CODEWITHSUNDEEP
matlabparallel-processing
David P
David P

Reputation: 51

Convolution along singleton dimensions on GPU in Matlab

I am trying to implement a convolution along one dimension of a large N-D array in parallel in Matlab. The convolution filter is small enough that it doesn't make sense to use an FFT to do it. For very long 1D signals, I know I can call conv(gpuArrayA,gpuArrayB), and for many functions I can use bsxfun to apply a function in parallel across a single dimension of a gpuArray, however when I try

bsxfun(@conv,gpuArrayA,gpuArrayB)

I get the error

Error using gpuArray/bsxfun
Use of 'conv' is not supported.

So then is there a way to implement this in Matlab without having to write custom CUDA code?

Upvotes: 1

Views: 451

Answers (1)

Matt J
Matt J

Reputation: 1137

Couldn't you just call convn() with a 1D convolution kernel, e.g.,

 A=gpuArray(rand(100,100,100));
 kernel=gpuArray([1;1;1]);

 result = convn(A,kernel,'same');  % convolution along dim=1

I find I get faster results with this than an equivalent CPU version, so there must be some GPU parallelization built into this.

Upvotes: 1

Related Questions