Fixes for resolving Rob's suggestions

- Releasing GPU resources when not needed.
- Adding VTK_EXEC to TreeNode and SplitProperties constructors

vtkm/cont/BoundingIntervalHierarchy.h

@@ -84,6 +84,8 @@ public:
   BoundingIntervalHierarchy(vtkm::IdComponent numPlanes = 4, vtkm::IdComponent maxLeafSize = 5)
     : NumPlanes(numPlanes)
     , MaxLeafSize(maxLeafSize)
+    , Nodes()
+    , ProcessedCellIds()
   {
   }
 

vtkm/cont/BoundingIntervalHierarchy.hxx

@@ -155,20 +155,27 @@ VTKM_CONT void BoundingIntervalHierarchy::CalculatePlaneSplitCost(
   vtkm::cont::ArrayHandle<vtkm::Id> pointsToRight;
   Algorithms::ReduceByKey(segmentIds, isROfSplitPlane, discardKeys, pointsToRight, vtkm::Add());
 
-  // Calculate Lmax and Rmin
-  vtkm::cont::ArrayHandle<vtkm::Range> leqRanges;
-  vtkm::cont::ArrayHandle<vtkm::Range> rRanges;
-  vtkm::worklet::DispatcherMapField<vtkm::worklet::spatialstructure::FilterRanges<true>>().Invoke(
-    coords, splitPlanes, ranges, leqRanges);
-  vtkm::worklet::DispatcherMapField<vtkm::worklet::spatialstructure::FilterRanges<false>>().Invoke(
-    coords, splitPlanes, ranges, rRanges);
+  isLEQOfSplitPlane.ReleaseResourcesExecution();
+  isROfSplitPlane.ReleaseResourcesExecution();
 
+  // Calculate Lmax and Rmin
   vtkm::cont::ArrayHandle<vtkm::Range> lMaxRanges;
+  {
+    vtkm::cont::ArrayHandle<vtkm::Range> leqRanges;
+    vtkm::worklet::DispatcherMapField<vtkm::worklet::spatialstructure::FilterRanges<true>>().Invoke(
+      coords, splitPlanes, ranges, leqRanges);
+    Algorithms::ReduceByKey(
+      segmentIds, leqRanges, discardKeys, lMaxRanges, vtkm::worklet::spatialstructure::RangeAdd());
+  }
+
   vtkm::cont::ArrayHandle<vtkm::Range> rMinRanges;
-  Algorithms::ReduceByKey(
-    segmentIds, leqRanges, discardKeys, lMaxRanges, vtkm::worklet::spatialstructure::RangeAdd());
-  Algorithms::ReduceByKey(
-    segmentIds, rRanges, discardKeys, rMinRanges, vtkm::worklet::spatialstructure::RangeAdd());
+  {
+    vtkm::cont::ArrayHandle<vtkm::Range> rRanges;
+    vtkm::worklet::DispatcherMapField<vtkm::worklet::spatialstructure::FilterRanges<false>>()
+      .Invoke(coords, splitPlanes, ranges, rRanges);
+    Algorithms::ReduceByKey(
+      segmentIds, rRanges, discardKeys, rMinRanges, vtkm::worklet::spatialstructure::RangeAdd());
+  }
 
   vtkm::cont::ArrayHandle<vtkm::FloatDefault> segmentedSplitPlanes;
   Algorithms::ReduceByKey(
@@ -177,6 +184,7 @@ VTKM_CONT void BoundingIntervalHierarchy::CalculatePlaneSplitCost(
   // Calculate costs
   vtkm::worklet::spatialstructure::SplitPropertiesCalculator splitPropertiesCalculator(
     index, NumPlanes + 1);
+
   vtkm::worklet::DispatcherMapField<vtkm::worklet::spatialstructure::SplitPropertiesCalculator>(
     splitPropertiesCalculator)
     .Invoke(pointsToLeft, pointsToRight, lMaxRanges, rMinRanges, segmentedSplitPlanes, splits);
@@ -313,6 +321,10 @@ public:
         segmentZRanges, zRanges, centerZs, segmentIds, zSplits, DeviceAdapter());
       //PRINT_TIMER("2.2", s22);
 
+      segmentXRanges.ReleaseResourcesExecution();
+      segmentYRanges.ReleaseResourcesExecution();
+      segmentZRanges.ReleaseResourcesExecution();
+
       //START_TIMER(s23);
       // Select best split plane and dimension across X, Y, Z dimension, per segment
       SplitArrayHandle segmentSplits;

vtkm/worklet/spatialstructure/BoundingIntervalHierarchy.h

@@ -51,6 +51,7 @@ struct TreeNode
   vtkm::FloatDefault RMin;
   vtkm::IdComponent Dimension;
 
+  VTKM_EXEC
   TreeNode()
     : LMax()
     , RMin()
@@ -68,6 +69,7 @@ struct SplitProperties
   vtkm::FloatDefault RMin;
   vtkm::FloatDefault Cost;
 
+  VTKM_EXEC
   SplitProperties()
     : Plane()
     , NumLeftPoints()
@@ -138,7 +140,10 @@ public:
 }; // struct LEQWorklet
 
 template <bool LEQ>
-struct FilterRanges : public vtkm::worklet::WorkletMapField
+struct FilterRanges;
+
+template <>
+struct FilterRanges<true> : public vtkm::worklet::WorkletMapField
 {
 public:
   typedef void ControlSignature(FieldIn<>, FieldIn<>, FieldIn<>, FieldOut<>);
@@ -151,14 +156,25 @@ public:
                   const vtkm::Range& cellBounds,
                   vtkm::Range& outBounds) const
   {
-    if (LEQ)
-    {
-      outBounds = (value <= planeValue) ? cellBounds : vtkm::Range();
-    }
-    else
-    {
-      outBounds = (value > planeValue) ? cellBounds : vtkm::Range();
-    }
+    outBounds = (value <= planeValue) ? cellBounds : vtkm::Range();
+  }
+}; // struct FilterRanges
+
+template <>
+struct FilterRanges<false> : public vtkm::worklet::WorkletMapField
+{
+public:
+  typedef void ControlSignature(FieldIn<>, FieldIn<>, FieldIn<>, FieldOut<>);
+  typedef void ExecutionSignature(_1, _2, _3, _4);
+  using InputDomain = _1;
+
+  VTKM_EXEC
+  void operator()(const vtkm::FloatDefault& value,
+                  const vtkm::FloatDefault& planeValue,
+                  const vtkm::Range& cellBounds,
+                  vtkm::Range& outBounds) const
+  {
+    outBounds = (value > planeValue) ? cellBounds : vtkm::Range();
   }
 }; // struct FilterRanges
 

vtkm/worklet/testing/UnitTestBoundingIntervalHierarchy.cxx

@@ -100,7 +100,7 @@ void TestBoundingIntervalHierarchy(vtkm::cont::DataSet dataSet, vtkm::IdComponen
 
   std::cout << "Default stack size " << stackSizeBackup << "\n";
 
-  cudaDeviceSetLimit(cudaLimitStackSize, 1024 * 200);
+  cudaDeviceSetLimit(cudaLimitStackSize, 1024 * 50);
 #endif
 
   vtkm::worklet::DispatcherMapField<BoundingIntervalHierarchyTester>().Invoke(