SparseMatSuperLU.hpp

Go to the documentation of this file.

/*******************************************************************************
 * Copyright (C) 2017-2023 Theodore Chang
 *
 * This program is free software: you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.
 ******************************************************************************/
// ReSharper disable CppCStyleCast
#ifndef SPARSEMATSUPERLU_HPP
#define SPARSEMATSUPERLU_HPP
 
#include <superlu-mt/superlu-mt.h>
#include "SparseMat.hpp"
#include "csc_form.hpp"
 
template<sp_d T> class SparseMatSuperLU final : public SparseMat<T> {
    SuperMatrix A{}, L{}, U{}, B{};
 
#ifndef SUANPAN_SUPERLUMT
    superlu_options_t options{};
 
    SuperLUStat_t stat{};
#else
    const int ordering_num = 1;
 
    Gstat_t stat{};
#endif
 
    void* t_val = nullptr;
    int* t_row = nullptr;
    int* t_col = nullptr;
 
    int* perm_r = nullptr;
    int* perm_c = nullptr;
 
    bool allocated = false;
 
    template<sp_d ET> void alloc_supermatrix(csc_form<ET, int>&&);
    void dealloc_supermatrix();
 
    template<sp_d ET> void wrap_b(const Mat<ET>&);
    template<sp_d ET> void tri_solve(int&);
    template<sp_d ET> void full_solve(int&);
 
    int solve_trs(Mat<T>&, Mat<T>&&);
    int solve_trs(Mat<T>&, const Mat<T>&);
 
public:
    SparseMatSuperLU(uword, uword, uword = 0);
    SparseMatSuperLU(const SparseMatSuperLU&);
    SparseMatSuperLU(SparseMatSuperLU&&) noexcept = delete;
    SparseMatSuperLU& operator=(const SparseMatSuperLU&) = delete;
    SparseMatSuperLU& operator=(SparseMatSuperLU&&) noexcept = delete;
    ~SparseMatSuperLU() override;
 
    void zeros() override;
 
    unique_ptr<MetaMat<T>> make_copy() override;
 
    int direct_solve(Mat<T>&, Mat<T>&&) override;
    int direct_solve(Mat<T>&, const Mat<T>&) override;
};
 
template<sp_d T> template<sp_d ET> void SparseMatSuperLU<T>::alloc_supermatrix(csc_form<ET, int>&& in) {
    dealloc_supermatrix();
 
    auto t_size = sizeof(ET) * in.n_elem;
    t_val = superlu_malloc(t_size);
    memcpy(t_val, (void*)in.val_mem(), t_size);
 
    t_size = sizeof(int) * in.n_elem;
    t_row = (int*)superlu_malloc(t_size);
    memcpy(t_row, (void*)in.row_mem(), t_size);
 
    t_size = sizeof(int) * (in.n_cols + 1llu);
    t_col = (int*)superlu_malloc(t_size);
    memcpy(t_col, (void*)in.col_mem(), t_size);
 
    if(std::is_same_v<ET, double>) {
        using E = double;
        dCreate_CompCol_Matrix(&A, in.n_rows, in.n_cols, in.n_elem, (E*)t_val, t_row, t_col, Stype_t::SLU_NC, Dtype_t::SLU_D, Mtype_t::SLU_GE);
    }
    else {
        using E = float;
        sCreate_CompCol_Matrix(&A, in.n_rows, in.n_cols, in.n_elem, (E*)t_val, t_row, t_col, Stype_t::SLU_NC, Dtype_t::SLU_S, Mtype_t::SLU_GE);
    }
 
    perm_r = (int*)superlu_malloc(sizeof(int) * (this->n_rows + 1));
    perm_c = (int*)superlu_malloc(sizeof(int) * (this->n_cols + 1));
 
    allocated = true;
}
 
template<sp_d T> void SparseMatSuperLU<T>::dealloc_supermatrix() {
    if(!allocated) return;
 
    Destroy_SuperMatrix_Store(&A);
#ifdef SUANPAN_SUPERLUMT
    Destroy_SuperNode_SCP(&L);
    Destroy_CompCol_NCP(&U);
#else
    Destroy_SuperNode_Matrix(&L);
    Destroy_CompCol_Matrix(&U);
#endif
 
    if(t_val) superlu_free(t_val);
    if(t_row) superlu_free(t_row);
    if(t_col) superlu_free(t_col);
    if(perm_r) superlu_free(perm_r);
    if(perm_c) superlu_free(perm_c);
 
    allocated = false;
}
 
template<sp_d T> template<sp_d ET> void SparseMatSuperLU<T>::wrap_b(const Mat<ET>& in_mat) {
    if(std::is_same_v<ET, float>) {
        using E = float;
        sCreate_Dense_Matrix(&B, (int)in_mat.n_rows, (int)in_mat.n_cols, (E*)in_mat.memptr(), (int)in_mat.n_rows, Stype_t::SLU_DN, Dtype_t::SLU_S, Mtype_t::SLU_GE);
    }
    else {
        using E = double;
        dCreate_Dense_Matrix(&B, (int)in_mat.n_rows, (int)in_mat.n_cols, (E*)in_mat.memptr(), (int)in_mat.n_rows, Stype_t::SLU_DN, Dtype_t::SLU_D, Mtype_t::SLU_GE);
    }
}
 
template<sp_d T> template<sp_d ET> void SparseMatSuperLU<T>::tri_solve(int& flag) {
#ifdef SUANPAN_SUPERLUMT
    if(std::is_same_v<ET, float>) sgstrs(NOTRANS, &L, &U, perm_c, perm_r, &B, &stat, &flag);
    else dgstrs(NOTRANS, &L, &U, perm_c, perm_r, &B, &stat, &flag);
#else
    superlu::gstrs<ET>(options.Trans, &L, &U, perm_c, perm_r, &B, &stat, &flag);
#endif
 
    Destroy_SuperMatrix_Store(&B);
}
 
template<sp_d T> template<sp_d ET> void SparseMatSuperLU<T>::full_solve(int& flag) {
#ifdef SUANPAN_SUPERLUMT
    get_perm_c(ordering_num, &A, perm_c);
    if(std::is_same_v<ET, float>) psgssv(SUANPAN_NUM_THREADS, &A, perm_c, perm_r, &L, &U, &B, &flag);
    else pdgssv(SUANPAN_NUM_THREADS, &A, perm_c, perm_r, &L, &U, &B, &flag);
#else
    superlu::gssv<ET>(&options, &A, perm_c, perm_r, &L, &U, &B, &stat, &flag);
#endif
 
    Destroy_SuperMatrix_Store(&B);
}
 
template<sp_d T> SparseMatSuperLU<T>::SparseMatSuperLU(const uword in_row, const uword in_col, const uword in_elem)
    : SparseMat<T>(in_row, in_col, in_elem) {
#ifndef SUANPAN_SUPERLUMT
    set_default_options(&options);
    options.IterRefine = std::is_same_v<T, float> ? superlu::IterRefine_t::SLU_SINGLE : superlu::IterRefine_t::SLU_DOUBLE;
    options.Equil = superlu::yes_no_t::NO;
 
    arrayops::fill_zeros(reinterpret_cast<char*>(&stat), sizeof(SuperLUStat_t));
 
    StatInit(&stat);
#else
    StatAlloc(static_cast<int>(in_col), SUANPAN_NUM_THREADS, sp_ienv(1), sp_ienv(2), &stat);
    StatInit(static_cast<int>(in_col), SUANPAN_NUM_THREADS, &stat);
#endif
}
 
template<sp_d T> SparseMatSuperLU<T>::SparseMatSuperLU(const SparseMatSuperLU& other)
    : SparseMat<T>(other) {
#ifndef SUANPAN_SUPERLUMT
    set_default_options(&options);
    options.IterRefine = std::is_same_v<T, float> ? superlu::IterRefine_t::SLU_SINGLE : superlu::IterRefine_t::SLU_DOUBLE;
    options.Equil = superlu::yes_no_t::NO;
 
    arrayops::fill_zeros(reinterpret_cast<char*>(&stat), sizeof(SuperLUStat_t));
 
    StatInit(&stat);
#else
    StatAlloc(static_cast<int>(other.n_cols), SUANPAN_NUM_THREADS, sp_ienv(1), sp_ienv(2), &stat);
    StatInit(static_cast<int>(other.n_cols), SUANPAN_NUM_THREADS, &stat);
#endif
}
 
template<sp_d T> SparseMatSuperLU<T>::~SparseMatSuperLU() {
    dealloc_supermatrix();
    StatFree(&stat);
}
 
template<sp_d T> void SparseMatSuperLU<T>::zeros() {
    SparseMat<T>::zeros();
    dealloc_supermatrix();
}
 
template<sp_d T> unique_ptr<MetaMat<T>> SparseMatSuperLU<T>::make_copy() { return std::make_unique<SparseMatSuperLU<T>>(*this); }
 
template<sp_d T> int SparseMatSuperLU<T>::direct_solve(Mat<T>& out_mat, const Mat<T>& in_mat) {
    if(this->factored) return solve_trs(out_mat, in_mat);
 
    this->factored = true;
 
    auto flag = 0;
 
    if(std::is_same_v<T, float> || Precision::FULL == this->setting.precision) {
        alloc_supermatrix(csc_form<T, int>(this->triplet_mat));
 
        out_mat = in_mat;
 
        wrap_b(out_mat);
 
        full_solve<T>(flag);
 
        return flag;
    }
 
    alloc_supermatrix(csc_form<float, int>(this->triplet_mat));
 
    const fmat f_mat(arma::size(in_mat), fill::none);
 
    wrap_b(f_mat);
 
    full_solve<float>(flag);
 
    return 0 == flag ? solve_trs(out_mat, in_mat) : flag;
}
 
template<sp_d T> int SparseMatSuperLU<T>::solve_trs(Mat<T>& out_mat, const Mat<T>& in_mat) {
    auto flag = 0;
 
    if(std::is_same_v<T, float> || Precision::FULL == this->setting.precision) {
        out_mat = in_mat;
 
        wrap_b(out_mat);
 
        tri_solve<T>(flag);
 
        return flag;
    }
 
    out_mat.zeros(arma::size(in_mat));
 
    mat full_residual = in_mat;
 
    auto multiplier = norm(full_residual);
 
    auto counter = 0u;
    while(counter++ < this->setting.iterative_refinement) {
        if(multiplier < this->setting.tolerance) break;
 
        auto residual = conv_to<fmat>::from(full_residual / multiplier);
 
        wrap_b(residual);
 
        tri_solve<float>(flag);
 
        if(0 != flag) break;
 
        const mat incre = multiplier * conv_to<mat>::from(residual);
 
        out_mat += incre;
 
        suanpan_debug("Mixed precision algorithm multiplier: {:.5E}.\n", multiplier = norm(full_residual -= this->operator*(incre)));
    }
 
    return flag;
}
 
template<sp_d T> int SparseMatSuperLU<T>::direct_solve(Mat<T>& out_mat, Mat<T>&& in_mat) {
    if(this->factored) return solve_trs(out_mat, std::forward<Mat<T>>(in_mat));
 
    this->factored = true;
 
    auto flag = 0;
 
    if(std::is_same_v<T, float> || Precision::FULL == this->setting.precision) {
        alloc_supermatrix(csc_form<T, int>(this->triplet_mat));
 
        wrap_b(in_mat);
 
        full_solve<T>(flag);
 
        out_mat = std::move(in_mat);
 
        return flag;
    }
 
    alloc_supermatrix(csc_form<float, int>(this->triplet_mat));
 
    const fmat f_mat(arma::size(in_mat), fill::none);
 
    wrap_b(f_mat);
 
    full_solve<float>(flag);
 
    return 0 == flag ? solve_trs(out_mat, std::forward<Mat<T>>(in_mat)) : flag;
}
 
template<sp_d T> int SparseMatSuperLU<T>::solve_trs(Mat<T>& out_mat, Mat<T>&& in_mat) {
    auto flag = 0;
 
    if(std::is_same_v<T, float> || Precision::FULL == this->setting.precision) {
        wrap_b(in_mat);
 
        tri_solve<T>(flag);
 
        out_mat = std::move(in_mat);
 
        return flag;
    }
 
    out_mat.zeros(arma::size(in_mat));
 
    auto multiplier = arma::norm(in_mat);
 
    auto counter = 0u;
    while(counter++ < this->setting.iterative_refinement) {
        if(multiplier < this->setting.tolerance) break;
 
        auto residual = conv_to<fmat>::from(in_mat / multiplier);
 
        wrap_b(residual);
 
        tri_solve<float>(flag);
 
        if(0 != flag) break;
 
        const mat incre = multiplier * conv_to<mat>::from(residual);
 
        out_mat += incre;
 
        suanpan_debug("Mixed precision algorithm multiplier: {:.5E}.\n", multiplier = norm(in_mat -= this->operator*(incre)));
    }
 
    return flag;
}
#endif