void DynamicLimit::init(float k, Type t, uint32 uLimit) {
resetRun();
std::memset(&adjust, 0, sizeof(adjust));
adjust.limit = uLimit;
adjust.rk = k;
adjust.type = t;
}
bool VMCSingleOMP::run()
{
resetRun();
//start the main estimator
Estimators->start(nBlocks);
#pragma omp parallel
{
int now=0;
#pragma omp for
for(int ip=0; ip<NumThreads; ++ip)
Movers[ip]->startRun(nBlocks,false);
for(int block=0;block<nBlocks; ++block)
{
#pragma omp for
for(int ip=0; ip<NumThreads; ++ip)
{
IndexType updatePeriod=(QMCDriverMode[QMC_UPDATE_MODE])?Period4CheckProperties:(nBlocks+1)*nSteps;
//assign the iterators and resuse them
MCWalkerConfiguration::iterator wit(W.begin()+wPerNode[ip]), wit_end(W.begin()+wPerNode[ip+1]);
Movers[ip]->startBlock(nSteps);
int now_loc=now;
for(int step=0; step<nSteps;++step)
{
Movers[ip]->advanceWalkers(wit,wit_end,false);
Movers[ip]->accumulate(wit,wit_end);
++now_loc;
if(now_loc%updatePeriod==0) Movers[ip]->updateWalkers(wit,wit_end);
if(now_loc%myPeriod4WalkerDump==0) wClones[ip]->saveEnsemble(wit,wit_end);
}
Movers[ip]->stopBlock();
}//end-of-parallel for
//increase now
now+=nSteps;
#pragma omp master
{
CurrentStep+=nSteps;
Estimators->stopBlock(estimatorClones);
recordBlock(block+1);
}//end of mater
}//block
}//end of parallel
Estimators->stop(estimatorClones);
//copy back the random states
for(int ip=0; ip<NumThreads; ++ip)
*(RandomNumberControl::Children[ip])=*(Rng[ip]);
//finalize a qmc section
return finalize(nBlocks);
}
//
// Perform resetRun on all the reset commands in the list, using
// initiator and initiator_type
void resetRunOn(LIST *list, void *initiator, int initiator_type,
const char *locale) {
if(listSize(list) > 0) {
LIST_ITERATOR *list_i = newListIterator(list);
RESET_DATA *reset = NULL;
ITERATE_LIST(reset, list_i)
resetRun(reset, initiator, initiator_type, locale);
deleteListIterator(list_i);
}
}
bool VMCSingle::run() {
resetRun();
Mover->startRun(nBlocks,true);
IndexType block = 0;
IndexType nAcceptTot = 0;
IndexType nRejectTot = 0;
do {
Mover->startBlock(nSteps);
IndexType step = 0;
do
{
++step;++CurrentStep;
Mover->advanceWalkers(W.begin(),W.end(),true); //step==nSteps);
Estimators->accumulate(W);
//save walkers for optimization
if(QMCDriverMode[QMC_OPTIMIZE]&&CurrentStep%Period4WalkerDump==0) W.saveEnsemble();
} while(step<nSteps);
Mover->stopBlock();
nAcceptTot += Mover->nAccept;
nRejectTot += Mover->nReject;
++block;
recordBlock(block);
////periodically re-evaluate everything for pbyp
//if(QMCDriverMode[QMC_UPDATE_MODE] && CurrentStep%100 == 0)
// Mover->updateWalkers(W.begin(),W.end());
} while(block<nBlocks);
Mover->stopRun();
//finalize a qmc section
return finalize(block);
}
bool VMCcuda::run() {
if (UseDrift == "yes")
return runWithDrift();
resetRun();
IndexType block = 0;
IndexType nAcceptTot = 0;
IndexType nRejectTot = 0;
IndexType updatePeriod= (QMCDriverMode[QMC_UPDATE_MODE])
? Period4CheckProperties
: (nBlocks+1)*nSteps;
int nat = W.getTotalNum();
int nw = W.getActiveWalkers();
vector<RealType> LocalEnergy(nw);
vector<PosType> delpos(nw);
vector<PosType> newpos(nw);
vector<ValueType> ratios(nw);
vector<GradType> oldG(nw), newG(nw);
vector<ValueType> oldL(nw), newL(nw);
vector<Walker_t*> accepted(nw);
Matrix<ValueType> lapl(nw, nat);
Matrix<GradType> grad(nw, nat);
double Esum;
// First do warmup steps
for (int step=0; step<myWarmupSteps; step++) {
for(int iat=0; iat<nat; ++iat) {
//create a 3N-Dimensional Gaussian with variance=1
makeGaussRandomWithEngine(delpos,Random);
for(int iw=0; iw<nw; ++iw) {
PosType G = W[iw]->Grad[iat];
newpos[iw]=W[iw]->R[iat] + m_sqrttau*delpos[iw];
ratios[iw] = 1.0;
}
W.proposeMove_GPU(newpos, iat);
Psi.ratio(W,iat,ratios,newG, newL);
accepted.clear();
vector<bool> acc(nw, false);
for(int iw=0; iw<nw; ++iw) {
if(ratios[iw]*ratios[iw] > Random()) {
accepted.push_back(W[iw]);
nAccept++;
W[iw]->R[iat] = newpos[iw];
acc[iw] = true;
}
else
nReject++;
}
W.acceptMove_GPU(acc);
if (accepted.size())
Psi.update(accepted,iat);
}
}
do {
IndexType step = 0;
nAccept = nReject = 0;
Esum = 0.0;
Estimators->startBlock(nSteps);
do
{
++step;
++CurrentStep;
for (int isub=0; isub<nSubSteps; isub++) {
for(int iat=0; iat<nat; ++iat) {
//create a 3N-Dimensional Gaussian with variance=1
makeGaussRandomWithEngine(delpos,Random);
for(int iw=0; iw<nw; ++iw) {
PosType G = W[iw]->Grad[iat];
newpos[iw]=W[iw]->R[iat] + m_sqrttau*delpos[iw];
ratios[iw] = 1.0;
}
W.proposeMove_GPU(newpos, iat);
Psi.ratio(W,iat,ratios,newG, newL);
accepted.clear();
vector<bool> acc(nw, false);
for(int iw=0; iw<nw; ++iw) {
if(ratios[iw]*ratios[iw] > Random()) {
accepted.push_back(W[iw]);
nAccept++;
W[iw]->R[iat] = newpos[iw];
acc[iw] = true;
}
else
nReject++;
}
W.acceptMove_GPU(acc);
if (accepted.size())
Psi.update(accepted,iat);
}
}
Psi.gradLapl(W, grad, lapl);
H.evaluate (W, LocalEnergy);
if (myPeriod4WalkerDump && (CurrentStep % myPeriod4WalkerDump)==0)
W.saveEnsemble();
Estimators->accumulate(W);
} while(step<nSteps);
Psi.recompute(W);
// vector<RealType> logPsi(W.WalkerList.size(), 0.0);
// Psi.evaluateLog(W, logPsi);
double accept_ratio = (double)nAccept/(double)(nAccept+nReject);
Estimators->stopBlock(accept_ratio);
nAcceptTot += nAccept;
nRejectTot += nReject;
++block;
recordBlock(block);
} while(block<nBlocks);
//Mover->stopRun();
//finalize a qmc section
return finalize(block);
}
bool VMCcuda::runWithDrift()
{
resetRun();
IndexType block = 0;
IndexType nAcceptTot = 0;
IndexType nRejectTot = 0;
int nat = W.getTotalNum();
int nw = W.getActiveWalkers();
vector<RealType> LocalEnergy(nw), oldScale(nw), newScale(nw);
vector<PosType> delpos(nw);
vector<PosType> dr(nw);
vector<PosType> newpos(nw);
vector<ValueType> ratios(nw), rplus(nw), rminus(nw);
vector<PosType> oldG(nw), newG(nw);
vector<ValueType> oldL(nw), newL(nw);
vector<Walker_t*> accepted(nw);
Matrix<ValueType> lapl(nw, nat);
Matrix<GradType> grad(nw, nat);
// First, do warmup steps
for (int step=0; step<myWarmupSteps; step++) {
for(int iat=0; iat<nat; iat++) {
Psi.getGradient (W, iat, oldG);
//create a 3N-Dimensional Gaussian with variance=1
makeGaussRandomWithEngine(delpos,Random);
for(int iw=0; iw<nw; iw++) {
oldScale[iw] = getDriftScale(m_tauovermass,oldG[iw]);
dr[iw] = (m_sqrttau*delpos[iw]) + (oldScale[iw]*oldG[iw]);
newpos[iw]=W[iw]->R[iat] + dr[iw];
ratios[iw] = 1.0;
}
W.proposeMove_GPU(newpos, iat);
Psi.ratio(W,iat,ratios,newG, newL);
accepted.clear();
vector<bool> acc(nw, false);
for(int iw=0; iw<nw; ++iw) {
PosType drOld =
newpos[iw] - (W[iw]->R[iat] + oldScale[iw]*oldG[iw]);
RealType logGf = -m_oneover2tau * dot(drOld, drOld);
newScale[iw] = getDriftScale(m_tauovermass,newG[iw]);
PosType drNew =
(newpos[iw] + newScale[iw]*newG[iw]) - W[iw]->R[iat];
RealType logGb = -m_oneover2tau * dot(drNew, drNew);
RealType x = logGb - logGf;
RealType prob = ratios[iw]*ratios[iw]*std::exp(x);
if(Random() < prob) {
accepted.push_back(W[iw]);
nAccept++;
W[iw]->R[iat] = newpos[iw];
acc[iw] = true;
}
else
nReject++;
}
W.acceptMove_GPU(acc);
if (accepted.size())
Psi.update(accepted,iat);
}
}
// Now do data collection steps
do {
IndexType step = 0;
nAccept = nReject = 0;
Estimators->startBlock(nSteps);
do {
step++;
CurrentStep++;
for (int isub=0; isub<nSubSteps; isub++) {
for(int iat=0; iat<nat; iat++) {
Psi.getGradient (W, iat, oldG);
//create a 3N-Dimensional Gaussian with variance=1
makeGaussRandomWithEngine(delpos,Random);
for(int iw=0; iw<nw; iw++) {
oldScale[iw] = getDriftScale(m_tauovermass,oldG[iw]);
dr[iw] = (m_sqrttau*delpos[iw]) + (oldScale[iw]*oldG[iw]);
newpos[iw]=W[iw]->R[iat] + dr[iw];
ratios[iw] = 1.0;
}
W.proposeMove_GPU(newpos, iat);
Psi.ratio(W,iat,ratios,newG, newL);
accepted.clear();
vector<bool> acc(nw, false);
for(int iw=0; iw<nw; ++iw) {
PosType drOld =
newpos[iw] - (W[iw]->R[iat] + oldScale[iw]*oldG[iw]);
// if (dot(drOld, drOld) > 25.0)
// cerr << "Large drift encountered! Old drift = " << drOld << endl;
RealType logGf = -m_oneover2tau * dot(drOld, drOld);
newScale[iw] = getDriftScale(m_tauovermass,newG[iw]);
PosType drNew =
(newpos[iw] + newScale[iw]*newG[iw]) - W[iw]->R[iat];
// if (dot(drNew, drNew) > 25.0)
// cerr << "Large drift encountered! Drift = " << drNew << endl;
RealType logGb = -m_oneover2tau * dot(drNew, drNew);
RealType x = logGb - logGf;
RealType prob = ratios[iw]*ratios[iw]*std::exp(x);
if(Random() < prob) {
accepted.push_back(W[iw]);
nAccept++;
W[iw]->R[iat] = newpos[iw];
acc[iw] = true;
}
else
nReject++;
}
W.acceptMove_GPU(acc);
if (accepted.size())
Psi.update(accepted,iat);
}
// cerr << "Rank = " << myComm->rank() <<
// " CurrentStep = " << CurrentStep << " isub = " << isub << endl;
}
Psi.gradLapl(W, grad, lapl);
H.evaluate (W, LocalEnergy);
if (myPeriod4WalkerDump && (CurrentStep % myPeriod4WalkerDump)==0)
W.saveEnsemble();
Estimators->accumulate(W);
} while(step<nSteps);
Psi.recompute(W);
double accept_ratio = (double)nAccept/(double)(nAccept+nReject);
Estimators->stopBlock(accept_ratio);
nAcceptTot += nAccept;
nRejectTot += nReject;
++block;
recordBlock(block);
} while(block<nBlocks);
//finalize a qmc section
if (!myComm->rank())
gpu::cuda_memory_manager.report();
return finalize(block);
}
void DynamicLimit::reset() {
std::memset(&global, 0, sizeof(global));
resetRun();
}
bool DMCcuda::run()
{
bool NLmove = NonLocalMove == "yes";
bool scaleweight = ScaleWeight == "yes";
if (NLmove)
app_log() << " Using Casula nonlocal moves in DMCcuda.\n";
if (scaleweight)
app_log() << " Scaling weight per Umrigar/Nightengale.\n";
resetRun();
Mover->MaxAge = 1;
IndexType block = 0;
IndexType nAcceptTot = 0;
IndexType nRejectTot = 0;
int nat = W.getTotalNum();
int nw = W.getActiveWalkers();
vector<RealType> LocalEnergy(nw), LocalEnergyOld(nw),
oldScale(nw), newScale(nw);
vector<PosType> delpos(nw);
vector<PosType> dr(nw);
vector<PosType> newpos(nw);
vector<ValueType> ratios(nw), rplus(nw), rminus(nw), R2prop(nw), R2acc(nw);
vector<PosType> oldG(nw), newG(nw);
vector<ValueType> oldL(nw), newL(nw);
vector<Walker_t*> accepted(nw);
Matrix<ValueType> lapl(nw, nat);
Matrix<GradType> grad(nw, nat);
vector<ValueType> V2(nw), V2bar(nw);
vector<vector<NonLocalData> > Txy(nw);
for (int iw=0; iw<nw; iw++)
W[iw]->Weight = 1.0;
do {
IndexType step = 0;
nAccept = nReject = 0;
Estimators->startBlock(nSteps);
do {
step++;
CurrentStep++;
nw = W.getActiveWalkers();
ResizeTimer.start();
LocalEnergy.resize(nw); oldScale.resize(nw);
newScale.resize(nw); delpos.resize(nw);
dr.resize(nw); newpos.resize(nw);
ratios.resize(nw); rplus.resize(nw);
rminus.resize(nw); oldG.resize(nw);
newG.resize(nw); oldL.resize(nw);
newL.resize(nw); accepted.resize(nw);
lapl.resize(nw, nat); grad.resize(nw, nat);
R2prop.resize(nw,0.0); R2acc.resize(nw,0.0);
V2.resize(nw,0.0); V2bar.resize(nw,0.0);
W.updateLists_GPU();
ResizeTimer.stop();
if (NLmove) {
Txy.resize(nw);
for (int iw=0; iw<nw; iw++) {
Txy[iw].clear();
Txy[iw].push_back(NonLocalData(-1, 1.0, PosType()));
}
}
for (int iw=0; iw<nw; iw++)
W[iw]->Age++;
DriftDiffuseTimer.start();
for(int iat=0; iat<nat; iat++) {
Psi.getGradient (W, iat, oldG);
//create a 3N-Dimensional Gaussian with variance=1
makeGaussRandomWithEngine(delpos,Random);
for(int iw=0; iw<nw; iw++) {
delpos[iw] *= m_sqrttau;
oldScale[iw] = getDriftScale(m_tauovermass,oldG[iw]);
dr[iw] = delpos[iw] + (oldScale[iw]*oldG[iw]);
newpos[iw]=W[iw]->R[iat] + dr[iw];
ratios[iw] = 1.0;
R2prop[iw] += dot(delpos[iw], delpos[iw]);
}
W.proposeMove_GPU(newpos, iat);
Psi.ratio(W,iat,ratios,newG, newL);
accepted.clear();
vector<bool> acc(nw, false);
for(int iw=0; iw<nw; ++iw) {
PosType drOld =
newpos[iw] - (W[iw]->R[iat] + oldScale[iw]*oldG[iw]);
RealType logGf = -m_oneover2tau * dot(drOld, drOld);
newScale[iw] = getDriftScale(m_tauovermass,newG[iw]);
PosType drNew =
(newpos[iw] + newScale[iw]*newG[iw]) - W[iw]->R[iat];
RealType logGb = -m_oneover2tau * dot(drNew, drNew);
RealType x = logGb - logGf;
RealType prob = ratios[iw]*ratios[iw]*std::exp(x);
if(Random() < prob && ratios[iw] > 0.0) {
accepted.push_back(W[iw]);
nAccept++;
W[iw]->R[iat] = newpos[iw];
W[iw]->Age = 0;
acc[iw] = true;
R2acc[iw] += dot(delpos[iw], delpos[iw]);
V2[iw] += m_tauovermass * m_tauovermass * dot(newG[iw],newG[iw]);
V2bar[iw] += newScale[iw] * newScale[iw] * dot(newG[iw],newG[iw]);
}
else {
nReject++;
V2[iw] += m_tauovermass * m_tauovermass * dot(oldG[iw],oldG[iw]);
V2bar[iw] += oldScale[iw] * oldScale[iw] * dot(oldG[iw],oldG[iw]);
}
}
W.acceptMove_GPU(acc);
if (accepted.size())
Psi.update(accepted,iat);
}
DriftDiffuseTimer.stop();
// Psi.recompute(W, false);
Psi.gradLapl(W, grad, lapl);
HTimer.start();
if (NLmove) H.evaluate (W, LocalEnergy, Txy);
else H.evaluate (W, LocalEnergy);
HTimer.stop();
// for (int iw=0; iw<nw; iw++) {
// branchEngine->clampEnergy(LocalEnergy[iw]);
// W[iw]->getPropertyBase()[LOCALENERGY] = LocalEnergy[iw];
// }
if (CurrentStep == 1)
LocalEnergyOld = LocalEnergy;
if (NLmove) {
// Now, attempt nonlocal move
accepted.clear();
vector<int> iatList;
vector<PosType> accPos;
for (int iw=0; iw<nw; iw++) {
/// HACK HACK HACK
// if (LocalEnergy[iw] < -2300.0) {
// cerr << "Walker " << iw << " has energy "
// << LocalEnergy[iw] << endl;;
// double maxWeight = 0.0;
// int elMax = -1;
// PosType posMax;
// for (int j=1; j<Txy[iw].size(); j++)
// if (std::fabs(Txy[iw][j].Weight) > std::fabs(maxWeight)) {
// maxWeight = Txy[iw][j].Weight;
// elMax = Txy[iw][j].PID;
// posMax = W[iw]->R[elMax] + Txy[iw][j].Delta;
// }
// cerr << "Maximum weight is " << maxWeight << " for electron "
// << elMax << " at position " << posMax << endl;
// PosType unit = W.Lattice.toUnit(posMax);
// unit[0] -= round(unit[0]);
// unit[1] -= round(unit[1]);
// unit[2] -= round(unit[2]);
// cerr << "Reduced position = " << unit << endl;
// }
int ibar = NLop.selectMove(Random(), Txy[iw]);
if (ibar) {
accepted.push_back(W[iw]);
int iat = Txy[iw][ibar].PID;
iatList.push_back(iat);
accPos.push_back(W[iw]->R[iat] + Txy[iw][ibar].Delta);
}
}
if (accepted.size()) {
Psi.ratio(accepted,iatList, accPos, ratios, newG, newL);
Psi.update(accepted,iatList);
for (int i=0; i<accepted.size(); i++)
accepted[i]->R[iatList[i]] = accPos[i];
W.NLMove_GPU (accepted, accPos, iatList);
// HACK HACK HACK
// Recompute the kinetic energy
// Psi.gradLapl(W, grad, lapl);
// H.evaluate (W, LocalEnergy);
//W.copyWalkersToGPU();
}
}
// Now branch
BranchTimer.start();
for (int iw=0; iw<nw; iw++) {
RealType v2=0.0, v2bar=0.0;
for(int iat=0; iat<nat; iat++) {
v2 += dot(W.G[iat],W.G[iat]);
RealType newscale = getDriftScale(m_tauovermass,newG[iw]);
v2 += m_tauovermass * m_tauovermass * dot(newG[iw],newG[iw]);
v2bar += newscale * newscale * dot(newG[iw],newG[iw]);
}
//RealType scNew = std::sqrt(V2bar[iw] / V2[iw]);
RealType scNew = std::sqrt(v2bar/v2);
RealType scOld = (CurrentStep == 1) ? scNew : W[iw]->getPropertyBase()[DRIFTSCALE];
W[iw]->getPropertyBase()[DRIFTSCALE] = scNew;
// fprintf (stderr, "iw = %d scNew = %1.8f scOld = %1.8f\n", iw, scNew, scOld);
RealType tauRatio = R2acc[iw] / R2prop[iw];
if (tauRatio < 0.5)
cerr << " tauRatio = " << tauRatio << endl;
RealType taueff = m_tauovermass * tauRatio;
if (scaleweight)
W[iw]->Weight *= branchEngine->branchWeightTau
(LocalEnergy[iw], LocalEnergyOld[iw], scNew, scOld, taueff);
else
W[iw]->Weight *= branchEngine->branchWeight
(LocalEnergy[iw], LocalEnergyOld[iw]);
W[iw]->getPropertyBase()[R2ACCEPTED] = R2acc[iw];
W[iw]->getPropertyBase()[R2PROPOSED] = R2prop[iw];
}
Mover->setMultiplicity(W.begin(), W.end());
branchEngine->branch(CurrentStep,W);
nw = W.getActiveWalkers();
LocalEnergyOld.resize(nw);
for (int iw=0; iw<nw; iw++)
LocalEnergyOld[iw] = W[iw]->getPropertyBase()[LOCALENERGY];
BranchTimer.stop();
} while(step<nSteps);
Psi.recompute(W, true);
double accept_ratio = (double)nAccept/(double)(nAccept+nReject);
Estimators->stopBlock(accept_ratio);
nAcceptTot += nAccept;
nRejectTot += nReject;
++block;
recordBlock(block);
} while(block<nBlocks);
//finalize a qmc section
return finalize(block);
}
bool DMCcuda::runWithNonlocal()
{
resetRun();
Mover->MaxAge = 1;
IndexType block = 0;
IndexType nAcceptTot = 0;
IndexType nRejectTot = 0;
int nat = W.getTotalNum();
int nw = W.getActiveWalkers();
vector<RealType> LocalEnergy(nw), LocalEnergyOld(nw),
oldScale(nw), newScale(nw);
vector<PosType> delpos(nw);
vector<PosType> dr(nw);
vector<PosType> newpos(nw);
vector<ValueType> ratios(nw), rplus(nw), rminus(nw), R2prop(nw), R2acc(nw);
vector<PosType> oldG(nw), newG(nw);
vector<ValueType> oldL(nw), newL(nw);
vector<Walker_t*> accepted(nw);
Matrix<ValueType> lapl(nw, nat);
Matrix<GradType> grad(nw, nat);
vector<vector<NonLocalData> > Txy(nw);
for (int iw=0; iw<nw; iw++)
W[iw]->Weight = 1.0;
do {
IndexType step = 0;
nAccept = nReject = 0;
Estimators->startBlock(nSteps);
do {
step++;
CurrentStep++;
nw = W.getActiveWalkers();
LocalEnergy.resize(nw); oldScale.resize(nw);
newScale.resize(nw); delpos.resize(nw);
dr.resize(nw); newpos.resize(nw);
ratios.resize(nw); rplus.resize(nw);
rminus.resize(nw); oldG.resize(nw);
newG.resize(nw); oldL.resize(nw);
newL.resize(nw); accepted.resize(nw);
lapl.resize(nw, nat); grad.resize(nw, nat);
R2prop.resize(nw,0.0); R2acc.resize(nw,0.0);
W.updateLists_GPU();
Txy.resize(nw);
for (int iw=0; iw<nw; iw++) {
Txy[iw].clear();
Txy[iw].push_back(NonLocalData(-1, 1.0, PosType()));
W[iw]->Age++;
}
for(int iat=0; iat<nat; iat++) {
Psi.getGradient (W, iat, oldG);
//create a 3N-Dimensional Gaussian with variance=1
makeGaussRandomWithEngine(delpos,Random);
for(int iw=0; iw<nw; iw++) {
delpos[iw] *= m_sqrttau;
oldScale[iw] = getDriftScale(m_tauovermass,oldG[iw]);
dr[iw] = delpos[iw] + (oldScale[iw]*oldG[iw]);
newpos[iw]=W[iw]->R[iat] + dr[iw];
ratios[iw] = 1.0;
R2prop[iw] += dot(delpos[iw], delpos[iw]);
}
W.proposeMove_GPU(newpos, iat);
Psi.ratio(W,iat,ratios,newG, newL);
accepted.clear();
vector<bool> acc(nw, false);
for(int iw=0; iw<nw; ++iw) {
PosType drOld =
newpos[iw] - (W[iw]->R[iat] + oldScale[iw]*oldG[iw]);
RealType logGf = -m_oneover2tau * dot(drOld, drOld);
newScale[iw] = getDriftScale(m_tauovermass,newG[iw]);
PosType drNew =
(newpos[iw] + newScale[iw]*newG[iw]) - W[iw]->R[iat];
RealType logGb = -m_oneover2tau * dot(drNew, drNew);
RealType x = logGb - logGf;
RealType prob = ratios[iw]*ratios[iw]*std::exp(x);
if(Random() < prob && ratios[iw] > 0.0) {
accepted.push_back(W[iw]);
nAccept++;
W[iw]->R[iat] = newpos[iw];
W[iw]->Age = 0;
acc[iw] = true;
R2acc[iw] += dot(delpos[iw], delpos[iw]);
}
else
nReject++;
}
W.acceptMove_GPU(acc);
if (accepted.size())
Psi.update(accepted,iat);
}
for (int iw=0; iw < nw; iw++)
if (W[iw]->Age)
cerr << "Encountered stuck walker with iw=" << iw << endl;
// Psi.recompute(W, false);
Psi.gradLapl(W, grad, lapl);
H.evaluate (W, LocalEnergy, Txy);
if (CurrentStep == 1)
LocalEnergyOld = LocalEnergy;
// Now, attempt nonlocal move
accepted.clear();
vector<int> iatList;
vector<PosType> accPos;
for (int iw=0; iw<nw; iw++) {
int ibar = NLop.selectMove(Random(), Txy[iw]);
// cerr << "Txy[iw].size() = " << Txy[iw].size() << endl;
if (ibar) {
accepted.push_back(W[iw]);
int iat = Txy[iw][ibar].PID;
iatList.push_back(iat);
accPos.push_back(W[iw]->R[iat] + Txy[iw][ibar].Delta);
}
}
if (accepted.size()) {
// W.proposeMove_GPU(newpos, iatList);
Psi.ratio(accepted,iatList, accPos, ratios, newG, newL);
Psi.update(accepted,iatList);
for (int i=0; i<accepted.size(); i++)
accepted[i]->R[iatList[i]] = accPos[i];
W.copyWalkersToGPU();
}
// Now branch
for (int iw=0; iw<nw; iw++) {
W[iw]->Weight *= branchEngine->branchWeight(LocalEnergy[iw], LocalEnergyOld[iw]);
W[iw]->getPropertyBase()[R2ACCEPTED] = R2acc[iw];
W[iw]->getPropertyBase()[R2PROPOSED] = R2prop[iw];
}
Mover->setMultiplicity(W.begin(), W.end());
branchEngine->branch(CurrentStep,W);
nw = W.getActiveWalkers();
LocalEnergyOld.resize(nw);
for (int iw=0; iw<nw; iw++)
LocalEnergyOld[iw] = W[iw]->getPropertyBase()[LOCALENERGY];
} while(step<nSteps);
Psi.recompute(W, true);
double accept_ratio = (double)nAccept/(double)(nAccept+nReject);
Estimators->stopBlock(accept_ratio);
nAcceptTot += nAccept;
nRejectTot += nReject;
++block;
recordBlock(block);
} while(block<nBlocks);
//finalize a qmc section
return finalize(block);
}
bool WFMCSingleOMP::run()
{
resetRun();
//start the main estimator
Estimators->start(nBlocks);
///Load a single walkers position into the walker.
MCWalkerConfiguration Keeper(W);
Keeper.createWalkers(W.begin(),W.end());
MCWalkerConfiguration::iterator Kit=(Keeper.begin()), Kit_end(Keeper.end());
int block=0;
while ((Kit!=Kit_end)&&(nBlocks>block))
{
MCWalkerConfiguration::iterator Wit(W.begin()), Wit_end(W.end());
while ((Wit!=Wit_end))
{
(*Wit)->R=(*Kit)->R;
(*Wit)->G=(*Kit)->G;
(*Wit)->L=(*Kit)->L;
//(*Wit)->Drift=(*Kit)->Drift;
(*Wit)->reset();
(*Wit)->resetPropertyHistory();
++Wit;
}
#pragma omp parallel for
for (int ip=0; ip<NumThreads; ++ip)
Movers[ip]->initWalkers(W.begin()+wPerNode[ip],W.begin()+wPerNode[ip+1]);
Wit=W.begin();
while ((Wit!=Wit_end))
{
// app_log()<<std::exp((*Wit)->Properties(LOGPSI))<<endl;
(*Wit)->PropertyHistory[0][0]=std::exp((*Wit)->Properties(LOGPSI));
++Wit;
}
#pragma omp parallel
{
#pragma omp for
for (int ip=0; ip<NumThreads; ++ip)
Movers[ip]->startRun(nBlocks,false);
#pragma omp for
for (int ip=0; ip<NumThreads; ++ip)
{
//assign the iterators and resuse them
MCWalkerConfiguration::iterator wit(W.begin()+wPerNode[ip]), wit_end(W.begin()+wPerNode[ip+1]);
Movers[ip]->startBlock(nSteps);
for (int step=0; step<nSteps; ++step)
{
Movers[ip]->advanceWalkers(wit,wit_end,false);
// Movers[ip]->updateWalkers(wit,wit_end);
// wClones[ip]->saveEnsemble(wit,wit_end);
}
Movers[ip]->accumulate(wit,wit_end);
Movers[ip]->stopBlock();
}
#pragma omp master
{
Estimators->stopBlock(estimatorClones);
recordBlock(block+1);
block++;
++Kit;
}
}
}//end of parallel
Estimators->stop(estimatorClones);
//copy back the random states
for (int ip=0; ip<NumThreads; ++ip)
*(RandomNumberControl::Children[ip])=*(Rng[ip]);
//finalize a qmc section
return finalize(nBlocks);
}
bool VMCSingleOMP::run()
{
resetRun();
//start the main estimator
Estimators->start(nBlocks);
for (int ip=0; ip<NumThreads; ++ip)
Movers[ip]->startRun(nBlocks,false);
Traces->startRun(nBlocks,traceClones);
const bool has_collectables=W.Collectables.size();
ADIOS_PROFILE::profile_adios_init(nBlocks);
for (int block=0; block<nBlocks; ++block)
{
ADIOS_PROFILE::profile_adios_start_comp(block);
#pragma omp parallel
{
int ip=omp_get_thread_num();
//IndexType updatePeriod=(QMCDriverMode[QMC_UPDATE_MODE])?Period4CheckProperties:(nBlocks+1)*nSteps;
IndexType updatePeriod=(QMCDriverMode[QMC_UPDATE_MODE])?Period4CheckProperties:0;
//assign the iterators and resuse them
MCWalkerConfiguration::iterator wit(W.begin()+wPerNode[ip]), wit_end(W.begin()+wPerNode[ip+1]);
Movers[ip]->startBlock(nSteps);
int now_loc=CurrentStep;
RealType cnorm=1.0/static_cast<RealType>(wPerNode[ip+1]-wPerNode[ip]);
for (int step=0; step<nSteps; ++step)
{
Movers[ip]->set_step(now_loc);
//collectables are reset, it is accumulated while advancing walkers
wClones[ip]->resetCollectables();
Movers[ip]->advanceWalkers(wit,wit_end,false);
if(has_collectables)
wClones[ip]->Collectables *= cnorm;
Movers[ip]->accumulate(wit,wit_end);
++now_loc;
//if (updatePeriod&& now_loc%updatePeriod==0) Movers[ip]->updateWalkers(wit,wit_end);
if (Period4WalkerDump&& now_loc%Period4WalkerDump==0)
wClones[ip]->saveEnsemble(wit,wit_end);
// if(storeConfigs && (now_loc%storeConfigs == 0))
// ForwardWalkingHistory.storeConfigsForForwardWalking(*wClones[ip]);
}
Movers[ip]->stopBlock(false);
}//end-of-parallel for
//Estimators->accumulateCollectables(wClones,nSteps);
CurrentStep+=nSteps;
Estimators->stopBlock(estimatorClones);
ADIOS_PROFILE::profile_adios_end_comp(block);
ADIOS_PROFILE::profile_adios_start_trace(block);
Traces->write_buffers(traceClones, block);
ADIOS_PROFILE::profile_adios_end_trace(block);
ADIOS_PROFILE::profile_adios_start_checkpoint(block);
if(storeConfigs)
recordBlock(block);
ADIOS_PROFILE::profile_adios_end_checkpoint(block);
}//block
ADIOS_PROFILE::profile_adios_finalize(myComm, nBlocks);
Estimators->stop(estimatorClones);
for (int ip=0; ip<NumThreads; ++ip)
Movers[ip]->stopRun2();
Traces->stopRun();
//copy back the random states
for (int ip=0; ip<NumThreads; ++ip)
*(RandomNumberControl::Children[ip])=*(Rng[ip]);
///write samples to a file
bool wrotesamples=DumpConfig;
if(DumpConfig)
{
wrotesamples=W.dumpEnsemble(wClones,wOut,myComm->size(),nBlocks);
if(wrotesamples)
app_log() << " samples are written to the config.h5" << endl;
}
//finalize a qmc section
return finalize(nBlocks,!wrotesamples);
}
