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.
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_);
}
"be All, Local or NonLocal.");
}
GMX_ASSERT(atomsStartAt >= 0,
- "The first elemtnt to copy has negative index. Probably, the GPU propagator state "
+ "The first element to copy has negative index. Probably, the GPU propagator state "
"was not initialized.");
GMX_ASSERT(numAtomsToCopy >= 0,
"Number of atoms to copy is negative. Probably, the GPU propagator state was not "
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);
}
// 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 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_OPENCL)
- {
- return nullptr;
- }
+
if (atomLocality == AtomLocality::NonLocal && stepWork.useGpuXHalo)
{
GMX_ASSERT(gpuCoordinateHaloLaunched != nullptr,