Merge pull request #458 from pdziekan/master

[skip ci] avg_edge_sclr: sharedmem barrier at start and at the end

Author: pdziekan

libmpdata++/solvers/detail/solver_1d.hpp

@@ -71,8 +71,10 @@ namespace libmpdataxx
 
         virtual void avg_edge_sclr(typename parent_t::arr_t &arr) final
         {
+          this->mem->barrier();
           for (auto &bc : this->bcs[0]) bc->copy_edge_sclr_to_halo1_cyclic(arr);
           for (auto &bc : this->bcs[0]) bc->avg_edge_and_halo1_sclr_cyclic(arr);
+          this->mem->barrier();
         }
 
         real_t courant_number(const arrvec_t<typename parent_t::arr_t> &arrvec) final

libmpdata++/solvers/detail/solver_2d.hpp

@@ -189,12 +189,12 @@ namespace libmpdataxx
                        const idx_t<2> &range_ijk
         ) final
         {
+          this->mem->barrier();
           for (auto &bc : this->bcs[0]) bc->copy_edge_sclr_to_halo1_cyclic(arr, range_ijk[1]);
           for (auto &bc : this->bcs[0]) bc->avg_edge_and_halo1_sclr_cyclic(arr, range_ijk[1]);
 
           for (auto &bc : this->bcs[1]) bc->copy_edge_sclr_to_halo1_cyclic(arr, range_ijk[0]);
           for (auto &bc : this->bcs[1]) bc->avg_edge_and_halo1_sclr_cyclic(arr, range_ijk[0]);
-
           this->mem->barrier();
         }
 

libmpdata++/solvers/detail/solver_3d.hpp

@@ -232,6 +232,7 @@ namespace libmpdataxx
                        const idx_t<3> &range_ijk
         ) final
         {
+          this->mem->barrier();
           for (auto &bc : this->bcs[0]) bc->copy_edge_sclr_to_halo1_cyclic(arr, range_ijk[1], range_ijk[2]);
           for (auto &bc : this->bcs[0]) bc->avg_edge_and_halo1_sclr_cyclic(arr, range_ijk[1], range_ijk[2]);
 
@@ -240,6 +241,7 @@ namespace libmpdataxx
 
           for (auto &bc : this->bcs[2]) bc->copy_edge_sclr_to_halo1_cyclic(arr, range_ijk[0], range_ijk[1]);
           for (auto &bc : this->bcs[2]) bc->avg_edge_and_halo1_sclr_cyclic(arr, range_ijk[0], range_ijk[1]);
+          this->mem->barrier();
         }
 
         void hook_ante_loop(const typename parent_t::advance_arg_t nt) // TODO: this nt conflicts in fact with multiple-advance()-call logic!