Track the cost of impl-side painting.

Track the rendering & painting cost of impl-side painting.
We record how long the rendering thread takes to render each PicturePile,
accumulate in TileManager, and report through LayerTreeHostImpl;
telemetry queries this information and dumps the output.
Note that non-impl-side-painting doesn't yet report this correctly?

BUG=163324
TEST=manual
TBR=nduca,reveman

Committed: https://src.chromium.org/viewvc/chrome?view=rev&revision=170274

[Tom Hudson, 2012-11-29 16:47:57]

Nat, is this what you were asking for this morning?

Numbers on the Linux desktop suggest we're rendering half as many pixels, and
spending half as much commit time, but *without* --enable-impl-side-painting
paint & render time aren't recording, so I can't get that.

[nduca, 2012-11-29 18:16:30]

epic. Small tweaks

https://codereview.chromium.org/11434033/diff/1009/cc/picture_pile.cc
File cc/picture_pile.cc (right):

https://codereview.chromium.org/11434033/diff/1009/cc/picture_pile.cc#newcode55
cc/picture_pile.cc:55: stats.totalPaintTimeInSeconds += (base::TimeTicks::Now()
-
Lets move this into accounting Picture::Record

https://codereview.chromium.org/11434033/diff/1009/cc/picture_pile.cc#newcode56
cc/picture_pile.cc:56: paintBeginTime).InSecondsF();
i think we also need to track the number of pixels recorded...

https://codereview.chromium.org/11434033/diff/1009/cc/picture_pile_impl.cc
File cc/picture_pile_impl.cc (left):

https://codereview.chromium.org/11434033/diff/1009/cc/picture_pile_impl.cc#ol...
cc/picture_pile_impl.cc:31: void PicturePileImpl::Raster(SkCanvas* canvas,
gfx::Rect rect) {
Lets pass in a ref to Stats object and accumulate on it. Then tile_manager can
have a stats object. Then TileManager::renderingStats() can copy the relevant
fields to the passed-in reingeringStats pointer.

https://codereview.chromium.org/11434033/diff/1009/cc/tile_manager.cc
File cc/tile_manager.cc (right):

https://codereview.chromium.org/11434033/diff/1009/cc/tile_manager.cc#newcode222
cc/tile_manager.cc:222: rasterize_time_in_seconds_ = 0;
we also need number of pixels rasterized

https://codereview.chromium.org/11434033/diff/1009/cc/tile_manager.h
File cc/tile_manager.h (right):

https://codereview.chromium.org/11434033/diff/1009/cc/tile_manager.h#newcode110
cc/tile_manager.h:110: 
Here, you could consider just a RenderingStats object, by value.

Then we can copy the rendering time value into the rendering stats pointer
during ::renderingStats  call. Or you could make a TileManagerStats object that
we put both rasterized time and pixel count inside.

[Tom Hudson, 2012-11-29 19:16:11]

https://codereview.chromium.org/11434033/diff/1009/cc/picture_pile.cc
File cc/picture_pile.cc (right):

https://codereview.chromium.org/11434033/diff/1009/cc/picture_pile.cc#newcode55
cc/picture_pile.cc:55: stats.totalPaintTimeInSeconds += (base::TimeTicks::Now()
-
On 2012/11/29 18:16:30, nduca wrote:
> Lets move this into accounting Picture::Record

D'oh! Picture::Record already does it, so this was double-counting.

https://codereview.chromium.org/11434033/diff/1009/cc/picture_pile.cc#newcode56
cc/picture_pile.cc:56: paintBeginTime).InSecondsF();
On 2012/11/29 18:16:30, nduca wrote:
> i think we also need to track the number of pixels recorded...

Done.

https://codereview.chromium.org/11434033/diff/1009/cc/picture_pile_impl.cc
File cc/picture_pile_impl.cc (left):

https://codereview.chromium.org/11434033/diff/1009/cc/picture_pile_impl.cc#ol...
cc/picture_pile_impl.cc:31: void PicturePileImpl::Raster(SkCanvas* canvas,
gfx::Rect rect) {
On 2012/11/29 18:16:30, nduca wrote:
> Lets pass in a ref to Stats object and accumulate on it. Then tile_manager can
> have a stats object. Then TileManager::renderingStats() can copy the relevant
> fields to the passed-in reingeringStats pointer.

I hit "Done" when I was working in PicturePileImpl, but then I got to
TileManager. RasterizeTile is not a member function, and it's running in the
other thread. How do we do this safely?

[reveman, 2012-11-29 19:27:02]

https://codereview.chromium.org/11434033/diff/1009/cc/picture_pile_impl.cc
File cc/picture_pile_impl.cc (left):

https://codereview.chromium.org/11434033/diff/1009/cc/picture_pile_impl.cc#ol...
cc/picture_pile_impl.cc:31: void PicturePileImpl::Raster(SkCanvas* canvas,
gfx::Rect rect) {
On 2012/11/29 19:16:11, Tom Hudson wrote:
> On 2012/11/29 18:16:30, nduca wrote:
> > Lets pass in a ref to Stats object and accumulate on it. Then tile_manager
can
> > have a stats object. Then TileManager::renderingStats() can copy the
relevant
> > fields to the passed-in reingeringStats pointer.
> 
> I hit "Done" when I was working in PicturePileImpl, but then I got to
> TileManager. RasterizeTile is not a member function, and it's running in the
> other thread. How do we do this safely?

you can allocate storage for the stat in TileManager::DispatchOneRasterTask,
pass a pointer to RasterizeTile, then collect the result and free the storage in
TileManager::OnRasterTaskCompleted.

[Tom Hudson, 2012-11-29 20:30:31]

I think I've addressed everything, and cleaned up the lint.

I'm not sure I trust the numbers yet: 1/3rd as much painting, 10x as much
rendering?

It's also clear from watching the tests run that --enable-impl-side-painting
works just fine on some sites (modulo swapping the R & B channels) and not at
all on others, but I haven't dumped on the patch Daniel sent me yet.

I'm out for the night, feel free to CQ if you're happy or to steal from me & do
it right yourself if you aren't.

https://codereview.chromium.org/11434033/diff/1009/cc/tile_manager.cc
File cc/tile_manager.cc (right):

https://codereview.chromium.org/11434033/diff/1009/cc/tile_manager.cc#newcode222
cc/tile_manager.cc:222: rasterize_time_in_seconds_ = 0;
On 2012/11/29 18:16:30, nduca wrote:
> we also need number of pixels rasterized

Done.

https://codereview.chromium.org/11434033/diff/1009/cc/tile_manager.h
File cc/tile_manager.h (right):

https://codereview.chromium.org/11434033/diff/1009/cc/tile_manager.h#newcode110
cc/tile_manager.h:110: 
On 2012/11/29 18:16:30, nduca wrote:
> Here, you could consider just a RenderingStats object, by value.
> 
> Then we can copy the rendering time value into the rendering stats pointer
> during ::renderingStats  call. Or you could make a TileManagerStats object
that
> we put both rasterized time and pixel count inside.

Done.

[nduca, 2012-11-29 21:51:43]

https://codereview.chromium.org/11434033/diff/6004/cc/tile_manager.cc
File cc/tile_manager.cc (right):

https://codereview.chromium.org/11434033/diff/6004/cc/tile_manager.cc#newcode223
cc/tile_manager.cc:223: stats->totalPixelsRasterized +=
rendering_stats_.totalPixelsRasterized;
dont add, just set.

https://codereview.chromium.org/11434033/diff/6004/cc/tile_manager.cc#newcode224
cc/tile_manager.cc:224: rendering_stats_.totalRasterizeTimeInSeconds = 0;
dont zero these...

[nduca, 2012-11-29 21:55:42]

Ima land this.