lpsolver 0.1.0 → 0.2.1

data/ext/lpsolver-highs/highs/ipm/hipo/factorhighs/FactorHiGHS.h

@@ -0,0 +1,112 @@
+#ifndef FACTOR_HIGHS_H
+#define FACTOR_HIGHS_H
+
+#include "CliqueStack.h"
+#include "DataCollector.h"
+#include "Numeric.h"
+#include "Symbolic.h"
+#include "ipm/hipo/auxiliary/Log.h"
+
+/*
+
+Direct solver for IPM matrices.
+
+Consider a sparse symmetric matrix M in CSC format.
+Only its lower triangular part is used; entries in the upper triangle are
+ignored.
+The matrix has n rows/cols and nz nonzero entries in the lower triangle.
+It is stored using three arrays:
+- ptr, column pointers, of length n+1;
+- rows, row indices, of length nz;
+- vals, values, of length nz.
+
+The direct solver uses the following objects:
+- Symbolic, to store the symbolic factorization;
+- FHsolver, to perform analyse and factorise phases.
+
+Define a vector signs that contains the expected sign of each pivot (1 or -1).
+Define a right-hand side rhs, which will be overwritten with the solution of
+M^{-1} * rhs. The pre-computed fill-reducing ordering to use is stored in the
+vector perm.
+
+Then, the factorization is performed as follows.
+
+    Symbolic S;
+    FHsolver FH;
+    FH.analyse(S, rows, ptr, signs, perm);
+    FH.factorise(S, rows, ptr, val);
+    FH.solve(x);
+
+Printing to screen is achieved using the interface in auxiliary/Log.h. Pass an
+object of type Log for normal printing:
+    ...
+    Log log;
+    FHsolver FH(&log);
+    ...
+Pass an object of type LogHighs for Highs logging.
+Pass nothing to suppress all logging.
+
+To add static regularisation when the pivots are selected, use
+setRegularisation(reg_p,reg_d) to choose values of primal and dual
+regularisation. If regularisation is already added to the matrix, ignore.
+
+The default block size is 128. To set a different block size, pass it as second
+input to the constructor.
+
+*/
+
+namespace hipo {
+
+class FHsolver {
+  const Log* log_;
+  DataCollector data_;
+  Regul regul_;
+  Numeric N_;
+  CliqueStack serial_stack_;
+
+  const Int nb_;  // block size
+  static const Int default_nb_ = 128;
+
+  // Columns of factorisation, stored by supernode.
+  // This memory is allocated the first time that it is used. Subsequent
+  // factorisations reuse the same memory.
+  std::vector<std::vector<double>> sn_columns_;
+
+ public:
+  // Create object and initialise DataCollector
+  FHsolver(const Log* log = nullptr, Int block_size = default_nb_);
+
+  // Print collected data (if any) and terminate DataCollector
+  ~FHsolver();
+
+  // Perform analyse phase of matrix with sparsity pattern given by rows and
+  // ptr, and store symbolic factorisation in object S.
+  // See ReturnValues.h for errors.
+  Int analyse(Symbolic& S, const std::vector<Int>& rows,
+              const std::vector<Int>& ptr, const std::vector<Int>& signs,
+              const std::vector<Int>& perm);
+
+  // Perform factorise phase of matrix given by rows, ptr, vals, and store
+  // numerical factorisation in object N. See ReturnValues.h for errors.
+  Int factorise(const Symbolic& S, const std::vector<Int>& rows,
+                const std::vector<Int>& ptr, const std::vector<double>& vals);
+
+  // Perform solve phase with rhs given by x, which is overwritten with the
+  // solution.
+  Int solve(std::vector<double>& x);
+
+  // If multiple factorisation are performed, call newIter() before each
+  // factorisation. This is used only to collect data for debugging, if
+  // expensive data collection is turned on at compile time.
+  void newIter();
+
+  // Set values for static regularisation to be added when a pivot is selected.
+  // If regularisation is already added to the matrix, ignore.
+  void setRegularisation(double reg_p, double reg_d);
+
+  void getRegularisation(std::vector<double>& reg);
+};
+
+}  // namespace hipo
+
+#endif

data/ext/lpsolver-highs/highs/ipm/hipo/factorhighs/FactorHiGHSSettings.h

@@ -0,0 +1,63 @@
+#ifndef FACTOR_HIGHS_SETTINGS_H
+#define FACTOR_HIGHS_SETTINGS_H
+
+#include <cmath>
+
+#include "ipm/hipo/auxiliary/IntConfig.h"
+
+// ===========================================================================
+// SWITCHES
+// ===========================================================================
+
+// Switch on/off pivoting. It uses a static variation of Bunch-Kaufman pivoting,
+// with potential dynamic regularisation. If pivoting is switched off, only
+// static regularisation is applied.
+#define HIPO_PIVOTING
+
+// Collect data during regularisation, e.g. number of regularised pivots, 2x2
+// pivots, pivot swaps, pivots with wrong sign, min and max entry of L and D.
+// This can be quite expensive and should only be used for debugging.
+// #define HIPO_COLLECT_EXPENSIVE_DATA
+
+// Choose level of timing:
+// - 0: no timing
+// - 1: basic timing
+// - 2: advanced timing
+// - 3: extreme timing (timing of each BLAS call, considerably slower)
+#define HIPO_TIMING_LEVEL 0
+
+// ===========================================================================
+// PARAMETERS
+// ===========================================================================
+
+namespace hipo {
+
+// supernode amalgamation
+const Int kStartThreshRelax = 256;
+const double kUpperRatioRelax = 0.02;
+const double kLowerRatioRelax = 0.01;
+const Int kMaxIterRelax = 20;
+const Int kSnSizeRelax = 16;
+const double kSpopsWeightSn = 50.0;
+
+// dense factorisation
+const double kAlphaBK = 0.01;  //(sqrt(17.0) + 1.0) / 8.0;
+const Int kBlockGrainSize = 1;
+const Int kBlockParallelThreshold = 5;
+
+const Int kMinConsecutiveSums = 1;
+
+// regularisation
+const double kDynamicDiagCoeff = 1e-24;
+
+// metis
+const Int kMetisSeed = 42;
+
+struct Regul {
+  double primal{};
+  double dual{};
+};
+
+}  // namespace hipo
+
+#endif

data/ext/lpsolver-highs/highs/ipm/hipo/factorhighs/Factorise.cpp

@@ -0,0 +1,405 @@
+#include "Factorise.h"
+
+#include <algorithm>
+#include <fstream>
+
+#include "DataCollector.h"
+#include "FactorHiGHSSettings.h"
+#include "FormatHandler.h"
+#include "HybridHybridFormatHandler.h"
+#include "ReturnValues.h"
+#include "ipm/hipo/auxiliary/Auxiliary.h"
+#include "ipm/hipo/auxiliary/Log.h"
+#include "parallel/HighsParallel.h"
+
+namespace hipo {
+
+Factorise::Factorise(const Symbolic& S, const std::vector<Int>& rowsM,
+                     const std::vector<Int>& ptrM,
+                     const std::vector<double>& valM, const Regul& regul,
+                     const Log* log, DataCollector& data,
+                     std::vector<std::vector<double>>& sn_columns,
+                     CliqueStack* stack)
+    : S_{S},
+      sn_columns_{sn_columns},
+      regul_{regul},
+      log_{log},
+      data_{data},
+      stack_{stack} {
+  // Input the symmetric matrix to be factorised in CSC format and the symbolic
+  // factorisation coming from Analyse.
+  // Only the lower triangular part of the matrix is used.
+
+  n_ = ptrM.size() - 1;
+
+  if (n_ != S_.size()) {
+    if (log_)
+      log_->printDevInfo(
+          "Matrix provided to Factorise has size incompatible with symbolic "
+          "object.\n");
+    return;
+  }
+
+  // Make a copy of the matrix to be factorised
+  rowsM_ = rowsM;
+  valM_ = valM;
+  ptrM_ = ptrM;
+
+  // Permute the matrix.
+  // This also removes any entry not in the lower triangle.
+  permute(S_.iperm());
+
+  nzM_ = ptrM_.back();
+
+  // Double transpose to sort columns
+  std::vector<Int> temp_ptr(n_ + 1);
+  std::vector<Int> temp_rows(nzM_);
+  std::vector<double> temp_val(nzM_);
+  transpose(ptrM_, rowsM_, valM_, temp_ptr, temp_rows, temp_val);
+  transpose(temp_ptr, temp_rows, temp_val, ptrM_, rowsM_, valM_);
+
+  // create linked lists of children in supernodal elimination tree
+  childrenLinkedList(S_.snParent(), first_child_, next_child_);
+
+  // create reverse linked lists of children
+  first_child_reverse_ = first_child_;
+  next_child_reverse_ = next_child_;
+  reverseLinkedList(first_child_reverse_, next_child_reverse_);
+
+  // compute largest diagonal entry in absolute value
+  max_diag_ = 0.0;
+  min_diag_ = kHighsInf;
+  for (Int col = 0; col < n_; ++col) {
+    double val = std::abs(valM_[ptrM_[col]]);
+    max_diag_ = std::max(max_diag_, val);
+    min_diag_ = std::min(min_diag_, val);
+  }
+
+  // one norm of columns of M
+  std::vector<double> one_norm_cols(n_, 0.0);
+  for (Int col = 0; col < n_; ++col) {
+    for (Int el = ptrM_[col]; el < ptrM_[col + 1]; ++el) {
+      Int row = rowsM_[el];
+      double val = valM_[el];
+      one_norm_cols[col] += std::abs(val);
+      if (row != col) one_norm_cols[row] += std::abs(val);
+    }
+  }
+  M_norm1_ = *std::max_element(one_norm_cols.begin(), one_norm_cols.end());
+
+  data_.setNorms(M_norm1_, max_diag_);
+}
+
+void Factorise::permute(const std::vector<Int>& iperm) {
+  // Symmetric permutation of the lower triangular matrix M based on inverse
+  // permutation iperm.
+  // The resulting matrix is lower triangular, regardless of the input matrix.
+
+  std::vector<Int> work(n_, 0);
+
+  // go through the columns to count the nonzeros
+  for (Int j = 0; j < n_; ++j) {
+    // get new index of column
+    const Int col = iperm[j];
+
+    // go through elements of column
+    for (Int el = ptrM_[j]; el < ptrM_[j + 1]; ++el) {
+      const Int i = rowsM_[el];
+
+      // ignore potential entries in upper triangular part
+      if (i < j) continue;
+
+      // get new index of row
+      const Int row = iperm[i];
+
+      // since only lower triangular part is used, col is smaller than row
+      Int actual_col = std::min(row, col);
+      ++work[actual_col];
+    }
+  }
+
+  std::vector<Int> new_ptr(n_ + 1);
+
+  // get column pointers by summing the count of nonzeros in each column.
+  // copy column pointers into work
+  counts2Ptr(new_ptr, work);
+
+  std::vector<Int> new_rows(new_ptr.back());
+  std::vector<double> new_val(new_ptr.back());
+
+  // go through the columns to assign row indices
+  for (Int j = 0; j < n_; ++j) {
+    // get new index of column
+    const Int col = iperm[j];
+
+    // go through elements of column
+    for (Int el = ptrM_[j]; el < ptrM_[j + 1]; ++el) {
+      const Int i = rowsM_[el];
+
+      // ignore potential entries in upper triangular part
+      if (i < j) continue;
+
+      // get new index of row
+      const Int row = iperm[i];
+
+      // since only lower triangular part is used, col is smaller than row
+      const Int actual_col = std::min(row, col);
+      const Int actual_row = std::max(row, col);
+
+      Int pos = work[actual_col]++;
+      new_rows[pos] = actual_row;
+      new_val[pos] = valM_[el];
+    }
+  }
+
+  ptrM_ = std::move(new_ptr);
+  rowsM_ = std::move(new_rows);
+  valM_ = std::move(new_val);
+}
+
+void Factorise::processSupernode(Int sn) {
+  // Assemble frontal matrix for supernode sn, perform partial factorisation and
+  // store the result.
+
+  TaskGroupSpecial tg;
+  HIPO_CLOCK_CREATE;
+
+  const bool parallel = S_.parTree();
+  const bool serial = !parallel;
+
+  if (flag_stop_.load(std::memory_order_relaxed)) return;
+
+  if (parallel) {
+    // spawn children of this supernode in forward order
+    Int child_to_spawn = first_child_[sn];
+    while (child_to_spawn != -1) {
+      tg.spawn([=]() { processSupernode(child_to_spawn); });
+      child_to_spawn = next_child_[child_to_spawn];
+    }
+
+    // wait for first child to finish, before starting the parent (if there is a
+    // first child)
+    if (first_child_[sn] != -1) tg.sync();
+  }
+
+  // ===================================================
+  // Supernode information
+  // ===================================================
+  HIPO_CLOCK_START(2);
+  // first and last+1 column of the supernodes
+  const Int sn_begin = S_.snStart(sn);
+  const Int sn_end = S_.snStart(sn + 1);
+  const Int sn_size = sn_end - sn_begin;
+
+  // When the tree is processed in serial, use CliqueStack to store the cliques.
+  // Otherwise, use local storage in FormatHandler.
+  double* clique_ptr = nullptr;
+  if (serial) {
+    bool reallocation = false;
+    clique_ptr = stack_->setup(S_.cliqueSize(sn), reallocation);
+    if (reallocation && log_)
+      log_->printDevInfo("Reallocation of CliqueStack\n");
+  }
+
+  // initialise the format handler
+  // this also allocates space for the frontal matrix and schur complement
+  std::unique_ptr<FormatHandler> FH(new HybridHybridFormatHandler(
+      S_, sn, regul_, data_, sn_columns_[sn], clique_ptr));
+
+  HIPO_CLOCK_STOP(2, data_, kTimeFactorisePrepare);
+
+  // ===================================================
+  // Assemble original matrix M into frontal
+  // ===================================================
+  HIPO_CLOCK_START(2);
+  // j is relative column index in the frontal matrix
+  for (Int j = 0; j < sn_size; ++j) {
+    // column index in the original matrix
+    const Int col = sn_begin + j;
+
+    // go through the column
+    for (Int el = ptrM_[col]; el < ptrM_[col + 1]; ++el) {
+      // relative row index in the frontal matrix
+      const Int i = S_.relindCols(el);
+
+      FH->assembleFrontal(i, j, valM_[el]);
+    }
+  }
+  HIPO_CLOCK_STOP(2, data_, kTimeFactoriseAssembleOriginal);
+
+  // ===================================================
+  // Assemble frontal matrices of children
+  // ===================================================
+  Int child_sn = first_child_reverse_[sn];
+  while (child_sn != -1) {
+    // Child contribution is found:
+    // - in cliquestack, if we are processing the tree in serial.
+    // - in schur_contribution_ if we are processing the tree in parallel.
+    // Children are always summed from last to first.
+
+    const double* child_clique;
+
+    if (parallel) {
+      // sync with spawned child, apart from the first one
+      if (child_sn != first_child_reverse_[sn]) tg.sync();
+
+      if (flag_stop_.load(std::memory_order_relaxed)) return;
+
+      child_clique = schur_contribution_[child_sn].data();
+
+      if (!child_clique) {
+        if (log_) log_->printDevInfo("Missing child supernode contribution\n");
+        flag_stop_.store(true, std::memory_order_relaxed);
+        return;
+      }
+    } else {
+      Int child;
+      child_clique = stack_->getChild(child);
+      assert(child == child_sn);
+    }
+
+    // determine size of clique of child
+    const Int child_begin = S_.snStart(child_sn);
+    const Int child_end = S_.snStart(child_sn + 1);
+
+    // number of nodes in child sn
+    const Int child_size = child_end - child_begin;
+
+    // size of clique of child sn
+    const Int nc = S_.ptr(child_sn + 1) - S_.ptr(child_sn) - child_size;
+
+    // ASSEMBLE INTO FRONTAL
+    HIPO_CLOCK_START(2);
+    // go through the columns of the contribution of the child
+    for (Int col = 0; col < nc; ++col) {
+      // relative index of column in the frontal matrix
+      Int j = S_.relindClique(child_sn, col);
+
+      if (j < sn_size) {
+        // assemble into frontal
+
+        // go through the rows of the contribution of the child
+        Int row = col;
+        while (row < nc) {
+          // relative index of the entry in the matrix frontal
+          const Int i = S_.relindClique(child_sn, row);
+
+          // how many entries to sum
+          Int consecutive = S_.consecutiveSums(child_sn, row);
+
+          FH->assembleFrontalMultiple(consecutive, child_clique, nc, child_sn,
+                                      row, col, i, j);
+
+          row += consecutive;
+        }
+      } else
+        break;
+    }
+    HIPO_CLOCK_STOP(2, data_, kTimeFactoriseAssembleChildrenFrontal);
+
+    // ASSEMBLE INTO CLIQUE
+    HIPO_CLOCK_START(2);
+    FH->assembleClique(child_clique, nc, child_sn);
+    HIPO_CLOCK_STOP(2, data_, kTimeFactoriseAssembleChildrenClique);
+
+    // Schur contribution of the child is no longer needed
+    if (parallel) {
+      freeVector(schur_contribution_[child_sn]);
+    } else {
+      stack_->popChild();
+    }
+
+    // move on to the next child
+    child_sn = next_child_reverse_[child_sn];
+  }
+
+  if (flag_stop_.load(std::memory_order_relaxed)) return;
+
+  // ===================================================
+  // Partial factorisation
+  // ===================================================
+  HIPO_CLOCK_START(2);
+  // threshold for regularisation
+  // const double reg_thresh = max_diag_ * kDynamicDiagCoeff;
+  const double reg_thresh = M_norm1_ * kDynamicDiagCoeff;
+
+  if (Int flag = FH->denseFactorise(reg_thresh)) {
+    flag_stop_.store(true, std::memory_order_relaxed);
+
+    if (log_ && flag == kRetInvalidInput)
+      log_->printDevInfo("DenseFact: invalid input\n");
+    else if (log_ && flag == kRetInvalidPivot)
+      log_->printDevInfo("DenseFact: invalid pivot\n");
+
+    return;
+  }
+  HIPO_CLOCK_STOP(2, data_, kTimeFactoriseDenseFact);
+
+  HIPO_CLOCK_START(2);
+  // compute largest elements in factorisation
+  FH->extremeEntries();
+
+  // terminate the format handler
+  FH->terminate(schur_contribution_[sn], total_reg_, swaps_[sn],
+                pivot_2x2_[sn]);
+
+  if (serial) stack_->pushWork(sn);
+
+  HIPO_CLOCK_STOP(2, data_, kTimeFactoriseTerminate);
+}
+
+bool Factorise::run(Numeric& num) {
+  HIPO_CLOCK_CREATE;
+
+  TaskGroupSpecial tg;
+
+  total_reg_.assign(n_, 0.0);
+
+  // allocate space
+  schur_contribution_.resize(S_.sn());
+  swaps_.resize(S_.sn());
+  pivot_2x2_.resize(S_.sn());
+
+  // This should actually allocate only the first time, then sn_columns_ reuses
+  // the memory of previous factorisations.
+  sn_columns_.resize(S_.sn());
+
+  if (S_.parTree()) {
+    Int spawned_roots{};
+    // spawn tasks for root supernodes
+    for (Int sn = 0; sn < S_.sn(); ++sn) {
+      if (S_.snParent(sn) == -1) {
+        tg.spawn([=]() { processSupernode(sn); });
+        ++spawned_roots;
+      }
+    }
+
+    // sync tasks for root supernodes
+    tg.taskWait();
+  } else {
+    // processing the tree in serial requires a CliqueStack
+    if (!stack_) return true;
+    if (stack_->empty()) stack_->init(S_.maxStackSize());
+
+    // go through each supernode serially
+    for (Int sn = 0; sn < S_.sn(); ++sn) {
+      processSupernode(sn);
+    }
+  }
+
+  if (flag_stop_.load(std::memory_order_relaxed)) return true;
+
+  // move factorisation to numerical object
+  num.S_ = &S_;
+  num.sn_columns_ = &sn_columns_;
+  num.total_reg_ = std::move(total_reg_);
+  num.swaps_ = std::move(swaps_);
+  num.pivot_2x2_ = std::move(pivot_2x2_);
+  num.data_ = &data_;
+
+  HIPO_CLOCK_STOP(1, data_, kTimeFactorise);
+
+  return false;
+}
+
+}  // namespace hipo

data/ext/lpsolver-highs/highs/ipm/hipo/factorhighs/Factorise.h

@@ -0,0 +1,85 @@
+#ifndef FACTORHIGHS_FACTORISE_H
+#define FACTORHIGHS_FACTORISE_H
+
+#include <cmath>
+
+#include "CliqueStack.h"
+#include "Numeric.h"
+#include "Symbolic.h"
+#include "ipm/hipo/auxiliary/IntConfig.h"
+#include "ipm/hipo/auxiliary/Log.h"
+
+namespace hipo {
+
+class Factorise {
+ public:
+  // matrix to factorise
+  std::vector<Int> rowsM_{};
+  std::vector<Int> ptrM_{};
+  std::vector<double> valM_{};
+  Int n_{};
+  Int nzM_{};
+
+  // symbolic factorisation
+  const Symbolic& S_;
+
+  // children in supernodal elimination tree
+  std::vector<Int> first_child_{};
+  std::vector<Int> next_child_{};
+
+  // reverse linked lists of chidlren
+  std::vector<Int> first_child_reverse_{};
+  std::vector<Int> next_child_reverse_{};
+
+  // generated elements, aka Schur complements.
+  std::vector<std::vector<double>> schur_contribution_{};
+
+  // columns of L, stored as dense supernodes
+  // This memory is managed outside of Factorise, so that it can be reused for
+  // all ipm iterations.
+  std::vector<std::vector<double>>& sn_columns_;
+
+  // swaps of columns for each supernode, ordered locally within a block
+  std::vector<std::vector<Int>> swaps_{};
+
+  // Information about 2x2 pivots.
+  // If pivot_2x2[sn][i] == 0, 1x1 pivot was used.
+  // If pivot_2x2[sn][i] != 0, 2x2 pivot was used and pivot_2x2[sn][i] stores
+  //  the off-diagonal pivot entry (of the 2x2 inverse).
+  std::vector<std::vector<double>> pivot_2x2_{};
+
+  // largest diagonal element in the original matrix and norms of columns
+  double max_diag_{};
+  double min_diag_{};
+  double M_norm1_{};
+
+  // regularisation
+  std::vector<double> total_reg_{};
+
+  // values for static regularisation
+  const Regul& regul_;
+
+  // flag to stop computation
+  std::atomic<bool> flag_stop_{false};
+
+  const Log* log_;
+  DataCollector& data_;
+
+  CliqueStack* stack_;
+
+ public:
+  void permute(const std::vector<Int>& iperm);
+  void processSupernode(Int sn);
+
+ public:
+  Factorise(const Symbolic& S, const std::vector<Int>& rowsM,
+            const std::vector<Int>& ptrM, const std::vector<double>& valM,
+            const Regul& regul, const Log* log, DataCollector& data,
+            std::vector<std::vector<double>>& sn_columns, CliqueStack* stack);
+
+  bool run(Numeric& num);
+};
+
+}  // namespace hipo
+
+#endif