2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017,2018,2019, 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.
36 * \brief Define CUDA implementation of nbnxn_gpu_data_mgmt.h
38 * \author Szilard Pall <pall.szilard@gmail.com>
47 #include "gromacs/gpu_utils/cudautils.cuh"
48 #include "gromacs/gpu_utils/gpu_utils.h"
49 #include "gromacs/gpu_utils/pmalloc_cuda.h"
50 #include "gromacs/hardware/gpu_hw_info.h"
51 #include "gromacs/math/vectypes.h"
52 #include "gromacs/mdlib/force_flags.h"
53 #include "gromacs/mdtypes/interaction_const.h"
54 #include "gromacs/mdtypes/md_enums.h"
55 #include "gromacs/nbnxm/gpu_data_mgmt.h"
56 #include "gromacs/nbnxm/nbnxm.h"
57 #include "gromacs/pbcutil/ishift.h"
58 #include "gromacs/timing/gpu_timing.h"
59 #include "gromacs/utility/basedefinitions.h"
60 #include "gromacs/utility/cstringutil.h"
61 #include "gromacs/utility/fatalerror.h"
62 #include "gromacs/utility/real.h"
63 #include "gromacs/utility/smalloc.h"
65 #include "nbnxm_cuda.h"
66 #include "nbnxm_cuda_types.h"
68 /* This is a heuristically determined parameter for the Kepler
69 * and Maxwell architectures for the minimum size of ci lists by multiplying
70 * this constant with the # of multiprocessors on the current device.
71 * Since the maximum number of blocks per multiprocessor is 16, the ideal
72 * count for small systems is 32 or 48 blocks per multiprocessor. Because
73 * there is a bit of fluctuations in the generated block counts, we use
74 * a target of 44 instead of the ideal value of 48.
76 static unsigned int gpu_min_ci_balanced_factor = 44;
79 static void nbnxn_cuda_clear_e_fshift(gmx_nbnxn_cuda_t *nb);
82 static void nbnxn_cuda_free_nbparam_table(cu_nbparam_t *nbparam);
84 /*! \brief Return whether combination rules are used.
86 * \param[in] pointer to nonbonded paramter struct
87 * \return true if combination rules are used in this run, false otherwise
89 static inline bool useLjCombRule(const cu_nbparam_t *nbparam)
91 return (nbparam->vdwtype == evdwCuCUTCOMBGEOM ||
92 nbparam->vdwtype == evdwCuCUTCOMBLB);
95 /*! \brief Initialized the Ewald Coulomb correction GPU table.
97 Tabulates the Ewald Coulomb force and initializes the size/scale
98 and the table GPU array. If called with an already allocated table,
99 it just re-uploads the table.
101 static void init_ewald_coulomb_force_table(const interaction_const_t *ic,
104 if (nbp->coulomb_tab != nullptr)
106 nbnxn_cuda_free_nbparam_table(nbp);
109 nbp->coulomb_tab_scale = ic->tabq_scale;
110 initParamLookupTable(nbp->coulomb_tab, nbp->coulomb_tab_texobj,
111 ic->tabq_coul_F, ic->tabq_size);
115 /*! Initializes the atomdata structure first time, it only gets filled at
117 static void init_atomdata_first(cu_atomdata_t *ad, int ntypes)
122 stat = cudaMalloc((void**)&ad->shift_vec, SHIFTS*sizeof(*ad->shift_vec));
123 CU_RET_ERR(stat, "cudaMalloc failed on ad->shift_vec");
124 ad->bShiftVecUploaded = false;
126 stat = cudaMalloc((void**)&ad->fshift, SHIFTS*sizeof(*ad->fshift));
127 CU_RET_ERR(stat, "cudaMalloc failed on ad->fshift");
129 stat = cudaMalloc((void**)&ad->e_lj, sizeof(*ad->e_lj));
130 CU_RET_ERR(stat, "cudaMalloc failed on ad->e_lj");
131 stat = cudaMalloc((void**)&ad->e_el, sizeof(*ad->e_el));
132 CU_RET_ERR(stat, "cudaMalloc failed on ad->e_el");
134 /* initialize to nullptr poiters to data that is not allocated here and will
135 need reallocation in nbnxn_cuda_init_atomdata */
139 /* size -1 indicates that the respective array hasn't been initialized yet */
144 /*! Selects the Ewald kernel type, analytical on SM 3.0 and later, tabulated on
145 earlier GPUs, single or twin cut-off. */
146 static int pick_ewald_kernel_type(bool bTwinCut)
148 bool bUseAnalyticalEwald, bForceAnalyticalEwald, bForceTabulatedEwald;
151 /* Benchmarking/development environment variables to force the use of
152 analytical or tabulated Ewald kernel. */
153 bForceAnalyticalEwald = (getenv("GMX_CUDA_NB_ANA_EWALD") != nullptr);
154 bForceTabulatedEwald = (getenv("GMX_CUDA_NB_TAB_EWALD") != nullptr);
156 if (bForceAnalyticalEwald && bForceTabulatedEwald)
158 gmx_incons("Both analytical and tabulated Ewald CUDA non-bonded kernels "
159 "requested through environment variables.");
162 /* By default use analytical Ewald. */
163 bUseAnalyticalEwald = true;
164 if (bForceAnalyticalEwald)
168 fprintf(debug, "Using analytical Ewald CUDA kernels\n");
171 else if (bForceTabulatedEwald)
173 bUseAnalyticalEwald = false;
177 fprintf(debug, "Using tabulated Ewald CUDA kernels\n");
181 /* Use twin cut-off kernels if requested by bTwinCut or the env. var.
182 forces it (use it for debugging/benchmarking only). */
183 if (!bTwinCut && (getenv("GMX_CUDA_NB_EWALD_TWINCUT") == nullptr))
185 kernel_type = bUseAnalyticalEwald ? eelCuEWALD_ANA : eelCuEWALD_TAB;
189 kernel_type = bUseAnalyticalEwald ? eelCuEWALD_ANA_TWIN : eelCuEWALD_TAB_TWIN;
195 /*! Copies all parameters related to the cut-off from ic to nbp */
196 static void set_cutoff_parameters(cu_nbparam_t *nbp,
197 const interaction_const_t *ic,
198 const NbnxnListParameters *listParams)
200 nbp->ewald_beta = ic->ewaldcoeff_q;
201 nbp->sh_ewald = ic->sh_ewald;
202 nbp->epsfac = ic->epsfac;
203 nbp->two_k_rf = 2.0 * ic->k_rf;
204 nbp->c_rf = ic->c_rf;
205 nbp->rvdw_sq = ic->rvdw * ic->rvdw;
206 nbp->rcoulomb_sq = ic->rcoulomb * ic->rcoulomb;
207 nbp->rlistOuter_sq = listParams->rlistOuter * listParams->rlistOuter;
208 nbp->rlistInner_sq = listParams->rlistInner * listParams->rlistInner;
209 nbp->useDynamicPruning = listParams->useDynamicPruning;
211 nbp->sh_lj_ewald = ic->sh_lj_ewald;
212 nbp->ewaldcoeff_lj = ic->ewaldcoeff_lj;
214 nbp->rvdw_switch = ic->rvdw_switch;
215 nbp->dispersion_shift = ic->dispersion_shift;
216 nbp->repulsion_shift = ic->repulsion_shift;
217 nbp->vdw_switch = ic->vdw_switch;
220 /*! Initializes the nonbonded parameter data structure. */
221 static void init_nbparam(cu_nbparam_t *nbp,
222 const interaction_const_t *ic,
223 const NbnxnListParameters *listParams,
224 const nbnxn_atomdata_t::Params &nbatParams)
228 ntypes = nbatParams.numTypes;
230 set_cutoff_parameters(nbp, ic, listParams);
232 /* The kernel code supports LJ combination rules (geometric and LB) for
233 * all kernel types, but we only generate useful combination rule kernels.
234 * We currently only use LJ combination rule (geometric and LB) kernels
235 * for plain cut-off LJ. On Maxwell the force only kernels speed up 15%
236 * with PME and 20% with RF, the other kernels speed up about half as much.
237 * For LJ force-switch the geometric rule would give 7% speed-up, but this
238 * combination is rarely used. LJ force-switch with LB rule is more common,
239 * but gives only 1% speed-up.
241 if (ic->vdwtype == evdwCUT)
243 switch (ic->vdw_modifier)
246 case eintmodPOTSHIFT:
247 switch (nbatParams.comb_rule)
250 nbp->vdwtype = evdwCuCUT;
253 nbp->vdwtype = evdwCuCUTCOMBGEOM;
256 nbp->vdwtype = evdwCuCUTCOMBLB;
259 gmx_incons("The requested LJ combination rule is not implemented in the CUDA GPU accelerated kernels!");
262 case eintmodFORCESWITCH:
263 nbp->vdwtype = evdwCuFSWITCH;
265 case eintmodPOTSWITCH:
266 nbp->vdwtype = evdwCuPSWITCH;
269 gmx_incons("The requested VdW interaction modifier is not implemented in the CUDA GPU accelerated kernels!");
272 else if (ic->vdwtype == evdwPME)
274 if (ic->ljpme_comb_rule == ljcrGEOM)
276 assert(nbatParams.comb_rule == ljcrGEOM);
277 nbp->vdwtype = evdwCuEWALDGEOM;
281 assert(nbatParams.comb_rule == ljcrLB);
282 nbp->vdwtype = evdwCuEWALDLB;
287 gmx_incons("The requested VdW type is not implemented in the CUDA GPU accelerated kernels!");
290 if (ic->eeltype == eelCUT)
292 nbp->eeltype = eelCuCUT;
294 else if (EEL_RF(ic->eeltype))
296 nbp->eeltype = eelCuRF;
298 else if ((EEL_PME(ic->eeltype) || ic->eeltype == eelEWALD))
300 /* Initially rcoulomb == rvdw, so it's surely not twin cut-off. */
301 nbp->eeltype = pick_ewald_kernel_type(false);
305 /* Shouldn't happen, as this is checked when choosing Verlet-scheme */
306 gmx_incons("The requested electrostatics type is not implemented in the CUDA GPU accelerated kernels!");
309 /* generate table for PME */
310 nbp->coulomb_tab = nullptr;
311 if (nbp->eeltype == eelCuEWALD_TAB || nbp->eeltype == eelCuEWALD_TAB_TWIN)
313 init_ewald_coulomb_force_table(ic, nbp);
316 /* set up LJ parameter lookup table */
317 if (!useLjCombRule(nbp))
319 initParamLookupTable(nbp->nbfp, nbp->nbfp_texobj,
320 nbatParams.nbfp.data(), 2*ntypes*ntypes);
323 /* set up LJ-PME parameter lookup table */
324 if (ic->vdwtype == evdwPME)
326 initParamLookupTable(nbp->nbfp_comb, nbp->nbfp_comb_texobj,
327 nbatParams.nbfp_comb.data(), 2*ntypes);
331 /*! Re-generate the GPU Ewald force table, resets rlist, and update the
332 * electrostatic type switching to twin cut-off (or back) if needed. */
333 void nbnxn_gpu_pme_loadbal_update_param(const nonbonded_verlet_t *nbv,
334 const interaction_const_t *ic,
335 const NbnxnListParameters *listParams)
337 if (!nbv || nbv->grp[0].kernel_type != nbnxnk8x8x8_GPU)
341 gmx_nbnxn_cuda_t *nb = nbv->gpu_nbv;
342 cu_nbparam_t *nbp = nb->nbparam;
344 set_cutoff_parameters(nbp, ic, listParams);
346 nbp->eeltype = pick_ewald_kernel_type(ic->rcoulomb != ic->rvdw);
348 init_ewald_coulomb_force_table(ic, nb->nbparam);
351 /*! Initializes the pair list data structure. */
352 static void init_plist(cu_plist_t *pl)
354 /* initialize to nullptr pointers to data that is not allocated here and will
355 need reallocation in nbnxn_gpu_init_pairlist */
361 /* size -1 indicates that the respective array hasn't been initialized yet */
368 pl->imask_nalloc = -1;
370 pl->excl_nalloc = -1;
371 pl->haveFreshList = false;
374 /*! Initializes the timer data structure. */
375 static void init_timers(cu_timers_t *t, bool bUseTwoStreams)
377 /* The non-local counters/stream (second in the array) are needed only with DD. */
378 for (int i = 0; i <= (bUseTwoStreams ? 1 : 0); i++)
380 t->didPairlistH2D[i] = false;
381 t->didPrune[i] = false;
382 t->didRollingPrune[i] = false;
386 /*! Initializes the timings data structure. */
387 static void init_timings(gmx_wallclock_gpu_nbnxn_t *t)
396 for (i = 0; i < 2; i++)
398 for (j = 0; j < 2; j++)
400 t->ktime[i][j].t = 0.0;
401 t->ktime[i][j].c = 0;
405 t->pruneTime.t = 0.0;
406 t->dynamicPruneTime.c = 0;
407 t->dynamicPruneTime.t = 0.0;
410 /*! Initializes simulation constant data. */
411 static void nbnxn_cuda_init_const(gmx_nbnxn_cuda_t *nb,
412 const interaction_const_t *ic,
413 const NbnxnListParameters *listParams,
414 const nbnxn_atomdata_t::Params &nbatParams)
416 init_atomdata_first(nb->atdat, nbatParams.numTypes);
417 init_nbparam(nb->nbparam, ic, listParams, nbatParams);
419 /* clear energy and shift force outputs */
420 nbnxn_cuda_clear_e_fshift(nb);
423 void nbnxn_gpu_init(gmx_nbnxn_cuda_t **p_nb,
424 const gmx_device_info_t *deviceInfo,
425 const interaction_const_t *ic,
426 const NbnxnListParameters *listParams,
427 const nbnxn_atomdata_t *nbat,
429 gmx_bool bLocalAndNonlocal)
432 gmx_nbnxn_cuda_t *nb;
441 snew(nb->nbparam, 1);
442 snew(nb->plist[eintLocal], 1);
443 if (bLocalAndNonlocal)
445 snew(nb->plist[eintNonlocal], 1);
448 nb->bUseTwoStreams = bLocalAndNonlocal;
450 nb->timers = new cu_timers_t();
451 snew(nb->timings, 1);
454 pmalloc((void**)&nb->nbst.e_lj, sizeof(*nb->nbst.e_lj));
455 pmalloc((void**)&nb->nbst.e_el, sizeof(*nb->nbst.e_el));
456 pmalloc((void**)&nb->nbst.fshift, SHIFTS * sizeof(*nb->nbst.fshift));
458 init_plist(nb->plist[eintLocal]);
460 /* set device info, just point it to the right GPU among the detected ones */
461 nb->dev_info = deviceInfo;
463 /* local/non-local GPU streams */
464 stat = cudaStreamCreate(&nb->stream[eintLocal]);
465 CU_RET_ERR(stat, "cudaStreamCreate on stream[eintLocal] failed");
466 if (nb->bUseTwoStreams)
468 init_plist(nb->plist[eintNonlocal]);
470 /* Note that the device we're running on does not have to support
471 * priorities, because we are querying the priority range which in this
472 * case will be a single value.
474 int highest_priority;
475 stat = cudaDeviceGetStreamPriorityRange(nullptr, &highest_priority);
476 CU_RET_ERR(stat, "cudaDeviceGetStreamPriorityRange failed");
478 stat = cudaStreamCreateWithPriority(&nb->stream[eintNonlocal],
481 CU_RET_ERR(stat, "cudaStreamCreateWithPriority on stream[eintNonlocal] failed");
484 /* init events for sychronization (timing disabled for performance reasons!) */
485 stat = cudaEventCreateWithFlags(&nb->nonlocal_done, cudaEventDisableTiming);
486 CU_RET_ERR(stat, "cudaEventCreate on nonlocal_done failed");
487 stat = cudaEventCreateWithFlags(&nb->misc_ops_and_local_H2D_done, cudaEventDisableTiming);
488 CU_RET_ERR(stat, "cudaEventCreate on misc_ops_and_local_H2D_done failed");
490 /* WARNING: CUDA timings are incorrect with multiple streams.
491 * This is the main reason why they are disabled by default.
493 // TODO: Consider turning on by default when we can detect nr of streams.
494 nb->bDoTime = (getenv("GMX_ENABLE_GPU_TIMING") != nullptr);
498 init_timers(nb->timers, nb->bUseTwoStreams);
499 init_timings(nb->timings);
502 /* set the kernel type for the current GPU */
503 /* pick L1 cache configuration */
504 nbnxn_cuda_set_cacheconfig();
506 nbnxn_cuda_init_const(nb, ic, listParams, nbat->params());
512 fprintf(debug, "Initialized CUDA data structures.\n");
516 void nbnxn_gpu_init_pairlist(gmx_nbnxn_cuda_t *nb,
517 const NbnxnPairlistGpu *h_plist,
521 bool bDoTime = (nb->bDoTime && !h_plist->sci.empty());
522 cudaStream_t stream = nb->stream[iloc];
523 cu_plist_t *d_plist = nb->plist[iloc];
525 if (d_plist->na_c < 0)
527 d_plist->na_c = h_plist->na_ci;
531 if (d_plist->na_c != h_plist->na_ci)
533 sprintf(sbuf, "In cu_init_plist: the #atoms per cell has changed (from %d to %d)",
534 d_plist->na_c, h_plist->na_ci);
541 nb->timers->pl_h2d[iloc].openTimingRegion(stream);
542 nb->timers->didPairlistH2D[iloc] = true;
545 Context context = nullptr;
547 reallocateDeviceBuffer(&d_plist->sci, h_plist->sci.size(),
548 &d_plist->nsci, &d_plist->sci_nalloc, context);
549 copyToDeviceBuffer(&d_plist->sci, h_plist->sci.data(), 0, h_plist->sci.size(),
550 stream, GpuApiCallBehavior::Async,
551 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
553 reallocateDeviceBuffer(&d_plist->cj4, h_plist->cj4.size(),
554 &d_plist->ncj4, &d_plist->cj4_nalloc, context);
555 copyToDeviceBuffer(&d_plist->cj4, h_plist->cj4.data(), 0, h_plist->cj4.size(),
556 stream, GpuApiCallBehavior::Async,
557 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
559 reallocateDeviceBuffer(&d_plist->imask, h_plist->cj4.size()*c_nbnxnGpuClusterpairSplit,
560 &d_plist->nimask, &d_plist->imask_nalloc, context);
562 reallocateDeviceBuffer(&d_plist->excl, h_plist->excl.size(),
563 &d_plist->nexcl, &d_plist->excl_nalloc, context);
564 copyToDeviceBuffer(&d_plist->excl, h_plist->excl.data(), 0, h_plist->excl.size(),
565 stream, GpuApiCallBehavior::Async,
566 bDoTime ? nb->timers->pl_h2d[iloc].fetchNextEvent() : nullptr);
570 nb->timers->pl_h2d[iloc].closeTimingRegion(stream);
573 /* the next use of thist list we be the first one, so we need to prune */
574 d_plist->haveFreshList = true;
577 void nbnxn_gpu_upload_shiftvec(gmx_nbnxn_cuda_t *nb,
578 const nbnxn_atomdata_t *nbatom)
580 cu_atomdata_t *adat = nb->atdat;
581 cudaStream_t ls = nb->stream[eintLocal];
583 /* only if we have a dynamic box */
584 if (nbatom->bDynamicBox || !adat->bShiftVecUploaded)
586 cu_copy_H2D_async(adat->shift_vec, nbatom->shift_vec.data(),
587 SHIFTS * sizeof(*adat->shift_vec), ls);
588 adat->bShiftVecUploaded = true;
592 /*! Clears the first natoms_clear elements of the GPU nonbonded force output array. */
593 static void nbnxn_cuda_clear_f(gmx_nbnxn_cuda_t *nb, int natoms_clear)
596 cu_atomdata_t *adat = nb->atdat;
597 cudaStream_t ls = nb->stream[eintLocal];
599 stat = cudaMemsetAsync(adat->f, 0, natoms_clear * sizeof(*adat->f), ls);
600 CU_RET_ERR(stat, "cudaMemsetAsync on f falied");
603 /*! Clears nonbonded shift force output array and energy outputs on the GPU. */
604 static void nbnxn_cuda_clear_e_fshift(gmx_nbnxn_cuda_t *nb)
607 cu_atomdata_t *adat = nb->atdat;
608 cudaStream_t ls = nb->stream[eintLocal];
610 stat = cudaMemsetAsync(adat->fshift, 0, SHIFTS * sizeof(*adat->fshift), ls);
611 CU_RET_ERR(stat, "cudaMemsetAsync on fshift falied");
612 stat = cudaMemsetAsync(adat->e_lj, 0, sizeof(*adat->e_lj), ls);
613 CU_RET_ERR(stat, "cudaMemsetAsync on e_lj falied");
614 stat = cudaMemsetAsync(adat->e_el, 0, sizeof(*adat->e_el), ls);
615 CU_RET_ERR(stat, "cudaMemsetAsync on e_el falied");
618 void nbnxn_gpu_clear_outputs(gmx_nbnxn_cuda_t *nb, int flags)
620 nbnxn_cuda_clear_f(nb, nb->atdat->natoms);
621 /* clear shift force array and energies if the outputs were
622 used in the current step */
623 if (flags & GMX_FORCE_VIRIAL)
625 nbnxn_cuda_clear_e_fshift(nb);
629 void nbnxn_gpu_init_atomdata(gmx_nbnxn_cuda_t *nb,
630 const nbnxn_atomdata_t *nbat)
635 bool bDoTime = nb->bDoTime;
636 cu_timers_t *timers = nb->timers;
637 cu_atomdata_t *d_atdat = nb->atdat;
638 cudaStream_t ls = nb->stream[eintLocal];
640 natoms = nbat->numAtoms();
645 /* time async copy */
646 timers->atdat.openTimingRegion(ls);
649 /* need to reallocate if we have to copy more atoms than the amount of space
650 available and only allocate if we haven't initialized yet, i.e d_atdat->natoms == -1 */
651 if (natoms > d_atdat->nalloc)
653 nalloc = over_alloc_small(natoms);
655 /* free up first if the arrays have already been initialized */
656 if (d_atdat->nalloc != -1)
658 freeDeviceBuffer(&d_atdat->f);
659 freeDeviceBuffer(&d_atdat->xq);
660 freeDeviceBuffer(&d_atdat->atom_types);
661 freeDeviceBuffer(&d_atdat->lj_comb);
664 stat = cudaMalloc((void **)&d_atdat->f, nalloc*sizeof(*d_atdat->f));
665 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->f");
666 stat = cudaMalloc((void **)&d_atdat->xq, nalloc*sizeof(*d_atdat->xq));
667 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->xq");
668 if (useLjCombRule(nb->nbparam))
670 stat = cudaMalloc((void **)&d_atdat->lj_comb, nalloc*sizeof(*d_atdat->lj_comb));
671 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->lj_comb");
675 stat = cudaMalloc((void **)&d_atdat->atom_types, nalloc*sizeof(*d_atdat->atom_types));
676 CU_RET_ERR(stat, "cudaMalloc failed on d_atdat->atom_types");
679 d_atdat->nalloc = nalloc;
683 d_atdat->natoms = natoms;
684 d_atdat->natoms_local = nbat->natoms_local;
686 /* need to clear GPU f output if realloc happened */
689 nbnxn_cuda_clear_f(nb, nalloc);
692 if (useLjCombRule(nb->nbparam))
694 cu_copy_H2D_async(d_atdat->lj_comb, nbat->params().lj_comb.data(),
695 natoms*sizeof(*d_atdat->lj_comb), ls);
699 cu_copy_H2D_async(d_atdat->atom_types, nbat->params().type.data(),
700 natoms*sizeof(*d_atdat->atom_types), ls);
705 timers->atdat.closeTimingRegion(ls);
709 static void nbnxn_cuda_free_nbparam_table(cu_nbparam_t *nbparam)
711 if (nbparam->eeltype == eelCuEWALD_TAB || nbparam->eeltype == eelCuEWALD_TAB_TWIN)
713 destroyParamLookupTable(nbparam->coulomb_tab, nbparam->coulomb_tab_texobj);
717 void nbnxn_gpu_free(gmx_nbnxn_cuda_t *nb)
720 cu_atomdata_t *atdat;
721 cu_nbparam_t *nbparam;
729 nbparam = nb->nbparam;
731 nbnxn_cuda_free_nbparam_table(nbparam);
733 stat = cudaEventDestroy(nb->nonlocal_done);
734 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->nonlocal_done");
735 stat = cudaEventDestroy(nb->misc_ops_and_local_H2D_done);
736 CU_RET_ERR(stat, "cudaEventDestroy failed on timers->misc_ops_and_local_H2D_done");
741 /* The non-local counters/stream (second in the array) are needed only with DD. */
742 for (int i = 0; i <= (nb->bUseTwoStreams ? 1 : 0); i++)
744 stat = cudaStreamDestroy(nb->stream[i]);
745 CU_RET_ERR(stat, "cudaStreamDestroy failed on stream");
749 if (!useLjCombRule(nb->nbparam))
751 destroyParamLookupTable(nbparam->nbfp, nbparam->nbfp_texobj);
755 if (nbparam->vdwtype == evdwCuEWALDGEOM || nbparam->vdwtype == evdwCuEWALDLB)
757 destroyParamLookupTable(nbparam->nbfp_comb, nbparam->nbfp_comb_texobj);
760 stat = cudaFree(atdat->shift_vec);
761 CU_RET_ERR(stat, "cudaFree failed on atdat->shift_vec");
762 stat = cudaFree(atdat->fshift);
763 CU_RET_ERR(stat, "cudaFree failed on atdat->fshift");
765 stat = cudaFree(atdat->e_lj);
766 CU_RET_ERR(stat, "cudaFree failed on atdat->e_lj");
767 stat = cudaFree(atdat->e_el);
768 CU_RET_ERR(stat, "cudaFree failed on atdat->e_el");
770 freeDeviceBuffer(&atdat->f);
771 freeDeviceBuffer(&atdat->xq);
772 freeDeviceBuffer(&atdat->atom_types);
773 freeDeviceBuffer(&atdat->lj_comb);
776 auto *plist = nb->plist[eintLocal];
777 freeDeviceBuffer(&plist->sci);
778 freeDeviceBuffer(&plist->cj4);
779 freeDeviceBuffer(&plist->imask);
780 freeDeviceBuffer(&plist->excl);
782 if (nb->bUseTwoStreams)
784 auto *plist_nl = nb->plist[eintNonlocal];
785 freeDeviceBuffer(&plist_nl->sci);
786 freeDeviceBuffer(&plist_nl->cj4);
787 freeDeviceBuffer(&plist_nl->imask);
788 freeDeviceBuffer(&plist_nl->excl);
793 pfree(nb->nbst.e_lj);
794 nb->nbst.e_lj = nullptr;
796 pfree(nb->nbst.e_el);
797 nb->nbst.e_el = nullptr;
799 pfree(nb->nbst.fshift);
800 nb->nbst.fshift = nullptr;
809 fprintf(debug, "Cleaned up CUDA data structures.\n");
813 //! This function is documented in the header file
814 gmx_wallclock_gpu_nbnxn_t *nbnxn_gpu_get_timings(gmx_nbnxn_cuda_t *nb)
816 return (nb != nullptr && nb->bDoTime) ? nb->timings : nullptr;
819 void nbnxn_gpu_reset_timings(nonbonded_verlet_t* nbv)
821 if (nbv->gpu_nbv && nbv->gpu_nbv->bDoTime)
823 init_timings(nbv->gpu_nbv->timings);
827 int nbnxn_gpu_min_ci_balanced(gmx_nbnxn_cuda_t *nb)
829 return nb != nullptr ?
830 gpu_min_ci_balanced_factor*nb->dev_info->prop.multiProcessorCount : 0;
834 gmx_bool nbnxn_gpu_is_kernel_ewald_analytical(const gmx_nbnxn_cuda_t *nb)
836 return ((nb->nbparam->eeltype == eelCuEWALD_ANA) ||
837 (nb->nbparam->eeltype == eelCuEWALD_ANA_TWIN));
840 void *nbnxn_gpu_get_command_stream(gmx_nbnxn_gpu_t *nb,
845 return static_cast<void *>(&nb->stream[iloc]);
848 void *nbnxn_gpu_get_xq(gmx_nbnxn_gpu_t *nb)
852 return static_cast<void *>(nb->atdat->xq);
855 void *nbnxn_gpu_get_f(gmx_nbnxn_gpu_t *nb)
859 return static_cast<void *>(nb->atdat->f);
862 rvec *nbnxn_gpu_get_fshift(gmx_nbnxn_gpu_t *nb)
866 return reinterpret_cast<rvec *>(nb->atdat->fshift);