On Mon, Mar 23, 2015 at 10:44:39PM +0300, Ilya Verbin wrote:
> If it is too late for such global changes (rework initialization in libgomp,
> change mic and ptx plugins), then here is a small workaround patch to fix
> offloading from libraries. Likely, it will not affect OpenACC programs with
> one
> image. make check-target-libgomp passed.
Sorry for not getting to this earlier, really busy with severe regressions
bugfixing lately.
Anyway, IMHO it is not too late to fixing it properly, after all,
the current code is majorly broken. As I've said earlier, e.g. the lack
of mutex guarding gomp_target_init (which is using pthread_once guaranteed
to be run just once) vs. concurrent GOMP_offload_register calls
(if those are run from ctors, then I guess something like dl_load_lock
ensures at least on glibc that multiple GOMP_offload_register calls aren't
performed at the same time) in accessing/reallocating offload_images
and num_offload_images and the lack of support to register further
images after the gomp_target_init call (if you dlopen further shared
libraries) is really bad. And it would be really nice to support the
unloading.
But I'm afraid I'm lost in what is the latest posted patch for that,
and how has it been tested (whether just on MIC or MIC emul, or also for
nvptx).
So can you please post a link to the latest full patch and how it has been
tested, and if it is still error prone if say one thread executes
GOMP_target the first time and another at the same time dlopens some shared
library that has offloading regions in it, fix that too?
We still have a week or so to get this sorted out.
Jakub
