From 8c8cfcc32ebf731dabf4e19e57a7d64cfc615b5b Mon Sep 17 00:00:00 2001
From: btracey <bdtinternet@gmail.com>
Date: Mon, 12 Jan 2015 10:46:47 -0800
Subject: [PATCH 1/2] Optimized Inner with the new assembly code

---
 mat64/bench_test.go | 10 ++++++++++
 mat64/inner.go      |  6 ++++--
 mat64/inner_test.go | 31 +++++++++++++++++++++++++++++++
 3 files changed, 45 insertions(+), 2 deletions(-)
 create mode 100644 mat64/bench_test.go

diff --git a/mat64/bench_test.go b/mat64/bench_test.go
new file mode 100644
index 0000000..4a91a0e
--- /dev/null
+++ b/mat64/bench_test.go
@@ -0,0 +1,10 @@
+package mat64
+
+import "github.com/gonum/blas/testblas"
+
+const (
+	Sm   = testblas.SmallMat
+	Med  = testblas.MediumMat
+	Lg   = testblas.LargeMat
+	Huge = testblas.HugeMat
+)
diff --git a/mat64/inner.go b/mat64/inner.go
index 154acad..63144b2 100644
--- a/mat64/inner.go
+++ b/mat64/inner.go
@@ -4,6 +4,8 @@
 
 package mat64
 
+import "github.com/gonum/internal/asm"
+
 // Inner computes the generalized inner product
 //   x^T A y
 // between vectors x and y with matrix A. This is only a true inner product if
@@ -28,8 +30,8 @@ func Inner(x []float64, A Matrix, y []float64) float64 {
 	case RawMatrixer:
 		bmat := b.RawMatrix()
 		for i, xi := range x {
-			for j, yj := range y {
-				sum += xi * bmat.Data[i*bmat.Stride+j] * yj
+			if xi != 0 {
+				sum += xi * asm.DdotUnitary(bmat.Data[i*bmat.Stride:i*bmat.Stride+n], y)
 			}
 		}
 	default:
diff --git a/mat64/inner_test.go b/mat64/inner_test.go
index 829f90a..af3b3d9 100644
--- a/mat64/inner_test.go
+++ b/mat64/inner_test.go
@@ -5,6 +5,9 @@
 package mat64
 
 import (
+	"testing"
+
+	"github.com/gonum/blas/blas64"
 	"gopkg.in/check.v1"
 )
 
@@ -73,3 +76,31 @@ func (s *S) TestInner(c *check.C) {
 		c.Check(want, check.Equals, got, check.Commentf("Test %v: want %v, got %v", i, want, got))
 	}
 }
+
+func benchmarkInner(b *testing.B, m, n int) {
+	x := make([]float64, m)
+	randomSlice(x)
+	y := make([]float64, n)
+	randomSlice(y)
+	data := make([]float64, m*n)
+	randomSlice(data)
+	for i := 0; i < b.N; i++ {
+		Inner(x, &Dense{blas64.General{Rows: m, Cols: n, Stride: n, Data: data}}, y)
+	}
+}
+
+func BenchmarkInnerSmSm(b *testing.B) {
+	benchmarkInner(b, Sm, Sm)
+}
+
+func BenchmarkInnerMedMed(b *testing.B) {
+	benchmarkInner(b, Med, Med)
+}
+
+func BenchmarkInnerLgLg(b *testing.B) {
+	benchmarkInner(b, Lg, Lg)
+}
+
+func BenchmarkInnerLgSm(b *testing.B) {
+	benchmarkInner(b, Lg, Sm)
+}

From 07a5cb547e252dc8e13ed8a6450d3a8a4ce72903 Mon Sep 17 00:00:00 2001
From: btracey <bdtinternet@gmail.com>
Date: Mon, 12 Jan 2015 10:52:07 -0800
Subject: [PATCH 2/2] Allocate the matrix in advance and reset the timer for a
 fairer benchmark

---
 mat64/inner_test.go | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/mat64/inner_test.go b/mat64/inner_test.go
index af3b3d9..a26d214 100644
--- a/mat64/inner_test.go
+++ b/mat64/inner_test.go
@@ -84,8 +84,10 @@ func benchmarkInner(b *testing.B, m, n int) {
 	randomSlice(y)
 	data := make([]float64, m*n)
 	randomSlice(data)
+	mat := &Dense{blas64.General{Rows: m, Cols: n, Stride: n, Data: data}}
+	b.ResetTimer()
 	for i := 0; i < b.N; i++ {
-		Inner(x, &Dense{blas64.General{Rows: m, Cols: n, Stride: n, Data: data}}, y)
+		Inner(x, mat, y)
 	}
 }