?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Introduce SimulatorBuilder
[gromacs.git] / src / gromacs / nbnxm / kernels_reference / kernel_gpu_ref.cpp
bloba23bcc6c145d3c527dca56a9bb2e3d7a50600ab9
1 /*
2 * This file is part of the GROMACS molecular simulation package.
3 *
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.
8 *
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.
13 *
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.
18 *
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.
23 *
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.
31 *
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.
34 */
35 #include "gmxpre.h"
36
37 #include "kernel_gpu_ref.h"
38
39 #include <cmath>
40
41 #include <algorithm>
42
43 #include "gromacs/math/functions.h"
44 #include "gromacs/math/utilities.h"
45 #include "gromacs/math/vec.h"
46 #include "gromacs/mdlib/force_flags.h"
47 #include "gromacs/mdtypes/md_enums.h"
48 #include "gromacs/nbnxm/atomdata.h"
49 #include "gromacs/nbnxm/nbnxm.h"
50 #include "gromacs/nbnxm/pairlist.h"
51 #include "gromacs/pbcutil/ishift.h"
52 #include "gromacs/utility/fatalerror.h"
53
54 static const int c_numClPerSupercl = c_nbnxnGpuNumClusterPerSupercluster;
55 static const int c_clSize = c_nbnxnGpuClusterSize;
56
57 void
58 nbnxn_kernel_gpu_ref(const NbnxnPairlistGpu *nbl,
59 const nbnxn_atomdata_t *nbat,
60 const interaction_const_t *iconst,
61 rvec *shift_vec,
62 int force_flags,
63 int clearF,
64 gmx::ArrayRef<real> f,
65 real * fshift,
66 real * Vc,
67 real * Vvdw)
68 {
69 gmx_bool bEner;
70 gmx_bool bEwald;
71 const real *Ftab = nullptr;
72 real rcut2, rvdw2, rlist2;
73 int ntype;
74 real facel;
75 int ish3;
76 int sci;
77 int cj4_ind0, cj4_ind1, cj4_ind;
78 int ci, cj;
79 int ic, jc, ia, ja, is, ifs, js, jfs, im, jm;
80 int n0;
81 int ggid;
82 real shX, shY, shZ;
83 real fscal, tx, ty, tz;
84 real rinvsq;
85 real iq;
86 real qq, vcoul = 0, krsq, vctot;
87 int nti;
88 int tj;
89 real rt, r, eps;
90 real rinvsix;
91 real Vvdwtot;
92 real Vvdw_rep, Vvdw_disp;
93 real ix, iy, iz, fix, fiy, fiz;
94 real jx, jy, jz;
95 real dx, dy, dz, rsq, rinv;
96 int int_bit;
97 real fexcl;
98 real c6, c12;
99 const nbnxn_excl_t *excl[2];
100
101 int npair_tot, npair;
102 int nhwu, nhwu_pruned;
103
104 if (nbl->na_ci != c_clSize)
105 {
106 gmx_fatal(FARGS, "The neighborlist cluster size in the GPU reference kernel is %d, expected it to be %d", nbl->na_ci, c_clSize);
107 }
108
109 if (clearF == enbvClearFYes)
110 {
111 for (real &elem : f)
112 {
113 elem = 0;
114 }
115 }
116
117 bEner = ((force_flags & GMX_FORCE_ENERGY) != 0);
118
119 bEwald = EEL_FULL(iconst->eeltype);
120 if (bEwald)
121 {
122 Ftab = iconst->coulombEwaldTables->tableF.data();
123 }
124
125 rcut2 = iconst->rcoulomb*iconst->rcoulomb;
126 rvdw2 = iconst->rvdw*iconst->rvdw;
127
128 rlist2 = nbl->rlist*nbl->rlist;
129
130 const int *type = nbat->params().type.data();
131 facel = iconst->epsfac;
132 const real *shiftvec = shift_vec[0];
133 const real *vdwparam = nbat->params().nbfp.data();
134 ntype = nbat->params().numTypes;
135
136 const real *x = nbat->x().data();
137
138 npair_tot = 0;
139 nhwu = 0;
140 nhwu_pruned = 0;
141
142 for (const nbnxn_sci_t &nbln : nbl->sci)
143 {
144 ish3 = 3*nbln.shift;
145 shX = shiftvec[ish3];
146 shY = shiftvec[ish3+1];
147 shZ = shiftvec[ish3+2];
148 cj4_ind0 = nbln.cj4_ind_start;
149 cj4_ind1 = nbln.cj4_ind_end;
150 sci = nbln.sci;
151 vctot = 0;
152 Vvdwtot = 0;
153
154 if (nbln.shift == CENTRAL &&
155 nbl->cj4[cj4_ind0].cj[0] == sci*c_numClPerSupercl)
156 {
157 /* we have the diagonal:
158 * add the charge self interaction energy term
159 */
160 for (im = 0; im < c_numClPerSupercl; im++)
161 {
162 ci = sci*c_numClPerSupercl + im;
163 for (ic = 0; ic < c_clSize; ic++)
164 {
165 ia = ci*c_clSize + ic;
166 iq = x[ia*nbat->xstride+3];
167 vctot += iq*iq;
168 }
169 }
170 if (!bEwald)
171 {
172 vctot *= -facel*0.5*iconst->c_rf;
173 }
174 else
175 {
176 /* last factor 1/sqrt(pi) */
177 vctot *= -facel*iconst->ewaldcoeff_q*M_1_SQRTPI;
178 }
179 }
180
181 for (cj4_ind = cj4_ind0; (cj4_ind < cj4_ind1); cj4_ind++)
182 {
183 excl[0] = &nbl->excl[nbl->cj4[cj4_ind].imei[0].excl_ind];
184 excl[1] = &nbl->excl[nbl->cj4[cj4_ind].imei[1].excl_ind];
185
186 for (jm = 0; jm < c_nbnxnGpuJgroupSize; jm++)
187 {
188 cj = nbl->cj4[cj4_ind].cj[jm];
189
190 for (im = 0; im < c_numClPerSupercl; im++)
191 {
192 /* We're only using the first imask,
193 * but here imei[1].imask is identical.
194 */
195 if ((nbl->cj4[cj4_ind].imei[0].imask >> (jm*c_numClPerSupercl + im)) & 1)
196 {
197 gmx_bool within_rlist;
198
199 ci = sci*c_numClPerSupercl + im;
200
201 within_rlist = FALSE;
202 npair = 0;
203 for (ic = 0; ic < c_clSize; ic++)
204 {
205 ia = ci*c_clSize + ic;
206
207 is = ia*nbat->xstride;
208 ifs = ia*nbat->fstride;
209 ix = shX + x[is+0];
210 iy = shY + x[is+1];
211 iz = shZ + x[is+2];
212 iq = facel*x[is+3];
213 nti = ntype*2*type[ia];
214
215 fix = 0;
216 fiy = 0;
217 fiz = 0;
218
219 for (jc = 0; jc < c_clSize; jc++)
220 {
221 ja = cj*c_clSize + jc;
222
223 if (nbln.shift == CENTRAL &&
224 ci == cj && ja <= ia)
225 {
226 continue;
227 }
228
229 constexpr int clusterPerSplit = c_nbnxnGpuClusterSize/c_nbnxnGpuClusterpairSplit;
230 int_bit = ((excl[jc/clusterPerSplit]->pair[(jc & (clusterPerSplit - 1))*c_clSize + ic]
231 >> (jm*c_numClPerSupercl + im)) & 1);
232
233 js = ja*nbat->xstride;
234 jfs = ja*nbat->fstride;
235 jx = x[js+0];
236 jy = x[js+1];
237 jz = x[js+2];
238 dx = ix - jx;
239 dy = iy - jy;
240 dz = iz - jz;
241 rsq = dx*dx + dy*dy + dz*dz;
242 if (rsq < rlist2)
243 {
244 within_rlist = TRUE;
245 }
246 if (rsq >= rcut2)
247 {
248 continue;
249 }
250
251 if (type[ia] != ntype-1 && type[ja] != ntype-1)
252 {
253 npair++;
254 }
255
256 // Ensure distance do not become so small that r^-12 overflows
257 rsq = std::max(rsq, NBNXN_MIN_RSQ);
258
259 rinv = gmx::invsqrt(rsq);
260 rinvsq = rinv*rinv;
261
262 qq = iq*x[js+3];
263 if (!bEwald)
264 {
265 /* Reaction-field */
266 krsq = iconst->k_rf*rsq;
267 fscal = qq*(int_bit*rinv - 2*krsq)*rinvsq;
268 if (bEner)
269 {
270 vcoul = qq*(int_bit*rinv + krsq - iconst->c_rf);
271 }
272 }
273 else
274 {
275 r = rsq*rinv;
276 rt = r*iconst->coulombEwaldTables->scale;
277 n0 = static_cast<int>(rt);
278 eps = rt - n0;
279
280 fexcl = (1 - eps)*Ftab[n0] + eps*Ftab[n0+1];
281
282 fscal = qq*(int_bit*rinvsq - fexcl)*rinv;
283
284 if (bEner)
285 {
286 vcoul = qq*((int_bit - std::erf(iconst->ewaldcoeff_q*r))*rinv - int_bit*iconst->sh_ewald);
287 }
288 }
289
290 if (rsq < rvdw2)
291 {
292 tj = nti + 2*type[ja];
293
294 /* Vanilla Lennard-Jones cutoff */
295 c6 = vdwparam[tj];
296 c12 = vdwparam[tj+1];
297
298 rinvsix = int_bit*rinvsq*rinvsq*rinvsq;
299 Vvdw_disp = c6*rinvsix;
300 Vvdw_rep = c12*rinvsix*rinvsix;
301 fscal += (Vvdw_rep - Vvdw_disp)*rinvsq;
302
303 if (bEner)
304 {
305 vctot += vcoul;
306
307 Vvdwtot +=
308 (Vvdw_rep + int_bit*c12*iconst->repulsion_shift.cpot)/12 -
309 (Vvdw_disp + int_bit*c6*iconst->dispersion_shift.cpot)/6;
310 }
311 }
312
313 tx = fscal*dx;
314 ty = fscal*dy;
315 tz = fscal*dz;
316 fix = fix + tx;
317 fiy = fiy + ty;
318 fiz = fiz + tz;
319 f[jfs+0] -= tx;
320 f[jfs+1] -= ty;
321 f[jfs+2] -= tz;
322 }
323
324 f[ifs+0] += fix;
325 f[ifs+1] += fiy;
326 f[ifs+2] += fiz;
327 fshift[ish3] = fshift[ish3] + fix;
328 fshift[ish3+1] = fshift[ish3+1] + fiy;
329 fshift[ish3+2] = fshift[ish3+2] + fiz;
330
331 /* Count in half work-units.
332 * In CUDA one work-unit is 2 warps.
333 */
334 if ((ic+1) % (c_clSize/c_nbnxnGpuClusterpairSplit) == 0)
335 {
336 npair_tot += npair;
337
338 nhwu++;
339 if (within_rlist)
340 {
341 nhwu_pruned++;
342 }
343
344 within_rlist = FALSE;
345 npair = 0;
346 }
347 }
348 }
349 }
350 }
351 }
352
353 if (bEner)
354 {
355 ggid = 0;
356 Vc[ggid] = Vc[ggid] + vctot;
357 Vvdw[ggid] = Vvdw[ggid] + Vvdwtot;
358 }
359 }
360
361 if (debug)
362 {
363 fprintf(debug, "number of half %dx%d atom pairs: %d after pruning: %d fraction %4.2f\n",
364 nbl->na_ci, nbl->na_ci,
365 nhwu, nhwu_pruned, nhwu_pruned/static_cast<double>(nhwu));
366 fprintf(debug, "generic kernel pair interactions: %d\n",
367 nhwu*nbl->na_ci/2*nbl->na_ci);
368 fprintf(debug, "generic kernel post-prune pair interactions: %d\n",
369 nhwu_pruned*nbl->na_ci/2*nbl->na_ci);
370 fprintf(debug, "generic kernel non-zero pair interactions: %d\n",
371 npair_tot);
372 fprintf(debug, "ratio non-zero/post-prune pair interactions: %4.2f\n",
373 npair_tot/static_cast<double>(nhwu_pruned*gmx::exactDiv(nbl->na_ci, 2)*nbl->na_ci));
374 }
375 }
The ultimate molecular dynamics simulation package