From 99ce10873df41ba650e51c8069280a32a1f57961 Mon Sep 17 00:00:00 2001
From: Michael Niedermayer <michaelni@gmx.at>
Date: Sun, 8 Oct 2006 11:24:37 +0000
Subject: [PATCH] 5% faster get_cabac()

Originally committed as revision 6586 to svn://svn.ffmpeg.org/ffmpeg/trunk
---
 libavcodec/cabac.c | 16 ++++++++--------
 libavcodec/cabac.h | 24 ++++++++++++++++++++----
 libavcodec/h264.c  |  4 ++--
 3 files changed, 30 insertions(+), 14 deletions(-)

diff --git a/libavcodec/cabac.c b/libavcodec/cabac.c
index 86eb4696cb..8dd00a3cd9 100644
--- a/libavcodec/cabac.c
+++ b/libavcodec/cabac.c
@@ -133,19 +133,19 @@ void ff_init_cabac_states(CABACContext *c, uint8_t const (*lps_range)[4],
 
     for(i=0; i<state_count; i++){
         for(j=0; j<4; j++){ //FIXME check if this is worth the 1 shift we save
-            c->lps_range[2*i+0][j+4]=
-            c->lps_range[2*i+1][j+4]= lps_range[i][j];
+            c->lps_range[2*i+2][j+4]=
+            c->lps_range[2*i+3][j+4]= lps_range[i][j];
         }
 
-        c->mps_state[2*i+0]= 2*mps_state[i];
-        c->mps_state[2*i+1]= 2*mps_state[i]+1;
+        c->mps_state[2*i+2]= 2*mps_state[i]+2;
+        c->mps_state[2*i+3]= 2*mps_state[i]+3;
 
         if( i ){
-            c->lps_state[2*i+0]= 2*lps_state[i];
-            c->lps_state[2*i+1]= 2*lps_state[i]+1;
+            c->lps_state[2*i+2]= 2*lps_state[i]+2;
+            c->lps_state[2*i+3]= 2*lps_state[i]+3;
         }else{
-            c->lps_state[2*i+0]= 1;
-            c->lps_state[2*i+1]= 0;
+            c->lps_state[2*i+2]= 3;
+            c->lps_state[2*i+3]= 2;
         }
     }
 }
diff --git a/libavcodec/cabac.h b/libavcodec/cabac.h
index 18fd51d56a..a71b75a0f4 100644
--- a/libavcodec/cabac.h
+++ b/libavcodec/cabac.h
@@ -39,9 +39,9 @@ typedef struct CABACContext{
 #ifdef STRICT_LIMITS
     int symCount;
 #endif
-    uint8_t lps_range[2*65][4];   ///< rangeTabLPS
-    uint8_t lps_state[2*64];      ///< transIdxLPS
-    uint8_t mps_state[2*64];      ///< transIdxMPS
+    uint8_t lps_range[2*66][4];   ///< rangeTabLPS
+    uint8_t lps_state[2*65];      ///< transIdxLPS
+    uint8_t mps_state[2*65];      ///< transIdxMPS
     const uint8_t *bytestream_start;
     const uint8_t *bytestream;
     const uint8_t *bytestream_end;
@@ -376,7 +376,23 @@ static int get_cabac(CABACContext *c, uint8_t * const state){
 #if 1
     if(c->low < c->range){
         bit= s&1;
-        *state= c->mps_state[s];
+#ifdef ARCH_X86
+    //P3:627
+asm(
+        "addb $2, %b0       \n\t"
+        " js 1f             \n\t"
+        "movb %b0, %1       \n\t"
+        "1:                 \n\t"
+        : "+q"(s), "=m"(*state)
+);
+#else
+        *state= c->mps_state[s]; //P3:655
+/*        if(s<126) //P3:657
+            *state= s+2;*/
+        s+=2; //P3:631
+        if(s<128)
+            *state= s;
+#endif
         renorm_cabac_decoder_once(c);
     }else{
 //        int shift= ff_h264_norm_shift[RangeLPS>>17];
diff --git a/libavcodec/h264.c b/libavcodec/h264.c
index 5afda47f64..060d38699e 100644
--- a/libavcodec/h264.c
+++ b/libavcodec/h264.c
@@ -7399,9 +7399,9 @@ static int decode_slice(H264Context *h){
                 pre = clip( ((cabac_context_init_PB[h->cabac_init_idc][i][0] * s->qscale) >>4 ) + cabac_context_init_PB[h->cabac_init_idc][i][1], 1, 126 );
 
             if( pre <= 63 )
-                h->cabac_state[i] = 2 * ( 63 - pre ) + 0;
+                h->cabac_state[i] = 2 * ( 63 - pre ) + 2;
             else
-                h->cabac_state[i] = 2 * ( pre - 64 ) + 1;
+                h->cabac_state[i] = 2 * ( pre - 64 ) + 3;
         }
 
         for(;;){