Matrix low size limit

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)
Post Reply
paoloviviani
Posts: 14
Joined: Mon Nov 09, 2015 12:17 pm

Matrix low size limit

Post by paoloviviani » Mon Feb 13, 2017 12:52 pm

Dear all,
is there a chance that, for matrices smaller than a certain size, MAGMA decides to NOT use the GPU at all and reverts to the CPU lapack implementation (e.g. openblas)?

Thanks in advance,
Paolo

paoloviviani
Posts: 14
Joined: Mon Nov 09, 2015 12:17 pm

Re: Matrix low size limit

Post by paoloviviani » Tue Feb 21, 2017 6:43 am

I'm asking this since I'm comparing Openblas ssyevd against Magma ssyevd and I see that, at least until the matrix size is 128x128, the timing are exactly the same, while for bigger matrices the behavior changes.

mgates3
Posts: 842
Joined: Fri Jan 06, 2012 2:13 pm

Re: Matrix low size limit

Post by mgates3 » Wed Feb 22, 2017 12:33 am

Yes, small matrices are simply handed to LAPACK. There is not enough work for the GPU to be efficient and overcome the cost of copying the matrix to the GPU.
-mark

Post Reply