2013-03-28 01:09:41 +04:00
|
|
|
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
|
2007-03-03 02:19:02 +03:00
|
|
|
/*
|
|
|
|
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
|
|
|
|
* University Research and Technology
|
|
|
|
* Corporation. All rights reserved.
|
2015-02-15 22:47:27 +03:00
|
|
|
* Copyright (c) 2004-2015 The University of Tennessee and The University
|
2007-03-03 02:19:02 +03:00
|
|
|
* of Tennessee Research Foundation. All rights
|
|
|
|
* reserved.
|
|
|
|
* Copyright (c) 2004-2005 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.
|
2013-03-28 01:09:41 +04:00
|
|
|
* Copyright (c) 2013 Los Alamos National Security, LLC. All rights
|
|
|
|
* reserved.
|
2007-03-03 02:19:02 +03:00
|
|
|
* $COPYRIGHT$
|
|
|
|
*
|
|
|
|
* Additional copyrights may follow
|
|
|
|
*
|
|
|
|
* $HEADER$
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include "ompi_config.h"
|
|
|
|
|
|
|
|
#include "mpi.h"
|
|
|
|
#include "ompi/constants.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"
|
2007-03-03 02:19:02 +03:00
|
|
|
#include "ompi/communicator/communicator.h"
|
|
|
|
#include "ompi/mca/coll/coll.h"
|
|
|
|
#include "ompi/mca/coll/base/coll_tags.h"
|
|
|
|
#include "ompi/mca/pml/pml.h"
|
2015-02-15 22:47:27 +03:00
|
|
|
#include "ompi/mca/coll/base/coll_base_functions.h"
|
|
|
|
#include "coll_base_topo.h"
|
|
|
|
#include "coll_base_util.h"
|
2007-03-03 02:19:02 +03:00
|
|
|
|
|
|
|
int
|
2015-02-15 22:47:27 +03:00
|
|
|
ompi_coll_base_scatter_intra_binomial(void *sbuf, int scount,
|
2012-04-06 19:48:07 +04:00
|
|
|
struct ompi_datatype_t *sdtype,
|
|
|
|
void *rbuf, int rcount,
|
|
|
|
struct ompi_datatype_t *rdtype,
|
|
|
|
int root,
|
|
|
|
struct ompi_communicator_t *comm,
|
|
|
|
mca_coll_base_module_t *module)
|
2007-03-03 02:19:02 +03:00
|
|
|
{
|
2012-04-06 19:48:07 +04:00
|
|
|
int line = -1, i, rank, vrank, size, total_send = 0, err;
|
|
|
|
char *ptmp, *tempbuf = NULL;
|
2007-03-03 02:19:02 +03:00
|
|
|
ompi_coll_tree_t* bmtree;
|
|
|
|
MPI_Status status;
|
2015-02-15 22:47:27 +03:00
|
|
|
MPI_Aint sextent, slb, strue_lb, strue_extent;
|
2007-03-03 02:19:02 +03:00
|
|
|
MPI_Aint rextent, rlb, rtrue_lb, rtrue_extent;
|
2015-02-15 22:47:27 +03:00
|
|
|
mca_coll_base_module_t *base_module = (mca_coll_base_module_t*) module;
|
|
|
|
mca_coll_base_comm_t *data = base_module->base_data;
|
2007-03-03 02:19:02 +03:00
|
|
|
|
|
|
|
size = ompi_comm_size(comm);
|
|
|
|
rank = ompi_comm_rank(comm);
|
|
|
|
|
2015-02-15 22:47:27 +03:00
|
|
|
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
|
|
|
|
"ompi_coll_base_scatter_intra_binomial rank %d", rank));
|
2007-03-03 02:19:02 +03:00
|
|
|
|
|
|
|
/* create the binomial tree */
|
2015-02-15 22:47:27 +03:00
|
|
|
COLL_BASE_UPDATE_IN_ORDER_BMTREE( comm, base_module, root );
|
2007-08-19 07:37:49 +04:00
|
|
|
bmtree = data->cached_in_order_bmtree;
|
2007-03-03 02:19:02 +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_get_extent(sdtype, &slb, &sextent);
|
|
|
|
ompi_datatype_get_true_extent(sdtype, &strue_lb, &strue_extent);
|
|
|
|
ompi_datatype_get_extent(rdtype, &rlb, &rextent);
|
|
|
|
ompi_datatype_get_true_extent(rdtype, &rtrue_lb, &rtrue_extent);
|
2007-03-03 02:19:02 +03:00
|
|
|
|
|
|
|
vrank = (rank - root + size) % size;
|
2012-04-06 19:48:07 +04:00
|
|
|
ptmp = (char *) rbuf; /* by default suppose leaf nodes, just use rbuf */
|
2007-03-03 02:19:02 +03:00
|
|
|
|
|
|
|
if (rank == root) {
|
2012-04-06 19:48:07 +04:00
|
|
|
if (0 == root) {
|
|
|
|
/* root on 0, just use the send buffer */
|
|
|
|
ptmp = (char *) sbuf;
|
|
|
|
if (rbuf != MPI_IN_PLACE) {
|
|
|
|
/* local copy to rbuf */
|
|
|
|
err = ompi_datatype_sndrcv(sbuf, scount, sdtype,
|
|
|
|
rbuf, rcount, rdtype);
|
|
|
|
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
/* root is not on 0, allocate temp buffer for send */
|
|
|
|
tempbuf = (char *) malloc(strue_extent + ((ptrdiff_t)scount * (ptrdiff_t)size - 1) * sextent);
|
|
|
|
if (NULL == tempbuf) {
|
|
|
|
err = OMPI_ERR_OUT_OF_RESOURCE; line = __LINE__; goto err_hndl;
|
|
|
|
}
|
|
|
|
|
2014-11-14 07:22:01 +03:00
|
|
|
ptmp = tempbuf - strue_lb;
|
2012-04-06 19:48:07 +04:00
|
|
|
|
|
|
|
/* and rotate data so they will eventually in the right place */
|
|
|
|
err = ompi_datatype_copy_content_same_ddt(sdtype, (ptrdiff_t)scount * (ptrdiff_t)(size - root),
|
|
|
|
ptmp, (char *) sbuf + sextent * (ptrdiff_t)root * (ptrdiff_t)scount);
|
|
|
|
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
|
|
|
|
|
|
|
|
|
|
|
|
err = ompi_datatype_copy_content_same_ddt(sdtype, (ptrdiff_t)scount * (ptrdiff_t)root,
|
|
|
|
ptmp + sextent * (ptrdiff_t)scount * (ptrdiff_t)(size - root), (char *)sbuf);
|
|
|
|
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
|
|
|
|
|
|
|
|
if (rbuf != MPI_IN_PLACE) {
|
|
|
|
/* local copy to rbuf */
|
|
|
|
err = ompi_datatype_sndrcv(ptmp, scount, sdtype,
|
|
|
|
rbuf, rcount, rdtype);
|
|
|
|
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
|
|
|
|
}
|
|
|
|
}
|
|
|
|
total_send = scount;
|
2007-03-03 02:19:02 +03:00
|
|
|
} else if (!(vrank % 2)) {
|
2012-04-06 19:48:07 +04:00
|
|
|
/* non-root, non-leaf nodes, allocte temp buffer for recv
|
|
|
|
* the most we need is rcount*size/2 */
|
|
|
|
tempbuf = (char *) malloc(rtrue_extent + ((ptrdiff_t)rcount * (ptrdiff_t)size - 1) * rextent);
|
|
|
|
if (NULL == tempbuf) {
|
|
|
|
err= OMPI_ERR_OUT_OF_RESOURCE; line = __LINE__; goto err_hndl;
|
|
|
|
}
|
|
|
|
|
2014-11-14 07:22:01 +03:00
|
|
|
ptmp = tempbuf - rtrue_lb;
|
2012-04-06 19:48:07 +04:00
|
|
|
|
|
|
|
sdtype = rdtype;
|
|
|
|
scount = rcount;
|
|
|
|
sextent = rextent;
|
|
|
|
total_send = scount;
|
2007-03-03 02:19:02 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
if (!(vrank % 2)) {
|
2012-04-06 19:48:07 +04:00
|
|
|
if (rank != root) {
|
|
|
|
/* recv from parent on non-root */
|
|
|
|
err = MCA_PML_CALL(recv(ptmp, (ptrdiff_t)rcount * (ptrdiff_t)size, rdtype, bmtree->tree_prev,
|
|
|
|
MCA_COLL_BASE_TAG_SCATTER, comm, &status));
|
|
|
|
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
|
|
|
|
/* local copy to rbuf */
|
|
|
|
err = ompi_datatype_sndrcv(ptmp, scount, sdtype,
|
|
|
|
rbuf, rcount, rdtype);
|
|
|
|
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
|
|
|
|
}
|
|
|
|
/* send to children on all non-leaf */
|
|
|
|
for (i = 0; i < bmtree->tree_nextsize; i++) {
|
|
|
|
size_t mycount = 0;
|
2012-03-06 02:23:44 +04:00
|
|
|
int vkid;
|
2012-04-06 19:48:07 +04:00
|
|
|
/* figure out how much data I have to send to this child */
|
|
|
|
vkid = (bmtree->tree_next[i] - root + size) % size;
|
|
|
|
mycount = vkid - vrank;
|
|
|
|
if( (int)mycount > (size - vkid) )
|
|
|
|
mycount = size - vkid;
|
|
|
|
mycount *= scount;
|
|
|
|
|
|
|
|
err = MCA_PML_CALL(send(ptmp + (ptrdiff_t)total_send * sextent, mycount, sdtype,
|
|
|
|
bmtree->tree_next[i],
|
|
|
|
MCA_COLL_BASE_TAG_SCATTER,
|
|
|
|
MCA_PML_BASE_SEND_STANDARD, comm));
|
|
|
|
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
|
|
|
|
|
|
|
|
total_send += mycount;
|
|
|
|
}
|
|
|
|
|
2015-02-15 22:47:27 +03:00
|
|
|
if (NULL != tempbuf)
|
2012-04-06 19:48:07 +04:00
|
|
|
free(tempbuf);
|
2007-03-03 02:19:02 +03:00
|
|
|
} else {
|
2012-04-06 19:48:07 +04:00
|
|
|
/* recv from parent on leaf nodes */
|
|
|
|
err = MCA_PML_CALL(recv(ptmp, rcount, rdtype, bmtree->tree_prev,
|
|
|
|
MCA_COLL_BASE_TAG_SCATTER, comm, &status));
|
|
|
|
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
|
2007-03-03 02:19:02 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
return MPI_SUCCESS;
|
|
|
|
|
|
|
|
err_hndl:
|
|
|
|
if (NULL != tempbuf)
|
2012-04-06 19:48:07 +04:00
|
|
|
free(tempbuf);
|
2007-03-03 02:19:02 +03:00
|
|
|
|
2015-02-15 22:47:27 +03:00
|
|
|
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, "%s:%4d\tError occurred %d, rank %2d",
|
2012-04-06 19:48:07 +04:00
|
|
|
__FILE__, line, err, rank));
|
2007-04-02 23:11:06 +04:00
|
|
|
return err;
|
2007-03-03 02:19:02 +03:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Linear functions are copied from the BASIC coll module
|
|
|
|
* they do not segment the message and are simple implementations
|
2015-02-15 22:47:27 +03:00
|
|
|
* but for some small number of nodes and/or small data sizes they
|
|
|
|
* are just as fast as base/tree based segmenting operations
|
2007-03-03 02:19:02 +03:00
|
|
|
* and as such may be selected by the decision functions
|
|
|
|
* These are copied into this module due to the way we select modules
|
|
|
|
* in V1. i.e. in V2 we will handle this differently and so will not
|
|
|
|
* have to duplicate code.
|
2015-02-15 22:47:27 +03:00
|
|
|
* JPG following the examples from other coll_base implementations. Dec06.
|
2007-03-03 02:19:02 +03:00
|
|
|
*/
|
|
|
|
|
|
|
|
/* copied function (with appropriate renaming) starts here */
|
|
|
|
/*
|
|
|
|
* scatter_intra
|
|
|
|
*
|
|
|
|
* Function: - basic scatter operation
|
|
|
|
* Accepts: - same arguments as MPI_Scatter()
|
|
|
|
* Returns: - MPI_SUCCESS or error code
|
|
|
|
*/
|
|
|
|
int
|
2015-02-15 22:47:27 +03:00
|
|
|
ompi_coll_base_scatter_intra_basic_linear(void *sbuf, int scount,
|
2015-02-16 01:59:18 +03:00
|
|
|
struct ompi_datatype_t *sdtype,
|
|
|
|
void *rbuf, int rcount,
|
|
|
|
struct ompi_datatype_t *rdtype,
|
|
|
|
int root,
|
|
|
|
struct ompi_communicator_t *comm,
|
|
|
|
mca_coll_base_module_t *module)
|
2007-03-03 02:19:02 +03:00
|
|
|
{
|
|
|
|
int i, rank, size, err;
|
|
|
|
ptrdiff_t lb, incr;
|
2012-04-06 19:48:07 +04:00
|
|
|
char *ptmp;
|
2007-03-03 02:19:02 +03:00
|
|
|
|
|
|
|
/* Initialize */
|
|
|
|
|
|
|
|
rank = ompi_comm_rank(comm);
|
|
|
|
size = ompi_comm_size(comm);
|
|
|
|
|
|
|
|
/* If not root, receive data. */
|
|
|
|
|
|
|
|
if (rank != root) {
|
|
|
|
err = MCA_PML_CALL(recv(rbuf, rcount, rdtype, root,
|
|
|
|
MCA_COLL_BASE_TAG_SCATTER,
|
|
|
|
comm, MPI_STATUS_IGNORE));
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* I am the root, loop sending data. */
|
|
|
|
|
- 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_get_extent(sdtype, &lb, &incr);
|
2007-03-03 02:19:02 +03:00
|
|
|
if (OMPI_SUCCESS != err) {
|
|
|
|
return OMPI_ERROR;
|
|
|
|
}
|
|
|
|
|
|
|
|
incr *= scount;
|
|
|
|
for (i = 0, ptmp = (char *) sbuf; i < size; ++i, ptmp += incr) {
|
|
|
|
|
|
|
|
/* simple optimization */
|
|
|
|
|
|
|
|
if (i == rank) {
|
|
|
|
if (MPI_IN_PLACE != rbuf) {
|
|
|
|
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
|
|
|
ompi_datatype_sndrcv(ptmp, scount, sdtype, rbuf, rcount,
|
2012-04-06 19:48:07 +04:00
|
|
|
rdtype);
|
2007-03-03 02:19:02 +03:00
|
|
|
}
|
|
|
|
} else {
|
|
|
|
err = MCA_PML_CALL(send(ptmp, scount, sdtype, i,
|
|
|
|
MCA_COLL_BASE_TAG_SCATTER,
|
|
|
|
MCA_PML_BASE_SEND_STANDARD, comm));
|
|
|
|
}
|
|
|
|
if (MPI_SUCCESS != err) {
|
|
|
|
return err;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* All done */
|
|
|
|
|
|
|
|
return MPI_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/* copied function (with appropriate renaming) ends here */
|