Don't refcount the predefined datatypes. #2393
Conversation
bosilca
force-pushed
the
topic/no_predefined_ddt_refcount
branch
from
6fdb361 to
805eaec
Compare
ompi/datatype/ompi_datatype.h
Outdated
| #define OMPI_DATATYPE_RELEASE(ddt) \ | ||
| { \ | ||
| if( !ompi_datatype_is_predefined((ddt)) ) { \ | ||
| OBJ_RELEASE((ddt)); \ |
There was a problem hiding this comment.
@bosilca you need to OPAL_OUTPUT_VERBOSE __before``` OBJ_RELEASE
also, is this the right verbosity level here ?
There was a problem hiding this comment.
Thanks for catching this up. I increased the debug level to 100.
bosilca
force-pushed
the
topic/no_predefined_ddt_refcount
branch
from
805eaec to
6981757
Compare
| #define OMPI_DATATYPE_RELEASE(ddt) \ | ||
| { \ | ||
| if( !ompi_datatype_is_predefined((ddt)) ) { \ | ||
| OPAL_OUTPUT_VERBOSE((0, 100, "Datatype %p [%s] refcount %d in file %s:%d\n", \ |
There was a problem hiding this comment.
@bosilca fwiw, you might or might not want to print
(ddt)->super.super.obj_reference_count - 1 instead
i do not really have a strong opinion on that though
There was a problem hiding this comment.
This message reports non-thread safe data, with the exception of developers nobody should ever use it. Let's keep it as is.
|
bot:retest |
Signed-off-by: George Bosilca <[email protected]>
bosilca
force-pushed
the
topic/no_predefined_ddt_refcount
branch
from
6981757 to
c2cd717
Compare
jsquyres
dismissed
ggouaillardet’s stale review
Gilles indicated that he didn't have a strong opinion about the change he requested.
|
@ggouaillardet Can you re-review this PR? |
|
@ggouaillardet ping |
5 participants
Skip the refcount update via an atomic operation for all predefined datatypes. This saves 2 atomics on the critical path for all communications, as long as the datatype is predefined.