Hi,
I'm sorry it took me more than a month to come up with another round
of patches aiming at merging the HSA branch into the trunk. Keeping
up-to date with the latest changes in the OpenMP 4.5 area was
strenuous and we have discovered and fixed a few bugs as I intensified
my testing efforts.
While those are the main areas where this patch set differs from the
previous one, I have of course addressed the feedback I got the last
time, including implementing device-specific OpenMP target arguments,
moving kernel grid size from gimple class fields to new artificial
clauses and disabling the vectorizer for HSA functions using
DECL_FUNCTION_SPECIFIC_OPTIMIZATION rather than extra code in
respective pass gates.
Because I have not been able to come up with any solution to failing
libgomp/testsuite/libgomp.c++/target-2.C, I have disabled use of
dynamic parallelism in this merge (I keep it on the branch) and
therefore entirely rely on the gridification process to run loops on
the accelerator, because gridified constructs do not have this issue
(passing private symbols by reference).
HSA tests are still missing, I would need some guidance as to how to
best implement them (specially to test gridification which of course
does not happen for other accelerators). There are no failing
testcases if HSA is not configured. If it is, there are some, all of
which fall into one the following categories:
1) HSA cannot compile a function for one reason or another (most
common cause is inability of HSA to take an address of a function
or make an indirect call) and gives a warning, which is regarded
as an "excess error" by dejagnu.
2) When HSA is not emitted for a function, libgomp runs a host
fallback instead of it. When the test queries
omp_is_initial_device and asserts it returns false, the test
fails.
3) There are still a few failing OpenACC tests, but those just
should not be run.
Of course, the patch set bootstraps fine on x86_64-linux with or
without configured HSA.
Any feedback is welcome. Thanks,
Martin
