Remove excluxive_scan(space, view) as non-standard

aprokop · Dec 27, 2023 · aed278b · aed278b
1 parent 287ea35
commit aed278b
Show file tree

Hide file tree

Showing 10 changed files with 15 additions and 23 deletions.
diff --git a/benchmarks/dbscan/dbscan_timpl.hpp b/benchmarks/dbscan/dbscan_timpl.hpp
@@ -120,7 +120,7 @@ void sortAndFilterClusters(ExecutionSpace const &exec_space,
       },
       num_clusters);
   Kokkos::resize(Kokkos::WithoutInitializing, cluster_offset, num_clusters + 1);
-  KokkosExt::exclusive_scan(exec_space, cluster_offset);
+  KokkosExt::exclusive_scan(exec_space, cluster_offset, cluster_offset);
 
   auto cluster_starts = KokkosExt::clone(exec_space, cluster_offset);
   KokkosExt::reallocWithoutInitializing(

diff --git a/src/details/ArborX_DetailsCrsGraphWrapperImpl.hpp b/src/details/ArborX_DetailsCrsGraphWrapperImpl.hpp
@@ -206,7 +206,7 @@ void queryImpl(ExecutionSpace const &space, Tree const &tree,
       "ArborX::CrsGraphWrapper::copy_counts_to_offsets",
       Kokkos::RangePolicy<ExecutionSpace>(space, 0, n_queries),
       KOKKOS_LAMBDA(int const i) { permuted_offset(i) = counts(i); });
-  KokkosExt::exclusive_scan(space, offset);
+  KokkosExt::exclusive_scan(space, offset, offset);
 
   int const n_results = KokkosExt::lastElement(space, offset);
 
@@ -302,7 +302,7 @@ allocateAndInitializeStorage(Tag, ExecutionSpace const &space,
 
   if (buffer_size != 0)
   {
-    KokkosExt::exclusive_scan(space, offset);
+    KokkosExt::exclusive_scan(space, offset, offset);
 
     // Use calculation for the size to avoid calling lastElement(space, offset)
     // as it will launch an extra kernel to copy to host.
@@ -325,7 +325,7 @@ allocateAndInitializeStorage(Tag, ExecutionSpace const &space,
       "scan_queries_for_numbers_of_nearest_neighbors",
       Kokkos::RangePolicy<ExecutionSpace>(space, 0, n_queries),
       KOKKOS_LAMBDA(int i) { offset(i) = getK(predicates(i)); });
-  KokkosExt::exclusive_scan(space, offset);
+  KokkosExt::exclusive_scan(space, offset, offset);
 
   KokkosExt::reallocWithoutInitializing(space, out,
                                         KokkosExt::lastElement(space, offset));

diff --git a/src/details/ArborX_DetailsDistributedTreeImpl.hpp b/src/details/ArborX_DetailsDistributedTreeImpl.hpp
@@ -353,7 +353,7 @@ void DistributedTreeImpl<DeviceType>::deviseStrategy(
         }
       });
 
-  KokkosExt::exclusive_scan(space, new_offset);
+  KokkosExt::exclusive_scan(space, new_offset, new_offset);
 
   // Truncate results so that queries will only be forwarded to as many local
   // trees as necessary to find k neighbors.
@@ -685,7 +685,7 @@ void DistributedTreeImpl<DeviceType>::countResults(
         Kokkos::atomic_increment(&offset(query_ids(i)));
       });
 
-  KokkosExt::exclusive_scan(space, offset);
+  KokkosExt::exclusive_scan(space, offset, offset);
 }
 
 template <typename DeviceType>
@@ -893,7 +893,7 @@ void DistributedTreeImpl<DeviceType>::filterResults(
         new_offset(q) = min(offset(q + 1) - offset(q), getK(queries(q)));
       });
 
-  KokkosExt::exclusive_scan(space, new_offset);
+  KokkosExt::exclusive_scan(space, new_offset, new_offset);
 
   int const n_truncated_results = KokkosExt::lastElement(space, new_offset);
   Kokkos::View<int *, DeviceType> new_indices(

diff --git a/src/details/ArborX_DetailsExpandHalfToFull.hpp b/src/details/ArborX_DetailsExpandHalfToFull.hpp
@@ -41,7 +41,7 @@ void expandHalfToFull(ExecutionSpace const &space, Offsets &offsets,
           Kokkos::atomic_increment(&offsets(k));
         }
       });
-  KokkosExt::exclusive_scan(space, offsets);
+  KokkosExt::exclusive_scan(space, offsets, offsets);
 
   auto const m = KokkosExt::lastElement(space, offsets);
   KokkosExt::reallocWithoutInitializing(space, indices, m);

diff --git a/src/details/ArborX_DetailsTreeTraversal.hpp b/src/details/ArborX_DetailsTreeTraversal.hpp
@@ -158,7 +158,7 @@ struct TreeTraversal<BVH, Predicates, Callback, NearestPredicateTag>
         "scan_queries_for_numbers_of_neighbors",
         Kokkos::RangePolicy<ExecutionSpace>(space, 0, n_queries),
         KOKKOS_CLASS_LAMBDA(int i) { offset(i) = getK(_predicates(i)); });
-    KokkosExt::exclusive_scan(space, offset);
+    KokkosExt::exclusive_scan(space, offset, offset);
     int const buffer_size = KokkosExt::lastElement(space, offset);
     // Allocate buffer over which to perform heap operations in
     // TreeTraversal::nearestQuery() to store nearest leaf nodes found so far.

diff --git a/src/details/ArborX_NeighborList.hpp b/src/details/ArborX_NeighborList.hpp
@@ -57,7 +57,7 @@ void findHalfNeighborList(ExecutionSpace const &space,
       space, bvh,
       KOKKOS_LAMBDA(int, int j) { Kokkos::atomic_increment(&offsets(j)); },
       NeighborListPredicateGetter{radius});
-  KokkosExt::exclusive_scan(space, offsets);
+  KokkosExt::exclusive_scan(space, offsets, offsets);
   KokkosExt::reallocWithoutInitializing(space, indices,
                                         KokkosExt::lastElement(space, offsets));
 
@@ -105,7 +105,7 @@ void findFullNeighborList(ExecutionSpace const &space,
         Kokkos::atomic_increment(&offsets(j));
       },
       NeighborListPredicateGetter{radius});
-  KokkosExt::exclusive_scan(space, offsets);
+  KokkosExt::exclusive_scan(space, offsets, offsets);
   KokkosExt::reallocWithoutInitializing(space, indices,
                                         KokkosExt::lastElement(space, offsets));
 

diff --git a/src/kokkos_ext/ArborX_DetailsKokkosExtStdAlgorithms.hpp b/src/kokkos_ext/ArborX_DetailsKokkosExtStdAlgorithms.hpp
@@ -72,14 +72,6 @@ void exclusive_scan(ExecutionSpace &&space, Kokkos::View<ST, SP...> const &src,
                         ExclusiveScanFunctor<ValueType, DeviceType>(src, dst));
 }
 
-template <typename ExecutionSpace, typename T, typename... P>
-inline std::enable_if_t<
-    Kokkos::is_execution_space<std::remove_reference_t<ExecutionSpace>>::value>
-exclusive_scan(ExecutionSpace &&space, Kokkos::View<T, P...> const &v)
-{
-  exclusive_scan(std::forward<ExecutionSpace>(space), v, v);
-}
-
 template <typename ExecutionSpace, typename ViewType>
 typename ViewType::non_const_value_type
 reduce(ExecutionSpace &&space, ViewType const &v,

diff --git a/test/ArborX_BoostRTreeHelpers.hpp b/test/ArborX_BoostRTreeHelpers.hpp
@@ -201,7 +201,7 @@ performQueries(RTree<Indexable> const &rtree, InputView const &queries)
                             std::back_inserter(returned_values));
   using ExecutionSpace = typename InputView::execution_space;
   ExecutionSpace space;
-  KokkosExt::exclusive_scan(space, offset);
+  KokkosExt::exclusive_scan(space, offset, offset);
   auto const n_results = KokkosExt::lastElement(space, offset);
   OutputView indices("indices", n_results);
   for (int i = 0; i < n_queries; ++i)
@@ -230,7 +230,7 @@ performQueries(ParallelRTree<Indexable> const &rtree, InputView const &queries)
                             std::back_inserter(returned_values));
   using ExecutionSpace = typename InputView::execution_space;
   ExecutionSpace space;
-  KokkosExt::exclusive_scan(space, offset);
+  KokkosExt::exclusive_scan(space, offset, offset);
   auto const n_results = KokkosExt::lastElement(space, offset);
   OutputView1 values("values", n_results);
   for (int i = 0; i < n_queries; ++i)

diff --git a/test/tstDetailsCrsGraphWrapperImpl.cpp b/test/tstDetailsCrsGraphWrapperImpl.cpp
@@ -73,7 +73,7 @@ BOOST_AUTO_TEST_CASE_TEMPLATE(query_impl, DeviceType, ARBORX_DEVICE_TYPES)
   ExecutionSpace space;
   KokkosExt::iota(space, permute);
 
-  KokkosExt::exclusive_scan(space, offset);
+  KokkosExt::exclusive_scan(space, offset, offset);
   Kokkos::realloc(indices, KokkosExt::lastElement(space, offset));
   ArborX::Details::CrsGraphWrapperImpl::queryImpl(
       space, Test1{}, predicates, ArborX::Details::DefaultCallback{}, indices,

diff --git a/test/tstDetailsUtils.cpp b/test/tstDetailsUtils.cpp
@@ -95,7 +95,7 @@ BOOST_AUTO_TEST_CASE_TEMPLATE(exclusive_scan, DeviceType, ARBORX_DEVICE_TYPES)
   v_host(1) = 1.;
   v_host(2) = 0.;
   Kokkos::deep_copy(v, v_host);
-  KokkosExt::exclusive_scan(space, v);
+  KokkosExt::exclusive_scan(space, v, v);
   Kokkos::deep_copy(v_host, v);
   std::vector<double> v_ref = {0., 1., 2.};
   BOOST_TEST(v_host == v_ref, tt::per_element());