311571 - Optimize canvas

Status: RESOLVED INCOMPLETE

(Core :: Graphics: Canvas2D, defect, P3)

Description

[Boris Zbarsky [:bzbarsky]]

This is a tracking bug for various canvas performance stuff.  At the moment,
I've been profiling http://www.hixie.ch/tests/adhoc/perf/video/002.html and will
attach those profiles.

Comment 1

[Boris Zbarsky [:bzbarsky]]

Attachment: Profile of the testcase (zipped HTML)

Note the amount of time spent under nsIFrame::Invalidate here...

Comment 2

[Boris Zbarsky [:bzbarsky]]

Attachment: Profile of just the cairo_fill part

Note that we spend about 25-30% of total time in malloc and free...  Would be
nice not to do that.  ;)

Comment 3

[Jo Hermans]

Maybe we can use nsRecycleAllocator for this ?

Comment 4

[Boris Zbarsky [:bzbarsky]]

For which?  The memory thrashing is in Cairo, not in Mozilla code.

Comment 5

[Jo Hermans]

(In reply to comment #4)
> For which?  The memory thrashing is in Cairo, not in Mozilla code.

I was talking about a similar optimization, not necesarily calling the
Mozilla-code directly. I see that a lot of malloc/frees are done, but no cacheis
used. A pool would only work if you're allocating blocks of the same size over
and over again. Some parts could also be allocated on the stack
(pixman_region_create/pixman_region_destroy seems to be a good candidate for
instance). But that's after a 10 minute walk through the profile data and code.

Comment 6

[Boris Zbarsky [:bzbarsky]]

Right, but that would be something to talk about in a bug we file on cairo
(which someone should do, for that matter).

Comment 7

[Boris Zbarsky [:bzbarsky]]

On current trunk, _moz_cairo_fill is about 8% of total time.

We still spend a lot of time on frame invalidation, but there is a whole lot less memory thrashing.  total time in malloc/free is down to 4% or so.

Comment 8

[Lucas Malor (mail: c6kfnkn2uc AT snkmail DOT c0m)]

I've done the same comparison as I wrote in Bug 297959 Comment 18. The results for this testcase is:

Fx3:

Elapsed wall-clock time: 3554ms
Elapsed non-idle time: 2914ms
Speed: 4.22fps

Opera:

Elapsed wall-clock time: 1984ms
Elapsed non-idle time: 1344ms
Speed: 7.56fps

Considerations I made in Bug 297959 Comment 18 are still valid here.

Comment 9

[Boris Zbarsky [:bzbarsky]]

The correct comparison is to ideal fps, not to other browsers.

Comment 10

[Joe Drew (not getting mail)]

This is a metabug that hasn't seen any use in the past 3 years; I think its time has passed.