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;
71 template <bool usePBC>
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;
98 /*! \brief unpack non-local force data buffer on the GPU using pre-populated "map" containing index information
99 * \param[out] data full array of force values
100 * \param[in] dataPacked packed array of force values to be transferred
101 * \param[in] map array of indices defining mapping from full to packed array
102 * \param[in] mapSize number of elements in map array
104 template <bool accumulate>
105 __global__ void unpackRecvBufKernel(float3 * __restrict__ data,
106 const float3 * __restrict__ dataPacked,
107 const int * __restrict__ map,
111 int threadIndex = blockIdx.x*blockDim.x+threadIdx.x;
112 const float3 *gm_dataSrc = &dataPacked[threadIndex];
113 float3 *gm_dataDest = &data[map[threadIndex]];
115 if (threadIndex < mapSize)
119 *gm_dataDest += *gm_dataSrc;
123 *gm_dataDest = *gm_dataSrc;
130 void GpuHaloExchange::Impl::reinitHalo(float3 *d_coordinatesBuffer,
131 float3 *d_forcesBuffer)
134 d_x_ = d_coordinatesBuffer;
135 d_f_ = d_forcesBuffer;
137 cudaStream_t stream = nonLocalStream_;
139 const gmx_domdec_comm_t &comm = *dd_->comm;
140 const gmx_domdec_comm_dim_t &cd = comm.cd[0];
141 const gmx_domdec_ind_t &ind = cd.ind[0];
142 int newSize = ind.nsend[nzone+1];
144 GMX_RELEASE_ASSERT(cd.numPulses() == 1, "Multiple pulses are not yet supported in GPU halo exchange");
145 GMX_ASSERT(cd.receiveInPlace, "Out-of-place receive is not yet supported in GPU halo exchange");
147 // reallocates only if needed
148 h_indexMap_.resize(newSize);
149 // reallocate on device only if needed
150 if (newSize > maxPackedBufferSize_)
152 reallocateDeviceBuffer(&d_indexMap_, newSize, &indexMapSize_, &indexMapSizeAlloc_, nullptr);
153 reallocateDeviceBuffer(&d_sendBuf_, newSize, &sendBufSize_, &sendBufSizeAlloc_, nullptr);
154 reallocateDeviceBuffer(&d_recvBuf_, newSize, &recvBufSize_, &recvBufSizeAlloc_, nullptr);
155 maxPackedBufferSize_ = newSize;
158 xSendSize_ = newSize;
160 MPI_Sendrecv(&xSendSize_, sizeof(int), MPI_BYTE, sendRankX_, 0,
161 &xRecvSize_, sizeof(int), MPI_BYTE, recvRankX_, 0,
162 mpi_comm_mysim_, MPI_STATUS_IGNORE);
164 fSendSize_ = xRecvSize_;
165 fRecvSize_ = xSendSize_;
167 numHomeAtoms_ = comm.atomRanges.numHomeAtoms(); //offset for data recieved by this rank
169 GMX_ASSERT(ind.index.size() == h_indexMap_.size(), "Size mismatch");
170 std::copy(ind.index.begin(), ind.index.end(), h_indexMap_.begin());
172 copyToDeviceBuffer(&d_indexMap_, h_indexMap_.data(), 0, newSize, stream, GpuApiCallBehavior::Async, nullptr);
174 // This rank will push data to its neighbor, so needs to know
175 // the remote receive address and similarly send its receive
176 // address to other neighbour. We can do this here in reinit fn
177 // since the pointers will not change until the next NS step.
179 //Coordinates buffer:
181 void* recvPtr = static_cast<void*> (&d_coordinatesBuffer[numHomeAtoms_]);
182 MPI_Sendrecv(&recvPtr, sizeof(void*), MPI_BYTE, recvRankX_, 0,
183 &remoteXPtr_, sizeof(void*), MPI_BYTE, sendRankX_, 0,
184 mpi_comm_mysim_, MPI_STATUS_IGNORE);
187 recvPtr = static_cast<void*> (d_recvBuf_);
188 MPI_Sendrecv(&recvPtr, sizeof(void*), MPI_BYTE, recvRankF_, 0,
189 &remoteFPtr_, sizeof(void*), MPI_BYTE, sendRankF_, 0,
190 mpi_comm_mysim_, MPI_STATUS_IGNORE);
197 // The following method be called after local setCoordinates (which records the coordinatesOnDeviceEvent_
198 // event when the coordinate data has been copied to the device).
199 void GpuHaloExchange::Impl::communicateHaloCoordinates(const matrix box)
202 //ensure stream waits until coordinate data is available on device
203 coordinatesOnDeviceEvent_->enqueueWaitEvent(nonLocalStream_);
205 // launch kernel to pack send buffer
206 KernelLaunchConfig config;
207 config.blockSize[0] = c_threadsPerBlock;
208 config.blockSize[1] = 1;
209 config.blockSize[2] = 1;
210 config.gridSize[0] = (xSendSize_+c_threadsPerBlock-1)/c_threadsPerBlock;
211 config.gridSize[1] = 1;
212 config.gridSize[2] = 1;
213 config.sharedMemorySize = 0;
214 config.stream = nonLocalStream_;
216 const float3 *sendBuf = d_sendBuf_;
217 const float3 *d_x = d_x_;
218 const int *indexMap = d_indexMap_;
219 const int size = xSendSize_;
220 // The coordinateShift changes between steps when we have
221 // performed a DD partition, or have updated the box e.g. when
222 // performing pressure coupling. So, for simplicity, the the box
223 // is used every step to pass the shift vector as an argument of
224 // the packing kernel.
226 // Because only one-dimensional DD is supported, the coordinate
227 // shift only needs to handle that dimension.
228 const int dimensionIndex = dd_->dim[0];
229 const float3 coordinateShift {
230 box[dimensionIndex][XX], box[dimensionIndex][YY], box[dimensionIndex][ZZ]
233 // Avoid launching kernel when there is no work to do
236 auto kernelFn = usePBC_ ? packSendBufKernel<true> : packSendBufKernel<false>;
238 const auto kernelArgs = prepareGpuKernelArguments(kernelFn, config, &sendBuf, &d_x, &indexMap,
239 &size, &coordinateShift);
241 launchGpuKernel(kernelFn, config, nullptr, "Domdec GPU Apply X Halo Exchange", kernelArgs);
244 communicateHaloData(d_x_, HaloQuantity::HaloCoordinates);
249 // The following method should be called after non-local buffer operations,
250 // and before the local buffer operations. It operates in the non-local stream.
251 void GpuHaloExchange::Impl::communicateHaloForces(bool accumulateForces)
254 // Communicate halo data (in non-local stream)
255 communicateHaloData(d_f_, HaloQuantity::HaloForces);
259 if (!accumulateForces)
261 //Clear local portion of force array (in local stream)
262 cudaMemsetAsync(d_f, 0, numHomeAtoms_*sizeof(rvec), localStream_);
265 // ensure non-local stream waits for local stream, due to dependence on
266 // the previous H2D copy of CPU forces (if accumulateForces is true)
267 // or the above clearing.
268 // TODO remove this dependency on localStream - edmine issue #3093
269 GpuEventSynchronizer eventLocal;
270 eventLocal.markEvent(localStream_);
271 eventLocal.enqueueWaitEvent(nonLocalStream_);
273 //Unpack halo buffer into force array
275 KernelLaunchConfig config;
276 config.blockSize[0] = c_threadsPerBlock;
277 config.blockSize[1] = 1;
278 config.blockSize[2] = 1;
279 config.gridSize[0] = (fRecvSize_+c_threadsPerBlock-1)/c_threadsPerBlock;
280 config.gridSize[1] = 1;
281 config.gridSize[2] = 1;
282 config.sharedMemorySize = 0;
283 config.stream = nonLocalStream_;
285 const float3 *recvBuf = d_recvBuf_;
286 const int *indexMap = d_indexMap_;
287 const int size = fRecvSize_;
291 auto kernelFn = accumulateForces ? unpackRecvBufKernel<true> : unpackRecvBufKernel<false>;
293 const auto kernelArgs = prepareGpuKernelArguments(kernelFn, config, &d_f,
297 launchGpuKernel(kernelFn, config, nullptr, "Domdec GPU Apply F Halo Exchange", kernelArgs);
302 void GpuHaloExchange::Impl::communicateHaloData(float3 * d_ptr,
303 HaloQuantity haloQuantity)
312 if (haloQuantity == HaloQuantity::HaloCoordinates)
314 sendPtr = static_cast<void*> (d_sendBuf_);
315 sendSize = xSendSize_;
316 remotePtr = remoteXPtr_;
317 sendRank = sendRankX_;
318 recvRank = recvRankX_;
321 //Wait for signal from receiving task that it is ready, and similarly send signal to task that will push data to this task
322 char thisTaskIsReady, remoteTaskIsReady;
323 MPI_Sendrecv(&thisTaskIsReady, sizeof(char), MPI_BYTE, recvRank, 0,
324 &remoteTaskIsReady, sizeof(char), MPI_BYTE, sendRank, 0,
325 mpi_comm_mysim_, MPI_STATUS_IGNORE);
330 sendPtr = static_cast<void*> (&(d_ptr[numHomeAtoms_]));
331 sendSize = fSendSize_;
332 remotePtr = remoteFPtr_;
333 sendRank = sendRankF_;
334 recvRank = recvRankF_;
337 communicateHaloDataWithCudaDirect(sendPtr, sendSize, sendRank, remotePtr, recvRank);
340 void GpuHaloExchange::Impl::communicateHaloDataWithCudaDirect(void *sendPtr,
348 cudaStream_t stream = nonLocalStream_;
350 // We asynchronously push data to remote rank. The remote
351 // destination pointer has already been set in the init fn. We
352 // don't need to worry about overwriting data the remote ranks
353 // still needs since the halo exchange is just done once per
354 // timestep, for each of X and F.
356 // send data to neighbor, if any data exists to send
359 stat = cudaMemcpyAsync(remotePtr, sendPtr, sendSize*DIM*sizeof(float), cudaMemcpyDeviceToDevice, stream);
360 CU_RET_ERR(stat, "cudaMemcpyAsync on GPU Domdec CUDA direct data transfer failed");
364 //ensure pushed data has arrived before remote rank progresses
365 // This rank records an event and sends it to the remote rank which has just been pushed data.
366 // This rank recieves event from remote rank which has pushed data here, and enqueues that event to
368 GpuEventSynchronizer *haloDataTransferRemote;
370 haloDataTransferLaunched_->markEvent(stream);
372 MPI_Sendrecv(&haloDataTransferLaunched_, sizeof(GpuEventSynchronizer*), MPI_BYTE, sendRank, 0,
373 &haloDataTransferRemote, sizeof(GpuEventSynchronizer*), MPI_BYTE, recvRank, 0,
374 mpi_comm_mysim_, MPI_STATUS_IGNORE);
376 haloDataTransferRemote->enqueueWaitEvent(stream);
378 GMX_UNUSED_VALUE(sendRank);
379 GMX_UNUSED_VALUE(recvRank);
384 /*! \brief Create Domdec GPU object */
385 GpuHaloExchange::Impl::Impl(gmx_domdec_t *dd,
386 MPI_Comm mpi_comm_mysim,
388 void * nonLocalStream,
389 void * coordinatesOnDeviceEvent)
391 sendRankX_(dd->neighbor[0][1]),
392 recvRankX_(dd->neighbor[0][0]),
393 sendRankF_(dd->neighbor[0][0]),
394 recvRankF_(dd->neighbor[0][1]),
395 usePBC_(dd->ci[dd->dim[0]] == 0),
396 haloDataTransferLaunched_(new GpuEventSynchronizer()),
397 mpi_comm_mysim_(mpi_comm_mysim),
398 localStream_(*static_cast<cudaStream_t*> (localStream)),
399 nonLocalStream_(*static_cast<cudaStream_t*> (nonLocalStream)),
400 coordinatesOnDeviceEvent_(static_cast<GpuEventSynchronizer*> (coordinatesOnDeviceEvent))
403 GMX_RELEASE_ASSERT(GMX_THREAD_MPI, "GPU Halo exchange is currently only supported with thread-MPI enabled");
407 gmx_fatal(FARGS, "Error: dd->ndim > 1 is not yet supported in GPU halo exchange");
410 if (usePBC_ && dd->unitCellInfo.haveScrewPBC)
412 gmx_fatal(FARGS, "Error: screw is not yet supported in GPU halo exchange\n");
415 changePinningPolicy(&h_indexMap_, gmx::PinningPolicy::PinnedIfSupported);
417 allocateDeviceBuffer(&d_fShift_, 1, nullptr);
421 GpuHaloExchange::Impl::~Impl()
423 freeDeviceBuffer(&d_indexMap_);
424 freeDeviceBuffer(&d_sendBuf_);
425 freeDeviceBuffer(&d_recvBuf_);
426 freeDeviceBuffer(&d_fShift_);
427 delete haloDataTransferLaunched_;
430 GpuHaloExchange::GpuHaloExchange(gmx_domdec_t *dd,
431 MPI_Comm mpi_comm_mysim,
433 void *nonLocalStream,
434 void *coordinatesOnDeviceEvent)
435 : impl_(new Impl(dd, mpi_comm_mysim, localStream, nonLocalStream, coordinatesOnDeviceEvent))
439 GpuHaloExchange::~GpuHaloExchange() = default;
441 void GpuHaloExchange::reinitHalo(DeviceBuffer<float> d_coordinatesBuffer,
442 DeviceBuffer<float> d_forcesBuffer)
444 impl_->reinitHalo(reinterpret_cast<float3*>(d_coordinatesBuffer), reinterpret_cast<float3*>(d_forcesBuffer));
447 void GpuHaloExchange::communicateHaloCoordinates(const matrix box)
449 impl_->communicateHaloCoordinates(box);
452 void GpuHaloExchange::communicateHaloForces(bool accumulateForces)
454 impl_->communicateHaloForces(accumulateForces);