Re: [eigen] GPU with eigen

[Karl Rupp <rupp@xxxxxxxxxxxxxxxx>]

Hi Billy,

> I would like to test using the GPU with Eigen. I was thinking of using
> ViennaCL or GPUMatrix. Although GPUMatrix seems a bit new. Has anyone
> have experience which can be shared about using Eigen on GPU?

I'm certainly biased, so I'm not giving you a recommendation in favor of 
one of the packages.

Nevertheless, I can give you general advice on what to expect and what 
not to expect from GPUs. In simple terms, make sure that you have 'large 
operations' with sufficient parallelism to be executed on the GPU. As a 
rule of thumb, make sure that vectors participating in vector operations 
have at least ~10k-50k entries, otherwise the kernel launch overhead (in 
the range of 10 us, which is about 20k CPU cycles!) will constitute 
significant overhead. Similarly, sparse matrices should have a similar 
number of rows/columns to make GPUs efficient. For dense matrices, their 
size should be at least 500x500 for the same reason. For more 
complicated algorithms such as QR factorizations, this 'minimum useful 
size' moves further towards larger sizes because of additional 
PCI-Express communications.

Hence, if you mostly operate on small vectors and matrices, for which 
Eigen is really great, GPUs may give you a really hard time. In some 
cases it's possible to pack data together, but even then manual tweaking 
would be required to mitigate the costs of data rearrangements. I tend 
not to recommend GPUs at all in such case, unless the operations are 
embarrassingly parallel by nature.

Hope that helps :-)

Best regards,
Karli