lpsolver 0.1.0 → 0.2.1

data/ext/lpsolver-highs/highs/ipm/ipx/diagonal_precond.cc

@@ -0,0 +1,70 @@
+// Copyright (c) 2018-2019 ERGO-Code. See license.txt for license.
+
+#include "ipm/ipx/diagonal_precond.h"
+#include <cassert>
+#include <cmath>
+#include <vector>
+#include "ipm/ipx/timer.h"
+
+namespace ipx {
+
+DiagonalPrecond::DiagonalPrecond(const Model& model) : model_(model) {
+    const Int m = model_.rows();
+    diagonal_.resize(m);
+}
+
+void DiagonalPrecond::Factorize(const double* W, Info* info) {
+    const Int m = model_.rows();
+    const Int n = model_.cols();
+    const SparseMatrix& AI = model_.AI();
+
+    factorized_ = false;
+
+    // Build diagonal of normal matrix.
+    if (W) {
+        for (Int i = 0; i < m; i++)
+            diagonal_[i] = W[n+i];
+        for (Int j = 0; j < n; j++) {
+            double w = W[j];
+            for (Int p = AI.begin(j); p < AI.end(j); p++)
+                diagonal_[AI.index(p)] += AI.value(p) * w * AI.value(p);
+        }
+    } else {
+        diagonal_ = 0.0;        // rightmost m columns have weight zero
+        for (Int j = 0; j < n; j++) {
+            for (Int p = AI.begin(j); p < AI.end(j); p++)
+                diagonal_[AI.index(p)] += AI.value(p) * AI.value(p);
+        }
+    }
+
+    factorized_ = true;
+}
+
+double DiagonalPrecond::time() const {
+    return time_;
+}
+
+void DiagonalPrecond::reset_time() {
+    time_ = 0.0;
+}
+
+void DiagonalPrecond::_Apply(const Vector& rhs, Vector& lhs,
+                              double* rhs_dot_lhs) {
+    const Int m = model_.rows();
+    double rldot = 0.0;
+    Timer timer;
+
+    assert(factorized_);
+    assert(lhs.size() == static_cast<size_t>(m));
+    assert(rhs.size() == static_cast<size_t>(m));
+
+    for (Int i = 0; i < m; i++) {
+        lhs[i] = rhs[i] / diagonal_[i];
+        rldot += lhs[i] * rhs[i];
+    }
+    if (rhs_dot_lhs)
+        *rhs_dot_lhs = rldot;
+    time_ += timer.Elapsed();
+}
+
+}  // namespace ipx

data/ext/lpsolver-highs/highs/ipm/ipx/diagonal_precond.h

@@ -0,0 +1,45 @@
+// Copyright (c) 2018-2019 ERGO-Code. See license.txt for license.
+
+#ifndef IPX_DIAGONAL_PRECOND_H_
+#define IPX_DIAGONAL_PRECOND_H_
+
+#include "ipm/ipx/linear_operator.h"
+#include "ipm/ipx/model.h"
+#include "ipm/ipx/sparse_matrix.h"
+
+namespace ipx {
+
+// DiagonalPrecond provides inverse operations with the diagonal matrix
+//
+//   diag(AI*W*AI').                                (1)
+//
+// Here AI is the m-by-(n+m) matrix defined by the model, and W is a diagonal
+// (weight) matrix that is provided by the user.
+
+class DiagonalPrecond : public LinearOperator {
+public:
+    // Constructor stores a reference to the model. No data is copied. The model
+    // must be valid as long as the preconditioner is used.
+    explicit DiagonalPrecond(const Model& model);
+
+    // Factorizes the preconditioner. W must either hold n+m entries, or be
+    // NULL, in which case the first n entries are assumed 1.0 and the last
+    // m entries are assumed 0.0.
+    void Factorize(const double* W, Info* info);
+
+    // Returns computation time for calls to Apply() since last reset_time().
+    double time() const;
+    void reset_time();
+
+private:
+    void _Apply(const Vector& rhs, Vector& lhs, double* rhs_dot_lhs) override;
+
+    const Model& model_;
+    bool factorized_{false};    // preconditioner factorized?
+    Vector diagonal_;           // diagonal of normal matrix
+    double time_{0.0};
+};
+
+}  // namespace ipx
+
+#endif  // IPX_DIAGONAL_PRECOND_H_

data/ext/lpsolver-highs/highs/ipm/ipx/forrest_tomlin.cc

@@ -0,0 +1,360 @@
+#include "ipm/ipx/forrest_tomlin.h"
+#include <algorithm>
+#include <cassert>
+#include <utility>
+#include "ipm/ipx/utils.h"
+
+namespace ipx {
+
+ForrestTomlin::ForrestTomlin(const Control& control, Int dim,
+                             std::unique_ptr<LuFactorization>& lu) :
+    control_(control), dim_(dim) {
+    work_.resize(dim_ + kMaxUpdates);
+    lu_ = std::move(lu);
+}
+
+Int ForrestTomlin::_Factorize(const Int* Bbegin, const Int* Bend, const Int* Bi,
+                              const double* Bx, bool strict_abs_pivottol) {
+    // Reset updates.
+    R_.resize(dim_, 0);
+    replaced_.clear();
+    replace_next_ = -1;
+    have_btran_ = false;
+    have_ftran_ = false;
+
+    lu_->Factorize(dim_, Bbegin, Bend, Bi, Bx, pivottol_, strict_abs_pivottol,
+                   &L_, &U_, &rowperm_, &colperm_, &dependent_cols_);
+    rowperm_inv_ = InversePerm(rowperm_);
+    colperm_inv_ = InversePerm(colperm_);
+
+    // Compute fill factor.
+    Int bnz = 0;
+    for (Int j = 0; j < dim_; j++)
+        bnz += Bend[j]-Bbegin[j];
+    fill_factor_ = 1.0 * (L_.entries()+U_.entries()) / bnz;
+
+    if (control_.Debug(3)) {
+        double normLinv = NormestInverse(L_, "lower", 1);
+        double normUinv = NormestInverse(U_, "upper", 0);
+        control_.Debug(3)
+            << " normLinv = " << sci2(normLinv) << ','
+            << " normUinv = " << sci2(normUinv) << ','
+            << " stability = " << sci2(lu_->stability()) << '\n';
+    }
+
+    Int ret = 0;
+    if (lu_->stability() > kLuStabilityThreshold)
+        ret |= 1;
+    if (!dependent_cols_.empty())
+        ret |= 2;
+    return ret;
+}
+
+void ForrestTomlin::_GetFactors(SparseMatrix* L, SparseMatrix* U, Int* rowperm,
+                                Int* colperm, std::vector<Int>* dependent_cols){
+    if (L) *L = L_;
+    if (U) *U = U_;
+    if (rowperm)
+        std::copy(rowperm_.begin(), rowperm_.end(), rowperm);
+    if (colperm)
+        std::copy(colperm_.begin(), colperm_.end(), colperm);
+    if (dependent_cols)
+        *dependent_cols = dependent_cols_;
+}
+
+void ForrestTomlin::_SolveDense(const Vector& rhs, Vector& lhs, char trans) {
+    if (trans == 't' || trans == 'T') {
+        PermuteBack(colperm_, rhs, work_);
+        SolvePermuted(work_, 'T');
+        Permute(rowperm_, work_, lhs);
+    }
+    else {
+        PermuteBack(rowperm_, rhs, work_);
+        SolvePermuted(work_, 'N');
+        Permute(colperm_, work_, lhs);
+    }
+}
+
+void ForrestTomlin::_FtranForUpdate(Int nz, const Int* bi, const double* bx) {
+    ComputeSpike(nz, bi, bx);
+}
+
+void ForrestTomlin::_FtranForUpdate(Int nb, const Int* bi, const double* bx,
+                                   IndexedVector& lhs) {
+    ComputeSpike(nb, bi, bx);
+    TriangularSolve(U_, work_, 'n', "upper", 0);
+
+    // Move extra variables from updates to replaced positions.
+    Int num_updates = replaced_.size();
+    for (Int k = num_updates-1; k >= 0; k--)
+        work_[replaced_[k]] = work_[dim_+k];
+
+    // Return lhs without pattern.
+    for (Int p = 0; p < dim_; p++)
+        lhs[colperm_[p]] = work_[p];
+    lhs.InvalidatePattern();
+}
+
+void ForrestTomlin::_BtranForUpdate(Int j) {
+    ComputeEta(j);
+}
+
+void ForrestTomlin::_BtranForUpdate(Int j, IndexedVector& lhs) {
+    ComputeEta(j);
+
+    // Apply update etas and solve with L'.
+    Int num_updates = replaced_.size();
+    for (Int k = num_updates-1; k >= 0; k--) {
+        ScatterColumn(R_, k, -work_[dim_+k], work_);
+        work_[replaced_[k]] = work_[dim_+k];
+        work_[dim_+k] = 0.0;
+    }
+    TriangularSolve(L_, work_, 't', "lower", 1);
+
+    // Return lhs without pattern.
+    for (Int p = 0; p < dim_; p++)
+        lhs[rowperm_[p]] = work_[p];
+    lhs.InvalidatePattern();
+}
+
+// Dot product of entries in queues of A1 and A2. Indices must be sorted in
+// increasing order.
+static double SparseDot(const SparseMatrix& A1, const SparseMatrix& A2) {
+    Int q1 = A1.queue_size();
+    Int q2 = A2.queue_size();
+    Int p1 = 0;
+    Int p2 = 0;
+    double d = 0.0;
+    while (p1 < q1 && p2 < q2) {
+        Int i1 = A1.qindex(p1);
+        Int i2 = A2.qindex(p2);
+        if (i1 == i2) {
+            d += A1.qvalue(p1) * A2.qvalue(p2);
+            p1++; p2++;
+        } else if (i1 < i2) {
+            p1++;
+        } else {
+            p2++;
+        }
+    }
+    return d;
+}
+
+Int ForrestTomlin::_Update(double pivot) {
+    Int num_updates = replaced_.size();
+    assert(have_ftran_);
+    assert(have_btran_);
+
+    // Find the position of the entry with row index replace_next_ in spike.
+    // If not present, we will have where == qend.
+    assert(replace_next_ >= 0);
+    Int qend = U_.queue_size();
+    Int where = 0;
+    while (where < qend && U_.qindex(where) != replace_next_)
+        where++;
+
+    // Compute new diagonal entry of U. newdiag1 will be inserted into U;
+    // newdiag2 would be the same in exact arithmetic and is for monitoring
+    // numerical stability.
+    double olddiag = U_.value(U_.end(replace_next_)-1);
+    double newdiag1 = pivot * olddiag;
+    double newdiag2 = (where == qend ? 0.0 : U_.qvalue(where)) -
+        SparseDot(U_, R_);
+    double newdiag_err = std::abs(newdiag1-newdiag2);
+    double rel_newdiag_err = newdiag_err / std::abs(newdiag1);
+
+    // Put new diagonal entry at end of spike.
+    if (where < qend) {
+        for (Int l = where; l < qend-1; l++) {
+            U_.qindex(l) = U_.qindex(l+1);
+            U_.qvalue(l) = U_.qvalue(l+1);
+        }
+        U_.qindex(qend-1) = dim_+num_updates;
+        U_.qvalue(qend-1) = newdiag1;
+    } else {
+        U_.push_back(dim_+num_updates, newdiag1);
+    }
+
+    // Overwrite replaced column by unit column in U.
+    Int end = U_.end(replace_next_);
+    for (Int l = U_.begin(replace_next_); l < end-1; l++)
+        U_.value(l) = 0.0;
+    U_.value(end-1) = 1.0;
+
+    // Finish update.
+    U_.add_column();
+    R_.add_column();
+    replaced_.push_back(replace_next_);
+    replace_next_ = -1;
+    have_btran_ = false;
+    have_ftran_ = false;
+
+    if (newdiag1 == 0.0)
+        return -1;
+
+    // Print a debugging message if a new eta entry is large.
+    double max_eta = 0.0;
+    for (Int l = R_.begin(num_updates); l < R_.end(num_updates); l++)
+        max_eta = std::max(max_eta, std::abs(R_.value(l)));
+    if (max_eta > 1e10)
+        control_.Debug(3) << " max eta = " << sci2(max_eta) << '\n';
+
+    // stability check
+    if (rel_newdiag_err > kFtDiagErrorTol) {
+        control_.Debug(3)
+            << " relative error in new diagonal entry of U = "
+            << sci2(rel_newdiag_err) << '\n';
+        return 1;
+    }
+    return 0;
+}
+
+bool ForrestTomlin::_NeedFreshFactorization() {
+    Int num_updates = replaced_.size();
+    Int Rnz = R_.entries();        // nnz in accumulated row etas
+    Int Lnz = L_.entries() + dim_; // nnz(L) incl. diagonal
+    Int Unz = U_.entries();        // nnz(U) incl. zeroed out columns
+    Int U0nz = U_.begin(dim_);     // nnz(U) after factorization
+
+    if (num_updates == kMaxUpdates)
+        return true;
+    if (num_updates < 100)
+        return false;
+    if (Rnz > 1.0 * Lnz) {
+        // control_.Debug()
+        //     << " Rnz = " << Rnz << ", Lnz = " << Lnz
+        //     << " after " << num_updates << " updates\n";
+        return true;
+    }
+    if (Unz > 1.7 * U0nz) {
+        // control_.Debug()
+        //     << " Unz = " << Unz << ", U0nz = " << U0nz
+        //     << " after " << num_updates << " updates\n";
+        return true;
+    }
+    return false;
+}
+
+double ForrestTomlin::_fill_factor() const {
+    return fill_factor_;
+}
+
+double ForrestTomlin::_pivottol() const {
+    return pivottol_;
+}
+
+void ForrestTomlin::_pivottol(double new_pivottol) {
+    pivottol_ = new_pivottol;
+}
+
+void ForrestTomlin::SolvePermuted(Vector& lhs, char trans) {
+    Int num_updates = replaced_.size();
+    assert(U_.cols() == dim_+num_updates);
+
+    // Require num_updates elements workspace at end of lhs.
+    assert((Int)lhs.size() >= dim_+num_updates);
+
+    if (trans == 't' || trans == 'T') {
+        // Move replaced entries to the end of the pivot sequence and zero out
+        // their old position. Because the corresponding columns of U are unit
+        // columns now, the replaced positions remain zero when solving with U'.
+        // This is crucial because we neglected to zero out the corresponding
+        // rows of U in the update, so there are entries in U which actually
+        // should not be there.
+        for (Int k = 0; k < num_updates; k++) {
+            lhs[dim_+k] = lhs[replaced_[k]];
+            lhs[replaced_[k]] = 0.0;
+        }
+        TriangularSolve(U_, lhs, 't', "upper", 0);
+        // Solve backwards with row eta matrices (leading scatter operations)
+        // and put the entry from the end of the pivot sequence back into the
+        // position that its update replaced.
+        for (Int k = num_updates-1; k >= 0; k--) {
+            ScatterColumn(R_, k, -lhs[dim_+k], lhs);
+            lhs[replaced_[k]] = lhs[dim_+k];
+            lhs[dim_+k] = 0.0;
+        }
+        TriangularSolve(L_, lhs, 't', "lower", 1);
+    }
+    else {
+        TriangularSolve(L_, lhs, 'n', "lower", 1);
+        // Solve forward with row eta matrices (leading gather operations) and
+        // put the newly computed entry at the end of the pivot sequence. We
+        // actually would not need to zero out the replaced positions here
+        // because they will not be accessed by any subsequent eta matrix.
+        for (Int k = 0; k < num_updates; k++) {
+            lhs[dim_+k] = lhs[replaced_[k]] - DotColumn(R_, k, lhs);
+            lhs[replaced_[k]] = 0.0;
+        }
+        // The triangular solve with U fills the replaced position with garbage.
+        // This does not hurt because by the fact that the corresponding columns
+        // of U have no nonzero off-diagonal entries, the computed values are
+        // not propagated further. We finally overwrite them with their actual
+        // values from the end of the pivot sequence.
+        TriangularSolve(U_, lhs, 'n', "upper", 0);
+        for (Int k = num_updates-1; k >= 0; k--) {
+            lhs[replaced_[k]] = lhs[dim_+k];
+            lhs[dim_+k] = 0.0;
+        }
+    }
+}
+
+void ForrestTomlin::ComputeSpike(Int nb, const Int* bi, const double* bx) {
+    Int num_updates = replaced_.size();
+
+    // Solve L*lhs=b.
+    work_ = 0.0;
+    for (Int p = 0; p < nb; p++)
+        work_[rowperm_inv_[bi[p]]] = bx[p];
+    TriangularSolve(L_, work_, 'n', "lower", 1);
+
+    // Apply update etas.
+    for (Int k = 0; k < num_updates; k++) {
+        work_[dim_+k] = work_[replaced_[k]] - DotColumn(R_, k, work_);
+        work_[replaced_[k]] = 0.0;
+    }
+
+    // Store spike in U. Indices are sorted, which is required for the sparse
+    // dot product in Update().
+    U_.clear_queue();
+    for (Int p = 0; p < dim_+num_updates; p++) {
+        if (work_[p] != 0.0)
+            U_.push_back(p, work_[p]);
+    }
+    have_ftran_ = true;
+}
+
+void ForrestTomlin::ComputeEta(Int j) {
+    Int num_updates = replaced_.size();
+    assert(U_.cols() == dim_+num_updates);
+
+    // Find permuted position of j.
+    Int pos = colperm_inv_[j];
+    for (Int k = 0; k < num_updates; k++)
+        if (replaced_[k] == pos)
+            pos = dim_+k;
+
+    // Solve lhs'U=e_pos'. Replaced position must remain zero (by the fact that
+    // the corresponding columns of U are unit), so replaced indices cannot
+    // occur in the new eta matrix.
+    work_ = 0.0;
+    work_[pos] = 1.0;
+    TriangularSolve(U_, work_, 't', "upper", 0);
+    #ifndef NDEBUG
+    for (Int k = 0; k < num_updates; k++)
+        assert(work_[replaced_[k]] == 0.0);
+    #endif
+
+    // Queue eta at end of R. Indices are sorted, which is required for the
+    // sparse dot product in Update().
+    R_.clear_queue();
+    double pivot = work_[pos];
+    for (Int i = pos+1; i < dim_+num_updates; i++) {
+        if (work_[i] != 0.0)
+            R_.push_back(i, -work_[i]/pivot);
+    }
+    have_btran_ = true;
+    replace_next_ = pos;
+}
+
+}  // namespace ipx

data/ext/lpsolver-highs/highs/ipm/ipx/forrest_tomlin.h

@@ -0,0 +1,102 @@
+#ifndef IPX_FORREST_TOMLIN_H_
+#define IPX_FORREST_TOMLIN_H_
+
+#include <memory>
+#include <vector>
+#include "ipm/ipx/control.h"
+#include "ipm/ipx/lu_factorization.h"
+#include "ipm/ipx/lu_update.h"
+
+namespace ipx {
+
+// Generic implementation of the Forrest-Tomlin update [1] that can be used with
+// any LU factorization. The implementation does not exploit hypersparsity,
+// which excludes its use for such problems. For non-hypersparse problems the
+// implementation is better suited than BASICLU, however, because it stores L
+// and U in compressed form with permuted indices; hence solving triangular
+// systems with a dense rhs/lhs accesses memory contiguously. BASICLU could not
+// be implemented in that form due to re-permuting U to triangular form in the
+// updates. Hypersparsity support could be added to the present implementation
+// when required.
+//
+// [1] J.J.H. Forrest and J.A. Tomlin, "Updated triangular factors of the basis
+//     to maintain sparsity in the product form simplex method", Math.
+//     Programming 2 (1972)
+
+class ForrestTomlin : public LuUpdate {
+public:
+    // Constructor takes ownership of the LuFactorization object by a move from
+    // @lu.
+    ForrestTomlin(const Control& control, Int dim,
+                  std::unique_ptr<LuFactorization>& lu);
+    ~ForrestTomlin() = default;
+
+private:
+    Int _Factorize(const Int* Bbegin, const Int* Bend, const Int* Bi,
+                   const double* Bx, bool strict_abs_pivottol) override;
+    void _GetFactors(SparseMatrix* L, SparseMatrix* U, Int* rowperm,
+                     Int* colperm, std::vector<Int>* dependent_cols) override;
+    void _SolveDense(const Vector& rhs, Vector& lhs, char trans) override;
+    void _FtranForUpdate(Int nz, const Int* bi, const double* bx) override;
+    void _FtranForUpdate(Int nz, const Int* bi, const double* bx,
+                         IndexedVector& lhs) override;
+    void _BtranForUpdate(Int j) override;
+    void _BtranForUpdate(Int j, IndexedVector& lhs) override;
+    Int _Update(double pivot) override;
+    bool _NeedFreshFactorization() override;
+    double _fill_factor() const override;
+    double _pivottol() const override;
+    void _pivottol(double new_pivottol) override;
+
+    // Maximum # updates before refactorization is required.
+    static constexpr Int kMaxUpdates = 5000;
+
+    // Solves a linear system with the basis matrix. On entry @x holds the
+    // permuted right-hand side, on return it holds the permuted solution.
+    // @x must have dimension at least dim_ + # computed updates; the additional
+    // components are used as workspace.
+    // @trans: 't' or 'T' for transposed system.
+    void SolvePermuted(Vector& x, char trans);
+
+    // Computes the spike column for the FT update from
+    //   R_k^{-1} * ... * R_1^{-1} * L^{-1} * b
+    // and stores it in compressed form at the end of U. The spike is also
+    // returned as a full vector in work_.
+    // @nb, @bi, @bx: b in compressed form.
+    void ComputeSpike(Int nb, const Int* bi, const double* bx);
+
+    // Computes the partial BTRAN solution r = ep' * U^{-1}, where ep is the
+    // p-th unit vector and p the position of column @j in the pivot sequence.
+    // The row eta vector -r/r[p] without the unit diagonal entry is stored in
+    // compressed form at end of R. r is returned as a full vector in work_.
+    void ComputeEta(Int j);
+
+    const Control& control_;
+    const Int dim_;
+    std::unique_ptr<LuFactorization> lu_;
+
+    std::vector<Int> rowperm_;     // row permutation from factorization
+    std::vector<Int> colperm_;     // col permutation from factorization
+    std::vector<Int> rowperm_inv_; // inverse permutation of rowperm_
+    std::vector<Int> colperm_inv_; // inverse permutation of colperm_
+    std::vector<Int> dependent_cols_;
+                             // passed trough from _Factorize() to _GetFactors()
+
+    SparseMatrix L_;            // L from factorization
+    SparseMatrix U_;            // U from factorization with spike cols appended
+    SparseMatrix R_;            // cols of R build row eta matrices from updates
+
+    // replaced_[k] == p if update k replaced position p in pivot sequence.
+    // replaced_.size() is the # updates performed.
+    std::vector<Int> replaced_;
+    Int replace_next_;          // position to be replaced in next update
+    bool have_btran_{false};    // true if row eta has been computed
+    bool have_ftran_{false};    // true if spike has been computed
+    double fill_factor_{0.0};   // fill factor from last factorization
+    double pivottol_{0.1};      // LU pivot tolerance for next factorization
+    Vector work_;               // size dim_ + kMaxUpdates workspace
+};
+
+}  // namespace ipx
+
+#endif  // IPX_FORREST_TOMLIN_H_