2004-01-12 00:26:55 +03:00
|
|
|
/*
|
2005-11-05 22:57:48 +03:00
|
|
|
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
|
|
|
|
* University Research and Technology
|
|
|
|
* Corporation. All rights reserved.
|
2006-08-24 20:38:08 +04:00
|
|
|
* Copyright (c) 2004-2006 The University of Tennessee and The University
|
2005-11-05 22:57:48 +03:00
|
|
|
* of Tennessee Research Foundation. All rights
|
|
|
|
* reserved.
|
2004-11-28 23:09:25 +03:00
|
|
|
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
|
|
|
|
* University of Stuttgart. All rights reserved.
|
2005-03-24 15:43:37 +03:00
|
|
|
* Copyright (c) 2004-2005 The Regents of the University of California.
|
|
|
|
* All rights reserved.
|
2012-04-04 19:11:03 +04:00
|
|
|
* Copyright (c) 2012 Oak Ridge National Labs. All rights reserved.
|
2004-11-22 04:38:40 +03:00
|
|
|
* $COPYRIGHT$
|
|
|
|
*
|
|
|
|
* Additional copyrights may follow
|
|
|
|
*
|
2004-01-12 00:26:55 +03:00
|
|
|
* $HEADER$
|
|
|
|
*/
|
|
|
|
|
2004-06-07 19:33:53 +04:00
|
|
|
#include "ompi_config.h"
|
2004-02-15 00:26:30 +03:00
|
|
|
#include "coll_basic.h"
|
2004-01-12 00:26:55 +03:00
|
|
|
|
|
|
|
#include "mpi.h"
|
- Split the datatype engine into two parts: an MPI specific part in
OMPI
and a language agnostic part in OPAL. The convertor is completely
moved into OPAL. This offers several benefits as described in RFC
http://www.open-mpi.org/community/lists/devel/2009/07/6387.php
namely:
- Fewer basic types (int* and float* types, boolean and wchar
- Fixing naming scheme to ompi-nomenclature.
- Usability outside of the ompi-layer.
- Due to the fixed nature of simple opal types, their information is
completely
known at compile time and therefore constified
- With fewer datatypes (22), the actual sizes of bit-field types may be
reduced
from 64 to 32 bits, allowing reorganizing the opal_datatype
structure, eliminating holes and keeping data required in convertor
(upon send/recv) in one cacheline...
This has implications to the convertor-datastructure and other parts
of the code.
- Several performance tests have been run, the netpipe latency does not
change with
this patch on Linux/x86-64 on the smoky cluster.
- Extensive tests have been done to verify correctness (no new
regressions) using:
1. mpi_test_suite on linux/x86-64 using clean ompi-trunk and
ompi-ddt:
a. running both trunk and ompi-ddt resulted in no differences
(except for MPI_SHORT_INT and MPI_TYPE_MIX_LB_UB do now run
correctly).
b. with --enable-memchecker and running under valgrind (one buglet
when run with static found in test-suite, commited)
2. ibm testsuite on linux/x86-64 using clean ompi-trunk and ompi-ddt:
all passed (except for the dynamic/ tests failed!! as trunk/MTT)
3. compilation and usage of HDF5 tests on Jaguar using PGI and
PathScale compilers.
4. compilation and usage on Scicortex.
- Please note, that for the heterogeneous case, (-m32 compiled
binaries/ompi), neither
ompi-trunk, nor ompi-ddt branch would successfully launch.
This commit was SVN r21641.
2009-07-13 08:56:31 +04:00
|
|
|
#include "ompi/datatype/ompi_datatype.h"
|
2006-02-12 04:33:29 +03:00
|
|
|
#include "ompi/constants.h"
|
2005-09-13 00:21:53 +04:00
|
|
|
#include "ompi/mca/coll/coll.h"
|
|
|
|
#include "ompi/mca/coll/base/coll_tags.h"
|
2004-01-12 00:26:55 +03:00
|
|
|
#include "coll_basic.h"
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
2004-06-29 04:02:25 +04:00
|
|
|
* allgatherv_intra
|
2004-01-12 00:26:55 +03:00
|
|
|
*
|
2005-01-30 03:24:05 +03:00
|
|
|
* Function: - allgatherv using other MPI collectives
|
2004-01-12 00:26:55 +03:00
|
|
|
* Accepts: - same as MPI_Allgatherv()
|
|
|
|
* Returns: - MPI_SUCCESS or error code
|
|
|
|
*/
|
2005-08-10 14:51:42 +04:00
|
|
|
int
|
|
|
|
mca_coll_basic_allgatherv_intra(void *sbuf, int scount,
|
2005-08-10 21:53:43 +04:00
|
|
|
struct ompi_datatype_t *sdtype,
|
|
|
|
void *rbuf, int *rcounts, int *disps,
|
|
|
|
struct ompi_datatype_t *rdtype,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module)
|
2004-01-12 00:26:55 +03:00
|
|
|
{
|
2006-11-07 21:07:55 +03:00
|
|
|
int i, size, rank ;
|
2005-08-10 14:51:42 +04:00
|
|
|
int err;
|
2005-09-15 22:21:18 +04:00
|
|
|
MPI_Aint extent;
|
|
|
|
MPI_Aint lb;
|
2005-09-17 14:54:48 +04:00
|
|
|
char *send_buf = NULL;
|
2006-11-07 21:07:55 +03:00
|
|
|
struct ompi_datatype_t *newtype, *send_type;
|
2004-01-12 00:26:55 +03:00
|
|
|
|
2005-08-10 14:51:42 +04:00
|
|
|
size = ompi_comm_size(comm);
|
2005-09-14 00:06:54 +04:00
|
|
|
rank = ompi_comm_rank(comm);
|
2006-11-07 21:07:55 +03:00
|
|
|
/*
|
|
|
|
* We don't have a root process defined. Arbitrarily assign root
|
|
|
|
* to process with rank 0 (OMPI convention)
|
|
|
|
*/
|
2005-09-15 22:21:18 +04:00
|
|
|
|
|
|
|
if (MPI_IN_PLACE == sbuf) {
|
- Split the datatype engine into two parts: an MPI specific part in
OMPI
and a language agnostic part in OPAL. The convertor is completely
moved into OPAL. This offers several benefits as described in RFC
http://www.open-mpi.org/community/lists/devel/2009/07/6387.php
namely:
- Fewer basic types (int* and float* types, boolean and wchar
- Fixing naming scheme to ompi-nomenclature.
- Usability outside of the ompi-layer.
- Due to the fixed nature of simple opal types, their information is
completely
known at compile time and therefore constified
- With fewer datatypes (22), the actual sizes of bit-field types may be
reduced
from 64 to 32 bits, allowing reorganizing the opal_datatype
structure, eliminating holes and keeping data required in convertor
(upon send/recv) in one cacheline...
This has implications to the convertor-datastructure and other parts
of the code.
- Several performance tests have been run, the netpipe latency does not
change with
this patch on Linux/x86-64 on the smoky cluster.
- Extensive tests have been done to verify correctness (no new
regressions) using:
1. mpi_test_suite on linux/x86-64 using clean ompi-trunk and
ompi-ddt:
a. running both trunk and ompi-ddt resulted in no differences
(except for MPI_SHORT_INT and MPI_TYPE_MIX_LB_UB do now run
correctly).
b. with --enable-memchecker and running under valgrind (one buglet
when run with static found in test-suite, commited)
2. ibm testsuite on linux/x86-64 using clean ompi-trunk and ompi-ddt:
all passed (except for the dynamic/ tests failed!! as trunk/MTT)
3. compilation and usage of HDF5 tests on Jaguar using PGI and
PathScale compilers.
4. compilation and usage on Scicortex.
- Please note, that for the heterogeneous case, (-m32 compiled
binaries/ompi), neither
ompi-trunk, nor ompi-ddt branch would successfully launch.
This commit was SVN r21641.
2009-07-13 08:56:31 +04:00
|
|
|
ompi_datatype_get_extent(rdtype, &lb, &extent);
|
2006-11-07 21:07:55 +03:00
|
|
|
send_type = rdtype;
|
2006-08-24 20:38:08 +04:00
|
|
|
send_buf = (char*)rbuf;
|
2005-09-15 22:21:18 +04:00
|
|
|
for (i = 0; i < rank; ++i) {
|
|
|
|
send_buf += (rcounts[i] * extent);
|
|
|
|
}
|
2006-11-07 21:07:55 +03:00
|
|
|
} else {
|
2006-11-08 19:52:48 +03:00
|
|
|
send_buf = (char*)sbuf;
|
2006-11-07 21:07:55 +03:00
|
|
|
send_type = sdtype;
|
2005-09-15 22:21:18 +04:00
|
|
|
}
|
2006-11-07 21:07:55 +03:00
|
|
|
|
|
|
|
err = comm->c_coll.coll_gatherv(send_buf,
|
|
|
|
rcounts[rank], send_type,rbuf,
|
|
|
|
rcounts, disps, rdtype, 0,
|
2007-10-04 20:29:24 +04:00
|
|
|
comm, comm->c_coll.coll_gatherv_module);
|
|
|
|
|
2006-11-07 21:07:55 +03:00
|
|
|
if (MPI_SUCCESS != err) {
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* we now have all the data in the root's rbuf. Need to
|
|
|
|
* broadcast the data out to the other processes
|
|
|
|
*
|
|
|
|
* Need to define a datatype that captures the different vectors
|
|
|
|
* from each process. MPI_TYPE_INDEXED with params
|
|
|
|
* size,rcount,displs,rdtype,newtype
|
|
|
|
* should do the trick.
|
2006-11-08 19:52:48 +03:00
|
|
|
* Use underlying ddt functions to create, and commit the
|
2006-11-07 21:07:55 +03:00
|
|
|
* new datatype on each process, then broadcast and destroy the
|
|
|
|
* datatype.
|
|
|
|
*/
|
|
|
|
|
- Split the datatype engine into two parts: an MPI specific part in
OMPI
and a language agnostic part in OPAL. The convertor is completely
moved into OPAL. This offers several benefits as described in RFC
http://www.open-mpi.org/community/lists/devel/2009/07/6387.php
namely:
- Fewer basic types (int* and float* types, boolean and wchar
- Fixing naming scheme to ompi-nomenclature.
- Usability outside of the ompi-layer.
- Due to the fixed nature of simple opal types, their information is
completely
known at compile time and therefore constified
- With fewer datatypes (22), the actual sizes of bit-field types may be
reduced
from 64 to 32 bits, allowing reorganizing the opal_datatype
structure, eliminating holes and keeping data required in convertor
(upon send/recv) in one cacheline...
This has implications to the convertor-datastructure and other parts
of the code.
- Several performance tests have been run, the netpipe latency does not
change with
this patch on Linux/x86-64 on the smoky cluster.
- Extensive tests have been done to verify correctness (no new
regressions) using:
1. mpi_test_suite on linux/x86-64 using clean ompi-trunk and
ompi-ddt:
a. running both trunk and ompi-ddt resulted in no differences
(except for MPI_SHORT_INT and MPI_TYPE_MIX_LB_UB do now run
correctly).
b. with --enable-memchecker and running under valgrind (one buglet
when run with static found in test-suite, commited)
2. ibm testsuite on linux/x86-64 using clean ompi-trunk and ompi-ddt:
all passed (except for the dynamic/ tests failed!! as trunk/MTT)
3. compilation and usage of HDF5 tests on Jaguar using PGI and
PathScale compilers.
4. compilation and usage on Scicortex.
- Please note, that for the heterogeneous case, (-m32 compiled
binaries/ompi), neither
ompi-trunk, nor ompi-ddt branch would successfully launch.
This commit was SVN r21641.
2009-07-13 08:56:31 +04:00
|
|
|
err = ompi_datatype_create_indexed(size,rcounts,disps,rdtype,&newtype);
|
2006-11-07 21:07:55 +03:00
|
|
|
if (MPI_SUCCESS != err) {
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
- Split the datatype engine into two parts: an MPI specific part in
OMPI
and a language agnostic part in OPAL. The convertor is completely
moved into OPAL. This offers several benefits as described in RFC
http://www.open-mpi.org/community/lists/devel/2009/07/6387.php
namely:
- Fewer basic types (int* and float* types, boolean and wchar
- Fixing naming scheme to ompi-nomenclature.
- Usability outside of the ompi-layer.
- Due to the fixed nature of simple opal types, their information is
completely
known at compile time and therefore constified
- With fewer datatypes (22), the actual sizes of bit-field types may be
reduced
from 64 to 32 bits, allowing reorganizing the opal_datatype
structure, eliminating holes and keeping data required in convertor
(upon send/recv) in one cacheline...
This has implications to the convertor-datastructure and other parts
of the code.
- Several performance tests have been run, the netpipe latency does not
change with
this patch on Linux/x86-64 on the smoky cluster.
- Extensive tests have been done to verify correctness (no new
regressions) using:
1. mpi_test_suite on linux/x86-64 using clean ompi-trunk and
ompi-ddt:
a. running both trunk and ompi-ddt resulted in no differences
(except for MPI_SHORT_INT and MPI_TYPE_MIX_LB_UB do now run
correctly).
b. with --enable-memchecker and running under valgrind (one buglet
when run with static found in test-suite, commited)
2. ibm testsuite on linux/x86-64 using clean ompi-trunk and ompi-ddt:
all passed (except for the dynamic/ tests failed!! as trunk/MTT)
3. compilation and usage of HDF5 tests on Jaguar using PGI and
PathScale compilers.
4. compilation and usage on Scicortex.
- Please note, that for the heterogeneous case, (-m32 compiled
binaries/ompi), neither
ompi-trunk, nor ompi-ddt branch would successfully launch.
This commit was SVN r21641.
2009-07-13 08:56:31 +04:00
|
|
|
err = ompi_datatype_commit(&newtype);
|
2006-11-07 21:07:55 +03:00
|
|
|
if(MPI_SUCCESS != err) {
|
|
|
|
return err;
|
2004-06-29 04:02:25 +04:00
|
|
|
}
|
2004-01-12 00:26:55 +03:00
|
|
|
|
2012-04-04 19:11:03 +04:00
|
|
|
err = comm->c_coll.coll_bcast( rbuf, 1 ,newtype,0,comm,
|
|
|
|
comm->c_coll.coll_bcast_module);
|
2006-11-07 21:07:55 +03:00
|
|
|
|
- Split the datatype engine into two parts: an MPI specific part in
OMPI
and a language agnostic part in OPAL. The convertor is completely
moved into OPAL. This offers several benefits as described in RFC
http://www.open-mpi.org/community/lists/devel/2009/07/6387.php
namely:
- Fewer basic types (int* and float* types, boolean and wchar
- Fixing naming scheme to ompi-nomenclature.
- Usability outside of the ompi-layer.
- Due to the fixed nature of simple opal types, their information is
completely
known at compile time and therefore constified
- With fewer datatypes (22), the actual sizes of bit-field types may be
reduced
from 64 to 32 bits, allowing reorganizing the opal_datatype
structure, eliminating holes and keeping data required in convertor
(upon send/recv) in one cacheline...
This has implications to the convertor-datastructure and other parts
of the code.
- Several performance tests have been run, the netpipe latency does not
change with
this patch on Linux/x86-64 on the smoky cluster.
- Extensive tests have been done to verify correctness (no new
regressions) using:
1. mpi_test_suite on linux/x86-64 using clean ompi-trunk and
ompi-ddt:
a. running both trunk and ompi-ddt resulted in no differences
(except for MPI_SHORT_INT and MPI_TYPE_MIX_LB_UB do now run
correctly).
b. with --enable-memchecker and running under valgrind (one buglet
when run with static found in test-suite, commited)
2. ibm testsuite on linux/x86-64 using clean ompi-trunk and ompi-ddt:
all passed (except for the dynamic/ tests failed!! as trunk/MTT)
3. compilation and usage of HDF5 tests on Jaguar using PGI and
PathScale compilers.
4. compilation and usage on Scicortex.
- Please note, that for the heterogeneous case, (-m32 compiled
binaries/ompi), neither
ompi-trunk, nor ompi-ddt branch would successfully launch.
This commit was SVN r21641.
2009-07-13 08:56:31 +04:00
|
|
|
ompi_datatype_destroy (&newtype);
|
2006-11-07 21:07:55 +03:00
|
|
|
|
2012-04-04 19:11:03 +04:00
|
|
|
return err;
|
2004-06-29 04:02:25 +04:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
* allgatherv_inter
|
|
|
|
*
|
2005-01-30 03:24:05 +03:00
|
|
|
* Function: - allgatherv using other MPI collectives
|
2004-06-29 04:02:25 +04:00
|
|
|
* Accepts: - same as MPI_Allgatherv()
|
|
|
|
* Returns: - MPI_SUCCESS or error code
|
|
|
|
*/
|
2005-08-10 14:51:42 +04:00
|
|
|
int
|
|
|
|
mca_coll_basic_allgatherv_inter(void *sbuf, int scount,
|
2005-08-10 21:53:43 +04:00
|
|
|
struct ompi_datatype_t *sdtype,
|
|
|
|
void *rbuf, int *rcounts, int *disps,
|
|
|
|
struct ompi_datatype_t *rdtype,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module)
|
2004-06-29 04:02:25 +04:00
|
|
|
{
|
2008-08-06 17:57:43 +04:00
|
|
|
int rsize, err, i;
|
2009-05-05 17:47:47 +04:00
|
|
|
int *scounts, *sdisps;
|
2004-06-29 04:02:25 +04:00
|
|
|
|
2005-08-10 14:51:42 +04:00
|
|
|
rsize = ompi_comm_remote_size(comm);
|
|
|
|
|
2009-05-05 17:47:47 +04:00
|
|
|
scounts = (int *) malloc(2 * rsize * sizeof(int));
|
|
|
|
sdisps = scounts + rsize;
|
|
|
|
if (NULL == scounts) {
|
2005-08-10 21:53:43 +04:00
|
|
|
return OMPI_ERR_OUT_OF_RESOURCE;
|
2004-08-04 02:12:57 +04:00
|
|
|
}
|
2005-08-10 14:51:42 +04:00
|
|
|
|
|
|
|
for (i = 0; i < rsize; i++) {
|
2005-08-10 21:53:43 +04:00
|
|
|
scounts[i] = scount;
|
2009-05-05 17:47:47 +04:00
|
|
|
sdisps[i] = 0;
|
2004-08-05 22:52:14 +04:00
|
|
|
}
|
2004-08-04 02:12:57 +04:00
|
|
|
|
2005-08-10 14:51:42 +04:00
|
|
|
err = comm->c_coll.coll_alltoallv(sbuf, scounts, sdisps, sdtype,
|
2007-10-04 20:29:24 +04:00
|
|
|
rbuf, rcounts, disps, rdtype, comm,
|
|
|
|
comm->c_coll.coll_alltoallv_module);
|
2004-08-04 02:12:57 +04:00
|
|
|
|
2005-08-10 14:51:42 +04:00
|
|
|
if (NULL != scounts) {
|
2005-08-10 21:53:43 +04:00
|
|
|
free(scounts);
|
2004-08-04 02:12:57 +04:00
|
|
|
}
|
2005-08-10 14:51:42 +04:00
|
|
|
|
2004-08-04 02:12:57 +04:00
|
|
|
return err;
|
2004-01-12 00:26:55 +03:00
|
|
|
}
|