(I tried to collect the cross-posts from the cfe and openmp list here,
unsure if I got the latest ones though. No more cross posts and/or
dropping a list please.)
Ravi's email to cfe-dev
Everytime we want to pass something new to the libomptarget runtime we
would need to add a new interface. Why not pass a pointer to a
structure, 1st field says how many valid fields this structure
contains and populate the fields with info or null. Run time can
check if the field is null or not and take appropriate action.
Struct {
Int Num_fields 2
Int * mapper;
Int * nowait_object
}
To support async (aka nowait) we need an object to call back to the
libomp library. This object can be passed in this struct.
Initially, I thought this was a similarly good design. While fairly
future prove, I have some problems I want to mention. First, however,
I'd like to point out that new API functions alone are not really
problematic though, especially if they are just wrappers around the
newest version.
Drawbacks of a "version struct":
- It makes the API backwards compatible, at least until runtime. I
mean, when you recompile a component of your application with a
(maybe newer) compiler it is unclear what version the rest of the
application, and the used structs, are using. If you assume the
struct passed around is at least of the current version, it will
break if it is not, silently at runtime. If you don't assume
anything, you need to emit dynamic checks and versioning when you
want to modify the code. The same situation with the "new API" model
will cause a link error if the runtime you link into is not as new
as the API calls you used. This allows you to rewrite the code based
on the API call semantic you know without versioning and fear of
silent miscompilation. I know this explanation is very abstract, let
me know if I should try to make it more concrete.
- We cannot remove arguments in a reasonable way. Please correct me if
I'm wrong here.
>
>
> I would like to bring your attention to the choice of 2 proposals for the
> declare mapper runtime interface:
>
> 1. The current design which creates new runtime functions for declare
> mappers. For example, right now we have `__tgt_target_teams(...)` which
> corresponds to the runtime interface for `omp target teams`. Now we add
> `__tgt_target_teams_mapper(..., void **mappers)` to replace it.
>
New feature, new library call seems reasonable. The current interface can
presumably be reimplemented as a call to (the internals of) the proposed
interface.
Yes, I mentioned that as well:
old(args) = new(args, nullptr)
> 2. Introduce a function `__tgt_push_mappers`, which should be called before
> every target function call (e.g., `__tgt_target_teams`) to pass the mapper
> argument for that function. The call of `__tgt_push_mappers` is implicitly
> bonded with the actual target call.
>
This seems error prone to use and likely to degrade performance.
Choice 1 seems more appealing to me.
So far, most people seem to prefer choice 1. We should go with that one
if there are no complaints soon.
Cheers,
Johannes