Homec4science
Diffusion akantu (master-stable) Browse

rAKA/src/synchronizer/communicator_mpi_inline_impl.cc
6a524c970887master-stable

Last Change
Raw File

communicator_mpi_inline_impl.cc
View Options

/**
* @file communicator_mpi_inline_impl.cc
*
* @author Nicolas Richart <nicolas.richart@epfl.ch>
*
* @date creation: Tue Nov 07 2017
* @date last modification: Mon Dec 18 2017
*
* @brief StaticCommunicatorMPI implementation
*
* @section LICENSE
*
* Copyright (©) 2016-2018 EPFL (Ecole Polytechnique Fédérale de Lausanne)
* Laboratory (LSMS - Laboratoire de Simulation en Mécanique des Solides)
*
* Akantu is free software: you can redistribute it and/or modify it under the
* terms of the GNU Lesser General Public License as published by the Free
* Software Foundation, either version 3 of the License, or (at your option) any
* later version.
*
* Akantu is distributed in the hope that it will be useful, but WITHOUT ANY
* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
* A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
* details.
*
* You should have received a copy of the GNU Lesser General Public License
* along with Akantu. If not, see <http://www.gnu.org/licenses/>.
*
*/
/* -------------------------------------------------------------------------- */
#include "aka_iterators.hh"
#include "communicator.hh"
#include "mpi_communicator_data.hh"
/* -------------------------------------------------------------------------- */
#include <memory>
#include <type_traits>
#include <unordered_map>
#include <vector>
/* -------------------------------------------------------------------------- */
#include <mpi.h>
/* -------------------------------------------------------------------------- */
#if (defined(__GNUC__) || defined(__GNUG__))
#if AKA_GCC_VERSION < 60000
namespace std {
template <> struct hash<akantu::SynchronizerOperation> {
using argument_type = akantu::SynchronizerOperation;
size_t operator()(const argument_type & e) const noexcept {
auto ue = underlying_type_t<argument_type>(e);
return uh(ue);
}
private:
const hash<underlying_type_t<argument_type>> uh{};
};
} // namespace std
#endif
#endif
namespace akantu {
class CommunicationRequestMPI : public InternalCommunicationRequest {
public:
CommunicationRequestMPI(UInt source, UInt dest)
: InternalCommunicationRequest(source, dest),
request(std::make_unique<MPI_Request>()) {}
MPI_Request & getMPIRequest() { return *request; };
private:
std::unique_ptr<MPI_Request> request;
};
namespace {
template <typename T> inline MPI_Datatype getMPIDatatype();
MPI_Op getMPISynchronizerOperation(const SynchronizerOperation & op) {
std::unordered_map<SynchronizerOperation, MPI_Op> _operations{
{SynchronizerOperation::_sum, MPI_SUM},
{SynchronizerOperation::_min, MPI_MIN},
{SynchronizerOperation::_max, MPI_MAX},
{SynchronizerOperation::_prod, MPI_PROD},
{SynchronizerOperation::_land, MPI_LAND},
{SynchronizerOperation::_band, MPI_BAND},
{SynchronizerOperation::_lor, MPI_LOR},
{SynchronizerOperation::_bor, MPI_BOR},
{SynchronizerOperation::_lxor, MPI_LXOR},
{SynchronizerOperation::_bxor, MPI_BXOR},
{SynchronizerOperation::_min_loc, MPI_MINLOC},
{SynchronizerOperation::_max_loc, MPI_MAXLOC},
{SynchronizerOperation::_null, MPI_OP_NULL}};
return _operations[op];
}
template <typename T> MPI_Datatype inline getMPIDatatype() {
return MPI_DATATYPE_NULL;
}
#define SPECIALIZE_MPI_DATATYPE(type, mpi_type) \
template <> MPI_Datatype inline getMPIDatatype<type>() { return mpi_type; }
#define COMMA ,
SPECIALIZE_MPI_DATATYPE(char, MPI_CHAR)
SPECIALIZE_MPI_DATATYPE(float, MPI_FLOAT)
SPECIALIZE_MPI_DATATYPE(double, MPI_DOUBLE)
SPECIALIZE_MPI_DATATYPE(long double, MPI_LONG_DOUBLE)
SPECIALIZE_MPI_DATATYPE(signed int, MPI_INT)
SPECIALIZE_MPI_DATATYPE(NodeType, getMPIDatatype<Int>())
SPECIALIZE_MPI_DATATYPE(unsigned int, MPI_UNSIGNED)
SPECIALIZE_MPI_DATATYPE(signed long int, MPI_LONG)
SPECIALIZE_MPI_DATATYPE(unsigned long int, MPI_UNSIGNED_LONG)
SPECIALIZE_MPI_DATATYPE(signed long long int, MPI_LONG_LONG)
SPECIALIZE_MPI_DATATYPE(unsigned long long int, MPI_UNSIGNED_LONG_LONG)
SPECIALIZE_MPI_DATATYPE(SCMinMaxLoc<double COMMA int>, MPI_DOUBLE_INT)
SPECIALIZE_MPI_DATATYPE(SCMinMaxLoc<float COMMA int>, MPI_FLOAT_INT)
SPECIALIZE_MPI_DATATYPE(bool, MPI_CXX_BOOL)
inline int getMPISource(int src) {
if (src == _any_source)
return MPI_ANY_SOURCE;
return src;
}
decltype(auto) convertRequests(std::vector<CommunicationRequest> & requests) {
std::vector<MPI_Request> mpi_requests(requests.size());
for (auto && request_pair : zip(requests, mpi_requests)) {
auto && req = std::get<0>(request_pair);
auto && mpi_req = std::get<1>(request_pair);
mpi_req = dynamic_cast<CommunicationRequestMPI &>(req.getInternal())
.getMPIRequest();
}
return mpi_requests;
}
} // namespace
// this is ugly but shorten the code a lot
#define MPIDATA \
(*reinterpret_cast<MPICommunicatorData *>(communicator_data.get()))
/* -------------------------------------------------------------------------- */
/* Implementation */
/* -------------------------------------------------------------------------- */
struct Communicator::private_member {
private_member(MPI_Comm comm) : comm(std::move(comm)) {}
MPI_Comm comm;
};
/* -------------------------------------------------------------------------- */
Communicator::Communicator(int & /*argc*/, char **& /*argv*/)
: communicator_data(std::make_unique<MPICommunicatorData>()) {
prank = MPIDATA.rank();
psize = MPIDATA.size();
}
/* -------------------------------------------------------------------------- */
Communicator::Communicator(const private_member & p)
: communicator_data(std::make_unique<MPICommunicatorData>(p.comm)) {
prank = MPIDATA.rank();
psize = MPIDATA.size();
}
/* -------------------------------------------------------------------------- */
template <typename T>
void Communicator::sendImpl(const T * buffer, Int size, Int receiver, Int tag,
const CommunicationMode & mode) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
MPI_Datatype type = getMPIDatatype<T>();
switch (mode) {
case CommunicationMode::_auto:
MPI_Send(buffer, size, type, receiver, tag, communicator);
break;
case CommunicationMode::_synchronous:
MPI_Ssend(buffer, size, type, receiver, tag, communicator);
break;
case CommunicationMode::_ready:
MPI_Rsend(buffer, size, type, receiver, tag, communicator);
break;
}
}
/* -------------------------------------------------------------------------- */
template <typename T>
void Communicator::receiveImpl(T * buffer, Int size, Int sender,
Int tag) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
MPI_Status status;
MPI_Datatype type = getMPIDatatype<T>();
MPI_Recv(buffer, size, type, getMPISource(sender), tag, communicator,
&status);
}
/* -------------------------------------------------------------------------- */
template <typename T>
CommunicationRequest
Communicator::asyncSendImpl(const T * buffer, Int size, Int receiver, Int tag,
const CommunicationMode & mode) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
auto * request = new CommunicationRequestMPI(prank, receiver);
MPI_Request & req = request->getMPIRequest();
MPI_Datatype type = getMPIDatatype<T>();
switch (mode) {
case CommunicationMode::_auto:
MPI_Isend(buffer, size, type, receiver, tag, communicator, &req);
break;
case CommunicationMode::_synchronous:
MPI_Issend(buffer, size, type, receiver, tag, communicator, &req);
break;
case CommunicationMode::_ready:
MPI_Irsend(buffer, size, type, receiver, tag, communicator, &req);
break;
}
return std::shared_ptr<InternalCommunicationRequest>(request);
}
/* -------------------------------------------------------------------------- */
template <typename T>
CommunicationRequest Communicator::asyncReceiveImpl(T * buffer, Int size,
Int sender, Int tag) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
auto * request = new CommunicationRequestMPI(sender, prank);
MPI_Datatype type = getMPIDatatype<T>();
MPI_Request & req = request->getMPIRequest();
MPI_Irecv(buffer, size, type, getMPISource(sender), tag, communicator, &req);
return std::shared_ptr<InternalCommunicationRequest>(request);
}
/* -------------------------------------------------------------------------- */
template <typename T>
void Communicator::probe(Int sender, Int tag,
CommunicationStatus & status) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
MPI_Status mpi_status;
MPI_Probe(getMPISource(sender), tag, communicator, &mpi_status);
MPI_Datatype type = getMPIDatatype<T>();
int count;
MPI_Get_count(&mpi_status, type, &count);
status.setSource(mpi_status.MPI_SOURCE);
status.setTag(mpi_status.MPI_TAG);
status.setSize(count);
}
/* -------------------------------------------------------------------------- */
template <typename T>
bool Communicator::asyncProbe(Int sender, Int tag,
CommunicationStatus & status) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
MPI_Status mpi_status;
int test;
MPI_Iprobe(getMPISource(sender), tag, communicator, &test, &mpi_status);
if (not test)
return false;
MPI_Datatype type = getMPIDatatype<T>();
int count;
MPI_Get_count(&mpi_status, type, &count);
status.setSource(mpi_status.MPI_SOURCE);
status.setTag(mpi_status.MPI_TAG);
status.setSize(count);
return true;
}
/* -------------------------------------------------------------------------- */
bool Communicator::test(CommunicationRequest & request) const {
MPI_Status status;
int flag;
auto & req_mpi =
dynamic_cast<CommunicationRequestMPI &>(request.getInternal());
MPI_Request & req = req_mpi.getMPIRequest();
#if !defined(AKANTU_NDEBUG)
int ret =
#endif
MPI_Test(&req, &flag, &status);
AKANTU_DEBUG_ASSERT(ret == MPI_SUCCESS, "Error in MPI_Test.");
return (flag != 0);
}
/* -------------------------------------------------------------------------- */
bool Communicator::testAll(std::vector<CommunicationRequest> & requests) const {
int are_finished;
auto && mpi_requests = convertRequests(requests);
MPI_Testall(mpi_requests.size(), mpi_requests.data(), &are_finished,
MPI_STATUSES_IGNORE);
return are_finished != 0 ? true : false;
}
/* -------------------------------------------------------------------------- */
void Communicator::wait(CommunicationRequest & request) const {
MPI_Status status;
auto & req_mpi =
dynamic_cast<CommunicationRequestMPI &>(request.getInternal());
MPI_Request & req = req_mpi.getMPIRequest();
MPI_Wait(&req, &status);
}
/* -------------------------------------------------------------------------- */
void Communicator::waitAll(std::vector<CommunicationRequest> & requests) const {
auto && mpi_requests = convertRequests(requests);
MPI_Waitall(mpi_requests.size(), mpi_requests.data(), MPI_STATUSES_IGNORE);
}
/* -------------------------------------------------------------------------- */
UInt Communicator::waitAny(std::vector<CommunicationRequest> & requests) const {
auto && mpi_requests = convertRequests(requests);
int pos;
MPI_Waitany(mpi_requests.size(), mpi_requests.data(), &pos,
MPI_STATUSES_IGNORE);
if (pos != MPI_UNDEFINED) {
return pos;
} else {
return UInt(-1);
}
}
/* -------------------------------------------------------------------------- */
void Communicator::barrier() const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
MPI_Barrier(communicator);
}
/* -------------------------------------------------------------------------- */
CommunicationRequest Communicator::asyncBarrier() const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
auto * request = new CommunicationRequestMPI(0, 0);
MPI_Request & req = request->getMPIRequest();
MPI_Ibarrier(communicator, &req);
return std::shared_ptr<InternalCommunicationRequest>(request);
}
/* -------------------------------------------------------------------------- */
template <typename T>
void Communicator::reduceImpl(T * values, int nb_values,
const SynchronizerOperation & op,
int root) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
MPI_Datatype type = getMPIDatatype<T>();
MPI_Reduce(MPI_IN_PLACE, values, nb_values, type,
getMPISynchronizerOperation(op), root, communicator);
}
/* -------------------------------------------------------------------------- */
template <typename T>
void Communicator::allReduceImpl(T * values, int nb_values,
const SynchronizerOperation & op) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
MPI_Datatype type = getMPIDatatype<T>();
MPI_Allreduce(MPI_IN_PLACE, values, nb_values, type,
getMPISynchronizerOperation(op), communicator);
}
/* -------------------------------------------------------------------------- */
template <typename T>
void Communicator::allGatherImpl(T * values, int nb_values) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
MPI_Datatype type = getMPIDatatype<T>();
MPI_Allgather(MPI_IN_PLACE, nb_values, type, values, nb_values, type,
communicator);
}
/* -------------------------------------------------------------------------- */
template <typename T>
void Communicator::allGatherVImpl(T * values, int * nb_values) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
std::vector<int> displs(psize);
displs[0] = 0;
for (int i = 1; i < psize; ++i) {
displs[i] = displs[i - 1] + nb_values[i - 1];
}
MPI_Datatype type = getMPIDatatype<T>();
MPI_Allgatherv(MPI_IN_PLACE, *nb_values, type, values, nb_values,
displs.data(), type, communicator);
}
/* -------------------------------------------------------------------------- */
template <typename T>
void Communicator::gatherImpl(T * values, int nb_values, int root) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
T *send_buf = nullptr, *recv_buf = nullptr;
if (prank == root) {
send_buf = (T *)MPI_IN_PLACE;
recv_buf = values;
} else {
send_buf = values;
}
MPI_Datatype type = getMPIDatatype<T>();
MPI_Gather(send_buf, nb_values, type, recv_buf, nb_values, type, root,
communicator);
}
/* -------------------------------------------------------------------------- */
template <typename T>
void Communicator::gatherImpl(T * values, int nb_values, T * gathered,
int nb_gathered) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
T * send_buf = values;
T * recv_buf = gathered;
if (nb_gathered == 0)
nb_gathered = nb_values;
MPI_Datatype type = getMPIDatatype<T>();
MPI_Gather(send_buf, nb_values, type, recv_buf, nb_gathered, type,
this->prank, communicator);
}
/* -------------------------------------------------------------------------- */
template <typename T>
void Communicator::gatherVImpl(T * values, int * nb_values, int root) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
int * displs = nullptr;
if (prank == root) {
displs = new int[psize];
displs[0] = 0;
for (int i = 1; i < psize; ++i) {
displs[i] = displs[i - 1] + nb_values[i - 1];
}
}
T *send_buf = nullptr, *recv_buf = nullptr;
if (prank == root) {
send_buf = (T *)MPI_IN_PLACE;
recv_buf = values;
} else
send_buf = values;
MPI_Datatype type = getMPIDatatype<T>();
MPI_Gatherv(send_buf, *nb_values, type, recv_buf, nb_values, displs, type,
root, communicator);
if (prank == root) {
delete[] displs;
}
}
/* -------------------------------------------------------------------------- */
template <typename T>
void Communicator::broadcastImpl(T * values, int nb_values, int root) const {
MPI_Comm communicator = MPIDATA.getMPICommunicator();
MPI_Datatype type = getMPIDatatype<T>();
MPI_Bcast(values, nb_values, type, root, communicator);
}
/* -------------------------------------------------------------------------- */
int Communicator::getMaxTag() const { return MPIDATA.getMaxTag(); }
int Communicator::getMinTag() const { return 0; }
/* -------------------------------------------------------------------------- */
} // namespace akantu
c4science · Help