HINLINE void operator()(
const RefWrapper<ThreadParams*> params,
const RefWrapper<FrameType> frame,
const size_t elements)
{
typedef T_Identifier Identifier;
typedef typename Identifier::type ValueType;
const uint32_t components = GetNComponents<ValueType>::value;
typedef typename GetComponentsType<ValueType>::type ComponentType;
log<picLog::INPUT_OUTPUT > ("ADIOS: (begin) write species attribute: %1%") % Identifier::getName();
ComponentType* tmpBfr = new ComponentType[elements];
for (uint32_t d = 0; d < components; d++)
{
ValueType* dataPtr = frame.get().getIdentifier(Identifier()).getPointer();
/* copy strided data from source to temporary buffer */
for (size_t i = 0; i < elements; ++i)
{
tmpBfr[i] = ((ComponentType*)dataPtr)[i * components];
}
int64_t adiosAttributeVarId = *(params.get()->adiosParticleAttrVarIds.begin());
params.get()->adiosParticleAttrVarIds.pop_front();
ADIOS_CMD(adios_write_byid(params.get()->adiosFileHandle, adiosAttributeVarId, tmpBfr));
}
__deleteArray(tmpBfr);
log<picLog::INPUT_OUTPUT > ("ADIOS: ( end ) write species attribute: %1%") %
Identifier::getName();
}
HINLINE void operator_impl(RefWrapper<ThreadParams*> params)
{
DataConnector &dc = Environment<>::get().DataConnector();
/*## update field ##*/
/*load FieldTmp without copy data to host*/
FieldTmp* fieldTmp = &(dc.getData<FieldTmp > (FieldTmp::getName(), true));
/*load particle without copy particle data to host*/
Species* speciesTmp = &(dc.getData<Species >(Species::FrameType::getName(), true));
fieldTmp->getGridBuffer().getDeviceBuffer().setValue(ValueType(0.0));
/*run algorithm*/
fieldTmp->computeValue < CORE + BORDER, Solver > (*speciesTmp, params.get()->currentStep);
EventTask fieldTmpEvent = fieldTmp->asyncCommunication(__getTransactionEvent());
__setTransactionEvent(fieldTmpEvent);
/* copy data to host that we can write same to disk*/
fieldTmp->getGridBuffer().deviceToHost();
dc.releaseData(Species::FrameType::getName());
/*## finish update field ##*/
params.get()->gridLayout = fieldTmp->getGridLayout();
/*write data to HDF5 file*/
Field::writeField(params.get(),
getName(),
getUnit(),
fieldTmp->getHostDataBox(),
ValueType());
dc.releaseData(FieldTmp::getName());
}
HINLINE void operator()(RefWrapper<ValueType> dest, RefWrapper<ValueType> src) const
{
type* ptr = NULL;
type* srcPtr = src.get().getIdentifier(T_Type()).getPointer();
if (srcPtr != NULL)
{
CUDA_CHECK(cudaHostGetDevicePointer(&ptr, srcPtr, 0));
}
dest.get().getIdentifier(T_Type()) =
VectorDataBox<type>(ptr);
}
HDINLINE void operator()(RefWrapper<ThreadParams*> params)
{
#ifndef __CUDA_ARCH__
DataConnector &dc = Environment<>::get().DataConnector();
T* field = &(dc.getData<T > (T::getName()));
params.get()->gridLayout = field->getGridLayout();
Field::writeField(params.get(),
T::getName(),
getUnit(),
field->getHostDataBox(),
ValueType());
dc.releaseData(T::getName());
#endif
}
HINLINE void operator()(RefWrapper<ValueType> v1, const size_t size) const
{
type* ptr = NULL;
if (size != 0)
{
CUDA_CHECK(cudaHostAlloc(&ptr, size * sizeof (type), cudaHostAllocMapped));
}
v1.get().getIdentifier(T_Type()) = VectorDataBox<type>(ptr);
}
HINLINE void operator()(RefWrapper<ValueType> value) const
{
type* ptr = value.get().getIdentifier(T_Type()).getPointer();
if (ptr != NULL)
CUDA_CHECK(cudaFreeHost(ptr));
}
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();
}
HDINLINE
void operator()(RefWrapper<T_T1> dest, const T_T2& src)
{
dest.get()[T_Key()]=src[T_Key()];
}
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();
}
