fix

trilinos · Sep 26, 2024 · 29fba51 · 29fba51
1 parent 54dcd9a
commit 29fba51
Show file tree

Hide file tree

Showing 5 changed files with 19 additions and 6 deletions.
diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_In_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TET_In_FEMDef.hpp
@@ -77,7 +77,7 @@ getValues(       OutputViewType output,
   }
   case OPERATOR_CURL: {
     const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
-    ptr += card*npts*spaceDim*get_dimension_scalar(work);
+    ptr += card*npts*spaceDim*get_dimension_scalar(input);
     const ViewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
 
     Impl::Basis_HGRAD_TET_Cn_FEM_ORTH::
@@ -593,7 +593,7 @@ Basis_HCURL_TET_In_FEM<DT,OT,PT>::getValues(
     const int numPoints = inputPoints.extent(0);
     using ScalarType = typename ScalarTraits<typename PointViewType::value_type>::scalar_type;
     using WorkViewType = Kokkos::DynRankView< ScalarType,typename DT::execution_space::scratch_memory_space,Kokkos::MemoryTraits<Kokkos::Unmanaged> >;
-    ordinal_type scalarSizePerPoint = (operatorType == OPERATOR_VALUE) ? this->basisCardinality_ : 5*this->basisCardinality_;
+    ordinal_type scalarSizePerPoint = (operatorType == OPERATOR_VALUE) ? this->basisCardinality_ : 7*this->basisCardinality_;
     ordinal_type sizePerPoint = scalarSizePerPoint*get_dimension_scalar(inputPoints);
     WorkViewType workView(scratchStorage, sizePerPoint*team_member.team_size());
     using range_type = Kokkos::pair<ordinal_type,ordinal_type>;

diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_In_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HCURL_TRI_In_FEMDef.hpp
@@ -75,7 +75,7 @@ namespace Intrepid2 {
       }
       case OPERATOR_CURL: {
         const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
-        ptr += card*npts*spaceDim*get_dimension_scalar(work);
+        ptr += card*npts*spaceDim*get_dimension_scalar(input);
         const ViewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
 
         Impl::Basis_HGRAD_TRI_Cn_FEM_ORTH::

diff --git a/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEMDef.hpp b/packages/intrepid2/src/Discretization/Basis/Intrepid2_HGRAD_TRI_Cn_FEMDef.hpp
@@ -106,7 +106,7 @@ getValues(       OutputViewType output,
   }
   case OPERATOR_CURL: { // only works in 2d. first component is -d/dy, second is d/dx
     const ViewType phis(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim);
-    ptr += card*npts*spaceDim*get_dimension_scalar(work);
+    ptr += card*npts*spaceDim*get_dimension_scalar(input);
     const ViewType workView(Kokkos::view_wrap(ptr, vcprop), card, npts, spaceDim+1);
 
 

diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_In_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_HEX_In_FEM/test_02.hpp
@@ -86,13 +86,20 @@ namespace Intrepid2 {
             int scratch_space_level =1;
             const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
             Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+            {
+              // avoid using a team size larger than needed, to reduce allocated scrach space memory
+              ordinal_type team_size = teamPolicy.team_size_recommended(functor, Kokkos::ParallelForTag());
+              *outStream << "Max Recommended team size: " << team_size << ", Requested team size: " << numPoints <<std::endl;
+              team_size = std::min(team_size, numPoints);
+              teamPolicy = Kokkos::TeamPolicy<typename DeviceType::execution_space>(numCells, team_size,vectorSize);
+            }
 
             { //compute values
               auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {
                   auto valsACell = Kokkos::subview(outputValuesA, team_member.league_rank(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
                   basisRawPtr_device->getValues(valsACell, inputPoints, OPERATOR_VALUE, team_member, team_member.team_scratch(scratch_space_level));
-              };              
-              
+              };   
+
               //Get the required size of the scratch space per team and per thread.
               int perThreadSpaceSize(0), perTeamSpaceSize(0);
               basisPtr->getScratchSpaceSize(perTeamSpaceSize,perThreadSpaceSize,inputPoints, OPERATOR_VALUE);

diff --git a/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_In_FEM/test_02.hpp b/packages/intrepid2/unit-test/Discretization/Basis/HCURL_TET_In_FEM/test_02.hpp
@@ -86,6 +86,12 @@ namespace Intrepid2 {
             int scratch_space_level =1;
             const int vectorSize = getVectorSizeForHierarchicalParallelism<PointValueType>();
             Kokkos::TeamPolicy<DeviceSpaceType> teamPolicy(ncells, Kokkos::AUTO,vectorSize);
+            { // avoid using a team size larger than needed, to reduce allocated scrach space memory
+              ordinal_type team_size = teamPolicy.team_size_recommended(functor, Kokkos::ParallelForTag());
+              *outStream << "Max Recommended team size: " << team_size << ", Requested team size: " << numPoints <<std::endl;
+              team_size = std::min(team_size, numPoints);
+              teamPolicy = Kokkos::TeamPolicy<typename DeviceType::execution_space>(numCells, team_size,vectorSize);
+            }
 
             { //compute values
               auto functor = KOKKOS_LAMBDA (typename Kokkos::TeamPolicy<DeviceSpaceType>::member_type team_member) {