-

diff --git a/example/bwmf/kldiv_loss.go b/example/bwmf/kldiv_loss.go
index 764a46d..3dbd771 100644
--- a/example/bwmf/kldiv_loss.go
+++ b/example/bwmf/kldiv_loss.go
@@ -3,6 +3,8 @@ package bwmf
 import (
 	"math"
 	"sort"
+	"time"
+	"fmt"
 
 	pb "github.com/taskgraph/taskgraph/example/bwmf/proto"
 	"github.com/taskgraph/taskgraph/op"
@@ -57,6 +59,28 @@ func (l *KLDivLoss) Evaluate(param op.Parameter, gradient op.Parameter) float32
 	H := param
 	op.Fill(gradient, 0.0)
 	value := float32(0.0)
+	start := time.Now()
+
+    fmt.Printf("m: %d, n: %d, k: %d\n", l.m, l.n, l.k)
+	for i := 0; i < l.m; i++ {
+		for j := 0; j < l.n; j++ {
+			wRow := l.GetWRow(i)
+			wh := float32(0.0)
+			for k, wk := range *wRow {
+				wh += wk * H.Get(j*l.k+int(k))
+			}
+		}
+	}
+
+	elapsed := time.Since(start)
+	fmt.Printf("Took %s\n", elapsed)
+
+	tot := 0
+	for i := 0; i < l.m; i++ {
+		tot += len(*l.GetWRow(i))
+	}
+	fmt.Printf("Tot: %d\n", tot)
+
 	for i := 0; i < l.m; i++ {
 		for j := 0; j < l.n; j++ {
 			v, ve := l.V.GetRow()[j].At[int32(i)]

Output

m: 1000, n: 2000, k: 10
Took 754.997378ms
Tot: 9972

m: 1000, n: 2000, k: 10
Took 884.45471ms
Tot: 9972

In `julia'

julia> tic(); rand(1000,10)*rand(10,2000); toc()
elapsed time: 0.012591754 seconds

We can make it parallel by using a fixed num of goroutines, which each takes charge of a proportion of the matrix.

CC @xiaoyunwu