Hmm I suspect the problem is that GPU is simply too old yes, but perhaps there is a simple enough workaround available in the code as you suggest. I will investigate further on Monday.Best regards,Jacob Faibussowitsch(Jacob Fai - booss - oh - vitch)On Jan 6, 2023, at 09:55, Mark Lohry
These cards do indeed not support cudaDeviceGetMemPool --
cudaDeviceGetAttribute on cudaDevAttrMemoryPoolsSupported return false,
meaning it doesn't support cudaMallocAsync, so the first point of failure
is the call to cudaDeviceGetMemPool in the initialization.
Would a workaround be to replace
It built+ran fine on a different system with an sm75 arch. Is there a
documented minimum version if that indeed is the cause?
One minor hiccup FYI -- compilation of hypre fails with cuda toolkit 12,
due to cusprase removing csrsv2Info_t (although it's still referenced in
their docs...) in favor
Jacob, is it because the cuda arch is too old?
--Junchao Zhang
On Thu, Jan 5, 2023 at 4:30 PM Mark Lohry wrote:
> I'm seeing the same thing on latest main with a different machine and
> -sm52 card, cuda 11.8. make check fails with the below, where the indicated
> line 249 corresponds to
I'm seeing the same thing on latest main with a different machine and -sm52
card, cuda 11.8. make check fails with the below, where the indicated line
249 corresponds to PetscCallCUPM(cupmDeviceGetMemPool(,
static_cast(device->deviceId))); in the initialize function.
Running check examples to
I'm trying to compile the cuda example
./config/examples/arch-ci-linux-cuda-double-64idx.py
--with-cudac=/usr/local/cuda-11.5/bin/nvcc
and running make test passes the test ok
diff-sys_objects_device_tests-ex1_host_with_device+nsize-1device_enable-lazy
but the eager variant fails, pasted below.