Task/rhornung67/device numeric limits

[rhornung67]

Summary

• This PR adds a type alias layer to be able to access C++ std::numeric_limits functionality in host and device code (CUDA and HIP) in a syntactically identical way.
• This will enable us to replace our usage of std::numeric_limits with axom::numeric_limits for consistency across the code and to prevent unseemly compiler warnings.

Resolves #1195

Also, see #1338

[rhornung67]

With the changes in this PR, is there any reason we need this stuff: https://github.com/LLNL/axom/blob/develop/src/axom/core/numerics/floating_point_limits.hpp ? We are currently not consistent across the code.

[publixsubfan]

Would it make sense to remove if defined(AXOM_DEVICE_CODE) and just have the AXOM_USE_CUDA guard? It's my impression that cuda::std should work on both the host and device.

[rhornung67]

If it does, that would be preferable. I need to look into that. @kennyweiss discussed this PR yesterday. That resulted in some concerns about several things in the code that we should discuss as a team. Unfortunately, that will have to wait for a couple of weeks as next Monday is a LLNL holiday and NECDC is the week after that.

[rhornung67]

@publixsubfan cuda::std does work in device and host code. I ran into some issues with some Axom tests where cuda::std::numeric_limits does not support long double. The intent of my change was to use std::numeric_limits in host code for all builds. However, it's not clear to me that we need to support long double. long double is automatically converted to double in device code and attempting to pass long double data between host and device code is problematic.

[gunney1]

Thanks @rhornung67 This looks useful. I hope the long double issue is not a blocker.

[rhornung67]

Folks, could I get another pair of eyeballs (at least) to review this PR?

[BradWhitlock]

Is the point of using CUDA-specific numeric limits to get host device on them? Or, does that not matter since they are probably all constexpr? Does HIP have a header like this? Thanks.

[rhornung67]

The main point of this is to prevent compiler warnings (calling host function from host-device function) that a few users have reported. It is not needed for HIP since the amdclang compiler is a unified host-device compiler and know how to sort everything out.