GPU path for SkMatrixConvolutionImageFilter

This patch implements GPU processing for SkMatrixConvolutionImageFilter.  All parameters from the raster path are supported.  Note that the premultiplying is done less efficiently than in the raster path:  it's done on each texture access, rather than as a pre-processing pass.   This was so I could do the filter as a single custom stage; will try the optimization separately.

Implementing the GPU path revealed a bug in the raster path:  when not convolving alpha, the convolved RGB should be clamped to 255, not to the original alpha, since it will be remultiplied against original alpha anyway.  In order to test this more thoroughly, I've added a gradient alpha to the source bitmap in the matrixconvolution GM, and added test cases for all the tiling modes in the !convolveAlpha case.

[bsalomon, 2012-10-03 17:48:18]

Can we add it to the unit test? (GR_DECLARE_CUSTOM_STAGE_TEST /
GR_DEFINE_CUSTOM_STAGE_TEST)

https://codereview.appspot.com/6585069/diff/1/src/effects/SkMatrixConvolution...
File src/effects/SkMatrixConvolutionImageFilter.cpp (right):

https://codereview.appspot.com/6585069/diff/1/src/effects/SkMatrixConvolution...
src/effects/SkMatrixConvolutionImageFilter.cpp:285: typedef
SkMatrixConvolutionImageFilter::TileMode TileMode;
Maybe these should either be made public or explicitly private?

[Stephen White, 2012-10-04 13:55:22]

Can we add it to the unit test? (GR_DECLARE_CUSTOM_STAGE_TEST
/GR_DEFINE_CUSTOM_STAGE_TEST)

Done.  This revealed that I was using too many bits in my StageKey, so thanks
for that!  Added a little hashing function to mash the kernel width and height
into 7 bits.

https://codereview.appspot.com/6585069/diff/1/src/effects/SkMatrixConvolution...
File src/effects/SkMatrixConvolutionImageFilter.cpp (right):

https://codereview.appspot.com/6585069/diff/1/src/effects/SkMatrixConvolution...
src/effects/SkMatrixConvolutionImageFilter.cpp:285: typedef
SkMatrixConvolutionImageFilter::TileMode TileMode;
On 2012/10/03 17:48:18, bsalomon wrote:
> Maybe these should either be made public or explicitly private?

Done.

[bsalomon, 2012-10-04 13:58:43]

LGTM

https://codereview.appspot.com/6585069/diff/9/src/effects/SkMatrixConvolution...
File src/effects/SkMatrixConvolutionImageFilter.cpp (right):

https://codereview.appspot.com/6585069/diff/9/src/effects/SkMatrixConvolution...
src/effects/SkMatrixConvolutionImageFilter.cpp:414: }
cool!

[TomH, 2012-10-04 14:04:55]

On 2012/10/04 13:55:22, Stephen White wrote:
> 
> Done.  This revealed that I was using too many bits in my StageKey, so thanks
> for that!  Added a little hashing function to mash the kernel width and height
> into 7 bits.

So, when will we know that we should make the StageKey larger?

[Stephen White, 2012-10-04 14:08:34]

On 2012/10/04 14:04:55, TomH wrote:
> On 2012/10/04 13:55:22, Stephen White wrote:
> > 
> > Done.  This revealed that I was using too many bits in my StageKey, so
thanks
> > for that!  Added a little hashing function to mash the kernel width and
height
> > into 7 bits.
> 
> So, when will we know that we should make the StageKey larger?

When the tests break?  :)

[Stephen White, 2012-10-05 14:37:44]

Landed as https://code.google.com/p/skia/source/detail?r=5809.  Closing.