#include "petscvec.h" PetscErrorCode VecScatterCreate(Vec xin,IS ix,Vec yin,IS iy,VecScatter *newctx)Collective on Vec
xin | - a vector that defines the shape (parallel data layout of the vector) of vectors from which we scatter | |
yin | - a vector that defines the shape (parallel data layout of the vector) of vectors to which we scatter | |
ix | - the indices of xin to scatter (if NULL scatters all values) | |
iy | - the indices of yin to hold results (if NULL fills entire vector yin) |
Options Database Keys: (uses regular MPI_Sends by default)
-vecscatter_view | - Prints detail of communications | |
-vecscatter_view ::ascii_info | - Print less details about communication | |
-vecscatter_ssend | - Uses MPI_Ssend_init() instead of MPI_Send_init() | |
-vecscatter_rsend | - use ready receiver mode for MPI sends | |
-vecscatter_merge | - VecScatterBegin() handles all of the communication, VecScatterEnd() is a nop eliminates the chance for overlap of computation and communication | |
-vecscatter_sendfirst | - Posts sends before receives | |
-vecscatter_packtogether | - Pack all messages before sending, receive all messages before unpacking | |
-vecscatter_alltoall | - Uses MPI all to all communication for scatter | |
-vecscatter_window | - Use MPI 2 window operations to move data | |
-vecscatter_nopack | - Avoid packing to work vector when possible (if used with -vecscatter_alltoall then will use MPI_Alltoallw() | |
-vecscatter_reproduce | - insure that the order of the communications are done the same for each scatter, this under certain circumstances will make the results of scatters deterministic when otherwise they are not (it may be slower also). |
--When packing is used--
MPI Datatypes (no packing) sendfirst merge packtogether persistent*
_nopack _sendfirst _merge _packtogether -vecscatter_
----------------------------------------------------------------------------------------------------------------------------
Message passing Send p X X X always
Ssend p X X X always _ssend
Rsend p nonsense X X always _rsend
AlltoAll v or w X nonsense always X nonsense _alltoall
MPI_Win p nonsense p p nonsense _window
Since persistent sends and receives require a constant memory address they can only be used when data is packed into the work vector
because the in and out array may be different for each call to VecScatterBegin/End().
p indicates possible, but not implemented. X indicates implemented
Currently the MPI_Send(), MPI_Ssend() and MPI_Rsend() all use PERSISTENT versions. (this unfortunately requires that the same in and out arrays be used for each use, this is why when not using MPI_alltoallw() we always need to pack the input into the work array before sending and unpack upon receeving instead of using MPI datatypes to avoid the packing/unpacking).
Both ix and iy cannot be NULL at the same time.
Level:intermediate
Location:src/vec/vec/utils/vscat.c
Index of all Vec routines
Table of Contents for all manual pages
Index of all manual pages