Add unit test for graph connectivity with data from the ECL_CC paper

data/data/third_party/ecl_cc/README

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f5e6e3dc559fefc7990daaec071fcd620f620e5ab8652dddaa6b43ca4ba08e7
+size 222

data/data/third_party/ecl_cc/internet.egr

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:892470e152ccd46ddcca70e26bcd88816c247f08c496319cea80864b6b94ce46
+size 3596536

data/data/third_party/ecl_cc/kron_g500-logn21.egr

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1058d22553440f11d3077e70ac8d3a06c6f59107f2dee3fa889fb5933a2a2483
+size 1465043532

vtkm/worklet/connectivities/GraphConnectivity.h

@@ -81,7 +81,7 @@ public:
     auto root_u = findRoot(parents, u);
     auto root_v = findRoot(parents, v);
 
-    // There is a potential concurrent write data race as it is possible for
+    // Case 3. There is a potential concurrent write data race as it is possible for
     // two threads to try to change the same old root to different new roots,
     // e.g. threadA calls parents.Set(root, rootB) while threadB calls
     // parents(root, rootB) where rootB < root and rootC < root (but the order

vtkm/worklet/connectivities/ImageConnectivity.h

@@ -63,11 +63,14 @@ public:
     // Case 1, Two threads calling Unite(u, v) concurrently.
     // Problem: One thread might attach u to v while the other thread attach
     // v to u, causing a cycle in the Union-Find data structure.
-    // Resolution: This is not necessary a data race issue. This is resolved by
-    // "linking by index" as in SV Jayanti et.al. with less than as the total order.
+    // Resolution: This is not so much a race condition as a problem with the
+    // consistency of the algorithm that can also happen in serial. This is
+    // resolved by "linking by index" as in SV Jayanti et.al. with less than
+    // as the total order.
     // The two threads will make the same decision on how to Unite the two tree
-    // (e.g. from root with larger id to root with smaller id.) This avoids cycles in
-    // the resulting graph and maintains the rooted forest structure of Union-Find.
+    // (e.g. from root with larger id to root with smaller id.) This avoids
+    // cycles in the resulting graph and maintains the rooted forest structure
+    // of Union-Find.
 
     // Case 2, T0 calling Unite(u, v), T1 calling Unite(u, w) and T2 calling
     // Unite(v, s) concurrently.
@@ -87,7 +90,7 @@ public:
     auto root_u = findRoot(parents, u);
     auto root_v = findRoot(parents, v);
 
-    // There is a potential concurrent write data race as it is possible for
+    // Case 3. There is a potential concurrent write data race as it is possible for
     // two threads to try to change the same old root to different new roots,
     // e.g. threadA calls parents.Set(root, rootB) while threadB calls
     // parents(root, rootB) where rootB < root and rootC < root (but the order

vtkm/worklet/testing/UnitTestGraphConnectivity.cxx

@@ -12,10 +12,108 @@
 
 #include <vtkm/worklet/connectivities/GraphConnectivity.h>
 
+class AdjacentDifference : public vtkm::worklet::WorkletMapField
+{
+public:
+  using ControlSignature = void(FieldIn index, WholeArrayIn counts, FieldOut outputCount);
+  using ExecutionSignature = void(_1, _2, _3);
+  using InputDomain = _1;
+
+  template <typename WholeArrayType>
+  VTKM_EXEC void operator()(const vtkm::Id& index,
+                            const WholeArrayType& counts,
+                            int& difference) const
+  {
+    difference = counts.Get(index + 1) - counts.Get(index);
+  }
+};
+
+class SameComponent : public vtkm::worklet::WorkletMapField
+{
+public:
+  using ControlSignature = void(FieldIn start,
+                                FieldIn degree,
+                                WholeArrayIn conns,
+                                WholeArrayIn comps,
+                                AtomicArrayInOut same);
+  using ExecutionSignature = void(WorkIndex, _1, _2, _3, _4, _5);
+
+  template <typename Conn, typename Comp, typename AtomicSame>
+  VTKM_EXEC void operator()(vtkm::Id index,
+                            int start,
+                            int degree,
+                            const Conn& conns,
+                            const Comp& comps,
+                            AtomicSame& same) const
+  {
+    for (vtkm::Id offset = start; offset < start + degree; ++offset)
+    {
+      vtkm::Id neighbor = conns.Get(offset);
+      if (comps.Get(index) != comps.Get(neighbor))
+      {
+        same.Set(0, 0);
+      }
+    }
+  }
+};
+
 class TestGraphConnectivity
 {
 public:
-  void operator()() const
+  void TestECL_CC(const std::string& filename, int ncomps) const
+  {
+    auto pathname =
+      vtkm::cont::testing::Testing::GetTestDataBasePath() + "/third_party/ecl_cc/" + filename;
+    std::ifstream stream(pathname, std::ios_base::in | std::ios_base::binary);
+
+    int nnodes;
+    stream.read(reinterpret_cast<char*>(&nnodes), sizeof(nnodes));
+
+    int nedges;
+    stream.read(reinterpret_cast<char*>(&nedges), sizeof(nedges));
+
+    // CSR, there is one more element in offsets thant the actual number of nodes.
+    std::vector<int> offsets(nnodes + 1);
+    std::vector<int> conns(nedges);
+
+    stream.read(reinterpret_cast<char*>(offsets.data()), (nnodes + 1) * sizeof(int));
+    stream.read(reinterpret_cast<char*>(conns.data()), nedges * sizeof(int));
+
+    vtkm::cont::ArrayHandle<int> counts_h;
+    vtkm::cont::Invoker invoke;
+    invoke(AdjacentDifference{},
+           vtkm::cont::make_ArrayHandleCounting(0, 1, nnodes),
+           vtkm::cont::make_ArrayHandle<int>(offsets, vtkm::CopyFlag::On),
+           counts_h);
+
+    offsets.pop_back();
+    vtkm::cont::ArrayHandle<int> offsets_h =
+      vtkm::cont::make_ArrayHandle(offsets, vtkm::CopyFlag::On);
+
+    vtkm::cont::ArrayHandle<int> conns_h = vtkm::cont::make_ArrayHandle(conns, vtkm::CopyFlag::Off);
+
+    vtkm::cont::ArrayHandle<vtkm::Id> comps_h;
+    vtkm::worklet::connectivity::GraphConnectivity().Run(counts_h, offsets_h, conns_h, comps_h);
+
+    VTKM_TEST_ASSERT(vtkm::cont::Algorithm::Reduce(comps_h, 0, vtkm::Maximum{}) == ncomps - 1,
+                     "number of components mismatch");
+
+    vtkm::cont::ArrayHandle<vtkm::UInt32> atomicSame;
+    atomicSame.Allocate(1);
+    atomicSame.WritePortal().Set(0, 1);
+
+    invoke(SameComponent{}, offsets_h, counts_h, conns_h, comps_h, atomicSame);
+    VTKM_TEST_ASSERT(atomicSame.ReadPortal().Get(0) == 1,
+                     "Neighboring nodes don't have the same component id");
+  }
+
+  void TestECL_CC_DataSets() const
+  {
+    TestECL_CC("internet.egr", 1);
+    TestECL_CC("kron_g500-logn21.egr", 553159);
+  }
+
+  void TestSimpleGraph() const
   {
     vtkm::cont::ArrayHandle<vtkm::Id> counts_h =
       vtkm::cont::make_ArrayHandle<vtkm::Id>({ 1, 1, 2, 2, 2 });
@@ -32,6 +130,12 @@ public:
       VTKM_TEST_ASSERT(comps.ReadPortal().Get(i) == 0, "Components has unexpected value.");
     }
   }
+
+  void operator()() const
+  {
+    TestSimpleGraph();
+    TestECL_CC_DataSets();
+  }
 };
 
 int UnitTestGraphConnectivity(int argc, char* argv[])