const bool accumulate = runScheduleWork->domainWork.haveCpuLocalForceWork
|| runScheduleWork->simulationWork.havePpDomainDecomposition;
const int atomStart = 0;
- fr->gpuForceReduction[gmx::AtomLocality::Local]->reinit(stateGpu->getForces(),
- nbv->getNumAtoms(AtomLocality::Local),
- nbv->getGridIndices(),
- atomStart,
- accumulate,
- stateGpu->fReducedOnDevice());
+ fr->gpuForceReduction[gmx::AtomLocality::Local]->reinit(
+ stateGpu->getForces(),
+ nbv->getNumAtoms(AtomLocality::Local),
+ nbv->getGridIndices(),
+ atomStart,
+ accumulate,
+ stateGpu->fReducedOnDevice(AtomLocality::Local));
// register forces and add dependencies
fr->gpuForceReduction[gmx::AtomLocality::Local]->registerNbnxmForce(Nbnxm::gpu_get_f(nbv->gpu_nbv));
}
}
- if (runScheduleWork->domainWork.haveCpuLocalForceWork && !runScheduleWork->simulationWork.useGpuHaloExchange)
+ if (runScheduleWork->domainWork.haveCpuLocalForceWork
+ || (runScheduleWork->simulationWork.havePpDomainDecomposition
+ && !runScheduleWork->simulationWork.useGpuHaloExchange))
{
- // in the DD case we use the same stream for H2D and reduction, hence no explicit dependency needed
- if (!runScheduleWork->simulationWork.havePpDomainDecomposition)
- {
- const bool useGpuForceBufferOps = true;
- fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
- stateGpu->getForcesReadyOnDeviceEvent(AtomLocality::All, useGpuForceBufferOps));
- }
+ fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
+ stateGpu->fReadyOnDevice(AtomLocality::Local));
}
if (runScheduleWork->simulationWork.useGpuHaloExchange)
const bool accumulate = runScheduleWork->domainWork.haveCpuBondedWork
|| runScheduleWork->domainWork.haveFreeEnergyWork;
const int atomStart = dd_numHomeAtoms(*cr->dd);
- fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->reinit(stateGpu->getForces(),
- nbv->getNumAtoms(AtomLocality::NonLocal),
- nbv->getGridIndices(),
- atomStart,
- accumulate);
+ fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->reinit(
+ stateGpu->getForces(),
+ nbv->getNumAtoms(AtomLocality::NonLocal),
+ nbv->getGridIndices(),
+ atomStart,
+ accumulate,
+ stateGpu->fReducedOnDevice(AtomLocality::NonLocal));
// register forces and add dependencies
- // in the DD case we use the same stream for H2D and reduction, hence no explicit dependency needed
fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->registerNbnxmForce(
Nbnxm::gpu_get_f(nbv->gpu_nbv));
+
+ if (runScheduleWork->domainWork.haveNonLocalForceContribInCpuBuffer)
+ {
+ fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->addDependency(
+ stateGpu->fReadyOnDevice(AtomLocality::NonLocal));
+ }
}
}
runScheduleWork->stepWork = setupStepWorkload(legacyFlags, inputrec.mtsLevels, step, simulationWork);
const StepWorkload& stepWork = runScheduleWork->stepWork;
+ if (stepWork.useGpuFHalo && !runScheduleWork->domainWork.haveCpuLocalForceWork)
+ {
+ // GPU Force halo exchange will set a subset of local atoms with remote non-local data
+ // First clear local portion of force array, so that untouched atoms are zero.
+ // The dependency for this is that forces from previous timestep have been consumed,
+ // which is satisfied when getCoordinatesReadyOnDeviceEvent has been marked.
+ stateGpu->clearForcesOnGpu(AtomLocality::Local,
+ stateGpu->getCoordinatesReadyOnDeviceEvent(
+ AtomLocality::Local, simulationWork, stepWork));
+ }
+
/* At a search step we need to start the first balancing region
* somewhere early inside the step after communication during domain
* decomposition (and not during the previous step as usual).
}
else
{
+ GpuEventSynchronizer* gpuCoordinateHaloLaunched = nullptr;
if (stepWork.useGpuXHalo)
{
// The following must be called after local setCoordinates (which records an event
// when the coordinate data has been copied to the device).
- communicateGpuHaloCoordinates(*cr, box, localXReadyOnDevice);
+ gpuCoordinateHaloLaunched = communicateGpuHaloCoordinates(*cr, box, localXReadyOnDevice);
if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
{
// non-local part of coordinate buffer must be copied back to host for CPU work
- stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
+ stateGpu->copyCoordinatesFromGpu(
+ x.unpaddedArrayRef(), AtomLocality::NonLocal, gpuCoordinateHaloLaunched);
}
}
else
{
stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
}
- nbv->convertCoordinatesGpu(AtomLocality::NonLocal,
- stateGpu->getCoordinates(),
- stateGpu->getCoordinatesReadyOnDeviceEvent(
- AtomLocality::NonLocal, simulationWork, stepWork));
+ nbv->convertCoordinatesGpu(
+ AtomLocality::NonLocal,
+ stateGpu->getCoordinates(),
+ stateGpu->getCoordinatesReadyOnDeviceEvent(
+ AtomLocality::NonLocal, simulationWork, stepWork, gpuCoordinateHaloLaunched));
}
else
{
{
// If there exist CPU forces, data from halo exchange should accumulate into these
bool accumulateForces = domainWork.haveCpuLocalForceWork;
- if (!accumulateForces)
- {
- // Force halo exchange will set a subset of local atoms with remote non-local data
- // First clear local portion of force array, so that untouched atoms are zero
- stateGpu->clearForcesOnGpu(AtomLocality::Local);
- }
- communicateGpuHaloForces(*cr, accumulateForces);
+ gmx::FixedCapacityVector<GpuEventSynchronizer*, 2> gpuForceHaloDependencies;
+ gpuForceHaloDependencies.push_back(stateGpu->fReadyOnDevice(AtomLocality::Local));
+ gpuForceHaloDependencies.push_back(stateGpu->fReducedOnDevice(AtomLocality::NonLocal));
+
+ communicateGpuHaloForces(*cr, accumulateForces, &gpuForceHaloDependencies);
}
else
{
// These should be unified.
if (domainWork.haveLocalForceContribInCpuBuffer && !stepWork.useGpuFHalo)
{
- // Note: AtomLocality::All is used for the non-DD case because, as in this
- // case copyForcesToGpu() uses a separate stream, it allows overlap of
- // CPU force H2D with GPU force tasks on all streams including those in the
- // local stream which would otherwise be implicit dependencies for the
- // transfer and would not overlap.
- auto locality = simulationWork.havePpDomainDecomposition ? AtomLocality::Local
- : AtomLocality::All;
-
- stateGpu->copyForcesToGpu(forceWithShift, locality);
+ stateGpu->copyForcesToGpu(forceWithShift, AtomLocality::Local);
}
if (stepWork.computeNonbondedForces)