Ilmir Usmanov wrote:
> On 19.03.2014 23:35, Thomas Schwinge wrote:
> > Do you intend to support loop constructs that are not nested in a
> > parallel or kernels construct? As I'm reading it, the specification is
> > not clear on this. (I guess I'll raise this question with the OpenACC
> > guys.)
>
> Yes, I do. There are three reasons to support OpenACC loop directive
> without enclosing construct:
> 1) The spec doesn't require the directive to be inside of a construct.
> 2) PGI compiler does support this.
> 3) GCC OpenMP implementation supports loop construct without parallel.
>
> However, I agree that there no sense to use OpenACC loop without a
> construct. So, should there be a warning, perhaps?
Could it happen that such a loop becomes active due to inlining, such
as:
f() __attribute__((inline)) {
#pragma acc loop
for (...) {
...
}
...
#pragma acc parallel
{
f()
}
Or is that not possible? With OpenMP I surely used a similar code, but
there one does no have to generate special accelerator code. I know that
the Cray compiler does even LTO inlining with OpenACC, which permits to
call functions in a kernels/parallel section, which are defined in a
different file. [That is without using the OpenACC 2.0 features of
specially tagging a function as one which also runs on accelerators.]
However, I do not know whether Cray or other OpenACC compilers will
honor the 'loop' in this case. I would assume that they do, but I
haven't tested it.
Tobias
