asm implementation of mid_pred.

20% faster huffyuv decoding, 4% faster ffv1.


git-svn-id: file:///var/local/repositories/ffmpeg/trunk@6254 9553f0bf-9b14-0410-a0b8-cfaf0461ba5b

libavutil/common.h

@@ -235,7 +235,22 @@ static inline int av_log2_16bit(unsigned int v)
 /* median of 3 */
 static inline int mid_pred(int a, int b, int c)
 {
-#if 0
+#if (defined(ARCH_X86) && __CPU__ >= 686 || defined(ARCH_X86_64)) && !defined(RUNTIME_CPUDETECT)
+    int i=a, j=a;
+    asm volatile(
+        "cmp    %4, %2 \n\t"
+        "cmovg  %4, %0 \n\t"
+        "cmovl  %4, %1 \n\t"
+        "cmp    %4, %3 \n\t"
+        "cmovg  %3, %0 \n\t"
+        "cmovl  %3, %1 \n\t"
+        "cmp    %3, %2 \n\t"
+        "cmovl  %1, %0 \n\t"
+        :"+&r"(i), "+&r"(j)
+        :"r"(a), "r"(b), "r"(c)
+    );
+    return i;
+#elif 0
     int t= (a-b)&((a-b)>>31);
     a-=t;
     b+=t;