[flang][OpenMP] Support host_eval for target teams loop

[ergawy]

Extends host_eval support for the currently supported form of the generic loop directive.

[skatrak]

Thank you Kareem for working on this! I've got a couple of small comments, but I think it's almost there.

[ergawy]

Thanks for the review @skatrak. I handled your comments (hopefully) and added a small offloading test. But I wanted to ask: where in the spec do we extract the info that a certain config is generic, SPMD, or generic-SPMD?

[skatrak]

Thank you Kareem, and sorry for the delay getting back to this! LGTM, just minor nits.

[skatrak]

Suggested change

	// Detect SPMD: target-teams-loop[-simd].
	// Detect SPMD: target-teams-loop.

[skatrak]

Suggested change

	if (!isa<DistributeOp>(innermostWrapper) && !isa<LoopOp>(innermostWrapper))
	if (!isa<DistributeOp, LoopOp>(innermostWrapper))

[skatrak]

Thanks for the review @skatrak. I handled your comments (hopefully) and added a small offloading test. But I wanted to ask: where in the spec do we extract the info that a certain config is generic, SPMD, or generic-SPMD?

That differentiation is not really in the spec, but it's instead part of the device runtime. It's an implementation detail that allows us to provide more efficient specializations for specific kinds of kernels. SPMD, for example, is used to identify kernels for which the number of teams and threads and the trip count of the single loop being run across all threads and teams remains constant during execution and is known in advance.

A generic kernel might have multiple loops, none at all, or violate any of the conditions above (which are not mandated by the spec), and for that case the generated code implements a state machine that reduces performance on GPUs. Generic-SPMD seems to fall somewhere in the middle, and I don't know much about it apart from the fact that it seems to represent target teams distribute, so something inside of a loop that is distributed across teams. I got that understanding some time ago looking at uses of the OMPTgtExecModeFlags, but I wouldn't be surprised if I didn't understand Generic-SPMD fully.

[ergawy]

That differentiation is not really in the spec, but it's instead part of the device runtime. It's an implementation detail that allows us to provide more efficient specializations for specific kinds of kernels. SPMD, for example, is used to identify kernels for which the number of teams and threads and the trip count of the single loop being run across all threads and teams remains constant during execution and is known in advance.

A generic kernel might have multiple loops, none at all, or violate any of the conditions above (which are not mandated by the spec), and for that case the generated code implements a state machine that reduces performance on GPUs. Generic-SPMD seems to fall somewhere in the middle, and I don't know much about it apart from the fact that it seems to represent target teams distribute, so something inside of a loop that is distributed across teams. I got that understanding some time ago looking at uses of the OMPTgtExecModeFlags, but I wouldn't be surprised if I didn't understand Generic-SPMD fully.

That clarifies things quite a bit. Thanks!