Merge pull request #2 from NDari/master

Add benchmarking script

.gitignore

@@ -22,3 +22,6 @@ _testmain.go
 *.exe
 *.test
 *.prof
+
+vendor
+vendor/*

Gopkg.toml

@@ -0,0 +1,7 @@
+[[constraint]]
+  name = "github.com/chewxy/math32"
+  version = "1.0.0"
+
+[[constraint]]
+  name = "github.com/stretchr/testify"
+  version = "1.1.4"

arith_bench_test.go

@@ -0,0 +1,151 @@
+// +build !sse,!avx
+
+package vecf32
+
+import (
+	"testing"
+
+	"github.com/chewxy/math32"
+)
+
+/* BENCHMARKS */
+
+func _vanillaVecAdd(a, b []float32) {
+	for i := range a {
+		a[i] += b[i]
+	}
+}
+
+func BenchmarkVecAdd(b *testing.B) {
+	x := Range(0, niceprime)
+	y := Range(niceprime, 2*niceprime)
+
+	for n := 0; n < b.N; n++ {
+		Add(x, y)
+	}
+}
+
+func BenchmarkVanillaVecAdd(b *testing.B) {
+	x := Range(0, niceprime)
+	y := Range(niceprime, 2*niceprime)
+
+	for n := 0; n < b.N; n++ {
+		_vanillaVecAdd(x, y)
+	}
+}
+
+func _vanillaVecSub(a, b []float32) {
+	for i := range a {
+		a[i] -= b[i]
+	}
+}
+
+func BenchmarkVecSub(b *testing.B) {
+	x := Range(0, niceprime)
+	y := Range(niceprime, 2*niceprime)
+
+	for n := 0; n < b.N; n++ {
+		Sub(x, y)
+	}
+}
+
+func BenchmarkVanillaVecSub(b *testing.B) {
+	x := Range(0, niceprime)
+	y := Range(niceprime, 2*niceprime)
+
+	for n := 0; n < b.N; n++ {
+		_vanillaVecSub(x, y)
+	}
+}
+
+func _vanillaVecMul(a, b []float32) {
+	for i := range a {
+		a[i] *= b[i]
+	}
+}
+
+func BenchmarkVecMul(b *testing.B) {
+	x := Range(0, niceprime)
+	y := Range(niceprime, 2*niceprime)
+
+	for n := 0; n < b.N; n++ {
+		Mul(x, y)
+	}
+}
+
+func BenchmarkVanillaVecMul(b *testing.B) {
+	x := Range(0, niceprime)
+	y := Range(niceprime, 2*niceprime)
+
+	for n := 0; n < b.N; n++ {
+		_vanillaVecMul(x, y)
+	}
+}
+
+func _vanillaVecDiv(a, b []float32) {
+	for i := range a {
+		a[i] /= b[i]
+	}
+}
+
+func BenchmarkVecDiv(b *testing.B) {
+	x := Range(0, niceprime)
+	y := Range(niceprime, 2*niceprime)
+
+	for n := 0; n < b.N; n++ {
+		Div(x, y)
+	}
+}
+
+func BenchmarkVanillaVecDiv(b *testing.B) {
+	x := Range(0, niceprime)
+	y := Range(niceprime, 2*niceprime)
+
+	for n := 0; n < b.N; n++ {
+		_vanillaVecDiv(x, y)
+	}
+}
+
+func _vanillaVecSqrt(a []float32) {
+	for i, v := range a {
+		a[i] = math32.Sqrt(v)
+	}
+}
+
+func BenchmarkVecSqrt(b *testing.B) {
+	x := Range(0, niceprime)
+
+	for n := 0; n < b.N; n++ {
+		Sqrt(x)
+	}
+}
+
+func BenchmarkVanillaVecSqrt(b *testing.B) {
+	x := Range(0, niceprime)
+
+	for n := 0; n < b.N; n++ {
+		_vanillaVecSqrt(x)
+	}
+}
+
+func _vanillaVecInverseSqrt(a []float32) {
+	for i, v := range a {
+		a[i] = 1.0 / math32.Sqrt(v)
+	}
+}
+
+func BenchmarkVecInvSqrt(b *testing.B) {
+	x := Range(0, niceprime)
+
+	for n := 0; n < b.N; n++ {
+		InvSqrt(x)
+	}
+}
+
+func BenchmarkVanillaVecInvSqrt(b *testing.B) {
+	x := Range(0, niceprime)
+
+	for n := 0; n < b.N; n++ {
+		_vanillaVecInverseSqrt(x)
+	}
+}

arith_test.go

@@ -298,7 +298,7 @@ func TestMinOf(t *testing.T) {
 	correct := float32(0)
 	got := MinOf(a)
 	if got != correct {
-		t.Error("Expected %f. Got %v instead", correct, got)
+		t.Errorf("Expected %f. Got %v instead", correct, got)
 	}
 
 	a = []float32{}
@@ -313,7 +313,7 @@ func TestArgmax(t *testing.T) {
 	correct := 3
 	got := Argmax(a)
 	if got != correct {
-		t.Error("Expected argmax to be %v. Got %v instead", correct, got)
+		t.Errorf("Expected argmax to be %v. Got %v instead", correct, got)
 	}
 
 	a = []float32{math32.Inf(-1), 2, 3, 4}

bench.sh

@@ -0,0 +1,17 @@
+set -ex
+
+benchtime=${1:-1s}
+
+go test -bench . -benchtime $benchtime
+go test -tags='sse' -bench . -benchtime $benchtime
+go test -tags='avx' -bench . -benchtime $benchtime
+
+# travis compiles commands in script and then executes in bash.  By adding
+# set -e we are changing the travis build script's behavior, and the set
+# -e lives on past the commands we are providing it.  Some of the travis
+# commands are supposed to exit with non zero status, but then continue
+# executing.  set -x makes the travis log files extremely verbose and
+# difficult to understand.
+#
+# see travis-ci/travis-ci#5120
+set +ex