[cairo] Performance work update

[Carl Worth]

On Mon Aug 15 11:12:13 2005,  wrote:
>   I put together a quick set of benchmarks for some cases I care about:
> compositing, line drawing, and image scaling (up and down).
> 
>     http://www.freedesktop.org/~vektor/cairo-benchmarks-0.1.tar.gz

Billy, thanks for putting this together!

Is there any reason this shouldn't go straight into CVS? I'll use my
creative naming skills to suggest a module location of:

	cairo-benchmarks

>   I'd like to hear feedback on what operations people are finding slow
> and maybe put together a more comprehensive set of benchmarks.  Please
> email me if you have any ideas or code.

One idea that I think would be helpful is a way to run the suite of
tests and get some tabulated results. Maybe something with rows for
each test and columns for each backend. And hopefully a table small
enough to be easy to send back and forth in an email message, say.

>   - I did an implementation of fbCompositeSrc_8888x8888mmx which makes
>     my "over" test twice as fast.
...
>   - I also found that mmxCombineMaskU and mmxCombineOverU missed special
>     cases for 0 alpha and 0xff alpha.  Adding those checks gave a 2x
>     speedup for my line test.

Keep up the good work! Are these new things all in cairo already?

>   - I am getting some strange results with Xfake.  My "lines" test is
>     slower (at least 50%) with the MMX code enabled than with it
>     disabled.  This is not reproducable with pixman + fbmmx, where the
>     MMX code is much faster.

I have no insight on this one.

>   - Some of the xlib paths with certain servers are terribly slow.  With
>     XFree86 4.3 on my Debian machine, the "over" test is 100x slower
>     than the pixman version.  I think we may need a better strategy for
>     dealing with this problem.

I seem to recall Owen sending out an email not too long ago that
outlined a plan for dealing with the "old X server" problem. I can't
seem to find the mail now, but pieces of the plan involve deciding how
old an X server we should plan to support, and deciding whether the
right answer is to just always use the image backend rather than doing
some part-Render part-pixman approach.

-Carl
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://lists.freedesktop.org/archives/cairo/attachments/20050818/17ec0442/attachment-0001.pgp