reuse motion vectors/mb types/field select values of the source video, if the...

reuse motion vectors/mb types/field select values of the source video, if the SSE for a macroblock which is predicted with these values is below me_threshold
currently works only with mpeg1/2 source or some luck
may need -sync 0 as otherwise framedrops could lead to extreemly long b frame sequences

Originally committed as revision 3042 to svn://svn.ffmpeg.org/ffmpeg/trunk

ffmpeg.c

@@ -159,6 +159,7 @@ static int noise_reduction = 0;
 static int sc_threshold = 0;
 static int debug = 0;
 static int debug_mv = 0;
+static int me_threshold = 0;
 extern int loop_input; /* currently a hack */
 
 static int gop_size = 12;
@@ -187,6 +188,7 @@ static int bitexact = 0;
 static char *pass_logfilename = NULL;
 static int audio_stream_copy = 0;
 static int video_stream_copy = 0;
+static int sync_method= 1;
 
 static int rate_emu = 0;
 
@@ -577,6 +579,7 @@ static void do_video_out(AVFormatContext *s,
     /* NOTE: the A/V sync is always done by considering the audio is
        the master clock. It is suffisant for transcoding or playing,
        but not for the general case */
+  if(sync_method){
     if (audio_sync) {
         /* compute the A-V delay and duplicate/remove frames if needed */
         double adelta, vdelta, av_delay;
@@ -588,11 +591,11 @@ static void do_video_out(AVFormatContext *s,
             s->pts_num / s->pts_den);
 
         av_delay = adelta - vdelta;
-        //            printf("delay=%f\n", av_delay);
         if (av_delay < -AV_DELAY_MAX)
             nb_frames = 2;
         else if (av_delay > AV_DELAY_MAX)
             nb_frames = 0;
+//        printf("delay=%f nb=%d (A)\n", av_delay, nb_frames);
     } else {
         double vdelta;
 
@@ -607,7 +610,9 @@ static void do_video_out(AVFormatContext *s,
             if (!ost->sync_ipts_offset)
                 ost->sync_ipts_offset = 0.000001; /* one microsecond */
         }
+//        printf("delay=%f nb=%d (V)\n",vdelta, nb_frames);
     }
+  }
     
 #if defined(AVSYNC_DEBUG)
     {
@@ -793,7 +798,8 @@ static void do_video_out(AVFormatContext *s,
                 big_picture.quality = ist->st->quality;
             }else
                 big_picture.quality = ost->st->quality;
-            big_picture.pict_type = 0;
+            if(!me_threshold)
+                big_picture.pict_type = 0;
             big_picture.pts = AV_NOPTS_VALUE; //FIXME
             ret = avcodec_encode_video(enc, 
                                        video_buffer, VIDEO_BUFFER_SIZE,
@@ -1853,6 +1859,11 @@ static void opt_idct_algo(const char *arg)
     idct_algo = atoi(arg);
 }
 
+static void opt_me_threshold(const char *arg)
+{
+    me_threshold = atoi(arg);
+}
+
 
 static void opt_error_resilience(const char *arg)
 {
@@ -1880,6 +1891,11 @@ static void opt_verbose(const char *arg)
     av_log_set_level(atoi(arg));
 }
 
+static void opt_sync_method(const char *arg)
+{
+    sync_method = atoi(arg);
+}
+
 static void opt_frame_rate(const char *arg)
 {
     if (parse_frame_rate(&frame_rate, &frame_rate_base, arg) < 0) {
@@ -2554,6 +2570,8 @@ static void opt_input_file(const char *filename)
             enc->debug_mv = debug_mv;            
             if(bitexact)
                 enc->flags|= CODEC_FLAG_BITEXACT;
+            if(me_threshold)
+                enc->debug |= FF_DEBUG_MV;
 
             assert(enc->frame_rate_base == rfps_base); // should be true for now
             if (enc->frame_rate != rfps) { 
@@ -2844,6 +2862,7 @@ static void opt_output_file(const char *filename)
                 video_enc->inter_quant_bias = video_inter_quant_bias;
                 video_enc->dct_algo = dct_algo;
                 video_enc->idct_algo = idct_algo;
+                video_enc->me_threshold= me_threshold;
                 video_enc->strict_std_compliance = strict;
                 video_enc->error_rate = error_rate;
                 video_enc->noise_reduction= noise_reduction;
@@ -3415,6 +3434,7 @@ const OptionDef options[] = {
     { "v", HAS_ARG, {(void*)opt_verbose}, "control amount of logging", "verbose" },
     { "target", HAS_ARG, {(void*)opt_target}, "specify target file type (\"vcd\", \"svcd\" or \"dvd\")", "type" },
     { "threads", HAS_ARG | OPT_EXPERT, {(void*)opt_thread_count}, "thread count", "count" },
+    { "sync", HAS_ARG | OPT_EXPERT, {(void*)opt_sync_method}, "sync method", "" },
 
     /* video options */
     { "b", HAS_ARG | OPT_VIDEO, {(void*)opt_video_bitrate}, "set video bitrate (in kbit/s)", "bitrate" },
@@ -3461,6 +3481,7 @@ const OptionDef options[] = {
       "method" },
     { "dct_algo", HAS_ARG | OPT_EXPERT | OPT_VIDEO, {(void*)opt_dct_algo}, "set dct algo",  "algo" },
     { "idct_algo", HAS_ARG | OPT_EXPERT | OPT_VIDEO, {(void*)opt_idct_algo}, "set idct algo",  "algo" },
+    { "me_threshold", HAS_ARG | OPT_EXPERT | OPT_VIDEO, {(void*)opt_me_threshold}, "motion estimaton threshold",  "" },
     { "er", HAS_ARG | OPT_EXPERT | OPT_VIDEO, {(void*)opt_error_resilience}, "set error resilience",  "n" },
     { "ec", HAS_ARG | OPT_EXPERT | OPT_VIDEO, {(void*)opt_error_concealment}, "set error concealment",  "bit_mask" },
     { "bf", HAS_ARG | OPT_EXPERT | OPT_VIDEO, {(void*)opt_b_frames}, "use 'frames' B frames", "frames" },

libavcodec/motion_est.c

@@ -973,6 +973,119 @@ static int interlaced_search(MpegEncContext *s, int ref_index,
     }
 }
 
+static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
+    MotionEstContext * const c= &s->me;
+    Picture *p= s->current_picture_ptr;
+    int mb_xy= mb_x + mb_y*s->mb_stride;
+    int xy= 2*mb_x + 2*mb_y*s->b8_stride;
+    int mb_type= s->current_picture.mb_type[mb_xy];
+    int flags= c->flags;
+    int shift= (flags&FLAG_QPEL) + 1;
+    int mask= (1<<shift)-1;
+    int x, y;
+    int d=0;
+    me_cmp_func cmpf= s->dsp.sse[0];
+    me_cmp_func chroma_cmpf= s->dsp.sse[1];
+    
+    assert(p_type==0 || !USES_LIST(mb_type, 1));
+    assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
+    
+    if(IS_INTERLACED(mb_type)){
+        int xy2= xy  + s->b8_stride;
+        s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
+        c->stride<<=1;
+        c->uvstride<<=1;
+        c->ref[1][0] = c->ref[0][0] + s->linesize;
+        c->ref[3][0] = c->ref[2][0] + s->linesize;
+        c->src[1][0] = c->src[0][0] + s->linesize;
+        if(c->flags & FLAG_CHROMA){
+            c->ref[1][1] = c->ref[0][1] + s->uvlinesize;
+            c->ref[1][2] = c->ref[0][2] + s->uvlinesize;
+            c->ref[3][1] = c->ref[2][1] + s->uvlinesize;
+            c->ref[3][2] = c->ref[2][2] + s->uvlinesize;
+            c->src[1][1] = c->src[0][1] + s->uvlinesize;
+            c->src[1][2] = c->src[0][2] + s->uvlinesize;
+        }
+        if(USES_LIST(mb_type, 0)){
+            int field_select0= p->ref_index[0][xy ];
+            int field_select1= p->ref_index[0][xy2];
+            assert(field_select0==0 ||field_select0==1);
+            assert(field_select1==0 ||field_select1==1);
+            if(p_type){
+                s->p_field_select_table[0][mb_xy]= field_select0;
+                s->p_field_select_table[1][mb_xy]= field_select1;
+                *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
+                *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
+                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
+            }else{
+                s->b_field_select_table[0][0][mb_xy]= field_select0;
+                s->b_field_select_table[0][1][mb_xy]= field_select1;
+                *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
+                *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
+                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
+            }
+
+            x= p->motion_val[0][xy ][0]; 
+            y= p->motion_val[0][xy ][1];
+            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
+            x= p->motion_val[0][xy2][0]; 
+            y= p->motion_val[0][xy2][1];
+            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
+        }
+        if(USES_LIST(mb_type, 1)){
+            int field_select0= p->ref_index[1][xy ];
+            int field_select1= p->ref_index[1][xy2];
+            assert(field_select0==0 ||field_select0==1);
+            assert(field_select1==0 ||field_select1==1);
+            s->b_field_select_table[1][0][mb_xy]= field_select0;
+            s->b_field_select_table[1][1][mb_xy]= field_select1;
+            *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
+            *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
+            if(USES_LIST(mb_type, 0)){
+                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
+            }else{
+                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
+            }
+
+            x= p->motion_val[1][xy ][0]; 
+            y= p->motion_val[1][xy ][1];
+            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
+            x= p->motion_val[1][xy2][0]; 
+            y= p->motion_val[1][xy2][1];
+            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
+            //FIXME bidir scores
+        }
+        c->stride>>=1;
+        c->uvstride>>=1;
+    }else{
+        if(USES_LIST(mb_type, 0)){
+            if(p_type){
+                *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
+                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
+            }else if(USES_LIST(mb_type, 1)){
+                *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
+                *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
+                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
+            }else{
+                *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
+                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
+            }
+            x= p->motion_val[0][xy][0]; 
+            y= p->motion_val[0][xy][1];
+            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
+        }else if(USES_LIST(mb_type, 1)){
+            *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
+            s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
+           
+            x= p->motion_val[1][xy][0]; 
+            y= p->motion_val[1][xy][1];
+            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
+        }else
+            s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
+    }
+    return d;
+}
+
 void ff_estimate_p_frame_motion(MpegEncContext * s,
                                 int mb_x, int mb_y)
 {
@@ -999,6 +1112,28 @@ void ff_estimate_p_frame_motion(MpegEncContext * s,
     get_limits(s, 16*mb_x, 16*mb_y);
     s->me.skip=0;
 
+    if(s->avctx->me_threshold){
+        vard= (check_input_motion(s, mb_x, mb_y, 1)+128)>>8;
+        
+        if(vard<s->avctx->me_threshold){
+            pix = c->src[0][0];
+            sum = s->dsp.pix_sum(pix, s->linesize);
+            varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
+        
+            pic->mb_var   [s->mb_stride * mb_y + mb_x] = varc;
+            pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
+            pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
+            s->mb_var_sum_temp    += varc;
+            s->mc_mb_var_sum_temp += vard;
+            if (vard <= 64 || vard < varc) { //FIXME
+                s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
+            }else{
+                s->scene_change_score+= s->qscale;
+            }
+            return;
+        }
+    }
+
     switch(s->me_method) {
     case ME_ZERO:
     default:
@@ -1555,6 +1690,28 @@ void ff_estimate_b_frame_motion(MpegEncContext * s,
     init_mc(s, 0, s->me.flags);
     
     s->me.skip=0;
+    if(s->avctx->me_threshold){
+        int vard= (check_input_motion(s, mb_x, mb_y, 0)+128)>>8;
+        
+        if(vard<s->avctx->me_threshold){
+//            pix = c->src[0][0];
+//            sum = s->dsp.pix_sum(pix, s->linesize);
+//            varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
+        
+//            pic->mb_var   [s->mb_stride * mb_y + mb_x] = varc;
+             s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
+/*            pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
+            s->mb_var_sum_temp    += varc;*/
+            s->mc_mb_var_sum_temp += vard;
+/*            if (vard <= 64 || vard < varc) {
+                s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
+            }else{
+                s->scene_change_score+= s->qscale;
+            }*/
+            return;
+        }
+    }
+
     if (s->codec_id == CODEC_ID_MPEG4)
         dmin= direct_search(s, mb_x, mb_y);
     else

libavcodec/mpegvideo.c

@@ -295,7 +295,8 @@ static void copy_picture_attributes(MpegEncContext *s, AVFrame *dst, AVFrame *sr
         if(!src->ref_index[0])
             av_log(s->avctx, AV_LOG_ERROR, "AVFrame.ref_index not set!\n");
         if(src->motion_subsample_log2 != dst->motion_subsample_log2)
-            av_log(s->avctx, AV_LOG_ERROR, "AVFrame.motion_subsample_log2 doesnt match!\n");
+            av_log(s->avctx, AV_LOG_ERROR, "AVFrame.motion_subsample_log2 doesnt match! (%d!=%d)\n",
+            src->motion_subsample_log2, dst->motion_subsample_log2);
 
         memcpy(dst->mb_type, src->mb_type, s->mb_stride * s->mb_height * sizeof(dst->mb_type[0]));
         
@@ -2055,11 +2056,12 @@ static void select_input_picture(MpegEncContext *s){
                 s->reordered_input_picture[0]->data[i]= NULL;
             s->reordered_input_picture[0]->type= 0;
             
-            copy_picture_attributes(s, (AVFrame*)pic, (AVFrame*)s->reordered_input_picture[0]);
             pic->reference              = s->reordered_input_picture[0]->reference;
             
             alloc_picture(s, pic, 0);
 
+            copy_picture_attributes(s, (AVFrame*)pic, (AVFrame*)s->reordered_input_picture[0]);
+
             s->current_picture_ptr= pic;
         }else{
             // input is not a shared pix -> reuse buffer for current_pix
@@ -4676,7 +4678,7 @@ static void encode_picture(MpegEncContext *s, int picture_number)
     
     /* Estimate motion for every MB */
     if(s->pict_type != I_TYPE){
-        if(s->pict_type != B_TYPE){
+        if(s->pict_type != B_TYPE && s->avctx->me_threshold==0){
             if((s->avctx->pre_me && s->last_non_b_pict_type==I_TYPE) || s->avctx->pre_me==2){
                 s->avctx->execute(s->avctx, pre_estimate_motion_thread, (void**)&(s->thread_context[0]), NULL, s->avctx->thread_count);
             }