4 files changed, 58 insertions, 0 deletions

diff --git a/avx.h b/avx.h
index 3c7129c..eea76a1 100644
--- a/avx.h
+++ b/avx.h
@@ -122,3 +122,7 @@
 /* VFMADD231PD (256 bit) */
 #define VFMADD231PD_256(s0, s1, d)	VEX3(0,0,0,VEX_m_0F38,1,(s0),VEX_L_256,VEX_p_66);	\
 			VOP(0xB8, 0x3, (d), (s1))
+
+/* VINSERTF128 */
+#define VINSERTF128(i, s0, s1, d)	VEX3(0,0,0,VEX_m_0F3A,0,(s0),VEX_L_256,VEX_p_66);	\
+			VOPi(0x18, 0x3, (d), (s1), (i))
diff --git a/bench/main.c b/bench/main.c
index 811c471..1567e8c 100644
--- a/bench/main.c
+++ b/bench/main.c
@@ -14,6 +14,7 @@ Point3 crossvec3_sse(Point3, Point3);
 double hsubpd(double, double);
 double fma(double, double, double);
 Point2 addpt2_avx(Point2, Point2);
+Point3 addpt3_avx(Point3, Point3);
 
 double
 fmin(double a, double b)
@@ -258,6 +259,37 @@ baddpt2(int fd)
 	benchfreegr(&g);
 }
 
+static void
+baddpt3(int fd)
+{
+	Bgr g;
+	B *b0, *b1;
+	Point3 a, b;
+	int i;
+
+	benchinitgr(&g, "3d point sum");
+	b0 = benchadd(&g, "addpt3");
+	b1 = benchadd(&g, "addpt3_avx");
+
+	while(b0->n > 0 || b1->n > 0){
+		a = Pt3(truerand()*frand(), truerand()*frand(), truerand()*frand(), truerand()*frand());
+		b = Pt3(truerand()*frand(), truerand()*frand(), truerand()*frand(), truerand()*frand());
+
+		benchin(b0);
+		for(i = 0; i < 1e6; i++)
+			addpt3(a, b);
+		benchout(b0);
+
+		benchin(b1);
+		for(i = 0; i < 1e6; i++)
+			addpt3_avx(a, b);
+		benchout(b1);
+	}
+
+	benchprintgr(&g, fd);
+	benchfreegr(&g);
+}
+
 void
 threadmain(int argc, char **argv)
 {
@@ -280,6 +312,8 @@ threadmain(int argc, char **argv)
 	bfma(1);
 	bseparator(1);
 	baddpt2(1);
+	bseparator(1);
+	baddpt3(1);
 
 	threadexitsall(nil);
 }
diff --git a/dppd.s b/dppd.s
index b746117..db805ff 100644
--- a/dppd.s
+++ b/dppd.s
@@ -99,6 +99,7 @@ TEXT fma(SB), 1, $0
 	VFMADD231SD(rX1, rX2, rX0)
 	RET
 
+/* TODO: write only 24 bytes */
 TEXT addpt2_avx(SB), 1, $0
 	MOVQ SP, AX
 	ADDQ $8, AX
@@ -107,3 +108,12 @@ TEXT addpt2_avx(SB), 1, $0
 	VADDPD_256rr(rX1, rX0, rX0)
 	VMOVDQU_256rm(rX0, rAX)
 	RET
+
+TEXT addpt3_avx(SB), 1, $0
+	MOVQ SP, AX
+	ADDQ $8, AX
+	VMOVDQU_256mr(8, rAX, rX0)
+	VMOVDQU_256mr(40, rAX, rX1)
+	VADDPD_256rr(rX1, rX0, rX0)
+	VMOVDQU_256rm(rX0, rAX)
+	RET
diff --git a/main.c b/main.c
index 274daf4..02fc11c 100644
--- a/main.c
+++ b/main.c
@@ -12,6 +12,7 @@ Point3 crossvec3_sse(Point3, Point3);
 double hsubpd(double, double);
 double fma(double, double, double);
 Point2 addpt2_avx(Point2, Point2);
+Point3 addpt3_avx(Point3, Point3);
 
 double
 fmin(double a, double b)
@@ -96,5 +97,14 @@ main(int argc, char *argv[])
 	pr = addpt2_avx(p0, p1);
 	print("addpt2_avx(%v, %v) = %v\n", p0, p1, pr);
 
+	print("\n");
+
+	p0t = Pt3(a, 1, 1, b);
+	p1t = Pt3(b, 3, 1, a);
+	prt = addpt3(p0t, p1t);
+	print("addpt3(%V, %V) = %V\n", p0t, p1t, prt);
+	prt = addpt3_avx(p0t, p1t);
+	print("addpt3_avx(%V, %V) = %V\n", p0t, p1t, prt);
+
 	exits(nil);
 }