implement memory aligned versions of some functions.

1 files changed, 55 insertions, 2 deletions

diff --git a/bench/main.c b/bench/main.c
index 6f4886f..514fb7c 100644
--- a/bench/main.c
+++ b/bench/main.c
@@ -8,8 +8,12 @@ double min(double, double);
 double dotvec2_sse(Point2, Point2);
 double dotvec2_sse4(Point2, Point2);
 double dotvec2_avx(Point2, Point2);
+double dotvec2_sse_a(Point2*, Point2*);
+double dotvec2_sse4_a(Point2*, Point2*);
+double dotvec2_avx_a(Point2*, Point2*);
 double dotvec3_sse4(Point3, Point3);
 double dotvec3_avx(Point3, Point3);
+double dotvec3_sse4_a(Point3*, Point3*);
 Point2 Pt2b(double, double, double);
 Point3 crossvec3_sse(Point3, Point3);
 double hsubpd(double, double);
@@ -18,6 +22,21 @@ Point2 addpt2_sse(Point2, Point2);
 Point2 addpt2_avx(Point2, Point2);
 Point3 addpt3_avx(Point3, Point3);
 
+void *
+amalloc(ulong n, ulong a)
+{
+	void *p;
+
+	assert(a > 1 && (a&1) == 0);
+
+	a--;
+	p = malloc(n+a);
+	if(p == nil)
+		sysfatal("malloc: %r");
+	p = (void*)(((uintptr)p + a)&~a);
+	return p;
+}
+
 double
 fmin(double a, double b)
 {
@@ -65,8 +84,9 @@ static void
 bdotvec2(int fd)
 {
 	Bgr g;
-	B *b0, *b1, *b2, *b3;
+	B *b0, *b1, *b2, *b3, *b4, *b5, *b6;
 	Point2 a, b;
+	Point2 *aa, *bb;
 	int i;
 
 	benchinitgr(&g, "2d dot product");
@@ -74,10 +94,17 @@ bdotvec2(int fd)
 	b1 = benchadd(&g, "dotvec2_sse");
 	b2 = benchadd(&g, "dotvec2_sse4");
 	b3 = benchadd(&g, "dotvec2_avx");
+	b4 = benchadd(&g, "dotvec2_sse_a");
+	b5 = benchadd(&g, "dotvec2_sse4_a");
+	b6 = benchadd(&g, "dotvec2_avx_a");
 
 	while(b0->n > 0 || b1->n > 0){
 		a = Vec2(truerand()*frand(), truerand()*frand());
 		b = Vec2(truerand()*frand(), truerand()*frand());
+		aa = amalloc(sizeof(Point2), 16);
+		bb = amalloc(sizeof(Point2), 16);
+		*aa = a;
+		*bb = b;
 
 		benchin(b0);
 		for(i = 0; i < 1e6; i++)
@@ -98,6 +125,21 @@ bdotvec2(int fd)
 		for(i = 0; i < 1e6; i++)
 			dotvec2_avx(a, b);
 		benchout(b3);
+
+		benchin(b4);
+		for(i = 0; i < 1e6; i++)
+			dotvec2_sse_a(aa, bb);
+		benchout(b4);
+
+		benchin(b5);
+		for(i = 0; i < 1e6; i++)
+			dotvec2_sse4_a(aa, bb);
+		benchout(b5);
+
+		benchin(b6);
+		for(i = 0; i < 1e6; i++)
+			dotvec2_avx_a(aa, bb);
+		benchout(b6);
 	}
 
 	benchprintgr(&g, fd);
@@ -108,18 +150,24 @@ static void
 bdotvec3(int fd)
 {
 	Bgr g;
-	B *b0, *b1, *b2;
+	B *b0, *b1, *b2, *b3;
 	Point3 a, b;
+	Point3 *aa, *bb;
 	int i;
 
 	benchinitgr(&g, "3d dot product");
 	b0 = benchadd(&g, "dotvec3");
 	b1 = benchadd(&g, "dotvec3_sse4");
 	b2 = benchadd(&g, "dotvec3_avx");
+	b3 = benchadd(&g, "dotvec3_sse4_a");
 
 	while(b0->n > 0 || b1->n > 0){
 		a = Vec3(truerand()*frand(), truerand()*frand(), truerand()*frand());
 		b = Vec3(truerand()*frand(), truerand()*frand(), truerand()*frand());
+		aa = amalloc(sizeof(Point3), 16);
+		bb = amalloc(sizeof(Point3), 16);
+		*aa = a;
+		*bb = b;
 
 		benchin(b0);
 		for(i = 0; i < 1e6; i++)
@@ -135,6 +183,11 @@ bdotvec3(int fd)
 		for(i = 0; i < 1e6; i++)
 			dotvec3_avx(a, b);
 		benchout(b2);
+
+		benchin(b3);
+		for(i = 0; i < 1e6; i++)
+			dotvec3_sse4_a(aa, bb);
+		benchout(b3);
 	}
 
 	benchprintgr(&g, fd);