add more avx instructions and a bench9 benchmark file.

4 files changed, 182 insertions, 0 deletions

diff --git a/bench/main.c b/bench/main.c
new file mode 100644
index 0000000..9d00719
--- /dev/null
+++ b/bench/main.c
@@ -0,0 +1,134 @@
+#include <u.h>
+#include <libc.h>
+#include <thread.h>
+#include <geometry.h>
+#include "b.h"
+
+double dppd(Point2, Point2);
+double dppda(Point2, Point2);
+double dppd3(Point3, Point3);
+double dppd3a(Point3, Point3);
+Point3 xvec3(Point3, Point3);
+
+static void
+bdotvec2(int fd)
+{
+	Bgr g;
+	B *b0, *b1, *b2;
+	Point2 a, b;
+	int i;
+
+	benchinitgr(&g, "2d dot product");
+	b0 = benchadd(&g, "dotvec2");
+	b1 = benchadd(&g, "dotvec2_simd");
+	b2 = benchadd(&g, "dotvec2_avx");
+
+	while(b0->n > 0 || b1->n > 0){
+		a = Vec2(truerand()*frand(), truerand()*frand());
+		b = Vec2(truerand()*frand(), truerand()*frand());
+
+		benchin(b0);
+		for(i = 0; i < 1e6; i++)
+			dotvec2(a, b);
+		benchout(b0);
+
+		benchin(b1);
+		for(i = 0; i < 1e6; i++)
+			dppd(a, b);
+		benchout(b1);
+
+		benchin(b2);
+		for(i = 0; i < 1e6; i++)
+			dppda(a, b);
+		benchout(b2);
+	}
+
+	benchprintgr(&g, fd);
+	benchfreegr(&g);
+}
+
+static void
+bdotvec3(int fd)
+{
+	Bgr g;
+	B *b0, *b1, *b2;
+	Point3 a, b;
+	int i;
+
+	benchinitgr(&g, "3d dot product");
+	b0 = benchadd(&g, "dotvec3");
+	b1 = benchadd(&g, "dotvec3_simd");
+	b2 = benchadd(&g, "dotvec3_avx");
+
+	while(b0->n > 0 || b1->n > 0){
+		a = Vec3(truerand()*frand(), truerand()*frand(), truerand()*frand());
+		b = Vec3(truerand()*frand(), truerand()*frand(), truerand()*frand());
+
+		benchin(b0);
+		for(i = 0; i < 1e6; i++)
+			dotvec3(a, b);
+		benchout(b0);
+
+		benchin(b1);
+		for(i = 0; i < 1e6; i++)
+			dppd3(a, b);
+		benchout(b1);
+
+		benchin(b2);
+		for(i = 0; i < 1e6; i++)
+			dppd3a(a, b);
+		benchout(b2);
+	}
+
+	benchprintgr(&g, fd);
+	benchfreegr(&g);
+}
+
+static void
+bcrossvec3(int fd)
+{
+	Bgr g;
+	B *b0, *b1;
+	Point3 a, b;
+	int i;
+
+	benchinitgr(&g, "3d cross product");
+	b0 = benchadd(&g, "crossvec3");
+	b1 = benchadd(&g, "crossvec3_simd");
+
+	while(b0->n > 0 || b1->n > 0){
+		a = Vec3(truerand()*frand(), truerand()*frand(), truerand()*frand());
+		b = Vec3(truerand()*frand(), truerand()*frand(), truerand()*frand());
+
+		benchin(b0);
+		for(i = 0; i < 1e6; i++)
+			crossvec3(a, b);
+		benchout(b0);
+
+		benchin(b1);
+		for(i = 0; i < 1e6; i++)
+			xvec3(a, b);
+		benchout(b1);
+	}
+
+	benchprintgr(&g, fd);
+	benchfreegr(&g);
+}
+
+void
+threadmain(int argc, char **argv)
+{
+	ARGBEGIN{
+	}ARGEND
+
+	if(benchwire(0) != 0)
+		fprint(2, "failed to wire: %r\n");
+
+	bdotvec2(1);
+	bseparator(1);
+	bdotvec3(1);
+	bseparator(1);
+	bcrossvec3(1);
+
+	threadexitsall(nil);
+}
diff --git a/dppd.s b/dppd.s
index 907a437..7239d9f 100644
--- a/dppd.s
+++ b/dppd.s
@@ -12,6 +12,13 @@ TEXT dppd(SB), 1, $0
 	DPPD(rX1, rX0)		/* DPPD $0x31, X1, X0 */
 	RET
 
+TEXT dppda(SB), 1, $0
+	MOVQ SP, AX
+	VMOVUPD_128mr(8, rAX, rX0)	/* VMOVUPD a+0(FP), X0 */
+	VMOVUPD_128mr(32, rAX, rX1)	/* VMOVUPD b+24(FP), X1 */
+	VDPPD(rX1, rX0, rX0)		/* VDPPD $0x31, X1, X0, X0 */
+	RET
+
 TEXT dppd3(SB), 1, $0
 	MOVQ SP, AX
 	MOVLPD(8, rAX, rX0)	/* MOVLPD a+0(FP), X0 */
@@ -25,6 +32,16 @@ TEXT dppd3(SB), 1, $0
 	DPPD(rX1, rX0)		/* DPPD $0x31, X1, X0 */
 	RET
 
+TEXT dppd3a(SB), 1, $0
+	MOVQ SP, AX
+	VMOVUPD_128mr(8, rAX, rX0)	/* VMOVUPD a+0(FP), X0 */
+	VMOVUPD_128mr(40, rAX, rX1)	/* VMOVUPD b+32(FP), X1 */
+	VDPPD(rX1, rX0, rX0)		/* VDPPD $0x31, X1, X0, X0 */
+	MOVSD a+16(FP), X1
+	MOVSD b+48(FP), X2
+	VFMADD231SD(rX1, rX2, rX0)
+	RET
+
 TEXT Pt2b(SB), 1, $0
 	MOVQ BP, DI
 	MOVSD x+8(FP), X0
diff --git a/main.c b/main.c
index fe51889..f14a7ed 100644
--- a/main.c
+++ b/main.c
@@ -5,7 +5,9 @@
 uvlong nanosec(void);
 double min(double, double);
 double dppd(Point2, Point2);
+double dppda(Point2, Point2);
 double dppd3(Point3, Point3);
+double dppd3a(Point3, Point3);
 Point2 Pt2b(double, double, double);
 Point3 xvec3(Point3, Point3);
 double hsubpd(double, double);
@@ -41,6 +43,8 @@ main(int argc, char *argv[])
 	t1 = nanosec();
 	print("min(%g, %g) = %g\ttook %lludns\n", a, b, r, t1-t0);
 
+	print("\n");
+
 	p0 = Pt2b(a, 1, 1);
 	p1 = Pt2b(b, 3, 1);
 	t0 = nanosec();
@@ -51,6 +55,12 @@ main(int argc, char *argv[])
 	r = dotvec2(p0, p1);
 	t1 = nanosec();
 	print("dotvec2(%v, %v) = %g\ttook %lludns\n", p0, p1, r, t1-t0);
+	t0 = nanosec();
+	r = dppda(p0, p1);
+	t1 = nanosec();
+	print("dppda(%v, %v) = %g\ttook %lludns\n", p0, p1, r, t1-t0);
+
+	print("\n");
 
 	p0t = Pt3(a, 1, 9, 1);
 	p1t = Pt3(b, 3, 4, 1);
@@ -62,12 +72,20 @@ main(int argc, char *argv[])
 	r = dotvec3(p0t, p1t);
 	t1 = nanosec();
 	print("dotvec3(%V, %V) = %g\ttook %lludns\n", p0t, p1t, r, t1-t0);
+	t0 = nanosec();
+	r = dppd3a(p0t, p1t);
+	t1 = nanosec();
+	print("dppd3a(%V, %V) = %g\ttook %lludns\n", p0t, p1t, r, t1-t0);
+
+	print("\n");
 
 	t0 = nanosec();
 	r = hsubpd(a, b);
 	t1 = nanosec();
 	print("hsubpd(%g, %g) = %g\ttook %lludns\n", a, b, r, t1-t0);
 
+	print("\n");
+
 	p0t = Pt3(a, 1, 9, 1);
 	p1t = Pt3(b, 3, 4, 1);
 	t0 = nanosec();
@@ -79,6 +97,8 @@ main(int argc, char *argv[])
 	t1 = nanosec();
 	print("crossvec3(%V, %V) = %V\ttook %lludns\n", p0t, p1t, pr, t1-t0);
 
+	print("\n");
+
 	t0 = nanosec();
 	r = fma(a, b, 21);
 	t1 = nanosec();
diff --git a/sse.h b/sse.h
index ffad543..6a03c9e 100644
--- a/sse.h
+++ b/sse.h
@@ -41,6 +41,8 @@
 			BYTE $(((~r)<<7)|((~v)<<3)|((l)<<2)|(p))
 #define VOP(o, m, ro, rm)	BYTE $(o);	\
 			BYTE $(((m)<<6)|((ro)<<3)|(rm))
+#define VOPi(o, m, ro, rm, i)	VOP((o), (m), (ro), (rm));	\
+			BYTE $(i)
 
 /* MOVLPD */
 //opcode = 660F12
@@ -65,6 +67,15 @@
 //imm8   = 0011 0001
 #define DPPD(s, d) OP4i(0x413A, 0x3, (d), (s), 0x31)
 
+/* VMOVAPD */
+#define VMOVUPD_128mr(off, s, d) VEX3(0,0,0,VEX_m_0F,0,0,VEX_L_128,VEX_p_66);			\
+			VOPi(0x10, 0x1, (d), (s), (off))
+#define VMOVAPD_128rr(s, d) VEX3(0,0,0,VEX_m_0F,0,0,VEX_L_128,VEX_p_66);			\
+			VOP(0x28, 0x3, (d), (s))
+/* VDPPD */
+#define VDPPD(s0, s1, d) VEX3(0,0,0,VEX_m_0F3A,0,(s0),VEX_L_128,VEX_p_66);		\
+			VOPi(0x41, 0x3, (d), (s1), 0x31)
+
 /* VFMADD231SD (128 bit) */
 #define VFMADD231SD(s0, s1, d) VEX3(0,0,0,VEX_m_0F38,1,(s0),VEX_L_128,VEX_p_66);	\
 			VOP(0xB9, 0x3, (d), (s1))