1
1
openmpi/ompi/op/op.h

545 строки
18 KiB
C
Исходник Обычный вид История

/* -*- Mode: C; c-basic-offset:4 ; -*- */
/*
* Copyright (c) 2004-2006 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2007 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2007 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2008 UT-Battelle, LLC
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
* Copyright (c) 2008-2009 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*
* Public interface for the MPI_Op handle.
*/
#ifndef OMPI_OP_H
#define OMPI_OP_H
#include "ompi_config.h"
#include <stdio.h>
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
#include "mpi.h"
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
#include "opal/class/opal_object.h"
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
#include "ompi/datatype/datatype.h"
#include "ompi/mpi/f77/fint_2_int.h"
#include "ompi/mca/op/op.h"
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
BEGIN_C_DECLS
/**
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
* Typedef for C op functions for user-defined MPI_Ops.
*
* We don't use MPI_User_function because this would create a
* confusing dependency loop between this file and mpi.h. So this is
* repeated code, but it's better this way (and this typedef will
* never change, so there's not much of a maintenance worry).
*/
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
typedef void (ompi_op_c_handler_fn_t)(void *, void *, int *,
struct ompi_datatype_t **);
/**
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
* Typedef for fortran user-defined MPI_Ops.
*/
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
typedef void (ompi_op_fortran_handler_fn_t)(void *, void *,
MPI_Fint *, MPI_Fint *);
/**
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
* Typedef for C++ op functions intercept (used for user-defined
* MPI::Ops).
*
* See the lengthy explanation for why this is different than the C
* intercept in ompi/mpi/cxx/intercepts.cc in the
* ompi_mpi_cxx_op_intercept() function.
*/
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
typedef void (ompi_op_cxx_handler_fn_t)(void *, void *, int *,
struct ompi_datatype_t **,
MPI_User_function * op);
/*
* Flags for MPI_Op
*/
/** Set if the MPI_Op is a built-in operation */
#define OMPI_OP_FLAGS_INTRINSIC 0x0001
/** Set if the callback function is in Fortran */
#define OMPI_OP_FLAGS_FORTRAN_FUNC 0x0002
/** Set if the callback function is in C++ */
#define OMPI_OP_FLAGS_CXX_FUNC 0x0004
/** Set if the callback function is associative (MAX and SUM will both
have ASSOC set -- in fact, it will only *not* be set if we
implement some extensions to MPI, because MPI says that all
MPI_Op's should be associative, so this flag is really here for
future expansion) */
#define OMPI_OP_FLAGS_ASSOC 0x0008
/** Set if the callback function is associative for floating point
operands (e.g., MPI_SUM will have ASSOC set, but will *not* have
FLOAT_ASSOC set) */
#define OMPI_OP_FLAGS_FLOAT_ASSOC 0x0010
/** Set if the callback function is communative */
#define OMPI_OP_FLAGS_COMMUTE 0x0020
/**
* Back-end type of MPI_Op
*/
struct ompi_op_t {
/** Parent class, for reference counting */
opal_object_t super;
/** Name, for debugging purposes */
char o_name[MPI_MAX_OBJECT_NAME];
/** Flags about the op */
uint32_t o_flags;
/** Index in Fortran <-> C translation array */
int o_f_to_c_index;
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
/** Union holding (2-buffer functions):
1. Function pointers for all supported datatypes when this op
is intrinsic
2. Function pointers for when this op is user-defined (only
need one function pointer for this; we call it for *all*
datatypes, even intrinsics)
*/
union {
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
/** Function/module pointers for intrinsic ops */
ompi_op_base_op_fns_t intrinsic;
/** C handler function pointer */
ompi_op_c_handler_fn_t *c_fn;
/** Fortran handler function pointer */
ompi_op_fortran_handler_fn_t *fort_fn;
/** C++ intercept function data -- see lengthy comment in
ompi/mpi/cxx/intercepts.cc::ompi_mpi_cxx_op_intercept() for
an explanation */
struct {
/* The user's function (it's the wrong type, but that's ok) */
ompi_op_c_handler_fn_t *user_fn;
/* The OMPI C++ callback/intercept function */
ompi_op_cxx_handler_fn_t *intercept_fn;
} cxx_data;
} o_func;
/** 3-buffer functions, which is only for intrinsic ops. No need
for the C/C++/Fortran user-defined functions. */
ompi_op_base_op_3buff_fns_t o_3buff_intrinsic;
};
/**
* Convenience typedef
*/
typedef struct ompi_op_t ompi_op_t;
OMPI_DECLSPEC OBJ_CLASS_DECLARATION(ompi_op_t);
/**
* Array to map ddt->id values to the corresponding position in the op
* function array.
*
* NOTE: It is possible to have an implementation without this map.
* There are basically 3 choices for implementing "how to find the
* right position in the op array based on the datatype":
*
* 1. Use the exact same ordering as ddt->id in the op map. This is
* nice in that it's always a direct lookup via one memory
* de-reference. But it makes a sparse op array, and it's at least
* somewhat wasteful. It also chains the ddt and op implementations
* together. If the ddt ever changes its ordering, op is screwed. It
* seemed safer from a maintenance point of view not to do it that
* way.
*
* 2. Re-arrange the ddt ID values so that all the reducable types are
* at the beginning. This means that we can have a dense array here
* in op, but then we have the same problem as number one -- and so
* this didn't seem like a good idea from a maintenance point of view.
*
* 3. Create a mapping between the ddt->id values and the position in
* the op array. This allows a nice dense op array, and if we make
* the map based on symbolic values, then if ddt ever changes its
* ordering, it won't matter to op. This seemed like the safest thing
* to do from a maintenance perspective, and since it only costs one
* extra lookup, and that lookup is way cheaper than the function call
* to invoke the reduction operation, it seemed like the best idea.
*/
OMPI_DECLSPEC extern int ompi_op_ddt_map[DT_MAX_PREDEFINED];
/**
* Global variable for MPI_OP_NULL
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_null;
/**
* Global variable for MPI_MAX
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_max;
/**
* Global variable for MPI_MIN
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_min;
/**
* Global variable for MPI_SUM
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_sum;
/**
* Global variable for MPI_PROD
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_prod;
/**
* Global variable for MPI_LAND
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_land;
/**
* Global variable for MPI_BAND
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_band;
/**
* Global variable for MPI_LOR
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_lor;
/**
* Global variable for MPI_BOR
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_bor;
/**
* Global variable for MPI_LXOR
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_lxor;
/**
* Global variable for MPI_BXOR
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_bxor;
/**
* Global variable for MPI_MAXLOC
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_maxloc;
/**
* Global variable for MPI_MINLOC
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_minloc;
/**
* Global variable for MPI_REPLACE
*/
OMPI_DECLSPEC extern ompi_op_t ompi_mpi_op_replace;
/**
* Table for Fortran <-> C op handle conversion
*/
extern struct opal_pointer_array_t *ompi_op_f_to_c_table;
/**
* Initialize the op interface.
*
* @returns OMPI_SUCCESS Upon success
* @returns OMPI_ERROR Otherwise
*
* Invoked from ompi_mpi_init(); sets up the op interface, creates
* the predefined MPI operations, and creates the corresopnding F2C
* translation table.
*/
int ompi_op_init(void);
/**
* Finalize the op interface.
*
* @returns OMPI_SUCCESS Always
*
* Invokes from ompi_mpi_finalize(); tears down the op interface, and
* destroys the F2C translation table.
*/
int ompi_op_finalize(void);
/**
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
* Create a ompi_op_t with a user-defined callback (vs. creating an
* intrinsic ompi_op_t).
*
* @param commute Boolean indicating whether the operation is
* communative or not
* @param func Function pointer of the error handler
*
* @returns op Pointer to the ompi_op_t that will be
* created and returned
*
* This function is called as the back-end of all the MPI_OP_CREATE
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
* function. It creates a new ompi_op_t object, initializes it to the
* correct object type, and sets the callback function on it.
*
* The type of the function pointer is (arbitrarily) the fortran
* function handler type. Since this function has to accept 2
* different function pointer types (lest we have 2 different
* functions to create errhandlers), the fortran one was picked
* arbitrarily. Note that (void*) is not sufficient because at
* least theoretically, a sizeof(void*) may not necessarily be the
* same as sizeof(void(*)).
*
* NOTE: It *always* sets the "fortran" flag to false. The Fortran
* wrapper for MPI_OP_CREATE is expected to reset this flag to true
* manually.
*/
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
ompi_op_t *ompi_op_create_user(bool commute,
ompi_op_fortran_handler_fn_t func);
/**
* Mark an MPI_Op as holding a C++ callback function, and cache
* that function in the MPI_Op. See a lenghty comment in
* ompi/mpi/cxx/op.c::ompi_mpi_cxx_op_intercept() for a full
* expalantion.
*/
OMPI_DECLSPEC void ompi_op_set_cxx_callback(ompi_op_t * op,
MPI_User_function * fn);
/**
* Check to see if an op is intrinsic.
*
* @param op The op to check
*
* @returns true If the op is intrinsic
* @returns false If the op is not intrinsic
*
* Self-explanitory. This is needed in a few top-level MPI functions;
* this function is provided to hide the internal structure field
* names.
*/
static inline bool ompi_op_is_intrinsic(ompi_op_t * op)
{
return (bool) (0 != (op->o_flags & OMPI_OP_FLAGS_INTRINSIC));
}
/**
* Check to see if an op is communative or not
*
* @param op The op to check
*
* @returns true If the op is communative
* @returns false If the op is not communative
*
* Self-explanitory. This is needed in a few top-level MPI functions;
* this function is provided to hide the internal structure field
* names.
*/
static inline bool ompi_op_is_commute(ompi_op_t * op)
{
return (bool) (0 != (op->o_flags & OMPI_OP_FLAGS_COMMUTE));
}
/**
* Check to see if an op is floating point associative or not
*
* @param op The op to check
*
* @returns true If the op is floating point associative
* @returns false If the op is not floating point associative
*
* Self-explanitory. This is needed in a few top-level MPI functions;
* this function is provided to hide the internal structure field
* names.
*/
static inline bool ompi_op_is_float_assoc(ompi_op_t * op)
{
return (bool) (0 != (op->o_flags & OMPI_OP_FLAGS_FLOAT_ASSOC));
}
/**
* Check to see if an op is valid on a given datatype
*
* @param op The op to check
* @param ddt The datatype to check
*
* @returns true If the op is valid on that datatype
* @returns false If the op is not valid on that datatype
*
* Self-explanitory. This is needed in a few top-level MPI functions;
* this function is provided to hide the internal structure field
* names.
*/
static inline bool ompi_op_is_valid(ompi_op_t * op, ompi_datatype_t * ddt,
char **msg, const char *func)
{
/* Check:
- non-intrinsic ddt's cannot be invoked on intrinsic op's
- if intrinsic ddt invoked on intrinsic op:
- ensure the datatype is defined in the op map
- ensure we have a function pointer for that combination
*/
if (ompi_op_is_intrinsic(op)) {
if (ompi_ddt_is_predefined(ddt)) {
/* Intrinsic ddt on intrinsic op */
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
if (-1 == ompi_op_ddt_map[ddt->id] ||
NULL == op->o_func.intrinsic.fns[ompi_op_ddt_map[ddt->id]]) {
asprintf(msg,
"%s: the reduction operation %s is not defined on the %s datatype",
func, op->o_name, ddt->name);
return false;
}
} else {
/* Non-intrinsic ddt on intrinsic op */
if ('\0' != ddt->name[0]) {
asprintf(msg,
"%s: the reduction operation %s is not defined for non-intrinsic datatypes (attempted with datatype named \"%s\")",
func, op->o_name, ddt->name);
} else {
asprintf(msg,
"%s: the reduction operation %s is not defined for non-intrinsic datatypes",
func, op->o_name);
}
return false;
}
}
/* All other cases ok */
return true;
}
/**
* Perform a reduction operation.
*
* @param op The operation (IN)
* @param source Source (input) buffer (IN)
* @param target Target (output) buffer (IN/OUT)
* @param count Number of elements (IN)
* @param dtype MPI datatype (IN)
*
* @returns void As with MPI user-defined reduction functions, there
* is no return code from this function.
*
* Perform a reduction operation with count elements of type dtype in
* the buffers source and target. The target buffer obtains the
* result (i.e., the original values in the target buffer are reduced
* with the values in the source buffer and the result is stored in
* the target buffer).
*
* This function figures out which reduction operation function to
* invoke and whether to invoke it with C- or Fortran-style invocation
* methods. If the op is intrinsic and has the operation defined for
* dtype, the appropriate back-end function will be invoked.
* Otherwise, the op is assumed to be a user op and the first function
* pointer in the op array will be used.
*
* NOTE: This function assumes that a correct combination will be
* given to it; it makes no provision for errors (in the name of
* optimization). If you give it an intrinsic op with a datatype that
* is not defined to have that operation, it is likely to seg fault.
*/
static inline void ompi_op_reduce(ompi_op_t * op, void *source,
void *target, int count,
ompi_datatype_t * dtype)
{
MPI_Fint f_dtype, f_count;
/*
* Call the reduction function. Two dimensions: a) if both the op
* and the datatype are intrinsic, we have a series of predefined
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
* functions for each datatype (that are *only* in C -- not
* Fortran or C++!), or b) the op is user-defined, and therefore
* we have to check whether to invoke the callback with the C,
* C++, or Fortran callback signature (see lengthy description of
* the C++ callback in ompi/mpi/cxx/intercepts.cc).
*
* NOTE: We *assume* the following:
*
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
* 1. If the op is intrinsic, the op is pre-defined
* 2. That we will get a valid result back from the
* ompi_op_ddt_map[] (and not -1).
*
* Failures in these assumptions should have been caught by the
* upper layer (i.e., they should never have called this
* function). If either of these assumptions are wrong, it's
* likely that the MPI API function parameter checking is turned
* off, then it's an erroneous program and it's the user's fault.
* :-)
*/
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
/* For intrinsics, we also pass the corresponding op module */
if (0 != (op->o_flags & OMPI_OP_FLAGS_INTRINSIC)) {
op->o_func.intrinsic.fns[ompi_op_ddt_map[dtype->id]](source, target,
&count, &dtype,
op->o_func.intrinsic.modules[ompi_op_ddt_map[dtype->id]]);
}
/* User-defined function */
else if (0 != (op->o_flags & OMPI_OP_FLAGS_FORTRAN_FUNC)) {
f_dtype = OMPI_INT_2_FINT(dtype->d_f_to_c_index);
f_count = OMPI_INT_2_FINT(count);
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
op->o_func.fort_fn(source, target, &f_count, &f_dtype);
} else if (0 != (op->o_flags & OMPI_OP_FLAGS_CXX_FUNC)) {
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
op->o_func.cxx_data.intercept_fn(source, target, &count, &dtype,
op->o_func.cxx_data.user_fn);
} else {
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
op->o_func.c_fn(source, target, &count, &dtype);
}
}
/**
* Perform a reduction operation.
*
* @param op The operation (IN)
* @param source Source1 (input) buffer (IN)
* @param source Source2 (input) buffer (IN)
* @param target Target (output) buffer (IN/OUT)
* @param count Number of elements (IN)
* @param dtype MPI datatype (IN)
*
* @returns void As with MPI user-defined reduction functions, there
* is no return code from this function.
*
* Perform a reduction operation with count elements of type dtype in
* the buffers source and target. The target buffer obtains the
* result (i.e., the original values in the target buffer are reduced
* with the values in the source buffer and the result is stored in
* the target buffer).
*
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
* This function will *only* be invoked on intrinsic MPI_Ops.
*
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
* Otherwise, this function is the same as ompi_op_reduce.
*/
static inline void ompi_3buff_op_reduce(ompi_op_t * op, void *source1,
void *source2, void *target,
int count, ompi_datatype_t * dtype)
{
void *restrict src1;
void *restrict src2;
void *restrict tgt;
src1 = source1;
src2 = source2;
tgt = target;
Two major things in this commit: * New "op" MPI layer framework * Addition of the MPI_REDUCE_LOCAL proposed function (for MPI-2.2) = Op framework = Add new "op" framework in the ompi layer. This framework replaces the hard-coded MPI_Op back-end functions for (MPI_Op, MPI_Datatype) tuples for pre-defined MPI_Ops, allowing components and modules to provide the back-end functions. The intent is that components can be written to take advantage of hardware acceleration (GPU, FPGA, specialized CPU instructions, etc.). Similar to other frameworks, components are intended to be able to discover at run-time if they can be used, and if so, elect themselves to be selected (or disqualify themselves from selection if they cannot run). If specialized hardware is not available, there is a default set of functions that will automatically be used. This framework is ''not'' used for user-defined MPI_Ops. The new op framework is similar to the existing coll framework, in that the final set of function pointers that are used on any given intrinsic MPI_Op can be a mixed bag of function pointers, potentially coming from multiple different op modules. This allows for hardware that only supports some of the operations, not all of them (e.g., a GPU that only supports single-precision operations). All the hard-coded back-end MPI_Op functions for (MPI_Op, MPI_Datatype) tuples still exist, but unlike coll, they're in the framework base (vs. being in a separate "basic" component) and are automatically used if no component is found at runtime that provides a module with the necessary function pointers. There is an "example" op component that will hopefully be useful to those writing meaningful op components. It is currently .ompi_ignore'd so that it doesn't impinge on other developers (it's somewhat chatty in terms of opal_output() so that you can tell when its functions have been invoked). See the README file in the example op component directory. Developers of new op components are encouraged to look at the following wiki pages: https://svn.open-mpi.org/trac/ompi/wiki/devel/Autogen https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateComponent https://svn.open-mpi.org/trac/ompi/wiki/devel/CreateFramework = MPI_REDUCE_LOCAL = Part of the MPI-2.2 proposal listed here: https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24 is to add a new function named MPI_REDUCE_LOCAL. It is very easy to implement, so I added it (also because it makes testing the op framework pretty easy -- you can do it in serial rather than via parallel reductions). There's even a man page! This commit was SVN r20280.
2009-01-15 02:44:31 +03:00
op->o_3buff_intrinsic.fns[ompi_op_ddt_map[dtype->id]](src1, src2,
tgt, &count,
&dtype,
op->o_3buff_intrinsic.modules[ompi_op_ddt_map[dtype->id]]);
}
END_C_DECLS
#endif /* OMPI_OP_H */