2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2008, The GROMACS development team.
6 * Copyright (c) 2012,2014, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
43 #include "gromacs/math/vec.h"
44 #include "gromacs/topology/topology.h"
45 #include "gromacs/utility/fatalerror.h"
47 #include "gromacs/legacyheaders/perf_est.h"
48 #include "gromacs/legacyheaders/types/commrec.h"
49 #include "nbnxn_search.h"
50 #include "nbnxn_consts.h"
52 /* Computational cost of bonded, non-bonded and PME calculations.
53 * This will be machine dependent.
54 * The numbers here are accurate for Intel Core2 and AMD Athlon 64
55 * in single precision. In double precision PME mesh is slightly cheaper,
56 * although not so much that the numbers need to be adjusted.
59 /* Cost of a pair interaction in the "group" cut-off scheme */
61 #define C_GR_QLJ_CUT 1.5
62 #define C_GR_QLJ_TAB 2.0
63 #define C_GR_LJ_CUT 1.0
64 #define C_GR_LJ_TAB 1.75
65 /* Cost of 1 water with one Q/LJ atom */
66 #define C_GR_QLJW_CUT 2.0
67 #define C_GR_QLJW_TAB 2.25
68 /* Cost of 1 water with one Q atom or with 1/3 water (LJ negligible) */
71 /* Cost of a pair interaction in the "Verlet" cut-off scheme, QEXP is Ewald */
73 #define C_VT_QRF_LJ 0.40
75 #define C_VT_QEXP_LJ 0.55
76 #define C_VT_QEXP 0.50
77 /* Extra cost for expensive LJ interaction, e.g. pot-switch or LJ-PME */
78 #define C_VT_LJEXP_ADD 0.20
80 /* Cost of PME, with all components running with SSE instructions */
81 /* Cost of particle reordering and redistribution */
82 #define C_PME_REDIST 12.0
83 /* Cost of q spreading and force interpolation per charge (mainly memory) */
84 #define C_PME_SPREAD 0.30
85 /* Cost of fft's, will be multiplied with N log(N) */
86 #define C_PME_FFT 0.20
87 /* Cost of pme_solve, will be multiplied with N */
88 #define C_PME_SOLVE 0.50
90 /* Cost of a bonded interaction divided by the number of (pbc_)dx nrequired */
93 int n_bonded_dx(gmx_mtop_t *mtop, gmx_bool bExcl)
95 int mb, nmol, ftype, ndxb, ndx_excl;
99 /* Count the number of pbc_rvec_sub calls required for bonded interactions.
100 * This number is also roughly proportional to the computational cost.
104 #if __ICC == 1400 || __ICL == 1400
105 #pragma novector /* Work-around for incorrect vectorization */
107 for (mb = 0; mb < mtop->nmolblock; mb++)
109 molt = &mtop->moltype[mtop->molblock[mb].type];
110 nmol = mtop->molblock[mb].nmol;
111 for (ftype = 0; ftype < F_NRE; ftype++)
113 if (interaction_function[ftype].flags & IF_BOND)
118 case F_FBPOSRES: ndxb = 1; break;
119 case F_CONNBONDS: ndxb = 0; break;
120 default: ndxb = NRAL(ftype) - 1; break;
122 ndx += nmol*ndxb*molt->ilist[ftype].nr/(1 + NRAL(ftype));
127 ndx_excl += nmol*(molt->excls.nra - molt->atoms.nr)/2;
137 fprintf(debug, "ndx bonded %d exclusions %d\n", ndx, ndx_excl);
145 static void pp_group_load(gmx_mtop_t *mtop, t_inputrec *ir, matrix box,
146 int *nq_tot, int *nlj_tot,
148 gmx_bool *bChargePerturbed, gmx_bool *bTypePerturbed)
151 int mb, nmol, atnr, cg, a, a0, ncqlj, ncq, nclj;
152 gmx_bool bBHAM, bLJcut, bWater, bQ, bLJ;
153 int nw, nqlj, nq, nlj;
154 float fq, fqlj, flj, fljtab, fqljw, fqw;
158 bBHAM = (mtop->ffparams.functype[0] == F_BHAM);
160 bLJcut = ((ir->vdwtype == evdwCUT) && !bBHAM);
162 /* Computational cost of bonded, non-bonded and PME calculations.
163 * This will be machine dependent.
164 * The numbers here are accurate for Intel Core2 and AMD Athlon 64
165 * in single precision. In double precision PME mesh is slightly cheaper,
166 * although not so much that the numbers need to be adjusted.
169 fqlj = (bLJcut ? C_GR_QLJ_CUT : C_GR_QLJ_TAB);
170 flj = (bLJcut ? C_GR_LJ_CUT : C_GR_LJ_TAB);
171 /* Cost of 1 water with one Q/LJ atom */
172 fqljw = (bLJcut ? C_GR_QLJW_CUT : C_GR_QLJW_TAB);
173 /* Cost of 1 water with one Q atom or with 1/3 water (LJ negligible) */
176 iparams = mtop->ffparams.iparams;
177 atnr = mtop->ffparams.atnr;
182 *bChargePerturbed = FALSE;
183 for (mb = 0; mb < mtop->nmolblock; mb++)
185 molt = &mtop->moltype[mtop->molblock[mb].type];
186 atom = molt->atoms.atom;
187 nmol = mtop->molblock[mb].nmol;
189 for (cg = 0; cg < molt->cgs.nr; cg++)
196 while (a < molt->cgs.index[cg+1])
198 bQ = (atom[a].q != 0 || atom[a].qB != 0);
199 bLJ = (iparams[(atnr+1)*atom[a].type].lj.c6 != 0 ||
200 iparams[(atnr+1)*atom[a].type].lj.c12 != 0);
201 if (atom[a].q != atom[a].qB)
203 *bChargePerturbed = TRUE;
205 if (atom[a].type != atom[a].typeB)
207 *bTypePerturbed = TRUE;
209 /* This if this atom fits into water optimization */
210 if (!((a == a0 && bQ && bLJ) ||
211 (a == a0+1 && bQ && !bLJ) ||
212 (a == a0+2 && bQ && !bLJ && atom[a].q == atom[a-1].q) ||
213 (a == a0+3 && !bQ && bLJ)))
247 *nq_tot = nq + nqlj + nw*3;
248 *nlj_tot = nlj + nqlj + nw;
252 fprintf(debug, "nw %d nqlj %d nq %d nlj %d\n", nw, nqlj, nq, nlj);
255 /* For the PP non-bonded cost it is (unrealistically) assumed
256 * that all atoms are distributed homogeneously in space.
257 * Factor 3 is used because a water molecule has 3 atoms
258 * (and TIP4P effectively has 3 interactions with (water) atoms)).
260 *cost_pp = 0.5*(fqljw*nw*nqlj +
261 fqw *nw*(3*nw + nq) +
263 fq *nq*(3*nw + nqlj + nq) +
264 flj *nlj*(nw + nqlj + nlj))
265 *4/3*M_PI*ir->rlist*ir->rlist*ir->rlist/det(box);
268 static void pp_verlet_load(gmx_mtop_t *mtop, t_inputrec *ir, matrix box,
269 int *nq_tot, int *nlj_tot,
271 gmx_bool *bChargePerturbed, gmx_bool *bTypePerturbed)
274 int mb, nmol, atnr, cg, a, a0, nqlj, nq, nlj;
279 double c_qlj, c_q, c_lj;
281 /* Conversion factor for reference vs SIMD kernel performance.
282 * The factor is about right for SSE2/4, but should be 2 higher for AVX256.
285 const real nbnxn_refkernel_fac = 4.0;
287 const real nbnxn_refkernel_fac = 8.0;
290 bQRF = (EEL_RF(ir->coulombtype) || ir->coulombtype == eelCUT);
292 iparams = mtop->ffparams.iparams;
293 atnr = mtop->ffparams.atnr;
296 *bChargePerturbed = FALSE;
297 for (mb = 0; mb < mtop->nmolblock; mb++)
299 molt = &mtop->moltype[mtop->molblock[mb].type];
300 atom = molt->atoms.atom;
301 nmol = mtop->molblock[mb].nmol;
303 for (a = 0; a < molt->atoms.nr; a++)
305 if (atom[a].q != 0 || atom[a].qB != 0)
307 if (iparams[(atnr+1)*atom[a].type].lj.c6 != 0 ||
308 iparams[(atnr+1)*atom[a].type].lj.c12 != 0)
317 if (atom[a].q != atom[a].qB)
319 *bChargePerturbed = TRUE;
321 if (atom[a].type != atom[a].typeB)
323 *bTypePerturbed = TRUE;
328 nlj = mtop->natoms - nqlj - nq;
331 *nlj_tot = nqlj + nlj;
333 /* Effective cut-off for cluster pair list of 4x4 atoms */
334 r_eff = ir->rlist + nbnxn_get_rlist_effective_inc(NBNXN_CPU_CLUSTER_I_SIZE, mtop->natoms/det(box));
338 fprintf(debug, "nqlj %d nq %d nlj %d rlist %.3f r_eff %.3f\n",
339 nqlj, nq, nlj, ir->rlist, r_eff);
342 /* Determine the cost per pair interaction */
343 c_qlj = (bQRF ? C_VT_QRF_LJ : C_VT_QEXP_LJ);
344 c_q = (bQRF ? C_VT_QRF : C_VT_QEXP);
346 if (ir->vdw_modifier == eintmodPOTSWITCH || EVDW_PME(ir->vdwtype))
348 c_qlj += C_VT_LJEXP_ADD;
349 c_lj += C_VT_LJEXP_ADD;
351 if (EVDW_PME(ir->vdwtype) && ir->ljpme_combination_rule == eljpmeLB)
353 /* We don't have LJ-PME LB comb. rule kernels, we use slow kernels */
354 c_qlj *= nbnxn_refkernel_fac;
355 c_q *= nbnxn_refkernel_fac;
356 c_lj *= nbnxn_refkernel_fac;
359 /* For the PP non-bonded cost it is (unrealistically) assumed
360 * that all atoms are distributed homogeneously in space.
362 /* Convert mtop->natoms to double to avoid int overflow */
364 *cost_pp = 0.5*nat*(nqlj*c_qlj + nq*c_q + nlj*c_lj)
365 *4/3*M_PI*r_eff*r_eff*r_eff/det(box);
368 float pme_load_estimate(gmx_mtop_t *mtop, t_inputrec *ir, matrix box)
371 int mb, nmol, atnr, cg, a, a0, nq_tot, nlj_tot, f;
372 gmx_bool bBHAM, bLJcut, bChargePerturbed, bTypePerturbed;
373 gmx_bool bWater, bQ, bLJ;
374 double cost_bond, cost_pp, cost_redist, cost_spread, cost_fft, cost_solve, cost_pme;
379 /* Computational cost of bonded, non-bonded and PME calculations.
380 * This will be machine dependent.
381 * The numbers here are accurate for Intel Core2 and AMD Athlon 64
382 * in single precision. In double precision PME mesh is slightly cheaper,
383 * although not so much that the numbers need to be adjusted.
386 iparams = mtop->ffparams.iparams;
387 atnr = mtop->ffparams.atnr;
389 cost_bond = C_BOND*n_bonded_dx(mtop, TRUE);
391 if (ir->cutoff_scheme == ecutsGROUP)
393 pp_group_load(mtop, ir, box,
394 &nq_tot, &nlj_tot, &cost_pp,
395 &bChargePerturbed, &bTypePerturbed);
399 pp_verlet_load(mtop, ir, box,
400 &nq_tot, &nlj_tot, &cost_pp,
401 &bChargePerturbed, &bTypePerturbed);
409 if (EEL_PME(ir->coulombtype))
411 f = ((ir->efep != efepNO && bChargePerturbed) ? 2 : 1);
412 cost_redist += C_PME_REDIST*nq_tot;
413 cost_spread += f*C_PME_SPREAD*nq_tot*pow(ir->pme_order, 3);
414 cost_fft += f*C_PME_FFT*ir->nkx*ir->nky*ir->nkz*log(ir->nkx*ir->nky*ir->nkz);
415 cost_solve += f*C_PME_SOLVE*ir->nkx*ir->nky*ir->nkz;
418 if (EVDW_PME(ir->vdwtype))
420 f = ((ir->efep != efepNO && bTypePerturbed) ? 2 : 1);
421 if (ir->ljpme_combination_rule == eljpmeLB)
423 /* LB combination rule: we have 7 mesh terms */
426 cost_redist += C_PME_REDIST*nlj_tot;
427 cost_spread += f*C_PME_SPREAD*nlj_tot*pow(ir->pme_order, 3);
428 cost_fft += f*C_PME_FFT*ir->nkx*ir->nky*ir->nkz*log(ir->nkx*ir->nky*ir->nkz);
429 cost_solve += f*C_PME_SOLVE*ir->nkx*ir->nky*ir->nkz;
432 cost_pme = cost_redist + cost_spread + cost_fft + cost_solve;
434 ratio = cost_pme/(cost_bond + cost_pp + cost_pme);
445 cost_bond, cost_pp, cost_redist, cost_spread, cost_fft, cost_solve);
447 fprintf(debug, "Estimate for relative PME load: %.3f\n", ratio);