Use forced key frames when encoding with libxvid.
[ffmpeg.git] / libavcodec / ratecontrol.c
index 59dd654..4261678 100644 (file)
 /*
-    Copyright (C) 2002 Michael Niedermayer <michaelni@gmx.at>
-
-    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
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License
-    along with this program; if not, write to the Free Software
-    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
-*/
-
-#include <inttypes.h>
-#include <stdio.h>
-
-#include "common.h" //needed for mpegvideo.h to compile
-#include "dsputil.h" //needed for mpegvideo.h to compile
+ * Rate control for video encoders
+ *
+ * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file
+ * Rate control for video encoders.
+ */
+
+#include "libavutil/intmath.h"
 #include "avcodec.h"
+#include "dsputil.h"
+#include "ratecontrol.h"
 #include "mpegvideo.h"
+#include "libavutil/eval.h"
+
+#undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
+#include <assert.h>
 
-#define STATS_FILE "lavc_stats.txt"
+#ifndef M_E
+#define M_E 2.718281828
+#endif
 
 static int init_pass2(MpegEncContext *s);
+static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num);
 
 void ff_write_pass1_stats(MpegEncContext *s){
-    RateControlContext *rcc= &s->rc_context;
-//    fprintf(c->stats_file, "type:%d q:%d icount:%d pcount:%d scount:%d itex:%d ptex%d mv:%d misc:%d fcode:%d bcode:%d\")
-    fprintf(rcc->stats_file, "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d fcode:%d bcode:%d\n",
-            s->picture_number, s->input_picture_number - s->max_b_frames, s->pict_type, 
-            s->qscale, s->i_tex_bits, s->p_tex_bits, s->mv_bits, s->misc_bits, s->f_code, s->b_code);
+    snprintf(s->avctx->stats_out, 256, "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 skipcount:%d hbits:%d;\n",
+            s->current_picture_ptr->display_picture_number, s->current_picture_ptr->coded_picture_number, s->pict_type,
+            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, s->skip_count, s->header_bits);
+}
+
+static inline double qp2bits(RateControlEntry *rce, double qp){
+    if(qp<=0.0){
+        av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
+    }
+    return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ qp;
+}
+
+static inline double bits2qp(RateControlEntry *rce, double bits){
+    if(bits<0.9){
+        av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
+    }
+    return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ bits;
 }
 
 int ff_rate_control_init(MpegEncContext *s)
 {
     RateControlContext *rcc= &s->rc_context;
+    int i, res;
+    static const char * const const_names[]={
+        "PI",
+        "E",
+        "iTex",
+        "pTex",
+        "tex",
+        "mv",
+        "fCode",
+        "iCount",
+        "mcVar",
+        "var",
+        "isI",
+        "isP",
+        "isB",
+        "avgQP",
+        "qComp",
+/*        "lastIQP",
+        "lastPQP",
+        "lastBQP",
+        "nextNonBQP",*/
+        "avgIITex",
+        "avgPITex",
+        "avgPPTex",
+        "avgBPTex",
+        "avgTex",
+        NULL
+    };
+    static double (* const func1[])(void *, double)={
+        (void *)bits2qp,
+        (void *)qp2bits,
+        NULL
+    };
+    static const char * const func1_names[]={
+        "bits2qp",
+        "qp2bits",
+        NULL
+    };
     emms_c();
 
-    if(s->flags&CODEC_FLAG_PASS1){
-        rcc->stats_file= fopen(STATS_FILE, "w");
-        if(!rcc->stats_file){
-            fprintf(stderr, "failed to open " STATS_FILE "\n");
-            return -1;
-        }
-    } else if(s->flags&CODEC_FLAG_PASS2){
-        int size;
+    res = av_parse_expr(&rcc->rc_eq_eval, s->avctx->rc_eq ? s->avctx->rc_eq : "tex^qComp", const_names, func1_names, func1, NULL, NULL, 0, s->avctx);
+    if (res < 0) {
+        av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->avctx->rc_eq);
+        return res;
+    }
+
+    for(i=0; i<5; i++){
+        rcc->pred[i].coeff= FF_QP2LAMBDA * 7.0;
+        rcc->pred[i].count= 1.0;
+
+        rcc->pred[i].decay= 0.4;
+        rcc->i_cplx_sum [i]=
+        rcc->p_cplx_sum [i]=
+        rcc->mv_bits_sum[i]=
+        rcc->qscale_sum [i]=
+        rcc->frame_count[i]= 1; // 1 is better because of 1/0 and such
+        rcc->last_qscale_for[i]=FF_QP2LAMBDA * 5;
+    }
+    rcc->buffer_index= s->avctx->rc_initial_buffer_occupancy;
+
+    if(s->flags&CODEC_FLAG_PASS2){
         int i;
+        char *p;
 
-        rcc->stats_file= fopen(STATS_FILE, "r");
-        if(!rcc->stats_file){
-            fprintf(stderr, "failed to open " STATS_FILE "\n");
-            return -1;
+        /* find number of pics */
+        p= s->avctx->stats_in;
+        for(i=-1; p; i++){
+            p= strchr(p+1, ';');
         }
+        i+= s->max_b_frames;
+        if(i<=0 || i>=INT_MAX / sizeof(RateControlEntry))
+            return -1;
+        rcc->entry = av_mallocz(i*sizeof(RateControlEntry));
+        rcc->num_entries= i;
 
-        /* find number of pics without reading the file twice :) */
-        fseek(rcc->stats_file, 0, SEEK_END);
-        size= ftell(rcc->stats_file);
-        fseek(rcc->stats_file, 0, SEEK_SET);
-
-        size/= 64; // we need at least 64 byte to store a line ...
-        rcc->entry = (RateControlEntry*)av_mallocz(size*sizeof(RateControlEntry));
+        /* init all to skipped p frames (with b frames we might have a not encoded frame at the end FIXME) */
+        for(i=0; i<rcc->num_entries; i++){
+            RateControlEntry *rce= &rcc->entry[i];
+            rce->pict_type= rce->new_pict_type=FF_P_TYPE;
+            rce->qscale= rce->new_qscale=FF_QP2LAMBDA * 2;
+            rce->misc_bits= s->mb_num + 10;
+            rce->mb_var_sum= s->mb_num*100;
+        }
 
-        for(i=0; !feof(rcc->stats_file); i++){
+        /* read stats */
+        p= s->avctx->stats_in;
+        for(i=0; i<rcc->num_entries - s->max_b_frames; i++){
             RateControlEntry *rce;
             int picture_number;
             int e;
-            
-            e= fscanf(rcc->stats_file, "in:%d ", &picture_number);
+            char *next;
+
+            next= strchr(p, ';');
+            if(next){
+                (*next)=0; //sscanf in unbelievably slow on looong strings //FIXME copy / do not write
+                next++;
+            }
+            e= sscanf(p, " in:%d ", &picture_number);
+
+            assert(picture_number >= 0);
+            assert(picture_number < rcc->num_entries);
             rce= &rcc->entry[picture_number];
-            e+=fscanf(rcc->stats_file, "out:%*d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d fcode:%*d bcode:%*d\n",
-                   &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits, &rce->mv_bits, &rce->misc_bits);
-            if(e!=7){
-                fprintf(stderr, STATS_FILE " is damaged\n");
+
+            e+=sscanf(p, " 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 skipcount:%d hbits:%d",
+                   &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits, &rce->mv_bits, &rce->misc_bits,
+                   &rce->f_code, &rce->b_code, &rce->mc_mb_var_sum, &rce->mb_var_sum, &rce->i_count, &rce->skip_count, &rce->header_bits);
+            if(e!=14){
+                av_log(s->avctx, AV_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
                 return -1;
             }
+
+            p= next;
         }
-        rcc->num_entries= i;
-        
+
         if(init_pass2(s) < 0) return -1;
+
+        //FIXME maybe move to end
+        if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) {
+#if CONFIG_LIBXVID
+            return ff_xvid_rate_control_init(s);
+#else
+            av_log(s->avctx, AV_LOG_ERROR, "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n");
+            return -1;
+#endif
+        }
+    }
+
+    if(!(s->flags&CODEC_FLAG_PASS2)){
+
+        rcc->short_term_qsum=0.001;
+        rcc->short_term_qcount=0.001;
+
+        rcc->pass1_rc_eq_output_sum= 0.001;
+        rcc->pass1_wanted_bits=0.001;
+
+        if(s->avctx->qblur > 1.0){
+            av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
+            return -1;
+        }
+        /* init stuff with the user specified complexity */
+        if(s->avctx->rc_initial_cplx){
+            for(i=0; i<60*30; i++){
+                double bits= s->avctx->rc_initial_cplx * (i/10000.0 + 1.0)*s->mb_num;
+                RateControlEntry rce;
+
+                if     (i%((s->gop_size+3)/4)==0) rce.pict_type= FF_I_TYPE;
+                else if(i%(s->max_b_frames+1))    rce.pict_type= FF_B_TYPE;
+                else                              rce.pict_type= FF_P_TYPE;
+
+                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   = FF_QP2LAMBDA * 2;
+                rce.f_code   = 2;
+                rce.b_code   = 1;
+                rce.misc_bits= 1;
+
+                if(s->pict_type== FF_I_TYPE){
+                    rce.i_count   = s->mb_num;
+                    rce.i_tex_bits= bits;
+                    rce.p_tex_bits= 0;
+                    rce.mv_bits= 0;
+                }else{
+                    rce.i_count   = 0; //FIXME we do know this approx
+                    rce.i_tex_bits= 0;
+                    rce.p_tex_bits= bits*0.9;
+                    rce.mv_bits= bits*0.1;
+                }
+                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;
+                rcc->frame_count[rce.pict_type] ++;
+
+                get_qscale(s, &rce, rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum, i);
+                rcc->pass1_wanted_bits+= s->bit_rate/(1/av_q2d(s->avctx->time_base)); //FIXME misbehaves a little for variable fps
+            }
+        }
+
     }
-     
-    /* no 2pass stuff, just normal 1-pass */
-    //initial values, they dont really matter as they will be totally different within a few frames
-    s->i_pred.coeff= s->p_pred.coeff= 7.0;
-    s->i_pred.count= s->p_pred.count= 1.0;
-    
-    s->i_pred.decay= s->p_pred.decay= 0.4;
-    
-    // use more bits at the beginning, otherwise high motion at the begin will look like shit
-    s->qsum=100 * s->qmin;
-    s->qcount=100;
-
-    s->short_term_qsum=0.001;
-    s->short_term_qcount=0.001;
 
     return 0;
 }
@@ -106,24 +254,272 @@ void ff_rate_control_uninit(MpegEncContext *s)
     RateControlContext *rcc= &s->rc_context;
     emms_c();
 
-    if(rcc->stats_file) 
-        fclose(rcc->stats_file);
-    rcc->stats_file = NULL;
+    av_free_expr(rcc->rc_eq_eval);
     av_freep(&rcc->entry);
+
+#if CONFIG_LIBXVID
+    if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
+        ff_xvid_rate_control_uninit(s);
+#endif
+}
+
+int ff_vbv_update(MpegEncContext *s, int frame_size){
+    RateControlContext *rcc= &s->rc_context;
+    const double fps= 1/av_q2d(s->avctx->time_base);
+    const int buffer_size= s->avctx->rc_buffer_size;
+    const double min_rate= s->avctx->rc_min_rate/fps;
+    const double max_rate= s->avctx->rc_max_rate/fps;
+
+//printf("%d %f %d %f %f\n", buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
+    if(buffer_size){
+        int left;
+
+        rcc->buffer_index-= frame_size;
+        if(rcc->buffer_index < 0){
+            av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
+            rcc->buffer_index= 0;
+        }
+
+        left= buffer_size - rcc->buffer_index - 1;
+        rcc->buffer_index += av_clip(left, min_rate, max_rate);
+
+        if(rcc->buffer_index > buffer_size){
+            int stuffing= ceil((rcc->buffer_index - buffer_size)/8);
+
+            if(stuffing < 4 && s->codec_id == CODEC_ID_MPEG4)
+                stuffing=4;
+            rcc->buffer_index -= 8*stuffing;
+
+            if(s->avctx->debug & FF_DEBUG_RC)
+                av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
+
+            return stuffing;
+        }
+    }
+    return 0;
+}
+
+/**
+ * modifies the bitrate curve from pass1 for one frame
+ */
+static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num){
+    RateControlContext *rcc= &s->rc_context;
+    AVCodecContext *a= s->avctx;
+    double q, bits;
+    const int pict_type= rce->new_pict_type;
+    const double mb_num= s->mb_num;
+    int i;
+
+    double const_values[]={
+        M_PI,
+        M_E,
+        rce->i_tex_bits*rce->qscale,
+        rce->p_tex_bits*rce->qscale,
+        (rce->i_tex_bits + rce->p_tex_bits)*(double)rce->qscale,
+        rce->mv_bits/mb_num,
+        rce->pict_type == FF_B_TYPE ? (rce->f_code + rce->b_code)*0.5 : rce->f_code,
+        rce->i_count/mb_num,
+        rce->mc_mb_var_sum/mb_num,
+        rce->mb_var_sum/mb_num,
+        rce->pict_type == FF_I_TYPE,
+        rce->pict_type == FF_P_TYPE,
+        rce->pict_type == FF_B_TYPE,
+        rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
+        a->qcompress,
+/*        rcc->last_qscale_for[FF_I_TYPE],
+        rcc->last_qscale_for[FF_P_TYPE],
+        rcc->last_qscale_for[FF_B_TYPE],
+        rcc->next_non_b_qscale,*/
+        rcc->i_cplx_sum[FF_I_TYPE] / (double)rcc->frame_count[FF_I_TYPE],
+        rcc->i_cplx_sum[FF_P_TYPE] / (double)rcc->frame_count[FF_P_TYPE],
+        rcc->p_cplx_sum[FF_P_TYPE] / (double)rcc->frame_count[FF_P_TYPE],
+        rcc->p_cplx_sum[FF_B_TYPE] / (double)rcc->frame_count[FF_B_TYPE],
+        (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
+        0
+    };
+
+    bits = av_eval_expr(rcc->rc_eq_eval, const_values, rce);
+    if (isnan(bits)) {
+        av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq);
+        return -1;
+    }
+
+    rcc->pass1_rc_eq_output_sum+= bits;
+    bits*=rate_factor;
+    if(bits<0.0) bits=0.0;
+    bits+= 1.0; //avoid 1/0 issues
+
+    /* user override */
+    for(i=0; i<s->avctx->rc_override_count; i++){
+        RcOverride *rco= s->avctx->rc_override;
+        if(rco[i].start_frame > frame_num) continue;
+        if(rco[i].end_frame   < frame_num) continue;
+
+        if(rco[i].qscale)
+            bits= qp2bits(rce, rco[i].qscale); //FIXME move at end to really force it?
+        else
+            bits*= rco[i].quality_factor;
+    }
+
+    q= bits2qp(rce, bits);
+
+    /* I/B difference */
+    if     (pict_type==FF_I_TYPE && s->avctx->i_quant_factor<0.0)
+        q= -q*s->avctx->i_quant_factor + s->avctx->i_quant_offset;
+    else if(pict_type==FF_B_TYPE && s->avctx->b_quant_factor<0.0)
+        q= -q*s->avctx->b_quant_factor + s->avctx->b_quant_offset;
+    if(q<1) q=1;
+
+    return q;
+}
+
+static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q){
+    RateControlContext *rcc= &s->rc_context;
+    AVCodecContext *a= s->avctx;
+    const int pict_type= rce->new_pict_type;
+    const double last_p_q    = rcc->last_qscale_for[FF_P_TYPE];
+    const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
+
+    if     (pict_type==FF_I_TYPE && (a->i_quant_factor>0.0 || rcc->last_non_b_pict_type==FF_P_TYPE))
+        q= last_p_q    *FFABS(a->i_quant_factor) + a->i_quant_offset;
+    else if(pict_type==FF_B_TYPE && a->b_quant_factor>0.0)
+        q= last_non_b_q*    a->b_quant_factor  + a->b_quant_offset;
+    if(q<1) q=1;
+
+    /* last qscale / qdiff stuff */
+    if(rcc->last_non_b_pict_type==pict_type || pict_type!=FF_I_TYPE){
+        double last_q= rcc->last_qscale_for[pict_type];
+        const int maxdiff= FF_QP2LAMBDA * 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 cannot do that after blurring
+
+    if(pict_type!=FF_B_TYPE)
+        rcc->last_non_b_pict_type= pict_type;
+
+    return q;
+}
+
+/**
+ * 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->lmin;
+    int qmax= s->avctx->lmax;
+
+    assert(qmin <= qmax);
+
+    if(pict_type==FF_B_TYPE){
+        qmin= (int)(qmin*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
+        qmax= (int)(qmax*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
+    }else if(pict_type==FF_I_TYPE){
+        qmin= (int)(qmin*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
+        qmax= (int)(qmax*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
+    }
+
+    qmin= av_clip(qmin, 1, FF_LAMBDA_MAX);
+    qmax= av_clip(qmax, 1, FF_LAMBDA_MAX);
+
+    if(qmax<qmin) qmax= qmin;
+
+    *qmin_ret= qmin;
+    *qmax_ret= qmax;
+}
+
+static double modify_qscale(MpegEncContext *s, RateControlEntry *rce, double q, int frame_num){
+    RateControlContext *rcc= &s->rc_context;
+    int qmin, qmax;
+    const int pict_type= rce->new_pict_type;
+    const double buffer_size= s->avctx->rc_buffer_size;
+    const double fps= 1/av_q2d(s->avctx->time_base);
+    const double min_rate= s->avctx->rc_min_rate / fps;
+    const double max_rate= s->avctx->rc_max_rate / fps;
+
+    get_qminmax(&qmin, &qmax, s, pict_type);
+
+    /* modulation */
+    if(s->avctx->rc_qmod_freq && frame_num%s->avctx->rc_qmod_freq==0 && pict_type==FF_P_TYPE)
+        q*= s->avctx->rc_qmod_amp;
+
+//printf("q:%f\n", q);
+    /* buffer overflow/underflow protection */
+    if(buffer_size){
+        double expected_size= rcc->buffer_index;
+        double q_limit;
+
+        if(min_rate){
+            double d= 2*(buffer_size - expected_size)/buffer_size;
+            if(d>1.0) d=1.0;
+            else if(d<0.0001) d=0.0001;
+            q*= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
+
+            q_limit= bits2qp(rce, FFMAX((min_rate - buffer_size + rcc->buffer_index) * s->avctx->rc_min_vbv_overflow_use, 1));
+            if(q > q_limit){
+                if(s->avctx->debug&FF_DEBUG_RC){
+                    av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
+                }
+                q= q_limit;
+            }
+        }
+
+        if(max_rate){
+            double d= 2*expected_size/buffer_size;
+            if(d>1.0) d=1.0;
+            else if(d<0.0001) d=0.0001;
+            q/= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
+
+            q_limit= bits2qp(rce, FFMAX(rcc->buffer_index * s->avctx->rc_max_available_vbv_use, 1));
+            if(q < q_limit){
+                if(s->avctx->debug&FF_DEBUG_RC){
+                    av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
+                }
+                q= q_limit;
+            }
+        }
+    }
+//printf("q:%f max:%f min:%f size:%f index:%d bits:%f agr:%f\n", q,max_rate, min_rate, buffer_size, rcc->buffer_index, bits, s->avctx->rc_buffer_aggressivity);
+    if(s->avctx->rc_qsquish==0.0 || qmin==qmax){
+        if     (q<qmin) q=qmin;
+        else if(q>qmax) q=qmax;
+    }else{
+        double min2= log(qmin);
+        double max2= log(qmax);
+
+        q= log(q);
+        q= (q - min2)/(max2-min2) - 0.5;
+        q*= -4.0;
+        q= 1.0/(1.0 + exp(q));
+        q= q*(max2-min2) + min2;
+
+        q= exp(q);
+    }
+
+    return q;
 }
 
 //----------------------------------
 // 1 Pass Code
 
-static double predict(Predictor *p, double q, double var)
+static double predict_size(Predictor *p, double q, double var)
 {
      return p->coeff*var / (q*p->count);
 }
 
+/*
+static double predict_qp(Predictor *p, double size, double var)
+{
+//printf("coeff:%f, count:%f, var:%f, size:%f//\n", p->coeff, p->count, var, size);
+     return p->coeff*var / (size*p->count);
+}
+*/
+
 static void update_predictor(Predictor *p, double q, double var, double size)
 {
     double new_coeff= size*q / (var + 1);
-    if(var<1000) return;
+    if(var<10) return;
 
     p->count*= p->decay;
     p->coeff*= p->decay;
@@ -131,93 +527,294 @@ static void update_predictor(Predictor *p, double q, double var, double size)
     p->coeff+= new_coeff;
 }
 
-int ff_rate_estimate_qscale(MpegEncContext *s)
+static void adaptive_quantization(MpegEncContext *s, double q){
+    int i;
+    const float lumi_masking= s->avctx->lumi_masking / (128.0*128.0);
+    const float dark_masking= s->avctx->dark_masking / (128.0*128.0);
+    const float temp_cplx_masking= s->avctx->temporal_cplx_masking;
+    const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
+    const float p_masking = s->avctx->p_masking;
+    const float border_masking = s->avctx->border_masking;
+    float bits_sum= 0.0;
+    float cplx_sum= 0.0;
+    float cplx_tab[s->mb_num];
+    float bits_tab[s->mb_num];
+    const int qmin= s->avctx->mb_lmin;
+    const int qmax= s->avctx->mb_lmax;
+    Picture * const pic= &s->current_picture;
+    const int mb_width = s->mb_width;
+    const int mb_height = s->mb_height;
+
+    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]); //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;
+        int mb_x = mb_xy % s->mb_stride;
+        int mb_y = mb_xy / s->mb_stride;
+        int mb_distance;
+        float mb_factor = 0.0;
+#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]&CANDIDATE_MB_TYPE_INTRA)){//FIXME hq mode
+            cplx= spat_cplx;
+            factor= 1.0 + p_masking;
+        }else{
+            cplx= temp_cplx;
+            factor= pow(temp_cplx, - temp_cplx_masking);
+        }
+        factor*=pow(spat_cplx, - spatial_cplx_masking);
+
+        if(lumi>127)
+            factor*= (1.0 - (lumi-128)*(lumi-128)*lumi_masking);
+        else
+            factor*= (1.0 - (lumi-128)*(lumi-128)*dark_masking);
+
+        if(mb_x < mb_width/5){
+            mb_distance = mb_width/5 - mb_x;
+            mb_factor = (float)mb_distance / (float)(mb_width/5);
+        }else if(mb_x > 4*mb_width/5){
+            mb_distance = mb_x - 4*mb_width/5;
+            mb_factor = (float)mb_distance / (float)(mb_width/5);
+        }
+        if(mb_y < mb_height/5){
+            mb_distance = mb_height/5 - mb_y;
+            mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
+        }else if(mb_y > 4*mb_height/5){
+            mb_distance = mb_y - 4*mb_height/5;
+            mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
+        }
+
+        factor*= 1.0 - border_masking*mb_factor;
+
+        if(factor<0.00001) factor= 0.00001;
+
+        bits= cplx*factor;
+        cplx_sum+= cplx;
+        bits_sum+= bits;
+        cplx_tab[i]= cplx;
+        bits_tab[i]= bits;
+    }
+
+    /* handle qmin/qmax clipping */
+    if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
+        float factor= bits_sum/cplx_sum;
+        for(i=0; i<s->mb_num; i++){
+            float newq= q*cplx_tab[i]/bits_tab[i];
+            newq*= factor;
+
+            if     (newq > qmax){
+                bits_sum -= bits_tab[i];
+                cplx_sum -= cplx_tab[i]*q/qmax;
+            }
+            else if(newq < qmin){
+                bits_sum -= bits_tab[i];
+                cplx_sum -= cplx_tab[i]*q/qmin;
+            }
+        }
+        if(bits_sum < 0.001) bits_sum= 0.001;
+        if(cplx_sum < 0.001) cplx_sum= 0.001;
+    }
+
+    for(i=0; i<s->mb_num; i++){
+        const int mb_xy= s->mb_index2xy[i];
+        float newq= q*cplx_tab[i]/bits_tab[i];
+        int intq;
+
+        if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
+            newq*= bits_sum/cplx_sum;
+        }
+
+        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]));
+        s->lambda_table[mb_xy]= intq;
+    }
+}
+
+void ff_get_2pass_fcode(MpegEncContext *s){
+    RateControlContext *rcc= &s->rc_context;
+    int picture_number= s->picture_number;
+    RateControlEntry *rce;
+
+    rce= &rcc->entry[picture_number];
+    s->f_code= rce->f_code;
+    s->b_code= rce->b_code;
+}
+
+//FIXME rd or at least approx for dquant
+
+float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
 {
-    int qmin= s->qmin;
-    int qmax= s->qmax;
-    int rate_q=5;
     float q;
-    int qscale;
+    int qmin, qmax;
     float br_compensation;
     double diff;
     double short_term_q;
-    double long_term_q;
     double fps;
-    int picture_number= s->input_picture_number - s->max_b_frames;
+    int picture_number= s->picture_number;
     int64_t wanted_bits;
+    RateControlContext *rcc= &s->rc_context;
+    AVCodecContext *a= s->avctx;
+    RateControlEntry local_rce, *rce;
+    double bits;
+    double rate_factor;
+    int var;
+    const int pict_type= s->pict_type;
+    Picture * const pic= &s->current_picture;
     emms_c();
 
-    fps= (double)s->frame_rate / FRAME_RATE_BASE;
-    wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps);
-//    printf("%d %d %d\n", picture_number, (int)wanted_bits, (int)s->total_bits);
-    
-    if(s->pict_type==B_TYPE){
-        qmin= (int)(qmin*s->b_quant_factor+0.5);
-        qmax= (int)(qmax*s->b_quant_factor+0.5);
-    }
-    if(qmin<1) qmin=1;
-    if(qmax>31) qmax=31;
-    if(qmax<=qmin) qmax= qmin;
+#if CONFIG_LIBXVID
+    if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
+        return ff_xvid_rate_estimate_qscale(s, dry_run);
+#endif
 
+    get_qminmax(&qmin, &qmax, s, pict_type);
+
+    fps= 1/av_q2d(s->avctx->time_base);
+//printf("input_pic_num:%d pic_num:%d frame_rate:%d\n", s->input_picture_number, s->picture_number, s->frame_rate);
         /* update predictors */
-    if(picture_number>2){
-        if(s->pict_type!=B_TYPE && s->last_non_b_pict_type == P_TYPE){
-//printf("%d %d %d %f\n", s->qscale, s->last_mc_mb_var, s->frame_bits, s->p_pred.coeff);
-            update_predictor(&s->p_pred, s->last_non_b_qscale, s->last_non_b_mc_mb_var, s->pb_frame_bits);
-        }
+    if(picture_number>2 && !dry_run){
+        const int last_var= s->last_pict_type == FF_I_TYPE ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum;
+        update_predictor(&rcc->pred[s->last_pict_type], rcc->last_qscale, sqrt(last_var), s->frame_bits);
     }
 
-    if(s->pict_type == I_TYPE){
-        short_term_q= s->short_term_qsum/s->short_term_qcount;
-    
-        long_term_q= s->qsum/s->qcount*(s->total_bits+1)/(wanted_bits+1); //+1 to avoid nan & 0
+    if(s->flags&CODEC_FLAG_PASS2){
+        assert(picture_number>=0);
+        assert(picture_number<rcc->num_entries);
+        rce= &rcc->entry[picture_number];
+        wanted_bits= rce->expected_bits;
+    }else{
+        Picture *dts_pic;
+        rce= &local_rce;
+
+        //FIXME add a dts field to AVFrame and ensure its set and use it here instead of reordering
+        //but the reordering is simpler for now until h.264 b pyramid must be handeld
+        if(s->pict_type == FF_B_TYPE || s->low_delay)
+            dts_pic= s->current_picture_ptr;
+        else
+            dts_pic= s->last_picture_ptr;
+
+//if(dts_pic)
+//            av_log(NULL, AV_LOG_ERROR, "%Ld %Ld %Ld %d\n", s->current_picture_ptr->pts, s->user_specified_pts, dts_pic->pts, picture_number);
+
+        if(!dts_pic || dts_pic->pts == AV_NOPTS_VALUE)
+            wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps);
+        else
+            wanted_bits= (uint64_t)(s->bit_rate*(double)dts_pic->pts/fps);
+    }
 
-        q= 1/((1/long_term_q - 1/short_term_q)*s->qcompress + 1/short_term_q);
-    }else if(s->pict_type==B_TYPE){
-        q= (int)(s->last_non_b_qscale*s->b_quant_factor+0.5);
-    }else{ //P Frame
-        int i;
-        int diff, best_diff=1000000000;
-        for(i=1; i<=31; i++){
-            diff= predict(&s->p_pred, i, s->mc_mb_var) - (double)s->bit_rate/fps;
-            if(diff<0) diff= -diff;
-            if(diff<best_diff){
-                best_diff= diff;
-                rate_q= i;
-            }
+    diff= s->total_bits - wanted_bits;
+    br_compensation= (a->bit_rate_tolerance - diff)/a->bit_rate_tolerance;
+    if(br_compensation<=0.0) br_compensation=0.001;
+
+    var= pict_type == FF_I_TYPE ? pic->mb_var_sum : pic->mc_mb_var_sum;
+
+    short_term_q = 0; /* avoid warning */
+    if(s->flags&CODEC_FLAG_PASS2){
+        if(pict_type!=FF_I_TYPE)
+            assert(pict_type == rce->new_pict_type);
+
+        q= rce->new_qscale / br_compensation;
+//printf("%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale, br_compensation, s->frame_bits, var, pict_type);
+    }else{
+        rce->pict_type=
+        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   = FF_QP2LAMBDA * 2;
+        rce->f_code   = s->f_code;
+        rce->b_code   = s->b_code;
+        rce->misc_bits= 1;
+
+        bits= predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
+        if(pict_type== FF_I_TYPE){
+            rce->i_count   = s->mb_num;
+            rce->i_tex_bits= bits;
+            rce->p_tex_bits= 0;
+            rce->mv_bits= 0;
+        }else{
+            rce->i_count   = 0; //FIXME we do know this approx
+            rce->i_tex_bits= 0;
+            rce->p_tex_bits= bits*0.9;
+
+            rce->mv_bits= bits*0.1;
         }
-        s->short_term_qsum*=s->qblur;
-        s->short_term_qcount*=s->qblur;
+        rcc->i_cplx_sum [pict_type] += rce->i_tex_bits*rce->qscale;
+        rcc->p_cplx_sum [pict_type] += rce->p_tex_bits*rce->qscale;
+        rcc->mv_bits_sum[pict_type] += rce->mv_bits;
+        rcc->frame_count[pict_type] ++;
 
-        s->short_term_qsum+= rate_q;
-        s->short_term_qcount++;
-        short_term_q= s->short_term_qsum/s->short_term_qcount;
-    
-        long_term_q= s->qsum/s->qcount*(s->total_bits+1)/(wanted_bits+1); //+1 to avoid nan & 0
+        bits= rce->i_tex_bits + rce->p_tex_bits;
+        rate_factor= rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum * br_compensation;
 
-//    q= (long_term_q - short_term_q)*s->qcompress + short_term_q;
-        q= 1/((1/long_term_q - 1/short_term_q)*s->qcompress + 1/short_term_q);
+        q= get_qscale(s, rce, rate_factor, picture_number);
+        if (q < 0)
+            return -1;
+
+        assert(q>0.0);
+//printf("%f ", q);
+        q= get_diff_limited_q(s, rce, q);
+//printf("%f ", q);
+        assert(q>0.0);
+
+        if(pict_type==FF_P_TYPE || s->intra_only){ //FIXME type dependent blur like in 2-pass
+            rcc->short_term_qsum*=a->qblur;
+            rcc->short_term_qcount*=a->qblur;
+
+            rcc->short_term_qsum+= q;
+            rcc->short_term_qcount++;
+//printf("%f ", q);
+            q= short_term_q= rcc->short_term_qsum/rcc->short_term_qcount;
+//printf("%f ", q);
+        }
+        assert(q>0.0);
+
+        q= modify_qscale(s, rce, q, picture_number);
+
+        rcc->pass1_wanted_bits+= s->bit_rate/fps;
+
+        assert(q>0.0);
     }
 
-    diff= s->total_bits - wanted_bits;
-    br_compensation= (s->bit_rate_tolerance - diff)/s->bit_rate_tolerance;
-    if(br_compensation<=0.0) br_compensation=0.001;
-    q/=br_compensation;
-//printf("%f %f %f\n", q, br_compensation, short_term_q);
-    qscale= (int)(q + 0.5);
-    if     (qscale<qmin) qscale=qmin;
-    else if(qscale>qmax) qscale=qmax;
-    
-    if(s->pict_type!=B_TYPE){
-        s->qsum+= qscale;
-        s->qcount++;
-        if     (qscale<s->last_non_b_qscale-s->max_qdiff) qscale=s->last_non_b_qscale-s->max_qdiff;
-        else if(qscale>s->last_non_b_qscale+s->max_qdiff) qscale=s->last_non_b_qscale+s->max_qdiff;
+    if(s->avctx->debug&FF_DEBUG_RC){
+        av_log(s->avctx, AV_LOG_DEBUG, "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f size:%d var:%d/%d br:%d fps:%d\n",
+        av_get_pict_type_char(pict_type), qmin, q, qmax, picture_number, (int)wanted_bits/1000, (int)s->total_bits/1000,
+        br_compensation, short_term_q, s->frame_bits, pic->mb_var_sum, pic->mc_mb_var_sum, s->bit_rate/1000, (int)fps
+        );
+    }
+
+    if     (q<qmin) q=qmin;
+    else if(q>qmax) q=qmax;
+
+    if(s->adaptive_quant)
+        adaptive_quantization(s, q);
+    else
+        q= (int)(q + 0.5);
+
+    if(!dry_run){
+        rcc->last_qscale= q;
+        rcc->last_mc_mb_var_sum= pic->mc_mb_var_sum;
+        rcc->last_mb_var_sum= pic->mb_var_sum;
     }
-//printf("q:%d diff:%d comp:%f rate_q:%d st_q:%f fvar:%d last_size:%d\n", qscale, (int)diff, br_compensation, 
-//       rate_q, short_term_q, s->mc_mb_var, s->frame_bits);
-//printf("%d %d\n", s->bit_rate, (int)fps);
-    return qscale;
+#if 0
+{
+    static int mvsum=0, texsum=0;
+    mvsum += s->mv_bits;
+    texsum += s->i_tex_bits + s->p_tex_bits;
+    printf("%d %d//\n\n", mvsum, texsum);
+}
+#endif
+    return q;
 }
 
 //----------------------------------------------
@@ -226,180 +823,138 @@ int ff_rate_estimate_qscale(MpegEncContext *s)
 static int init_pass2(MpegEncContext *s)
 {
     RateControlContext *rcc= &s->rc_context;
-    int i;
-    double fps= (double)s->frame_rate / FRAME_RATE_BASE;
+    AVCodecContext *a= s->avctx;
+    int i, toobig;
+    double fps= 1/av_q2d(s->avctx->time_base);
     double complexity[5]={0,0,0,0,0};   // aproximate bits at quant=1
-    double avg_quantizer[5];
-    uint64_t const_bits[5]={0,0,0,0,0}; // quantizer idependant bits
-    uint64_t available_bits[5];
+    uint64_t const_bits[5]={0,0,0,0,0}; // quantizer independent bits
     uint64_t all_const_bits;
     uint64_t all_available_bits= (uint64_t)(s->bit_rate*(double)rcc->num_entries/fps);
-    int num_frames[5]={0,0,0,0,0};
     double rate_factor=0;
     double step;
-    int last_i_frame=-10000000;
+    //int last_i_frame=-10000000;
+    const int filter_size= (int)(a->qblur*4) | 1;
+    double expected_bits;
+    double *qscale, *blurred_qscale, qscale_sum;
 
     /* find complexity & const_bits & decide the pict_types */
     for(i=0; i<rcc->num_entries; i++){
         RateControlEntry *rce= &rcc->entry[i];
-        
-        if(s->b_frame_strategy==0 || s->max_b_frames==0){
-            rce->new_pict_type= rce->pict_type;
-        }else{
-            int j;
-            int next_non_b_type=P_TYPE;
 
-            switch(rce->pict_type){
-            case I_TYPE:
-                if(i-last_i_frame>s->gop_size/2){ //FIXME this is not optimal
-                    rce->new_pict_type= I_TYPE;
-                    last_i_frame= i;
-                }else{
-                    rce->new_pict_type= P_TYPE; // will be caught by the scene detection anyway
-                }
-                break;
-            case P_TYPE:
-                rce->new_pict_type= P_TYPE;
-                break;
-            case B_TYPE:
-                for(j=i+1; j<i+s->max_b_frames+2 && j<rcc->num_entries; j++){
-                    if(rcc->entry[j].pict_type != B_TYPE){
-                        next_non_b_type= rcc->entry[j].pict_type;
-                        break;
-                    }
-                }
-                if(next_non_b_type==I_TYPE)
-                    rce->new_pict_type= P_TYPE;
-                else
-                    rce->new_pict_type= B_TYPE;
-                break;
-            }
-        }
+        rce->new_pict_type= rce->pict_type;
+        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;
+        rcc->frame_count[rce->pict_type] ++;
 
         complexity[rce->new_pict_type]+= (rce->i_tex_bits+ rce->p_tex_bits)*(double)rce->qscale;
         const_bits[rce->new_pict_type]+= rce->mv_bits + rce->misc_bits;
-        num_frames[rce->new_pict_type]++;
     }
-    all_const_bits= const_bits[I_TYPE] + const_bits[P_TYPE] + const_bits[B_TYPE];
-    
+    all_const_bits= const_bits[FF_I_TYPE] + const_bits[FF_P_TYPE] + const_bits[FF_B_TYPE];
+
     if(all_available_bits < all_const_bits){
-        fprintf(stderr, "requested bitrate is to low\n");
+        av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
         return -1;
     }
 
-//    avg_complexity= complexity/rcc->num_entries;
-    avg_quantizer[P_TYPE]= 
-    avg_quantizer[I_TYPE]=   (complexity[I_TYPE]+complexity[P_TYPE] + complexity[B_TYPE]/s->b_quant_factor) 
-                           / (all_available_bits - all_const_bits);
-    avg_quantizer[B_TYPE]= avg_quantizer[P_TYPE]*s->b_quant_factor;
-//printf("avg quantizer: %f %f\n", avg_quantizer[P_TYPE], avg_quantizer[B_TYPE]);
+    qscale= av_malloc(sizeof(double)*rcc->num_entries);
+    blurred_qscale= av_malloc(sizeof(double)*rcc->num_entries);
+    toobig = 0;
 
-    for(i=0; i<5; i++){
-        available_bits[i]= const_bits[i] + complexity[i]/avg_quantizer[i];
-    }
-//printf("%lld %lld %lld %lld\n", available_bits[I_TYPE], available_bits[P_TYPE], available_bits[B_TYPE], all_available_bits);
-    
     for(step=256*256; step>0.0000001; step*=0.5){
-        uint64_t expected_bits=0;
+        expected_bits=0;
         rate_factor+= step;
+
+        rcc->buffer_index= s->avctx->rc_buffer_size/2;
+
         /* find qscale */
         for(i=0; i<rcc->num_entries; i++){
+            qscale[i]= get_qscale(s, &rcc->entry[i], rate_factor, i);
+        }
+        assert(filter_size%2==1);
+
+        /* fixed I/B QP relative to P mode */
+        for(i=rcc->num_entries-1; i>=0; i--){
             RateControlEntry *rce= &rcc->entry[i];
-            double short_term_q, q, bits_left;
-            const int pict_type= rce->new_pict_type;
-            int qmin= s->qmin;
-            int qmax= s->qmax;
 
-            if(pict_type==B_TYPE){
-                qmin= (int)(qmin*s->b_quant_factor+0.5);
-                qmax= (int)(qmax*s->b_quant_factor+0.5);
-            }
-            if(qmin<1) qmin=1;
-            if(qmax>31) qmax=31;
-            if(qmax<=qmin) qmax= qmin;
-            
-            switch(s->rc_strategy){
-            case 0:
-                bits_left= available_bits[pict_type]/num_frames[pict_type]*rate_factor - rce->misc_bits - rce->mv_bits;
-                if(bits_left<1.0) bits_left=1.0;
-                short_term_q= rce->qscale*(rce->i_tex_bits + rce->p_tex_bits)/bits_left;
-                break;
-            case 1:
-                bits_left= (available_bits[pict_type] - const_bits[pict_type])/num_frames[pict_type]*rate_factor;
-                if(bits_left<1.0) bits_left=1.0;
-                short_term_q= rce->qscale*(rce->i_tex_bits + rce->p_tex_bits)/bits_left;
-                break;
-            case 2:
-                bits_left= available_bits[pict_type]/num_frames[pict_type]*rate_factor;
-                if(bits_left<1.0) bits_left=1.0;
-                short_term_q= rce->qscale*(rce->i_tex_bits + rce->p_tex_bits + rce->misc_bits + rce->mv_bits)/bits_left;
-                break;
-            default:
-                fprintf(stderr, "unknown strategy\n");
-                short_term_q=3; //gcc warning fix
-            }
+            qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
+        }
 
-            if(short_term_q>31.0) short_term_q=31.0;
-            else if (short_term_q<1.0) short_term_q=1.0;
+        /* smooth curve */
+        for(i=0; i<rcc->num_entries; i++){
+            RateControlEntry *rce= &rcc->entry[i];
+            const int pict_type= rce->new_pict_type;
+            int j;
+            double q=0.0, sum=0.0;
 
-            q= 1/((1/avg_quantizer[pict_type] - 1/short_term_q)*s->qcompress + 1/short_term_q);
-            if     (q<qmin) q=qmin;
-            else if(q>qmax) q=qmax;
-//printf("lq:%f, sq:%f t:%f q:%f\n", avg_quantizer[rce->pict_type], short_term_q, bits_left, q);
-            rce->new_qscale= q;
+            for(j=0; j<filter_size; j++){
+                int index= i+j-filter_size/2;
+                double d= index-i;
+                double coeff= a->qblur==0 ? 1.0 : exp(-d*d/(a->qblur * a->qblur));
+
+                if(index < 0 || index >= rcc->num_entries) continue;
+                if(pict_type != rcc->entry[index].new_pict_type) continue;
+                q+= qscale[index] * coeff;
+                sum+= coeff;
+            }
+            blurred_qscale[i]= q/sum;
         }
 
-        /* smooth curve */
-    
         /* find expected bits */
         for(i=0; i<rcc->num_entries; i++){
             RateControlEntry *rce= &rcc->entry[i];
-            double factor= rce->qscale / rce->new_qscale;
-            
+            double bits;
+            rce->new_qscale= modify_qscale(s, rce, blurred_qscale[i], i);
+            bits= qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
+//printf("%d %f\n", rce->new_bits, blurred_qscale[i]);
+            bits += 8*ff_vbv_update(s, bits);
+
             rce->expected_bits= expected_bits;
-            expected_bits += (int)(rce->misc_bits + rce->mv_bits + (rce->i_tex_bits + rce->p_tex_bits)*factor + 0.5);
+            expected_bits += bits;
         }
 
-//        printf("%d %d %f\n", (int)expected_bits, (int)all_available_bits, rate_factor);
-        if(expected_bits > all_available_bits) rate_factor-= step;
+        /*
+        av_log(s->avctx, AV_LOG_INFO,
+            "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
+            expected_bits, (int)all_available_bits, rate_factor);
+        */
+        if(expected_bits > all_available_bits) {
+            rate_factor-= step;
+            ++toobig;
+        }
     }
+    av_free(qscale);
+    av_free(blurred_qscale);
 
-    return 0;
-}
-
-int ff_rate_estimate_qscale_pass2(MpegEncContext *s)
-{
-    int qmin= s->qmin;
-    int qmax= s->qmax;
-    float q;
-    int qscale;
-    float br_compensation;
-    double diff;
-    int picture_number= s->picture_number;
-    RateControlEntry *rce= &s->rc_context.entry[picture_number];
-    int64_t wanted_bits= rce->expected_bits;
-    emms_c();
-
-//    printf("%d %d %d\n", picture_number, (int)wanted_bits, (int)s->total_bits);
-    
-    if(s->pict_type==B_TYPE){
-        qmin= (int)(qmin*s->b_quant_factor+0.5);
-        qmax= (int)(qmax*s->b_quant_factor+0.5);
+    /* check bitrate calculations and print info */
+    qscale_sum = 0.0;
+    for(i=0; i<rcc->num_entries; i++){
+        /* av_log(s->avctx, AV_LOG_DEBUG, "[lavc rc] entry[%d].new_qscale = %.3f  qp = %.3f\n",
+            i, rcc->entry[i].new_qscale, rcc->entry[i].new_qscale / FF_QP2LAMBDA); */
+        qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA, s->avctx->qmin, s->avctx->qmax);
+    }
+    assert(toobig <= 40);
+    av_log(s->avctx, AV_LOG_DEBUG,
+        "[lavc rc] requested bitrate: %d bps  expected bitrate: %d bps\n",
+        s->bit_rate,
+        (int)(expected_bits / ((double)all_available_bits/s->bit_rate)));
+    av_log(s->avctx, AV_LOG_DEBUG,
+        "[lavc rc] estimated target average qp: %.3f\n",
+        (float)qscale_sum / rcc->num_entries);
+    if (toobig == 0) {
+        av_log(s->avctx, AV_LOG_INFO,
+            "[lavc rc] Using all of requested bitrate is not "
+            "necessary for this video with these parameters.\n");
+    } else if (toobig == 40) {
+        av_log(s->avctx, AV_LOG_ERROR,
+            "[lavc rc] Error: bitrate too low for this video "
+            "with these parameters.\n");
+        return -1;
+    } else if (fabs(expected_bits/all_available_bits - 1.0) > 0.01) {
+        av_log(s->avctx, AV_LOG_ERROR,
+            "[lavc rc] Error: 2pass curve failed to converge\n");
+        return -1;
     }
-    if(qmin<1) qmin=1;
-    if(qmax>31) qmax=31;
-    if(qmax<=qmin) qmax= qmin;
-
-    q= rce->new_qscale;
-
-    diff= s->total_bits - wanted_bits;
-    br_compensation= (s->bit_rate_tolerance - diff)/s->bit_rate_tolerance;
-    if(br_compensation<=0.0) br_compensation=0.001;
-    q/=br_compensation;
 
-    qscale= (int)(q + 0.5);
-    if     (qscale<qmin) qscale=qmin;
-    else if(qscale>qmax) qscale=qmax;
-//    printf("%d %d %d %d type:%d\n", qmin, qscale, qmax, picture_number, s->pict_type); fflush(stdout);
-    return qscale;
+    return 0;
 }