[Mplayer-cvslog] CVS: main/libavcodec dsputil.c,1.3,1.4

[Jürgen Keil]

Update of /cvsroot/mplayer/main/libavcodec
In directory mplayer:/var/tmp.root/cvs-serv28069

Modified Files:
	dsputil.c 
Log Message:
Another small optimization for ffmpeg:  Copy DCT blocks using 32-bit moves
on x86 cpus.  Reduces video decoding time by another 10%.


Index: dsputil.c
===================================================================
RCS file: /cvsroot/mplayer/main/libavcodec/dsputil.c,v
retrieving revision 1.3
retrieving revision 1.4
diff -u -r1.3 -r1.4
--- dsputil.c	10 Jul 2001 21:34:14 -0000	1.3
+++ dsputil.c	12 Jul 2001 18:11:03 -0000	1.4
@@ -22,6 +22,29 @@
 #include "dsputil.h"
 #include "mpegvideo.h"
 
+#ifdef ARCH_X86
+/*
+ * x86 CPUs can access unaligned 32-bit data, so we can speed up copying
+ * DCT blocks a bit by using 32-bit data moves instead of 8-bit moves
+ * (fewer instructions in the inner loop)
+ */
+static void put_pixels_long(UINT8 *block, const UINT8 *pixels, int line_size, int h)
+{
+    UINT32 *p;
+    const UINT32 *pix;
+
+    p   = (UINT32*)block;
+    pix = (UINT32*)pixels;
+    do {
+        p[0] = pix[0];
+        p[1] = pix[1];
+        pix = (UINT32*) ((char*)pix + line_size);
+        p =   (UINT32*) ((char*)p   + line_size);
+    } while (--h);
+}
+#endif
+
+
 #ifndef HAVE_MMX
 static UINT8 cropTbl[256 + 2 * MAX_NEG_CROP];
 #endif
@@ -40,6 +63,11 @@
     for(i=0;i<512;i++) {
         squareTbl[i] = (i - 256) * (i - 256);
     }
+
+#ifdef ARCH_X86
+    put_pixels_tab[0] = put_pixels_long;
+    put_no_rnd_pixels_tab[0] = put_pixels_long;
+#endif
 }
 
 void get_pixels(DCTELEM *block, const UINT8 *pixels, int line_size)