2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2019,2020,2021, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
37 * \brief Implements GPU halo exchange using CUDA.
40 * \author Alan Gray <alang@nvidia.com>
42 * \ingroup module_domdec
46 #include "gpuhaloexchange_impl.cuh"
55 #include "gromacs/domdec/domdec.h"
56 #include "gromacs/domdec/domdec_struct.h"
57 #include "gromacs/domdec/gpuhaloexchange.h"
58 #include "gromacs/gpu_utils/cudautils.cuh"
59 #include "gromacs/gpu_utils/device_context.h"
60 #include "gromacs/gpu_utils/devicebuffer.h"
61 #include "gromacs/gpu_utils/gpueventsynchronizer.h"
62 #include "gromacs/gpu_utils/typecasts.cuh"
63 #include "gromacs/gpu_utils/vectype_ops.cuh"
64 #include "gromacs/math/vectypes.h"
65 #include "gromacs/pbcutil/ishift.h"
66 #include "gromacs/timing/wallcycle.h"
67 #include "gromacs/utility/gmxmpi.h"
69 #include "domdec_internal.h"
74 //! Number of CUDA threads in a block
75 // TODO Optimize this through experimentation
76 constexpr static int c_threadsPerBlock = 256;
79 __global__ void packSendBufKernel(float3* __restrict__ dataPacked,
80 const float3* __restrict__ data,
81 const int* __restrict__ map,
83 const float3 coordinateShift)
85 int threadIndex = blockIdx.x * blockDim.x + threadIdx.x;
86 float3* gm_dataDest = &dataPacked[threadIndex];
87 const float3* gm_dataSrc = &data[map[threadIndex]];
89 if (threadIndex < mapSize)
93 *gm_dataDest = *gm_dataSrc + coordinateShift;
97 *gm_dataDest = *gm_dataSrc;
102 /*! \brief unpack non-local force data buffer on the GPU using pre-populated "map" containing index
103 * information \param[out] data full array of force values \param[in] dataPacked packed
104 * array of force values to be transferred \param[in] map array of indices defining mapping
105 * from full to packed array \param[in] mapSize number of elements in map array
107 template<bool accumulate>
108 __global__ void unpackRecvBufKernel(float3* __restrict__ data,
109 const float3* __restrict__ dataPacked,
110 const int* __restrict__ map,
114 int threadIndex = blockIdx.x * blockDim.x + threadIdx.x;
115 const float3* gm_dataSrc = &dataPacked[threadIndex];
116 float3* gm_dataDest = &data[map[threadIndex]];
118 if (threadIndex < mapSize)
122 *gm_dataDest += *gm_dataSrc;
126 *gm_dataDest = *gm_dataSrc;
131 void GpuHaloExchange::Impl::reinitHalo(float3* d_coordinatesBuffer, float3* d_forcesBuffer)
133 wallcycle_start(wcycle_, WallCycleCounter::Domdec);
134 wallcycle_sub_start(wcycle_, WallCycleSubCounter::DDGpu);
136 d_x_ = d_coordinatesBuffer;
137 d_f_ = d_forcesBuffer;
139 const gmx_domdec_comm_t& comm = *dd_->comm;
140 const gmx_domdec_comm_dim_t& cd = comm.cd[dimIndex_];
141 const gmx_domdec_ind_t& ind = cd.ind[pulse_];
143 numHomeAtoms_ = comm.atomRanges.numHomeAtoms(); // offset for data recieved by this rank
145 // Determine receive offset for the dimension index and pulse of this halo exchange object
148 int numAtomsTotal = numHomeAtoms_;
149 for (int i = 0; i <= dimIndex_; i++)
151 int pulseMax = (i == dimIndex_) ? pulse_ : (comm.cd[i].numPulses() - 1);
152 for (int p = 0; p <= pulseMax; p++)
154 atomOffset_ = numAtomsTotal;
155 const gmx_domdec_ind_t& indTemp = comm.cd[i].ind[p];
156 numAtomsTotal += indTemp.nrecv[numZoneTemp + 1];
158 numZone = numZoneTemp;
159 numZoneTemp += numZoneTemp;
162 int newSize = ind.nsend[numZone + 1];
164 GMX_ASSERT(cd.receiveInPlace, "Out-of-place receive is not yet supported in GPU halo exchange");
166 // reallocates only if needed
167 h_indexMap_.resize(newSize);
168 // reallocate on device only if needed
169 if (newSize > maxPackedBufferSize_)
171 reallocateDeviceBuffer(&d_indexMap_, newSize, &indexMapSize_, &indexMapSizeAlloc_, deviceContext_);
172 reallocateDeviceBuffer(&d_sendBuf_, newSize, &sendBufSize_, &sendBufSizeAlloc_, deviceContext_);
173 reallocateDeviceBuffer(&d_recvBuf_, newSize, &recvBufSize_, &recvBufSizeAlloc_, deviceContext_);
174 maxPackedBufferSize_ = newSize;
177 xSendSize_ = newSize;
179 MPI_Sendrecv(&xSendSize_,
192 fSendSize_ = xRecvSize_;
193 fRecvSize_ = xSendSize_;
197 GMX_ASSERT(ind.index.size() == h_indexMap_.size(),
198 "Size mismatch between domain decomposition communication index array and GPU "
199 "halo exchange index mapping array");
200 std::copy(ind.index.begin(), ind.index.end(), h_indexMap_.begin());
203 &d_indexMap_, h_indexMap_.data(), 0, newSize, nonLocalStream_, GpuApiCallBehavior::Async, nullptr);
207 // Exchange of remote addresses from neighboring ranks is needed only with CUDA-direct as cudamemcpy needs both src/dst pointer
208 // MPI calls such as MPI_send doesn't worry about receiving address, that is taken care by MPI_recv call in neighboring rank
211 // This rank will push data to its neighbor, so needs to know
212 // the remote receive address and similarly send its receive
213 // address to other neighbour. We can do this here in reinit fn
214 // since the pointers will not change until the next NS step.
216 // Coordinates buffer:
217 float3* recvPtr = &d_x_[atomOffset_];
218 MPI_Sendrecv(&recvPtr,
232 recvPtr = d_recvBuf_;
233 MPI_Sendrecv(&recvPtr,
248 wallcycle_sub_stop(wcycle_, WallCycleSubCounter::DDGpu);
249 wallcycle_stop(wcycle_, WallCycleCounter::Domdec);
252 void GpuHaloExchange::Impl::enqueueWaitRemoteCoordinatesReadyEvent(GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
254 GMX_ASSERT(coordinatesReadyOnDeviceEvent != nullptr,
255 "Co-ordinate Halo exchange requires valid co-ordinate ready event");
257 // Wait for event from receiving task that remote coordinates are ready, and enqueue that event to stream used
258 // for subsequent data push. This avoids a race condition with the remote data being written in the previous timestep.
259 // Similarly send event to task that will push data to this task.
260 GpuEventSynchronizer* remoteCoordinatesReadyOnDeviceEvent;
261 MPI_Sendrecv(&coordinatesReadyOnDeviceEvent,
262 sizeof(GpuEventSynchronizer*), //NOLINT(bugprone-sizeof-expression)
266 &remoteCoordinatesReadyOnDeviceEvent,
267 sizeof(GpuEventSynchronizer*), //NOLINT(bugprone-sizeof-expression)
273 remoteCoordinatesReadyOnDeviceEvent->enqueueWaitEvent(nonLocalStream_);
276 void GpuHaloExchange::Impl::communicateHaloCoordinates(const matrix box,
277 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
280 wallcycle_start(wcycle_, WallCycleCounter::LaunchGpu);
283 // ensure stream waits until coordinate data is available on device
284 coordinatesReadyOnDeviceEvent->enqueueWaitEvent(nonLocalStream_);
287 wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuMoveX);
289 // launch kernel to pack send buffer
290 KernelLaunchConfig config;
291 config.blockSize[0] = c_threadsPerBlock;
292 config.blockSize[1] = 1;
293 config.blockSize[2] = 1;
294 config.gridSize[0] = (xSendSize_ + c_threadsPerBlock - 1) / c_threadsPerBlock;
295 config.gridSize[1] = 1;
296 config.gridSize[2] = 1;
297 config.sharedMemorySize = 0;
299 const float3* sendBuf = d_sendBuf_;
300 const float3* d_x = d_x_;
301 const int* indexMap = d_indexMap_;
302 const int size = xSendSize_;
303 // The coordinateShift changes between steps when we have
304 // performed a DD partition, or have updated the box e.g. when
305 // performing pressure coupling. So, for simplicity, the box
306 // is used every step to pass the shift vector as an argument of
307 // the packing kernel.
308 const int boxDimensionIndex = dd_->dim[dimIndex_];
309 const float3 coordinateShift{ box[boxDimensionIndex][XX],
310 box[boxDimensionIndex][YY],
311 box[boxDimensionIndex][ZZ] };
313 // Avoid launching kernel when there is no work to do
316 auto kernelFn = usePBC_ ? packSendBufKernel<true> : packSendBufKernel<false>;
318 const auto kernelArgs = prepareGpuKernelArguments(
319 kernelFn, config, &sendBuf, &d_x, &indexMap, &size, &coordinateShift);
322 kernelFn, config, nonLocalStream_, nullptr, "Domdec GPU Apply X Halo Exchange", kernelArgs);
325 wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuMoveX);
326 wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
328 // Consider time spent in communicateHaloData as Comm.X counter
329 // ToDo: We need further refinement here as communicateHaloData includes launch time for cudamemcpyasync
330 wallcycle_start(wcycle_, WallCycleCounter::MoveX);
332 // wait for remote co-ordinates is implicit with process-MPI as non-local stream is synchronized before MPI calls
333 // and MPI_Waitall call makes sure both neighboring ranks' non-local stream is synchronized before data transfer is initiated
334 if (GMX_THREAD_MPI && pulse_ == 0)
336 enqueueWaitRemoteCoordinatesReadyEvent(coordinatesReadyOnDeviceEvent);
339 float3* recvPtr = GMX_THREAD_MPI ? remoteXPtr_ : &d_x_[atomOffset_];
340 communicateHaloData(d_sendBuf_, xSendSize_, sendRankX_, recvPtr, xRecvSize_, recvRankX_);
342 wallcycle_stop(wcycle_, WallCycleCounter::MoveX);
345 // The following method should be called after non-local buffer operations,
346 // and before the local buffer operations. It operates in the non-local stream.
347 void GpuHaloExchange::Impl::communicateHaloForces(bool accumulateForces)
349 // Consider time spent in communicateHaloData as Comm.F counter
350 // ToDo: We need further refinement here as communicateHaloData includes launch time for cudamemcpyasync
351 wallcycle_start(wcycle_, WallCycleCounter::MoveF);
353 float3* recvPtr = GMX_THREAD_MPI ? remoteFPtr_ : d_recvBuf_;
355 // Communicate halo data (in non-local stream)
356 communicateHaloData(&(d_f_[atomOffset_]), fSendSize_, sendRankF_, recvPtr, fRecvSize_, recvRankF_);
358 wallcycle_stop(wcycle_, WallCycleCounter::MoveF);
360 wallcycle_start_nocount(wcycle_, WallCycleCounter::LaunchGpu);
361 wallcycle_sub_start(wcycle_, WallCycleSubCounter::LaunchGpuMoveF);
364 // If this is the last pulse and index (noting the force halo
365 // exchanges across multiple pulses and indices are called in
366 // reverse order) then perform the following preparation
368 if ((pulse_ == (dd_->comm->cd[dimIndex_].numPulses() - 1)) && (dimIndex_ == (dd_->ndim - 1)))
370 // ensure non-local stream waits for local stream, due to dependence on
371 // the previous H2D copy of CPU forces (if accumulateForces is true)
372 // or local force clearing.
373 GpuEventSynchronizer eventLocal;
374 eventLocal.markEvent(localStream_);
375 eventLocal.enqueueWaitEvent(nonLocalStream_);
378 // Unpack halo buffer into force array
380 KernelLaunchConfig config;
381 config.blockSize[0] = c_threadsPerBlock;
382 config.blockSize[1] = 1;
383 config.blockSize[2] = 1;
384 config.gridSize[0] = (fRecvSize_ + c_threadsPerBlock - 1) / c_threadsPerBlock;
385 config.gridSize[1] = 1;
386 config.gridSize[2] = 1;
387 config.sharedMemorySize = 0;
389 const float3* recvBuf = d_recvBuf_;
390 const int* indexMap = d_indexMap_;
391 const int size = fRecvSize_;
393 if (pulse_ > 0 || dd_->ndim > 1)
395 // We need to accumulate rather than set, since it is possible
396 // that, in this pulse/dim, a value could be written to a location
397 // corresponding to the halo region of a following pulse/dim.
398 accumulateForces = true;
403 auto kernelFn = accumulateForces ? unpackRecvBufKernel<true> : unpackRecvBufKernel<false>;
405 const auto kernelArgs =
406 prepareGpuKernelArguments(kernelFn, config, &d_f, &recvBuf, &indexMap, &size);
409 kernelFn, config, nonLocalStream_, nullptr, "Domdec GPU Apply F Halo Exchange", kernelArgs);
414 fReadyOnDevice_.markEvent(nonLocalStream_);
417 wallcycle_sub_stop(wcycle_, WallCycleSubCounter::LaunchGpuMoveF);
418 wallcycle_stop(wcycle_, WallCycleCounter::LaunchGpu);
421 void GpuHaloExchange::Impl::communicateHaloData(float3* sendPtr,
430 // no need to explicitly sync with GMX_THREAD_MPI as all operations are
431 // anyway launched in correct stream
432 communicateHaloDataWithCudaDirect(sendPtr, sendSize, sendRank, recvPtr, recvRank);
436 communicateHaloDataWithCudaMPI(sendPtr, sendSize, sendRank, recvPtr, recvSize, recvRank);
440 void GpuHaloExchange::Impl::communicateHaloDataWithCudaMPI(float3* sendPtr,
447 // no need to wait for haloDataReadyOnDevice event if this rank is not sending any data
450 // wait for non local stream to complete all outstanding
451 // activities, to ensure that buffer is up-to-date in GPU memory
452 // before transferring to remote rank
454 // ToDo: Replace stream synchronize with event synchronize
455 nonLocalStream_.synchronize();
458 // perform halo exchange directly in device buffers
462 // recv remote data into halo region
463 MPI_Irecv(recvPtr, recvSize * DIM, MPI_FLOAT, recvRank, 0, mpi_comm_mysim_, &request);
465 // send data to remote halo region
466 MPI_Send(sendPtr, sendSize * DIM, MPI_FLOAT, sendRank, 0, mpi_comm_mysim_);
468 MPI_Wait(&request, MPI_STATUS_IGNORE);
472 void GpuHaloExchange::Impl::communicateHaloDataWithCudaDirect(float3* sendPtr,
481 // We asynchronously push data to remote rank. The remote
482 // destination pointer has already been set in the init fn. We
483 // don't need to worry about overwriting data the remote ranks
484 // still needs since the halo exchange is just done once per
485 // timestep, for each of X and F.
487 // send data to neighbor, if any data exists to send
490 stat = cudaMemcpyAsync(remotePtr,
492 sendSize * DIM * sizeof(float),
493 cudaMemcpyDeviceToDevice,
494 nonLocalStream_.stream());
496 CU_RET_ERR(stat, "cudaMemcpyAsync on GPU Domdec CUDA direct data transfer failed");
500 // ensure pushed data has arrived before remote rank progresses
501 // This rank records an event and sends it to the remote rank which has just been pushed data.
502 // This rank recieves event from remote rank which has pushed data here, and enqueues that event
504 GpuEventSynchronizer* haloDataTransferRemote;
506 GMX_ASSERT(haloDataTransferLaunched_ != nullptr,
507 "Halo exchange requires valid event to synchronize data transfer initiated in "
509 haloDataTransferLaunched_->markEvent(nonLocalStream_);
511 MPI_Sendrecv(&haloDataTransferLaunched_,
512 sizeof(GpuEventSynchronizer*), //NOLINT(bugprone-sizeof-expression)
516 &haloDataTransferRemote,
517 sizeof(GpuEventSynchronizer*), //NOLINT(bugprone-sizeof-expression)
524 haloDataTransferRemote->enqueueWaitEvent(nonLocalStream_);
526 GMX_UNUSED_VALUE(sendRank);
527 GMX_UNUSED_VALUE(recvRank);
531 GpuEventSynchronizer* GpuHaloExchange::Impl::getForcesReadyOnDeviceEvent()
533 return &fReadyOnDevice_;
536 /*! \brief Create Domdec GPU object */
537 GpuHaloExchange::Impl::Impl(gmx_domdec_t* dd,
539 MPI_Comm mpi_comm_mysim,
540 const DeviceContext& deviceContext,
541 const DeviceStream& localStream,
542 const DeviceStream& nonLocalStream,
544 gmx_wallcycle* wcycle) :
546 sendRankX_(dd->neighbor[dimIndex][1]),
547 recvRankX_(dd->neighbor[dimIndex][0]),
548 sendRankF_(dd->neighbor[dimIndex][0]),
549 recvRankF_(dd->neighbor[dimIndex][1]),
550 usePBC_(dd->ci[dd->dim[dimIndex]] == 0),
551 haloDataTransferLaunched_(GMX_THREAD_MPI ? new GpuEventSynchronizer() : nullptr),
552 mpi_comm_mysim_(mpi_comm_mysim),
553 deviceContext_(deviceContext),
554 localStream_(localStream),
555 nonLocalStream_(nonLocalStream),
560 if (usePBC_ && dd->unitCellInfo.haveScrewPBC)
562 gmx_fatal(FARGS, "Error: screw is not yet supported in GPU halo exchange\n");
565 changePinningPolicy(&h_indexMap_, gmx::PinningPolicy::PinnedIfSupported);
567 allocateDeviceBuffer(&d_fShift_, 1, deviceContext_);
570 GpuHaloExchange::Impl::~Impl()
572 freeDeviceBuffer(&d_indexMap_);
573 freeDeviceBuffer(&d_sendBuf_);
574 freeDeviceBuffer(&d_recvBuf_);
575 freeDeviceBuffer(&d_fShift_);
576 delete haloDataTransferLaunched_;
579 GpuHaloExchange::GpuHaloExchange(gmx_domdec_t* dd,
581 MPI_Comm mpi_comm_mysim,
582 const DeviceContext& deviceContext,
583 const DeviceStream& localStream,
584 const DeviceStream& nonLocalStream,
586 gmx_wallcycle* wcycle) :
587 impl_(new Impl(dd, dimIndex, mpi_comm_mysim, deviceContext, localStream, nonLocalStream, pulse, wcycle))
591 GpuHaloExchange::GpuHaloExchange(GpuHaloExchange&&) noexcept = default;
593 GpuHaloExchange& GpuHaloExchange::operator=(GpuHaloExchange&& other) noexcept
595 std::swap(impl_, other.impl_);
599 GpuHaloExchange::~GpuHaloExchange() = default;
601 void GpuHaloExchange::reinitHalo(DeviceBuffer<RVec> d_coordinatesBuffer, DeviceBuffer<RVec> d_forcesBuffer)
603 impl_->reinitHalo(asFloat3(d_coordinatesBuffer), asFloat3(d_forcesBuffer));
606 void GpuHaloExchange::communicateHaloCoordinates(const matrix box,
607 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
609 impl_->communicateHaloCoordinates(box, coordinatesReadyOnDeviceEvent);
612 void GpuHaloExchange::communicateHaloForces(bool accumulateForces)
614 impl_->communicateHaloForces(accumulateForces);
617 GpuEventSynchronizer* GpuHaloExchange::getForcesReadyOnDeviceEvent()
619 return impl_->getForcesReadyOnDeviceEvent();