exchangeoperator.h

Go to the documentation of this file.

#ifndef SRC_APPS_CHEM_EXCHANGEOPERATOR_H_
#define SRC_APPS_CHEM_EXCHANGEOPERATOR_H_
 
#include<madness.h>
#include<madness/world/cloud.h>
#include<madness/mra/macrotaskq.h>
#include<madness/chem/SCFOperators.h>
 
namespace madness {
 
// forward declaration
class SCF;
class Nemo;
 
 
template<typename T, std::size_t NDIM>
class Exchange<T,NDIM>::ExchangeImpl {
    typedef Function<T, NDIM> functionT;
    typedef std::vector<functionT> vecfuncT;
 
    static inline std::atomic<long> apply_timer;
    static inline std::atomic<long> mul2_timer;
    static inline std::atomic<long> mul1_timer; ///< timing
    static inline double elapsed_time;
 
    static void reset_timer() {
        mul1_timer = 0l;
        mul2_timer = 0l;
        apply_timer = 0l;
        elapsed_time = 0.0;
    }
 
public:
    nlohmann::json gather_timings(World& world) const {
        double t1 = double(mul1_timer) * 0.001;
        double t2 = double(apply_timer) * 0.001;
        double t3 = double(mul2_timer) * 0.001;
        world.gop.sum(t1);
        world.gop.sum(t2);
        world.gop.sum(t3);
        nlohmann::json j;
        j["multiply1"] = t1;
        j["apply"] = t2;
        j["multiply2"] = t3;
        j["total"] = elapsed_time;
        return j;
    }
 
    void print_timer(World& world) const {
        auto timings= gather_timings(world);
        if (world.rank() == 0) {
            printf(" cpu time spent in multiply1   %8.2fs\n", timings["multiply1"].template get<double>());
            printf(" cpu time spent in apply       %8.2fs\n", timings["apply"].template get<double>());
            printf(" cpu time spent in multiply2   %8.2fs\n", timings["multiply2"].template get<double>());
            printf(" total wall time               %8.2fs\n", timings["total"].template get<double>());
        }
    }
 
 
    typedef Exchange<T,NDIM>::ExchangeAlgorithm Algorithm;
    Algorithm algorithm_ = multiworld_efficient_row;
    MacroTaskInfo macro_task_info = MacroTaskInfo::preset("default");
 
    /// default ctor
    ExchangeImpl(World& world, const double lo, const double thresh) : world(world), lo(lo), thresh(thresh) {}
 
    /// ctor with a conventional calculation
    ExchangeImpl(World& world, const SCF *calc, const int ispin) ;
 
    /// ctor with a nemo calculation
    ExchangeImpl(World& world, const Nemo *nemo, const int ispin);
 
    /// set the bra and ket orbital spaces, and the occupation
 
    /// @param[in]  bra     bra space, must be provided as complex conjugate
    /// @param[in]  ket     ket space
    void set_bra_and_ket(const vecfuncT& bra, const vecfuncT& ket) {
        mo_bra = copy(world, bra);
        mo_ket = copy(world, ket);
    }
 
    std::string info() const {return "K";}
 
    static auto set_poisson(World& world, const double lo, const double econv = FunctionDefaults<3>::get_thresh()) {
        return std::shared_ptr<real_convolution_3d>(CoulombOperatorPtr(world, lo, econv));
    }
 
    /// apply the exchange operator on a vector of functions
 
    /// note that only one spin is used (either alpha or beta orbitals)
    /// @param[in]  vket       the orbitals |i> that the operator is applied on
    /// @return     a vector of orbitals  K| i>
    vecfuncT operator()(const vecfuncT& vket) const;
 
    bool is_symmetric() const { return symmetric_; }
 
    ExchangeImpl& set_taskq(std::shared_ptr<MacroTaskQ> taskq1) {
        this->taskq=taskq1;
        return *this;
    }
 
    ExchangeImpl& symmetric(const bool flag) {
        symmetric_ = flag;
        return *this;
    }
 
    ExchangeImpl& set_macro_task_info(const MacroTaskInfo& info) {
        macro_task_info = info;
        return *this;
    }
 
    ExchangeImpl& set_algorithm(const Algorithm& alg) {
        algorithm_ = alg;
        return *this;
    }
 
    ExchangeImpl& set_printlevel(const long& level) {
        printlevel=level;
        return *this;
    }
 
    std::shared_ptr<MacroTaskQ> get_taskq() const {return taskq;}
 
    World& get_world() const {return world;}
 
    nlohmann::json get_statistics() const {return statistics;}
 
    /// return some statistics about the current settings
    nlohmann::json gather_statistics() const {
        nlohmann::json j;
        j["symmetric"] = symmetric_;
        j["lo"] = lo;
        j["thresh"] = thresh;
        j["mul_tol"] = mul_tol;
        j["printlevel"] = printlevel;
        j["algorithm"] = to_string(algorithm_);
        j["macro_task_info"] = macro_task_info.to_json();
        auto timings = gather_timings(world);
        j.update(timings);
        return j;
    }
 
private:
 
    /// exchange using macrotasks, i.e. apply K on a function in individual worlds
    vecfuncT K_macrotask_efficient(const vecfuncT& vket, const double mul_tol = 0.0) const;
 
    /// exchange using macrotasks, i.e. apply K on a function in individual worlds row-wise
    vecfuncT K_macrotask_efficient_row(const vecfuncT& vket, const double mul_tol = 0.0) const;
 
    /// computing the full square of the double sum (over vket and the K orbitals)
    vecfuncT K_small_memory(const vecfuncT& vket, const double mul_tol = 0.0) const;
 
    /// computing the upper triangle of the double sum (over vket and the K orbitals)
    vecfuncT K_large_memory(const vecfuncT& vket, const double mul_tol = 0.0) const;
 
    /// computing the upper triangle of the double sum (over vket and the K orbitals)
    static vecfuncT compute_K_tile(World& world, const vecfuncT& mo_bra, const vecfuncT& mo_ket,
                                   const vecfuncT& vket, std::shared_ptr<real_convolution_3d> poisson,
                                   const bool symmetric, const double mul_tol = 0.0);
 
    inline bool printdebug() const {return printlevel >= 10; }
    inline bool printprogress() const {return (printlevel>=4) and (not (printdebug()));}
    inline bool printtimings() const {return printlevel>=3;}
    inline bool printtimings_detail() const {return printlevel>=4;}
 
    World& world;
    std::shared_ptr<MacroTaskQ> taskq;
    bool symmetric_ = false;      /// is the exchange matrix symmetric? K phi_i = \sum_k \phi_k \int \phi_k \phi_i
    vecfuncT mo_bra, mo_ket;    ///< MOs for bra and ket
    double lo = 1.e-4;
    double thresh = FunctionDefaults<NDIM>::get_thresh();
    long printlevel = 0;
    double mul_tol = FunctionDefaults<NDIM>::get_thresh()*0.1;
 
    mutable nlohmann::json statistics;  ///< statistics of the Cloud (timings, memory)  and of the parameters of this run
 
    class MacroTaskExchangeSimple : public MacroTaskOperationBase {
 
        long nresult;
        double lo = 1.e-4;
        double mul_tol = 1.e-7;
        bool symmetric = false;
 
        /// custom partitioning for the exchange operator in exchangeoperator.h
 
        /// arguments are: result[i] += sum_k vket[k] \int 1/r vbra[k] f[i]
        /// with f and vbra being batched, result and vket being passed on as a whole
        class MacroTaskPartitionerExchange : public MacroTaskPartitioner {
        public:
            MacroTaskPartitionerExchange(const bool symmetric) : symmetric(symmetric) {
                max_batch_size=30;
            }
 
            bool symmetric = false;
 
            partitionT do_partitioning(const std::size_t& vsize1, const std::size_t& vsize2,
                                       const std::string policy) const override {
 
                partitionT partition1 = do_1d_partition(vsize1, policy);
                partitionT partition2 = do_1d_partition(vsize2, policy);
                partitionT result;
                for (auto i = partition1.begin(); i != partition1.end(); ++i) {
                    if (symmetric) {
                        for (auto j = i; j != partition1.end(); ++j) {
                            Batch batch(i->first.input[0], j->first.input[0], _);
                            double priority=compute_priority(batch);
                            result.push_back(std::make_pair(batch,priority));
                        }
                    } else {
                        for (auto j = partition2.begin(); j != partition2.end(); ++j) {
                            Batch batch(i->first.input[0], j->first.input[0], _);
                            double priority=compute_priority(batch);
                            result.push_back(std::make_pair(batch,priority));
                        }
                    }
                }
                return result;
            }
 
            /// compute the priority of this task for non-dumb scheduling
 
            /// \return the priority as double number (no limits)
            double compute_priority(const Batch& batch) const override {
                MADNESS_CHECK(batch.input.size() == 2);   // must be quadratic batches
                long nrow = batch.input[0].size();
                long ncol = batch.input[1].size();
                return double(nrow * ncol);
            }
        };
 
    public:
        MacroTaskExchangeSimple(const long nresult, const double lo, const double mul_tol, const bool symmetric)
                : nresult(nresult), lo(lo), mul_tol(mul_tol), symmetric(symmetric) {
            partitioner.reset(new MacroTaskPartitionerExchange(symmetric));
        }
 
 
        // you need to define the exact argument(s) of operator() as tuple
        typedef std::tuple<const std::vector<Function<T, NDIM>>&,
                const std::vector<Function<T, NDIM>>&,
                const std::vector<Function<T, NDIM>>&> argtupleT;
 
        using resultT = std::vector<Function<T, NDIM>>;
 
        // you need to define an empty constructor for the result
        // resultT must implement operator+=(const resultT&)
        resultT allocator(World& world, const argtupleT& argtuple) const {
            std::size_t n = std::get<0>(argtuple).size();
            resultT result = zero_functions_compressed<T, NDIM>(world, n);
            return result;
        }
 
        std::vector<Function<T, NDIM>>
        operator()(const std::vector<Function<T, NDIM>>& vf_batch,     // will be batched (column)
                   const std::vector<Function<T, NDIM>>& bra_batch,    // will be batched (row)
                   const std::vector<Function<T, NDIM>>& vket) {       // will not be batched
 
            World& world = vf_batch.front().world();
            resultT Kf = zero_functions_compressed<T, NDIM>(world, nresult);
 
            bool diagonal_block = batch.input[0] == batch.input[1];
            auto& bra_range = batch.input[1];    // corresponds to vbra
            auto& vf_range = batch.input[0];       // corresponds to vf_batch
 
            if (vf_range.is_full_size()) vf_range.end = vf_batch.size();
            if (bra_range.is_full_size()) bra_range.end = bra_batch.size();
 
            MADNESS_CHECK(vf_range.end <= nresult);
            if (symmetric) MADNESS_CHECK(bra_range.end <= nresult);
 
            if (symmetric and diagonal_block) {
                auto ket_batch = bra_range.copy_batch(vket);
                vecfuncT resultcolumn = compute_diagonal_batch_in_symmetric_matrix(world, ket_batch, bra_batch,
                                                                                   vf_batch);
 
                for (int i = vf_range.begin; i < vf_range.end; ++i){
                    Kf[i] += resultcolumn[i - vf_range.begin];}
 
            } else if (symmetric and not diagonal_block) {
                auto[resultcolumn, resultrow]=compute_offdiagonal_batch_in_symmetric_matrix(world, vket, bra_batch,
                                                                                            vf_batch);
 
                for (int i = bra_range.begin; i < bra_range.end; ++i){
                    Kf[i] += resultcolumn[i - bra_range.begin];}
                for (int i = vf_range.begin; i < vf_range.end; ++i){
                    Kf[i] += resultrow[i - vf_range.begin];}
            } else {
                auto ket_batch = bra_range.copy_batch(vket);
                vecfuncT resultcolumn = compute_batch_in_asymmetric_matrix(world, ket_batch, bra_batch, vf_batch);
                for (int i = vf_range.begin; i < vf_range.end; ++i)
                    Kf[i] += resultcolumn[i - vf_range.begin];
            }
            return Kf;
        }
 
        /// compute a batch of the exchange matrix, with identical ranges, exploiting the matrix symmetry
 
        /// \param subworld     the world we're computing in
        /// \param cloud        where to store the results
        /// \param bra_batch    the bra batch of orbitals (including the nuclear correlation factor square)
        /// \param ket_batch    the ket batch of orbitals, i.e. the orbitals to premultiply with
        /// \param vf_batch     the argument of the exchange operator
        vecfuncT compute_diagonal_batch_in_symmetric_matrix(World& subworld,
                                                            const vecfuncT& ket_batch,      // is batched
                                                            const vecfuncT& bra_batch,      // is batched
                                                            const vecfuncT& vf_batch        // is batched
        ) const {
            double mul_tol = 0.0;
            double symmetric = true;
            auto poisson = Exchange<double, 3>::ExchangeImpl::set_poisson(subworld, lo);
            return Exchange<T, NDIM>::ExchangeImpl::compute_K_tile(subworld, bra_batch, ket_batch, vf_batch, poisson, symmetric,
                                                     mul_tol);
        }
 
        /// compute a batch of the exchange matrix, with non-identical ranges
 
        /// \param subworld     the world we're computing in
        /// \param cloud        where to store the results
        /// \param bra_batch    the bra batch of orbitals (including the nuclear correlation factor square)
        /// \param ket_batch    the ket batch of orbitals, i.e. the orbitals to premultiply with
        /// \param vf_batch     the argument of the exchange operator
        vecfuncT compute_batch_in_asymmetric_matrix(World& subworld,
                                                    const vecfuncT& ket_batch,
                                                    const vecfuncT& bra_batch,
                                                    const vecfuncT& vf_batch) const {
            double mul_tol = 0.0;
            double symmetric = false;
            auto poisson = Exchange<double, 3>::ExchangeImpl::set_poisson(subworld, lo);
            return Exchange<T, NDIM>::ExchangeImpl::compute_K_tile(subworld, bra_batch, ket_batch, vf_batch, poisson, symmetric,
                                                     mul_tol);
        }
 
        /// compute a batch of the exchange matrix, with non-identical ranges
 
        /// \param subworld     the world we're computing in
        /// \param cloud        where to store the results
        /// \param bra_batch    the bra batch of orbitals (including the nuclear correlation factor square)
        /// \param ket_batch    the ket batch of orbitals, i.e. the orbitals to premultiply with
        /// \param vf_batch     the argument of the exchange operator
        std::pair<vecfuncT, vecfuncT> compute_offdiagonal_batch_in_symmetric_matrix(World& subworld,
                                                                                    const vecfuncT& ket, // not batched
                                                                                    const vecfuncT& bra_batch, // batched
                                                                                    const vecfuncT& vf_batch) const; // batched
 
    };
 
    class MacroTaskExchangeRow : public MacroTaskOperationBase {
 
        long nresult;
        double lo = 1.e-4;
        double mul_tol = 1.e-7;
        bool symmetric = false;
        Algorithm algorithm_;
 
        /// custom partitioning for the exchange operator in exchangeoperator.h
        class MacroTaskPartitionerRow : public MacroTaskPartitioner {
        public:
            MacroTaskPartitionerRow() {
              max_batch_size=1;
            }       
        };
 
    public:
        MacroTaskExchangeRow(const long nresult, const double lo, const double mul_tol, const Algorithm algorithm)
                : nresult(nresult), lo(lo), mul_tol(mul_tol),  algorithm_(algorithm) {
            partitioner.reset(new MacroTaskPartitionerRow());
            name="MacroTaskExchangeRow";
        }
 
        // you need to define the exact argument(s) of operator() as tuple
        typedef std::tuple<const std::vector<Function<T, NDIM>>&,
                           const std::vector<Function<T, NDIM>>&,
                           const std::vector<Function<T, NDIM>>&> argtupleT;
 
        using resultT = std::vector<Function<T, NDIM>>;
 
        // you need to define an empty constructor for the result
        // resultT must implement operator+=(const resultT&)
        resultT allocator(World& world, const argtupleT& argtuple) const {
            std::size_t n = std::get<0>(argtuple).size();
            resultT result = zero_functions_compressed<T, NDIM>(world, n);
            return result;
        }
 
        /// compute exchange row-wise for a fixed orbital phi_i of vket
 
        /// create 2 worlds: one fetches the function coefficients from the universe, the other
        /// does the computation, then swap. The result is copied back to the universe
        std::vector<Function<T, NDIM>>
        operator()(const std::vector<Function<T, NDIM>>& vket,
                   const std::vector<Function<T, NDIM>>& mo_bra, 
                   const std::vector<Function<T, NDIM>>& mo_ket) {
            std::vector<Function<T,NDIM>> result;
            if (algorithm_==fetch_compute) {
                result=row_fetch_compute(vket,mo_bra,mo_ket);
            } else if (algorithm_==multiworld_efficient_row) {
                result=row(vket,mo_bra,mo_ket);
            } else {
                MADNESS_EXCEPTION("unknown algorithm in Exchange::MacroTaskExchangeRow::operator()",1);
            }
            return result;
        }
 
        std::vector<Function<T,NDIM>>
        row(const std::vector<Function<T, NDIM>>& vket,
            const std::vector<Function<T, NDIM>>& mo_bra,
            const std::vector<Function<T, NDIM>>& mo_ket) {
 
            double cpu0, cpu1;
            World& world = vket.front().world();
            mul_tol = 0.0;
 
            resultT Kf = zero_functions_compressed<T, NDIM>(world, 1);
            vecfuncT psif = zero_functions_compressed<T,NDIM>(world, mo_bra.size());
            auto poisson = Exchange<double, 3>::ExchangeImpl::set_poisson(world, lo);
 
            // !! NO !! vket is batched, starts at batch.input[0].begin
            // auto& i = batch.input[0].begin;
            long i=0;
            MADNESS_CHECK_THROW(vket.size()==1,"out-of-bounds error in Exchange::MacroTaskExchangeRow::operator()");
            size_t min_tile = 10;
            size_t ntile = std::min(mo_bra.size(), min_tile);
 
            for (size_t ilo=0; ilo<mo_bra.size(); ilo+=ntile){
                cpu0 = cpu_time();
                size_t iend = std::min(ilo+ntile,mo_bra.size());
                vecfuncT tmp_mo_bra(mo_bra.begin()+ilo,mo_bra.begin()+iend);
                auto tmp_psif = mul_sparse(world, vket[i], tmp_mo_bra, mul_tol);
                truncate(world, tmp_psif);
                cpu1 = cpu_time();
                mul1_timer += long((cpu1 - cpu0) * 1000l);
 
                cpu0 = cpu_time();
                tmp_psif = apply(world, *poisson.get(), tmp_psif);
                truncate(world, tmp_psif);
                cpu1 = cpu_time();
                apply_timer += long((cpu1 - cpu0) * 1000l);
 
                cpu0 = cpu_time();
                vecfuncT tmp_mo_ket(mo_ket.begin()+ilo,mo_ket.begin()+iend);
                auto tmp_Kf = dot(world, tmp_mo_ket, tmp_psif);
                cpu1 = cpu_time();
                mul2_timer += long((cpu1 - cpu0) * 1000l);
 
                Kf[0] += tmp_Kf;
                truncate(world, Kf);
            }
 
            return Kf;
        }
 
        std::vector<Function<T,NDIM>>
        row_fetch_compute(const std::vector<Function<T, NDIM>>& vket,
            const std::vector<Function<T, NDIM>>& mo_bra,
            const std::vector<Function<T, NDIM>>& mo_ket) {
 
            io_redirect_cout();
            double total_execution_time=0.0;
            double total_fetch_time=0.0;
            double total_fetch_spawn_time=0.0;
 
            resultT Kf = zero_functions_compressed<T, NDIM>(*subworld_ptr, 1);
            {
                // create the two worlds that will be used for fetching and computing
                // std::shared_ptr<World> executing_world(subworld_ptr);
                double cpu0=cpu_time();
                SafeMPI::Intracomm comm = subworld_ptr->mpi.comm();
                std::shared_ptr<World> fetching_world(new World(comm.Clone()));
                std::shared_ptr<World> executing_world(new World(comm.Clone()));
                double cpu1=cpu_time();
                print("time to create two worlds:",cpu1-cpu0,"seconds");
                print("executing_world.id()",executing_world->id(),"fetching_world.id()",fetching_world->id(),"in MacroTaskExchangeRow");
 
                {
                    auto poisson1 = Exchange<double, 3>::ExchangeImpl::set_poisson(*executing_world, lo);
                    auto poisson2 = Exchange<double, 3>::ExchangeImpl::set_poisson(*fetching_world, lo);
 
                    functionT phi1=copy(*executing_world,vket[0]);
                    functionT phi2=copy(*fetching_world,vket[0]);
 
                    // !! NO !! vket is batched, starts at batch.input[0].begin
                    // auto& i = batch.input[0].begin;
                    MADNESS_CHECK_THROW(vket.size()==1,"out-of-bounds error in Exchange::MacroTaskExchangeRow::operator()");
                    size_t min_tile = 10;
                    size_t ntile = std::min(mo_bra.size(), min_tile);
 
                    struct Tile {
                        size_t ilo;
                        size_t iend;
                    };
 
 
                    // copy the data from the universe bra and ket to subworld bra and ket
                    // returns a pair of vectors in the subworld which are still awaiting the function coefficients
                    auto fetch_data = [&](World& world, const Tile& tile) {
                        MADNESS_CHECK_THROW(mo_bra.size()==mo_ket.size(),
                                            "bra and ket size mismatch in Exchange::MacroTaskExchangeRow::execute()");
 
                        std::size_t sz=tile.iend-tile.ilo;
                        vecfuncT subworld_bra(sz);
                        vecfuncT subworld_ket;
                        for (size_t i=tile.ilo; i<tile.iend; ++i) {
                            auto f=copy(world,mo_bra[i],false);
                            subworld_bra[i-tile.ilo]=f;
                            subworld_ket.push_back(copy(world, mo_ket[i],false));
                        }
                        return std::make_pair(subworld_bra,subworld_ket);
                    };
 
                    // apply the exchange operator on phi for a a tile of mo_bra and mo_ket
                    auto execute = [&](World& world, auto poisson, const functionT& phi, const vecfuncT& mo_bra, const vecfuncT& mo_ket) {
                        MADNESS_CHECK_THROW(mo_bra.size()==mo_ket.size(),
                                            "bra and ket size mismatch in Exchange::MacroTaskExchangeRow::execute()");
 
                        auto world_id=world.id();
                        auto phi_id=phi.world().id();
                        auto bra_id=mo_bra.front().world().id();
                        auto ket_id=mo_ket.front().world().id();
                        std::string msg="world mismatch in Exchange::MacroTaskExchangeRow::execute(): ";
                        msg+="world.id()="+std::to_string(world_id)+", ";
                        msg+="phi.world().id()="+std::to_string(phi_id)+", ";
                        msg+="bra.world().id()="+std::to_string(bra_id)+", ";
                        msg+="ket.world().id()="+std::to_string(ket_id);
                        if (not (world_id==phi_id && world_id==bra_id && world_id==ket_id)) {
                            print(msg);
                        }
                        MADNESS_CHECK_THROW(world_id==phi_id && world_id==bra_id && world_id==ket_id,msg.c_str());
 
                        double cpu0 = cpu_time();
                        auto tmp_psif = mul_sparse(world, phi, mo_bra, mul_tol);
                        truncate(world, tmp_psif);
                        double cpu1 = cpu_time();
                        mul1_timer += long((cpu1 - cpu0) * 1000l);
 
                        cpu0 = cpu_time();
                        tmp_psif = apply(world, *poisson.get(), tmp_psif);
                        truncate(world, tmp_psif);
                        cpu1 = cpu_time();
                        apply_timer += long((cpu1 - cpu0) * 1000l);
 
                        cpu0 = cpu_time();
                        auto tmp_Kf = dot(world, mo_ket, tmp_psif);
                        cpu1 = cpu_time();
                        mul2_timer += long((cpu1 - cpu0) * 1000l);
 
                        return tmp_Kf.truncate();
 
                    };
 
                    std::vector<Tile> tiles;
                    for (size_t ilo=0; ilo<mo_bra.size(); ilo+=ntile) {
                        tiles.push_back(Tile{ilo,std::min(ilo+ntile,mo_bra.size())});
                    }
 
                    vecfuncT tmp_mo_bra1,tmp_mo_ket1;
                    vecfuncT tmp_mo_bra2,tmp_mo_ket2;
 
                    for (size_t itile=0; itile<tiles.size(); ++itile) {
                        Tile& tile = tiles[itile];
 
                        if (itile==0) {
                            double t0=cpu_time();
                            print("fetching tile",tile.ilo,"into world",executing_world->id());
                            std::tie(tmp_mo_bra1,tmp_mo_ket1)=fetch_data(*executing_world,tiles[itile]);
                            fetching_world->gop.set_forbid_fence(false);
                            double t2=cpu_time();
                            executing_world->gop.fence();
                            double t1=cpu_time();
                            total_fetch_time += (t1 - t0);
                            total_fetch_spawn_time += (t2 - t0);
                        }
 
                        if (itile>=0) {
                            double t0=cpu_time();
                            fetching_world->gop.set_forbid_fence(true);
                            if (itile<tiles.size()-1) {
                                // fetch data into fetching_world while computing in executing_world
                                print("fetching tile",tiles[itile+1].ilo,"into world",fetching_world->id()," at time ",wall_time());
                                std::tie(tmp_mo_bra2,tmp_mo_ket2)=fetch_data(*fetching_world,tiles[itile+1]);
                            }
                            fetching_world->gop.set_forbid_fence(false);
                            double t2=cpu_time();
                            // uncomment the next line to enforce that fetching is finished before executing
                            // fetching_world->gop.fence();
                            double t1=cpu_time();
                            total_fetch_time += (t1 - t0);
                            total_fetch_spawn_time += (t2 - t0);
 
                            print("executing tile",tile.ilo,"in world",executing_world->id());
                            double dpu0=cpu_time();
                            Kf[0]+=execute(*executing_world,poisson1,phi1,tmp_mo_bra1,tmp_mo_ket1);
                            double dpu1=cpu_time();
                            print("time to execute tile",tile.ilo,"in world",executing_world->id(),dpu1-dpu0,"seconds");
                            total_execution_time += dpu1-dpu0;
 
                            fetching_world->gop.fence();
 
                            // change roles of the two worlds
                            std::swap(poisson1,poisson2);
                            std::swap(phi1,phi2);
                            std::swap(tmp_mo_bra2,tmp_mo_bra1);
                            std::swap(tmp_mo_ket2,tmp_mo_ket1);
                            std::swap(executing_world,fetching_world);
                        }
                    }
                } // objects living in the two worlds must be destroyed before the worlds are freed
 
                // deferred destruction of WorldObjects happens here
                fetching_world->gop.fence();
                executing_world->gop.fence();
                double cpu2=cpu_time();
                print("overall time: ",cpu2-cpu0,"seconds");
                print("total execution time:",total_execution_time,"seconds");
                print("total fetch time:",total_fetch_time,"seconds");
                print("total fetch spawn time:",total_fetch_spawn_time,"seconds");
            } // worlds are destroyed here
 
            return Kf;
        }
    };
};
 
} /* namespace madness */
 
#endif /* SRC_APPS_CHEM_EXCHANGEOPERATOR_H_ */