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) 2017,2018,2019,2020,2021, 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/hardware/device_information.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;
82 initParamLookupTable(&nbp->coulomb_tab, &nbp->coulomb_tab_texobj, tables.tableF.data(),
83 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(), environmentVariableSuffix.c_str(),
95 environmentVariableSuffix.c_str());
99 int nbnxn_gpu_pick_ewald_kernel_type(const interaction_const_t& ic,
100 const DeviceInformation gmx_unused& deviceInfo)
102 bool bTwinCut = (ic.rcoulomb != ic.rvdw);
105 /* Benchmarking/development environment variables to force the use of
106 analytical or tabulated Ewald kernel. */
108 // Remove these when old environment variables are deprecated
109 const bool forceAnalyticalEwaldLegacy = (getenv("GMX_CUDA_NB_ANA_EWALD") != nullptr)
110 || (getenv("GMX_OCL_NB_ANA_EWALD") != nullptr);
111 const bool forceTabulatedEwaldLegacy = (getenv("GMX_CUDA_NB_TAB_EWALD") != nullptr)
112 || (getenv("GMX_OCL_NB_TAB_EWALD") != nullptr);
113 const bool forceTwinCutoffEwaldLegacy = (getenv("GMX_CUDA_NB_EWALD_TWINCUT") != nullptr)
114 || (getenv("GMX_OCL_NB_EWALD_TWINCUT") != nullptr);
116 printEnvironmentVariableDeprecationMessage(forceAnalyticalEwaldLegacy, "NB_ANA_EWALD");
117 printEnvironmentVariableDeprecationMessage(forceTabulatedEwaldLegacy, "NB_TAB_EWALD");
118 printEnvironmentVariableDeprecationMessage(forceTwinCutoffEwaldLegacy, "NB_EWALD_TWINCUT");
120 const bool forceAnalyticalEwald =
121 (getenv("GMX_GPU_NB_ANA_EWALD") != nullptr) || forceAnalyticalEwaldLegacy;
122 const bool forceTabulatedEwald =
123 (getenv("GMX_GPU_NB_TAB_EWALD") != nullptr) || forceTabulatedEwaldLegacy;
124 const bool forceTwinCutoffEwald =
125 (getenv("GMX_GPU_NB_EWALD_TWINCUT") != nullptr) || forceTwinCutoffEwaldLegacy;
127 if (forceAnalyticalEwald && forceTabulatedEwald)
130 "Both analytical and tabulated Ewald GPU non-bonded kernels "
131 "requested through environment variables.");
134 /* By default, use analytical Ewald except with CUDA on NVIDIA CC 7.0 and 8.0.
136 const bool c_useTabulatedEwaldDefault =
138 (deviceInfo.prop.major == 7 && deviceInfo.prop.minor == 0)
139 || (deviceInfo.prop.major == 8 && deviceInfo.prop.minor == 0);
143 bool bUseAnalyticalEwald = !c_useTabulatedEwaldDefault;
144 if (forceAnalyticalEwald)
146 bUseAnalyticalEwald = true;
149 fprintf(debug, "Using analytical Ewald GPU kernels\n");
152 else if (forceTabulatedEwald)
154 bUseAnalyticalEwald = false;
158 fprintf(debug, "Using tabulated Ewald GPU kernels\n");
162 /* Use twin cut-off kernels if requested by bTwinCut or the env. var.
163 forces it (use it for debugging/benchmarking only). */
164 if (!bTwinCut && !forceTwinCutoffEwald)
166 kernel_type = bUseAnalyticalEwald ? eelTypeEWALD_ANA : eelTypeEWALD_TAB;
170 kernel_type = bUseAnalyticalEwald ? eelTypeEWALD_ANA_TWIN : eelTypeEWALD_TAB_TWIN;
176 void set_cutoff_parameters(NBParamGpu* nbp, const interaction_const_t* ic, const PairlistParams& listParams)
178 nbp->ewald_beta = ic->ewaldcoeff_q;
179 nbp->sh_ewald = ic->sh_ewald;
180 nbp->epsfac = ic->epsfac;
181 nbp->two_k_rf = 2.0 * ic->k_rf;
182 nbp->c_rf = ic->c_rf;
183 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
184 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
185 nbp->rlistOuter_sq = listParams.rlistOuter * listParams.rlistOuter;
186 nbp->rlistInner_sq = listParams.rlistInner * listParams.rlistInner;
187 nbp->useDynamicPruning = listParams.useDynamicPruning;
189 nbp->sh_lj_ewald = ic->sh_lj_ewald;
190 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
192 nbp->rvdw_switch = ic->rvdw_switch;
193 nbp->dispersion_shift = ic->dispersion_shift;
194 nbp->repulsion_shift = ic->repulsion_shift;
195 nbp->vdw_switch = ic->vdw_switch;
198 void gpu_pme_loadbal_update_param(const nonbonded_verlet_t* nbv, const interaction_const_t* ic)
200 if (!nbv || !nbv->useGpu())
204 NbnxmGpu* nb = nbv->gpu_nbv;
205 NBParamGpu* nbp = nb->nbparam;
207 set_cutoff_parameters(nbp, ic, nbv->pairlistSets().params());
209 nbp->eeltype = nbnxn_gpu_pick_ewald_kernel_type(*ic, nb->deviceContext_->deviceInfo());
211 GMX_RELEASE_ASSERT(ic->coulombEwaldTables, "Need valid Coulomb Ewald correction tables");
212 init_ewald_coulomb_force_table(*ic->coulombEwaldTables, nbp, *nb->deviceContext_);
215 void init_plist(gpu_plist* pl)
217 /* initialize to nullptr pointers to data that is not allocated here and will
218 need reallocation in nbnxn_gpu_init_pairlist */
224 /* size -1 indicates that the respective array hasn't been initialized yet */
231 pl->imask_nalloc = -1;
233 pl->excl_nalloc = -1;
234 pl->haveFreshList = false;
237 void init_timings(gmx_wallclock_gpu_nbnxn_t* t)
246 for (i = 0; i < 2; i++)
248 for (j = 0; j < 2; j++)
250 t->ktime[i][j].t = 0.0;
251 t->ktime[i][j].c = 0;
255 t->pruneTime.t = 0.0;
256 t->dynamicPruneTime.c = 0;
257 t->dynamicPruneTime.t = 0.0;
260 //! This function is documented in the header file
261 void gpu_init_pairlist(NbnxmGpu* nb, const NbnxnPairlistGpu* h_plist, const InteractionLocality iloc)
264 // Timing accumulation should happen only if there was work to do
265 // because getLastRangeTime() gets skipped with empty lists later
266 // which leads to the counter not being reset.
267 bool bDoTime = (nb->bDoTime && !h_plist->sci.empty());
268 const DeviceStream& deviceStream = *nb->deviceStreams[iloc];
269 gpu_plist* d_plist = nb->plist[iloc];
271 if (d_plist->na_c < 0)
273 d_plist->na_c = h_plist->na_ci;
277 if (d_plist->na_c != h_plist->na_ci)
279 sprintf(sbuf, "In init_plist: the #atoms per cell has changed (from %d to %d)",
280 d_plist->na_c, h_plist->na_ci);
285 gpu_timers_t::Interaction& iTimers = nb->timers->interaction[iloc];
289 iTimers.pl_h2d.openTimingRegion(deviceStream);
290 iTimers.didPairlistH2D = true;
293 // TODO most of this function is same in CUDA and OpenCL, move into the header
294 const DeviceContext& deviceContext = *nb->deviceContext_;
296 reallocateDeviceBuffer(&d_plist->sci, h_plist->sci.size(), &d_plist->nsci, &d_plist->sci_nalloc,
298 copyToDeviceBuffer(&d_plist->sci, h_plist->sci.data(), 0, h_plist->sci.size(), deviceStream,
299 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
301 reallocateDeviceBuffer(&d_plist->cj4, h_plist->cj4.size(), &d_plist->ncj4, &d_plist->cj4_nalloc,
303 copyToDeviceBuffer(&d_plist->cj4, h_plist->cj4.data(), 0, h_plist->cj4.size(), deviceStream,
304 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
306 reallocateDeviceBuffer(&d_plist->imask, h_plist->cj4.size() * c_nbnxnGpuClusterpairSplit,
307 &d_plist->nimask, &d_plist->imask_nalloc, deviceContext);
309 reallocateDeviceBuffer(&d_plist->excl, h_plist->excl.size(), &d_plist->nexcl,
310 &d_plist->excl_nalloc, deviceContext);
311 copyToDeviceBuffer(&d_plist->excl, h_plist->excl.data(), 0, h_plist->excl.size(), deviceStream,
312 GpuApiCallBehavior::Async, bDoTime ? iTimers.pl_h2d.fetchNextEvent() : nullptr);
316 iTimers.pl_h2d.closeTimingRegion(deviceStream);
319 /* need to prune the pair list during the next step */
320 d_plist->haveFreshList = true;
323 //! This function is documented in the header file
324 gmx_wallclock_gpu_nbnxn_t* gpu_get_timings(NbnxmGpu* nb)
326 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
329 //! This function is documented in the header file
330 void gpu_reset_timings(nonbonded_verlet_t* nbv)
332 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
334 init_timings(nbv->gpu_nbv->timings);
338 bool gpu_is_kernel_ewald_analytical(const NbnxmGpu* nb)
340 return ((nb->nbparam->eeltype == eelTypeEWALD_ANA) || (nb->nbparam->eeltype == eelTypeEWALD_ANA_TWIN));