First pass at accelerating gpu-based AA clipping

This delivery adds two small enhancements to accelerate gpu AA clipping. The ClipMaskManager now attempts to reuse the cached texture and delays creation of the temp texture until it is needed. These two changes make the AA path go from 66-400x slower (than the stencil buffer path) to being 3-17x slower. This delivery also adds a new bench test case for profiling.

[bsalomon, 2012-05-07 13:52:44]

http://codereview.appspot.com/6195051/diff/1/src/gpu/GrClipMaskManager.cpp
File src/gpu/GrClipMaskManager.cpp (right):

http://codereview.appspot.com/6195051/diff/1/src/gpu/GrClipMaskManager.cpp#ne...
src/gpu/GrClipMaskManager.cpp:372: if (fAACache.getLastMaskWidth() >=
desc.fWidth &&
Do we want to have some limit on how oversized the reused texture can be (e.g.
the texture is 2K x 2K but the clip is 100x100).  Would it make sense to use the
scratch texture mechanism provided by GrContext?

[robertphillips, 2012-05-07 17:29:20]

http://codereview.appspot.com/6195051/diff/1/src/gpu/GrClipMaskManager.cpp
File src/gpu/GrClipMaskManager.cpp (right):

http://codereview.appspot.com/6195051/diff/1/src/gpu/GrClipMaskManager.cpp#ne...
src/gpu/GrClipMaskManager.cpp:372: if (fAACache.getLastMaskWidth() >=
desc.fWidth &&
Given the current usage pattern I don't think there needs to be a limit - only
one of these is being generated and cached at a time. Once we move to > 1 cached
mask we will need a better system. I will look into the scratch texture
mechanism.

[robertphillips, 2012-05-07 17:31:47]

One other item to consider is that Mike brought up the possibility of pushing
Skia's save/restore information down to the clip mask cache level. Skia knows
the lifetime of each clip mask whereas the gpu alone would have to guess via
some LRU scheme. The additional information would come in handy when the scene
consists of a lot of nested clips that are constantly being pushed & popped.

[bsalomon, 2012-05-07 18:02:47]

On 2012/05/07 17:31:47, robertphillips wrote:
> One other item to consider is that Mike brought up the possibility of pushing
> Skia's save/restore information down to the clip mask cache level. Skia knows
> the lifetime of each clip mask whereas the gpu alone would have to guess via
> some LRU scheme. The additional information would come in handy when the scene
> consists of a lot of nested clips that are constantly being pushed & popped.

LGTM

[robertphillips, 2012-05-07 18:08:01]

committed as r3853