2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2018,2019,2020, 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 bonded lists for CUDA
39 * \author Berk Hess <hess@kth.se>
40 * \author Szilárd Páll <pall.szilard@gmail.com>
41 * \author Mark Abraham <mark.j.abraham@gmail.com>
42 * \author Magnus Lundborg <lundborg.magnus@gmail.com>
44 * \ingroup module_listed_forces
49 #include "gpubonded_impl.h"
51 #include "gromacs/gpu_utils/cuda_arch_utils.cuh"
52 #include "gromacs/gpu_utils/cudautils.cuh"
53 #include "gromacs/gpu_utils/device_context.h"
54 #include "gromacs/gpu_utils/devicebuffer.h"
55 #include "gromacs/gpu_utils/typecasts.cuh"
56 #include "gromacs/mdtypes/enerdata.h"
57 #include "gromacs/timing/wallcycle.h"
58 #include "gromacs/topology/forcefieldparameters.h"
65 // ---- GpuBonded::Impl
67 GpuBonded::Impl::Impl(const gmx_ffparams_t& ffparams,
68 const DeviceContext& deviceContext,
70 gmx_wallcycle* wcycle) :
71 deviceContext_(deviceContext)
73 stream_ = *static_cast<CommandStream*>(streamPtr);
76 allocateDeviceBuffer(&d_forceParams_, ffparams.numTypes(), deviceContext_);
77 // This could be an async transfer (if the source is pinned), so
78 // long as it uses the same stream as the kernels and we are happy
79 // to consume additional pinned pages.
80 copyToDeviceBuffer(&d_forceParams_, ffparams.iparams.data(), 0, ffparams.numTypes(), stream_,
81 GpuApiCallBehavior::Sync, nullptr);
83 allocateDeviceBuffer(&d_vTot_, F_NRE, deviceContext_);
84 clearDeviceBufferAsync(&d_vTot_, 0, F_NRE, stream_);
86 kernelParams_.d_forceParams = d_forceParams_;
87 kernelParams_.d_xq = d_xq_;
88 kernelParams_.d_f = d_f_;
89 kernelParams_.d_fShift = d_fShift_;
90 kernelParams_.d_vTot = d_vTot_;
91 for (int i = 0; i < numFTypesOnGpu; i++)
93 kernelParams_.d_iatoms[i] = nullptr;
94 kernelParams_.fTypeRangeStart[i] = 0;
95 kernelParams_.fTypeRangeEnd[i] = -1;
99 GpuBonded::Impl::~Impl()
101 for (int fType : fTypesOnGpu)
103 if (d_iLists_[fType].iatoms)
105 freeDeviceBuffer(&d_iLists_[fType].iatoms);
106 d_iLists_[fType].iatoms = nullptr;
110 freeDeviceBuffer(&d_forceParams_);
111 freeDeviceBuffer(&d_vTot_);
114 //! Return whether function type \p fType in \p idef has perturbed interactions
115 static bool fTypeHasPerturbedEntries(const InteractionDefinitions& idef, int fType)
117 GMX_ASSERT(idef.ilsort == ilsortNO_FE || idef.ilsort == ilsortFE_SORTED,
118 "Perturbed interations should be sorted here");
120 const InteractionList& ilist = idef.il[fType];
122 return (idef.ilsort != ilsortNO_FE && idef.numNonperturbedInteractions[fType] != ilist.size());
125 //! Converts \p src with atom indices in state order to \p dest in nbnxn order
126 static void convertIlistToNbnxnOrder(const InteractionList& src,
127 HostInteractionList* dest,
128 int numAtomsPerInteraction,
129 ArrayRef<const int> nbnxnAtomOrder)
131 GMX_ASSERT(src.size() == 0 || !nbnxnAtomOrder.empty(), "We need the nbnxn atom order");
133 dest->iatoms.resize(src.size());
135 // TODO use OpenMP to parallelise this loop
136 for (int i = 0; i < src.size(); i += 1 + numAtomsPerInteraction)
138 dest->iatoms[i] = src.iatoms[i];
139 for (int a = 0; a < numAtomsPerInteraction; a++)
141 dest->iatoms[i + 1 + a] = nbnxnAtomOrder[src.iatoms[i + 1 + a]];
146 //! Returns \p input rounded up to the closest multiple of \p factor.
147 static inline int roundUpToFactor(const int input, const int factor)
149 GMX_ASSERT(factor > 0, "The factor to round up to must be > 0.");
151 int remainder = input % factor;
157 return (input + (factor - remainder));
160 // TODO Consider whether this function should be a factory method that
161 // makes an object that is the only one capable of the device
162 // operations needed for the lifetime of an interaction list. This
163 // would be harder to misuse than GpuBonded, and exchange the problem
164 // of naming this method for the problem of what to name the
165 // BondedDeviceInteractionListHandler type.
167 /*! Divides bonded interactions over threads and GPU.
168 * The bonded interactions are assigned by interaction type to GPU threads. The intereaction
169 * types are assigned in blocks sized as <warp_size>. The beginning and end (thread index) of each
170 * interaction type are stored in kernelParams_. Pointers to the relevant data structures on the
171 * GPU are also stored in kernelParams_.
173 * \todo Use DeviceBuffer for the d_xqPtr.
175 void GpuBonded::Impl::updateInteractionListsAndDeviceBuffers(ArrayRef<const int> nbnxnAtomOrder,
176 const InteractionDefinitions& idef,
178 DeviceBuffer<RVec> d_fPtr,
179 DeviceBuffer<RVec> d_fShiftPtr)
181 // TODO wallcycle sub start
182 haveInteractions_ = false;
183 int fTypesCounter = 0;
185 for (int fType : fTypesOnGpu)
187 auto& iList = iLists_[fType];
189 /* Perturbation is not implemented in the GPU bonded kernels.
190 * But instead of doing all interactions on the CPU, we can
191 * still easily handle the types that have no perturbed
192 * interactions on the GPU. */
193 if (!idef.il[fType].empty() && !fTypeHasPerturbedEntries(idef, fType))
195 haveInteractions_ = true;
197 convertIlistToNbnxnOrder(idef.il[fType], &iList, NRAL(fType), nbnxnAtomOrder);
201 iList.iatoms.clear();
204 // Update the device if necessary. This can leave some
205 // allocations on the device when the host size decreases to
206 // zero, which is OK, since we deallocate everything at the
208 if (iList.size() > 0)
210 t_ilist& d_iList = d_iLists_[fType];
212 reallocateDeviceBuffer(&d_iList.iatoms, iList.size(), &d_iList.nr, &d_iList.nalloc,
215 copyToDeviceBuffer(&d_iList.iatoms, iList.iatoms.data(), 0, iList.size(), stream_,
216 GpuApiCallBehavior::Async, nullptr);
218 kernelParams_.fTypesOnGpu[fTypesCounter] = fType;
219 kernelParams_.numFTypeIAtoms[fTypesCounter] = iList.size();
220 int numBonds = iList.size() / (interaction_function[fType].nratoms + 1);
221 kernelParams_.numFTypeBonds[fTypesCounter] = numBonds;
222 kernelParams_.d_iatoms[fTypesCounter] = d_iLists_[fType].iatoms;
223 if (fTypesCounter == 0)
225 kernelParams_.fTypeRangeStart[fTypesCounter] = 0;
229 kernelParams_.fTypeRangeStart[fTypesCounter] =
230 kernelParams_.fTypeRangeEnd[fTypesCounter - 1] + 1;
232 kernelParams_.fTypeRangeEnd[fTypesCounter] = kernelParams_.fTypeRangeStart[fTypesCounter]
233 + roundUpToFactor(numBonds, warp_size) - 1;
235 GMX_ASSERT(numBonds > 0
236 || kernelParams_.fTypeRangeEnd[fTypesCounter]
237 <= kernelParams_.fTypeRangeStart[fTypesCounter],
238 "Invalid GPU listed forces setup. numBonds must be > 0 if there are threads "
239 "allocated to do work on that interaction function type.");
240 GMX_ASSERT(kernelParams_.fTypeRangeStart[fTypesCounter] % warp_size == 0
241 && (kernelParams_.fTypeRangeEnd[fTypesCounter] + 1) % warp_size == 0,
242 "The bonded interactions must be assigned to the GPU in blocks of warp size.");
247 d_xq_ = static_cast<float4*>(d_xqPtr);
248 d_f_ = asFloat3(d_fPtr);
249 d_fShift_ = asFloat3(d_fShiftPtr);
251 kernelParams_.d_xq = d_xq_;
252 kernelParams_.d_f = d_f_;
253 kernelParams_.d_fShift = d_fShift_;
254 kernelParams_.d_forceParams = d_forceParams_;
255 kernelParams_.d_vTot = d_vTot_;
257 // TODO wallcycle sub stop
260 bool GpuBonded::Impl::haveInteractions() const
262 return haveInteractions_;
265 void GpuBonded::Impl::launchEnergyTransfer()
267 GMX_ASSERT(haveInteractions_,
268 "No GPU bonded interactions, so no energies will be computed, so transfer should "
270 wallcycle_sub_start_nocount(wcycle_, ewcsLAUNCH_GPU_BONDED);
271 // TODO add conditional on whether there has been any compute (and make sure host buffer doesn't contain garbage)
272 float* h_vTot = vTot_.data();
273 copyFromDeviceBuffer(h_vTot, &d_vTot_, 0, F_NRE, stream_, GpuApiCallBehavior::Async, nullptr);
274 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_GPU_BONDED);
277 void GpuBonded::Impl::waitAccumulateEnergyTerms(gmx_enerdata_t* enerd)
279 GMX_ASSERT(haveInteractions_,
280 "No GPU bonded interactions, so no energies will be computed or transferred, so "
281 "accumulation should not occur");
283 wallcycle_start(wcycle_, ewcWAIT_GPU_BONDED);
284 cudaError_t stat = cudaStreamSynchronize(stream_);
285 CU_RET_ERR(stat, "D2H transfer of bonded energies failed");
286 wallcycle_stop(wcycle_, ewcWAIT_GPU_BONDED);
288 for (int fType : fTypesOnGpu)
290 if (fType != F_LJ14 && fType != F_COUL14)
292 enerd->term[fType] += vTot_[fType];
296 // Note: We do not support energy groups here
297 gmx_grppairener_t* grppener = &enerd->grpp;
298 GMX_RELEASE_ASSERT(grppener->nener == 1, "No energy group support for bondeds on the GPU");
299 grppener->ener[egLJ14][0] += vTot_[F_LJ14];
300 grppener->ener[egCOUL14][0] += vTot_[F_COUL14];
303 void GpuBonded::Impl::clearEnergies()
305 wallcycle_start_nocount(wcycle_, ewcLAUNCH_GPU);
306 wallcycle_sub_start_nocount(wcycle_, ewcsLAUNCH_GPU_BONDED);
307 clearDeviceBufferAsync(&d_vTot_, 0, F_NRE, stream_);
308 wallcycle_sub_stop(wcycle_, ewcsLAUNCH_GPU_BONDED);
309 wallcycle_stop(wcycle_, ewcLAUNCH_GPU);
314 GpuBonded::GpuBonded(const gmx_ffparams_t& ffparams,
315 const DeviceContext& deviceContext,
317 gmx_wallcycle* wcycle) :
318 impl_(new Impl(ffparams, deviceContext, streamPtr, wcycle))
322 GpuBonded::~GpuBonded() = default;
324 void GpuBonded::updateInteractionListsAndDeviceBuffers(ArrayRef<const int> nbnxnAtomOrder,
325 const InteractionDefinitions& idef,
327 DeviceBuffer<RVec> d_f,
328 DeviceBuffer<RVec> d_fShift)
330 impl_->updateInteractionListsAndDeviceBuffers(nbnxnAtomOrder, idef, d_xq, d_f, d_fShift);
333 bool GpuBonded::haveInteractions() const
335 return impl_->haveInteractions();
338 void GpuBonded::launchEnergyTransfer()
340 impl_->launchEnergyTransfer();
343 void GpuBonded::waitAccumulateEnergyTerms(gmx_enerdata_t* enerd)
345 impl_->waitAccumulateEnergyTerms(enerd);
348 void GpuBonded::clearEnergies()
350 impl_->clearEnergies();