2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 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.
38 * Stubs of functions that must be defined by nbnxm sycl implementation.
40 * \ingroup module_nbnxm
44 #include "gromacs/gpu_utils/device_stream_manager.h"
45 #include "gromacs/gpu_utils/pmalloc.h"
46 #include "gromacs/hardware/device_information.h"
47 #include "gromacs/mdtypes/interaction_const.h"
48 #include "gromacs/nbnxm/atomdata.h"
49 #include "gromacs/nbnxm/gpu_data_mgmt.h"
50 #include "gromacs/nbnxm/nbnxm_gpu.h"
51 #include "gromacs/nbnxm/nbnxm_gpu_data_mgmt.h"
52 #include "gromacs/pbcutil/ishift.h"
53 #include "gromacs/utility/exceptions.h"
55 #include "nbnxm_sycl_types.h"
60 //! This function is documented in the header file
61 void gpu_clear_outputs(NbnxmGpu* nb, bool computeVirial)
63 NBAtomData* adat = nb->atdat;
64 const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
66 clearDeviceBufferAsync(&adat->f, 0, nb->atdat->numAtoms, localStream);
67 // Clear shift force array and energies if the outputs were used in the current step
70 clearDeviceBufferAsync(&adat->fShift, 0, SHIFTS, localStream);
71 clearDeviceBufferAsync(&adat->eLJ, 0, 1, localStream);
72 clearDeviceBufferAsync(&adat->eElec, 0, 1, localStream);
76 /*! \brief Initialize \p atomdata first time; it only gets filled at pair-search. */
77 static void initAtomdataFirst(NbnxmGpu* nb, int numTypes, const DeviceContext& deviceContext)
79 const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
80 NBAtomData* atomdata = nb->atdat;
81 atomdata->numTypes = numTypes;
82 allocateDeviceBuffer(&atomdata->shiftVec, SHIFTS, deviceContext);
83 atomdata->shiftVecUploaded = false;
85 allocateDeviceBuffer(&atomdata->fShift, SHIFTS, deviceContext);
86 allocateDeviceBuffer(&atomdata->eLJ, 1, deviceContext);
87 allocateDeviceBuffer(&atomdata->eElec, 1, deviceContext);
89 clearDeviceBufferAsync(&atomdata->fShift, 0, SHIFTS, localStream);
90 clearDeviceBufferAsync(&atomdata->eElec, 0, 1, localStream);
91 clearDeviceBufferAsync(&atomdata->eLJ, 0, 1, localStream);
93 /* initialize to nullptr pointers to data that is not allocated here and will
94 need reallocation in later */
95 atomdata->xq = nullptr;
96 atomdata->f = nullptr;
98 /* size -1 indicates that the respective array hasn't been initialized yet */
99 atomdata->numAtoms = -1;
100 atomdata->numAtomsAlloc = -1;
103 /*! \brief Initialize the nonbonded parameter data structure. */
104 static void initNbparam(NBParamGpu* nbp,
105 const interaction_const_t& ic,
106 const PairlistParams& listParams,
107 const nbnxn_atomdata_t::Params& nbatParams,
108 const DeviceContext& deviceContext)
110 const int numTypes = nbatParams.numTypes;
112 set_cutoff_parameters(nbp, &ic, listParams);
114 nbp->vdwType = nbnxmGpuPickVdwKernelType(&ic, nbatParams.ljCombinationRule);
115 nbp->elecType = nbnxmGpuPickElectrostaticsKernelType(&ic, deviceContext.deviceInfo());
117 /* generate table for PME */
118 nbp->coulomb_tab = nullptr;
119 if (nbp->elecType == ElecType::EwaldTab || nbp->elecType == ElecType::EwaldTabTwin)
121 GMX_RELEASE_ASSERT(ic.coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
122 init_ewald_coulomb_force_table(*ic.coulombEwaldTables, nbp, deviceContext);
125 /* set up LJ parameter lookup table */
126 if (!useLjCombRule(nbp->vdwType))
128 static_assert(sizeof(decltype(nbp->nbfp)) == 2 * sizeof(decltype(*nbatParams.nbfp.data())),
129 "Mismatch in the size of host / device data types");
130 initParamLookupTable(&nbp->nbfp,
132 reinterpret_cast<const Float2*>(nbatParams.nbfp.data()),
137 /* set up LJ-PME parameter lookup table */
138 if (ic.vdwtype == VanDerWaalsType::Pme)
140 static_assert(sizeof(decltype(nbp->nbfp_comb))
141 == 2 * sizeof(decltype(*nbatParams.nbfp_comb.data())),
142 "Mismatch in the size of host / device data types");
143 initParamLookupTable(&nbp->nbfp_comb,
144 &nbp->nbfp_comb_texobj,
145 reinterpret_cast<const Float2*>(nbatParams.nbfp_comb.data()),
151 NbnxmGpu* gpu_init(const gmx::DeviceStreamManager& deviceStreamManager,
152 const interaction_const_t* ic,
153 const PairlistParams& listParams,
154 const nbnxn_atomdata_t* nbat,
155 const bool bLocalAndNonlocal)
157 auto* nb = new NbnxmGpu();
158 nb->deviceContext_ = &deviceStreamManager.context();
159 nb->atdat = new NBAtomData;
160 nb->nbparam = new NBParamGpu;
161 nb->plist[InteractionLocality::Local] = new Nbnxm::gpu_plist;
162 if (bLocalAndNonlocal)
164 nb->plist[InteractionLocality::NonLocal] = new Nbnxm::gpu_plist;
167 nb->bUseTwoStreams = bLocalAndNonlocal;
169 nb->timers = nullptr;
170 nb->timings = nullptr;
173 pmalloc(reinterpret_cast<void**>(&nb->nbst.eLJ), sizeof(*nb->nbst.eLJ));
174 pmalloc(reinterpret_cast<void**>(&nb->nbst.eElec), sizeof(*nb->nbst.eElec));
175 pmalloc(reinterpret_cast<void**>(&nb->nbst.fShift), SHIFTS * sizeof(*nb->nbst.fShift));
177 init_plist(nb->plist[InteractionLocality::Local]);
179 /* local/non-local GPU streams */
180 GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedLocal),
181 "Local non-bonded stream should be initialized to use GPU for non-bonded.");
182 nb->deviceStreams[InteractionLocality::Local] =
183 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedLocal);
184 // In general, it's not strictly necessary to use 2 streams for SYCL, since they are
185 // out-of-order. But for the time being, it will be less disruptive to keep them.
186 if (nb->bUseTwoStreams)
188 init_plist(nb->plist[InteractionLocality::NonLocal]);
190 GMX_RELEASE_ASSERT(deviceStreamManager.streamIsValid(gmx::DeviceStreamType::NonBondedNonLocal),
191 "Non-local non-bonded stream should be initialized to use GPU for "
192 "non-bonded with domain decomposition.");
193 nb->deviceStreams[InteractionLocality::NonLocal] =
194 &deviceStreamManager.stream(gmx::DeviceStreamType::NonBondedNonLocal);
199 const nbnxn_atomdata_t::Params& nbatParams = nbat->params();
200 const DeviceContext& deviceContext = *nb->deviceContext_;
202 initNbparam(nb->nbparam, *ic, listParams, nbatParams, deviceContext);
203 initAtomdataFirst(nb, nbatParams.numTypes, deviceContext);
208 void gpu_upload_shiftvec(NbnxmGpu* nb, const nbnxn_atomdata_t* nbatom)
210 NBAtomData* adat = nb->atdat;
211 const DeviceStream& localStream = *nb->deviceStreams[InteractionLocality::Local];
213 /* only if we have a dynamic box */
214 if (nbatom->bDynamicBox || !adat->shiftVecUploaded)
216 GMX_ASSERT(adat->shiftVec.elementSize() == sizeof(nbatom->shift_vec[0]),
217 "Sizes of host- and device-side shift vectors should be the same.");
218 copyToDeviceBuffer(&adat->shiftVec,
219 reinterpret_cast<const Float3*>(nbatom->shift_vec.data()),
223 GpuApiCallBehavior::Async,
225 adat->shiftVecUploaded = true;
229 void gpu_free(NbnxmGpu* nb)
236 NBAtomData* atdat = nb->atdat;
237 NBParamGpu* nbparam = nb->nbparam;
239 if ((!nbparam->coulomb_tab)
240 && (nbparam->elecType == ElecType::EwaldTab || nbparam->elecType == ElecType::EwaldTabTwin))
242 destroyParamLookupTable(&nbparam->coulomb_tab, nbparam->coulomb_tab_texobj);
245 if (!useLjCombRule(nb->nbparam->vdwType))
247 destroyParamLookupTable(&nbparam->nbfp, nbparam->nbfp_texobj);
250 if (nbparam->vdwType == VdwType::EwaldGeom || nbparam->vdwType == VdwType::EwaldLB)
252 destroyParamLookupTable(&nbparam->nbfp_comb, nbparam->nbfp_comb_texobj);
256 auto* plist = nb->plist[InteractionLocality::Local];
258 if (nb->bUseTwoStreams)
260 auto* plist_nl = nb->plist[InteractionLocality::NonLocal];
266 nb->nbst.eLJ = nullptr;
268 pfree(nb->nbst.eElec);
269 nb->nbst.eElec = nullptr;
271 pfree(nb->nbst.fShift);
272 nb->nbst.fShift = nullptr;
279 int gpu_min_ci_balanced(NbnxmGpu* nb)
281 // SYCL-TODO: Logic and magic values taken from OpenCL
282 static constexpr unsigned int balancedFactor = 50;
287 const cl::sycl::device device = nb->deviceContext_->deviceInfo().syclDevice;
288 const int numComputeUnits = device.get_info<cl::sycl::info::device::max_compute_units>();
289 return balancedFactor * numComputeUnits;