Merge pull request #208 from gonum/c64/dot

asm/c64: Adding dot product asm functions.
gonum · Sep 13, 2017 · 7acac0b · 7acac0b
2 parents 00ece78 + 9523cc3
commit 7acac0b
Show file tree

Hide file tree

Showing 12 changed files with 1,156 additions and 138 deletions.
diff --git a/internal/asm/c64/benchDot_test.go b/internal/asm/c64/benchDot_test.go
@@ -0,0 +1,59 @@
+// Copyright ©2017 The gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build go1.7
+
+package c64
+
+import (
+	"fmt"
+	"testing"
+)
+
+var benchSink complex64
+
+func BenchmarkDotUnitary(t *testing.B) {
+	for _, tst := range []struct {
+		name string
+		f    func(x, y []complex64) complex64
+	}{
+		{"DotcUnitary", DotcUnitary},
+		{"DotuUnitary", DotuUnitary},
+	} {
+		for _, v := range []int64{1, 2, 3, 4, 5, 10, 100, 1e3, 5e3, 1e4, 5e4} {
+			t.Run(fmt.Sprintf("%s-%d", tst.name, v), func(b *testing.B) {
+				x, y := x[:v], y[:v]
+				b.SetBytes(128 * v)
+				for i := 0; i < b.N; i++ {
+					benchSink = tst.f(x, y)
+				}
+			})
+		}
+	}
+}
+
+func BenchmarkDotInc(t *testing.B) {
+	for _, tst := range []struct {
+		name string
+		f    func(x, y []complex64, n, incX, incY, ix, iy uintptr) complex64
+	}{
+		{"DotcInc", DotcInc},
+		{"DotuInc", DotuInc},
+	} {
+		for _, ln := range []int{1, 2, 3, 4, 5, 10, 100, 1e3, 5e3, 1e4, 5e4} {
+			for _, inc := range []int{1, 2, 4, 10, -1, -2, -4, -10} {
+				t.Run(fmt.Sprintf("%s-%d-inc%d", tst.name, ln, inc), func(b *testing.B) {
+					b.SetBytes(int64(128 * ln))
+					var idx int
+					if inc < 0 {
+						idx = (-ln + 1) * inc
+					}
+					for i := 0; i < b.N; i++ {
+						benchSink = tst.f(x, y, uintptr(ln), uintptr(inc), uintptr(inc), uintptr(idx), uintptr(idx))
+					}
+				})
+			}
+		}
+	}
+}
diff --git a/internal/asm/c64/dot_test.go b/internal/asm/c64/dot_test.go
@@ -0,0 +1,178 @@
+// Copyright ©2017 The gonum Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package c64
+
+import (
+	"fmt"
+	"math"
+	"testing"
+)
+
+const (
+	msgVal   = "%v: unexpected value at %v Got: %v Expected: %v"
+	msgGuard = "%v: Guard violated in %s vector %v %v"
+)
+
+var (
+	inf = float32(math.Inf(1))
+)
+
+var dotTests = []struct {
+	x, y               []complex64
+	wantu, wantc       complex64
+	wantuRev, wantcRev complex64
+	n                  int
+}{
+	{
+		x:     []complex64{},
+		y:     []complex64{},
+		n:     0,
+		wantu: 0, wantc: 0,
+		wantuRev: 0, wantcRev: 0,
+	},
+	{
+		x:     []complex64{1 + 1i},
+		y:     []complex64{1 + 1i},
+		n:     1,
+		wantu: 0 + 2i, wantc: 2,
+		wantuRev: 0 + 2i, wantcRev: 2,
+	},
+	{
+		x:     []complex64{1 + 2i},
+		y:     []complex64{1 + 1i},
+		n:     1,
+		wantu: -1 + 3i, wantc: 3 - 1i,
+		wantuRev: -1 + 3i, wantcRev: 3 - 1i,
+	},
+	{
+		x:     []complex64{1 + 2i, 3 + 4i, 5 + 6i, 7 + 8i, 9 + 10i, 11 + 12i, 13 + 14i, 15 + 16i, 17 + 18i, 19 + 20i},
+		y:     []complex64{1 + 2i, 3 + 4i, 5 + 6i, 7 + 8i, 9 + 10i, 11 + 12i, 13 + 14i, 15 + 16i, 17 + 18i, 19 + 20i},
+		n:     10,
+		wantu: -210 + 2860i, wantc: 2870 + 0i,
+		wantuRev: -210 + 1540i, wantcRev: 1550 + 0i,
+	},
+	{
+		x:     []complex64{1 + 1i, 1 + 1i, 1 + 2i, 1 + 1i, 1 + 1i, 1 + 1i, 1 + 3i, 1 + 1i, 1 + 1i, 1 + 4i},
+		y:     []complex64{1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i},
+		n:     10,
+		wantu: -22 + 36i, wantc: 42 + 4i,
+		wantuRev: -22 + 36i, wantcRev: 42 + 4i,
+	},
+	{
+		x:     []complex64{1 + 1i, 1 + 1i, 2 + 1i, 1 + 1i, 1 + 1i, 1 + 1i, 1 + 1i, 1 + 1i, 1 + 1i, 2 + 1i},
+		y:     []complex64{1 + 2i, 1 + 2i, 1 + 3i, 1 + 2i, 1 + 3i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i},
+		n:     10,
+		wantu: -10 + 37i, wantc: 34 + 17i,
+		wantuRev: -10 + 36i, wantcRev: 34 + 16i,
+	},
+	{
+		x:     []complex64{1 + 1i, 1 + 1i, 1 + 1i, 1 + 1i, complex(inf, 1), 1 + 1i, 1 + 1i, 1 + 1i, 1 + 1i, 1 + 1i},
+		y:     []complex64{1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i, 1 + 2i},
+		n:     10,
+		wantu: complex(inf, inf), wantc: complex(inf, inf),
+		wantuRev: complex(inf, inf), wantcRev: complex(inf, inf),
+	},
+}
+
+func TestDotcUnitary(t *testing.T) {
+	const gd = 1 + 5i
+	for i, test := range dotTests {
+		for _, align := range align2 {
+			prefix := fmt.Sprintf("Test %v (x:%v y:%v)", i, align.x, align.y)
+			xgLn, ygLn := 4+align.x, 4+align.y
+			xg, yg := guardVector(test.x, gd, xgLn), guardVector(test.y, gd, ygLn)
+			x, y := xg[xgLn:len(xg)-xgLn], yg[ygLn:len(yg)-ygLn]
+			res := DotcUnitary(x, y)
+			if !same(res, test.wantc) {
+				t.Errorf(msgVal, prefix, i, res, test.wantc)
+			}
+			if !isValidGuard(xg, gd, xgLn) {
+				t.Errorf(msgGuard, prefix, "x", xg[:xgLn], xg[len(xg)-xgLn:])
+			}
+			if !isValidGuard(yg, gd, ygLn) {
+				t.Errorf(msgGuard, prefix, "y", yg[:ygLn], yg[len(yg)-ygLn:])
+			}
+		}
+	}
+}
+
+func TestDotcInc(t *testing.T) {
+	const gd, gdLn = 2 + 5i, 4
+	for i, test := range dotTests {
+		for _, inc := range newIncSet(1, 2, 5, 10, -1, -2, -5, -10) {
+			xg, yg := guardIncVector(test.x, gd, inc.x, gdLn), guardIncVector(test.y, gd, inc.y, gdLn)
+			x, y := xg[gdLn:len(xg)-gdLn], yg[gdLn:len(yg)-gdLn]
+			want := test.wantc
+			var ix, iy int
+			if inc.x < 0 {
+				ix, want = -inc.x*(test.n-1), test.wantcRev
+			}
+			if inc.y < 0 {
+				iy, want = -inc.y*(test.n-1), test.wantcRev
+			}
+			prefix := fmt.Sprintf("Test %v (x:%v y:%v) (ix:%v iy:%v)", i, inc.x, inc.y, ix, iy)
+			res := DotcInc(x, y, uintptr(test.n), uintptr(inc.x), uintptr(inc.y), uintptr(ix), uintptr(iy))
+			if inc.x*inc.y > 0 {
+				want = test.wantc
+			}
+			if !same(res, want) {
+				t.Errorf(msgVal, prefix, i, res, want)
+				t.Error(x, y)
+			}
+			checkValidIncGuard(t, xg, gd, inc.x, gdLn)
+			checkValidIncGuard(t, yg, gd, inc.y, gdLn)
+		}
+	}
+}
+
+func TestDotuUnitary(t *testing.T) {
+	const gd = 1 + 5i
+	for i, test := range dotTests {
+		for _, align := range align2 {
+			prefix := fmt.Sprintf("Test %v (x:%v y:%v)", i, align.x, align.y)
+			xgLn, ygLn := 4+align.x, 4+align.y
+			xg, yg := guardVector(test.x, gd, xgLn), guardVector(test.y, gd, ygLn)
+			x, y := xg[xgLn:len(xg)-xgLn], yg[ygLn:len(yg)-ygLn]
+			res := DotuUnitary(x, y)
+			if !same(res, test.wantu) {
+				t.Errorf(msgVal, prefix, i, res, test.wantu)
+			}
+			if !isValidGuard(xg, gd, xgLn) {
+				t.Errorf(msgGuard, prefix, "x", xg[:xgLn], xg[len(xg)-xgLn:])
+			}
+			if !isValidGuard(yg, gd, ygLn) {
+				t.Errorf(msgGuard, prefix, "y", yg[:ygLn], yg[len(yg)-ygLn:])
+			}
+		}
+	}
+}
+
+func TestDotuInc(t *testing.T) {
+	const gd, gdLn = 1 + 5i, 4
+	for i, test := range dotTests {
+		for _, inc := range newIncSet(1, 2, 5, 10, -1, -2, -5, -10) {
+			prefix := fmt.Sprintf("Test %v (x:%v y:%v)", i, inc.x, inc.y)
+			xg, yg := guardIncVector(test.x, gd, inc.x, gdLn), guardIncVector(test.y, gd, inc.y, gdLn)
+			x, y := xg[gdLn:len(xg)-gdLn], yg[gdLn:len(yg)-gdLn]
+			want := test.wantc
+			var ix, iy int
+			if inc.x < 0 {
+				ix, want = -inc.x*(test.n-1), test.wantuRev
+			}
+			if inc.y < 0 {
+				iy, want = -inc.y*(test.n-1), test.wantuRev
+			}
+			res := DotuInc(x, y, uintptr(test.n), uintptr(inc.x), uintptr(inc.y), uintptr(ix), uintptr(iy))
+			if inc.x*inc.y > 0 {
+				want = test.wantu
+			}
+			if !same(res, want) {
+				t.Errorf(msgVal, prefix, i, res, want)
+			}
+			checkValidIncGuard(t, xg, gd, inc.x, gdLn)
+			checkValidIncGuard(t, yg, gd, inc.y, gdLn)
+		}
+	}
+}
diff --git a/internal/asm/c64/dotc.go b/internal/asm/c64/dotc.go