How to execute fft's with gpu,cuda in parallel (spmd)?

I have IxJxL single type data volume. I want to execute 1-D fft's of L length IxJ times.

Win7-64,GFX670,MatlabR2012a

data_device = gpuArray(data);

Nested 'for' loop on GPU works slower than on CPU.

for ii = 1:I
  for jj = 1:J
     data_device(ii,jj,:) = fft(data_device(ii,jj,:));
  end
end

'Parfor' works even slower than simple 'for' loop.

1 Is there a way to use spmd method on GPU alike its done on CPU? How?

2 What is the optimal size of data to be send to GPU, considering params gpuDevice() function retuns?

Answer 1

[Edric Ellis]

MATLAB's FFT function can operate along any single dimension. So you can simply do:

data_device = fft(data_device, [], 3);

rather than having a loop. See the FFT reference page for more.

[Greg]

Very true, indeed, thank you!

Still, Edric, GFX670 has 7 cores, so I thought, I could split data_device matrix into 7 smaller matrices, so each core would work on its own fft(data_device_(1/7)_part, [], 3); in parallel... Is that possible?

[Ben Tordoff]

Hi Greg, assuming you mean a GTX670, it actually has 1344 cores (7 processors, each with 192 cores). However, you cannot address a core individually. Infact you can't even address a processor individually. You need to send the GPU one big operation to perform and let it work out how to split it over the heirarchy of symmetric multi-processors and the cores they contain.

Our recommendation is always to send as much work (and data) in a single command as you can.

[Greg]

Ben, thanks a lot.