Change Atlas recycling to track current flush count and recycle if Atlas not used in current flush.

Change Atlas recycling to track current flush count and recycle if Atlas not used in current flush.

BUG=

Committed: http://code.google.com/p/skia/source/detail?r=10777

[jvanverth1, 2013-08-13 20:58:09]

No perf hit or gm change on Mac with this. Big change is are the upload and
textstrike purge counts (repeating 100 times):

New:
running bench [1257 3080] tabl_worldjournal.skp   tile_256x256_gpu: cmsecs = 
61.39
Num uploads: 53308
Num purges: 0
running bench [1027 23155] zh.wikipedia.org.skp   tile_256x256_gpu: cmsecs =
239.80
Num uploads: 263809
Num purges: 0

Old:
running bench [1257 3080] tabl_worldjournal.skp   tile_256x256_gpu: cmsecs = 
63.23
Num uploads: 84027
Num purges: 403
running bench [1027 23155] zh.wikipedia.org.skp   tile_256x256_gpu: cmsecs =
234.76
Num uploads: 276921
Num purges: 1009

[bsalomon, 2013-08-13 21:18:28]

https://codereview.chromium.org/23120004/diff/4001/include/gpu/GrContext.h
File include/gpu/GrContext.h (right):

https://codereview.chromium.org/23120004/diff/4001/include/gpu/GrContext.h#ne...
include/gpu/GrContext.h:827: uint64_t getFlushCount() const { return
fFlushCount; }
I think this guy needs some documentation even though it is for internal use.
Let's talk a little tomorrow about this. Now that I see this I have some
thoughts about flush counts and supporting them as the code evolves (perhaps
having partial flushes in the future). We might want something close to the GL
fence model.

https://codereview.chromium.org/23120004/diff/4001/src/gpu/GrTextStrike.cpp
File src/gpu/GrTextStrike.cpp (right):

https://codereview.chromium.org/23120004/diff/4001/src/gpu/GrTextStrike.cpp#n...
src/gpu/GrTextStrike.cpp:184: static uint64_t gCurrentFlushCount = 0;
wha? There can be multiple GrContexts

[jvanverth1, 2013-08-14 13:24:32]

https://codereview.chromium.org/23120004/diff/4001/src/gpu/GrTextStrike.cpp
File src/gpu/GrTextStrike.cpp (right):

https://codereview.chromium.org/23120004/diff/4001/src/gpu/GrTextStrike.cpp#n...
src/gpu/GrTextStrike.cpp:184: static uint64_t gCurrentFlushCount = 0;
On 2013/08/13 21:18:28, bsalomon wrote:
> wha? There can be multiple GrContexts

A bit of a hack to get around the fact that visitAll() only takes one argument.
But I agree it's not thread-safe (and a hack). Changed to something cleaner,
below.

[jvanverth1, 2013-08-14 18:52:48]

Added a FlushToken to track the flush state for a given Atlas. This can be used
to check whether an Atlas is still current or the state that used it has already
been flushed. For a generic DrawTarget it's assumed that it's never current
(i.e. the draw calls go right through); however, the InOrderDrawBuffer will
track flush ids.

[bsalomon, 2013-08-14 19:04:27]

https://codereview.chromium.org/23120004/diff/29001/src/gpu/GrDrawTarget.h
File src/gpu/GrDrawTarget.h (right):

https://codereview.chromium.org/23120004/diff/29001/src/gpu/GrDrawTarget.h#ne...
src/gpu/GrDrawTarget.h:631: bool isCurrent() { return NULL != fDrawTarget &&
fDrawTarget->isCurrentFlush(fFlushID); }
I think we should reverse the sense of this: isFlushed(), isIssued(), ...


It seems more well defined for a wider class of subclasses that aren't a strict
queue like the in order guy, or even that are a queue but can perform partial
flushing (e.g. flush up through the last draw to such and such render target).

https://codereview.chromium.org/23120004/diff/29001/src/gpu/GrInOrderDrawBuff...
File src/gpu/GrInOrderDrawBuffer.h (right):

https://codereview.chromium.org/23120004/diff/29001/src/gpu/GrInOrderDrawBuff...
src/gpu/GrInOrderDrawBuffer.h:71: virtual bool       isCurrentFlush(uint32_t
flushID) { return flushID == fFlushID; }
shouldn't this be private?

[jvanverth1, 2013-08-15 14:49:57]

https://codereview.chromium.org/23120004/diff/29001/src/gpu/GrDrawTarget.h
File src/gpu/GrDrawTarget.h (right):

https://codereview.chromium.org/23120004/diff/29001/src/gpu/GrDrawTarget.h#ne...
src/gpu/GrDrawTarget.h:631: bool isCurrent() { return NULL != fDrawTarget &&
fDrawTarget->isCurrentFlush(fFlushID); }
On 2013/08/14 19:04:27, bsalomon wrote:
> I think we should reverse the sense of this: isFlushed(), isIssued(), ...
> 
> 
> It seems more well defined for a wider class of subclasses that aren't a
strict
> queue like the in order guy, or even that are a queue but can perform partial
> flushing (e.g. flush up through the last draw to such and such render target).


Done.

https://codereview.chromium.org/23120004/diff/29001/src/gpu/GrInOrderDrawBuff...
File src/gpu/GrInOrderDrawBuffer.h (right):

https://codereview.chromium.org/23120004/diff/29001/src/gpu/GrInOrderDrawBuff...
src/gpu/GrInOrderDrawBuffer.h:71: virtual bool       isCurrentFlush(uint32_t
flushID) { return flushID == fFlushID; }
On 2013/08/14 19:04:27, bsalomon wrote:
> shouldn't this be private?

Done.

[bsalomon, 2013-08-15 14:59:33]

I like the new name (DrawToken)!
lgtm

[commit-bot: I haz the power, 2013-08-16 13:17:46]

CQ is trying da patch. Follow status at
https://skia-tree-status.appspot.com/cq/jvanverth@google.com/23120004/42001

[commit-bot: I haz the power, 2013-08-16 15:53:53]

Change committed as 10777