smcuda: fixes when using enable-mca-dso #11443
Conversation
hppritcha
force-pushed
the
fix_for_issue_11354
branch
from
7b6c570
to
6ad92c2
Compare
|
rebasing on top of main to pull in pmix_fence usage patch |
| @@ -79,6 +83,8 @@ int mca_btl_smcuda_accelerator_init(void) | |||
| goto cleanup_and_error; | |||
| } | |||
|
|
|||
| opal_finalize_register_cleanup(mca_btl_smcuda_accelerator_fini); | |||
There was a problem hiding this comment.
I believe you when you say this will fix the issue, but do you mind explaining why it needs to happen this way since I don't quite understand. The pml ob1 component does something similar and I'm wondering if it requires the same fix
There was a problem hiding this comment.
@wckzhang does this comment block help answer your questions?
There was a problem hiding this comment.
Yes, I think we need to do the same thing with ob1 then
There was a problem hiding this comment.
isn't in theory this the case for all components that use the accelerator framework functionality?
There was a problem hiding this comment.
@wckzhang where would you want to make a change in the ob1 to use the opal cleanup callback mechanism?
There was a problem hiding this comment.
@edgargabriel if you find your consumer of the accelerator framework is doing something in its component close method that involves calls into the accelerator framework then you are probably right.
There was a problem hiding this comment.
I take it back about ob1 because it already does it in a fini function instead of component close
There was a problem hiding this comment.
@edgargabriel Yes, though the other components aren't doing any accelerator functionality during finalization. The reason ob1 and smcuda are susceptible to this is because they create event arrays for the purpose of their asynchronous progress engine which need to be cleaned up during finalization.
There was a problem hiding this comment.
what about the rcache/grdma component? I think that one does something similar, isn't it?
There was a problem hiding this comment.
No as far as I can tell it doesn't access the accelerator framework in its closing sequence, just during its register path.
|
np i'll put some verbiage in around this call. |
related to open-mpi#11354 Signed-off-by: Howard Pritchard <howardp@lanl.gov>
hppritcha
force-pushed
the
fix_for_issue_11354
branch
from
6ad92c2
to
f7803dd
Compare
related to #11354