CODEWITHSUNDEEP
matlaboptimizationfft
Mikhail
Mikhail

Reputation: 8028

Avenues for overcomming FFT bottleneck?

I spent some time optimizing my algorithm and my quasi-serial (no explicit parralelization) code spends 95% of the time on a line that performs the fftn and dense single(float) matrix multiplication

for k=1:10
    q = q +  x{k}.* fftn( mArray{k}.* ifftn( mOther{k} .* z ) );

I tried adding some wisdoms for the FFT although the performance increase was negligible.

I am at a loss for ways to speed up this code, do you think compiling FFTW could result in a performance increase? I am using Matlab 2012b for a 3rd generation i7.

Edit

I seem to have made a typo, x depends on k, it would have been too easy otherwise. I was hoping somebody could speak to optimizing the actual fft.

    q = q +  x.* fftn( mArray{k}.* ifftn( mOther{k} .* z ) );
    q = q +  x{k}.* fftn( mArray{k}.* ifftn( mOther{k} .* z ) );

Upvotes: 1

Views: 136

Answers (2)

Shai
Shai

Reputation: 114856

Have you considered using convolutions instead of back and forth FFTs?

Upvotes: 0

Andreas H.
Andreas H.

Reputation: 6095

You should add the q's together before fftn transforming and multiplying by x. e.g.

A = 0;
for k=1:10
     A = A + mArray{k}.* ifftn( mOther{k} .* z );
end
q = q + x.*fftn(A);

IMO this should be equivalent.

Upvotes: 2

Related Questions