HINLINE void operator()(RefWrapper<ThreadParams*> params,
std::string prefix,
const DomainInformation domInfo,const Space particleOffset)
{
log<picLog::INPUT_OUTPUT > ("HDF5: write species: %1%") % Hdf5FrameType::getName();
DataConnector &dc = DataConnector::getInstance();
/*load particle without copy particle data to host*/
ThisSpecies* speciesTmp = &(dc.getData<ThisSpecies >(ThisSpecies::FrameType::CommunicationTag, true));
// count total number of particles on the device
uint64_cu totalNumParticles = 0;
PMACC_AUTO(simBox, SubGrid<simDim>::getInstance().getSimulationBox());
log<picLog::INPUT_OUTPUT > ("HDF5: count particles: %1%") % Hdf5FrameType::getName();
totalNumParticles = PMacc::CountParticles::countOnDevice < CORE + BORDER > (
*speciesTmp,
*(params.get()->cellDescription),
domInfo.localDomainOffset,
domInfo.domainSize);
log<picLog::INPUT_OUTPUT > ("HDF5: Finish count particles: %1% = %2%") % Hdf5FrameType::getName() % totalNumParticles;
Hdf5FrameType hostFrame;
log<picLog::INPUT_OUTPUT > ("HDF5: malloc mapped memory: %1%") % Hdf5FrameType::getName();
/*malloc mapped memory*/
ForEach<typename Hdf5FrameType::ValueTypeSeq, MallocMemory<void> > mallocMem;
mallocMem(byRef(hostFrame), totalNumParticles);
log<picLog::INPUT_OUTPUT > ("HDF5: Finish malloc mapped memory: %1%") % Hdf5FrameType::getName();
if (totalNumParticles != 0)
{
log<picLog::INPUT_OUTPUT > ("HDF5: get mapped memory device pointer: %1%") % Hdf5FrameType::getName();
/*load device pointer of mapped memory*/
Hdf5FrameType deviceFrame;
ForEach<typename Hdf5FrameType::ValueTypeSeq, GetDevicePtr<void> > getDevicePtr;
getDevicePtr(byRef(deviceFrame), byRef(hostFrame));
log<picLog::INPUT_OUTPUT > ("HDF5: Finish get mapped memory device pointer: %1%") % Hdf5FrameType::getName();
log<picLog::INPUT_OUTPUT > ("HDF5: copy particle to host: %1%") % Hdf5FrameType::getName();
typedef bmpl::vector< PositionFilter3D<> > usedFilters;
typedef typename FilterFactory<usedFilters>::FilterType MyParticleFilter;
MyParticleFilter filter;
/*activeate filter pipline if moving window is activated*/
filter.setStatus(MovingWindow::getInstance().isSlidingWindowActive());
filter.setWindowPosition(domInfo.localDomainOffset, domInfo.domainSize);
dim3 block(TILE_SIZE);
DataSpace<simDim> superCells = speciesTmp->getParticlesBuffer().getSuperCellsCount();
GridBuffer<int, DIM1> counterBuffer(DataSpace<DIM1>(1));
AreaMapping < CORE + BORDER, MappingDesc > mapper(*(params.get()->cellDescription));
__cudaKernel(copySpecies)
(mapper.getGridDim(), block)
(counterBuffer.getDeviceBuffer().getPointer(),
deviceFrame, speciesTmp->getDeviceParticlesBox(),
filter,
particleOffset, /*relative to data domain (not to physical domain)*/
mapper
);
counterBuffer.deviceToHost();
log<picLog::INPUT_OUTPUT > ("HDF5: memcpy particle counter to host: %1%") % Hdf5FrameType::getName();
__getTransactionEvent().waitForFinished();
log<picLog::INPUT_OUTPUT > ("HDF5: all events are finish: %1%") % Hdf5FrameType::getName();
/*this cost a little bit of time but hdf5 writing is slower^^*/
assert((uint64_cu) counterBuffer.getHostBuffer().getDataBox()[0] == totalNumParticles);
}
/*dump to hdf5 file*/
ForEach<typename Hdf5FrameType::ValueTypeSeq, hdf5::ParticleAttribute<void> > writeToHdf5;
writeToHdf5(params, byRef(hostFrame), prefix + FrameType::getName(), domInfo, totalNumParticles);
/*free host memory*/
ForEach<typename Hdf5FrameType::ValueTypeSeq, FreeMemory<void> > freeMem;
freeMem(byRef(hostFrame));
log<picLog::INPUT_OUTPUT > ("HDF5: Finish write species: %1%") % Hdf5FrameType::getName();
}
/** Load species from HDF5 checkpoint file
*
* @param params thread params with domainwriter, ...
* @param restartChunkSize number of particles processed in one kernel call
*/
HINLINE void operator()(ThreadParams* params, const uint32_t restartChunkSize)
{
log<picLog::INPUT_OUTPUT > ("HDF5: (begin) load species: %1%") % Hdf5FrameType::getName();
DataConnector &dc = Environment<>::get().DataConnector();
GridController<simDim> &gc = Environment<simDim>::get().GridController();
std::string subGroup = std::string("particles/") + FrameType::getName();
const PMacc::Selection<simDim>& localDomain = Environment<simDim>::get().SubGrid().getLocalDomain();
/* load particle without copying particle data to host */
ThisSpecies* speciesTmp = &(dc.getData<ThisSpecies >(ThisSpecies::FrameType::getName(), true));
/* count total number of particles on the device */
uint64_cu totalNumParticles = 0;
/* load particles info table entry for this process
particlesInfo is (part-count, scalar pos, x, y, z) */
typedef uint64_t uint64Quint[5];
uint64Quint particlesInfo[gc.getGlobalSize()];
Dimensions particlesInfoSizeRead;
params->dataCollector->read(params->currentStep,
(std::string(subGroup) + std::string("/particles_info")).c_str(),
particlesInfoSizeRead,
particlesInfo);
assert(particlesInfoSizeRead[0] == gc.getGlobalSize());
/* search my entry (using my scalar position) in particlesInfo */
uint64_t particleOffset = 0;
uint64_t myScalarPos = gc.getScalarPosition();
for (size_t i = 0; i < particlesInfoSizeRead[0]; ++i)
{
if (particlesInfo[i][1] == myScalarPos)
{
totalNumParticles = particlesInfo[i][0];
break;
}
particleOffset += particlesInfo[i][0];
}
log<picLog::INPUT_OUTPUT > ("Loading %1% particles from offset %2%") %
(long long unsigned) totalNumParticles % (long long unsigned) particleOffset;
if (totalNumParticles != 0)
{
Hdf5FrameType hostFrame;
log<picLog::INPUT_OUTPUT > ("HDF5: malloc mapped memory: %1%") % Hdf5FrameType::getName();
/*malloc mapped memory*/
ForEach<typename Hdf5FrameType::ValueTypeSeq, MallocMemory<bmpl::_1> > mallocMem;
mallocMem(forward(hostFrame), totalNumParticles);
log<picLog::INPUT_OUTPUT > ("HDF5: get mapped memory device pointer: %1%") % Hdf5FrameType::getName();
/*load device pointer of mapped memory*/
Hdf5FrameType deviceFrame;
ForEach<typename Hdf5FrameType::ValueTypeSeq, GetDevicePtr<bmpl::_1> > getDevicePtr;
getDevicePtr(forward(deviceFrame), forward(hostFrame));
ForEach<typename Hdf5FrameType::ValueTypeSeq, LoadParticleAttributesFromHDF5<bmpl::_1> > loadAttributes;
loadAttributes(forward(params), forward(hostFrame), subGroup, particleOffset, totalNumParticles);
dim3 block(PMacc::math::CT::volume<SuperCellSize>::type::value);
/* counter is used to apply for work, count used frames and count loaded particles
* [0] -> offset for loading particles
* [1] -> number of loaded particles
* [2] -> number of used frames
*
* all values are zero after initialization
*/
GridBuffer<uint32_t, DIM1> counterBuffer(DataSpace<DIM1>(3));
const uint32_t cellsInSuperCell = PMacc::math::CT::volume<SuperCellSize>::type::value;
const uint32_t iterationsForLoad = ceil(double(totalNumParticles) / double(restartChunkSize));
uint32_t leftOverParticles = totalNumParticles;
__startAtomicTransaction(__getTransactionEvent());
for (uint32_t i = 0; i < iterationsForLoad; ++i)
{
/* only load a chunk of particles per iteration to avoid blow up of frame usage
*/
uint32_t currentChunkSize = std::min(leftOverParticles, restartChunkSize);
log<picLog::INPUT_OUTPUT > ("HDF5: load particles on device chunk offset=%1%; chunk size=%2%; left particles %3%") %
(i * restartChunkSize) % currentChunkSize % leftOverParticles;
__cudaKernel(copySpeciesGlobal2Local)
(ceil(double(currentChunkSize) / double(cellsInSuperCell)), cellsInSuperCell)
(counterBuffer.getDeviceBuffer().getDataBox(),
speciesTmp->getDeviceParticlesBox(), deviceFrame,
(int) totalNumParticles,
localDomain.offset, /*relative to data domain (not to physical domain)*/
*(params->cellDescription)
);
speciesTmp->fillAllGaps();
leftOverParticles -= currentChunkSize;
}
__setTransactionEvent(__endTransaction());
counterBuffer.deviceToHost();
log<picLog::INPUT_OUTPUT > ("HDF5: wait for last processed chunk: %1%") % Hdf5FrameType::getName();
__getTransactionEvent().waitForFinished();
log<picLog::INPUT_OUTPUT > ("HDF5: used frames to load particles: %1%") % counterBuffer.getHostBuffer().getDataBox()[2];
if ((uint64_cu) counterBuffer.getHostBuffer().getDataBox()[1] != totalNumParticles)
{
log<picLog::INPUT_OUTPUT >("HDF5: error load species | counter is %1% but should %2%") % counterBuffer.getHostBuffer().getDataBox()[1] % totalNumParticles;
}
assert((uint64_cu) counterBuffer.getHostBuffer().getDataBox()[1] == totalNumParticles);
/*free host memory*/
ForEach<typename Hdf5FrameType::ValueTypeSeq, FreeMemory<bmpl::_1> > freeMem;
freeMem(forward(hostFrame));
log<picLog::INPUT_OUTPUT > ("HDF5: ( end ) load species: %1%") % Hdf5FrameType::getName();
}
}
HINLINE void operator()(RefWrapper<ThreadParams*> params,
std::string subGroup,
const Space particleOffset)
{
log<picLog::INPUT_OUTPUT > ("HDF5: (begin) write species: %1%") % Hdf5FrameType::getName();
DataConnector &dc = Environment<>::get().DataConnector();
/* load particle without copy particle data to host */
ThisSpecies* speciesTmp = &(dc.getData<ThisSpecies >(ThisSpecies::FrameType::getName(), true));
/* count total number of particles on the device */
uint64_cu totalNumParticles = 0;
log<picLog::INPUT_OUTPUT > ("HDF5: (begin) count particles: %1%") % Hdf5FrameType::getName();
totalNumParticles = PMacc::CountParticles::countOnDevice < CORE + BORDER > (
*speciesTmp,
*(params.get()->cellDescription),
params.get()->localWindowToDomainOffset,
params.get()->window.localDimensions.size);
log<picLog::INPUT_OUTPUT > ("HDF5: ( end ) count particles: %1% = %2%") % Hdf5FrameType::getName() % totalNumParticles;
Hdf5FrameType hostFrame;
log<picLog::INPUT_OUTPUT > ("HDF5: (begin) malloc mapped memory: %1%") % Hdf5FrameType::getName();
/*malloc mapped memory*/
ForEach<typename Hdf5FrameType::ValueTypeSeq, MallocMemory<bmpl::_1> > mallocMem;
mallocMem(byRef(hostFrame), totalNumParticles);
log<picLog::INPUT_OUTPUT > ("HDF5: ( end ) malloc mapped memory: %1%") % Hdf5FrameType::getName();
if (totalNumParticles != 0)
{
log<picLog::INPUT_OUTPUT > ("HDF5: (begin) get mapped memory device pointer: %1%") % Hdf5FrameType::getName();
/*load device pointer of mapped memory*/
Hdf5FrameType deviceFrame;
ForEach<typename Hdf5FrameType::ValueTypeSeq, GetDevicePtr<bmpl::_1> > getDevicePtr;
getDevicePtr(byRef(deviceFrame), byRef(hostFrame));
log<picLog::INPUT_OUTPUT > ("HDF5: ( end ) get mapped memory device pointer: %1%") % Hdf5FrameType::getName();
log<picLog::INPUT_OUTPUT > ("HDF5: (begin) copy particle to host: %1%") % Hdf5FrameType::getName();
typedef bmpl::vector< typename GetPositionFilter<simDim>::type > usedFilters;
typedef typename FilterFactory<usedFilters>::FilterType MyParticleFilter;
MyParticleFilter filter;
/* activate filter pipeline if moving window is activated */
filter.setStatus(MovingWindow::getInstance().isSlidingWindowActive());
filter.setWindowPosition(params.get()->localWindowToDomainOffset,
params.get()->window.localDimensions.size);
dim3 block(TILE_SIZE);
GridBuffer<int, DIM1> counterBuffer(DataSpace<DIM1>(1));
AreaMapping < CORE + BORDER, MappingDesc > mapper(*(params.get()->cellDescription));
__cudaKernel(copySpecies)
(mapper.getGridDim(), block)
(counterBuffer.getDeviceBuffer().getPointer(),
deviceFrame, speciesTmp->getDeviceParticlesBox(),
filter,
particleOffset, /*relative to data domain (not to physical domain)*/
mapper
);
counterBuffer.deviceToHost();
log<picLog::INPUT_OUTPUT > ("HDF5: ( end ) copy particle to host: %1%") % Hdf5FrameType::getName();
__getTransactionEvent().waitForFinished();
log<picLog::INPUT_OUTPUT > ("HDF5: all events are finished: %1%") % Hdf5FrameType::getName();
/*this cost a little bit of time but hdf5 writing is slower^^*/
assert((uint64_cu) counterBuffer.getHostBuffer().getDataBox()[0] == totalNumParticles);
}
/*dump to hdf5 file*/
ForEach<typename Hdf5FrameType::ValueTypeSeq, hdf5::ParticleAttribute<bmpl::_1> > writeToHdf5;
writeToHdf5(params, byRef(hostFrame), std::string("particles/") + FrameType::getName() + std::string("/") + subGroup,
totalNumParticles);
/* write meta attributes for species */
writeMetaAttributes(params.get());
/*write species counter table to hdf5 file*/
log<picLog::INPUT_OUTPUT > ("HDF5: (begin) writing particle index table for %1%") % Hdf5FrameType::getName();
{
ColTypeUInt64_5Array ctUInt64_5;
GridController<simDim>& gc = Environment<simDim>::get().GridController();
const size_t pos_offset = 2;
/* particlesMetaInfo = (num particles, scalar position, particle offset x, y, z) */
uint64_t particlesMetaInfo[5] = {totalNumParticles, gc.getScalarPosition(), 0, 0, 0};
for (size_t d = 0; d < simDim; ++d)
particlesMetaInfo[pos_offset + d] = particleOffset[d];
/* prevent that top (y) gpus have negative value here */
if (gc.getPosition().y() == 0)
particlesMetaInfo[pos_offset + 1] = 0;
if (particleOffset[1] < 0) // 1 == y
particlesMetaInfo[pos_offset + 1] = 0;
params.get()->dataCollector->write(
params.get()->currentStep,
Dimensions(gc.getGlobalSize(), 1, 1),
Dimensions(gc.getGlobalRank(), 0, 0),
ctUInt64_5, 1,
Dimensions(1, 1, 1),
(std::string("particles/") + FrameType::getName() + std::string("/") +
subGroup + std::string("/particles_info")).c_str(),
particlesMetaInfo);
}
log<picLog::INPUT_OUTPUT > ("HDF5: ( end ) writing particle index table for %1%") % Hdf5FrameType::getName();
/*free host memory*/
ForEach<typename Hdf5FrameType::ValueTypeSeq, FreeMemory<bmpl::_1> > freeMem;
freeMem(byRef(hostFrame));
log<picLog::INPUT_OUTPUT > ("HDF5: ( end ) writing species: %1%") % Hdf5FrameType::getName();
}
