code review 6734045: image/jpeg: skip TestDCT on arm hosts

image/jpeg: skip TestDCT on arm hosts

For arm5 machines, without floating point hardware, TestDCT
takes 2 to 3 minutes. Skip the test for all arm hosts.

[dfc, 2012-10-18 05:09:47]

Hello golang-dev@googlegroups.com,

I'd like you to review this change to
https://code.google.com/p/go

[remyoudompheng, 2012-10-18 07:04:36]

Can we make the test faster instead?

Try introducing a global var cosines [16]float64 holding the value of
cos(kπ/16). and compute the cosines only once for slowIDCT/slowFDCT.

Rémy.

[remyoudompheng, 2012-10-18 07:09:33]

On amd64 It got 6x faster using this diff, I expect a huge improvement on arm.

diff -r 6b67a925abcc src/pkg/image/jpeg/dct_test.go
--- a/src/pkg/image/jpeg/dct_test.go	Tue Oct 16 07:22:33 2012 +0200
+++ b/src/pkg/image/jpeg/dct_test.go	Thu Oct 18 09:08:50 2012 +0200
@@ -112,6 +112,14 @@
 	return math.Sqrt2
 }

+var cosines [32]float64 // cosines[k] = cos(π/2 * k/8)
+
+func init() {
+	for k := range cosines {
+		cosines[k] = math.Cos(math.Pi * float64(k) / 16)
+	}
+}
+
 // slowFDCT performs the 8*8 2-dimensional forward discrete cosine transform:
 //
 //	dst[u,v] = (1/8) * Σ_x Σ_y alpha(u) * alpha(v) * src[x,y] *
@@ -129,8 +137,8 @@
 			for y := 0; y < 8; y++ {
 				for x := 0; x < 8; x++ {
 					sum += alpha(u) * alpha(v) * float64(b[8*y+x]) *
-						math.Cos(math.Pi*float64((2*x+1)*u)/16) *
-						math.Cos(math.Pi*float64((2*y+1)*v)/16)
+						cosines[((2*x+1)*u)%32] *
+						cosines[((2*y+1)*v)%32]
 				}
 			}
 			dst[8*v+u] = sum / 8
@@ -159,8 +167,8 @@
 			for v := 0; v < 8; v++ {
 				for u := 0; u < 8; u++ {
 					sum += alpha(u) * alpha(v) * float64(b[8*v+u]) *
-						math.Cos(math.Pi*float64((2*x+1)*u)/16) *
-						math.Cos(math.Pi*float64((2*y+1)*v)/16)
+						cosines[((2*x+1)*u)%32] *
+						cosines[((2*y+1)*v)%32]
 				}
 			}
 			dst[8*y+x] = sum / 8

[dfc, 2012-10-18 07:20:47]

Thanks Rémy,

before; ~ 132seconds on a 2Ghz arm5

after:

=== RUN TestDCT
--- PASS: TestDCT (12.05 seconds)
PASS
ok      image/jpeg      12.073s

I'm going to abandon this proposal, you should definitely propose your fix.

On Thu, Oct 18, 2012 at 6:09 PM, Rémy Oudompheng
<remyoudompheng@gmail.com> wrote:
> On amd64 It got 6x faster using this diff, I expect a huge improvement on
arm.
>
> diff -r 6b67a925abcc src/pkg/image/jpeg/dct_test.go
> --- a/src/pkg/image/jpeg/dct_test.go    Tue Oct 16 07:22:33 2012 +0200
> +++ b/src/pkg/image/jpeg/dct_test.go    Thu Oct 18 09:08:50 2012 +0200
> @@ -112,6 +112,14 @@
>         return math.Sqrt2
>  }
>
> +var cosines [32]float64 // cosines[k] = cos(π/2 * k/8)
> +
> +func init() {
> +       for k := range cosines {
> +               cosines[k] = math.Cos(math.Pi * float64(k) / 16)
> +       }
> +}
> +
>  // slowFDCT performs the 8*8 2-dimensional forward discrete cosine
transform:
>  //
>  //     dst[u,v] = (1/8) * Σ_x Σ_y alpha(u) * alpha(v) * src[x,y] *
> @@ -129,8 +137,8 @@
>                         for y := 0; y < 8; y++ {
>                                 for x := 0; x < 8; x++ {
>                                         sum += alpha(u) * alpha(v) *
float64(b[8*y+x]) *
> -                                              
math.Cos(math.Pi*float64((2*x+1)*u)/16) *
> -                                              
math.Cos(math.Pi*float64((2*y+1)*v)/16)
> +                                               cosines[((2*x+1)*u)%32] *
> +                                               cosines[((2*y+1)*v)%32]
>                                 }
>                         }
>                         dst[8*v+u] = sum / 8
> @@ -159,8 +167,8 @@
>                         for v := 0; v < 8; v++ {
>                                 for u := 0; u < 8; u++ {
>                                         sum += alpha(u) * alpha(v) *
float64(b[8*v+u]) *
> -                                              
math.Cos(math.Pi*float64((2*x+1)*u)/16) *
> -                                              
math.Cos(math.Pi*float64((2*y+1)*v)/16)
> +                                               cosines[((2*x+1)*u)%32] *
> +                                               cosines[((2*y+1)*v)%32]
>                                 }
>                         }
>                         dst[8*y+x] = sum / 8

[dfc, 2012-10-18 07:21:07]

*** Abandoned ***