olb::ConcreteBlockCommunicator< ConcreteBlockLattice< T, DESCRIPTOR, Platform::GPU_CUDA > >::SendTask Class Reference

Wrapper for a non-blocking block propagation send request. More...

#include <communicator.hh>

Collaboration diagram for olb::ConcreteBlockCommunicator< ConcreteBlockLattice< T, DESCRIPTOR, Platform::GPU_CUDA > >::SendTask:

Public Member Functions
	SendTask (MPI_Comm comm, int tag, int rank, const std::vector< std::type_index > &fields, const std::vector< CellID > &cells, ConcreteBlockLattice< T, DESCRIPTOR, Platform::GPU_CUDA > &block)
	~SendTask ()
void	prepare ()
void	send ()
void	wait ()

Detailed Description

template<typename T, typename DESCRIPTOR>
class olb::ConcreteBlockCommunicator< ConcreteBlockLattice< T, DESCRIPTOR, Platform::GPU_CUDA > >::SendTask

Wrapper for a non-blocking block propagation send request.

Definition at line 351 of file communicator.hh.

Constructor & Destructor Documentation

SendTask()

template<typename T , typename DESCRIPTOR >

olb::ConcreteBlockCommunicator< ConcreteBlockLattice< T, DESCRIPTOR, Platform::GPU_CUDA > >::SendTask::SendTask ( MPI_Comm comm, int tag, int rank, const std::vector< std::type_index > & fields, const std::vector< CellID > & cells, ConcreteBlockLattice< T, DESCRIPTOR, Platform::GPU_CUDA > & block )

inline

Definition at line 366 of file communicator.hh.

                                                                      :
    _fields(block.getDataRegistry().deviceFieldArrays(fields)),
    _onlyPopulationField(fields.size() == 1 && fields[0] == typeid(descriptors::POPULATION)),
    _cells(cells),
    _source(block),
    _stream(std::make_unique<gpu::cuda::device::Stream>(cudaStreamNonBlocking))
  {
    std::size_t size = 0;
    for (auto& field : fields) {
      size += _source.getCommunicatable(field).size(cells);
    }
    _buffer = gpu::cuda::device::malloc<std::uint8_t>(size);
    _request = std::make_unique<MpiSendRequest>(
      _buffer.get(), size, rank, tag, comm);
  }

References olb::gpu::cuda::device::unique_ptr< T >::get(), olb::ConcreteBlockLattice< T, DESCRIPTOR, PLATFORM >::getCommunicatable(), and olb::Communicatable::size().

Here is the call graph for this function:

~SendTask()

template<typename T , typename DESCRIPTOR >

olb::ConcreteBlockCommunicator< ConcreteBlockLattice< T, DESCRIPTOR, Platform::GPU_CUDA > >::SendTask::~SendTask ( )

inline

Definition at line 385 of file communicator.hh.

  {
    _stream->synchronize();
    wait();
  }

References olb::ConcreteBlockCommunicator< BLOCK >::wait().

Here is the call graph for this function:

Member Function Documentation

prepare()

template<typename T , typename DESCRIPTOR >

void olb::ConcreteBlockCommunicator< ConcreteBlockLattice< T, DESCRIPTOR, Platform::GPU_CUDA > >::SendTask::prepare ( )

inline

Definition at line 391 of file communicator.hh.

  {
    if (_onlyPopulationField) {
      gpu::cuda::DeviceContext<T,DESCRIPTOR> lattice(_source);
      gpu::cuda::async_gather_field<descriptors::POPULATION>(_stream->get(), lattice, _cells, _buffer.get());
    } else {
      gpu::cuda::async_gather_any_fields(_stream->get(), _fields, _cells, _buffer.get());
    }
  }

References olb::gpu::cuda::async_gather_any_fields(), and olb::gpu::cuda::device::unique_ptr< T >::get().

Here is the call graph for this function:

send()

template<typename T , typename DESCRIPTOR >

void olb::ConcreteBlockCommunicator< ConcreteBlockLattice< T, DESCRIPTOR, Platform::GPU_CUDA > >::SendTask::send ( )

inline

Definition at line 401 of file communicator.hh.

  {
    _stream->synchronize();
    _request->start();
  }

wait()

template<typename T , typename DESCRIPTOR >

void olb::ConcreteBlockCommunicator< ConcreteBlockLattice< T, DESCRIPTOR, Platform::GPU_CUDA > >::SendTask::wait ( )

inline

Definition at line 407 of file communicator.hh.

  {
    _request->wait();
  }

---

The documentation for this class was generated from the following file:

• src/core/platform/gpu/cuda/communicator.hh