; + rce.i_count = 0; + rce.s_count = 0; + rce.qscale = 1; + rce.pict_type = pict_type; + + if( h->param.rc.i_rf_constant ) + { + q = get_qscale( h, &rce, rcc->rate_factor_constant, h->fenc->i_frame ); + overflow = 1; + } + else + { + q = get_qscale( h, &rce, rcc->wanted_bits_window / rcc->cplxr_sum, h->fenc->i_frame ); + + wanted_bits = h->fenc->i_frame * rcc->bitrate / rcc->fps; + abr_buffer *= X264_MAX( 1, sqrt(h->fenc->i_frame/25) ); + overflow = x264_clip3f( 1.0 + (total_bits - wanted_bits) / abr_buffer, .5, 2 ); + q *= overflow; + } - rce = &rcc->entry[picture_number]; + if( pict_type == SLICE_TYPE_I && h->param.i_keyint_max > 1 + /* should test _next_ pict type, but that isn't decided yet */ + && rcc->last_non_b_pict_type != SLICE_TYPE_I ) + { + q = qp2qscale( rcc->accum_p_qp / rcc->accum_p_norm ); + q /= fabs( h->param.rc.f_ip_factor ); + q = clip_qscale( h, pict_type, q ); + } + else + { + if( h->stat.i_slice_count[h->param.i_keyint_max > 1 ? SLICE_TYPE_P : SLICE_TYPE_I] < 5 ) + { + float w = h->stat.i_slice_count[SLICE_TYPE_P] / 5.; + float q2 = qp2qscale(ABR_INIT_QP); + q = q*w + q2*(1-w); + } - if(pict_type!=SLICE_TYPE_I) - assert(pict_type == rce->new_pict_type); + /* Asymmetric clipping, because symmetric would prevent + * overflow control in areas of rapidly oscillating complexity */ + lmin = rcc->last_qscale_for[pict_type] / rcc->lstep; + lmax = rcc->last_qscale_for[pict_type] * rcc->lstep; + if( overflow > 1.1 ) + lmax *= rcc->lstep; + else if( overflow < 0.9 ) + lmin /= rcc->lstep; + + q = x264_clip3f(q, lmin, lmax); + q = clip_qscale(h, pict_type, q); + //FIXME use get_diff_limited_q() ? + } + } - diff = (int64_t)total_bits - (int64_t)rce->expected_bits; - br_compensation = (rcc->buffer_size - diff) / rcc->buffer_size; - if(br_compensation<=0.0) br_compensation=0.001; + rcc->last_qscale_for[pict_type] = + rcc->last_qscale = q; - q = rce->new_qscale / br_compensation; - q = x264_clip3f(q, lmin, lmax); - rcc->last_qscale = q; - return q; + return q; + } } static int init_pass2( x264_t *h ) @@ -793,7 +1008,6 @@ static int init_pass2( x264_t *h ) double qblur = h->param.rc.f_qblur; double cplxblur = h->param.rc.f_complexity_blur; const int filter_size = (int)(qblur*4) | 1; - const int cplx_filter_size = (int)(cplxblur*2); double expected_bits; double *qscale, *blurred_qscale; int i; @@ -801,12 +1015,11 @@ static int init_pass2( x264_t *h ) /* find total/average complexity & const_bits */ for(i=0; inum_entries; i++){ ratecontrol_entry_t *rce = &rcc->entry[i]; - rce->new_pict_type = rce->pict_type; - all_const_bits += rce->mv_bits + rce->misc_bits; - rcc->i_cplx_sum[rce->new_pict_type] += rce->i_tex_bits * rce->qscale; - rcc->p_cplx_sum[rce->new_pict_type] += rce->p_tex_bits * rce->qscale; - rcc->mv_bits_sum[rce->new_pict_type] += rce->mv_bits; - rcc->frame_count[rce->new_pict_type] ++; + all_const_bits += rce->misc_bits; + rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale; + rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale; + rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits * rce->qscale; + rcc->frame_count[rce->pict_type] ++; } if( all_available_bits < all_const_bits) @@ -816,32 +1029,32 @@ static int init_pass2( x264_t *h ) return -1; } - if(cplxblur<.01) - cplxblur = .01; + /* Blur complexities, to reduce local fluctuation of QP. + * We don't blur the QPs directly, because then one very simple frame + * could drag down the QP of a nearby complex frame and give it more + * bits than intended. */ for(i=0; inum_entries; i++){ ratecontrol_entry_t *rce = &rcc->entry[i]; - double weight, weight_sum = 0; + double weight_sum = 0; double cplx_sum = 0; - double mask = 1.0; + double weight = 1.0; int j; /* weighted average of cplx of future frames */ - for(j=1; jnum_entries-i; j++){ + for(j=1; jnum_entries-i; j++){ ratecontrol_entry_t *rcj = &rcc->entry[i+j]; - mask *= (double)(rcc->nmb + rcj->i_count) / rcc->nmb; - weight = mask * exp(-j*j/(cplxblur*cplxblur)); + weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 ); if(weight < .0001) break; weight_sum += weight; - cplx_sum += weight * (rcj->i_tex_bits + rcj->p_tex_bits) * rce->qscale; + cplx_sum += weight * qscale2bits(rcj, 1); } /* weighted average of cplx of past frames */ - mask = 1.0; - for(j=0; jentry[i-j]; - weight = mask * exp(-j*j/(cplxblur*cplxblur)); weight_sum += weight; - cplx_sum += weight * (rcj->i_tex_bits + rcj->p_tex_bits) * rce->qscale; - mask *= (double)(rcc->nmb + rcj->i_count) / rcc->nmb; + cplx_sum += weight * qscale2bits(rcj, 1); + weight *= 1 - pow( (float)rcj->i_count / rcc->nmb, 2 ); if(weight < .0001) break; } @@ -854,9 +1067,16 @@ static int init_pass2( x264_t *h ) else blurred_qscale = qscale; + /* Search for a factor which, when multiplied by the RCEQ values from + * each frame, adds up to the desired total size. + * There is no exact closed-form solution because of VBV constraints and + * because qscale2bits is not invertible, but we can start with the simple + * approximation of scaling the 1st pass by the ratio of bitrates. + * The search range is probably overkill, but speed doesn't matter here. */ + expected_bits = 1; for(i=0; inum_entries; i++) - expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0)); + expected_bits += qscale2bits(&rcc->entry[i], get_qscale(h, &rcc->entry[i], 1.0, i)); step_mult = all_available_bits / expected_bits; rate_factor = 0; @@ -864,12 +1084,17 @@ static int init_pass2( x264_t *h ) expected_bits = 0; rate_factor += step; + rcc->last_non_b_pict_type = -1; + rcc->last_accum_p_norm = 1; + rcc->accum_p_norm = 0; + rcc->buffer_fill = rcc->buffer_size * h->param.rc.f_vbv_buffer_init; + /* find qscale */ for(i=0; inum_entries; i++){ - qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor); + qscale[i] = get_qscale(h, &rcc->entry[i], rate_factor, i); } - /* fixed I/B QP relative to P mode */ + /* fixed I/B qscale relative to P */ for(i=rcc->num_entries-1; i>=0; i--){ qscale[i] = get_diff_limited_q(h, &rcc->entry[i], qscale[i]); assert(qscale[i] >= 0); @@ -880,7 +1105,6 @@ static int init_pass2( x264_t *h ) assert(filter_size%2==1); for(i=0; inum_entries; i++){ ratecontrol_entry_t *rce = &rcc->entry[i]; - const int pict_type = rce->new_pict_type; int j; double q=0.0, sum=0.0; @@ -889,7 +1113,7 @@ static int init_pass2( x264_t *h ) double d = index-i; double coeff = qblur==0 ? 1.0 : exp(-d*d/(qblur*qblur)); if(index < 0 || index >= rcc->num_entries) continue; - if(pict_type != rcc->entry[index].new_pict_type) continue; + if(rce->pict_type != rcc->entry[index].pict_type) continue; q += qscale[index] * coeff; sum += coeff; } @@ -901,12 +1125,13 @@ static int init_pass2( x264_t *h ) for(i=0; inum_entries; i++){ ratecontrol_entry_t *rce = &rcc->entry[i]; double bits; - rce->new_qscale = clip_qscale(h, rce, blurred_qscale[i]); + rce->new_qscale = clip_qscale(h, rce->pict_type, blurred_qscale[i]); assert(rce->new_qscale >= 0); - bits = qscale2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits; + bits = qscale2bits(rce, rce->new_qscale) + rce->misc_bits; rce->expected_bits = expected_bits; expected_bits += bits; + update_vbv(h, bits); } //printf("expected:%llu available:%llu factor:%lf avgQ:%lf\n", (uint64_t)expected_bits, all_available_bits, rate_factor); @@ -925,12 +1150,26 @@ static int init_pass2( x264_t *h ) avgq = qscale2qp(avgq / rcc->num_entries); x264_log(h, X264_LOG_ERROR, "Error: 2pass curve failed to converge\n"); - x264_log(h, X264_LOG_ERROR, "expected bits: %llu, available: %llu, avg QP: %.4lf\n", (uint64_t)expected_bits, all_available_bits, avgq); - if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 1) - x264_log(h, X264_LOG_ERROR, "try reducing bitrate or reducing qp_min\n"); - if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_min - 1) - x264_log(h, X264_LOG_ERROR, "try increasing bitrate or increasing qp_max\n"); - return -1; + x264_log(h, X264_LOG_ERROR, "target: %.2f kbit/s, expected: %.2f kbit/s, avg QP: %.4f\n", + (float)h->param.rc.i_bitrate, + expected_bits * rcc->fps / (rcc->num_entries * 1000.), + avgq); + if(expected_bits < all_available_bits && avgq < h->param.rc.i_qp_min + 2) + { + if(h->param.rc.i_qp_min > 0) + x264_log(h, X264_LOG_ERROR, "try reducing target bitrate or reducing qp_min (currently %d)\n", h->param.rc.i_qp_min); + else + x264_log(h, X264_LOG_ERROR, "try reducing target bitrate\n"); + } + else if(expected_bits > all_available_bits && avgq > h->param.rc.i_qp_max - 2) + { + if(h->param.rc.i_qp_max < 51) + x264_log(h, X264_LOG_ERROR, "try increasing target bitrate or increasing qp_max (currently %d)\n", h->param.rc.i_qp_max); + else + x264_log(h, X264_LOG_ERROR, "try increasing target bitrate\n"); + } + else + x264_log(h, X264_LOG_ERROR, "internal error\n"); } return 0;