Yielding GPU to windows briefly to prevent resets?

[Dustin Moore]

Is there nice way to take a @jit function and add in a breakpoint to allow

windows to use the gpu for any business it needs to and then return to

numba? I have calculations that will need many minutes to run on the

GPU and windows resets the gpu driver if anything takes more than

a few seconds. Thanks,

[Stanley Seibert]

Unfortunately, this has been a limitation of CUDA since it was created.  (And so Numba can't really do anything about it.)  There is no way for a running CUDA kernel to yield execution to the graphics display driver without terminating.  The only solution is to break up your calculation into multiple kernel launches that each take less than a few seconds.

--
You received this message because you are subscribed to the Google Groups "Numba Public Discussion - Public" group.
To unsubscribe from this group and stop receiving emails from it, send an email to numba-users...@continuum.io.
To post to this group, send email to numba...@continuum.io.
To view this discussion on the web visit https://groups.google.com/a/continuum.io/d/msgid/numba-users/62c55156-3a90-466a-9893-f785dbc881ad%40continuum.io.
For more options, visit https://groups.google.com/a/continuum.io/d/optout.

[Graham Markall]

Alternatively, if this is a possibility in your situation, you could install a second graphics card for driving the display and leave the GPU unconnected to any displays - this will remove the 5 second kernel execution time limit.

To view this discussion on the web visit https://groups.google.com/a/continuum.io/d/msgid/numba-users/CADv3RKTswP3tPk8BY2SsicMtwYD_q8nzwu3iXwjkDDvGckHExg%40mail.gmail.com.