Improve SLIC macros when axom is configured with CUDA

kennyweiss commented 4 years ago

A recent PR (https://github.com/LLNL/axom/pull/156) adds preliminary support for slic macros when axom is configured with cuda support.

Specifically, it converts all calls to SLIC_ASSERT* and SLIC_CHECK* to a CUDA assert, and drops the message.

There are several improvements/upgrades that we should consider adding in the near future:

[ ] Unit tests that specifically test the slic macros in different spaces (host vs. device) and configurations (e.g. raja, cuda, openmp, ...)
[ ] CUDA assert should only be called on devices that support this. See: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#assertion
[ ] SLIC_CHECK should respect the axom::slic::debug::checksAreErrors flag
[ ] Support should be added for SLIC_ERROR*. This should respect the slic::isAbortOnErrorsEnabled() flag
[ ] Support should be added for SLIC_WARNING*. This should respect the slic::isAbortOnWarningsEnabled() flag
[ ] We should consider how we can support the *_MSG variants of the slic macros. E.g. perhaps we cannot support the full ostream operator within these messsages, but we might be able to support a simpler printf-style output
[ ] We should consider if/how we can support simple message output (e.g. via SLIC_INFO) within kernels. One idea would be to just printf the message. The team has discussed another idea brought up by a user to have errors trigger predefined message (currently tracked by JIRA issue ATK-1299).

rhornung67 commented 2 years ago

Slic macros are also not thread-safe for OpenMP. This should also be addressed.

rhornung67 commented 2 years ago

Thus requires more discussion and distillation of use cases, etc.

LLNL / axom