rate distortion optimal cbp support (h263/mpeg4 non intra only)

Originally committed as revision 2323 to svn://svn.ffmpeg.org/ffmpeg/trunk

1 files changed, 110 insertions, 11 deletions

diff --git a/libavcodec/h263.c b/libavcodec/h263.c
index c74bee8752..845b03e600 100644
--- a/libavcodec/h263.c
+++ b/libavcodec/h263.c
@@ -29,6 +29,8 @@
  */
  
 //#define DEBUG
+#include <limits.h>
+
 #include "common.h"
 #include "dsputil.h"
 #include "avcodec.h"
@@ -560,6 +562,106 @@ void ff_h263_update_motion_val(MpegEncContext * s){
 }
 
 #ifdef CONFIG_ENCODERS
+
+static inline int get_p_cbp(MpegEncContext * s,
+                      DCTELEM block[6][64],
+                      int motion_x, int motion_y){
+    int cbp, i;
+
+    if(s->flags & CODEC_FLAG_CBP_RD){
+        int best_cbpy_score= INT_MAX;
+        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; 
+
+        for(i=0; i<4; i++){
+            int score= inter_MCBPC_bits[i + offset] * lambda;
+            if(i&1) score += s->coded_score[5];
+            if(i&2) score += s->coded_score[4];
+
+            if(score < best_cbpc_score){
+                best_cbpc_score= score;
+                cbpc= i;
+            }
+        }
+
+        for(i=0; i<16; i++){
+            int score= cbpy_tab[i ^ 0xF][1] * lambda;
+            if(i&1) score += s->coded_score[3];
+            if(i&2) score += s->coded_score[2];
+            if(i&4) score += s->coded_score[1];
+            if(i&8) score += s->coded_score[0];
+
+            if(score < best_cbpy_score){
+                best_cbpy_score= score;
+                cbpy= i;
+            }
+        }
+        cbp= cbpc + 4*cbpy;
+        if ((motion_x | motion_y | s->dquant) == 0 && s->mv_type==MV_TYPE_16X16){
+            if(best_cbpy_score + best_cbpc_score + 2*lambda >= 0)
+                cbp= 0;
+        }
+
+        for (i = 0; i < 6; i++) {
+            if (s->block_last_index[i] >= 0 && ((cbp >> (5 - i))&1)==0 ){
+                s->block_last_index[i]= -1;
+                memset(s->block[i], 0, sizeof(DCTELEM)*64);
+            }
+        }
+    }else{
+        cbp= 0;
+        for (i = 0; i < 6; i++) {
+            if (s->block_last_index[i] >= 0)
+                cbp |= 1 << (5 - i);
+        }
+    }
+    return cbp;
+}
+
+static inline int get_b_cbp(MpegEncContext * s, DCTELEM block[6][64],
+                            int motion_x, int motion_y, int mb_type){
+    int cbp=0, i;
+
+    if(s->flags & CODEC_FLAG_CBP_RD){
+        int score=0;
+        const int lambda= (s->qscale*s->qscale*64*105 + 64)>>7; 
+        
+        for(i=0; i<6; i++){
+            if(s->coded_score[i] < 0){
+                score += s->coded_score[i];
+                cbp |= 1 << (5 - i);
+            }
+        }
+        
+        if(cbp){
+            int zero_score= -6;
+            if ((motion_x | motion_y | s->dquant | mb_type) == 0){
+                zero_score-= 4; //2*MV + mb_type + cbp bit
+            }
+
+            zero_score*= lambda;
+            if(zero_score <= score){
+                cbp=0;
+            }
+        }
+
+        for (i = 0; i < 6; i++) {
+            if (s->block_last_index[i] >= 0 && ((cbp >> (5 - i))&1)==0 ){
+                s->block_last_index[i]= -1;
+                memset(s->block[i], 0, sizeof(DCTELEM)*64);
+            }
+        }
+    }else{
+        for (i = 0; i < 6; i++) {
+            if (s->block_last_index[i] >= 0)
+                cbp |= 1 << (5 - i);
+        }
+    }
+    return cbp;
+}
+
 void mpeg4_encode_mb(MpegEncContext * s,
 		    DCTELEM block[6][64],
 		    int motion_x, int motion_y)
@@ -574,12 +676,8 @@ void mpeg4_encode_mb(MpegEncContext * s,
     //    printf("**mb x=%d y=%d\n", s->mb_x, s->mb_y);
     if (!s->mb_intra) {
         /* compute cbp */
-        int i, cbp = 0;
-        for (i = 0; i < 6; i++) {
-            if (s->block_last_index[i] >= 0)
-                cbp |= 1 << (5 - i);
-        }
-
+        int i, cbp;
+        
         if(s->pict_type==B_TYPE){
             static const int mb_type_table[8]= {-1, 2, 3, 1,-1,-1,-1, 0}; /* convert from mv_dir to type */
             int mb_type=  mb_type_table[s->mv_dir];
@@ -609,6 +707,8 @@ void mpeg4_encode_mb(MpegEncContext * s,
                 return;
             }
             
+            cbp= get_b_cbp(s, block, motion_x, motion_y, mb_type);
+            
             if ((cbp | motion_x | motion_y | mb_type) ==0) {
                 /* direct MB with MV={0,0} */
                 assert(s->dquant==0);
@@ -699,6 +799,8 @@ void mpeg4_encode_mb(MpegEncContext * s,
                 s->p_tex_bits+= get_bits_diff(s);
             }
         }else{ /* s->pict_type==B_TYPE */
+            cbp= get_p_cbp(s, block, motion_x, motion_y);
+        
             if ((cbp | motion_x | motion_y | s->dquant) == 0 && s->mv_type==MV_TYPE_16X16) {
                 /* check if the B frames can skip it too, as we must skip it if we skip here 
                    why didnt they just compress the skip-mb bits instead of reusing them ?! */
@@ -938,11 +1040,8 @@ void h263_encode_mb(MpegEncContext * s,
     //printf("**mb x=%d y=%d\n", s->mb_x, s->mb_y);
     if (!s->mb_intra) {
         /* compute cbp */
-        cbp = 0;
-        for (i = 0; i < 6; i++) {
-            if (s->block_last_index[i] >= 0)
-                cbp |= 1 << (5 - i);
-        }
+        cbp= get_p_cbp(s, block, motion_x, motion_y);
+
         if ((cbp | motion_x | motion_y | s->dquant) == 0) {
             /* skip macroblock */
             put_bits(&s->pb, 1, 1);