Files
lammps/src/min_cg.cpp

650 lines
18 KiB
C++

/* ----------------------------------------------------------------------
LAMMPS - Large-scale Atomic/Molecular Massively Parallel Simulator
http://lammps.sandia.gov, Sandia National Laboratories
Steve Plimpton, sjplimp@sandia.gov
Copyright (2003) Sandia Corporation. Under the terms of Contract
DE-AC04-94AL85000 with Sandia Corporation, the U.S. Government retains
certain rights in this software. This software is distributed under
the GNU General Public License.
See the README file in the top-level LAMMPS directory.
------------------------------------------------------------------------- */
/* ----------------------------------------------------------------------
Sources: Numerical Recipes frprmn routine
"Conjugate Gradient Method Without the Agonizing Pain" by
JR Shewchuk, http://www-2.cs.cmu.edu/~jrs/jrspapers.html#cg
------------------------------------------------------------------------- */
#include "math.h"
#include "string.h"
#include "mpi.h"
#include "min_cg.h"
#include "neighbor.h"
#include "domain.h"
#include "comm.h"
#include "atom.h"
#include "force.h"
#include "pair.h"
#include "bond.h"
#include "angle.h"
#include "dihedral.h"
#include "improper.h"
#include "kspace.h"
#include "output.h"
#include "thermo.h"
#include "update.h"
#include "modify.h"
#include "compute.h"
#include "fix_minimize.h"
#include "thermo.h"
#include "timer.h"
#include "memory.h"
#include "error.h"
using namespace LAMMPS_NS;
#define MIN(A,B) ((A) < (B)) ? (A) : (B)
#define MAX(A,B) ((A) > (B)) ? (A) : (B)
#define EPS 1.0e-6
#define SCAN_FACTOR 2.0
#define SECANT_EPS 1.0e-3
#define SCAN 0 // same as in min.cpp
#define SECANT 1
/* ---------------------------------------------------------------------- */
MinCG::MinCG(LAMMPS *lmp) : Min(lmp) {}
/* ---------------------------------------------------------------------- */
void MinCG::init()
{
// create fix needed for storing atom-based gradient vectors
// will delete it at end of run
char **fixarg = new char*[3];
fixarg[0] = (char *) "MINIMIZE";
fixarg[1] = (char *) "all";
fixarg[2] = (char *) "MINIMIZE";
modify->add_fix(3,fixarg);
delete [] fixarg;
fix_minimize = (FixMinimize *) modify->fix[modify->nfix-1];
// zero gradient vectors before first atom exchange
setup_vectors();
for (int i = 0; i < ndof; i++) h[i] = g[i] = 0.0;
// virial_style:
// 1 if computed explicitly by pair->compute via sum over pair interactions
// 2 if computed implicitly by pair->virial_compute via sum over ghost atoms
if (force->newton_pair) virial_style = 2;
else virial_style = 1;
// setup lists of computes for global and per-atom PE and pressure
ev_setup();
// set flags for what arrays to clear in force_clear()
// need to clear torques if array exists
torqueflag = 0;
if (atom->torque) torqueflag = 1;
// orthogonal vs triclinic simulation box
triclinic = domain->triclinic;
// reset reneighboring criteria if necessary
neigh_every = neighbor->every;
neigh_delay = neighbor->delay;
neigh_dist_check = neighbor->dist_check;
if (neigh_every != 1 || neigh_delay != 0 || neigh_dist_check != 1) {
if (comm->me == 0)
error->warning("Resetting reneighboring criteria during minimization");
}
neighbor->every = 1;
neighbor->delay = 0;
neighbor->dist_check = 1;
// set ptr to linemin function
if (linestyle == SCAN) linemin = &MinCG::linemin_scan;
else if (linestyle == SECANT) linemin = &MinCG::linemin_secant;
}
/* ----------------------------------------------------------------------
perform minimization, with setup first
------------------------------------------------------------------------- */
void MinCG::run()
{
double tmp,*f;
// set initial force & energy
// normalize energy if thermo PE does
setup();
setup_vectors();
int id = modify->find_compute("thermo_pe");
if (id < 0) error->all("Minimization could not find thermo_pe compute");
pe_compute = modify->compute[id];
ecurrent = pe_compute->compute_scalar();
if (output->thermo->normflag) ecurrent /= atom->natoms;
// stats for Finish to print
einitial = ecurrent;
f = atom->f[0];
tmp = 0.0;
for (int i = 0; i < ndof; i++) tmp += f[i]*f[i];
MPI_Allreduce(&tmp,&gnorm2_init,1,MPI_DOUBLE,MPI_SUM,world);
gnorm2_init = sqrt(gnorm2_init);
tmp = 0.0;
for (int i = 0; i < ndof; i++) tmp = MAX(fabs(f[i]),tmp);
MPI_Allreduce(&tmp,&gnorminf_init,1,MPI_DOUBLE,MPI_MAX,world);
// minimizer iterations
timer->barrier_start(TIME_LOOP);
iterate(update->nsteps);
// account for early exit from iterate loop due to convergence
// set niter/nsteps for Finish stats to print
// set output->next values to this timestep
// call eng_force to insure vflag is set when forces computed
// output->write does final output for thermo, dump, restart files
// add ntimestep to ALL computes that store invocation times
// since just hardwired call to thermo/dumps and they may not be ready
if (niter < update->nsteps) {
niter++;
update->nsteps = niter;
for (int idump = 0; idump < output->ndump; idump++)
output->next_dump[idump] = update->ntimestep;
output->next_dump_any = update->ntimestep;
if (output->restart_every) output->next_restart = update->ntimestep;
output->next_thermo = update->ntimestep;
modify->addstep_compute_all(update->ntimestep);
int ntmp;
double *xtmp,*htmp,etmp;
eng_force(&ntmp,&xtmp,&htmp,&etmp);
output->write(update->ntimestep);
}
timer->barrier_stop(TIME_LOOP);
// delete fix at end of run, so its atom arrays won't persist
modify->delete_fix("MINIMIZE");
// reset reneighboring criteria
neighbor->every = neigh_every;
neighbor->delay = neigh_delay;
neighbor->dist_check = neigh_dist_check;
// stats for Finish to print
efinal = ecurrent;
f = atom->f[0];
tmp = 0.0;
for (int i = 0; i < ndof; i++) tmp += f[i]*f[i];
MPI_Allreduce(&tmp,&gnorm2_final,1,MPI_DOUBLE,MPI_SUM,world);
gnorm2_final = sqrt(gnorm2_final);
tmp = 0.0;
for (int i = 0; i < ndof; i++) tmp = MAX(fabs(f[i]),tmp);
MPI_Allreduce(&tmp,&gnorminf_final,1,MPI_DOUBLE,MPI_MAX,world);
}
/* ----------------------------------------------------------------------
setup before run
------------------------------------------------------------------------- */
void MinCG::setup()
{
if (comm->me == 0 && screen) fprintf(screen,"Setting up minimization ...\n");
// setup domain, communication and neighboring
// acquire ghosts
// build neighbor lists
// reset gradient vector ptrs
if (triclinic) domain->x2lamda(atom->nlocal);
domain->pbc();
domain->reset_box();
comm->setup();
if (neighbor->style) neighbor->setup_bins();
comm->exchange();
comm->borders();
if (triclinic) domain->lamda2x(atom->nlocal+atom->nghost);
neighbor->build();
neighbor->ncalls = 0;
setup_vectors();
// compute all forces
ev_set(update->ntimestep);
force_clear();
if (force->pair) force->pair->compute(eflag,vflag);
if (atom->molecular) {
if (force->bond) force->bond->compute(eflag,vflag);
if (force->angle) force->angle->compute(eflag,vflag);
if (force->dihedral) force->dihedral->compute(eflag,vflag);
if (force->improper) force->improper->compute(eflag,vflag);
}
if (force->kspace) {
force->kspace->setup();
force->kspace->compute(eflag,vflag);
}
if (force->newton) comm->reverse_communicate();
modify->setup();
output->setup(1);
}
/* ----------------------------------------------------------------------
minimization via conjugate gradient iterations
Polak-Ribiere formulation
------------------------------------------------------------------------- */
void MinCG::iterate(int n)
{
int i,gradsearch,fail,ntimestep;
double alpha,beta,gg,dot[2],dotall[2];
double *f;
f = atom->f[0];
for (int i = 0; i < ndof; i++) h[i] = g[i] = f[i];
dot[0] = 0.0;
for (i = 0; i < ndof; i++) dot[0] += f[i]*f[i];
MPI_Allreduce(dot,&gg,1,MPI_DOUBLE,MPI_SUM,world);
neval = 0;
gradsearch = 1;
for (niter = 0; niter < n; niter++) {
ntimestep = ++update->ntimestep;
// line minimization along direction h from current atom->x
eprevious = ecurrent;
fail = (this->*linemin)(ndof,atom->x[0],h,ecurrent,dmin,dmax,alpha,neval);
// if max_eval exceeded, all done
// if linemin failed or energy did not decrease sufficiently:
// if searched in grad direction, then all done
// else force next search to be in grad direction (CG restart)
if (neval >= update->max_eval) break;
if (fail || fabs(ecurrent-eprevious) <=
update->tolerance * 0.5*(fabs(ecurrent) + fabs(eprevious) + EPS)) {
if (gradsearch == 1) break;
gradsearch = -1;
}
// update h from new f = -Grad(x) and old g
// old g,h must have migrated with atoms to do this correctly
// done if size sq of grad vector < EPS
// force new search dir to be grad dir if need to restart CG
// set gradsearch to 1 if will search in grad dir on next iteration
f = atom->f[0];
dot[0] = dot[1] = 0.0;
for (i = 0; i < ndof; i++) {
dot[0] += f[i]*f[i];
dot[1] += f[i]*g[i];
}
MPI_Allreduce(dot,dotall,2,MPI_DOUBLE,MPI_SUM,world);
beta = MAX(0.0,(dotall[0] - dotall[1])/gg);
gg = dotall[0];
if (gg < EPS) break;
if (gradsearch == -1) beta = 0.0;
if (beta == 0.0) gradsearch = 1;
else gradsearch = 0;
for (i = 0; i < ndof; i++) {
g[i] = f[i];
h[i] = g[i] + beta*h[i];
}
// output for thermo, dump, restart files
if (output->next == ntimestep) {
timer->stamp();
output->write(ntimestep);
timer->stamp(TIME_OUTPUT);
}
}
}
/* ----------------------------------------------------------------------
set ndof and vector pointers after atoms have migrated
------------------------------------------------------------------------- */
void MinCG::setup_vectors()
{
ndof = 3 * atom->nlocal;
if (ndof) g = fix_minimize->gradient[0];
else g = NULL;
if (ndof) h = fix_minimize->searchdir[0];
else h = NULL;
}
/* ----------------------------------------------------------------------
evaluate potential energy and forces
may migrate atoms
new energy stored in ecurrent and returned (in case caller not in class)
negative gradient will be stored in atom->f
------------------------------------------------------------------------- */
void MinCG::eng_force(int *pndof, double **px, double **ph, double *peng)
{
// check for reneighboring
// always communicate since minimizer moved atoms
// if reneighbor, have to setup_vectors() since atoms migrated
int nflag = neighbor->decide();
if (nflag == 0) {
timer->stamp();
comm->communicate();
timer->stamp(TIME_COMM);
} else {
if (triclinic) domain->x2lamda(atom->nlocal);
domain->pbc();
if (domain->box_change) {
domain->reset_box();
comm->setup();
if (neighbor->style) neighbor->setup_bins();
}
timer->stamp();
comm->exchange();
comm->borders();
if (triclinic) domain->lamda2x(atom->nlocal+atom->nghost);
timer->stamp(TIME_COMM);
neighbor->build();
timer->stamp(TIME_NEIGHBOR);
setup_vectors();
}
ev_set(update->ntimestep);
force_clear();
timer->stamp();
if (force->pair) {
force->pair->compute(eflag,vflag);
timer->stamp(TIME_PAIR);
}
if (atom->molecular) {
if (force->bond) force->bond->compute(eflag,vflag);
if (force->angle) force->angle->compute(eflag,vflag);
if (force->dihedral) force->dihedral->compute(eflag,vflag);
if (force->improper) force->improper->compute(eflag,vflag);
timer->stamp(TIME_BOND);
}
if (force->kspace) {
force->kspace->compute(eflag,vflag);
timer->stamp(TIME_KSPACE);
}
if (force->newton) {
comm->reverse_communicate();
timer->stamp(TIME_COMM);
}
// fixes that affect minimization
if (modify->n_min_post_force) modify->min_post_force(vflag);
// compute potential energy of system
// normalize if thermo PE does
ecurrent = pe_compute->compute_scalar();
if (output->thermo->normflag) ecurrent /= atom->natoms;
// return updated ptrs to caller since atoms may have migrated
*pndof = ndof;
*px = atom->x[0];
*ph = h;
*peng = ecurrent;
}
/* ----------------------------------------------------------------------
clear force on own & ghost atoms
setup and clear other arrays as needed
------------------------------------------------------------------------- */
void MinCG::force_clear()
{
int i;
// clear global force array
// nall includes ghosts only if either newton flag is set
int nall;
if (force->newton) nall = atom->nlocal + atom->nghost;
else nall = atom->nlocal;
double **f = atom->f;
for (i = 0; i < nall; i++) {
f[i][0] = 0.0;
f[i][1] = 0.0;
f[i][2] = 0.0;
}
if (torqueflag) {
double **torque = atom->torque;
for (i = 0; i < nall; i++) {
torque[i][0] = 0.0;
torque[i][1] = 0.0;
torque[i][2] = 0.0;
}
}
}
/* ----------------------------------------------------------------------
line minimization methods
find minimum-energy starting at x along dir direction
input: n = # of degrees of freedom on this proc
x = ptr to atom->x[0] as vector
dir = search direction as vector
eng = current energy at initial x
min/max dist = min/max distance to move any atom coord
output: return 0 if successful move, set alpha
return 1 if failed, no move, no need to set alpha
alpha = distance moved along dir to set x to min-eng config
caller has several quantities set via last call to eng_force()
INSURE last call to eng_force() is consistent with returns
if fail, eng_force() of original x
if succeed, eng_force() at x + alpha*dir
atom->x = coords at new configuration
atom->f = force (-Grad) is evaulated at new configuration
ecurrent = energy of new configuration
NOTE: when call eng_force: n,x,dir,eng may change due to atom migration
updated values are returned by eng_force()
these routines CANNOT store atom-based quantities b/c of migration
------------------------------------------------------------------------- */
/* ----------------------------------------------------------------------
linemin: scan forward by larger and larger steps (SCAN_FACTOR)
uses no gradient info, but should be very robust
start at mindist, continue until maxdist
quit as soon as energy starts to rise
------------------------------------------------------------------------- */
int MinCG::linemin_scan(int n, double *x, double *dir, double eng,
double mindist, double maxdist,
double &alpha, int &nfunc)
{
int i;
double fmax,fme,elowest,alphamin,alphamax,alphalast;
// alphamin = step that moves some atom coord by mindist
// alphamax = step that moves some atom coord by maxdist
fme = 0.0;
for (i = 0; i < n; i++) fme = MAX(fme,fabs(dir[i]));
MPI_Allreduce(&fme,&fmax,1,MPI_DOUBLE,MPI_MAX,world);
if (fmax == 0.0) return 1;
alphamin = mindist/fmax;
alphamax = maxdist/fmax;
// if minstep is already uphill, fail
// if eng increases, stop and return previous alpha
// if alphamax, stop and return alphamax
elowest = eng;
alpha = alphamin;
while (1) {
for (i = 0; i < n; i++) x[i] += alpha*dir[i];
eng_force(&n,&x,&dir,&eng);
nfunc++;
if (alpha == alphamin && eng >= elowest) {
for (i = 0; i < n; i++) x[i] -= alpha*dir[i];
eng_force(&n,&x,&dir,&eng);
nfunc++;
return 1;
}
if (eng > elowest) {
for (i = 0; i < n; i++) x[i] += (alphalast-alpha)*dir[i];
eng_force(&n,&x,&dir,&eng);
nfunc++;
alpha = alphalast;
return 0;
}
if (alpha == alphamax) return 0;
elowest = eng;
alphalast = alpha;
alpha *= SCAN_FACTOR;
if (alpha > alphamax) alpha = alphamax;
}
}
/* ----------------------------------------------------------------------
linemin: use secant approximation to estimate parabola minimum at each step
should converge more quickly/accurately than "scan", but can be less robust
------------------------------------------------------------------------- */
int MinCG::linemin_secant(int n, double *x, double *dir, double eng,
double mindist, double maxdist,
double &alpha, int &nfunc)
{
int i,iter;
double eta,eta_prev,alphamin,alphamax,alphadelta,fme,fmax,dsq,e0,tmp;
double *f;
double epssq = SECANT_EPS * SECANT_EPS;
// stopping criterion for secant iterations
fme = 0.0;
for (i = 0; i < n; i++) fme += dir[i]*dir[i];
MPI_Allreduce(&fme,&dsq,1,MPI_DOUBLE,MPI_SUM,world);
// alphamin = smallest allowed step of mindist
// alphamax = largest allowed step (in single iteration) of maxdist
fme = 0.0;
for (i = 0; i < n; i++) fme = MAX(fme,fabs(dir[i]));
MPI_Allreduce(&fme,&fmax,1,MPI_DOUBLE,MPI_MAX,world);
if (fmax == 0.0) return 1;
alphamin = mindist/fmax;
alphamax = maxdist/fmax;
// eval func at alphamin
// exit if minstep is already uphill
e0 = eng;
for (i = 0; i < n; i++) x[i] += alphamin*dir[i];
eng_force(&n,&x,&dir,&eng);
nfunc++;
if (eng >= e0) {
for (i = 0; i < n; i++) x[i] -= alphamin*dir[i];
eng_force(&n,&x,&dir,&eng);
nfunc++;
return 1;
}
// secant iterations
// alphadelta = new increment to move, alpha = accumulated move
// first step is alpha = 0, first previous step is at mindist
// prevent func evals for alpha outside mindist to maxdist
// if happens on 1st iteration and alpha < mindist
// secant approx is likely searching
// for a maximum (negative alpha), so reevaluate at alphamin
// if happens on 1st iteration and alpha > maxdist
// wants to take big step, so reevaluate at alphamax
f = atom->f[0];
tmp = 0.0;
for (i = 0; i < n; i++) tmp -= f[i]*dir[i];
MPI_Allreduce(&tmp,&eta_prev,1,MPI_DOUBLE,MPI_SUM,world);
alpha = alphamin;
alphadelta = -alphamin;
for (iter = 0; iter < lineiter; iter++) {
alpha += alphadelta;
for (i = 0; i < n; i++) x[i] += alphadelta*dir[i];
eng_force(&n,&x,&dir,&eng);
nfunc++;
f = atom->f[0];
tmp = 0.0;
for (i = 0; i < n; i++) tmp -= f[i]*dir[i];
MPI_Allreduce(&tmp,&eta,1,MPI_DOUBLE,MPI_SUM,world);
alphadelta *= eta / (eta_prev - eta);
eta_prev = eta;
if (alphadelta*alphadelta*dsq <= epssq) break;
if (alpha+alphadelta < alphamin || alpha+alphadelta > alphamax) {
if (iter == 0) {
if (alpha+alphadelta < alphamin) alpha = alphamin;
else alpha = alphamax;
for (i = 0; i < n; i++) x[i] += alpha*dir[i];
eng_force(&n,&x,&dir,&eng);
nfunc++;
}
break;
}
}
return 0;
}