Mysterious warning compiling RAJA + CUDA on Lassen

[kab163]

Mysterious warning compiling RAJA + CUDA on Lassen

[rhornung67]

warnings like this tend to be somewhat inconsistent across compilers and versions. The warning doesn't show up when using nvcc 11.2.0 for example (same gcc 8.3.1). IMO, nvcc 10.2.89 is complaining about something that it should not.

[rhornung67]

The stack size warnings are common and can probably be ignored if the code runs correctly. For example: kokkos/kokkos#2039

[MrBurmark]

I have seen these warnings as well. I'm pretty sure the compiler shouldn't be complaining as std::move is allowed in device code. I have been ignoring them and hoping things turn out ok, the device sort test pass and they do use this code on the device.

[MrBurmark]

Maybe we could work around these warnings by making our own version of move or using a static cast instead.

[kab163]

warnings like this tend to be somewhat inconsistent across compilers and versions. The warning doesn't show up when using nvcc 11.2.0 for example (same gcc 8.3.1). IMO, nvcc 10.2.89 is complaining about something that it should not.

Let me try cuda 11. If I can reproduce that (having the warning go away) then I think it would be reasonable to assume it's a non-issue. Thanks, Rich and Jason!

[kab163]

Ok, that warning disappeared when using cuda/11.2.0.

Do we care about the ptxas warnings? I get a bunch of these:

ptxas warning : Stack size for entry function '_Z18forone_cuda_globalIZN14IntroSortPairsI11forone_cuda11RunOnDeviceEclIN4RAJA4SpanIPdlEENS6_IPllEENS5_9operators4lessIddEEEEvOT_OT0_T1_EUlvE_EvT_' cannot be statically determined

But ya, we can close this issue then. Thanks again!