MATLAB Answers

0

Checking if persistent variable exists on GPU: check if 'data no longer exists on the GPU' for a variable.

Asked by D. Plotnick on 6 Dec 2018
Latest activity Edited by D. Plotnick on 6 Dec 2018
Hello,
I am running into an issue if I have persistent variables that I am storing as GPU arrays. This is especially useful if they are somewhat large, static variables, since I don't need to move them on/off the device or recalculate them on every function call. However, if I have foolishly reset my gpuDevice, I run into issues. Typically I will have code of the form:
persistent x
if isempty(x)
x = calculateX(argin)
end
I then run into an issue on isempty, which unsurprisingly throws an error 'The data no longer exists on the device.' The isempty boolean never returns, the program returns an error, and the persistent variable is never assigned.
What I really need is a logical check as to whether a variable has been cleared on the device, but I cannot seem to find any obvious solutions. Everything else I have thought of seems too kludgy:
(A) use a try-catch statement around the isempty statement (yuck).
(B) have a custom reset(gpuDevice) function that also calls clear on all of the functions that use persistent gpuArrays (more yuck)
(C) have a try-catch statement at the header of each function that will reset all persistent variables if it throws an error.
Since the obvious solution is just a logical telling me whether a variable has been cleared, these do not seem like the best options. There, of course, may be good reasons to avoid storing persistent variables on the GPU, but I haven't seen them.
Thanks,
-Dan

  0 Comments

Sign in to comment.

1 Answer

Answer by Edric Ellis
on 6 Dec 2018
 Accepted Answer

You need existsOnGPU, like so:
x = gpuArray(1);
assert(existsOnGPU(x));
reset(gpuDevice);
assert(~existsOnGPU(x));

  1 Comment

Thanks, that is exactly the command I needed.
Note (this may be intentional) but
doc existsOnGPU
sends you to the gpuArray help document. You then have to access existsOnGPU via the See Also links.

Sign in to comment.