Vector::DeleteAt on device using InclusiveScan

# Conflicts:
#	mesh/mesh.cpp

.github/workflows/builds-and-tests.yml

@@ -168,10 +168,13 @@ jobs:
           env
         shell: bash
 
+      # For info on Xcode see:
+      #  - https://github.com/actions/runner-images/issues/12541
+      #  - https://github.com/actions/runner-images/blob/releases/macos-15-arm64/20250811/images/macos/macos-15-arm64-Readme.md#xcode
       - name: Xcode version setup (MacOS)
         if: matrix.os == 'macos-latest'
         run: |
-          XCODE_PATH="/Applications/Xcode_15.3.app"
+          XCODE_PATH="/Applications/Xcode_16.4.app"
           echo "> sudo xcode-select -s ${XCODE_PATH}"
           sudo xcode-select -s ${XCODE_PATH}
           echo "> g++ -v"
@@ -289,10 +292,10 @@ jobs:
         run: |
           export HOMEBREW_NO_INSTALL_CLEANUP=1
           brew update
-          brew install llvm@19 enzyme
-          echo "LLVM_PREFIX=$(brew --prefix llvm@19)" >> $GITHUB_ENV
-          echo "OMPI_CC=$(brew --prefix llvm@19)/bin/clang" >> $GITHUB_ENV
-          echo "OMPI_CXX=$(brew --prefix llvm@19)/bin/clang++" >> $GITHUB_ENV
+          brew install llvm@20 enzyme
+          echo "LLVM_PREFIX=$(brew --prefix llvm@20)" >> $GITHUB_ENV
+          echo "OMPI_CC=$(brew --prefix llvm@20)/bin/clang" >> $GITHUB_ENV
+          echo "OMPI_CXX=$(brew --prefix llvm@20)/bin/clang++" >> $GITHUB_ENV
 
       # MFEM build and test
       - name: build

.gitignore

@@ -211,7 +211,7 @@ miniapps/electromagnetics/joule
 miniapps/electromagnetics/Volta-AMR*
 miniapps/electromagnetics/Tesla-AMR*
 miniapps/electromagnetics/Maxwell-Parallel*
-miniapps/electromagnetics/Joule_*
+miniapps/electromagnetics/Joule_[0-9]*
 
 miniapps/gslib/field-diff
 miniapps/gslib/field-interp
@@ -267,9 +267,9 @@ miniapps/meshing/bounding-box*
 miniapps/meshing/jacobian-determinant*
 
 miniapps/mtop/parheat
-miniapps/mtop/ParHeat*
+miniapps/mtop/ParHeat/*
 miniapps/mtop/seqheat
-miniapps/mtop/SeqHeat*
+miniapps/mtop/SeqHeat/*
 
 miniapps/autodiff/paradiff
 miniapps/autodiff/seqadiff
@@ -277,7 +277,7 @@ miniapps/autodiff/seqtest
 miniapps/autodiff/par_example
 miniapps/autodiff/seq_example
 miniapps/autodiff/seq_test
-miniapps/autodiff/Exampl*
+miniapps/autodiff/Example/*
 
 miniapps/navier/navier_mms
 miniapps/navier/navier_kovasznay

.gitlab-ci.yml

@@ -22,7 +22,7 @@ include:
 # the "needs" keyword and express the DAG of jobs for more efficiency.
 # - We use setup and setup_baseline phases to download content outside of mfem
 #   directory.
-# - Allocate/Release is where ruby resource are allocated/released once for all.
+# - Allocate/Release is where Dane resource are allocated/released once for all.
 # - Build and Test is where we build and MFEM for multiple toolchains.
 # - Baseline_checks gathers baseline-type test suites execution
 # - Baseline_publish, only available on master, allows to update baseline
@@ -53,7 +53,7 @@ variables:
   AUTOTEST_COMMIT: "YES"
 
 # Trigger subpipelines:
-ruby-build-and-test:
+dane-build-and-test:
   stage: sub-pipelines
   variables:
     # Explicitly pass down values that we want to be able to set when triggering
@@ -61,10 +61,10 @@ ruby-build-and-test:
     AUTOTEST: "${AUTOTEST}"
     AUTOTEST_COMMIT: "${AUTOTEST_COMMIT}"
   trigger:
-    include: .gitlab/ruby-build-and-test.yml
+    include: .gitlab/dane-build-and-test.yml
     strategy: depend
 
-ruby-baseline:
+dane-baseline:
   stage: sub-pipelines
   variables:
     # Explicitly pass down values that we want to be able to set when triggering
@@ -73,7 +73,7 @@ ruby-baseline:
     AUTOTEST: "${AUTOTEST}"
     AUTOTEST_COMMIT: "${AUTOTEST_COMMIT}"
   trigger:
-    include: .gitlab/ruby-baseline.yml
+    include: .gitlab/dane-baseline.yml
     strategy: depend
 
 lassen-build-and-test:

.gitlab/README.md

@@ -24,7 +24,7 @@ and `test type`.
 
 Machines typically include:
 
-* Ruby: 2nd Gen Intel Xeon (Cascade Lake)
+* Dane: Intel Sapphire Rapids
 * Lassen: Power9 + Nvidia GPU
 * Corona: AMD GPU
 
@@ -76,13 +76,13 @@ with a spack spec of MFEM, within the limits permitted by the MFEM spack
 package.
 
 In any build-and-test sub-pipeline a job basically consists in defining the
-spack spec to use. Adding a job on ruby for example resumes to:
+spack spec to use. Adding a job on Dane for example resumes to:
 
 ```yaml
 <job_name>:
   variables:
     SPEC: "<spack_spec>"
-  extends: .build_and_test_on_ruby
+  extends: .build_and_test_on_dane
 ```
 
 The remaining and non trivial work is to make sure this spec is working. To

.gitlab/configs/common.yml

@@ -24,7 +24,7 @@ variables:
 # TODO: add a clean-up mechanism
   BUILD_ROOT: ${USER_CI_TOP_DIR}/${CI_PROJECT_NAME}-${MACHINE_NAME}-pipeline-${CI_PIPELINE_ID}
 
-# On LLNL's ruby, there is only one allocation shared among jobs in order to
+# On LLNL's Dane, there is only one allocation shared among jobs in order to
 # save time and resource. This allocation has to be uniquely named so that we
 # are sure to retrieve it.
   ALLOC_NAME: ${CI_PROJECT_NAME}_ci_${CI_PIPELINE_ID}

.gitlab/configs/dane-config.yml

@@ -9,17 +9,17 @@
 # terms of the BSD-3 license. We welcome feedback and contributions, see file
 # CONTRIBUTING.md for details.
 
-# GitLab pipelines configurations for the Ruby machine at LLNL
+# GitLab pipelines configurations for the Dane machine at LLNL
 variables:
-  MACHINE_NAME: ruby
+  MACHINE_NAME: dane
 
-.on_ruby:
+.on_dane:
   tags:
     - shell
-    - ruby
+    - dane
   rules:
-    # Don't run ruby jobs if...
-    - if: '$CI_COMMIT_BRANCH =~ /_qnone/ || $ON_RUBY == "OFF"'
+    # Don't run dane jobs if...
+    - if: '$CI_COMMIT_BRANCH =~ /_qnone/ || $ON_DANE == "OFF"'
       when: never
     # Don't run autotest update if...
     - if: '$CI_JOB_NAME =~ /report/ && $AUTOTEST != "YES"'
@@ -40,16 +40,17 @@ variables:
     - when: on_success
 
 # Spack helped builds
-# Generic ruby build job, extending build script
-.build_and_test_on_ruby:
-  extends: [.on_ruby]
+# Generic dane build job, extending build script
+.build_and_test_on_dane:
+  extends: [.on_dane]
   stage: build_and_test
   script:
     # THREADS is used by 'tests/gitlab/build_and_test', run below
-    - export THREADS=16
+    # Dane has 224 threads/node and we run 7 separate jobs: 224=7*32
+    - export THREADS=28
     - echo ${ALLOC_NAME}
     - export JOBID=$(squeue -h --name=${ALLOC_NAME} --format=%A)
     - echo ${JOBID}
     - echo ${MFEM_DATA_DIR}
     - echo ${SPEC}
-    - srun $( [[ -n "${JOBID}" ]] && echo "--jobid=${JOBID}" ) --reservation=ci -t 45 -N 1 tests/gitlab/build_and_test --spec "${SPEC}" --data-dir "${MFEM_DATA_DIR}" --data
+    - srun $( [[ -n "${JOBID}" ]] && echo "--jobid=${JOBID}" ) --reservation=ci -t 60 -N 1 tests/gitlab/build_and_test --spec "${SPEC}" --data-dir "${MFEM_DATA_DIR}" --data

.gitlab/configs/setup-baseline.yml

@@ -18,7 +18,7 @@
 setup_baseline:
   tags:
     - shell
-    - ruby
+    - dane
   stage: setup
   variables:
     GIT_STRATEGY: none

.gitlab/configs/setup-build-and-test.yml

@@ -16,7 +16,7 @@
 setup:
   tags:
     - shell
-    - ruby
+    - dane
   stage: setup
   variables:
     GIT_STRATEGY: none

.gitlab/dane-baseline.yml

@@ -19,8 +19,8 @@ stages:
   - cleanup
   - baseline_publish
 
-baselinecheck_mfem_intel_ruby:
-  extends: [.on_ruby]
+baselinecheck_mfem_intel_dane:
+  extends: [.on_dane]
   stage: baseline_check
   variables:
     # TPLS_DIR is used in .gitlab/scripts/baseline to provide the tpls location
@@ -31,8 +31,8 @@ baselinecheck_mfem_intel_ruby:
   script:
     - echo ${BUILD_ROOT}
     - echo ${TPLS_DIR}
-    # Used by the tests in MFEM/tests:
-    - export MFEM_TEST_NP=48
+    # Used by the tests in MFEM/tests, dane has 224 threads/node:
+    - export MFEM_TEST_NP=192
     # The next script uses the following environment variables:
     # * BASELINE_TEST, SYS_TYPE, CI_PROJECT_DIR, ARTIFACTS_DIR,
     # * BUILD_ROOT, TPLS_DIR, MACHINE_NAME
@@ -44,7 +44,7 @@ baselinecheck_mfem_intel_ruby:
   allow_failure: true
 
 cleanup:
-  extends: .on_ruby
+  extends: .on_dane
   stage: cleanup
   variables:
     GIT_STRATEGY: none
@@ -53,7 +53,7 @@ cleanup:
     - rm -rf "${BUILD_ROOT}" || true
 
 report_baseline:
-  extends: [.on_ruby]
+  extends: [.on_dane]
   stage: baseline_report
   script:
     - echo ${MACHINE_NAME}
@@ -113,8 +113,8 @@ report_baseline:
         exit $err
       ) 9> autotest.lock
 
-baselinepublish_mfem_ruby:
-  extends: [.on_ruby]
+baselinepublish_mfem_dane:
+  extends: [.on_dane]
   stage: baseline_publish
   rules:
     # - if: '$CI_COMMIT_BRANCH == "master" || $REBASELINE == "YES"'
@@ -129,5 +129,5 @@ baselinepublish_mfem_ruby:
 
 include:
   - local: .gitlab/configs/common.yml
-  - local: .gitlab/configs/ruby-config.yml
+  - local: .gitlab/configs/dane-config.yml
   - local: .gitlab/configs/setup-baseline.yml

.gitlab/dane-build-and-test.yml

@@ -19,54 +19,54 @@ stages:
 allocate_resource:
   variables:
     GIT_STRATEGY: none
-  extends: .on_ruby
+  extends: .on_dane
   stage: allocate_resource
   script:
     - echo ${ALLOC_NAME}
     - salloc --exclusive --nodes=1 --reservation=ci --time=60 --no-shell --job-name=${ALLOC_NAME}
   timeout: 6h
 
-# GitLab jobs for the Ruby machine at LLNL
+# GitLab jobs for the Dane machine at LLNL
 debug_ser_gcc_10:
   variables:
     SPEC: "%gcc@10.3.1 +debug~mpi"
-  extends: .build_and_test_on_ruby
+  extends: .build_and_test_on_dane
 
 debug_par_gcc_10:
   variables:
     SPEC: "%gcc@10.3.1 +debug+mpi"
-  extends: .build_and_test_on_ruby
+  extends: .build_and_test_on_dane
 
 opt_ser_gcc_10:
   variables:
     SPEC: "%gcc@10.3.1 ~mpi"
-  extends: .build_and_test_on_ruby
+  extends: .build_and_test_on_dane
 
 opt_par_gcc_10:
   variables:
     SPEC: "%gcc@10.3.1"
-  extends: .build_and_test_on_ruby
+  extends: .build_and_test_on_dane
 
 opt_par_gcc_10_sundials:
   variables:
     SPEC: "%gcc@10.3.1 +sundials"
-  extends: .build_and_test_on_ruby
+  extends: .build_and_test_on_dane
 
 opt_par_gcc_10_petsc:
   variables:
     SPEC: "%gcc@10.3.1 +petsc ^petsc+mumps~superlu-dist"
-  extends: .build_and_test_on_ruby
+  extends: .build_and_test_on_dane
 
 opt_par_gcc_10_pumi:
   variables:
     SPEC: "%gcc@10.3.1 +pumi"
-  extends: .build_and_test_on_ruby
+  extends: .build_and_test_on_dane
 
 # Release
 release_resource:
   variables:
     GIT_STRATEGY: none
-  extends: .on_ruby
+  extends: .on_dane
   stage: release_resource_and_report
   script:
     - echo ${ALLOC_NAME}
@@ -78,17 +78,17 @@ release_resource:
 report_job_success:
   stage: release_resource_and_report
   extends:
-    - .on_ruby
+    - .on_dane
     - .report_job_success
 
 report_job_failure:
   stage: release_resource_and_report
   extends:
-    - .on_ruby
+    - .on_dane
     - .report_job_failure
 
 include:
   - local: .gitlab/configs/common.yml
-  - local: .gitlab/configs/ruby-config.yml
+  - local: .gitlab/configs/dane-config.yml
   - local: .gitlab/configs/setup-build-and-test.yml
   - local: .gitlab/configs/report-build-and-test.yml

.gitlab/scripts/baseline

@@ -14,7 +14,7 @@
 # locals
 glob_err=${BASELINE_TEST}.err
 base=${BASELINE_TEST}-${SYS_TYPE}
-if [[ "${MACHINE_NAME}" == "ruby" ]]; then
+if [[ "${MACHINE_NAME}" == "dane" ]]; then
   base="${BASELINE_TEST}-${MACHINE_NAME}"
 fi
 base_diff=${base}.diff
@@ -31,7 +31,7 @@ cd tests
 mkdir _${BASELINE_TEST} && cd _${BASELINE_TEST}
 
 # run
-if [[ "${MACHINE_NAME}" == "ruby" ]]; then
+if [[ "${MACHINE_NAME}" == "dane" ]]; then
   salloc --nodes=1 --exclusive --reservation=ci ../runtest ../../mfem "${BASELINE_TEST} ${TPLS_DIR}"
 elif [[ ${MACHINE_NAME} == "corona" ]]; then
   salloc --nodes=1 -t 60 -p pbatch ../runtest ../../mfem "${BASELINE_TEST} ${TPLS_DIR}"

.gitlab/scripts/rebaseline

@@ -11,7 +11,7 @@
 # terms of the BSD-3 license. We welcome feedback and contributions, see file
 # CONTRIBUTING.md for details.
 
-# There will be collision between corona and ruby baselines.
+# There will be collision between corona and dane baselines.
 # Once the corresponding files have been generated, we can switch to machine
 # specific ref.
 ARTIFACT_PATH=${CI_PROJECT_DIR}/${ARTIFACTS_DIR}/baseline-${SYS_TYPE}
@@ -21,7 +21,7 @@ PATCH_FILE=${ARTIFACT_PATH}.patch
 FULL_FILE=${ARTIFACT_PATH}.out
 DIFF_FILE=${ARTIFACT_PATH}.diff
 
-# There will be collision between corona and ruby baselines.
+# There will be collision between corona and dane baselines.
 # Once the corresponding files have been generated, we can switch to machine
 # specific ref.
 SAVED_NAME=baseline-${SYS_TYPE}.saved

CHANGELOG

@@ -46,6 +46,11 @@ GPU computing
   where host execution is always needed (e.g. when the DOFs array is small).
 - Introduced MFEM_FOREACH_THREAD_DIRECT, which directly maps loop tasks to GPU
   threads, assigning one task per thread.
+- Implemented a GPU-accelerated matrix-free AMR derefinement `GridFunction`
+  update operator. This supports mixed geometry meshes and variable order
+  spaces, and is the default derefinement operator constructed by
+  `FiniteElementSpace::Update` and `ParFiniteElementSpace::Update`.
+  The operator requires `FiniteElementSpace::Nonconforming() == true`.
 
 New and updated examples and miniapps
 -------------------------------------

CMakeLists.txt

@@ -598,14 +598,20 @@ set(MFEM_TPLS OPENMP HYPRE LAPACK BLAS SuperLUDist STRUMPACK METIS SuiteSparse
     BENCHMARK PARELAG TRIBOL MPI_CXX HIP HIPBLAS HIPSPARSE MOONOLITH BLITZ
     ALGOIM ENZYME)
 
-# Add all *_FOUND libraries in the variable TPL_LIBRARIES.
+# Add all created targets and *_FOUND libraries in the variables TPL_TARGETS and
+# TPL_LIBRARIES, respectively.
+set(TPL_TARGETS)
 set(TPL_LIBRARIES "")
 set(TPL_INCLUDE_DIRS "")
 foreach(TPL IN LISTS MFEM_TPLS)
-  if (${TPL}_FOUND)
+  if (${TPL}_FOUND OR TARGET ${TPL})
     message(STATUS "MFEM: using package ${TPL}")
-    list(APPEND TPL_LIBRARIES ${${TPL}_LIBRARIES})
-    list(APPEND TPL_INCLUDE_DIRS ${${TPL}_INCLUDE_DIRS})
+    if (TARGET ${TPL})
+      list(APPEND TPL_TARGETS ${TPL})
+    else()
+      list(APPEND TPL_LIBRARIES ${${TPL}_LIBRARIES})
+      list(APPEND TPL_INCLUDE_DIRS ${${TPL}_INCLUDE_DIRS})
+    endif()
   endif()
 endforeach(TPL)
 list(REVERSE TPL_LIBRARIES)
@@ -680,7 +686,10 @@ set(MFEM_INSTALL_DIR ${CMAKE_INSTALL_PREFIX})
 # Declaring the library
 mfem_add_library(mfem ${SOURCES} ${HEADERS} ${MASTER_HEADERS})
 # message(STATUS "TPL_LIBRARIES = ${TPL_LIBRARIES}")
-target_link_libraries(mfem PUBLIC ${TPL_LIBRARIES})
+target_link_libraries(mfem PUBLIC ${TPL_LIBRARIES} ${TPL_TARGETS})
+if (TPL_TARGETS)
+  add_dependencies(mfem ${TPL_TARGETS})
+endif()
 if (MINGW)
   target_link_libraries(mfem PRIVATE ws2_32)
 endif()

INSTALL

@@ -121,6 +121,11 @@ Parallel build:
    make -j 4
 (For METIS 5, see https://mfem.org/building/#parallel-build-using-metis-5)
 
+Parallel build with fetching of hypre and METIS:
+   mkdir <mfem-buil-dir> ; cd <mfem-build-dir>
+   cmake <mfem-source-dir> -DMFEM_USE_MPI=YES -DFETCH_TPLS=YES
+   make -j 4
+
 CUDA build:
    (this build requires CMake 3.17 or newer)
    mkdir <mfem-build-dir> ; cd <mfem-build-dir>
@@ -663,6 +668,7 @@ The specific libraries and their options are:
 - OpenMP (optional), usually part of compiler, used when either MFEM_USE_OPENMP
   or MFEM_USE_LEGACY_OPENMP is set to YES.
   Options: OPENMP_OPT, OPENMP_LIB.
+  Versions: OpenMP >= 3.1 when MFEM_USE_OPENMP=YES.
 
 - High-resolution POSIX clocks: when using MFEM_TIMER_TYPE = 2, it may be
   necessary to link with a system library (e.g. librt.so).
@@ -842,6 +848,7 @@ The specific libraries and their options are:
 - HIP (optional), used when MFEM_USE_HIP = YES.
   URL: https://rocmdocs.amd.com
   Options: HIP_CXX, HIP_ARCH, HIP_OPT, HIP_LIB.
+  Versions: ROCm >= 5.6.1.
 
 - OCCA (optional), used when MFEM_USE_OCCA = YES.
   URL: https://libocca.org
@@ -1074,6 +1081,9 @@ The following options are CMake specific:
 MFEM_ENABLE_TESTING  - Enable the ctest framework for testing.
 MFEM_ENABLE_EXAMPLES - Build all of the examples by default.
 MFEM_ENABLE_MINIAPPS - Build all of the miniapps by default.
+FETCH_TPLS           - Enable fetching of all supported third-party libraries.
+HYPRE_FETCH          - Enable fetching of hypre.
+METIS_FETCH          - Enable fetching of metis.
 
 External libraries (CMake):
 ---------------------------
@@ -1135,6 +1145,12 @@ The following built-in CMake packages are also used:
       set the <LIBNAME>_LIBRARIES option directly; the configuration option
       <LIBNAME>_DIR is not supported.
 
+The MFEM CMake build system also provides fetching (automated building) for the
+packages/libraries listed below. Note that when fetching is enabled, any related
+auto-detection functionality is disabled.
+
+ - HYPRE
+ - METIS
 
 Building without GNU make or CMake
 ==================================

config/cmake/modules/FindHYPRE.cmake

@@ -9,15 +9,18 @@
 # terms of the BSD-3 license. We welcome feedback and contributions, see file
 # CONTRIBUTING.md for details.
 
-# Defines the following variables:
+# Defines the following variables if fetching of TPLs is disabled (default):
 #   - HYPRE_FOUND
 #   - HYPRE_LIBRARIES
 #   - HYPRE_INCLUDE_DIRS
 #   - HYPRE_VERSION
 #   - HYPRE_USING_CUDA (internal)
 #   - HYPRE_USING_HIP (internal)
+# otherwise, the following are defined:
+#   - HYPRE (imported library target)
+#   - HYPRE_VERSION (cache variable)
 
-if (HYPRE_FOUND)
+if (HYPRE_FOUND OR TARGET HYPRE)
   if (HYPRE_USING_CUDA)
     find_package(CUDAToolkit REQUIRED)
   endif()
@@ -33,6 +36,55 @@ if (HYPRE_FOUND)
   endif()
 endif()
 
+if (HYPRE_FETCH OR FETCH_TPLS)
+  set(HYPRE_FETCH_VERSION 2.33.0)
+  add_library(HYPRE STATIC IMPORTED)
+  # set options and associated dependencies
+  set(CMAKE_OPTIONS)
+  list(APPEND CMAKE_OPTIONS -DCMAKE_BUILD_TYPE:STRING=${CMAKE_BUILD_TYPE})
+  if (MFEM_USE_CUDA)
+    list(APPEND CMAKE_OPTIONS -DHYPRE_WITH_CUDA:BOOL=ON)
+    find_package(CUDAToolkit REQUIRED)
+    target_link_libraries(HYPRE INTERFACE CUDA::cusparse CUDA::curand CUDA::cublas)
+  elseif (MFEM_USE_HIP)
+    list(APPEND CMAKE_OPTIONS -DHYPRE_WITH_HIP:BOOL=ON)
+    find_package(rocsparse REQUIRED)
+    find_package(rocrand REQUIRED)
+    target_link_libraries(HYPRE INTERFACE rocsparse rocrand)
+  endif()
+  if (MFEM_USE_SINGLE)
+    list(APPEND CMAKE_OPTIONS -DHYPRE_ENABLE_SINGLE:BOOL=ON)
+  endif()
+  # define external project and create future include directory so it is present
+  # to pass CMake checks at end of MFEM configuration step
+  message(STATUS "Will fetch HYPRE ${HYPRE_FETCH_VERSION} to be built with ${CMAKE_OPTIONS}")
+  set(PREFIX ${CMAKE_BINARY_DIR}/fetch/hypre)
+  include(ExternalProject)
+  ExternalProject_Add(hypre
+    GIT_REPOSITORY https://github.com/hypre-space/hypre.git
+    GIT_TAG v${HYPRE_FETCH_VERSION}
+    GIT_SHALLOW TRUE
+    UPDATE_DISCONNECTED TRUE
+    SOURCE_SUBDIR src
+    PREFIX ${PREFIX}
+    CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${PREFIX} -DCMAKE_INSTALL_LIBDIR:PATH=lib ${CMAKE_OPTIONS})
+  file(MAKE_DIRECTORY ${PREFIX}/include)
+  # set imported library target properties
+  add_dependencies(HYPRE hypre)
+  set_target_properties(HYPRE PROPERTIES
+    IMPORTED_LOCATION ${PREFIX}/lib/libHYPRE.a
+    INTERFACE_INCLUDE_DIRECTORIES ${PREFIX}/include)
+  # convert HYPRE version to integer
+  string(REGEX MATCHALL "[0-9]+" HYPRE_SPLIT_VERSION ${HYPRE_FETCH_VERSION})
+  list(GET HYPRE_SPLIT_VERSION 0 HYPRE_MAJOR_VERSION)
+  list(GET HYPRE_SPLIT_VERSION 1 HYPRE_MINOR_VERSION)
+  list(GET HYPRE_SPLIT_VERSION 2 HYPRE_PATCH_VERSION)
+  math(EXPR HYPRE_VERSION "10000*${HYPRE_MAJOR_VERSION} + 100*${HYPRE_MINOR_VERSION} + ${HYPRE_PATCH_VERSION}")
+  # set cache variables that would otherwise be set after mfem_find_package call
+  set(HYPRE_VERSION ${HYPRE_VERSION} CACHE STRING "HYPRE version." FORCE)
+  return()
+endif()
+
 include(MfemCmakeUtilities)
 mfem_find_package(HYPRE HYPRE HYPRE_DIR "include" "HYPRE.h" "lib" "HYPRE"
   "Paths to headers required by HYPRE." "Libraries required by HYPRE."

config/cmake/modules/FindMETIS.cmake

@@ -9,10 +9,38 @@
 # terms of the BSD-3 license. We welcome feedback and contributions, see file
 # CONTRIBUTING.md for details.
 
-# Defines the following variables:
+# Defines the following variables if fetching of TPLs is disabled (default):
 #   - METIS_FOUND
 #   - METIS_LIBRARIES
 #   - METIS_INCLUDE_DIRS
+#   - METIS_VERSION_5
+# otherwise, the following are defined:
+#   - METIS (imported library target)
+#   - METIS_VERSION_5 (cache variable)
+
+if (METIS_FETCH OR FETCH_TPLS)
+  set(METIS_FETCH_VERSION 4.0.3)
+  add_library(METIS STATIC IMPORTED)
+  # define external project
+  message(STATUS "Will fetch METIS ${METIS_FETCH_VERSION} to be built with default options")
+  set(PREFIX ${CMAKE_BINARY_DIR}/fetch/metis)
+  include(ExternalProject)
+  ExternalProject_Add(metis
+    GIT_REPOSITORY https://github.com/mfem/tpls
+    GIT_TAG b60352fbe9675d374b00828055e55be4584c7995 # tag from 1/16/25
+    GIT_SHALLOW TRUE
+    UPDATE_DISCONNECTED TRUE
+    PREFIX ${PREFIX}
+    CONFIGURE_COMMAND tar -xzf ../metis/metis-${METIS_FETCH_VERSION}-mac.tgz --strip=1
+    INSTALL_COMMAND mkdir -p ${PREFIX}/lib && cp libmetis.a ${PREFIX}/lib/)
+  # set imported library target properties
+  add_dependencies(METIS metis)
+  set_target_properties(METIS PROPERTIES
+    IMPORTED_LOCATION ${PREFIX}/lib/libmetis.a)
+  # set cache variables that would otherwise be set after mfem_find_package call
+  set(METIS_VERSION_5 FALSE CACHE BOOL "Is METIS version 5?")
+  return()
+endif()
 
 include(MfemCmakeUtilities)
 mfem_find_package(METIS METIS METIS_DIR "include;Lib" "metis.h"

config/config.hpp

@@ -23,11 +23,14 @@
 #include "_config.hpp"
 #endif
 
+#include <cstdint>
+#include <climits>
+
 namespace mfem
 {
 
 #if (defined(MFEM_USE_CUDA) && defined(__CUDACC__)) || \
-    (defined(MFEM_USE_HIP) && defined(__HIPCC__))
+    (defined(MFEM_USE_HIP) && defined(__HIP__))
 #define MFEM_HOST_DEVICE __host__ __device__
 #else
 #define MFEM_HOST_DEVICE

config/defaults.cmake

@@ -89,6 +89,12 @@ option(MFEM_ENABLE_EXAMPLES "Build all of the examples" OFF)
 option(MFEM_ENABLE_MINIAPPS "Build all of the miniapps" OFF)
 option(MFEM_ENABLE_BENCHMARKS "Build all of the benchmarks" OFF)
 
+# Allow a user to specify fetching of certain third-party libraries instead of
+# searching for existing installations.
+option(FETCH_TPLS "Enable fetching of all supported third-party libraries" OFF)
+option(HYPRE_FETCH "Enable fetching of hypre" OFF)
+option(METIS_FETCH "Enable fetching of METIS" OFF)
+
 # Setting CXX/MPICXX on the command line or in user.cmake will overwrite the
 # autodetected C++ compiler.
 # set(CXX g++)

fem/CMakeLists.txt

@@ -82,6 +82,8 @@ set(SRCS
   fe/fe_ser.cpp
   fe_coll.cpp
   fespace.cpp
+  derefmat_op.cpp
+  pderefmat_op.cpp
   geom.cpp
   gridfunc.cpp
   hybridization.cpp

fem/bilinearform.cpp

@@ -1275,6 +1275,22 @@ void BilinearForm::Update(FiniteElementSpace *nfes)
    height = width = fes->GetVSize();
 
    if (ext) { ext->Update(); }
+   for (int k = 0; k < domain_integs.Size(); ++k)
+   {
+      domain_integs[k]->Update();
+   }
+   for (int k = 0; k < boundary_integs.Size(); ++k)
+   {
+      boundary_integs[k]->Update();
+   }
+   for (int k = 0; k < interior_face_integs.Size(); ++k)
+   {
+      interior_face_integs[k]->Update();
+   }
+   for (int k = 0; k < boundary_integs.Size(); ++k)
+   {
+      boundary_face_integs[k]->Update();
+   }
 }
 
 void BilinearForm::SetDiagonalPolicy(DiagonalPolicy policy)
@@ -2337,6 +2353,31 @@ void MixedBilinearForm::Update()
    height = test_fes->GetVSize();
    width = trial_fes->GetVSize();
    if (ext) { ext->Update(); }
+
+   for (int k = 0; k < domain_integs.Size(); ++k)
+   {
+      domain_integs[k]->Update();
+   }
+   for (int k = 0; k < boundary_integs.Size(); ++k)
+   {
+      boundary_integs[k]->Update();
+   }
+   for (int k = 0; k < interior_face_integs.Size(); ++k)
+   {
+      interior_face_integs[k]->Update();
+   }
+   for (int k = 0; k < boundary_integs.Size(); ++k)
+   {
+      boundary_face_integs[k]->Update();
+   }
+   for (int k = 0; k < trace_face_integs.Size(); ++k)
+   {
+      trace_face_integs[k]->Update();
+   }
+   for (int k = 0; k < boundary_trace_face_integs.Size(); ++k)
+   {
+      boundary_trace_face_integs[k]->Update();
+   }
 }
 
 MixedBilinearForm::~MixedBilinearForm()

fem/bilinearform_ext.cpp

@@ -266,11 +266,7 @@ void PABilinearFormExtension::SetupRestrictionOperators(const L2FaceValues m)
 
       // Gather the attributes on the host from all the elements
       const Mesh &mesh = *trial_fes->GetMesh();
-      elem_attributes.SetSize(mesh.GetNE());
-      for (int i = 0; i < mesh.GetNE(); ++i)
-      {
-         elem_attributes[i] = mesh.GetAttribute(i);
-      }
+      elem_attributes = &mesh.GetElementAttributes();
    }
 
    // Construct face restriction operators only if the bilinear form has
@@ -329,45 +325,7 @@ void PABilinearFormExtension::SetupRestrictionOperators(const L2FaceValues m)
          bdr_face_dYdn.SetSize(bdr_face_restrict_lex->Height());
       }
 
-      const Mesh &mesh = *trial_fes->GetMesh();
-      // See LinearFormExtension::Update for explanation of f_to_be logic.
-      std::unordered_map<int,int> f_to_be;
-      for (int i = 0; i < mesh.GetNBE(); ++i)
-      {
-         const int f = mesh.GetBdrElementFaceIndex(i);
-         f_to_be[f] = i;
-      }
-      const int nf_bdr = trial_fes->GetNFbyType(FaceType::Boundary);
-      bdr_attributes.SetSize(nf_bdr);
-      int f_ind = 0;
-      int missing_bdr_elems = 0;
-      for (int f = 0; f < mesh.GetNumFaces(); ++f)
-      {
-         if (!mesh.GetFaceInformation(f).IsOfFaceType(FaceType::Boundary))
-         {
-            continue;
-         }
-         int attribute = 1; // default value
-         if (f_to_be.find(f) != f_to_be.end())
-         {
-            const int be = f_to_be[f];
-            attribute = mesh.GetBdrAttribute(be);
-         }
-         else
-         {
-            // If a boundary face does not correspond to the a boundary element,
-            // we assign it the default attribute of 1. We also generate a
-            // warning at runtime with the number of such missing elements.
-            ++missing_bdr_elems;
-         }
-         bdr_attributes[f_ind] = attribute;
-         ++f_ind;
-      }
-      if (missing_bdr_elems)
-      {
-         MFEM_WARNING("Missing " << missing_bdr_elems << " boundary elements "
-                      "for boundary faces.");
-      }
+      bdr_face_attributes = &trial_fes->GetMesh()->GetBdrFaceAttributes();
    }
 }
 
@@ -429,7 +387,7 @@ void PABilinearFormExtension::AssembleDiagonal(Vector &y) const
          mfem::forall(ne, [=] MFEM_HOST_DEVICE (int e)
          {
             const int attr = d_attr[e];
-            if (d_m[attr - 1] == 0)
+            if (attr <= 0 || d_m[attr - 1] == 0)
             {
                for (int i = 0; i < nd; ++i)
                {
@@ -450,7 +408,7 @@ void PABilinearFormExtension::AssembleDiagonal(Vector &y) const
          for (int i = 0; i < iSz; ++i)
          {
             assemble_diagonal_with_markers(*integrators[i], elem_markers[i],
-                                           elem_attributes, localY);
+                                           *elem_attributes, localY);
          }
          const ElementRestriction* H1elem_restrict =
             dynamic_cast<const ElementRestriction*>(elem_restrict);
@@ -476,7 +434,7 @@ void PABilinearFormExtension::AssembleDiagonal(Vector &y) const
       for (int i = 0; i < iSz; ++i)
       {
          assemble_diagonal_with_markers(*integrators[i], elem_markers[i],
-                                        elem_attributes, y);
+                                        *elem_attributes, y);
       }
    }
 
@@ -489,7 +447,7 @@ void PABilinearFormExtension::AssembleDiagonal(Vector &y) const
       for (int i = 0; i < n_bdr_integs; ++i)
       {
          assemble_diagonal_with_markers(*bdr_integs[i], bdr_markers[i],
-                                        bdr_attributes, bdr_face_Y);
+                                        *bdr_face_attributes, bdr_face_Y);
       }
       bdr_face_restrict_lex->AddAbsMultTranspose(bdr_face_Y, y);
    }
@@ -588,7 +546,7 @@ void PABilinearFormExtension::MultInternal(const Vector &x, Vector &y,
          for (int i = 0; i < iSz; ++i)
          {
             AddMultWithMarkers(*integrators[i], localX, elem_markers[i],
-                               elem_attributes, false, localY, useAbs);
+                               *elem_attributes, false, localY, useAbs);
          }
          if (H1elem_restrict && useAbs)
          {
@@ -690,8 +648,8 @@ void PABilinearFormExtension::MultInternal(const Vector &x, Vector &y,
          }
          for (int i = 0; i < n_bdr_integs; ++i)
          {
-            AddMultWithMarkers(*bdr_integs[i], bdr_face_X, bdr_markers[i], bdr_attributes,
-                               false, bdr_face_Y);
+            AddMultWithMarkers(*bdr_integs[i], bdr_face_X, bdr_markers[i],
+                               *bdr_face_attributes, false, bdr_face_Y);
          }
          for (int i = 0; i < n_bdr_face_integs; ++i)
          {
@@ -699,12 +657,14 @@ void PABilinearFormExtension::MultInternal(const Vector &x, Vector &y,
             {
                AddMultNormalDerivativesWithMarkers(
                   *bdr_face_integs[i], bdr_face_X, bdr_face_dXdn,
-                  bdr_face_markers[i], bdr_attributes, bdr_face_Y, bdr_face_dYdn);
+                  bdr_face_markers[i], *bdr_face_attributes, bdr_face_Y,
+                  bdr_face_dYdn);
             }
             else
             {
-               AddMultWithMarkers(*bdr_face_integs[i], bdr_face_X, bdr_face_markers[i],
-                                  bdr_attributes, false, bdr_face_Y);
+               AddMultWithMarkers(*bdr_face_integs[i], bdr_face_X,
+                                  bdr_face_markers[i], *bdr_face_attributes, false,
+                                  bdr_face_Y);
             }
          }
          bdr_face_restrict_lex->AddMultTransposeInPlace(bdr_face_Y, y);
@@ -727,7 +687,7 @@ void PABilinearFormExtension::MultTranspose(const Vector &x, Vector &y) const
       localY = 0.0;
       for (int i = 0; i < iSz; ++i)
       {
-         AddMultWithMarkers(*integrators[i], localX, elem_markers[i], elem_attributes,
+         AddMultWithMarkers(*integrators[i], localX, elem_markers[i], *elem_attributes,
                             true, localY);
       }
       elem_restrict->MultTranspose(localY, y);
@@ -774,13 +734,14 @@ void PABilinearFormExtension::MultTranspose(const Vector &x, Vector &y) const
          bdr_face_Y = 0.0;
          for (int i = 0; i < n_bdr_integs; ++i)
          {
-            AddMultWithMarkers(*bdr_integs[i], bdr_face_X, bdr_markers[i], bdr_attributes,
-                               true, bdr_face_Y);
+            AddMultWithMarkers(*bdr_integs[i], bdr_face_X, bdr_markers[i],
+                               *bdr_face_attributes, true, bdr_face_Y);
          }
          for (int i = 0; i < n_bdr_face_integs; ++i)
          {
-            AddMultWithMarkers(*bdr_face_integs[i], bdr_face_X, bdr_face_markers[i],
-                               bdr_attributes, true, bdr_face_Y);
+            AddMultWithMarkers(*bdr_face_integs[i], bdr_face_X,
+                               bdr_face_markers[i], *bdr_face_attributes, true,
+                               bdr_face_Y);
          }
          bdr_face_restrict_lex->AddMultTransposeInPlace(bdr_face_Y, y);
       }
@@ -804,7 +765,7 @@ static void AddWithMarkers_(
    mfem::forall(ne, [=] MFEM_HOST_DEVICE (int e)
    {
       const int attr = d_attr[e];
-      if (d_m[attr - 1] == 0) { return; }
+      if (attr <= 0 || d_m[attr - 1] == 0) { return; }
       for (int i = 0; i < nd; ++i)
       {
          d_y(i, e) += d_x(i, e);
@@ -920,7 +881,8 @@ void EABilinearFormExtension::Assemble()
       {
          const int i = idx % sz;
          const int e = idx / sz;
-         const real_t val = d_m[d_a[e] - 1] ? d_ea_1(i, e) : 0.0;
+         const real_t val =
+            d_a[e] > 0 ? (d_m[d_a[e] - 1] ? d_ea_1(i, e) : 0) : 0;
          if (add)
          {
             d_ea_2(i, e) += val;
@@ -953,7 +915,7 @@ void EABilinearFormExtension::Assemble()
             ea_data_tmp.SetSize(ea_data.Size());
             integrators[i]->AssembleEA(*a->FESpace(), ea_data_tmp, false);
             add_with_markers(ea_data_tmp, ea_data, ne, *markers,
-                             elem_attributes, add);
+                             *elem_attributes, add);
          }
       }
    }
@@ -982,7 +944,7 @@ void EABilinearFormExtension::Assemble()
             ea_data_tmp.SetSize(ea_data_bdr.Size());
             bdr_integs[i]->AssembleEABoundary(*a->FESpace(), ea_data_tmp, add);
             add_with_markers(ea_data_tmp, ea_data_bdr, nf_bdr, *markers,
-                             bdr_attributes, add);
+                             *bdr_face_attributes, add);
          }
       }
    }
@@ -1031,7 +993,7 @@ void EABilinearFormExtension::Assemble()
                                                         ea_data_tmp,
                                                         add);
             add_with_markers(ea_data_tmp, ea_data_bdr, nf_bdr, *markers,
-                             bdr_attributes, add);
+                             *bdr_face_attributes, add);
          }
       }
    }

fem/bilinearform_ext.hpp

@@ -69,7 +69,8 @@ class PABilinearFormExtension : public BilinearFormExtension
 protected:
    const FiniteElementSpace *trial_fes, *test_fes; // Not owned
    /// Attributes of all mesh elements.
-   Array<int> elem_attributes, bdr_attributes;
+   const Array<int> *elem_attributes; // Not owned
+   const Array<int> *bdr_face_attributes; // Not owned
    mutable Vector tmp_evec; // Work array
    mutable Vector localX, localY;
    mutable Vector int_face_X, int_face_Y;

fem/bilininteg.cpp

@@ -21,6 +21,11 @@ using namespace std;
 namespace mfem
 {
 
+void BilinearFormIntegrator::Update()
+{
+   // default no-op
+}
+
 void BilinearFormIntegrator::AssemblePA(const FiniteElementSpace&)
 {
    MFEM_ABORT("BilinearFormIntegrator::AssemblePA(fes)\n"
@@ -3460,6 +3465,12 @@ real_t ElasticityIntegrator::ComputeFluxEnergy(const FiniteElement &fluxelem,
    return energy;
 }
 
+void DGTraceIntegrator::Update()
+{
+   qspace[0].reset();
+   qspace[1].reset();
+}
+
 void DGTraceIntegrator::AssembleFaceMatrix(const FiniteElement &el1,
                                            const FiniteElement &el2,
                                            FaceElementTransformations &Trans,

fem/bilininteg.hpp

@@ -23,6 +23,8 @@
 
 namespace mfem
 {
+class QuadratureSpace;
+class FaceQuadratureSpace;
 
 /// Abstract base class BilinearFormIntegrator
 class BilinearFormIntegrator : public NonlinearFormIntegrator
@@ -44,6 +46,10 @@ public:
    // make sense for the action of the nonlinear operator (but they all make
    // sense for its Jacobian).
 
+   /// Signal this integrator that something about either the trial or test space has changed.
+   virtual void Update();
+
+
    /// Method defining partial assembly.
    /** The result of the partial assembly is stored internally so that it can be
        used later in the methods AddMultPA() and AddMultTransposePA(). */
@@ -812,7 +818,7 @@ protected:
       const FiniteElement & test_fe) const
    {
       return (trial_fe.GetDim() == 1 && test_fe.GetDim() == 1 &&
-              trial_fe.GetDerivType() == mfem::FiniteElement::GRAD  &&
+              trial_fe.GetDerivType() == mfem::FiniteElement::GRAD &&
               test_fe.GetRangeType()  == mfem::FiniteElement::SCALAR );
    }
 
@@ -884,7 +890,7 @@ protected:
       const FiniteElement & trial_fe,
       const FiniteElement & test_fe) const
    {
-      return (trial_fe.GetDerivType() == mfem::FiniteElement::DIV  &&
+      return (trial_fe.GetDerivType() == mfem::FiniteElement::DIV &&
               test_fe.GetRangeType()  == mfem::FiniteElement::SCALAR );
    }
 
@@ -919,7 +925,7 @@ protected:
       const FiniteElement & trial_fe,
       const FiniteElement & test_fe) const
    {
-      return (trial_fe.GetDerivType() == mfem::FiniteElement::DIV  &&
+      return (trial_fe.GetDerivType() == mfem::FiniteElement::DIV &&
               test_fe.GetRangeType()  == mfem::FiniteElement::VECTOR );
    }
 
@@ -1600,7 +1606,7 @@ public:
    {
       return (trial_fe.GetCurlDim() == 3 && test_fe.GetRangeDim() == 3 &&
               trial_fe.GetRangeType() == mfem::FiniteElement::VECTOR &&
-              trial_fe.GetDerivType() == mfem::FiniteElement::CURL   &&
+              trial_fe.GetDerivType() == mfem::FiniteElement::CURL  &&
               test_fe.GetRangeType()  == mfem::FiniteElement::VECTOR );
    }
 
@@ -1635,7 +1641,7 @@ public:
    {
       return (trial_fe.GetDim() == 2 && test_fe.GetDim() == 2 &&
               trial_fe.GetRangeType() == mfem::FiniteElement::VECTOR &&
-              trial_fe.GetDerivType() == mfem::FiniteElement::CURL   &&
+              trial_fe.GetDerivType() == mfem::FiniteElement::CURL  &&
               test_fe.GetRangeType()  == mfem::FiniteElement::VECTOR );
    }
 
@@ -1669,7 +1675,7 @@ public:
    {
       return (trial_fe.GetDim() == 2 && test_fe.GetDim() == 2 &&
               trial_fe.GetRangeType() == mfem::FiniteElement::SCALAR &&
-              trial_fe.GetDerivType() == mfem::FiniteElement::GRAD   &&
+              trial_fe.GetDerivType() == mfem::FiniteElement::GRAD  &&
               test_fe.GetRangeType()  == mfem::FiniteElement::SCALAR );
    }
 
@@ -1760,7 +1766,7 @@ public:
       const FiniteElement & test_fe) const
    {
       return (trial_fe.GetRangeType() == mfem::FiniteElement::SCALAR &&
-              trial_fe.GetDerivType() == mfem::FiniteElement::GRAD   &&
+              trial_fe.GetDerivType() == mfem::FiniteElement::GRAD  &&
               test_fe.GetRangeType()  == mfem::FiniteElement::SCALAR );
    }
 
@@ -1793,7 +1799,7 @@ public:
       const FiniteElement & test_fe) const
    {
       return (trial_fe.GetRangeType() == mfem::FiniteElement::SCALAR &&
-              trial_fe.GetDerivType() == mfem::FiniteElement::GRAD   &&
+              trial_fe.GetDerivType() == mfem::FiniteElement::GRAD  &&
               test_fe.GetRangeType()  == mfem::FiniteElement::VECTOR &&
               test_fe.GetDerivType()  == mfem::FiniteElement::DIV   );
    }
@@ -1832,7 +1838,7 @@ public:
       const FiniteElement & test_fe) const
    {
       return (trial_fe.GetRangeType() == mfem::FiniteElement::VECTOR &&
-              trial_fe.GetDerivType() == mfem::FiniteElement::DIV    &&
+              trial_fe.GetDerivType() == mfem::FiniteElement::DIV   &&
               test_fe.GetRangeType()  == mfem::FiniteElement::SCALAR &&
               test_fe.GetDerivType()  == mfem::FiniteElement::GRAD
              );
@@ -1973,7 +1979,7 @@ protected:
       const FiniteElement & test_fe) const override
    {
       return (trial_fe.GetCurlDim() == 3 && test_fe.GetRangeDim() == 3 &&
-              trial_fe.GetDerivType() == mfem::FiniteElement::CURL  &&
+              trial_fe.GetDerivType() == mfem::FiniteElement::CURL &&
               test_fe.GetRangeType()  == mfem::FiniteElement::VECTOR );
    }
 
@@ -3311,6 +3317,7 @@ protected:
    VectorCoefficient *u;
    real_t alpha, beta;
    // PA extension
+   std::unique_ptr<FaceQuadratureSpace> qspace[2];
    Vector pa_data;
    const DofToQuad *maps;             ///< Not owned
    const FaceGeometricFactors *geom;  ///< Not owned
@@ -3333,6 +3340,8 @@ public:
                      real_t a, real_t b)
    { rho = &rho_; u = &u_; alpha = a; beta = b; }
 
+   void Update() override;
+
    using BilinearFormIntegrator::AssembleFaceMatrix;
    void AssembleFaceMatrix(const FiniteElement &el1,
                            const FiniteElement &el2,

fem/conduitdatacollection.cpp

@@ -912,7 +912,7 @@ ConduitDataCollection::GridFunctionToBlueprintField(mfem::GridFunction *gf,
 
    if (vdim == 1) // scalar case
    {
-      n_field["values"].set_external(gf->GetData(),
+      n_field["values"].set_external(const_cast<real_t *>(gf->HostRead()),
                                      ndofs);
    }
    else // vector case
@@ -925,18 +925,18 @@ ConduitDataCollection::GridFunctionToBlueprintField(mfem::GridFunction *gf,
       int vdim_stride  = (ordering == Ordering::byNODES ? ndofs : 1);
 
       index_t offset = 0;
-      index_t stride = sizeof(double) * entry_stride;
+      index_t stride = sizeof(real_t) * entry_stride;
 
       for (int d = 0;  d < vdim; d++)
       {
          std::ostringstream oss;
          oss << "v" << d;
          std::string comp_name = oss.str();
-         n_field["values"][comp_name].set_external(gf->GetData(),
+         n_field["values"][comp_name].set_external(const_cast<real_t *>(gf->HostRead()),
                                                    ndofs,
                                                    offset,
                                                    stride);
-         offset +=  sizeof(double) * vdim_stride;
+         offset +=  sizeof(real_t) * vdim_stride;
       }
    }
 

fem/derefmat_op.cpp

@@ -0,0 +1,266 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#include "derefmat_op.hpp"
+#include "fes_kernels.hpp"
+
+/// \cond DO_NOT_DOCUMENT
+namespace mfem
+{
+
+namespace internal
+{
+template <Ordering::Type Order, bool Atomic>
+static void DerefMultKernelImpl(const DerefineMatrixOp &op, const Vector &x,
+                                Vector &y)
+{
+   DerefineMatrixOpMultFunctor<Order, Atomic> func;
+   func.xptr = x.Read();
+   y.UseDevice();
+   y = 0.;
+   func.yptr = y.ReadWrite();
+   func.bsptr = op.block_storage.Read();
+   func.boptr = op.block_offsets.Read();
+   func.brptr = op.block_row_idcs_offsets.Read();
+   func.bcptr = op.block_col_idcs_offsets.Read();
+   func.rptr = op.row_idcs.Read();
+   func.cptr = op.col_idcs.Read();
+   func.vdims = op.fespace->GetVDim();
+   func.nblocks = op.block_offsets.Size();
+   func.width = op.Width() / func.vdims;
+   func.height = op.Height() / func.vdims;
+   func.Run(op.max_rows);
+}
+
+} // namespace internal
+
+DerefineMatrixOp::DerefineMatrixOp(FiniteElementSpace &fespace_, int old_ndofs,
+                                   const Table *old_elem_dof,
+                                   const Table *old_elem_fos)
+   : Operator(fespace_.GetVSize(), old_ndofs * fespace_.GetVDim()),
+     fespace(&fespace_)
+{
+   static Kernels kernels;
+   constexpr int max_team_size = 256;
+   /// TODO: Implement DofTransformation support
+
+   MFEM_VERIFY(fespace->Nonconforming(),
+               "Not implemented for conforming meshes.");
+   MFEM_VERIFY(old_ndofs, "Missing previous (finer) space.");
+   MFEM_VERIFY(fespace->GetNDofs() <= old_ndofs,
+               "Previous space is not finer.");
+
+   const CoarseFineTransformations &dtrans =
+      fespace->GetMesh()->ncmesh->GetDerefinementTransforms();
+
+   MFEM_ASSERT(dtrans.embeddings.Size() == old_elem_dof->Size(), "");
+
+   const bool is_dg = fespace->FEColl()->GetContType()
+                      == FiniteElementCollection::DISCONTINUOUS;
+   DenseMatrix localRVO; // for variable-order only
+
+   DenseTensor localR[Geometry::NumGeom];
+   int total_rows = 0;
+   int total_cols = 0;
+   block_offsets.SetSize(dtrans.embeddings.Size());
+   block_offsets.HostWrite();
+   if (fespace->IsVariableOrder())
+   {
+      // TODO: any potential for some compression here?
+      // determine storage size and offsets
+      block_offsets[0] = 0;
+      int total_size = 0;
+      for (int k = 0; k < dtrans.embeddings.Size(); ++k)
+      {
+         const Embedding &emb = dtrans.embeddings[k];
+         const FiniteElement *fe = fespace->GetFE(emb.parent);
+         const int ldof = fe->GetDof();
+         if (k + 1 < dtrans.embeddings.Size())
+         {
+            block_offsets[k + 1] = block_offsets[k] + ldof * ldof;
+         }
+         total_rows += ldof;
+         total_cols += ldof;
+         total_size += ldof * ldof;
+      }
+      block_storage.SetSize(total_size);
+   }
+   else
+   {
+      // compression scheme:
+      // block_offsets is the start of each block, potentially repeated
+      // only need to store localR for used shapes
+      Mesh::GeometryList elem_geoms(*fespace->GetMesh());
+
+      int geom_offsets[Geometry::NumGeom];
+      {
+         int size = 0;
+         for (int i = 0; i < elem_geoms.Size(); ++i)
+         {
+            fespace->GetLocalDerefinementMatrices(elem_geoms[i],
+                                                  localR[elem_geoms[i]]);
+            geom_offsets[elem_geoms[i]] = size;
+            size += localR[elem_geoms[i]].TotalSize();
+         }
+         block_storage.SetSize(size);
+         // copy blocks into block_storage
+         auto bs_ptr = block_storage.HostWrite();
+         for (int i = 0; i < elem_geoms.Size(); ++i)
+         {
+            std::copy(localR[elem_geoms[i]].Data(),
+                      localR[elem_geoms[i]].Data()
+                      + localR[elem_geoms[i]].TotalSize(),
+                      bs_ptr);
+            bs_ptr += localR[elem_geoms[i]].TotalSize();
+         }
+      }
+      for (int k = 0; k < dtrans.embeddings.Size(); ++k)
+      {
+         const Embedding &emb = dtrans.embeddings[k];
+         Geometry::Type geom =
+            fespace->GetMesh()->GetElementBaseGeometry(emb.parent);
+
+         auto size = localR[geom].SizeI() * localR[geom].SizeJ();
+         total_rows += localR[geom].SizeI();
+         total_cols += localR[geom].SizeJ();
+         // set block offsets and sizes
+         block_offsets[k] = geom_offsets[geom] + size * emb.matrix;
+      }
+   }
+   row_idcs.SetSize(total_rows);
+   row_idcs.HostWrite();
+   col_idcs.SetSize(total_cols);
+   col_idcs.HostWrite();
+   block_row_idcs_offsets.SetSize(dtrans.embeddings.Size() + 1);
+   block_row_idcs_offsets.HostWrite();
+   block_col_idcs_offsets.SetSize(dtrans.embeddings.Size() + 1);
+   block_col_idcs_offsets.HostWrite();
+   block_row_idcs_offsets[0] = 0;
+   block_col_idcs_offsets[0] = 0;
+
+   // compute index information
+   Array<int> dofs, old_dofs;
+   max_rows = 1;
+
+   {
+      Array<int> mark(fespace->GetNDofs());
+      mark = 0;
+      auto bs_ptr = block_storage.HostWrite();
+      int ridx = 0;
+      int cidx = 0;
+      int num_marked = 0;
+      for (int k = 0; k < dtrans.embeddings.Size(); k++)
+      {
+         const Embedding &emb = dtrans.embeddings[k];
+         Geometry::Type geom =
+            fespace->GetMesh()->GetElementBaseGeometry(emb.parent);
+
+         if (fespace->IsVariableOrder())
+         {
+            const FiniteElement *fe = fespace->GetFE(emb.parent);
+            const DenseTensor &pmats = dtrans.point_matrices[geom];
+            const int ldof = fe->GetDof();
+
+            IsoparametricTransformation isotr;
+            isotr.SetIdentityTransformation(geom);
+
+            localRVO.SetSize(ldof, ldof);
+            isotr.SetPointMat(pmats(emb.matrix));
+            // Local restriction is size ldofxldof assuming that the parent
+            // and child are of same polynomial order.
+            fe->GetLocalRestriction(isotr, localRVO);
+            // copy block
+            auto size = localRVO.Height() * localRVO.Width();
+            std::copy(localRVO.Data(), localRVO.Data() + size, bs_ptr);
+            bs_ptr += size;
+         }
+         DenseMatrix &lR =
+            fespace->IsVariableOrder() ? localRVO : localR[geom](emb.matrix);
+         block_row_idcs_offsets[k + 1] =
+            block_row_idcs_offsets[k] + lR.Height();
+         block_col_idcs_offsets[k + 1] = block_col_idcs_offsets[k] + lR.Width();
+         max_rows = std::max(lR.Height(), max_rows);
+         // index information
+         fespace->elem_dof->GetRow(emb.parent, dofs);
+         old_elem_dof->GetRow(k, old_dofs);
+         MFEM_VERIFY(old_dofs.Size() == dofs.Size(),
+                     "Parent and child must have same #dofs.");
+         for (int i = 0; i < lR.Height(); ++i, ++ridx)
+         {
+            if (!std::isfinite(lR(i, 0)))
+            {
+               row_idcs[ridx] = INT_MAX;
+               continue;
+            }
+            int r = dofs[i];
+            int m = (r >= 0) ? r : (-1 - r);
+            if (is_dg || !mark[m])
+            {
+               row_idcs[ridx] = r;
+               mark[m] = 1;
+               ++num_marked;
+            }
+            else
+            {
+               row_idcs[ridx] = INT_MAX;
+            }
+         }
+         for (int i = 0; i < lR.Width(); ++i, ++cidx)
+         {
+            col_idcs[cidx] = old_dofs[i];
+         }
+      }
+      if (!is_dg && !fespace->IsVariableOrder())
+      {
+         MFEM_VERIFY(num_marked * fespace->GetVDim() == Height(),
+                     "internal error: not all rows were set.");
+      }
+   }
+   // if not using GPU, set max_rows/max_cols to zero
+   if (Device::Allows(Backend::DEVICE_MASK))
+   {
+      max_rows = std::min(max_rows, max_team_size);
+   }
+   else
+   {
+      max_rows = 1;
+   }
+}
+
+void DerefineMatrixOp::Mult(const Vector &x, Vector &y) const
+{
+   const bool is_dg = fespace->FEColl()->GetContType()
+                      == FiniteElementCollection::DISCONTINUOUS;
+   // DG needs atomic summation
+   MultKernel::Run(fespace->GetOrdering(), is_dg, *this, x, y);
+}
+
+DerefineMatrixOp::Kernels::Kernels()
+{
+   MultKernel::Specialization<Ordering::byNODES, false>::Add();
+   MultKernel::Specialization<Ordering::byVDIM, false>::Add();
+   MultKernel::Specialization<Ordering::byNODES, true>::Add();
+   MultKernel::Specialization<Ordering::byVDIM, true>::Add();
+}
+
+template <Ordering::Type Order, bool Atomic>
+DerefineMatrixOp::MultKernelType DerefineMatrixOp::MultKernel::Kernel()
+{
+   return internal::DerefMultKernelImpl<Order, Atomic>;
+}
+
+DerefineMatrixOp::MultKernelType
+DerefineMatrixOp::MultKernel::Fallback(Ordering::Type, bool)
+{
+   MFEM_ABORT("invalid MultKernel parameters");
+}
+} // namespace mfem
+/// \endcond DO_NOT_DOCUMENT

fem/derefmat_op.hpp

@@ -0,0 +1,65 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#ifndef MFEM_DEREFMAT_OP
+#define MFEM_DEREFMAT_OP
+
+#include "fespace.hpp"
+
+#include "kernel_dispatch.hpp"
+
+namespace mfem
+{
+
+/// \cond DO_NOT_DOCUMENT
+
+struct DerefineMatrixOp : public Operator
+{
+   FiniteElementSpace *fespace;
+   /// offsets into block_storage
+   Array<int> block_offsets;
+   /// offsets into row_idcs
+   Array<int> block_row_idcs_offsets;
+   /// offsets into col_idcs
+   Array<int> block_col_idcs_offsets;
+   /// mapping for row dofs, INT_MAX indicates the block row should be ignored.
+   /// negative means the row data should be negated.
+   Array<int> row_idcs;
+   /// mapping for col dofs, negative means the col data should be negated.
+   Array<int> col_idcs;
+   /// dense block matrices which can be reused to construct the full matrix
+   /// operation. These are stored contiguously and blocks have no restrictions
+   /// on shape (can be rectangle and differ from block to block).
+   Vector block_storage;
+   /// maximum height of any block in block_storage for GPU
+   /// parallelization, or 1 for CPU runs.
+   int max_rows;
+
+   using MultKernelType = void (*)(const DerefineMatrixOp &, const Vector &,
+                                   Vector &);
+   /// template args: ordering, atomic
+   MFEM_REGISTER_KERNELS(MultKernel, MultKernelType, (Ordering::Type, bool));
+
+   struct Kernels
+   {
+      Kernels();
+   };
+
+   void Mult(const Vector &x, Vector &y) const;
+
+   DerefineMatrixOp(FiniteElementSpace &fespace_, int old_ndofs,
+                    const Table *old_elem_dof, const Table *old_elem_fos);
+};
+
+/// \endcond DO_NOT_DOCUMENT
+
+} // namespace mfem
+#endif

fem/dfem/util.hpp

@@ -327,8 +327,8 @@ void print_mpi_sync(const std::string& msg)
    // First gather string lengths
    size_t msg_len = msg.length();
    std::vector<size_t> lengths(nranks);
-   MPI_Gather(&msg_len, 1, MPI_INT,
-              lengths.data(), 1, MPI_INT,
+   MPI_Gather(&msg_len, 1, MPITypeMap<size_t>::mpi_type,
+              lengths.data(), 1, MPITypeMap<size_t>::mpi_type,
               0, MPI_COMM_WORLD);
 
    if (myrank == 0)
@@ -568,7 +568,7 @@ struct ThreadBlocks
    int z = 1;
 };
 
-#if (defined(MFEM_USE_CUDA) || defined(MFEM_USE_HIP))
+#if defined(MFEM_USE_CUDA_OR_HIP)
 template <typename func_t>
 __global__ void forall_kernel_shmem(func_t f, int n)
 {
@@ -591,7 +591,7 @@ void forall(func_t f,
    if (Device::Allows(Backend::CUDA_MASK) ||
        Device::Allows(Backend::HIP_MASK))
    {
-#if (defined(MFEM_USE_CUDA) || defined(MFEM_USE_HIP))
+#if defined(MFEM_USE_CUDA_OR_HIP)
       // int gridsize = (N + Z - 1) / Z;
       int num_bytes = num_shmem * sizeof(decltype(shmem));
       dim3 block_size(blocks.x, blocks.y, blocks.z);

fem/fes_kernels.hpp

@@ -0,0 +1,249 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#ifndef MFEM_FES_KERNELS_HPP
+#define MFEM_FES_KERNELS_HPP
+
+#include "../general/forall.hpp"
+
+#include <climits>
+
+namespace mfem
+{
+/// \cond DO_NOT_DOCUMENT
+namespace internal
+{
+
+///
+/// Implements matrix-vector multiply $y = A x$ for a sparse matrix composed of
+/// a sum of smaller dense blocks. There is additional permutation/sign
+/// information associated with each block. The base class only implements
+/// helper routines such as computing block widths, index into x, index into y,
+/// and column in A given sub-block information.
+/// @sa DerefineMatrixOpMultFunctor
+///
+/// @tparam Order vdim ordering for x and y. Note that for Diag = false this is
+/// ignored for x as x has a special interleaved order.
+/// @tparam Base used for the curious recurring template pattern (CRTP) so the
+/// base class can access child class fields without virtual functions
+/// @tparam Diag true if this corresponds to the diagonal block (coarse element
+/// and fine element are on our rank), false otherwise (coarse element is on our
+/// rank, fine element is on a different rank).
+///
+template <Ordering::Type Order, class Base, bool Diag = true>
+struct DerefineMatrixOpFunctorBase;
+
+template <class Base>
+struct DerefineMatrixOpFunctorBase<Ordering::byNODES, Base, true>
+{
+   /// block column indices offsets
+   const int *bcptr;
+   /// column indices
+   const int *cptr;
+
+   int MFEM_HOST_DEVICE BlockWidth(int k) const
+   {
+      return bcptr[k + 1] - bcptr[k];
+   }
+
+   void MFEM_HOST_DEVICE Col(int j, int k, int &col, int &sign) const
+   {
+      col = cptr[bcptr[k] + j];
+      if (col < 0)
+      {
+         col = -1 - col;
+         sign = -sign;
+      }
+   }
+
+   int MFEM_HOST_DEVICE IndexX(int col, int vdim, int) const
+   {
+      return col + vdim * static_cast<const Base *>(this)->width;
+   }
+   int MFEM_HOST_DEVICE IndexY(int row, int vdim) const
+   {
+      return row + vdim * static_cast<const Base *>(this)->height;
+   }
+};
+
+template <class Base>
+struct DerefineMatrixOpFunctorBase<Ordering::byVDIM, Base, true>
+{
+   /// block column indices offsets
+   const int *bcptr;
+   /// column indices
+   const int *cptr;
+
+   int MFEM_HOST_DEVICE BlockWidth(int k) const
+   {
+      return bcptr[k + 1] - bcptr[k];
+   }
+
+   void MFEM_HOST_DEVICE Col(int j, int k, int &col, int &sign) const
+   {
+      col = cptr[bcptr[k] + j];
+      if (col < 0)
+      {
+         col = -1 - col;
+         sign = -sign;
+      }
+   }
+
+   int MFEM_HOST_DEVICE IndexX(int col, int vdim, int) const
+   {
+      return vdim + col * static_cast<const Base *>(this)->vdims;
+   }
+   int MFEM_HOST_DEVICE IndexY(int row, int vdim) const
+   {
+      return vdim + row * static_cast<const Base *>(this)->vdims;
+   }
+};
+
+template <class Base>
+struct DerefineMatrixOpFunctorBase<Ordering::byNODES, Base, false>
+{
+   /// receive segment offsets
+   const int *segptr;
+   /// receive segment index
+   const int *rsptr;
+   /// off-diagonal block column offsets
+   const int *coptr;
+   /// off-diagonal block widths
+   const int *bwptr;
+
+   int MFEM_HOST_DEVICE BlockWidth(int k) const { return bwptr[k]; }
+
+   void MFEM_HOST_DEVICE Col(int j, int k, int &col, int &sign) const
+   {
+      col = coptr[k] + j;
+   }
+
+   int MFEM_HOST_DEVICE IndexX(int col, int vdim, int k) const
+   {
+      int tmp = rsptr[k];
+      int segwidth = segptr[tmp + 1] - segptr[tmp];
+      return segptr[tmp] * static_cast<const Base *>(this)->vdims + col +
+             vdim * segwidth;
+   }
+   int MFEM_HOST_DEVICE IndexY(int row, int vdim) const
+   {
+      return row + vdim * static_cast<const Base *>(this)->height;
+   }
+};
+
+template <class Base>
+struct DerefineMatrixOpFunctorBase<Ordering::byVDIM, Base, false>
+{
+   /// receive segment offsets
+   const int *segptr;
+   /// receive segment index
+   const int *rsptr;
+   /// off-diagonal block column offsets
+   const int *coptr;
+   /// off-diagonal block widths
+   const int *bwptr;
+
+   int MFEM_HOST_DEVICE BlockWidth(int k) const { return bwptr[k]; }
+
+   void MFEM_HOST_DEVICE Col(int j, int k, int &col, int &sign) const
+   {
+      col = coptr[k] + j;
+   }
+
+   int MFEM_HOST_DEVICE IndexX(int col, int vdim, int k) const
+   {
+      int tmp = rsptr[k];
+      int segwidth = segptr[tmp + 1] - segptr[tmp];
+      return segptr[tmp] * static_cast<const Base *>(this)->vdims + col +
+             vdim * segwidth;
+   }
+   int MFEM_HOST_DEVICE IndexY(int row, int vdim) const
+   {
+      return vdim + row * static_cast<const Base *>(this)->vdims;
+   }
+};
+
+/// internally used to implement the derefinement operator Mult diagonal
+/// block
+template <Ordering::Type Order, bool Atomic, bool Diag = true>
+struct DerefineMatrixOpMultFunctor
+   : public DerefineMatrixOpFunctorBase<
+     Order, DerefineMatrixOpMultFunctor<Order, Atomic, Diag>, Diag>
+{
+   const real_t *xptr;
+   real_t *yptr;
+   /// block storage
+   const real_t *bsptr;
+   /// block offsets
+   const int *boptr;
+   /// block row index offsets
+   const int *brptr;
+   /// row indices
+   const int *rptr;
+
+   // number of blocks
+   int nblocks;
+   // number of components
+   int vdims;
+   /// overall operator height (for vdim = 1)
+   int height;
+   /// overall operator width (for vdim = 1)
+   int width;
+   void MFEM_HOST_DEVICE operator()(int kidx) const
+   {
+      int k = kidx % nblocks;
+      int vdim = kidx / nblocks;
+
+      int block_height = brptr[k + 1] - brptr[k];
+      int block_width = this->BlockWidth(k);
+      MFEM_FOREACH_THREAD(i, x, block_height)
+      {
+         int row = rptr[brptr[k] + i];
+         int rsign = 1;
+         if (row < 0)
+         {
+            row = -1 - row;
+            rsign = -1;
+         }
+         if (row < INT_MAX)
+         {
+            // row not marked as unused
+            real_t sum = 0;
+            for (int j = 0; j < block_width; ++j)
+            {
+               int col, sign = rsign;
+               this->Col(j, k, col, sign);
+               sum += sign * bsptr[boptr[k] + i + j * block_height] *
+                      xptr[this->IndexX(col, vdim, k)];
+            }
+#if defined(__CUDA_ARCH__) or defined(__HIP_DEVICE_COMPILE__)
+            if (Atomic)
+            {
+               atomicAdd(yptr + this->IndexY(row, vdim), sum);
+            }
+            else
+#endif
+            {
+               yptr[this->IndexY(row, vdim)] += sum;
+            }
+         }
+      }
+   }
+
+   /// N is the max block row size (doesn't have to be a power of 2)
+   void Run(int N) const { forall_2D(nblocks * vdims, N, 1, *this); }
+};
+
+} // namespace internal
+/// \endcond DO_NOT_DOCUMENT
+} // namespace mfem
+
+#endif

fem/fespace.cpp

@@ -17,6 +17,9 @@
 #include "fem.hpp"
 #include "ceed/interface/util.hpp"
 
+#include "derefmat_op.hpp"
+
+#include <algorithm>
 #include <cmath>
 #include <cstdarg>
 
@@ -24,9 +27,9 @@ using namespace std;
 
 namespace mfem
 {
-
-template <> void Ordering::
-DofsToVDofs<Ordering::byNODES>(int ndofs, int vdim, Array<int> &dofs)
+template <>
+void Ordering::DofsToVDofs<Ordering::byNODES>(int ndofs, int vdim,
+                                              Array<int> &dofs)
 {
    // static method
    int size = dofs.Size();
@@ -40,8 +43,9 @@ DofsToVDofs<Ordering::byNODES>(int ndofs, int vdim, Array<int> &dofs)
    }
 }
 
-template <> void Ordering::
-DofsToVDofs<Ordering::byVDIM>(int ndofs, int vdim, Array<int> &dofs)
+template <>
+void Ordering::DofsToVDofs<Ordering::byVDIM>(int ndofs, int vdim,
+                                             Array<int> &dofs)
 {
    // static method
    int size = dofs.Size();
@@ -55,7 +59,6 @@ DofsToVDofs<Ordering::byVDIM>(int ndofs, int vdim, Array<int> &dofs)
    }
 }
 
-
 FiniteElementSpace::FiniteElementSpace()
    : mesh(NULL), fec(NULL), vdim(0), ordering(Ordering::byNODES),
      ndofs(0), nvdofs(0), nedofs(0), nfdofs(0), nbdofs(0),
@@ -4244,7 +4247,11 @@ void FiniteElementSpace::Update(bool want_transform)
          case Mesh::DEREFINE:
          {
             BuildConformingInterpolation();
+#if 0
             Th.Reset(DerefinementMatrix(old_ndofs, old_elem_dof, old_elem_fos));
+#else
+            Th.Reset(new DerefineMatrixOp(*this, old_ndofs, old_elem_dof, old_elem_fos));
+#endif
             if (IsVariableOrder())
             {
                if (cP && cR_hp)

fem/fespace.hpp

@@ -113,7 +113,7 @@ class QuadratureSpace;
 class QuadratureInterpolator;
 class FaceQuadratureInterpolator;
 class PRefinementTransferOperator;
-
+struct DerefineMatrixOp;
 
 /** @brief Class FiniteElementSpace - responsible for providing FEM view of the
     mesh, mainly managing the set of degrees of freedom.
@@ -246,6 +246,7 @@ class FiniteElementSpace
    friend class PRefinementTransferOperator;
    friend void Mesh::Swap(Mesh &, bool);
    friend class LORBase;
+   friend struct DerefineMatrixOp;
 
 protected:
    /// The mesh that FE space lives on (not owned).

fem/integ/bilininteg_dgtrace_pa.cpp

@@ -139,8 +139,6 @@ void DGTraceIntegrator::SetupPA(const FiniteElementSpace &fes, FaceType type)
    const MemoryType mt = (pa_mt == MemoryType::DEFAULT) ?
                          Device::GetDeviceMemoryType() : pa_mt;
 
-   nf = fes.GetNFbyType(type);
-   if (nf==0) { return; }
    // Assumes tensor-product elements
    Mesh *mesh = fes.GetMesh();
    const FiniteElement &el = *fes.GetTypicalTraceElement();
@@ -148,6 +146,17 @@ void DGTraceIntegrator::SetupPA(const FiniteElementSpace &fes, FaceType type)
                                IntRule:
                                &GetRule(el.GetGeomType(), el.GetOrder(),
                                         *mesh->GetTypicalElementTransformation());
+
+   if (!qspace[static_cast<int>(type)])
+   {
+      qspace[static_cast<int>(type)].reset(
+         new FaceQuadratureSpace(*mesh, *ir, type));
+   }
+
+   FaceQuadratureSpace& qs = *qspace[static_cast<int>(type)];
+   nf = qs.GetNumFaces();
+   if (nf==0) { return; }
+
    const int symmDims = 4;
    nq = ir->GetNPoints();
    dim = mesh->Dimension();
@@ -159,8 +168,6 @@ void DGTraceIntegrator::SetupPA(const FiniteElementSpace &fes, FaceType type)
    dofs1D = maps->ndof;
    quad1D = maps->nqpt;
    pa_data.SetSize(symmDims * nq * nf, Device::GetMemoryType());
-
-   FaceQuadratureSpace qs(*mesh, *ir, type);
    CoefficientVector vel(*u, qs, CoefficientStorage::COMPRESSED);
 
    CoefficientVector r(qs, CoefficientStorage::COMPRESSED);

fem/integ/bilininteg_mass_pa.cpp

@@ -59,26 +59,23 @@ void MassIntegrator::AssemblePA(const FiniteElementSpace &fes)
 
    QuadratureSpace qs(*mesh, *ir);
    CoefficientVector coeff(Q, qs, CoefficientStorage::COMPRESSED);
-
-   const int NE = ne;
-   const int Q1D = quad1D;
-   const int NQ = static_cast<int>(std::pow(Q1D, dim));
-   const bool const_c = coeff.Size() == 1;
-   const bool by_val = map_type == FiniteElement::VALUE;
-   const auto W = Reshape(ir->GetWeights().Read(), NQ);
-   const auto J = Reshape(geom->detJ.Read(), NQ, NE);
-   const auto C = const_c ? Reshape(coeff.Read(), 1, 1) :
-                  Reshape(coeff.Read(), NQ,NE);
-   auto v = Reshape(pa_data.Write(), NQ, NE);
-   mfem::forall_2D(NE, NQ, 1, [=] MFEM_HOST_DEVICE (int e)
    {
-      MFEM_FOREACH_THREAD(i, x, NQ)
+      const int NE = ne;
+      const int NQ = nq;
+      const bool const_c = coeff.Size() == 1;
+      const bool by_val = map_type == FiniteElement::VALUE;
+      const auto W = Reshape(ir->GetWeights().Read(), NQ);
+      const auto J = Reshape(geom->detJ.Read(), NQ, NE);
+      const auto C =
+         const_c ? Reshape(coeff.Read(), 1, 1) : Reshape(coeff.Read(), NQ, NE);
+      auto v = Reshape(pa_data.Write(), NQ, NE);
+      mfem::forall(NQ, NE, [=] MFEM_HOST_DEVICE(int q, int e)
       {
-         const real_t detJ = J(i,e);
-         const real_t coeff = const_c ? C(0,0) : C(i,e);
-         v(i,e) =  W(i) * coeff * (by_val ? detJ : 1.0/detJ);
-      }
-   });
+         const real_t detJ = J(q, e);
+         const real_t coeff = const_c ? C(0, 0) : C(q, e);
+         v(q, e) = W(q) * coeff * (by_val ? detJ : 1.0 / detJ);
+      });
+   }
 }
 
 void MassIntegrator::AssemblePABoundary(const FiniteElementSpace &fes)
@@ -109,50 +106,22 @@ void MassIntegrator::AssemblePABoundary(const FiniteElementSpace &fes)
    CoefficientVector coeff(Q, qs, CoefficientStorage::COMPRESSED);
 
    const int NE = ne;
-   const int Q1D = quad1D;
+   const int NQ = nq;
    const bool const_c = coeff.Size() == 1;
    const bool by_val = map_type == FiniteElement::VALUE;
-   if (dim==1)
    {
-      const auto W = Reshape(ir->GetWeights().Read(), Q1D);
-      const auto J = Reshape(face_geom->detJ.Read(), Q1D, NE);
-      const auto C = const_c ? Reshape(coeff.Read(), 1, 1) :
-                     Reshape(coeff.Read(), Q1D, NE);
-      auto v = Reshape(pa_data.Write(), Q1D, NE);
-      mfem::forall_2D(NE, Q1D, 1, [=] MFEM_HOST_DEVICE (int e)
+      const auto W = Reshape(ir->GetWeights().Read(), NQ);
+      const auto J = Reshape(face_geom->detJ.Read(), NQ, NE);
+      const auto C = const_c ? Reshape(coeff.Read(), 1, 1)
+                     : Reshape(coeff.Read(), NQ, NE);
+      auto v = Reshape(pa_data.Write(), NQ, NE);
+      mfem::forall(NQ, NE, [=] MFEM_HOST_DEVICE(int q, int e)
       {
-         MFEM_FOREACH_THREAD(qx,x,Q1D)
-         {
-            const real_t detJ = J(qx,e);
-            const real_t coeff = const_c ? C(0,0) : C(qx,e);
-            v(qx,e) =  W(qx) * coeff * (by_val ? detJ : 1.0/detJ);
-         }
+         const real_t detJ = J(q, e);
+         const real_t coeff = const_c ? C(0, 0) : C(q, e);
+         v(q, e) = W(q) * coeff * (by_val ? detJ : 1.0 / detJ);
       });
    }
-   else if (dim==2)
-   {
-      const auto W = Reshape(ir->GetWeights().Read(), Q1D,Q1D);
-      const auto J = Reshape(face_geom->detJ.Read(), Q1D,Q1D,NE);
-      const auto C = const_c ? Reshape(coeff.Read(), 1,1,1) :
-                     Reshape(coeff.Read(), Q1D,Q1D,NE);
-      auto v = Reshape(pa_data.Write(), Q1D,Q1D, NE);
-      mfem::forall_2D(NE, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
-      {
-         MFEM_FOREACH_THREAD(qx,x,Q1D)
-         {
-            MFEM_FOREACH_THREAD(qy,y,Q1D)
-            {
-               const real_t detJ = J(qx,qy,e);
-               const real_t coeff = const_c ? C(0,0,0) : C(qx,qy,e);
-               v(qx,qy,e) =  W(qx,qy) * coeff * (by_val ? detJ : 1.0/detJ);
-            }
-         }
-      });
-   }
-   else
-   {
-      MFEM_ABORT("Not supported.");
-   }
 }
 
 void MassIntegrator::AssembleDiagonalPA(Vector &diag)

fem/intrules_cut.hpp

@@ -346,13 +346,13 @@ private:
       template<typename T>
       T operator() (const blitz::TinyVector<T,3>& x) const
       {
-         int el_order=el->GetOrder();
-         T u1[el_order+1];
-         T u2[el_order+1];
-         T u3[el_order+1];
-         TmplPoly_1D::CalcBernstein(el_order, x[0], u1);
-         TmplPoly_1D::CalcBernstein(el_order, x[1], u2);
-         TmplPoly_1D::CalcBernstein(el_order, x[2], u3);
+         const int el_order = el->GetOrder();
+         std::vector<T> u1(el_order+1);
+         std::vector<T> u2(el_order+1);
+         std::vector<T> u3(el_order+1);
+         TmplPoly_1D::CalcBernstein(el_order, x[0], u1.data());
+         TmplPoly_1D::CalcBernstein(el_order, x[1], u2.data());
+         TmplPoly_1D::CalcBernstein(el_order, x[2], u3.data());
 
          const Array<int>& dof_map=el->GetDofMap();
 
@@ -370,17 +370,17 @@ private:
       template<typename T>
       blitz::TinyVector<T,3> grad(const blitz::TinyVector<T,3>& x) const
       {
-         int el_order=el->GetOrder();
-         T u1[el_order+1];
-         T u2[el_order+1];
-         T u3[el_order+1];
-         T d1[el_order+1];
-         T d2[el_order+1];
-         T d3[el_order+1];
+         const int el_order = el->GetOrder();
+         std::vector<T> u1(el_order+1);
+         std::vector<T> u2(el_order+1);
+         std::vector<T> u3(el_order+1);
+         std::vector<T> d1(el_order+1);
+         std::vector<T> d2(el_order+1);
+         std::vector<T> d3(el_order+1);
 
-         TmplPoly_1D::CalcBernstein(el_order,x[0], u1, d1);
-         TmplPoly_1D::CalcBernstein(el_order,x[1], u2, d2);
-         TmplPoly_1D::CalcBernstein(el_order,x[2], u3, d3);
+         TmplPoly_1D::CalcBernstein(el_order,x[0], u1.data(), d1.data());
+         TmplPoly_1D::CalcBernstein(el_order,x[1], u2.data(), d2.data());
+         TmplPoly_1D::CalcBernstein(el_order,x[2], u3.data(), d3.data());
 
          blitz::TinyVector<T,3> res(T(0.0),T(0.0),T(0.0));
 
@@ -415,11 +415,11 @@ private:
       template<typename T>
       T operator() (const blitz::TinyVector<T,2>& x) const
       {
-         int el_order=el->GetOrder();
-         T u1[el_order+1];
-         T u2[el_order+1];
-         TmplPoly_1D::CalcBernstein(el_order, x[0], u1);
-         TmplPoly_1D::CalcBernstein(el_order, x[1], u2);
+         const int el_order = el->GetOrder();
+         std::vector<T> u1(el_order+1);
+         std::vector<T> u2(el_order+1);
+         TmplPoly_1D::CalcBernstein(el_order, x[0], u1.data());
+         TmplPoly_1D::CalcBernstein(el_order, x[1], u2.data());
 
          const Array<int>& dof_map=el->GetDofMap();
 
@@ -437,14 +437,14 @@ private:
       template<typename T>
       blitz::TinyVector<T,2> grad(const blitz::TinyVector<T,2>& x) const
       {
-         int el_order=el->GetOrder();
-         T u1[el_order+1];
-         T u2[el_order+1];
-         T d1[el_order+1];
-         T d2[el_order+1];
+         const int el_order = el->GetOrder();
+         std::vector<T> u1(el_order+1);
+         std::vector<T> u2(el_order+1);
+         std::vector<T> d1(el_order+1);
+         std::vector<T> d2(el_order+1);
 
-         TmplPoly_1D::CalcBernstein(el_order,x[0], u1, d1);
-         TmplPoly_1D::CalcBernstein(el_order,x[1], u2, d2);
+         TmplPoly_1D::CalcBernstein(el_order,x[0], u1.data(), d1.data());
+         TmplPoly_1D::CalcBernstein(el_order,x[1], u2.data(), d2.data());
 
          blitz::TinyVector<T,2> res(T(0.0),T(0.0));
 

fem/linearform_ext.cpp

@@ -51,7 +51,7 @@ void LinearFormExtension::Assemble()
       {
          // scan the attributes to set the markers to 0 or 1
          const int NE = fes.GetNE();
-         const auto attr = attributes.Read();
+         const auto attr = attributes->Read();
          const auto dimk = domain_integs_marker_k->Read();
          auto markers_w = markers.Write();
          mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
@@ -93,13 +93,14 @@ void LinearFormExtension::Assemble()
       else
       {
          // scan the attributes to set the markers to 0 or 1
-         const int NBE = bdr_attributes.Size();
-         const auto attr = bdr_attributes.Read();
+         const int NBE = bdr_face_attributes->Size();
+         const auto attr = bdr_face_attributes->Read();
          const auto attr_markers = boundary_integs_marker_k->Read();
          auto markers_w = bdr_markers.Write();
-         mfem::forall(NBE, [=] MFEM_HOST_DEVICE (int e)
+         mfem::forall(NBE, [=] MFEM_HOST_DEVICE(int e)
          {
-            markers_w[e] = attr_markers[attr[e]-1] == 1;
+            markers_w[e] =
+               attr[e] > 0 ? (attr_markers[attr[e] - 1] == 1) : false;
          });
       }
 
@@ -125,8 +126,7 @@ void LinearFormExtension::Update()
       //markers.UseDevice(true);
 
       // Gather the attributes on the host from all the elements
-      attributes.SetSize(NE);
-      for (int i = 0; i < NE; ++i) { attributes[i] = mesh.GetAttribute(i); }
+      attributes = &mesh.GetElementAttributes();
 
       elem_restrict_lex = fes.GetElementRestriction(ordering);
       MFEM_VERIFY(elem_restrict_lex, "Element restriction not available");
@@ -136,34 +136,12 @@ void LinearFormExtension::Update()
 
    if (lf->boundary_integs.Size() > 0)
    {
-      const int nf_bdr = fes.GetNFbyType(FaceType::Boundary);
+      bdr_face_attributes = &mesh.GetBdrFaceAttributes();
+
+      const int nf_bdr = bdr_face_attributes->Size();
       bdr_markers.SetSize(nf_bdr);
       // bdr_markers.UseDevice(true);
 
-      // The face restriction will give us "face E-vectors" on the boundary that
-      // are numbered in the order of the faces of mesh. This numbering will be
-      // different than the numbering of the boundary elements. We compute
-      // mappings so that the array `bdr_attributes[i]` gives the boundary
-      // attribute of the `i`th boundary face in the mesh face order.
-      std::unordered_map<int,int> f_to_be;
-      for (int i = 0; i < mesh.GetNBE(); ++i)
-      {
-         const int f = mesh.GetBdrElementFaceIndex(i);
-         f_to_be[f] = i;
-      }
-      MFEM_VERIFY(size_t(nf_bdr) == f_to_be.size(), "Incompatible sizes");
-      bdr_attributes.SetSize(nf_bdr);
-      int f_ind = 0;
-      for (int f = 0; f < mesh.GetNumFaces(); ++f)
-      {
-         if (f_to_be.find(f) != f_to_be.end())
-         {
-            const int be = f_to_be[f];
-            bdr_attributes[f_ind] = mesh.GetBdrAttribute(be);
-            ++f_ind;
-         }
-      }
-
       bdr_restrict_lex =
          dynamic_cast<const FaceRestriction*>(
             fes.GetFaceRestriction(ordering, FaceType::Boundary,

fem/linearform_ext.hpp

@@ -25,7 +25,8 @@ class LinearForm;
 class LinearFormExtension
 {
    /// Attributes of all mesh elements.
-   Array<int> attributes, bdr_attributes;
+   const Array<int> *attributes; // Not owned
+   const Array<int> *bdr_face_attributes; // Not owned
 
    /// Temporary markers for device kernels.
    Array<int> markers, bdr_markers;

fem/pderefmat_op.cpp

@@ -0,0 +1,591 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#include "pderefmat_op.hpp"
+
+#ifdef MFEM_USE_MPI
+
+#include "fes_kernels.hpp"
+/// \cond DO_NOT_DOCUMENT
+namespace mfem
+{
+namespace internal
+{
+template <Ordering::Type Order, bool Atomic>
+static void ParDerefMultKernelImpl(const ParDerefineMatrixOp &op,
+                                   const Vector &x, Vector &y)
+{
+   // pack sends
+   if (op.xghost_send.Size())
+   {
+      auto src = x.Read();
+      auto idcs = op.send_permutations.Read();
+      auto dst = Device::GetGPUAwareMPI() ? op.xghost_send.Write()
+                 : op.xghost_send.HostWrite();
+      auto vdims = op.fespace->GetVDim();
+      auto sptr = op.send_segment_idcs.Read();
+      auto lptr = op.send_segments.Read();
+      auto old_ndofs = x.Size() / vdims;
+
+      forall(op.send_permutations.Size(), [=] MFEM_HOST_DEVICE(int i)
+      {
+         int seg = sptr[i];
+         int width = lptr[seg + 1] - lptr[seg];
+         auto tdst = dst + i + lptr[seg] * vdims;
+         int sign = 1;
+         int col = idcs[i];
+         if (col < 0)
+         {
+            sign = -1;
+            col = -1 - col;
+         }
+         for (int vdim = 0; vdim < vdims; ++vdim)
+         {
+            tdst[vdim * width] =
+               sign
+               * src[Order == Ordering::byNODES ? (col + vdim * old_ndofs)
+                           : (col * vdims + vdim)];
+         }
+      });
+      // TODO: is this needed so we can send the packed data correctly?
+      // unclear for GPU-aware MPI, definitely required otherwise
+      MFEM_DEVICE_SYNC;
+   }
+   // initialize off-diagonal receive and send
+   op.requests.clear();
+   if (op.xghost_recv.Size())
+   {
+      auto vdims = op.fespace->GetVDim();
+      auto rcv = Device::GetGPUAwareMPI() ? op.xghost_recv.Write()
+                 : op.xghost_recv.HostWrite();
+      for (int i = 0; i < op.recv_ranks.Size(); ++i)
+      {
+         op.requests.emplace_back();
+         MPI_Irecv(rcv + op.recv_segments[i] * vdims,
+                   (op.recv_segments[i + 1] - op.recv_segments[i]) * vdims,
+                   MPITypeMap<real_t>::mpi_type, op.recv_ranks[i],
+                   MessageTag::DEREFINEMENT_MATRIX_CONSTRUCTION_DATA,
+                   op.fespace->GetComm(), &op.requests.back());
+      }
+   }
+   if (op.xghost_send.Size())
+   {
+      auto vdims = op.fespace->GetVDim();
+      // only is a GPU mem ptr if GPU-aware MPI is enabled
+      auto dst = Device::GetGPUAwareMPI() ? op.xghost_send.Write()
+                 : op.xghost_send.HostWrite();
+      for (int i = 0; i < op.send_ranks.Size(); ++i)
+      {
+         op.requests.emplace_back();
+         MPI_Isend(dst + op.send_segments[i] * vdims,
+                   (op.send_segments[i + 1] - op.send_segments[i]) * vdims,
+                   MPITypeMap<real_t>::mpi_type, op.send_ranks[i],
+                   MessageTag::DEREFINEMENT_MATRIX_CONSTRUCTION_DATA,
+                   op.fespace->GetComm(), &op.requests.back());
+      }
+   }
+   {
+      // diagonal
+      DerefineMatrixOpMultFunctor<Order, Atomic, true> func;
+      func.xptr = x.Read();
+      y.UseDevice();
+      y = 0.;
+      func.yptr = y.ReadWrite();
+      func.bsptr = op.block_storage.Read();
+      func.boptr = op.block_offsets.Read();
+      func.brptr = op.block_row_idcs_offsets.Read();
+      func.bcptr = op.block_col_idcs_offsets.Read();
+      func.rptr = op.row_idcs.Read();
+      func.cptr = op.col_idcs.Read();
+      func.vdims = op.fespace->GetVDim();
+      func.nblocks = op.block_offsets.Size();
+      func.width = op.Width() / func.vdims;
+      func.height = op.Height() / func.vdims;
+      func.Run(op.max_rows);
+   }
+   // wait for comm to finish, if any
+   if (op.requests.size())
+   {
+      MPI_Waitall(op.requests.size(), op.requests.data(), MPI_STATUSES_IGNORE);
+      if (op.xghost_recv.Size())
+      {
+         // off-diagonal kernel
+         DerefineMatrixOpMultFunctor<Order, Atomic, false> func;
+         // directly read from host-pinned memory if not using GPU-aware MPI
+         func.xptr = Device::GetGPUAwareMPI() ? op.xghost_recv.Read()
+                     : op.xghost_recv.HostRead();
+         func.yptr = y.ReadWrite();
+         func.bsptr = op.block_storage.Read();
+         func.boptr = op.off_diag_block_offsets.Read();
+         func.brptr = op.block_off_diag_row_idcs_offsets.Read();
+         func.rsptr = op.recv_segment_idcs.Read();
+         func.segptr = op.recv_segments.Read();
+         func.coptr = op.block_off_diag_col_offsets.Read();
+         func.bwptr = op.block_off_diag_widths.Read();
+         func.rptr = op.row_off_diag_idcs.Read();
+         func.vdims = op.fespace->GetVDim();
+         func.nblocks = op.off_diag_block_offsets.Size();
+         func.width = op.xghost_recv.Size() / func.vdims;
+         func.height = op.Height() / func.vdims;
+         func.Run(op.max_rows);
+      }
+   }
+}
+} // namespace internal
+
+template <Ordering::Type Order, bool Atomic>
+ParDerefineMatrixOp::MultKernelType ParDerefineMatrixOp::MultKernel::Kernel()
+{
+   return internal::ParDerefMultKernelImpl<Order, Atomic>;
+}
+
+ParDerefineMatrixOp::MultKernelType
+ParDerefineMatrixOp::MultKernel::Fallback(Ordering::Type, bool)
+{
+   MFEM_ABORT("invalid MultKernel parameters");
+}
+
+ParDerefineMatrixOp::Kernels::Kernels()
+{
+   MultKernel::Specialization<Ordering::byNODES, false>::Add();
+   MultKernel::Specialization<Ordering::byVDIM, false>::Add();
+   MultKernel::Specialization<Ordering::byNODES, true>::Add();
+   MultKernel::Specialization<Ordering::byVDIM, true>::Add();
+}
+
+void ParDerefineMatrixOp::Mult(const Vector &x, Vector &y) const
+{
+   const bool is_dg = fespace->FEColl()->GetContType()
+                      == FiniteElementCollection::DISCONTINUOUS;
+   // DG needs atomic summation
+   MultKernel::Run(fespace->GetOrdering(), is_dg, *this, x, y);
+   // use this to prevent xghost* from being re-purposed for subsequent Mult
+   // calls
+   MFEM_DEVICE_SYNC;
+}
+
+ParDerefineMatrixOp::ParDerefineMatrixOp(ParFiniteElementSpace &fespace_,
+                                         int old_ndofs,
+                                         const Table *old_elem_dof,
+                                         const Table *old_elem_fos)
+   : Operator(fespace_.GetVSize(), old_ndofs * fespace_.GetVDim()),
+     fespace(&fespace_)
+{
+   static Kernels kernels;
+   constexpr int max_team_size = 256;
+
+   const int NRanks = fespace->GetNRanks();
+
+   const int nrk = HYPRE_AssumedPartitionCheck() ? 2 : NRanks;
+
+   MFEM_VERIFY(fespace->Nonconforming(),
+               "Not implemented for conforming meshes.");
+   MFEM_VERIFY(fespace->old_dof_offsets[nrk],
+               "Missing previous (finer) space.");
+
+   const int MyRank = fespace->GetMyRank();
+   ParNCMesh *old_pncmesh = fespace->GetParMesh()->pncmesh;
+   const CoarseFineTransformations &dtrans =
+      old_pncmesh->GetDerefinementTransforms();
+   const Array<int> &old_ranks = old_pncmesh->GetDerefineOldRanks();
+
+   const bool is_dg = fespace->FEColl()->GetContType()
+                      == FiniteElementCollection::DISCONTINUOUS;
+   DenseMatrix localRVO; // for variable-order only
+
+   DenseTensor localR[Geometry::NumGeom];
+   int diag_rows = 0;
+   int off_diag_rows = 0;
+   int diag_cols = 0;
+
+   auto get_ldofs = [&](int k) -> int
+   {
+      const Embedding &emb = dtrans.embeddings[k];
+      if (fespace->IsVariableOrder())
+      {
+         const FiniteElement *fe = fespace->GetFE(emb.parent);
+         return fe->GetDof();
+      }
+      else
+      {
+         Geometry::Type geom =
+         fespace->GetParMesh()->GetElementBaseGeometry(emb.parent);
+         return fespace->FEColl()->FiniteElementForGeometry(geom)->GetDof();
+      }
+   };
+   Array<int> dofs, old_dofs;
+   max_rows = 1;
+   // first pass:
+   // - determine memory block lengths
+   // - identify dofs in x we need to send/receive
+   // don't need to send the indices, fine rank will re-arrange and sign
+   // change x before transmitting the ghost data
+
+   // key: coarse rank to send to
+   // value: old dofs to send (with sign)
+   std::map<int, std::vector<int>> to_send;
+   // key: fine rank
+   // value: indices into dtrans.embeddings
+   std::map<int, std::vector<int>> od_ks;
+   // key: fine rank
+   // value: recv segment length
+   std::map<int, int> od_seg_lens;
+   int send_len = 0;
+   int recv_len = 0;
+   // size of block_storage, if fespace->IsVariableOrder()
+   // otherwise unused
+   int total_size = 0;
+   int num_diagonal_blocks = 0;
+   int num_offdiagonal_blocks = 0;
+   for (int k = 0; k < dtrans.embeddings.Size(); ++k)
+   {
+      const Embedding &emb = dtrans.embeddings[k];
+      int fine_rank = old_ranks[k];
+      int coarse_rank = (emb.parent < 0) ? (-1 - emb.parent)
+                        : old_pncmesh->ElementRank(emb.parent);
+      if (coarse_rank != MyRank && fine_rank == MyRank)
+      {
+         // this rank needs to send data in x to course_rank
+         old_elem_dof->GetRow(k, old_dofs);
+         auto &tmp = to_send[coarse_rank];
+         send_len += old_dofs.Size();
+         for (int i = 0; i < old_dofs.Size(); ++i)
+         {
+            tmp.emplace_back(old_dofs[i]);
+         }
+      }
+      else if (coarse_rank == MyRank && fine_rank != MyRank)
+      {
+         // this rank needs to receive data in x from fine_rank
+         MFEM_ASSERT(emb.parent >= 0, "");
+         auto ldofs = get_ldofs(k);
+         off_diag_rows += ldofs;
+         recv_len += ldofs;
+         od_ks[fine_rank].emplace_back(k);
+         od_seg_lens[fine_rank] += ldofs;
+         ++num_offdiagonal_blocks;
+         if (fespace->IsVariableOrder())
+         {
+            total_size += ldofs * ldofs;
+         }
+      }
+      else if (coarse_rank == MyRank && fine_rank == MyRank)
+      {
+         MFEM_ASSERT(emb.parent >= 0, "");
+         // diagonal
+         ++num_diagonal_blocks;
+         auto ldofs = get_ldofs(k);
+         diag_rows += ldofs;
+         diag_cols += ldofs;
+         if (fespace->IsVariableOrder())
+         {
+            total_size += ldofs * ldofs;
+         }
+      }
+   }
+   send_segments.SetSize(to_send.size() + 1);
+   send_segments.HostWrite();
+   send_ranks.SetSize(to_send.size());
+   send_ranks.HostWrite();
+   {
+      int idx = 0;
+      send_segments[0] = 0;
+      for (auto &tmp : to_send)
+      {
+         send_ranks[idx] = tmp.first;
+         send_segments[idx + 1] = send_segments[idx] + tmp.second.size();
+         ++idx;
+      }
+   }
+   recv_segment_idcs.SetSize(off_diag_rows);
+   recv_segment_idcs.HostWrite();
+   recv_segments.SetSize(od_ks.size() + 1);
+   recv_segments.HostWrite();
+   recv_ranks.SetSize(od_ks.size());
+   recv_ranks.HostWrite();
+
+   // set sizes
+   row_idcs.SetSize(diag_rows);
+   row_idcs.HostWrite();
+   row_off_diag_idcs.SetSize(off_diag_rows);
+   row_off_diag_idcs.HostWrite();
+   col_idcs.SetSize(diag_cols);
+   col_idcs.HostWrite();
+   block_row_idcs_offsets.SetSize(num_diagonal_blocks + 1);
+   block_row_idcs_offsets.HostWrite();
+   block_col_idcs_offsets.SetSize(num_diagonal_blocks + 1);
+   block_col_idcs_offsets.HostWrite();
+   block_off_diag_row_idcs_offsets.SetSize(num_offdiagonal_blocks + 1);
+   block_off_diag_row_idcs_offsets.HostWrite();
+   block_off_diag_col_offsets.SetSize(num_offdiagonal_blocks);
+   block_off_diag_col_offsets.HostWrite();
+   block_off_diag_widths.SetSize(num_offdiagonal_blocks);
+   block_off_diag_widths.HostWrite();
+   pack_col_idcs.SetSize(send_len);
+   // memory manager doesn't appear to have a graceful fallback for
+   // HOST_PINNED if not built with CUDA or HIP
+#if defined(MFEM_USE_CUDA) or defined(MFEM_USE_HIP)
+   xghost_send.SetSize(send_len * fespace->GetVDim(),
+                       Device::GetGPUAwareMPI() ? MemoryType::DEFAULT
+                       : MemoryType::HOST_PINNED);
+   xghost_recv.SetSize(recv_len * fespace->GetVDim(),
+                       Device::GetGPUAwareMPI() ? MemoryType::DEFAULT
+                       : MemoryType::HOST_PINNED);
+#else
+   xghost_send.SetSize(send_len * fespace->GetVDim());
+   xghost_recv.SetSize(recv_len * fespace->GetVDim());
+#endif
+   send_permutations.SetSize(send_len);
+   send_segment_idcs.SetSize(send_len);
+   block_offsets.SetSize(num_diagonal_blocks);
+   block_offsets.HostWrite();
+   off_diag_block_offsets.SetSize(num_offdiagonal_blocks);
+   off_diag_block_offsets.HostWrite();
+   int geom_offsets[Geometry::NumGeom];
+   real_t *bs_ptr;
+
+   if (fespace->IsVariableOrder())
+   {
+      block_storage.SetSize(total_size);
+      bs_ptr = block_storage.HostWrite();
+      // compute block data later
+   }
+   else
+   {
+      // compression scheme:
+      // block_offsets is the start of each block, potentially repeated
+      // only need to store localR for used shapes
+      Mesh::GeometryList elem_geoms(*fespace->GetMesh());
+
+      int size = 0;
+      for (int i = 0; i < elem_geoms.Size(); ++i)
+      {
+         fespace->GetLocalDerefinementMatrices(elem_geoms[i],
+                                               localR[elem_geoms[i]]);
+         geom_offsets[elem_geoms[i]] = size;
+         size += localR[elem_geoms[i]].TotalSize();
+      }
+      block_storage.SetSize(size);
+      bs_ptr = block_storage.HostWrite();
+      // copy blocks into block_storage
+      for (int i = 0; i < elem_geoms.Size(); ++i)
+      {
+         std::copy(localR[elem_geoms[i]].Data(),
+                   localR[elem_geoms[i]].Data()
+                   + localR[elem_geoms[i]].TotalSize(),
+                   bs_ptr);
+         bs_ptr += localR[elem_geoms[i]].TotalSize();
+      }
+   }
+
+   // second pass:
+   // - initialize buffers
+
+   {
+      auto ptr = send_permutations.HostWrite();
+      auto ptr2 = send_segment_idcs.HostWrite();
+      int i = 0;
+      for (auto &v : to_send)
+      {
+         ptr = std::copy(v.second.begin(), v.second.end(), ptr);
+         for (size_t idx = 0; idx < v.second.size(); ++idx)
+         {
+            *ptr2 = i;
+            ++ptr2;
+         }
+         ++i;
+      }
+   }
+
+   block_row_idcs_offsets[0] = 0;
+   block_col_idcs_offsets[0] = 0;
+   block_off_diag_row_idcs_offsets[0] = 0;
+   Array<int> mark(fespace->GetNDofs());
+   mark = 0;
+   {
+      int idx = 0;
+      recv_segments[0] = 0;
+      for (auto &v : od_seg_lens)
+      {
+         recv_ranks[idx] = v.first;
+         recv_segments[idx + 1] = recv_segments[idx] + v.second;
+         ++idx;
+      }
+   }
+   // key: index into dtrans.embeddings
+   // value: off-diagonal block offset, od_ridx, seg id
+   std::unordered_map<int, std::array<int, 3>> ks_map;
+   {
+      int od_ridx = 0;
+      int seg_id = 0;
+      for (auto &v1 : od_ks)
+      {
+         for (auto k : v1.second)
+         {
+            auto &tmp = ks_map[k];
+            tmp[0] = ks_map.size() - 1;
+            tmp[1] = od_ridx;
+            tmp[2] = seg_id;
+            od_ridx += get_ldofs(k);
+         }
+         ++seg_id;
+      }
+   }
+   int diag_idx = 0;
+   int var_offset = 0;
+   int ridx = 0;
+   int cidx = 0;
+   // can't break this up into separate diagonals/off-diagonals loops because
+   // of mark
+   for (int k = 0; k < dtrans.embeddings.Size(); ++k)
+   {
+      const Embedding &emb = dtrans.embeddings[k];
+      if (emb.parent < 0)
+      {
+         continue;
+      }
+      int fine_rank = old_ranks[k];
+      int coarse_rank = (emb.parent < 0) ? (-1 - emb.parent)
+                        : old_pncmesh->ElementRank(emb.parent);
+      if (coarse_rank == MyRank)
+      {
+         // either diagonal or off-diagonal
+         Geometry::Type geom =
+            fespace->GetMesh()->GetElementBaseGeometry(emb.parent);
+         if (fespace->IsVariableOrder())
+         {
+            const FiniteElement *fe = fespace->GetFE(emb.parent);
+            const DenseTensor &pmats = dtrans.point_matrices[geom];
+            const int ldof = fe->GetDof();
+
+            IsoparametricTransformation isotr;
+            isotr.SetIdentityTransformation(geom);
+
+            localRVO.SetSize(ldof, ldof);
+            isotr.SetPointMat(pmats(emb.matrix));
+            // Local restriction is size ldofxldof assuming that the parent
+            // and child are of same polynomial order.
+            fe->GetLocalRestriction(isotr, localRVO);
+            // copy block
+            auto s = localRVO.Height() * localRVO.Width();
+            std::copy(localRVO.Data(), localRVO.Data() + s, bs_ptr);
+            bs_ptr += s;
+         }
+         DenseMatrix &lR =
+            fespace->IsVariableOrder() ? localRVO : localR[geom](emb.matrix);
+         max_rows = std::max(lR.Height(), max_rows);
+         auto size = lR.Height() * lR.Width();
+         fespace->elem_dof->GetRow(emb.parent, dofs);
+         if (fine_rank == MyRank)
+         {
+            // diagonal
+            old_elem_dof->GetRow(k, old_dofs);
+            MFEM_VERIFY(old_dofs.Size() == dofs.Size(),
+                        "Parent and child must have same #dofs.");
+            block_row_idcs_offsets[diag_idx + 1] =
+               block_row_idcs_offsets[diag_idx] + lR.Height();
+            block_col_idcs_offsets[diag_idx + 1] =
+               block_col_idcs_offsets[diag_idx] + lR.Width();
+
+            if (fespace->IsVariableOrder())
+            {
+               block_offsets[diag_idx] = var_offset;
+               var_offset += size;
+            }
+            else
+            {
+               block_offsets[diag_idx] = geom_offsets[geom] + size * emb.matrix;
+            }
+            for (int i = 0; i < lR.Height(); ++i, ++ridx)
+            {
+               if (!std::isfinite(lR(i, 0)))
+               {
+                  row_idcs[ridx] = INT_MAX;
+                  continue;
+               }
+               int r = dofs[i];
+               int m = (r >= 0) ? r : (-1 - r);
+               if (is_dg || !mark[m])
+               {
+                  row_idcs[ridx] = r;
+                  mark[m] = 1;
+               }
+               else
+               {
+                  row_idcs[ridx] = INT_MAX;
+               }
+            }
+            for (int i = 0; i < lR.Width(); ++i, ++cidx)
+            {
+               col_idcs[cidx] = old_dofs[i];
+            }
+            ++diag_idx;
+         }
+         else
+         {
+            // off-diagonal
+            auto &tmp = ks_map.at(k);
+            auto od_idx = tmp[0];
+            auto od_ridx = tmp[1];
+            block_off_diag_row_idcs_offsets[od_idx + 1] =
+               block_off_diag_row_idcs_offsets[od_idx] + lR.Height();
+            block_off_diag_col_offsets[od_idx] = od_ridx;
+            block_off_diag_widths[od_idx] = lR.Width();
+            recv_segment_idcs[od_idx] = tmp[2];
+
+            if (fespace->IsVariableOrder())
+            {
+               off_diag_block_offsets[od_idx] = var_offset;
+               var_offset += size;
+            }
+            else
+            {
+               off_diag_block_offsets[od_idx] =
+                  geom_offsets[geom] + size * emb.matrix;
+            }
+            for (int i = 0; i < lR.Height(); ++i, ++od_ridx)
+            {
+               if (!std::isfinite(lR(i, 0)))
+               {
+                  row_off_diag_idcs[od_ridx] = INT_MAX;
+                  continue;
+               }
+               int r = dofs[i];
+               int m = (r >= 0) ? r : (-1 - r);
+               if (is_dg || !mark[m])
+               {
+                  row_off_diag_idcs[od_ridx] = r;
+                  mark[m] = 1;
+               }
+               else
+               {
+                  row_off_diag_idcs[od_ridx] = INT_MAX;
+               }
+            }
+            ++od_idx;
+         }
+      }
+   }
+   // if not using GPU, set max_rows/max_cols to zero
+   if (Device::Allows(Backend::DEVICE_MASK))
+   {
+      max_rows = std::min(max_rows, max_team_size);
+   }
+   else
+   {
+      max_rows = 1;
+   }
+   requests.reserve(recv_ranks.Size() + send_ranks.Size());
+}
+} // namespace mfem
+/// \endcond DO_NOT_DOCUMENT
+
+#endif

fem/pderefmat_op.hpp

@@ -0,0 +1,111 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#ifndef MFEM_PDEREFMAT_OP
+#define MFEM_PDEREFMAT_OP
+
+#include "../config/config.hpp"
+
+#ifdef MFEM_USE_MPI
+
+#include "pfespace.hpp"
+
+#include "kernel_dispatch.hpp"
+
+#include <vector>
+
+namespace mfem
+{
+
+/// \cond DO_NOT_DOCUMENT
+struct ParDerefineMatrixOp : public Operator
+{
+   ParFiniteElementSpace *fespace;
+   /// offsets into block_storage for diagonal
+   Array<int> block_offsets;
+   /// offsets into row_idcs for diagonal
+   Array<int> block_row_idcs_offsets;
+   /// offsets into col_idcs for diagonal
+   Array<int> block_col_idcs_offsets;
+
+   /// offsets into block_storage for off-diagonal
+   Array<int> off_diag_block_offsets;
+   /// offsets into row_idcs for off-diagonal
+   Array<int> block_off_diag_row_idcs_offsets;
+   Array<int> block_off_diag_col_offsets;
+   Array<int> block_off_diag_widths;
+   /// mapping for row dofs, INT_MAX indicates the block row should be ignored.
+   /// negative means the row data should be negated.
+   /// only for diagonal blocks
+   Array<int> row_idcs;
+   /// mapping for col dofs, negative means the col data should be negated.
+   /// only for diagonal blocks
+   Array<int> col_idcs;
+
+   Array<int> pack_col_idcs;
+
+   /// mapping for row dofs, INT_MAX indicates the block row should be ignored.
+   /// negative means the row data should be negated.
+   /// only for off-diagonal blocks
+   Array<int> row_off_diag_idcs;
+   /// dense block matrices which can be reused to construct the full matrix
+   /// operation. These are stored contiguously and blocks have no restrictions
+   /// on shape (can be rectangle and differ from block to block).
+   /// This is only for the diagonal block.
+   Vector block_storage;
+   /// maximum height of any block in block_storage for GPU
+   /// parallelization, or 1 for CPU runs.
+   int max_rows;
+
+   /// quasi Ordering::byNODES, broken into sections by ranks we need to send
+   /// the data to
+   mutable Vector xghost_send;
+   /// quasi Ordering::byNODES, broken into sections by ranks we received
+   /// the data from
+   mutable Vector xghost_recv;
+   /// maps off-diagonal k to segment
+   Array<int> recv_segment_idcs;
+   /// cumulative count of dofs which will be received from other ranks
+   Array<int> recv_segments;
+   /// Source rank of each recv segment
+   Array<int> recv_ranks;
+   /// What send segment each entry in send_permutations corresponds to
+   Array<int> send_segment_idcs;
+   /// cumulative count of dofs which will be sent to other ranks
+   Array<int> send_segments;
+   /// Destination rank of each send segment
+   Array<int> send_ranks;
+   /// how to permute/sign change values from our local x to send to other ranks
+   Array<int> send_permutations;
+   /// internal buffer for MPI requests
+   mutable std::vector<MPI_Request> requests;
+
+   using MultKernelType = void (*)(const ParDerefineMatrixOp &, const Vector &,
+                                   Vector &);
+   /// template args: ordering, atomic
+   MFEM_REGISTER_KERNELS(MultKernel, MultKernelType, (Ordering::Type, bool));
+
+   struct Kernels
+   {
+      Kernels();
+   };
+
+   void Mult(const Vector &x, Vector &y) const;
+
+   ParDerefineMatrixOp(ParFiniteElementSpace &fespace_, int old_ndofs,
+                       const Table *old_elem_dof, const Table *old_elem_fos);
+};
+/// \endcond DO_NOT_DOCUMENT
+} // namespace mfem
+
+#endif
+
+#endif

fem/pfespace.cpp

@@ -22,12 +22,13 @@
 #include "../mesh/mesh_headers.hpp"
 #include "../general/binaryio.hpp"
 
+#include "pderefmat_op.hpp"
+
 #include <limits>
 #include <list>
 
 namespace mfem
 {
-
 ParFiniteElementSpace::ParFiniteElementSpace(
    const ParFiniteElementSpace &orig, ParMesh *pmesh,
    const FiniteElementCollection *fec)
@@ -4487,13 +4488,6 @@ ParFiniteElementSpace::RebalanceMatrix(int old_ndofs,
    return M;
 }
 
-
-struct DerefDofMessage
-{
-   std::vector<HYPRE_BigInt> dofs;
-   MPI_Request request;
-};
-
 HypreParMatrix*
 ParFiniteElementSpace::ParallelDerefinementMatrix(int old_ndofs,
                                                   const Table* old_elem_dof,
@@ -4536,7 +4530,13 @@ ParFiniteElementSpace::ParallelDerefinementMatrix(int old_ndofs,
       old_pncmesh->GetDerefinementTransforms();
    const Array<int> &old_ranks = old_pncmesh->GetDerefineOldRanks();
 
-   std::map<int, DerefDofMessage> messages;
+   // key: other rank
+   // value: send or recieve buffer
+   std::map<int, std::vector<HYPRE_BigInt>> to_send;
+   std::map<int, std::vector<HYPRE_BigInt>> to_recv;
+   // key: index into dtrans.embeddings
+   // value: [start, stop]
+   std::unordered_map<int, std::array<size_t, 2>> recv_messages;
 
    HYPRE_BigInt old_offset = HYPRE_AssumedPartitionCheck()
                              ? old_dof_offsets[0] : old_dof_offsets[MyRank];
@@ -4556,30 +4556,46 @@ ParFiniteElementSpace::ParallelDerefinementMatrix(int old_ndofs,
          old_elem_dof->GetRow(k, dofs);
          DofsToVDofs(dofs, old_ndofs);
 
-         DerefDofMessage &msg = messages[k];
-         msg.dofs.resize(dofs.Size());
+         std::vector<HYPRE_BigInt>& send_buf = to_send[coarse_rank];
+         auto pos = send_buf.size();
+         send_buf.resize(pos + dofs.Size());
          for (int i = 0; i < dofs.Size(); i++)
          {
-            msg.dofs[i] = old_offset + dofs[i];
+            send_buf[pos + i] = old_offset + dofs[i];
          }
-
-         MPI_Isend(&msg.dofs[0], static_cast<int>(msg.dofs.size()), HYPRE_MPI_BIG_INT,
-                   coarse_rank, 291, MyComm, &msg.request);
       }
       else if (coarse_rank == MyRank && fine_rank != MyRank)
       {
          MFEM_ASSERT(emb.parent >= 0, "");
          Geometry::Type geom = mesh->GetElementBaseGeometry(emb.parent);
 
-         DerefDofMessage &msg = messages[k];
-         msg.dofs.resize(ldof[geom]*vdim);
-
-         MPI_Irecv(&msg.dofs[0], ldof[geom]*vdim, HYPRE_MPI_BIG_INT,
-                   fine_rank, 291, MyComm, &msg.request);
+         std::vector<HYPRE_BigInt>& recv_buf = to_recv[fine_rank];
+         auto& msg = recv_messages[k];
+         msg[0] = recv_buf.size();
+         recv_buf.resize(recv_buf.size() + ldof[geom] * vdim);
+         msg[1] = recv_buf.size();
       }
-      // TODO: coalesce Isends/Irecvs to the same rank. Typically, on uniform
-      // derefinement, there should be just one send to MyRank-1 and one recv
-      // from MyRank+1
+   }
+
+   // assume embedding orders are consistent (i.e. what we expect to receive
+   // first from a given rank is sent first, etc.)
+   std::vector<MPI_Request> requests;
+   requests.reserve(to_send.size() + to_recv.size());
+   // enqueue recvs
+   for (auto &v : to_recv)
+   {
+      requests.emplace_back();
+      MPI_Irecv(v.second.data(), v.second.size(), HYPRE_MPI_BIG_INT, v.first,
+                MessageTag::DEREFINEMENT_MATRIX_CONSTRUCTION_DATA, MyComm,
+                &requests.back());
+   }
+   // enqueue sends
+   for (auto &v : to_send)
+   {
+      requests.emplace_back();
+      MPI_Isend(v.second.data(), v.second.size(), HYPRE_MPI_BIG_INT, v.first,
+                MessageTag::DEREFINEMENT_MATRIX_CONSTRUCTION_DATA, MyComm,
+                &requests.back());
    }
 
    DenseTensor localR[Geometry::NumGeom];
@@ -4637,10 +4653,7 @@ ParFiniteElementSpace::ParallelDerefinementMatrix(int old_ndofs,
    diag->Finalize();
 
    // wait for all sends/receives to complete
-   for (auto it = messages.begin(); it != messages.end(); ++it)
-   {
-      MPI_Wait(&it->second.request, MPI_STATUS_IGNORE);
-   }
+   MPI_Waitall(requests.size(), requests.data(), MPI_STATUSES_IGNORE);
 
    // create the off-diagonal part of the derefinement matrix
    SparseMatrix *offd = new SparseMatrix(ndofs*vdim, 1);
@@ -4661,13 +4674,14 @@ ParFiniteElementSpace::ParallelDerefinementMatrix(int old_ndofs,
 
          elem_dof->GetRow(emb.parent, dofs);
 
-         DerefDofMessage &msg = messages[k];
-         MFEM_ASSERT(msg.dofs.size(), "");
+         auto& odofs = to_recv.at(fine_rank);
+         auto &msg = recv_messages[k];
+         MFEM_ASSERT(msg[1] > msg[0], "");
 
          for (int vd = 0; vd < vdim; vd++)
          {
             MFEM_ASSERT(ldof[geom], "");
-            HYPRE_BigInt* remote_dofs = &msg.dofs[vd*ldof[geom]];
+            HYPRE_BigInt *remote_dofs = odofs.data() + msg[0] + vd * ldof[geom];
 
             for (int i = 0; i < lR.Height(); i++)
             {
@@ -4694,7 +4708,6 @@ ParFiniteElementSpace::ParallelDerefinementMatrix(int old_ndofs,
       }
    }
 
-   messages.clear();
    offd->Finalize(0);
    offd->SetWidth(static_cast<int>(col_map.size()));
 
@@ -4946,8 +4959,13 @@ void ParFiniteElementSpace::Update(bool want_transform)
 
          case Mesh::DEREFINE:
          {
+#if 0
             Th.Reset(ParallelDerefinementMatrix(old_ndofs, old_elem_dof,
                                                 old_elem_fos));
+#else
+            Th.Reset(new ParDerefineMatrixOp(*this, old_ndofs, old_elem_dof,
+                                             old_elem_fos));
+#endif
             if (Nonconforming())
             {
                Th.SetOperatorOwner(false);

fem/pfespace.hpp

@@ -24,9 +24,12 @@
 namespace mfem
 {
 
+struct ParDerefineMatrixOp;
+
 /// Abstract parallel finite element space.
 class ParFiniteElementSpace : public FiniteElementSpace
 {
+   friend struct ParDerefineMatrixOp;
 private:
    /// MPI data.
    MPI_Comm MyComm;

fem/qfunction.hpp

@@ -214,16 +214,16 @@ public:
 inline void QuadratureFunction::GetValues(
    int idx, Vector &values)
 {
-   const int s_offset = qspace->offsets[idx];
-   const int sl_size = qspace->offsets[idx+1] - s_offset;
+   const int s_offset = qspace->Offset(idx);
+   const int sl_size = qspace->Offset(idx + 1) - s_offset;
    values.MakeRef(*this, vdim*s_offset, vdim*sl_size);
 }
 
 inline void QuadratureFunction::GetValues(
    int idx, Vector &values) const
 {
-   const int s_offset = qspace->offsets[idx];
-   const int sl_size = qspace->offsets[idx+1] - s_offset;
+   const int s_offset = qspace->Offset(idx);
+   const int sl_size = qspace->Offset(idx + 1) - s_offset;
    values.SetSize(vdim*sl_size);
    values.HostWrite();
    const real_t *q = HostRead() + vdim*s_offset;
@@ -236,14 +236,14 @@ inline void QuadratureFunction::GetValues(
 inline void QuadratureFunction::GetValues(
    int idx, const int ip_num, Vector &values)
 {
-   const int s_offset = qspace->offsets[idx] * vdim + ip_num * vdim;
+   const int s_offset = qspace->Offset(idx) * vdim + ip_num * vdim;
    values.MakeRef(*this, s_offset, vdim);
 }
 
 inline void QuadratureFunction::GetValues(
    int idx, const int ip_num, Vector &values) const
 {
-   const int s_offset = qspace->offsets[idx] * vdim + ip_num * vdim;
+   const int s_offset = qspace->Offset(idx) * vdim + ip_num * vdim;
    values.SetSize(vdim);
    values.HostWrite();
    const real_t *q = HostRead() + s_offset;
@@ -256,8 +256,8 @@ inline void QuadratureFunction::GetValues(
 inline void QuadratureFunction::GetValues(
    int idx, DenseMatrix &values)
 {
-   const int s_offset = qspace->offsets[idx];
-   const int sl_size = qspace->offsets[idx+1] - s_offset;
+   const int s_offset = qspace->Offset(idx);
+   const int sl_size = qspace->Offset(idx + 1) - s_offset;
    // Make the values matrix memory an alias of the quadrature function memory
    Memory<real_t> &values_mem = values.GetMemory();
    values_mem.Delete();
@@ -268,8 +268,8 @@ inline void QuadratureFunction::GetValues(
 inline void QuadratureFunction::GetValues(
    int idx, DenseMatrix &values) const
 {
-   const int s_offset = qspace->offsets[idx];
-   const int sl_size = qspace->offsets[idx+1] - s_offset;
+   const int s_offset = qspace->Offset(idx);
+   const int sl_size = qspace->Offset(idx + 1) - s_offset;
    values.SetSize(vdim, sl_size);
    values.HostWrite();
    const real_t *q = HostRead() + vdim*s_offset;

fem/qspace.cpp

@@ -17,8 +17,9 @@ namespace mfem
 {
 
 QuadratureSpaceBase::QuadratureSpaceBase(Mesh &mesh_, Geometry::Type geom,
-                                         const IntegrationRule &ir)
-   : mesh(mesh_), order(ir.GetOrder())
+                                         const IntegrationRule &ir,
+                                         QSpaceStorage storage)
+   : mesh(mesh_), order(ir.GetOrder()), storage(storage)
 {
    for (int g = 0; g < Geometry::NumGeom; g++)
    {
@@ -96,27 +97,37 @@ void QuadratureSpaceBase::Integrate(VectorCoefficient &coeff,
 
 void QuadratureSpace::ConstructOffsets()
 {
-   const int num_elem = mesh.GetNE();
-   offsets.SetSize(num_elem + 1);
-   int offset = 0;
-   for (int i = 0; i < num_elem; i++)
+   const int num_elem = ne;
+
+   if (storage == QSpaceStorage::COMPRESSED &&
+       mesh.GetNumGeometries(mesh.Dimension()) == 1)
    {
-      offsets[i] = offset;
-      int geom = mesh.GetElementBaseGeometry(i);
-      MFEM_ASSERT(int_rule[geom] != NULL, "Missing integration rule.");
-      offset += int_rule[geom]->GetNPoints();
+      Array<Geometry::Type> geoms;
+      mesh.GetGeometries(mesh.Dimension(), geoms);
+      offsets.SetSize(1);
+      offsets.HostWrite();
+      offsets[0] = int_rule[geoms[0]]->GetNPoints();
+      size = num_elem * offsets[0];
+   }
+   else
+   {
+      offsets.SetSize(num_elem + 1);
+      int offset = 0;
+      for (int i = 0; i < num_elem; i++)
+      {
+         offsets[i] = offset;
+         const Geometry::Type geom = mesh.GetElementBaseGeometry(i);
+         MFEM_ASSERT(int_rule[geom] != NULL, "Missing integration rule.");
+         offset += int_rule[geom]->GetNPoints();
+      }
+      offsets[num_elem] = offset;
+      size = offsets.Last();
    }
-   offsets[num_elem] = size = offset;
 }
 
-void QuadratureSpace::Construct()
-{
-   ConstructIntRules(mesh.Dimension());
-   ConstructOffsets();
-}
-
-QuadratureSpace::QuadratureSpace(Mesh *mesh_, std::istream &in)
-   : QuadratureSpaceBase(*mesh_)
+QuadratureSpace::QuadratureSpace(Mesh *mesh_, std::istream &in,
+                                 QSpaceStorage storage)
+   : QuadratureSpaceBase(*mesh_, 0, storage)
 {
    const char *msg = "invalid input stream";
    std::string ident;
@@ -135,15 +146,24 @@ QuadratureSpace::QuadratureSpace(Mesh *mesh_, std::istream &in)
       return;
    }
 
-   Construct();
+   ne = mesh.GetNE();
+   ConstructIntRules(mesh.Dimension());
 }
 
-QuadratureSpace::QuadratureSpace(Mesh &mesh_, const IntegrationRule &ir)
-   : QuadratureSpaceBase(mesh_, mesh_.GetTypicalElementGeometry(), ir)
+QuadratureSpace::QuadratureSpace(Mesh *mesh_, int order_, QSpaceStorage storage)
+   : QuadratureSpaceBase(*mesh_, order_, storage)
+{
+   ne = mesh.GetNE();
+   ConstructIntRules(mesh.Dimension());
+}
+
+QuadratureSpace::QuadratureSpace(Mesh &mesh_, const IntegrationRule &ir,
+                                 QSpaceStorage storage)
+   : QuadratureSpaceBase(mesh_, mesh_.GetTypicalElementGeometry(), ir, storage)
 {
    MFEM_VERIFY(mesh.GetNumGeometries(mesh.Dimension()) <= 1,
                "Constructor not valid for mixed meshes");
-   ConstructOffsets();
+   ne = mesh.GetNE();
 }
 
 void QuadratureSpace::Save(std::ostream &os) const
@@ -165,54 +185,53 @@ const Vector &QuadratureSpace::GetGeometricFactorWeights() const
 }
 
 FaceQuadratureSpace::FaceQuadratureSpace(Mesh &mesh_, int order_,
-                                         FaceType face_type_)
-   : QuadratureSpaceBase(mesh_, order_),
-     face_type(face_type_),
-     num_faces(mesh.GetNFbyType(face_type))
+                                         FaceType face_type_,
+                                         QSpaceStorage storage)
+   : QuadratureSpaceBase(mesh_, order_, storage), face_type(face_type_),
+     face_indices(mesh.GetFaceIndices(face_type_)),
+     face_indices_inv(mesh.GetInvFaceIndices(face_type_))
 {
-   Construct();
+   ne = face_indices.Size();
+   ConstructIntRules(mesh.Dimension() - 1);
 }
 
 FaceQuadratureSpace::FaceQuadratureSpace(Mesh &mesh_, const IntegrationRule &ir,
-                                         FaceType face_type_)
-   : QuadratureSpaceBase(mesh_, mesh_.GetTypicalFaceGeometry(), ir),
-     face_type(face_type_),
-     num_faces(mesh.GetNFbyType(face_type))
+                                         FaceType face_type_,
+                                         QSpaceStorage storage)
+   : QuadratureSpaceBase(mesh_, mesh_.GetTypicalFaceGeometry(), ir, storage),
+     face_type(face_type_), face_indices(mesh.GetFaceIndices(face_type_)),
+     face_indices_inv(mesh.GetInvFaceIndices(face_type_))
 {
    MFEM_VERIFY(mesh.GetNumGeometries(mesh.Dimension() - 1) <= 1,
                "Constructor not valid for mixed meshes");
-   ConstructOffsets();
+   ne = face_indices.Size();
 }
 
 void FaceQuadratureSpace::ConstructOffsets()
 {
-   face_indices.SetSize(num_faces);
-   offsets.SetSize(num_faces + 1);
-   int offset = 0;
-   int f_idx = 0;
-   for (int i = 0; i < mesh.GetNumFacesWithGhost(); i++)
+   if (storage == QSpaceStorage::COMPRESSED &&
+       mesh.GetNumGeometries(mesh.Dimension() - 1) == 1)
    {
-      const Mesh::FaceInformation face = mesh.GetFaceInformation(i);
-      if (face.IsNonconformingCoarse() || !face.IsOfFaceType(face_type))
-      {
-         continue;
-      }
-      face_indices[f_idx] = i;
-      face_indices_inv[i] = f_idx;
-      offsets[f_idx] = offset;
-      Geometry::Type geom = mesh.GetFaceGeometry(i);
-      MFEM_ASSERT(int_rule[geom] != NULL, "Missing integration rule");
-      offset += int_rule[geom]->GetNPoints();
-
-      f_idx++;
+      Array<Geometry::Type> geoms;
+      mesh.GetGeometries(mesh.Dimension() - 1, geoms);
+      offsets.SetSize(1);
+      offsets.HostWrite();
+      offsets[0] = int_rule[geoms[0]]->GetNPoints();
+      size = ne * offsets[0];
+   }
+   else
+   {
+      offsets.SetSize(face_indices.Size() + 1);
+      int offset = 0;
+      for (int i = 0; i < mesh.GetNFbyType(face_type); ++i)
+      {
+         offsets[i] = offset;
+         Geometry::Type geom = mesh.GetFaceGeometry(face_indices[i]);
+         MFEM_ASSERT(int_rule[geom] != NULL, "Missing integration rule");
+         offset += int_rule[geom]->GetNPoints();
+      }
+      offsets[face_indices.Size()] = size = offset;
    }
-   offsets[num_faces] = size = offset;
-}
-
-void FaceQuadratureSpace::Construct()
-{
-   ConstructIntRules(mesh.Dimension() - 1);
-   ConstructOffsets();
 }
 
 int FaceQuadratureSpace::GetPermutedIndex(int idx, int iq) const

fem/qspace.hpp

@@ -19,35 +19,49 @@
 namespace mfem
 {
 
+enum class QSpaceStorage
+{
+   FULL,
+   COMPRESSED
+};
+
 /// Abstract base class for QuadratureSpace and FaceQuadratureSpace.
 /** This class represents the storage layout for QuadratureFunction%s, that may
     be defined either on mesh elements or mesh faces. */
 class QuadratureSpaceBase
 {
 protected:
-   friend class QuadratureFunction; // Uses the offsets.
-
    Mesh &mesh; ///< The underlying mesh.
    int order; ///< The order of integration rule.
-   int size; ///< Total number of quadrature points.
+   int size = -1; ///< Total number of quadrature points. -1 indicates
+   ///< offsets/size not computed yet.
+   int ne; ///< Actual number of entities
    mutable Vector weights; ///< Integration weights.
    mutable long nodes_sequence = 0; ///< Nodes counter for cache invalidation.
 
-   /// @brief Entity quadrature point offset array, of size num_entities + 1.
+   QSpaceStorage storage;
+
+   /// @brief Entity quadrature point offset array.
+   ///
+   /// Supports a constant compression scheme for meshes which have a single
+   /// geometry type. When compressed, will have a single value. The true offset
+   /// can be computed as i * offsets[0], where i is the entity index. Otherwise
+   /// has size num_entities + 1. Lazily constructed.
    ///
-   /// The quadrature point values for entity i are stored in the indices between
-   /// offsets[i] and offsets[i+1].
    Array<int> offsets;
    /// The quadrature rules used for each geometry type.
    const IntegrationRule *int_rule[Geometry::NumGeom];
 
    /// Protected constructor. Used by derived classes.
-   QuadratureSpaceBase(Mesh &mesh_, int order_ = 0)
-      : mesh(mesh_), order(order_) { }
+   QuadratureSpaceBase(Mesh &mesh_, int order_ = 0,
+                       QSpaceStorage storage = QSpaceStorage::COMPRESSED)
+      : mesh(mesh_), order(order_), storage(storage)
+   {}
 
    /// Protected constructor. Used by derived classes.
    QuadratureSpaceBase(Mesh &mesh_, Geometry::Type geom,
-                       const IntegrationRule &ir);
+                       const IntegrationRule &ir,
+                       QSpaceStorage storage = QSpaceStorage::COMPRESSED);
 
    /// Fill the @ref int_rule array for each geometry type using @ref order.
    void ConstructIntRules(int dim);
@@ -58,15 +72,55 @@ protected:
    /// Compute the integration weights.
    void ConstructWeights() const;
 
+   virtual void ConstructOffsets() = 0;
+
 public:
+   QSpaceStorage StorageType() const { return storage; }
+
+   /// @brief Gets the offset for a given entity @a idx.
+   ///
+   /// The quadrature point values for entity i are stored in the indices
+   /// between Offset(i) and Offset(i+1)
+   int Offset(int idx) const
+   {
+      if (size < 0)
+      {
+         const_cast<QuadratureSpaceBase *>(this)->ConstructOffsets();
+      }
+      return (offsets.Size() == 1) ? (idx * offsets[0]) : offsets[idx];
+   }
+
+   /// @brief Entity quadrature point offset array.
+   ///
+   /// Supports a constant compression scheme for meshes which have a single
+   /// geometry type. When compressed, will have a single value. The true offset
+   /// can be computed as i * offsets[0], where i is the entity index. Otherwise
+   /// has size num_entities + 1.
+   ///
+   const Array<int> &Offsets() const
+   {
+      if (size < 0)
+      {
+         const_cast<QuadratureSpaceBase *>(this)->ConstructOffsets();
+      }
+      return offsets;
+   }
+
    /// Return the total number of quadrature points.
-   int GetSize() const { return size; }
+   int GetSize() const
+   {
+      if (size < 0)
+      {
+         const_cast<QuadratureSpaceBase *>(this)->ConstructOffsets();
+      }
+      return size;
+   }
 
    /// Return the order of the quadrature rule(s) used by all elements.
    int GetOrder() const { return order; }
 
    /// Return the number of entities.
-   int GetNE() const { return offsets.Size() - 1; }
+   int GetNE() const { return ne; }
 
    /// Returns the mesh.
    inline Mesh *GetMesh() const { return &mesh; }
@@ -120,19 +174,20 @@ class QuadratureSpace : public QuadratureSpaceBase
 {
 protected:
    const Vector &GetGeometricFactorWeights() const override;
-   void ConstructOffsets();
-   void Construct();
+   void ConstructOffsets() override;
 public:
    /// Create a QuadratureSpace based on the global rules from #IntRules.
-   QuadratureSpace(Mesh *mesh_, int order_)
-      : QuadratureSpaceBase(*mesh_, order_) { Construct(); }
+   QuadratureSpace(Mesh *mesh_, int order_,
+                   QSpaceStorage storage = QSpaceStorage::COMPRESSED);
 
    /// @brief Create a QuadratureSpace with an IntegrationRule, valid only when
    /// the mesh has one element type.
-   QuadratureSpace(Mesh &mesh_, const IntegrationRule &ir);
+   QuadratureSpace(Mesh &mesh_, const IntegrationRule &ir,
+                   QSpaceStorage storage = QSpaceStorage::COMPRESSED);
 
    /// Read a QuadratureSpace from the stream @a in.
-   QuadratureSpace(Mesh *mesh_, std::istream &in);
+   QuadratureSpace(Mesh *mesh_, std::istream &in,
+                   QSpaceStorage storage = QSpaceStorage::COMPRESSED);
 
    /// Returns number of elements in the mesh.
    inline int GetNE() const { return mesh.GetNE(); }
@@ -169,29 +224,29 @@ public:
 class FaceQuadratureSpace : public QuadratureSpaceBase
 {
    FaceType face_type; ///< Is the space defined on interior or boundary faces?
-   const int num_faces; ///< Number of faces.
 
    /// Map from boundary or interior face indices to mesh face indices.
-   Array<int> face_indices;
+   const Array<int> &face_indices;
 
    /// Inverse of the map @a face_indices.
-   std::unordered_map<int,int> face_indices_inv;
+   const std::unordered_map<int,int> &face_indices_inv;
 
    const Vector &GetGeometricFactorWeights() const override;
-   void ConstructOffsets();
-   void Construct();
+   void ConstructOffsets() override;
 
 public:
    /// Create a FaceQuadratureSpace based on the global rules from #IntRules.
-   FaceQuadratureSpace(Mesh &mesh_, int order_, FaceType face_type_);
+   FaceQuadratureSpace(Mesh &mesh_, int order_, FaceType face_type_,
+                       QSpaceStorage storage = QSpaceStorage::COMPRESSED);
 
    /// @brief Create a FaceQuadratureSpace with an IntegrationRule, valid only
    /// when the mesh has one type of face geometry.
    FaceQuadratureSpace(Mesh &mesh_, const IntegrationRule &ir,
-                       FaceType face_type_);
+                       FaceType face_type_,
+                       QSpaceStorage storage = QSpaceStorage::COMPRESSED);
 
    /// Returns number of faces in the mesh.
-   inline int GetNumFaces() const { return num_faces; }
+   inline int GetNumFaces() const { return face_indices.Size(); }
 
    /// Returns the face type (boundary or interior).
    FaceType GetFaceType() const { return face_type; }

general/array.hpp

@@ -211,6 +211,9 @@ public:
    /// Delete the first entry with value == 'el'.
    inline void DeleteFirst(const T &el);
 
+   /// Delete entries at @a indices, and resize.
+   inline void DeleteAt(const Array<int> &indices);
+
    /// Delete the whole array.
    inline void DeleteAll();
 
@@ -935,6 +938,30 @@ inline void Array<T>::DeleteFirst(const T &el)
    }
 }
 
+template <class T>
+inline void Array<T>::DeleteAt(const Array<int> &indices)
+{
+   // Make a copy of the indices, sorted.
+   Array<int> sorted_indices(indices);
+   sorted_indices.Sort();
+
+   int rm_count = 0;
+   for (int i = 0; i < size; i++)
+   {
+      if (rm_count < sorted_indices.Size() && i == sorted_indices[rm_count])
+      {
+         rm_count++;
+      }
+      else
+      {
+         data[i-rm_count] = data[i]; // shift data rm_count
+      }
+   }
+
+   // Resize to remove tail
+   size -= rm_count;
+}
+
 template <class T>
 inline void Array<T>::DeleteAll()
 {

general/backends.hpp

@@ -14,7 +14,7 @@
 
 #include "../config/config.hpp"
 
-#ifdef MFEM_USE_CUDA
+#if defined(MFEM_USE_CUDA) && defined(__CUDACC__)
 #include <cusparse.h>
 #include <library_types.h>
 #include <cuda_runtime.h>
@@ -22,7 +22,7 @@
 #endif
 #include "cuda.hpp"
 
-#ifdef MFEM_USE_HIP
+#if defined(MFEM_USE_HIP) && defined(__HIP__)
 #include <hip/hip_runtime.h>
 #endif
 #include "hip.hpp"
@@ -43,7 +43,7 @@
 #endif
 #endif
 
-#if !(defined(MFEM_USE_CUDA) || defined(MFEM_USE_HIP))
+#if !defined(MFEM_USE_CUDA_OR_HIP)
 #define MFEM_DEVICE
 #define MFEM_HOST
 #define MFEM_LAMBDA
@@ -55,7 +55,7 @@
 #endif
 
 #if !((defined(MFEM_USE_CUDA) && defined(__CUDA_ARCH__)) || \
-      (defined(MFEM_USE_HIP)  && defined(__HIP_DEVICE_COMPILE__)))
+      (defined(MFEM_USE_HIP) && defined(__HIP_DEVICE_COMPILE__)))
 #define MFEM_SHARED
 #define MFEM_SYNC_THREAD
 #define MFEM_BLOCK_ID(k) 0
@@ -66,7 +66,7 @@
 #endif
 
 // 'double' and 'float' atomicAdd implementation for previous versions of CUDA
-#if defined(MFEM_USE_CUDA) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 600
+#if defined(MFEM_USE_CUDA) && defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 600)
 MFEM_DEVICE inline mfem::real_t atomicAdd(mfem::real_t *add, mfem::real_t val)
 {
    unsigned long long int *ptr = (unsigned long long int *) add;
@@ -94,7 +94,7 @@ template <typename T>
 MFEM_HOST_DEVICE T AtomicAdd(T &add, const T val)
 {
 #if ((defined(MFEM_USE_CUDA) && defined(__CUDA_ARCH__)) || \
-     (defined(MFEM_USE_HIP)  && defined(__HIP_DEVICE_COMPILE__)))
+     (defined(MFEM_USE_HIP) && defined(__HIP_DEVICE_COMPILE__)))
    return atomicAdd(&add,val);
 #else
    T old = add;

general/communication.cpp

@@ -347,13 +347,23 @@ void GroupTopology::Swap(GroupTopology &other)
    mfem::Swap(group_mgroup, other.group_mgroup);
 }
 
+/// \cond DO_NOT_DOCUMENT
 // Initialize the static mpi_type for the specializations of MPITypeMap:
+const MPI_Datatype MPITypeMap<bool>::mpi_type = MFEM_MPI_CXX_BOOL;
+const MPI_Datatype MPITypeMap<char>::mpi_type = MPI_CHAR;
+const MPI_Datatype MPITypeMap<unsigned char>::mpi_type = MPI_UNSIGNED_CHAR;
+const MPI_Datatype MPITypeMap<short>::mpi_type = MPI_SHORT;
+const MPI_Datatype MPITypeMap<unsigned short>::mpi_type = MPI_UNSIGNED_SHORT;
 const MPI_Datatype MPITypeMap<int>::mpi_type = MPI_INT;
+const MPI_Datatype MPITypeMap<unsigned int>::mpi_type = MPI_UNSIGNED;
+const MPI_Datatype MPITypeMap<long>::mpi_type = MPI_LONG;
+const MPI_Datatype MPITypeMap<unsigned long>::mpi_type = MPI_UNSIGNED_LONG;
+const MPI_Datatype MPITypeMap<long long>::mpi_type = MPI_LONG_LONG;
+const MPI_Datatype MPITypeMap<unsigned long long>::mpi_type =
+   MPI_UNSIGNED_LONG_LONG;
 const MPI_Datatype MPITypeMap<float>::mpi_type = MPI_FLOAT;
 const MPI_Datatype MPITypeMap<double>::mpi_type = MPI_DOUBLE;
-const MPI_Datatype MPITypeMap<int64_t>::mpi_type = MPI_INT64_T;
-const MPI_Datatype MPITypeMap<uint64_t>::mpi_type = MPI_UINT64_T;
-
+/// \endcond DO_NOT_DOCUMENT
 
 GroupCommunicator::GroupCommunicator(const GroupTopology &gt, Mode m)
    : gtopo(gt), mode(m)

general/communication.hpp

@@ -425,16 +425,24 @@ public:
    ~GroupCommunicator();
 };
 
+/// General MPI message tags used by MFEM
+enum MessageTag
+{
+   DEREFINEMENT_MATRIX_CONSTRUCTION_DATA =
+      291, /// ParFiniteElementSpace ParallelDerefinementMatrix and
+   /// ParDerefineMatrixOp
+};
+
 enum VarMessageTag
 {
-   NEIGHBOR_ELEMENT_RANK_VM,  ///< NeighborElementRankMessage
-   NEIGHBOR_ORDER_VM,         ///< NeighborOrderMessage
-   NEIGHBOR_DEREFINEMENT_VM,  ///< NeighborDerefinementMessage
-   NEIGHBOR_REFINEMENT_VM,    ///< NeighborRefinementMessage
-   NEIGHBOR_PREFINEMENT_VM,   ///< NeighborPRefinementMessage
-   NEIGHBOR_ROW_VM,           ///< NeighborRowMessage
-   REBALANCE_VM,              ///< RebalanceMessage
-   REBALANCE_DOF_VM           ///< RebalanceDofMessage
+   NEIGHBOR_ELEMENT_RANK_VM, ///< NeighborElementRankMessage
+   NEIGHBOR_ORDER_VM,        ///< NeighborOrderMessage
+   NEIGHBOR_DEREFINEMENT_VM, ///< NeighborDerefinementMessage
+   NEIGHBOR_REFINEMENT_VM,   ///< NeighborRefinementMessage
+   NEIGHBOR_PREFINEMENT_VM,  ///< NeighborPRefinementMessage
+   NEIGHBOR_ROW_VM,          ///< NeighborRowMessage
+   REBALANCE_VM,             ///< RebalanceMessage
+   REBALANCE_DOF_VM,         ///< RebalanceDofMessage
 };
 
 /// \brief Variable-length MPI message containing unspecific binary data.
@@ -599,10 +607,50 @@ protected:
 template <typename Type> struct MPITypeMap;
 
 // Specializations of MPITypeMap; mpi_type initialized in communication.cpp:
+template<> struct MPITypeMap<bool>
+{
+   static MFEM_EXPORT const MPI_Datatype mpi_type;
+};
+template<> struct MPITypeMap<char>
+{
+   static MFEM_EXPORT const MPI_Datatype mpi_type;
+};
+template<> struct MPITypeMap<unsigned char>
+{
+   static MFEM_EXPORT const MPI_Datatype mpi_type;
+};
+template<> struct MPITypeMap<short>
+{
+   static MFEM_EXPORT const MPI_Datatype mpi_type;
+};
+template<> struct MPITypeMap<unsigned short>
+{
+   static MFEM_EXPORT const MPI_Datatype mpi_type;
+};
 template<> struct MPITypeMap<int>
 {
    static MFEM_EXPORT const MPI_Datatype mpi_type;
 };
+template<> struct MPITypeMap<unsigned int>
+{
+   static MFEM_EXPORT const MPI_Datatype mpi_type;
+};
+template<> struct MPITypeMap<long>
+{
+   static MFEM_EXPORT const MPI_Datatype mpi_type;
+};
+template<> struct MPITypeMap<unsigned long>
+{
+   static MFEM_EXPORT const MPI_Datatype mpi_type;
+};
+template<> struct MPITypeMap<long long>
+{
+   static MFEM_EXPORT const MPI_Datatype mpi_type;
+};
+template<> struct MPITypeMap<unsigned long long>
+{
+   static MFEM_EXPORT const MPI_Datatype mpi_type;
+};
 template<> struct MPITypeMap<double>
 {
    static MFEM_EXPORT const MPI_Datatype mpi_type;
@@ -611,15 +659,6 @@ template<> struct MPITypeMap<float>
 {
    static MFEM_EXPORT const MPI_Datatype mpi_type;
 };
-template<> struct MPITypeMap<int64_t>
-{
-   static MFEM_EXPORT const MPI_Datatype mpi_type;
-};
-template<> struct MPITypeMap<uint64_t>
-{
-   static MFEM_EXPORT const MPI_Datatype mpi_type;
-};
-
 
 /** Reorder MPI ranks to follow the Z-curve within the physical machine topology
     (provided that functions to query physical node coordinates are available).

general/cuda.cpp

@@ -24,7 +24,7 @@ void mfem_cuda_error(cudaError_t err, const char *expr, const char *func,
                      const char *file, int line)
 {
    mfem::err << "\n\nCUDA error: (" << expr << ") failed with error:\n --> "
-             << cudaGetErrorString(err)
+             << cudaGetErrorString(err) << " [code: " << (int)err << ']'
              << "\n ... in function: " << func
              << "\n ... in file: " << file << ':' << line << '\n';
    mfem_error();

general/cuda.hpp

@@ -18,7 +18,7 @@
 // CUDA block size used by MFEM.
 #define MFEM_CUDA_BLOCKS 256
 
-#ifdef MFEM_USE_CUDA
+#if defined(MFEM_USE_CUDA) && defined(__CUDACC__)
 #define MFEM_USE_CUDA_OR_HIP
 #define MFEM_DEVICE __device__
 #define MFEM_HOST __host__
@@ -37,10 +37,9 @@
                               __FILE__, __LINE__);                             \
     }                                                                          \
   } while (0)
-#endif // MFEM_USE_CUDA
 
 // Define the MFEM inner threading macros
-#if defined(MFEM_USE_CUDA) && defined(__CUDA_ARCH__)
+#if defined(__CUDA_ARCH__)
 #define MFEM_SHARED __shared__
 #define MFEM_SYNC_THREAD __syncthreads()
 #define MFEM_BLOCK_ID(k) blockIdx.k
@@ -48,12 +47,13 @@
 #define MFEM_THREAD_SIZE(k) blockDim.k
 #define MFEM_FOREACH_THREAD(i,k,N) for(int i=threadIdx.k; i<N; i+=blockDim.k)
 #define MFEM_FOREACH_THREAD_DIRECT(i,k,N) if(const int i=threadIdx.k; i<N)
-#endif
+#endif // defined(__CUDA_ARCH__)
+#endif // defined(MFEM_USE_CUDA) && defined(__CUDACC__)
 
 namespace mfem
 {
 
-#ifdef MFEM_USE_CUDA
+#if defined(MFEM_USE_CUDA) && defined(__CUDACC__)
 // Function used by the macro MFEM_GPU_CHECK.
 void mfem_cuda_error(cudaError_t err, const char *expr, const char *func,
                      const char *file, int line);

general/device.cpp

@@ -615,7 +615,7 @@ void Device::Setup(const std::string &device_option, const int device_id)
    if (Allows(Backend::DEBUG_DEVICE)) { ngpu = 1; }
 }
 
-MemoryType Device::QueryMemoryType(void *ptr)
+MemoryType Device::QueryMemoryType(const void* ptr)
 {
    // from HYPRE's hypre_GetPointerLocation
    MemoryType res = MemoryType::HOST;

general/device.hpp

@@ -297,9 +297,9 @@ public:
    /// Get the status of GPU-aware MPI flag.
    static bool GetGPUAwareMPI() { return Get().mpi_gpu_aware; }
 
-   /** @brief Query the device driver for what memory type a given @a ptr is
-       allocated with. */
-   static MemoryType QueryMemoryType(void *ptr);
+   /** Query the device driver for what memory type a given @a ptr is allocated
+    * with. */
+   static MemoryType QueryMemoryType(const void* ptr);
 
    /** @brief The number of hardware compute units/streaming multiprocessors
        available on a given compute device @a device_id. */

general/error.hpp

@@ -176,7 +176,7 @@ __device__ void abort_msg(T & msg)
       printf(__VA_ARGS__);     \
       asm("trap;");            \
    }
-#elif defined(MFEM_USE_HIP)
+#elif defined(__HIP_DEVICE_COMPILE__)
 #define MFEM_ABORT_KERNEL(...) \
    {                           \
       printf(__VA_ARGS__);     \

general/forall.hpp

@@ -158,8 +158,8 @@ private:
 #define MFEM_PRAGMA(X) _Pragma(#X)
 
 // MFEM_UNROLL pragma macro that can be used inside MFEM_FORALL macros.
-#if defined(MFEM_USE_CUDA) && defined(__CUDA_ARCH__)
-#ifdef __NVCC__
+#if defined(MFEM_USE_CUDA) && defined(__CUDA_ARCH__) // Clang cuda or nvcc
+#ifdef __NVCC__ // nvcc specifically
 #define MFEM_UNROLL(N) MFEM_PRAGMA(unroll(N))
 #else // Assuming Clang CUDA
 #define MFEM_UNROLL(N) MFEM_PRAGMA(unroll N)
@@ -169,12 +169,12 @@ private:
 #endif
 
 // MFEM_GPU_FORALL: "parallel for" executed with CUDA or HIP based on the MFEM
-// build-time configuration (MFEM_USE_CUDA or MFEM_USE_HIP). If neither CUDA nor
-// HIP is enabled, this macro is a no-op.
-#if defined(MFEM_USE_CUDA)
+// build-time configuration (MFEM_USE_CUDA or MFEM_USE_HIP), and if compiling
+// with CUDA/HIP language. Otherwise, this macro is a no-op.
+#if defined(MFEM_USE_CUDA) && defined(__CUDACC__)
 #define MFEM_GPU_FORALL(i, N,...) CuWrap1D(N, [=] MFEM_DEVICE      \
                                        (int i) {__VA_ARGS__})
-#elif defined(MFEM_USE_HIP)
+#elif defined(MFEM_USE_HIP) && defined(__HIP__)
 #define MFEM_GPU_FORALL(i, N,...) HipWrap1D(N, [=] MFEM_DEVICE     \
                                         (int i) {__VA_ARGS__})
 #else
@@ -225,6 +225,52 @@ void OmpWrap(const int N, HBODY &&h_body)
 #endif
 }
 
+template <typename HBODY>
+void OmpWrap2D(const int Nx, const int Ny, HBODY &&h_body)
+{
+#ifdef MFEM_USE_OPENMP
+   // requires OpenMP 3.1
+   #pragma omp parallel for collapse(2)
+   for (int j = 0; j < Ny; j++)
+   {
+      for (int i = 0; i < Nx; i++)
+      {
+         h_body(i, j);
+      }
+   }
+#else
+   MFEM_CONTRACT_VAR(Nx);
+   MFEM_CONTRACT_VAR(Ny);
+   MFEM_CONTRACT_VAR(h_body);
+   MFEM_ABORT("OpenMP requested for MFEM but OpenMP is not enabled!");
+#endif
+}
+
+template <typename HBODY>
+void OmpWrap3D(const int Nx, const int Ny, const int Nz, HBODY &&h_body)
+{
+#ifdef MFEM_USE_OPENMP
+   // requires OpenMP 3.1
+   #pragma omp parallel for collapse(3)
+   for (int k = 0; k < Nz; k++)
+   {
+      for (int j = 0; j < Ny; j++)
+      {
+         for (int i = 0; i < Nx; i++)
+         {
+            h_body(i, j, k);
+         }
+      }
+   }
+#else
+   MFEM_CONTRACT_VAR(Nx);
+   MFEM_CONTRACT_VAR(Ny);
+   MFEM_CONTRACT_VAR(Nz);
+   MFEM_CONTRACT_VAR(h_body);
+   MFEM_ABORT("OpenMP requested for MFEM but OpenMP is not enabled!");
+#endif
+}
+
 
 /// RAJA Cuda and Hip backends
 #if defined(MFEM_USE_RAJA) && defined(RAJA_ENABLE_CUDA)
@@ -454,6 +500,42 @@ void RajaOmpWrap(const int N, HBODY &&h_body)
    RAJA::forall<RAJA::omp_parallel_for_exec>(RAJA::RangeSegment(0,N), h_body);
 }
 
+template <typename HBODY>
+void RajaOmpWrap2D(const int Nx, const int Ny, HBODY &&h_body)
+{
+   using omp_launch_policy = RAJA::LaunchPolicy<RAJA::omp_launch_t>;
+   using global_thread_xy = RAJA::LoopPolicy<RAJA::omp_for_exec>;
+   RAJA::RangeSegment xrange(0, Nx);
+   RAJA::RangeSegment yrange(0, Ny);
+   RAJA::launch<omp_launch_policy>(RAJA::ExecPlace::HOST, RAJA::LaunchParams(),
+                                   [=](RAJA::LaunchContext ctx)
+   {
+      // contiguous in x
+      RAJA::expt::loop<global_thread_xy>(ctx, xrange, yrange, [&](int i, int j)
+      {
+         h_body(i, j);
+      });
+   });
+}
+
+template <typename HBODY>
+void RajaOmpWrap3D(const int Nx, const int Ny, const int Nz, HBODY &&h_body)
+{
+   using omp_launch_policy = RAJA::LaunchPolicy<RAJA::omp_launch_t>;
+   using global_thread_xyz = RAJA::LoopPolicy<RAJA::omp_for_exec>;
+   RAJA::RangeSegment xrange(0, Nx);
+   RAJA::RangeSegment yrange(0, Ny);
+   RAJA::RangeSegment zrange(0, Nz);
+   RAJA::launch<omp_launch_policy>(RAJA::ExecPlace::HOST, RAJA::LaunchParams(),
+                                   [=](RAJA::LaunchContext ctx)
+   {
+      // contiguous in x
+      RAJA::expt::loop<global_thread_xyz>(ctx, xrange, yrange, zrange,
+                                          [&](int i, int j, int k)
+      { h_body(i, j, k); });
+   });
+}
+
 #endif
 
 
@@ -481,7 +563,7 @@ void RajaSeqWrap(const int N, HBODY &&h_body)
 
 
 /// CUDA backend
-#ifdef MFEM_USE_CUDA
+#if defined(MFEM_USE_CUDA) && defined(__CUDACC__)
 
 template <typename BODY> __global__ static
 void CuKernel1D(const int N, BODY body)
@@ -573,11 +655,11 @@ struct CuWrap<3>
    }
 };
 
-#endif // MFEM_USE_CUDA
+#endif // defined(MFEM_USE_CUDA) && defined(__CUDACC__)
 
 
 /// HIP backend
-#ifdef MFEM_USE_HIP
+#if defined(MFEM_USE_HIP) && defined(__HIP__)
 
 template <typename BODY> __global__ static
 void HipKernel1D(const int N, BODY body)
@@ -668,7 +750,7 @@ struct HipWrap<3>
    }
 };
 
-#endif // MFEM_USE_HIP
+#endif // defined(MFEM_USE_HIP) && defined(__HIP__)
 
 
 /// The forall kernel body wrapper
@@ -701,7 +783,7 @@ inline void ForallWrap(const bool use_dev, const int N,
    }
 #endif
 
-#ifdef MFEM_USE_CUDA
+#if defined(MFEM_USE_CUDA) && defined(__CUDACC__)
    // If Backend::CUDA is allowed, use it
    if (Device::Allows(Backend::CUDA))
    {
@@ -709,7 +791,7 @@ inline void ForallWrap(const bool use_dev, const int N,
    }
 #endif
 
-#ifdef MFEM_USE_HIP
+#if defined(MFEM_USE_HIP) && defined(__HIP__)
    // If Backend::HIP is allowed, use it
    if (Device::Allows(Backend::HIP))
    {
@@ -753,6 +835,83 @@ inline void ForallWrap(const bool use_dev, const int N, lambda &&body,
 template<typename lambda>
 inline void forall(int N, lambda &&body) { ForallWrap<1>(true, N, body); }
 
+template<typename lambda>
+inline void forall(int Nx, int Ny, lambda &&body)
+{
+   if (Device::Allows(Backend::DEVICE_MASK))
+   {
+      forall(Nx * Ny, [=] MFEM_HOST_DEVICE(int idx)
+      {
+         int j = idx / Nx;
+         int i = idx % Nx;
+         body(i, j);
+      });
+   }
+#if defined(MFEM_USE_RAJA) && defined(RAJA_ENABLE_OPENMP)
+   else if (Device::Allows(Backend::RAJA_OMP))
+   {
+      return RajaOmpWrap2D(Nx, Ny, body);
+   }
+#endif
+#ifdef MFEM_USE_OPENMP
+   else if (Device::Allows(Backend::OMP))
+   {
+      return OmpWrap2D(Nx, Ny, body);
+   }
+#endif
+   else
+   {
+      for (int j = 0; j < Ny; ++j)
+      {
+         for (int i = 0; i < Nx; ++i)
+         {
+            body(i, j);
+         }
+      }
+   }
+}
+
+template<typename lambda>
+inline void forall(int Nx, int Ny, int Nz, lambda &&body)
+{
+   if (Device::Allows(Backend::DEVICE_MASK))
+   {
+      forall(Nx * Ny * Nz, [=] MFEM_HOST_DEVICE(int idx)
+      {
+         int i = idx % Nx;
+         int j = idx / Nx;
+         int k = j / Ny;
+         j = j % Ny;
+         body(i, j, k);
+      });
+   }
+#if defined(MFEM_USE_RAJA) && defined(RAJA_ENABLE_OPENMP)
+   else if (Device::Allows(Backend::RAJA_OMP))
+   {
+      return RajaOmpWrap3D(Nx, Ny, Nz, body);
+   }
+#endif
+#ifdef MFEM_USE_OPENMP
+   else if (Device::Allows(Backend::OMP))
+   {
+      return OmpWrap3D(Nx, Ny, Nz, body);
+   }
+#endif
+   else
+   {
+      for (int k = 0; k < Nz; ++k)
+      {
+         for (int j = 0; j < Ny; ++j)
+         {
+            for (int i = 0; i < Nx; ++i)
+            {
+               body(i, j, k);
+            }
+         }
+      }
+   }
+}
+
 template<typename lambda>
 inline void forall_switch(bool use_dev, int N, lambda &&body)
 {

general/hip.cpp

@@ -24,7 +24,7 @@ void mfem_hip_error(hipError_t err, const char *expr, const char *func,
                     const char *file, int line)
 {
    mfem::err << "\n\nHIP error: (" << expr << ") failed with error:\n --> "
-             << hipGetErrorString(err)
+             << hipGetErrorString(err) << " [code: " << (int)err << ']'
              << "\n ... in function: " << func
              << "\n ... in file: " << file << ':' << line << '\n';
    mfem_error();

general/hip.hpp

@@ -18,7 +18,7 @@
 // HIP block size used by MFEM.
 #define MFEM_HIP_BLOCKS 256
 
-#ifdef MFEM_USE_HIP
+#if defined(MFEM_USE_HIP) && defined(__HIP__)
 #define MFEM_USE_CUDA_OR_HIP
 #define MFEM_DEVICE __device__
 #define MFEM_HOST __host__
@@ -37,10 +37,9 @@
                              __FILE__, __LINE__);                              \
     }                                                                          \
   } while (0)
-#endif // MFEM_USE_HIP
 
 // Define the MFEM inner threading macros
-#if defined(MFEM_USE_HIP) && defined(__HIP_DEVICE_COMPILE__)
+#if defined(__HIP_DEVICE_COMPILE__)
 #define MFEM_SHARED __shared__
 #define MFEM_SYNC_THREAD __syncthreads()
 #define MFEM_BLOCK_ID(k) hipBlockIdx_ ##k
@@ -50,7 +49,8 @@
    for(int i=hipThreadIdx_ ##k; i<N; i+=hipBlockDim_ ##k)
 #define MFEM_FOREACH_THREAD_DIRECT(i,k,N) \
    if(const int i=hipThreadIdx_ ##k; i<N)
-#endif
+#endif // defined(__HIP_DEVICE_COMPILE__)
+#endif // defined(MFEM_USE_HIP) && defined(__HIP__)
 
 namespace mfem
 {

general/mem_manager.cpp

@@ -513,10 +513,7 @@ public:
    void *HtoD(void *dst, const void *src, size_t bytes) override
    { return HipMemcpyHtoD(dst, src, bytes); }
    void *DtoD(void* dst, const void* src, size_t bytes) override
-   // Unlike cudaMemcpy(DtoD), hipMemcpy(DtoD) causes a host-side synchronization so
-   // instead we use hipMemcpyAsync to get similar behavior.
-   // for more info see: https://github.com/mfem/mfem/pull/2780
-   { return HipMemcpyDtoDAsync(dst, src, bytes); }
+   { return HipMemcpyDtoD(dst, src, bytes); }
    void *DtoH(void *dst, const void *src, size_t bytes) override
    { return HipMemcpyDtoH(dst, src, bytes); }
 };
@@ -658,10 +655,7 @@ public:
       return CuMemcpyDtoD(dst, src, bytes);
 #endif
 #ifdef MFEM_USE_HIP
-      // Unlike cudaMemcpy(DtoD), hipMemcpy(DtoD) causes a host-side synchronization so
-      // instead we use hipMemcpyAsync to get similar behavior.
-      // for more info see: https://github.com/mfem/mfem/pull/2780
-      return HipMemcpyDtoDAsync(dst, src, bytes);
+      return HipMemcpyDtoD(dst, src, bytes);
 #endif
       // rm.copy(dst, const_cast<void*>(src), bytes); return dst;
    }

general/mem_manager.hpp

@@ -896,6 +896,7 @@ inline HYPRE_MemoryLocation GetHypreMemoryLocation()
 #elif MFEM_HYPRE_VERSION < 23100
    return HYPRE_MEMORY_DEVICE;
 #else // HYPRE_USING_GPU is defined and MFEM_HYPRE_VERSION >= 23100
+   if (!HYPRE_Initialized()) { return HYPRE_MEMORY_HOST; }
    HYPRE_MemoryLocation loc;
    HYPRE_GetMemoryLocation(&loc);
    return loc;
@@ -1057,7 +1058,8 @@ inline void Memory<T>::MakeAlias(const Memory &base, int offset, int size)
          // register the 'base' if the MemoryManager::Exists():
          MemoryManager::Exists()
 #else // HYPRE_USING_GPU is defined and MFEM_HYPRE_VERSION >= 23100
-         MemoryManager::Exists() && HypreUsingGPU()
+         IsDeviceMemory(MemoryManager::GetDeviceMemoryType()) ||
+         (MemoryManager::Exists() && HypreUsingGPU())
 #endif
       )
       {

general/reducers.hpp

@@ -537,7 +537,7 @@ void reduce(int N, T &res, B &&body, const R &reducer, bool use_dev,
       return;
    }
 
-#if defined(MFEM_USE_HIP) || defined(MFEM_USE_CUDA)
+#if defined(MFEM_USE_CUDA_OR_HIP)
    if (use_dev &&
        mfem::Device::Allows(Backend::CUDA | Backend::HIP | Backend::RAJA_CUDA |
                             Backend::RAJA_HIP))

general/scan.hpp

@@ -0,0 +1,176 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#ifndef MFEM_SCAN_HPP
+#define MFEM_SCAN_HPP
+
+#ifdef MFEM_USE_CUDA
+#include <cub/device/device_scan.cuh>
+#define MFEM_CUB_NAMESPACE cub
+#elif MFEM_USE_HIP
+#include <hipcub/device/device_scan.hpp>
+#define MFEM_CUB_NAMESPACE hipcub
+#endif
+
+#include <functional>
+#include <numeric>
+
+namespace mfem
+{
+/// Equivalent to InclusiveScan(use_dev, d_in, d_out, num_items, workspace,
+/// std::plus<>{})
+template <class InputIt, class OutputIt>
+void InclusiveScan(bool use_dev, InputIt d_in, OutputIt d_out, size_t num_items,
+                   Array<char> &workspace)
+{
+   // forward to InclusiveSum for potentially faster kernels
+#if defined(MFEM_USE_CUDA) || defined(MFEM_USE_HIP)
+   if (use_dev && mfem::Device::Allows(Backend::CUDA_MASK | Backend::HIP_MASK))
+   {
+      size_t bytes = workspace.Size();
+      if (bytes)
+      {
+         auto err = MFEM_CUB_NAMESPACE::DeviceScan::InclusiveSum(
+                       workspace.Write(), bytes, d_in, d_out, num_items);
+#if defined(MFEM_USE_CUDA)
+         if (err == cudaSuccess)
+         {
+            return;
+         }
+#elif defined(MFEM_USE_HIP)
+         if (err == hipSuccess)
+         {
+            return;
+         }
+#endif
+      }
+      // try allocating a larger buffer
+      bytes = 0;
+      MFEM_GPU_CHECK(MFEM_CUB_NAMESPACE::DeviceScan::InclusiveSum(
+                        nullptr, bytes, d_in, d_out, num_items));
+      workspace.SetSize(bytes);
+      MFEM_GPU_CHECK(MFEM_CUB_NAMESPACE::DeviceScan::InclusiveSum(
+                        workspace.Write(), bytes, d_in, d_out, num_items));
+      return;
+   }
+#endif
+   std::inclusive_scan(d_in, d_in + num_items, d_out);
+}
+
+/// Performs an inclusive scan of [d_in, d_in+num_items) -> [d_out,
+/// d_out+num_items). This call is potentially asynchronous on the device.
+/// @a d_in input start.
+/// @a d_out output start. Can perform in-place scans with d_out = d_in
+/// @a workspace temporary workspace used for device scans. TODO: replace with
+/// internal temporary workspace once that's added to the memory manager.
+/// @a scan_op binary scan functor. Must be associative. If only weakly
+/// associative (i.e. floating point addition) results are not deterministic. On
+/// device this must also be commutative.
+template <class InputIt, class OutputIt, class ScanOp>
+void InclusiveScan(bool use_dev, InputIt d_in, OutputIt d_out, size_t num_items,
+                   Array<char> &workspace, ScanOp scan_op)
+{
+#if defined(MFEM_USE_CUDA) || defined(MFEM_USE_HIP)
+   if (use_dev && mfem::Device::Allows(Backend::CUDA_MASK | Backend::HIP_MASK))
+   {
+      size_t bytes = workspace.Size();
+      if (bytes)
+      {
+         auto err = MFEM_CUB_NAMESPACE::DeviceScan::InclusiveScan(
+                       workspace.Write(), bytes, d_in, d_out, scan_op, num_items);
+#if defined(MFEM_USE_CUDA)
+         if (err == cudaSuccess)
+         {
+            return;
+         }
+#elif defined(MFEM_USE_HIP)
+         if (err == hipSuccess)
+         {
+            return;
+         }
+#endif
+      }
+      // try allocating a larger buffer
+      bytes = 0;
+      MFEM_GPU_CHECK(MFEM_CUB_NAMESPACE::DeviceScan::InclusiveScan(
+                        nullptr, bytes, d_in, d_out, scan_op, num_items));
+      workspace.SetSize(bytes);
+      MFEM_GPU_CHECK(MFEM_CUB_NAMESPACE::DeviceScan::InclusiveScan(
+                        workspace.Write(), bytes, d_in, d_out, scan_op, num_items));
+      return;
+   }
+#endif
+   std::inclusive_scan(d_in, d_in + num_items, d_out, scan_op);
+}
+
+/// Performs an exclusive scan of [d_in, d_in+num_items) -> [d_out,
+/// d_out+num_items). This call is potentially asynchronous on the device.
+/// @a d_in input start.
+/// @a d_out output start. Can perform in-place scans with d_out = d_in
+/// @a workspace temporary workspace used for device scans. TODO: replace with
+/// internal temporary workspace once that's added to the memory manager.
+/// @a scan_op binary scan functor. Must be associative. If only weakly
+/// associative (i.e. floating point addition) results are not deterministic. On
+/// device this must also be commutative.
+template <class InputIt, class OutputIt, class T, class ScanOp>
+void ExclusiveScan(bool use_dev, InputIt d_in, OutputIt d_out, size_t num_items,
+                   T init_value, Array<char> &workspace, ScanOp scan_op)
+{
+#if defined(MFEM_USE_CUDA) || defined(MFEM_USE_HIP)
+   if (use_dev && mfem::Device::Allows(Backend::CUDA_MASK | Backend::HIP_MASK))
+   {
+      size_t bytes = workspace.Size();
+      if (bytes)
+      {
+         auto err = MFEM_CUB_NAMESPACE::DeviceScan::ExclusiveScan(
+                       workspace.Write(), bytes, d_in, d_out, scan_op, init_value,
+                       num_items);
+#if defined(MFEM_USE_CUDA)
+         if (err == cudaSuccess)
+         {
+            return;
+         }
+#elif defined(MFEM_USE_HIP)
+         if (err == hipSuccess)
+         {
+            return;
+         }
+#endif
+      }
+      // try allocating a larger buffer
+      bytes = 0;
+      MFEM_GPU_CHECK(MFEM_CUB_NAMESPACE::DeviceScan::ExclusiveScan(
+                        nullptr, bytes, d_in, d_out, scan_op, init_value, num_items));
+      workspace.SetSize(bytes);
+      MFEM_GPU_CHECK(MFEM_CUB_NAMESPACE::DeviceScan::ExclusiveScan(
+                        workspace.Write(), bytes, d_in, d_out, scan_op, init_value,
+                        num_items));
+      return;
+   }
+#endif
+   std::exclusive_scan(d_in, d_in + num_items, d_out, init_value, scan_op);
+}
+
+/// Equivalent to ExclusiveScan(use_dev, d_in, d_out, num_items, init_value,
+/// workspace, std::plus<>{})
+template <class InputIt, class OutputIt, class T>
+void ExclusiveScan(bool use_dev, InputIt d_in, OutputIt d_out, size_t num_items,
+                   T init_value, Array<char> &workspace)
+{
+   ExclusiveScan(use_dev, d_in, d_out, num_items, init_value, workspace,
+                 std::plus<> {});
+}
+
+} // namespace mfem
+
+#undef MFEM_CUB_NAMESPACE
+
+#endif

linalg/vector.cpp

@@ -14,6 +14,7 @@
 #include "../general/forall.hpp"
 #include "../general/reducers.hpp"
 #include "../general/hash.hpp"
+#include "../general/scan.hpp"
 #include "vector.hpp"
 
 #ifdef MFEM_USE_OPENMP
@@ -117,7 +118,7 @@ Vector::Vector(const Vector &v)
    UseDevice(v.UseDevice());
 }
 
-Vector::Vector(Vector &&v)
+Vector::Vector(Vector &&v) : Vector()
 {
    *this = std::move(v);
 }
@@ -1252,4 +1253,44 @@ real_t Vector::Sum() const
    return res;
 }
 
+void Vector::DeleteAt(const Array<int> &indices)
+{
+   const bool use_dev = UseDevice();
+
+   Array<int> flag(size);
+   const auto d_flag = flag.Write(use_dev);
+   mfem::forall_switch(use_dev, size, [=] MFEM_HOST_DEVICE (int i)
+   {
+      d_flag[i] = true;
+   });
+   const auto d_indices = indices.Read(use_dev);
+   mfem::forall_switch(use_dev, indices.Size(), [=] MFEM_HOST_DEVICE (int i)
+   {
+      d_flag[d_indices[i]] = false;
+   });
+
+   Array<int> out_idx(size);
+   auto d_out_idx = out_idx.Write(use_dev);
+   Array<char> workspace;
+   // Perform inclusive scan so that the last entry is the new size.
+   InclusiveScan(use_dev, d_flag, d_out_idx, size, workspace);
+
+   Vector copy(*this);
+   auto d_in = copy.Read(use_dev);
+   auto d_out = Write(use_dev);
+   mfem::forall_switch(use_dev, size, [=] MFEM_HOST_DEVICE (int i)
+   {
+      if (d_flag[i])
+      {
+         // Transform inclusive scan to exclusive by shifting.
+         const int j = (i > 0) ? d_out_idx[i - 1] : 0;
+         d_out[j] = d_in[i];
+      }
+   });
+
+   // Get the new size of the vector. Copy only the last entry.
+   Memory<int> submem(out_idx.GetMemory(), out_idx.Size() - 1, 1);
+   size = submem.Read(MemoryClass::HOST, 1)[0];
+}
+
 } // namespace mfem

linalg/vector.hpp

@@ -171,6 +171,12 @@ public:
    /// Resize the vector to size @a s using the MemoryType of @a v.
    void SetSize(int s, const Vector &v) { SetSize(s, v.GetMemory().GetMemoryType()); }
 
+   /// Update \ref Capacity() to @a res (if less than current), keeping existing entries.
+   void Reserve(int res);
+
+   /// Delete entries at @a indices and resize vector accordingly.
+   void DeleteAt(const Array<int> &indices);
+
    /// Set the Vector data.
    /// @warning This method should be called only when OwnsData() is false.
    void SetData(real_t *d) { data.Wrap(d, data.Capacity(), false); }
@@ -621,6 +627,18 @@ inline void Vector::SetSize(int s, MemoryType mt)
    data.UseDevice(use_dev);
 }
 
+inline void Vector::Reserve(int res)
+{
+   if (res > Capacity())
+   {
+      Memory<real_t> p(res, data.GetMemoryType());
+      p.CopyFrom(data, size);
+      p.UseDevice(data.UseDevice());
+      data.Delete();
+      data = p;
+   }
+}
+
 inline void Vector::NewMemoryAndSize(const Memory<real_t> &mem, int s,
                                      bool own_mem)
 {

mesh/mesh.cpp

@@ -32,6 +32,7 @@
 #include <cstring>
 #include <ctime>
 #include <functional>
+#include <set>
 #include <numeric>
 #include <unordered_map>
 #include <unordered_set>
@@ -921,6 +922,103 @@ const FaceGeometricFactors* Mesh::GetFaceGeometricFactors(
    return gf;
 }
 
+const Array<int>& Mesh::GetBdrFaceAttributes() const
+{
+   if (bdr_face_attrs_cache.Size() == 0)
+   {
+      std::unordered_map<int, int> f_to_be;
+      for (int i = 0; i < GetNBE(); ++i)
+      {
+         const int f = GetBdrElementFaceIndex(i);
+         f_to_be[f] = i;
+      }
+      const int nf_bdr = GetNFbyType(FaceType::Boundary);
+      // MFEM_VERIFY(size_t(nf_bdr) == f_to_be.size(), "Incompatible sizes");
+      bdr_face_attrs_cache.SetSize(nf_bdr);
+      int f_ind = 0;
+      const int nf = GetNumFaces();
+      for (int f = 0; f < nf; ++f)
+      {
+         if (!GetFaceInformation(f).IsOfFaceType(FaceType::Boundary))
+         {
+            continue;
+         }
+         int attribute = -1; // default value
+         auto iter = f_to_be.find(f);
+         if (iter != f_to_be.end())
+         {
+            const int be = iter->second;
+            attribute = GetBdrAttribute(be);
+         }
+         else
+         {
+            // If a boundary face does not correspond to the a boundary element,
+            // we assign it the default attribute of -1.
+         }
+         bdr_face_attrs_cache[f_ind] = attribute;
+         ++f_ind;
+      }
+   }
+   return bdr_face_attrs_cache;
+}
+
+const Array<int>& Mesh::GetElementAttributes() const
+{
+   if (elem_attrs_cache.Size() == 0)
+   {
+      // re-compute cache
+      elem_attrs_cache.SetSize(GetNE());
+      elem_attrs_cache.HostWrite();
+      for (int i = 0; i < GetNE(); ++i)
+      {
+         elem_attrs_cache[i] = GetAttribute(i);
+         MFEM_ASSERT(elem_attrs_cache[i] > 0,
+                     "Negative attribute on element " << i);
+      }
+   }
+   return elem_attrs_cache;
+}
+
+void Mesh::ComputeFaceInfo(FaceType ftype) const
+{
+   auto &fidcs = face_indices[static_cast<int>(ftype)];
+   auto &ifidcs = inv_face_indices[static_cast<int>(ftype)];
+   fidcs.SetSize(GetNFbyType(ftype));
+   fidcs.HostWrite();
+   ifidcs.reserve(fidcs.Size());
+   int f_idx = 0;
+   for (int i = 0; i < GetNumFacesWithGhost(); ++i)
+   {
+      const FaceInformation face = GetFaceInformation(i);
+      if (face.IsNonconformingCoarse() || !face.IsOfFaceType(ftype))
+      {
+         continue;
+      }
+      fidcs[f_idx] = i;
+      ifidcs[i] = f_idx;
+      ++f_idx;
+   }
+}
+
+const Array<int> &Mesh::GetFaceIndices(FaceType ftype) const
+{
+   if (face_indices[static_cast<int>(ftype)].Size() == 0)
+   {
+      ComputeFaceInfo(ftype);
+   }
+   return face_indices[static_cast<int>(ftype)];
+}
+
+const std::unordered_map<int, int> &
+Mesh::GetInvFaceIndices(FaceType ftype) const
+{
+   if (inv_face_indices[static_cast<int>(ftype)].empty())
+   {
+      ComputeFaceInfo(ftype);
+   }
+   return inv_face_indices[static_cast<int>(ftype)];
+}
+
 void Mesh::DeleteGeometricFactors()
 {
    for (int i = 0; i < geom_factors.Size(); i++)
@@ -1804,8 +1902,15 @@ void Mesh::Destroy()
    TetMemory.Clear();
 #endif
 
+   elem_attrs_cache.DeleteAll();
+   bdr_face_attrs_cache.DeleteAll();
    attributes.DeleteAll();
    bdr_attributes.DeleteAll();
+
+   face_indices[0].DeleteAll();
+   face_indices[1].DeleteAll();
+   inv_face_indices[0] = std::unordered_map<int, int>();
+   inv_face_indices[1] = std::unordered_map<int, int>();
 }
 
 void Mesh::ResetLazyData()
@@ -1817,36 +1922,48 @@ void Mesh::ResetLazyData()
    DeleteGeometricFactors();
    nbInteriorFaces = -1;
    nbBoundaryFaces = -1;
+   // set size to 0 so re-computations can potentially avoid a new allocation
+   bdr_face_attrs_cache.SetSize(0);
+   elem_attrs_cache.SetSize(0);
 }
 
-void Mesh::SetAttributes()
+void Mesh::SetAttributes(bool elem_attrs_changed, bool bdr_face_attrs_changed)
 {
-   Array<int> attribs;
+   if (bdr_face_attrs_changed)
+   {
+      bdr_face_attrs_cache.SetSize(0); // Invalidate the cache
 
-   attribs.SetSize(GetNBE());
-   for (int i = 0; i < attribs.Size(); i++)
-   {
-      attribs[i] = GetBdrAttribute(i);
-   }
-   attribs.Sort();
-   attribs.Unique();
-   attribs.Copy(bdr_attributes);
-   if (bdr_attributes.Size() > 0 && bdr_attributes[0] <= 0)
-   {
-      MFEM_WARNING("Non-positive attributes on the boundary!");
+      // Get sorted list of unique boundary element attributes
+      std::set<int> attribs;
+      for (int i = 0; i < GetNBE(); i++)
+      {
+         attribs.emplace(GetBdrAttribute(i));
+      }
+
+      bdr_attributes.SetSize(attribs.size());
+      bdr_attributes.HostWrite();
+      std::copy(attribs.begin(), attribs.end(), bdr_attributes.begin());
+      if (bdr_attributes.Size() > 0 && bdr_attributes[0] <= 0)
+      {
+         MFEM_WARNING("Non-positive attributes on the boundary!");
+      }
    }
 
-   attribs.SetSize(GetNE());
-   for (int i = 0; i < attribs.Size(); i++)
+   if (elem_attrs_changed)
    {
-      attribs[i] = GetAttribute(i);
-   }
-   attribs.Sort();
-   attribs.Unique();
-   attribs.Copy(attributes);
-   if (attributes.Size() > 0 && attributes[0] <= 0)
-   {
-      MFEM_WARNING("Non-positive attributes in the domain!");
+      // Re-compute the attributes cache
+      elem_attrs_cache.SetSize(0);
+      GetElementAttributes();
+      // Get sorted list of unique element attributes
+      std::set<int> attribs(elem_attrs_cache.begin(), elem_attrs_cache.end());
+      attributes.SetSize(attribs.size());
+      attributes.HostWrite();
+      std::copy(attribs.begin(), attribs.end(), attributes.begin());
+
+      if (attributes.Size() > 0 && attributes[0] <= 0)
+      {
+         MFEM_WARNING("Non-positive attributes in the domain!");
+      }
    }
 }
 
@@ -4454,6 +4571,10 @@ Mesh::Mesh(const Mesh &mesh, bool copy_nodes)
       Nodes = mesh.Nodes;
       own_nodes = 0;
    }
+
+   // copy attribute caches
+   elem_attrs_cache = mesh.elem_attrs_cache;
+   bdr_face_attrs_cache = mesh.bdr_face_attrs_cache;
 }
 
 Mesh::Mesh(Mesh &&mesh) : Mesh()
@@ -7721,6 +7842,12 @@ void Mesh::GetBdrElementAdjacentElement2(
 void Mesh::SetAttribute(int i, int attr)
 {
   elements[i]->SetAttribute(attr);
+  if (elem_attrs_cache.Size() == GetNE())
+  {
+     // update the existing cache instead of deleting it
+     elem_attrs_cache.HostReadWrite();
+     elem_attrs_cache[i] = attr;
+  }
   if (ncmesh) ncmesh->SetAttribute(i, attr);
 }
 
@@ -8055,6 +8182,12 @@ void Mesh::GenerateFaces()
       FreeElement(f);
    }
 
+   // delete caches
+   face_indices[0].SetSize(0);
+   face_indices[1].SetSize(0);
+   inv_face_indices[0].clear();
+   inv_face_indices[1].clear();
+
    // (re)generate the interior faces and the info for them
    faces.SetSize(nfaces);
    faces_info.SetSize(nfaces);
@@ -10850,6 +10983,15 @@ void Mesh::Swap(Mesh& other, bool non_geometry)
       mfem::Swap(nodes_sequence, other.nodes_sequence);
       mfem::Swap(last_operation, other.last_operation);
    }
+
+   // copy attribute caches
+   mfem::Swap(elem_attrs_cache, other.elem_attrs_cache);
+   mfem::Swap(bdr_face_attrs_cache, other.bdr_face_attrs_cache);
+
+   mfem::Swap(face_indices[0], other.face_indices[0]);
+   mfem::Swap(face_indices[1], other.face_indices[1]);
+   inv_face_indices[0].swap(other.inv_face_indices[0]);
+   inv_face_indices[1].swap(other.inv_face_indices[1]);
 }
 
 void Mesh::GetElementData(const Array<Element*> &elem_array, int geom,
@@ -13257,7 +13399,7 @@ void Mesh::ScaleElements(real_t sf)
    delete [] vn;
 }
 
-void Mesh::Transform(void (*f)(const Vector&, Vector&))
+void Mesh::Transform(std::function<void(const Vector &, Vector&)> f)
 {
    // TODO: support for different new spaceDim.
    if (Nodes == NULL)
@@ -13270,7 +13412,7 @@ void Mesh::Transform(void (*f)(const Vector&, Vector&))
             vold(j) = vertices[i](j);
          }
          vnew.SetData(vertices[i]());
-         (*f)(vold, vnew);
+         f(vold, vnew);
       }
    }
    else

mesh/mesh.hpp

@@ -107,6 +107,11 @@ protected:
    Array<Element *> boundary;
    Array<Element *> faces;
 
+   /// internal cache for element attributes
+   mutable Array<int> elem_attrs_cache;
+   /// internal cache for boundary element attributes
+   mutable Array<int> bdr_face_attrs_cache;
+
    /** @brief This structure stores the low level information necessary to
        interpret the configuration of elements on a specific face. This
        information can be accessed using methods like GetFaceElements(),
@@ -273,6 +278,13 @@ protected:
 
    // used during NC mesh initialization only
    Array<Triple<int, int, int> > tmp_vertex_parents;
+   /// cache for FaceIndices(ftype)
+   mutable Array<int> face_indices[2];
+   /// cache for FaceIndices(ftype)
+   mutable std::unordered_map<int, int> inv_face_indices[2];
+
+   /// compute face_indices[ftype] and inv_face_indices[type]
+   void ComputeFaceInfo(FaceType ftype) const;
 
 public:
    typedef Geometry::Constants<Geometry::SEGMENT>     seg_t;
@@ -307,6 +319,11 @@ public:
    // (true) is set in mesh_readers.cpp.
    static bool remove_unused_vertices;
 
+   /// Map from boundary or interior face indices to mesh face indices.
+   const Array<int>& GetFaceIndices(FaceType ftype) const;
+   /// Inverse of the map FaceIndices(ftype)
+   const std::unordered_map<int, int>& GetInvFaceIndices(FaceType ftype) const;
+
 protected:
    Operation last_operation;
 
@@ -1122,13 +1139,14 @@ public:
        Mesh vertices or nodes are set. */
    virtual void Finalize(bool refine = false, bool fix_orientation = false);
 
-   /// @brief Determine the sets of unique attribute values in domain and
-   /// boundary elements.
+   /// @brief Determine the sets of unique attribute values in domain if @a
+   /// elem_attrs_changed and boundary elements if @a bdr_face_attrs_changed.
    ///
    /// Separately scan the domain and boundary elements to generate unique,
    /// sorted sets of the element attribute values present in the mesh and
    /// store these in the Mesh::attributes and Mesh::bdr_attributes arrays.
-   virtual void SetAttributes();
+   virtual void SetAttributes(bool elem_attrs_changed = true,
+                              bool bdr_face_attrs_changed = true);
 
    /// Check (and optionally attempt to fix) the orientation of the elements
    /** @param[in] fix_it  If `true`, attempt to fix the orientations of some
@@ -2254,7 +2272,7 @@ public:
    void ScaleSubdomains (real_t sf);
    void ScaleElements (real_t sf);
 
-   void Transform(void (*f)(const Vector&, Vector&));
+   void Transform(std::function<void(const Vector &, Vector&)> f);
    void Transform(VectorCoefficient &deformation);
 
    /** @brief This function should be called after the mesh node coordinates
@@ -2267,6 +2285,35 @@ public:
        method does not modify the nodes. */
    void NodesUpdated() { DeleteGeometricFactors(); }
 
+   /// @brief Returns the attributes for all elements in this mesh. The i'th
+   /// entry of the array is the attribute of the i'th element of the mesh.
+   ///
+   /// The returned array points to an internal object that may be invalidated
+   /// by mesh operations such as refinement or any element attributes are
+   /// modified. Since not all such modifications can be tracked by the Mesh
+   /// class (e.g. if a user calls GetElement() then changes the element
+   /// attribute directly), one needs to account for such changes by calling the
+   /// method SetAttributes().
+   const Array<int>& GetElementAttributes() const;
+
+   /// @brief Returns the attributes for all boundary elements in this mesh.
+   ///
+   /// The face restriction will give "face E-vectors" on the boundary that
+   /// are numbered in the order of the faces of mesh. This numbering will be
+   /// different than the numbering of the boundary elements. We compute
+   /// mappings so that the array `bdr_attributes[i]` gives the boundary
+   /// attribute of the `i`th boundary face in the mesh face order.
+   /// Attributes <= 0 indicate there is no boundary element and should be
+   /// skipped.
+   ///
+   /// The returned array points to an internal object that may be invalidated
+   /// by mesh operations such as refinement or any element attributes are
+   /// modified. Since not all such modifications can be tracked by the Mesh
+   /// class (e.g. if a user calls GetElement() then changes the element
+   /// attribute directly), one needs to account for such changes by calling the
+   /// method SetAttributes().
+   const Array<int>& GetBdrFaceAttributes() const;
+
    /// @}
 
    /// @anchor mfem_Mesh_gf_nodes

mesh/pmesh.cpp

@@ -1590,21 +1590,27 @@ void ParMesh::DistributeAttributes(Array<int> &attr)
    delete [] glb_attr_marker;
 }
 
-void ParMesh::SetAttributes()
+void ParMesh::SetAttributes(bool elem_attrs_changed, bool bdr_attrs_changed)
 {
    // Determine the attributes occurring in local interior and boundary elements
-   Mesh::SetAttributes();
+   Mesh::SetAttributes(elem_attrs_changed, bdr_attrs_changed);
 
-   DistributeAttributes(bdr_attributes);
-   if (bdr_attributes.Size() > 0 && bdr_attributes[0] <= 0)
+   if (bdr_attrs_changed)
    {
-      MFEM_WARNING("Non-positive boundary element attributes found!");
+      DistributeAttributes(bdr_attributes);
+      if (bdr_attributes.Size() > 0 && bdr_attributes[0] <= 0)
+      {
+         MFEM_WARNING("Non-positive boundary element attributes found!");
+      }
    }
 
-   DistributeAttributes(attributes);
-   if (attributes.Size() > 0 && attributes[0] <= 0)
+   if (elem_attrs_changed)
    {
-      MFEM_WARNING("Non-positive element attributes found!");
+      DistributeAttributes(attributes);
+      if (attributes.Size() > 0 && attributes[0] <= 0)
+      {
+         MFEM_WARNING("Non-positive element attributes found!");
+      }
    }
 }
 

mesh/pmesh.hpp

@@ -394,7 +394,8 @@ public:
 
    void Finalize(bool refine = false, bool fix_orientation = false) override;
 
-   void SetAttributes() override;
+   void SetAttributes(bool elem_attrs_changed = true,
+                      bool bdr_attrs_changed = true) override;
 
    /// Checks if any rank in the mesh has boundary elements
    bool HasBoundaryElements() const override;

miniapps/performance/makefile

@@ -105,8 +105,11 @@ MFEM_PERF_CXXFLAGS_xlc = -mcpu=native
 # - Clang extra options:
 ifeq ($(MFEM_MACHINE),riscv64)
   MFEM_PERF_CXXFLAGS_clang += -march=rv64gc
-else ifneq ($(MFEM_MACHINE),arm64)
-  # -march=native is unavailable on clang/ARM64 as of 05/2021: support could be added later.
+else ifneq (,$(findstring ppc,$(MFEM_MACHINE)))
+  MFEM_PERF_CXXFLAGS_clang += -mcpu=native -mtune=native
+else ifeq ($(MFEM_MACHINE),arm64)
+  MFEM_PERF_CXXFLAGS_clang += -mcpu=native -mtune=native
+else
   MFEM_PERF_CXXFLAGS_clang += -march=native
 endif
 MFEM_PERF_CXXFLAGS_clang += $(PEDANTIC_FLAG) -Wall

tests/unit/CMakeLists.txt

@@ -22,6 +22,7 @@ set(UNIT_TESTS_SRCS
   dfem/test_mass.cpp
   general/test_array.cpp
   general/test_reduction.cpp
+  general/test_scan.cpp
   general/test_arrays_by_name.cpp
   general/test_error.cpp
   general/test_mem.cpp

tests/unit/fem/test_derefine.cpp

@@ -235,7 +235,7 @@ void test_derefine_L2_element(int order, Element::Type el_type, int basis_type)
    }
 }
 
-TEST_CASE("AMR Coarsen L2 Element","[AMR][Coarsen]")
+TEST_CASE("AMR Coarsen L2 Element","[AMR][Coarsen][CUDA]")
 {
    std::vector<int> orders_1d{0,1,2,3};
    std::vector<int> orders_2d{0,1,2,3};
@@ -420,7 +420,7 @@ void stress_parallel_coarsen(int order, Element::Type el_type, int basis_type)
 #endif
 }
 
-TEST_CASE("Parallel AMR Coarsen Stress Test", "[AMR][Coarsen][Parallel]")
+TEST_CASE("Parallel AMR Coarsen Stress Test", "[AMR][Coarsen][Parallel][CUDA]")
 {
    std::vector<int> orders_1d{0,1,2,3};
    std::vector<int> orders_2d{0,1,2,3};
@@ -482,7 +482,7 @@ TEST_CASE("Parallel AMR Coarsen Stress Test", "[AMR][Coarsen][Parallel]")
 
 }
 
-TEST_CASE("ParDerefine", "[Parallel]")
+TEST_CASE("ParDerefine", "[Parallel][CUDA]")
 {
    for (dimension = 2; dimension <= 3; ++dimension)
    {

tests/unit/general/test_array.cpp

@@ -124,3 +124,19 @@ TEST_CASE("Array stl-interactions", "[Array]")
       CHECK(x[i] == y[i]);
    }
 }
+
+TEST_CASE("Array delete at indices", "[Array]")
+{
+   Array<int>       test({0,1,2,3,4,5,6,7,8});
+   Array<int> rm_indices({0,    3,4,  6,  8});
+   Array<int>     result({  1,2,    5,  7  });
+
+   test.DeleteAt(rm_indices);
+
+   REQUIRE(test.Size() == result.Size());
+
+   for (int i = 0; i < test.Size(); i++)
+   {
+      CHECK(test[i] == result[i]);
+   }
+}

tests/unit/general/test_scan.cpp

@@ -0,0 +1,102 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#include <algorithm>
+#include <limits>
+
+#include "mfem.hpp"
+#include "unit_tests.hpp"
+
+// must be included after mfem.hpp
+#include "general/scan.hpp"
+
+using namespace mfem;
+
+TEST_CASE("Inclusive Scan", "[Scan],[GPU]")
+{
+   Array<char> workspace;
+   Array<int> a(10);
+
+   for (int use_dev = 0; use_dev < 2; ++use_dev)
+   {
+      CAPTURE(use_dev);
+      a.HostReadWrite();
+      for (int i = 0; i < a.Size(); ++i)
+      {
+         a[i] = i;
+      }
+      auto dptr = a.ReadWrite(use_dev);
+      InclusiveScan(use_dev, dptr, dptr, a.Size(), workspace);
+      a.HostRead();
+      for (int i = 0; i < a.Size(); ++i)
+      {
+         int expected = (i + 1) * i / 2;
+         CAPTURE(i);
+         REQUIRE(a[i] == expected);
+      }
+      a.HostReadWrite();
+      for (int i = 0; i < a.Size(); ++i)
+      {
+         a[i] = i + 1;
+      }
+      a.ReadWrite(use_dev);
+      InclusiveScan(use_dev, dptr, dptr, a.Size(), workspace, std::multiplies<> {});
+      a.HostRead();
+      int expected = 1;
+      for (int i = 0; i < a.Size(); ++i)
+      {
+         expected *= i + 1;
+         CAPTURE(i);
+         REQUIRE(a[i] == expected);
+      }
+   }
+}
+
+TEST_CASE("Exclusive Scan", "[Scan],[GPU]")
+{
+   Array<char> workspace;
+   Array<int> a(10);
+
+   for (int use_dev = 0; use_dev < 2; ++use_dev)
+   {
+      CAPTURE(use_dev);
+      a.HostReadWrite();
+      for (int i = 0; i < a.Size(); ++i)
+      {
+         a[i] = i;
+      }
+      auto dptr = a.ReadWrite(use_dev);
+      ExclusiveScan(use_dev, dptr, dptr, a.Size(), 5, workspace);
+      a.HostRead();
+      for (int i = 0; i < a.Size(); ++i)
+      {
+         int expected = (i + 1) * i / 2 - i + 5;
+         CAPTURE(i);
+         REQUIRE(a[i] == expected);
+      }
+      a.HostReadWrite();
+      for (int i = 0; i < a.Size(); ++i)
+      {
+         a[i] = i + 1;
+      }
+      a.ReadWrite(use_dev);
+      ExclusiveScan(use_dev, dptr, dptr, a.Size(), 5, workspace,
+                    std::multiplies<> {});
+      a.HostRead();
+      int expected = 5;
+      for (int i = 0; i < a.Size(); ++i)
+      {
+         CAPTURE(i);
+         REQUIRE(a[i] == expected);
+         expected *= i + 1;
+      }
+   }
+}

tests/unit/linalg/test_vector.cpp

@@ -247,3 +247,21 @@ TEST_CASE("Vector Sum", "[Vector],[GPU]")
 
    REQUIRE(sum_1 == MFEM_Approx(sum_2));
 }
+
+TEST_CASE("Vector delete at indices", "[Vector][GPU]")
+{
+   Vector           test({0,1,2,3,4,5,6,7,8});
+   Array<int> rm_indices({0,    3,4,  6,  8});
+   Vector         result({  1,2,    5,  7  });
+
+   test.UseDevice(true);
+   test.DeleteAt(rm_indices);
+
+   REQUIRE(test.Size() == result.Size());
+
+   test.HostReadWrite();
+   for (int i = 0; i < test.Size(); i++)
+   {
+      CHECK(test[i] == result[i]);
+   }
+}