Add GrMemoryPool and test

[bsalomon, 2012-06-15 14:39:54]

This class is intended to be used to handle op new and delete for GrCustomStage.
However, I'm holding back adding the new/delete overrides until after the M21
branch. This just adds the class and the unit test.

[TomH, 2012-06-18 19:25:32]

Can we get some idea of performance vs malloc? This looks like you have really
optimized for speed instead of memory efficiency, but I'd still like to be
confident that it's 10x or 100x faster than malloc()

http://codereview.appspot.com/6306090/diff/2001/include/gpu/GrTypes.h
File include/gpu/GrTypes.h (right):

http://codereview.appspot.com/6306090/diff/2001/include/gpu/GrTypes.h#newcode87
include/gpu/GrTypes.h:87: #define GR_CT_ROUND_UP(X, Y) (((X) + ((Y)-1)) / (Y))
GR_CT_DIV_ROUND_UP?

http://codereview.appspot.com/6306090/diff/2001/src/gpu/GrMemoryPool.cpp
File src/gpu/GrMemoryPool.cpp (right):

http://codereview.appspot.com/6306090/diff/2001/src/gpu/GrMemoryPool.cpp#newc...
src/gpu/GrMemoryPool.cpp:56: // we stash a point back the header before the
allocated space
nit: slight grammar error?

http://codereview.appspot.com/6306090/diff/2001/src/gpu/GrMemoryPool.h
File src/gpu/GrMemoryPool.h (right):

http://codereview.appspot.com/6306090/diff/2001/src/gpu/GrMemoryPool.h#newcode56
src/gpu/GrMemoryPool.h:56: BlockHeader* fNext;      // D-L list of blocks.
nit: please spell out D-L (doubly linked?)

[bsalomon, 2012-06-19 13:29:05]

I added two benchs. Both randomly alternate between deleting and allocating. One
deletes in random order and the other deletes in stack order. On my windows
machine the random order is 9.5x std new/delete and the stack order is 4.5x.

http://codereview.appspot.com/6306090/diff/2001/include/gpu/GrTypes.h
File include/gpu/GrTypes.h (right):

http://codereview.appspot.com/6306090/diff/2001/include/gpu/GrTypes.h#newcode87
include/gpu/GrTypes.h:87: #define GR_CT_ROUND_UP(X, Y) (((X) + ((Y)-1)) / (Y))
On 2012/06/18 19:25:32, TomH wrote:
> GR_CT_DIV_ROUND_UP?

Done.

http://codereview.appspot.com/6306090/diff/2001/src/gpu/GrMemoryPool.cpp
File src/gpu/GrMemoryPool.cpp (right):

http://codereview.appspot.com/6306090/diff/2001/src/gpu/GrMemoryPool.cpp#newc...
src/gpu/GrMemoryPool.cpp:56: // we stash a point back the header before the
allocated space
On 2012/06/18 19:25:32, TomH wrote:
> nit: slight grammar error?

Done.

http://codereview.appspot.com/6306090/diff/2001/src/gpu/GrMemoryPool.h
File src/gpu/GrMemoryPool.h (right):

http://codereview.appspot.com/6306090/diff/2001/src/gpu/GrMemoryPool.h#newcode56
src/gpu/GrMemoryPool.h:56: BlockHeader* fNext;      // D-L list of blocks.
On 2012/06/18 19:25:32, TomH wrote:
> nit: please spell out D-L (doubly linked?)

Done.

[TomH, 2012-06-19 13:36:32]

On 2012/06/19 13:29:05, bsalomon wrote:
> I added two benchs. Both randomly alternate between deleting and allocating.
One
> deletes in random order and the other deletes in stack order. On my windows
> machine the random order is 9.5x std new/delete and the stack order is 4.5x.

Not as good as we hoped, but hopefully it'll do. (Didn't you get a preliminary
estimate > 10% slowdown with system malloc?)

Nit: can we also get a sense of the % memory waste in those two benchmarks?

LGTM

[bsalomon, 2012-06-19 21:14:35]

On 2012/06/19 13:36:32, TomH wrote:
> On 2012/06/19 13:29:05, bsalomon wrote:
> > I added two benchs. Both randomly alternate between deleting and allocating.
> One
> > deletes in random order and the other deletes in stack order. On my windows
> > machine the random order is 9.5x std new/delete and the stack order is 4.5x.
> 
> Not as good as we hoped, but hopefully it'll do. (Didn't you get a preliminary
> estimate > 10% slowdown with system malloc?)
> 
> Nit: can we also get a sense of the % memory waste in those two benchmarks?
> 
> LGTM

Checked in with a third bench (deletes are in queue order) in r4282. I'll
consider adding memory consumption stats in the next go-round on this.