#ifdef NMZ_MIC_OFFLOAD

#pragma offload_attribute(push, target(mic))
#include "libnormaliz/offload_handler.h"
#include <offload.h>  // offload system header
#include "libnormaliz/matrix.h"
#include "libnormaliz/full_cone.h"
#include "libnormaliz/list_and_map_operations.h"
#include "libnormaliz/vector_operations.h"
#include "libnormaliz/my_omp.h"
#include "libnormaliz/HilbertSeries.h"
#include <cstring>  // for strcpy
#include <iostream>
#include <fstream>

namespace libnormaliz {

using namespace std;

// transferring vector
template <typename Integer>
void fill_vector(vector<Integer>& v, long size, Integer* data) {
    for (long i = 0; i < size; i++)
        v[i] = data[i];
}

template <typename Integer>
void fill_plain(Integer* data, long size, const vector<Integer>& v) {
    for (long i = 0; i < size; i++)
        data[i] = v[i];
}

// transferring Matrix
template <typename Integer>
void fill_matrix(Matrix<Integer>& M, long rows, long cols, Integer* data) {
    for (long i = 0; i < rows; i++)
        for (long j = 0; j < cols; j++)
            M[i][j] = data[i * cols + j];
}

template <typename Integer>
void fill_plain(Integer* data, long rows, long cols, const Matrix<Integer>& M) {
    for (long i = 0; i < rows; i++)
        for (long j = 0; j < cols; j++)
            data[i * cols + j] = M[i][j];
}

// transferring list<vector>
// the vectors may have different lengths
// fill_list_vector also creates the list entries and appendes them to the list!
template <typename Integer>
void fill_list_vector(list<vector<Integer> >& l, long plain_size, Integer* data) {
    Integer* data_end = data + plain_size;  // position after last entry
    while (data < data_end) {
        l.push_back(vector<Integer>(*data));
        fill_vector(l.back(), *data, data + 1);
        data += *data + 1;
    }
}

template <typename Integer>
void fill_plain(Integer* data, long size, const list<vector<Integer> >& l) {
    long v_size;
    for (const auto& it : l) {
        v_size = it.size();
        *data = v_size;
        fill_plain(++data, v_size, it);
        data += v_size;
    }
}

template <typename Integer>
long plain_size(const list<vector<Integer> >& l) {
    long size = 0;
    for (const auto& it : l)
        size += it.size() + 1;
    return size;
}
#pragma offload_attribute(pop)

//-------------------------- OffloadHandler ---------------------------------

template <typename Integer>
OffloadHandler<Integer>::OffloadHandler(Full_Cone<Integer>& fc, int mic_number)
    : mic_nr(mic_number), running(false), local_fc_ref(fc) {
    cout << "mic " << mic_nr << ": Offload and initialize Full_Cone..." << endl;
    create_full_cone();

    transfer_bools();
    transfer_support_hyperplanes();
    transfer_grading();             // including truncation and shift
    transfer_triangulation_info();  // extreme rays, deg1_triangulation, Order_Vector

    primal_algorithm_initialize();
    cout << "mic " << mic_nr << ": Full_Cone initialized." << endl;
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::create_full_cone() {
    const Matrix<Integer>& M = local_fc_ref.Generators;
    long nr = M.nr_of_rows();
    long nc = M.nr_of_columns();
    long size = nr * nc;
    Integer* data = new Integer[size];
    fill_plain(data, nr, nc, M);

//  cout << "mic " << mic_nr<< ": Offload Full_Cone..." << endl;
// offload to mic, copy data and free it afterwards, but keep a pointer to the created Full_Cone
#pragma offload target(mic : mic_nr) in(nr, nc)in(data : length(size) ONCE)
    {
        omp_set_num_threads(118);
        Matrix<Integer> gens(nr, nc);
        fill_matrix(gens, nr, nc, data);
        offload_fc_ptr = new Full_Cone<Integer>(gens);
    }
    //  cout << "mic " << mic_nr<< ": Offload Full_Cone completed." << endl;
    delete[] data;
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::transfer_bools() {
    //  cout << "mic " << mic_nr<< ": transfer_bools" << endl;
    Full_Cone<Integer>& foo_loc = local_fc_ref;  // prevents segfault
    // TODO segfaults should be resolved in intel compiler version 2015
    bool is_computed_pointed = local_fc_ref.isComputed(ConeProperty::IsPointed);
    bool inhomogeneous = foo_loc.inhomogeneous;
    bool do_Hilbert_basis = foo_loc.do_Hilbert_basis;
    bool do_h_vector = foo_loc.do_h_vector;
    bool keep_triangulation = foo_loc.keep_triangulation;
    bool do_multiplicity = foo_loc.do_multiplicity;
    bool do_determinants = foo_loc.do_determinants;
    bool do_triangulation = foo_loc.do_triangulation;
    bool do_deg1_elements = foo_loc.do_deg1_elements;
    bool do_Stanley_dec = foo_loc.do_Stanley_dec;
    bool do_approximation = foo_loc.do_approximation;
    bool do_default_mode = foo_loc.do_default_mode;
    bool deg1_generated = foo_loc.deg1_generated;
    bool pointed = foo_loc.pointed;
#pragma offload target(mic : mic_nr) in(mic_nr)
    {
        // bool foo = offload_fc_ptr->inhomogeneous;  // prevents segfault
        offload_fc_ptr->inhomogeneous = inhomogeneous;
        offload_fc_ptr->do_Hilbert_basis = do_Hilbert_basis;
        offload_fc_ptr->do_h_vector = do_h_vector;
        offload_fc_ptr->keep_triangulation = keep_triangulation;
        offload_fc_ptr->do_multiplicity = do_multiplicity;
        offload_fc_ptr->do_determinants = do_determinants;
        offload_fc_ptr->do_triangulation = do_triangulation;
        offload_fc_ptr->do_deg1_elements = do_deg1_elements;
        offload_fc_ptr->do_Stanley_dec = do_Stanley_dec;
        offload_fc_ptr->do_approximation = do_approximation;
        offload_fc_ptr->do_default_mode = do_default_mode;
        offload_fc_ptr->do_all_hyperplanes = false;
        // deg1_generated could be set more precise
        offload_fc_ptr->deg1_triangulation = deg1_generated;
        if (is_computed_pointed) {
            offload_fc_ptr->pointed = pointed;
            offload_fc_ptr->is_Computed.set(ConeProperty::IsPointed);
        }
        offload_fc_ptr->verbose = true;
        string fstr = "/tmp/mic";
        fstr.push_back(static_cast<char>('0' + mic_nr));
        fstr.append(".log");
        cout << "writing to " << fstr << endl;
        ofstream* fout = new ofstream(fstr.c_str());
        setVerboseOutput(*fout);
        setErrorOutput(*fout);
        verboseOutput() << "Start logging" << endl;
    }
    //  cout << "mic " << mic_nr<< ": transfer_bools done" << endl;
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::transfer_support_hyperplanes() {
    if (!local_fc_ref.isComputed(ConeProperty::SupportHyperplanes))
        return;

    //  cout << "mic " << mic_nr<< ": transfer_support_hyperplanes" << endl;
    const Matrix<Integer>& M = local_fc_ref.Support_Hyperplanes;
    long nr = M.nr_of_rows();
    long nc = M.nr_of_columns();
    long size = nr * nc;
    assert(size > 0);  // make sure there are support hyperplanes computed
    Integer* data = new Integer[size];
    fill_plain(data, nr, nc, M);

// offload to mic, copy data and create the C++ matrix in the offloaded Full_Cone
#pragma offload target(mic : mic_nr) in(nr, nc)in(data : length(size) ONCE)
    {
        offload_fc_ptr->Support_Hyperplanes = Matrix<Integer>(nr, nc);
        fill_matrix(offload_fc_ptr->Support_Hyperplanes, nr, nc, data);
        offload_fc_ptr->nrSupport_Hyperplanes = nr;
        offload_fc_ptr->is_Computed.set(ConeProperty::SupportHyperplanes);
    }
    delete[] data;

    //  cout << "mic " << mic_nr<< ": transfer_support_hyperplanes done" << endl;
    if (local_fc_ref.ExcludedFaces.nr_of_rows() > 0) {
        const Matrix<Integer>& ExFaces = local_fc_ref.ExcludedFaces;

        nr = ExFaces.nr_of_rows();
        nc = ExFaces.nr_of_columns();
        size = nr * nc;
        assert(size > 0);  // make sure there are support hyperplanes computed
        Integer* data = new Integer[size];
        fill_plain(data, nr, nc, ExFaces);

// offload to mic, copy data and create the C++ matrix in the offloaded Full_Cone
#pragma offload target(mic : mic_nr) in(nr, nc)in(data : length(size) ONCE)
        {
            offload_fc_ptr->ExcludedFaces = Matrix<Integer>(nr, nc);
            fill_matrix(offload_fc_ptr->ExcludedFaces, nr, nc, data);
        }
        delete[] data;
    }
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::transfer_grading() {
    //  cout << "mic " << mic_nr<< ": transfer_grading" << endl;
    long dim = local_fc_ref.dim;
    if (local_fc_ref.inhomogeneous) {
        assert(local_fc_ref.isComputed(ConeProperty::RecessionRank));
        Integer* data = new Integer[dim];
        fill_plain(data, dim, local_fc_ref.Truncation);
        long level0_dim = local_fc_ref.level0_dim;

#pragma offload target(mic : mic_nr) in(dim)in(level0_dim)in(data : length(dim) ONCE)
        {
            offload_fc_ptr->Truncation = vector<Integer>(dim);
            fill_vector(offload_fc_ptr->Truncation, dim, data);
            offload_fc_ptr->level0_dim = level0_dim;
        }
        delete[] data;
    }

    if (local_fc_ref.isComputed(ConeProperty::Grading)) {
        Integer* data = new Integer[dim];
        fill_plain(data, dim, local_fc_ref.Grading);

#pragma offload target(mic : mic_nr) in(dim)in(data : length(dim) ONCE)
        {
            offload_fc_ptr->Grading = vector<Integer>(dim);
            fill_vector(offload_fc_ptr->Grading, dim, data);
            offload_fc_ptr->is_Computed.set(ConeProperty::Grading);
        }
        delete[] data;
    }

    if (local_fc_ref.shift != 0) {
        auto shift = local_fc_ref.shift;
#pragma offload target(mic : mic_nr) in(shift)
        { offload_fc_ptr->shift = shift; }
    }

    //  cout << "mic " << mic_nr<< ": transfer_grading done" << endl;
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::transfer_triangulation_info() {
    //  cout << "mic " << mic_nr<< ": transfer_triangulation_info" << endl;
    long dim = local_fc_ref.dim;
    long nr_gen = local_fc_ref.nr_gen;

    if (local_fc_ref.isComputed(ConeProperty::ExtremeRays)) {
        bool* data = new bool[nr_gen];
        fill_plain(data, nr_gen, local_fc_ref.Extreme_Rays_Ind);

#pragma offload target(mic : mic_nr) in(nr_gen)in(data : length(nr_gen) ONCE)
        {
            offload_fc_ptr->Extreme_Rays_Ind = vector<bool>(nr_gen);
            fill_vector(offload_fc_ptr->Extreme_Rays_Ind, nr_gen, data);
            offload_fc_ptr->is_Computed.set(ConeProperty::ExtremeRays);
        }
        delete[] data;
    }

    // always transfer the order vector
    {
        Integer* data = new Integer[dim];
        fill_plain(data, dim, local_fc_ref.Order_Vector);

#pragma offload target(mic : mic_nr) in(dim)in(data : length(dim) ONCE)
        {
            offload_fc_ptr->Order_Vector = vector<Integer>(dim);
            fill_vector(offload_fc_ptr->Order_Vector, dim, data);
        }
        delete[] data;
    }

    if (!local_fc_ref.Comparisons.empty()) {
        long size = local_fc_ref.Comparisons.size();

        size_t* data = new size_t[size];
        fill_plain(data, size, local_fc_ref.Comparisons);

#pragma offload target(mic : mic_nr) in(size)in(data : length(size) ONCE)
        {
            offload_fc_ptr->Comparisons.resize(size);
            fill_vector(offload_fc_ptr->Comparisons, size, data);
            offload_fc_ptr->nrTotalComparisons = offload_fc_ptr->Comparisons[size - 1];
        }
        delete[] data;
    }
    //  cout << "mic " << mic_nr<< ": transfer_triangulation_info done" << endl;
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::primal_algorithm_initialize() {
#pragma offload target(mic : mic_nr) signal(&running)
    {
        offload_fc_ptr->do_vars_check(false);
        if (offload_fc_ptr->inhomogeneous)
            offload_fc_ptr->set_levels();
        offload_fc_ptr->set_degrees();
        offload_fc_ptr->primal_algorithm_initialize();

        cout << "mic " << mic_nr << ": create 3 mio empty simplices ..." << endl;
        SHORTSIMPLEX<Integer> simp;
        simp.key = vector<key_t>(offload_fc_ptr->dim);
        simp.height = 0;
        simp.vol = 0;
        offload_fc_ptr->FreeSimpl.insert(offload_fc_ptr->FreeSimpl.end(), 3000000, simp);
        cout << "mic " << mic_nr << ": creating simplices done." << endl;
    }
    running = true;
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::transfer_pyramids(const list<vector<key_t> >& pyramids) {
    long size = plain_size(pyramids);

    key_t* data = new key_t[size];
    fill_plain(data, size, pyramids);

    transfer_pyramids_inner(data, size);

    delete[] data;
    cout << "mic " << mic_nr << ": transferred " << pyramids.size()
         << " pyramids. avg. key size:" << static_cast<double>(size) / pyramids.size() - 1 << endl;
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::transfer_pyramids_inner(key_t* data, long size) {
    wait();
#pragma offload target(mic : mic_nr) in(size)in(data : length(size) ONCE)
    {
        fill_list_vector(offload_fc_ptr->Pyramids[0], size, data);
        offload_fc_ptr->nrPyramids[0] = offload_fc_ptr->Pyramids[0].size();
    }
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::evaluate_pyramids() {
    wait();
    cout << "mic " << mic_nr << ": evaluate_pyramids" << endl;
#pragma offload target(mic : mic_nr) signal(&running)
    {
        offload_fc_ptr->evaluate_stored_pyramids(0);
        offload_fc_ptr->evaluate_triangulation();
    }
    // cout << "Nach Start evaluate mic" << mic_nr << endl;
    running = true;
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::evaluate_triangulation() {
    wait();
    cout << "mic " << mic_nr << ": evaluate_triangulation" << endl;
#pragma offload target(mic : mic_nr) signal(&running)
    { offload_fc_ptr->evaluate_triangulation(); }
    running = true;
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::complete_evaluation() {
    wait();
    cout << "mic " << mic_nr << ": complete_evaluation" << endl;
#pragma offload target(mic : mic_nr) signal(&running)
    { offload_fc_ptr->primal_algorithm_finalize(); }
    running = true;
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::collect_data() {
    wait();
    cout << "mic " << mic_nr << ": collect_data" << endl;
    collect_integers();  // TriangulationSize, DetSum, Multiplicity, ...
    collect_hilbert_series();
    collect_candidates();  // Hilbert basis, degree 1 elements
    cout << "mic " << mic_nr << ": collect_data done" << endl;
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::collect_integers() {
    {
        size_t col_totalNrSimplices, col_nrSimplicialPyr, col_totalNrPyr;
        Integer col_detSum;
#pragma offload target(mic : mic_nr) out(col_totalNrSimplices) out(col_nrSimplicialPyr) out(col_totalNrPyr) out(col_detSum)
        {
            col_totalNrSimplices = offload_fc_ptr->totalNrSimplices;
            col_nrSimplicialPyr = offload_fc_ptr->nrSimplicialPyr;
            col_totalNrPyr = offload_fc_ptr->totalNrPyr;
            col_detSum = offload_fc_ptr->detSum;
        }
        local_fc_ref.totalNrSimplices += col_totalNrSimplices;
        local_fc_ref.nrSimplicialPyr += col_nrSimplicialPyr;
        local_fc_ref.totalNrPyr += col_totalNrPyr;
        local_fc_ref.detSum += col_detSum;
    }

    if (local_fc_ref.do_triangulation && local_fc_ref.do_evaluation && local_fc_ref.isComputed(ConeProperty::Grading)) {
        //    cout << "mic " << mic_nr<< ": collecting multiplicity ..." << endl;
        long size;
        std::string* str_ptr;
#pragma offload target(mic : mic_nr) out(size) nocopy(str_ptr : length(0))
        {
            str_ptr = new std::string(offload_fc_ptr->multiplicity.get_str());
            size = str_ptr->length() + 1;
        }

        char* c_str = new char[size];
#pragma offload target(mic : mic_nr) in(size)out(c_str : length(size)) nocopy(str_ptr : length(0))
        {
            std::strcpy(c_str, str_ptr->c_str());
            delete str_ptr;
        }
        mpq_class coll_mult(c_str);
        delete c_str;
        local_fc_ref.multiplicity += coll_mult;
        //    cout << "mic " << mic_nr<< ": collecting multiplicity done" << endl;
    }
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::collect_hilbert_series() {
    if (local_fc_ref.do_h_vector) {
        //    cout << "mic " << mic_nr<< ": collecting Hilbert series ..." << endl;
        long size;
        std::string* str_ptr;
#pragma offload target(mic : mic_nr) out(size) nocopy(str_ptr : length(0))
        {
            str_ptr = new std::string(offload_fc_ptr->Hilbert_Series.to_string_rep());
            size = str_ptr->length() + 1;
        }

        char* c_str = new char[size];
#pragma offload target(mic : mic_nr) in(size)out(c_str : length(size)) nocopy(str_ptr : length(0))
        {
            std::strcpy(c_str, str_ptr->c_str());
            delete str_ptr;
        }
        HilbertSeries col_HS = HilbertSeries(string(c_str));
        delete c_str;
        local_fc_ref.Hilbert_Series += col_HS;
        //    cout << "mic " << mic_nr<< ": collecting Hilbert series done" << endl;
    }
}

//---------------------------------------------------------------------------

#pragma offload_attribute(push, target(mic))
template <typename Integer>
bool is_ori_gen(const Candidate<Integer>& c) {
    return c.original_generator;
}
#pragma offload_attribute(push, target(mic))

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::collect_candidates() {
    if (local_fc_ref.do_Hilbert_basis) {
        //    cout << "mic " << mic_nr<< ": collect Hilbert basis" << endl;
        long size;

#pragma offload target(mic : mic_nr) out(size)
        {
            // remove all original generators, they are also inserted on the host
            offload_fc_ptr->OldCandidates.Candidates.remove_if(is_ori_gen<Integer>);
            // or in c++11 with lambda function
            // offload_fc_ptr->OldCandidates.Candidates.remove_if([](const Candidate<Integer>& c){ return c.original_generator;
            // });
            offload_fc_ptr->OldCandidates.extract(offload_fc_ptr->Hilbert_Basis);
            offload_fc_ptr->OldCandidates.Candidates.clear();

            // using the same methods as for pyramids
            // handling list of vectors of possible different lengths
            size = offload_fc_ptr->Hilbert_Basis.size() * (offload_fc_ptr->dim + 1);
        }
        if (size > 0) {
            Integer* data = new Integer[size];

#pragma offload target(mic : mic_nr) in(size)out(data : length(size) ONCE)
            { fill_plain(data, size, offload_fc_ptr->Hilbert_Basis); }
            //      fill_list_vector(local_fc_ref.Hilbert_Basis, size, data);
            list<vector<Integer> > coll_HB;
            fill_list_vector(coll_HB, size, data);
            delete[] data;
            CandidateList<Integer> cand_l;
            while (!coll_HB.empty()) {
                cand_l.push_back(Candidate<Integer>(coll_HB.front(), local_fc_ref));
                coll_HB.pop_front();
            }
            //      cout << "mic " << mic_nr<< ": CandidateList complete" << endl;
            //      #pragma omp critical(CANDIDATES)
            local_fc_ref.NewCandidates.splice(cand_l);

            local_fc_ref.NewCandidates.reduce_by(local_fc_ref.OldCandidates);
            local_fc_ref.update_reducers();

        }  // if (size > 0)
           //    cout << "mic " << mic_nr<< ": collect Hilbert basis done" << endl;
    }

    if (local_fc_ref.do_deg1_elements) {
        //    cout << "mic " << mic_nr<< ": collect degree 1 elements" << endl;
        long size;

#pragma offload target(mic : mic_nr) out(size)
        {
            // using the same methods as for pyramids
            // handling list of vectors of possible different lengths
            size = offload_fc_ptr->Deg1_Elements.size() * (offload_fc_ptr->dim + 1);
        }
        if (size > 0) {
            Integer* data = new Integer[size];

#pragma offload target(mic : mic_nr) in(size)out(data : length(size) ONCE)
            { fill_plain(data, size, offload_fc_ptr->Deg1_Elements); }
            list<vector<Integer> > coll_Deg1;
            fill_list_vector(coll_Deg1, size, data);
            delete[] data;
            local_fc_ref.Deg1_Elements.splice(local_fc_ref.Deg1_Elements.end(), coll_Deg1);
        }
        //    cout << "mic " << mic_nr<< ": collect degree 1 elements done" << endl;
    }
}

//---------------------------------------------------------------------------

template <typename Integer>
bool OffloadHandler<Integer>::is_running() {
#ifndef __MIC__
    if (running) {
        running = !_Offload_signaled(mic_nr, &running);
    }
#endif
    return running;
}

//---------------------------------------------------------------------------

template <typename Integer>
void OffloadHandler<Integer>::wait() {
    if (is_running()) {
        cout << "mic " << mic_nr << ": waiting ..." << endl;
#pragma offload_wait target(mic : mic_nr) wait(&running)
        running = false;
        cout << "mic " << mic_nr << ": waiting completed." << endl;
    }
}

//---------------------------------------------------------------------------

template <typename Integer>
OffloadHandler<Integer>::~OffloadHandler() {
    cout << "mic " << mic_nr << ": OffloadHandler destructor" << endl;
#pragma offload target(mic : mic_nr)
    { delete offload_fc_ptr; }
}

//-------------------------- MicOffloader -----------------------------------

template <typename Integer>
MicOffloader<Integer>::MicOffloader() : is_init(false), nr_mics(_Offload_number_of_devices()), nr_handlers(nr_mics) {
    handlers.resize(nr_handlers);
    // cout << "Constructor " << nr_handlers << endl;
}

//---------------------------------------------------------------------------

template <typename Integer>
MicOffloader<Integer>::~MicOffloader() {
    if (is_init)
        for (int i = 0; i < nr_handlers; ++i)
            delete handlers[i];
}

//---------------------------------------------------------------------------

template <typename Integer>
void MicOffloader<Integer>::init(Full_Cone<Integer>& fc) {
    if (!is_init) {
        // TODO check preconditions
        assert(fc.Order_Vector.size() == fc.dim);
        if (fc.do_Hilbert_basis) {
            fc.get_supphyps_from_copy(false);  // (bool from_scratch)
            fc.check_pointed();
        }

        // create handler
        for (int i = 0; i < nr_handlers; ++i)
            handlers[i] = new OffloadHandler<Integer>(fc, i % nr_mics);
        is_init = true;
        // wait for completed creation // TODO keep it?
        for (int i = 0; i < nr_handlers; ++i)
            handlers[i]->wait();
    }
}

//---------------------------------------------------------------------------

// template<typename T>
bool compare_sizes(const vector<key_t>& v, const vector<key_t>& w) {
    return v.size() < w.size();
}

template <typename Integer>
void MicOffloader<Integer>::offload_pyramids(Full_Cone<Integer>& fc, const size_t level) {
    if (!is_init)
        init(fc);
    // cout << "nr_handlers " << nr_handlers << endl;

    size_t fraction = 6;
    if (fc.start_from == fc.nr_gen) {  // all gens are done
        fraction = 20;
        if (level > 0)
            fraction = 30;
    }
    if (fraction < nr_handlers)
        fraction = nr_handlers;  // ensure every card can get some

    size_t nr_transfer = min(fc.nrPyramids[level] / fraction, 25000ul);
    // cout << "transfer " << nr_transfer << endl;
    if (nr_transfer == 0)
        return;

    bool at_least_one_idle = false;
    for (int i = 0; i < nr_handlers; ++i) {
        if (!handlers[i]->is_running()) {
            at_least_one_idle = true;
            /* cout << "Mixing ======================" << endl;
            random_order(fc.Pyramids[level]);
            auto py=fc.Pyramids[level].begin();
            for(;py!=fc.Pyramids[level].end();++py)
                cout << py->size() << " ";
            cout << "======================" << endl;*/
            break;
        }
    }
    if (!at_least_one_idle)
        return;

    if (fc.Pyramids_scrambled[level]) {
        fc.Pyramids[level].reverse();  // bring the big pyramids to the rear
    }
    else {
        fc.Pyramids_scrambled[level] = true;  // will not be scrambeld again
        fc.Pyramids[level].sort(compare_sizes);

        size_t size_sum = 0;
        for (const auto& p : fc.Pyramids[level])
            size_sum += p.size();
        size_t size_bound = 2 * size_sum / fc.nrPyramids[level];  // 2*average size

        for (const auto& p : fc.Pyramids[level]) {
            if (p.size() > size_bound)
                break;
        }

        random_order(fc.Pyramids[level], fc.Pyramids[level].begin(), p);
    }

    // offload some pyramids
    list<vector<key_t> > pyrs;
    vector<bool> started(nr_handlers, false);

    for (int i = 0; i < nr_handlers; ++i) {
        if (!handlers[i]->is_running()) {
            // cout << "Testing mic" << i << endl;
            started[i] = true;
            auto transfer_end(fc.Pyramids[level].begin());
            for (size_t j = 0; j < nr_transfer; ++j, ++transfer_end)
                ;
            pyrs.splice(pyrs.end(), fc.Pyramids[level], fc.Pyramids[level].begin(), transfer_end);
            fc.nrPyramids[level] -= nr_transfer;
            handlers[i]->transfer_pyramids(pyrs);
            pyrs.clear();
        }
    }

    fc.Pyramids[level].reverse();  // bring the big pyramids to the front

    // compute on mics
    for (int i = 0; i < nr_handlers; ++i)
        if (started[i])
            handlers[i]->evaluate_pyramids();
}

//---------------------------------------------------------------------------

template <typename Integer>
void MicOffloader<Integer>::evaluate_triangulation() {
    if (is_init) {
        for (int i = 0; i < nr_handlers; ++i)
            if (!handlers[i]->is_running())
                handlers[i]->evaluate_triangulation();
    }
}

//---------------------------------------------------------------------------

template <typename Integer>
void MicOffloader<Integer>::finalize() {
    if (is_init) {
        list<int> to_complete, to_collect;

        for (int i = 0; i < nr_handlers; ++i)
            to_complete.push_back(i);
        // first start it on all idle mics
        while (!to_complete.empty() || !to_collect.empty()) {
            for (auto it = to_complete.begin(); it != to_complete.end();) {
                if (!handlers[*it]->is_running()) {
                    handlers[*it]->complete_evaluation();
                    to_collect.push_back(*it);
                    it = to_complete.erase(it);
                }
                else {
                    ++it;
                }
            }

            for (auto it = to_collect.begin(); it != to_collect.end();) {
                if (!handlers[*it]->is_running()) {
                    handlers[*it]->collect_data();
                    delete handlers[*it];
                    handlers[*it] = NULL;
                    it = to_collect.erase(it);
                }
                else {
                    ++it;
                }
            }
        }
        is_init = false;
    }
}

/***************** Instantiation for template parameter long long *****************/

template class MicOffloader<long long int>;
template class OffloadHandler<long long int>;

}  // end namespace libnormaliz

#endif  // NMZ_MIC_OFFLOAD