From 999c62359046674117c0e0e1348072e002622c15 Mon Sep 17 00:00:00 2001 From: Guido Cossu Date: Mon, 18 Sep 2017 14:39:04 +0100 Subject: [PATCH] Solving a memory leak in Communicator_mpi --- lib/cartesian/Cartesian_base.h | 3 + lib/cartesian/Cartesian_full.h | 3 + lib/communicator/Communicator_base.h | 1 + lib/communicator/Communicator_mpi.cc | 7 + .../action/scalar/ScalarInteractionAction.h | 220 ++++++++++-------- 5 files changed, 135 insertions(+), 99 deletions(-) diff --git a/lib/cartesian/Cartesian_base.h b/lib/cartesian/Cartesian_base.h index f4f9a269..0c67e951 100644 --- a/lib/cartesian/Cartesian_base.h +++ b/lib/cartesian/Cartesian_base.h @@ -50,6 +50,9 @@ public: GridBase(const std::vector & processor_grid) : CartesianCommunicator(processor_grid) {}; + virtual ~GridBase() = default; + + // Physics Grid information. std::vector _simd_layout;// Which dimensions get relayed out over simd lanes. std::vector _fdimensions;// (full) Global dimensions of array prior to cb removal diff --git a/lib/cartesian/Cartesian_full.h b/lib/cartesian/Cartesian_full.h index 815e3b22..62481bb8 100644 --- a/lib/cartesian/Cartesian_full.h +++ b/lib/cartesian/Cartesian_full.h @@ -93,6 +93,7 @@ public: // Use a reduced simd grid _ldimensions[d] = _gdimensions[d] / _processors[d]; //local dimensions + //std::cout << _ldimensions[d] << " " << _gdimensions[d] << " " << _processors[d] << std::endl; assert(_ldimensions[d] * _processors[d] == _gdimensions[d]); _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; //overdecomposition @@ -137,6 +138,8 @@ public: block = block * _rdimensions[d]; } }; + + virtual ~GridCartesian() = default; }; } #endif diff --git a/lib/communicator/Communicator_base.h b/lib/communicator/Communicator_base.h index ac866ced..ada017b0 100644 --- a/lib/communicator/Communicator_base.h +++ b/lib/communicator/Communicator_base.h @@ -152,6 +152,7 @@ class CartesianCommunicator { // Constructor of any given grid //////////////////////////////////////////////// CartesianCommunicator(const std::vector &pdimensions_in); + virtual ~CartesianCommunicator(); //////////////////////////////////////////////////////////////////////////////////////// // Wraps MPI_Cart routines, or implements equivalent on other impls diff --git a/lib/communicator/Communicator_mpi.cc b/lib/communicator/Communicator_mpi.cc index bd2a62fb..a3427b00 100644 --- a/lib/communicator/Communicator_mpi.cc +++ b/lib/communicator/Communicator_mpi.cc @@ -75,6 +75,13 @@ CartesianCommunicator::CartesianCommunicator(const std::vector &processors) assert(Size==_Nprocessors); } + +CartesianCommunicator::~CartesianCommunicator(){ + if (communicator && !MPI::Is_finalized()) + MPI_Comm_free(&communicator); +} + + void CartesianCommunicator::GlobalSum(uint32_t &u){ int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator); assert(ierr==0); diff --git a/lib/qcd/action/scalar/ScalarInteractionAction.h b/lib/qcd/action/scalar/ScalarInteractionAction.h index 4d189352..a681b62c 100644 --- a/lib/qcd/action/scalar/ScalarInteractionAction.h +++ b/lib/qcd/action/scalar/ScalarInteractionAction.h @@ -30,119 +30,141 @@ directory #ifndef SCALAR_INT_ACTION_H #define SCALAR_INT_ACTION_H - // Note: this action can completely absorb the ScalarAction for real float fields // use the scalarObjs to generalise the structure -namespace Grid { - // FIXME drop the QCD namespace everywhere here +namespace Grid +{ +// FIXME drop the QCD namespace everywhere here - template - class ScalarInteractionAction : public QCD::Action { - public: - INHERIT_FIELD_TYPES(Impl); - private: - RealD mass_square; - RealD lambda; +template +class ScalarInteractionAction : public QCD::Action +{ +public: + INHERIT_FIELD_TYPES(Impl); +private: + RealD mass_square; + RealD lambda; - typedef typename Field::vector_object vobj; - typedef CartesianStencil Stencil; + typedef typename Field::vector_object vobj; + typedef CartesianStencil Stencil; - SimpleCompressor compressor; - int npoint = 2*Ndim; - std::vector directions;// = {0,1,2,3,0,1,2,3}; // forcing 4 dimensions - std::vector displacements;// = {1,1,1,1, -1,-1,-1,-1}; + SimpleCompressor compressor; + int npoint = 2 * Ndim; + std::vector directions; // = {0,1,2,3,0,1,2,3}; // forcing 4 dimensions + std::vector displacements; // = {1,1,1,1, -1,-1,-1,-1}; - - public: - - ScalarInteractionAction(RealD ms, RealD l) : mass_square(ms), lambda(l), displacements(2*Ndim,0), directions(2*Ndim,0){ - for (int mu = 0 ; mu < Ndim; mu++){ - directions[mu] = mu; directions[mu+Ndim] = mu; - displacements[mu] = 1; displacements[mu+Ndim] = -1; - } +public: + ScalarInteractionAction(RealD ms, RealD l) : mass_square(ms), lambda(l), displacements(2 * Ndim, 0), directions(2 * Ndim, 0) + { + for (int mu = 0; mu < Ndim; mu++) + { + directions[mu] = mu; + directions[mu + Ndim] = mu; + displacements[mu] = 1; + displacements[mu + Ndim] = -1; } + } - virtual std::string LogParameters() { - std::stringstream sstream; - sstream << GridLogMessage << "[ScalarAction] lambda : " << lambda << std::endl; - sstream << GridLogMessage << "[ScalarAction] mass_square : " << mass_square << std::endl; - return sstream.str(); - } + virtual std::string LogParameters() + { + std::stringstream sstream; + sstream << GridLogMessage << "[ScalarAction] lambda : " << lambda << std::endl; + sstream << GridLogMessage << "[ScalarAction] mass_square : " << mass_square << std::endl; + return sstream.str(); + } - virtual std::string action_name() {return "ScalarAction";} + virtual std::string action_name() { return "ScalarAction"; } - virtual void refresh(const Field &U, GridParallelRNG &pRNG) {} + virtual void refresh(const Field &U, GridParallelRNG &pRNG) {} - virtual RealD S(const Field &p) { - assert(p._grid->Nd() == Ndim); - static Stencil phiStencil(p._grid, npoint, 0, directions, displacements); - phiStencil.HaloExchange(p, compressor); - Field action(p._grid), pshift(p._grid), phisquared(p._grid); - phisquared = p*p; - action = (2.0*Ndim + mass_square)*phisquared - lambda/24.*phisquared*phisquared; - for (int mu = 0; mu < Ndim; mu++) { - // pshift = Cshift(p, mu, +1); // not efficient, implement with stencils - parallel_for (int i = 0; i < p._grid->oSites(); i++) { - int permute_type; - StencilEntry *SE; - vobj temp2; - const vobj *temp, *t_p; - - SE = phiStencil.GetEntry(permute_type, mu, i); - t_p = &p._odata[i]; - if ( SE->_is_local ) { - temp = &p._odata[SE->_offset]; - if ( SE->_permute ) { - permute(temp2, *temp, permute_type); - action._odata[i] -= temp2*(*t_p) + (*t_p)*temp2; - } else { - action._odata[i] -= (*temp)*(*t_p) + (*t_p)*(*temp); - } - } else { - action._odata[i] -= phiStencil.CommBuf()[SE->_offset]*(*t_p) + (*t_p)*phiStencil.CommBuf()[SE->_offset]; - } - } - // action -= pshift*p + p*pshift; - } - // NB the trace in the algebra is normalised to 1/2 - // minus sign coming from the antihermitian fields - return -(TensorRemove(sum(trace(action)))).real(); - }; - - virtual void deriv(const Field &p, Field &force) { - assert(p._grid->Nd() == Ndim); - force = (2.0*Ndim + mass_square)*p - lambda/12.*p*p*p; - // move this outside - static Stencil phiStencil(p._grid, npoint, 0, directions, displacements); - phiStencil.HaloExchange(p, compressor); - - //for (int mu = 0; mu < QCD::Nd; mu++) force -= Cshift(p, mu, -1) + Cshift(p, mu, 1); - for (int point = 0; point < npoint; point++) { - parallel_for (int i = 0; i < p._grid->oSites(); i++) { - const vobj *temp; - vobj temp2; - int permute_type; - StencilEntry *SE; - SE = phiStencil.GetEntry(permute_type, point, i); - - if ( SE->_is_local ) { - temp = &p._odata[SE->_offset]; - if ( SE->_permute ) { - permute(temp2, *temp, permute_type); - force._odata[i] -= temp2; - } else { - force._odata[i] -= *temp; - } - } else { - force._odata[i] -= phiStencil.CommBuf()[SE->_offset]; - } - } + virtual RealD S(const Field &p) + { + assert(p._grid->Nd() == Ndim); + static Stencil phiStencil(p._grid, npoint, 0, directions, displacements); + phiStencil.HaloExchange(p, compressor); + Field action(p._grid), pshift(p._grid), phisquared(p._grid); + phisquared = p * p; + action = (2.0 * Ndim + mass_square) * phisquared - lambda / 24. * phisquared * phisquared; + for (int mu = 0; mu < Ndim; mu++) + { + // pshift = Cshift(p, mu, +1); // not efficient, implement with stencils + parallel_for(int i = 0; i < p._grid->oSites(); i++) + { + int permute_type; + StencilEntry *SE; + vobj temp2; + const vobj *temp, *t_p; + + SE = phiStencil.GetEntry(permute_type, mu, i); + t_p = &p._odata[i]; + if (SE->_is_local) + { + temp = &p._odata[SE->_offset]; + if (SE->_permute) + { + permute(temp2, *temp, permute_type); + action._odata[i] -= temp2 * (*t_p) + (*t_p) * temp2; + } + else + { + action._odata[i] -= (*temp) * (*t_p) + (*t_p) * (*temp); + } + } + else + { + action._odata[i] -= phiStencil.CommBuf()[SE->_offset] * (*t_p) + (*t_p) * phiStencil.CommBuf()[SE->_offset]; + } } + // action -= pshift*p + p*pshift; } + // NB the trace in the algebra is normalised to 1/2 + // minus sign coming from the antihermitian fields + return -(TensorRemove(sum(trace(action)))).real(); }; - -} // namespace Grid -#endif // SCALAR_INT_ACTION_H + virtual void deriv(const Field &p, Field &force) + { + assert(p._grid->Nd() == Ndim); + force = (2.0 * Ndim + mass_square) * p - lambda / 12. * p * p * p; + // move this outside + static Stencil phiStencil(p._grid, npoint, 0, directions, displacements); + phiStencil.HaloExchange(p, compressor); + + //for (int mu = 0; mu < QCD::Nd; mu++) force -= Cshift(p, mu, -1) + Cshift(p, mu, 1); + for (int point = 0; point < npoint; point++) + { + parallel_for(int i = 0; i < p._grid->oSites(); i++) + { + const vobj *temp; + vobj temp2; + int permute_type; + StencilEntry *SE; + SE = phiStencil.GetEntry(permute_type, point, i); + + if (SE->_is_local) + { + temp = &p._odata[SE->_offset]; + if (SE->_permute) + { + permute(temp2, *temp, permute_type); + force._odata[i] -= temp2; + } + else + { + force._odata[i] -= *temp; + } + } + else + { + force._odata[i] -= phiStencil.CommBuf()[SE->_offset]; + } + } + } + } +}; + +} // namespace Grid + +#endif // SCALAR_INT_ACTION_H