BiCGStab is slower on R2021a using GPUs

Hello,

I have both matlab R2020a and R2021a. I have been using rtx 3090 GPU to do my calculations

I noticed preconditioned BiCGStab algorithm on R2020a is 3 times faster than the R2021a version using the GPU.

That doesn't make sense because R2020a has forward compatibility for my GPU where R2021a has native support for the ampere GPUs

If I'm missing something what could it be?

Answer 1

[Joss Knight]

There is a MATLAB bug report on this known issue:

https://www.mathworks.com/support/bugreports/details/2438221

You should make sure you are using the latest Update for R2021a.

[deniz]

I'm on update 3 but it hasn't been solved yet I believe

[deniz]

I can also provide more detail. I also wrote down my own regression algorithm for BiCGStab. I noticed the slow part wasn't the sparse matrix vector multiplication but it was the preconditioning phase where sparse triangular matrix solving is performed. I thought that was the problem. On the other hand, the bug report says the problem is the basic sparse operations like multiplication, summation and negation.

For some reason, the problem still persists for me even on the 3rd update. However, the bug report says it has been solved. I don't know if anybody has confirmed it yet. Or should I wait for more updates on this issue?

[Jakub Sikorowski]

You're right Deniz, that there are two issues in the gpu sparse iterative solvers. The first is described in the EBR provided by Joss. The second that affects the sparse triangular solvers used in GPU preconditioned solvers is under investigation. I'm afraid that currently, we do not know of any workaround other than using older Matlab than 21a.

[deniz]

It seems the problem still persists on Matlab R2021b

[Joss Knight]

The MATLAB bug report to track for this particular issue is:

https://www.mathworks.com/support/bugreports/2534618

The problem is proving difficult to work around due to changes in the underlying NVIDIA libraries, for which we apologise. We have to upgrade these libraries to support the latest hardware.

If you can manage without using a preconditioner then you can avoid the problem.