3 files changed, 76 insertions, 22 deletions

diff --git a/externals/grill/xsample/source/groove.cpp b/externals/grill/xsample/source/groove.cpp
index 4759812c..f70846f7 100644
--- a/externals/grill/xsample/source/groove.cpp
+++ b/externals/grill/xsample/source/groove.cpp
@@ -498,6 +498,11 @@ V xgroove::s_pos_loop(I n,S *const *invecs,S *const *outvecs)
 	S *pos = outvecs[outchns];
 	BL lpbang = false;
 
+#ifdef __VEC__
+	// prefetch cache
+	vec_dst(speed,GetPrefetchConstant(1,n>>2,0),0);
+#endif
+
 	const D smin = curmin,smax = curmax,plen = smax-smin; //curlen;
 
 	if(buf && plen > 0) {
@@ -529,6 +534,10 @@ V xgroove::s_pos_loop(I n,S *const *invecs,S *const *outvecs)
 	else 
 		s_pos_off(n,invecs,outvecs);
 		
+#ifdef __VEC__
+	vec_dss(0);
+#endif
+
 	if(lpbang) ToOutBang(outchns+3);
 }
 
diff --git a/externals/grill/xsample/source/inter.h b/externals/grill/xsample/source/inter.h
index 3f8ea5f8..d5f2b4ac 100755
--- a/externals/grill/xsample/source/inter.h
+++ b/externals/grill/xsample/source/inter.h
@@ -38,7 +38,7 @@ TMPLDEF V xinter::st_play1(const S *bdt,const I smin,const I smax,const I n,cons
 		}
 		else if(oint >= smax) {
 			// position > last sample ... take only last sample
-			fp = bdt+(smax-1)*BCHNS;
+			fp = bdt+(smin == smax?smin:smax-1)*BCHNS;
 		}
 		else {
 			// normal
@@ -116,50 +116,73 @@ TMPLDEF V xinter::st_play4(const S *bdt,const I smin,const I smax,const I n,cons
 
 	// position info are frame units
 	const S *pos = invecs[0];
+
+#ifdef __VEC__
+	// prefetch cache
+	vec_dst(pos,GetPrefetchConstant(1,n>>2,0),0);
+	const int pf = GetPrefetchConstant(BCHNS,1,16*BCHNS);
+#endif
+
 	S *const *sig = outvecs;
 	register I si = 0;
 	
 	// 4-point interpolation
 	// ---------------------
 	const I maxo = smax-1; // last sample in play region
+	const S *maxp = bdt+maxo*BCHNS;  // pointer to last sample
 
 	for(I i = 0; i < n; ++i,++si) {	
 		F o = *(pos++);
-		register I oint = (I)o,ointm,oint1,oint2;
+		register I oint = (I)o;
+		register F frac;
+		register const S *fa,*fb,*fc,*fd;
 
 		if(oint <= smin) { 
 			if(oint < smin) oint = smin,o = (float)smin;
-			// position is first simple
-			ointm = smin; // first sample 
-			oint1 = oint+1;
-			oint2 = oint1+1;
+			
+			fa = bdt+smin*BCHNS;  // position is first sample
+			fb = bdt+oint*BCHNS;
+						
+			frac = o-oint;
+			fc = fb+BCHNS;
+			fd = fc+BCHNS;
 		}
 		else if(oint >= maxo-2) { 
 			if(oint > maxo) oint = maxo,o = (float)smax;
-			ointm = oint-1;
-			oint1 = oint >= maxo?maxo:oint+1;
-			oint2 = oint1 >= maxo?maxo:oint1+1;
+			frac = o-oint;
+
+			fb = bdt+oint*BCHNS;
+			fa = fb-BCHNS;   // CACHE!
+			
+			// \TODO what about wrap-around???
+			fc = fb >= maxp?maxp:fb+BCHNS;	// ev. CACHE!		
+			fd = fc >= maxp?maxp:fc+BCHNS;	// ev. CACHE!
 		}
 		else {
-			ointm = oint-1;
-			oint1 = oint+1;
-			oint2 = oint1+1;
+			fa = bdt+oint*BCHNS-BCHNS;
+			frac = o-oint;
+			fb = fa+BCHNS;
+#ifdef __VEC__
+			vec_dst(fa,pf,0);
+#endif
+			fc = fb+BCHNS;
+			fd = fc+BCHNS;
 		}
 
-		register F frac = o-oint;
+		register F f1 = 0.5f*(frac-1.0f);
+		register F f3 = frac*3.0f-1.0f;
 		
-		register const S *fa = bdt+ointm*BCHNS;
-		register const S *fb = bdt+oint*BCHNS;
-		register const S *fc = bdt+oint1*BCHNS;
-		register const S *fd = bdt+oint2*BCHNS;
-
 		for(I ci = 0; ci < OCHNS; ++ci) {
+			const F amdf = (fa[ci]-fd[ci])*frac;
 			const F cmb = fc[ci]-fb[ci];
-			sig[ci][si] = fb[ci] + frac*( 
-				cmb - 0.5f*(frac-1.0f) * ((fa[ci]-fd[ci]+3.0f*cmb)*frac + (fb[ci]-fa[ci]-cmb))
-			);
+			const F bma = fb[ci]-fa[ci];
+			sig[ci][si] = fb[ci] + frac*( cmb - f1 * ( amdf+bma+cmb*f3 ) );
 		}
 	}
+	
+#ifdef __VEC__
+	vec_dss(0);
+#endif
 
 	// clear rest of output channels (if buffer has less channels)
 	for(I ci = OCHNS; ci < outchns; ++ci) ZeroSamples(sig[ci],n);
diff --git a/externals/grill/xsample/source/main.h b/externals/grill/xsample/source/main.h
index b3b8a499..61c0bf4d 100644
--- a/externals/grill/xsample/source/main.h
+++ b/externals/grill/xsample/source/main.h
@@ -12,7 +12,7 @@ WARRANTIES, see the file, "license.txt," in this distribution.
 #define __XSAMPLE_H
 
 
-#define XSAMPLE_VERSION "0.3.0pre12"
+#define XSAMPLE_VERSION "0.3.0pre13"
 
 
 #define FLEXT_ATTRIBUTES 1 
@@ -72,6 +72,28 @@ WARRANTIES, see the file, "license.txt," in this distribution.
 	#define STD
 #endif
 
+#ifdef __ALTIVEC__
+#if FLEXT_CPU == FLEXT_CPU_PPC && defined(__MWERKS__)
+	#pragma altivec_model on
+	#include <vBasicOps.h>
+	#include <vectorOps.h>
+#elif FLEXT_CPU == FLEXT_CPU_PPC && defined(__GNUG__)
+	#include <vecLib/vBasicOps.h>
+	#include <vecLib/vectorOps.h>
+#endif
+
+	// Initialize a prefetch constant for use with vec_dst(), vec_dstt(), vec_dstst or vec_dststt
+	// Taken from the "AltiVec tutorial" by Ian Ollmann, Ph.D. 
+	inline UInt32 GetPrefetchConstant( int blockSizeInVectors,int blockCount,int blockStride )
+	{
+//		FLEXT_ASSERT( blockSizeInVectors > 0 && blockSizeInVectors <= 32 );
+//		FLEXT_ASSERT( blockCount > 0 && blockCount <= 256 );
+//		FLEXT_ASSERT( blockStride > MIN_SHRT && blockStride <= MAX_SHRT );
+		return ((blockSizeInVectors << 24) & 0x1F000000) |
+			((blockCount << 16) && 0x00FF0000) |
+			(blockStride & 0xFFFF);
+	}
+#endif
 
 class xsample:
 	public flext_dsp