How to parallelize many small SVD's on a GPU

I'm computing SVD's of millions of small (e.g. 5 x 10) matrices. It is surprisingly fast as a for-loop on the CPU, however I need the computation to be even faster and would like to know if parallelization can be leveraged on a GPU. SVD doesn't currently appear to be a function supported by pagefun for gpuArray objects. I'm using R2020a and the Parallel Computing toolbox. Thanks

Answer 1

[Christine Tobler]

As of R2021b, there is a function pagesvd which is supported on both CPU and distributed arrays. It's not supported for gpuArray as of now.

[Tim]

Thanks for letting me know; unfortunately the distributed case doesn't really help me much since I assume SVD is already multithreaded and I'm limited to a single laptop with a high-power GPU. But I have a couple questions:

• Does pagesvd offer much in the way of a speedup vs. a loop for the CPU case?
• Are there plans to extend the pagesvd function to gpuArrays?

[Matt J]

Does pagesvd offer much in the way of a speedup vs. a loop for the CPU case?

We can try it right here:

A=rand(5,10,1e6);

tic
 pagesvd(A);
toc

Elapsed time is 2.531096 seconds.

tic
for i=1:size(A,3)

     svd(A(:,:,i));

end
toc

Elapsed time is 8.363056 seconds.

[Tim]

Brilliant - I know we can write formatted code in the answer/comment boxes but somehow it skipped me that we can actually evaluate the code as well. 3x is a nontrivial improvement but what I would really like to know is if there is a GPU implementation coming soon.

[Christine Tobler]

Thanks for your interest, I've passed this information along to the gpuArray team.

[Matt J]

On my laptop ( Intel(R) Core(TM) i5-7300HQ CPU @ 2.50GHz, 2501 Mhz, 4 Core(s), 4 Logical Processor(s) )

Elapsed time is 1.379264 seconds. %pagesvd
Elapsed time is 7.861928 seconds. %for loop

[Joss Knight]

gpuArray/pagesvd coming next year, watch this space...

[Tim]

Great news, thank you Joss

Answer 2

[Matt J]

SVD doesn't currently appear to be a function supported by pagefun for gpuArray objects

Perhaps not for gpuArrays, but it is supported for distributed arrays:

https://www.mathworks.com/help/parallel-computing/pagefun.html#mw_df48aa75-069c-4420-b3b2-933cd686b577