July 2013 - commits - mailman3.cels.anl.gov

[mpich] MPICH primary repository branch, master, updated. v3.0.4-411-g27f6340
by noreply＠mpich.org 31 Jul '13

31 Jul '13

This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "MPICH primary repository". The branch, master has been updated via 27f6340573fe2e7bc0f2a387ab791fa95ec1479d (commit) from b54d7e0831d6f617c3971e05b643a47d06413423 (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- http://git.mpich.org/mpich.git/commitdiff/27f6340573fe2e7bc0f2a387ab791fa95… commit 27f6340573fe2e7bc0f2a387ab791fa95ec1479d Author: Pavan Balaji <balaji(a)mcs.anl.gov> Date: Wed Jul 31 10:51:54 2013 -0500 Error in the buffer attached to dynamic windows. In MPICH, Win_attach/detach are essentially no-ops. The actual accumulate calls don't actually check if the actual buffer is attached or not. This is technically a user error. The win_dynamic_acc test had a wrong buffer attached, but was still passing because of this reason. This patch only corrects the test program. Thanks to Michael Raymond @ SGI for reporting the error. Signed-off-by: Xin Zhao <xinzhao3(a)illinois.edu> diff --git a/test/mpi/rma/win_dynamic_acc.c b/test/mpi/rma/win_dynamic_acc.c index ebb0a35..ee9f725 100644 --- a/test/mpi/rma/win_dynamic_acc.c +++ b/test/mpi/rma/win_dynamic_acc.c @@ -34,7 +34,7 @@ int main(int argc, char **argv) { MPI_COMM_WORLD); MPI_Win_create_dynamic(MPI_INFO_NULL, MPI_COMM_WORLD, &dyn_win); - MPI_Win_attach(dyn_win, &one, sizeof(int)); + MPI_Win_attach(dyn_win, &val, sizeof(int)); for (i = 0; i < ITER; i++) { MPI_Win_fence(MPI_MODE_NOPRECEDE, dyn_win); @@ -50,7 +50,7 @@ int main(int argc, char **argv) { printf("%d -- Got %d, expected %d\n", rank, val, ITER); } - MPI_Win_detach(dyn_win, &one); + MPI_Win_detach(dyn_win, &val); MPI_Win_free(&dyn_win); MPI_Reduce(&errors, &all_errors, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD); ----------------------------------------------------------------------- Summary of changes: test/mpi/rma/win_dynamic_acc.c | 4 ++-- 1 files changed, 2 insertions(+), 2 deletions(-) hooks/post-receive -- MPICH primary repository

1 0

[mpich] MPICH primary repository branch, master, updated. v3.0.4-410-gb54d7e0
by noreply＠mpich.org 31 Jul '13

31 Jul '13

This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "MPICH primary repository". The branch, master has been updated via b54d7e0831d6f617c3971e05b643a47d06413423 (commit) from b6a06cd3e6cee9d351bee7cb9ca35c422a5f73a0 (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- http://git.mpich.org/mpich.git/commitdiff/b54d7e0831d6f617c3971e05b643a47d0… commit b54d7e0831d6f617c3971e05b643a47d06413423 Author: Ken Raffenetti <raffenet(a)mcs.anl.gov> Date: Tue Jul 30 17:31:27 2013 -0500 Ensure fault tolerance tests only run with hydra. Fault tolerance is only supported under hydra, so we configure the tests to run only when it is the primary PM. Fixes #1914. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/test/mpi/Makefile.am b/test/mpi/Makefile.am index 0ffd429..c314e09 100644 --- a/test/mpi/Makefile.am +++ b/test/mpi/Makefile.am @@ -11,7 +11,7 @@ include $(top_srcdir)/Makefile.mtest ACLOCAL_AMFLAGS = -I confdb static_subdirs = util attr basic datatype coll comm errhan group info init \ - pt2pt rma topo errors manual perf mpi_t ft + pt2pt rma topo errors manual perf mpi_t all_lang_subdirs = f77 cxx f90 # DIST_SUBDIRS must be a superset of SUBDIRS, and automake must be able to @@ -19,7 +19,7 @@ all_lang_subdirs = f77 cxx f90 # duplication because automake is able to "see" into simple variable # assignments that are not driven by a configure @-substitution variable. DIST_SUBDIRS = $(static_subdirs) io $(all_lang_subdirs) threads spawn . -SUBDIRS = $(static_subdirs) $(iodir) $(otherlangs) $(threadsdir) $(spawndir) $(ckpointdir) . +SUBDIRS = $(static_subdirs) $(iodir) $(otherlangs) $(threadsdir) $(spawndir) $(ckpointdir) $(ftdir) . EXTRA_DIST = maint/common.defn maint/f77tof90.in maint/testmerge.in maint/updatefiles testlist.in diff --git a/test/mpi/configure.ac b/test/mpi/configure.ac index 55e1790..565842a 100644 --- a/test/mpi/configure.ac +++ b/test/mpi/configure.ac @@ -177,6 +177,57 @@ AC_ARG_WITH(mpi, [Use the selected MPI; compilation scripts for mpicc, mpif77 and mpicxx should be in dir/bin])],,) +AC_ARG_WITH(pm, + AC_HELP_STRING([--with-pm=name], + [Specify the process manager for MPICH. "no" or "none" are + valid values. Multiple process managers may be specified as + long as they all use the same pmi interface by separating them + with colons. The mpiexec for the first named process manager + will be installed. Example: "--with-pm=hydra:mpd:gforker" + builds the three process managers hydra, mpd and gforker; + only the mpiexec from hydra is installed into the bin + directory.]),,with_pm=default) + +if test "$with_pm" = "none" ; then + # add "none" as synonym for "no" to agree with older erroneous docs + with_pm="no" +fi +if test "$MPID_NO_PM" = yes ; then + if test "$with_pm" != "default" -a "$with_pm" != no ; then + AC_MSG_ERROR([The PM chosen ($with_pm) is is not valid for the selected device ($with_device)]) + fi + # This is used to change with_pm=default to with_pm=no in the case + # where the device does not want a PM + with_pm=no +fi +if test -z "$with_pm" ; then + with_pm="no" +fi +if test "$with_pmi" = "uni" -a "$with_pm" = "default" ; then + with_pm="no" +fi +if test "$with_pm" = "default" -o "$with_pm" = "yes" ; then + if test ! -z "$MPID_DEFAULT_PM" ; then + with_pm=${MPID_DEFAULT_PM} + else + with_pm=hydra + fi +fi + +# Get the first pm specified +if test "$with_pm" != "no" ; then + first_pm="`echo $with_pm | sed -e 's/:.*//' -e 's/,.*//'`" +else + first_pm="" +fi + +# Only run the fault tolerance tests if hydra is the primary PM +ftdir="#ft" +if test "$first_pm" = "hydra"; then + ftdir="ft" +fi +AC_SUBST(ftdir) + AC_ARG_WITH(config-args, [AC_HELP_STRING([--with-config-args=filename], [Specify configure argument file that contains the diff --git a/test/mpi/testlist.in b/test/mpi/testlist.in index 5d32b47..f822165 100644 --- a/test/mpi/testlist.in +++ b/test/mpi/testlist.in @@ -22,4 +22,4 @@ topo @errordir@ @threadsdir@ @ckpointdir@ -ft +@ftdir@ ----------------------------------------------------------------------- Summary of changes: test/mpi/Makefile.am | 4 +- test/mpi/configure.ac | 51 +++++++++++++++++++++++++++++++++++++++++++++++++ test/mpi/testlist.in | 2 +- 3 files changed, 54 insertions(+), 3 deletions(-) hooks/post-receive -- MPICH primary repository

1 0

[mpich] MPICH primary repository branch, master, updated. v3.0.4-409-gb6a06cd
by noreply＠mpich.org 30 Jul '13

30 Jul '13

This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "MPICH primary repository". The branch, master has been updated via b6a06cd3e6cee9d351bee7cb9ca35c422a5f73a0 (commit) from 1f307dfa1ce322081457139cdbfbc28e71f9043b (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- http://git.mpich.org/mpich.git/commitdiff/b6a06cd3e6cee9d351bee7cb9ca35c422… commit b6a06cd3e6cee9d351bee7cb9ca35c422a5f73a0 Author: William Gropp <wgropp(a)illinois.edu> Date: Mon Jul 29 10:16:18 2013 -0500 Recent versions of Cray Fortran may require a special arg to create modules While testing on the Blue Waters development system, builds failed because the code for determining the Fortran Module options was unable to figure out the Cray compiler (naturally, this used to work). Investigation showed that the compiler needs a "-em" option to create separate module files. This change adds a test for that, if the compiler is unable to create module files without that option. This also cleans up several other minor problems with this code. diff --git a/confdb/aclocal_fc.m4 b/confdb/aclocal_fc.m4 index 02a5d5d..7c9ef94 100644 --- a/confdb/aclocal_fc.m4 +++ b/confdb/aclocal_fc.m4 @@ -171,6 +171,10 @@ dnl Special characteristics that have no autoconf counterpart but that dnl we need as part of the Fortran 90 support. To distinquish these, they dnl have a [PAC] prefix. dnl +dnl At least one version of the Cray compiler needs the option -em to +dnl generate a separate module file, rather than including the module +dnl information in the object (.o) file. +dnl dnl dnl PAC_FC_MODULE_EXT(action if found,action if not found) dnl @@ -187,11 +191,11 @@ AC_COMPILE_IFELSE([ end module conftest ]) ],[ - dnl Look for module name - dnl First, try to find known names. This avoids confusion caused by - dnl additional files (like <name>.stb created by some versions of pgf90) - dnl Early versions of the Intel compiler used "d" as the module extension; - dnl we include that just to finish the test as early as possible. + # Look for module name + # First, try to find known names. This avoids confusion caused by + # additional files (like <name>.stb created by some versions of pgf90) + # Early versions of the Intel compiler used d as the module extension; + # we include that just to finish the test as early as possible. for name in conftest CONFTEST ; do for ext in mod MOD d ; do if test -s $name.$ext ; then @@ -208,14 +212,14 @@ AC_COMPILE_IFELSE([ if test -n "$pac_cv_fc_module_ext" ; then break ; fi done if test -z "$pac_MOD" ; then - pac_MOD=`ls conftest.* 2>&1 | grep -v conftest.${ac_fc_srcext} | grep -v conftest.o` + # The test on .err is needed for Cray Fortran. + pac_MOD=`ls conftest.* 2>&1 | grep -v conftest.${ac_fc_srcext} | grep -v conftest.o | grep -v conftest.err` pac_MOD=`echo $pac_MOD | sed -e 's/conftest\.//g'` pac_cv_fc_module_case="lower" if test "X$pac_MOD" = "X" ; then - pac_MOD=`ls CONFTEST* 2>&1 | grep -v CONFTEST.f | grep -v CONFTEST.o` + pac_MOD=`ls CONFTEST* 2>&1 | grep -v CONFTEST.${ac_fc_srcext} | grep -v CONFTEST.o | grep -v CONFTEST.err` pac_MOD=`echo $pac_MOD | sed -e 's/CONFTEST\.//g'` if test -n "$pac_MOD" -a -s "CONFTEST.$pac_MOD" ; then - testname="CONFTEST" pac_cv_fc_module_case="upper" else # Clear because we must have gotten an error message @@ -231,11 +235,40 @@ AC_COMPILE_IFELSE([ ],[ pac_cv_fc_module_ext="unknown" ]) + +if test "$pac_cv_fc_module_ext" = "unknown" ; then + # Try again, but with an -em option. Abbreviated, because we're + # just looking for the Cray option + saveFCFLAGS=$FCFLAGS + FCFLAGS="$FCFLAGS -em" + AC_COMPILE_IFELSE([ + AC_LANG_SOURCE([ + module conftest + integer n + parameter (n=1) + end module conftest + ]) +],[ + if test -s conftest.mod ; then + pac_cv_fc_module_ext="mod" + pac_cv_fc_module_case="lower" + elif test -s CONFTEST.mod ; then + pac_cv_fc_module_ext="mod" + pac_cv_fc_module_case="upper" + fi +],[ + : + # do nothing - already have the unknown default value +]) + if test "$pac_cv_fc_module_ext" = "unknown" ; then + # The additional command line option did not help - restore + # the original flags. + FCFLAGS=$saveFCFLAGS + fi +fi AC_LANG_POP(Fortran) ]) -dnl -dnl -dnl +# AC_SUBST(FCMODEXT) if test "$pac_cv_fc_module_ext" = "unknown" ; then ifelse($2,,:,[$2]) @@ -400,10 +433,14 @@ AC_COMPILE_IFELSE([],[ # pgf90 -module ${dir} # ifort -module ${dir} # nagfor -mdir ${dir} -# ftn -J ${dir} ## the Cray fortran compiler +# ftn -J ${dir} ## the Cray fortran compiler +# ftn -em -J${dir} ## the Cray fortran compiler (crayftn, in 2013) +# For this above case, we must have added -em to FCFLAGS, since other +# module tests do not always use the module output flag. See +# FC_MODULE_EXT , where this is determined. # f95 -YMOD_OUT_DIR=${dir} ## the Absoft fortran compiler -# lf95 -Am -mod ${dir} ## the Lahey/Fujitsu fortran compiler -# f90 -moddir=${dir} ## the Sun f90 compiler +# lf95 -Am -mod ${dir} ## the Lahey/Fujitsu fortran compiler +# f90 -moddir=${dir} ## the Sun f90 compiler # g95 -fmod=${dir} # # If there are any compilers still out there that are totally brain-dead and ----------------------------------------------------------------------- Summary of changes: confdb/aclocal_fc.m4 | 65 +++++++++++++++++++++++++++++++++++++++---------- 1 files changed, 51 insertions(+), 14 deletions(-) hooks/post-receive -- MPICH primary repository

1 0

[mpich] MPICH primary repository branch, master, updated. v3.0.4-408-g1f307df
by noreply＠mpich.org 30 Jul '13

30 Jul '13

This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "MPICH primary repository". The branch, master has been updated via 1f307dfa1ce322081457139cdbfbc28e71f9043b (commit) from cfaa95cc36e05cc5df38e6a895c0b9e3abc90128 (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- http://git.mpich.org/mpich.git/commitdiff/1f307dfa1ce322081457139cdbfbc28e7… commit 1f307dfa1ce322081457139cdbfbc28e71f9043b Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Tue Jul 30 10:38:54 2013 -0500 Fixed touch command option order Fix touch command in autogen.sh that generates a random file and reruns configure during first make. Some systems (MACs) require the options to be specified prior to the target file for the touch command. This commit fixes the option order in the touch command in autogen.sh introduced to fix ticket #1870. Signed-off-by: Antonio J. Pena <apenya(a)mcs.anl.gov> diff --git a/autogen.sh b/autogen.sh index d79c451..5dc2a7e 100755 --- a/autogen.sh +++ b/autogen.sh @@ -917,7 +917,7 @@ if [ "$do_build_configure" = "yes" ] ; then # Remove possible leftovers, which don't imply a failure rm -f $amdir/confdb/libtool.m4.orig # Reset libtool.m4 timestamps to avoid confusing make - touch $amdir/confdb/libtool.m4 -r $amdir/confdb/ltversion.m4 + touch -r $amdir/confdb/ltversion.m4 $amdir/confdb/libtool.m4 else echo "failed" fi ----------------------------------------------------------------------- Summary of changes: autogen.sh | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-) hooks/post-receive -- MPICH primary repository

1 0

[mpich] MPICH primary repository branch, master, updated. v3.0.4-407-gcfaa95c
by noreply＠mpich.org 29 Jul '13

29 Jul '13

This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "MPICH primary repository". The branch, master has been updated via cfaa95cc36e05cc5df38e6a895c0b9e3abc90128 (commit) from c565e663b28b5304f6d08ab83af34c565ece9e49 (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- http://git.mpich.org/mpich.git/commitdiff/cfaa95cc36e05cc5df38e6a895c0b9e3a… commit cfaa95cc36e05cc5df38e6a895c0b9e3abc90128 Author: Ken Raffenetti <raffenet(a)mcs.anl.gov> Date: Mon Jul 29 09:12:22 2013 -0500 Adds fault tolerance tests. Adds tests for when -disable-auto-cleanup is enabled and a process dies or calls MPI_Abort. There is no communication in these cases. Signed-off-by: Wesley Bland <wbland(a)mcs.anl.gov> diff --git a/test/mpi/Makefile.am b/test/mpi/Makefile.am index ed82f0e..0ffd429 100644 --- a/test/mpi/Makefile.am +++ b/test/mpi/Makefile.am @@ -11,7 +11,7 @@ include $(top_srcdir)/Makefile.mtest ACLOCAL_AMFLAGS = -I confdb static_subdirs = util attr basic datatype coll comm errhan group info init \ - pt2pt rma topo errors manual perf mpi_t + pt2pt rma topo errors manual perf mpi_t ft all_lang_subdirs = f77 cxx f90 # DIST_SUBDIRS must be a superset of SUBDIRS, and automake must be able to diff --git a/test/mpi/configure.ac b/test/mpi/configure.ac index 7339192..55e1790 100644 --- a/test/mpi/configure.ac +++ b/test/mpi/configure.ac @@ -1427,6 +1427,7 @@ AC_OUTPUT(maint/testmerge \ errors/f90/Makefile \ errors/f90/errhan/Makefile \ ckpoint/Makefile \ + ft/Makefile \ manual/Makefile \ manual/manyconnect \ manual/mpi_t/Makefile \ diff --git a/test/mpi/ft/Makefile.am b/test/mpi/ft/Makefile.am new file mode 100644 index 0000000..40061d3 --- /dev/null +++ b/test/mpi/ft/Makefile.am @@ -0,0 +1,13 @@ +# -*- Mode: Makefile; -*- +# vim: set ft=automake : +# +# (C) 2011 by Argonne National Laboratory. +# See COPYRIGHT in top-level directory. +# + +include $(top_srcdir)/Makefile.mtest + +## for all programs that are just built from the single corresponding source +## file, we don't need per-target _SOURCES rules, automake will infer them +## correctly +noinst_PROGRAMS = die abort diff --git a/test/mpi/ft/abort.c b/test/mpi/ft/abort.c new file mode 100644 index 0000000..0578fd4 --- /dev/null +++ b/test/mpi/ft/abort.c @@ -0,0 +1,27 @@ +/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil ; -*- */ +/* + * + * (C) 2003 by Argonne National Laboratory. + * See COPYRIGHT in top-level directory. + */ +#include "mpi.h" +#include <stdio.h> + +int main(int argc, char **argv) +{ + int rank, size; + + MPI_Init(&argc, &argv); + MPI_Comm_rank(MPI_COMM_WORLD, &rank); + + printf("No Errors\n"); + + if (rank == 0) + MPI_Abort(MPI_COMM_WORLD, MPI_SUCCESS); + + while(1) + ; + + MPI_Finalize(); + return 0; +} diff --git a/test/mpi/ft/die.c b/test/mpi/ft/die.c new file mode 100644 index 0000000..337f14b --- /dev/null +++ b/test/mpi/ft/die.c @@ -0,0 +1,33 @@ +/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil ; -*- */ +/* + * + * (C) 2003 by Argonne National Laboratory. + * See COPYRIGHT in top-level directory. + */ +#include <mpi.h> +#include <stdio.h> +#include <sys/types.h> +#include <unistd.h> +#include <signal.h> + +int main(int argc, char **argv) +{ + int rank, size; + pid_t pid; + + MPI_Init(&argc, &argv); + MPI_Comm_rank(MPI_COMM_WORLD, &rank); + + if (rank == 1) { + pid = getpid(); + kill(pid, SIGSTOP); + } + + MPI_Finalize(); + + if (rank == 0) { + printf("No Errors\n"); + } + + return 0; +} diff --git a/test/mpi/ft/testlist b/test/mpi/ft/testlist new file mode 100644 index 0000000..00187ae --- /dev/null +++ b/test/mpi/ft/testlist @@ -0,0 +1,2 @@ +die 4 mpiexecarg=-disable-auto-cleanup timeLimit=10 +abort 2 mpiexecarg=-disable-auto-cleanup timeLimit=10 diff --git a/test/mpi/testlist.in b/test/mpi/testlist.in index 0a77991..5d32b47 100644 --- a/test/mpi/testlist.in +++ b/test/mpi/testlist.in @@ -22,3 +22,4 @@ topo @errordir@ @threadsdir@ @ckpointdir@ +ft ----------------------------------------------------------------------- Summary of changes: test/mpi/Makefile.am | 2 +- test/mpi/configure.ac | 1 + test/mpi/{ckpoint => ft}/Makefile.am | 2 +- test/mpi/{manual/segfault.c => ft/abort.c} | 24 +++++++++++++----------- test/mpi/{manual/segfault.c => ft/die.c} | 28 ++++++++++++++++++---------- test/mpi/ft/testlist | 2 ++ test/mpi/testlist.in | 1 + 7 files changed, 37 insertions(+), 23 deletions(-) copy test/mpi/{ckpoint => ft}/Makefile.am (91%) copy test/mpi/{manual/segfault.c => ft/abort.c} (51%) copy test/mpi/{manual/segfault.c => ft/die.c} (52%) create mode 100644 test/mpi/ft/testlist hooks/post-receive -- MPICH primary repository

1 0

[mpich] MPICH primary repository branch, master, updated. v3.0.4-406-gc565e66
by noreply＠mpich.org 29 Jul '13

29 Jul '13

This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "MPICH primary repository". The branch, master has been updated via c565e663b28b5304f6d08ab83af34c565ece9e49 (commit) from eb42872a7e4d97853e9e5b8e70a7a37f9dcab2bc (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- http://git.mpich.org/mpich.git/commitdiff/c565e663b28b5304f6d08ab83af34c565… commit c565e663b28b5304f6d08ab83af34c565ece9e49 Author: Wesley Bland <wbland(a)mcs.anl.gov> Date: Mon Jul 29 10:49:25 2013 -0500 Add missing end error handling tag. No reviewer necessary. diff --git a/src/mpi/coll/helper_fns.c b/src/mpi/coll/helper_fns.c index 12079be..26e1798 100644 --- a/src/mpi/coll/helper_fns.c +++ b/src/mpi/coll/helper_fns.c @@ -346,6 +346,7 @@ int MPIC_Recv(void *buf, int count, MPI_Datatype datatype, int source, int tag, /* --BEGIN ERROR HANDLING-- */ if (request_ptr) MPID_Request_release(request_ptr); goto fn_exit; + /* --END ERROR HANDLING-- */ } #undef FUNCNAME ----------------------------------------------------------------------- Summary of changes: src/mpi/coll/helper_fns.c | 1 + 1 files changed, 1 insertions(+), 0 deletions(-) hooks/post-receive -- MPICH primary repository

1 0

[mpich] MPICH primary repository branch, master, updated. v3.0.4-405-geb42872
by noreply＠mpich.org 28 Jul '13

28 Jul '13

This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "MPICH primary repository". The branch, master has been updated via eb42872a7e4d97853e9e5b8e70a7a37f9dcab2bc (commit) via b56e194829859aa69ac16cddba02c784d8c60c09 (commit) from cd5986fc1cfaa68fb4abb6c6245c82e319be77ee (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- http://git.mpich.org/mpich.git/commitdiff/eb42872a7e4d97853e9e5b8e70a7a37f9… commit eb42872a7e4d97853e9e5b8e70a7a37f9dcab2bc Author: Pavan Balaji <balaji(a)mcs.anl.gov> Date: Sun Jul 28 17:00:37 2013 -0500 Update CHANGES to reflect the improvement to MPI_WIN_ALLOCATE. diff --git a/CHANGES b/CHANGES index 7c2ca44..42a3041 100644 --- a/CHANGES +++ b/CHANGES @@ -15,6 +15,10 @@ time. The option MPIR_PARAM_ENABLE_COLL_FT_RET has been removed as it is no longer necessary. + # Improvements to MPI_WIN_ALLOCATE to internally allocate shared + memory between processes on the same node (the "alloc_shm" info + argument needs to be provided). + # Enable shared library builds by default. # Upgraded hwloc to 1.7.1. http://git.mpich.org/mpich.git/commitdiff/b56e194829859aa69ac16cddba02c784d… commit b56e194829859aa69ac16cddba02c784d8c60c09 Author: Pavan Balaji <balaji(a)mcs.anl.gov> Date: Sun Jul 28 16:57:08 2013 -0500 Update CHANGES to not describe the class of change. diff --git a/CHANGES b/CHANGES index a637edf..7c2ca44 100644 --- a/CHANGES +++ b/CHANGES @@ -2,25 +2,24 @@ Changes in 3.1 =============================================================================== - # OVERALL: Integrated MPICH-PAMI code base for Blue Gene/Q and other - IBM platforms. + # Integrated MPICH-PAMI code base for Blue Gene/Q and other IBM + platforms. - # OVERALL: Several improvements to the SCIF netmod. (code - contribution from Intel). + # Several improvements to the SCIF netmod. (code contribution from + Intel). - # OVERALL: Added environment variables to control a lot more - capabilities for collectives. See the README.envvar file for more - information. + # Added environment variables to control a lot more capabilities for + collectives. See the README.envvar file for more information. - # OVERALL: Allow non-blocking collectives and fault tolerance at the same - time. The option MPIR_PARAM_ENABLE_COLL_FT_RET has been removed as it is no - longer necessary. + # Allow non-blocking collectives and fault tolerance at the same + time. The option MPIR_PARAM_ENABLE_COLL_FT_RET has been removed as + it is no longer necessary. - # BUILD: Enable shared library builds by default. + # Enable shared library builds by default. - # PM/PMI: Upgraded hwloc to 1.7.1. + # Upgraded hwloc to 1.7.1. - # PM/PMI: Several improvements to the Hydra-SLURM integration. + # Several improvements to the Hydra-SLURM integration. # Several other minor bug fixes, memory leak fixes, and code cleanup. A full list of changes is available at the following link: ----------------------------------------------------------------------- Summary of changes: CHANGES | 29 ++++++++++++++++------------- 1 files changed, 16 insertions(+), 13 deletions(-) hooks/post-receive -- MPICH primary repository

1 0

[mpich] MPICH primary repository branch, master, updated. v3.0.4-403-gcd5986f
by noreply＠mpich.org 28 Jul '13

28 Jul '13

This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "MPICH primary repository". The branch, master has been updated via cd5986fc1cfaa68fb4abb6c6245c82e319be77ee (commit) via f038a277715f77c3b746215c6ac527cb1ddfebad (commit) via f0e27d1d65541d4a69a1152cd318a72e8c038bfa (commit) via 62cd51ca10890e27bcb13491d876c9d610b52876 (commit) via 5467e5004174421ae581dd08253adfe5c86a0000 (commit) via 447c1d5585135a41fd6341dcb9760ba945261f11 (commit) via 34dd7c1ad295d223119dbad3d032e38fa7f94c8e (commit) via 55cb0f4c1cdc5c1e7017d487bc2f13871dcf766e (commit) via b668bc1ed3168790d8361ff14e1976581349cb81 (commit) via 384d96b7b7f193963a178e67ebb64d750751f3cc (commit) via 44d9c98dc934631af208c62113703c135bd86064 (commit) via 382bdfb2a2d35025b906033bc63fea98e3c551e3 (commit) via e397148092dc2377b0671ebb22238a1a3456ea90 (commit) via 5c6b971af446d9ffaafdb35da449c78b82b92d8e (commit) from d698823cf1db9bcacb69b1456953c06ec074a460 (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- http://git.mpich.org/mpich.git/commitdiff/cd5986fc1cfaa68fb4abb6c6245c82e31… commit cd5986fc1cfaa68fb4abb6c6245c82e319be77ee Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Wed Jul 24 22:47:45 2013 -0500 Set alloc_shared_noncontig to 1 by default if alloc_shm is passed to MPI_Win_allocate. Because shared memory region is not exposed to the user but just an runtime optimization, it is not useful to allocate a contiguous memory region. If user set alloc_shared_noncontig to "false", alloc_shared_noncontig will be set to 0. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/src/mpid/ch3/channels/nemesis/src/ch3_win_fns.c b/src/mpid/ch3/channels/nemesis/src/ch3_win_fns.c index 6b4fdce..c867138 100644 --- a/src/mpid/ch3/channels/nemesis/src/ch3_win_fns.c +++ b/src/mpid/ch3/channels/nemesis/src/ch3_win_fns.c @@ -62,15 +62,22 @@ static int MPIDI_CH3I_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Inf MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_WIN_ALLOCATE_SHARED); + /* If create flavor is MPI_WIN_FLAVOR_ALLOCATE, alloc_shared_noncontig is set to 1 by default. */ + if ((*win_ptr)->create_flavor == MPI_WIN_FLAVOR_ALLOCATE) + (*win_ptr)->info_args.alloc_shared_noncontig = 1; + /* Check if we are allowed to allocate space non-contiguously */ if (info != NULL) { int alloc_shared_nctg_flag = 0; char alloc_shared_nctg_value[MPI_MAX_INFO_VAL+1]; MPIR_Info_get_impl(info, "alloc_shared_noncontig", MPI_MAX_INFO_VAL, alloc_shared_nctg_value, &alloc_shared_nctg_flag); - if ((alloc_shared_nctg_flag == 1) && (!strncmp(alloc_shared_nctg_value, "true", - strlen("true")))) - (*win_ptr)->info_args.alloc_shared_noncontig = 1; + if ((alloc_shared_nctg_flag == 1)) { + if (!strncmp(alloc_shared_nctg_value, "true", strlen("true"))) + (*win_ptr)->info_args.alloc_shared_noncontig = 1; + if (!strncmp(alloc_shared_nctg_value, "false", strlen("false"))) + (*win_ptr)->info_args.alloc_shared_noncontig = 0; + } } /* see if we can allocate all windows contiguously */ http://git.mpich.org/mpich.git/commitdiff/f038a277715f77c3b746215c6ac527cb1… commit f038a277715f77c3b746215c6ac527cb1ddfebad Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Mon Jul 22 13:02:16 2013 -0500 add alloc_mem test in win_info.c test. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/test/mpi/rma/win_info.c b/test/mpi/rma/win_info.c index 44286a9..3a2222d 100644 --- a/test/mpi/rma/win_info.c +++ b/test/mpi/rma/win_info.c @@ -57,6 +57,9 @@ int main(int argc, char **argv) { MPI_Info_get(info_out, "same_size", MPI_MAX_INFO_VAL, buf, &flag); if (flag && VERBOSE) printf("%d: same_size = %s\n", rank, buf); + MPI_Info_get(info_out, "alloc_shm", MPI_MAX_INFO_VAL, buf, &flag); + if (flag && VERBOSE) printf("%d: alloc_shm = %s\n", rank, buf); + MPI_Info_free(&info_in); MPI_Info_free(&info_out); MPI_Win_free(&win); http://git.mpich.org/mpich.git/commitdiff/f0e27d1d65541d4a69a1152cd318a72e8… commit f0e27d1d65541d4a69a1152cd318a72e8c038bfa Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Mon Jul 22 13:01:49 2013 -0500 add alloc_mem test in req_example.c test. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/test/mpi/rma/Makefile.am b/test/mpi/rma/Makefile.am index 2ccccb0..8c88221 100644 --- a/test/mpi/rma/Makefile.am +++ b/test/mpi/rma/Makefile.am @@ -111,6 +111,7 @@ noinst_PROGRAMS = \ flush \ reqops \ req_example \ + req_example_shm \ win_info \ pscw_ordering \ mutex_bench \ @@ -160,6 +161,9 @@ get_accumulate_long_derived_SOURCES = get_accumulate.c get_accumulate_double_derived_SOURCES = get_accumulate.c get_accumulate_int_derived_SOURCES = get_accumulate.c +req_example_shm_CPPFLAGS = -DUSE_WIN_ALLOC_SHM $(AM_CPPFLAGS) +req_example_shm_SOURCES = req_example.c + mutex_bench_SOURCES = mutex_bench.c mcs-mutex.c mcs-mutex.h mutex_bench_shared_CPPFLAGS = -DUSE_WIN_SHARED $(AM_CPPFLAGS) mutex_bench_shared_SOURCES = mutex_bench.c mcs-mutex.c mcs-mutex.h diff --git a/test/mpi/rma/req_example.c b/test/mpi/rma/req_example.c index 571325c..5788972 100644 --- a/test/mpi/rma/req_example.c +++ b/test/mpi/rma/req_example.c @@ -42,6 +42,7 @@ int main( int argc, char *argv[] ) MPI_Request get_req; double *baseptr; double data[M][N]; /* M buffers of length N */ + MPI_Info win_info; MPI_Init(&argc, &argv); MPI_Comm_rank(MPI_COMM_WORLD, &rank); @@ -49,7 +50,15 @@ int main( int argc, char *argv[] ) assert(M < NSTEPS); - MPI_Win_allocate(NSTEPS*N*sizeof(double), sizeof(double), MPI_INFO_NULL, + MPI_Info_create(&win_info); + +#ifdef USE_WIN_ALLOC_SHM + MPI_Info_set(win_info, "alloc_shm", "true"); +#else + MPI_Info_set(win_info, "alloc_shm", "false"); +#endif + + MPI_Win_allocate(NSTEPS*N*sizeof(double), sizeof(double), win_info, MPI_COMM_WORLD, &baseptr, &win); MPI_Win_lock_all(0, win); @@ -80,6 +89,8 @@ int main( int argc, char *argv[] ) MPI_Win_free(&win); + MPI_Info_free(&win_info); + MPI_Reduce(&errors, &all_errors, 1, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD); if (rank == 0 && all_errors == 0) diff --git a/test/mpi/rma/testlist b/test/mpi/rma/testlist index 522466c..d69d4ed 100644 --- a/test/mpi/rma/testlist +++ b/test/mpi/rma/testlist @@ -91,6 +91,7 @@ get_accumulate_short_derived 4 mpiversion=3.0 flush 4 mpiversion=3.0 reqops 4 mpiversion=3.0 req_example 4 mpiversion=3.0 +req_example_shm 4 mpiversion=3.0 win_info 4 mpiversion=3.0 linked_list_lockall 4 mpiversion=3.0 pscw_ordering 4 mpiversion=3.0 http://git.mpich.org/mpich.git/commitdiff/62cd51ca10890e27bcb13491d876c9d61… commit 62cd51ca10890e27bcb13491d876c9d610b52876 Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Mon Jul 22 13:01:20 2013 -0500 add alloc_shm test in mcs-mutex.c test. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/test/mpi/rma/Makefile.am b/test/mpi/rma/Makefile.am index 1c33a1b..2ccccb0 100644 --- a/test/mpi/rma/Makefile.am +++ b/test/mpi/rma/Makefile.am @@ -115,6 +115,7 @@ noinst_PROGRAMS = \ pscw_ordering \ mutex_bench \ mutex_bench_shared \ + mutex_bench_shm \ rma-contig \ badrma @@ -162,6 +163,8 @@ get_accumulate_int_derived_SOURCES = get_accumulate.c mutex_bench_SOURCES = mutex_bench.c mcs-mutex.c mcs-mutex.h mutex_bench_shared_CPPFLAGS = -DUSE_WIN_SHARED $(AM_CPPFLAGS) mutex_bench_shared_SOURCES = mutex_bench.c mcs-mutex.c mcs-mutex.h +mutex_bench_shm_CPPFLAGS = -DUSE_WIN_ALLOC_SHM $(AM_CPPFLAGS) +mutex_bench_shm_SOURCES = mutex_bench.c mcs-mutex.c mcs-mutex.h linked_list_bench_lock_shr_nocheck_SOURCES = linked_list_bench_lock_shr.c linked_list_bench_lock_shr_nocheck_CPPFLAGS = -DUSE_MODE_NOCHECK $(AM_CPPFLAGS) diff --git a/test/mpi/rma/mcs-mutex.c b/test/mpi/rma/mcs-mutex.c index 533fbdb..835bddf 100644 --- a/test/mpi/rma/mcs-mutex.c +++ b/test/mpi/rma/mcs-mutex.c @@ -42,9 +42,16 @@ int MCS_Mutex_create(int tail_rank, MPI_Comm comm, MCS_Mutex * hdl_out) MPI_Win_allocate_shared(2*sizeof(int), sizeof(int), MPI_INFO_NULL, hdl->comm, &hdl->base, &hdl->window); #else +#ifdef USE_WIN_ALLOC_SHM + MPI_Info_create(&hdl->win_info); + MPI_Info_set(hdl->win_info, "alloc_shm", "true"); + MPI_Win_allocate(2*sizeof(int), sizeof(int), hdl->win_info, hdl->comm, + &hdl->base, &hdl->window); +#else MPI_Win_allocate(2*sizeof(int), sizeof(int), MPI_INFO_NULL, hdl->comm, &hdl->base, &hdl->window); #endif +#endif MPI_Win_lock_all(0, hdl->window); @@ -73,6 +80,9 @@ int MCS_Mutex_free(MCS_Mutex * hdl_ptr) MPI_Win_free(&hdl->window); MPI_Comm_free(&hdl->comm); +#ifdef USE_WIN_ALLOC_SHM + MPI_Info_free(&hdl->win_info); +#endif free(hdl); hdl_ptr = NULL; diff --git a/test/mpi/rma/mcs-mutex.h b/test/mpi/rma/mcs-mutex.h index c8d8843..e67fc5d 100644 --- a/test/mpi/rma/mcs-mutex.h +++ b/test/mpi/rma/mcs-mutex.h @@ -22,6 +22,7 @@ struct mcs_mutex_s { MPI_Comm comm; MPI_Win window; int *base; + MPI_Info win_info; }; typedef struct mcs_mutex_s * MCS_Mutex; diff --git a/test/mpi/rma/testlist b/test/mpi/rma/testlist index 4da3c0f..522466c 100644 --- a/test/mpi/rma/testlist +++ b/test/mpi/rma/testlist @@ -100,5 +100,6 @@ linked_list_bench_lock_shr 4 mpiversion=3.0 linked_list_bench_lock_shr_nocheck 4 mpiversion=3.0 mutex_bench 4 mpiversion=3.0 mutex_bench_shared 4 mpiversion=3.0 +mutex_bench_shm 4 mpiversion=3.0 rma-contig 4 mpiversion=3.0 timeLimit=600 badrma 2 mpiversion=3.0 http://git.mpich.org/mpich.git/commitdiff/5467e5004174421ae581dd08253adfe5c… commit 5467e5004174421ae581dd08253adfe5c86a0000 Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Mon Jul 22 12:59:18 2013 -0500 do SHM_MUTEX_LOCK/UNLOCK in packet handlers on target. If "alloc_shm" is set, it may happen that the target process is doing a RMA operation from a remote process concurrently with a local process is also doing a RMA operation on the same target and on overlapping memory location. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/src/mpid/ch3/src/ch3u_handle_recv_req.c b/src/mpid/ch3/src/ch3u_handle_recv_req.c index 872cfb8..3425aa5 100644 --- a/src/mpid/ch3/src/ch3u_handle_recv_req.c +++ b/src/mpid/ch3/src/ch3u_handle_recv_req.c @@ -142,16 +142,22 @@ int MPIDI_CH3_ReqHandler_PutAccumRespComplete( MPIDI_VC_t *vc, rreq->dev.resp_request_handle = MPI_REQUEST_NULL; } + MPID_Win_get_ptr(rreq->dev.target_win_handle, win_ptr); + if (MPIDI_Request_get_type(rreq) == MPIDI_REQUEST_TYPE_ACCUM_RESP) { + + if (win_ptr->shm_allocated == TRUE) + MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); /* accumulate data from tmp_buf into user_buf */ mpi_errno = do_accumulate_op(rreq); + if (win_ptr->shm_allocated == TRUE) + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } } - MPID_Win_get_ptr(rreq->dev.target_win_handle, win_ptr); - mpi_errno = MPIDI_CH3_Finish_rma_op_target(vc, win_ptr, TRUE, rreq->dev.flags, rreq->dev.source_win_handle); if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } @@ -430,7 +436,11 @@ int MPIDI_CH3_ReqHandler_SinglePutAccumComplete( MPIDI_VC_t *vc, lock_queue_entry->pt_single_op->datatype); } else { + if (win_ptr->shm_allocated == TRUE) + MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); mpi_errno = do_simple_accumulate(lock_queue_entry->pt_single_op); + if (win_ptr->shm_allocated == TRUE) + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); } if (mpi_errno) { @@ -510,12 +520,18 @@ int MPIDI_CH3_ReqHandler_FOPComplete( MPIDI_VC_t *vc, MPIU_Memcpy( resp_req->dev.tmpbuf, rreq->dev.real_user_buf, len ); } + MPID_Win_get_ptr(rreq->dev.target_win_handle, win_ptr); + /* Apply the op */ if (rreq->dev.op != MPI_NO_OP) { uop = MPIR_OP_HDL_TO_FN(rreq->dev.op); one = 1; + if (win_ptr->shm_allocated == TRUE) + MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); (*uop)(rreq->dev.user_buf, rreq->dev.real_user_buf, &one, &rreq->dev.datatype); + if (win_ptr->shm_allocated == TRUE) + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); } /* Send back the original data. We do this here to ensure that the @@ -552,8 +568,6 @@ int MPIDI_CH3_ReqHandler_FOPComplete( MPIDI_VC_t *vc, /* There are additional steps to take if this is a passive target RMA or the last operation from the source */ - MPID_Win_get_ptr(rreq->dev.target_win_handle, win_ptr); - mpi_errno = MPIDI_CH3_Finish_rma_op_target(vc, win_ptr, TRUE, rreq->dev.flags, rreq->dev.source_win_handle); if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } @@ -958,7 +972,11 @@ int MPIDI_CH3I_Release_lock(MPID_Win *win_ptr) single_op->datatype); } else if (single_op->type == MPIDI_RMA_ACCUMULATE) { + if (win_ptr->shm_allocated == TRUE) + MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); mpi_errno = do_simple_accumulate(single_op); + if (win_ptr->shm_allocated == TRUE) + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); } else if (single_op->type == MPIDI_RMA_GET) { mpi_errno = do_simple_get(win_ptr, lock_queue); diff --git a/src/mpid/ch3/src/ch3u_rma_sync.c b/src/mpid/ch3/src/ch3u_rma_sync.c index 4eb0834..df097d0 100644 --- a/src/mpid/ch3/src/ch3u_rma_sync.c +++ b/src/mpid/ch3/src/ch3u_rma_sync.c @@ -3900,6 +3900,8 @@ int MPIDI_CH3_PktHandler_Accumulate_Immed( MPIDI_VC_t *vc, MPIDI_CH3_Pkt_t *pkt, } else { MPIU_INSTR_DURATION_START(rmapkt_acc_immed_op); + if (win_ptr->shm_allocated == TRUE) + MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); /* Data is already present */ if (accum_pkt->op == MPI_REPLACE) { /* no datatypes required */ @@ -3920,6 +3922,8 @@ int MPIDI_CH3_PktHandler_Accumulate_Immed( MPIDI_VC_t *vc, MPIDI_CH3_Pkt_t *pkt, "**opnotpredefined %d", accum_pkt->op ); } } + if (win_ptr->shm_allocated == TRUE) + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); MPIU_INSTR_DURATION_END(rmapkt_acc_immed_op); @@ -3982,8 +3986,20 @@ int MPIDI_CH3_PktHandler_CAS( MPIDI_VC_t *vc, MPIDI_CH3_Pkt_t *pkt, /* Copy old value into the response packet */ MPID_Datatype_get_size_macro(cas_pkt->datatype, len); MPIU_Assert(len <= sizeof(MPIDI_CH3_CAS_Immed_u)); + + if (win_ptr->shm_allocated == TRUE) + MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); + MPIU_Memcpy( (void *)&cas_resp_pkt->data, cas_pkt->addr, len ); + /* Compare and replace if equal */ + if (MPIR_Compare_equal(&cas_pkt->compare_data, cas_pkt->addr, cas_pkt->datatype)) { + MPIU_Memcpy(cas_pkt->addr, &cas_pkt->origin_data, len); + } + + if (win_ptr->shm_allocated == TRUE) + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + /* Send the response packet */ MPIU_THREAD_CS_ENTER(CH3COMM,vc); mpi_errno = MPIDI_CH3_iStartMsg(vc, cas_resp_pkt, sizeof(*cas_resp_pkt), &req); @@ -3995,10 +4011,6 @@ int MPIDI_CH3_PktHandler_CAS( MPIDI_VC_t *vc, MPIDI_CH3_Pkt_t *pkt, MPID_Request_release(req); } - /* Compare and replace if equal */ - if (MPIR_Compare_equal(&cas_pkt->compare_data, cas_pkt->addr, cas_pkt->datatype)) { - MPIU_Memcpy(cas_pkt->addr, &cas_pkt->origin_data, len); - } /* There are additional steps to take if this is a passive target RMA or the last operation from the source */ http://git.mpich.org/mpich.git/commitdiff/447c1d5585135a41fd6341dcb9760ba94… commit 447c1d5585135a41fd6341dcb9760ba945261f11 Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Fri Jul 26 19:41:01 2013 -0500 Delete decrementing ref count in SHM RMA operations, but add conditions in operaiton issue routines. In RMA operation issue routines, judge if shm_allocate == 1 and target vc is local, if so, do not add reference count on datatypes, because they will not be referenced by the progress engine, but will be completed directly by origin. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/src/mpid/ch3/src/ch3u_rma_acc_ops.c b/src/mpid/ch3/src/ch3u_rma_acc_ops.c index 82b1d2d..0253095 100644 --- a/src/mpid/ch3/src/ch3u_rma_acc_ops.c +++ b/src/mpid/ch3/src/ch3u_rma_acc_ops.c @@ -71,6 +71,7 @@ int MPIDI_Get_accumulate(const void *origin_addr, int origin_count, else { MPIDI_RMA_Ops_list_t *ops_list = MPIDI_CH3I_RMA_Get_ops_list(win_ptr, target_rank); MPIDI_RMA_Op_t *new_ptr = NULL; + MPIDI_VC_t *vc = NULL; /* Append the operation to the window's RMA ops queue */ MPIU_INSTR_DURATION_START(rmaqueue_alloc); @@ -97,19 +98,27 @@ int MPIDI_Get_accumulate(const void *origin_addr, int origin_count, new_ptr->op = op; MPIU_INSTR_DURATION_END(rmaqueue_set); - /* if source or target datatypes are derived, increment their - reference counts */ - if (!origin_predefined) { - MPID_Datatype_get_ptr(origin_datatype, dtp); - MPID_Datatype_add_ref(dtp); - } - if (!result_predefined) { - MPID_Datatype_get_ptr(result_datatype, dtp); - MPID_Datatype_add_ref(dtp); - } - if (!target_predefined) { - MPID_Datatype_get_ptr(target_datatype, dtp); - MPID_Datatype_add_ref(dtp); + /* check if target is local and shared memory is allocated on window, + if so, we do not need to increment reference counts on datatype. This is + because this operation will be directly done on shared memory region, instead + of sending and receiving through the progress engine, therefore datatype + will not be referenced by the progress engine */ + MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &vc); + if (!(win_ptr->shm_allocated == TRUE && vc->ch.is_local)) { + /* if source or target datatypes are derived, increment their + reference counts */ + if (!origin_predefined) { + MPID_Datatype_get_ptr(origin_datatype, dtp); + MPID_Datatype_add_ref(dtp); + } + if (!result_predefined) { + MPID_Datatype_get_ptr(result_datatype, dtp); + MPID_Datatype_add_ref(dtp); + } + if (!target_predefined) { + MPID_Datatype_get_ptr(target_datatype, dtp); + MPID_Datatype_add_ref(dtp); + } } } diff --git a/src/mpid/ch3/src/ch3u_rma_ops.c b/src/mpid/ch3/src/ch3u_rma_ops.c index b57283c..b7d6d49 100644 --- a/src/mpid/ch3/src/ch3u_rma_ops.c +++ b/src/mpid/ch3/src/ch3u_rma_ops.c @@ -190,6 +190,7 @@ int MPIDI_Put(const void *origin_addr, int origin_count, MPI_Datatype { MPIDI_RMA_Ops_list_t *ops_list = MPIDI_CH3I_RMA_Get_ops_list(win_ptr, target_rank); MPIDI_RMA_Op_t *new_ptr = NULL; + MPIDI_VC_t *vc = NULL; /* queue it up */ MPIU_INSTR_DURATION_START(rmaqueue_alloc); @@ -212,20 +213,28 @@ int MPIDI_Put(const void *origin_addr, int origin_count, MPI_Datatype new_ptr->target_datatype = target_datatype; MPIU_INSTR_DURATION_END(rmaqueue_set); - /* if source or target datatypes are derived, increment their - reference counts */ - MPIDI_CH3I_DATATYPE_IS_PREDEFINED(origin_datatype, predefined); - if (!predefined) - { - MPID_Datatype_get_ptr(origin_datatype, dtp); - MPID_Datatype_add_ref(dtp); - } - MPIDI_CH3I_DATATYPE_IS_PREDEFINED(target_datatype, predefined); - if (!predefined) - { - MPID_Datatype_get_ptr(target_datatype, dtp); - MPID_Datatype_add_ref(dtp); - } + /* check if target is local and shared memory is allocated on window, + if so, we do not need to increment reference counts on datatype. This is + because this operation will be directly done on shared memory region, instead + of sending and receiving through the progress engine, therefore datatype + will not be referenced by the progress engine */ + MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &vc); + if (!(win_ptr->shm_allocated == TRUE && vc->ch.is_local)) { + /* if source or target datatypes are derived, increment their + reference counts */ + MPIDI_CH3I_DATATYPE_IS_PREDEFINED(origin_datatype, predefined); + if (!predefined) + { + MPID_Datatype_get_ptr(origin_datatype, dtp); + MPID_Datatype_add_ref(dtp); + } + MPIDI_CH3I_DATATYPE_IS_PREDEFINED(target_datatype, predefined); + if (!predefined) + { + MPID_Datatype_get_ptr(target_datatype, dtp); + MPID_Datatype_add_ref(dtp); + } + } } fn_exit: @@ -288,6 +297,7 @@ int MPIDI_Get(void *origin_addr, int origin_count, MPI_Datatype { MPIDI_RMA_Ops_list_t *ops_list = MPIDI_CH3I_RMA_Get_ops_list(win_ptr, target_rank); MPIDI_RMA_Op_t *new_ptr = NULL; + MPIDI_VC_t *vc = NULL; /* queue it up */ MPIU_INSTR_DURATION_START(rmaqueue_alloc); @@ -307,20 +317,28 @@ int MPIDI_Get(void *origin_addr, int origin_count, MPI_Datatype new_ptr->target_datatype = target_datatype; MPIU_INSTR_DURATION_END(rmaqueue_set); - /* if source or target datatypes are derived, increment their - reference counts */ - MPIDI_CH3I_DATATYPE_IS_PREDEFINED(origin_datatype, predefined); - if (!predefined) - { - MPID_Datatype_get_ptr(origin_datatype, dtp); - MPID_Datatype_add_ref(dtp); - } - MPIDI_CH3I_DATATYPE_IS_PREDEFINED(target_datatype, predefined); - if (!predefined) - { - MPID_Datatype_get_ptr(target_datatype, dtp); - MPID_Datatype_add_ref(dtp); - } + /* check if target is local and shared memory is allocated on window, + if so, we do not need to increment reference counts on datatype. This is + because this operation will be directly done on shared memory region, instead + of sending and receiving through the progress engine, therefore datatype + will not be referenced by the progress engine */ + MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &vc); + if (!(win_ptr->shm_allocated == TRUE && vc->ch.is_local)) { + /* if source or target datatypes are derived, increment their + reference counts */ + MPIDI_CH3I_DATATYPE_IS_PREDEFINED(origin_datatype, predefined); + if (!predefined) + { + MPID_Datatype_get_ptr(origin_datatype, dtp); + MPID_Datatype_add_ref(dtp); + } + MPIDI_CH3I_DATATYPE_IS_PREDEFINED(target_datatype, predefined); + if (!predefined) + { + MPID_Datatype_get_ptr(target_datatype, dtp); + MPID_Datatype_add_ref(dtp); + } + } } fn_exit: @@ -389,6 +407,7 @@ int MPIDI_Accumulate(const void *origin_addr, int origin_count, MPI_Datatype { MPIDI_RMA_Ops_list_t *ops_list = MPIDI_CH3I_RMA_Get_ops_list(win_ptr, target_rank); MPIDI_RMA_Op_t *new_ptr = NULL; + MPIDI_VC_t *vc = NULL; /* queue it up */ MPIU_INSTR_DURATION_START(rmaqueue_alloc); @@ -429,18 +448,26 @@ int MPIDI_Accumulate(const void *origin_addr, int origin_count, MPI_Datatype new_ptr->op = op; MPIU_INSTR_DURATION_END(rmaqueue_set); - /* if source or target datatypes are derived, increment their - reference counts */ - if (!origin_predefined) - { - MPID_Datatype_get_ptr(origin_datatype, dtp); - MPID_Datatype_add_ref(dtp); - } - if (!target_predefined) - { - MPID_Datatype_get_ptr(target_datatype, dtp); - MPID_Datatype_add_ref(dtp); - } + /* check if target is local and shared memory is allocated on window, + if so, we do not need to increment reference counts on datatype. This is + because this operation will be directly done on shared memory region, instead + of sending and receiving through the progress engine, therefore datatype + will not be referenced by the progress engine */ + MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &vc); + if (!(win_ptr->shm_allocated == TRUE && vc->ch.is_local)) { + /* if source or target datatypes are derived, increment their + reference counts */ + if (!origin_predefined) + { + MPID_Datatype_get_ptr(origin_datatype, dtp); + MPID_Datatype_add_ref(dtp); + } + if (!target_predefined) + { + MPID_Datatype_get_ptr(target_datatype, dtp); + MPID_Datatype_add_ref(dtp); + } + } } fn_exit: diff --git a/src/mpid/ch3/src/ch3u_rma_sync.c b/src/mpid/ch3/src/ch3u_rma_sync.c index 4852465..4eb0834 100644 --- a/src/mpid/ch3/src/ch3u_rma_sync.c +++ b/src/mpid/ch3/src/ch3u_rma_sync.c @@ -208,27 +208,11 @@ static int create_datatype(const MPIDI_RMA_dtype_info *dtype_info, do { \ switch ((op_ptr_)->type) \ { \ - int predefined_; \ - MPID_Datatype *dtp_; \ case (MPIDI_RMA_PUT): \ (err_) = MPIDI_CH3I_Shm_put_op((op_ptr_)->origin_addr, (op_ptr_)->origin_count, (op_ptr_)->origin_datatype, \ (op_ptr_)->target_rank, (op_ptr_)->target_disp, (op_ptr_)->target_count, \ (op_ptr_)->target_datatype, (win_ptr_)); \ if (err_) {MPIU_ERR_POP(err_);} \ - /* if source or target datatypes are derived, decrement their \ - reference counts because they will not be referenced by progress engine */ \ - MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->origin_datatype, predefined_); \ - if (!predefined_) \ - { \ - MPID_Datatype_get_ptr((op_ptr_)->origin_datatype, dtp_); \ - MPID_Datatype_release(dtp_); \ - } \ - MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->target_datatype, predefined_); \ - if (!predefined_) \ - { \ - MPID_Datatype_get_ptr((op_ptr_)->target_datatype, dtp_); \ - MPID_Datatype_release(dtp_); \ - } \ break; \ case (MPIDI_RMA_ACCUMULATE): \ case (MPIDI_RMA_ACC_CONTIG): \ @@ -236,20 +220,6 @@ static int create_datatype(const MPIDI_RMA_dtype_info *dtype_info, (op_ptr_)->target_rank, (op_ptr_)->target_disp, (op_ptr_)->target_count, \ (op_ptr_)->target_datatype, (op_ptr_)->op, (win_ptr_)); \ if (err_) {MPIU_ERR_POP(err_);} \ - /* if source or target datatypes are derived, decrement their \ - reference counts because they will not be referenced by progress engine */ \ - MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->origin_datatype, predefined_); \ - if (!predefined_) \ - { \ - MPID_Datatype_get_ptr((op_ptr_)->origin_datatype, dtp_); \ - MPID_Datatype_release(dtp_); \ - } \ - MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->target_datatype, predefined_); \ - if (!predefined_) \ - { \ - MPID_Datatype_get_ptr((op_ptr_)->target_datatype, dtp_); \ - MPID_Datatype_release(dtp_); \ - } \ break; \ case (MPIDI_RMA_GET_ACCUMULATE): \ (err_) = MPIDI_CH3I_Shm_get_acc_op((op_ptr_)->origin_addr, (op_ptr_)->origin_count, (op_ptr_)->origin_datatype, \ @@ -257,49 +227,12 @@ static int create_datatype(const MPIDI_RMA_dtype_info *dtype_info, (op_ptr_)->target_rank, (op_ptr_)->target_disp, (op_ptr_)->target_count, \ (op_ptr_)->target_datatype, (op_ptr_)->op, (win_ptr_)); \ if (err_) {MPIU_ERR_POP(err_);} \ - /* if source or target datatypes are derived, decrement their \ - reference counts because they will not be referenced by progress engine */ \ - predefined_ = TRUE; \ - if ((op_ptr_)->op != MPI_NO_OP) { \ - MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->origin_datatype, predefined_); \ - } \ - if (!predefined_) \ - { \ - MPID_Datatype_get_ptr((op_ptr_)->origin_datatype, dtp_); \ - MPID_Datatype_release(dtp_); \ - } \ - MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->result_datatype, predefined_); \ - if (!predefined_) \ - { \ - MPID_Datatype_get_ptr((op_ptr_)->result_datatype, dtp_); \ - MPID_Datatype_release(dtp_); \ - } \ - MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->target_datatype, predefined_); \ - if (!predefined_) \ - { \ - MPID_Datatype_get_ptr((op_ptr_)->target_datatype, dtp_); \ - MPID_Datatype_release(dtp_); \ - } \ break; \ case (MPIDI_RMA_GET): \ (err_) = MPIDI_CH3I_Shm_get_op((op_ptr_)->origin_addr, (op_ptr_)->origin_count, (op_ptr_)->origin_datatype, \ (op_ptr_)->target_rank, (op_ptr_)->target_disp, (op_ptr_)->target_count, \ (op_ptr_)->target_datatype, (win_ptr_)); \ if (err_) {MPIU_ERR_POP(err_);} \ - /* if source or target datatypes are derived, decrement their \ - reference counts because they will not be referenced by progress engine */ \ - MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->origin_datatype, predefined_); \ - if (!predefined_) \ - { \ - MPID_Datatype_get_ptr((op_ptr_)->origin_datatype, dtp_); \ - MPID_Datatype_release(dtp_); \ - } \ - MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->target_datatype, predefined_); \ - if (!predefined_) \ - { \ - MPID_Datatype_get_ptr((op_ptr_)->target_datatype, dtp_); \ - MPID_Datatype_release(dtp_); \ - } \ break; \ case (MPIDI_RMA_COMPARE_AND_SWAP): \ (err_) = MPIDI_CH3I_Shm_cas_op((op_ptr_)->origin_addr, (op_ptr_)->compare_addr, (op_ptr_)->result_addr, \ http://git.mpich.org/mpich.git/commitdiff/34dd7c1ad295d223119dbad3d032e38fa… commit 34dd7c1ad295d223119dbad3d032e38fa7f94c8e Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Fri Jul 26 09:10:32 2013 -0500 In SHM RMA operations, add condition for "alloc_shm". If shared memory is allocated for window, and target vc is local, do SHM RMA operations. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/src/mpid/ch3/include/mpidrma.h b/src/mpid/ch3/include/mpidrma.h index ecb59c3..3221c6a 100644 --- a/src/mpid/ch3/include/mpidrma.h +++ b/src/mpid/ch3/include/mpidrma.h @@ -325,13 +325,17 @@ static inline int MPIDI_CH3I_Shm_put_op(const void *origin_addr, int origin_coun MPID_Win *win_ptr) { int mpi_errno = MPI_SUCCESS; + MPIDI_VC_t *vc = NULL; void *base = NULL; int disp_unit; MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SHM_PUT_OP); MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_SHM_PUT_OP); - if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { + MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &vc); + + if ((win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) || + (win_ptr->shm_allocated == TRUE && vc->ch.is_local)) { base = win_ptr->shm_base_addrs[target_rank]; disp_unit = win_ptr->disp_units[target_rank]; } @@ -369,6 +373,7 @@ static inline int MPIDI_CH3I_Shm_acc_op(const void *origin_addr, int origin_coun int origin_predefined, target_predefined; MPI_User_function *uop = NULL; MPID_Datatype *dtp; + MPIDI_VC_t *vc = NULL; int mpi_errno = MPI_SUCCESS; MPIU_CHKLMEM_DECL(2); MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SHM_ACC_OP); @@ -378,7 +383,10 @@ static inline int MPIDI_CH3I_Shm_acc_op(const void *origin_addr, int origin_coun MPIDI_CH3I_DATATYPE_IS_PREDEFINED(origin_datatype, origin_predefined); MPIDI_CH3I_DATATYPE_IS_PREDEFINED(target_datatype, target_predefined); - if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { + MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &vc); + + if ((win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) || + (win_ptr->shm_allocated == TRUE && vc->ch.is_local)) { shm_op = 1; base = win_ptr->shm_base_addrs[target_rank]; disp_unit = win_ptr->disp_units[target_rank]; @@ -521,6 +529,7 @@ static inline int MPIDI_CH3I_Shm_get_acc_op(const void *origin_addr, int origin_ MPI_User_function *uop = NULL; MPID_Datatype *dtp; int origin_predefined, result_predefined, target_predefined; + MPIDI_VC_t *vc = NULL; int mpi_errno = MPI_SUCCESS; MPIU_CHKLMEM_DECL(2); MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SHM_GET_ACC_OP); @@ -534,7 +543,10 @@ static inline int MPIDI_CH3I_Shm_get_acc_op(const void *origin_addr, int origin_ MPIDI_CH3I_DATATYPE_IS_PREDEFINED(result_datatype, result_predefined); MPIDI_CH3I_DATATYPE_IS_PREDEFINED(target_datatype, target_predefined); - if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { + MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &vc); + + if ((win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) || + (win_ptr->shm_allocated == TRUE && vc->ch.is_local)) { base = win_ptr->shm_base_addrs[target_rank]; disp_unit = win_ptr->disp_units[target_rank]; MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); @@ -690,12 +702,16 @@ static inline int MPIDI_CH3I_Shm_get_op(void *origin_addr, int origin_count, MPI { void *base = NULL; int disp_unit; + MPIDI_VC_t *vc = NULL; int mpi_errno = MPI_SUCCESS; MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SHM_GET_OP); MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_SHM_GET_OP); - if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { + MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &vc); + + if ((win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) || + (win_ptr->shm_allocated == TRUE && vc->ch.is_local)) { base = win_ptr->shm_base_addrs[target_rank]; disp_unit = win_ptr->disp_units[target_rank]; } @@ -730,12 +746,16 @@ static inline int MPIDI_CH3I_Shm_cas_op(const void *origin_addr, const void *com void *base = NULL, *dest_addr = NULL; int disp_unit; int len, shm_locked = 0; + MPIDI_VC_t *vc = NULL; int mpi_errno = MPI_SUCCESS; MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SHM_CAS_OP); MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_SHM_CAS_OP); - if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { + MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &vc); + + if ((win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) || + (win_ptr->shm_allocated == TRUE && vc->ch.is_local)) { base = win_ptr->shm_base_addrs[target_rank]; disp_unit = win_ptr->disp_units[target_rank]; @@ -786,12 +806,16 @@ static inline int MPIDI_CH3I_Shm_fop_op(const void *origin_addr, void *result_ad MPI_User_function *uop = NULL; int disp_unit; int len, one, shm_locked = 0; + MPIDI_VC_t *vc = NULL; int mpi_errno = MPI_SUCCESS; MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SHM_FOP_OP); MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_SHM_FOP_OP); - if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { + MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &vc); + + if ((win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) || + (win_ptr->shm_allocated == TRUE && vc->ch.is_local)) { base = win_ptr->shm_base_addrs[target_rank]; disp_unit = win_ptr->disp_units[target_rank]; http://git.mpich.org/mpich.git/commitdiff/55cb0f4c1cdc5c1e7017d487bc2f13871… commit 55cb0f4c1cdc5c1e7017d487bc2f13871dcf766e Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Fri Jul 26 08:56:28 2013 -0500 Call DO_SHM_RMA_OP in fence/unlock/complete. decrement datatype reference counts in SHM RMA operations. because they will not be referenced by the progress engine, but be completed directly by origin. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/src/mpid/ch3/src/ch3u_rma_sync.c b/src/mpid/ch3/src/ch3u_rma_sync.c index f05a5bb..4852465 100644 --- a/src/mpid/ch3/src/ch3u_rma_sync.c +++ b/src/mpid/ch3/src/ch3u_rma_sync.c @@ -202,6 +202,122 @@ static int create_datatype(const MPIDI_RMA_dtype_info *dtype_info, } while (0) +/* Perform RMA operation asynchronously if window of processes on the same node + is allocated on shared memory */ +#define MPIDI_CH3I_DO_SHM_OP(op_ptr_, win_ptr_, err_) \ + do { \ + switch ((op_ptr_)->type) \ + { \ + int predefined_; \ + MPID_Datatype *dtp_; \ + case (MPIDI_RMA_PUT): \ + (err_) = MPIDI_CH3I_Shm_put_op((op_ptr_)->origin_addr, (op_ptr_)->origin_count, (op_ptr_)->origin_datatype, \ + (op_ptr_)->target_rank, (op_ptr_)->target_disp, (op_ptr_)->target_count, \ + (op_ptr_)->target_datatype, (win_ptr_)); \ + if (err_) {MPIU_ERR_POP(err_);} \ + /* if source or target datatypes are derived, decrement their \ + reference counts because they will not be referenced by progress engine */ \ + MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->origin_datatype, predefined_); \ + if (!predefined_) \ + { \ + MPID_Datatype_get_ptr((op_ptr_)->origin_datatype, dtp_); \ + MPID_Datatype_release(dtp_); \ + } \ + MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->target_datatype, predefined_); \ + if (!predefined_) \ + { \ + MPID_Datatype_get_ptr((op_ptr_)->target_datatype, dtp_); \ + MPID_Datatype_release(dtp_); \ + } \ + break; \ + case (MPIDI_RMA_ACCUMULATE): \ + case (MPIDI_RMA_ACC_CONTIG): \ + (err_) = MPIDI_CH3I_Shm_acc_op((op_ptr_)->origin_addr, (op_ptr_)->origin_count, (op_ptr_)->origin_datatype, \ + (op_ptr_)->target_rank, (op_ptr_)->target_disp, (op_ptr_)->target_count, \ + (op_ptr_)->target_datatype, (op_ptr_)->op, (win_ptr_)); \ + if (err_) {MPIU_ERR_POP(err_);} \ + /* if source or target datatypes are derived, decrement their \ + reference counts because they will not be referenced by progress engine */ \ + MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->origin_datatype, predefined_); \ + if (!predefined_) \ + { \ + MPID_Datatype_get_ptr((op_ptr_)->origin_datatype, dtp_); \ + MPID_Datatype_release(dtp_); \ + } \ + MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->target_datatype, predefined_); \ + if (!predefined_) \ + { \ + MPID_Datatype_get_ptr((op_ptr_)->target_datatype, dtp_); \ + MPID_Datatype_release(dtp_); \ + } \ + break; \ + case (MPIDI_RMA_GET_ACCUMULATE): \ + (err_) = MPIDI_CH3I_Shm_get_acc_op((op_ptr_)->origin_addr, (op_ptr_)->origin_count, (op_ptr_)->origin_datatype, \ + (op_ptr_)->result_addr, (op_ptr_)->result_count, (op_ptr_)->result_datatype, \ + (op_ptr_)->target_rank, (op_ptr_)->target_disp, (op_ptr_)->target_count, \ + (op_ptr_)->target_datatype, (op_ptr_)->op, (win_ptr_)); \ + if (err_) {MPIU_ERR_POP(err_);} \ + /* if source or target datatypes are derived, decrement their \ + reference counts because they will not be referenced by progress engine */ \ + predefined_ = TRUE; \ + if ((op_ptr_)->op != MPI_NO_OP) { \ + MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->origin_datatype, predefined_); \ + } \ + if (!predefined_) \ + { \ + MPID_Datatype_get_ptr((op_ptr_)->origin_datatype, dtp_); \ + MPID_Datatype_release(dtp_); \ + } \ + MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->result_datatype, predefined_); \ + if (!predefined_) \ + { \ + MPID_Datatype_get_ptr((op_ptr_)->result_datatype, dtp_); \ + MPID_Datatype_release(dtp_); \ + } \ + MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->target_datatype, predefined_); \ + if (!predefined_) \ + { \ + MPID_Datatype_get_ptr((op_ptr_)->target_datatype, dtp_); \ + MPID_Datatype_release(dtp_); \ + } \ + break; \ + case (MPIDI_RMA_GET): \ + (err_) = MPIDI_CH3I_Shm_get_op((op_ptr_)->origin_addr, (op_ptr_)->origin_count, (op_ptr_)->origin_datatype, \ + (op_ptr_)->target_rank, (op_ptr_)->target_disp, (op_ptr_)->target_count, \ + (op_ptr_)->target_datatype, (win_ptr_)); \ + if (err_) {MPIU_ERR_POP(err_);} \ + /* if source or target datatypes are derived, decrement their \ + reference counts because they will not be referenced by progress engine */ \ + MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->origin_datatype, predefined_); \ + if (!predefined_) \ + { \ + MPID_Datatype_get_ptr((op_ptr_)->origin_datatype, dtp_); \ + MPID_Datatype_release(dtp_); \ + } \ + MPIDI_CH3I_DATATYPE_IS_PREDEFINED((op_ptr_)->target_datatype, predefined_); \ + if (!predefined_) \ + { \ + MPID_Datatype_get_ptr((op_ptr_)->target_datatype, dtp_); \ + MPID_Datatype_release(dtp_); \ + } \ + break; \ + case (MPIDI_RMA_COMPARE_AND_SWAP): \ + (err_) = MPIDI_CH3I_Shm_cas_op((op_ptr_)->origin_addr, (op_ptr_)->compare_addr, (op_ptr_)->result_addr, \ + (op_ptr_)->origin_datatype, (op_ptr_)->target_rank, (op_ptr_)->target_disp, \ + (win_ptr_)); \ + if (err_) {MPIU_ERR_POP(err_);} \ + break; \ + case (MPIDI_RMA_FETCH_AND_OP): \ + (err_) = MPIDI_CH3I_Shm_fop_op((op_ptr_)->origin_addr, (op_ptr_)->result_addr, (op_ptr_)->origin_datatype, \ + (op_ptr_)->target_rank, (op_ptr_)->target_disp, (op_ptr_)->op, (win_ptr_)); \ + if (err_) {MPIU_ERR_POP(err_);} \ + break; \ + default: \ + MPIU_ERR_SETANDJUMP(err_,MPI_ERR_OTHER,"**winInvalidOp"); \ + } \ + } while (0) + + #undef FUNCNAME #define FUNCNAME MPIDI_Win_fence #undef FCNAME @@ -309,9 +425,13 @@ int MPIDI_Win_fence(int assert, MPID_Win *win_ptr) curr_ptr = MPIDI_CH3I_RMA_Ops_head(ops_list); while (curr_ptr != NULL) { - total_op_count++; - rma_target_proc[curr_ptr->target_rank] = 1; - nops_to_proc[curr_ptr->target_rank]++; + MPIDI_VC_t *vc = NULL; + MPIDI_Comm_get_vc(win_ptr->comm_ptr, curr_ptr->target_rank, &vc); + if (!(win_ptr->shm_allocated == TRUE && vc->ch.is_local)) { + total_op_count++; + rma_target_proc[curr_ptr->target_rank] = 1; + nops_to_proc[curr_ptr->target_rank]++; + } curr_ptr = curr_ptr->next; } @@ -347,6 +467,14 @@ int MPIDI_Win_fence(int assert, MPID_Win *win_ptr) curr_ptr = MPIDI_CH3I_RMA_Ops_head(ops_list); while (curr_ptr != NULL) { + MPIDI_VC_t *vc = NULL; + MPIDI_Comm_get_vc(win_ptr->comm_ptr, curr_ptr->target_rank, &vc); + if (win_ptr->shm_allocated == TRUE && vc->ch.is_local) { + MPIDI_CH3I_DO_SHM_OP(curr_ptr, win_ptr, mpi_errno); + MPIDI_CH3I_RMA_Ops_free_and_next(ops_list, &curr_ptr); + } + else { + MPIDI_CH3_Pkt_flags_t flags = MPIDI_CH3_PKT_FLAG_NONE; /* The completion counter at the target is decremented only on @@ -388,6 +516,7 @@ int MPIDI_Win_fence(int assert, MPID_Win *win_ptr) nRequestNew = nRequest; } } + } /* end of else */ } MPIU_INSTR_DURATION_END(winfence_issue); @@ -1628,8 +1757,12 @@ int MPIDI_Win_complete(MPID_Win *win_ptr) curr_ptr = MPIDI_CH3I_RMA_Ops_head(ops_list); while (curr_ptr != NULL) { - nops_to_proc[curr_ptr->target_rank]++; - total_op_count++; + MPIDI_VC_t *vc = NULL; + MPIDI_Comm_get_vc(win_ptr->comm_ptr, curr_ptr->target_rank, &vc); + if (!(win_ptr->shm_allocated == TRUE && vc->ch.is_local)) { + nops_to_proc[curr_ptr->target_rank]++; + total_op_count++; + } curr_ptr = curr_ptr->next; } @@ -1648,6 +1781,14 @@ int MPIDI_Win_complete(MPID_Win *win_ptr) curr_ptr = MPIDI_CH3I_RMA_Ops_head(ops_list); while (curr_ptr != NULL) { + MPIDI_VC_t *vc = NULL; + MPIDI_Comm_get_vc(win_ptr->comm_ptr, curr_ptr->target_rank, &vc); + if (win_ptr->shm_allocated == TRUE && vc->ch.is_local) { + MPIDI_CH3I_DO_SHM_OP(curr_ptr, win_ptr, mpi_errno); + MPIDI_CH3I_RMA_Ops_free_and_next(ops_list, &curr_ptr); + } + else { + MPIDI_CH3_Pkt_flags_t flags = MPIDI_CH3_PKT_FLAG_NONE; /* The completion counter at the target is decremented only on @@ -1683,6 +1824,7 @@ int MPIDI_Win_complete(MPID_Win *win_ptr) nRequestNew = nRequest; } } + } /* end of else */ } MPIU_INSTR_DURATION_END(wincomplete_issue); @@ -2509,6 +2651,7 @@ static int MPIDI_CH3I_Do_passive_target_rma(MPID_Win *win_ptr, int target_rank, MPIDI_RMA_Op_t *curr_ptr; MPI_Win source_win_handle = MPI_WIN_NULL, target_win_handle = MPI_WIN_NULL; int nRequest=0, nRequestNew=0; + MPIDI_VC_t *vc = NULL; MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_DO_PASSIVE_TARGET_RMA); MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_DO_PASSIVE_TARGET_RMA); @@ -2518,6 +2661,20 @@ static int MPIDI_CH3I_Do_passive_target_rma(MPID_Win *win_ptr, int target_rank, (win_ptr->targets[target_rank].remote_lock_state == MPIDI_CH3_WIN_LOCK_CALLED && win_ptr->targets[target_rank].remote_lock_assert & MPI_MODE_NOCHECK)); + /* if alloc_shm is enabled and target process is on the same node, + directly perform RMA operations at the origin side and remove them + from passive RMA operation list */ + + MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &vc); + if (win_ptr->shm_allocated == TRUE && vc->ch.is_local) { + curr_ptr = MPIDI_CH3I_RMA_Ops_head(&win_ptr->targets[target_rank].rma_ops_list); + while (curr_ptr != NULL) { + MPIU_Assert(curr_ptr->target_rank == target_rank); + MPIDI_CH3I_DO_SHM_OP(curr_ptr, win_ptr, mpi_errno); + MPIDI_CH3I_RMA_Ops_free_and_next(&win_ptr->targets[target_rank].rma_ops_list, &curr_ptr); + } + } + if (win_ptr->targets[target_rank].remote_lock_mode == MPI_LOCK_EXCLUSIVE && win_ptr->targets[target_rank].remote_lock_state != MPIDI_CH3_WIN_LOCK_CALLED) { /* Exclusive lock already held -- no need to wait for rma done pkt at http://git.mpich.org/mpich.git/commitdiff/b668bc1ed3168790d8361ff14e1976581… commit b668bc1ed3168790d8361ff14e1976581349cb81 Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Mon Jul 22 10:18:29 2013 -0500 add "alloc_shm" info to MPIDI_Win_get_info. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/src/mpid/ch3/src/ch3u_win_fns.c b/src/mpid/ch3/src/ch3u_win_fns.c index b44bae2..ef7d84c 100644 --- a/src/mpid/ch3/src/ch3u_win_fns.c +++ b/src/mpid/ch3/src/ch3u_win_fns.c @@ -382,6 +382,13 @@ int MPIDI_Win_get_info(MPID_Win *win, MPID_Info **info_used) if (mpi_errno != MPI_SUCCESS) { MPIU_ERR_POP(mpi_errno); } + if (win->info_args.alloc_shm == TRUE) + mpi_errno = MPIR_Info_set_impl(*info_used, "alloc_shm", "true"); + else + mpi_errno = MPIR_Info_set_impl(*info_used, "alloc_shm", "false"); + + if (mpi_errno != MPI_SUCCESS) { MPIU_ERR_POP(mpi_errno); } + if (win->create_flavor == MPI_WIN_FLAVOR_SHARED) { if (win->info_args.alloc_shared_noncontig) mpi_errno = MPIR_Info_set_impl(*info_used, "alloc_shared_noncontig", "true"); http://git.mpich.org/mpich.git/commitdiff/384d96b7b7f193963a178e67ebb64d750… commit 384d96b7b7f193963a178e67ebb64d750751f3cc Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Mon Jul 22 10:17:30 2013 -0500 Add "alloc_shm" info to MPI_Win_allocate. Add "alloc_shm" to window's info arguments and initialize it to FALSE. In MPID_Win_allocate, if "alloc_shm" is set to true, call ALLOCATE_SHARED, otherwise call ALLOCATE. Free window memory only when SHM region is not allocated, therwise it is already freed in MPIDI_CH3I_SHM_Win_free. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/src/mpid/ch3/include/mpidpre.h b/src/mpid/ch3/include/mpidpre.h index 35ddac4..8210c97 100644 --- a/src/mpid/ch3/include/mpidpre.h +++ b/src/mpid/ch3/include/mpidpre.h @@ -244,6 +244,7 @@ struct MPIDI_Win_info_args { int accumulate_ops; int same_size; /* valid flavor = allocate */ int alloc_shared_noncontig; /* valid flavor = allocate shared */ + int alloc_shm; /* valid flavor = allocate */ }; struct MPIDI_RMA_op; /* forward decl from mpidrma.h */ diff --git a/src/mpid/ch3/src/ch3u_rma_ops.c b/src/mpid/ch3/src/ch3u_rma_ops.c index 5f6fb4c..b57283c 100644 --- a/src/mpid/ch3/src/ch3u_rma_ops.c +++ b/src/mpid/ch3/src/ch3u_rma_ops.c @@ -67,7 +67,8 @@ int MPIDI_Win_free(MPID_Win **win_ptr) /* Free the attached buffer for windows created with MPI_Win_allocate() */ if ((*win_ptr)->create_flavor == MPI_WIN_FLAVOR_ALLOCATE && (*win_ptr)->size > 0) { - MPIU_Free((*win_ptr)->base); + if ((*win_ptr)->shm_allocated != TRUE) + MPIU_Free((*win_ptr)->base); } MPIU_Object_release_ref(*win_ptr, &in_use); diff --git a/src/mpid/ch3/src/mpid_rma.c b/src/mpid/ch3/src/mpid_rma.c index 48c27e8..d330933 100644 --- a/src/mpid/ch3/src/mpid_rma.c +++ b/src/mpid/ch3/src/mpid_rma.c @@ -109,8 +109,22 @@ int MPID_Win_allocate(MPI_Aint size, int disp_unit, MPID_Info *info, mpi_errno = win_init(size, disp_unit, MPI_WIN_FLAVOR_ALLOCATE, MPI_WIN_UNIFIED, comm_ptr, win_ptr); if (mpi_errno != MPI_SUCCESS) { MPIU_ERR_POP(mpi_errno); } - mpi_errno = MPIDI_CH3U_Win_fns.allocate(size, disp_unit, info, comm_ptr, baseptr, win_ptr); - if (mpi_errno != MPI_SUCCESS) { MPIU_ERR_POP(mpi_errno); } + if (info != NULL) { + int alloc_shm_flag = 0; + char shm_alloc_value[MPI_MAX_INFO_VAL+1]; + MPIR_Info_get_impl(info, "alloc_shm", MPI_MAX_INFO_VAL, shm_alloc_value, &alloc_shm_flag); + if ((alloc_shm_flag == 1) && (!strncmp(shm_alloc_value, "true", sizeof("true")))) + (*win_ptr)->info_args.alloc_shm = TRUE; + } + + if((*win_ptr)->info_args.alloc_shm == TRUE) { + mpi_errno = MPIDI_CH3U_Win_fns.allocate_shared(size, disp_unit, info, comm_ptr, baseptr, win_ptr); + if (mpi_errno != MPI_SUCCESS) MPIU_ERR_POP(mpi_errno); + } + else { + mpi_errno = MPIDI_CH3U_Win_fns.allocate(size, disp_unit, info, comm_ptr, baseptr, win_ptr); + if (mpi_errno != MPI_SUCCESS) { MPIU_ERR_POP(mpi_errno); } + } fn_fail: MPIDI_FUNC_EXIT(MPID_STATE_MPID_WIN_ALLOCATE); @@ -298,6 +312,7 @@ static int win_init(MPI_Aint size, int disp_unit, int create_flavor, int model, (*win_ptr)->info_args.accumulate_ops = MPIDI_ACC_OPS_SAME_OP_NO_OP; (*win_ptr)->info_args.same_size = 0; (*win_ptr)->info_args.alloc_shared_noncontig = 0; + (*win_ptr)->info_args.alloc_shm = FALSE; MPID_WIN_FTABLE_SET_DEFAULTS(win_ptr); http://git.mpich.org/mpich.git/commitdiff/44d9c98dc934631af208c62113703c135… commit 44d9c98dc934631af208c62113703c135bd86064 Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Mon Jul 22 10:08:58 2013 -0500 Refactoring the code of SHM RMA operations from operation routines to inline functions. The code in inline functions is moved from operation routines in ch3u_rma_ops.c and ch3u_rma_acc_ops.c. By moving them in inline functions, both operation routines and synchronization routines can call them. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/src/mpid/ch3/include/mpidrma.h b/src/mpid/ch3/include/mpidrma.h index 8abc78d..ecb59c3 100644 --- a/src/mpid/ch3/include/mpidrma.h +++ b/src/mpid/ch3/include/mpidrma.h @@ -7,6 +7,7 @@ #define MPICH_MPIDRMA_H_INCLUDED #include "mpl_utlist.h" +#include "mpidi_ch3_impl.h" #ifdef USE_MPIU_INSTR MPIU_INSTR_DURATION_EXTERN_DECL(wincreate_allgather); @@ -308,6 +309,528 @@ static inline MPIDI_RMA_Ops_list_t *MPIDI_CH3I_RMA_Get_ops_list(MPID_Win *win_pt } +/* ------------------------------------------------------------------------ */ +/* + * Followings are new routines for origin completion for RMA operations. + */ +/* ------------------------------------------------------------------------ */ + +#undef FUNCNAME +#define FUNCNAME MPIDI_CH3I_Shm_put_op +#undef FCNAME +#define FCNAME MPIDI_QUOTE(FUNCNAME) +static inline int MPIDI_CH3I_Shm_put_op(const void *origin_addr, int origin_count, MPI_Datatype + origin_datatype, int target_rank, MPI_Aint target_disp, + int target_count, MPI_Datatype target_datatype, + MPID_Win *win_ptr) +{ + int mpi_errno = MPI_SUCCESS; + void *base = NULL; + int disp_unit; + MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SHM_PUT_OP); + + MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_SHM_PUT_OP); + + if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { + base = win_ptr->shm_base_addrs[target_rank]; + disp_unit = win_ptr->disp_units[target_rank]; + } + else { + base = win_ptr->base; + disp_unit = win_ptr->disp_unit; + } + + mpi_errno = MPIR_Localcopy(origin_addr, origin_count, origin_datatype, + (char *) base + disp_unit * target_disp, + target_count, target_datatype); + if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } + + fn_exit: + MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_SHM_PUT_OP); + return mpi_errno; + /* --BEGIN ERROR HANDLING-- */ + fn_fail: + goto fn_exit; + /* --END ERROR HANDLING-- */ +} + + +#undef FUNCNAME +#define FUNCNAME MPIDI_CH3I_Shm_acc_op +#undef FCNAME +#define FCNAME MPIDI_QUOTE(FUNCNAME) +static inline int MPIDI_CH3I_Shm_acc_op(const void *origin_addr, int origin_count, MPI_Datatype + origin_datatype, int target_rank, MPI_Aint target_disp, + int target_count, MPI_Datatype target_datatype, MPI_Op op, + MPID_Win *win_ptr) +{ + void *base = NULL; + int disp_unit, shm_op = 0; + int origin_predefined, target_predefined; + MPI_User_function *uop = NULL; + MPID_Datatype *dtp; + int mpi_errno = MPI_SUCCESS; + MPIU_CHKLMEM_DECL(2); + MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SHM_ACC_OP); + + MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_SHM_ACC_OP); + + MPIDI_CH3I_DATATYPE_IS_PREDEFINED(origin_datatype, origin_predefined); + MPIDI_CH3I_DATATYPE_IS_PREDEFINED(target_datatype, target_predefined); + + if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { + shm_op = 1; + base = win_ptr->shm_base_addrs[target_rank]; + disp_unit = win_ptr->disp_units[target_rank]; + } + else { + base = win_ptr->base; + disp_unit = win_ptr->disp_unit; + } + + if (op == MPI_REPLACE) + { + if (shm_op) MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); + mpi_errno = MPIR_Localcopy(origin_addr, origin_count, + origin_datatype, + (char *) base + disp_unit * target_disp, + target_count, target_datatype); + if (shm_op) MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } + goto fn_exit; + } + + MPIU_ERR_CHKANDJUMP1((HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN), + mpi_errno, MPI_ERR_OP, "**opnotpredefined", + "**opnotpredefined %d", op ); + + /* get the function by indexing into the op table */ + uop = MPIR_OP_HDL_TO_FN(op); + + if (origin_predefined && target_predefined) + { + /* Cast away const'ness for origin_address in order to + * avoid changing the prototype for MPI_User_function */ + if (shm_op) MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); + (*uop)((void *) origin_addr, (char *) base + disp_unit*target_disp, + &target_count, &target_datatype); + if (shm_op) MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + } + else + { + /* derived datatype */ + + MPID_Segment *segp; + DLOOP_VECTOR *dloop_vec; + MPI_Aint first, last; + int vec_len, i, type_size, count; + MPI_Datatype type; + MPI_Aint true_lb, true_extent, extent; + void *tmp_buf=NULL, *target_buf; + const void *source_buf; + + if (origin_datatype != target_datatype) + { + /* first copy the data into a temporary buffer with + the same datatype as the target. Then do the + accumulate operation. */ + + MPIR_Type_get_true_extent_impl(target_datatype, &true_lb, &true_extent); + MPID_Datatype_get_extent_macro(target_datatype, extent); + + MPIU_CHKLMEM_MALLOC(tmp_buf, void *, + target_count * (MPIR_MAX(extent,true_extent)), + mpi_errno, "temporary buffer"); + /* adjust for potential negative lower bound in datatype */ + tmp_buf = (void *)((char*)tmp_buf - true_lb); + + mpi_errno = MPIR_Localcopy(origin_addr, origin_count, + origin_datatype, tmp_buf, + target_count, target_datatype); + if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } + } + + if (target_predefined) { + /* target predefined type, origin derived datatype */ + + if (shm_op) MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); + (*uop)(tmp_buf, (char *) base + disp_unit * target_disp, + &target_count, &target_datatype); + if (shm_op) MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + } + else { + + segp = MPID_Segment_alloc(); + MPIU_ERR_CHKANDJUMP1((!segp), mpi_errno, MPI_ERR_OTHER, + "**nomem","**nomem %s","MPID_Segment_alloc"); + MPID_Segment_init(NULL, target_count, target_datatype, segp, 0); + first = 0; + last = SEGMENT_IGNORE_LAST; + + MPID_Datatype_get_ptr(target_datatype, dtp); + vec_len = dtp->max_contig_blocks * target_count + 1; + /* +1 needed because Rob says so */ + MPIU_CHKLMEM_MALLOC(dloop_vec, DLOOP_VECTOR *, + vec_len * sizeof(DLOOP_VECTOR), + mpi_errno, "dloop vector"); + + MPID_Segment_pack_vector(segp, first, &last, dloop_vec, &vec_len); + + source_buf = (tmp_buf != NULL) ? tmp_buf : origin_addr; + target_buf = (char *) base + disp_unit * target_disp; + type = dtp->eltype; + type_size = MPID_Datatype_get_basic_size(type); + if (shm_op) MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); + for (i=0; i<vec_len; i++) + { + count = (dloop_vec[i].DLOOP_VECTOR_LEN)/type_size; + (*uop)((char *)source_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF), + (char *)target_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF), + &count, &type); + } + if (shm_op) MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + + MPID_Segment_free(segp); + } + } + + fn_exit: + MPIU_CHKLMEM_FREEALL(); + MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_SHM_ACC_OP); + return mpi_errno; + /* --BEGIN ERROR HANDLING-- */ + fn_fail: + if (shm_op) MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + goto fn_exit; + /* --END ERROR HANDLING-- */ +} + + +#undef FUNCNAME +#define FUNCNAME MPIDI_CH3I_Shm_get_acc_op +#undef FCNAME +#define FCNAME MPIDI_QUOTE(FUNCNAME) +static inline int MPIDI_CH3I_Shm_get_acc_op(const void *origin_addr, int origin_count, MPI_Datatype + origin_datatype, void *result_addr, int result_count, + MPI_Datatype result_datatype, int target_rank, MPI_Aint + target_disp, int target_count, MPI_Datatype target_datatype, + MPI_Op op, MPID_Win *win_ptr) +{ + int disp_unit, shm_locked = 0; + void *base = NULL; + MPI_User_function *uop = NULL; + MPID_Datatype *dtp; + int origin_predefined, result_predefined, target_predefined; + int mpi_errno = MPI_SUCCESS; + MPIU_CHKLMEM_DECL(2); + MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SHM_GET_ACC_OP); + + MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_SHM_GET_ACC_OP); + + origin_predefined = TRUE; /* quiet uninitialized warnings (b/c goto) */ + if (op != MPI_NO_OP) { + MPIDI_CH3I_DATATYPE_IS_PREDEFINED(origin_datatype, origin_predefined); + } + MPIDI_CH3I_DATATYPE_IS_PREDEFINED(result_datatype, result_predefined); + MPIDI_CH3I_DATATYPE_IS_PREDEFINED(target_datatype, target_predefined); + + if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { + base = win_ptr->shm_base_addrs[target_rank]; + disp_unit = win_ptr->disp_units[target_rank]; + MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); + shm_locked = 1; + } + else { + base = win_ptr->base; + disp_unit = win_ptr->disp_unit; + } + + /* Perform the local get first, then the accumulate */ + mpi_errno = MPIR_Localcopy((char *) base + disp_unit * target_disp, + target_count, target_datatype, + result_addr, result_count, result_datatype); + if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } + + /* NO_OP: Don't perform the accumulate */ + if (op == MPI_NO_OP) { + if (shm_locked) { + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + shm_locked = 0; + } + + goto fn_exit; + } + + if (op == MPI_REPLACE) { + mpi_errno = MPIR_Localcopy(origin_addr, origin_count, origin_datatype, + (char *) base + disp_unit * target_disp, + target_count, target_datatype); + + if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } + + if (shm_locked) { + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + shm_locked = 0; + } + + goto fn_exit; + } + + MPIU_ERR_CHKANDJUMP1((HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN), + mpi_errno, MPI_ERR_OP, "**opnotpredefined", + "**opnotpredefined %d", op ); + + /* get the function by indexing into the op table */ + uop = MPIR_OP_HDL_TO_FN(op); + + if (origin_predefined && target_predefined) { + /* Cast away const'ness for origin_address in order to + * avoid changing the prototype for MPI_User_function */ + (*uop)((void *) origin_addr, (char *) base + disp_unit*target_disp, + &target_count, &target_datatype); + } + else { + /* derived datatype */ + + MPID_Segment *segp; + DLOOP_VECTOR *dloop_vec; + MPI_Aint first, last; + int vec_len, i, type_size, count; + MPI_Datatype type; + MPI_Aint true_lb, true_extent, extent; + void *tmp_buf=NULL, *target_buf; + const void *source_buf; + + if (origin_datatype != target_datatype) { + /* first copy the data into a temporary buffer with + the same datatype as the target. Then do the + accumulate operation. */ + + MPIR_Type_get_true_extent_impl(target_datatype, &true_lb, &true_extent); + MPID_Datatype_get_extent_macro(target_datatype, extent); + + MPIU_CHKLMEM_MALLOC(tmp_buf, void *, + target_count * (MPIR_MAX(extent,true_extent)), + mpi_errno, "temporary buffer"); + /* adjust for potential negative lower bound in datatype */ + tmp_buf = (void *)((char*)tmp_buf - true_lb); + + mpi_errno = MPIR_Localcopy(origin_addr, origin_count, + origin_datatype, tmp_buf, + target_count, target_datatype); + if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } + } + + if (target_predefined) { + /* target predefined type, origin derived datatype */ + + (*uop)(tmp_buf, (char *) base + disp_unit * target_disp, + &target_count, &target_datatype); + } + else { + + segp = MPID_Segment_alloc(); + MPIU_ERR_CHKANDJUMP1((!segp), mpi_errno, MPI_ERR_OTHER, + "**nomem","**nomem %s","MPID_Segment_alloc"); + MPID_Segment_init(NULL, target_count, target_datatype, segp, 0); + first = 0; + last = SEGMENT_IGNORE_LAST; + + MPID_Datatype_get_ptr(target_datatype, dtp); + vec_len = dtp->max_contig_blocks * target_count + 1; + /* +1 needed because Rob says so */ + MPIU_CHKLMEM_MALLOC(dloop_vec, DLOOP_VECTOR *, + vec_len * sizeof(DLOOP_VECTOR), + mpi_errno, "dloop vector"); + + MPID_Segment_pack_vector(segp, first, &last, dloop_vec, &vec_len); + + source_buf = (tmp_buf != NULL) ? tmp_buf : origin_addr; + target_buf = (char *) base + disp_unit * target_disp; + type = dtp->eltype; + type_size = MPID_Datatype_get_basic_size(type); + + for (i=0; i<vec_len; i++) { + count = (dloop_vec[i].DLOOP_VECTOR_LEN)/type_size; + (*uop)((char *)source_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF), + (char *)target_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF), + &count, &type); + } + + MPID_Segment_free(segp); + } + } + + if (shm_locked) { + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + shm_locked = 0; + } + + fn_exit: + MPIU_CHKLMEM_FREEALL(); + MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_SHM_GET_ACC_OP); + return mpi_errno; + /* --BEGIN ERROR HANDLING-- */ + fn_fail: + if (shm_locked) { + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + } + goto fn_exit; + /* --END ERROR HANDLING-- */ +} + + +#undef FUNCNAME +#define FUNCNAME MPIDI_CH3I_Shm_get_op +#undef FCNAME +#define FCNAME MPIDI_QUOTE(FUNCNAME) +static inline int MPIDI_CH3I_Shm_get_op(void *origin_addr, int origin_count, MPI_Datatype origin_datatype, + int target_rank, MPI_Aint target_disp, int target_count, + MPI_Datatype target_datatype, MPID_Win *win_ptr) +{ + void *base = NULL; + int disp_unit; + int mpi_errno = MPI_SUCCESS; + MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SHM_GET_OP); + + MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_SHM_GET_OP); + + if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { + base = win_ptr->shm_base_addrs[target_rank]; + disp_unit = win_ptr->disp_units[target_rank]; + } + else { + base = win_ptr->base; + disp_unit = win_ptr->disp_unit; + } + + mpi_errno = MPIR_Localcopy((char *) base + disp_unit * target_disp, + target_count, target_datatype, origin_addr, + origin_count, origin_datatype); + if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } + + fn_exit: + MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_SHM_GET_OP); + return mpi_errno; + /* --BEGIN ERROR HANDLING-- */ + fn_fail: + goto fn_exit; + /* --END ERROR HANDLING-- */ +} + + +#undef FUNCNAME +#define FUNCNAME MPIDI_CH3I_Shm_cas_op +#undef FCNAME +#define FCNAME MPIDI_QUOTE(FUNCNAME) +static inline int MPIDI_CH3I_Shm_cas_op(const void *origin_addr, const void *compare_addr, + void *result_addr, MPI_Datatype datatype, int target_rank, + MPI_Aint target_disp, MPID_Win *win_ptr) +{ + void *base = NULL, *dest_addr = NULL; + int disp_unit; + int len, shm_locked = 0; + int mpi_errno = MPI_SUCCESS; + MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SHM_CAS_OP); + + MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_SHM_CAS_OP); + + if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { + base = win_ptr->shm_base_addrs[target_rank]; + disp_unit = win_ptr->disp_units[target_rank]; + + MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); + shm_locked = 1; + } + else { + base = win_ptr->base; + disp_unit = win_ptr->disp_unit; + } + + dest_addr = (char *) base + disp_unit * target_disp; + + MPID_Datatype_get_size_macro(datatype, len); + MPIU_Memcpy(result_addr, dest_addr, len); + + if (MPIR_Compare_equal(compare_addr, dest_addr, datatype)) { + MPIU_Memcpy(dest_addr, origin_addr, len); + } + + if (shm_locked) { + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + shm_locked = 0; + } + + fn_exit: + MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_SHM_CAS_OP); + return mpi_errno; + /* --BEGIN ERROR HANDLING-- */ + fn_fail: + if (shm_locked) { + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + } + goto fn_exit; + /* --END ERROR HANDLING-- */ +} + + +#undef FUNCNAME +#define FUNCNAME MPIDI_CH3I_Shm_fop_op +#undef FCNAME +#define FCNAME MPIDI_QUOTE(FUNCNAME) +static inline int MPIDI_CH3I_Shm_fop_op(const void *origin_addr, void *result_addr, + MPI_Datatype datatype, int target_rank, + MPI_Aint target_disp, MPI_Op op, MPID_Win *win_ptr) +{ + void *base = NULL, *dest_addr = NULL; + MPI_User_function *uop = NULL; + int disp_unit; + int len, one, shm_locked = 0; + int mpi_errno = MPI_SUCCESS; + MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_SHM_FOP_OP); + + MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_SHM_FOP_OP); + + if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { + base = win_ptr->shm_base_addrs[target_rank]; + disp_unit = win_ptr->disp_units[target_rank]; + + MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); + shm_locked = 1; + } + else { + base = win_ptr->base; + disp_unit = win_ptr->disp_unit; + } + + dest_addr = (char *) base + disp_unit * target_disp; + + MPID_Datatype_get_size_macro(datatype, len); + MPIU_Memcpy(result_addr, dest_addr, len); + + uop = MPIR_OP_HDL_TO_FN(op); + one = 1; + + (*uop)((void *) origin_addr, dest_addr, &one, &datatype); + + if (shm_locked) { + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + shm_locked = 0; + } + + fn_exit: + MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_SHM_FOP_OP); + return mpi_errno; + /* --BEGIN ERROR HANDLING-- */ + fn_fail: + if (shm_locked) { + MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); + } + goto fn_exit; + /* --END ERROR HANDLING-- */ +} + + #undef FUNCNAME #define FUNCNAME MPIDI_CH3I_Wait_for_pt_ops_finish #undef FCNAME diff --git a/src/mpid/ch3/src/ch3u_rma_acc_ops.c b/src/mpid/ch3/src/ch3u_rma_acc_ops.c index f50e6b1..82b1d2d 100644 --- a/src/mpid/ch3/src/ch3u_rma_acc_ops.c +++ b/src/mpid/ch3/src/ch3u_rma_acc_ops.c @@ -4,7 +4,6 @@ * See COPYRIGHT in top-level directory. */ -#include "mpidi_ch3_impl.h" #include "mpidrma.h" #ifdef USE_MPIU_INSTR @@ -25,7 +24,6 @@ int MPIDI_Get_accumulate(const void *origin_addr, int origin_count, int mpi_errno = MPI_SUCCESS; MPIDI_msg_sz_t data_sz; int rank, origin_predefined, result_predefined, target_predefined; - int shm_locked = 0; int dt_contig ATTRIBUTE((unused)); MPI_Aint dt_true_lb ATTRIBUTE((unused)); MPID_Datatype *dtp; @@ -64,141 +62,11 @@ int MPIDI_Get_accumulate(const void *origin_addr, int origin_count, /* Do =! rank first (most likely branch?) */ if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { - MPI_User_function *uop; - void *base; - int disp_unit; - - if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { - base = win_ptr->shm_base_addrs[target_rank]; - disp_unit = win_ptr->disp_units[target_rank]; - MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); - shm_locked = 1; - } - else { - base = win_ptr->base; - disp_unit = win_ptr->disp_unit; - } - - /* Perform the local get first, then the accumulate */ - mpi_errno = MPIR_Localcopy((char *) base + disp_unit * target_disp, - target_count, target_datatype, - result_addr, result_count, result_datatype); - if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } - - /* NO_OP: Don't perform the accumulate */ - if (op == MPI_NO_OP) { - if (shm_locked) { - MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); - shm_locked = 0; - } - - goto fn_exit; - } - - if (op == MPI_REPLACE) { - mpi_errno = MPIR_Localcopy(origin_addr, origin_count, origin_datatype, - (char *) base + disp_unit * target_disp, - target_count, target_datatype); - - if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } - - if (shm_locked) { - MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); - shm_locked = 0; - } - - goto fn_exit; - } - - MPIU_ERR_CHKANDJUMP1((HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN), - mpi_errno, MPI_ERR_OP, "**opnotpredefined", - "**opnotpredefined %d", op ); - - /* get the function by indexing into the op table */ - uop = MPIR_OP_HDL_TO_FN(op); - - if (origin_predefined && target_predefined) { - /* Cast away const'ness for origin_address in order to - * avoid changing the prototype for MPI_User_function */ - (*uop)((void *) origin_addr, (char *) base + disp_unit*target_disp, - &target_count, &target_datatype); - } - else { - /* derived datatype */ - - MPID_Segment *segp; - DLOOP_VECTOR *dloop_vec; - MPI_Aint first, last; - int vec_len, i, type_size, count; - MPI_Datatype type; - MPI_Aint true_lb, true_extent, extent; - void *tmp_buf=NULL, *target_buf; - const void *source_buf; - - if (origin_datatype != target_datatype) { - /* first copy the data into a temporary buffer with - the same datatype as the target. Then do the - accumulate operation. */ - - MPIR_Type_get_true_extent_impl(target_datatype, &true_lb, &true_extent); - MPID_Datatype_get_extent_macro(target_datatype, extent); - - MPIU_CHKLMEM_MALLOC(tmp_buf, void *, - target_count * (MPIR_MAX(extent,true_extent)), - mpi_errno, "temporary buffer"); - /* adjust for potential negative lower bound in datatype */ - tmp_buf = (void *)((char*)tmp_buf - true_lb); - - mpi_errno = MPIR_Localcopy(origin_addr, origin_count, - origin_datatype, tmp_buf, - target_count, target_datatype); - if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } - } - - if (target_predefined) { - /* target predefined type, origin derived datatype */ - - (*uop)(tmp_buf, (char *) base + disp_unit * target_disp, - &target_count, &target_datatype); - } - else { - - segp = MPID_Segment_alloc(); - MPIU_ERR_CHKANDJUMP1((!segp), mpi_errno, MPI_ERR_OTHER, - "**nomem","**nomem %s","MPID_Segment_alloc"); - MPID_Segment_init(NULL, target_count, target_datatype, segp, 0); - first = 0; - last = SEGMENT_IGNORE_LAST; - - MPID_Datatype_get_ptr(target_datatype, dtp); - vec_len = dtp->max_contig_blocks * target_count + 1; - /* +1 needed because Rob says so */ - MPIU_CHKLMEM_MALLOC(dloop_vec, DLOOP_VECTOR *, - vec_len * sizeof(DLOOP_VECTOR), - mpi_errno, "dloop vector"); - - MPID_Segment_pack_vector(segp, first, &last, dloop_vec, &vec_len); - - source_buf = (tmp_buf != NULL) ? tmp_buf : origin_addr; - target_buf = (char *) base + disp_unit * target_disp; - type = dtp->eltype; - type_size = MPID_Datatype_get_basic_size(type); - - for (i=0; i<vec_len; i++) { - count = (dloop_vec[i].DLOOP_VECTOR_LEN)/type_size; - (*uop)((char *)source_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF), - (char *)target_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF), - &count, &type); - } - - MPID_Segment_free(segp); - } - } - - if (shm_locked) { - MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); - shm_locked = 0; - } + mpi_errno = MPIDI_CH3I_Shm_get_acc_op(origin_addr, origin_count, origin_datatype, + result_addr, result_count, result_datatype, + target_rank, target_disp, target_count, target_datatype, + op, win_ptr); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); } else { MPIDI_RMA_Ops_list_t *ops_list = MPIDI_CH3I_RMA_Get_ops_list(win_ptr, target_rank); @@ -252,9 +120,6 @@ int MPIDI_Get_accumulate(const void *origin_addr, int origin_count, /* --BEGIN ERROR HANDLING-- */ fn_fail: - if (shm_locked) { - MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); - } goto fn_exit; /* --END ERROR HANDLING-- */ } @@ -269,7 +134,6 @@ int MPIDI_Compare_and_swap(const void *origin_addr, const void *compare_addr, MPI_Aint target_disp, MPID_Win *win_ptr) { int mpi_errno = MPI_SUCCESS; - int shm_locked = 0; int rank; MPIDI_STATE_DECL(MPID_STATE_MPIDI_COMPARE_AND_SWAP); @@ -296,35 +160,9 @@ int MPIDI_Compare_and_swap(const void *origin_addr, const void *compare_addr, * that uses a processor atomic operation. */ if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { - void *base, *dest_addr; - int disp_unit; - int len; - - if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { - base = win_ptr->shm_base_addrs[target_rank]; - disp_unit = win_ptr->disp_units[target_rank]; - - MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); - shm_locked = 1; - } - else { - base = win_ptr->base; - disp_unit = win_ptr->disp_unit; - } - - dest_addr = (char *) base + disp_unit * target_disp; - - MPID_Datatype_get_size_macro(datatype, len); - MPIU_Memcpy(result_addr, dest_addr, len); - - if (MPIR_Compare_equal(compare_addr, dest_addr, datatype)) { - MPIU_Memcpy(dest_addr, origin_addr, len); - } - - if (shm_locked) { - MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); - shm_locked = 0; - } + mpi_errno = MPIDI_CH3I_Shm_cas_op(origin_addr, compare_addr, result_addr, + datatype, target_rank, target_disp, win_ptr); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); } else { MPIDI_RMA_Ops_list_t *ops_list = MPIDI_CH3I_RMA_Get_ops_list(win_ptr, target_rank); @@ -359,9 +197,6 @@ fn_exit: return mpi_errno; /* --BEGIN ERROR HANDLING-- */ fn_fail: - if (shm_locked) { - MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); - } goto fn_exit; /* --END ERROR HANDLING-- */ } @@ -376,7 +211,6 @@ int MPIDI_Fetch_and_op(const void *origin_addr, void *result_addr, MPI_Aint target_disp, MPI_Op op, MPID_Win *win_ptr) { int mpi_errno = MPI_SUCCESS; - int shm_locked = 0; int rank; MPIDI_STATE_DECL(MPID_STATE_MPIDI_FETCH_AND_OP); @@ -402,37 +236,9 @@ int MPIDI_Fetch_and_op(const void *origin_addr, void *result_addr, * that uses a processor atomic operation. */ if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { - MPI_User_function *uop; - void *base, *dest_addr; - int disp_unit; - int len, one; - - if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { - base = win_ptr->shm_base_addrs[target_rank]; - disp_unit = win_ptr->disp_units[target_rank]; - - MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); - shm_locked = 1; - } - else { - base = win_ptr->base; - disp_unit = win_ptr->disp_unit; - } - - dest_addr = (char *) base + disp_unit * target_disp; - - MPID_Datatype_get_size_macro(datatype, len); - MPIU_Memcpy(result_addr, dest_addr, len); - - uop = MPIR_OP_HDL_TO_FN(op); - one = 1; - - (*uop)((void *) origin_addr, dest_addr, &one, &datatype); - - if (shm_locked) { - MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); - shm_locked = 0; - } + mpi_errno = MPIDI_CH3I_Shm_fop_op(origin_addr, result_addr, datatype, + target_rank, target_disp, op, win_ptr); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); } else { MPIDI_RMA_Ops_list_t *ops_list = MPIDI_CH3I_RMA_Get_ops_list(win_ptr, target_rank); @@ -465,9 +271,6 @@ fn_exit: return mpi_errno; /* --BEGIN ERROR HANDLING-- */ fn_fail: - if (shm_locked) { - MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); - } goto fn_exit; /* --END ERROR HANDLING-- */ } diff --git a/src/mpid/ch3/src/ch3u_rma_ops.c b/src/mpid/ch3/src/ch3u_rma_ops.c index 7189513..5f6fb4c 100644 --- a/src/mpid/ch3/src/ch3u_rma_ops.c +++ b/src/mpid/ch3/src/ch3u_rma_ops.c @@ -4,7 +4,6 @@ * See COPYRIGHT in top-level directory. */ -#include "mpidi_ch3_impl.h" #include "mpidrma.h" static int enableShortACC=1; @@ -182,22 +181,9 @@ int MPIDI_Put(const void *origin_addr, int origin_count, MPI_Datatype /* If the put is a local operation, do it here */ if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { - void *base; - int disp_unit; - - if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { - base = win_ptr->shm_base_addrs[target_rank]; - disp_unit = win_ptr->disp_units[target_rank]; - } - else { - base = win_ptr->base; - disp_unit = win_ptr->disp_unit; - } - - mpi_errno = MPIR_Localcopy(origin_addr, origin_count, origin_datatype, - (char *) base + disp_unit * target_disp, - target_count, target_datatype); - if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } + mpi_errno = MPIDI_CH3I_Shm_put_op(origin_addr, origin_count, origin_datatype, target_rank, + target_disp, target_count, target_datatype, win_ptr); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); } else { @@ -293,22 +279,9 @@ int MPIDI_Get(void *origin_addr, int origin_count, MPI_Datatype /* If the get is a local operation, do it here */ if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { - void *base; - int disp_unit; - - if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { - base = win_ptr->shm_base_addrs[target_rank]; - disp_unit = win_ptr->disp_units[target_rank]; - } - else { - base = win_ptr->base; - disp_unit = win_ptr->disp_unit; - } - - mpi_errno = MPIR_Localcopy((char *) base + disp_unit * target_disp, - target_count, target_datatype, origin_addr, - origin_count, origin_datatype); - if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } + mpi_errno = MPIDI_CH3I_Shm_get_op(origin_addr, origin_count, origin_datatype, target_rank, + target_disp, target_count, target_datatype, win_ptr); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); } else { @@ -406,125 +379,10 @@ int MPIDI_Accumulate(const void *origin_addr, int origin_count, MPI_Datatype /* Do =! rank first (most likely branch?) */ if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { - MPI_User_function *uop; - void *base; - int disp_unit, shm_op = 0; - - if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) { - shm_op = 1; - base = win_ptr->shm_base_addrs[target_rank]; - disp_unit = win_ptr->disp_units[target_rank]; - } - else { - base = win_ptr->base; - disp_unit = win_ptr->disp_unit; - } - - if (op == MPI_REPLACE) - { - if (shm_op) MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); - mpi_errno = MPIR_Localcopy(origin_addr, origin_count, - origin_datatype, - (char *) base + disp_unit * target_disp, - target_count, target_datatype); - if (shm_op) MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); - if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } - goto fn_exit; - } - - MPIU_ERR_CHKANDJUMP1((HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN), - mpi_errno, MPI_ERR_OP, "**opnotpredefined", - "**opnotpredefined %d", op ); - - /* get the function by indexing into the op table */ - uop = MPIR_OP_HDL_TO_FN(op); - - if (origin_predefined && target_predefined) - { - /* Cast away const'ness for origin_address in order to - * avoid changing the prototype for MPI_User_function */ - if (shm_op) MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); - (*uop)((void *) origin_addr, (char *) base + disp_unit*target_disp, - &target_count, &target_datatype); - if (shm_op) MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); - } - else - { - /* derived datatype */ - - MPID_Segment *segp; - DLOOP_VECTOR *dloop_vec; - MPI_Aint first, last; - int vec_len, i, type_size, count; - MPI_Datatype type; - MPI_Aint true_lb, true_extent, extent; - void *tmp_buf=NULL, *target_buf; - const void *source_buf; - - if (origin_datatype != target_datatype) - { - /* first copy the data into a temporary buffer with - the same datatype as the target. Then do the - accumulate operation. */ - - MPIR_Type_get_true_extent_impl(target_datatype, &true_lb, &true_extent); - MPID_Datatype_get_extent_macro(target_datatype, extent); - - MPIU_CHKLMEM_MALLOC(tmp_buf, void *, - target_count * (MPIR_MAX(extent,true_extent)), - mpi_errno, "temporary buffer"); - /* adjust for potential negative lower bound in datatype */ - tmp_buf = (void *)((char*)tmp_buf - true_lb); - - mpi_errno = MPIR_Localcopy(origin_addr, origin_count, - origin_datatype, tmp_buf, - target_count, target_datatype); - if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } - } - - if (target_predefined) { - /* target predefined type, origin derived datatype */ - - if (shm_op) MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); - (*uop)(tmp_buf, (char *) base + disp_unit * target_disp, - &target_count, &target_datatype); - if (shm_op) MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); - } - else { - - segp = MPID_Segment_alloc(); - MPIU_ERR_CHKANDJUMP1((!segp), mpi_errno, MPI_ERR_OTHER, - "**nomem","**nomem %s","MPID_Segment_alloc"); - MPID_Segment_init(NULL, target_count, target_datatype, segp, 0); - first = 0; - last = SEGMENT_IGNORE_LAST; - - MPID_Datatype_get_ptr(target_datatype, dtp); - vec_len = dtp->max_contig_blocks * target_count + 1; - /* +1 needed because Rob says so */ - MPIU_CHKLMEM_MALLOC(dloop_vec, DLOOP_VECTOR *, - vec_len * sizeof(DLOOP_VECTOR), - mpi_errno, "dloop vector"); - - MPID_Segment_pack_vector(segp, first, &last, dloop_vec, &vec_len); - - source_buf = (tmp_buf != NULL) ? tmp_buf : origin_addr; - target_buf = (char *) base + disp_unit * target_disp; - type = dtp->eltype; - type_size = MPID_Datatype_get_basic_size(type); - if (shm_op) MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr); - for (i=0; i<vec_len; i++) - { - count = (dloop_vec[i].DLOOP_VECTOR_LEN)/type_size; - (*uop)((char *)source_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF), - (char *)target_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF), - &count, &type); - } - if (shm_op) MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr); - - MPID_Segment_free(segp); - } - } + mpi_errno = MPIDI_CH3I_Shm_acc_op(origin_addr, origin_count, origin_datatype, + target_rank, target_disp, target_count, target_datatype, + op, win_ptr); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); } else { http://git.mpich.org/mpich.git/commitdiff/382bdfb2a2d35025b906033bc63fea98e… commit 382bdfb2a2d35025b906033bc63fea98e3c551e3 Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Mon Jul 22 10:03:49 2013 -0500 modify SHM_Win_allocate_shared and SHM_Win_free to accomodate global comm. MPIDI_CH3I_Win_allocate_shared can be called by both MPI_Win_allocate_shared and MPI_Win_allocate. If it is called by MPI_Win_allocate, we need node_comm, node_sizes and node_shm_base_addrs to allocate shm segment region, hence we need to copy from win_ptr->sizes to node_sizes at beginning and copy from node_shm_base_addrs to win_ptr->shm_base_addrs at last. If it is called by MPI_Win_allocate_shared, these copies can be eliminated. If there is only one process on this node, node_comm is NULL, we use comm_self instead. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/src/mpid/ch3/channels/nemesis/src/ch3_rma_shm.c b/src/mpid/ch3/channels/nemesis/src/ch3_rma_shm.c index 9e3f7ef..6045f54 100644 --- a/src/mpid/ch3/channels/nemesis/src/ch3_rma_shm.c +++ b/src/mpid/ch3/channels/nemesis/src/ch3_rma_shm.c @@ -84,8 +84,19 @@ int MPIDI_CH3_SHM_Win_free(MPID_Win **win_ptr) /* Free shared process mutex memory region */ if ((*win_ptr)->shm_mutex && (*win_ptr)->shm_segment_len > 0) { - - if ((*win_ptr)->comm_ptr->rank == 0) { + MPID_Comm *node_comm_ptr = NULL; + + /* When allocating shared memory region segment, we need comm of processes + that are on the same node as this process (node_comm). + If node_comm == NULL, this process is the only one on this node, therefore + we use comm_self as node comm. */ + if ((*win_ptr)->comm_ptr->node_comm != NULL) + node_comm_ptr = (*win_ptr)->comm_ptr->node_comm; + else + node_comm_ptr = MPIR_Process.comm_self; + MPIU_Assert(node_comm_ptr != NULL); + + if (node_comm_ptr->rank == 0) { MPIDI_CH3I_SHM_MUTEX_DESTROY(*win_ptr); } diff --git a/src/mpid/ch3/channels/nemesis/src/ch3_win_fns.c b/src/mpid/ch3/channels/nemesis/src/ch3_win_fns.c index 8942ed2..6b4fdce 100644 --- a/src/mpid/ch3/channels/nemesis/src/ch3_win_fns.c +++ b/src/mpid/ch3/channels/nemesis/src/ch3_win_fns.c @@ -49,11 +49,15 @@ static int MPIDI_CH3I_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Inf { int mpi_errno = MPI_SUCCESS; int i, k, comm_size, rank; + int node_size, node_rank; + MPID_Comm *node_comm_ptr; + MPI_Aint *node_sizes; + void **node_shm_base_addrs; MPI_Aint *tmp_buf; int errflag = FALSE; int noncontig = FALSE; MPIU_CHKPMEM_DECL(6); - MPIU_CHKLMEM_DECL(1); + MPIU_CHKLMEM_DECL(3); MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_WIN_ALLOCATE_SHARED); MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_WIN_ALLOCATE_SHARED); @@ -77,6 +81,18 @@ static int MPIDI_CH3I_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Inf comm_size = (*win_ptr)->comm_ptr->local_size; rank = (*win_ptr)->comm_ptr->rank; + /* When allocating shared memory region segment, we need comm of processes + that are on the same node as this process (node_comm). + If node_comm == NULL, this process is the only one on this node, therefore + we use comm_self as node comm. */ + if ((*win_ptr)->comm_ptr->node_comm != NULL) + node_comm_ptr = (*win_ptr)->comm_ptr->node_comm; + else + node_comm_ptr = MPIR_Process.comm_self; + MPIU_Assert(node_comm_ptr != NULL); + node_size = node_comm_ptr->local_size; + node_rank = node_comm_ptr->rank; + MPIU_INSTR_DURATION_START(wincreate_allgather); /* allocate memory for the base addresses, disp_units, and completion counters of all processes */ @@ -119,6 +135,14 @@ static int MPIDI_CH3I_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Inf if (mpi_errno) MPIU_ERR_POP(mpi_errno); MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail"); + if ((*win_ptr)->create_flavor != MPI_WIN_FLAVOR_SHARED) { + MPIU_CHKLMEM_MALLOC(node_sizes, MPI_Aint *, node_size*sizeof(MPI_Aint), mpi_errno, "node_sizes"); + for (i = 0; i < node_size; i++) node_sizes[i] = 0; + } + else { + node_sizes = (*win_ptr)->sizes; + } + (*win_ptr)->shm_segment_len = 0; k = 0; for (i = 0; i < comm_size; ++i) { @@ -126,11 +150,24 @@ static int MPIDI_CH3I_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Inf (*win_ptr)->disp_units[i] = (int) tmp_buf[k++]; (*win_ptr)->all_win_handles[i] = (MPI_Win) tmp_buf[k++]; + if ((*win_ptr)->create_flavor != MPI_WIN_FLAVOR_SHARED) { + /* If create flavor is not MPI_WIN_FLAVOR_SHARED, all processes on this + window may not be on the same node. Because we only need the sizes of local + processes (in order), we copy their sizes to a seperate array and keep them + in order, fur purpose of future use of calculating shm_base_addrs. */ + if ((*win_ptr)->comm_ptr->intranode_table[i] >= 0) { + MPIU_Assert((*win_ptr)->comm_ptr->intranode_table[i] < node_size); + node_sizes[(*win_ptr)->comm_ptr->intranode_table[i]] = (*win_ptr)->sizes[i]; + } + } + } + + for (i = 0; i < node_size; i++) { if (noncontig) /* Round up to next page size */ - (*win_ptr)->shm_segment_len += MPIDI_CH3_ROUND_UP_PAGESIZE((*win_ptr)->sizes[i]); + (*win_ptr)->shm_segment_len += MPIDI_CH3_ROUND_UP_PAGESIZE(node_sizes[i]); else - (*win_ptr)->shm_segment_len += (*win_ptr)->sizes[i]; + (*win_ptr)->shm_segment_len += node_sizes[i]; } if ((*win_ptr)->shm_segment_len == 0) { @@ -141,7 +178,7 @@ static int MPIDI_CH3I_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Inf mpi_errno = MPIU_SHMW_Hnd_init(&(*win_ptr)->shm_segment_handle); if (mpi_errno) MPIU_ERR_POP(mpi_errno); - if (rank == 0) { + if (node_rank == 0) { char *serialized_hnd_ptr = NULL; /* create shared memory region for all processes in win and map */ @@ -153,12 +190,12 @@ static int MPIDI_CH3I_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Inf mpi_errno = MPIU_SHMW_Hnd_get_serialized_by_ref((*win_ptr)->shm_segment_handle, &serialized_hnd_ptr); if (mpi_errno) MPIU_ERR_POP(mpi_errno); - mpi_errno = MPIR_Bcast_impl(serialized_hnd_ptr, MPIU_SHMW_GHND_SZ, MPI_CHAR, 0, (*win_ptr)->comm_ptr, &errflag); + mpi_errno = MPIR_Bcast_impl(serialized_hnd_ptr, MPIU_SHMW_GHND_SZ, MPI_CHAR, 0, node_comm_ptr, &errflag); if (mpi_errno) MPIU_ERR_POP(mpi_errno); MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail"); /* wait for other processes to attach to win */ - mpi_errno = MPIR_Barrier_impl((*win_ptr)->comm_ptr, &errflag); + mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag); if (mpi_errno) MPIU_ERR_POP(mpi_errno); MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail"); @@ -170,7 +207,7 @@ static int MPIDI_CH3I_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Inf char serialized_hnd[MPIU_SHMW_GHND_SZ] = {0}; /* get serialized handle from rank 0 and deserialize it */ - mpi_errno = MPIR_Bcast_impl(serialized_hnd, MPIU_SHMW_GHND_SZ, MPI_CHAR, 0, (*win_ptr)->comm_ptr, &errflag); + mpi_errno = MPIR_Bcast_impl(serialized_hnd, MPIU_SHMW_GHND_SZ, MPI_CHAR, 0, node_comm_ptr, &errflag); if (mpi_errno) MPIU_ERR_POP(mpi_errno); MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail"); @@ -182,7 +219,7 @@ static int MPIDI_CH3I_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Inf (char **)&(*win_ptr)->shm_base_addr, 0); if (mpi_errno) MPIU_ERR_POP(mpi_errno); - mpi_errno = MPIR_Barrier_impl((*win_ptr)->comm_ptr, &errflag); + mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag); if (mpi_errno) MPIU_ERR_POP(mpi_errno); MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail"); } @@ -191,7 +228,7 @@ static int MPIDI_CH3I_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Inf mpi_errno = MPIU_SHMW_Hnd_init(&(*win_ptr)->shm_mutex_segment_handle); if (mpi_errno) MPIU_ERR_POP(mpi_errno); - if (rank == 0) { + if (node_rank == 0) { char *serialized_hnd_ptr = NULL; /* create shared memory region for all processes in win and map */ @@ -205,12 +242,12 @@ static int MPIDI_CH3I_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Inf mpi_errno = MPIU_SHMW_Hnd_get_serialized_by_ref((*win_ptr)->shm_mutex_segment_handle, &serialized_hnd_ptr); if (mpi_errno) MPIU_ERR_POP(mpi_errno); - mpi_errno = MPIR_Bcast_impl(serialized_hnd_ptr, MPIU_SHMW_GHND_SZ, MPI_CHAR, 0, (*win_ptr)->comm_ptr, &errflag); + mpi_errno = MPIR_Bcast_impl(serialized_hnd_ptr, MPIU_SHMW_GHND_SZ, MPI_CHAR, 0, node_comm_ptr, &errflag); if (mpi_errno) MPIU_ERR_POP(mpi_errno); MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail"); /* wait for other processes to attach to win */ - mpi_errno = MPIR_Barrier_impl((*win_ptr)->comm_ptr, &errflag); + mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag); if (mpi_errno) MPIU_ERR_POP(mpi_errno); MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail"); @@ -221,7 +258,7 @@ static int MPIDI_CH3I_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Inf char serialized_hnd[MPIU_SHMW_GHND_SZ] = {0}; /* get serialized handle from rank 0 and deserialize it */ - mpi_errno = MPIR_Bcast_impl(serialized_hnd, MPIU_SHMW_GHND_SZ, MPI_CHAR, 0, (*win_ptr)->comm_ptr, &errflag); + mpi_errno = MPIR_Bcast_impl(serialized_hnd, MPIU_SHMW_GHND_SZ, MPI_CHAR, 0, node_comm_ptr, &errflag); if (mpi_errno) MPIU_ERR_POP(mpi_errno); MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail"); @@ -233,25 +270,51 @@ static int MPIDI_CH3I_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Inf (char **)&(*win_ptr)->shm_mutex, 0); if (mpi_errno) MPIU_ERR_POP(mpi_errno); - mpi_errno = MPIR_Barrier_impl((*win_ptr)->comm_ptr, &errflag); + mpi_errno = MPIR_Barrier_impl(node_comm_ptr, &errflag); if (mpi_errno) MPIU_ERR_POP(mpi_errno); MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail"); } /* compute the base addresses of each process within the shared memory segment */ { + if ((*win_ptr)->create_flavor != MPI_WIN_FLAVOR_SHARED) { + /* If create flavor is not MPI_WIN_FLAVOR_SHARED, all processes on this + window may not be on the same node. Because we only need to calculate + local processes' shm_base_addrs using local processes's sizes, + we allocate a temporary array to place results and copy results + back to shm_base_addrs on the window at last. */ + MPIU_CHKLMEM_MALLOC(node_shm_base_addrs, void **, node_size*sizeof(void*), + mpi_errno, "node_shm_base_addrs"); + } + else { + node_shm_base_addrs = (*win_ptr)->shm_base_addrs; + } + char *cur_base = (*win_ptr)->shm_base_addr; - (*win_ptr)->shm_base_addrs[0] = (*win_ptr)->shm_base_addr; - for (i = 1; i < comm_size; ++i) { - if ((*win_ptr)->sizes[i]) { + node_shm_base_addrs[0] = (*win_ptr)->shm_base_addr; + for (i = 1; i < node_size; ++i) { + if (node_sizes[i]) { if (noncontig) { - (*win_ptr)->shm_base_addrs[i] = cur_base + MPIDI_CH3_ROUND_UP_PAGESIZE((*win_ptr)->sizes[i-1]); + node_shm_base_addrs[i] = cur_base + MPIDI_CH3_ROUND_UP_PAGESIZE(node_sizes[i-1]); } else { - (*win_ptr)->shm_base_addrs[i] = cur_base + (*win_ptr)->sizes[i-1]; + node_shm_base_addrs[i] = cur_base + node_sizes[i-1]; } - cur_base = (*win_ptr)->shm_base_addrs[i]; + cur_base = node_shm_base_addrs[i]; } else { - (*win_ptr)->shm_base_addrs[i] = NULL; /* FIXME: Is this right? */ + node_shm_base_addrs[i] = NULL; /* FIXME: Is this right? */ + } + } + + if ((*win_ptr)->create_flavor != MPI_WIN_FLAVOR_SHARED) { + /* if MPI_WIN_FLAVOR_SHARED is not set, copy from node_shm_base_addrs to + (*win_ptr)->shm_base_addrs */ + for (i = 0; i < comm_size; i++) { + if ((*win_ptr)->comm_ptr->intranode_table[i] >= 0) { + MPIU_Assert((*win_ptr)->comm_ptr->intranode_table[i] < node_size); + (*win_ptr)->shm_base_addrs[i] = node_shm_base_addrs[(*win_ptr)->comm_ptr->intranode_table[i]]; + } + else + (*win_ptr)->shm_base_addrs[i] = NULL; } } } http://git.mpich.org/mpich.git/commitdiff/e397148092dc2377b0671ebb22238a1a3… commit e397148092dc2377b0671ebb22238a1a3456ea90 Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Fri Jul 26 15:42:25 2013 -0500 Waiting for passive RMA operations to finish in MPIDI_CH3_SHM_Win_free. This is for the optimization of allocating shared memory region in MPI_Win_allocate. In this case MPIDI_CH3_SHM_Win_free must first wait for passive RMA operations to finish before free the shared memory region. Note that because MPIDI_CH3_SHM_Win_free calls MPIDI_Win_free at last, and MPIDI_Win_free will also call inline function of waiting for passive RMA operation to finish, in this case the inline function will be called twice. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/src/mpid/ch3/channels/nemesis/src/ch3_rma_shm.c b/src/mpid/ch3/channels/nemesis/src/ch3_rma_shm.c index 08c6a1d..9e3f7ef 100644 --- a/src/mpid/ch3/channels/nemesis/src/ch3_rma_shm.c +++ b/src/mpid/ch3/channels/nemesis/src/ch3_rma_shm.c @@ -64,6 +64,9 @@ int MPIDI_CH3_SHM_Win_free(MPID_Win **win_ptr) MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3_SHM_WIN_FREE); + mpi_errno = MPIDI_CH3I_Wait_for_pt_ops_finish(*win_ptr); + if (mpi_errno) MPIU_ERR_POP(mpi_errno); + /* Free shared memory region */ if ((*win_ptr)->shm_allocated) { /* free shm_base_addrs that's only used for shared memory windows */ http://git.mpich.org/mpich.git/commitdiff/5c6b971af446d9ffaafdb35da449c78b8… commit 5c6b971af446d9ffaafdb35da449c78b82b92d8e Author: Xin Zhao <xinzhao3(a)illinois.edu> Date: Fri Jul 26 15:40:58 2013 -0500 Refactoring code of waiting passive RMA operations. Moving code of waiting finish of passive RMA operations from MPIDI_Win_free to an inline function in mpidrma.h. Signed-off-by: Pavan Balaji <balaji(a)mcs.anl.gov> diff --git a/src/mpid/ch3/include/mpidrma.h b/src/mpid/ch3/include/mpidrma.h index 26d1b17..8abc78d 100644 --- a/src/mpid/ch3/include/mpidrma.h +++ b/src/mpid/ch3/include/mpidrma.h @@ -8,6 +8,12 @@ #include "mpl_utlist.h" +#ifdef USE_MPIU_INSTR +MPIU_INSTR_DURATION_EXTERN_DECL(wincreate_allgather); +MPIU_INSTR_DURATION_EXTERN_DECL(winfree_rs); +MPIU_INSTR_DURATION_EXTERN_DECL(winfree_complete); +#endif + typedef enum MPIDI_RMA_Op_type { MPIDI_RMA_PUT = 23, MPIDI_RMA_GET = 24, @@ -301,6 +307,59 @@ static inline MPIDI_RMA_Ops_list_t *MPIDI_CH3I_RMA_Get_ops_list(MPID_Win *win_pt } } + +#undef FUNCNAME +#define FUNCNAME MPIDI_CH3I_Wait_for_pt_ops_finish +#undef FCNAME +#define FCNAME MPIDI_QUOTE(FUNCNAME) +static inline int MPIDI_CH3I_Wait_for_pt_ops_finish(MPID_Win *win_ptr) +{ + int mpi_errno = MPI_SUCCESS, total_pt_rma_puts_accs; + MPID_Comm *comm_ptr; + int errflag = FALSE; + MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_WAIT_FOR_PT_OPS_FINISH); + + MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_WAIT_FOR_PT_OPS_FINISH); + + comm_ptr = win_ptr->comm_ptr; + MPIU_INSTR_DURATION_START(winfree_rs); + mpi_errno = MPIR_Reduce_scatter_block_impl(win_ptr->pt_rma_puts_accs, + &total_pt_rma_puts_accs, 1, + MPI_INT, MPI_SUM, comm_ptr, &errflag); + if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } + MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail"); + MPIU_INSTR_DURATION_END(winfree_rs); + + if (total_pt_rma_puts_accs != win_ptr->my_pt_rma_puts_accs) + { + MPID_Progress_state progress_state; + + /* poke the progress engine until the two are equal */ + MPIU_INSTR_DURATION_START(winfree_complete); + MPID_Progress_start(&progress_state); + while (total_pt_rma_puts_accs != win_ptr->my_pt_rma_puts_accs) + { + mpi_errno = MPID_Progress_wait(&progress_state); + /* --BEGIN ERROR HANDLING-- */ + if (mpi_errno != MPI_SUCCESS) + { + MPID_Progress_end(&progress_state); + MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**winnoprogress"); + } + /* --END ERROR HANDLING-- */ + } + MPID_Progress_end(&progress_state); + MPIU_INSTR_DURATION_END(winfree_complete); + } + + fn_exit: + MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_WAIT_FOR_PT_OPS_FINISH); + return mpi_errno; + fn_fail: + goto fn_exit; +} + + #undef FUNCNAME #undef FCNAME diff --git a/src/mpid/ch3/src/ch3u_rma_ops.c b/src/mpid/ch3/src/ch3u_rma_ops.c index 95d47b9..7189513 100644 --- a/src/mpid/ch3/src/ch3u_rma_ops.c +++ b/src/mpid/ch3/src/ch3u_rma_ops.c @@ -10,9 +10,6 @@ static int enableShortACC=1; #ifdef USE_MPIU_INSTR -MPIU_INSTR_DURATION_EXTERN_DECL(wincreate_allgather); -MPIU_INSTR_DURATION_EXTERN_DECL(winfree_rs); -MPIU_INSTR_DURATION_EXTERN_DECL(winfree_complete); MPIU_INSTR_DURATION_EXTERN_DECL(rmaqueue_alloc); MPIU_INSTR_DURATION_EXTERN_DECL(rmaqueue_set); extern void MPIDI_CH3_RMA_InitInstr(void); @@ -55,37 +52,10 @@ int MPIDI_Win_free(MPID_Win **win_ptr) MPIU_ERR_CHKANDJUMP((*win_ptr)->epoch_state != MPIDI_EPOCH_NONE, mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync"); - comm_ptr = (*win_ptr)->comm_ptr; - MPIU_INSTR_DURATION_START(winfree_rs); - mpi_errno = MPIR_Reduce_scatter_block_impl((*win_ptr)->pt_rma_puts_accs, - &total_pt_rma_puts_accs, 1, - MPI_INT, MPI_SUM, comm_ptr, &errflag); - if (mpi_errno) { MPIU_ERR_POP(mpi_errno); } - MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail"); - MPIU_INSTR_DURATION_END(winfree_rs); - - if (total_pt_rma_puts_accs != (*win_ptr)->my_pt_rma_puts_accs) - { - MPID_Progress_state progress_state; - - /* poke the progress engine until the two are equal */ - MPIU_INSTR_DURATION_START(winfree_complete); - MPID_Progress_start(&progress_state); - while (total_pt_rma_puts_accs != (*win_ptr)->my_pt_rma_puts_accs) - { - mpi_errno = MPID_Progress_wait(&progress_state); - /* --BEGIN ERROR HANDLING-- */ - if (mpi_errno != MPI_SUCCESS) - { - MPID_Progress_end(&progress_state); - MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**winnoprogress"); - } - /* --END ERROR HANDLING-- */ - } - MPID_Progress_end(&progress_state); - MPIU_INSTR_DURATION_END(winfree_complete); - } + mpi_errno = MPIDI_CH3I_Wait_for_pt_ops_finish(*win_ptr); + if(mpi_errno) MPIU_ERR_POP(mpi_errno); + comm_ptr = (*win_ptr)->comm_ptr; mpi_errno = MPIR_Comm_free_impl(comm_ptr); if (mpi_errno) MPIU_ERR_POP(mpi_errno); ----------------------------------------------------------------------- Summary of changes: src/mpid/ch3/channels/nemesis/src/ch3_rma_shm.c | 18 +- src/mpid/ch3/channels/nemesis/src/ch3_win_fns.c | 116 ++++- src/mpid/ch3/include/mpidpre.h | 1 + src/mpid/ch3/include/mpidrma.h | 606 +++++++++++++++++++++++ src/mpid/ch3/src/ch3u_handle_recv_req.c | 26 +- src/mpid/ch3/src/ch3u_rma_acc_ops.c | 254 ++--------- src/mpid/ch3/src/ch3u_rma_ops.c | 308 +++--------- src/mpid/ch3/src/ch3u_rma_sync.c | 120 +++++- src/mpid/ch3/src/ch3u_win_fns.c | 7 + src/mpid/ch3/src/mpid_rma.c | 19 +- test/mpi/rma/Makefile.am | 7 + test/mpi/rma/mcs-mutex.c | 10 + test/mpi/rma/mcs-mutex.h | 1 + test/mpi/rma/req_example.c | 13 +- test/mpi/rma/testlist | 2 + test/mpi/rma/win_info.c | 3 + 16 files changed, 1023 insertions(+), 488 deletions(-) hooks/post-receive -- MPICH primary repository

1 0

[mpich] MPICH primary repository branch, master, updated. v3.0.4-389-gd698823
by noreply＠mpich.org 26 Jul '13

26 Jul '13

This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "MPICH primary repository". The branch, master has been updated via d698823cf1db9bcacb69b1456953c06ec074a460 (commit) from 55b86f37201cd9aa9e3dcbf2141504552634ba8d (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- http://git.mpich.org/mpich.git/commitdiff/d698823cf1db9bcacb69b1456953c06ec… commit d698823cf1db9bcacb69b1456953c06ec074a460 Author: Pavan Balaji <balaji(a)mcs.anl.gov> Date: Fri Jul 26 17:56:18 2013 -0500 Mention the Intel SCIF contribution in the CHANGES file. diff --git a/CHANGES b/CHANGES index a1924dd..a637edf 100644 --- a/CHANGES +++ b/CHANGES @@ -5,6 +5,9 @@ # OVERALL: Integrated MPICH-PAMI code base for Blue Gene/Q and other IBM platforms. + # OVERALL: Several improvements to the SCIF netmod. (code + contribution from Intel). + # OVERALL: Added environment variables to control a lot more capabilities for collectives. See the README.envvar file for more information. ----------------------------------------------------------------------- Summary of changes: CHANGES | 3 +++ 1 files changed, 3 insertions(+), 0 deletions(-) hooks/post-receive -- MPICH primary repository

1 0

[mpich] MPICH primary repository branch, master, updated. v3.0.4-388-g55b86f3
by noreply＠mpich.org 26 Jul '13

26 Jul '13

This is an automated email from the git hooks/post-receive script. It was generated because a ref change was pushed to the repository containing the project "MPICH primary repository". The branch, master has been updated via 55b86f37201cd9aa9e3dcbf2141504552634ba8d (commit) from fab816f0cdfdf82e44aaddc3737e1f38562bc858 (commit) Those revisions listed above that are new to this repository have not appeared on any other notification email; so we list those revisions in full, below. - Log ----------------------------------------------------------------- http://git.mpich.org/mpich.git/commitdiff/55b86f37201cd9aa9e3dcbf2141504552… commit 55b86f37201cd9aa9e3dcbf2141504552634ba8d Author: Ken Raffenetti <raffenet(a)mcs.anl.gov> Date: Fri Jul 26 13:54:49 2013 -0500 Amended write_ckpoint test. Added checking to see if ckpoint context files exist and are non-empty. diff --git a/test/mpi/ckpoint/write_ckpoint.c b/test/mpi/ckpoint/write_ckpoint.c index eadd43c..1ddd46d 100644 --- a/test/mpi/ckpoint/write_ckpoint.c +++ b/test/mpi/ckpoint/write_ckpoint.c @@ -6,12 +6,16 @@ #include "mpi.h" #include <stdio.h> +#include <unistd.h> +#include <sys/stat.h> +#include <sys/types.h> int main(int argc,char *argv[]) { int numprocs, myid, i; int namelen; char processor_name[MPI_MAX_PROCESSOR_NAME]; + struct stat fileStat; MPI_Init(&argc,&argv); MPI_Comm_size(MPI_COMM_WORLD,&numprocs); @@ -24,7 +28,13 @@ int main(int argc,char *argv[]) } if (myid == 0) { - printf("No Errors\n"); + if (stat("/tmp/context-num2-0-0",&fileStat) < 0) { + printf("failed to find ckpoint file\n"); + } else if (fileStat.st_size == 0) { + printf("ckpoint file is empty\n"); + } else { + printf("No Errors\n"); + } } MPI_Finalize(); ----------------------------------------------------------------------- Summary of changes: test/mpi/ckpoint/write_ckpoint.c | 12 +++++++++++- 1 files changed, 11 insertions(+), 1 deletions(-) hooks/post-receive -- MPICH primary repository

1 0