Require Kokkos 4.1

.github/workflows/windows.yml

@@ -23,7 +23,7 @@ jobs:
     - uses: actions/checkout@v3
       with:
         repository: kokkos/kokkos
-        ref: 4.0.00
+        ref: 4.1.00
         path: ${GITHUB_WORKSPACE}/../kokkos
     - name: Install Kokkos
       shell: bash

.gitlab-ci.yml

@@ -36,7 +36,7 @@ BuildKokkos:
   extends: .LoadModules
   stage: buildDependencies
   script:
-    - git clone --depth=1 --branch 4.0.00 https://github.com/kokkos/kokkos.git &&
+    - git clone --depth=1 --branch 4.1.00 https://github.com/kokkos/kokkos.git &&
       cd kokkos &&
       mkdir build &&
       cd build &&

.jenkins/continuous.groovy

@@ -42,7 +42,7 @@ pipeline {
                         dockerfile {
                             filename "Dockerfile"
                             dir "docker"
-                            additionalBuildArgs '--build-arg BASE=nvidia/cuda:11.5.2-devel-ubuntu20.04 --build-arg KOKKOS_VERSION=4.1.00 --build-arg KOKKOS_OPTIONS="-DCMAKE_CXX_EXTENSIONS=OFF -DKokkos_ENABLE_SERIAL=ON -DKokkos_ENABLE_CUDA=ON -DKokkos_ENABLE_CUDA_LAMBDA=ON -DKokkos_ARCH_VOLTA70=ON" --build-arg CUDA_AWARE_MPI=1'
+                            additionalBuildArgs '--build-arg BASE=nvidia/cuda:11.5.2-devel-ubuntu20.04 --build-arg KOKKOS_VERSION=4.1.00 --build-arg KOKKOS_OPTIONS="-DCMAKE_CXX_EXTENSIONS=OFF -DKokkos_ENABLE_SERIAL=ON -DKokkos_ENABLE_CUDA=ON -DKokkos_ARCH_VOLTA70=ON" --build-arg CUDA_AWARE_MPI=1'
                             args '-v /tmp/ccache:/tmp/ccache --env NVIDIA_VISIBLE_DEVICES=${NVIDIA_VISIBLE_DEVICES}'
                             label 'NVIDIA_Tesla_V100-PCIE-32GB && nvidia-docker'
                         }
@@ -102,7 +102,7 @@ pipeline {
                         dockerfile {
                             filename "Dockerfile"
                             dir "docker"
-                            additionalBuildArgs '--build-arg BASE=nvidia/cuda:11.0.3-devel-ubuntu20.04 --build-arg KOKKOS_OPTIONS="-DCMAKE_CXX_EXTENSIONS=OFF -DKokkos_ENABLE_SERIAL=ON -DKokkos_ENABLE_OPENMP=ON -DKokkos_ENABLE_CUDA=ON -DKokkos_ENABLE_CUDA_LAMBDA=ON -DKokkos_ARCH_VOLTA70=ON"'
+                            additionalBuildArgs '--build-arg BASE=nvidia/cuda:11.0.3-devel-ubuntu20.04 --build-arg KOKKOS_OPTIONS="-DCMAKE_CXX_EXTENSIONS=OFF -DKokkos_ENABLE_SERIAL=ON -DKokkos_ENABLE_OPENMP=ON -DKokkos_ENABLE_CUDA=ON -DKokkos_ARCH_VOLTA70=ON"'
                             args '-v /tmp/ccache:/tmp/ccache --env NVIDIA_VISIBLE_DEVICES=${NVIDIA_VISIBLE_DEVICES}'
                             label 'NVIDIA_Tesla_V100-PCIE-32GB && nvidia-docker'
                         }
@@ -161,7 +161,7 @@ pipeline {
                         dockerfile {
                             filename "Dockerfile"
                             dir "docker"
-                            additionalBuildArgs '--build-arg BASE=nvidia/cuda:11.0.3-devel-ubuntu18.04 --build-arg KOKKOS_VERSION="4.1.00" --build-arg KOKKOS_OPTIONS="-DCMAKE_CXX_EXTENSIONS=OFF -DCMAKE_CXX_COMPILER=clang++ -DKokkos_ENABLE_THREADS=ON -DKokkos_ENABLE_CUDA=ON -DKokkos_ENABLE_CUDA_LAMBDA=ON -DKokkos_ARCH_VOLTA70=ON -DCMAKE_PREFIX_PATH=/usr/lib/x86_64-linux-gnu"'
+                            additionalBuildArgs '--build-arg BASE=nvidia/cuda:11.0.3-devel-ubuntu18.04 --build-arg KOKKOS_VERSION="4.1.00" --build-arg KOKKOS_OPTIONS="-DCMAKE_CXX_EXTENSIONS=OFF -DCMAKE_CXX_COMPILER=clang++ -DKokkos_ENABLE_THREADS=ON -DKokkos_ENABLE_CUDA=ON -DKokkos_ARCH_VOLTA70=ON -DCMAKE_PREFIX_PATH=/usr/lib/x86_64-linux-gnu"'
                             args '-v /tmp/ccache:/tmp/ccache --env NVIDIA_VISIBLE_DEVICES=${NVIDIA_VISIBLE_DEVICES}'
                             label 'NVIDIA_Tesla_V100-PCIE-32GB && nvidia-docker'
                         }

.jenkins/nightly.groovy

@@ -26,7 +26,7 @@ pipeline {
                         dir('source-kokkos') {
                             sh 'git rev-parse --short HEAD'
                         }
-                        sh 'cmake -S source-kokkos -B build-kokkos -D CMAKE_INSTALL_PREFIX=$PWD/install-kokkos $CMAKE_OPTIONS -D Kokkos_ENABLE_CUDA=ON -D Kokkos_ENABLE_CUDA_LAMBDA=ON'
+                        sh 'cmake -S source-kokkos -B build-kokkos -D CMAKE_INSTALL_PREFIX=$PWD/install-kokkos $CMAKE_OPTIONS -D Kokkos_ENABLE_CUDA=ON'
                         sh 'cmake --build build-kokkos --parallel 8'
                         sh 'cmake --install build-kokkos'
                         // Disable tests as Ubuntu 22.04 comes with Boost 1.74 which causes build issues with CUDA
@@ -66,7 +66,7 @@ pipeline {
                         dir('source-kokkos') {
                             sh 'git rev-parse --short HEAD'
                         }
-                        sh 'cmake -S source-kokkos -B build-kokkos -D CMAKE_INSTALL_PREFIX=$PWD/install-kokkos $CMAKE_OPTIONS -D Kokkos_ENABLE_CUDA=ON -D Kokkos_ENABLE_CUDA_LAMBDA=ON'
+                        sh 'cmake -S source-kokkos -B build-kokkos -D CMAKE_INSTALL_PREFIX=$PWD/install-kokkos $CMAKE_OPTIONS -D Kokkos_ENABLE_CUDA=ON'
                         sh 'cmake --build build-kokkos --parallel 8'
                         sh 'cmake --install build-kokkos'
                         // Disable tests as Ubuntu 22.04 comes with Boost 1.74 which causes build issues with CUDA

CMakeLists.txt

@@ -1,12 +1,8 @@
 cmake_minimum_required(VERSION 3.16)
 project(ArborX CXX)
 
-find_package(Kokkos 4.0 REQUIRED CONFIG)
+find_package(Kokkos 4.1 REQUIRED CONFIG)
 message(STATUS "Found Kokkos: ${Kokkos_DIR} (version \"${Kokkos_VERSION}\")")
-if(Kokkos_ENABLE_CUDA AND Kokkos_VERSION VERSION_LESS 4.1)
-  # CUDA_LAMBDA is deprecated starting 4.1
-  kokkos_check(OPTIONS CUDA_LAMBDA)
-endif()
 
 add_library(ArborX INTERFACE)
 target_link_libraries(ArborX INTERFACE Kokkos::kokkos)

docker/Dockerfile

@@ -131,8 +131,8 @@ RUN SCRATCH_DIR=/scratch && mkdir -p ${SCRATCH_DIR} && cd ${SCRATCH_DIR} && \
 ENV LD_LIBRARY_PATH=/usr/local/cuda/targets/x86_64-linux/lib:${LD_LIBRARY_PATH}
 
 # Install Kokkos
-ARG KOKKOS_VERSION=4.0.00
-ARG KOKKOS_OPTIONS="-DKokkos_ENABLE_SERIAL=ON -DKokkos_ENABLE_OPENMP=ON -DKokkos_ENABLE_CUDA=ON -DKokkos_ENABLE_CUDA_LAMBDA=ON"
+ARG KOKKOS_VERSION=4.2.00
+ARG KOKKOS_OPTIONS="-DKokkos_ENABLE_SERIAL=ON -DKokkos_ENABLE_OPENMP=ON -DKokkos_ENABLE_CUDA=ON"
 ENV KOKKOS_DIR=/opt/kokkos
 RUN KOKKOS_URL=https://github.com/kokkos/kokkos/archive/${KOKKOS_VERSION}.tar.gz && \
     KOKKOS_ARCHIVE=kokkos-${KOKKOS_HASH}.tar.gz && \

docker/Dockerfile.hipcc

@@ -89,7 +89,7 @@ RUN SCRATCH_DIR=/scratch && mkdir -p ${SCRATCH_DIR} && cd ${SCRATCH_DIR} && \
     rm -rf ${SCRATCH_DIR}
 
 # Install Kokkos
-ARG KOKKOS_VERSION=4.0.00
+ARG KOKKOS_VERSION=4.2.00
 ARG KOKKOS_ARCH=Kokkos_ARCH_VEGA906
 ARG KOKKOS_OPTIONS="-DKokkos_ENABLE_HIP=ON -DKokkos_ENABLE_OPENMP=ON -D${KOKKOS_ARCH}=ON"
 ENV KOKKOS_DIR=/opt/kokkos

docker/Dockerfile.sycl

@@ -109,7 +109,7 @@ RUN SCRATCH_DIR=/scratch && mkdir -p ${SCRATCH_DIR} && cd ${SCRATCH_DIR} && \
     rm -rf ${SCRATCH_DIR}
 
 # Install Kokkos
-ARG KOKKOS_VERSION=4.1.00
+ARG KOKKOS_VERSION=4.2.00
 ARG KOKKOS_OPTIONS="-DKokkos_ENABLE_SYCL=ON -DCMAKE_CXX_FLAGS=-Wno-unknown-cuda-version -DKokkos_ENABLE_UNSUPPORTED_ARCHS=ON -DKokkos_ENABLE_DEPRECATED_CODE_3=OFF -DKokkos_ARCH_VOLTA70=ON -DCMAKE_CXX_STANDARD=17 -DCMAKE_CXX_FLAGS=-w"
 ENV KOKKOS_DIR=/opt/kokkos
 RUN . /opt/intel/oneapi/setvars.sh --include-intel-llvm && \

docker/README.md

@@ -3,7 +3,7 @@ in the environment file.  Its value will be prepended along with the service
 name to the container on start up.
 
 
-You may use multiple compose files to customize your container.  For instance, you
+You may use multiple compose files to customize your container. For instance, you
 could reproduce the configuration from one of the automated builds by providing
 the following a `docker-compose.override.yml` file:
 ```
@@ -13,6 +13,6 @@ services:
     build:
       args:
         - BASE=nvidia/cuda:11.5.2-devel-ubuntu20.04
-        - KOKKOS_VERSION=4.0.00
-        - KOKKOS_OPTIONS=-DCMAKE_CXX_STANDARD=17 -DKokkos_ENABLE_OPENMP=ON -DKokkos_ENABLE_CUDA=ON -DKokkos_ENABLE_CUDA_LAMBDA=ON -DKokkos_ARCH_SNB=ON -DKokkos_ARCH_VOLTA70=ON
+        - KOKKOS_VERSION=4.2.00
+        - KOKKOS_OPTIONS=-DCMAKE_CXX_STANDARD=17 -DKokkos_ENABLE_OPENMP=ON -DKokkos_ENABLE_CUDA=ON -DKokkos_ARCH_SNB=ON -DKokkos_ARCH_VOLTA70=ON
 ```

src/ArborX_BruteForce.hpp

@@ -18,11 +18,11 @@
 #include <ArborX_DetailsBruteForceImpl.hpp>
 #include <ArborX_DetailsCrsGraphWrapperImpl.hpp>
 #include <ArborX_DetailsKokkosExtAccessibilityTraits.hpp>
-#include <ArborX_DetailsKokkosExtScopedProfileRegion.hpp>
 #include <ArborX_DetailsLegacy.hpp>
 #include <ArborX_IndexableGetter.hpp>
 
 #include <Kokkos_Core.hpp>
+#include <Kokkos_Profiling_ScopedRegion.hpp>
 
 namespace ArborX
 {
@@ -69,7 +69,7 @@ class BasicBruteForce
   query(ExecutionSpace const &space, Predicates const &predicates,
         CallbackOrView &&callback_or_view, View &&view, Args &&...args) const
   {
-    KokkosExt::ScopedProfileRegion guard("ArborX::BruteForce::query_crs");
+    Kokkos::Profiling::ScopedRegion guard("ArborX::BruteForce::query_crs");
 
     Details::CrsGraphWrapperImpl::
         check_valid_callback_if_first_argument_is_not_a_view<value_type>(
@@ -156,7 +156,7 @@ class BruteForce
     }
     else
     {
-      KokkosExt::ScopedProfileRegion guard("ArborX::BruteForce::query_crs");
+      Kokkos::Profiling::ScopedRegion guard("ArborX::BruteForce::query_crs");
 
       Kokkos::View<int *, MemorySpace> indices(
           "ArborX::CrsGraphWrapper::query::indices", 0);
@@ -194,7 +194,7 @@ BasicBruteForce<MemorySpace, Value, IndexableGetter, BoundingVolume>::
                                               ExecutionSpace>::value,
                 "Values must be accessible from the execution space");
 
-  KokkosExt::ScopedProfileRegion guard("ArborX::BruteForce::BruteForce");
+  Kokkos::Profiling::ScopedRegion guard("ArborX::BruteForce::BruteForce");
 
   if (empty())
   {

src/ArborX_HDBSCAN.hpp

@@ -12,9 +12,10 @@
 #define ARBORX_HDBSCAN_HPP
 
 #include <ArborX_Dendrogram.hpp>
-#include <ArborX_DetailsKokkosExtScopedProfileRegion.hpp>
 #include <ArborX_MinimumSpanningTree.hpp>
 
+#include <Kokkos_Profiling_ScopedRegion.hpp>
+
 namespace ArborX::Experimental
 {
 
@@ -24,7 +25,7 @@ auto hdbscan(ExecutionSpace const &exec_space, Primitives const &primitives,
              DendrogramImplementation dendrogram_impl =
                  DendrogramImplementation::BORUVKA)
 {
-  KokkosExt::ScopedProfileRegion guard("ArborX::HDBSCAN");
+  Kokkos::Profiling::ScopedRegion guard("ArborX::HDBSCAN");
 
   using namespace ArborX::Details;
 

src/ArborX_LinearBVH.hpp

@@ -19,7 +19,6 @@
 #include <ArborX_DetailsBatchedQueries.hpp>
 #include <ArborX_DetailsCrsGraphWrapperImpl.hpp>
 #include <ArborX_DetailsKokkosExtAccessibilityTraits.hpp>
-#include <ArborX_DetailsKokkosExtScopedProfileRegion.hpp>
 #include <ArborX_DetailsLegacy.hpp>
 #include <ArborX_DetailsNode.hpp>
 #include <ArborX_DetailsPermutedData.hpp>
@@ -32,6 +31,7 @@
 #include <ArborX_TraversalPolicy.hpp>
 
 #include <Kokkos_Core.hpp>
+#include <Kokkos_Profiling_ScopedRegion.hpp>
 
 namespace ArborX
 {
@@ -93,7 +93,7 @@ class BasicBoundingVolumeHierarchy
   query(ExecutionSpace const &space, Predicates const &predicates,
         CallbackOrView &&callback_or_view, View &&view, Args &&...args) const
   {
-    KokkosExt::ScopedProfileRegion guard("ArborX::BVH::query_crs");
+    Kokkos::Profiling::ScopedRegion guard("ArborX::BVH::query_crs");
 
     Details::CrsGraphWrapperImpl::
         check_valid_callback_if_first_argument_is_not_a_view<value_type>(
@@ -206,7 +206,7 @@ class BoundingVolumeHierarchy
     }
     else
     {
-      KokkosExt::ScopedProfileRegion guard("ArborX::BVH::query_crs");
+      Kokkos::Profiling::ScopedRegion guard("ArborX::BVH::query_crs");
 
       Kokkos::View<int *, MemorySpace> indices(
           "ArborX::CrsGraphWrapper::query::indices", 0);
@@ -266,7 +266,7 @@ BasicBoundingVolumeHierarchy<MemorySpace, Value, IndexableGetter,
 
   Details::check_valid_space_filling_curve<DIM>(curve);
 
-  KokkosExt::ScopedProfileRegion guard("ArborX::BVH::BVH");
+  Kokkos::Profiling::ScopedRegion guard("ArborX::BVH::BVH");
 
   if (empty())
   {

src/details/ArborX_DetailsDistributedTreeImpl.hpp

@@ -17,7 +17,6 @@
 #include <ArborX_DetailsDistributor.hpp>
 #include <ArborX_DetailsHappyTreeFriends.hpp>
 #include <ArborX_DetailsKokkosExtMinMaxOperations.hpp>
-#include <ArborX_DetailsKokkosExtScopedProfileRegion.hpp>
 #include <ArborX_DetailsKokkosExtViewHelpers.hpp>
 #include <ArborX_DetailsPriorityQueue.hpp>
 #include <ArborX_DetailsUtils.hpp>
@@ -28,6 +27,7 @@
 #include <ArborX_Sphere.hpp>
 
 #include <Kokkos_Core.hpp>
+#include <Kokkos_Profiling_ScopedRegion.hpp>
 
 #include <mpi.h>
 
@@ -249,7 +249,7 @@ DistributedTreeImpl<DeviceType>::sendAcrossNetwork(
     ExecutionSpace const &space, Distributor<DeviceType> const &distributor,
     View exports, typename View::non_const_type imports)
 {
-  KokkosExt::ScopedProfileRegion guard(
+  Kokkos::Profiling::ScopedRegion guard(
       "ArborX::DistributedTree::sendAcrossNetwork (" + exports.label() + ")");
 
   ARBORX_ASSERT((exports.extent(0) == distributor.getTotalSendLength()) &&
@@ -319,7 +319,7 @@ void DistributedTreeImpl<DeviceType>::deviseStrategy(
     ExecutionSpace const &space, Predicates const &queries,
     DistributedTree const &tree, Indices &indices, Offset &offset, Distances &)
 {
-  KokkosExt::ScopedProfileRegion guard(
+  Kokkos::Profiling::ScopedRegion guard(
       "ArborX::DistributedTree::deviseStrategy");
 
   auto const &top_tree = tree._top_tree;
@@ -381,7 +381,7 @@ void DistributedTreeImpl<DeviceType>::reassessStrategy(
     DistributedTree const &tree, Indices &indices, Offset &offset,
     Distances &distances)
 {
-  KokkosExt::ScopedProfileRegion guard(
+  Kokkos::Profiling::ScopedRegion guard(
       "ArborX::DistributedTree::reassessStrategy");
 
   auto const &top_tree = tree._top_tree;
@@ -487,7 +487,7 @@ DistributedTreeImpl<DeviceType>::queryDispatchImpl(
     ExecutionSpace const &space, Predicates const &queries, Indices &indices,
     Offset &offset, Ranks &ranks, Distances *distances_ptr)
 {
-  KokkosExt::ScopedProfileRegion guard(
+  Kokkos::Profiling::ScopedRegion guard(
       "ArborX::DistributedTree::query::nearest");
 
   auto const &bottom_tree = tree._bottom_tree;
@@ -590,7 +590,7 @@ DistributedTreeImpl<DeviceType>::queryDispatch(
     ExecutionSpace const &space, Predicates const &queries,
     Callback const &callback, OutputView &out, OffsetView &offset)
 {
-  KokkosExt::ScopedProfileRegion guard(
+  Kokkos::Profiling::ScopedRegion guard(
       "ArborX::DistributedTree::query::spatial");
 
   auto const &top_tree = tree._top_tree;
@@ -701,7 +701,7 @@ void DistributedTreeImpl<DeviceType>::forwardQueries(
     Kokkos::View<Query *, DeviceType> &fwd_queries,
     Kokkos::View<int *, DeviceType> &fwd_ids, Ranks &fwd_ranks)
 {
-  KokkosExt::ScopedProfileRegion guard(
+  Kokkos::Profiling::ScopedRegion guard(
       "ArborX::DistributedTree::forwardQueries");
 
   int comm_rank;
@@ -794,7 +794,7 @@ void DistributedTreeImpl<DeviceType>::communicateResultsBack(
     Kokkos::View<int *, DeviceType> offset, Ranks &ranks,
     Kokkos::View<int *, DeviceType> &ids, Distances *distances_ptr)
 {
-  KokkosExt::ScopedProfileRegion guard(
+  Kokkos::Profiling::ScopedRegion guard(
       "ArborX::DistributedTree::communicateResultsBack");
 
   int comm_rank;
@@ -882,7 +882,7 @@ void DistributedTreeImpl<DeviceType>::filterResults(
     Kokkos::View<float *, DeviceType> distances, Indices &indices,
     Offset &offset, Ranks &ranks)
 {
-  KokkosExt::ScopedProfileRegion guard(
+  Kokkos::Profiling::ScopedRegion guard(
       "ArborX::DistributedTree::filterResults");
 
   using Access = AccessTraits<Predicates, PredicatesTag>;

src/details/ArborX_DetailsMinimumSpanningTree.hpp

@@ -15,15 +15,15 @@
 #include <ArborX_DetailsAlgorithms.hpp>
 #include <ArborX_DetailsHappyTreeFriends.hpp>
 #include <ArborX_DetailsKokkosExtArithmeticTraits.hpp>
-#include <ArborX_DetailsKokkosExtBitManipulation.hpp>
 #include <ArborX_DetailsKokkosExtMinMaxOperations.hpp>
-#include <ArborX_DetailsKokkosExtScopedProfileRegion.hpp>
 #include <ArborX_DetailsKokkosExtSwap.hpp>
 #include <ArborX_DetailsUtils.hpp>
 #include <ArborX_DetailsWeightedEdge.hpp>
 
+#include <Kokkos_BitManipulation.hpp>
 #include <Kokkos_Core.hpp>
 #include <Kokkos_MathematicalFunctions.hpp> // isfinite, signbit
+#include <Kokkos_Profiling_ScopedRegion.hpp>
 
 namespace ArborX::Details
 {
@@ -503,7 +503,7 @@ void updateSidedParents(ExecutionSpace const &space, Labels const &labels,
                         SidedParents &sided_parents, int edges_start,
                         int edges_end)
 {
-  KokkosExt::ScopedProfileRegion guard("ArborX::MST::update_sided_parents");
+  Kokkos::Profiling::ScopedRegion guard("ArborX::MST::update_sided_parents");
 
   // Same as dendrogram alpha's standalone "updateSidedParents"
   Kokkos::parallel_for(
@@ -559,7 +559,7 @@ template <typename ExecutionSpace, typename Edges, typename SidedParents,
 void computeParents(ExecutionSpace const &space, Edges const &edges,
                     SidedParents const &sided_parents, Parents &parents)
 {
-  KokkosExt::ScopedProfileRegion guard("ArborX::MST::compute_edge_parents");
+  Kokkos::Profiling::ScopedRegion guard("ArborX::MST::compute_edge_parents");
 
   using MemorySpace = typename SidedParents::memory_space;
 
@@ -614,7 +614,7 @@ void computeParents(ExecutionSpace const &space, Edges const &edges,
         static_assert(sizeof(int) == sizeof(float));
         KOKKOS_ASSERT(Kokkos::isfinite(edge.weight) &&
                       Kokkos::signbit(edge.weight) == 0);
-        keys(e) = (key << shift) + KokkosExt::bit_cast<int>(edge.weight);
+        keys(e) = (key << shift) + Kokkos::bit_cast<int>(edge.weight);
       });
 
   auto permute = sortObjects(space, keys);

src/details/ArborX_MinimumSpanningTree.hpp

@@ -13,7 +13,6 @@
 #define ARBORX_MINIMUM_SPANNING_TREE_HPP
 
 #include <ArborX_AccessTraits.hpp>
-#include <ArborX_DetailsKokkosExtScopedProfileRegion.hpp>
 #include <ArborX_DetailsKokkosExtViewHelpers.hpp>
 #include <ArborX_DetailsMinimumSpanningTree.hpp>
 #include <ArborX_DetailsMutualReachabilityDistance.hpp>
@@ -22,6 +21,9 @@
 #include <ArborX_DetailsWeightedEdge.hpp>
 #include <ArborX_LinearBVH.hpp>
 
+#include <Kokkos_Core.hpp>
+#include <Kokkos_Profiling_ScopedRegion.hpp>
+
 namespace ArborX::Details
 {
 
@@ -230,7 +232,7 @@ struct MinimumSpanningTree
                              edges_start, edges_end);
         else
         {
-          KokkosExt::ScopedProfileRegion guard(
+          Kokkos::Profiling::ScopedRegion guard(
               "ArborX::MST::compute_vertex_parents");
           assignVertexParents(space, labels, component_out_edges, edges_mapping,
                               bvh, dendrogram_parents);