Update Kokkos version to 4.4.1 (#1191)

* bump Kokkos version * changelog * Update Kokkos version to 4.4.1 * Fix View of View dealloc * Move view_alloc to host * Temp disable rocthrust on amd ci container * Fix doc creation. Rollback to ubuntu22-04 * Fix more view of views * Chasing more ViewOfView on host * Fix ViewAlloc func and add doc * Disable Ascent in testing * Revert "Disable Ascent in testing" This reverts commit ebbcc17. * Disable CUDA IPC * Fix more view of views * Fix missig update of private var * Make linter happy * Add ParArray#DRaw and hide Kokkos interface more * Add doc * Debug build in check compilers without symbols * Only check compilers prior to merge * Fix typo * Bump rocm image and change arch for debug hip build * Fix comment that broke the command line --------- Co-authored-by: Jonah Miller <[email protected]> Co-authored-by: Patrick Mullen <[email protected]> Co-authored-by: Philipp Grete <[email protected]>
parthenon-hpc-lab · Nov 22, 2024 · b559452 · b559452
1 parent cad5ccd
commit b559452
Show file tree

Hide file tree

Showing 22 changed files with 199 additions and 95 deletions.
diff --git a/.github/workflows/check-compilers.yml b/.github/workflows/check-compilers.yml
@@ -1,6 +1,14 @@
 name: Check compilers
 
-on: [push, pull_request]
+on:
+  # run every day at 06:00 UTC
+  schedule:
+    - cron: '0 6 * * *'
+  # when triggered manually
+  workflow_dispatch:
+  # when auto merge is enabled (hack to make sure it's run before merging)
+  pull_request:
+    types: [auto_merge_enabled]
 
 # Cancel "duplicated" workflows triggered by pushes to internal
 # branches with associated PRs.
@@ -14,7 +22,7 @@ jobs:
     strategy:
       matrix:
         cxx: ['g++', 'clang++-15']
-        cmake_build_type: ['Release', 'Debug']
+        cmake_build_type: ['Release', 'DbgNoSym']
         device: ['cuda', 'host']
         parallel: ['serial', 'mpi']
         exclude:
@@ -23,7 +31,7 @@ jobs:
           # https://github.com/lanl/parthenon/issues/630
           - cxx: clang++-15
             device: cuda
-            cmake_build_type: Debug
+            cmake_build_type: DbgNoSym
     runs-on: ubuntu-latest
     container:
       image: ghcr.io/parthenon-hpc-lab/cuda11.6-mpi-hdf5-ascent
@@ -48,25 +56,29 @@ jobs:
     strategy:
       matrix:
         cxx: ['hipcc']
-        cmake_build_type: ['Release', 'Debug']
+        cmake_build_type: ['Release', 'DbgNoSym']
         device: ['hip']
         parallel: ['serial', 'mpi']
     runs-on: ubuntu-latest
     container:
-      image: ghcr.io/parthenon-hpc-lab/rocm5.4.3-mpi-hdf5
+      image: ghcr.io/parthenon-hpc-lab/rocm6.2-mpi-hdf5
     env:
       CMAKE_GENERATOR: Ninja
     steps:
       - uses: actions/checkout@v2
         with:
           submodules: 'true'
       - name: CMake
+        # Manually chaning the arch for this (debug) build as the
+        # -O0 option causes compiler issue for the navi 1030 GPU at
+        # compile time, see https://github.com/parthenon-hpc-lab/parthenon/pull/1191#issuecomment-2492035364
         run: |
           cmake -B builddir \
             -DCMAKE_CXX_COMPILER=${{ matrix.cxx }} \
             -DCMAKE_BUILD_TYPE=${{ matrix.cmake_build_type }} \
             -DMACHINE_CFG=${PWD}/cmake/machinecfg/GitHubActions.cmake \
-            -DMACHINE_VARIANT=${{ matrix.device }}_${{ matrix.parallel }}
+            -DMACHINE_VARIANT=${{ matrix.device }}_${{ matrix.parallel }} \
+            -DKokkos_ARCH_AMD_GFX90A=ON -DKokkos_ARCH_NAVI1030=OFF
       - name: Build
         run: |
           cmake --build builddir --parallel 2
diff --git a/.github/workflows/ci-extended.yml b/.github/workflows/ci-extended.yml
@@ -21,6 +21,8 @@ env:
   CMAKE_BUILD_PARALLEL_LEVEL: 5 # num threads for build
   MACHINE_CFG: cmake/machinecfg/CI.cmake
   OMPI_MCA_mpi_common_cuda_event_max: 1000
+  # CUDA IPC within docker repeated seem to cause issue on the CI machine
+  OMPI_MCA_btl_smcuda_use_cuda_ipc: 0
   # https://github.com/open-mpi/ompi/issues/4948#issuecomment-395468231
   OMPI_MCA_btl_vader_single_copy_mechanism: none
 
@@ -34,7 +36,7 @@ jobs:
     container:
       image: ghcr.io/parthenon-hpc-lab/cuda11.6-mpi-hdf5-ascent
       # map to local user id on CI  machine to allow writing to build cache
-      options: --user 1001
+      options: --user 1001 --cap-add CAP_SYS_PTRACE --shm-size="8g" --ulimit memlock=134217728
     steps:
       - uses: actions/checkout@v3
         with:

diff --git a/.github/workflows/ci-short.yml b/.github/workflows/ci-short.yml
@@ -13,6 +13,8 @@ env:
   CMAKE_BUILD_PARALLEL_LEVEL: 5 # num threads for build
   MACHINE_CFG: cmake/machinecfg/CI.cmake
   OMPI_MCA_mpi_common_cuda_event_max: 1000
+  # CUDA IPC within docker repeated seem to cause issue on the CI machine
+  OMPI_MCA_btl_smcuda_use_cuda_ipc: 0
   # https://github.com/open-mpi/ompi/issues/4948#issuecomment-395468231
   OMPI_MCA_btl_vader_single_copy_mechanism: none
 
@@ -22,7 +24,7 @@ jobs:
     container:
       image: ghcr.io/parthenon-hpc-lab/cuda11.6-mpi-hdf5-ascent
       # map to local user id on CI  machine to allow writing to build cache
-      options: --user 1001
+      options: --user 1001 --cap-add CAP_SYS_PTRACE --shm-size="8g" --ulimit memlock=134217728
     steps:
       - uses: actions/checkout@v3
         with:
@@ -47,7 +49,7 @@ jobs:
     container:
       image: ghcr.io/parthenon-hpc-lab/cuda11.6-mpi-hdf5-ascent
       # map to local user id on CI  machine to allow writing to build cache
-      options: --user 1001
+      options: --user 1001 --cap-add CAP_SYS_PTRACE --shm-size="8g" --ulimit memlock=134217728
     steps:
       - uses: actions/checkout@v3
         with:
@@ -79,7 +81,7 @@ jobs:
     container:
       image: ghcr.io/parthenon-hpc-lab/cuda11.6-mpi-hdf5-ascent
       # map to local user id on CI  machine to allow writing to build cache
-      options: --user 1001
+      options: --user 1001 --cap-add CAP_SYS_PTRACE --shm-size="8g" --ulimit memlock=134217728
     steps:
       - uses: actions/checkout@v3
         with:

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -4,20 +4,20 @@
 
 ### Added (new features/APIs/variables/...)
 - [[PR 1103]](https://github.com/parthenon-hpc-lab/parthenon/pull/1103) Add sparsity to vector wave equation test
-- [[PR 1185]](https://github.com/parthenon-hpc-lab/parthenon/pull/1185/files) Bugfix to particle defragmentation
+- [[PR 1185]](https://github.com/parthenon-hpc-lab/parthenon/pull/1185) Bugfix to particle defragmentation
 - [[PR 1184]](https://github.com/parthenon-hpc-lab/parthenon/pull/1184) Fix swarm block neighbor indexing in 1D, 2D
 - [[PR 1183]](https://github.com/parthenon-hpc-lab/parthenon/pull/1183) Fix particle leapfrog example initialization data
 - [[PR 1179]](https://github.com/parthenon-hpc-lab/parthenon/pull/1179) Make a global variable for whether simulation is a restart
 - [[PR 1171]](https://github.com/parthenon-hpc-lab/parthenon/pull/1171) Add PARTHENON_USE_SYSTEM_PACKAGES build option
 - [[PR 1161]](https://github.com/parthenon-hpc-lab/parthenon/pull/1161) Make flux field Metadata accessible, add Metadata::CellMemAligned flag, small perfomance upgrades
 
 ### Changed (changing behavior/API/variables/...)
+- [[PR 1191]](https://github.com/parthenon-hpc-lab/parthenon/pull/1191) Update Kokkos version to 4.4.1
 - [[PR 1209]](https://github.com/parthenon-hpc-lab/parthenon/pull/1209) Ordered history output
 - [[PR 1206]](https://github.com/parthenon-hpc-lab/parthenon/pull/1206) Leapfrog fix
-- [[PR1203]](https://github.com/parthenon-hpc-lab/parthenon/pull/1203) Pin Ubuntu CI image
-- [[PR1177]](https://github.com/parthenon-hpc-lab/parthenon/pull/1177) Make mesh-level boundary conditions usable without the "user" flag
+- [[PR 1203]](https://github.com/parthenon-hpc-lab/parthenon/pull/1203) Pin Ubuntu CI image
+- [[PR 1177]](https://github.com/parthenon-hpc-lab/parthenon/pull/1177) Make mesh-level boundary conditions usable without the "user" flag
 - [[PR 1187]](https://github.com/parthenon-hpc-lab/parthenon/pull/1187) Make DataCollection::Add safer and generalize MeshBlockData::Initialize
-- [[Issue 1165]](https://github.com/parthenon-hpc-lab/parthenon/issues/1165) Bump Kokkos submodule to 4.4.1
 - [[PR 1171]](https://github.com/parthenon-hpc-lab/parthenon/pull/1171) Add PARTHENON_USE_SYSTEM_PACKAGES build option
 - [[PR 1172]](https://github.com/parthenon-hpc-lab/parthenon/pull/1172) Make parthenon manager robust against external MPI init and finalize calls
 
@@ -35,7 +35,7 @@
 
 
 ### Incompatibilities (i.e. breaking changes)
-- [[PR1177]](https://github.com/parthenon-hpc-lab/parthenon/pull/1177) Make mesh-level boundary conditions usable without the "user" flag
+- [[PR 1177]](https://github.com/parthenon-hpc-lab/parthenon/pull/1177) Make mesh-level boundary conditions usable without the "user" flag
 
 ## Release 24.08
 Date: 2024-08-30
@@ -159,12 +159,12 @@ Date: 2024-03-21
 - [[PR 973]](https://github.com/parthenon-hpc-lab/parthenon/pull/973) Multigrid performance upgrades
 
 ### Fixed (not changing behavior/API/variables/...)
-- [[PR1023]](https://github.com/parthenon-hpc-lab/parthenon/pull/1023) Fix broken param of a scalar bool
-- [[PR1012]](https://github.com/parthenon-hpc-lab/parthenon/pull/1012) Remove accidentally duplicated code
-- [[PR992]](https://github.com/parthenon-hpc-lab/parthenon/pull/992) Allow custom PR ops with sparse pools
-- [[PR988]](https://github.com/parthenon-hpc-lab/parthenon/pull/988) Fix bug in neighbor finding routine for small, periodic, refined meshes
-- [[PR986]](https://github.com/parthenon-hpc-lab/parthenon/pull/986) Fix bug in sparse boundary communication BndInfo cacheing
-- [[PR978]](https://github.com/parthenon-hpc-lab/parthenon/pull/978) remove erroneous sparse check
+- [[PR 1023]](https://github.com/parthenon-hpc-lab/parthenon/pull/1023) Fix broken param of a scalar bool
+- [[PR 1012]](https://github.com/parthenon-hpc-lab/parthenon/pull/1012) Remove accidentally duplicated code
+- [[PR 992]](https://github.com/parthenon-hpc-lab/parthenon/pull/992) Allow custom PR ops with sparse pools
+- [[PR 988]](https://github.com/parthenon-hpc-lab/parthenon/pull/988) Fix bug in neighbor finding routine for small, periodic, refined meshes
+- [[PR 986]](https://github.com/parthenon-hpc-lab/parthenon/pull/986) Fix bug in sparse boundary communication BndInfo cacheing
+- [[PR 978]](https://github.com/parthenon-hpc-lab/parthenon/pull/978) remove erroneous sparse check
 
 ### Infrastructure (changes irrelevant to downstream codes)
 - [[PR 1027]](https://github.com/parthenon-hpc-lab/parthenon/pull/1027) Refactor RestartReader as abstract class
@@ -231,7 +231,7 @@ Date: 2023-11-16
 - [[PR 901]](https://github.com/parthenon-hpc-lab/parthenon/pull/901) Implement shared element ownership model
 
 ### Removed (removing behavior/API/varaibles/...)
-- [[PR 930](https://github.com/parthenon-hpc-lab/parthenon/pull/930) Remove ParthenonManager::ParthenonInit as it is error-prone and the split functions are the recommended usage.
+- [[PR 930]](https://github.com/parthenon-hpc-lab/parthenon/pull/930) Remove ParthenonManager::ParthenonInit as it is error-prone and the split functions are the recommended usage.
 
 
 ## Release 0.8.0

diff --git a/README.md b/README.md
@@ -30,7 +30,7 @@ Parthenon -- a performance portable block-structured adaptive mesh refinement fr
 
 * CMake 3.16 or greater
 * C++17 compatible compiler
-* Kokkos 4.0.1 or greater
+* Kokkos 4.4.1 or greater
 
 ## Optional (enabling features)
 

diff --git a/cmake/machinecfg/GitHubActions.cmake b/cmake/machinecfg/GitHubActions.cmake
@@ -19,6 +19,9 @@ message(STATUS "Loading machine configuration for GitHub Actions CI. ")
 
 # common options
 set(NUM_MPI_PROC_TESTING "2" CACHE STRING "CI runs tests with 2 MPI ranks")
+set(Kokkos_ENABLE_ROCTHRUST OFF CACHE BOOL "Temporarily disabled as the container needs to be updated to the `-complete` base image.")
+
+set(CMAKE_CXX_FLAGS_DBGNOSYM "-O0" CACHE STRING "Debug build without symbols")
 
 set(MACHINE_CXX_FLAGS "")
 if (${MACHINE_VARIANT} MATCHES "cuda")

diff --git a/doc/sphinx/src/development.rst b/doc/sphinx/src/development.rst
@@ -21,7 +21,10 @@ Kokkos wrappers/abstractions
 -  ``par_for`` wrappers use inclusive bounds, i.e., the loop will
    include the last index given
 -  ``ParArrayND`` arrays by default allocate on the *device* using
-   default precision configured
+   default precision configured and come with a `State` that can
+   be used to store additional metadata.
+- ``ParArray#DRaw`` directly map to Kokkos ``Views`` that are allocated
+   on *device* using default precision.
 -  To create an array on the host with identical layout to the device
    array either use
 
@@ -62,6 +65,42 @@ parallelism interface that is needed for managing memory cached in
 tightly nested loops. The wrappers are documented
 :ref:`here <nested par for>`.
 
+View of Views
+-------------
+
+Special care needs to be taken when working with a ``View`` of ``Views``.
+
+To repeat the Kokkos documenation: `Don't use them <https://kokkos.org/kokkos-core-wiki/ProgrammingGuide/View.html#can-i-make-a-view-of-views>`__
+
+But if you have to (which is the case in some places inside Parthenon)
+then follow this pattern:
+
+.. code:: c++
+
+   ParArray1DRaw<ParArray1D<Real>> view_of_pararrays(parthenon::ViewOfViewAlloc("myname"), 10);
+
+The ``ViewOfViewAlloc`` ensures that the ``Kokkos::SequentialHostInit`` property is added,
+which results in the (inner ``View`` ) deallocators being called on the host (rather than on
+the device by default).
+Also note the use of the "raw" ``ParArray1DRaw``, which directly maps to a Kokkos ``View``
+(that is required to process the allocation property as this interface is not exposed
+in the more generic ``ParArrayND``).
+
+Similarly, when you create a host mirror of said ``View`` of ``View`` add the additional
+property for the same reason.
+
+.. code:: c++
+
+   // explicit theoretical example -- don't use this
+   auto view_of_pararrays_h =
+        Kokkos::create_mirror_view(Kokkos::view_alloc(Kokkos::SequentialHostInit), view_of_pararrays);
+
+   // but instead use this interface provided by Parthenon:
+   auto view_of_pararrays_h = create_view_of_view_mirror(view_of_pararrays);
+
+
+Note that the ``SequentialHostInit`` was only added in Kokkos 4.4.1 (which is now the default in Parthenon).
+
 The need for reductions within function handling ``MeshBlock`` data
 -------------------------------------------------------------------
 

diff --git a/external/Kokkos b/external/Kokkos
diff --git a/scripts/docker/Dockerfile.hip-rocm b/scripts/docker/Dockerfile.hip-rocm
@@ -1,4 +1,4 @@
-FROM rocm/dev-ubuntu-20.04:5.4.3
+FROM rocm/dev-ubuntu-24.04:6.2
 
 RUN apt-get clean && apt-get update -y && \
     DEBIAN_FRONTEND="noninteractive" TZ=America/New_York apt-get install -y --no-install-recommends git python3-minimal libpython3-stdlib bc hwloc wget openssh-client python3-numpy python3-h5py python3-matplotlib lcov curl cmake ninja-build openmpi-bin libopenmpi-dev && \
@@ -14,12 +14,10 @@ RUN cd /tmp && \
     cd / && \
     rm -rf /tmp/hdf5-1.10.8*
 
-# "mpic++ --showme" forgets open-pal in Ubuntu 20.04 + OpenMPI 4.0.3
-#   https://bugs.launchpad.net/ubuntu/+source/openmpi/+bug/1941786
-#   https://github.com/open-mpi/ompi/issues/9317
-ENV LDFLAGS="-lopen-pal"
-
 RUN update-alternatives --install /usr/bin/python python /usr/bin/python3 10
 
-# uid 1000 maps to the one running the container on the CI host
-RUN useradd --create-home --shell /bin/bash -u 1000 -G render ci
+# Latest image has default user with uid 1000 (which maps to the one running the container on the CI host
+# Need to add user to the group that can access the GPU
+RUN usermod -a -G render ubuntu
+
+WORKDIR /home/ubuntu
diff --git a/src/bvals/bvals.hpp b/src/bvals/bvals.hpp
@@ -101,9 +101,6 @@ class BoundarySwarm : public BoundaryCommunication {
   explicit BoundarySwarm(std::weak_ptr<MeshBlock> pmb, const std::string &label);
   ~BoundarySwarm() = default;
 
-  std::vector<ParArrayND<int>> vars_int;
-  std::vector<ParArrayND<Real>> vars_real;
-
   // (usuallly the std::size_t unsigned integer type)
   std::vector<BoundaryCommunication *>::size_type bswarm_index;
 

diff --git a/src/bvals/comms/bnd_info.cpp b/src/bvals/comms/bnd_info.cpp
@@ -40,8 +40,8 @@
 namespace parthenon {
 
 void ProResCache_t::Initialize(int n_regions, StateDescriptor *pkg) {
-  prores_info = ParArray1D<ProResInfo>("prores_info", n_regions);
-  prores_info_h = Kokkos::create_mirror_view(prores_info);
+  prores_info = ProResInfoArr_t(ViewOfViewAlloc("prores_info"), n_regions);
+  prores_info_h = create_view_of_view_mirror(prores_info);
   int nref_funcs = pkg->NumRefinementFuncs();
   // Note that assignment of Kokkos views resets them, but
   // buffer_subset_sizes is a std::vector. It must be cleared, then

diff --git a/src/bvals/comms/bnd_info.hpp b/src/bvals/comms/bnd_info.hpp
@@ -26,6 +26,7 @@
 #include "bvals/neighbor_block.hpp"
 #include "coordinates/coordinates.hpp"
 #include "interface/variable_state.hpp"
+#include "kokkos_abstraction.hpp"
 #include "mesh/domain.hpp"
 #include "mesh/forest/logical_coordinate_transformation.hpp"
 #include "utils/communication_buffer.hpp"
@@ -127,11 +128,11 @@ struct ProResInfo {
 int GetBufferSize(MeshBlock *pmb, const NeighborBlock &nb,
                   std::shared_ptr<Variable<Real>> v);
 
-using BndInfoArr_t = ParArray1D<BndInfo>;
+using BndInfoArr_t = ParArray1DRaw<BndInfo>;
 using BndInfoArrHost_t = typename BndInfoArr_t::HostMirror;
 
-using ProResInfoArr_t = ParArray1D<ProResInfo>;
-using ProResInfoArrHost_t = typename ParArray1D<ProResInfo>::HostMirror;
+using ProResInfoArr_t = ParArray1DRaw<ProResInfo>;
+using ProResInfoArrHost_t = typename ProResInfoArr_t::HostMirror;
 class StateDescriptor;
 struct ProResCache_t {
   ProResInfoArr_t prores_info{};

diff --git a/src/bvals/comms/bvals_utils.hpp b/src/bvals/comms/bvals_utils.hpp
@@ -28,6 +28,7 @@
 #include "bvals/comms/bnd_info.hpp"
 #include "bvals/comms/bvals_in_one.hpp"
 #include "interface/variable.hpp"
+#include "kokkos_abstraction.hpp"
 #include "mesh/domain.hpp"
 #include "mesh/mesh.hpp"
 #include "mesh/meshblock.hpp"
@@ -215,8 +216,8 @@ inline void RebuildBufferCache(std::shared_ptr<MeshData<Real>> md, int nbound,
   using namespace loops;
   using namespace loops::shorthands;
   BvarsSubCache_t &cache = md->GetBvarsCache().GetSubCache(BOUND_TYPE, SENDER);
-  cache.bnd_info = BndInfoArr_t("bnd_info", nbound);
-  cache.bnd_info_h = Kokkos::create_mirror_view(cache.bnd_info);
+  cache.bnd_info = BndInfoArr_t(ViewOfViewAlloc("bnd_info"), nbound);
+  cache.bnd_info_h = create_view_of_view_mirror(cache.bnd_info);
 
   // prolongation/restriction sub-sets
   // TODO(JMM): Right now I exclude fluxcorrection boundaries but if

diff --git a/src/interface/mesh_data.hpp b/src/interface/mesh_data.hpp
@@ -26,6 +26,7 @@
 #include "interface/sparse_pack_base.hpp"
 #include "interface/swarm_pack_base.hpp"
 #include "interface/variable_pack.hpp"
+#include "kokkos_abstraction.hpp"
 #include "mesh/domain.hpp"
 #include "mesh/meshblock.hpp"
 #include "mesh/meshblock_pack.hpp"
@@ -149,8 +150,8 @@ const MeshBlockPack<P> &PackOnMesh(M &map, BlockDataList_t<Real> &block_data_,
   }
 
   if (make_new_pack) {
-    ParArray1D<P> packs("MeshData::PackVariables::packs", nblocks);
-    auto packs_host = Kokkos::create_mirror_view(packs);
+    ParArray1DRaw<P> packs(ViewOfViewAlloc("MeshData::PackVariables::packs"), nblocks);
+    auto packs_host = create_view_of_view_mirror(packs);
 
     for (size_t i = 0; i < nblocks; i++) {
       const auto &pack = packing_function(block_data_[i], this_map, this_key);

diff --git a/src/interface/sparse_pack_base.cpp b/src/interface/sparse_pack_base.cpp
@@ -30,6 +30,7 @@
 #include "interface/sparse_pack_base.hpp"
 #include "interface/state_descriptor.hpp"
 #include "interface/variable.hpp"
+#include "kokkos_abstraction.hpp"
 #include "utils/utils.hpp"
 namespace parthenon {
 namespace impl {
@@ -151,8 +152,8 @@ SparsePackBase SparsePackBase::Build(T *pmd, const PackDescriptor &desc,
   } else if (contains_face_or_edge) {
     leading_dim += 2;
   }
-  pack.pack_ = pack_t("data_ptr", leading_dim, pack.nblocks_, max_size);
-  pack.pack_h_ = Kokkos::create_mirror_view(pack.pack_);
+  pack.pack_ = pack_t(ViewOfViewAlloc("data_ptr"), leading_dim, pack.nblocks_, max_size);
+  pack.pack_h_ = create_view_of_view_mirror(pack.pack_);
 
   // For non-flat packs, shape of pack is type x block x var x k x j x i
   // where type here might be a flux.
@@ -167,8 +168,8 @@ SparsePackBase SparsePackBase::Build(T *pmd, const PackDescriptor &desc,
   pack.block_props_ = block_props_t("block_props", nblocks, 27 + 1);
   pack.block_props_h_ = Kokkos::create_mirror_view(pack.block_props_);
 
-  pack.coords_ = coords_t("coords", desc.flat ? max_size : nblocks);
-  auto coords_h = Kokkos::create_mirror_view(pack.coords_);
+  pack.coords_ = coords_t(ViewOfViewAlloc("coords"), desc.flat ? max_size : nblocks);
+  auto coords_h = create_view_of_view_mirror(pack.coords_);
 
   // Fill the views
   int idx = 0;