/*
* This file is part of the GROMACS molecular simulation package.
*
- * Copyright (c) 2020, by the GROMACS development team, led by
+ * Copyright (c) 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.
*/
/*! \internal \file
*
- * \brief May be used to implement force reduction interfaces for non-GPU builds.
+ * \brief Implements backend-agnostic GPU Force Reduction functions
*
* \author Alan Gray <alang@nvidia.com>
*
#include "gmxpre.h"
-#include "config.h"
+#include "gpuforcereduction_impl.h"
-#include "gpuforcereduction.h"
-
-#if !GMX_GPU_CUDA
+#include "gromacs/gpu_utils/device_stream.h"
+#include "gromacs/gpu_utils/devicebuffer.h"
+#include "gromacs/gpu_utils/gpueventsynchronizer.h"
+#include "gromacs/mdlib/gpuforcereduction_impl_internal.h"
+#include "gromacs/utility/gmxassert.h"
namespace gmx
{
-class GpuForceReduction::Impl
+GpuForceReduction::Impl::Impl(const DeviceContext& deviceContext,
+ const DeviceStream& deviceStream,
+ gmx_wallcycle* wcycle) :
+ baseForce_(),
+ deviceContext_(deviceContext),
+ deviceStream_(deviceStream),
+ nbnxmForceToAdd_(),
+ rvecForceToAdd_(),
+ wcycle_(wcycle)
+{
+}
+
+void GpuForceReduction::Impl::reinit(DeviceBuffer<Float3> baseForcePtr,
+ const int numAtoms,
+ ArrayRef<const int> cell,
+ const int atomStart,
+ const bool accumulate,
+ GpuEventSynchronizer* completionMarker)
+{
+ GMX_ASSERT(baseForcePtr, "Input base force for reduction has no data");
+ baseForce_ = baseForcePtr;
+ numAtoms_ = numAtoms;
+ atomStart_ = atomStart;
+ accumulate_ = accumulate;
+ completionMarker_ = completionMarker;
+ cellInfo_.cell = cell.data();
+
+ wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
+ reallocateDeviceBuffer(
+ &cellInfo_.d_cell, numAtoms_, &cellInfo_.cellSize, &cellInfo_.cellSizeAlloc, deviceContext_);
+ copyToDeviceBuffer(&cellInfo_.d_cell,
+ &(cellInfo_.cell[atomStart]),
+ 0,
+ numAtoms_,
+ deviceStream_,
+ GpuApiCallBehavior::Async,
+ nullptr);
+ wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
+
+ dependencyList_.clear();
+};
+
+void GpuForceReduction::Impl::registerNbnxmForce(DeviceBuffer<RVec> forcePtr)
+{
+ GMX_ASSERT(forcePtr, "Input force for reduction has no data");
+ nbnxmForceToAdd_ = forcePtr;
+};
+
+void GpuForceReduction::Impl::registerRvecForce(DeviceBuffer<RVec> forcePtr)
{
+ GMX_ASSERT(forcePtr, "Input force for reduction has no data");
+ rvecForceToAdd_ = forcePtr;
};
-GpuForceReduction::GpuForceReduction(const DeviceContext& /* deviceContext */,
- const DeviceStream& /* deviceStream */,
- gmx_wallcycle* /*wcycle*/) :
- impl_(nullptr)
+void GpuForceReduction::Impl::addDependency(GpuEventSynchronizer* dependency)
{
- GMX_ASSERT(false, "A CPU stub has been called instead of the correct implementation.");
+ dependencyList_.push_back(dependency);
}
-// NOLINTNEXTLINE readability-convert-member-functions-to-static
-void GpuForceReduction::reinit(DeviceBuffer<RVec> /*baseForcePtr*/,
- const int /*numAtoms*/,
- ArrayRef<const int> /*cell*/,
- const int /*atomStart*/,
- const bool /*accumulate*/,
- GpuEventSynchronizer* /*completionMarker*/)
+void GpuForceReduction::Impl::execute()
{
- GMX_ASSERT(false, "A CPU stub has been called instead of the correct implementation.");
+ wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
+ wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuNBFBufOps);
+
+ if (numAtoms_ != 0)
+ {
+ GMX_ASSERT(nbnxmForceToAdd_, "Nbnxm force for reduction has no data");
+
+ // Enqueue wait on all dependencies passed
+ for (auto* synchronizer : dependencyList_)
+ {
+ synchronizer->enqueueWaitEvent(deviceStream_);
+ }
+
+ const bool addRvecForce = static_cast<bool>(rvecForceToAdd_); // True iff initialized
+
+ launchForceReductionKernel(numAtoms_,
+ atomStart_,
+ addRvecForce,
+ accumulate_,
+ nbnxmForceToAdd_,
+ rvecForceToAdd_,
+ baseForce_,
+ cellInfo_.d_cell,
+ deviceStream_);
+ }
+ else
+ {
+ /* In case we have nothing to do, but still have dependencies, we need
+ * to consume them and mark our own event.
+ * Happens sometimes in MdrunVsitesTest.
+ * Issue #3988, #4227. */
+ for (auto* synchronizer : dependencyList_)
+ {
+ synchronizer->consume();
+ }
+ }
+
+ /* Mark that kernel has been launched.
+ * Even if we have no work to do and have not launched the kernel, we still mark the event
+ * in order to ensure proper marking/consumption balance, see Issue #3988, #4227. */
+ if (completionMarker_ != nullptr)
+ {
+ completionMarker_->markEvent(deviceStream_);
+ }
+
+ wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuNBFBufOps);
+ wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
}
-// NOLINTNEXTLINE readability-convert-member-functions-to-static
-void GpuForceReduction::registerNbnxmForce(void* /* forcePtr */)
+GpuForceReduction::GpuForceReduction(const DeviceContext& deviceContext,
+ const DeviceStream& deviceStream,
+ gmx_wallcycle* wcycle) :
+ impl_(new Impl(deviceContext, deviceStream, wcycle))
{
- GMX_ASSERT(false, "A CPU stub has been called instead of the correct implementation.");
}
-// NOLINTNEXTLINE readability-convert-member-functions-to-static
-void GpuForceReduction::registerRvecForce(void* /* forcePtr */)
+void GpuForceReduction::registerNbnxmForce(DeviceBuffer<RVec> forcePtr)
{
- GMX_ASSERT(false, "A CPU stub has been called instead of the correct implementation.");
+ impl_->registerNbnxmForce(forcePtr);
}
-// NOLINTNEXTLINE readability-convert-member-functions-to-static
-void GpuForceReduction::addDependency(GpuEventSynchronizer* const /* dependency */)
+void GpuForceReduction::registerRvecForce(DeviceBuffer<RVec> forcePtr)
{
- GMX_ASSERT(false, "A CPU stub has been called instead of the correct implementation.");
+ impl_->registerRvecForce(forcePtr);
}
-// NOLINTNEXTLINE readability-convert-member-functions-to-static
+void GpuForceReduction::addDependency(GpuEventSynchronizer* dependency)
+{
+ impl_->addDependency(dependency);
+}
+
+void GpuForceReduction::reinit(DeviceBuffer<RVec> baseForcePtr,
+ const int numAtoms,
+ ArrayRef<const int> cell,
+ const int atomStart,
+ const bool accumulate,
+ GpuEventSynchronizer* completionMarker)
+{
+ impl_->reinit(baseForcePtr, numAtoms, cell, atomStart, accumulate, completionMarker);
+}
void GpuForceReduction::execute()
{
- GMX_ASSERT(false, "A CPU stub has been called instead of the correct implementation.");
+ impl_->execute();
}
GpuForceReduction::~GpuForceReduction() = default;
} // namespace gmx
-
-#endif
biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff