# HG changeset patch
# User michael
# Date 1267141482 0
# Node ID 411ab09ada91e6b2a58a3a51b0fe0809e1aea859
# Parent  1527e25ec9d49ab99f03caad7c0190bd92e988ad
Get rid of mb2b8_xy and b8_stride, change arrays organized based on b8_stride to
ones based on mb_stride in h264.
about 20 cpu cycles faster overall per MB

diff -r 1527e25ec9d4 -r 411ab09ada91 error_resilience.c
--- a/error_resilience.c	Thu Feb 25 18:41:48 2010 +0000
+++ b/error_resilience.c	Thu Feb 25 23:44:42 2010 +0000
@@ -692,7 +692,7 @@
         av_log(s->avctx, AV_LOG_ERROR, "Warning MVs not available\n");
 
         for(i=0; i<2; i++){
-            pic->ref_index[i]= av_mallocz(size * sizeof(uint8_t));
+            pic->ref_index[i]= av_mallocz(s->mb_stride * s->mb_height * 4 * sizeof(uint8_t));
             pic->motion_val_base[i]= av_mallocz((size+4) * 2 * sizeof(uint16_t));
             pic->motion_val[i]= pic->motion_val_base[i]+4;
         }
diff -r 1527e25ec9d4 -r 411ab09ada91 h264.c
--- a/h264.c	Thu Feb 25 18:41:48 2010 +0000
+++ b/h264.c	Thu Feb 25 23:44:42 2010 +0000
@@ -661,7 +661,6 @@
 
     av_freep(&h->mb2b_xy);
     av_freep(&h->mb2br_xy);
-    av_freep(&h->mb2b8_xy);
 
     for(i = 0; i < MAX_THREADS; i++) {
         hx = h->thread_context[i];
@@ -764,16 +763,13 @@
 
     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
-    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b8_xy , big_mb_num * sizeof(uint32_t), fail);
     for(y=0; y<s->mb_height; y++){
         for(x=0; x<s->mb_width; x++){
             const int mb_xy= x + y*s->mb_stride;
             const int b_xy = 4*x + 4*y*h->b_stride;
-            const int b8_xy= 2*x + 2*y*h->b8_stride;
 
             h->mb2b_xy [mb_xy]= b_xy;
             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
-            h->mb2b8_xy[mb_xy]= b8_xy;
         }
     }
 
@@ -798,7 +794,6 @@
     dst->cbp_table                = src->cbp_table;
     dst->mb2b_xy                  = src->mb2b_xy;
     dst->mb2br_xy                 = src->mb2br_xy;
-    dst->mb2b8_xy                 = src->mb2b8_xy;
     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
     dst->mvd_table[0]             = src->mvd_table[0];
     dst->mvd_table[1]             = src->mvd_table[1];
@@ -1768,7 +1763,6 @@
     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
 
     h->b_stride=  s->mb_width*4;
-    h->b8_stride= s->mb_width*2;
 
     s->width = 16*s->mb_width - 2*FFMIN(h->sps.crop_right, 7);
     if(h->sps.frame_mbs_only_flag)
diff -r 1527e25ec9d4 -r 411ab09ada91 h264.h
--- a/h264.h	Thu Feb 25 18:41:48 2010 +0000
+++ b/h264.h	Thu Feb 25 23:44:42 2010 +0000
@@ -347,9 +347,7 @@
 
     uint32_t *mb2b_xy; //FIXME are these 4 a good idea?
     uint32_t *mb2br_xy;
-    uint32_t *mb2b8_xy;
     int b_stride; //FIXME use s->b4_stride
-    int b8_stride;
 
     int mb_linesize;   ///< may be equal to s->linesize or s->linesize*2, for mbaff
     int mb_uvlinesize;
@@ -990,12 +988,11 @@
 
             if(USES_LIST(top_type, list)){
                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
-                const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
                 AV_COPY128(h->mv_cache[list][scan8[0] + 0 - 1*8], s->current_picture.motion_val[list][b_xy + 0]);
                     h->ref_cache[list][scan8[0] + 0 - 1*8]=
-                    h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
+                    h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][4*top_xy + 2];
                     h->ref_cache[list][scan8[0] + 2 - 1*8]=
-                    h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
+                    h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][4*top_xy + 3];
             }else{
                 AV_ZERO128(h->mv_cache[list][scan8[0] + 0 - 1*8]);
                 AV_WN32A(&h->ref_cache[list][scan8[0] + 0 - 1*8], ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101);
@@ -1005,11 +1002,11 @@
                 int cache_idx = scan8[0] - 1 + i*2*8;
                 if(USES_LIST(left_type[i], list)){
                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
-                    const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
+                    const int b8_xy= 4*left_xy[i] + 1;
                     AV_COPY32(h->mv_cache[list][cache_idx  ], s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[0+i*2]]);
                     AV_COPY32(h->mv_cache[list][cache_idx+8], s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[1+i*2]]);
-                        h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
-                        h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
+                        h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + (left_block[0+i*2]&~1)];
+                        h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + (left_block[1+i*2]&~1)];
                 }else{
                     AV_ZERO32(h->mv_cache [list][cache_idx  ]);
                     AV_ZERO32(h->mv_cache [list][cache_idx+8]);
@@ -1020,7 +1017,7 @@
 
             if(USES_LIST(topleft_type, list)){
                 const int b_xy = h->mb2b_xy [topleft_xy] + 3 + h->b_stride + (h->topleft_partition & 2*h->b_stride);
-                const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (h->topleft_partition & h->b8_stride);
+                const int b8_xy= 4*topleft_xy + 1 + (h->topleft_partition & 2);
                 AV_COPY32(h->mv_cache[list][scan8[0] - 1 - 1*8], s->current_picture.motion_val[list][b_xy]);
                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
             }else{
@@ -1030,9 +1027,8 @@
 
             if(USES_LIST(topright_type, list)){
                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
-                const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
                 AV_COPY32(h->mv_cache[list][scan8[0] + 4 - 1*8], s->current_picture.motion_val[list][b_xy]);
-                h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
+                h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][4*topright_xy + 2];
             }else{
                 AV_ZERO32(h->mv_cache [list][scan8[0] + 4 - 1*8]);
                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
@@ -1241,12 +1237,12 @@
                 continue;
             }
 
-            ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
+            ref = &s->current_picture.ref_index[list][4*mb_xy];
             {
                 int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
                 AV_WN32A(&h->ref_cache[list][scan8[ 0]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
                 AV_WN32A(&h->ref_cache[list][scan8[ 2]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
-                ref += h->b8_stride;
+                ref += 2;
                 AV_WN32A(&h->ref_cache[list][scan8[ 8]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
                 AV_WN32A(&h->ref_cache[list][scan8[10]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
             }
@@ -1319,7 +1315,7 @@
         for(list=0; list<h->list_count; list++){
             if(USES_LIST(top_type, list)){
                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
-                const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
+                const int b8_xy= 4*top_xy + 2;
                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
                 AV_COPY128(h->mv_cache[list][scan8[0] + 0 - 1*8], s->current_picture.motion_val[list][b_xy + 0]);
                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
@@ -1334,16 +1330,16 @@
             if(!IS_INTERLACED(mb_type^left_type[0])){
                 if(USES_LIST(left_type[0], list)){
                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
-                    const int b8_xy= h->mb2b8_xy[left_xy[0]] + 1;
+                    const int b8_xy= 4*left_xy[0] + 1;
                     int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[0]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 0 ], s->current_picture.motion_val[list][b_xy + h->b_stride*0]);
                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 8 ], s->current_picture.motion_val[list][b_xy + h->b_stride*1]);
                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 +16 ], s->current_picture.motion_val[list][b_xy + h->b_stride*2]);
                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 +24 ], s->current_picture.motion_val[list][b_xy + h->b_stride*3]);
                     h->ref_cache[list][scan8[0] - 1 + 0 ]=
-                    h->ref_cache[list][scan8[0] - 1 + 8 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + h->b8_stride*0]];
+                    h->ref_cache[list][scan8[0] - 1 + 8 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*0]];
                     h->ref_cache[list][scan8[0] - 1 +16 ]=
-                    h->ref_cache[list][scan8[0] - 1 +24 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + h->b8_stride*1]];
+                    h->ref_cache[list][scan8[0] - 1 +24 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*1]];
                 }else{
                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 0 ]);
                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 8 ]);
@@ -1388,12 +1384,12 @@
 
 static inline void write_back_motion(H264Context *h, int mb_type){
     MpegEncContext * const s = &h->s;
-    const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
-    const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
+    const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride; //try mb2b(8)_xy
+    const int b8_xy= 4*h->mb_xy;
     int list;
 
     if(!USES_LIST(mb_type, 0))
-        fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
+        fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, 2, (uint8_t)LIST_NOT_USED, 1);
 
     for(list=0; list<h->list_count; list++){
         int y, b_stride;
@@ -1424,10 +1420,10 @@
 
         {
             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
-            ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
-            ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
-            ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
-            ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
+            ref_index[0+0*2]= h->ref_cache[list][scan8[0]];
+            ref_index[1+0*2]= h->ref_cache[list][scan8[4]];
+            ref_index[0+1*2]= h->ref_cache[list][scan8[8]];
+            ref_index[1+1*2]= h->ref_cache[list][scan8[12]];
         }
     }
 
diff -r 1527e25ec9d4 -r 411ab09ada91 h264_direct.c
--- a/h264_direct.c	Thu Feb 25 18:41:48 2010 +0000
+++ b/h264_direct.c	Thu Feb 25 23:44:42 2010 +0000
@@ -142,7 +142,7 @@
 
 static void pred_spatial_direct_motion(H264Context * const h, int *mb_type){
     MpegEncContext * const s = &h->s;
-    int b8_stride = h->b8_stride;
+    int b8_stride = 2;
     int b4_stride = h->b_stride;
     int mb_xy = h->mb_xy;
     int mb_type_col[2];
@@ -228,7 +228,7 @@
             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
-            b8_stride *= 3;
+            b8_stride = 2+4*s->mb_stride;
             b4_stride *= 6;
 
             sub_mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
@@ -262,12 +262,12 @@
 
     l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
     l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
-    l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
-    l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
+    l1ref0 = &h->ref_list[1][0].ref_index [0][4*mb_xy];
+    l1ref1 = &h->ref_list[1][0].ref_index [1][4*mb_xy];
     if(!b8_stride){
         if(s->mb_y&1){
-            l1ref0 += h->b8_stride;
-            l1ref1 += h->b8_stride;
+            l1ref0 += 2;
+            l1ref1 += 2;
             l1mv0  +=  2*b4_stride;
             l1mv1  +=  2*b4_stride;
         }
@@ -342,11 +342,12 @@
                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
 
+                assert(b8_stride==2);
                 /* col_zero_flag */
-                if(!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref && (   l1ref0[x8 + y8*b8_stride] == 0
-                                              || (l1ref0[x8 + y8*b8_stride] < 0 && l1ref1[x8 + y8*b8_stride] == 0
+                if(!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref && (   l1ref0[i8] == 0
+                                              || (l1ref0[i8] < 0 && l1ref1[i8] == 0
                                                   && h->x264_build>33U))){
-                    const int16_t (*l1mv)[2]= l1ref0[x8 + y8*b8_stride] == 0 ? l1mv0 : l1mv1;
+                    const int16_t (*l1mv)[2]= l1ref0[i8] == 0 ? l1mv0 : l1mv1;
                     if(IS_SUB_8X8(sub_mb_type)){
                         const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
@@ -381,7 +382,7 @@
 
 static void pred_temp_direct_motion(H264Context * const h, int *mb_type){
     MpegEncContext * const s = &h->s;
-    int b8_stride = h->b8_stride;
+    int b8_stride = 2;
     int b4_stride = h->b_stride;
     int mb_xy = h->mb_xy;
     int mb_type_col[2];
@@ -406,7 +407,7 @@
             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
-            b8_stride *= 3;
+            b8_stride = 2+4*s->mb_stride;
             b4_stride *= 6;
 
             sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
@@ -441,12 +442,12 @@
 
     l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
     l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
-    l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
-    l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
+    l1ref0 = &h->ref_list[1][0].ref_index [0][4*mb_xy];
+    l1ref1 = &h->ref_list[1][0].ref_index [1][4*mb_xy];
     if(!b8_stride){
         if(s->mb_y&1){
-            l1ref0 += h->b8_stride;
-            l1ref1 += h->b8_stride;
+            l1ref0 += 2;
+            l1ref1 += 2;
             l1mv0  +=  2*b4_stride;
             l1mv1  +=  2*b4_stride;
         }
@@ -549,11 +550,12 @@
                     continue;
                 }
 
-                ref0 = l1ref0[x8 + y8*b8_stride];
+                assert(b8_stride == 2);
+                ref0 = l1ref0[i8];
                 if(ref0 >= 0)
                     ref0 = map_col_to_list0[0][ref0 + ref_offset];
                 else{
-                    ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
+                    ref0 = map_col_to_list0[1][l1ref1[i8] + ref_offset];
                     l1mv= l1mv1;
                 }
                 scale = dist_scale_factor[ref0];
diff -r 1527e25ec9d4 -r 411ab09ada91 h264_mvpred.h
--- a/h264_mvpred.h	Thu Feb 25 18:41:48 2010 +0000
+++ b/h264_mvpred.h	Thu Feb 25 23:44:42 2010 +0000
@@ -43,15 +43,15 @@
      * make mbaff happy, so we can't move all this logic to fill_caches */
     if(FRAME_MBAFF){
 
-#define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
-                const int x4 = X4, y4 = Y4;\
-                const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
+#define SET_DIAG_MV(MV_OP, REF_OP, XY, Y4)\
+                const int xy = XY, y4 = Y4;\
+                const int mb_type = mb_types[xy+(y4>>2)*s->mb_stride];\
                 if(!USES_LIST(mb_type,list))\
                     return LIST_NOT_USED;\
-                mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
+                mv = s->current_picture_ptr->motion_val[list][h->mb2b_xy[xy]+3 + y4*h->b_stride];\
                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
-                return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
+                return s->current_picture_ptr->ref_index[list][4*xy+1 + (y4&~1)] REF_OP;
 
         if(topright_ref == PART_NOT_AVAILABLE
            && i >= scan8[0]+8 && (i&7)==4
@@ -63,12 +63,13 @@
 
             if(!MB_FIELD
                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
-                SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
+                SET_DIAG_MV(*2, >>1, h->left_mb_xy[0]+s->mb_stride, (s->mb_y&1)*2+(i>>4)-1);
+                assert(h->left_mb_xy[0] == h->left_mb_xy[1]);
             }
             if(MB_FIELD
                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
                 // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
-                SET_DIAG_MV(/2, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
+                SET_DIAG_MV(/2, <<1, h->left_mb_xy[i>=36], (- 1 + ((i-scan8[0])>>3)*2)&3);
             }
         }
 #undef SET_DIAG_MV
diff -r 1527e25ec9d4 -r 411ab09ada91 mpegvideo.c
--- a/mpegvideo.c	Thu Feb 25 18:41:48 2010 +0000
+++ b/mpegvideo.c	Thu Feb 25 23:44:42 2010 +0000
@@ -257,7 +257,7 @@
             for(i=0; i<2; i++){
                 FF_ALLOCZ_OR_GOTO(s->avctx, pic->motion_val_base[i], 2 * (b4_array_size+4)  * sizeof(int16_t), fail)
                 pic->motion_val[i]= pic->motion_val_base[i]+4;
-                FF_ALLOCZ_OR_GOTO(s->avctx, pic->ref_index[i], b8_array_size * sizeof(uint8_t), fail)
+                FF_ALLOCZ_OR_GOTO(s->avctx, pic->ref_index[i], 4*mb_array_size * sizeof(uint8_t), fail)
             }
             pic->motion_subsample_log2= 2;
         }else if(s->out_format == FMT_H263 || s->encoding || (s->avctx->debug&FF_DEBUG_MV) || (s->avctx->debug_mv)){