Fix a performance issue in mpi::schur_pressure_correction

ddemidov · Aug 6, 2018 · 4a2fb52 · 4a2fb52
1 parent 386cac0
commit 4a2fb52
Show file tree

Hide file tree

Showing 2 changed files with 9 additions and 3 deletions.
diff --git a/amgcl/mpi/distributed_matrix.hpp b/amgcl/mpi/distributed_matrix.hpp
@@ -437,6 +437,7 @@ class distributed_matrix {
         }
 
         void move_to_backend(const backend_params &bprm = backend_params()) {
+            AMGCL_TIC("move to backend");
             if (!A_loc) {
                 A_loc = Backend::copy_matrix(a_loc, bprm);
             }
@@ -450,6 +451,7 @@ class distributed_matrix {
 
             a_loc.reset();
             a_rem.reset();
+            AMGCL_TOC("move to backend");
         }
 
         template <class A, class VecX, class B, class VecY>

diff --git a/amgcl/mpi/schur_pressure_correction.hpp b/amgcl/mpi/schur_pressure_correction.hpp
@@ -407,18 +407,22 @@ class schur_pressure_correction {
             tmp = backend_type::create_vector(nu, bprm);
 
             if (prm.approx_schur) {
-                M = backend_type::create_vector(nu, bprm);
-
+                AMGCL_TIC("approx_schur");
+                auto m = std::make_shared<backend::numa_vector<value_type>>(nu, false);
 #pragma omp parallel
                 for(ptrdiff_t i = 0; i < nu; ++i) {
                     value_type v = math::zero<value_type>();
                     for(ptrdiff_t j = Kuu_loc->ptr[i], e = Kuu_loc->ptr[i+1]; j < e; ++j) {
                         if (Kuu_loc->col[j] == i) {
                             v = math::inverse(Kuu_loc->val[j]);
+                            break;
                         }
                     }
-                    (*M)[i] = v;
+                    (*m)[i] = v;
                 }
+
+                M = Backend::copy_vector(m, bprm);
+                AMGCL_TOC("approx_schur");
             }
             AMGCL_TOC("other");