Support expt-relaxed-constexpr same as expt-extended-lambda

[PhilipFackler]

The Kokkos CMake (and the makefile generator) support the --expt-extended-lambda option through the Kokkos_ENABLE_CUDA_LAMBDA option (or the "enable_lambda" cuda option). This gets propagated automatically to my project's compiler options.

However, I have to manually add --expt-relaxed-constexpr to my compiler options. Can you add an option to support this flag in the same way?

[dhollman]

I would further argue that it should be enabled by default.

[crtrott]

I don't want to enable experimental features by default. We have enough failures in-field from non-experimental capabilities ...

[dhollman]

But it's not actually an experimental feature...

[dhollman]

As evidence, I offer the last three tweets of this twitter conversation: https://twitter.com/__simt__/status/1164743863279341568?s=20

[DavidPoliakoff]

@dhollman (memeing) yeah that would be the first time somebody said something on Twitter that didn't come to pass, LGTM

[dhollman]

@DavidPoliakoff sure, but the relevant part is that that's someone pretty high up at NVIDIA.

[DavidPoliakoff]

@dhollman : to provide some actual content, it would be great if it could be enabled by default but with a warning. Generally I like your idea of letting people opt out of things accidentally working (see also: non __device__ marked functions getting inlined and thus not being a problem)

[crtrott]

Look it is nice that Olivier thinks these features are kinda real stuff. But considering how much issues we find with features which are actually technically supported, I don't care. If Olivier can convince the compiler team to take the "experimental" off and NVIDIA promises us that if we find issues with it they fix them then fine. Until then it is NOT going to be the default.

[dalg24]

I don't want to enable experimental features by default. We have enough failures in-field from non-experimental capabilities ...

Then you might want to comment on #2412

[ibaned]

Team discussion today, despite strong support from several team members to enable these experimental options by default, there was enough dissent among other members that we decided to still not enable them by default. The argument was that NVIDIA does not give these features good enough first-class support that we feel confident in having customers silently depend on them by default.

[ibaned]

To clarify, this is not quite the original question in the issue above. I don't think we voted on giving constexpr equal footing to lambda.

[mhoemmen]

@crtrott You realize the whole of Sierra turns this lambda option on and uses it right

[mhoemmen]

If it's experimental it would be good to know sooner rather than later whether we should use it

[crtrott]

Mark: I always told you guys that it is experimental in NVIDIA, but seems to work well enough if you know what you are doing. I remember very distinctly voicing my unease with the decision to require lambda support in Tpetra. More on this in an email ...