release-1.8: Revert "codegen: explicitly handle Float16 intrinsics (#45249)"

[KristofferC]

This reverts commit eb82f18.

This PR had some issues in that it e.g. broke PowerPC (#45556 (comment)) and made some performance regressions on M1 (#45542). Since the fix is not fixing a regression (the same is true for 1.7.4), it seems better to me to let the status quo be until it can be properly fixed without the adverse effects, possibly in 1.8.1.

[KristofferC]

This was discussed during triage and it was concluded that the original PR should be conditioned so that it is not applied on hardware where native F16 exists. However, it is not a release blocker since it only applies to tier 2 and tier 1 systems. So whenever someone implements the check it can be put into 1.8.x (even 1.8.0 if it happens quickly).

[ctkelley]

I'm a bit confused. It's the bottom line that the poor F16 performance will persist and that I should stick with 1.7.2 I'm the one who raised 45542

[vchuravy]

This also destroys performance for Float16 on x86? X86 has native support for converting Float16 to Float32 and vice versa, which we need to use to have a semblance of performance.

@maleadt did you manage to check if CUDA jl Float16 support still works?

[maleadt]

Yeah this kills performance on x86 too. CUDA isn't affected as we don't run the float16demote pass.

[KristofferC]

Okay can someone help to fix this then? I doubt @vtjnash is interested in prioritising this since he doesn't consider it blocking so someone need to roll up their sleeves.

[vchuravy]

x-ref #45249 and #45644

[vchuravy]

Okay can someone help to fix this then?

I proposed an alternative in #45649, but I will need to do the prove to show that it suffices.

[KristofferC]

Since this is not reverted on master, it has to be reverted on the backport branch as well.

[ViralBShah]

Will this be applied to master too?

[vchuravy]

See #45644 and #45736 (comment)