shithub: amd64-simd

Info  •  Files  •  Log  •  Branches

add more benchmarks. extend the mkfile. use mallocalign(2).

--- a/bench/main.c

+++ b/bench/main.c

@@ -5,6 +5,7 @@

 #include "../bench9/b.h"

 double min(double, double);

+

 double dotvec2_sse(Point2, Point2);

 double dotvec2_sse4(Point2, Point2);

 double dotvec2_avx(Point2, Point2);

@@ -11,32 +12,24 @@

 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*);

+double dotvec3_avx_a(Point3*, Point3*);

+

 Point2 Pt2b(double, double, double);

+

 Point3 crossvec3_sse(Point3, Point3);

+

 double hsubpd(double, double);

+

 double fma(double, double, double);

+

 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)

 {

@@ -49,6 +42,18 @@

 	return a + b*c;

 }

+double

+dotvec2_p(Point2 *a, Point2 *b)

+{

+	return a->x*b->x + a->y*b->y;

+}

+

+double

+dotvec3_p(Point3 *a, Point3 *b)

+{

+	return a->x*b->x + a->y*b->y + a->z*b->z;

+}

+

 static void

 bmin(int fd)

 {

@@ -84,7 +89,7 @@

 bdotvec2(int fd)

 {

 	Bgr g;

-	B *b0, *b1, *b2, *b3, *b4, *b5, *b6;

+	B *b0, *b1, *b2, *b3, *b4, *b5, *b6, *b7;

 	Point2 a, b;

 	Point2 *aa, *bb;

 	int i;

@@ -97,12 +102,13 @@

 	b4 = benchadd(&g, "dotvec2_sse_a");

 	b5 = benchadd(&g, "dotvec2_sse4_a");

 	b6 = benchadd(&g, "dotvec2_avx_a");

+	b7 = benchadd(&g, "dotvec2_p");

 	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 = mallocalign(sizeof(Point2), 16, 0, 0);

+		bb = mallocalign(sizeof(Point2), 16, 0, 0);

 		*aa = a;

 		*bb = b;

@@ -140,6 +146,11 @@

 		for(i = 0; i < 1e6; i++)

 			dotvec2_avx_a(aa, bb);

 		benchout(b6);

+

+		benchin(b7);

+		for(i = 0; i < 1e6; i++)

+			dotvec2_p(aa, bb);

+		benchout(b7);

 	}

 	benchprintgr(&g, fd);

@@ -150,7 +161,7 @@

 bdotvec3(int fd)

 {

 	Bgr g;

-	B *b0, *b1, *b2, *b3;

+	B *b0, *b1, *b2, *b3, *b4, *b5;

 	Point3 a, b;

 	Point3 *aa, *bb;

 	int i;

@@ -160,12 +171,14 @@

 	b1 = benchadd(&g, "dotvec3_sse4");

 	b2 = benchadd(&g, "dotvec3_avx");

 	b3 = benchadd(&g, "dotvec3_sse4_a");

+	b4 = benchadd(&g, "dotvec3_avx_a");

+	b5 = benchadd(&g, "dotvec3_p");

 	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 = mallocalign(sizeof(Point3), 16, 0, 0);

+		bb = mallocalign(sizeof(Point3), 16, 0, 0);

 		*aa = a;

 		*bb = b;

@@ -188,6 +201,16 @@

 		for(i = 0; i < 1e6; i++)

 			dotvec3_sse4_a(aa, bb);

 		benchout(b3);

+

+		benchin(b4);

+		for(i = 0; i < 1e6; i++)

+			dotvec3_avx_a(aa, bb);

+		benchout(b4);

+

+		benchin(b5);

+		for(i = 0; i < 1e6; i++)

+			dotvec3_p(aa, bb);

+		benchout(b5);

 	}

 	benchprintgr(&g, fd);

--- a/dppd.s

+++ b/dppd.s

@@ -96,6 +96,17 @@

 	ADDSD X1, X0

 	RET

+TEXT dotvec3_avx_a(SB), 1, $0

+	MOVQ b+8(FP), DX

+	VMOVAPD_128mr(0, rDX, rX0)

+	VMOVAPD_128mr(0, rBP, rX1)

+	VDPPD(rX1, rX0, rX0)		/* VDPPD $0x31, X1, X0, X0 */

+	MOVSD 16(DX), X1

+	MOVSD 16(BP), X2

+	VFMADD231SD(rX1, rX2, rX0)

+	VZEROUPPER

+	RET

+

 TEXT Pt2b(SB), 1, $0

 	MOVQ BP, DI

 	MOVSD x+8(FP), X0

--- a/main.c

+++ b/main.c

@@ -3,6 +3,7 @@

 #include <geometry.h>

 double min(double, double);

+

 double dotvec2_sse(Point2, Point2);

 double dotvec2_sse4(Point2, Point2);

 double dotvec2_avx(Point2, Point2);

@@ -9,34 +10,28 @@

 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*);

+double dotvec3_avx_a(Point3*, Point3*);

+

 Point2 Pt2b(double, double, double);

+

 Point3 crossvec3_sse(Point3, Point3);

+

 double hsubpd(double, double);

+

 double fma(double, double, double);

+

 Point2 addpt2_sse(Point2, Point2);

 Point2 addpt2_avx(Point2, Point2);

 Point3 addpt3_avx(Point3, Point3);

+

 void addsub_sse(double*,double*);

+

 double round(double);

-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;

-}

-

 void

 addsub(double *a, double *b)

 {

@@ -56,6 +51,12 @@

 	return a + b*c;

 }

+double

+dotvec2_p(Point2 *a, Point2 *b)

+{

+	return a->x*b->x + a->y*b->y;

+}

+

 void

 main(int argc, char *argv[])

 {

@@ -73,13 +74,13 @@

 	a = strtod(argv[0], nil);

 	b = strtod(argv[1], nil);

-	ap0 = amalloc(sizeof(Point2), 16);

-	ap1 = amalloc(sizeof(Point2), 16);

-	apr = amalloc(sizeof(Point2), 16);

+	ap0 = mallocalign(sizeof(Point2), 16, 0, 0);

+	ap1 = mallocalign(sizeof(Point2), 16, 0, 0);

+	apr = mallocalign(sizeof(Point2), 16, 0, 0);

-	ap0t = amalloc(sizeof(Point3), 16);

-	ap1t = amalloc(sizeof(Point3), 16);

-	aprt = amalloc(sizeof(Point3), 16);

+	ap0t = mallocalign(sizeof(Point3), 16, 0, 0);

+	ap1t = mallocalign(sizeof(Point3), 16, 0, 0);

+	aprt = mallocalign(sizeof(Point3), 16, 0, 0);

 	r = 0;

 	r = fmin(a, b);

@@ -96,6 +97,9 @@

 	r = dotvec2(p0, p1);

 	print("dotvec2(%v, %v) = %g\n", p0, p1, r);

 	r = 0;

+	r = dotvec2_p(&p0, &p1);

+	print("dotvec2_p(%v, %v) = %g\n", p0, p1, r);

+	r = 0;

 	r = dotvec2_sse(p0, p1);

 	print("dotvec2_sse(%v, %v) = %g\n", p0, p1, r);

 	r = 0;

@@ -140,6 +144,9 @@

 	r = 0;

 	r = dotvec3_sse4_a(ap0t, ap1t);

 	print("dotvec3_sse4_a(%V, %V) = %g\n", *ap0t, *ap1t, r);

+	r = 0;

+	r = dotvec3_avx_a(ap0t, ap1t);

+	print("dotvec3_avx_a(%V, %V) = %g\n", *ap0t, *ap1t, r);

 	print("\n");

--- a/mkfile

+++ b/mkfile

@@ -17,3 +17,9 @@

 pulldeps:VQ:

 	git/clone git://shithub.us/sigrid/bench9 || \

 	git/clone https://git.sr.ht/~ft/bench9

+

+verify:VQ: $O.out

+	$O.out 13 31

+

+bench:VQ:

+	@{cd bench; mk && $O.out}