# HG changeset patch
# User michael
# Date 1263771863 0
# Node ID f4cf3960b8c638aa771bbfb11ce5b8e6b0be11e5
# Parent  28840dfd4b52a23c6f7aa5285d9e006f467fcaec
Reorganize how values are stored in h->non_zero_count.
~1% faster

diff -r 28840dfd4b52 -r f4cf3960b8c6 h264.h
--- a/h264.h	Sun Jan 17 22:05:36 2010 +0000
+++ b/h264.h	Sun Jan 17 23:44:23 2010 +0000
@@ -300,6 +300,13 @@
      * is 64 if not available.
      */
     DECLARE_ALIGNED_8(uint8_t, non_zero_count_cache[6*8]);
+
+    /*
+    .UU.YYYY
+    .UU.YYYY
+    .vv.YYYY
+    .VV.YYYY
+    */
     uint8_t (*non_zero_count)[32];
 
     /**
@@ -727,11 +734,11 @@
     const uint8_t * left_block;
     int topleft_partition= -1;
     int i;
-    static const uint8_t left_block_options[4][8]={
-        {0,1,2,3,7,10,8,11},
-        {2,2,3,3,8,11,8,11},
-        {0,0,1,1,7,10,7,10},
-        {0,2,0,2,7,10,7,10}
+    static const uint8_t left_block_options[4][16]={
+        {0,1,2,3,7,10,8,11,7+0*8, 7+1*8, 7+2*8, 7+3*8, 2+0*8, 2+3*8, 2+1*8, 2+2*8},
+        {2,2,3,3,8,11,8,11,7+2*8, 7+2*8, 7+3*8, 7+3*8, 2+1*8, 2+2*8, 2+1*8, 2+2*8},
+        {0,0,1,1,7,10,7,10,7+0*8, 7+0*8, 7+1*8, 7+1*8, 2+0*8, 2+3*8, 2+0*8, 2+3*8},
+        {0,2,0,2,7,10,7,10,7+0*8, 7+2*8, 7+0*8, 7+2*8, 2+0*8, 2+3*8, 2+0*8, 2+3*8}
     };
 
     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
@@ -788,6 +795,12 @@
     h->left_mb_xy[0] = left_xy[0];
     h->left_mb_xy[1] = left_xy[1];
     if(for_deblock){
+        *((uint64_t*)&h->non_zero_count_cache[0+8*1])= *((uint64_t*)&h->non_zero_count[mb_xy][ 0]);
+        *((uint64_t*)&h->non_zero_count_cache[0+8*2])= *((uint64_t*)&h->non_zero_count[mb_xy][ 8]);
+        *((uint32_t*)&h->non_zero_count_cache[0+8*5])= *((uint32_t*)&h->non_zero_count[mb_xy][16]);
+        *((uint32_t*)&h->non_zero_count_cache[4+8*3])= *((uint32_t*)&h->non_zero_count[mb_xy][20]);
+        *((uint64_t*)&h->non_zero_count_cache[0+8*4])= *((uint64_t*)&h->non_zero_count[mb_xy][24]);
+
         topleft_type = 0;
         topright_type = 0;
         top_type     = h->slice_table[top_xy     ] < 0xFFFF ? s->current_picture.mb_type[top_xy]     : 0;
@@ -922,16 +935,13 @@
 */
 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
     if(top_type){
-        h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
-        h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
-        h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
-        h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
+        *(uint32_t*)&h->non_zero_count_cache[4+8*0]= *(uint32_t*)&h->non_zero_count[top_xy][4+3*8];
 
-        h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
-        h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
+        h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][1+1*8];
+        h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][2+1*8];
 
-        h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
-        h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
+        h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][1+2*8];
+        h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][2+2*8];
 
     }else{
         h->non_zero_count_cache[4+8*0]=
@@ -949,10 +959,10 @@
 
     for (i=0; i<2; i++) {
         if(left_type[i]){
-            h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
-            h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
-            h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
-            h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
+            h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[8+0+2*i]];
+            h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[8+1+2*i]];
+            h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[8+4+2*i]];
+            h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[8+5+2*i]];
         }else{
             h->non_zero_count_cache[3+8*1 + 2*8*i]=
             h->non_zero_count_cache[3+8*2 + 2*8*i]=
@@ -1204,38 +1214,11 @@
 static inline void write_back_non_zero_count(H264Context *h){
     const int mb_xy= h->mb_xy;
 
-    h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
-    h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
-    h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
-    h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
-    h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
-    h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
-    h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
-
-    h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
-    h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
-    h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
-
-    h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
-    h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
-    h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
-
-    //FIXME sort better how things are stored in non_zero_count
-
-
-    h->non_zero_count[mb_xy][13]= h->non_zero_count_cache[6+8*1];
-    h->non_zero_count[mb_xy][14]= h->non_zero_count_cache[6+8*2];
-    h->non_zero_count[mb_xy][15]= h->non_zero_count_cache[6+8*3];
-    h->non_zero_count[mb_xy][16]= h->non_zero_count_cache[5+8*1];
-    h->non_zero_count[mb_xy][17]= h->non_zero_count_cache[5+8*2];
-    h->non_zero_count[mb_xy][18]= h->non_zero_count_cache[5+8*3];
-    h->non_zero_count[mb_xy][19]= h->non_zero_count_cache[4+8*1];
-    h->non_zero_count[mb_xy][20]= h->non_zero_count_cache[4+8*2];
-    h->non_zero_count[mb_xy][21]= h->non_zero_count_cache[4+8*3];
-
-    h->non_zero_count[mb_xy][22]= h->non_zero_count_cache[1+8*1];
-    h->non_zero_count[mb_xy][23]= h->non_zero_count_cache[1+8*4];
-
+    *((uint64_t*)&h->non_zero_count[mb_xy][ 0]) = *((uint64_t*)&h->non_zero_count_cache[0+8*1]);
+    *((uint64_t*)&h->non_zero_count[mb_xy][ 8]) = *((uint64_t*)&h->non_zero_count_cache[0+8*2]);
+    *((uint32_t*)&h->non_zero_count[mb_xy][16]) = *((uint32_t*)&h->non_zero_count_cache[0+8*5]);
+    *((uint32_t*)&h->non_zero_count[mb_xy][20]) = *((uint32_t*)&h->non_zero_count_cache[4+8*3]);
+    *((uint64_t*)&h->non_zero_count[mb_xy][24]) = *((uint64_t*)&h->non_zero_count_cache[0+8*4]);
 }
 
 static inline void write_back_motion(H264Context *h, int mb_type){
diff -r 28840dfd4b52 -r f4cf3960b8c6 h264_loopfilter.c
--- a/h264_loopfilter.c	Sun Jan 17 22:05:36 2010 +0000
+++ b/h264_loopfilter.c	Sun Jan 17 23:44:23 2010 +0000
@@ -472,7 +472,6 @@
         // be done twice (one each of the field) even if we are in a
         // frame macroblock.
         //
-        static const int nnz_idx[4] = {4,5,6,3};
         unsigned int tmp_linesize   = 2 *   linesize;
         unsigned int tmp_uvlinesize = 2 * uvlinesize;
         int mbn_xy = mb_xy - 2 * s->mb_stride;
@@ -488,7 +487,7 @@
                 const uint8_t *mbn_nnz = h->non_zero_count[mbn_xy];
                 for( i = 0; i < 4; i++ ) {
                     if( h->non_zero_count_cache[scan8[0]+i] != 0 ||
-                        mbn_nnz[nnz_idx[i]] != 0 )
+                        mbn_nnz[i+4+3*8] != 0 )
                         bS[i] = 2;
                     else
                         bS[i] = 1;
@@ -663,36 +662,6 @@
             return;
         }
     }
-
-    h->non_zero_count_cache[7+8*1]=h->non_zero_count[mb_xy][0];
-    h->non_zero_count_cache[7+8*2]=h->non_zero_count[mb_xy][1];
-    h->non_zero_count_cache[7+8*3]=h->non_zero_count[mb_xy][2];
-    h->non_zero_count_cache[7+8*4]=h->non_zero_count[mb_xy][3];
-    h->non_zero_count_cache[4+8*4]=h->non_zero_count[mb_xy][4];
-    h->non_zero_count_cache[5+8*4]=h->non_zero_count[mb_xy][5];
-    h->non_zero_count_cache[6+8*4]=h->non_zero_count[mb_xy][6];
-
-    h->non_zero_count_cache[1+8*2]=h->non_zero_count[mb_xy][9];
-    h->non_zero_count_cache[2+8*2]=h->non_zero_count[mb_xy][8];
-    h->non_zero_count_cache[2+8*1]=h->non_zero_count[mb_xy][7];
-
-    h->non_zero_count_cache[1+8*5]=h->non_zero_count[mb_xy][12];
-    h->non_zero_count_cache[2+8*5]=h->non_zero_count[mb_xy][11];
-    h->non_zero_count_cache[2+8*4]=h->non_zero_count[mb_xy][10];
-
-    h->non_zero_count_cache[6+8*1]=h->non_zero_count[mb_xy][13];
-    h->non_zero_count_cache[6+8*2]=h->non_zero_count[mb_xy][14];
-    h->non_zero_count_cache[6+8*3]=h->non_zero_count[mb_xy][15];
-    h->non_zero_count_cache[5+8*1]=h->non_zero_count[mb_xy][16];
-    h->non_zero_count_cache[5+8*2]=h->non_zero_count[mb_xy][17];
-    h->non_zero_count_cache[5+8*3]=h->non_zero_count[mb_xy][18];
-    h->non_zero_count_cache[4+8*1]=h->non_zero_count[mb_xy][19];
-    h->non_zero_count_cache[4+8*2]=h->non_zero_count[mb_xy][20];
-    h->non_zero_count_cache[4+8*3]=h->non_zero_count[mb_xy][21];
-
-    h->non_zero_count_cache[1+8*1]=h->non_zero_count[mb_xy][22];
-    h->non_zero_count_cache[1+8*4]=h->non_zero_count[mb_xy][23];
-
     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
     if(!h->pps.cabac && h->pps.transform_8x8_mode){
         int top_type, left_type[2];
@@ -762,7 +731,7 @@
                          ((!h->pps.cabac && IS_8x8DCT(s->current_picture.mb_type[mbn_xy])) ?
                             (h->cbp_table[mbn_xy] & ((MB_FIELD ? (i&2) : (mb_y&1)) ? 8 : 2))
                                                                        :
-                            h->non_zero_count[mbn_xy][MB_FIELD ? i&3 : (i>>2)+(mb_y&1)*2]))
+                            h->non_zero_count[mbn_xy][7+(MB_FIELD ? (i&3) : (i>>2)+(mb_y&1)*2)*8]))
                     bS[i] = 2;
                 else
                     bS[i] = 1;