Panzer: update extreme value response for cuda

trilinos · Apr 16, 2024 · 8bbfbdc · 8bbfbdc
1 parent af5c146
commit 8bbfbdc
Show file tree

Hide file tree

Showing 2 changed files with 47 additions and 21 deletions.
diff --git a/packages/panzer/disc-fe/src/evaluators/Panzer_CellExtreme_impl.hpp b/packages/panzer/disc-fe/src/evaluators/Panzer_CellExtreme_impl.hpp
@@ -118,26 +118,50 @@ void
 CellExtreme<EvalT, Traits>::
 evaluateFields(
   typename Traits::EvalData workset)
-{ 
-  for (index_t cell = 0; cell < workset.num_cells; ++cell) {
-
-    for (std::size_t qp = 0; qp < num_qp; ++qp) {
-      ScalarT current= multiplier * scalar(cell,qp);
-      for (typename std::vector<PHX::MDField<const ScalarT,Cell,IP> >::iterator field = field_multipliers.begin();
-	   field != field_multipliers.end(); ++field)
-        current *= (*field)(cell,qp);  
-
-      // take first quad point value
-      if(qp==0)
-        extreme(cell) = current;
-
-      // take largest value in the cell
-      if(use_max)
-        extreme(cell) = extreme(cell)<current ? current : extreme(cell);
-      else // use_min
-        extreme(cell) = extreme(cell)>current ? current : extreme(cell);
+{
+    double mult = this->multiplier;
+    std::size_t num_pt = this->num_qp;
+    bool extreme_max = this->use_max;
+    auto scalar_view = scalar.get_view();
+    auto extreme_view = extreme.get_view();
+
+    // compute field weighted with multipliers
+    PHX::View<ScalarT**> mult_field("Multiplied Field", workset.num_cells, scalar.extent(1));
+
+    // initialize to scalar value
+    Kokkos::parallel_for("Initialize Muliplier Field", workset.num_cells, KOKKOS_LAMBDA( const int cell) {
+      for (std::size_t qp = 0; qp < num_pt; ++qp) {
+        mult_field(cell, qp) = mult * scalar_view(cell, qp);
+      }
+    });
+
+    // multiply by field values
+    for (std::size_t field_num = 0; field_num < field_multipliers.size(); ++field_num)
+    {
+      auto field = field_multipliers[field_num];
+      Kokkos::parallel_for("CellExtreme: Multiply Fields", workset.num_cells, KOKKOS_LAMBDA( const int cell) {
+        for (std::size_t qp = 0; qp < num_pt; ++qp) {
+          mult_field(cell, qp) *= field(cell, qp);
+        }
+      });
     }
-  }
+
+    // take extreme over points in each cell
+    Kokkos::parallel_for ("CellExtreme", workset.num_cells, KOKKOS_LAMBDA( const int cell) {
+      for (std::size_t qp = 0; qp < num_pt; ++qp) {
+        auto current = mult_field(cell, qp);
+
+        // take first point
+        if (qp == 0)
+          extreme_view(cell) = current;
+
+        // take largest value in the cell
+        if(extreme_max)
+          extreme_view(cell) = extreme_view(cell)<current ? current : extreme_view(cell);
+        else // use_min
+          extreme_view(cell) = extreme_view(cell)>current ? current : extreme_view(cell);
+      }
+    });
 }
 
 //**********************************************************************

diff --git a/packages/panzer/disc-fe/src/responses/Panzer_ResponseScatterEvaluator_ExtremeValue_impl.hpp b/packages/panzer/disc-fe/src/responses/Panzer_ResponseScatterEvaluator_ExtremeValue_impl.hpp
@@ -137,11 +137,13 @@ template<typename EvalT, typename Traits>
 void ResponseScatterEvaluator_ExtremeValue<EvalT,Traits>::
 evaluateFields(typename Traits::EvalData d)
 {
+  auto hostField = Kokkos::create_mirror_view(cellExtremeValue_.get_view());
+  Kokkos::deep_copy(hostField, cellExtremeValue_.get_view());
   for(index_t i=0;i<d.num_cells;i++) {
     if(useMax_)
-      responseObj_->value = (responseObj_->value < cellExtremeValue_(i)) ? cellExtremeValue_(i) : responseObj_->value;
+      responseObj_->value = (responseObj_->value < hostField(i)) ? hostField(i) : responseObj_->value;
     else
-      responseObj_->value = (responseObj_->value > cellExtremeValue_(i)) ? cellExtremeValue_(i) : responseObj_->value;
+      responseObj_->value = (responseObj_->value > hostField(i)) ? hostField(i) : responseObj_->value;
   }
 }