https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88370
Bug ID: 88370
Summary: acc_get_cuda_stream/acc_set_cuda_stream:
acc_async_sync, acc_async_noval
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Keywords: openacc, patch
Severity: normal
Priority: P3
Component: libgomp
Assignee: tschwinge at gcc dot gnu.org
Reporter: tschwinge at gcc dot gnu.org
CC: cltang at gcc dot gnu.org, jakub at gcc dot gnu.org
Target Milestone: ---
Per my reading of the OpenACC specification (and I submitted to them a patch to
clarify that, <https://github.com/OpenACC/openacc-spec/pull/165>), it's valid
to call "acc_get_cuda_stream"/"acc_set_cuda_stream" also with "acc_async_sync",
"acc_async_noval" arguments, not just with the nonnegative values as currently
implemented.
This is also supported by secondary documentation, such as code examples, or
presentations, which especially makes use of
"acc_get_cuda_stream(acc_async_sync)". For example:
<https://github.com/fomics/EuroHack15/wiki/CUDA-Streams-and-OpenACC-async-interoperability>,
or the PGI "OpenACC Getting Started Guide" (Version 2017) stating that "the
input argument is an async number or a pre-defined value such as
acc_async_sync".
This will need to be fixed on all release branches.
