/// Broadcasts an InterpolationObject from rank 0 to all other ranks.
///
/// It is commonly the case that the data needed to create the
/// interpolation table is available on only one task (for example, only
/// one task has read the data from a file). Broadcasting the table
/// eliminates the need to put broadcast code in multiple table readers.
///
/// \see eamBcastPotential
void bcastInterpolationObject(InterpolationObject** table)
{
struct
{
int n;
real_t x0, invDx;
} buf;
if (getMyRank() == 0)
{
buf.n = (*table)->n;
buf.x0 = (*table)->x0;
buf.invDx = (*table)->invDx;
}
bcastParallel(&buf, sizeof(buf), 0);
if (getMyRank() != 0)
{
assert(*table == NULL);
*table = comdMalloc(sizeof(InterpolationObject));
(*table)->n = buf.n;
(*table)->x0 = buf.x0;
(*table)->invDx = buf.invDx;
(*table)->values = comdMalloc(sizeof(real_t) * (buf.n+3) );
(*table)->values++;
}
int valuesSize = sizeof(real_t) * ((*table)->n+3);
bcastParallel((*table)->values-1, valuesSize, 0);
}
/// Allocate and initialize the EAM potential data structure.
///
/// \param [in] dir The directory in which potential table files are found.
/// \param [in] file The name of the potential table file.
/// \param [in] type The file format of the potential file (setfl or funcfl).
BasePotential* initEamPot(const char* dir, const char* file, const char* type)
{
EamPotential* pot = comdMalloc(sizeof(EamPotential));
assert(pot);
pot->force = eamForce;
pot->print = eamPrint;
pot->destroy = eamDestroy;
pot->phi = NULL;
pot->rho = NULL;
pot->f = NULL;
// Initialization of the next three items requires information about
// the parallel decomposition and link cells that isn't available
// with the potential is initialized. Hence, we defer their
// initialization until the first time we call the force routine.
pot->dfEmbed = NULL;
pot->rhobar = NULL;
pot->forceExchange = NULL;
if (getMyRank() == 0)
{
if (strcmp(type, "setfl" ) == 0)
eamReadSetfl(pot, dir, file);
else if (strcmp(type,"funcfl") == 0)
eamReadFuncfl(pot, dir, file);
else
typeNotSupported("initEamPot", type);
}
eamBcastPotential(pot);
return (BasePotential*) pot;
}
/// Broadcasts an EamPotential from rank 0 to all other ranks.
/// If the table coefficients are read from a file only rank 0 does the
/// read. Hence we need to broadcast the potential to all other ranks.
void eamBcastPotential(EamPotential* pot)
{
assert(pot);
struct
{
real_t cutoff, mass, lat;
char latticeType[8];
char name[3];
int atomicNo;
} buf;
if (getMyRank() == 0)
{
buf.cutoff = pot->cutoff;
buf.mass = pot->mass;
buf.lat = pot->lat;
buf.atomicNo = pot->atomicNo;
strcpy(buf.latticeType, pot->latticeType);
strcpy(buf.name, pot->name);
}
bcastParallel(&buf, sizeof(buf), 0);
pot->cutoff = buf.cutoff;
pot->mass = buf.mass;
pot->lat = buf.lat;
pot->atomicNo = buf.atomicNo;
strcpy(pot->latticeType, buf.latticeType);
strcpy(pot->name, buf.name);
bcastInterpolationObject(&pot->phi);
bcastInterpolationObject(&pot->rho);
bcastInterpolationObject(&pot->f);
}
/// \param [in] xproc x-size of domain decomposition grid.
/// \param [in] yproc y-size of domain decomposition grid.
/// \param [in] zproc z-size of domain decomposition grid.
/// \param [in] globalExtent Size of the simulation domain (in Angstroms).
Domain* initDecomposition(int xproc, int yproc, int zproc, real3 globalExtent)
{
assert( xproc * yproc * zproc == getNRanks());
Domain* dd = comdMalloc(sizeof(Domain));
dd->procGrid[0] = xproc;
dd->procGrid[1] = yproc;
dd->procGrid[2] = zproc;
// calculate grid coordinates i,j,k for this processor
int myRank = getMyRank();
dd->procCoord[0] = myRank % dd->procGrid[0];
myRank /= dd->procGrid[0];
dd->procCoord[1] = myRank % dd->procGrid[1];
dd->procCoord[2] = myRank / dd->procGrid[1];
// initialialize global bounds
for (int i = 0; i < 3; i++)
{
dd->globalMin[i] = 0;
dd->globalMax[i] = globalExtent[i];
dd->globalExtent[i] = dd->globalMax[i] - dd->globalMin[i];
}
// initialize local bounds on this processor
for (int i = 0; i < 3; i++)
{
dd->localExtent[i] = dd->globalExtent[i] / dd->procGrid[i];
dd->localMin[i] = dd->globalMin[i] + dd->procCoord[i] * dd->localExtent[i];
dd->localMax[i] = dd->globalMin[i] + (dd->procCoord[i]+1) * dd->localExtent[i];
}
return dd;
}
void Filtro2::GetParameter () {
string parameter;
parameter = getArgument ( "a" );
if ( parameter.length () != 0 ) {
char rank[5];
sprintf ( rank, "%d", getMyRank () );
inputFile = parameter;
}
parameter = getArgument ( "e" );
if ( parameter.length () != 0 ) {
groupsOut = parameter;
}
parameter = getArgument ( "f" );
if ( parameter.length () != 0 ) {
deletedOut = parameter;
}
parameter = getArgument ( "g" );
if ( parameter.length () != 0 ) {
maximalOut = parameter;
}
}
/// \details
/// The report contains two blocks. The upper block is performance
/// information for the printRank. The lower block is statistical
/// information over all ranks.
void printPerformanceResults(int nGlobalAtoms, int printRate)
{
// Collect timer statistics overall and across ranks
timerStats();
if (!printRank())
return;
// only print timers with non-zero values.
double tick = getTick();
double loopTime = perfTimer[loopTimer].total*tick;
fprintf(screenOut, "\n\nTimings for Rank %d\n", getMyRank());
fprintf(screenOut, " Timer # Calls Avg/Call (s) Total (s) %% Loop\n");
fprintf(screenOut, "___________________________________________________________________\n");
/* for (int ii=0; ii<numberOfTimers; ++ii)
{
double totalTime = perfTimer[ii].total*tick;
if (perfTimer[ii].count > 0)
fprintf(screenOut, "%-16s%12"PRIu64" %8.4f %8.4f %8.2f\n",
timerName[ii],
perfTimer[ii].count,
totalTime/(double)perfTimer[ii].count,
totalTime,
totalTime/loopTime*100.0);
}*/
fprintf(screenOut, "\nTiming Statistics Across %d Ranks:\n", getNRanks());
fprintf(screenOut, " Timer Rank: Min(s) Rank: Max(s) Avg(s) Stdev(s)\n");
fprintf(screenOut, "_____________________________________________________________________________\n");
/* for (int ii = 0; ii < numberOfTimers; ++ii)
{
if (perfTimer[ii].count > 0)
fprintf(screenOut, "%-16s%6d:%10.4f %6d:%10.4f %10.4f %10.4f\n",
timerName[ii],
perfTimer[ii].minRank, perfTimer[ii].minValue*tick,
perfTimer[ii].maxRank, perfTimer[ii].maxValue*tick,
perfTimer[ii].average*tick, perfTimer[ii].stdev*tick);
}*/
double atomsPerTask = nGlobalAtoms/(real_t)getNRanks();
perfGlobal.atomRate = perfTimer[timestepTimer].average * tick * 1e6 /
(atomsPerTask * perfTimer[timestepTimer].count * printRate);
perfGlobal.atomAllRate = perfTimer[timestepTimer].average * tick * 1e6 /
(nGlobalAtoms * perfTimer[timestepTimer].count * printRate);
perfGlobal.atomsPerUSec = 1.0 / perfGlobal.atomAllRate;
fprintf(screenOut, "\n---------------------------------------------------\n");
// fprintf(screenOut, " Average atom update rate: %6.2f us/atom/task\n", perfGlobal.atomRate);
fprintf(screenOut, "---------------------------------------------------\n\n");
fprintf(screenOut, "\n---------------------------------------------------\n");
// fprintf(screenOut, " Average all atom update rate: %6.2f us/atom\n", perfGlobal.atomAllRate);
fprintf(screenOut, "---------------------------------------------------\n\n");
fprintf(screenOut, "\n---------------------------------------------------\n");
// fprintf(screenOut, " Average atom rate: %6.2f atoms/us\n", perfGlobal.atomsPerUSec);
fprintf(screenOut, "---------------------------------------------------\n\n");
}
void printPerformanceResultsYaml(FILE* file)
{
if (! printRank())
return;
double tick = getTick();
double loopTime = perfTimer[loopTimer].total*tick;
fprintf(file,"\nPerformance Results:\n");
fprintf(file, " TotalRanks: %d\n", getNRanks());
fprintf(file, " ReportingTimeUnits: seconds\n");
fprintf(file, "Performance Results For Rank %d:\n", getMyRank());
for (int ii = 0; ii < numberOfTimers; ii++)
{
if (perfTimer[ii].count > 0)
{
double totalTime = perfTimer[ii].total*tick;
fprintf(file, " Timer: %s\n", timerName[ii]);
fprintf(file, " CallCount: %"PRIu64"\n", perfTimer[ii].count);
fprintf(file, " AvgPerCall: %8.4f\n", totalTime/(double)perfTimer[ii].count);
fprintf(file, " Total: %8.4f\n", totalTime);
fprintf(file, " PercentLoop: %8.2f\n", totalTime/loopTime*100);
}
}
fprintf(file, "Performance Results Across Ranks:\n");
for (int ii = 0; ii < numberOfTimers; ii++)
{
if (perfTimer[ii].count > 0)
{
fprintf(file, " Timer: %s\n", timerName[ii]);
fprintf(file, " MinRank: %d\n", perfTimer[ii].minRank);
fprintf(file, " MinTime: %8.4f\n", perfTimer[ii].minValue*tick);
fprintf(file, " MaxRank: %d\n", perfTimer[ii].maxRank);
fprintf(file, " MaxTime: %8.4f\n", perfTimer[ii].maxValue*tick);
fprintf(file, " AvgTime: %8.4f\n", perfTimer[ii].average*tick);
fprintf(file, " StdevTime: %8.4f\n", perfTimer[ii].stdev*tick);
}
}
fprintf(file,"Performance Global Update Rates:\n");
fprintf(file, " AtomUpdateRate:\n");
fprintf(file, " AverageRate: %6.2f\n", perfGlobal.atomRate);
fprintf(file, " Units: us/atom/task\n");
fprintf(file, " AllAtomUpdateRate:\n");
fprintf(file, " AverageRate: %6.2f\n", perfGlobal.atomAllRate);
fprintf(file, " Units: us/atom\n");
fprintf(file, " AtomRate:\n");
fprintf(file, " AverageRate: %6.2f\n", perfGlobal.atomsPerUSec);
fprintf(file, " Units: atoms/us\n");
fprintf(file, "\n");
}
void Filtro2::printDeleted(set<set<int > > deleted) {
int count=0;
cout << "GetMyRank: " << getMyRank();
cout << " Deleted: " << endl;
for(set < set< int> >::iterator it1=deleted.begin();it1!=deleted.end();it1++) {
for(set < int >::iterator it2=(*it1).begin();it2!=(*it1).end();it2++) {
cout << (*it2) << " ";
++count;
}
cout << " " << endl;
}
}
/// Collect timer statistics across ranks.
void timerStats(void)
{
double sendBuf[numberOfTimers], recvBuf[numberOfTimers];
// Determine average of each timer across ranks
for (int ii = 0; ii < numberOfTimers; ii++)
sendBuf[ii] = (double)perfTimer[ii].total;
addDoubleParallel(sendBuf, recvBuf, numberOfTimers);
for (int ii = 0; ii < numberOfTimers; ii++)
perfTimer[ii].average = recvBuf[ii] / (double)getNRanks();
// Determine min and max across ranks and which rank
RankReduceData reduceSendBuf[numberOfTimers], reduceRecvBuf[numberOfTimers];
for (int ii = 0; ii < numberOfTimers; ii++)
{
reduceSendBuf[ii].val = (double)perfTimer[ii].total;
reduceSendBuf[ii].rank = getMyRank();
}
minRankDoubleParallel(reduceSendBuf, reduceRecvBuf, numberOfTimers);
for (int ii = 0; ii < numberOfTimers; ii++)
{
perfTimer[ii].minValue = reduceRecvBuf[ii].val;
perfTimer[ii].minRank = reduceRecvBuf[ii].rank;
}
maxRankDoubleParallel(reduceSendBuf, reduceRecvBuf, numberOfTimers);
for (int ii = 0; ii < numberOfTimers; ii++)
{
perfTimer[ii].maxValue = reduceRecvBuf[ii].val;
perfTimer[ii].maxRank = reduceRecvBuf[ii].rank;
}
// Determine standard deviation
for (int ii = 0; ii < numberOfTimers; ii++)
{
double temp = (double)perfTimer[ii].total - perfTimer[ii].average;
sendBuf[ii] = temp * temp;
}
addDoubleParallel(sendBuf, recvBuf, numberOfTimers);
for (int ii = 0; ii < numberOfTimers; ii++)
{
perfTimer[ii].stdev = sqrt(recvBuf[ii] / (double) getNRanks());
}
}
int main( int argc, char** argv )
{
bool localRank = false;
bool myRank = false;
bool totalRank = false;
po::options_description desc( "Allowed options" );
desc.add_options( )
( "help,h", "produce help message" )
( "mpi_host_rank", po::value<bool > ( &localRank )->zero_tokens( ), "get local mpi rank" )
( "mpi_rank", po::value<bool > ( &myRank )->zero_tokens( ), "get mpi rank" )
( "mpi_size", po::value<bool > ( &totalRank )->zero_tokens( ), "get count of mpi ranks" );
// parse command line options and config file and store values in vm
po::variables_map vm;
po::store( boost::program_options::parse_command_line( argc, argv, desc ), vm );
po::notify( vm );
// print help message and quit simulation
if ( vm.count( "help" ) )
{
std::cerr << desc << "\n";
return false;
}
MPI_CHECK( MPI_Init( &argc, &argv ) );
if ( localRank )
std::cout << "mpi_host_rank: " << getHostRank( ) << std::endl;
if ( myRank )
std::cout << "mpi_rank: " << getMyRank( ) << std::endl;
if ( totalRank )
std::cout << "mpi_size: " << getTotalRanks( ) << std::endl;
MPI_CHECK( MPI_Finalize( ) );
return 0;
}
