/*
* This file is part of the GROMACS molecular simulation package.
*
- * Copyright (c) 2019, by the GROMACS development team, led by
+ * Copyright (c) 2019,2020,2021, by the GROMACS development team, led by
* Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
* and including many others, as listed in the AUTHORS file in the
* top-level source directory and at http://www.gromacs.org.
#include "config.h"
-#if GMX_GPU != GMX_GPU_NONE
+#if GMX_GPU
-# if GMX_GPU == GMX_GPU_CUDA
-# include "gromacs/gpu_utils/cudautils.cuh"
-# endif
+# include "gromacs/gpu_utils/device_stream_manager.h"
# include "gromacs/gpu_utils/devicebuffer.h"
-# if GMX_GPU == GMX_GPU_OPENCL
-# include "gromacs/gpu_utils/oclutils.h"
-# endif
+# include "gromacs/gpu_utils/gpueventsynchronizer.h"
# include "gromacs/math/vectypes.h"
# include "gromacs/mdtypes/state_propagator_data_gpu.h"
# include "gromacs/timing/wallcycle.h"
namespace gmx
{
-StatePropagatorDataGpu::Impl::Impl(const void* pmeStream,
- const void* localStream,
- const void* nonLocalStream,
- const void* deviceContext,
- GpuApiCallBehavior transferKind,
- int paddingSize,
- gmx_wallcycle* wcycle) :
+StatePropagatorDataGpu::Impl::Impl(const DeviceStreamManager& deviceStreamManager,
+ GpuApiCallBehavior transferKind,
+ int allocationBlockSizeDivisor,
+ gmx_wallcycle* wcycle) :
+ deviceContext_(deviceStreamManager.context()),
transferKind_(transferKind),
- paddingSize_(paddingSize),
+ allocationBlockSizeDivisor_(allocationBlockSizeDivisor),
wcycle_(wcycle)
{
- static_assert(GMX_GPU != GMX_GPU_NONE,
- "This object should only be constructed on the GPU code-paths.");
+ static_assert(
+ GMX_GPU,
+ "GPU state propagator data object should only be constructed on the GPU code-paths.");
- // TODO: Refactor when the StreamManager is introduced.
- if (GMX_GPU == GMX_GPU_OPENCL)
- {
- GMX_ASSERT(deviceContext != nullptr, "GPU context should be set in OpenCL builds.");
- GMX_ASSERT(pmeStream != nullptr, "GPU PME stream should be set in OpenCL builds.");
-
- // The update stream is set to the PME stream in OpenCL, since PME stream is the only stream created in the PME context.
- pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
- updateStream_ = *static_cast<const CommandStream*>(pmeStream);
- deviceContext_ = *static_cast<const DeviceContext*>(deviceContext);
- GMX_UNUSED_VALUE(localStream);
- GMX_UNUSED_VALUE(nonLocalStream);
- }
-
- if (GMX_GPU == GMX_GPU_CUDA)
- {
- if (pmeStream != nullptr)
- {
- pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
- }
- if (localStream != nullptr)
- {
- localStream_ = *static_cast<const CommandStream*>(localStream);
- }
- if (nonLocalStream != nullptr)
- {
- nonLocalStream_ = *static_cast<const CommandStream*>(nonLocalStream);
- }
-
- // TODO: The update stream should be created only when it is needed.
-# if (GMX_GPU == GMX_GPU_CUDA)
- cudaError_t stat;
- stat = cudaStreamCreate(&updateStream_);
- CU_RET_ERR(stat, "CUDA stream creation failed in StatePropagatorDataGpu");
-# endif
- GMX_UNUSED_VALUE(deviceContext);
- }
+ // We need to keep local copies for re-initialization.
+ 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);
// 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.
// Note, that nullptr stream is used here to indicate that the copy is not supported.
xCopyStreams_[AtomLocality::Local] = updateStream_;
xCopyStreams_[AtomLocality::NonLocal] = nonLocalStream_;
- xCopyStreams_[AtomLocality::All] = updateStream_;
+ xCopyStreams_[AtomLocality::All] = nullptr;
vCopyStreams_[AtomLocality::Local] = updateStream_;
vCopyStreams_[AtomLocality::NonLocal] = nullptr;
- vCopyStreams_[AtomLocality::All] = updateStream_;
+ vCopyStreams_[AtomLocality::All] = nullptr;
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 void* pmeStream,
- const void* deviceContext,
- GpuApiCallBehavior transferKind,
- int paddingSize,
- gmx_wallcycle* wcycle) :
+StatePropagatorDataGpu::Impl::Impl(const DeviceStream* pmeStream,
+ const DeviceContext& deviceContext,
+ GpuApiCallBehavior transferKind,
+ int allocationBlockSizeDivisor,
+ gmx_wallcycle* wcycle) :
+ deviceContext_(deviceContext),
transferKind_(transferKind),
- paddingSize_(paddingSize),
+ allocationBlockSizeDivisor_(allocationBlockSizeDivisor),
wcycle_(wcycle)
{
- static_assert(GMX_GPU != GMX_GPU_NONE,
- "This object should only be constructed on the GPU code-paths.");
-
- if (GMX_GPU == GMX_GPU_OPENCL)
- {
- GMX_ASSERT(deviceContext != nullptr, "GPU context should be set in OpenCL builds.");
- deviceContext_ = *static_cast<const DeviceContext*>(deviceContext);
- }
-
- GMX_ASSERT(pmeStream != nullptr, "GPU PME stream should be set.");
- pmeStream_ = *static_cast<const CommandStream*>(pmeStream);
+ static_assert(
+ GMX_GPU,
+ "GPU state propagator data object should only be constructed on the GPU code-paths.");
- localStream_ = nullptr;
+ GMX_ASSERT(pmeStream->isValid(), "GPU PME stream should be valid.");
+ pmeStream_ = pmeStream;
+ localStream_ = pmeStream; // For clearing the force buffer
nonLocalStream_ = nullptr;
updateStream_ = nullptr;
// This it temporary measure to make it safe to use this class in those cases.
xCopyStreams_[AtomLocality::Local] = pmeStream_;
xCopyStreams_[AtomLocality::NonLocal] = nullptr;
- xCopyStreams_[AtomLocality::All] = pmeStream_;
+ xCopyStreams_[AtomLocality::All] = nullptr;
vCopyStreams_[AtomLocality::Local] = nullptr;
vCopyStreams_[AtomLocality::NonLocal] = nullptr;
void StatePropagatorDataGpu::Impl::reinit(int numAtomsLocal, int numAtomsAll)
{
- wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
- wallcycle_sub_start_nocount(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
+ wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
+ wallcycle_sub_start_nocount(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
numAtomsLocal_ = numAtomsLocal;
numAtomsAll_ = numAtomsAll;
int numAtomsPadded;
- if (paddingSize_ > 0)
+ if (allocationBlockSizeDivisor_ > 0)
{
- numAtomsPadded = ((numAtomsAll_ + paddingSize_ - 1) / paddingSize_) * paddingSize_;
+ numAtomsPadded = ((numAtomsAll_ + allocationBlockSizeDivisor_ - 1) / allocationBlockSizeDivisor_)
+ * allocationBlockSizeDivisor_;
}
else
{
numAtomsPadded = numAtomsAll_;
}
- reallocateDeviceBuffer(&d_x_, DIM * numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
+ reallocateDeviceBuffer(&d_x_, numAtomsPadded, &d_xSize_, &d_xCapacity_, deviceContext_);
const size_t paddingAllocationSize = numAtomsPadded - numAtomsAll_;
if (paddingAllocationSize > 0)
{
// The PME stream is used here because the padding region of d_x_ is only in the PME task.
- clearDeviceBufferAsync(&d_x_, DIM * numAtomsAll_, DIM * paddingAllocationSize, pmeStream_);
+ clearDeviceBufferAsync(&d_x_, numAtomsAll_, paddingAllocationSize, *pmeStream_);
}
- reallocateDeviceBuffer(&d_v_, DIM * numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
+ reallocateDeviceBuffer(&d_v_, numAtomsAll_, &d_vSize_, &d_vCapacity_, deviceContext_);
const int d_fOldCapacity = d_fCapacity_;
- reallocateDeviceBuffer(&d_f_, DIM * numAtomsAll_, &d_fSize_, &d_fCapacity_, deviceContext_);
+ 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 == 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.
+ if ((GMX_GPU_CUDA || GMX_GPU_SYCL) && d_fCapacity_ != d_fOldCapacity)
{
- clearDeviceBufferAsync(&d_f_, 0, d_fCapacity_, localStream_);
+ clearDeviceBufferAsync(&d_f_, 0, d_fCapacity_, *localStream_);
}
- wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
- wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
+ wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
+ wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
}
-std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
+std::tuple<int, int> StatePropagatorDataGpu::Impl::getAtomRangesFromAtomLocality(AtomLocality atomLocality) const
{
int atomsStartAt = 0;
int numAtomsToCopy = 0;
return std::make_tuple(atomsStartAt, numAtomsToCopy);
}
-void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<float> d_data,
+void StatePropagatorDataGpu::Impl::copyToDevice(DeviceBuffer<RVec> d_data,
const gmx::ArrayRef<const gmx::RVec> h_data,
int dataSize,
AtomLocality atomLocality,
- CommandStream commandStream)
+ const DeviceStream& deviceStream)
{
GMX_UNUSED_VALUE(dataSize);
+ GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
+
GMX_ASSERT(dataSize >= 0, "Trying to copy to device buffer before it was allocated.");
- wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
- wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
+ GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
int atomsStartAt, numAtomsToCopy;
std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
- int elementsStartAt = atomsStartAt * DIM;
- int numElementsToCopy = numAtomsToCopy * DIM;
-
if (numAtomsToCopy != 0)
{
- GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize,
+ GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
"The device allocation is smaller than requested copy range.");
GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
"The host buffer is smaller than the requested copy range.");
- copyToDeviceBuffer(&d_data, reinterpret_cast<const float*>(&h_data.data()[atomsStartAt]),
- elementsStartAt, numElementsToCopy, commandStream, transferKind_, nullptr);
+ copyToDeviceBuffer(&d_data,
+ reinterpret_cast<const RVec*>(&h_data.data()[atomsStartAt]),
+ atomsStartAt,
+ numAtomsToCopy,
+ deviceStream,
+ transferKind_,
+ nullptr);
}
-
- wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
- wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
}
void StatePropagatorDataGpu::Impl::copyFromDevice(gmx::ArrayRef<gmx::RVec> h_data,
- DeviceBuffer<float> d_data,
+ DeviceBuffer<RVec> d_data,
int dataSize,
AtomLocality atomLocality,
- CommandStream commandStream)
+ const DeviceStream& deviceStream)
{
GMX_UNUSED_VALUE(dataSize);
+ GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
+
GMX_ASSERT(dataSize >= 0, "Trying to copy from device buffer before it was allocated.");
- wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
- wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
+ GMX_ASSERT(deviceStream.isValid(), "No stream is valid for copying with given atom locality.");
int atomsStartAt, numAtomsToCopy;
std::tie(atomsStartAt, numAtomsToCopy) = getAtomRangesFromAtomLocality(atomLocality);
- int elementsStartAt = atomsStartAt * DIM;
- int numElementsToCopy = numAtomsToCopy * DIM;
-
if (numAtomsToCopy != 0)
{
- GMX_ASSERT(elementsStartAt + numElementsToCopy <= dataSize,
+ GMX_ASSERT(atomsStartAt + numAtomsToCopy <= dataSize,
"The device allocation is smaller than requested copy range.");
GMX_ASSERT(atomsStartAt + numAtomsToCopy <= h_data.ssize(),
"The host buffer is smaller than the requested copy range.");
- copyFromDeviceBuffer(reinterpret_cast<float*>(&h_data.data()[atomsStartAt]), &d_data,
- elementsStartAt, numElementsToCopy, commandStream, transferKind_, nullptr);
+ copyFromDeviceBuffer(reinterpret_cast<RVec*>(&h_data.data()[atomsStartAt]),
+ &d_data,
+ atomsStartAt,
+ numAtomsToCopy,
+ deviceStream,
+ transferKind_,
+ nullptr);
}
+}
+
+void StatePropagatorDataGpu::Impl::clearOnDevice(DeviceBuffer<RVec> d_data,
+ int dataSize,
+ AtomLocality atomLocality,
+ const DeviceStream& deviceStream) const
+{
+ GMX_UNUSED_VALUE(dataSize);
+
+ GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
+
+ GMX_ASSERT(dataSize >= 0, "Trying to clear to device buffer before it was allocated.");
+
+ GMX_ASSERT(deviceStream.isValid(), "No stream is valid for clearing with given atom locality.");
+
+ int atomsStartAt, numAtomsToClear;
+ std::tie(atomsStartAt, numAtomsToClear) = getAtomRangesFromAtomLocality(atomLocality);
+
+ if (numAtomsToClear != 0)
+ {
+ GMX_ASSERT(atomsStartAt + numAtomsToClear <= dataSize,
+ "The device allocation is smaller than requested clear range.");
- wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
- wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
+ clearDeviceBufferAsync(&d_data, atomsStartAt, numAtomsToClear, deviceStream);
+ }
}
-DeviceBuffer<float> StatePropagatorDataGpu::Impl::getCoordinates()
+DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getCoordinates()
{
return d_x_;
}
void StatePropagatorDataGpu::Impl::copyCoordinatesToGpu(const gmx::ArrayRef<const gmx::RVec> h_x,
AtomLocality atomLocality)
{
- GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
- CommandStream commandStream = xCopyStreams_[atomLocality];
- GMX_ASSERT(commandStream != nullptr,
+ GMX_ASSERT(atomLocality < AtomLocality::All,
+ formatString("Wrong atom locality. Only Local and NonLocal are allowed for "
+ "coordinate transfers, passed value is \"%s\"",
+ enumValueToString(atomLocality))
+ .c_str());
+
+ const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
+ GMX_ASSERT(deviceStream != nullptr,
"No stream is valid for copying positions with given atom locality.");
- wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
- wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
+ wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
+ wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
- copyToDevice(d_x_, h_x, d_xSize_, atomLocality, commandStream);
+ 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 == GMX_GPU_CUDA)
+ if (GMX_GPU_CUDA || GMX_GPU_SYCL)
{
- xReadyOnDevice_[atomLocality].markEvent(commandStream);
+ xReadyOnDevice_[atomLocality].markEvent(*deviceStream);
}
- wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
- wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
+ wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
+ 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.
+ // the copy event. Non-local coordinates are provided by the GPU halo exchange (if active), otherwise by H2D copy.
//
// In OpenCL no events are used as coordinate sync is not necessary
- if (GMX_GPU == GMX_GPU_OPENCL)
+ if (GMX_GPU_OPENCL)
{
return nullptr;
}
+ if (atomLocality == AtomLocality::NonLocal && stepWork.useGpuXHalo)
+ {
+ 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)
{
- return &xUpdatedOnDevice_;
+ GMX_ASSERT(xUpdatedOnDeviceEvent_ != nullptr, "The event synchronizer can not be nullptr.");
+ return xUpdatedOnDeviceEvent_;
}
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];
}
}
void StatePropagatorDataGpu::Impl::waitCoordinatesCopiedToDevice(AtomLocality atomLocality)
{
- wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
+ wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
xReadyOnDevice_[atomLocality].waitForEvent();
- wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
+ wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
}
-GpuEventSynchronizer* StatePropagatorDataGpu::Impl::xUpdatedOnDevice()
+void StatePropagatorDataGpu::Impl::consumeCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
{
- return &xUpdatedOnDevice_;
+ GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
+ xReadyOnDevice_[atomLocality].consume();
}
-void StatePropagatorDataGpu::Impl::copyCoordinatesFromGpu(gmx::ArrayRef<gmx::RVec> h_x, AtomLocality atomLocality)
+void StatePropagatorDataGpu::Impl::resetCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
{
GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
- CommandStream commandStream = xCopyStreams_[atomLocality];
- GMX_ASSERT(commandStream != nullptr,
+ 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,
+ GpuEventSynchronizer* dependency)
+{
+ GMX_ASSERT(atomLocality < AtomLocality::All,
+ formatString("Wrong atom locality. Only Local and NonLocal are allowed for "
+ "coordinate transfers, passed value is \"%s\"",
+ enumValueToString(atomLocality))
+ .c_str());
+ const DeviceStream* deviceStream = xCopyStreams_[atomLocality];
+ GMX_ASSERT(deviceStream != nullptr,
"No stream is valid for copying positions with given atom locality.");
- wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
- wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
+ if (dependency != nullptr)
+ {
+ dependency->enqueueWaitEvent(*deviceStream);
+ }
+
+ wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
+ wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
- copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, commandStream);
+ copyFromDevice(h_x, d_x_, d_xSize_, atomLocality, *deviceStream);
// Note: unlike copyCoordinatesToGpu this is not used in OpenCL, and the conditional is not needed.
- xReadyOnHost_[atomLocality].markEvent(commandStream);
+ xReadyOnHost_[atomLocality].markEvent(*deviceStream);
- wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
- wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
+ wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
+ wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
}
void StatePropagatorDataGpu::Impl::waitCoordinatesReadyOnHost(AtomLocality atomLocality)
{
- wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
+ wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
xReadyOnHost_[atomLocality].waitForEvent();
- wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
+ wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
}
-DeviceBuffer<float> StatePropagatorDataGpu::Impl::getVelocities()
+DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getVelocities()
{
return d_v_;
}
void StatePropagatorDataGpu::Impl::copyVelocitiesToGpu(const gmx::ArrayRef<const gmx::RVec> h_v,
AtomLocality atomLocality)
{
- GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
- CommandStream commandStream = vCopyStreams_[atomLocality];
- GMX_ASSERT(commandStream != nullptr,
+ GMX_ASSERT(atomLocality == AtomLocality::Local,
+ formatString("Wrong atom locality. Only Local is allowed for "
+ "velocity transfers, passed value is \"%s\"",
+ enumValueToString(atomLocality))
+ .c_str());
+ const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
+ GMX_ASSERT(deviceStream != nullptr,
"No stream is valid for copying velocities with given atom locality.");
- wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
- wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
+ wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
+ wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
- copyToDevice(d_v_, h_v, d_vSize_, atomLocality, commandStream);
- vReadyOnDevice_[atomLocality].markEvent(commandStream);
+ copyToDevice(d_v_, h_v, d_vSize_, atomLocality, *deviceStream);
+ /* Not marking the event, because it is not used anywhere.
+ * Since we only use velocities on the device for update, and we launch the copy in
+ * the "update" stream, that should be safe.
+ */
- wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
- wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
-}
-
-GpuEventSynchronizer* StatePropagatorDataGpu::Impl::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
-{
- return &vReadyOnDevice_[atomLocality];
+ wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
+ wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
}
-
void StatePropagatorDataGpu::Impl::copyVelocitiesFromGpu(gmx::ArrayRef<gmx::RVec> h_v, AtomLocality atomLocality)
{
- GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
- CommandStream commandStream = vCopyStreams_[atomLocality];
- GMX_ASSERT(commandStream != nullptr,
+ GMX_ASSERT(atomLocality == AtomLocality::Local,
+ formatString("Wrong atom locality. Only Local is allowed for "
+ "velocity transfers, passed value is \"%s\"",
+ enumValueToString(atomLocality))
+ .c_str());
+ const DeviceStream* deviceStream = vCopyStreams_[atomLocality];
+ GMX_ASSERT(deviceStream != nullptr,
"No stream is valid for copying velocities with given atom locality.");
- wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
- wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
+ wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
+ wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
- copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, commandStream);
- vReadyOnHost_[atomLocality].markEvent(commandStream);
+ copyFromDevice(h_v, d_v_, d_vSize_, atomLocality, *deviceStream);
+ vReadyOnHost_[atomLocality].markEvent(*deviceStream);
- wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
- wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
+ wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
+ wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
}
void StatePropagatorDataGpu::Impl::waitVelocitiesReadyOnHost(AtomLocality atomLocality)
{
- wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
+ wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
vReadyOnHost_[atomLocality].waitForEvent();
- wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
+ wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
}
-DeviceBuffer<float> StatePropagatorDataGpu::Impl::getForces()
+DeviceBuffer<RVec> StatePropagatorDataGpu::Impl::getForces()
{
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.");
- CommandStream commandStream = fCopyStreams_[atomLocality];
- GMX_ASSERT(commandStream != nullptr,
+ DeviceStream* deviceStream = copyInStream_.get();
+ GMX_ASSERT(deviceStream != nullptr,
"No stream is valid for copying forces with given atom locality.");
- wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
- wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
+ wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
+ wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
+
+ copyToDevice(d_f_, h_f, d_fSize_, atomLocality, *deviceStream);
+ fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
+
+ wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
+ wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
+}
+
+void StatePropagatorDataGpu::Impl::clearForcesOnGpu(AtomLocality atomLocality, GpuEventSynchronizer* dependency)
+{
+ GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
+ 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.");
+
+ wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
+ wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
+
+ clearOnDevice(d_f_, d_fSize_, atomLocality, *deviceStream);
- copyToDevice(d_f_, h_f, d_fSize_, atomLocality, commandStream);
- fReadyOnDevice_[atomLocality].markEvent(commandStream);
+ fReadyOnDevice_[atomLocality].markEvent(*deviceStream);
- wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
- wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
+ 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_;
+ return &fReducedOnDevice_[atomLocality];
+}
+
+void StatePropagatorDataGpu::Impl::consumeForcesReducedOnDeviceEvent(AtomLocality atomLocality)
+{
+ fReducedOnDevice_[atomLocality].consume();
+}
+
+GpuEventSynchronizer* StatePropagatorDataGpu::Impl::fReadyOnDevice(AtomLocality atomLocality)
+{
+ return &fReadyOnDevice_[atomLocality];
}
void StatePropagatorDataGpu::Impl::copyForcesFromGpu(gmx::ArrayRef<gmx::RVec> h_f, AtomLocality atomLocality)
{
GMX_ASSERT(atomLocality < AtomLocality::Count, "Wrong atom locality.");
- CommandStream commandStream = fCopyStreams_[atomLocality];
- GMX_ASSERT(commandStream != nullptr,
+ const DeviceStream* deviceStream = fCopyStreams_[atomLocality];
+ GMX_ASSERT(deviceStream != nullptr,
"No stream is valid for copying forces with given atom locality.");
- wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
- wallcycle_sub_start(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
+ wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
+ wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
- copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, commandStream);
- fReadyOnHost_[atomLocality].markEvent(commandStream);
+ copyFromDevice(h_f, d_f_, d_fSize_, atomLocality, *deviceStream);
+ fReadyOnHost_[atomLocality].markEvent(*deviceStream);
- wallcycle_sub_stop(wcycle_, ewcsLAUNCH_STATE_PROPAGATOR_DATA);
- wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
+ wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchStatePropagatorData);
+ wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
}
void StatePropagatorDataGpu::Impl::waitForcesReadyOnHost(AtomLocality atomLocality)
{
- wallcycle_start(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
+ wallcycle_start(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
fReadyOnHost_[atomLocality].waitForEvent();
- wallcycle_stop(wcycle_, ewcWAIT_GPU_STATE_PROPAGATOR_DATA);
+ wallcycle_stop(wcycle_, WallCycleCounter::WaitGpuStatePropagatorData);
}
-void* StatePropagatorDataGpu::Impl::getUpdateStream()
+const DeviceStream* StatePropagatorDataGpu::Impl::getUpdateStream()
{
- return &updateStream_;
+ return updateStream_;
}
-int StatePropagatorDataGpu::Impl::numAtomsLocal()
+int StatePropagatorDataGpu::Impl::numAtomsLocal() const
{
return numAtomsLocal_;
}
-int StatePropagatorDataGpu::Impl::numAtomsAll()
+int StatePropagatorDataGpu::Impl::numAtomsAll() const
{
return numAtomsAll_;
}
-StatePropagatorDataGpu::StatePropagatorDataGpu(const void* pmeStream,
- const void* localStream,
- const void* nonLocalStream,
- const void* deviceContext,
- GpuApiCallBehavior transferKind,
- int paddingSize,
- gmx_wallcycle* wcycle) :
- impl_(new Impl(pmeStream, localStream, nonLocalStream, deviceContext, transferKind, paddingSize, wcycle))
+StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStreamManager& deviceStreamManager,
+ GpuApiCallBehavior transferKind,
+ int allocationBlockSizeDivisor,
+ gmx_wallcycle* wcycle) :
+ impl_(new Impl(deviceStreamManager, transferKind, allocationBlockSizeDivisor, wcycle))
{
}
-StatePropagatorDataGpu::StatePropagatorDataGpu(const void* pmeStream,
- const void* deviceContext,
- GpuApiCallBehavior transferKind,
- int paddingSize,
- gmx_wallcycle* wcycle) :
- impl_(new Impl(pmeStream, deviceContext, transferKind, paddingSize, wcycle))
+StatePropagatorDataGpu::StatePropagatorDataGpu(const DeviceStream* pmeStream,
+ const DeviceContext& deviceContext,
+ GpuApiCallBehavior transferKind,
+ int allocationBlockSizeDivisor,
+ gmx_wallcycle* wcycle) :
+ impl_(new Impl(pmeStream, deviceContext, transferKind, allocationBlockSizeDivisor, wcycle))
{
}
return impl_->reinit(numAtomsLocal, numAtomsAll);
}
-std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality)
+std::tuple<int, int> StatePropagatorDataGpu::getAtomRangesFromAtomLocality(AtomLocality atomLocality) const
{
return impl_->getAtomRangesFromAtomLocality(atomLocality);
}
-DeviceBuffer<float> StatePropagatorDataGpu::getCoordinates()
+DeviceBuffer<RVec> StatePropagatorDataGpu::getCoordinates()
{
return impl_->getCoordinates();
}
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);
}
-GpuEventSynchronizer* StatePropagatorDataGpu::xUpdatedOnDevice()
+void StatePropagatorDataGpu::consumeCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
+{
+ return impl_->consumeCoordinatesCopiedToDeviceEvent(atomLocality);
+}
+
+void StatePropagatorDataGpu::resetCoordinatesCopiedToDeviceEvent(AtomLocality atomLocality)
+{
+ return impl_->resetCoordinatesCopiedToDeviceEvent(atomLocality);
+}
+
+void StatePropagatorDataGpu::setXUpdatedOnDeviceEvent(GpuEventSynchronizer* xUpdatedOnDeviceEvent)
{
- return impl_->xUpdatedOnDevice();
+ 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)
}
-DeviceBuffer<float> StatePropagatorDataGpu::getVelocities()
+DeviceBuffer<RVec> StatePropagatorDataGpu::getVelocities()
{
return impl_->getVelocities();
}
return impl_->copyVelocitiesToGpu(h_v, atomLocality);
}
-GpuEventSynchronizer* StatePropagatorDataGpu::getVelocitiesReadyOnDeviceEvent(AtomLocality atomLocality)
-{
- return impl_->getVelocitiesReadyOnDeviceEvent(atomLocality);
-}
-
void StatePropagatorDataGpu::copyVelocitiesFromGpu(gmx::ArrayRef<RVec> h_v, AtomLocality atomLocality)
{
return impl_->copyVelocitiesFromGpu(h_v, atomLocality);
}
-DeviceBuffer<float> StatePropagatorDataGpu::getForces()
+DeviceBuffer<RVec> StatePropagatorDataGpu::getForces()
{
return impl_->getForces();
}
return impl_->copyForcesToGpu(h_f, atomLocality);
}
-GpuEventSynchronizer* StatePropagatorDataGpu::getForcesReadyOnDeviceEvent(AtomLocality atomLocality,
- bool useGpuFBufferOps)
+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_->fReducedOnDevice(atomLocality);
+}
+
+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)
}
-void* StatePropagatorDataGpu::getUpdateStream()
+const DeviceStream* StatePropagatorDataGpu::getUpdateStream()
{
return impl_->getUpdateStream();
}
-int StatePropagatorDataGpu::numAtomsLocal()
+int StatePropagatorDataGpu::numAtomsLocal() const
{
return impl_->numAtomsLocal();
}
-int StatePropagatorDataGpu::numAtomsAll()
+int StatePropagatorDataGpu::numAtomsAll() const
{
return impl_->numAtomsAll();
}
} // namespace gmx
-#endif // GMX_GPU == GMX_GPU_NONE
+#endif // GMX_GPU