Add constraints and warm start dual simplex from current basis

cpp/include/cuopt/linear_programming/constants.h

@@ -57,6 +57,7 @@
 #define CUOPT_MIP_HEURISTICS_ONLY         "mip_heuristics_only"
 #define CUOPT_MIP_SCALING                 "mip_scaling"
 #define CUOPT_MIP_PRESOLVE                "mip_presolve"
+#define CUOPT_MIP_CUT_PASSES              "mip_cut_passes"
 #define CUOPT_SOLUTION_FILE               "solution_file"
 #define CUOPT_NUM_CPU_THREADS             "num_cpu_threads"
 #define CUOPT_NUM_GPUS                    "num_gpus"

cpp/include/cuopt/linear_programming/mip/solver_settings.hpp

@@ -81,6 +81,7 @@ class mip_solver_settings_t {
   f_t time_limit       = std::numeric_limits<f_t>::infinity();
   bool heuristics_only = false;
   i_t num_cpu_threads  = -1;  // -1 means use default number of threads in branch and bound
+  i_t max_cut_passes   = 10;  // number of cut passes to make
   i_t num_gpus         = 1;
   bool log_to_console  = true;
   std::string log_file;

cpp/src/dual_simplex/CMakeLists.txt

@@ -10,6 +10,7 @@ set(DUAL_SIMPLEX_SRC_FILES
   ${CMAKE_CURRENT_SOURCE_DIR}/basis_updates.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/bound_flipping_ratio_test.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/branch_and_bound.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/cuts.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/crossover.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/folding.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/initial_basis.cpp
@@ -34,7 +35,7 @@ set(DUAL_SIMPLEX_SRC_FILES
   )
 
 # Uncomment to enable debug info
-#set_source_files_properties(${DUAL_SIMPLEX_SRC_FILES} DIRECTORY ${CMAKE_SOURCE_DIR} PROPERTIES COMPILE_OPTIONS "-g1")
+set_source_files_properties(${DUAL_SIMPLEX_SRC_FILES} DIRECTORY ${CMAKE_SOURCE_DIR} PROPERTIES COMPILE_OPTIONS "-g1")
 
 set(CUOPT_SRC_FILES ${CUOPT_SRC_FILES}
   ${DUAL_SIMPLEX_SRC_FILES} PARENT_SCOPE)

cpp/src/dual_simplex/basis_updates.cpp

@@ -1108,6 +1108,212 @@ i_t basis_update_t<i_t, f_t>::lower_triangular_multiply(const csc_matrix_t<i_t,
   return new_nz;
 }
 
+// Start of middle product form: basis_update_mpf_t
+
+template <typename i_t, typename f_t>
+i_t basis_update_mpf_t<i_t, f_t>::append_cuts(const csr_matrix_t<i_t, f_t>& cuts_basic)
+{
+  const i_t m = L0_.m;
+
+  // Solve for U^T W^T = C_B^T
+  // We do this one row at a time of C_B
+  csc_matrix_t<i_t, f_t> WT(m, cuts_basic.m, 0);
+
+  i_t WT_nz = 0;
+  for (i_t k = 0; k < cuts_basic.m; k++) {
+    sparse_vector_t<i_t, f_t> rhs(cuts_basic, k);
+    u_transpose_solve(rhs);
+    WT.col_start[k] = WT_nz;
+    for (i_t q = 0; q < rhs.i.size(); q++) {
+      WT.i.push_back(rhs.i[q]);
+      WT.x.push_back(rhs.x[q]);
+      WT_nz++;
+    }
+  }
+  WT.col_start[cuts_basic.m] = WT_nz;
+
+
+#ifdef CHECK_W
+  {
+    for (i_t k = 0; k < cuts_basic.m; k++) {
+      std::vector<f_t> WT_col(m, 0.0);
+      WT.load_a_column(k, WT_col);
+      std::vector<f_t> CBT_col(m, 0.0);
+      matrix_transpose_vector_multiply(U0_, 1.0, WT_col, 0.0, CBT_col);
+      sparse_vector_t<i_t, f_t> CBT_col_sparse(cuts_basic, k);
+      std::vector<f_t> CBT_col_dense(m);
+      CBT_col_sparse.to_dense(CBT_col_dense);
+      for (i_t h = 0; h < m; h++) {
+        if (std::abs(CBT_col_dense[h] - CBT_col[h]) > 1e-6) {
+          printf("col %d CBT_col_dense[%d] = %e CBT_col[%d] = %e\n", k, h, CBT_col_dense[h], h, CBT_col[h]);
+          exit(1);
+        }
+      }
+    }
+  }
+#endif
+
+  csc_matrix_t<i_t, f_t> V(cuts_basic.m, m, 0);
+  if (num_updates_ > 0) {
+    // W = V T_0 ... T_{num_updates_ - 1}
+    // or V = W T_{num_updates_ - 1}^{-1} ... T_0^{-1}
+    // or V^T = T_0^{-T} ... T_{num_updates_ - 1}^{-T} W^T
+    // We can compute V^T column by column so that we have
+    // V^T(:, h) = T_0^{-T} ... T_{num_updates_ - 1}^{-T} W^T(:, h)
+    // or
+    // V(h, :) = T_0^{-T} ... T_{num_updates_ - 1}^{-T} W^T(:, h)
+    // So we can form V row by row in CSR and then covert it to CSC
+    // for appending to L0
+
+    csr_matrix_t<i_t, f_t> V_row(cuts_basic.m, m, 0);
+    i_t V_nz           = 0;
+    const f_t zero_tol = 1e-13;
+    for (i_t h = 0; h < cuts_basic.m; h++) {
+      sparse_vector_t rhs(WT, h);
+      scatter_into_workspace(rhs);
+      i_t nz = rhs.i.size();
+      for (i_t k = num_updates_ - 1; k >= 0; --k) {
+        // T_k^{-T} = ( I - v u^T/(1 + u^T v))
+        // T_k^{-T} * b = b - v * (u^T * b) / (1 + u^T * v) = b - theta * v, theta = u^T b / mu
+
+        const i_t u_col = 2 * k;
+        const i_t v_col = 2 * k + 1;
+        const f_t mu    = mu_values_[k];
+
+        // dot = u^T * b
+        f_t dot         = dot_product(u_col, xi_workspace_, x_workspace_);
+        const f_t theta = dot / mu;
+        if (std::abs(theta) > zero_tol) {
+          add_sparse_column(S_, v_col, -theta, xi_workspace_, nz, x_workspace_);
+        }
+      }
+      gather_into_sparse_vector(nz, rhs);
+      V_row.row_start[h] = V_nz;
+      for (i_t q = 0; q < rhs.i.size(); q++) {
+        V_row.j.push_back(rhs.i[q]);
+        V_row.x.push_back(rhs.x[q]);
+        V_nz++;
+      }
+    }
+    V_row.row_start[cuts_basic.m] = V_nz;
+
+    V_row.to_compressed_col(V);
+
+
+#ifdef CHECK_V
+    csc_matrix_t<i_t, f_t> CB_col(cuts_basic.m, m, 0);
+    cuts_basic.to_compressed_col(CB_col);
+    for (i_t k = 0; k < m; k++) {
+      std::vector<f_t> U_col(m, 0.0);
+      U0_.load_a_column(k, U_col);
+      for (i_t h = num_updates_ - 1; h >= 0; --h) {
+        // T_h = ( I + u_h v_h^T)
+        // T_h * x = x + u_h * v_h^T * x = x + theta * u_h
+        const i_t u_col = 2 * h;
+        const i_t v_col = 2 * h + 1;
+        f_t theta = dot_product(v_col, U_col);
+        const i_t col_start = S_.col_start[u_col];
+        const i_t col_end = S_.col_start[u_col + 1];
+        for (i_t p = col_start; p < col_end; ++p) {
+          const i_t i = S_.i[p];
+          U_col[i] += theta * S_.x[p];
+        }
+      }
+      std::vector<f_t> CB_column(cuts_basic.m, 0.0);
+      matrix_vector_multiply(V, 1.0, U_col, 0.0, CB_column);
+      std::vector<f_t> CB_col_dense(cuts_basic.m);
+      CB_col.load_a_column(k, CB_col_dense);
+      for (i_t l = 0; l < cuts_basic.m; l++) {
+        if (std::abs(CB_col_dense[l] - CB_column[l]) > 1e-6) {
+          printf("col %d CB_col_dense[%d] = %e CB_column[%d] = %e\n", k, l, CB_col_dense[l], l, CB_column[l]);
+          exit(1);
+        }
+      }
+    }
+#endif
+  } else {
+    // W = V
+    WT.transpose(V);
+  }
+
+  // Extend u_i, v_i for i = 0, ..., num_updates_ - 1
+  S_.m += cuts_basic.m;
+
+  // Adjust L and U
+  // L = [ L0  0 ]
+  //     [ V   I ]
+
+  i_t V_nz = V.col_start[m];
+  i_t L_nz = L0_.col_start[m];
+  csc_matrix_t<i_t, f_t> new_L(m + cuts_basic.m, m + cuts_basic.m, L_nz + V_nz + cuts_basic.m);
+  i_t predicted_nz = L_nz + V_nz + cuts_basic.m;
+  L_nz = 0;
+  for (i_t j = 0; j < m; ++j) {
+    new_L.col_start[j]  = L_nz;
+    const i_t col_start = L0_.col_start[j];
+    const i_t col_end   = L0_.col_start[j + 1];
+    for (i_t p = col_start; p < col_end; ++p) {
+      new_L.i[L_nz] = L0_.i[p];
+      new_L.x[L_nz] = L0_.x[p];
+      L_nz++;
+    }
+    const i_t V_col_start = V.col_start[j];
+    const i_t V_col_end   = V.col_start[j + 1];
+    for (i_t p = V_col_start; p < V_col_end; ++p) {
+      new_L.i[L_nz] = V.i[p] + m;
+      new_L.x[L_nz] = V.x[p];
+      L_nz++;
+    }
+  }
+  for (i_t j = m; j < m + cuts_basic.m; ++j) {
+    new_L.col_start[j] = L_nz;
+    new_L.i[L_nz]      = j;
+    new_L.x[L_nz]      = 1.0;
+    L_nz++;
+  }
+  new_L.col_start[m + cuts_basic.m] = L_nz;
+  if (L_nz != predicted_nz) {
+    printf("L_nz %d predicted_nz %d\n", L_nz, predicted_nz);
+    exit(1);
+  }
+
+  L0_ = new_L;
+
+  // Adjust U
+  // U = [ U0 0 ]
+  //     [ 0  I ]
+
+  i_t U_nz = U0_.col_start[m];
+  U0_.col_start.resize(m + cuts_basic.m + 1);
+  U0_.i.resize(U_nz + cuts_basic.m);
+  U0_.x.resize(U_nz + cuts_basic.m);
+  for (i_t k = m; k < m + cuts_basic.m; ++k) {
+    U0_.col_start[k] = U_nz;
+    U0_.i[U_nz]      = k;
+    U0_.x[U_nz]      = 1.0;
+    U_nz++;
+  }
+  U0_.col_start[m + cuts_basic.m] = U_nz;
+  U0_.n                           = m + cuts_basic.m;
+  U0_.m                           = m + cuts_basic.m;
+
+  compute_transposes();
+
+  // Adjust row_permutation_ and inverse_row_permutation_
+  row_permutation_.resize(m + cuts_basic.m);
+  inverse_row_permutation_.resize(m + cuts_basic.m);
+  for (i_t k = m; k < m + cuts_basic.m; ++k) {
+    row_permutation_[k] = k;
+  }
+  inverse_permutation(row_permutation_, inverse_row_permutation_);
+
+  // Adjust workspace sizes
+  xi_workspace_.resize(2 * (m + cuts_basic.m), 0);
+  x_workspace_.resize(m + cuts_basic.m, 0.0);
+
+  return 0;
+}
+
 template <typename i_t, typename f_t>
 void basis_update_mpf_t<i_t, f_t>::gather_into_sparse_vector(i_t nz,
                                                              sparse_vector_t<i_t, f_t>& out) const
@@ -2065,7 +2271,7 @@ int basis_update_mpf_t<i_t, f_t>::refactor_basis(
                       q,
                       deficient,
                       slacks_needed) == -1) {
-    settings.log.debug("Initial factorization failed\n");
+    settings.log.printf("Initial factorization failed\n");
     basis_repair(A, settings, deficient, slacks_needed, basic_list, nonbasic_list, vstatus);
 
 #ifdef CHECK_BASIS_REPAIR

cpp/src/dual_simplex/basis_updates.hpp

@@ -291,6 +291,8 @@ class basis_update_mpf_t {
     reset_stats();
   }
 
+  i_t append_cuts(const csr_matrix_t<i_t, f_t>& cuts_basic);
+
   f_t estimate_solution_density(f_t rhs_nz, f_t sum, i_t& num_calls, bool& use_hypersparse) const
   {
     num_calls++;