suN_hmc_stout_comp.cpp

#include "hila.h"
#include "gauge/polyakov.h"
#include "gauge/wilson_plaquette_action.h"
 
#ifndef NCOLOR
#define NCOLOR 3
#endif
 
#ifdef HMCACTION
#define HMCS HMCACTION
#else
#define HMCS 1
#endif
 
#if HMCS == 1
#include "gauge/bulk_prevention_action.h"
#elif HMCS > 1
#include "gauge/wilson_line_and_force.h"
#if HMCS < 5
#include "gauge/improved_action.h"
#elif HMCS == 5
#include "gauge/log_plaquette_action.h"
#endif
#endif
 
 
#include "gauge/gradient_flow.h"
#include "tools/string_format.h"
#include "tools/floating_point_epsilon.h"
 
#ifdef STOUTSMEAR
hila::timer sms_timer("Smearing (action)");
hila::timer smf_timer("Smearing (force)");
#if defined STOUTMODE && STOUTMODE>0
#if STOUTMODE==1
#include "gauge/stout_smear_nch.h"
#elif STOUTMODE==2
#include "gauge/stout_smear_nchm.h"
#else
#include "gauge/stout_smear.h"
#endif
#else
#define STOUTMODE 0
#include "gauge/stout_smear.h"
#endif
#define STOUTSTEPS STOUTSMEAR
#else
#define STOUTSTEPS 0
#endif
 
 
using ftype = double;
using mygroup = SU<NCOLOR, ftype>;
 
ftype stoutc = 0.5;
int stout_nsteps = STOUTSTEPS;
 
 
// define a struct to hold the input parameters: this
// makes it simpler to pass the values around
struct parameters {
    ftype beta;         // inverse gauge coupling
    ftype dt;           // HMC time step
    int trajlen;        // number of HMC time steps per trajectory
    int n_traj;         // number of trajectories to generate
    int n_therm;        // number of thermalization trajectories (counts only accepted traj.)
    int gflow_freq;     // number of trajectories between gflow measurements
    ftype gflow_max_l;  // flow scale at which gradient flow stops
    ftype gflow_l_step; // flow scale interval between flow measurements
    ftype gflow_a_accu; // desired absolute accuracy of gradient flow integration steps
    ftype gflow_r_accu; // desired relative accuracy of gradient flow integration steps
    int n_save;         // number of trajectories between config. check point
    std::string config_file;
    ftype time_offset;
};
 
 
///////////////////////////////////////////////////////////////////////////////////
// HMC functions
 
template <typename group, typename atype = hila::arithmetic_type<group>>
double measure_s(const GaugeField<group> &U) {
    // compute the value of the chosen gauge action (without beta/N factor)
#if STOUTSTEPS > 0
    sms_timer.start();
    GaugeField<group> tU;
#if STOUTMODE == 0
    stout_smear(U, tU, stoutc, stout_nsteps);
#elif STOUTMODE == 1
    nch_stout_smear(U, tU, stoutc, stout_nsteps);
#elif STOUTMODE == 2
    std::vector<std::array<Field<int>, NDIM>> niterl(stout_nsteps);
    nchm_stout_smear(U, tU, niterl, stoutc);
    int maxiter = 0;
    for(int i=0; i<niterl.size(); ++i) {
        maxiter = 0;
        Field<double> titerl;
        foralldir(d) {
            onsites(ALL) {
                titerl[X] = (double)niterl[i][d][X];
            }
            int tmaxiter = titerl.max();
            if(tmaxiter>maxiter) {
                maxiter = tmaxiter;
            }
        }
        hila::out0 << "stout step "<<i<<" exp niter: " << maxiter << "\n";
    }
#else
    stout_smear(U, tU, stoutc, stout_nsteps);
#endif
    sms_timer.stop();
#else
 
    const GaugeField<group> &tU = U;
 
#endif
 
#if HMCS == 1 // BP
    double res = measure_s_bp(tU);
#elif HMCS == 2 // LW
    atype c12 = -1.0 / 12.0; // rectangle weight
    atype c11 = 1.0 - 8.0 * c12; // plaquette weight
    double res = measure_s_impr(tU, c11, c12);
#elif HMCS == 3 // IWASAKI
    atype c12 = -0.331;          // rectangle weight
    atype c11 = 1.0 - 8.0 * c12; // plaquette weight
    double res = measure_s_impr(tU, c11, c12);
#elif HMCS == 4 // DBW2
    atype c12 = -1.4088;         // rectangle weight
    atype c11 = 1.0 - 8.0 * c12; // plaquette weight
    double res = measure_s_impr(tU, c11, c12);
#elif HMCS == 5 // LOG-PLAQUETTE
    double res = measure_s_log_plaq(tU);
#else           // WILSON
    double res = measure_s_wplaq(tU);
#endif // END HMCS
 
    return res;
}
 
template <typename group, typename atype = hila::arithmetic_type<group>>
void update_E(const GaugeField<group> &U, VectorField<Algebra<group>> &E, atype delta) {
    // compute the force for the chosen action and use it to evolve E
    atype eps = delta / group::size();
 
#if STOUTSTEPS > 0
    sms_timer.start();
    std::vector<GaugeField<group>> tUl(stout_nsteps + 1);
    std::vector<VectorField<group>> tstapl(stout_nsteps);
 
#if STOUTMODE == 0
    stout_smear(U, tUl, tstapl, stoutc);
#elif STOUTMODE == 1
    nch_stout_smear(U, tUl, tstapl, stoutc);
#elif STOUTMODE == 2
    std::vector<std::array<Field<int>, NDIM>> niterl(stout_nsteps);
    nchm_stout_smear(U, tUl, tstapl, niterl, stoutc);
#else
    std::vector<VectorField<group>> tUKl(stout_nsteps);
    stout_smeark(U, tUl, tstapl, tUKl, stoutc);
#endif
 
    VectorField<Algebra<group>> tE;
 
    foralldir(d1) onsites(ALL) tE[d1][X] = 0;
 
    GaugeField<group> &tU = tUl[stout_nsteps];
    sms_timer.stop();
 
#else // STOUTSTEPS==0
 
    const GaugeField<group> &tU = U;
    VectorField<Algebra<group>> &tE = E;
 
#endif // END STOUTSTEPS
 
#if HMCS == 1 // BP
    get_force_bp_add(tU, tE, eps);
#elif HMCS == 2 // LW
    atype c12 = -1.0 / 12.0;     // rectangle weight
    atype c11 = 1.0 - 8.0 * c12; // plaquette weight
    get_force_impr_add(tU, tE, eps * c11, eps * c12);
#elif HMCS == 3 // IWASAKI
    atype c12 = -0.331;          // rectangle weight
    atype c11 = 1.0 - 8.0 * c12; // plaquette weight
    get_force_impr_add(tU, tE, eps * c11, eps * c12);
#elif HMCS == 4 // DBW2
    atype c12 = -1.4088;         // rectangle weight
    atype c11 = 1.0 - 8.0 * c12; // plaquette weight
    get_force_impr_add(tU, tE, eps * c11, eps * c12);
#elif HMCS == 5 // LOG-PLAQUETTE
    get_force_log_plaq_add(tU, tE, eps);
#else           // WILSON
    get_force_wplaq_add(tU, tE, eps);
#endif          // END HMCS
 
#if STOUTSTEPS > 0
    smf_timer.start();
    VectorField<Algebra<group>> KS;
 
#if STOUTMODE == 0
    stout_smear_force(tUl, tstapl, tE, KS, stoutc);
#elif STOUTMODE == 1
    nch_stout_smear_force(tUl, tstapl, tE, KS, stoutc);
#elif STOUTMODE == 2
    nchm_stout_smear_force(tUl, tstapl, tE, KS, niterl, stoutc);
#else
    stout_smeark_force(tUl, tstapl, tUKl, tE, KS, stoutc);
#endif
 
    foralldir(d1) {
        onsites(ALL) E[d1][X] += KS[d1][X];
    }
    smf_timer.stop();
#endif // END STOUTSTEPS
 
}
 
template <typename group, typename atype = hila::arithmetic_type<group>>
void update_U(GaugeField<group> &U, const VectorField<Algebra<group>> &E, atype delta) {
    // evolve U with momentum E over time step delta
    foralldir(d) {
        onsites(ALL) U[d][X] = chexp(E[d][X].expand_scaled(delta)) * U[d][X];
    }
}
 
template <typename group>
double measure_e2(const VectorField<Algebra<group>> &E) {
    // compute gauge kinetic energy from momentum field E
    Reduction<double> e2 = 0;
    e2.allreduce(false).delayed(true);
    foralldir(d) {
        onsites(ALL) e2 += E[d][X].squarenorm();
    }
    return e2.value() / 2;
}
 
template <typename group>
double measure_action(const GaugeField<group> &U, const VectorField<Algebra<group>> &E,
                     const parameters &p, double &plaq) {
    // measure the total action, consisting of plaquette and momentum term
    plaq = p.beta * measure_s(U);
    double e2 = measure_e2(E);
    return plaq + e2 / 2;
}
 
template <typename group>
double measure_action(const GaugeField<group> &U, const VectorField<Algebra<group>> &E,
                     const parameters &p) {
    // measure the total action, consisting of plaquette and momentum term
    double plaq = 0;
    return measure_action(U, E, p, plaq);
}
 
 
template <typename group, typename atype = hila::arithmetic_type<group>>
void do_trajectory(GaugeField<group> &U, VectorField<Algebra<group>> &E, const parameters &p) {
    // leap frog integration for normal action
 
    // start trajectory: advance U by half a time step
    update_U(U, E, p.dt / 2);
    // main trajectory integration:
    for (int n = 0; n < p.trajlen - 1; ++n) {
        update_E(U, E, p.beta * p.dt);
        update_U(U, E, p.dt);
    }
    // end trajectory: bring U and E to the same time
    update_E(U, E, p.beta * p.dt);
    update_U(U, E, p.dt / 2);
 
    U.reunitarize_gauge();
}
 
// end HMC functions
///////////////////////////////////////////////////////////////////////////////////
// measurement functions
 
template <typename group>
void measure_stuff(const GaugeField<group> &U, const VectorField<Algebra<group>> &E) {
    // perform measurements on current gauge and momentum pair (U, E) and
    // print results in formatted form to standard output
    static bool first = true;
    if (first) {
        // print legend for measurement output
        hila::out0
            << "LMEAS:     s-local          plaq           E^2        P.real        P.imag\n";
        first = false;
    }
    auto slocal = measure_s(U) / (lattice.volume() * NDIM * (NDIM - 1) / 2);
    auto plaq = measure_s_wplaq(U) / (lattice.volume() * NDIM * (NDIM - 1) / 2);
    auto e2 = measure_e2(E) / (lattice.volume() * NDIM);
    auto poly = measure_polyakov(U, e_t);
    hila::out0 << string_format("MEAS % 0.6e % 0.6e % 0.6e % 0.6e % 0.6e", slocal, plaq, e2,
                                poly.real(), poly.imag())
               << '\n';
}
 
// end measurement functions
///////////////////////////////////////////////////////////////////////////////////
// load/save config functions
 
template <typename group>
void checkpoint(const GaugeField<group> &U, int trajectory, const parameters &p) {
    double t = hila::gettime();
    // name of config with extra suffix
    std::string config_file =
        p.config_file + "_" + std::to_string(abs((trajectory + 1) / p.n_save) % 2);
    // save config
    U.config_write(config_file);
    // write run_status file
    if (hila::myrank() == 0) {
        std::ofstream outf;
        outf.open("run_status", std::ios::out | std::ios::trunc);
        outf << "trajectory  " << trajectory + 1 << '\n';
        outf << "seed        " << static_cast<uint64_t>(hila::random() * (1UL << 61)) << '\n';
        outf << "time        " << hila::gettime() << '\n';
        // write config name to status file:
        outf << "config name  " << config_file << '\n';
        outf.close();
    }
    std::stringstream msg;
    msg << "Checkpointing, time " << hila::gettime() - t;
    hila::timestamp(msg.str().c_str());
}
 
template <typename group>
bool restore_checkpoint(GaugeField<group> &U, int &trajectory, parameters &p) {
    uint64_t seed;
    bool ok = true;
    p.time_offset = 0;
    hila::input status;
    if (status.open("run_status", false, false)) {
        hila::out0 << "RESTORING FROM CHECKPOINT:\n";
        trajectory = status.get("trajectory");
        seed = status.get("seed");
        p.time_offset = status.get("time");
        // get config name with suffix from status file:
        std::string config_file = status.get("config name");
        status.close();
        hila::seed_random(seed);
        U.config_read(config_file);
        ok = true;
    } else {
        std::ifstream in;
        in.open(p.config_file, std::ios::in | std::ios::binary);
        if (in.is_open()) {
            in.close();
            hila::out0 << "READING initial config\n";
            U.config_read(p.config_file);
            ok = true;
        } else {
            ok = false;
        }
    }
    return ok;
}
 
// end load/save config functions
///////////////////////////////////////////////////////////////////////////////////
 
 
int main(int argc, char **argv) {
 
    // hila::initialize should be called as early as possible
    hila::initialize(argc, argv);
 
    // hila provides an input class hila::input, which is
    // a convenient way to read in parameters from input files.
    // parameters are presented as key - value pairs, as an example
    //  " lattice size  64, 64, 64, 64"
    // is read below.
    //
    // Values are broadcast to all MPI nodes.
    //
    // .get() -method can read many different input types,
    // see file "input.h" for documentation
 
#if HMCS == 1
    hila::out0 << "SU(" << mygroup::size() << ") HMC with bulk-prevention gauge action\n";
#elif HMCS == 2
    hila::out0 << "SU(" << mygroup::size() << ") HMC with Luscher-Weisz gauge action\n";
#elif HMCS == 3
    hila::out0 << "SU(" << mygroup::size() << ") HMC with Iwasaki gauge action\n";
#elif HMCS == 4
    hila::out0 << "SU(" << mygroup::size() << ") HMC with DBW2 gauge action\n";
#elif HMCS == 5
    hila::out0 << "SU(" << mygroup::size() << ") HMC with log-plaquette gauge action\n";
#elif HMCS == 6
    hila::out0 << "SU(" << mygroup::size() << ") HMC with topology suppressing gauge action\n";
#else
    hila::out0 << "SU(" << mygroup::size() << ") HMC with Wilson's plaquette gauge action\n";
#endif
 
    hila::out0 << "Using floating point epsilon: " << fp<ftype>::epsilon << "\n";
 
    parameters p;
 
    hila::input par("parameters");
 
    CoordinateVector lsize;
    // reads NDIM numbers
    lsize = par.get("lattice size");
    // inverse gauge coupling
    p.beta = par.get("beta");
    // HMC step size
    p.dt = par.get("dt");
    // trajectory length in steps
    p.trajlen = par.get("trajectory length");
    // number of trajectories
    p.n_traj = par.get("number of trajectories");
    // number of thermalization trajectories
    p.n_therm = par.get("thermalization trajs");
    // wilson flow frequency (number of traj. between w. flow measurement)
    p.gflow_freq = par.get("gflow freq");
    // wilson flow max. flow-distance
    p.gflow_max_l = par.get("gflow max lambda");
    // wilson flow flow-distance step size
    p.gflow_l_step = par.get("gflow lambda step");
    // wilson flow absolute accuracy (per integration step)
    p.gflow_a_accu = par.get("gflow abs. accuracy");
    // wilson flow relative accuracy (per integration step)
    p.gflow_r_accu = par.get("gflow rel. accuracy");
    // random seed = 0 -> get seed from time
    long seed = par.get("random seed");
    // save config and checkpoint
    p.n_save = par.get("trajs/saved");
    // measure surface properties and print "profile"
    p.config_file = par.get("config name");
 
    par.close(); // file is closed also when par goes out of scope
 
#if STOUTSTEPS > 0
    hila::out0 << "using stout smearing: nsteps=" << stout_nsteps << "  , c=" << stoutc << "  , ";
#if STOUTMODE == 0
    hila::out0 << "mode=CH\n";
#elif STOUTMODE == 1
    hila::out0 << "mode=NCH\n";
#elif STOUTMODE == 2
    hila::out0 << "mode=NCHM\n";
#else
    hila::out0 << "mode=CHK\n";
#endif
#endif
 
    // set up the lattice
    lattice.setup(lsize);
 
    // We need random number here
    hila::seed_random(seed);
 
    // Alloc gauge field and momenta (E)
    GaugeField<mygroup> U;
    VectorField<Algebra<mygroup>> E;
 
    int start_traj = 0;
    if (!restore_checkpoint(U, start_traj, p)) {
        U = 1;
        hila::out0 << "cold start: initial link variables set to identity\n";
    }
 
    hila::timer update_timer("Updates");
    hila::timer measure_timer("Measurements");
    hila::timer gf_timer("Gradient Flow");
 
    auto orig_dt = p.dt;
    auto orig_trajlen = p.trajlen;
    GaugeField<mygroup> U_old;
    int nreject = 0;
    ftype t_step0 = 0.0;
    double g_act_old, act_old, g_act_new, act_new;
    g_act_old = p.beta * measure_s(U);
 
    for (int trajectory = start_traj; trajectory < p.n_traj; ++trajectory) {
        if (trajectory < p.n_therm) {
            // during thermalization: start with 10% of normal step size (and trajectory length)
            // and increse linearly with number of accepted thermalization trajectories. normal
            // step size is reached after 3/4*p.n_therm accepted thermalization trajectories.
            if (trajectory < p.n_therm * 3.0 / 4.0) {
                p.dt = orig_dt * (0.1 + 0.9 * 4.0 / 3.0 * trajectory / p.n_therm);
                p.trajlen = orig_trajlen;
            } else {
                p.dt = orig_dt;
                p.trajlen = orig_trajlen;
            }
            if (nreject > 1) {
                // if two consecutive thermalization trajectories are rejected, decrese the
                // step size by factor 0.5 (but keeping trajectory length constant, since
                // thermalization becomes inefficient if trajectory length decreases)
                for (int irej = 0; irej < nreject - 1; ++irej) {
                    p.dt *= 0.5;
                    p.trajlen *= 2;
                }
                hila::out0 << " thermalization step size(reduzed due to multiple reject) dt="
                           << std::setprecision(8) << p.dt << '\n';
            } else {
                hila::out0 << " thermalization step size dt=" << std::setprecision(8) << p.dt
                           << '\n';
            }
        } else if (trajectory == p.n_therm) {
            p.dt = orig_dt;
            p.trajlen = orig_trajlen;
            hila::out0 << " normal stepsize dt=" << std::setprecision(8) << p.dt << '\n';
        }
 
        update_timer.start();
 
        U_old = U;
        ftype ttime = hila::gettime();
 
        foralldir(d) onsites(ALL) E[d][X].gaussian_random();
 
        act_old = g_act_old + measure_e2(E) / 2;
 
        do_trajectory(U, E, p);
 
        act_new = measure_action(U, E, p, g_act_new);
 
 
        bool reject = hila::broadcast(exp(act_old - act_new) < hila::random());
 
        if (trajectory < p.n_therm) {
            // during thermalization: keep track of number of rejected trajectories
            if (reject) {
                ++nreject;
                --trajectory;
            } else {
                if (nreject > 0) {
                    --nreject;
                }
            }
        }
 
        hila::out0 << std::setprecision(12) << "HMC " << trajectory << " S_TOT_start " << act_old
                   << " dS_TOT " << std::setprecision(6) << act_new - act_old
                   << std::setprecision(12);
        if (reject) {
            hila::out0 << " REJECT" << " --> S_GAUGE " << g_act_old;
            U = U_old;
        } else {
            hila::out0 << " ACCEPT" << " --> S_GAUGE " << g_act_new;
            g_act_old = g_act_new;
        }
        update_timer.stop();
 
        hila::out0 << "  time " << std::setprecision(3) << hila::gettime() - ttime << '\n';
 
 
        hila::out0 << "Measure_start " << trajectory << '\n';
 
        measure_timer.start();
 
        measure_stuff(U, E);
 
        measure_timer.stop();
 
        hila::out0 << "Measure_end " << trajectory << '\n';
 
        if (trajectory >= p.n_therm) {
 
            if (p.gflow_freq > 0 && trajectory % p.gflow_freq == 0) {
                int gtrajectory = trajectory / p.gflow_freq;
                if (p.gflow_l_step > 0) {
 
                    gf_timer.start();
 
                    int nflow_steps = (int)(p.gflow_max_l / p.gflow_l_step);
 
                    ftype gftime = hila::gettime();
                    hila::out0 << "Gflow_start " << gtrajectory << '\n';
 
                    GaugeField<mygroup> V = U;
 
                    ftype t_step = t_step0;
                    measure_gradient_flow_stuff(V, (ftype)0.0, t_step);
                    t_step = do_gradient_flow_adapt(V, (ftype)0.0, p.gflow_l_step, p.gflow_a_accu,
                                                    p.gflow_r_accu, t_step);
                    measure_gradient_flow_stuff(V, p.gflow_l_step, t_step);
                    t_step0 = t_step;
 
                    for (int i = 1; i < nflow_steps; ++i) {
 
                        t_step =
                            do_gradient_flow_adapt(V, i * p.gflow_l_step, (i + 1) * p.gflow_l_step,
                                                   p.gflow_a_accu, p.gflow_r_accu, t_step);
 
                        measure_gradient_flow_stuff(V, (i + 1) * p.gflow_l_step, t_step);
                    }
 
                    gf_timer.stop();
 
                    hila::out0 << "Gflow_end " << gtrajectory << "    time " << std::setprecision(3)
                               << hila::gettime() - gftime << '\n';
                }
            }
        }
 
        if (p.n_save > 0 && trajectory>=0 && (trajectory + 1) % p.n_save == 0) {
            checkpoint(U, trajectory, p);
        }
    }
 
    hila::finishrun();
}