MacResearch

A couple of things to try.

1) Have you profiled your code with Shark? It's part of the CHUD tools available from the Apple site.

2) There is a mailing list that Apple hosts called performance-optimization (lists.apple.com) if you aren't on it already, you may wish to post there. A lot of experts monitor that list (including people from the performance group at Apple) so you can often get good answers there. Of course the first thing they'll ask is to see a Shark trace :)

3) In terms of compilers. What are you using? The IBM compilers in some cases (many in my hands) will produce code that is sometimes 2x faster than gcc. Although between gcc 4.0 and XLC I've found the performance gains to be less than between gcc 3.3 and XLC.

But the first thing to profile your application in Shark (if you haven't already got one).

Hope that helps,

Dave

Hi Stuart,

Regarding:

2) Correct. XLF/XLC are not supported un 10.4. Both will work (XLF fully and XLC will compile but not link). However, you can use XLF to link XLC compiled code. There are other things you can do to "fix" the XLC portion, but that requires modifying some system files, which you may not want to do.

3) I would highly recommend using Shark (as part of CHUD), since as is often the case, the places where we think a program is spending its time, it's really not. Shark is probably one of the best (and easiest to use) profiling tools for perf op.

Before spending any time trying to parallelize via threading or MPI I'd very much recommend making sure that all obvious things in your scalar code are running as well as possible. Again, Shark is your friend and what we think may be an issue often is not, and things we don't think to look at are big problems.

Regards,

Dave

Have you seen these?

http://pythonmac.org/wiki/XcodeIntegration

http://ulaluma.com/pyx/archives/2004/02/running_python.html

Hi Stuart,

Not all is lost if you are finding that trig functions are slowing you down. If you can rearrange your code (or if you are already doing a lot of trig functions at the same time) you may be able to vectorize your code. Or better yet the Accelerate library already has many high performance versions of trig functions that operate on single vector quantities or if you are on 10.4 vForce (also part of Accelerate) can be used to operate on arrays of data to do a lot of similar trig functions at once.

If you require double precision or greater, then Accelerate won't work necessarily. But there are still options (including writing your own trancendentals that can be pipelined/inlined, as opposed to using the standard math library versions (which usually have to guarantee a right answer sometimes at the expense of speed).

Just something to think about.

Regards,

Dave