582489 - Much slower on particle demo due to canvas performance issues with fillRect

Status: RESOLVED INACTIVE

(Core :: Graphics: Canvas2D, defect)

Description

[Boris Zbarsky [:bzbarsky]]

BUILD: Current trunk mac build

STEPS TO REPRODUCE:
1)  Load the attachment.
2)  Click the "Start" button.
3)  Wait a bit for the FPS to stabilize

ACTUAL RESULTS: 8fps or so on my machine

EXPECTED RESULTS: 20+ fps like Safari and Chrome (note that Opera doesn't support fillText so doesn't run the testcase).

ADDITIONAL INFORMATION:  Shark says that 85% of our time is spent under nsCanvasRenderingContext2D::FillRect.  This breaks down as (all percentages are of total time, and are times under, not in):

2% ~gfxPath
2% free under nsCanvasRenderingContext2D::DrawRect
2% gfxContext::CopyPath
5% gfxContext::Rectangle 
1% gfxContext::UserToDevice (under nsCanvasRenderingContext2D::DrawPath)
3% gfxContext::GetUserPathExtent
68% _cairo_quartz_surface_fill

The fill stuff is:

CGContextFillPath: 25%
_cairo_quartz_setup_state: 24% (calling CGContextSetRGBStroke/FillColor and
                                CGContextSaveGState)
_cairo_quartz_teardown_state: 7% (all CGContextRestoreGState)
_cairo_quartz_cairo_path_to_quartz_context: 7% (all under
                                                _cairo_path_fixed_interpret).

I'm having a real hard time believing that rectangle fills need all that gunk (complete with the _cairo_path_fixed_interpret!)....  Is this really expected?

Comment 1

[Boris Zbarsky [:bzbarsky]]

Attachment: The testcase

Comment 2

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

sounds like we need to an optimization to recognize when the path we're drawing is just a rectangle...

Comment 3

[Boris Zbarsky [:bzbarsky]]

Note that I've seen this pop up before; fillRect is often somewhat of an issue in canvas profiles.  This one was just more egregious than most.  ;)  Point being, it's not that this particular testcase is somehow doing something weird... at least as far as I can tell.

I also don't know whether there are similar issues on non-Mac platforms.

Comment 4

[Joe Drew (not getting mail)]

I don't like not blocking on this, but I also can't justify it if it's not a regression. :( Would gladly take a patch!

Comment 5

[Boris Zbarsky [:bzbarsky]]

OK.  Is there a way to track this as a must-fix for post-2.0?  :(

Comment 6

[Bas Schouten (:bas.schouten)]

This appears fine on windows, do we have a plan to attack this Jeff? Or would you say we're going at this through Azure?

Comment 7

[Mayank Bansal]

This performs worse with gpu-canvas on Windows : https://share.firefox.dev/3WQYX7U
With d2d-canvas, the testcase is smoother than Chrome.