mpi_solver.cpp

#include <iostream>
#include <vector>
#include <string>

#include <boost/program_options.hpp>
#include <boost/property_tree/ptree.hpp>
#include <boost/property_tree/json_parser.hpp>
#include <boost/preprocessor/seq/for_each.hpp>

#include <amgcl/backend/builtin.hpp>
#include <amgcl/value_type/static_matrix.hpp>
#include <amgcl/adapter/crs_tuple.hpp>
#include <amgcl/adapter/block_matrix.hpp>

#if defined(SOLVER_BACKEND_VEXCL)
#  include <amgcl/backend/vexcl.hpp>
#  include <amgcl/backend/vexcl_static_matrix.hpp>
#elif defined(SOLVER_BACKEND_CUDA)
#  include <amgcl/backend/cuda.hpp>
#  include <amgcl/relaxation/cusparse_ilu0.hpp>
#else
#  ifndef SOLVER_BACKEND_BUILTIN
#    define SOLVER_BACKEND_BUILTIN
#  endif
#endif

#include <amgcl/mpi/util.hpp>
#include <amgcl/mpi/make_solver.hpp>
#include <amgcl/mpi/preconditioner.hpp>
#include <amgcl/mpi/solver/runtime.hpp>

#include <amgcl/io/mm.hpp>
#include <amgcl/io/binary.hpp>
#include <amgcl/profiler.hpp>

#ifndef AMGCL_BLOCK_SIZES
#  define AMGCL_BLOCK_SIZES (3)(4)
#endif

namespace amgcl {
    profiler<> prof;
}

namespace math = amgcl::math;

//---------------------------------------------------------------------------
ptrdiff_t assemble_poisson3d(amgcl::mpi::communicator comm,
        ptrdiff_t n, int block_size,
        std::vector<ptrdiff_t> &ptr,
        std::vector<ptrdiff_t> &col,
        std::vector<double>    &val,
        std::vector<double>    &rhs)
{
    ptrdiff_t n3 = n * n * n;

    ptrdiff_t chunk = (n3 + comm.size - 1) / comm.size;
    if (chunk % block_size != 0) {
        chunk += block_size - chunk % block_size;
    }
    ptrdiff_t row_beg = std::min(n3, chunk * comm.rank);
    ptrdiff_t row_end = std::min(n3, row_beg + chunk);
    chunk = row_end - row_beg;

    ptr.clear(); ptr.reserve(chunk + 1);
    col.clear(); col.reserve(chunk * 7);
    val.clear(); val.reserve(chunk * 7);

    rhs.resize(chunk);
    std::fill(rhs.begin(), rhs.end(), 1.0);

    const double h2i = (n - 1) * (n - 1);
    ptr.push_back(0);

    for (ptrdiff_t idx = row_beg; idx < row_end; ++idx) {
        ptrdiff_t k = idx / (n * n);
        ptrdiff_t j = (idx / n) % n;
        ptrdiff_t i = idx % n;

        if (k > 0)  {
            col.push_back(idx - n * n);
            val.push_back(-h2i);
        }

        if (j > 0)  {
            col.push_back(idx - n);
            val.push_back(-h2i);
        }

        if (i > 0) {
            col.push_back(idx - 1);
            val.push_back(-h2i);
        }

        col.push_back(idx);
        val.push_back(6 * h2i);

        if (i + 1 < n) {
            col.push_back(idx + 1);
            val.push_back(-h2i);
        }

        if (j + 1 < n) {
            col.push_back(idx + n);
            val.push_back(-h2i);
        }

        if (k + 1 < n) {
            col.push_back(idx + n * n);
            val.push_back(-h2i);
        }

        ptr.push_back( col.size() );
    }

    return chunk;
}

//---------------------------------------------------------------------------
ptrdiff_t read_matrix_market(
        amgcl::mpi::communicator comm,
        const std::string &A_file, const std::string &rhs_file, int block_size,
        std::vector<ptrdiff_t> &ptr,
        std::vector<ptrdiff_t> &col,
        std::vector<double>    &val,
        std::vector<double>    &rhs)
{
    amgcl::io::mm_reader A_mm(A_file);
    ptrdiff_t n = A_mm.rows();

    ptrdiff_t chunk = (n + comm.size - 1) / comm.size;
    if (chunk % block_size != 0) {
        chunk += block_size - chunk % block_size;
    }

    ptrdiff_t row_beg = std::min(n, chunk * comm.rank);
    ptrdiff_t row_end = std::min(n, row_beg + chunk);

    chunk = row_end - row_beg;

    A_mm(ptr, col, val, row_beg, row_end);

    if (rhs_file.empty()) {
        rhs.resize(chunk);
        std::fill(rhs.begin(), rhs.end(), 1.0);
    } else {
        amgcl::io::mm_reader rhs_mm(rhs_file);
        rhs_mm(rhs, row_beg, row_end);
    }

    return chunk;
}

//---------------------------------------------------------------------------
ptrdiff_t read_binary(
        amgcl::mpi::communicator comm,
        const std::string &A_file, const std::string &rhs_file, int block_size,
        std::vector<ptrdiff_t> &ptr,
        std::vector<ptrdiff_t> &col,
        std::vector<double>    &val,
        std::vector<double>    &rhs)
{
    ptrdiff_t n = amgcl::io::crs_size<ptrdiff_t>(A_file);

    ptrdiff_t chunk = (n + comm.size - 1) / comm.size;
    if (chunk % block_size != 0) {
        chunk += block_size - chunk % block_size;
    }

    ptrdiff_t row_beg = std::min(n, chunk * comm.rank);
    ptrdiff_t row_end = std::min(n, row_beg + chunk);

    chunk = row_end - row_beg;

    amgcl::io::read_crs(A_file, n, ptr, col, val, row_beg, row_end);

    if (rhs_file.empty()) {
        rhs.resize(chunk);
        std::fill(rhs.begin(), rhs.end(), 1.0);
    } else {
        ptrdiff_t rows, cols;
        amgcl::io::read_dense(rhs_file, rows, cols, rhs, row_beg, row_end);
    }

    return chunk;
}

//---------------------------------------------------------------------------
template <class Backend, class Matrix>
std::shared_ptr< amgcl::mpi::distributed_matrix<Backend> >
partition(amgcl::mpi::communicator comm, const Matrix &Astrip,
        typename Backend::vector &rhs, const typename Backend::params &bprm,
        amgcl::runtime::mpi::partition::type ptype, int block_size = 1)
{
    typedef typename Backend::value_type val_type;
    typedef typename amgcl::math::rhs_of<val_type>::type rhs_type;
    typedef amgcl::mpi::distributed_matrix<Backend> DMatrix;

    using amgcl::prof;

    auto A = std::make_shared<DMatrix>(comm, Astrip);

    if (comm.size == 1 || ptype == amgcl::runtime::mpi::partition::merge)
        return A;

    prof.tic("partition");
    boost::property_tree::ptree prm;
    prm.put("type", ptype);
    amgcl::runtime::mpi::partition::wrapper<Backend> part(prm);

    auto I = part(*A, block_size);
    auto J = transpose(*I);
    A = product(*J, *product(*A, *I));

#if defined(SOLVER_BACKEND_BUILTIN)
    amgcl::backend::numa_vector<rhs_type> new_rhs(J->loc_rows());
#elif defined(SOLVER_BACKEND_VEXCL)
    vex::vector<rhs_type> new_rhs(bprm.q, J->loc_rows());
#elif defined(SOLVER_BACKEND_CUDA)
    thrust::device_vector<rhs_type> new_rhs(J->loc_rows());
#endif

    J->move_to_backend(bprm);

    amgcl::backend::spmv(1, *J, rhs, 0, new_rhs);
    rhs.swap(new_rhs);
    prof.toc("partition");

    return A;
}

//---------------------------------------------------------------------------
#if defined(SOLVER_BACKEND_BUILTIN) || defined(SOLVER_BACKEND_VEXCL)
template <int B>
void solve_block(
        amgcl::mpi::communicator comm,
        ptrdiff_t chunk,
        const std::vector<ptrdiff_t>      &ptr,
        const std::vector<ptrdiff_t>      &col,
        const std::vector<double>         &val,
        const boost::property_tree::ptree &prm,
        const std::vector<double>         &f,
        amgcl::runtime::mpi::partition::type ptype
        )
{
    typedef amgcl::static_matrix<double, B, B> val_type;
    typedef amgcl::static_matrix<double, B, 1> rhs_type;

#if defined(SOLVER_BACKEND_BUILTIN)
    typedef amgcl::backend::builtin<val_type> Backend;
#elif defined(SOLVER_BACKEND_VEXCL)
    typedef amgcl::backend::vexcl<val_type> Backend;
#endif

    typedef
        amgcl::mpi::make_solver<
            amgcl::runtime::mpi::preconditioner<Backend>,
            amgcl::runtime::mpi::solver::wrapper<Backend>
            >
        Solver;

    using amgcl::prof;

    typename Backend::params bprm;

#if defined(SOLVER_BACKEND_BUILTIN)
    amgcl::backend::numa_vector<rhs_type> rhs(
            reinterpret_cast<const rhs_type*>(&f[0]),
            reinterpret_cast<const rhs_type*>(&f[0]) + chunk / B
            );
#elif defined(SOLVER_BACKEND_VEXCL)
    vex::Context ctx(vex::Filter::Env);
    bprm.q = ctx;

    vex::scoped_program_header header(ctx,
            amgcl::backend::vexcl_static_matrix_declaration<double,B>());

    if (comm.rank == 0) std::cout << ctx << std::endl;

    vex::vector<rhs_type> rhs(ctx, chunk / B, reinterpret_cast<const rhs_type*>(&f[0]));
#endif

    prof.tic("setup");
    std::shared_ptr<Solver> solve;
    if (ptype) {
        auto A = partition<Backend>(comm,
                amgcl::adapter::block_matrix<val_type>(std::tie(chunk, ptr, col, val)),
                rhs, bprm, ptype, prm.get("precond.coarsening.aggr.block_size", 1));

        solve = std::make_shared<Solver>(comm, A, prm, bprm);
        chunk = A->loc_rows();
    } else {
        solve = std::make_shared<Solver>(comm,
                amgcl::adapter::block_matrix<val_type>(std::tie(chunk, ptr, col, val)),
                prm, bprm);
    }
    prof.toc("setup");

    if (comm.rank == 0) {
        std::cout << *solve << std::endl;
    }

#if defined(SOLVER_BACKEND_BUILTIN)
    amgcl::backend::numa_vector<rhs_type> x(chunk);
#elif defined(SOLVER_BACKEND_VEXCL)
    vex::vector<rhs_type> x(ctx, chunk);
    x = math::zero<rhs_type>();
#endif

    int    iters;
    double error;

    prof.tic("solve");
    std::tie(iters, error) = (*solve)(rhs, x);
    prof.toc("solve");

    if (comm.rank == 0) {
        std::cout
            << "Iterations: " << iters << std::endl
            << "Error:      " << error << std::endl
            << prof << std::endl;
    }
}
#endif

//---------------------------------------------------------------------------
void solve_scalar(
        amgcl::mpi::communicator comm,
        ptrdiff_t chunk,
        const std::vector<ptrdiff_t> &ptr,
        const std::vector<ptrdiff_t> &col,
        const std::vector<double> &val,
        const boost::property_tree::ptree &prm,
        const std::vector<double> &f,
        amgcl::runtime::mpi::partition::type ptype
        )
{
#if defined(SOLVER_BACKEND_BUILTIN)
    typedef amgcl::backend::builtin<double> Backend;
#elif defined(SOLVER_BACKEND_VEXCL)
    typedef amgcl::backend::vexcl<double> Backend;
#elif defined(SOLVER_BACKEND_CUDA)
    typedef amgcl::backend::cuda<double> Backend;
#endif

    typedef
        amgcl::mpi::make_solver<
            amgcl::runtime::mpi::preconditioner<Backend>,
            amgcl::runtime::mpi::solver::wrapper<Backend>
            >
        Solver;

    using amgcl::prof;

    typename Backend::params bprm;

#if defined(SOLVER_BACKEND_BUILTIN)
    amgcl::backend::numa_vector<double> rhs(f);
#elif defined(SOLVER_BACKEND_VEXCL)
    vex::Context ctx(vex::Filter::Env);
    bprm.q = ctx;

    if (comm.rank == 0) std::cout << ctx << std::endl;

    vex::vector<double> rhs(ctx, f);
#elif defined(SOLVER_BACKEND_CUDA)
    cusparseCreate(&bprm.cusparse_handle);
    thrust::device_vector<double> rhs(f);
#endif

    prof.tic("setup");
    std::shared_ptr<Solver> solve;
    if (ptype) {
        auto A = partition<Backend>(comm,
                std::tie(chunk, ptr, col, val), rhs, bprm, ptype,
                prm.get("precond.coarsening.aggr.block_size", 1));

        solve = std::make_shared<Solver>(comm, A, prm, bprm);
        chunk = A->loc_rows();
    } else {
        solve = std::make_shared<Solver>(comm, std::tie(chunk, ptr, col, val), prm, bprm);
    }
    prof.toc("setup");

    if (comm.rank == 0) {
        std::cout << *solve << std::endl;
    }

#if defined(SOLVER_BACKEND_BUILTIN)
    amgcl::backend::numa_vector<double> x(chunk);
#elif defined(SOLVER_BACKEND_VEXCL)
    vex::vector<double> x(ctx, chunk);
    x = 0.0;
#elif defined(SOLVER_BACKEND_CUDA)
    thrust::device_vector<double> x(chunk, 0.0);
#endif

    int    iters;
    double error;

    prof.tic("solve");
    std::tie(iters, error) = (*solve)(rhs, x);
    prof.toc("solve");

    if (comm.rank == 0) {
        std::cout
            << "Iterations: " << iters << std::endl
            << "Error:      " << error << std::endl
            << prof << std::endl;
    }
}

//---------------------------------------------------------------------------
int main(int argc, char *argv[]) {
    amgcl::mpi::init_thread mpi(&argc, &argv);
    amgcl::mpi::communicator comm(MPI_COMM_WORLD);

    if (comm.rank == 0)
        std::cout << "World size: " << comm.size << std::endl;

    using amgcl::prof;

    // Read configuration from command line
    namespace po = boost::program_options;
    po::options_description desc("Options");

    desc.add_options()
        ("help,h", "show help")
        ("matrix,A",
         po::value<std::string>(),
         "System matrix in the MatrixMarket format. "
         "When not specified, a Poisson problem in 3D unit cube is assembled. "
        )
        (
         "rhs,f",
         po::value<std::string>()->default_value(""),
         "The RHS vector in the MatrixMarket format. "
         "When omitted, a vector of ones is used by default. "
         "Should only be provided together with a system matrix. "
        )
        (
         "Ap",
         po::value< std::vector<std::string> >()->multitoken(),
         "Pre-partitioned matrix (single file per MPI process)"
        )
        (
         "fp",
         po::value< std::vector<std::string> >()->multitoken(),
         "Pre-partitioned RHS (single file per MPI process)"
        )
        (
         "binary,B",
         po::bool_switch()->default_value(false),
         "When specified, treat input files as binary instead of as MatrixMarket. "
         "It is assumed the files were converted to binary format with mm2bin utility. "
        )
        (
         "block-size,b",
         po::value<int>()->default_value(1),
         "The block size of the system matrix. "
         "When specified, the system matrix is assumed to have block-wise structure. "
         "This usually is the case for problems in elasticity, structural mechanics, "
         "for coupled systems of PDE (such as Navier-Stokes equations), etc. "
        )
        (
         "partitioner,r",
         po::value<amgcl::runtime::mpi::partition::type>()->default_value(
#if defined(AMGCL_HAVE_SCOTCH)
             amgcl::runtime::mpi::partition::ptscotch
#elif defined(AMGCL_HAVE_PARMETIS)
             amgcl::runtime::mpi::partition::parmetis
#else
             amgcl::runtime::mpi::partition::merge
#endif
             ),
         "Repartition the system matrix"
        )
        (
         "size,n",
         po::value<ptrdiff_t>()->default_value(128),
         "domain size"
        )
        ("prm-file,P",
         po::value<std::string>(),
         "Parameter file in json format. "
        )
        (
         "prm,p",
         po::value< std::vector<std::string> >()->multitoken(),
         "Parameters specified as name=value pairs. "
         "May be provided multiple times. Examples:\n"
         "  -p solver.tol=1e-3\n"
         "  -p precond.coarse_enough=300"
        )
        ;

    po::positional_options_description p;
    p.add("prm", -1);

    po::variables_map vm;
    po::store(po::command_line_parser(argc, argv).options(desc).positional(p).run(), vm);
    po::notify(vm);

    if (vm.count("help")) {
        if (comm.rank == 0) std::cout << desc << std::endl;
        return 0;
    }

    boost::property_tree::ptree prm;
    if (vm.count("prm-file")) {
        read_json(vm["prm-file"].as<std::string>(), prm);
    }

    if (vm.count("prm")) {
        for(const std::string &v : vm["prm"].as<std::vector<std::string> >()) {
            amgcl::put(prm, v);
        }
    }

    ptrdiff_t n;
    std::vector<ptrdiff_t> ptr;
    std::vector<ptrdiff_t> col;
    std::vector<double>    val;
    std::vector<double>    rhs;

    int block_size = vm["block-size"].as<int>();
    int aggr_block = prm.get("precond.coarsening.aggr.block_size", 1);

    bool binary = vm["binary"].as<bool>();
    amgcl::runtime::mpi::partition::type ptype = vm["partitioner"].as<amgcl::runtime::mpi::partition::type>();

    if (vm.count("matrix")) {
        prof.tic("read");
        if (binary) {
            n = read_binary(comm,
                    vm["matrix"].as<std::string>(),
                    vm["rhs"].as<std::string>(),
                    block_size * aggr_block, ptr, col, val, rhs);
        } else {
            n = read_matrix_market(comm,
                    vm["matrix"].as<std::string>(),
                    vm["rhs"].as<std::string>(),
                    block_size * aggr_block, ptr, col, val, rhs);
        }
        prof.toc("read");
    } else if (vm.count("Ap")) {
        prof.tic("read");
        ptype = static_cast<amgcl::runtime::mpi::partition::type>(0);

        std::vector<std::string> Aparts = vm["Ap"].as<std::vector<std::string>>();
        comm.check(Aparts.size() == static_cast<size_t>(comm.size),
                "--Ap should have single entry per MPI process");

        if (binary) {
            amgcl::io::read_crs(Aparts[comm.rank], n, ptr, col, val);
        } else {
            ptrdiff_t m;
            std::tie(n, m) = amgcl::io::mm_reader(Aparts[comm.rank])(ptr, col, val);
        }

        if (vm.count("fp")) {
            std::vector<std::string> fparts = vm["fp"].as<std::vector<std::string>>();
            comm.check(fparts.size() == static_cast<size_t>(comm.size),
                    "--fp should have single entry per MPI process");

            ptrdiff_t rows;
            ptrdiff_t cols;

            if (binary) {
                amgcl::io::read_dense(fparts[comm.rank], rows, cols, rhs);
            } else {
                std::tie(rows, cols) = amgcl::io::mm_reader(fparts[comm.rank])(rhs);
            }

            comm.check(rhs.size() == static_cast<size_t>(n), "Wrong RHS size");
        } else {
            rhs.resize(n, 1);
        }
        prof.toc("read");
    } else {
        prof.tic("assemble");
        n = assemble_poisson3d(comm,
                vm["size"].as<ptrdiff_t>(),
                block_size * aggr_block, ptr, col, val, rhs);
        prof.toc("assemble");
    }

    switch(block_size) {

#if defined(SOLVER_BACKEND_BUILTIN) || defined(SOLVER_BACKEND_VEXCL)
#  define AMGCL_CALL_BLOCK_SOLVER(z, data, B)                        \
        case B:                                                      \
            solve_block<B>(comm, n, ptr, col, val, prm, rhs, ptype); \
            break;

        BOOST_PP_SEQ_FOR_EACH(AMGCL_CALL_BLOCK_SOLVER, ~, AMGCL_BLOCK_SIZES)

#  undef AMGCL_CALL_BLOCK_SOLVER
#endif

        case 1:
            solve_scalar(comm, n, ptr, col, val, prm, rhs, ptype);
            break;
        default:
            if (comm.rank == 0)
                std::cout << "Unsupported block size!" << std::endl;
    }
}