ratecontrol: fix get_fps() argument
[ffmpeg.git] / libavcodec / ratecontrol.c
index 196ba18..bc399b4 100644 (file)
@@ -3,20 +3,20 @@
  *
  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
  *
- * This file is part of Libav.
+ * This file is part of FFmpeg.
  *
- * Libav is free software; you can redistribute it and/or
+ * 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.
  *
- * Libav is distributed in the hope that it will be useful,
+ * 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 Libav; if not, write to the Free Software
+ * License along with FFmpeg; if not, write to the Free Software
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
@@ -44,9 +44,13 @@ static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_f
 
 void ff_write_pass1_stats(MpegEncContext *s){
     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);
+             s->current_picture_ptr->f.display_picture_number, s->current_picture_ptr->f.coded_picture_number, s->pict_type,
+             s->current_picture.f.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 double get_fps(AVCodecContext *avctx){
+    return 1.0 / av_q2d(avctx->time_base) / FFMAX(avctx->ticks_per_frame, 1);
 }
 
 static inline double qp2bits(RateControlEntry *rce, double qp){
@@ -106,6 +110,13 @@ int ff_rate_control_init(MpegEncContext *s)
     };
     emms_c();
 
+    if (!s->avctx->rc_max_available_vbv_use && s->avctx->rc_buffer_size) {
+        if (s->avctx->rc_max_rate) {
+            s->avctx->rc_max_available_vbv_use = av_clipf(s->avctx->rc_max_rate/(s->avctx->rc_buffer_size*get_fps(s->avctx)), 1.0/0.3, 1.0);
+        } else
+            s->avctx->rc_max_available_vbv_use = 1.0;
+    }
+
     res = av_expr_parse(&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);
@@ -125,6 +136,8 @@ int ff_rate_control_init(MpegEncContext *s)
         rcc->last_qscale_for[i]=FF_QP2LAMBDA * 5;
     }
     rcc->buffer_index= s->avctx->rc_initial_buffer_occupancy;
+    if (!rcc->buffer_index)
+        rcc->buffer_index = s->avctx->rc_buffer_size * 3 / 4;
 
     if(s->flags&CODEC_FLAG_PASS2){
         int i;
@@ -144,7 +157,7 @@ int ff_rate_control_init(MpegEncContext *s)
         /* 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->pict_type= rce->new_pict_type=AV_PICTURE_TYPE_P;
             rce->qscale= rce->new_qscale=FF_QP2LAMBDA * 2;
             rce->misc_bits= s->mb_num + 10;
             rce->mb_var_sum= s->mb_num*100;
@@ -211,9 +224,9 @@ int ff_rate_control_init(MpegEncContext *s)
                 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;
+                if     (i%((s->gop_size+3)/4)==0) rce.pict_type= AV_PICTURE_TYPE_I;
+                else if(i%(s->max_b_frames+1))    rce.pict_type= AV_PICTURE_TYPE_B;
+                else                              rce.pict_type= AV_PICTURE_TYPE_P;
 
                 rce.new_pict_type= rce.pict_type;
                 rce.mc_mb_var_sum= bits*s->mb_num/100000;
@@ -223,7 +236,7 @@ int ff_rate_control_init(MpegEncContext *s)
                 rce.b_code   = 1;
                 rce.misc_bits= 1;
 
-                if(s->pict_type== FF_I_TYPE){
+                if(s->pict_type== AV_PICTURE_TYPE_I){
                     rce.i_count   = s->mb_num;
                     rce.i_tex_bits= bits;
                     rce.p_tex_bits= 0;
@@ -240,7 +253,7 @@ int ff_rate_control_init(MpegEncContext *s)
                 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
+                rcc->pass1_wanted_bits+= s->bit_rate/get_fps(s->avctx); //FIXME misbehaves a little for variable fps
             }
         }
 
@@ -265,7 +278,7 @@ void ff_rate_control_uninit(MpegEncContext *s)
 
 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 double fps= get_fps(s->avctx);
     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;
@@ -286,7 +299,7 @@ int ff_vbv_update(MpegEncContext *s, int frame_size){
         if(rcc->buffer_index > buffer_size){
             int stuffing= ceil((rcc->buffer_index - buffer_size)/8);
 
-            if(stuffing < 4 && s->codec_id == CODEC_ID_MPEG4)
+            if(stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
                 stuffing=4;
             rcc->buffer_index -= 8*stuffing;
 
@@ -300,7 +313,7 @@ int ff_vbv_update(MpegEncContext *s, int frame_size){
 }
 
 /**
- * modifies the bitrate curve from pass1 for one frame
+ * Modify 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;
@@ -317,23 +330,23 @@ static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_f
         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->pict_type == AV_PICTURE_TYPE_B ? (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,
+        rce->pict_type == AV_PICTURE_TYPE_I,
+        rce->pict_type == AV_PICTURE_TYPE_P,
+        rce->pict_type == AV_PICTURE_TYPE_B,
         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->last_qscale_for[AV_PICTURE_TYPE_I],
+        rcc->last_qscale_for[AV_PICTURE_TYPE_P],
+        rcc->last_qscale_for[AV_PICTURE_TYPE_B],
         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[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
+        rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
+        rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
+        rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
         (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
         0
     };
@@ -364,9 +377,9 @@ static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_f
     q= bits2qp(rce, bits);
 
     /* I/B difference */
-    if     (pict_type==FF_I_TYPE && s->avctx->i_quant_factor<0.0)
+    if     (pict_type==AV_PICTURE_TYPE_I && 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)
+    else if(pict_type==AV_PICTURE_TYPE_B && s->avctx->b_quant_factor<0.0)
         q= -q*s->avctx->b_quant_factor + s->avctx->b_quant_offset;
     if(q<1) q=1;
 
@@ -377,17 +390,17 @@ static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, doubl
     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_p_q    = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
     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))
+    if     (pict_type==AV_PICTURE_TYPE_I && (a->i_quant_factor>0.0 || rcc->last_non_b_pict_type==AV_PICTURE_TYPE_P))
         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)
+    else if(pict_type==AV_PICTURE_TYPE_B && 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){
+    if(rcc->last_non_b_pict_type==pict_type || pict_type!=AV_PICTURE_TYPE_I){
         double last_q= rcc->last_qscale_for[pict_type];
         const int maxdiff= FF_QP2LAMBDA * a->max_qdiff;
 
@@ -397,14 +410,14 @@ static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, doubl
 
     rcc->last_qscale_for[pict_type]= q; //Note we cannot do that after blurring
 
-    if(pict_type!=FF_B_TYPE)
+    if(pict_type!=AV_PICTURE_TYPE_B)
         rcc->last_non_b_pict_type= pict_type;
 
     return q;
 }
 
 /**
- * gets the qmin & qmax for pict_type
+ * Get 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;
@@ -412,10 +425,10 @@ static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pic
 
     assert(qmin <= qmax);
 
-    if(pict_type==FF_B_TYPE){
+    if(pict_type==AV_PICTURE_TYPE_B){
         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){
+    }else if(pict_type==AV_PICTURE_TYPE_I){
         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);
     }
@@ -434,14 +447,14 @@ static double modify_qscale(MpegEncContext *s, RateControlEntry *rce, double q,
     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 fps= get_fps(s->avctx);
     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)
+    if(s->avctx->rc_qmod_freq && frame_num%s->avctx->rc_qmod_freq==0 && pict_type==AV_PICTURE_TYPE_P)
         q*= s->avctx->rc_qmod_amp;
 
 //printf("q:%f\n", q);
@@ -508,14 +521,6 @@ 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);
@@ -537,8 +542,8 @@ static void adaptive_quantization(MpegEncContext *s, double q){
     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];
+    float *cplx_tab = s->cplx_tab;
+    float *bits_tab = s->bits_tab;
     const int qmin= s->avctx->mb_lmin;
     const int qmax= s->avctx->mb_lmax;
     Picture * const pic= &s->current_picture;
@@ -555,10 +560,6 @@ static void adaptive_quantization(MpegEncContext *s, double q){
         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
 
@@ -680,12 +681,13 @@ float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
 
     get_qminmax(&qmin, &qmax, s, pict_type);
 
-    fps= 1/av_q2d(s->avctx->time_base);
+    fps= get_fps(s->avctx);
 //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 && !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);
+        const int last_var= s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum;
+        av_assert1(s->frame_bits >= s->stuffing_bits);
+        update_predictor(&rcc->pred[s->last_pict_type], rcc->last_qscale, sqrt(last_var), s->frame_bits - s->stuffing_bits);
     }
 
     if(s->flags&CODEC_FLAG_PASS2){
@@ -699,29 +701,29 @@ float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
 
         //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)
+        if(s->pict_type == AV_PICTURE_TYPE_B || 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);
+//            av_log(NULL, AV_LOG_ERROR, "%"PRId64" %"PRId64" %"PRId64" %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)
+        if (!dts_pic || dts_pic->f.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);
+            wanted_bits = (uint64_t)(s->bit_rate*(double)dts_pic->f.pts / fps);
     }
 
     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;
+    var= pict_type == AV_PICTURE_TYPE_I ? 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)
+        if(pict_type!=AV_PICTURE_TYPE_I)
             assert(pict_type == rce->new_pict_type);
 
         q= rce->new_qscale / br_compensation;
@@ -737,7 +739,7 @@ float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
         rce->misc_bits= 1;
 
         bits= predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
-        if(pict_type== FF_I_TYPE){
+        if(pict_type== AV_PICTURE_TYPE_I){
             rce->i_count   = s->mb_num;
             rce->i_tex_bits= bits;
             rce->p_tex_bits= 0;
@@ -767,7 +769,7 @@ float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
 //printf("%f ", q);
         assert(q>0.0);
 
-        if(pict_type==FF_P_TYPE || s->intra_only){ //FIXME type dependent blur like in 2-pass
+        if(pict_type==AV_PICTURE_TYPE_P || s->intra_only){ //FIXME type dependent blur like in 2-pass
             rcc->short_term_qsum*=a->qblur;
             rcc->short_term_qcount*=a->qblur;
 
@@ -817,7 +819,7 @@ static int init_pass2(MpegEncContext *s)
     RateControlContext *rcc= &s->rc_context;
     AVCodecContext *a= s->avctx;
     int i, toobig;
-    double fps= 1/av_q2d(s->avctx->time_base);
+    double fps= get_fps(s->avctx);
     double complexity[5]={0,0,0,0,0};   // aproximate bits at quant=1
     uint64_t const_bits[5]={0,0,0,0,0}; // quantizer independent bits
     uint64_t all_const_bits;
@@ -842,7 +844,7 @@ static int init_pass2(MpegEncContext *s)
         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;
     }
-    all_const_bits= const_bits[FF_I_TYPE] + const_bits[FF_P_TYPE] + const_bits[FF_B_TYPE];
+    all_const_bits= const_bits[AV_PICTURE_TYPE_I] + const_bits[AV_PICTURE_TYPE_P] + const_bits[AV_PICTURE_TYPE_B];
 
     if(all_available_bits < all_const_bits){
         av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
@@ -861,11 +863,19 @@ static int init_pass2(MpegEncContext *s)
 
         /* find qscale */
         for(i=0; i<rcc->num_entries; i++){
+            RateControlEntry *rce= &rcc->entry[i];
             qscale[i]= get_qscale(s, &rcc->entry[i], rate_factor, i);
+            rcc->last_qscale_for[rce->pict_type] = qscale[i];
         }
         assert(filter_size%2==1);
 
         /* fixed I/B QP relative to P mode */
+        for(i=FFMAX(0, rcc->num_entries-300); i<rcc->num_entries; i++){
+            RateControlEntry *rce= &rcc->entry[i];
+
+            qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
+        }
+
         for(i=rcc->num_entries-1; i>=0; i--){
             RateControlEntry *rce= &rcc->entry[i];