pmeStream_ = &deviceStreamManager.stream(DeviceStreamType::Pme);
localStream_ = &deviceStreamManager.stream(DeviceStreamType::NonBondedLocal);
nonLocalStream_ = &deviceStreamManager.stream(DeviceStreamType::NonBondedNonLocal);
- // PME stream is used in OpenCL for H2D coordinate transfer
- updateStream_ = &deviceStreamManager.stream(
- GMX_GPU_OPENCL ? DeviceStreamType::Pme : DeviceStreamType::UpdateAndConstraints);
+ updateStream_ = &deviceStreamManager.stream(DeviceStreamType::UpdateAndConstraints);
// Map the atom locality to the stream that will be used for coordinates,
// velocities and forces transfers. Same streams are used for H2D and D2H copies.
fCopyStreams_[AtomLocality::Local] = localStream_;
fCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
fCopyStreams_[AtomLocality::All] = updateStream_;
+
+ copyInStream_ = std::make_unique<DeviceStream>(deviceContext_, DeviceStreamPriority::Normal, false);
+ memsetStream_ = std::make_unique<DeviceStream>(deviceContext_, DeviceStreamPriority::Normal, false);
}
StatePropagatorDataGpu::Impl::Impl(const DeviceStream* pmeStream,
nonLocalStream_ = nullptr;
updateStream_ = nullptr;
+ isPmeOnly_ = true;
// Only local/all coordinates are allowed to be copied in PME-only rank/ PME tests.
// This it temporary measure to make it safe to use this class in those cases.
reallocateDeviceBuffer(&d_f_, numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
// Clearing of the forces can be done in local stream since the nonlocal stream cannot reach
- // the force accumulation stage before syncing with the local stream. Only done in CUDA,
- // since the force buffer ops are not implemented in OpenCL.
- if (GMX_GPU_CUDA && d_fCapacity_ != d_fOldCapacity)
+ // the force accumulation stage before syncing with the local stream. Only done in CUDA and
+ // SYCL, since the force buffer ops are not implemented in OpenCL.
+ static constexpr bool sc_haveGpuFBufferOps = ((GMX_GPU_CUDA != 0) || (GMX_GPU_SYCL != 0));
+ if (sc_haveGpuFBufferOps && d_fCapacity_ != d_fOldCapacity)
{
clearDeviceBufferAsync(&d_f_, 0, d_fCapacity_, *localStream_);
}
copyToDevice(d_x_, h_x, d_xSize_, atomLocality, *deviceStream);
- // markEvent is skipped in OpenCL as:
- // - it's not needed, copy is done in the same stream as the only consumer task (PME)
- // - we don't consume the events in OpenCL which is not allowed by GpuEventSynchronizer (would leak memory).
- // TODO: remove this by adding an event-mark free flavor of this function
- if (GMX_GPU_CUDA)
+ // marking is skipped on the PME-rank mode as everything is on the same stream
+ if (!isPmeOnly_)
{
xReadyOnDevice_[atomLocality].markEvent(*deviceStream);
}
wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
}
-GpuEventSynchronizer*
-StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
- const SimulationWorkload& simulationWork,
- const StepWorkload& stepWork)
+GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getCoordinatesReadyOnDeviceEvent(
+ AtomLocality atomLocality,
+ const SimulationWorkload& simulationWork,
+ const StepWorkload& stepWork,
+ GpuEventSynchronizer* gpuCoordinateHaloLaunched)
{
// The provider of the coordinates may be different for local atoms. If the update is offloaded
// and this is not a neighbor search step, then the consumer needs to wait for the update
// to complete. Otherwise, the coordinates are copied from the host and we need to wait for
- // the copy event. Non-local coordinates are always provided by the H2D copy.
- //
- // TODO: This should be reconsidered to support the halo exchange.
- //
- // In OpenCL no events are used as coordinate sync is not necessary
- if (GMX_GPU_OPENCL)
+ // the copy event. Non-local coordinates are provided by the GPU halo exchange (if active), otherwise by H2D copy.
+
+ if (atomLocality == AtomLocality::NonLocal && stepWork.useGpuXHalo)
{
- return nullptr;
+ GMX_ASSERT(gpuCoordinateHaloLaunched != nullptr,
+ "GPU halo exchange is active but its completion event is null.");
+ return gpuCoordinateHaloLaunched;
}
if (atomLocality == AtomLocality::Local && simulationWork.useGpuUpdate && !stepWork.doNeighborSearch)
{
}
else
{
+ if (stepWork.doNeighborSearch && xUpdatedOnDeviceEvent_)
+ {
+ /* On search steps, we do not consume the result of the GPU update
+ * but rather that of a H2D transfer. So, we reset the event triggered after
+ * update to avoid leaving it unconsumed.
+ * Unfortunately, we don't always have the event marked either (e.g., on the
+ * first step) so we just reset it here.
+ * See Issue #3988. */
+ xUpdatedOnDeviceEvent_->reset();
+ }
return &xReadyOnDevice_[atomLocality];
}
}
wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
}
+void StatePropagatorDataGpu::Impl::consumeCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
+{
+ GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
+ xReadyOnDevice_[atomLocality].consume();
+}
+
+void StatePropagatorDataGpu::Impl::resetCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
+{
+ GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
+ xReadyOnDevice_[atomLocality].reset();
+}
+
void StatePropagatorDataGpu::Impl::setXUpdatedOnDeviceEvent(GpuEventSynchronizer* xUpdatedOnDeviceEvent)
{
GMX_ASSERT(xUpdatedOnDeviceEvent != nullptr, "The event synchronizer can not be nullptr.");
xUpdatedOnDeviceEvent_ = xUpdatedOnDeviceEvent;
}
-void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x, AtomLocality atomLocality)
+void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x,
+ AtomLocality atomLocality,
+ GpuEventSynchronizer* dependency)
{
GMX_ASSERT(atomLocality < AtomLocality::All,
formatString("Wrong atom locality. Only Local and NonLocal are allowed for "
GMX_ASSERT(deviceStream != nullptr,
"No stream is valid for copying positions with given atom locality.");
+ if (dependency != nullptr)
+ {
+ dependency->enqueueWaitEvent(*deviceStream);
+ }
+
wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
return d_f_;
}
+// Copy CPU forces to GPU using stream internal to this module to allow overlap
+// with GPU force calculations.
void StatePropagatorDataGpu::Impl::copyForcesToGpu(const gmx::ArrayRef<const gmx::RVec> h_f,
AtomLocality atomLocality)
{
GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
- const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
+ DeviceStream* deviceStream = copyInStream_.get();
GMX_ASSERT(deviceStream != nullptr,
"No stream is valid for copying forces with given atom locality.");
wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
}
-void StatePropagatorDataGpu::Impl::clearForcesOnGpu(AtomLocality atomLocality)
+void StatePropagatorDataGpu::Impl::clearForcesOnGpu(AtomLocality atomLocality, GpuEventSynchronizer* dependency)
{
GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
- const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
+ DeviceStream* deviceStream = memsetStream_.get();
+
+ GMX_ASSERT(dependency != nullptr, "Dependency is not valid for clearing forces.");
+ dependency->enqueueWaitEvent(*deviceStream);
+
GMX_ASSERT(deviceStream != nullptr,
"No stream is valid for clearing forces with given atom locality.");
clearOnDevice(d_f_, d_fSize_, atomLocality, *deviceStream);
+ fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
+
wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
}
-GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
- bool useGpuFBufferOps)
+GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getLocalForcesReadyOnDeviceEvent(StepWorkload stepWork,
+ SimulationWorkload simulationWork)
{
- if ((atomLocality == AtomLocality::Local || atomLocality == AtomLocality::NonLocal) && useGpuFBufferOps)
+ if (stepWork.useGpuFBufferOps && !simulationWork.useCpuPmePpCommunication)
{
- return &fReducedOnDevice_;
+ return &fReducedOnDevice_[AtomLocality::Local];
}
else
{
- return &fReadyOnDevice_[atomLocality];
+ return &fReadyOnDevice_[AtomLocality::Local];
}
}
-GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice()
+GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReducedOnDevice(AtomLocality atomLocality)
+{
+ return &fReducedOnDevice_[atomLocality];
+}
+
+void StatePropagatorDataGpu::Impl::consumeForcesReducedOnDeviceEvent(AtomLocality atomLocality)
+{
+ fReducedOnDevice_[atomLocality].consume();
+}
+
+GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReadyOnDevice(AtomLocality atomLocality)
{
- return &fReducedOnDevice_;
+ return &fReadyOnDevice_[atomLocality];
}
void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality)
GpuEventSynchronizer*
StatePropagatorDataGpu::getCoordinatesReadyOnDeviceEvent(AtomLocality atomLocality,
const SimulationWorkload& simulationWork,
- const StepWorkload& stepWork)
+ const StepWorkload& stepWork,
+ GpuEventSynchronizer* gpuCoordinateHaloLaunched)
{
- return impl_->getCoordinatesReadyOnDeviceEvent(atomLocality, simulationWork, stepWork);
+ return impl_->getCoordinatesReadyOnDeviceEvent(
+ atomLocality, simulationWork, stepWork, gpuCoordinateHaloLaunched);
}
void StatePropagatorDataGpu::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
return impl_->waitCoordinatesCopiedToDevice(atomLocality);
}
+void StatePropagatorDataGpu::consumeCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
+{
+ return impl_->consumeCoordinatesCopiedToDeviceEvent(atomLocality);
+}
+
+void StatePropagatorDataGpu::resetCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
+{
+ return impl_->resetCoordinatesCopiedToDeviceEvent(atomLocality);
+}
+
void StatePropagatorDataGpu::setXUpdatedOnDeviceEvent(GpuEventSynchronizer* xUpdatedOnDeviceEvent)
{
impl_->setXUpdatedOnDeviceEvent(xUpdatedOnDeviceEvent);
}
-void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec> h_x, AtomLocality atomLocality)
+void StatePropagatorDataGpu::copyCoordinatesFromGpu(gmx::ArrayRef<RVec> h_x,
+ AtomLocality atomLocality,
+ GpuEventSynchronizer* dependency)
{
- return impl_->copyCoordinatesFromGpu(h_x, atomLocality);
+ return impl_->copyCoordinatesFromGpu(h_x, atomLocality, dependency);
}
void StatePropagatorDataGpu::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
return impl_->copyForcesToGpu(h_f, atomLocality);
}
-void StatePropagatorDataGpu::clearForcesOnGpu(AtomLocality atomLocality)
+void StatePropagatorDataGpu::clearForcesOnGpu(AtomLocality atomLocality, GpuEventSynchronizer* dependency)
+{
+ return impl_->clearForcesOnGpu(atomLocality, dependency);
+}
+
+GpuEventSynchronizer* StatePropagatorDataGpu::getLocalForcesReadyOnDeviceEvent(StepWorkload stepWork,
+ SimulationWorkload simulationWork)
+{
+ return impl_->getLocalForcesReadyOnDeviceEvent(stepWork, simulationWork);
+}
+
+GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice(AtomLocality atomLocality)
{
- return impl_->clearForcesOnGpu(atomLocality);
+ return impl_->fReducedOnDevice(atomLocality);
}
-GpuEventSynchronizer* StatePropagatorDataGpu::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
- bool useGpuFBufferOps)
+void StatePropagatorDataGpu::consumeForcesReducedOnDeviceEvent(AtomLocality atomLocality)
{
- return impl_->getForcesReadyOnDeviceEvent(atomLocality, useGpuFBufferOps);
+ impl_->consumeForcesReducedOnDeviceEvent(atomLocality);
}
-GpuEventSynchronizer* StatePropagatorDataGpu::fReducedOnDevice()
+GpuEventSynchronizer* StatePropagatorDataGpu::fReadyOnDevice(AtomLocality atomLocality)
{
- return impl_->fReducedOnDevice();
+ return impl_->fReadyOnDevice(atomLocality);
}
void StatePropagatorDataGpu::copyForcesFromGpu(gmx::ArrayRef<RVec> h_f, AtomLocality atomLocality)