Unroll main loop so the edge==0 case is seperate.
authorMichael Niedermayer <michaelni@gmx.at>
Thu, 28 Jan 2010 01:24:25 +0000 (01:24 +0000)
committerMichael Niedermayer <michaelni@gmx.at>
Thu, 28 Jan 2010 01:24:25 +0000 (01:24 +0000)
This allows many things to be simplified away.
h264 decoder is overall 1% faster with a mbaff sample and
0.1% slower with the cathedral sample, probably because the slow loop
filter code must be loaded into the code cache for each first MB of each
row but isnt used for the following MBs.

Originally committed as revision 21493 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/h264_loopfilter.c

index a09e316..155a323 100644 (file)
@@ -476,24 +476,128 @@ static av_always_inline void filter_mb_dir(H264Context *h, int mb_x, int mb_y, u
     }
 
     /* Calculate bS */
-    for( edge = start; edge < edges; edge++ ) {
-        /* mbn_xy: neighbor macroblock */
-        const int mbn_xy = edge > 0 ? mb_xy : mbm_xy;
-        const int mbn_type = s->current_picture.mb_type[mbn_xy];
+    if(start==0) {
+        DECLARE_ALIGNED_8(int16_t, bS)[4];
+        int qp;
+
+        if( IS_INTRA(mb_type|mbm_type)) {
+            *(uint64_t*)bS= 0x0003000300030003ULL;
+            if (   (!IS_INTERLACED(mb_type|mbm_type))
+                || ((FRAME_MBAFF || (s->picture_structure != PICT_FRAME)) && (dir == 0))
+            )
+                *(uint64_t*)bS= 0x0004000400040004ULL;
+        } else {
+            int i, l;
+            int mv_done;
+
+            if( FRAME_MBAFF && IS_INTERLACED(mb_type ^ mbm_type)) { //FIXME not posible left
+                *(uint64_t*)bS= 0x0001000100010001ULL;
+                mv_done = 1;
+            }
+            else if( mask_par0 && ((mbm_type & (MB_TYPE_16x16 | (MB_TYPE_8x16 >> dir)))) ) {
+                int b_idx= 8 + 4;
+                int bn_idx= b_idx - (dir ? 8:1);
+                int v = 0;
+
+                for( l = 0; !v && l < h->list_count; l++ ) {
+                    v |= h->ref_cache[l][b_idx] != h->ref_cache[l][bn_idx] |
+                         h->mv_cache[l][b_idx][0] - h->mv_cache[l][bn_idx][0] + 3 >= 7U |
+                         FFABS( h->mv_cache[l][b_idx][1] - h->mv_cache[l][bn_idx][1] ) >= mvy_limit;
+                }
+
+                if(h->list_count==2 && v){
+                    v=0;
+                    for( l = 0; !v && l < 2; l++ ) {
+                        int ln= 1-l;
+                        v |= h->ref_cache[l][b_idx] != h->ref_cache[ln][bn_idx] |
+                            h->mv_cache[l][b_idx][0] - h->mv_cache[ln][bn_idx][0] + 3 >= 7U |
+                            FFABS( h->mv_cache[l][b_idx][1] - h->mv_cache[ln][bn_idx][1] ) >= mvy_limit;
+                    }
+                }
+
+                bS[0] = bS[1] = bS[2] = bS[3] = v;
+                mv_done = 1;
+            }
+            else
+                mv_done = 0;
+
+            for( i = 0; i < 4; i++ ) {
+                int x = dir == 0 ? 0 : i;
+                int y = dir == 0 ? i    : 0;
+                int b_idx= 8 + 4 + x + 8*y;
+                int bn_idx= b_idx - (dir ? 8:1);
+
+                if( h->non_zero_count_cache[b_idx] |
+                    h->non_zero_count_cache[bn_idx] ) {
+                    bS[i] = 2;
+                }
+                else if(!mv_done)
+                {
+                    bS[i] = 0;
+                    for( l = 0; l < h->list_count; l++ ) {
+                        if( h->ref_cache[l][b_idx] != h->ref_cache[l][bn_idx] |
+                            h->mv_cache[l][b_idx][0] - h->mv_cache[l][bn_idx][0] + 3 >= 7U |
+                            FFABS( h->mv_cache[l][b_idx][1] - h->mv_cache[l][bn_idx][1] ) >= mvy_limit ) {
+                            bS[i] = 1;
+                            break;
+                        }
+                    }
+
+                    if(h->list_count == 2 && bS[i]){
+                        bS[i] = 0;
+                        for( l = 0; l < 2; l++ ) {
+                            int ln= 1-l;
+                            if( h->ref_cache[l][b_idx] != h->ref_cache[ln][bn_idx] |
+                                h->mv_cache[l][b_idx][0] - h->mv_cache[ln][bn_idx][0] + 3 >= 7U |
+                                FFABS( h->mv_cache[l][b_idx][1] - h->mv_cache[ln][bn_idx][1] ) >= mvy_limit ) {
+                                bS[i] = 1;
+                                break;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        /* Filter edge */
+        // Do not use s->qscale as luma quantizer because it has not the same
+        // value in IPCM macroblocks.
+        if(bS[0]+bS[1]+bS[2]+bS[3]){
+        qp = ( s->current_picture.qscale_table[mb_xy] + s->current_picture.qscale_table[mbm_xy] + 1 ) >> 1;
+        //tprintf(s->avctx, "filter mb:%d/%d dir:%d edge:%d, QPy:%d, QPc:%d, QPcn:%d\n", mb_x, mb_y, dir, edge, qp, h->chroma_qp[0], s->current_picture.qscale_table[mbn_xy]);
+        tprintf(s->avctx, "filter mb:%d/%d dir:%d edge:%d, QPy:%d ls:%d uvls:%d", mb_x, mb_y, dir, edge, qp, linesize, uvlinesize);
+        //{ int i; for (i = 0; i < 4; i++) tprintf(s->avctx, " bS[%d]:%d", i, bS[i]); tprintf(s->avctx, "\n"); }
+        if( dir == 0 ) {
+            filter_mb_edgev( &img_y[0], linesize, bS, qp, h );
+            {
+                int qp= ( h->chroma_qp[0] + get_chroma_qp( h, 0, s->current_picture.qscale_table[mbm_xy] ) + 1 ) >> 1;
+                filter_mb_edgecv( &img_cb[0], uvlinesize, bS, qp, h);
+                if(h->pps.chroma_qp_diff)
+                    qp= ( h->chroma_qp[1] + get_chroma_qp( h, 1, s->current_picture.qscale_table[mbm_xy] ) + 1 ) >> 1;
+                filter_mb_edgecv( &img_cr[0], uvlinesize, bS, qp, h);
+            }
+        } else {
+            filter_mb_edgeh( &img_y[0], linesize, bS, qp, h );
+            {
+                int qp= ( h->chroma_qp[0] + get_chroma_qp( h, 0, s->current_picture.qscale_table[mbm_xy] ) + 1 ) >> 1;
+                filter_mb_edgech( &img_cb[0], uvlinesize, bS, qp, h);
+                if(h->pps.chroma_qp_diff)
+                    qp= ( h->chroma_qp[1] + get_chroma_qp( h, 1, s->current_picture.qscale_table[mbm_xy] ) + 1 ) >> 1;
+                filter_mb_edgech( &img_cr[0], uvlinesize, bS, qp, h);
+            }
+        }
+        }
+    }
+    /* Calculate bS */
+    for( edge = 1; edge < edges; edge++ ) {
         DECLARE_ALIGNED_8(int16_t, bS)[4];
         int qp;
 
         if( IS_8x8DCT(mb_type & (edge<<24)) ) // (edge&1) && IS_8x8DCT(mb_type)
             continue;
 
-        if( IS_INTRA(mb_type|mbn_type)) {
+        if( IS_INTRA(mb_type)) {
             *(uint64_t*)bS= 0x0003000300030003ULL;
-            if (edge == 0) {
-                if (   (!IS_INTERLACED(mb_type|mbm_type))
-                    || ((FRAME_MBAFF || (s->picture_structure != PICT_FRAME)) && (dir == 0))
-                )
-                    *(uint64_t*)bS= 0x0004000400040004ULL;
-            }
         } else {
             int i, l;
             int mv_done;
@@ -502,11 +606,7 @@ static av_always_inline void filter_mb_dir(H264Context *h, int mb_x, int mb_y, u
                 *(uint64_t*)bS= 0;
                 mv_done = 1;
             }
-            else if( FRAME_MBAFF && IS_INTERLACED(mb_type ^ mbn_type)) {
-                *(uint64_t*)bS= 0x0001000100010001ULL;
-                mv_done = 1;
-            }
-            else if( mask_par0 && (edge || (mbn_type & (MB_TYPE_16x16 | (MB_TYPE_8x16 >> dir)))) ) {
+            else if( mask_par0 ) {
                 int b_idx= 8 + 4 + edge * (dir ? 8:1);
                 int bn_idx= b_idx - (dir ? 8:1);
                 int v = 0;
@@ -577,27 +677,21 @@ static av_always_inline void filter_mb_dir(H264Context *h, int mb_x, int mb_y, u
         /* Filter edge */
         // Do not use s->qscale as luma quantizer because it has not the same
         // value in IPCM macroblocks.
-        qp = ( s->current_picture.qscale_table[mb_xy] + s->current_picture.qscale_table[mbn_xy] + 1 ) >> 1;
+        qp = s->current_picture.qscale_table[mb_xy];
         //tprintf(s->avctx, "filter mb:%d/%d dir:%d edge:%d, QPy:%d, QPc:%d, QPcn:%d\n", mb_x, mb_y, dir, edge, qp, h->chroma_qp[0], s->current_picture.qscale_table[mbn_xy]);
         tprintf(s->avctx, "filter mb:%d/%d dir:%d edge:%d, QPy:%d ls:%d uvls:%d", mb_x, mb_y, dir, edge, qp, linesize, uvlinesize);
         //{ int i; for (i = 0; i < 4; i++) tprintf(s->avctx, " bS[%d]:%d", i, bS[i]); tprintf(s->avctx, "\n"); }
         if( dir == 0 ) {
             filter_mb_edgev( &img_y[4*edge], linesize, bS, qp, h );
             if( (edge&1) == 0 ) {
-                int qp= ( h->chroma_qp[0] + get_chroma_qp( h, 0, s->current_picture.qscale_table[mbn_xy] ) + 1 ) >> 1;
-                filter_mb_edgecv( &img_cb[2*edge], uvlinesize, bS, qp, h);
-                if(h->pps.chroma_qp_diff)
-                    qp= ( h->chroma_qp[1] + get_chroma_qp( h, 1, s->current_picture.qscale_table[mbn_xy] ) + 1 ) >> 1;
-                filter_mb_edgecv( &img_cr[2*edge], uvlinesize, bS, qp, h);
+                filter_mb_edgecv( &img_cb[2*edge], uvlinesize, bS, h->chroma_qp[0], h);
+                filter_mb_edgecv( &img_cr[2*edge], uvlinesize, bS, h->chroma_qp[1], h);
             }
         } else {
             filter_mb_edgeh( &img_y[4*edge*linesize], linesize, bS, qp, h );
             if( (edge&1) == 0 ) {
-                int qp= ( h->chroma_qp[0] + get_chroma_qp( h, 0, s->current_picture.qscale_table[mbn_xy] ) + 1 ) >> 1;
-                filter_mb_edgech( &img_cb[2*edge*uvlinesize], uvlinesize, bS, qp, h);
-                if(h->pps.chroma_qp_diff)
-                    qp= ( h->chroma_qp[1] + get_chroma_qp( h, 1, s->current_picture.qscale_table[mbn_xy] ) + 1 ) >> 1;
-                filter_mb_edgech( &img_cr[2*edge*uvlinesize], uvlinesize, bS, qp, h);
+                filter_mb_edgech( &img_cb[2*edge*uvlinesize], uvlinesize, bS, h->chroma_qp[0], h);
+                filter_mb_edgech( &img_cr[2*edge*uvlinesize], uvlinesize, bS, h->chroma_qp[1], h);
             }
         }
     }