[x264-devel] commit: Cache ref costs and use more accurate MV costs ( Jason Garrett-Glaser )

[git version control]

x264 | branch: master | Jason Garrett-Glaser <darkshikari at gmail.com> | Tue Jan 13 21:11:50 2009 -0500| [e46f64824a718ca146724d6cfa104aaba16eb169] | committer: Jason Garrett-Glaser 

Cache ref costs and use more accurate MV costs
New MV costs should improve quality slightly by improving the smoothness of the field of MV costs (and they're closer to CABAC's actual costs).
Despite being optimized for CABAC, they still help under CAVLC, albeit less.
MV cost change by Loren Merritt

> http://git.videolan.org/gitweb.cgi/x264.git/?a=commit;h=e46f64824a718ca146724d6cfa104aaba16eb169
---

 encoder/analyse.c |   15 +++++++++++++--
 1 files changed, 13 insertions(+), 2 deletions(-)

diff --git a/encoder/analyse.c b/encoder/analyse.c
index 4ed9a71..99ca6cc 100644
--- a/encoder/analyse.c
+++ b/encoder/analyse.c
@@ -22,6 +22,7 @@
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
  *****************************************************************************/
 
+#define _ISOC99_SOURCE
 #include <math.h>
 #include <limits.h>
 #ifndef _MSC_VER
@@ -29,6 +30,7 @@
 #endif
 
 #include "common/common.h"
+#include "common/cpu.h"
 #include "macroblock.h"
 #include "me.h"
 #include "ratecontrol.h"
@@ -77,6 +79,8 @@ typedef struct
     int i_lambda2;
     int i_qp;
     int16_t *p_cost_mv;
+    uint16_t *p_cost_ref0;
+    uint16_t *p_cost_ref1;
     int i_mbrd;
 
 
@@ -168,6 +172,7 @@ static const int i_sub_mb_p_cost_table[4] = {
 static void x264_analyse_update_cache( x264_t *h, x264_mb_analysis_t *a );
 
 uint16_t *x264_cost_mv_fpel[52][4];
+uint16_t x264_cost_ref[52][3][33];
 
 /* initialize an array of lambda*nbits for all possible mvs */
 static void x264_mb_analyse_load_costs( x264_t *h, x264_mb_analysis_t *a )
@@ -177,6 +182,7 @@ static void x264_mb_analyse_load_costs( x264_t *h, x264_mb_analysis_t *a )
 
     if( !p_cost_mv[a->i_qp] )
     {
+        x264_emms();
         /* could be faster, but isn't called many times */
         /* factor of 4 from qpel, 2 from sign, and 2 because mv can be opposite from mvp */
         p_cost_mv[a->i_qp] = x264_malloc( (4*4*2048 + 1) * sizeof(int16_t) );
@@ -184,10 +190,15 @@ static void x264_mb_analyse_load_costs( x264_t *h, x264_mb_analysis_t *a )
         for( i = 0; i <= 2*4*2048; i++ )
         {
             p_cost_mv[a->i_qp][-i] =
-            p_cost_mv[a->i_qp][i]  = a->i_lambda * bs_size_se( i );
+            p_cost_mv[a->i_qp][i]  = a->i_lambda * (log2f(i+1)*2 + 0.718f + !!i) + .5f;
         }
+        for( i = 0; i < 3; i++ )
+            for( j = 0; j < 33; j++ )
+                x264_cost_ref[a->i_qp][i][j] = a->i_lambda * bs_size_te( i, j );
     }
     a->p_cost_mv = p_cost_mv[a->i_qp];
+    a->p_cost_ref0 = x264_cost_ref[a->i_qp][x264_clip3(h->sh.i_num_ref_idx_l0_active-1,0,2)];
+    a->p_cost_ref1 = x264_cost_ref[a->i_qp][x264_clip3(h->sh.i_num_ref_idx_l1_active-1,0,2)];
 
     /* FIXME is this useful for all me methods? */
     if( h->param.analyse.i_me_method >= X264_ME_ESA && !x264_cost_mv_fpel[a->i_qp][0] )
@@ -1038,7 +1049,7 @@ static void x264_intra_rd_refine( x264_t *h, x264_mb_analysis_t *a )
     (m)->integral = &h->mb.pic.p_integral[list][ref][(xoff)+(yoff)*(m)->i_stride[0]];
 
 #define REF_COST(list, ref) \
-    (a->i_lambda * bs_size_te( h->sh.i_num_ref_idx_l##list##_active - 1, ref ))
+    (a->p_cost_ref##list[ref])
 
 static void x264_mb_analyse_inter_p16x16( x264_t *h, x264_mb_analysis_t *a )
 {