Hi Thomas,
Thanks for your response.
> "Changes init, shutdown, and set directives to accept just a single
> device_type argument rather than a list".) So, no need for you to
> invent a libgomp function that is accepting a list of device types.
Oh, actually my current implementation emits the corresponding libgomp
builtin call as many devices are inside the list, then I have a
(rookie?) question:
- Is it better to implement the builtin with variable number of
parameters rather than emiting separate builtin calls? Why?
> In Fortran ('gcc/fortran/openmp.cc:gfc_match_omp_clauses'), you just
> re-use 'OMP_CLAUSE_DEVICE_TYPE' also for OpenACC, and disambiguate
> the two cases with 'if (openacc)' (as done, for example, for
> 'OMP_CLAUSE_DEVICE', which also is shared for the OpenACC and OpenMP
> 'device' clauses, with different semantics).
Thanks! I saw the specific code to handle the OMP_CLAUSE_DEVICE_TYPE
and I realized that I was overthinking it. Oops.
> Therefore, in clause parsing, do a simple '!strcmp ("nvidia",
> [...])' etc. (see, for example, the OpenACC/OpenMP 'default' clause
> parsing), and then, for your 'GOACC_[...]' function calls, you may
> use 'GOMP_DEVICE_NVIDIA_PTX' etc. (see 'include/gomp-constants.h',
> which defines the "Communication between GCC and libgomp").
I'll change the implementation following this.
Thanks for your clarifications and your time!
Best,
Sebastian
