void RGFlow<Two_scale>::solve()
{
check_setup();
unsigned int max_iterations = get_max_iterations();
if (models.empty() || max_iterations == 0)
return;
initial_guess();
iteration = 0;
bool accuracy_reached = false;
while (iteration < max_iterations && !accuracy_reached) {
update_running_precision();
run_up();
run_down();
accuracy_reached = accuracy_goal_reached();
++iteration;
}
apply_lowest_constraint();
if (!accuracy_reached)
throw NoConvergenceError(max_iterations);
VERBOSE_MSG("convergence reached after " << iteration << " iterations");
}
/// main for SCF
int main (int argc, char **argv)
{
// init MPI
int myrank;
int nprocs;
int provided;
#if defined (USE_ELEMENTAL)
ElInitialize( &argc, &argv );
ElMPICommRank( MPI_COMM_WORLD, &myrank );
ElMPICommSize( MPI_COMM_WORLD, &nprocs );
MPI_Query_thread(&provided);
#else
MPI_Init_thread(&argc, &argv, MPI_THREAD_MULTIPLE, &provided);
MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
#endif
if (myrank == 0) {
printf("MPI thread support: %s\n", MPI_THREAD_STRING(provided));
}
#if 0
char hostname[1024];
gethostname (hostname, 1024);
printf ("Rank %d of %d running on node %s\n", myrank, nprocs, hostname);
#endif
// create basis set
BasisSet_t basis;
CInt_createBasisSet(&basis);
// input parameters and load basis set
int nprow_fock;
int npcol_fock;
int nblks_fock;
int nprow_purif;
int nshells;
int natoms;
int nfunctions;
int niters;
if (myrank == 0) {
if (argc != 8) {
usage(argv[0]);
MPI_Finalize();
exit(0);
}
// init parameters
nprow_fock = atoi(argv[3]);
npcol_fock = atoi(argv[4]);
nprow_purif = atoi(argv[5]);
nblks_fock = atoi(argv[6]);
niters = atoi(argv[7]);
assert(nprow_fock * npcol_fock == nprocs);
assert(nprow_purif * nprow_purif * nprow_purif <= nprocs);
assert(niters > 0);
CInt_loadBasisSet(basis, argv[1], argv[2]);
nshells = CInt_getNumShells(basis);
natoms = CInt_getNumAtoms(basis);
nfunctions = CInt_getNumFuncs(basis);
assert(nprow_fock <= nshells && npcol_fock <= nshells);
assert(nprow_purif <= nfunctions && nprow_purif <= nfunctions);
printf("Job information:\n");
char *fname;
fname = basename(argv[2]);
printf(" molecule: %s\n", fname);
fname = basename(argv[1]);
printf(" basisset: %s\n", fname);
printf(" charge = %d\n", CInt_getTotalCharge(basis));
printf(" #atoms = %d\n", natoms);
printf(" #shells = %d\n", nshells);
printf(" #functions = %d\n", nfunctions);
printf(" fock build uses %d (%dx%d) nodes\n",
nprow_fock * npcol_fock, nprow_fock, npcol_fock);
printf(" purification uses %d (%dx%dx%d) nodes\n",
nprow_purif * nprow_purif * nprow_purif,
nprow_purif, nprow_purif, nprow_purif);
printf(" #tasks = %d (%dx%d)\n",
nblks_fock * nblks_fock * nprow_fock * nprow_fock,
nblks_fock * nprow_fock, nblks_fock * nprow_fock);
int nthreads = omp_get_max_threads();
printf(" #nthreads_cpu = %d\n", nthreads);
}
int btmp[8];
btmp[0] = nprow_fock;
btmp[1] = npcol_fock;
btmp[2] = nprow_purif;
btmp[3] = nblks_fock;
btmp[4] = niters;
btmp[5] = natoms;
btmp[6] = nshells;
btmp[7] = nfunctions;
MPI_Bcast(btmp, 8, MPI_INT, 0, MPI_COMM_WORLD);
nprow_fock = btmp[0];
npcol_fock = btmp[1];
nprow_purif = btmp[2];
nblks_fock = btmp[3];
niters = btmp[4];
natoms = btmp[5];
nshells = btmp[6];
nfunctions = btmp[7];
// broadcast basis set
void *bsbuf;
int bsbufsize;
if (myrank == 0) {
CInt_packBasisSet(basis, &bsbuf, &bsbufsize);
MPI_Bcast(&bsbufsize, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Bcast(bsbuf, bsbufsize, MPI_CHAR, 0, MPI_COMM_WORLD);
}
else {
MPI_Bcast(&bsbufsize, 1, MPI_INT, 0, MPI_COMM_WORLD);
bsbuf = (void *)malloc(bsbufsize);
assert(bsbuf != NULL);
MPI_Bcast(bsbuf, bsbufsize, MPI_CHAR, 0, MPI_COMM_WORLD);
CInt_unpackBasisSet(basis, bsbuf);
free(bsbuf);
}
// init PFock
if (myrank == 0) {
printf("Initializing pfock ...\n");
}
PFock_t pfock;
PFock_create(basis, nprow_fock, npcol_fock, nblks_fock, 1e-11,
MAX_NUM_D, IS_SYMM, &pfock);
if (myrank == 0) {
double mem_cpu;
PFock_getMemorySize(pfock, &mem_cpu);
printf(" CPU uses %.3f MB\n", mem_cpu / 1024.0 / 1024.0);
printf(" Done\n");
}
// init purif
purif_t *purif = create_purif(basis, nprow_purif, nprow_purif, nprow_purif);
init_oedmat(basis, pfock, purif, nprow_fock, npcol_fock);
// compute SCF
if (myrank == 0) {
printf("Computing SCF ...\n");
}
int rowstart = purif->srow_purif;
int rowend = purif->nrows_purif + rowstart - 1;
int colstart = purif->scol_purif;
int colend = purif->ncols_purif + colstart - 1;
double energy0 = -1.0;
double totaltime = 0.0;
double purif_flops = 2.0 * nfunctions * nfunctions * nfunctions;
double diis_flops;
// set initial guess
if (myrank == 0) {
printf(" initialing D ...\n");
}
PFock_setNumDenMat(NUM_D, pfock);
initial_guess(pfock, basis, purif->runpurif,
rowstart, rowend, colstart, colend,
purif->D_block, purif->ldx);
MPI_Barrier(MPI_COMM_WORLD);
// compute nuc energy
double ene_nuc = CInt_getNucEnergy(basis);
if (myrank == 0) {
printf(" nuc energy = %.10f\n", ene_nuc);
}
MPI_Barrier(MPI_COMM_WORLD);
// main loop
double t1, t2, t3, t4;
for (int iter = 0; iter < niters; iter++) {
if (myrank == 0) {
printf(" iter %d\n", iter);
}
t3 = MPI_Wtime();
// fock matrix construction
t1 = MPI_Wtime();
fock_build(pfock, basis, purif->runpurif,
rowstart, rowend, colstart, colend,
purif->ldx, purif->D_block, purif->F_block);
if (myrank == 0) {
printf("After fock build \n");
}
// compute energy
double energy = compute_energy(purif, purif->F_block, purif->D_block);
t2 = MPI_Wtime();
if (myrank == 0) {
printf(" fock build takes %.3f secs\n", t2 - t1);
if (iter > 0) {
printf(" energy %.10f (%.10f), %le\n",
energy + ene_nuc, energy, fabs (energy - energy0));
}
else {
printf(" energy %.10f (%.10f)\n", energy + ene_nuc,
energy);
}
}
if (iter > 0 && fabs (energy - energy0) < 1e-11) {
niters = iter + 1;
break;
}
energy0 = energy;
// compute DIIS
t1 = MPI_Wtime();
compute_diis(pfock, purif, purif->D_block, purif->F_block, iter);
t2 = MPI_Wtime();
if (myrank == 0) {
if (iter > 1) {
diis_flops = purif_flops * 6.0;
} else {
diis_flops = purif_flops * 2.0;
}
printf(" diis takes %.3f secs, %.3lf Gflops\n",
t2 - t1, diis_flops / (t2 - t1) / 1e9);
}
#ifdef __SCF_OUT__
if (myrank == 0) {
double outbuf[nfunctions];
char fname[1024];
sprintf(fname, "XFX_%d_%d.dat", nfunctions, iter);
FILE *fp = fopen(fname, "w+");
assert(fp != NULL);
for (int i = 0; i < nfunctions; i++) {
PFock_getMat(pfock, PFOCK_MAT_TYPE_F, USE_D_ID,
i, i, USE_D_ID, nfunctions - 1,
outbuf, nfunctions);
for (int j = 0; j < nfunctions; j++) {
fprintf(fp, "%.10e\n", outbuf[j]);
}
}
fclose(fp);
}
#endif
// purification
MPI_Barrier(MPI_COMM_WORLD);
t1 = MPI_Wtime();
int it = compute_purification(purif, purif->F_block, purif->D_block);
t2 = MPI_Wtime();
MPI_Barrier(MPI_COMM_WORLD);
if (myrank == 0) {
printf(" purification takes %.3f secs,"
" %d iterations, %.3f Gflops\n",
t2 - t1, it,
(it * 2.0 + 4.0) * purif_flops / (t2 - t1) / 1e9);
}
/*
#if defined(USE_ELEMENTAL)
ElGlobalArraysPrint_d( eldga, pfock->ga_D[USE_D_ID] );
#else
GA_Print (pfock->ga_D[USE_D_ID]);
#endif
*/
t4 = MPI_Wtime ();
totaltime += t4 - t3;
#ifdef __SCF_TIMING__
PFock_getStatistics(pfock);
double purif_timedgemm;
double purif_timepdgemm;
double purif_timepass;
double purif_timetr;
MPI_Reduce(&purif->timedgemm, &purif_timedgemm,
1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
MPI_Reduce(&purif->timepdgemm, &purif_timepdgemm,
1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
MPI_Reduce(&purif->timepass, &purif_timepass,
1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
MPI_Reduce(&purif->timetr, &purif_timetr,
1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
if (myrank == 0) {
printf(" Purification Statistics:\n");
printf(" average totaltime = %.3f\n"
" average timetr = %.3f\n"
" average timedgemm = %.3f, %.3f Gflops\n"
" average timepdgemm = %.3f, %.3f Gflops\n",
purif_timepass / purif->np_purif,
purif_timetr / purif->np_purif,
purif_timedgemm / purif->np_purif,
(it * 2.0 + 4.0) *
purif_flops / (purif_timedgemm / purif->np_purif) / 1e9,
purif_timepdgemm / purif->np_purif,
(it * 2.0 + 4.0) *
purif_flops / (purif_timepdgemm / purif->np_purif) / 1e9);
}
#endif
} /* for (iter = 0; iter < NITERATIONS; iter++) */
if (myrank == 0) {
printf(" totally takes %.3f secs: %.3f secs/iters\n",
totaltime, totaltime / niters);
printf(" Done\n");
}
destroy_purif(purif);
PFock_destroy(pfock);
CInt_destroyBasisSet(basis);
MPI_Finalize();
return 0;
}
int main(int argc, char **argv)
{
USING_NAMESPACE_ACADO
const double KKT_tol = 1e-6;
//READ THE DEMO LENGTHS & nBF FROM THE COMMAND LINE
std::deque<std::string> args(argv + 1, argv + argc + !argc);
const int nBF=atoi(args[0].c_str());
args.pop_front();
const int nD=(int)args.size();
int nS=0;
for(int i=0; i<nD;i++)
nS+=atoi(args[i].c_str());
//READ DATA
std::string path=DATA_PATH;
Matrix D = readFromFile((path+"demos.dat").c_str()); //d(:,0)=time, d(:,1)=x, d(:,2)=dx, d(:,3)=ddx;
Vector pI = readFromFile((path+"initial_guess.dat").c_str());
Matrix A = readFromFile((path+"inequality_constraint_matrix.dat").c_str());
Vector b = readFromFile((path+"inequality_constraint_vector.dat").c_str());
Matrix S = readFromFile((path+"scale_matrix.dat").c_str());
//RELEVANT INDEX SETS
std::vector<std::vector<int> > d_ind=getDemoInd(args);
std::vector<int> a_ind=getAInd(nBF,nD);
std::vector<int> b_ind=getBInd(nBF,nD);
std::vector<std::vector<int> > w_ind=getWInd(nBF,nD);
std::vector<int> r_ind=getRInd(nBF,nD);
std::vector<int> c_ind=getCInd(nBF,nD);
//PARAMETER & OBJECTIVE FUNCTION
Parameter p(2*nD+nBF*(2+nD)+1,1);
Matrix BM(nS,2*nD+nBF*(2+nD)+1); BM.setZero();
Expression B(BM);
double t,x,dx;
for (int d=0; d<nD; d++)
for(int s=0;s<(int)d_ind[d].size();s++)
{
t=D(d_ind[d][s],0);
x=D(d_ind[d][s],1);
dx=D(d_ind[d][s],2);
B(d_ind[d][s],a_ind[d])=x;
B(d_ind[d][s],b_ind[d])=dx;
for(int n=0;n<nBF;n++){
B(d_ind[d][s],w_ind[d][n])=(-0.5*(t-p(c_ind[n])*S(c_ind[n],c_ind[n])).getPowInt(2)/(p(r_ind[n])*p(r_ind[n])*S(r_ind[n],r_ind[n])*S(r_ind[n],r_ind[n]))).getExp();
// std::cout<<d<<std::endl;
//std::cout<< S(r_ind[d],r_ind[d])<<std::endl;
}
}
Expression f;
f<<B*S*p-D.getCol(3);
Expression ez(nS);
for (int i=0; i<nS; i++)
ez(i)=p(2*nD+nBF*(2+nD));
Vector e(nS); e.setAll(1.0);
Vector null(nS); null.setAll(0.0);
NLP nlp;
nlp.minimize(p(2*nD+nBF*(2+nD)));
nlp.subjectTo(f - ez <= null);
nlp.subjectTo(f + ez >= null);
//nlp.subjectTo(A*S*p <= b);
//ALGORITHM
ParameterEstimationAlgorithm algorithm(nlp);
VariablesGrid initial_guess(2*nD+nBF*(2+nD)+1,0.0,0.0,1 );
initial_guess.setVector( 0,S.getInverse()*pI );
algorithm.initializeParameters(initial_guess);
// OPTIONS
algorithm.set( KKT_TOLERANCE, KKT_tol);
algorithm.set( ABSOLUTE_TOLERANCE, 1e-4);
algorithm.set( PRINTLEVEL,HIGH);
algorithm.set( MAX_NUM_ITERATIONS, 2000 );
algorithm.set (PRINT_COPYRIGHT, NO);
// algorithm.set (PRINT_SCP_METHOD_PROFILE, YES);
algorithm.set( HESSIAN_APPROXIMATION, EXACT_HESSIAN );
algorithm.set(GLOBALIZATION_STRATEGY, GS_LINESEARCH );
algorithm.set(LINESEARCH_TOLERANCE, 1e-2 );
algorithm.set(INFEASIBLE_QP_HANDLING,IQH_RELAX_L2);
algorithm.set(FEASIBILITY_CHECK,BT_TRUE);
// LOGGING
LogRecord logRecord( LOG_AT_EACH_ITERATION,(path+"log.dat").c_str(),PS_PLAIN);
logRecord << LOG_OBJECTIVE_VALUE;
algorithm << logRecord;
//SOLVING
double clock1 = clock();
algorithm.solve();
double clock2 = clock();
Vector solution;
algorithm.getParameters(solution);
// solution.print("optimal solution \n");
solution.printToFile((path+"solution.dat").c_str(),"",PS_PLAIN);
printf("\n computation time (ACADO) = %.16e \n", (clock2-clock1)/CLOCKS_PER_SEC);
return 0;
}
