405894 - [Cairo][Mac PPC] 131% perf regression when scrolling pages with large background images

Status: RESOLVED FIXED

(Core :: Graphics, defect, P4)

Description

[philippe (part-time)]

test case: https://bugzilla.mozilla.org/attachment.cgi?id=212976

spin off from bug 364221, for the evolution of the numbers see that bug.

Gecko 1.9     Gecko 1.8     Safari 3.04
17220         7417          10089
17280         7347          10003
17070         7475          9906
17056         7311          9907
17236         7533          9895
-----         -----         -----
17,172.4      7,424.6       9,960 (average over 5 runs)

Gecko 1.9: 2007112816 Minefield/3.0b2pre
Gecko 1.8: 20071128 BonEcho/2.0.0.11pre

Tested on a PowerBook G4, 1.5Mhz, 1.5Gb ram

Comment 1

[Mike Schroepfer]

Vlad - what do you want to do with this?

Comment 2

[Mike Schroepfer]

(In reply to comment #1)
> Vlad - what do you want to do with this?
> 

Ping?

Comment 3

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

Are those tests running with the image tiling or not?  That is, all the windows need to be the same size for the numbers to have any meaning; ideally they'd be run with tiling and without tiling (to get without tiling, just shrink the window until you don't see the image repeat).

Either way, I have a patch in cairo that will help in the tiling case, but under 10.5-only; in the non-tiling case, as best I can tell we're going straight to Quartz, so I'm not sure why the ppc numbers are slower.

Comment 4

[Damon Sicore (:damons)]

+'ing with P4.

Comment 5

[philippe (part-time)]

(In reply to comment #3)
> Are those tests running with the image tiling or not?  That is, all the windows
> need to be the same size for the numbers to have any meaning; ideally they'd be
> run with tiling and without tiling (to get without tiling, just shrink the
> window until you don't see the image repeat).

All tests I did ran within windows approx. 1100px wide. To get the image to tile, I'd need a much larger monitor (my PowerBook G4 has 1280px screenwidth, the image in the test is 1280px wide).

Comment 6

[philippe (part-time)]

Ok so I forced my window to be wide: about 1600px. Average of 5 runs for the test: 53156ms on current trunk build. yes: 53156ms.

I then made a second test case, using a tall/narrow image (500px by 2500px) (I'll it attach next).
With an ~1100px wide window (and 700px tall):
background-image set to no-repeat: average of 13593ms
background-image set to repeat-y: 13465ms
background-image set to repeat-x: 50147ms  <---------- !
(with repeat-x, the is image visible 3 times)

Comment 7

[philippe (part-time)]

Attachment: image for test case

Comment 8

[philippe (part-time)]

Attachment: test case using tall narrow image

Comment 9

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

bug 414685 should help significantly with this, especially on 10.4.

Comment 10

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

philippe, could you retest this on your ppc mac with the latest nightly?

Comment 11

[philippe (part-time)]

Nice improvement :-).

before		after
19826ms		12896ms
20037ms		12857ms
19738ms		12728ms
20230ms		12747ms
19602ms		12813ms
--------	--------
19886.6		12808.2

before: 2008012717 Minefield/3.0b3pre
after: 2008020615 Minefield/3.0b4pre

When the image is tiled (test case attached to this bug with repeat-x), I get an average of 16683ms (vs 50147ms noted in comment 6).

On a PowerBook G4, 1.5Ghz, 1.25 Gb.
The numbers 'before' are somewhat slower than those in comment 0; I've used a slower machine (less RAM). My own PowerBook died over the new year. RIP.

Comment 12

[Mike Schroepfer]

> On a PowerBook G4, 1.5Ghz, 1.25 Gb.
> The numbers 'before' are somewhat slower than those in comment 0; I've used a
> slower machine (less RAM). My own PowerBook died over the new year. RIP.
> 

Can you compare against 1.8 for same machine?

Comment 13

[philippe (part-time)]

(In reply to comment #12)

With Fx 2.0.0.11: 9719.4ms
the tiled image test: 10160ms
both average for 5 runs.

Comment 14

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

There's a very good chance that this was fixed completely after bug 418311 landed -- could you retest?

Comment 15

[philippe (part-time)]

(In reply to comment #14)
> There's a very good chance that this was fixed completely after bug 418311
> landed -- could you retest?
> 
I wish I could -(. I don't have permanent access to a PPC machine anymore. The owner of the PowerBook I used for testing in comment 11 is on the road 'till the end of the week. If anyone else can test this...

Comment 16

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

Ok, lets call this done for now -- if someone gets a chance to retest, please reopen.