Compiling FFTW with --enable-avx

[Viral Shah]

Does anyone have experience with using FFTW with --enable-avx, and the kind of performance gains one gets? I am experimenting with enabling threads and such in BLAS and FFTW to start with on the local_blas branch.

-viral

[Steven G. Johnson]

Viral Shah wrote:
> Does anyone have experience with using FFTW with --enable-avx, and the kind of performance gains one gets? I am experimenting with enabling threads and such in BLAS and FFTW to start with on the local_blas branch.

It is definitely worth it; last I heard we were getting 11-17gflops on a
single Sandy Bridge core with AVX in double precision for small to
moderate sized transforms. More in single precision.

Note that FFTW checks at runtime to determine whether the CPU supports
AVX instructions (and SSE etcetera), and disables any unsupported
kernels. So it is safe to compile with --enable-avx --enable-sse2 ...
the resulting binary should still work on fairly ancient machines.

--SGJ

[Stefan Karpinski]

That sounds like an all-around yes to me. Seems like something we should definitely do then.

--SGJ

--

[Viral Shah]

Fantastic. I believe we used to do that at one point, and it was disabled due to build issues. The issue, if I remember correctly was that compilers that did not have AVX support would barf altogether. This is also the reason that OpenBLAS is no longer built with DYNAMIC_ARCH=1.

We need to figure out how to make this work.

-viral

 

[Steven G. Johnson]

On 11/26/12 3:20 PM, Viral Shah wrote:
> Fantastic. I believe we used to do that at one point, and it was
> disabled due to build issues. The issue, if I remember correctly was
> that compilers that did not have AVX support would barf altogether. This
> is also the reason that OpenBLAS is no longer built with DYNAMIC_ARCH=1.

Maybe something like:

./configure --enable-avx --enable-sse2 || ./configure --enable-sse2

? The configure script checks if -mavx is accepted by the compiler and
exits otherwise.

Steven

[Jeff Bezanson]

Welcome, Steve. Great to have you on the list and thanks for answering this stuff.

Steven

--

[Viral Shah]

That should work. I'll try it out in the next couple of days, unless someone beats me to it.

Also, Steve, great to see you on the list!

-viral

[Elliot Saba]

Hey Steve, is there a way I can double check to make sure that FFTW is indeed using AVX?  Some symbol that's exported I can check with `nm` or some function I can call?

Thanks!

-viral

--

[Steven G. Johnson]

On 12/9/12 9:36 PM, Elliot Saba wrote:
> Hey Steve, is there a way I can double check to make sure that FFTW is
> indeed using AVX? Some symbol that's exported I can check with `nm` or
> some function I can call?

The symbol fftw_have_simd_avx (or fftwf_have_simd_avx in single
precision) is exported if FFTW is compiled with AVX enabled.

(This defines a function fftw_have_simd_avx() that checks at runtime if
AVX is supported on the current CPU and returns 1 if AVX is available.)

This is not documented, however, so I can't guarantee that it won't
change in future FFTW versions. If this is not a one-time thing and
you need some permanent functionality of this sort, could you let us
know why?

--SGJ

[Elliot Saba]

This is a one-time thing, I just want to verify that FFTW is able to use AVX on my processor to make sure the build process is compiling what we want it to.

Thanks!

-E

--SGJ

--