Assuming it works for other platforms, I think it would be neater to just move the constants to program scope and avoid the conditional includes. I'm open to suggestions!

Would MSL support these defined globally but as inline constexpr instead of static? That would also remove the duplication that occurs if these headers are used in multiple TUs.

In D13241#349593, @Jesse Yurkovich (deadpin) wrote:

Would MSL support these defined globally but as inline constexpr instead of static? That would also remove the duplication that occurs if these headers are used in multiple TUs.

Specifying inline is a no-op in MSL and gives a warning ("inline variables are a C++17 extension"), however MSL does require that any program scope variable must reside in constant address space.

We should probably hide the qualifiers behind a macro which can be defined on a per-platform basis.

In D13241#349529, @Brecht Van Lommel (brecht) wrote:

These were actually at program scope until recently, so should be fine. rB76de3ac4ce43: Cleanup: Remove data duplication from various lookup tables in Cycles.

The previous constants' locations (as in rB76de3ac4ce43: Cleanup: Remove data duplication from various lookup tables in Cycles) would in fact be at class scope on Metal due to the context class. The simplest solution to this would be to include the constants in a monolithic block before including Metal's "context_begin.h", at the loss of some locality of reference (although I'm happy to consider other approaches if they seem better).

I think it would be fine to move all the tables into a single file that we can include ahead of context_begin.h.

Looks fine, but I think we could try the suggestion from @Jesse Yurkovich (deadpin)?

Maybe rename ccl_static_constant to ccl_inline_constant, and then define that as inline constexpr for the CPU device, and leave it unchanged for GPU devices?

I'm not sure if it's better to use /* clang-format off */ or just add a line break after #include "kernel/tables.h".
(Just an observation).