Mercurial > libavcodec.hg
changeset 1505:010f76d07a27 libavcodec
use lagrange multipler instead of qp for ratecontrol, this may break some things, tell me ASAP if u notice anything broken
quality which was 1..31 float is now a 1..FF_LAMBDA_MAX int, and FF_QP2LAMBDA * qp can be used to convert to the new range
| author | michaelni |
|---|---|
| date | Tue, 07 Oct 2003 11:32:40 +0000 |
| parents | df7ab60d1ee0 |
| children | e0c02032d0f2 |
| files | avcodec.h h263.c mjpeg.c mpegvideo.c mpegvideo.h ratecontrol.c utils.c |
| diffstat | 7 files changed, 111 insertions(+), 82 deletions(-) [+] |
line wrap: on
line diff
--- a/avcodec.h Sun Oct 05 21:48:16 2003 +0000 +++ b/avcodec.h Tue Oct 07 11:32:40 2003 +0000 @@ -15,7 +15,7 @@ #define FFMPEG_VERSION_INT 0x000408 #define FFMPEG_VERSION "0.4.8" -#define LIBAVCODEC_BUILD 4683 +#define LIBAVCODEC_BUILD 4684 #define LIBAVCODEC_VERSION_INT FFMPEG_VERSION_INT #define LIBAVCODEC_VERSION FFMPEG_VERSION @@ -316,15 +316,15 @@ int display_picture_number;\ \ /**\ - * quality (between 1 (good) and 31 (bad)) \ + * quality (between 1 (good) and FF_LAMBDA_MAX (bad)) \ * - encoding: set by lavc for coded_picture (and set by user for input)\ * - decoding: set by lavc\ */\ - float quality; \ + int quality; \ \ /**\ * buffer age (1->was last buffer and dint change, 2->..., ...).\ - * set to something large if the buffer has not been used yet \ + * set to INT_MAX if the buffer has not been used yet \ * - encoding: unused\ * - decoding: MUST be set by get_buffer()\ */\ @@ -1181,8 +1181,13 @@ * Dont touch, used by lavc default_get_buffer() */ void *internal_buffer; - -#define FF_QUALITY_SCALE 256 + +#define FF_LAMBDA_SHIFT 7 +#define FF_LAMBDA_SCALE (1<<FF_LAMBDA_SHIFT) +#define FF_QP2LAMBDA 118 ///< factor to convert from H.263 QP to lambda +#define FF_LAMBDA_MAX (256*128-1) + +#define FF_QUALITY_SCALE FF_LAMBDA_SCALE //FIXME maybe remove /** * global quality for codecs which cannot change it per frame. * this should be proportional to MPEG1/2/4 qscale. @@ -1263,6 +1268,20 @@ * - decoding: unused */ int scenechange_threshold; + + /** + * minimum lagrange multipler + * - encoding: set by user. + * - decoding: unused + */ + int lmin; + + /** + * maximum lagrange multipler + * - encoding: set by user. + * - decoding: unused + */ + int lmax; } AVCodecContext;
--- a/h263.c Sun Oct 05 21:48:16 2003 +0000 +++ b/h263.c Tue Oct 07 11:32:40 2003 +0000 @@ -626,7 +626,7 @@ int best_cbpc_score= INT_MAX; int cbpc, cbpy; const int offset= (s->mv_type==MV_TYPE_16X16 ? 0 : 16) + (s->dquant ? 8 : 0); - const int lambda= (s->qscale*s->qscale*64*105 + 64)>>7; + const int lambda= s->lambda2 >> (FF_LAMBDA_SHIFT - 6); for(i=0; i<4; i++){ int score= inter_MCBPC_bits[i + offset] * lambda; @@ -679,7 +679,7 @@ if(s->flags & CODEC_FLAG_CBP_RD){ int score=0; - const int lambda= (s->qscale*s->qscale*64*105 + 64)>>7; + const int lambda= s->lambda2 >> (FF_LAMBDA_SHIFT - 6); for(i=0; i<6; i++){ if(s->coded_score[i] < 0){
--- a/mjpeg.c Sun Oct 05 21:48:16 2003 +0000 +++ b/mjpeg.c Tue Oct 07 11:32:40 2003 +0000 @@ -1877,7 +1877,8 @@ picture->qscale_table= s->qscale_table; memset(picture->qscale_table, picture->quality, (s->width+15)/16); if(avctx->debug & FF_DEBUG_QP) - printf("QP: %d\n", (int)picture->quality); + printf("QP: %d\n", picture->quality); + picture->quality*= FF_QP2LAMBDA; } goto the_end; @@ -2052,7 +2053,8 @@ picture->qscale_table= s->qscale_table; memset(picture->qscale_table, picture->quality, (s->width+15)/16); if(avctx->debug & FF_DEBUG_QP) - printf("QP: %f\n", picture->quality); + printf("QP: %d\n", picture->quality); + picture->quality*= FF_QP2LAMBDA; } return buf_ptr - buf;
--- a/mpegvideo.c Sun Oct 05 21:48:16 2003 +0000 +++ b/mpegvideo.c Tue Oct 07 11:32:40 2003 +0000 @@ -140,6 +140,13 @@ } } } + +static inline void update_qscale(MpegEncContext *s){ + s->qscale= (s->lambda*139 + FF_LAMBDA_SCALE*64) >> (FF_LAMBDA_SHIFT + 7); + s->qscale= clip(s->qscale, s->avctx->qmin, s->avctx->qmax); + + s->lambda2= (s->lambda*s->lambda + FF_LAMBDA_SCALE/2) >> FF_LAMBDA_SHIFT; +} #endif //CONFIG_ENCODERS void ff_init_scantable(uint8_t *permutation, ScanTable *st, const uint8_t *src_scantable){ @@ -424,6 +431,8 @@ /* Allocate MB type table */ CHECKED_ALLOCZ(s->mb_type , mb_array_size * sizeof(uint8_t)) //needed for encoding + + CHECKED_ALLOCZ(s->lambda_table, mb_array_size * sizeof(int)) } CHECKED_ALLOCZ(s->error_status_table, mb_array_size*sizeof(uint8_t)) @@ -539,6 +548,7 @@ av_freep(&s->ac_stats); av_freep(&s->error_status_table); av_freep(&s->mb_index2xy); + av_freep(&s->lambda_table); for(i=0; i<MAX_PICTURE_COUNT; i++){ free_picture(s, &s->picture[i]); @@ -1028,7 +1038,8 @@ } s->current_picture_ptr->pict_type= s->pict_type; - s->current_picture_ptr->quality= s->qscale; +// if(s->flags && CODEC_FLAG_QSCALE) + // s->current_picture_ptr->quality= s->new_picture_ptr->quality; s->current_picture_ptr->key_frame= s->pict_type == I_TYPE; s->current_picture= *s->current_picture_ptr; @@ -1602,10 +1613,6 @@ if(s->new_picture.data[0]){ s->pict_type= s->new_picture.pict_type; - if (s->fixed_qscale){ /* the ratecontrol needs the last qscale so we dont touch it for CBR */ - s->qscale= (int)(s->new_picture.quality+0.5); - assert(s->qscale); - } //emms_c(); //printf("qs:%f %f %d\n", s->new_picture.quality, s->current_picture.quality, s->qscale); MPV_frame_start(s, avctx); @@ -2815,22 +2822,23 @@ for(i=0; i<6; i++) skip_dct[i]=0; if(s->adaptive_quant){ - s->dquant= s->current_picture.qscale_table[mb_x + mb_y*s->mb_stride] - s->qscale; - - if(s->out_format==FMT_H263){ - if (s->dquant> 2) s->dquant= 2; - else if(s->dquant<-2) s->dquant=-2; - } + const int last_qp= s->qscale; + const int mb_xy= mb_x + mb_y*s->mb_stride; + + s->lambda= s->lambda_table[mb_xy]; + update_qscale(s); + s->dquant= s->qscale - last_qp; + + if(s->out_format==FMT_H263) + s->dquant= clip(s->dquant, -2, 2); //FIXME RD if(s->codec_id==CODEC_ID_MPEG4){ if(!s->mb_intra){ - if(s->mv_dir&MV_DIRECT) + if((s->mv_dir&MV_DIRECT) || s->mv_type==MV_TYPE_8X8) s->dquant=0; - - assert(s->dquant==0 || s->mv_type!=MV_TYPE_8X8); } } - s->qscale+= s->dquant; + s->qscale= last_qp + s->dquant; s->y_dc_scale= s->y_dc_scale_table[ s->qscale ]; s->c_dc_scale= s->c_dc_scale_table[ s->qscale ]; } @@ -2993,19 +3001,6 @@ } -#if 0 - { - float adap_parm; - - adap_parm = ((s->avg_mb_var << 1) + s->mb_var[s->mb_stride*mb_y+mb_x] + 1.0) / - ((s->mb_var[s->mb_stride*mb_y+mb_x] << 1) + s->avg_mb_var + 1.0); - - printf("\ntype=%c qscale=%2d adap=%0.2f dquant=%4.2f var=%4d avgvar=%4d", - (s->mb_type[s->mb_stride*mb_y+mb_x] > 0) ? 'I' : 'P', - s->qscale, adap_parm, s->qscale*adap_parm, - s->mb_var[s->mb_stride*mb_y+mb_x], s->avg_mb_var); - } -#endif /* DCT & quantize */ if(s->out_format==FMT_MJPEG){ for(i=0;i<6;i++) { @@ -3265,8 +3260,8 @@ if(s->avctx->mb_decision == FF_MB_DECISION_RD){ MPV_decode_mb(s, s->block); - score *= s->qscale * s->qscale * 109; - score += sse_mb(s) << 7; + score *= s->lambda2; + score += sse_mb(s) << FF_LAMBDA_SHIFT; } if(*next_block){ @@ -3351,7 +3346,7 @@ s->scene_change_score=0; - s->qscale= (int)(s->frame_qscale + 0.5); //FIXME qscale / ... stuff for ME ratedistoration + s->lambda= s->current_picture_ptr->quality; //FIXME qscale / ... stuff for ME ratedistoration if(s->pict_type==I_TYPE){ if(s->msmpeg4_version >= 3) s->no_rounding=1; @@ -3460,10 +3455,8 @@ } } - if (s->fixed_qscale) - s->frame_qscale = s->current_picture.quality; - else - s->frame_qscale = ff_rate_estimate_qscale(s); + if (!s->fixed_qscale) + s->current_picture.quality = ff_rate_estimate_qscale(s); if(s->adaptive_quant){ #ifdef CONFIG_RISKY @@ -3479,9 +3472,15 @@ } #endif - s->qscale= s->current_picture.qscale_table[0]; + s->lambda= s->lambda_table[0]; + //FIXME broken }else - s->qscale= (int)(s->frame_qscale + 0.5); + s->lambda= s->current_picture.quality; +//printf("%d %d\n", s->avctx->global_quality, s->current_picture.quality); + update_qscale(s); + + if(s->qscale < 3 && s->max_qcoeff<=128 && s->pict_type==I_TYPE && !(s->flags & CODEC_FLAG_QSCALE)) + s->qscale= 3; //reduce cliping problems if (s->out_format == FMT_MJPEG) { /* for mjpeg, we do include qscale in the matrix */ @@ -4020,12 +4019,14 @@ int not_coded_score= 0; int coeff[3][64]; int coeff_count[64]; - int lambda, qmul, qadd, start_i, last_non_zero, i, dc; + int qmul, qadd, start_i, last_non_zero, i, dc; const int esc_length= s->ac_esc_length; uint8_t * length; uint8_t * last_length; int score_limit=0; int left_limit= 0; + const int lambda= s->lambda2 >> (FF_LAMBDA_SHIFT - 6); + const int patch_table= s->out_format == FMT_MPEG1 && !s->mb_intra; s->dsp.fdct (block); @@ -4103,9 +4104,13 @@ return last_non_zero; } - lambda= (qscale*qscale*64*105 + 64)>>7; //FIXME finetune - score_tab[0]= 0; + + if(patch_table){ +// length[UNI_AC_ENC_INDEX(0, 63)]= +// length[UNI_AC_ENC_INDEX(0, 65)]= 2; + } + for(i=0; i<=last_non_zero - start_i; i++){ int level_index, run, j; const int dct_coeff= block[ scantable[i + start_i] ]; @@ -4212,10 +4217,13 @@ //Note: there is a vlc code in mpeg4 which is 1 bit shorter then another one with a shorter run and the same level while(score_tab[ left_limit ] > score_limit + lambda) left_limit++; + + if(patch_table){ +// length[UNI_AC_ENC_INDEX(0, 63)]= +// length[UNI_AC_ENC_INDEX(0, 65)]= 3; + } } - //FIXME add some cbp penalty - if(s->out_format != FMT_H263){ last_score= 256*256*256*120; for(i= left_limit; i<=last_non_zero - start_i + 1; i++){
--- a/mpegvideo.h Sun Oct 05 21:48:16 2003 +0000 +++ b/mpegvideo.h Tue Oct 07 11:32:40 2003 +0000 @@ -354,7 +354,9 @@ uint8_t *edge_emu_buffer; ///< points into the middle of allocated_edge_emu_buffer int qscale; ///< QP - float frame_qscale; ///< qscale from the frame level rc FIXME remove + int lambda; ///< lagrange multipler used in rate distortion + int lambda2; ///< (lambda*lambda) >> FF_LAMBDA_SHIFT + int *lambda_table; int adaptive_quant; ///< use adaptive quantization int dquant; ///< qscale difference to prev qscale int pict_type; ///< I_TYPE, P_TYPE, B_TYPE, ...
--- a/ratecontrol.c Sun Oct 05 21:48:16 2003 +0000 +++ b/ratecontrol.c Tue Oct 07 11:32:40 2003 +0000 @@ -38,9 +38,9 @@ static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num); void ff_write_pass1_stats(MpegEncContext *s){ - sprintf(s->avctx->stats_out, "in:%d out:%d type:%d q:%f itex:%d ptex:%d mv:%d misc:%d fcode:%d bcode:%d mc-var:%d var:%d icount:%d;\n", + sprintf(s->avctx->stats_out, "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d fcode:%d bcode:%d mc-var:%d var:%d icount:%d;\n", s->picture_number, s->input_picture_number - s->max_b_frames, s->pict_type, - s->frame_qscale, s->i_tex_bits, s->p_tex_bits, s->mv_bits, s->misc_bits, + s->current_picture.quality, s->i_tex_bits, s->p_tex_bits, s->mv_bits, s->misc_bits, s->f_code, s->b_code, s->current_picture.mc_mb_var_sum, s->current_picture.mb_var_sum, s->i_count); } @@ -51,7 +51,7 @@ emms_c(); for(i=0; i<5; i++){ - rcc->pred[i].coeff= 7.0; + rcc->pred[i].coeff= FF_QP2LAMBDA * 7.0; rcc->pred[i].count= 1.0; rcc->pred[i].decay= 0.4; @@ -60,7 +60,7 @@ rcc->mv_bits_sum[i]= rcc->qscale_sum [i]= rcc->frame_count[i]= 1; // 1 is better cuz of 1/0 and such - rcc->last_qscale_for[i]=5; + rcc->last_qscale_for[i]=FF_QP2LAMBDA * 5; } rcc->buffer_index= s->avctx->rc_buffer_size/2; @@ -81,7 +81,7 @@ for(i=0; i<rcc->num_entries; i++){ RateControlEntry *rce= &rcc->entry[i]; rce->pict_type= rce->new_pict_type=P_TYPE; - rce->qscale= rce->new_qscale=2; + rce->qscale= rce->new_qscale=FF_QP2LAMBDA * 2; rce->misc_bits= s->mb_num + 10; rce->mb_var_sum= s->mb_num*100; } @@ -140,7 +140,7 @@ rce.new_pict_type= rce.pict_type; rce.mc_mb_var_sum= bits*s->mb_num/100000; rce.mb_var_sum = s->mb_num; - rce.qscale = 2; + rce.qscale = FF_QP2LAMBDA * 2; rce.f_code = 2; rce.b_code = 1; rce.misc_bits= 1; @@ -339,9 +339,10 @@ /* last qscale / qdiff stuff */ if(rcc->last_non_b_pict_type==pict_type || pict_type!=I_TYPE){ double last_q= rcc->last_qscale_for[pict_type]; + const int maxdiff= FF_QP2LAMBDA * a->max_qdiff; - if (q > last_q + a->max_qdiff) q= last_q + a->max_qdiff; - else if(q < last_q - a->max_qdiff) q= last_q - a->max_qdiff; + if (q > last_q + maxdiff) q= last_q + maxdiff; + else if(q < last_q - maxdiff) q= last_q - maxdiff; } rcc->last_qscale_for[pict_type]= q; //Note we cant do that after blurring @@ -356,8 +357,8 @@ * gets the qmin & qmax for pict_type */ static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type){ - int qmin= s->avctx->qmin; - int qmax= s->avctx->qmax; + int qmin= s->avctx->lmin; + int qmax= s->avctx->lmax; assert(qmin <= qmax); @@ -369,12 +370,8 @@ qmax= (int)(qmax*ABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5); } - qmin= clip(qmin, 1, 31); - qmax= clip(qmax, 1, 31); - - if(qmin==1 && s->avctx->qmin>1) qmin=2; //avoid qmin=1 unless the user wants qmin=1 - - if(qmin<3 && s->max_qcoeff<=128 && pict_type==I_TYPE) qmin=3; //reduce cliping problems + qmin= clip(qmin, 1, FF_LAMBDA_MAX); + qmax= clip(qmax, 1, FF_LAMBDA_MAX); if(qmax<qmin) qmax= qmin; @@ -479,21 +476,23 @@ float cplx_sum= 0.0; float cplx_tab[s->mb_num]; float bits_tab[s->mb_num]; - const int qmin= s->avctx->mb_qmin; - const int qmax= s->avctx->mb_qmax; + const int qmin= s->avctx->lmin; + const int qmax= s->avctx->lmax; Picture * const pic= &s->current_picture; - int last_qscale=0; for(i=0; i<s->mb_num; i++){ const int mb_xy= s->mb_index2xy[i]; - float temp_cplx= sqrt(pic->mc_mb_var[mb_xy]); + float temp_cplx= sqrt(pic->mc_mb_var[mb_xy]); //FIXME merge in pow() float spat_cplx= sqrt(pic->mb_var[mb_xy]); const int lumi= pic->mb_mean[mb_xy]; float bits, cplx, factor; - +#if 0 if(spat_cplx < q/3) spat_cplx= q/3; //FIXME finetune if(temp_cplx < q/3) temp_cplx= q/3; //FIXME finetune - +#endif + if(spat_cplx < 4) spat_cplx= 4; //FIXME finetune + if(temp_cplx < 4) temp_cplx= 4; //FIXME finetune + if((s->mb_type[mb_xy]&MB_TYPE_INTRA)){//FIXME hq mode cplx= spat_cplx; factor= 1.0 + p_masking; @@ -543,19 +542,16 @@ newq*= bits_sum/cplx_sum; } - if(i && ABS(last_qscale - newq)<0.75) - intq= last_qscale; - else - intq= (int)(newq + 0.5); + intq= (int)(newq + 0.5); if (intq > qmax) intq= qmax; else if(intq < qmin) intq= qmin; //if(i%s->mb_width==0) printf("\n"); //printf("%2d%3d ", intq, ff_sqrt(s->mc_mb_var[i])); - last_qscale= - pic->qscale_table[mb_xy]= intq; + s->lambda_table[mb_xy]= intq; } } +//FIXME rd or at least approx for dquant float ff_rate_estimate_qscale(MpegEncContext *s) { @@ -614,7 +610,7 @@ rce->new_pict_type= pict_type; rce->mc_mb_var_sum= pic->mc_mb_var_sum; rce->mb_var_sum = pic-> mb_var_sum; - rce->qscale = 2; + rce->qscale = FF_QP2LAMBDA * 2; rce->f_code = s->f_code; rce->b_code = s->b_code; rce->misc_bits= 1;
--- a/utils.c Sun Oct 05 21:48:16 2003 +0000 +++ b/utils.c Tue Oct 07 11:32:40 2003 +0000 @@ -272,6 +272,8 @@ s->release_buffer= avcodec_default_release_buffer; s->get_format= avcodec_default_get_format; s->me_subpel_quality=8; + s->lmin= FF_QP2LAMBDA * s->qmin; + s->lmax= FF_QP2LAMBDA * s->qmax; s->intra_quant_bias= FF_DEFAULT_QUANT_BIAS; s->inter_quant_bias= FF_DEFAULT_QUANT_BIAS;