Being optimistic for better performance

If the matrix isn't singular (which is true most of the cases) this
PartialPivLU is much faster than FullPivLU.

Let's be optimistic, what can go wrong? ;)

diff --git a/src/kernel/lmm/bmf.cpp b/src/kernel/lmm/bmf.cpp
index 537e869..5f21896 100644 (file)
--- a/src/kernel/lmm/bmf.cpp
+++ b/src/kernel/lmm/bmf.cpp
@@ -184,7 +184,14 @@ Eigen::VectorXd BmfSolver::equilibrium(const allocation_map_t& alloc) const
   XBT_DEBUG("A':\n%s", debug_eigen(A_p).c_str());
 
   XBT_DEBUG("C':\n%s", debug_eigen(C_p).c_str());
-  Eigen::VectorXd rho = Eigen::FullPivLU<Eigen::MatrixXd>(A_p).solve(C_p);
+  /* Being optimist, PartialPivLU is much faster than FullPivLU but requires that the matrix is invertible
+   * FullPivLU however assures that it finds come solution even if the matrix is singular */
+  Eigen::VectorXd rho = Eigen::PartialPivLU<Eigen::MatrixXd>(A_p).solve(C_p);
+  if (rho.array().isNaN().any()) {
+    XBT_DEBUG("rho with nan values, falling back to FullPivLU, rho:\n%s", debug_eigen(rho).c_str());
+    rho = Eigen::FullPivLU<Eigen::MatrixXd>(A_p).solve(C_p);
+  }
+
   for (int p : bounded_players) {
     rho[p] = phi_[p];
   }