Move the v{Y,C}CoeffsBank vectors into the SwsContext, filling them in just

once when the scaler is initialized, instead of building them and freeing
them over and over. This gives massive performance improvements.
patch by Alan Curry, pacman*at*TheWorld*dot*com


git-svn-id: file:///var/local/repositories/mplayer/trunk/postproc@17589 b3059339-0415-0410-9bf9-f77b7e298cf2

swscale.c

@@ -2110,6 +2110,25 @@ SwsContext *sws_getContext(int srcW, int srcH, int origSrcFormat, int dstW, int
 				c->chrSrcH, c->chrDstH, filterAlign, (1<<12)-4,
 				(flags&SWS_BICUBLIN) ? (flags|SWS_BILINEAR) : flags,
 				srcFilter->chrV, dstFilter->chrV, c->param);
+
+#ifdef HAVE_ALTIVEC
+		c->vYCoeffsBank = memalign (16, sizeof (vector signed short)*c->vLumFilterSize*c->dstH);
+		c->vCCoeffsBank = memalign (16, sizeof (vector signed short)*c->vChrFilterSize*c->dstH);
+
+		for (i=0;i<c->vLumFilterSize*c->dstH;i++) {
+                  int j;
+		  short *p = (short *)&c->vYCoeffsBank[i];
+		  for (j=0;j<8;j++)
+		    p[j] = c->vLumFilter[i];
+		}
+
+		for (i=0;i<c->vChrFilterSize*c->dstH;i++) {
+                  int j;
+		  short *p = (short *)&c->vCCoeffsBank[i];
+		  for (j=0;j<8;j++)
+		    p[j] = c->vChrFilter[i];
+		}
+#endif
 	}
 
 	// Calculate Buffer Sizes so that they won't run out while handling these damn slices
@@ -2644,6 +2663,12 @@ void sws_freeContext(SwsContext *c){
 	c->hLumFilter = NULL;
 	if(c->hChrFilter) free(c->hChrFilter);
 	c->hChrFilter = NULL;
+#ifdef HAVE_ALTIVEC
+	if(c->vYCoeffsBank) free(c->vYCoeffsBank);
+	c->vYCoeffsBank = NULL;
+	if(c->vCCoeffsBank) free(c->vCCoeffsBank);
+	c->vCCoeffsBank = NULL;
+#endif
 
 	if(c->vLumFilterPos) free(c->vLumFilterPos);
 	c->vLumFilterPos = NULL;

swscale_internal.h

@@ -154,6 +154,7 @@ typedef struct SwsContext{
   vector signed short   CGV;
   vector signed short   OY;
   vector unsigned short CSHIFT;
+  vector signed short *vYCoeffsBank, *vCCoeffsBank;
 
 #endif
 

yuv2rgb_altivec.c

@@ -774,8 +774,6 @@ altivec_yuv2packedX (SwsContext *c,
 		       uint8_t *dest, int dstW, int dstY)
 {
   int i,j;
-  short tmp __attribute__((aligned (16)));
-  int16_t *p;
   short *f;
   vector signed short X,X0,X1,Y0,U0,V0,Y1,U1,V1,U,V;
   vector signed short R0,G0,B0,R1,G1,B1;
@@ -787,29 +785,10 @@ altivec_yuv2packedX (SwsContext *c,
   vector unsigned short SCL = vec_splat((vector unsigned short)AVV(4),0);
   unsigned long scratch[16] __attribute__ ((aligned (16)));
 
-  vector signed short *vYCoeffsBank, *vCCoeffsBank;
-
   vector signed short *YCoeffs, *CCoeffs;
 
-  vYCoeffsBank = memalign (16, sizeof (vector signed short)*lumFilterSize*c->dstH);
-  vCCoeffsBank = memalign (16, sizeof (vector signed short)*chrFilterSize*c->dstH);
-
-  for (i=0;i<lumFilterSize*c->dstH;i++) {
-    tmp = c->vLumFilter[i];
-    p = &vYCoeffsBank[i];
-    for (j=0;j<8;j++)
-      p[j] = tmp;
-  }
-
-  for (i=0;i<chrFilterSize*c->dstH;i++) {
-    tmp = c->vChrFilter[i];
-    p = &vCCoeffsBank[i];
-    for (j=0;j<8;j++)
-      p[j] = tmp;
-  }
-
-  YCoeffs = vYCoeffsBank+dstY*lumFilterSize;
-  CCoeffs = vCCoeffsBank+dstY*chrFilterSize;
+  YCoeffs = c->vYCoeffsBank+dstY*lumFilterSize;
+  CCoeffs = c->vCCoeffsBank+dstY*chrFilterSize;
 
   out = (vector unsigned char *)dest;
 
@@ -962,7 +941,4 @@ altivec_yuv2packedX (SwsContext *c,
     memcpy (&((uint32_t*)dest)[i], scratch, (dstW-i)/4);
   }
 
-  if (vYCoeffsBank) free (vYCoeffsBank);
-  if (vCCoeffsBank) free (vCCoeffsBank);
-
 }