Homec4science
Diffusion akantu (bugfixes/bug_fix_in_synchronization_of_mat_vec_mul_product_when_assembled_to_residual) Browse

rAKA/src/synchronizer/dof_synchronizer.cc
4bee089d6542bugfixes/bug_fix_in_syn…

Last Change
Raw File

dof_synchronizer.cc
View Options

/**
* @file dof_synchronizer.cc
*
* @author Nicolas Richart <nicolas.richart@epfl.ch>
* @author Aurelia Cuba Ramos <aurelia.cubaramos@epfl.ch>
*
* @date creation: Fri Jun 17 2011
* @date last modification: Thu Mar 27 2014
*
* @brief DOF synchronizing object implementation
*
* @section LICENSE
*
* Copyright (©) 2010-2012, 2014 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 "dof_synchronizer.hh"
#include "mesh.hh"
/* -------------------------------------------------------------------------- */
__BEGIN_AKANTU__
/* -------------------------------------------------------------------------- */
/**
* A DOFSynchronizer needs a mesh and the number of degrees of freedom
* per node to be created. In the constructor computes the local and global dof
* number for each dof. The member
* proc_informations (std vector) is resized with the number of mpi
* processes. Each entry in the vector is a PerProcInformations object
* that contains the interactions of the current mpi process (prank) with the
* mpi process corresponding to the position of that entry. Every
* ProcInformations object contains one array with the dofs that have
* to be sent to prank and a second one with dofs that willl be received form prank.
* This information is needed for the asychronous communications. The
* constructor sets up this information.
* @param mesh mesh discretizing the domain we want to analyze
* @param nb_degree_of_freedom number of degrees of freedom per node
*/
DOFSynchronizer::DOFSynchronizer(const Mesh & mesh, UInt nb_degree_of_freedom) :
global_dof_equation_numbers(0, 1, "global_equation_number"),
local_dof_equation_numbers(0, 1, "local_equation_number"),
dof_global_ids(0, 1, "global_ids"),
dof_types(0, 1, "types") {
gather_scatter_scheme_initialized = false;
prank = static_communicator->whoAmI();
psize = static_communicator->getNbProc();
proc_informations.resize(psize);
UInt nb_nodes = mesh.getNbNodes();
nb_dofs = nb_nodes * nb_degree_of_freedom;
dof_global_ids.resize(nb_dofs);
dof_types.resize(nb_dofs);
nb_global_dofs = mesh.getNbGlobalNodes() * nb_degree_of_freedom;
/// compute the global id for each dof and store the dof type (pure ghost, slave, master or local)
UInt * dof_global_id = dof_global_ids.storage();
Int * dof_type = dof_types.storage();
for(UInt n = 0, ld = 0; n < nb_nodes; ++n) {
UInt node_global_id = mesh.getNodeGlobalId(n);
UInt node_type = mesh.getNodeType(n);
for (UInt d = 0; d < nb_degree_of_freedom; ++d, ++ld) {
*dof_global_id = node_global_id * nb_degree_of_freedom + d;
global_dof_to_local[*dof_global_id] = ld;
*(dof_type++) = node_type;
dof_global_id++;
}
}
if(psize == 1) return;
/// creating communication scheme
dof_global_id = dof_global_ids.storage();
dof_type = dof_types.storage();
for (UInt n = 0; n < nb_dofs; ++n) {
/// check if dof is slave. In that case the value stored in
/// dof_type corresponds to the process that has the corresponding
/// master dof
if(*dof_type >= 0) {
/// has to receive n from proc[*dof_type]
proc_informations[*dof_type].master_dofs.push_back(n);
}
dof_type++;
}
/// at this point the master nodes in PerProcInfo are known
/// exchanging information
for (UInt p = 1; p < psize; ++p) {
UInt sendto = (psize + prank + p) % psize;
UInt recvfrom = (psize + prank - p) % psize;
/// send the master nodes
UInt nb_master_dofs = proc_informations[sendto].master_dofs.getSize();
UInt * master_dofs = proc_informations[sendto].master_dofs.storage();
UInt * send_buffer = new UInt[nb_master_dofs];
for (UInt d = 0; d < nb_master_dofs; ++d) {
send_buffer[d] = dof_global_id[master_dofs[d]];
}
UInt nb_slave_dofs = 0;
UInt * recv_buffer;
std::vector<CommunicationRequest *> requests;
requests.push_back(static_communicator->asyncSend(&nb_master_dofs, 1, sendto, 0));
if(nb_master_dofs != 0) {
AKANTU_DEBUG(dblInfo, "Sending "<< nb_master_dofs << " dofs to " << sendto + 1);
requests.push_back(static_communicator->asyncSend(send_buffer, nb_master_dofs, sendto, 1));
}
/// Receive the info and store them as slave nodes
static_communicator->receive(&nb_slave_dofs, 1, recvfrom, 0);
if(nb_slave_dofs != 0) {
AKANTU_DEBUG(dblInfo, "Receiving "<< nb_slave_dofs << " dofs from " << recvfrom + 1);
proc_informations[recvfrom].slave_dofs.resize(nb_slave_dofs);
recv_buffer = proc_informations[recvfrom].slave_dofs.storage();
static_communicator->receive(recv_buffer, nb_slave_dofs, recvfrom, 1);
}
for (UInt d = 0; d < nb_slave_dofs; ++d) {
recv_buffer[d] = global_dof_to_local[recv_buffer[d]];
}
static_communicator->waitAll(requests);
static_communicator->freeCommunicationRequest(requests);
requests.clear();
delete [] send_buffer;
}
}
/* -------------------------------------------------------------------------- */
DOFSynchronizer::~DOFSynchronizer() {
}
/* -------------------------------------------------------------------------- */
void DOFSynchronizer::initLocalDOFEquationNumbers() {
AKANTU_DEBUG_IN();
local_dof_equation_numbers.resize(nb_dofs);
Int * dof_equation_number = local_dof_equation_numbers.storage();
for (UInt d = 0; d < nb_dofs; ++d) {
*(dof_equation_number++) = d;
}
//local_dof_equation_numbers.resize(nb_dofs);
AKANTU_DEBUG_OUT();
}
/* -------------------------------------------------------------------------- */
void DOFSynchronizer::asynchronousSynchronize(DataAccessor & data_accessor,
SynchronizationTag tag) {
AKANTU_DEBUG_IN();
if (communications.find(tag) == communications.end())
computeBufferSize(data_accessor, tag);
Communication & communication = communications[tag];
AKANTU_DEBUG_ASSERT(communication.send_requests.size() == 0,
"There must be some pending sending communications. Tag is " << tag);
std::map<SynchronizationTag, UInt>::iterator t_it = tag_counter.find(tag);
UInt counter = 0;
if(t_it == tag_counter.end()) {
tag_counter[tag] = 0;
} else {
counter = ++(t_it->second);
}
for (UInt p = 0; p < psize; ++p) {
UInt ssize = communication.size_to_send[p];
if(p == prank || ssize == 0) continue;
CommunicationBuffer & buffer = communication.send_buffer[p];
buffer.resize(ssize);
#ifndef AKANTU_NDEBUG
UInt nb_dofs = proc_informations[p].slave_dofs.getSize();
AKANTU_DEBUG_INFO("Packing data for proc " << p
<< " (" << ssize << "/" << nb_dofs
<<" data to send/dofs)");
/// pack global equation numbers in debug mode
Array<UInt>::const_iterator<UInt> bit = proc_informations[p].slave_dofs.begin();
Array<UInt>::const_iterator<UInt> bend = proc_informations[p].slave_dofs.end();
for (; bit != bend; ++bit) {
buffer << global_dof_equation_numbers[*bit];
}
#endif
/// dof synchronizer needs to send the data corresponding to the
data_accessor.packDOFData(buffer, proc_informations[p].slave_dofs, tag);
AKANTU_DEBUG_ASSERT(buffer.getPackedSize() == ssize,
"a problem has been introduced with "
<< "false sent sizes declaration "
<< buffer.getPackedSize() << " != " << ssize);
AKANTU_DEBUG_INFO("Posting send to proc " << p
<< " (tag: " << tag << " - " << ssize << " data to send)"
<< " [" << Tag::genTag(prank, counter, tag) << "]");
communication.send_requests.push_back(static_communicator->asyncSend(buffer.storage(),
ssize,
p,
Tag::genTag(prank, counter, tag)));
}
AKANTU_DEBUG_ASSERT(communication.recv_requests.size() == 0,
"There must be some pending receive communications");
for (UInt p = 0; p < psize; ++p) {
UInt rsize = communication.size_to_receive[p];
if(p == prank || rsize == 0) continue;
CommunicationBuffer & buffer = communication.recv_buffer[p];
buffer.resize(rsize);
AKANTU_DEBUG_INFO("Posting receive from proc " << p
<< " (tag: " << tag << " - " << rsize << " data to receive) "
<< " [" << Tag::genTag(p, counter, tag) << "]");
communication.recv_requests.push_back(static_communicator->asyncReceive(buffer.storage(),
rsize,
p,
Tag::genTag(p, counter, tag)));
}
AKANTU_DEBUG_OUT();
}
/* -------------------------------------------------------------------------- */
void DOFSynchronizer::waitEndSynchronize(DataAccessor & data_accessor,
SynchronizationTag tag) {
AKANTU_DEBUG_IN();
AKANTU_DEBUG_ASSERT(communications.find(tag) != communications.end(), "No communication with the tag \""
<< tag <<"\" started");
Communication & communication = communications[tag];
std::vector<CommunicationRequest *> req_not_finished;
std::vector<CommunicationRequest *> * req_not_finished_tmp = &req_not_finished;
std::vector<CommunicationRequest *> * recv_requests_tmp = &(communication.recv_requests);
// static_communicator->waitAll(recv_requests);
while(!recv_requests_tmp->empty()) {
for (std::vector<CommunicationRequest *>::iterator req_it = recv_requests_tmp->begin();
req_it != recv_requests_tmp->end() ; ++req_it) {
CommunicationRequest * req = *req_it;
if(static_communicator->testRequest(req)) {
UInt proc = req->getSource();
AKANTU_DEBUG_INFO("Unpacking data coming from proc " << proc);
CommunicationBuffer & buffer = communication.recv_buffer[proc];
#ifndef AKANTU_NDEBUG
Array<UInt>::const_iterator<UInt> bit = proc_informations[proc].master_dofs.begin();
Array<UInt>::const_iterator<UInt> bend = proc_informations[proc].master_dofs.end();
for (; bit != bend; ++bit) {
Int global_dof_eq_nb_loc = global_dof_equation_numbers[*bit];
Int global_dof_eq_nb = 0;
buffer >> global_dof_eq_nb;
Real tolerance = Math::getTolerance();
if(std::abs(global_dof_eq_nb - global_dof_eq_nb_loc) <= tolerance) continue;
AKANTU_DEBUG_ERROR("Unpacking an unknown global dof equation number for dof: "
<< *bit
<< "(global dof equation number = " << global_dof_eq_nb
<< " and buffer = " << global_dof_eq_nb << ") ["
<< std::abs(global_dof_eq_nb - global_dof_eq_nb_loc)
<< "] - tag: " << tag);
}
#endif
data_accessor.unpackDOFData(buffer, proc_informations[proc].master_dofs, tag);
buffer.resize(0);
AKANTU_DEBUG_ASSERT(buffer.getLeftToUnpack() == 0,
"all data have not been unpacked: "
<< buffer.getLeftToUnpack() << " bytes left");
static_communicator->freeCommunicationRequest(req);
} else {
req_not_finished_tmp->push_back(req);
}
}
std::vector<CommunicationRequest *> * swap = req_not_finished_tmp;
req_not_finished_tmp = recv_requests_tmp;
recv_requests_tmp = swap;
req_not_finished_tmp->clear();
}
AKANTU_DEBUG_INFO("Waiting that every send requests are received");
static_communicator->waitAll(communication.send_requests);
for (std::vector<CommunicationRequest *>::iterator req_it = communication.send_requests.begin();
req_it != communication.send_requests.end() ; ++req_it) {
CommunicationRequest & req = *(*req_it);
if(static_communicator->testRequest(&req)) {
UInt proc = req.getDestination();
CommunicationBuffer & buffer = communication.send_buffer[proc];
buffer.resize(0);
static_communicator->freeCommunicationRequest(&req);
}
}
communication.send_requests.clear();
AKANTU_DEBUG_OUT();
}
/* -------------------------------------------------------------------------- */
/**
* This function computes the buffer size needed for in order to send data or receive data.
* @param data_accessor object of a class that needs to use the DOFSynchronizer. This class has to inherit from the * DataAccessor interface.
* @param tag synchronization tag: indicates what variable should be sychronized, e.g. mass, nodal residual, etc.
* for the SolidMechanicsModel
*/
void DOFSynchronizer::computeBufferSize(DataAccessor & data_accessor,
SynchronizationTag tag) {
AKANTU_DEBUG_IN();
communications[tag].resize(psize);
for (UInt p = 0; p < psize; ++p) {
/// initialize the size of data that we be sent and received
UInt ssend = 0;
UInt sreceive = 0;
if(p != prank) {
/// check if processor prank has to send dof information to p
if(proc_informations[p].slave_dofs.getSize() != 0) {
#ifndef AKANTU_NDEBUG
/// in debug mode increase buffer size to send the positions
/// of nodes in the direction that correspond to the dofs
ssend += proc_informations[p].slave_dofs.getSize() * sizeof(Int);
#endif
ssend += data_accessor.getNbDataForDOFs(proc_informations[p].slave_dofs, tag);
AKANTU_DEBUG_INFO("I have " << ssend << "(" << ssend / 1024.
<< "kB - "<< proc_informations[p].slave_dofs.getSize() <<" dof(s)) data to send to " << p << " for tag "
<< tag);
}
/// check if processor prank has to receive dof information from p
if(proc_informations[p].master_dofs.getSize() != 0) {
#ifndef AKANTU_NDEBUG
/// in debug mode increase buffer size to receive the
/// positions of nodes in the direction that correspond to the
/// dofs
sreceive += proc_informations[p].master_dofs.getSize() * sizeof(Int);
#endif
sreceive += data_accessor.getNbDataForDOFs(proc_informations[p].master_dofs, tag);
AKANTU_DEBUG_INFO("I have " << sreceive << "(" << sreceive / 1024.
<< "kB - "<< proc_informations[p].master_dofs.getSize() <<" dof(s)) data to receive for tag "
<< tag);
}
}
communications[tag].size_to_send [p] = ssend;
communications[tag].size_to_receive[p] = sreceive;
}
AKANTU_DEBUG_OUT();
}
/* -------------------------------------------------------------------------- */
void DOFSynchronizer::initGlobalDOFEquationNumbers() {
AKANTU_DEBUG_IN();
global_dof_equation_numbers.resize(nb_dofs);
Int * dof_type = dof_types.storage();
UInt * dof_global_id = dof_global_ids.storage();
Int * dof_equation_number = global_dof_equation_numbers.storage();
for(UInt d = 0; d < nb_dofs; ++d) {
/// if ghost dof the equation_number is greater than nb_global_dofs
Int global_eq_num = *dof_global_id + (*dof_type > -3 ? 0 : nb_global_dofs);
*(dof_equation_number) = global_eq_num;
global_dof_equation_number_to_local[global_eq_num] = d;
dof_equation_number++;
dof_global_id++;
dof_type++;
}
AKANTU_DEBUG_OUT();
}
/* -------------------------------------------------------------------------- */
void DOFSynchronizer::initScatterGatherCommunicationScheme() {
AKANTU_DEBUG_IN();
if(psize == 1 || gather_scatter_scheme_initialized) {
if(psize == 1) gather_scatter_scheme_initialized = true;
AKANTU_DEBUG_OUT();
return;
}
/// creating communication scheme
UInt * dof_global_id = dof_global_ids.storage();
Int * dof_type = dof_types.storage();
Array<UInt> * local_dofs = new Array<UInt>(0,2);
UInt local_dof_val[2];
nb_needed_dofs = 0;
nb_local_dofs = 0;
for (UInt n = 0; n < nb_dofs; ++n) {
if(*dof_type != -3) {
local_dof_val[0] = *dof_global_id;
if(*dof_type == -1 || *dof_type == -2) {
local_dof_val[1] = 0; // master for this node, shared or not
nb_local_dofs++;
} else if (*dof_type >= 0) {
nb_needed_dofs++;
local_dof_val[1] = 1; // slave node
}
local_dofs->push_back(local_dof_val);
}
dof_type++;
dof_global_id++;
}
Int * nb_dof_per_proc = new Int[psize];
nb_dof_per_proc[prank] = local_dofs->getSize();
static_communicator->allGather(nb_dof_per_proc, 1);
AKANTU_DEBUG(dblDebug, "I have " << local_dofs->getSize() << " not ghost dofs ("
<< nb_local_dofs << " local and " << nb_needed_dofs << " slave)");
UInt pos = 0;
for (UInt p = 0; p < prank; ++p) pos += nb_dof_per_proc[p];
UInt nb_total_dofs = pos;
for (UInt p = prank; p < psize; ++p) nb_total_dofs += nb_dof_per_proc[p];
int * nb_values = new int[psize];
for (unsigned int p = 0; p < psize; ++p) nb_values[p] = nb_dof_per_proc[p] * 2;
UInt * buffer = new UInt[2*nb_total_dofs];
memcpy(buffer + 2 * pos, local_dofs->storage(), local_dofs->getSize() * 2 * sizeof(UInt));
delete local_dofs;
static_communicator->allGatherV(buffer, nb_values);
UInt * tmp_buffer = buffer;
for (UInt p = 0; p < psize; ++p) {
UInt proc_p_nb_dof = nb_dof_per_proc[p];
if (p != prank){
AKANTU_DEBUG(dblDebug, "I get " << proc_p_nb_dof << "(" << nb_values[p] << ") dofs from " << p + 1);
proc_informations[p].dofs.resize(0);
for (UInt dd = 0; dd < proc_p_nb_dof; ++dd) {
UInt dof = tmp_buffer[2*dd];
if(tmp_buffer[2*dd + 1] == 0)
proc_informations[p].dofs.push_back(dof);
else
proc_informations[p].needed_dofs.push_back(dof);
}
AKANTU_DEBUG(dblDebug, "Proc " << p + 1 << " sends me "
<< proc_informations[p].dofs.getSize() << " local dofs, and "
<< proc_informations[p].needed_dofs.getSize() << " slave dofs");
}
tmp_buffer += 2*proc_p_nb_dof;
}
delete [] nb_dof_per_proc;
delete [] buffer;
delete [] nb_values;
gather_scatter_scheme_initialized = true;
AKANTU_DEBUG_OUT();
}
__END_AKANTU__
c4science · Help