1 // clang-format off
2 /* ----------------------------------------------------------------------
3 LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
4 https://www.lammps.org/, Sandia National Laboratories
5 Steve Plimpton, sjplimp@sandia.gov
6
7 Copyright (2003) Sandia Corporation. Under the terms of Contract
8 DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
9 certain rights in this software. This software is distributed under
10 the GNU General Public License.
11
12 See the README file in the top-level LAMMPS directory.
13 ------------------------------------------------------------------------- */
14
15 /* ----------------------------------------------------------------------
16 Contributing author: Trung Nguyen (Northwestern)
17 ------------------------------------------------------------------------- */
18
19 #include "pair_lj_expand_coul_long_gpu.h"
20
21 #include "atom.h"
22 #include "domain.h"
23 #include "error.h"
24 #include "force.h"
25 #include "gpu_extra.h"
26 #include "kspace.h"
27 #include "neigh_list.h"
28 #include "neigh_request.h"
29 #include "neighbor.h"
30 #include "suffix.h"
31
32 #include <cmath>
33
34 #define EWALD_F 1.12837917
35 #define EWALD_P 0.3275911
36 #define A1 0.254829592
37 #define A2 -0.284496736
38 #define A3 1.421413741
39 #define A4 -1.453152027
40 #define A5 1.061405429
41
42 using namespace LAMMPS_NS;
43
44 // External functions from cuda library for atom decomposition
45
46 int ljecl_gpu_init(const int ntypes, double **cutsq, double **host_lj1,
47 double **host_lj2, double **host_lj3, double **host_lj4,
48 double **offset, double **shift, double *special_lj,
49 const int nlocal, const int nall, const int max_nbors,
50 const int maxspecial, const double cell_size,
51 int &gpu_mode, FILE *screen, double **host_cut_ljsq,
52 double host_cut_coulsq, double *host_special_coul,
53 const double qqrd2e, const double g_ewald);
54 int ljecl_gpu_reinit(const int ntypes, double **cutsq, double **host_lj1,
55 double **host_lj2, double **host_lj3, double **host_lj4,
56 double **offset, double **shift, double **host_lj_cutsq);
57 void ljecl_gpu_clear();
58 int ** ljecl_gpu_compute_n(const int ago, const int inum,
59 const int nall, double **host_x, int *host_type,
60 double *sublo, double *subhi, tagint *tag,
61 int **nspecial, tagint **special, const bool eflag,
62 const bool vflag, const bool eatom, const bool vatom,
63 int &host_start, int **ilist, int **jnum,
64 const double cpu_time, bool &success, double *host_q,
65 double *boxlo, double *prd);
66 void ljecl_gpu_compute(const int ago, const int inum, const int nall,
67 double **host_x, int *host_type, int *ilist, int *numj,
68 int **firstneigh, const bool eflag, const bool vflag,
69 const bool eatom, const bool vatom, int &host_start,
70 const double cpu_time, bool &success, double *host_q,
71 const int nlocal, double *boxlo, double *prd);
72 double ljecl_gpu_bytes();
73
74 /* ---------------------------------------------------------------------- */
75
PairLJExpandCoulLongGPU(LAMMPS * lmp)76 PairLJExpandCoulLongGPU::PairLJExpandCoulLongGPU(LAMMPS *lmp) :
77 PairLJExpandCoulLong(lmp), gpu_mode(GPU_FORCE)
78 {
79 respa_enable = 0;
80 cpu_time = 0.0;
81 suffix_flag |= Suffix::GPU;
82 GPU_EXTRA::gpu_ready(lmp->modify, lmp->error);
83 }
84
85 /* ----------------------------------------------------------------------
86 free all arrays
87 ------------------------------------------------------------------------- */
88
~PairLJExpandCoulLongGPU()89 PairLJExpandCoulLongGPU::~PairLJExpandCoulLongGPU()
90 {
91 ljecl_gpu_clear();
92 }
93
94 /* ---------------------------------------------------------------------- */
95
compute(int eflag,int vflag)96 void PairLJExpandCoulLongGPU::compute(int eflag, int vflag)
97 {
98 ev_init(eflag,vflag);
99
100 int nall = atom->nlocal + atom->nghost;
101 int inum, host_start;
102
103 bool success = true;
104 int *ilist, *numneigh, **firstneigh;
105 if (gpu_mode != GPU_FORCE) {
106 double sublo[3],subhi[3];
107 if (domain->triclinic == 0) {
108 sublo[0] = domain->sublo[0];
109 sublo[1] = domain->sublo[1];
110 sublo[2] = domain->sublo[2];
111 subhi[0] = domain->subhi[0];
112 subhi[1] = domain->subhi[1];
113 subhi[2] = domain->subhi[2];
114 } else {
115 domain->bbox(domain->sublo_lamda,domain->subhi_lamda,sublo,subhi);
116 }
117 inum = atom->nlocal;
118 firstneigh = ljecl_gpu_compute_n(neighbor->ago, inum, nall, atom->x,
119 atom->type, sublo, subhi,
120 atom->tag, atom->nspecial, atom->special,
121 eflag, vflag, eflag_atom, vflag_atom,
122 host_start, &ilist, &numneigh, cpu_time,
123 success, atom->q, domain->boxlo,
124 domain->prd);
125 } else {
126 inum = list->inum;
127 ilist = list->ilist;
128 numneigh = list->numneigh;
129 firstneigh = list->firstneigh;
130 ljecl_gpu_compute(neighbor->ago, inum, nall, atom->x, atom->type,
131 ilist, numneigh, firstneigh, eflag, vflag, eflag_atom,
132 vflag_atom, host_start, cpu_time, success, atom->q,
133 atom->nlocal, domain->boxlo, domain->prd);
134 }
135 if (!success)
136 error->one(FLERR,"Insufficient memory on accelerator");
137
138 if (host_start<inum) {
139 cpu_time = MPI_Wtime();
140 cpu_compute(host_start, inum, eflag, vflag, ilist, numneigh, firstneigh);
141 cpu_time = MPI_Wtime() - cpu_time;
142 }
143 }
144
145 /* ----------------------------------------------------------------------
146 init specific to this pair style
147 ------------------------------------------------------------------------- */
148
init_style()149 void PairLJExpandCoulLongGPU::init_style()
150 {
151 cut_respa = nullptr;
152
153 if (!atom->q_flag)
154 error->all(FLERR,"Pair style lj/cut/coul/long/gpu requires atom attribute q");
155 if (force->newton_pair)
156 error->all(FLERR,"Pair style lj/cut/coul/long/gpu requires newton pair off");
157
158 // Repeat cutsq calculation because done after call to init_style
159 double maxcut = -1.0;
160 double cut;
161 for (int i = 1; i <= atom->ntypes; i++) {
162 for (int j = i; j <= atom->ntypes; j++) {
163 if (setflag[i][j] != 0 || (setflag[i][i] != 0 && setflag[j][j] != 0)) {
164 cut = init_one(i,j);
165 cut *= cut;
166 if (cut > maxcut)
167 maxcut = cut;
168 cutsq[i][j] = cutsq[j][i] = cut;
169 } else
170 cutsq[i][j] = cutsq[j][i] = 0.0;
171 }
172 }
173 double cell_size = sqrt(maxcut) + neighbor->skin;
174
175 cut_coulsq = cut_coul * cut_coul;
176
177 // insure use of KSpace long-range solver, set g_ewald
178
179 if (force->kspace == nullptr)
180 error->all(FLERR,"Pair style requires a KSpace style");
181 g_ewald = force->kspace->g_ewald;
182
183 // setup force tables
184
185 if (ncoultablebits) init_tables(cut_coul,cut_respa);
186
187 int maxspecial=0;
188 if (atom->molecular != Atom::ATOMIC)
189 maxspecial=atom->maxspecial;
190 int mnf = 5e-2 * neighbor->oneatom;
191 int success = ljecl_gpu_init(atom->ntypes+1, cutsq, lj1, lj2, lj3, lj4,
192 offset, shift, force->special_lj, atom->nlocal,
193 atom->nlocal+atom->nghost, mnf, maxspecial,
194 cell_size, gpu_mode, screen, cut_ljsq, cut_coulsq,
195 force->special_coul, force->qqrd2e, g_ewald);
196 GPU_EXTRA::check_flag(success,error,world);
197
198 if (gpu_mode == GPU_FORCE) {
199 int irequest = neighbor->request(this,instance_me);
200 neighbor->requests[irequest]->half = 0;
201 neighbor->requests[irequest]->full = 1;
202 }
203 }
204
205 /* ---------------------------------------------------------------------- */
206
reinit()207 void PairLJExpandCoulLongGPU::reinit()
208 {
209 Pair::reinit();
210
211 ljecl_gpu_reinit(atom->ntypes+1, cutsq, lj1, lj2, lj3, lj4, offset, shift, cut_ljsq);
212 }
213
214 /* ---------------------------------------------------------------------- */
215
memory_usage()216 double PairLJExpandCoulLongGPU::memory_usage()
217 {
218 double bytes = Pair::memory_usage();
219 return bytes + ljecl_gpu_bytes();
220 }
221
222 /* ---------------------------------------------------------------------- */
223
cpu_compute(int start,int inum,int eflag,int,int * ilist,int * numneigh,int ** firstneigh)224 void PairLJExpandCoulLongGPU::cpu_compute(int start, int inum, int eflag,
225 int /* vflag */, int *ilist,
226 int *numneigh, int **firstneigh)
227 {
228 int i,j,ii,jj,jnum,itype,jtype,itable;
229 double qtmp,xtmp,ytmp,ztmp,delx,dely,delz,evdwl,ecoul,fpair;
230 double fraction,table;
231 double r,r2inv,r6inv,forcecoul,forcelj,factor_coul,factor_lj;
232 double grij,expm2,prefactor,t,erfc;
233 double rsq,rshift,rshiftsq,rshift2inv;
234
235 int *jlist;
236
237 evdwl = ecoul = 0.0;
238
239 double **x = atom->x;
240 double **f = atom->f;
241 double *q = atom->q;
242 int *type = atom->type;
243 double *special_coul = force->special_coul;
244 double *special_lj = force->special_lj;
245 double qqrd2e = force->qqrd2e;
246
247 // loop over neighbors of my atoms
248
249 for (ii = start; ii < inum; ii++) {
250 i = ilist[ii];
251 qtmp = q[i];
252 xtmp = x[i][0];
253 ytmp = x[i][1];
254 ztmp = x[i][2];
255 itype = type[i];
256 jlist = firstneigh[i];
257 jnum = numneigh[i];
258
259 for (jj = 0; jj < jnum; jj++) {
260 j = jlist[jj];
261 factor_lj = special_lj[sbmask(j)];
262 factor_coul = special_coul[sbmask(j)];
263 j &= NEIGHMASK;
264
265 delx = xtmp - x[j][0];
266 dely = ytmp - x[j][1];
267 delz = ztmp - x[j][2];
268 rsq = delx*delx + dely*dely + delz*delz;
269 jtype = type[j];
270
271 if (rsq < cutsq[itype][jtype]) {
272 r2inv = 1.0/rsq;
273
274 if (rsq < cut_coulsq) {
275 if (!ncoultablebits || rsq <= tabinnersq) {
276 r = sqrt(rsq);
277 grij = g_ewald * r;
278 expm2 = exp(-grij*grij);
279 t = 1.0 / (1.0 + EWALD_P*grij);
280 erfc = t * (A1+t*(A2+t*(A3+t*(A4+t*A5)))) * expm2;
281 prefactor = qqrd2e * qtmp*q[j]/r;
282 forcecoul = prefactor * (erfc + EWALD_F*grij*expm2);
283 if (factor_coul < 1.0) forcecoul -= (1.0-factor_coul)*prefactor;
284 } else {
285 union_int_float_t rsq_lookup;
286 rsq_lookup.f = rsq;
287 itable = rsq_lookup.i & ncoulmask;
288 itable >>= ncoulshiftbits;
289 fraction = (rsq_lookup.f - rtable[itable]) * drtable[itable];
290 table = ftable[itable] + fraction*dftable[itable];
291 forcecoul = qtmp*q[j] * table;
292 if (factor_coul < 1.0) {
293 table = ctable[itable] + fraction*dctable[itable];
294 prefactor = qtmp*q[j] * table;
295 forcecoul -= (1.0-factor_coul)*prefactor;
296 }
297 }
298 } else forcecoul = 0.0;
299
300 if (rsq < cut_ljsq[itype][jtype]) {
301 r = sqrt(rsq);
302 rshift = r - shift[itype][jtype];
303 rshiftsq = rshift*rshift;
304 rshift2inv = 1.0/rshiftsq;
305 r6inv = rshift2inv*rshift2inv*rshift2inv;
306 forcelj = r6inv * (lj1[itype][jtype]*r6inv - lj2[itype][jtype]);
307 forcelj = factor_lj*forcelj/rshift/r;
308 } else forcelj = 0.0;
309
310 fpair = forcecoul*r2inv + forcelj;
311
312 f[i][0] += delx*fpair;
313 f[i][1] += dely*fpair;
314 f[i][2] += delz*fpair;
315
316 if (eflag) {
317 if (rsq < cut_coulsq) {
318 if (!ncoultablebits || rsq <= tabinnersq)
319 ecoul = prefactor*erfc;
320 else {
321 table = etable[itable] + fraction*detable[itable];
322 ecoul = qtmp*q[j] * table;
323 }
324 if (factor_coul < 1.0) ecoul -= (1.0-factor_coul)*prefactor;
325 } else ecoul = 0.0;
326
327 if (rsq < cut_ljsq[itype][jtype]) {
328 evdwl = r6inv*(lj3[itype][jtype]*r6inv-lj4[itype][jtype]) -
329 offset[itype][jtype];
330 evdwl *= factor_lj;
331 } else evdwl = 0.0;
332 }
333
334 if (evflag) ev_tally_full(i,evdwl,ecoul,fpair,delx,dely,delz);
335 }
336 }
337 }
338 }
339