[Numpy-discussion] OpenMP-fication of loops.

[Eric Carlson]

On 1/23/2011 10:36 AM, Vladimir Voznesensky wrote:
> My computer has 12 hyperthreaded cores.
> My application uses dot multiplication from Intel MKL, that accelerated
> it by ~ 5 times.
> After OpenMP-fication of loops.c.src, my app was accelerated by ~12-15
> times.
>

I was greatly disappointed in the parallel performance on a new 
workstation for some of my programs. I could not get better than about a 
factor of 5 on my dual xeon with 24 threads.

Last Fall, I stumbled across this example of OpenMP with f2py,

https://gist.github.com/226473

that I built on Ubuntu 10.04 x64 using (slightly different than the 
instructions):

f2py -c -m deemingomp periodogram.f90 --f90flags="-fopenmp " -lgomp 
-lf77blas -lcblas -latlas

On my machine for larger array sizes, I saw speed-ups of 20x over single 
thread in the example program. Indeed, the example serves as an 
excellent way to test the thermal stability of the workstation.

I did not get a chance to follow this up yet, but if you can get 12x 
improvement with normal numpy codes, I am very interested...

Cheers,
EC