2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016 by the GROMACS development team.
5 * Copyright (c) 2017,2018,2019,2020, by the GROMACS development team, led by
6 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 * and including many others, as listed in the AUTHORS file in the
8 * top-level source directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
37 * \brief Define common implementation of nbnxm_gpu_data_mgmt.h
39 * \author Anca Hamuraru <anca@streamcomputing.eu>
40 * \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
41 * \author Teemu Virolainen <teemu@streamcomputing.eu>
42 * \author Szilárd Páll <pall.szilard@gmail.com>
43 * \author Artem Zhmurov <zhmurov@gmail.com>
45 * \ingroup module_nbnxm
52 # include "cuda/nbnxm_cuda_types.h"
56 # include "opencl/nbnxm_ocl_types.h"
59 #include "nbnxm_gpu_data_mgmt.h"
61 #include "gromacs/mdtypes/interaction_const.h"
62 #include "gromacs/nbnxm/gpu_data_mgmt.h"
63 #include "gromacs/timing/gpu_timing.h"
64 #include "gromacs/utility/cstringutil.h"
66 #include "nbnxm_gpu.h"
67 #include "pairlistsets.h"
72 void init_ewald_coulomb_force_table(const EwaldCorrectionTables& tables,
74 const DeviceContext& deviceContext)
78 destroyParamLookupTable(&nbp->coulomb_tab, nbp->coulomb_tab_texobj);
81 nbp->coulomb_tab_scale = tables.scale;
83 &nbp->coulomb_tab, &nbp->coulomb_tab_texobj, tables.tableF.data(), tables.tableF.size(), deviceContext);
86 void inline printEnvironmentVariableDeprecationMessage(bool isEnvironmentVariableSet,
87 const std::string& environmentVariableSuffix)
89 if (isEnvironmentVariableSet)
92 "Environment variables GMX_CUDA_%s and GMX_OCL_%s are deprecated and will be\n"
93 "removed in release 2022, please use GMX_GPU_%s instead.",
94 environmentVariableSuffix.c_str(),
95 environmentVariableSuffix.c_str(),
96 environmentVariableSuffix.c_str());
100 enum ElecType nbnxn_gpu_pick_ewald_kernel_type(const interaction_const_t& ic)
102 bool bTwinCut = (ic.rcoulomb != ic.rvdw);
104 /* Benchmarking/development environment variables to force the use of
105 analytical or tabulated Ewald kernel. */
107 // Remove these when old environment variables are deprecated
108 const bool forceAnalyticalEwaldLegacy = (getenv("GMX_CUDA_NB_ANA_EWALD") != nullptr)
109 || (getenv("GMX_OCL_NB_ANA_EWALD") != nullptr);
110 const bool forceTabulatedEwaldLegacy = (getenv("GMX_CUDA_NB_TAB_EWALD") != nullptr)
111 || (getenv("GMX_OCL_NB_TAB_EWALD") != nullptr);
112 const bool forceTwinCutoffEwaldLegacy = (getenv("GMX_CUDA_NB_EWALD_TWINCUT") != nullptr)
113 || (getenv("GMX_OCL_NB_EWALD_TWINCUT") != nullptr);
115 printEnvironmentVariableDeprecationMessage(forceAnalyticalEwaldLegacy, "NB_ANA_EWALD");
116 printEnvironmentVariableDeprecationMessage(forceTabulatedEwaldLegacy, "NB_TAB_EWALD");
117 printEnvironmentVariableDeprecationMessage(forceTwinCutoffEwaldLegacy, "NB_EWALD_TWINCUT");
119 const bool forceAnalyticalEwald =
120 (getenv("GMX_GPU_NB_ANA_EWALD") != nullptr) || forceAnalyticalEwaldLegacy;
121 const bool forceTabulatedEwald =
122 (getenv("GMX_GPU_NB_TAB_EWALD") != nullptr) || forceTabulatedEwaldLegacy;
123 const bool forceTwinCutoffEwald =
124 (getenv("GMX_GPU_NB_EWALD_TWINCUT") != nullptr) || forceTwinCutoffEwaldLegacy;
126 if (forceAnalyticalEwald && forceTabulatedEwald)
129 "Both analytical and tabulated Ewald GPU non-bonded kernels "
130 "requested through environment variables.");
133 /* By default, use analytical Ewald
134 * TODO: tabulated does not work in OpenCL, it needs fixing, see init_nbparam() in nbnxn_ocl_data_mgmt.cpp
137 bool bUseAnalyticalEwald = true;
138 if (forceAnalyticalEwald)
142 fprintf(debug, "Using analytical Ewald GPU kernels\n");
145 else if (forceTabulatedEwald)
147 bUseAnalyticalEwald = false;
151 fprintf(debug, "Using tabulated Ewald GPU kernels\n");
155 /* Use twin cut-off kernels if requested by bTwinCut or the env. var.
156 forces it (use it for debugging/benchmarking only). */
157 if (!bTwinCut && !forceTwinCutoffEwald)
159 return bUseAnalyticalEwald ? ElecType::EwaldAna : ElecType::EwaldTab;
163 return bUseAnalyticalEwald ? ElecType::EwaldAnaTwin : ElecType::EwaldTabTwin;
167 void set_cutoff_parameters(NBParamGpu* nbp, const interaction_const_t* ic, const PairlistParams& listParams)
169 nbp->ewald_beta = ic->ewaldcoeff_q;
170 nbp->sh_ewald = ic->sh_ewald;
171 nbp->epsfac = ic->epsfac;
172 nbp->two_k_rf = 2.0 * ic->k_rf;
173 nbp->c_rf = ic->c_rf;
174 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
175 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
176 nbp->rlistOuter_sq = listParams.rlistOuter * listParams.rlistOuter;
177 nbp->rlistInner_sq = listParams.rlistInner * listParams.rlistInner;
178 nbp->useDynamicPruning = listParams.useDynamicPruning;
180 nbp->sh_lj_ewald = ic->sh_lj_ewald;
181 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
183 nbp->rvdw_switch = ic->rvdw_switch;
184 nbp->dispersion_shift = ic->dispersion_shift;
185 nbp->repulsion_shift = ic->repulsion_shift;
186 nbp->vdw_switch = ic->vdw_switch;
189 void gpu_pme_loadbal_update_param(const nonbonded_verlet_t* nbv, const interaction_const_t* ic)
191 if (!nbv || !nbv->useGpu())
195 NbnxmGpu* nb = nbv->gpu_nbv;
196 NBParamGpu* nbp = nb->nbparam;
198 set_cutoff_parameters(nbp, ic, nbv->pairlistSets().params());
200 nbp->elecType = nbnxn_gpu_pick_ewald_kernel_type(*ic);
202 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
203 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, *nb->deviceContext_);
206 void init_plist(gpu_plist* pl)
208 /* initialize to nullptr pointers to data that is not allocated here and will
209 need reallocation in nbnxn_gpu_init_pairlist */
215 /* size -1 indicates that the respective array hasn't been initialized yet */
222 pl->imask_nalloc = -1;
224 pl->excl_nalloc = -1;
225 pl->haveFreshList = false;
228 void init_timings(gmx_wallclock_gpu_nbnxn_t* t)
237 for (i = 0; i < 2; i++)
239 for (j = 0; j < 2; j++)
241 t->ktime[i][j].t = 0.0;
242 t->ktime[i][j].c = 0;
246 t->pruneTime.t = 0.0;
247 t->dynamicPruneTime.c = 0;
248 t->dynamicPruneTime.t = 0.0;
251 //! This function is documented in the header file
252 void gpu_init_pairlist(NbnxmGpu* nb, const NbnxnPairlistGpu* h_plist, const InteractionLocality iloc)
255 // Timing accumulation should happen only if there was work to do
256 // because getLastRangeTime() gets skipped with empty lists later
257 // which leads to the counter not being reset.
258 bool bDoTime = (nb->bDoTime && !h_plist->sci.empty());
259 const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
260 gpu_plist* d_plist = nb->plist[iloc];
262 if (d_plist->na_c < 0)
264 d_plist->na_c = h_plist->na_ci;
268 if (d_plist->na_c != h_plist->na_ci)
271 "In init_plist: the #atoms per cell has changed (from %d to %d)",
278 gpu_timers_t::Interaction& iTimers = nb->timers->interaction[iloc];
282 iTimers.pl_h2d.openTimingRegion(deviceStream);
283 iTimers.didPairlistH2D = true;
286 // TODO most of this function is same in CUDA and OpenCL, move into the header
287 const DeviceContext& deviceContext = *nb->deviceContext_;
289 reallocateDeviceBuffer(
290 &d_plist->sci, h_plist->sci.size(), &d_plist->nsci, &d_plist->sci_nalloc, deviceContext);
291 copyToDeviceBuffer(&d_plist->sci,
296 GpuApiCallBehavior::Async,
297 bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
299 reallocateDeviceBuffer(
300 &d_plist->cj4, h_plist->cj4.size(), &d_plist->ncj4, &d_plist->cj4_nalloc, deviceContext);
301 copyToDeviceBuffer(&d_plist->cj4,
306 GpuApiCallBehavior::Async,
307 bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
309 reallocateDeviceBuffer(&d_plist->imask,
310 h_plist->cj4.size() * c_nbnxnGpuClusterpairSplit,
312 &d_plist->imask_nalloc,
315 reallocateDeviceBuffer(
316 &d_plist->excl, h_plist->excl.size(), &d_plist->nexcl, &d_plist->excl_nalloc, deviceContext);
317 copyToDeviceBuffer(&d_plist->excl,
318 h_plist->excl.data(),
320 h_plist->excl.size(),
322 GpuApiCallBehavior::Async,
323 bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
327 iTimers.pl_h2d.closeTimingRegion(deviceStream);
330 /* need to prune the pair list during the next step */
331 d_plist->haveFreshList = true;
334 //! This function is documented in the header file
335 gmx_wallclock_gpu_nbnxn_t* gpu_get_timings(NbnxmGpu* nb)
337 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
340 //! This function is documented in the header file
341 void gpu_reset_timings(nonbonded_verlet_t* nbv)
343 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
345 init_timings(nbv->gpu_nbv->timings);
349 bool gpu_is_kernel_ewald_analytical(const NbnxmGpu* nb)
351 return ((nb->nbparam->elecType == ElecType::EwaldAna)
352 || (nb->nbparam->elecType == ElecType::EwaldAnaTwin));
355 enum ElecType nbnxmGpuPickElectrostaticsKernelType(const interaction_const_t* ic)
357 if (ic->eeltype == eelCUT)
359 return ElecType::Cut;
361 else if (EEL_RF(ic->eeltype))
365 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
367 return nbnxn_gpu_pick_ewald_kernel_type(*ic);
371 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
372 GMX_THROW(gmx::InconsistentInputError(
373 gmx::formatString("The requested electrostatics type %s (%d) is not implemented in "
374 "the GPU accelerated kernels!",
375 EELTYPE(ic->eeltype),
381 enum VdwType nbnxmGpuPickVdwKernelType(const interaction_const_t* ic, int combRule)
383 if (ic->vdwtype == evdwCUT)
385 switch (ic->vdw_modifier)
388 case eintmodPOTSHIFT:
391 case ljcrNONE: return VdwType::Cut;
392 case ljcrGEOM: return VdwType::CutCombGeom;
393 case ljcrLB: return VdwType::CutCombLB;
395 GMX_THROW(gmx::InconsistentInputError(gmx::formatString(
396 "The requested LJ combination rule %s (%d) is not implemented in "
397 "the GPU accelerated kernels!",
398 enum_name(combRule, ljcrNR, c_ljcrNames),
401 case eintmodFORCESWITCH: return VdwType::FSwitch;
402 case eintmodPOTSWITCH: return VdwType::PSwitch;
404 GMX_THROW(gmx::InconsistentInputError(
405 gmx::formatString("The requested VdW interaction modifier %s (%d) is not "
406 "implemented in the GPU accelerated kernels!",
407 INTMODIFIER(ic->vdw_modifier),
411 else if (ic->vdwtype == evdwPME)
413 if (ic->ljpme_comb_rule == ljcrGEOM)
415 assert(combRule == ljcrGEOM);
416 return VdwType::EwaldGeom;
420 assert(combRule == ljcrLB);
421 return VdwType::EwaldLB;
426 GMX_THROW(gmx::InconsistentInputError(gmx::formatString(
427 "The requested VdW type %s (%d) is not implemented in the GPU accelerated kernels!",
428 EVDWTYPE(ic->vdwtype),