Slow GPU performance with scan, slinalg, nlinalg

[Paul Baggenstoss]

Hello,

   I am a big Theano fan, although somewhat new to it.

 I am experiencing slow GPU performance for a Theano program. To be more specific,

it is about as fast as with CPU.   Because I have other programs

that get huge performance speed-ups relative to CPU, I know it is not an installation issue.

If Theano was paralelizing the operations over the independent samples

in a batch, then I should see increasing speed as I increase the batch size.

Using nvidia-smi, I checked that the GPU is not over-loaded.

I suspect it has to do with using  of the following functions:

   slinalg.Solve

   slinalg.Cholesky

   nlinalg.MatrixInverse

   nlinalg.Det

and there are used inside 'scan' loops.  Generally, I scan over the samples in a batch,

(i.e. the first index of the 'sequences' tensors are the samples in a batch. ).

I also use tensor.jacobian inside a scan loop.

Without getting into details of the program, is there any obvious reason that

Theano does not seem to be paralellizing my code over the GPU?

Many thanks in advance,

Paul

[Paul Baggenstoss]

I found the problem - My installation of scikit-cuda failed.
I needed to re-install using conda