From: bangerth Date: Mon, 7 Jun 2010 20:47:13 +0000 (+0000) Subject: Take over many of the changes in base/ and lac/ from branch_distributed_grids. X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=86dadaf67042063c530c538529e14d0c94a68fae;p=dealii-svn.git Take over many of the changes in base/ and lac/ from branch_distributed_grids. git-svn-id: https://svn.dealii.org/trunk@21162 0785d39b-7218-0410-832d-ea1e28bc413d --- diff --git a/deal.II/base/include/base/config.h.in b/deal.II/base/include/base/config.h.in index c5bb06ecc9..52a470e0b3 100644 --- a/deal.II/base/include/base/config.h.in +++ b/deal.II/base/include/base/config.h.in @@ -241,6 +241,9 @@ functionality is available. */ #undef DEAL_II_USE_MT_POSIX_NO_BARRIERS +/* Defined if a MUMPS installation was found and is going to be used */ +#undef DEAL_II_USE_MUMPS + /* Defined if a PETSc installation was found and is going to be used */ #undef DEAL_II_USE_PETSC @@ -254,9 +257,6 @@ /* Defined if a Trilinos installation was found and is going to be used */ #undef DEAL_II_USE_TRILINOS -/* Defined if a MUMPS installation was found and is going to be used */ -#undef DEAL_II_USE_MUMPS - /* Define if vector iterators are just plain pointers */ #undef DEAL_II_VECTOR_ITERATOR_IS_POINTER @@ -606,10 +606,9 @@ * compatibility; and (ii) DEAL_II_PETSC_VERSION_GTE is used to add * functionality to the PETScWrappers that does not exist in previous * versions of PETSc. Examples of usage can be found in - * lac/source/petsc_matrix_base.h. - * - * Note: SLEPcWrappers do not need their own anological macros, since - * SLEPc and PETSc must have identical version numbers anyways. + * lac/source/petsc_matrix_base.h. Note: SLEPcWrappers do not need + * their own anological macros, since SLEPc and PETSc must have + * identical version numbers anyways. */ #define DEAL_II_PETSC_VERSION_LT(major,minor,subminor) \ ((PETSC_VERSION_MAJOR * 10000 + \ @@ -626,6 +625,7 @@ (major)*10000 + (minor)*100 + (subminor)) #include +#include /** * If the compiler supports the upcoming C++1x standard, allow us to refer diff --git a/deal.II/base/include/base/index_set.h b/deal.II/base/include/base/index_set.h index c5dad3458b..19311fefb1 100644 --- a/deal.II/base/include/base/index_set.h +++ b/deal.II/base/include/base/index_set.h @@ -224,6 +224,24 @@ class IndexSet IndexSet get_view (const unsigned int begin, const unsigned int end) const; + + /** + * Removes all elements contained in @p + * other from this set. In other words, + * if $x$ is the current object and $o$ + * the argument, then we compute $x + * \leftarrow x \backslash o$. + */ + void subtract_set (const IndexSet & other); + + + /** + * Fills the given vector with all + * indices contained in this IndexSet. + */ + void fill_index_vector(std::vector & indices) const; + + /** * Outputs a text representation of this * IndexSet to the given stream. Used for diff --git a/deal.II/base/include/base/parallel.h b/deal.II/base/include/base/parallel.h index 3c0aa8dc0b..a2daa4ba43 100644 --- a/deal.II/base/include/base/parallel.h +++ b/deal.II/base/include/base/parallel.h @@ -37,13 +37,6 @@ DEAL_II_NAMESPACE_OPEN -/** - * A namespace in which we define a few algorithms that can run in parallel - * when deal.II is configured to use multiple threads. - * - * @ingroup threads - * @author Wolfgang Bangerth, 2008, 2009 - */ namespace parallel { namespace internal diff --git a/deal.II/base/include/base/timer.h b/deal.II/base/include/base/timer.h index 43462a7ee7..26775ee1a2 100644 --- a/deal.II/base/include/base/timer.h +++ b/deal.II/base/include/base/timer.h @@ -2,7 +2,7 @@ // $Id$ // Version: $Name$ // -// Copyright (C) 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2009 by the deal.II authors +// Copyright (C) 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2009, 2010 by the deal.II authors // // This file is subject to QPL and may not be distributed // without copyright and license information. Please refer @@ -19,6 +19,7 @@ #ifdef DEAL_II_COMPILER_SUPPORTS_MPI #include +#include #endif #include @@ -89,11 +90,89 @@ class Timer * * Starts the timer at 0 sec. * + * If @p sync_wall_time is true, the wall + * time is synchronized between all CPUs + * using a MPI_Barrier() and a collective + * operation. Note that this only works + * if you stop() the timer before + * querying for the wall time. The time + * for the MPI operations are not + * included in the timing but may slow + * down your program. + * * This constructor is only available * if the deal.II compiler is an MPI * compiler. */ - Timer (MPI_Comm mpi_communicator); + Timer (MPI_Comm mpi_communicator, + bool sync_wall_time = false); + + + /** + * Structure to save collective data + * measured by this timer class. Queried + * by get_data() or printed with + * print_data() after calling stop(). + */ + struct TimeMinMaxAvg + { + double sum; + double min; + double max; + unsigned int min_index; + unsigned int max_index; + double avg; + + /** + * Set the time values to @p + * val, and the MPI rank to + * the given @p rank. + */ + void set(const double val, + const unsigned int rank) + { + sum = min = max = val; + min_index = max_index = rank; + } + +#ifdef DEAL_II_COMPILER_SUPPORTS_MPI + /** + * Given two structures + * indicating timer + * information, do the + * reduction by choosing the + * one with the longer time + * interval. This is a max + * operation and therefore a + * reduction. + * + * Arguments are passed as + * void pointers to satisfy + * the MPI requirement of + * reduction operations. + */ + static void max_reduce ( const void * in_lhs_, + void * inout_rhs_, + int * len, + MPI_Datatype * ); +#endif + }; + + /** + * Returns a reference to the data + * structure with global timing + * information. Filled after calling + * stop(). + */ + const TimeMinMaxAvg & get_data() const; + + /** + * Prints the data to the given stream. + */ + template + void print_data(STREAM & stream) const; + + #endif /** @@ -200,6 +279,14 @@ class Timer * running. */ MPI_Comm mpi_communicator; + + /** + * Store whether the wall time is + * synchronized between machines. + */ + bool sync_wall_time; + + TimeMinMaxAvg mpi_data; #endif }; @@ -333,6 +420,10 @@ class TimerOutput ConditionalOStream &stream, const enum OutputFrequency output_frequency, const enum OutputType output_type); + + + + #endif /** @@ -472,6 +563,27 @@ class TimerOutput /* ---------------- inline functions ----------------- */ +#ifdef DEAL_II_COMPILER_SUPPORTS_MPI + +const Timer::TimeMinMaxAvg & Timer::get_data() const +{ + return mpi_data; +} + +template +void Timer::print_data(STREAM & stream) const +{ + unsigned int my_id = dealii::Utilities::System::get_this_mpi_process(mpi_communicator); + if (my_id==0) + stream << mpi_data.max << " wall," + << " max @" << mpi_data.max_index + << ", min=" << mpi_data.min << " @" << mpi_data.min_index + << ", avg=" << mpi_data.avg + << std::endl; +} + +#endif + inline void TimerOutput::enter_section (const std::string §ion_name) diff --git a/deal.II/base/include/base/utilities.h b/deal.II/base/include/base/utilities.h index a0c69f54f4..ee7d56796c 100644 --- a/deal.II/base/include/base/utilities.h +++ b/deal.II/base/include/base/utilities.h @@ -2,7 +2,7 @@ // $Id$ // Version: $Name$ // -// Copyright (C) 2005, 2006, 2007, 2008, 2009 by the deal.II authors +// Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010 by the deal.II authors // // This file is subject to QPL and may not be distributed // without copyright and license information. Please refer @@ -245,7 +245,7 @@ namespace Utilities * MPI. */ bool job_supports_mpi (); - + /** * Return the number of MPI processes * there exist in the given communicator @@ -266,6 +266,51 @@ namespace Utilities */ unsigned int get_this_mpi_process (const MPI_Comm &mpi_communicator); + + /** + * Consider an unstructured + * communication pattern where + * every process in an MPI + * universe wants to send some + * data to a subset of the other + * processors. To do that, the + * other processors need to know + * who to expect messages + * from. This function computes + * this information. + * + * @param mpi_comm A communicator + * that describes the processors + * that are going to communicate + * with each other. + * + * @param destinations The list + * of processors the current + * process wants to send + * information to. This list need + * not be sorted in any way. If + * it contains duplicate entries + * that means that multiple + * messages are intended for a + * given destination. + * + * @return A list of processors + * that have indicated that they + * want to send something to the + * current processor. The + * resulting list is not + * sorted. It may contain + * duplicate entries if + * processors enter the same + * destination more than once in + * their destinations list. + */ + std::vector + compute_point_to_point_communication_pattern (const MPI_Comm & mpi_comm, + const std::vector & destinations); + + + /** * Given a communicator, generate a new * communicator that contains the same @@ -502,7 +547,7 @@ namespace Utilities */ void destroy_communicator (Epetra_Comm &communicator); - + /** * Return the number of MPI processes * there exist in the given communicator diff --git a/deal.II/base/source/index_set.cc b/deal.II/base/source/index_set.cc index ac5e5cf71f..174c9ff591 100644 --- a/deal.II/base/source/index_set.cc +++ b/deal.II/base/source/index_set.cc @@ -2,7 +2,7 @@ // $Id$ // Version: $Name$ // -// Copyright (C) 2005, 2006, 2008, 2009 by the deal.II authors +// Copyright (C) 2005, 2006, 2008, 2009, 2010 by the deal.II authors // // This file is subject to QPL and may not be distributed // without copyright and license information. Please refer @@ -13,6 +13,7 @@ #include +#include #ifdef DEAL_II_USE_TRILINOS # ifdef DEAL_II_COMPILER_SUPPORTS_MPI @@ -75,6 +76,8 @@ IndexSet::do_compress () const i != ranges.end(); ++i) { + Assert(i->begin < i->end, ExcInternalError()); + i->nth_index_in_set = next_index; next_index += (i->end - i->begin); } @@ -197,6 +200,103 @@ IndexSet::get_view (const unsigned int begin, } +void +IndexSet::subtract_set (const IndexSet & other) +{ + compress(); + other.compress(); + is_compressed = false; + + + // we save new ranges to be added to our + // IndexSet in an temporary list and add + // all of them in one go at the end. This + // is necessary because a growing ranges + // vector invalidates iterators. + std::list temp_list; + + std::vector::iterator own_it = ranges.begin(); + std::vector::iterator other_it = other.ranges.begin(); + + while (own_it != ranges.end() && other_it != other.ranges.end()) + { + //advance own iterator until we get an + //overlap + if (own_it->end <= other_it->begin) + { + ++own_it; + continue; + } + //we are done with other_it, so advance + if (own_it->begin >= other_it->end) + { + ++other_it; + continue; + } + + //Now own_it and other_it overlap. + //First save the part of own_it that is + //before other_it (if not empty). + if (own_it->begin < other_it->begin) + { + Range r(own_it->begin, other_it->begin); + r.nth_index_in_set = 0; //fix warning of unused variable + temp_list.push_back(r); + } + // change own_it to the sub range + // behind other_it. Do not delete + // own_it in any case. As removal would + // invalidate iterators, we just shrink + // the range to an empty one. + own_it->begin = other_it->end; + if (own_it->begin > own_it->end) + { + own_it->begin = own_it->end; + ++own_it; + } + + // continue without advancing + // iterators, the right one will be + // advanced next. + } + + // Now delete all empty ranges we might + // have created. + for (std::vector::iterator it = ranges.begin(); + it != ranges.end(); ) + { + if (it->begin >= it->end) + it = ranges.erase(it); + else + ++it; + } + + // done, now add the temporary ranges + for (std::list::iterator it = temp_list.begin(); + it != temp_list.end(); + ++it) + add_range(it->begin, it->end); + + compress(); +} + + +void IndexSet::fill_index_vector(std::vector & indices) const +{ + compress(); + + indices.clear(); + indices.reserve(n_elements()); + + for (std::vector::iterator it = ranges.begin(); + it != ranges.end(); + ++it) + for (unsigned int i=it->begin; iend; ++i) + indices.push_back (i); + + Assert (indices.size() == n_elements(), ExcInternalError()); +} + #ifdef DEAL_II_USE_TRILINOS @@ -217,19 +317,14 @@ IndexSet::make_trilinos_map (const MPI_Comm &communicator, #endif else { - std::vector indices; - indices.reserve(n_elements()); - for (std::vector::iterator - i = ranges.begin(); - i != ranges.end(); - ++i) - for (unsigned int j=i->begin; jend; ++j) - indices.push_back (j); - Assert (indices.size() == n_elements(), ExcInternalError()); + std::vector indices; + fill_index_vector(indices); + + int * indices_ptr = reinterpret_cast(&indices[0]); return Epetra_Map (-1, n_elements(), - &indices[0], + indices_ptr, 0, #ifdef DEAL_II_COMPILER_SUPPORTS_MPI Epetra_MpiComm(communicator)); @@ -243,5 +338,4 @@ IndexSet::make_trilinos_map (const MPI_Comm &communicator, #endif - DEAL_II_NAMESPACE_CLOSE diff --git a/deal.II/base/source/timer.cc b/deal.II/base/source/timer.cc index 8677131357..a41ce678f6 100644 --- a/deal.II/base/source/timer.cc +++ b/deal.II/base/source/timer.cc @@ -14,10 +14,12 @@ #include #include +#include #include #include #include #include +#include // these includes should probably be properly // ./configure'd using the AC_HEADER_TIME macro: @@ -46,6 +48,7 @@ Timer::Timer() cumulative_wall_time (0.) #ifdef DEAL_II_COMPILER_SUPPORTS_MPI , mpi_communicator (MPI_COMM_SELF) + , sync_wall_time (false) #endif { start(); @@ -56,11 +59,13 @@ Timer::Timer() // in case we use an MPI compiler, use // the communicator given from input #ifdef DEAL_II_COMPILER_SUPPORTS_MPI -Timer::Timer(MPI_Comm mpi_communicator) +Timer::Timer(MPI_Comm mpi_communicator, + bool sync_wall_time_) : cumulative_time (0.), cumulative_wall_time (0.), - mpi_communicator (mpi_communicator) + mpi_communicator (mpi_communicator), + sync_wall_time(sync_wall_time_) { start(); } @@ -72,6 +77,11 @@ void Timer::start () { running = true; +#ifdef DEAL_II_COMPILER_SUPPORTS_MPI + if (sync_wall_time) + MPI_Barrier(mpi_communicator); +#endif + struct timeval wall_timer; gettimeofday(&wall_timer, NULL); start_wall_time = wall_timer.tv_sec + 1.e-6 * wall_timer.tv_usec; @@ -86,6 +96,33 @@ void Timer::start () } +#ifdef DEAL_II_COMPILER_SUPPORTS_MPI + +void Timer::TimeMinMaxAvg::max_reduce ( const void * in_lhs_, + void * inout_rhs_, + int * len, + MPI_Datatype * ) +{ + const Timer::TimeMinMaxAvg * in_lhs = static_cast(in_lhs_); + Timer::TimeMinMaxAvg * inout_rhs = static_cast(inout_rhs_); + + Assert(*len==1, ExcInternalError()); + + inout_rhs->sum += in_lhs->sum; + if (inout_rhs->min>in_lhs->min) + { + inout_rhs->min = in_lhs->min; + inout_rhs->min_index = in_lhs->min_index; + } + if (inout_rhs->maxmax) + { + inout_rhs->max = in_lhs->max; + inout_rhs->max_index = in_lhs->max_index; + } +} + + +#endif double Timer::stop () { @@ -105,8 +142,50 @@ double Timer::stop () struct timeval wall_timer; gettimeofday(&wall_timer, NULL); - cumulative_wall_time += wall_timer.tv_sec + 1.e-6 * wall_timer.tv_usec - - start_wall_time; + double time = wall_timer.tv_sec + 1.e-6 * wall_timer.tv_usec + - start_wall_time; + +#ifdef DEAL_II_COMPILER_SUPPORTS_MPI + if (sync_wall_time) + { + unsigned int my_id = dealii::Utilities::System::get_this_mpi_process(mpi_communicator); + + MPI_Op op; + int ierr = MPI_Op_create((MPI_User_function *)&Timer::TimeMinMaxAvg::max_reduce, + false, &op); + AssertThrow(ierr == MPI_SUCCESS, ExcInternalError()); + + TimeMinMaxAvg in; + in.set(time, my_id); + + MPI_Datatype type; + int lengths[]={3,2}; + MPI_Aint displacements[]={0,offsetof(TimeMinMaxAvg, min_index)}; + MPI_Datatype types[]={MPI_DOUBLE, MPI_INT}; + + ierr = MPI_Type_struct(2, lengths, displacements, types, &type); + AssertThrow(ierr == MPI_SUCCESS, ExcInternalError()); + + ierr = MPI_Type_commit(&type); + + ierr = MPI_Reduce ( &in, &this->mpi_data, 1, type, op, 0, mpi_communicator ); + AssertThrow(ierr == MPI_SUCCESS, ExcInternalError()); + + ierr = MPI_Type_free (&type); + AssertThrow(ierr == MPI_SUCCESS, ExcInternalError()); + + ierr = MPI_Op_free(&op); + AssertThrow(ierr == MPI_SUCCESS, ExcInternalError()); + + this->mpi_data.avg = this->mpi_data.sum / dealii::Utilities::System::get_n_mpi_processes(mpi_communicator); + + cumulative_wall_time += this->mpi_data.max; + } + else + cumulative_wall_time += time; +#else + cumulative_wall_time += time; +#endif } return cumulative_time; } diff --git a/deal.II/base/source/utilities.cc b/deal.II/base/source/utilities.cc index e543ee3838..ab8e94d807 100644 --- a/deal.II/base/source/utilities.cc +++ b/deal.II/base/source/utilities.cc @@ -2,7 +2,7 @@ // $Id$ // Version: $Name$ // -// Copyright (C) 2005, 2006, 2008, 2009 by the deal.II authors +// Copyright (C) 2005, 2006, 2008, 2009, 2010 by the deal.II authors // // This file is subject to QPL and may not be distributed // without copyright and license information. Please refer @@ -516,6 +516,74 @@ namespace Utilities } + std::vector + compute_point_to_point_communication_pattern (const MPI_Comm & mpi_comm, + const std::vector & destinations) + { + unsigned int myid = Utilities::System::get_this_mpi_process(mpi_comm); + unsigned int n_procs = Utilities::System::get_n_mpi_processes(mpi_comm); + + for (unsigned int i=0; i my_destinations(max_n_destinations, + numbers::invalid_unsigned_int); + std::copy (destinations.begin(), destinations.end(), + my_destinations.begin()); + + // now exchange these (we could + // communicate less data if we + // used MPI_Allgatherv, but + // we'd have to communicate + // my_n_destinations to all + // processors in this case, + // which is more expensive than + // the reduction operation + // above in MPI_Allreduce) + std::vector all_destinations (max_n_destinations * n_procs); + MPI_Allgather (&my_destinations[0], max_n_destinations, MPI_UNSIGNED, + &all_destinations[0], max_n_destinations, MPI_UNSIGNED, + mpi_comm); + + // now we know who is going to + // communicate with + // whom. collect who is going + // to communicate with us! + std::vector origins; + for (unsigned int i=0; i&6; } $as_echo "---------------- configuring additional libs ----------------" >&6; } - { $as_echo "$as_me:$LINENO: checking for PETSc library directory" >&5 $as_echo_n "checking for PETSc library directory... " >&6; } @@ -14011,6 +14032,117 @@ done + { $as_echo "$as_me:$LINENO: checking for MUMPS library directory" >&5 +$as_echo_n "checking for MUMPS library directory... " >&6; } + +# Check whether --with-mumps was given. +if test "${with_mumps+set}" = set; then + withval=$with_mumps; if test "x$withval" = "xno" ; then + { $as_echo "$as_me:$LINENO: result: explicitly disabled" >&5 +$as_echo "explicitly disabled" >&6; } + USE_CONTRIB_MUMPS=no + else + USE_CONTRIB_MUMPS=yes + DEAL_II_MUMPS_DIR="$withval" + { $as_echo "$as_me:$LINENO: result: $DEAL_II_MUMPS_DIR" >&5 +$as_echo "$DEAL_II_MUMPS_DIR" >&6; } + if test ! -d $DEAL_II_MUMPS_DIR \ + -o ! -d $DEAL_II_MUMPS_DIR/include \ + -o ! -d $DEAL_II_MUMPS_DIR/lib \ + ; then + { { $as_echo "$as_me:$LINENO: error: Path to MUMPS specified with --with-mumps does not point to a complete MUMPS installation" >&5 +$as_echo "$as_me: error: Path to MUMPS specified with --with-mumps does not point to a complete MUMPS installation" >&2;} + { (exit 1); exit 1; }; } + fi + fi + +else + USE_CONTRIB_MUMPS=no + +fi + + + if test "$USE_CONTRIB_MUMPS" = "yes" ; then + + { $as_echo "$as_me:$LINENO: checking for SCALAPACK library directory" >&5 +$as_echo_n "checking for SCALAPACK library directory... " >&6; } + +# Check whether --with-scalapack was given. +if test "${with_scalapack+set}" = set; then + withval=$with_scalapack; DEAL_II_SCALAPACK_DIR="$withval" + { $as_echo "$as_me:$LINENO: result: $DEAL_II_SCALAPACK_DIR" >&5 +$as_echo "$DEAL_II_SCALAPACK_DIR" >&6; } + if test ! -d $DEAL_II_SCALAPCK_DIR ; then + { { $as_echo "$as_me:$LINENO: error: The path to SCALAPACK specified with --with-scalapack does t point to a complete SCALAPACK installation" >&5 +$as_echo "$as_me: error: The path to SCALAPACK specified with --with-scalapack does t point to a complete SCALAPACK installation" >&2;} + { (exit 1); exit 1; }; } + fi + +else + { { $as_echo "$as_me:$LINENO: error: If MUMPS is used, the path to SCALAPACK must be specified with --with-scalapack" >&5 +$as_echo "$as_me: error: If MUMPS is used, the path to SCALAPACK must be specified with --with-scalapack" >&2;} + { (exit 1); exit 1; }; } + +fi + + + { $as_echo "$as_me:$LINENO: checking for BLACS library directory" >&5 +$as_echo_n "checking for BLACS library directory... " >&6; } + +# Check whether --with-blacs was given. +if test "${with_blacs+set}" = set; then + withval=$with_blacs; DEAL_II_BLACS_DIR="$withval" + { $as_echo "$as_me:$LINENO: result: $DEAL_II_BLACS_DIR" >&5 +$as_echo "$DEAL_II_BLACS_DIR" >&6; } + if test ! -d $DEAL_II_BLACS_DIR \ + -o ! -d $DEAL_II_BLACS_DIR/LIB ; then + { { $as_echo "$as_me:$LINENO: error: The path to BLACS specified with --with-blacs does not point to a complete BLACS installation" >&5 +$as_echo "$as_me: error: The path to BLACS specified with --with-blacs does not point to a complete BLACS installation" >&2;} + { (exit 1); exit 1; }; } + fi + +else + { { $as_echo "$as_me:$LINENO: error: If MUMPS is used, the path to BLACS must be specified with --with-blacs" >&5 +$as_echo "$as_me: error: If MUMPS is used, the path to BLACS must be specified with --with-blacs" >&2;} + { (exit 1); exit 1; }; } + +fi + + + { $as_echo "$as_me:$LINENO: checking for BLACS library architecture" >&5 +$as_echo_n "checking for BLACS library architecture... " >&6; } + BLACS_COMM=`cat $DEAL_II_BLACS_DIR/Bmake.inc \ + | grep "COMMLIB = " \ + | perl -pi -e 's/.*LIB =\s+//g;'` + BLACS_PLAT=`cat $DEAL_II_BLACS_DIR/Bmake.inc \ + | grep "PLAT = " \ + | perl -pi -e 's/.*PLAT =\s+//g;'` + BLACS_DEBUG=`cat $DEAL_II_BLACS_DIR/Bmake.inc \ + | grep "BLACSDBGLVL = " \ + | perl -pi -e 's/.*DBGLVL =\s+//g;'` + DEAL_II_BLACS_ARCH="$BLACS_COMM-$BLACS_PLAT-$BLACS_DEBUG" + { $as_echo "$as_me:$LINENO: result: $DEAL_II_BLACS_ARCH" >&5 +$as_echo "$DEAL_II_BLACS_ARCH" >&6; } + + fi + + if test "$USE_CONTRIB_MUMPS" = "yes" ; then + +cat >>confdefs.h <<\_ACEOF +#define DEAL_II_USE_MUMPS 1 +_ACEOF + + DEAL_II_DEFINE_DEAL_II_USE_MUMPS=DEAL_II_USE_MUMPS + if test "x$with_mumps" = "x" ; then + with_mumps="yes" + fi + fi + + + + + + if test "x$with_umfpack" != "x" -a "x$with_umfpack" != "xno" ; then @@ -16662,6 +16794,7 @@ fi + for ac_func in daxpy_ saxpy_ dgemv_ sgemv_ dgeev_ sgeev_ dgeevx_ sgeevx_ do as_ac_var=`$as_echo "ac_cv_func_$ac_func" | $as_tr_sh` diff --git a/deal.II/configure.in b/deal.II/configure.in index f34464ffaa..8ff1210d02 100644 --- a/deal.II/configure.in +++ b/deal.II/configure.in @@ -433,7 +433,7 @@ AC_SUBST(USE_CONTRIB_MUMPS) AC_SUBST(DEAL_II_MUMPS_DIR) AC_SUBST(DEAL_II_SCALAPACK_DIR) dnl MUMPS dependency AC_SUBST(DEAL_II_BLACS_DIR) dnl MUMPS dependency -AC_SUBST(DEAL_II_BLACS_ARCH) +AC_SUBST(DEAL_II_BLACS_ARCH) AC_SUBST(DEAL_II_DEFINE_DEAL_II_USE_MUMPS) dnl Make sure we configure for libraries used by other libraries. For @@ -727,6 +727,7 @@ AH_BOTTOM( (major)*10000 + (minor)*100 + (subminor)) #include +#include /** * If the compiler supports the upcoming C++1x standard, allow us to refer diff --git a/deal.II/doc/doxygen/headers/glossary.h b/deal.II/doc/doxygen/headers/glossary.h index cea9638436..dd2bc097ee 100644 --- a/deal.II/doc/doxygen/headers/glossary.h +++ b/deal.II/doc/doxygen/headers/glossary.h @@ -18,7 +18,7 @@ * documentation of classes of deal.II. The glossary often only gives * a microscopic view of a particular concept; if you struggle with * the bigger picture, it may therefore also be worth to consult the - * global overview of classes on the main/@ref index page. + * global overview of classes on the @ref index page. * *
* @@ -26,6 +26,29 @@ *
Mesh cells not refined any further in the hierarchy.
* * + * + *
@anchor GlossArtificialCell Artificial cells
+ *
+ * If a mesh is distributed across multiple MPI processes using the + * parallel::distributed::Triangulation class, each processor stores + * only the cells it owns, one layer of adjacent cells that are owned + * by other processors (called @ref GlossGhostCell "ghost cells"), all coarse level + * cells, and all cells that are necessary to maintain the invariant + * that adjacent cells must differ by at most one refinement + * level. The cells stored on each process that are not owned by this + * process and that are not ghost cells are called "artificial cells", + * and for these cells the predicate + * cell-@>is_artificial() returns true. Artificial cells + * are guaranteed to exist in the globally distributed mesh but they + * may be further refined on other processors. See the + * @ref distributed_paper "Distributed Computing paper" for more + * information. + * + * The concept of artificial cells has no meaning for triangulations + * that store the entire mesh on each processor, i.e. the + * dealii::Triangulation class.
+ * + * *
@anchor GlossBlockLA Block (linear algebra)
*
It is often convenient to treat a matrix or vector as a collection of @@ -267,10 +290,14 @@ * * The way out of a situation like this is to use one of the two following * ways: + * * - You tell the object that you want to compress what operation is - * intended. The TrilinosWrappers::VectorBase::compress() can take such an - * additional argument. Or, - * - You do a fake addition or set operation on the object in question. + * intended. The TrilinosWrappers::VectorBase::compress() can take + * such an additional argument. + * - You do a fake addition or set operation on the object in question. For + * example, you can add a zero to an element of the matrix or vector, + * which has no effect other than telling the object that the next + * compress operation should be in Add mode. * * Some of the objects are also indifferent and can figure out what to * do without being told. The TrilinosWrappers::SparseMatrix can do that, @@ -391,6 +418,22 @@ * flag. * * + *
@anchor distributed_paper + * Distributed computing paper
+ + *
The "distributed computing paper" is a paper by W. Bangerth, + * C. Burstedde, T. Heister and M. Kronbichler titled "Algorithms and Data + * Structures for Massively Parallel Generic Finite Element Codes" that + * described the implementation of parallel distributed computing in deal.II, + * i.e. computations where not only the linear system is split onto different + * machines as in, for example, step-18, but also the Triangulation and + * DoFHandler objects. In essence, it is a guide to the parallel::distributed + * namespace. + * + * The paper is currently in preparation. + *
+ * + * *
@anchor GlossFaceOrientation Face orientation
*
In a triangulation, the normal vector to a face * can be deduced from the face orientation by @@ -448,6 +491,28 @@ *
* * + *
@anchor GlossGhostCell Ghost cells
+ *
+ * If a mesh is distributed across multiple MPI processes using the + * parallel::distributed::Triangulation class, each processor stores + * only the cells it owns, one layer of adjacent cells that are owned + * by other processors, all coarse level cells, and all cells that are + * necessary to maintain the invariant that adjacent cells must differ + * by at most one refinement level. The cells stored on each process + * that are not owned by this process but that are adjacent to the + * ones owned by this process are called "ghost cells", and for these + * cells the predicate cell-@>is_ghost() returns + * true. Ghost cells are guaranteed to exist in the globally + * distributed mesh, i.e. these cells are actually owned by another + * process and are not further refined there. See the + * @ref distributed_paper "Distributed Computing paper" for more + * information. + * + * The concept of ghost cells has no meaning for triangulations that + * store the entire mesh on each processor, i.e. the + * dealii::Triangulation class.
+ * + * *
@anchor hp_paper %hp paper
*
The "hp paper" is a paper by W. Bangerth and O. Kayser-Herold, titled * "Data Structures and Requirements for hp Finite Element Software", that @@ -577,8 +642,9 @@ Article{JK10, * element shape functions are defined.
* * - *
@anchor GlossShape Shape functions
The restriction of - * the finite element basis functions to a single grid cell.
+ *
@anchor GlossShape Shape functions
+ *
The restriction of the finite element basis functions to a single + * grid cell.
* * *
@anchor GlossSubdomainId Subdomain id
@@ -600,8 +666,26 @@ Article{JK10, * coincides with the rank of the MPI process within the MPI * communicator). Partitioning is typically done using the * GridTools::partition() function, but any other method can also be used to - * do this though most other ideas will likely lead to less well balanced + * do this though most simple ideas will likely lead to less well balanced * numbers of degrees of freedom on the various subdomains. + * + * On the other hand, for programs that are parallelized using MPI but + * where meshes are held distributed across several processors using + * the parallel::distributed::Triangulation and + * parallel::distributed::DoFHandler classes, the subdomain id of + * cells are tied to the processor that owns the cell. In other words, + * querying the subdomain id of a cell tells you if the cell is owned + * by the current processor (i.e. if cell-@>subdomain_id() == + * triangulation.parallel::distributed::Triangulation::locally_owned_subdomain()) + * or by another processor. In the parallel distributed case, + * subdomain ids are only assigned to cells that the current processor + * owns as well as the immediately adjacent @ref GlossGhostCell "ghost cells". + * Cells further away are held on each processor to ensure + * that every MPI process has access to the full coarse grid as well + * as to ensure the invariant that neighboring cells differ by at most + * one refinement level. These cells are called "artificial" (see + * @ref GlossArtificialCell "here") and have the special subdomain id value + * types::artificial_subdomain_id. * * * diff --git a/deal.II/doc/doxygen/headers/multithreading.h b/deal.II/doc/doxygen/headers/multithreading.h index 4a3368bb64..895fee0116 100644 --- a/deal.II/doc/doxygen/headers/multithreading.h +++ b/deal.II/doc/doxygen/headers/multithreading.h @@ -12,7 +12,8 @@ //------------------------------------------------------------------------- /** - * @defgroup threads Parallel computing with multiple processors + * @defgroup threads Parallel computing with multiple processors accessing the shared memory + * @ingroup Parallel * * @brief A module discussing the use of parallelism on shared memory * machines. See the detailed documentation and diff --git a/deal.II/doc/doxygen/headers/parallel.h b/deal.II/doc/doxygen/headers/parallel.h new file mode 100644 index 0000000000..b61f9f58ea --- /dev/null +++ b/deal.II/doc/doxygen/headers/parallel.h @@ -0,0 +1,36 @@ +//------------------------------------------------------------------------- +// $Id$ +// Version: $Name$ +// +// Copyright (C) 2009 by the deal.II authors +// +// This file is subject to QPL and may not be distributed +// without copyright and license information. Please refer +// to the file deal.II/doc/license.html for the text and +// further information on this license. +// +//------------------------------------------------------------------------- + +/** + * @defgroup Parallel Parallel computing + * + * @brief A module discussing the use of multiple processor. + * + * This module contains information on %parallel computing. It is + * subdivided into parts on @ref threads and on @ref distributed. + */ + + +/** + * A namespace in which we define classes and algorithms that deal + * with running in %parallel on shared memory machines when deal.II is + * configured to use multiple threads (see @ref threads), as well as + * running things in %parallel on %distributed memory machines (see @ref + * distributed). + * + * @ingroup threads + * @author Wolfgang Bangerth, 2008, 2009 + */ +namespace parallel +{ +} diff --git a/deal.II/doc/news/changes.h b/deal.II/doc/news/changes.h index 9d0d4e17b2..2dbe0009f3 100644 --- a/deal.II/doc/news/changes.h +++ b/deal.II/doc/news/changes.h @@ -277,6 +277,19 @@ inconvenience this causes.

base

    +

  1. New: The Timer class can now accumulate and average run times of + pieces of code across multiple MPI processes. +
    + (Timo Heister 2010/06/07) +

    2. + +

  2. New: The Utilities::System::compute_point_to_point_communication_pattern + function can be used to compute who wants to send messages to the + current processor in unstructured point-to-point MPI communications. +
    + (WB 2010/06/07) +

    3. +

  3. New: The DataOutBase class (and all derived classes such as DataOut, MatrixOut, etc) can now produce the XML-based version of the VTK file format (the so-called VTU format). Furthermore, the @@ -436,10 +449,30 @@ inconvenience this causes.

    lac

      +

    1. New: The ConstraintMatrix class can now handle storing only + a subset of all constraints, for example only for degrees of + freedom that are relevant for the subdomain that is owned by one + process in an MPI universe. +
      + (Timo Heister, Martin Kronbichler 2010/06/07) +

      2. + +

    2. New: The PETScWrappers::MPI::Vector and TrilinosWrappers::MPI::Vector + classes can now handle ghost elements, i.e. elements that are not + owned by the current processor but are available for reading + anyway. The simplest form of ghosting would be to simply import + an entire vector to local memory, but the new function allow to + select the elements we need to support the case of computations + where importing all elements of even a single vector would + exceed available memory. +
      + (Timo Heister 2010/06/07) +

      3. +

    3. - New: A class SparseDirectMumps that provides an interface to - the MUltifrontal Massively Parallel sparse direct Solver (MUMPS). + New: A class SparseDirectMumps that provides an interface to + the MUltifrontal Massively Parallel sparse direct Solver (MUMPS).


      (Markus Buerg 2010/05/10) diff --git a/deal.II/doc/readme-petsc-trilinos.html b/deal.II/doc/readme-petsc-trilinos.html index 1484cf759a..a426571f59 100644 --- a/deal.II/doc/readme-petsc-trilinos.html +++ b/deal.II/doc/readme-petsc-trilinos.html @@ -200,6 +200,16 @@ make install want to use Trilinos with MPI on parallel machines, you also need to flip the value of the TPL_ENABLE_MPI flag above.

      +

      + Note: if the deal.II ./configure reports an error related to + HAVE_INTTYPES_H, edit /include/ml_config.h and comment out the + line + + 

      +#define HAVE_INTTYPES_H
+
      + +

      Configuring for installed Trilinos packages

      diff --git a/deal.II/lac/include/lac/block_matrix_base.h b/deal.II/lac/include/lac/block_matrix_base.h index d336d1d94c..8c63c06eea 100644 --- a/deal.II/lac/include/lac/block_matrix_base.h +++ b/deal.II/lac/include/lac/block_matrix_base.h @@ -97,7 +97,7 @@ struct IsBlockMatrix * indicates whether the template * argument to this class is a block * matrix (in fact whether the type is - * derived from BlockMatrix). + * derived from BlockMatrixBase). */ static const bool value = (sizeof(check_for_block_matrix ((MatrixType*)0)) diff --git a/deal.II/lac/include/lac/block_vector_base.h b/deal.II/lac/include/lac/block_vector_base.h index 632b8d1e6f..4bd235721b 100644 --- a/deal.II/lac/include/lac/block_vector_base.h +++ b/deal.II/lac/include/lac/block_vector_base.h @@ -33,6 +33,69 @@ DEAL_II_NAMESPACE_OPEN *@{ */ +template class BlockVectorBase; + + +/** + * A class that can be used to determine whether a given type is a block + * vector type or not. For example, + * @code + * IsBlockVector >::value + * @endcode + * has the value false, whereas + * @code + * IsBlockVector >::value + * @endcode + * is true. This is sometimes useful in template contexts where we may + * want to do things differently depending on whether a template type + * denotes a regular or a block vector type. + * + * @author Wolfgang Bangerth, 2010 + */ +template +struct IsBlockVector +{ + private: + struct yes_type { char c[1]; }; + struct no_type { char c[2]; }; + + /** + * Overload returning true if the class + * is derived from BlockVectorBase, + * which is what block vectors do. + */ + template + static yes_type check_for_block_vector (const BlockVectorBase *); + + /** + * Catch all for all other potential + * vector types that are not block + * matrices. + */ + static no_type check_for_block_vector (...); + + public: + /** + * A statically computable value that + * indicates whether the template + * argument to this class is a block + * vector (in fact whether the type is + * derived from BlockVectorBase). + */ + static const bool value = (sizeof(check_for_block_vector + ((VectorType*)0)) + == + sizeof(yes_type)); +}; + + +// instantiation of the static member +template +const bool IsBlockVector::value; + + + + namespace internal { diff --git a/deal.II/lac/include/lac/constraint_matrix.h b/deal.II/lac/include/lac/constraint_matrix.h index 9e745bb0cb..dd7a32f0c9 100644 --- a/deal.II/lac/include/lac/constraint_matrix.h +++ b/deal.II/lac/include/lac/constraint_matrix.h @@ -344,15 +344,28 @@ class ConstraintMatrix : public Subscriptor ConstraintMatrix (const ConstraintMatrix &constraint_matrix); /** - * Reinit the ConstraintMatrix - * object. This function is only relevant - * in the distributed case, to supply a - * different IndexSet. Otherwise this - * routine is equivalent to calling - * clear(). + * Reinit the ConstraintMatrix object and + * supply an IndexSet with lines that may + * be constrained. This function is only + * relevant in the distributed case, to + * supply a different IndexSet. Otherwise + * this routine is equivalent to calling + * clear(). Normally an IndexSet with all + * locally_active_dofs should be supplied + * here. */ void reinit (const IndexSet & local_constraints = IndexSet()); + /** + * Determines if we can store a + * constraint for the given @p + * line_index. This routine only matters + * in the distributed case and checks if + * the IndexSet allows storage of this + * line. Always returns true if not in + * the distributed case. + */ + bool can_store_line(unsigned int line_index) const; /** * This function copies the content of @p @@ -2093,7 +2106,9 @@ void ConstraintMatrix:: lines[lines_cache[calculate_line_index(*local_indices_begin)]]; for (unsigned int j=0; j &local_vector, lines[lines_cache[calculate_line_index(local_dof_indices[j])]]; for (unsigned int q=0; q &local_vector, // the entries of fixed dofs for (unsigned int j=0; jentries.size(); ++j) { - Assert (is_constrained(position->entries[j].first) == false, + Assert (!(!local_lines.size() + || local_lines.is_element(position->entries[j].first)) + || is_constrained(position->entries[j].first) == false, ExcMessage ("Tried to distribute to a fixed dof.")); global_vector(position->entries[j].first) += local_vector(i) * position->entries[j].second; diff --git a/deal.II/lac/include/lac/petsc_parallel_vector.h b/deal.II/lac/include/lac/petsc_parallel_vector.h index d05859da3c..337f94429e 100644 --- a/deal.II/lac/include/lac/petsc_parallel_vector.h +++ b/deal.II/lac/include/lac/petsc_parallel_vector.h @@ -2,7 +2,7 @@ // $Id$ // Version: $Name$ // -// Copyright (C) 2004, 2005, 2006, 2007, 2009 by the deal.II authors +// Copyright (C) 2004, 2005, 2006, 2007, 2009, 2010 by the deal.II authors // // This file is subject to QPL and may not be distributed // without copyright and license information. Please refer @@ -22,6 +22,7 @@ # include # include # include +# include DEAL_II_NAMESPACE_OPEN @@ -29,6 +30,7 @@ DEAL_II_NAMESPACE_OPEN // forward declaration template class Vector; +class IndexSet; /*! @addtogroup PETScWrappers @@ -186,6 +188,7 @@ namespace PETScWrappers const unsigned int n, const unsigned int local_size); + /** * Copy-constructor from deal.II * vectors. Sets the dimension to that @@ -206,6 +209,7 @@ namespace PETScWrappers const dealii::Vector &v, const unsigned int local_size); + /** * Copy-constructor the * values from a PETSc wrapper vector @@ -224,6 +228,29 @@ namespace PETScWrappers const VectorBase &v, const unsigned int local_size); + + /** + * Constructs a new parallel PETSc + * vector from an Indexset. Note that + * @p local must be contiguous and + * the global size of the vector is + * determined by local.size(). The + * global indices in @p ghost are + * sluppied as ghost indices that can + * also be read locally. Note that + * the @p ghost IndexSet may be empty + * and that any indices already + * contained in @p local are ignored + * during construction. That way you + * can construct with + * locally_relevent_dofs() for + * example. + */ + explicit Vector (const MPI_Comm &communicator, + const IndexSet & local, + const IndexSet & ghost = IndexSet(0)); + + /** * Copy the given vector. Resize the * present vector if necessary. Also @@ -232,6 +259,7 @@ namespace PETScWrappers */ Vector & operator = (const Vector &v); + /** * Copy the given sequential * (non-distributed) vector @@ -346,6 +374,16 @@ namespace PETScWrappers void reinit (const Vector &v, const bool fast = false); + /** + * Reinit as a ghosted vector. See + * constructor with same signature + * for more details. + */ + void reinit (const MPI_Comm &communicator, + const IndexSet & local, + const IndexSet & ghost = IndexSet(0)); + + /** * Return a reference to the MPI * communicator object in use with @@ -366,6 +404,21 @@ namespace PETScWrappers virtual void create_vector (const unsigned int n, const unsigned int local_size); + + + /** + * Create a vector of global length + * @p n, local size @p local_size and + * with the specified ghost + * indices. Note that you need to + * call update_ghost_values() before + * accessing those. + */ + virtual void create_vector (const unsigned int n, + const unsigned int local_size, + const IndexSet & ghostnodes); + + private: /** * Copy of the communicator object to diff --git a/deal.II/lac/include/lac/petsc_vector_base.h b/deal.II/lac/include/lac/petsc_vector_base.h index 30a2036778..919b2fccad 100644 --- a/deal.II/lac/include/lac/petsc_vector_base.h +++ b/deal.II/lac/include/lac/petsc_vector_base.h @@ -25,6 +25,7 @@ # include # include +# include DEAL_II_NAMESPACE_OPEN @@ -656,6 +657,14 @@ namespace PETScWrappers void ratio (const VectorBase &a, const VectorBase &b); + /** + * Updates the ghost values of this + * vector. This is necessary after any + * modification before reading ghost + * values. + */ + void update_ghost_values() const; + /** * Print to a * stream. @p precision denotes @@ -728,6 +737,23 @@ namespace PETScWrappers */ Vec vector; + /** + * Denotes if this vector has ghost + * indices associated with it. This + * means that at least one of the + * processes in a parallel programm has + * at least one ghost index. + */ + bool ghosted; + + /** + * This vector contains the global + * indices of the ghost values. The + * location in this vector denotes the + * local numbering, which is used in + * PETSc. + */ + IndexSet ghost_indices; /** * PETSc doesn't allow to mix additions @@ -783,6 +809,8 @@ namespace PETScWrappers const unsigned int *indices, const PetscScalar *values, const bool add_values); + + }; diff --git a/deal.II/lac/include/lac/sparsity_tools.h b/deal.II/lac/include/lac/sparsity_tools.h index 70f0a7ea46..442ac45bf7 100644 --- a/deal.II/lac/include/lac/sparsity_tools.h +++ b/deal.II/lac/include/lac/sparsity_tools.h @@ -2,7 +2,7 @@ // $Id$ // Version: $Name$ // -// Copyright (C) 2008, 2009 by the deal.II authors +// Copyright (C) 2008, 2009, 2010 by the deal.II authors // // This file is subject to QPL and may not be distributed // without copyright and license information. Please refer @@ -19,6 +19,11 @@ #include +#ifdef DEAL_II_COMPILER_SUPPORTS_MPI +#include +#include +#endif + DEAL_II_NAMESPACE_OPEN class SparsityPattern; @@ -190,6 +195,33 @@ namespace SparsityTools std::vector &new_indices, const std::vector &starting_indices = std::vector()); + +#ifdef DEAL_II_COMPILER_SUPPORTS_MPI + /** + * Communciate rows in a compressed + * sparsity pattern over MPI. The @param + * csp is modified inline. All entries in + * rows that belong to a different + * processor are send to them and added + * there. The ownership is determined by + * parameter @param rows_per_cpu. The + * IndexSet @param myrange should be the + * one used in the constructor of the + * CompressedSimpleSparsityPattern. All + * rows contained in @param myrange are + * checked in @param csp. This function + * needs to be used with + * PETScWrappers::MPI::SparseMatrix for it + * to work correctly. + */ + template + void distribute_sparsity_pattern(CSP_t & csp, + const std::vector & rows_per_cpu, + const MPI_Comm & mpi_comm, + const IndexSet & myrange); +#endif + + /** * Exception */ diff --git a/deal.II/lac/source/constraint_matrix.cc b/deal.II/lac/source/constraint_matrix.cc index 9c43849525..14bcc92661 100644 --- a/deal.II/lac/source/constraint_matrix.cc +++ b/deal.II/lac/source/constraint_matrix.cc @@ -1944,6 +1944,16 @@ template<> void ConstraintMatrix::distribute (TrilinosWrappers::MPI::Vector &vec) const { + //TODO: not implemented yet, we need to fix + //LocalRange() first to only include + //"owned" indices. For this we need to keep + //track of the owned indices, because + //Trilinos doesn't. Use same constructor + //interface as in PETSc with two IndexSets! + AssertThrow (vec.vector_partitioner().IsOneToOne(), + ExcMessage ("Distribute does not work on vectors with overlapping parallel partitioning.")); + + typedef std::vector::const_iterator constraint_iterator; ConstraintLine index_comparison; index_comparison.line = vec.local_range().first; @@ -2011,6 +2021,8 @@ ConstraintMatrix::distribute (TrilinosWrappers::MPI::Vector &vec) const it->entries[i].second); vec(it->line) = new_value; } + + vec.compress (); } @@ -2103,11 +2115,99 @@ ConstraintMatrix::distribute (TrilinosWrappers::MPI::BlockVector &vec) const vec(it->line) = new_value; } } + + vec.compress (); +} + +#endif + +#ifdef DEAL_II_USE_PETSC + + // this is a specialization for a + // parallel (non-block) PETSc + // vector. The basic idea is to just work + // on the local range of the vector. But + // we need access to values that the + // local nodes are constrained to. + +template<> +void +ConstraintMatrix::distribute (PETScWrappers::MPI::Vector &vec) const +{ + typedef std::vector::const_iterator constraint_iterator; + ConstraintLine index_comparison; + index_comparison.line = vec.local_range().first; + const constraint_iterator begin_my_constraints = + std::lower_bound (lines.begin(),lines.end(),index_comparison); + + index_comparison.line = vec.local_range().second; + const constraint_iterator end_my_constraints + = std::lower_bound(lines.begin(),lines.end(),index_comparison); + + // all indices we need to read from + IndexSet my_indices (vec.size()); + + const std::pair + local_range = vec.local_range(); + + my_indices.add_range (local_range.first, local_range.second); + + std::set individual_indices; + for (constraint_iterator it = begin_my_constraints; + it != end_my_constraints; ++it) + for (unsigned int i=0; ientries.size(); ++i) + if ((it->entries[i].first < local_range.first) + || + (it->entries[i].first >= local_range.second)) + individual_indices.insert (it->entries[i].first); + + my_indices.add_indices (individual_indices.begin(), + individual_indices.end()); + + IndexSet local_range_is (vec.size()); + local_range_is.add_range(local_range.first, local_range.second); + + + // create a vector and import those indices + PETScWrappers::MPI::Vector ghost_vec(vec.get_mpi_communicator(), + local_range_is, + my_indices); + ghost_vec = vec; + ghost_vec.update_ghost_values(); + + // finally do the distribution on own + // constraints + for (constraint_iterator it = begin_my_constraints; + it != end_my_constraints; ++it) + { + // fill entry in line + // next_constraint.line by adding the + // different contributions + double new_value = it->inhomogeneity; + for (unsigned int i=0; ientries.size(); ++i) + new_value += (ghost_vec(it->entries[i].first) * + it->entries[i].second); + vec(it->line) = new_value; + } + + // force every processor to write something + vec(local_range.first) = vec(local_range.first); + + vec.compress (); +} + + +template<> +void +ConstraintMatrix::distribute (PETScWrappers::MPI::BlockVector &/*vec*/) const +{ + AssertThrow (false, ExcNotImplemented()); } #endif + unsigned int ConstraintMatrix::n_constraints () const { return lines.size(); diff --git a/deal.II/lac/source/petsc_parallel_sparse_matrix.cc b/deal.II/lac/source/petsc_parallel_sparse_matrix.cc index 02328ba853..ac4163e11d 100644 --- a/deal.II/lac/source/petsc_parallel_sparse_matrix.cc +++ b/deal.II/lac/source/petsc_parallel_sparse_matrix.cc @@ -2,7 +2,7 @@ // $Id$ // Version: $Name$ // -// Copyright (C) 2004, 2005, 2006, 2008, 2009 by the deal.II authors +// Copyright (C) 2004, 2005, 2006, 2008, 2009, 2010 by the deal.II authors // // This file is subject to QPL and may not be distributed // without copyright and license information. Please refer @@ -302,6 +302,13 @@ namespace PETScWrappers } const unsigned int local_row_end = local_row_start + local_rows_per_process[this_process]; + +#if DEAL_II_PETSC_VERSION_LT(2,3,3) + //old version to create the matrix, we + //can skip calculating the row length + //at least starting from 2.3.3 (tested, + //see below) + const unsigned int local_col_end = local_col_start + local_columns_per_process[this_process]; @@ -345,6 +352,27 @@ namespace PETScWrappers &matrix); AssertThrow (ierr == 0, ExcPETScError(ierr)); +#else //PETSC_VERSION>=2.3.3 + // new version to create the matrix. We + // do not set row length but set the + // correct SparsityPattern later. + int ierr; + + ierr = MatCreate(communicator,&matrix); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + + ierr = MatSetSizes(matrix, + local_rows_per_process[this_process], + local_columns_per_process[this_process], + sparsity_pattern.n_rows(), + sparsity_pattern.n_cols()); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + + ierr = MatSetType(matrix,MATMPIAIJ); + AssertThrow (ierr == 0, ExcPETScError(ierr)); +#endif + + // next preset the exact given matrix // entries with zeros, if the user // requested so. this doesn't avoid any @@ -422,13 +450,24 @@ namespace PETScWrappers // now copy over the information // from the sparsity pattern. { - unsigned int index=0; +#ifdef PETSC_USE_64BIT_INDICES + PetscInt +#else + int +#endif + * ptr = & colnums_in_window[0]; + for (unsigned int i=local_row_start; i ghostindices; + ghostnodes.fill_index_vector(ghostindices); + + const PetscInt * ptr= (const PetscInt*)(&(ghostindices[0])); + + int ierr + = VecCreateGhost(communicator, + local_size, + PETSC_DETERMINE, + ghostindices.size(), + ptr, + &vector); + + AssertThrow (ierr == 0, ExcPETScError(ierr)); + + Assert (size() == n, + ExcDimensionMismatch (size(), n)); + +#if DEBUG + // test ghost allocation in debug mode + +#ifdef PETSC_USE_64BIT_INDICES + PetscInt +#else + int +#endif + begin, end; + + ierr = VecGetOwnershipRange (vector, &begin, &end); + + Assert(local_size==(unsigned int)(end-begin), ExcInternalError()); + + Vec l; + ierr = VecGhostGetLocalForm(vector, &l); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + + PetscInt lsize; + ierr = VecGetSize(l, &lsize); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + + ierr = VecGhostRestoreLocalForm(vector, &l); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + + Assert( lsize==end-begin+(PetscInt)ghost_indices.n_elements() ,ExcInternalError()); + +#endif + + } + + +} + } DEAL_II_NAMESPACE_CLOSE diff --git a/deal.II/lac/source/petsc_vector_base.cc b/deal.II/lac/source/petsc_vector_base.cc index ff60b8fa12..53c1df3e70 100644 --- a/deal.II/lac/source/petsc_vector_base.cc +++ b/deal.II/lac/source/petsc_vector_base.cc @@ -2,7 +2,7 @@ // $Id$ // Version: $Name$ // -// Copyright (C) 2004, 2005, 2006, 2007, 2008, 2009 by the deal.II authors +// Copyright (C) 2004, 2005, 2006, 2007, 2008, 2009, 2010 by the deal.II authors // // This file is subject to QPL and may not be distributed // without copyright and license information. Please refer @@ -31,63 +31,139 @@ namespace PETScWrappers Assert (index < vector.size(), ExcIndexRange (index, 0, vector.size())); - // this is clumsy: there is no simple - // way in PETSc to read an element from - // a vector, i.e. there is no function - // VecGetValue or so. The only way is - // to obtain a pointer to a contiguous + // old versions of PETSc appear to be + // missing the function VecGetValues(), + // so the workaround consists of + // obtaining a pointer to a contiguous // representation of the vector and - // read from it. Subsequently, the - // vector representation has to be - // restored. In addition, we can only - // get access to the local part of the - // vector, so we have to guard against - // that + // read from it. In addition, we can + // only get access to the local part of + // the vector, so we have to guard + // against that if (dynamic_cast(&vector) != 0) { + #if (PETSC_VERSION_MAJOR <= 2) && (PETSC_VERSION_MINOR < 3) PetscScalar *ptr; int ierr - = VecGetArray (static_cast(vector), &ptr); + = VecGetArray (vector.vector, &ptr); AssertThrow (ierr == 0, ExcPETScError(ierr)); const PetscScalar value = *(ptr+index); - ierr = VecRestoreArray (static_cast(vector), &ptr); + ierr = VecRestoreArray (vector.vector, &ptr); AssertThrow (ierr == 0, ExcPETScError(ierr)); - return value; +#else + PetscInt idx = index; + PetscScalar value; + int ierr = VecGetValues(vector.vector, 1, &idx, &value); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + return value; +#endif } else if (dynamic_cast(&vector) != 0) { + int ierr; + + if (vector.ghosted) + { +#ifdef PETSC_USE_64BIT_INDICES + PetscInt +#else + int +#endif + begin, end; + ierr = VecGetOwnershipRange (vector.vector, &begin, &end); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + + Vec l; + ierr = VecGhostGetLocalForm(vector.vector, &l); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + + PetscInt lsize; + ierr = VecGetSize(l, &lsize); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + + PetscScalar *ptr; + ierr = VecGetArray(l, &ptr); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + + PetscScalar value; + + if ( index>=static_cast(begin) + && index(end) ) + { + //local entry + value=*(ptr+index-begin); + } + else + { + //ghost entry + unsigned ghostidx + = vector.ghost_indices.index_within_set(index); + + Assert(ghostidx+end-begin<(unsigned int)lsize, ExcInternalError()); + value=*(ptr+ghostidx+end-begin); + + + } + + + ierr = VecRestoreArray(l, &ptr); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + + ierr = VecGhostRestoreLocalForm(vector.vector, &l); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + + return value; + } + + // first verify that the requested // element is actually locally // available - int ierr; + #ifdef PETSC_USE_64BIT_INDICES PetscInt #else int #endif begin, end; - ierr = VecGetOwnershipRange (static_cast(vector), - &begin, &end); + ierr = VecGetOwnershipRange (vector.vector, &begin, &end); AssertThrow (ierr == 0, ExcPETScError(ierr)); + + AssertThrow ((index >= static_cast(begin)) && (index < static_cast(end)), ExcAccessToNonlocalElement (index, begin, end-1)); + // old version which only work with + // VecGetArray()... +#if (PETSC_VERSION_MAJOR <= 2) && (PETSC_VERSION_MINOR < 3) + // then access it PetscScalar *ptr; - ierr = VecGetArray (static_cast(vector), &ptr); + ierr = VecGetArray (vector.vector, &ptr); AssertThrow (ierr == 0, ExcPETScError(ierr)); const PetscScalar value = *(ptr+index-begin); - ierr = VecRestoreArray (static_cast(vector), &ptr); + ierr = VecRestoreArray (vector.vector, &ptr); AssertThrow (ierr == 0, ExcPETScError(ierr)); return value; + +#else + //new version with VecGetValues() + PetscInt idx = index; + PetscScalar value; + ierr = VecGetValues(vector.vector, 1, &idx, &value); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + + return value; +#endif + } else // what? what other kind of vector @@ -99,6 +175,7 @@ namespace PETScWrappers VectorBase::VectorBase () : + ghosted(false), last_action (LastAction::none) {} @@ -107,6 +184,8 @@ namespace PETScWrappers VectorBase::VectorBase (const VectorBase &v) : Subscriptor (), + ghosted(v.ghosted), + ghost_indices(v.ghost_indices), last_action (LastAction::none) { int ierr = VecDuplicate (v.vector, &vector); @@ -1041,6 +1120,25 @@ namespace PETScWrappers last_action = LastAction::insert; } + + void + VectorBase::update_ghost_values() const + { + // generate an error for not ghosted + // vectors + if (!ghosted) + throw ExcInternalError(); + + int ierr; + + ierr = VecGhostUpdateBegin(vector, INSERT_VALUES, SCATTER_FORWARD); + AssertThrow (ierr == 0, ExcPETScError(ierr)); + ierr = VecGhostUpdateEnd(vector, INSERT_VALUES, SCATTER_FORWARD); + AssertThrow (ierr == 0, ExcPETScError(ierr)); +} + + + } DEAL_II_NAMESPACE_CLOSE diff --git a/deal.II/lac/source/sparsity_tools.cc b/deal.II/lac/source/sparsity_tools.cc index d4e0e1aca2..7dc9f41f26 100644 --- a/deal.II/lac/source/sparsity_tools.cc +++ b/deal.II/lac/source/sparsity_tools.cc @@ -18,6 +18,13 @@ #include +#ifdef DEAL_II_COMPILER_SUPPORTS_MPI +#include +#include +#include +#include +#endif + #ifdef DEAL_II_USE_METIS // This is sorta stupid. what we really would like to do here is this: // extern "C" { @@ -406,6 +413,149 @@ namespace SparsityTools (next_free_number == sparsity.n_rows()), ExcInternalError()); } + +#ifdef DEAL_II_COMPILER_SUPPORTS_MPI + template + void distribute_sparsity_pattern(CSP_t & csp, + const std::vector & rows_per_cpu, + const MPI_Comm & mpi_comm, + const IndexSet & myrange) + { + unsigned int myid = Utilities::System::get_this_mpi_process(mpi_comm); + std::vector start_index(rows_per_cpu.size()+1); + start_index[0]=0; + for (unsigned int i=0;i > map_vec_t; + + map_vec_t send_data; + + { + unsigned int dest_cpu=0; + + unsigned int n_local_rel_rows = myrange.n_elements(); + for (unsigned int row_idx=0;row_idx=start_index[dest_cpu+1]) + ++dest_cpu; + + //skip myself + if (dest_cpu==myid) + { + row_idx+=rows_per_cpu[myid]-1; + continue; } + unsigned int rlen = csp.row_length(row); + + //skip empty lines + if (!rlen) + continue; + + //save entries + std::vector & dst = send_data[dest_cpu]; + + dst.push_back(rlen); // number of entries + dst.push_back(row); // row index + for (unsigned int c=0; c send_to; + send_to.reserve(send_data.size()); + for (map_vec_t::iterator it=send_data.begin();it!=send_data.end();++it) + send_to.push_back(it->first); + + num_receive = + Utilities::System:: + compute_point_to_point_communication_pattern(mpi_comm, send_to).size(); + } + + std::vector requests(send_data.size()); + + + // send data + { + unsigned int idx=0; + for (map_vec_t::iterator it=send_data.begin();it!=send_data.end();++it, ++idx) + MPI_Isend(&(it->second[0]), + it->second.size(), + MPI_INT, + it->first, + 124, + mpi_comm, + &requests[idx]); + } + + { + //receive + std::vector recv_buf; + for (unsigned int index=0;index (SparsityType & csp, \ +const std::vector & rows_per_cpu,\ +const MPI_Comm & mpi_comm,\ +const IndexSet & myrange) + +#ifdef DEAL_II_COMPILER_SUPPORTS_MPI +SPARSITY_FUNCTIONS(CompressedSparsityPattern); +SPARSITY_FUNCTIONS(CompressedSimpleSparsityPattern); +#endif + +#undef SPARSITY_FUNCTIONS + DEAL_II_NAMESPACE_CLOSE