Is it possible to get MKL ERROR messages despite running on the GPU

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)
Post Reply
themightyoarfish
Posts: 1
Joined: Sat Jan 19, 2019 4:20 pm

Is it possible to get MKL ERROR messages despite running on the GPU

Post by themightyoarfish » Sat Jan 19, 2019 4:23 pm

I'm using SVD PyTorch, which on the GPU uses MAGMA dgesdd as far as I know. In some cases, my inputs contain NaNs, in which case I get
Intel MKL ERROR: Parameter 4 was incorrect on entry to SLASCL.
Now, this confuses me a bit, since MKL is a CPU library so far as I know, so how is it possible to get error messages from MKL when executing stuff on the GPU?

mgates3
Posts: 918
Joined: Fri Jan 06, 2012 2:13 pm

Re: Is it possible to get MKL ERROR messages despite running on the GPU

Post by mgates3 » Tue Apr 23, 2019 4:42 pm

MAGMA is a hybrid CPU-GPU library. It does some computation on the CPU, and some on the GPU, trying to overlap those two computations and leverage the strengths of each. Thus, it can run into MKL errors on the CPU.

As for your error, gesdd cannot work if there are NaNs. The SVD is not defined in that case.

-mark

[Sorry for the delay in posting. A misconfiguration left messages in the moderation queue.]

Post Reply