[MPlayer-dev-eng] [PATCH]: hqdn3d.c: refactorize LowPassMul inmacro, 10%~20% faster on Athlon X2

[Loren Merritt]

On Fri, 9 Jan 2009, Zhou Zongyi wrote:

> And any idea about SIMD optimization on this? I tried SSE2 on 
> deNoiseTemporal but it runs even slower than original C code.

SIMD would have to compute the gamma curve every time while scalar can use 
a LUT. That basically kills any such idea.


I get 18% speedup with the attached patch on conroe x86_64 gcc-4.2.3.
It just removes some sign-extends from the inner loop.

--Loren Merritt
-------------- next part --------------
Index: libmpcodecs/vf_hqdn3d.c
===================================================================
--- libmpcodecs/vf_hqdn3d.c	(revision 28192)
+++ libmpcodecs/vf_hqdn3d.c	(working copy)
@@ -70,7 +70,7 @@
 static inline unsigned int LowPassMul(unsigned int PrevMul, unsigned int CurrMul, int* Coef){
 //    int dMul= (PrevMul&0xFFFFFF)-(CurrMul&0xFFFFFF);
     int dMul= PrevMul-CurrMul;
-    int d=((dMul+0x10007FF)>>12);
+    unsigned int d=((dMul+0x10007FF)>>12);
     return CurrMul + Coef[d];
 }
 
@@ -81,7 +81,7 @@
                     int W, int H, int sStride, int dStride,
                     int *Temporal)
 {
-    int X, Y;
+    long X, Y;
     unsigned int PixelDst;
 
     for (Y = 0; Y < H; Y++){
@@ -103,8 +103,8 @@
                     int W, int H, int sStride, int dStride,
                     int *Horizontal, int *Vertical)
 {
-    int X, Y;
-    int sLineOffs = 0, dLineOffs = 0;
+    long X, Y;
+    long sLineOffs = 0, dLineOffs = 0;
     unsigned int PixelAnt;
     unsigned int PixelDst;
     
@@ -143,8 +143,8 @@
                     int W, int H, int sStride, int dStride,
                     int *Horizontal, int *Vertical, int *Temporal)
 {
-    int X, Y;
-    int sLineOffs = 0, dLineOffs = 0;
+    long X, Y;
+    long sLineOffs = 0, dLineOffs = 0;
     unsigned int PixelAnt;
     unsigned int PixelDst;
     unsigned short* FrameAnt=(*FrameAntPtr);