/**
* Starts copying data from host to device.
*/
void hostToDevice()
{
__startTransaction(__getTransactionEvent());
frames->hostToDevice();
EventTask ev1 = __endTransaction();
__startTransaction(__getTransactionEvent());
superCells->hostToDevice();
EventTask ev2 = __endTransaction();
__startTransaction(__getTransactionEvent());
nextFrames->hostToDevice();
EventTask ev3 = __endTransaction();
__startTransaction(__getTransactionEvent());
prevFrames->hostToDevice();
EventTask ev4 = __endTransaction();
__setTransactionEvent(ev1 + ev2 + ev3 + ev4);
}
EventTask asyncReceive(EventTask serialEvent, uint32_t recvEx)
{
if (hasReceiveExchange(recvEx))
{
__startAtomicTransaction(serialEvent + receiveEvents[recvEx]);
receiveEvents[recvEx] = receiveExchanges[recvEx]->startReceive();
__endTransaction();
return receiveEvents[recvEx];
}
return EventTask();
}
bool executeIntern()
{
switch (state)
{
case Init:
break;
case WaitForReceive:
if (NULL == Environment<>::get().Manager().getITaskIfNotFinished(lastReceiveEvent.getTaskId()))
{
state = InitInsert;
//bash is finished
__startTransaction();
lastSize = parBase.getParticlesBuffer().getReceiveExchangeStack(exchange).getHostParticlesCurrentSize();
parBase.insertParticles(exchange);
// std::cout<<"brecv = "<<parBase.getParticlesBuffer().getReceiveExchangeStack(exchange).getHostCurrentSize()<<std::endl;
tmpEvent = __endTransaction();
state = WaitForInsert;
}
break;
case InitInsert:
break;
case WaitForInsert:
if (NULL == Environment<>::get().Manager().getITaskIfNotFinished(tmpEvent.getTaskId()))
{
state=Wait;
assert(lastSize <= maxSize);
//check for next bash round
if (lastSize == maxSize)
{
std::cerr << "recv max size " << maxSize << " particles" << std::endl;
init(); //call init and run a full send cycle
}
else
{
state = Finished;
return true;
}
}
break;
case Wait:
break;
case Finished:
return true;
default:
return false;
}
return false;
}
EventTask asyncSend(EventTask serialEvent, uint32_t sendEx, EventTask &gpuFree)
{
if (hasSendExchange(sendEx))
{
__startAtomicTransaction(serialEvent + sendEvents[sendEx]);
sendEvents[sendEx] = sendExchanges[sendEx]->startSend(gpuFree);
__endTransaction();
/* add only the copy event, because all work on gpu can run after data is copyed
*/
return gpuFree;
}
return EventTask();
}
virtual void init()
{
state = Init;
EventTask serialEvent = __getTransactionEvent();
for (int i = 1; i < Exchanges; ++i)
{
if (buffer.getGridBuffer().hasSendExchange(i))
{
__startAtomicTransaction(serialEvent);
FieldFactory::getInstance().createTaskFieldSendExchange(buffer, i);
tmpEvent += __endTransaction();
}
}
state = WaitForSend;
}
void shiftParticles()
{
StrideMapping<AREA, DIM3, MappingDesc> mapper(this->cellDescription);
ParticlesBoxType pBox = particlesBuffer->getDeviceParticleBox();
__startTransaction(__getTransactionEvent());
do
{
__cudaKernel(kernelShiftParticles)
(mapper.getGridDim(), TileSize)
(pBox, mapper);
}
while (mapper.next());
__setTransactionEvent(__endTransaction());
}
HINLINE void operator()(
T_StorageTuple& tuple,
const uint32_t currentStep,
const T_Event eventInt,
T_Event& updateEvent,
T_Event& commEvent
) const
{
typedef typename HasFlag<FrameType, particlePusher<> >::type hasPusher;
if (hasPusher::value)
{
PMACC_AUTO(speciesPtr, tuple[SpeciesName()]);
__startTransaction(eventInt);
speciesPtr->update(currentStep);
commEvent += speciesPtr->asyncCommunication(__getTransactionEvent());
updateEvent += __endTransaction();
}
}
bool executeIntern()
{
switch (state)
{
case InitDone:
break;
case DeviceToHostFinished:
state = SendDone;
__startTransaction();
Environment<>::get().Factory().createTaskSendMPI(exchange, this);
__endTransaction(); //we need no blocking because we get a singnal if transaction is finished
break;
case SendDone:
break;
case Finish:
return true;
default:
return false;
}
return false;
}
virtual void init()
{
__startTransaction();
state = InitDone;
if (exchange->hasDeviceDoubleBuffer())
{
Environment<>::get().Factory().createTaskCopyDeviceToDevice(exchange->getDeviceBuffer(),
exchange->getDeviceDoubleBuffer()
);
copyEvent = Environment<>::get().Factory().createTaskCopyDeviceToHost(exchange->getDeviceDoubleBuffer(),
exchange->getHostBuffer(),
this);
}
else
{
copyEvent = Environment<>::get().Factory().createTaskCopyDeviceToHost(exchange->getDeviceBuffer(),
exchange->getHostBuffer(),
this);
}
__endTransaction(); //we need no blocking because we get a singnal if transaction is finished
}
bool executeIntern()
{
switch (state)
{
case WaitForReceived:
break;
case RunCopy:
state = WaitForFinish;
__startAtomicTransaction();
exchange->getHostBuffer().setCurrentSize(newBufferSize);
if (exchange->hasDeviceDoubleBuffer())
{
Factory::getInstance().createTaskCopyHostToDevice(exchange->getHostBuffer(),
exchange->getDeviceDoubleBuffer());
Factory::getInstance().createTaskCopyDeviceToDevice(exchange->getDeviceDoubleBuffer(),
exchange->getDeviceBuffer(),
this);
}
else
{
Factory::getInstance().createTaskCopyHostToDevice(exchange->getHostBuffer(),
exchange->getDeviceBuffer(),
this);
}
__endTransaction();
break;
case WaitForFinish:
break;
case Finish:
return true;
default:
return false;
}
return false;
}
void event(id_t, EventType type, IEventData* data)
{
switch (type)
{
case RECVFINISHED:
if (data != NULL)
{
__startTransaction(); //no blocking
EventDataReceive *rdata = static_cast<EventDataReceive*> (data);
// std::cout<<" data rec "<<rdata->getReceivedCount()/sizeof(TYPE)<<std::endl;
newBufferSize = rdata->getReceivedCount() / sizeof (TYPE);
__endTransaction();
state = RunCopy;
executeIntern();
}
break;
case COPYHOST2DEVICE:
case COPYDEVICE2DEVICE:
state = Finish;
break;
default:
return;
}
}
