osc/base: fix accumulate on derived datatypes #389
Conversation
With certain datatypes the opal_datatype_unpack method for performing the accumulate operation does not work. This commit modifies the accumulate code in the osc base to use opal_convertor_raw instead. Fixes open-mpi#385
|
Refer to this link for build results (access rights to CI server needed): |
|
Nathan, I think I spotted the issue here. If the op is called while using the optimized version of the datatype (the one ofter commit), instead of getting the original count you might get the number of bytes (because in a homogeneous environment the optimized datatype description merged everything as bytes in order to create the largest possible contiguous regions). |
|
@bosilca The code in question set the convertor to use the non-optimized description. Does something else need to be set? The new code works for all datatypes and isn't quite as complex as the unpack code. I have been using something similar to deal with datatypes in the new rdma osc component. |
|
Passing MTT. Merging. |
osc/base: fix accumulate on derived datatypes
More one sided bindings
3 participants
With certain datatypes the opal_datatype_unpack method for performing
the accumulate operation does not work. This commit modifies the
accumulate code in the osc base to use opal_convertor_raw instead.
Fixes #385