- 15 May, 2013 3 commits
-
-
Pavan Balaji authored
Reviewed by Mike Blocksome @ IBM.
-
Pavan Balaji authored
We now provide two types of function pointers: _sched and _req. The _sched functions just add instructions to the schedule, but no request is created. The _req functions return an MPID request. Reviewed by Mike Blocksome @ IBM.
-
The following collectives are updated: MPIX_Ibcast MPIX_Iallgather MPIX_Iallgatherv MPIX_Iallreduce MPIX_Ialltoall MPIX_Ialltoallv MPIX_Ialltoallw MPIX_Iexscan MPIX_Igather MPIX_Igatherv MPIX_Ireduce_scatter_block MPIX_Ireduce_scatter MPIX_Ireduce MPIX_Iscan MPIX_Iscatter MPIX_Iscatterv These collectives are updated similar to how the optimized MPIX_Ibarrier is implemented: | The previous MPIR_Ibarrier_impl() function forced all adi implementations | to create a MPID_Sched_t opaque object which was then passed in to the | specific ibarrier implementation via a function pointer table. | | The MPID_Sched_t object represents a completely new state machine that | must be advanced whenever mpi progress is made. | | The required construction of the MPID_Sched_t object and the required | advance of the schedule state machine would be extremely detrimental to | pamid performance. (ibm) 47b497e15cd87df666c6031a28f5f8a5ae53cece (ibm) aee872eb91f4300066957349e5a3d8103da33f40 Signed-off-by:
Charles Archer <archerc@us.ibm.com>
-
- 25 Oct, 2012 1 commit
-
-
Pavan Balaji authored
moved some error messages which should have been mpix to mpi. Reviewed by dinan.
-
- 19 Oct, 2012 1 commit
-
-
Pavan Balaji authored
MPI_. A slightly older version of this patch was reviewed by goodell.
-
- 20 Sep, 2012 1 commit
-
-
Pavan Balaji authored
In several places, after checking for a parameter (e.g., comm) we were directly using it assuming that the parameter is valid. Since the previous ERRTEST macros did not jump to fn_fail on an error, this could result in undefined behavior if the parameter was invalid. Now, since we jump on errors within the macros themselves, once the check is done, we know that the parameter values are valid. Reviewed by buntinas.
-
- 23 Aug, 2012 1 commit
-
-
James Dinan authored
Pulled bit manipulation arithmetic into macros for converting handles to indices in MPIR_Op and MPIR_Op_check_dtype function pointer tables. Reviewer: buntinas
-
- 25 Jul, 2012 1 commit
-
-
Pavan Balaji authored
since we don't need to maintain backward compatibility here. Hardcoded them to const, so they are better tested. Reviewed by goodell.
-
- 03 Jul, 2012 1 commit
-
-
Pavan Balaji authored
disabled for the public interfaces, but all internal functions use const. No reviewer.
-
- 19 Mar, 2012 1 commit
-
-
David Goodell authored
This avoids segfaults in some cases. Found by the clang static analyzer. No reviewer.
-
- 31 Mar, 2011 1 commit
-
-
David Goodell authored
This is a direct port of the recursive-doubling algorithm used for MPI_Exscan. Reviewed by buntinas@.
-
- 27 Jan, 2011 1 commit
-
-
David Goodell authored
This also allocates tags values per-comm, since NBC post order is only required to be consistent on a per-communicator basis. Reviewed by buntinas@.
-
- 20 Jan, 2011 3 commits
-
-
David Goodell authored
This change consists of four main parts: 1. Updating all of the MPIR_Ifoo_impl routines to assume that all nonblocking collective functions are set in the coll_fns field. 2. Defining the MPID_Sched_ portion of the ADI in mpir_nbc.h. 3. Implementing that interface in src/mpid/common/sched 4. Modifications to ch3:nemesis and ch3:sock in order to call the progress engine hook for the scheduling mechanism, as well as correctly including the various scheduling headers in CH3 headers. Reviewed by balaji@.
-
David Goodell authored
Reviewed by balaji@.
-
David Goodell authored
Reviewed by balaji@.
-
- 22 Oct, 2010 1 commit
-
-
David Goodell authored
These functions build but will yield an error at runtime if invoked. They will correctly call any collective function overrides specified in comm_ptr->coll_fns, permitting devices to experiment with NBC implementation before a complete implementation is available above the ADI. Reviewed by buntinas@.
-