2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2019, 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"
53 #include "gromacs/domdec/domdec.h"
54 #include "gromacs/domdec/domdec_struct.h"
55 #include "gromacs/domdec/gpuhaloexchange.h"
56 #include "gromacs/gpu_utils/cudautils.cuh"
57 #include "gromacs/gpu_utils/devicebuffer.h"
58 #include "gromacs/gpu_utils/gpueventsynchronizer.cuh"
59 #include "gromacs/gpu_utils/vectype_ops.cuh"
60 #include "gromacs/pbcutil/ishift.h"
62 #include "domdec_internal.h"
67 //! Number of CUDA threads in a block
68 // TODO Optimize this through experimentation
69 constexpr static int c_threadsPerBlock = 256;
72 __global__ void packSendBufKernel(float3* __restrict__ dataPacked,
73 const float3* __restrict__ data,
74 const int* __restrict__ map,
76 const float3 coordinateShift)
78 int threadIndex = blockIdx.x * blockDim.x + threadIdx.x;
79 float3* gm_dataDest = &dataPacked[threadIndex];
80 const float3* gm_dataSrc = &data[map[threadIndex]];
82 if (threadIndex < mapSize)
86 *gm_dataDest = *gm_dataSrc + coordinateShift;
90 *gm_dataDest = *gm_dataSrc;
97 /*! \brief unpack non-local force data buffer on the GPU using pre-populated "map" containing index
98 * information \param[out] data full array of force values \param[in] dataPacked packed
99 * array of force values to be transferred \param[in] map array of indices defining mapping
100 * from full to packed array \param[in] mapSize number of elements in map array
102 template<bool accumulate>
103 __global__ void unpackRecvBufKernel(float3* __restrict__ data,
104 const float3* __restrict__ dataPacked,
105 const int* __restrict__ map,
109 int threadIndex = blockIdx.x * blockDim.x + threadIdx.x;
110 const float3* gm_dataSrc = &dataPacked[threadIndex];
111 float3* gm_dataDest = &data[map[threadIndex]];
113 if (threadIndex < mapSize)
117 *gm_dataDest += *gm_dataSrc;
121 *gm_dataDest = *gm_dataSrc;
128 void GpuHaloExchange::Impl::reinitHalo(float3* d_coordinatesBuffer, float3* d_forcesBuffer)
131 d_x_ = d_coordinatesBuffer;
132 d_f_ = d_forcesBuffer;
134 cudaStream_t stream = nonLocalStream_;
136 const gmx_domdec_comm_t& comm = *dd_->comm;
137 const gmx_domdec_comm_dim_t& cd = comm.cd[0];
138 const gmx_domdec_ind_t& ind = cd.ind[0];
139 int newSize = ind.nsend[nzone + 1];
141 GMX_RELEASE_ASSERT(cd.numPulses() == 1,
142 "Multiple pulses are not yet supported in GPU halo exchange");
143 GMX_ASSERT(cd.receiveInPlace, "Out-of-place receive is not yet supported in GPU halo exchange");
145 // reallocates only if needed
146 h_indexMap_.resize(newSize);
147 // reallocate on device only if needed
148 if (newSize > maxPackedBufferSize_)
150 reallocateDeviceBuffer(&d_indexMap_, newSize, &indexMapSize_, &indexMapSizeAlloc_, nullptr);
151 reallocateDeviceBuffer(&d_sendBuf_, newSize, &sendBufSize_, &sendBufSizeAlloc_, nullptr);
152 reallocateDeviceBuffer(&d_recvBuf_, newSize, &recvBufSize_, &recvBufSizeAlloc_, nullptr);
153 maxPackedBufferSize_ = newSize;
156 xSendSize_ = newSize;
158 MPI_Sendrecv(&xSendSize_, sizeof(int), MPI_BYTE, sendRankX_, 0, &xRecvSize_, sizeof(int),
159 MPI_BYTE, recvRankX_, 0, mpi_comm_mysim_, MPI_STATUS_IGNORE);
161 fSendSize_ = xRecvSize_;
162 fRecvSize_ = xSendSize_;
164 numHomeAtoms_ = comm.atomRanges.numHomeAtoms(); // offset for data recieved by this rank
166 GMX_ASSERT(ind.index.size() == h_indexMap_.size(), "Size mismatch");
167 std::copy(ind.index.begin(), ind.index.end(), h_indexMap_.begin());
169 copyToDeviceBuffer(&d_indexMap_, h_indexMap_.data(), 0, newSize, stream,
170 GpuApiCallBehavior::Async, nullptr);
172 // This rank will push data to its neighbor, so needs to know
173 // the remote receive address and similarly send its receive
174 // address to other neighbour. We can do this here in reinit fn
175 // since the pointers will not change until the next NS step.
177 // Coordinates buffer:
179 void* recvPtr = static_cast<void*>(&d_coordinatesBuffer[numHomeAtoms_]);
180 MPI_Sendrecv(&recvPtr, sizeof(void*), MPI_BYTE, recvRankX_, 0, &remoteXPtr_, sizeof(void*),
181 MPI_BYTE, sendRankX_, 0, mpi_comm_mysim_, MPI_STATUS_IGNORE);
184 recvPtr = static_cast<void*>(d_recvBuf_);
185 MPI_Sendrecv(&recvPtr, sizeof(void*), MPI_BYTE, recvRankF_, 0, &remoteFPtr_, sizeof(void*),
186 MPI_BYTE, sendRankF_, 0, mpi_comm_mysim_, MPI_STATUS_IGNORE);
193 void GpuHaloExchange::Impl::communicateHaloCoordinates(const matrix box,
194 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
197 // ensure stream waits until coordinate data is available on device
198 coordinatesReadyOnDeviceEvent->enqueueWaitEvent(nonLocalStream_);
200 // launch kernel to pack send buffer
201 KernelLaunchConfig config;
202 config.blockSize[0] = c_threadsPerBlock;
203 config.blockSize[1] = 1;
204 config.blockSize[2] = 1;
205 config.gridSize[0] = (xSendSize_ + c_threadsPerBlock - 1) / c_threadsPerBlock;
206 config.gridSize[1] = 1;
207 config.gridSize[2] = 1;
208 config.sharedMemorySize = 0;
209 config.stream = nonLocalStream_;
211 const float3* sendBuf = d_sendBuf_;
212 const float3* d_x = d_x_;
213 const int* indexMap = d_indexMap_;
214 const int size = xSendSize_;
215 // The coordinateShift changes between steps when we have
216 // performed a DD partition, or have updated the box e.g. when
217 // performing pressure coupling. So, for simplicity, the box
218 // is used every step to pass the shift vector as an argument of
219 // the packing kernel.
221 // Because only one-dimensional DD is supported, the coordinate
222 // shift only needs to handle that dimension.
223 const int dimensionIndex = dd_->dim[0];
224 const float3 coordinateShift{ box[dimensionIndex][XX], box[dimensionIndex][YY],
225 box[dimensionIndex][ZZ] };
227 // Avoid launching kernel when there is no work to do
230 auto kernelFn = usePBC_ ? packSendBufKernel<true> : packSendBufKernel<false>;
232 const auto kernelArgs = prepareGpuKernelArguments(kernelFn, config, &sendBuf, &d_x,
233 &indexMap, &size, &coordinateShift);
235 launchGpuKernel(kernelFn, config, nullptr, "Domdec GPU Apply X Halo Exchange", kernelArgs);
238 communicateHaloData(d_x_, HaloQuantity::HaloCoordinates, coordinatesReadyOnDeviceEvent);
243 // The following method should be called after non-local buffer operations,
244 // and before the local buffer operations. It operates in the non-local stream.
245 void GpuHaloExchange::Impl::communicateHaloForces(bool accumulateForces)
248 // Communicate halo data (in non-local stream)
249 communicateHaloData(d_f_, HaloQuantity::HaloForces, nullptr);
253 if (!accumulateForces)
255 // Clear local portion of force array (in local stream)
256 cudaMemsetAsync(d_f, 0, numHomeAtoms_ * sizeof(rvec), localStream_);
259 // ensure non-local stream waits for local stream, due to dependence on
260 // the previous H2D copy of CPU forces (if accumulateForces is true)
261 // or the above clearing.
262 // TODO remove this dependency on localStream - edmine issue #3093
263 GpuEventSynchronizer eventLocal;
264 eventLocal.markEvent(localStream_);
265 eventLocal.enqueueWaitEvent(nonLocalStream_);
267 // Unpack halo buffer into force array
269 KernelLaunchConfig config;
270 config.blockSize[0] = c_threadsPerBlock;
271 config.blockSize[1] = 1;
272 config.blockSize[2] = 1;
273 config.gridSize[0] = (fRecvSize_ + c_threadsPerBlock - 1) / c_threadsPerBlock;
274 config.gridSize[1] = 1;
275 config.gridSize[2] = 1;
276 config.sharedMemorySize = 0;
277 config.stream = nonLocalStream_;
279 const float3* recvBuf = d_recvBuf_;
280 const int* indexMap = d_indexMap_;
281 const int size = fRecvSize_;
285 auto kernelFn = accumulateForces ? unpackRecvBufKernel<true> : unpackRecvBufKernel<false>;
287 const auto kernelArgs =
288 prepareGpuKernelArguments(kernelFn, config, &d_f, &recvBuf, &indexMap, &size);
290 launchGpuKernel(kernelFn, config, nullptr, "Domdec GPU Apply F Halo Exchange", kernelArgs);
295 void GpuHaloExchange::Impl::communicateHaloData(float3* d_ptr,
296 HaloQuantity haloQuantity,
297 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
306 if (haloQuantity == HaloQuantity::HaloCoordinates)
308 sendPtr = static_cast<void*>(d_sendBuf_);
309 sendSize = xSendSize_;
310 remotePtr = remoteXPtr_;
311 sendRank = sendRankX_;
312 recvRank = recvRankX_;
315 // Wait for event from receiving task that remote coordinates are ready, and enqueue that event to stream used
316 // for subsequent data push. This avoids a race condition with the remote data being written in the previous timestep.
317 // Similarly send event to task that will push data to this task.
318 GpuEventSynchronizer* remoteCoordinatesReadyOnDeviceEvent;
319 MPI_Sendrecv(&coordinatesReadyOnDeviceEvent, sizeof(GpuEventSynchronizer*), MPI_BYTE,
320 recvRank, 0, &remoteCoordinatesReadyOnDeviceEvent, sizeof(GpuEventSynchronizer*),
321 MPI_BYTE, sendRank, 0, mpi_comm_mysim_, MPI_STATUS_IGNORE);
322 remoteCoordinatesReadyOnDeviceEvent->enqueueWaitEvent(nonLocalStream_);
327 sendPtr = static_cast<void*>(&(d_ptr[numHomeAtoms_]));
328 sendSize = fSendSize_;
329 remotePtr = remoteFPtr_;
330 sendRank = sendRankF_;
331 recvRank = recvRankF_;
334 communicateHaloDataWithCudaDirect(sendPtr, sendSize, sendRank, remotePtr, recvRank);
337 void GpuHaloExchange::Impl::communicateHaloDataWithCudaDirect(void* sendPtr,
345 cudaStream_t stream = nonLocalStream_;
347 // We asynchronously push data to remote rank. The remote
348 // destination pointer has already been set in the init fn. We
349 // don't need to worry about overwriting data the remote ranks
350 // still needs since the halo exchange is just done once per
351 // timestep, for each of X and F.
353 // send data to neighbor, if any data exists to send
356 stat = cudaMemcpyAsync(remotePtr, sendPtr, sendSize * DIM * sizeof(float),
357 cudaMemcpyDeviceToDevice, stream);
358 CU_RET_ERR(stat, "cudaMemcpyAsync on GPU Domdec CUDA direct data transfer failed");
362 // ensure pushed data has arrived before remote rank progresses
363 // This rank records an event and sends it to the remote rank which has just been pushed data.
364 // This rank recieves event from remote rank which has pushed data here, and enqueues that event
366 GpuEventSynchronizer* haloDataTransferRemote;
368 haloDataTransferLaunched_->markEvent(stream);
370 MPI_Sendrecv(&haloDataTransferLaunched_, sizeof(GpuEventSynchronizer*), MPI_BYTE, sendRank, 0,
371 &haloDataTransferRemote, sizeof(GpuEventSynchronizer*), MPI_BYTE, recvRank, 0,
372 mpi_comm_mysim_, MPI_STATUS_IGNORE);
374 haloDataTransferRemote->enqueueWaitEvent(stream);
376 GMX_UNUSED_VALUE(sendRank);
377 GMX_UNUSED_VALUE(recvRank);
381 /*! \brief Create Domdec GPU object */
382 GpuHaloExchange::Impl::Impl(gmx_domdec_t* dd, MPI_Comm mpi_comm_mysim, void* localStream, void* nonLocalStream) :
384 sendRankX_(dd->neighbor[0][1]),
385 recvRankX_(dd->neighbor[0][0]),
386 sendRankF_(dd->neighbor[0][0]),
387 recvRankF_(dd->neighbor[0][1]),
388 usePBC_(dd->ci[dd->dim[0]] == 0),
389 haloDataTransferLaunched_(new GpuEventSynchronizer()),
390 mpi_comm_mysim_(mpi_comm_mysim),
391 localStream_(*static_cast<cudaStream_t*>(localStream)),
392 nonLocalStream_(*static_cast<cudaStream_t*>(nonLocalStream))
395 GMX_RELEASE_ASSERT(GMX_THREAD_MPI,
396 "GPU Halo exchange is currently only supported with thread-MPI enabled");
400 gmx_fatal(FARGS, "Error: dd->ndim > 1 is not yet supported in GPU halo exchange");
403 if (usePBC_ && dd->unitCellInfo.haveScrewPBC)
405 gmx_fatal(FARGS, "Error: screw is not yet supported in GPU halo exchange\n");
408 changePinningPolicy(&h_indexMap_, gmx::PinningPolicy::PinnedIfSupported);
410 allocateDeviceBuffer(&d_fShift_, 1, nullptr);
413 GpuHaloExchange::Impl::~Impl()
415 freeDeviceBuffer(&d_indexMap_);
416 freeDeviceBuffer(&d_sendBuf_);
417 freeDeviceBuffer(&d_recvBuf_);
418 freeDeviceBuffer(&d_fShift_);
419 delete haloDataTransferLaunched_;
422 GpuHaloExchange::GpuHaloExchange(gmx_domdec_t* dd, MPI_Comm mpi_comm_mysim, void* localStream, void* nonLocalStream) :
423 impl_(new Impl(dd, mpi_comm_mysim, localStream, nonLocalStream))
427 GpuHaloExchange::~GpuHaloExchange() = default;
429 void GpuHaloExchange::reinitHalo(DeviceBuffer<float> d_coordinatesBuffer, DeviceBuffer<float> d_forcesBuffer)
431 impl_->reinitHalo(reinterpret_cast<float3*>(d_coordinatesBuffer),
432 reinterpret_cast<float3*>(d_forcesBuffer));
435 void GpuHaloExchange::communicateHaloCoordinates(const matrix box,
436 GpuEventSynchronizer* coordinatesReadyOnDeviceEvent)
438 impl_->communicateHaloCoordinates(box, coordinatesReadyOnDeviceEvent);
441 void GpuHaloExchange::communicateHaloForces(bool accumulateForces)
443 impl_->communicateHaloForces(accumulateForces);