[cairo] Cairo 1.3 performance loss

[Jorn Baayen]

I attach a patch (incremental to Davids) that incorporates Behdad's
suggestion of repeating tests to run for a minimum of one second. This
seems to work quite nicely.

Thanks,

Jorn

On Wed, 2007-02-07 at 09:41 +0200, Jorn Baayen wrote:
> Hi,
> 
> This is exactly what I had in mind. It turns out though that this makes
> certain tests take ages (as expected), while others still run too fast
> to be able to take any measurement. With -m 10 cairo-perf already takes
> ~ 10 hours to run, so it's painful to experiment with higher values. 
> 
> Thanks,
> 
> Jorn
> 
> On Thu, 2007-02-01 at 16:46 +0100, David Turner wrote: 
> > Sorry again, could you try the following patch, it should do the
> > trick with the -m flag. However, I don't think it may be merged into
> > the git repository right now, since the changes do change a few
> > measurements.
> > 
> > Hope this helps,
> > 
> > - David
> > 
> > 
> > On Thu, 01 Feb 2007 13:10:25 +0200, "Jorn Baayen" <jorn at openedhand.com> said:
> > > Hi,
> > > 
> > > On Thu, 2007-02-01 at 11:51 +0100, David Turner wrote:
> > > > > This still times individual iterations (but runs X of them and takes the
> > > > > median). The problem is that on this machine gettimeofday()'s
> > > > > granularity is too coarse to register any difference in time for many of
> > > > > the tests.
> > > > > 
> > > > 
> > > > Sorry for the confusion, here's a patch that adds a '-m' flag to cairo-perf
> > > > that can be used to specify test case multipliers. You can also define
> > > > the CAIRO_PERF_MULTIPLIER environment variable.
> > > 
> > > Unfortunately this doesn't help. Every test case does this:
> > > 
> > > cairo_perf_start ();
> > > do_something ();
> > > cairo_perf_stop ();
> > > return cairo_perf_elapsed ();
> > > 
> > > Now on ARM _start() and _stop() record the time using gettimeofday(),
> > > and _elapsed() takes the difference. But many tests run so quickly that
> > > they are done in less than a centisecond, which is all the timer 
> > > granularity there is here. This means the times get returned as 0, and
> > > adding together many zeroes doesn't help. 
> > > 
> > > To add a 'time many iterations in one go' mode, a timer would need to be
> > > added around the whole iterator, and that's perhaps not very elegant
> > > as we also have timing code inside the tests themselves ..
> > > 
> > > Sorry for the confusion!
> > > 
> > > Thanks,
> > > 
> > > Jorn
> > > 
> > > > 
> > > > Hope this helps,
> > > > 
> > > > - David
> > > -- 
> > > OpenedHand Ltd.
> > > http://o-hand.com/
> > > 
-- 
OpenedHand Ltd.
http://o-hand.com/
-------------- next part --------------
A non-text attachment was scrubbed...
Name: cairo-perf.patch
Type: text/x-patch
Size: 2421 bytes
Desc: not available
Url : http://lists.freedesktop.org/archives/cairo/attachments/20070207/45e98cc7/cairo-perf.bin