Fix a small issue in miniapps/shifted/distance.cpp

in supported cases.

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

@@ -289,10 +289,10 @@ 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
@@ -411,6 +411,8 @@ miniapps/tribol/contact-patch-test
 miniapps/diag-smoothers/abs-l1-jacobi
 miniapps/diag-smoothers/mg-abs-l1-jacobi
 
+miniapps/benchmarks/ceed-solver-bps/solver-bp
+
 # Unit test binary and outputs
 tests/unit/output_meshes
 tests/unit/unit_tests

.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
 -------------------------------------
@@ -75,6 +80,13 @@ Miscellaneous
   variable is an alternative to calling 'Device::SetGPUAwareMPI(true)'.
 - Added parallel Address Sanitizer, serial and parallel Undefined Behavior
   Sanitizer and serial Memory Sanitizer GitHub actions tests on Ubuntu.
+- MFEM_PERF_* annotations: added options to enable GPU-stream- and
+  MPI-synchronizations at the start and at the end of annotation regions. These
+  synchronizations can be enabled or disabled (default) in code via the new
+  macros: MFEM_PERF_SYNC, MFEM_PERF_SYNC_STREAM, and MFEM_PERF_SYNC_MPI; the
+  environment variables with the same names can be set to 0/1 to control the
+  synchronization as well.
+
 
 Version 4.8, released on Apr 9, 2025
 ====================================

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++)

examples/ex26.cpp

@@ -78,6 +78,7 @@ private:
       opr.SetOperatorOwner(false);
 
       CGSolver* pcg = new CGSolver();
+      // pcg->iterative_mode = false; // the multigrid algorithm does this
       pcg->SetPrintLevel(-1);
       pcg->SetMaxIter(200);
       pcg->SetRelTol(sqrt(1e-4));
@@ -100,7 +101,8 @@ private:
       Vector diag(fespace.GetTrueVSize());
       bfs[level]->AssembleDiagonal(diag);
 
-      Solver* smoother = new OperatorChebyshevSmoother(*opr, diag, ess_tdof_list, 2);
+      Solver *smoother = new OperatorChebyshevSmoother(
+         *opr, diag, ess_tdof_list, 2);
       AddLevel(opr.Ptr(), smoother, true, true);
    }
 };

examples/ex26p.cpp

@@ -88,6 +88,7 @@ private:
       amg->SetPrintLevel(-1);
 
       CGSolver* pcg = new CGSolver(MPI_COMM_WORLD);
+      // pcg->iterative_mode = false; // the multigrid algorithm does this
       pcg->SetPrintLevel(-1);
       pcg->SetMaxIter(10);
       pcg->SetRelTol(sqrt(1e-4));

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_ext.cpp

@@ -255,6 +255,8 @@ PABilinearFormExtension::PABilinearFormExtension(BilinearForm *form)
 
 void PABilinearFormExtension::SetupRestrictionOperators(const L2FaceValues m)
 {
+   MFEM_PERF_FUNCTION;
+
    if ( Device::Allows(Backend::CEED_MASK) ) { return; }
    ElementDofOrdering ordering = GetEVectorOrdering(*a->FESpace());
    elem_restrict = trial_fes->GetElementRestriction(ordering);
@@ -266,11 +268,7 @@ 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
@@ -329,50 +327,14 @@ 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();
    }
 }
 
 void PABilinearFormExtension::Assemble()
 {
+   MFEM_PERF_FUNCTION;
+
    SetupRestrictionOperators(L2FaceValues::DoubleValued);
 
    Array<BilinearFormIntegrator*> &integrators = *a->GetDBFI();
@@ -429,7 +391,7 @@ 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 +412,7 @@ 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 +438,7 @@ 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 +451,7 @@ 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);
    }
@@ -529,6 +491,8 @@ void PABilinearFormExtension::FormLinearSystem(const Array<int> &ess_tdof_list,
 void PABilinearFormExtension::MultInternal(const Vector &x, Vector &y,
                                            const bool useAbs) const
 {
+   MFEM_PERF_FUNCTION;
+
    Array<BilinearFormIntegrator*> &integrators = *a->GetDBFI();
 
    const int iSz = integrators.Size();
@@ -588,7 +552,7 @@ 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 +654,8 @@ 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 +663,14 @@ 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 +693,7 @@ 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 +740,14 @@ 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 +771,7 @@ 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);
@@ -898,6 +865,7 @@ EABilinearFormExtension::EABilinearFormExtension(BilinearForm *form)
 
 void EABilinearFormExtension::Assemble()
 {
+   MFEM_PERF_FUNCTION;
    SetupRestrictionOperators(L2FaceValues::SingleValued);
 
    ne = trial_fes->GetMesh()->GetNE();
@@ -920,7 +888,8 @@ 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 +922,7 @@ 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 +951,7 @@ 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 +1000,7 @@ 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);
          }
       }
    }
@@ -1445,6 +1414,7 @@ FABilinearFormExtension::FABilinearFormExtension(BilinearForm *form)
 
 void FABilinearFormExtension::Assemble()
 {
+   MFEM_PERF_FUNCTION;
    EABilinearFormExtension::Assemble();
    FiniteElementSpace &fes = *a->FESpace();
    int width = fes.GetVSize();

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.hpp

@@ -812,7 +812,7 @@ 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 +884,7 @@ 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 +919,7 @@ 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 +1600,7 @@ 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 +1635,7 @@ 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 +1669,7 @@ 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 +1760,7 @@ 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 +1793,7 @@ 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 +1832,7 @@ 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 +1973,7 @@ 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 );
    }
 

fem/coefficient.cpp

@@ -50,6 +50,7 @@ ElementTransformation *RefinedToCoarse(
 
 void Coefficient::Project(QuadratureFunction &qf)
 {
+   MFEM_PERF_FUNCTION;
    QuadratureSpaceBase &qspace = *qf.GetSpace();
    const int ne = qspace.GetNE();
    Vector values;

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),
@@ -98,7 +101,10 @@ FiniteElementSpace::FiniteElementSpace(const FiniteElementSpace &orig,
 FiniteElementSpace::FiniteElementSpace(Mesh *mesh,
                                        const FiniteElementCollection *fec,
                                        int vdim, int ordering)
-{ Constructor(mesh, NULL, fec, vdim, ordering); }
+{
+   MFEM_PERF_FUNCTION;
+   Constructor(mesh, NULL, fec, vdim, ordering);
+}
 
 FiniteElementSpace::FiniteElementSpace(Mesh *mesh, NURBSExtension *ext,
                                        const FiniteElementCollection *fec,
@@ -390,6 +396,8 @@ void FiniteElementSpace::BuildElementToDofTable() const
 {
    if (elem_dof) { return; }
 
+   MFEM_PERF_FUNCTION;
+
    // TODO: can we call GetElementDofs only once per element?
    Table *el_dof = new Table;
    Table *el_fos = (mesh->Dimension() > 2) ? (new Table) : NULL;
@@ -2745,6 +2753,8 @@ void FiniteElementSpace::BuildNURBSFaceToDofTable() const
 
 void FiniteElementSpace::Construct()
 {
+   MFEM_PERF_FUNCTION;
+
    // This method should be used only for non-NURBS spaces.
    MFEM_VERIFY(!NURBSext, "internal error");
 
@@ -4244,7 +4254,11 @@ 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/gridfunc.cpp

@@ -19,6 +19,7 @@
 #include "../mesh/nurbs.hpp"
 #include "../mesh/vtkhdf.hpp"
 #include "../general/text.hpp"
+#include "../general/reducers.hpp"
 
 #ifdef MFEM_USE_MPI
 #include "pfespace.hpp"
@@ -3326,8 +3327,126 @@ real_t GridFunction::ComputeLpError(const real_t p, Coefficient &exsol,
                                     const IntegrationRule *irs[],
                                     const Array<int> *elems) const
 {
+   MFEM_PERF_FUNCTION;
+
+   MFEM_VERIFY(fes->GetVDim() == 1, "invalid vector dimension!");
+
    real_t error = 0.0;
-   const FiniteElement *fe;
+
+   bool device_eval = true;
+   // TODO: check for cases that are not supported on device:
+   //       * mixed meshes
+   //       * meshes with non-tensor-product elements can have negative weights
+   //       * variable orders
+   //       * weight is not NULL
+   //       * elems is not NULL
+   //       * map type is not VALUE
+   //       * ...
+   Mesh *mesh = fes->GetMesh();
+   const FiniteElement *fe = fes->GetTypicalFE();
+   if (mesh->GetNumGeometries(mesh->Dimension()) > 1 ||
+       (mesh->Dimension() > 1 && mesh->MeshGenerator() != 2) ||
+       fes->IsVariableOrder() ||
+       weight != nullptr ||
+       elems != nullptr ||
+       fe->GetMapType() != FiniteElement::MapType::VALUE)
+   {
+      device_eval = false;
+   }
+   if (device_eval)
+   {
+      Geometry::Type geom = mesh->GetTypicalElementGeometry();
+      const IntegrationRule *ir_p;
+      if (irs)
+      {
+         ir_p = irs[geom];
+      }
+      else
+      {
+         int intorder = 2*fe->GetOrder() + 3; // <----------
+         ir_p = &(IntRules.Get(geom, intorder));
+      }
+      const IntegrationRule &ir = *ir_p;
+      QuadratureSpace qs(*mesh, ir);
+      CoefficientVector coeff(exsol, qs, CoefficientStorage::FULL);
+
+      const QVectorLayout ql = QVectorLayout::byNODES;
+      const MemoryType d_mt = MemoryType::DEFAULT;
+      Vector q_vals;
+      // TODO: make this a method
+      {
+         // const FiniteElement *fe = fes->GetTypicalFE();
+         const int vdim = fes->GetVDim();
+         const int NE   = fes->GetNE();
+         const int ND   = fe->GetDof();
+         const int NQ   = ir.GetNPoints();
+         MemoryType my_d_mt = (d_mt != MemoryType::DEFAULT) ? d_mt :
+                              Device::GetDeviceMemoryType();
+         // byNODES : NQPT x VDIM x NE
+         // byVDIM  : VDIM x NQPT x NE
+         q_vals.SetSize(vdim*NQ*NE, my_d_mt);
+         const QuadratureInterpolator &qi = *fes->GetQuadratureInterpolator(ir);
+         qi.SetOutputLayout(ql);
+         const bool use_tensor_products = UsesTensorBasis(*fes);
+         qi.DisableTensorProducts(!use_tensor_products);
+         const ElementDofOrdering e_ordering =
+            use_tensor_products ?
+            ElementDofOrdering::LEXICOGRAPHIC :
+            ElementDofOrdering::NATIVE;
+         const Operator *elem_restr = fes->GetElementRestriction(e_ordering);
+         if (fe->GetMapType() == FiniteElement::MapType::INTEGRAL)
+         {
+            // Pre-compute the geometric factors in order to set the desired
+            // MemoryType they use:
+            fes->GetMesh()->GetGeometricFactors(
+               ir, GeometricFactors::DETERMINANTS, my_d_mt);
+         }
+         if (elem_restr)
+         {
+            Vector f_e(vdim*ND*NE, my_d_mt);
+            elem_restr->Mult(*this, f_e);
+            qi.PhysValues(f_e, q_vals);
+         }
+         else
+         {
+            qi.PhysValues(*this, q_vals);
+         }
+      }
+
+      const real_t *exact_d = coeff.Read();
+      const real_t *gridf_d = q_vals.Read();
+      // FIXME: reuse the workspace vector from vector.cpp?
+      static Array<real_t> workspace;
+      if (p < infinity())
+      {
+         MemoryType my_d_mt = (d_mt != MemoryType::DEFAULT) ? d_mt :
+                              Device::GetDeviceMemoryType();
+         const GeometricFactors *geom_factors =
+            fes->GetMesh()->GetGeometricFactors(
+               ir, GeometricFactors::DETERMINANTS, my_d_mt);
+         const real_t *detJ_d = geom_factors->detJ.Read();
+         const real_t *w_d = ir.GetWeights().Read();
+         const int NQ = ir.GetNPoints();
+         mfem::reduce(q_vals.Size(), error,
+                      [=] MFEM_HOST_DEVICE(int i, real_t &r)
+         {
+            const real_t diff = fabs(exact_d[i] - gridf_d[i]);
+            r += w_d[i%NQ] * detJ_d[i] * pow(diff, p);
+         }, SumReducer<real_t> {}, true, workspace);
+         error = pow(error, 1./p);
+      }
+      else
+      {
+         mfem::reduce(q_vals.Size(), error,
+                      [=] MFEM_HOST_DEVICE(int i, real_t &r)
+         {
+            const real_t diff = fabs(exact_d[i] - gridf_d[i]);
+            r = fmax(r, diff);
+         }, MaxReducer<real_t> {}, true, workspace);
+      }
+      return error;
+   }
+
    ElementTransformation *T;
    Vector vals;
 

fem/integ/bilininteg_diffusion_ea.cpp

@@ -161,7 +161,8 @@ static void EADiffusionAssemble3D(const int NE,
    auto B = Reshape(b.Read(), Q1D, D1D);
    auto G = Reshape(g.Read(), Q1D, D1D);
    auto D = Reshape(padata.Read(), Q1D, Q1D, Q1D, 6, NE);
-   auto A = Reshape(eadata.ReadWrite(), D1D, D1D, D1D, D1D, D1D, D1D, NE);
+   auto A = Reshape(add ? eadata.ReadWrite() : eadata.Write(),
+                    D1D, D1D, D1D, D1D, D1D, D1D, NE);
    mfem::forall_3D(NE, D1D, D1D, D1D, [=] MFEM_HOST_DEVICE (int e)
    {
       const int D1D = T_D1D ? T_D1D : d1d;
@@ -246,58 +247,60 @@ void DiffusionIntegrator::AssembleEA(const FiniteElementSpace &fes,
                                      Vector &ea_data,
                                      const bool add)
 {
+   MFEM_PERF_FUNCTION;
    AssemblePA(fes);
    ne = fes.GetMesh()->GetNE();
    const Array<real_t> &B = maps->B;
    const Array<real_t> &G = maps->G;
+   decltype(&EADiffusionAssemble1D<>) kernel = nullptr;
    if (dim == 1)
    {
       switch ((dofs1D << 4 ) | quad1D)
       {
-         case 0x22: return EADiffusionAssemble1D<2,2>(ne,B,G,pa_data,ea_data,add);
-         case 0x33: return EADiffusionAssemble1D<3,3>(ne,B,G,pa_data,ea_data,add);
-         case 0x44: return EADiffusionAssemble1D<4,4>(ne,B,G,pa_data,ea_data,add);
-         case 0x55: return EADiffusionAssemble1D<5,5>(ne,B,G,pa_data,ea_data,add);
-         case 0x66: return EADiffusionAssemble1D<6,6>(ne,B,G,pa_data,ea_data,add);
-         case 0x77: return EADiffusionAssemble1D<7,7>(ne,B,G,pa_data,ea_data,add);
-         case 0x88: return EADiffusionAssemble1D<8,8>(ne,B,G,pa_data,ea_data,add);
-         case 0x99: return EADiffusionAssemble1D<9,9>(ne,B,G,pa_data,ea_data,add);
-         default:   return EADiffusionAssemble1D(ne,B,G,pa_data,ea_data,add,
-                                                    dofs1D,quad1D);
+         case 0x22: kernel = EADiffusionAssemble1D<2,2>;
+         case 0x33: kernel = EADiffusionAssemble1D<3,3>;
+         case 0x44: kernel = EADiffusionAssemble1D<4,4>;
+         case 0x55: kernel = EADiffusionAssemble1D<5,5>;
+         case 0x66: kernel = EADiffusionAssemble1D<6,6>;
+         case 0x77: kernel = EADiffusionAssemble1D<7,7>;
+         case 0x88: kernel = EADiffusionAssemble1D<8,8>;
+         case 0x99: kernel = EADiffusionAssemble1D<9,9>;
+         default:   kernel = EADiffusionAssemble1D<>;
       }
    }
    else if (dim == 2)
    {
       switch ((dofs1D << 4 ) | quad1D)
       {
-         case 0x22: return EADiffusionAssemble2D<2,2>(ne,B,G,pa_data,ea_data,add);
-         case 0x33: return EADiffusionAssemble2D<3,3>(ne,B,G,pa_data,ea_data,add);
-         case 0x44: return EADiffusionAssemble2D<4,4>(ne,B,G,pa_data,ea_data,add);
-         case 0x55: return EADiffusionAssemble2D<5,5>(ne,B,G,pa_data,ea_data,add);
-         case 0x66: return EADiffusionAssemble2D<6,6>(ne,B,G,pa_data,ea_data,add);
-         case 0x77: return EADiffusionAssemble2D<7,7>(ne,B,G,pa_data,ea_data,add);
-         case 0x88: return EADiffusionAssemble2D<8,8>(ne,B,G,pa_data,ea_data,add);
-         case 0x99: return EADiffusionAssemble2D<9,9>(ne,B,G,pa_data,ea_data,add);
-         default:   return EADiffusionAssemble2D(ne,B,G,pa_data,ea_data,add,
-                                                    dofs1D,quad1D);
+         case 0x22: kernel = EADiffusionAssemble2D<2,2>;
+         case 0x33: kernel = EADiffusionAssemble2D<3,3>;
+         case 0x44: kernel = EADiffusionAssemble2D<4,4>;
+         case 0x55: kernel = EADiffusionAssemble2D<5,5>;
+         case 0x66: kernel = EADiffusionAssemble2D<6,6>;
+         case 0x77: kernel = EADiffusionAssemble2D<7,7>;
+         case 0x88: kernel = EADiffusionAssemble2D<8,8>;
+         case 0x99: kernel = EADiffusionAssemble2D<9,9>;
+         default:   kernel = EADiffusionAssemble2D<>;
       }
    }
    else if (dim == 3)
    {
       switch ((dofs1D << 4 ) | quad1D)
       {
-         case 0x23: return EADiffusionAssemble3D<2,3>(ne,B,G,pa_data,ea_data,add);
-         case 0x34: return EADiffusionAssemble3D<3,4>(ne,B,G,pa_data,ea_data,add);
-         case 0x45: return EADiffusionAssemble3D<4,5>(ne,B,G,pa_data,ea_data,add);
-         case 0x56: return EADiffusionAssemble3D<5,6>(ne,B,G,pa_data,ea_data,add);
-         case 0x67: return EADiffusionAssemble3D<6,7>(ne,B,G,pa_data,ea_data,add);
-         case 0x78: return EADiffusionAssemble3D<7,8>(ne,B,G,pa_data,ea_data,add);
-         case 0x89: return EADiffusionAssemble3D<8,9>(ne,B,G,pa_data,ea_data,add);
-         default:   return EADiffusionAssemble3D(ne,B,G,pa_data,ea_data,add,
-                                                    dofs1D,quad1D);
+         case 0x23: kernel = EADiffusionAssemble3D<2,3>;
+         case 0x34: kernel = EADiffusionAssemble3D<3,4>;
+         case 0x45: kernel = EADiffusionAssemble3D<4,5>;
+         case 0x56: kernel = EADiffusionAssemble3D<5,6>;
+         case 0x67: kernel = EADiffusionAssemble3D<6,7>;
+         case 0x78: kernel = EADiffusionAssemble3D<7,8>;
+         case 0x89: kernel = EADiffusionAssemble3D<8,9>;
+         default:   kernel = EADiffusionAssemble3D<>;
       }
    }
-   MFEM_ABORT("Unknown kernel.");
+   MFEM_VERIFY(kernel != nullptr, "Unknown kernel.");
+   kernel(ne,B,G,pa_data,ea_data,add,dofs1D,quad1D);
+   // Free the PA data:
+   pa_data.Destroy();
 }
 
 }

fem/integ/bilininteg_diffusion_pa.cpp

@@ -39,6 +39,8 @@ void DiffusionIntegrator::AssembleDiagonalPA(Vector &diag)
 // PA Diffusion Apply kernel
 void DiffusionIntegrator::AddMultPA(const Vector &x, Vector &y) const
 {
+   MFEM_PERF_FUNCTION;
+
    if (DeviceCanUseCeed())
    {
       ceedOp->AddMult(x, y);
@@ -88,6 +90,8 @@ void DiffusionIntegrator::AddMultTransposePA(const Vector &x, Vector &y) const
 
 void DiffusionIntegrator::AssemblePA(const FiniteElementSpace &fes)
 {
+   MFEM_PERF_FUNCTION;
+
    const MemoryType mt = (pa_mt == MemoryType::DEFAULT) ?
                          Device::GetDeviceMemoryType() : pa_mt;
    // Assuming the same element type

fem/integ/bilininteg_mass_pa.cpp

@@ -23,6 +23,8 @@ namespace mfem
 
 void MassIntegrator::AssemblePA(const FiniteElementSpace &fes)
 {
+   MFEM_PERF_FUNCTION;
+
    const MemoryType mt = (pa_mt == MemoryType::DEFAULT) ?
                          Device::GetDeviceMemoryType() : pa_mt;
 
@@ -59,26 +61,23 @@ 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 +108,22 @@ 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)
@@ -170,6 +141,8 @@ void MassIntegrator::AssembleDiagonalPA(Vector &diag)
 
 void MassIntegrator::AddMultPA(const Vector &x, Vector &y) const
 {
+   MFEM_PERF_FUNCTION;
+
    if (DeviceCanUseCeed())
    {
       ceedOp->AddMult(x, y);

fem/integ/lininteg_domain.cpp

@@ -242,6 +242,8 @@ void DomainLFIntegrator::AssembleDevice(const FiniteElementSpace &fes,
                                         const Array<int> &markers,
                                         Vector &b)
 {
+   MFEM_PERF_FUNCTION;
+
    const FiniteElement &fe = *fes.GetTypicalFE();
    const int qorder = oa * fe.GetOrder() + ob;
    const Geometry::Type gtype = fe.GetGeomType();

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.cpp

@@ -161,6 +161,7 @@ bool LinearForm::SupportsDevice() const
 
 void LinearForm::UseFastAssembly(bool use_fa)
 {
+   MFEM_PERF_FUNCTION;
    fast_assembly = use_fa;
 
    if (fast_assembly && SupportsDevice() && !ext)
@@ -171,6 +172,8 @@ void LinearForm::UseFastAssembly(bool use_fa)
 
 void LinearForm::Assemble()
 {
+   MFEM_PERF_FUNCTION;
+
    Array<int> vdofs;
    ElementTransformation *eltrans;
    Vector elemvect;

fem/linearform_ext.cpp

@@ -15,10 +15,16 @@
 namespace mfem
 {
 
-LinearFormExtension::LinearFormExtension(LinearForm *lf): lf(lf) { Update(); }
+LinearFormExtension::LinearFormExtension(LinearForm *lf): lf(lf)
+{
+   MFEM_PERF_FUNCTION;
+   Update();
+}
 
 void LinearFormExtension::Assemble()
 {
+   MFEM_PERF_FUNCTION;
+
    const FiniteElementSpace &fes = *lf->FESpace();
    MFEM_VERIFY(lf->SupportsDevice(), "Not supported.");
    MFEM_VERIFY(lf->Size() == fes.GetVSize(), "LinearForm size does not "
@@ -51,7 +57,7 @@ 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 +99,14 @@ 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;
          });
       }
 
@@ -112,6 +119,8 @@ void LinearFormExtension::Assemble()
 
 void LinearFormExtension::Update()
 {
+   MFEM_PERF_FUNCTION;
+
    const FiniteElementSpace &fes = *lf->FESpace();
    const Mesh &mesh = *fes.GetMesh();
    constexpr ElementDofOrdering ordering = ElementDofOrdering::LEXICOGRAPHIC;
@@ -125,8 +134,7 @@ 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 +144,12 @@ 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

@@ -14,6 +14,7 @@
 
 #include "../general/array.hpp"
 #include "../linalg/vector.hpp"
+#include "fespace.hpp"
 
 namespace mfem
 {
@@ -25,7 +26,8 @@ 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/lor/lor.cpp

@@ -365,6 +365,8 @@ FiniteElementSpace &LORBase::GetFESpace() const
 
 void LORBase::AssembleSystem(BilinearForm &a_ho, const Array<int> &ess_dofs)
 {
+   MFEM_PERF_FUNCTION;
+
    A.Clear();
    delete a;
    if (BatchedLORAssembly::FormIsSupported(a_ho))

fem/lor/lor_batched.cpp

@@ -360,6 +360,8 @@ void BatchedLORAssembly::FillJAndData(SparseMatrix &A) const
 
 void BatchedLORAssembly::SparseIJToCSR(OperatorHandle &A) const
 {
+   MFEM_PERF_FUNCTION;
+
    const int nvdof = fes_ho.GetVSize();
 
    // If A contains an existing SparseMatrix, reuse it (and try to reuse its
@@ -417,6 +419,8 @@ static void Assemble_(LOR_KERNEL &kernel, int dim, int sdim, int order)
 template <typename LOR_KERNEL>
 void BatchedLORAssembly::AssemblyKernel(BilinearForm &a)
 {
+   MFEM_PERF_FUNCTION;
+
    LOR_KERNEL kernel(a, fes_ho, X_vert, sparse_ij, sparse_mapping);
 
    const int dim = fes_ho.GetMesh()->Dimension();

fem/lor/lor_h1_impl.hpp

@@ -184,6 +184,8 @@ void BatchedLOR_H1::Assemble2D()
 template <int ORDER>
 void BatchedLOR_H1::Assemble3D()
 {
+   MFEM_PERF_FUNCTION;
+
    const int nel_ho = fes_ho.GetNE();
    static constexpr int nv = 8;
    static constexpr int dim = 3;

fem/multigrid.cpp

@@ -10,6 +10,7 @@
 // CONTRIBUTING.md for details.
 
 #include "multigrid.hpp"
+#include "../general/annotation.hpp"
 
 namespace mfem
 {
@@ -17,7 +18,10 @@ namespace mfem
 MultigridBase::MultigridBase()
    : cycleType(CycleType::VCYCLE), preSmoothingSteps(1), postSmoothingSteps(1),
      nrhs(0)
-{}
+{
+   coarse_solver = nullptr;
+   own_coarse_solver = false;
+}
 
 MultigridBase::MultigridBase(const Array<Operator*>& operators_,
                              const Array<Solver*>& smoothers_,
@@ -29,12 +33,18 @@ MultigridBase::MultigridBase(const Array<Operator*>& operators_,
 {
    operators_.Copy(operators);
    smoothers_.Copy(smoothers);
+   coarse_solver = nullptr;
    ownedOperators_.Copy(ownedOperators);
    ownedSmoothers_.Copy(ownedSmoothers);
+   own_coarse_solver = false;
 }
 
 MultigridBase::~MultigridBase()
 {
+   if (own_coarse_solver)
+   {
+      delete coarse_solver;
+   }
    for (int i = 0; i < operators.Size(); ++i)
    {
       if (ownedOperators[i])
@@ -56,16 +66,17 @@ void MultigridBase::InitVectors() const
    X.SetSize(M, nrhs);
    Y.SetSize(M, nrhs);
    R.SetSize(M, nrhs);
-   Z.SetSize(M, nrhs);
-   for (int i = 0; i < X.NumRows(); ++i)
+   for (int i = 0; i < M; ++i)
    {
       const int n = operators[i]->Height();
-      for (int j = 0; j < X.NumCols(); ++j)
+      for (int j = 0; j < nrhs; ++j)
       {
-         X(i, j) = new Vector(n);
-         Y(i, j) = new Vector(n);
+         if (i < M - 1)
+         {
+            X(i, j) = new Vector(n);
+            Y(i, j) = new Vector(n);
+         }
          R(i, j) = new Vector(n);
-         Z(i, j) = new Vector(n);
       }
    }
 }
@@ -76,10 +87,12 @@ void MultigridBase::EraseVectors() const
    {
       for (int j = 0; j < X.NumCols(); ++j)
       {
-         delete X(i, j);
-         delete Y(i, j);
+         if (i < X.NumRows() - 1)
+         {
+            delete X(i, j);
+            delete Y(i, j);
+         }
          delete R(i, j);
-         delete Z(i, j);
       }
    }
 }
@@ -95,6 +108,12 @@ void MultigridBase::AddLevel(Operator* op, Solver* smoother,
    ownedSmoothers.Append(ownSmoother);
 }
 
+void MultigridBase::AddCoarseSolver(Solver *c_solver, bool own_c_solver)
+{
+   coarse_solver = c_solver;
+   own_coarse_solver = own_c_solver;
+}
+
 void MultigridBase::SetCycleType(CycleType cycleType_, int preSmoothingSteps_,
                                  int postSmoothingSteps_)
 {
@@ -105,25 +124,24 @@ void MultigridBase::SetCycleType(CycleType cycleType_, int preSmoothingSteps_,
 
 void MultigridBase::Mult(const Vector& x, Vector& y) const
 {
-   Array<const Vector*> X_(1);
-   Array<Vector*> Y_(1);
-   X_[0] = &x;
-   Y_[0] = &y;
+   const Vector *x_array[1] = { &x };
+   Array<const Vector*> X_(x_array, 1); // no heap allocation
+
+   Vector *y_array[1] = { &y };
+   Array<Vector*> Y_(y_array, 1); // no heap allocation
+
    ArrayMult(X_, Y_);
 }
 
 void MultigridBase::ArrayMult(const Array<const Vector*>& X_,
                               Array<Vector*>& Y_) const
 {
+   MFEM_PERF_FUNCTION;
+
    MFEM_ASSERT(operators.Size() > 0,
                "Multigrid solver does not have operators set!");
    MFEM_ASSERT(X_.Size() == Y_.Size(),
                "Number of columns mismatch in MultigridBase::Mult!");
-   if (iterative_mode)
-   {
-      MFEM_WARNING("Multigrid solver does not use iterative_mode and ignores "
-                   "the initial guess!");
-   }
 
    // Add capacity as necessary
    nrhs = X_.Size();
@@ -134,96 +152,163 @@ void MultigridBase::ArrayMult(const Array<const Vector*>& X_,
    for (int j = 0; j < nrhs; ++j)
    {
       MFEM_ASSERT(X_[j] && Y_[j], "Missing Vector in MultigridBase::Mult!");
-      *X(M - 1, j) = *X_[j];
-      *Y(M - 1, j) = 0.0;
-   }
-   Cycle(M - 1);
-   for (int j = 0; j < nrhs; ++j)
-   {
-      *Y_[j] = *Y(M - 1, j);
+      X(M - 1, j) = const_cast<Vector*>(X_[j]);
+      Y(M - 1, j) = Y_[j];
    }
+   const bool zero = !iterative_mode;
+   Cycle(M - 1, zero);
 }
 
 void MultigridBase::SmoothingStep(int level, bool zero, bool transpose) const
 {
+   MFEM_PERF_FUNCTION;
+
    // y = y + S (x - A y) or y = y + S^T (x - A y)
+
+   // Note: 'zero' == true means that Y(level,*) are not initialized and we
+   // should assume that the input they typically provide to this call is zeros.
+
+   // We can't use the smoothers' iterative mode since we don't know if they
+   // actually support it, so we always turn the iterative mode off to properly
+   // use smoothers that do support it.
+   smoothers[level]->iterative_mode = false;
+
    if (zero)
    {
-      Array<Vector *> X_(X[level], nrhs), Y_(Y[level], nrhs);
-      GetSmootherAtLevel(level)->ArrayMult(X_, Y_);
+      MFEM_ASSERT(!transpose, "internal error!");
+      const Array<const Vector *> cX_((const Vector **)(X[level]), nrhs);
+      Array<Vector *> Y_(Y[level], nrhs);
+
+      GetSmootherAtLevel(level)->ArrayMult(cX_, Y_);
    }
    else
    {
-      Array<Vector *> Y_(Y[level], nrhs), R_(R[level], nrhs),
-            Z_(Z[level], nrhs);
+      const Array<const Vector *> cY_((const Vector **)(Y[level]), nrhs),
+            cR_((const Vector **)(R[level]), nrhs);
+      Array<Vector *> Y_(Y[level], nrhs), R_(R[level], nrhs);
+
+      GetOperatorAtLevel(level)->ArrayMult(cY_, R_);
       for (int j = 0; j < nrhs; ++j)
       {
-         *R_[j] = *X(level, j);
+         // *R_[j] = *X(level, j) - *R_[j]
+         subtract(*X(level, j), *R_[j], *R_[j]);
       }
-      GetOperatorAtLevel(level)->ArrayAddMult(Y_, R_, -1.0);
       if (transpose)
       {
-         GetSmootherAtLevel(level)->ArrayMultTranspose(R_, Z_);
+         GetSmootherAtLevel(level)->ArrayAddMultTranspose(cR_, Y_);
       }
       else
       {
-         GetSmootherAtLevel(level)->ArrayMult(R_, Z_);
-      }
-      for (int j = 0; j < nrhs; ++j)
-      {
-         *Y_[j] += *Z_[j];
+         GetSmootherAtLevel(level)->ArrayAddMult(cR_, Y_);
       }
    }
 }
 
-void MultigridBase::Cycle(int level) const
+void MultigridBase::CoarseSolve(bool zero) const
 {
-   // Coarse solve
-   if (level == 0)
+   MFEM_PERF_FUNCTION;
+
+   // See the comment about iterative mode in SmoothingStep()
+   coarse_solver->iterative_mode = false;
+
+   if (zero)
    {
-      SmoothingStep(0, true, false);
+      const Array<const Vector *> cX_((const Vector **)(X[0]), nrhs);
+      Array<Vector *> Y_(Y[0], nrhs);
+
+      coarse_solver->ArrayMult(cX_, Y_);
+   }
+   else
+   {
+      const Array<const Vector *> cY_((const Vector **)(Y[0]), nrhs),
+            cR_((const Vector **)(R[0]), nrhs);
+      Array<Vector *> Y_(Y[0], nrhs), R_(R[0], nrhs);
+
+      GetOperatorAtLevel(0)->ArrayMult(cY_, R_);
+      for (int j = 0; j < nrhs; ++j)
+      {
+         // *R_[j] = *X(0, j) - *R_[j]
+         subtract(*X(0, j), *R_[j], *R_[j]);
+      }
+      coarse_solver->ArrayAddMult(cR_, Y_);
+   }
+}
+
+void MultigridBase::Cycle(int level, bool zero) const
+{
+   // Note: 'zero' == true means that Y(level,*) are not initialized and we
+   // should assume that the input they typically provide to this call is zeros.
+
+   // Coarse solve
+   if (level == 0 && !coarse_solver)
+   {
+      SmoothingStep(0, zero, false);
       return;
    }
 
    // Pre-smooth
    for (int i = 0; i < preSmoothingSteps; ++i)
    {
-      SmoothingStep(level, (cycleType == CycleType::VCYCLE && i == 0), false);
+      SmoothingStep(level, zero && (i == 0), false);
+   }
+
+   // Coarse solve with 'coarse_solver'
+   if (level == 0)
+   {
+      CoarseSolve(preSmoothingSteps == 0 && zero);
+      goto mg_post_smooth;
    }
 
    // Compute residual and restrict
+   if (preSmoothingSteps == 0 && zero)
    {
-      Array<Vector *> Y_(Y[level], nrhs), R_(R[level], nrhs),
-            X_(X[level - 1], nrhs);
+      const Array<const Vector *> cX_l((const Vector **)(X[level]), nrhs);
+      Array<Vector *> X_lm1(X[level - 1], nrhs);
+
+      GetProlongationAtLevel(level - 1)->ArrayMultTranspose(cX_l, X_lm1);
+   }
+   else
+   {
+      const Array<const Vector *> cY_((const Vector **)(Y[level]), nrhs),
+            cR_((const Vector **)(R[level]), nrhs);
+      Array<Vector *> R_(R[level], nrhs), X_(X[level - 1], nrhs);
+
+      GetOperatorAtLevel(level)->ArrayMult(cY_, R_);
       for (int j = 0; j < nrhs; ++j)
       {
-         *R_[j] = *X(level, j);
-      }
-      GetOperatorAtLevel(level)->ArrayAddMult(Y_, R_, -1.0);
-      GetProlongationAtLevel(level - 1)->ArrayMultTranspose(R_, X_);
-      for (int j = 0; j < nrhs; ++j)
-      {
-         *Y(level - 1, j) = 0.0;
+         // *R_[j] = *X(level, j) - *R_[j]
+         subtract(*X(level, j), *R_[j], *R_[j]);
       }
+      GetProlongationAtLevel(level - 1)->ArrayMultTranspose(cR_, X_);
    }
 
    // Corrections
-   Cycle(level - 1);
+   Cycle(level - 1, true);
    if (cycleType == CycleType::WCYCLE)
    {
-      Cycle(level - 1);
+      // If the coarse solve at level 0 is "exact" solve, then we don't want to
+      // repeat it.
+      // To support multiple level 0 coarse-grid corrections, one can wrap that
+      // smoother in an SLI solver and use that instead.
+      if (level > 1) { Cycle(level - 1, false); }
    }
 
    // Prolongate and add
    {
-      Array<Vector *> Y_(Y[level - 1], nrhs), Z_(Z[level], nrhs);
-      GetProlongationAtLevel(level - 1)->ArrayMult(Y_, Z_);
-      for (int j = 0; j < nrhs; ++j)
+      const Array<const Vector *> cY_lm1((const Vector **)(Y[level - 1]), nrhs);
+      Array<Vector *> Y_l(Y[level], nrhs);
+
+      if (preSmoothingSteps == 0 && zero)
       {
-         *Y(level, j) += *Z_[j];
+         GetProlongationAtLevel(level - 1)->ArrayMult(cY_lm1, Y_l);
+      }
+      else
+      {
+         GetProlongationAtLevel(level - 1)->ArrayAddMult(cY_lm1, Y_l);
       }
    }
 
+mg_post_smooth:
    // Post-smooth
    for (int i = 0; i < postSmoothingSteps; ++i)
    {

fem/multigrid.hpp

@@ -36,12 +36,14 @@ protected:
    Array<Solver*> smoothers;
    Array<bool> ownedOperators;
    Array<bool> ownedSmoothers;
+   Solver *coarse_solver; /// can be NULL, see AddCoarseSolver()
+   bool own_coarse_solver;
 
    CycleType cycleType;
    int preSmoothingSteps;
    int postSmoothingSteps;
 
-   mutable Array2D<Vector*> X, Y, R, Z;
+   mutable Array2D<Vector*> X, Y, R;
    mutable int nrhs;
 
 public:
@@ -65,6 +67,16 @@ public:
    void AddLevel(Operator* op, Solver* smoother, bool ownOperator,
                  bool ownSmoother);
 
+   /// Adds a coarse solver for level 0 to work in tandem with the smoother
+   /** If this coarse solver is not given, the smoother at level 0 is used as
+       the coarse solver. When this coarse solver is given, the smoother at
+       level 0 is used similar to the smoothers at other levels. Thus, the
+       action at level 0 consists of:
+       - pre-smoothing steps with smoother 0,
+       - solve step with @a c_solver,
+       - post-smoothing steps with smoother 0. */
+   void AddCoarseSolver(Solver *c_solver, bool own_c_solver);
+
    /// Returns the number of levels
    int NumLevels() const { return operators.Size(); }
 
@@ -118,11 +130,14 @@ public:
 
 private:
    /// Application of a multigrid cycle at particular level
-   void Cycle(int level) const;
+   void Cycle(int level, bool zero) const;
 
    /// Application of a pre-/post-smoothing step at particular level
    void SmoothingStep(int level, bool zero, bool transpose) const;
 
+   /// Perform a coarse solve with 'coarse_solve' (must be non-NULL)
+   void CoarseSolve(bool zero) const;
+
    /// Allocate or destroy temporary storage
    void InitVectors() const;
    void EraseVectors() const;
@@ -202,6 +217,9 @@ public:
 
    /// Recover the solution of a linear system formed with FormFineLinearSystem()
    void RecoverFineFEMSolution(const Vector& X, const Vector& b, Vector& x);
+
+   const Array<int> &GetFineEssentialTrueDofs() const
+   { return *essentialTrueDofs.Last(); }
 };
 
 } // namespace mfem

fem/pbilinearform.cpp

@@ -124,6 +124,8 @@ void ParBilinearForm::pAllocMat()
 void ParBilinearForm::ParallelRAP(SparseMatrix &loc_A, OperatorHandle &A,
                                   bool steal_loc_A)
 {
+   MFEM_PERF_FUNCTION;
+
    ParFiniteElementSpace &pfespace = *ParFESpace();
 
    // Create a block diagonal parallel matrix

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)
@@ -62,9 +63,11 @@ ParFiniteElementSpace::ParFiniteElementSpace(
 
 ParFiniteElementSpace::ParFiniteElementSpace(
    ParMesh *pm, const FiniteElementCollection *f, int dim, int ordering)
-   : FiniteElementSpace(pm, f, dim, ordering)
+   : FiniteElementSpace((MFEM_PERF_BEGIN(_MFEM_FUNC_NAME), pm),
+                        f, dim, ordering)
 {
    ParInit(pm);
+   MFEM_PERF_END(_MFEM_FUNC_NAME);
 }
 
 ParFiniteElementSpace::ParFiniteElementSpace(
@@ -91,6 +94,7 @@ ParNURBSExtension *ParFiniteElementSpace::MakeLocalNURBSext(
 
 void ParFiniteElementSpace::ParInit(ParMesh *pm)
 {
+   MFEM_PERF_FUNCTION;
    pmesh = pm;
    pncmesh = nullptr;
 
@@ -180,6 +184,7 @@ void ParFiniteElementSpace::CommunicateGhostOrder()
 
 void ParFiniteElementSpace::Construct()
 {
+   MFEM_PERF_FUNCTION;
    if (NURBSext)
    {
       ConstructTrueNURBSDofs();
@@ -838,6 +843,8 @@ void ParFiniteElementSpace::Build_Dof_TrueDof_Matrix() const // matrix P
 
    if (P) { return; }
 
+   MFEM_PERF_FUNCTION;
+
    if (!nd_strias)
    {
       // Safe to assume 1-1 correspondence between shared dofs
@@ -1423,6 +1430,7 @@ const Operator *ParFiniteElementSpace::GetRestrictionOperator() const
       if (NRanks == 1)
       {
          R_transpose.reset(new IdentityOperator(GetTrueVSize()));
+         Rconf = new IdentityOperator(GetTrueVSize());
       }
       else
       {
@@ -1435,8 +1443,8 @@ const Operator *ParFiniteElementSpace::GetRestrictionOperator() const
             R_transpose.reset(
                new DeviceConformingProlongationOperator(*this, true));
          }
+         Rconf = new TransposeOperator(*R_transpose);
       }
-      Rconf = new TransposeOperator(*R_transpose);
       return Rconf;
    }
    else
@@ -4487,13 +4495,6 @@ 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 +4537,13 @@ 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 +4563,46 @@ 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 +4660,7 @@ 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 +4681,14 @@ 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 +4715,6 @@ ParFiniteElementSpace::ParallelDerefinementMatrix(int old_ndofs,
       }
    }
 
-   messages.clear();
    offd->Finalize(0);
    offd->SetWidth(static_cast<int>(col_map.size()));
 
@@ -4946,8 +4966,13 @@ 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/plinearform.cpp

@@ -45,6 +45,8 @@ void ParLinearForm::MakeRef(ParFiniteElementSpace *pf, Vector &v, int v_offset)
 
 void ParLinearForm::Assemble()
 {
+   MFEM_PERF_FUNCTION;
+
    LinearForm::Assemble();
 
    if (interior_face_integs.Size())

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

@@ -96,17 +96,33 @@ void QuadratureSpaceBase::Integrate(VectorCoefficient &coeff,
 
 void QuadratureSpace::ConstructOffsets()
 {
+   MFEM_PERF_FUNCTION;
    const int num_elem = mesh.GetNE();
-   offsets.SetSize(num_elem + 1);
-   int offset = 0;
-   for (int i = 0; i < num_elem; i++)
+   ne = num_elem;
+
+   if (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()
@@ -188,6 +204,7 @@ void FaceQuadratureSpace::ConstructOffsets()
 {
    face_indices.SetSize(num_faces);
    offsets.SetSize(num_faces + 1);
+   ne = num_faces;
    int offset = 0;
    int f_idx = 0;
    for (int i = 0; i < mesh.GetNumFacesWithGhost(); i++)

fem/qspace.hpp

@@ -30,13 +30,17 @@ protected:
    Mesh &mesh; ///< The underlying mesh.
    int order; ///< The order of integration rule.
    int size; ///< Total number of quadrature points.
+   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.
+   /// @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.
    ///
-   /// 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];
@@ -59,6 +63,24 @@ protected:
    void ConstructWeights() const;
 
 public:
+   /// @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
+   {
+      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 { return offsets; }
+
    /// Return the total number of quadrature points.
    int GetSize() const { return size; }
 
@@ -66,7 +88,7 @@ public:
    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; }

fem/quadinterpolator.cpp

@@ -503,6 +503,7 @@ void QuadratureInterpolator::Mult(const Vector &e_vec,
                                   Vector &q_der,
                                   Vector &q_det) const
 {
+   MFEM_PERF_FUNCTION;
    using namespace internal::quadrature_interpolator;
 
    const int ne = fespace->GetNE();

fem/restriction.cpp

@@ -25,7 +25,7 @@ namespace mfem
 
 ElementRestriction::ElementRestriction(const FiniteElementSpace &f,
                                        ElementDofOrdering e_ordering)
-   : fes(f),
+   : fes((MFEM_PERF_BEGIN(_MFEM_FUNC_NAME), f)),
      ne(fes.GetNE()),
      vdim(fes.GetVDim()),
      byvdim(fes.GetOrdering() == Ordering::byVDIM),
@@ -104,10 +104,13 @@ ElementRestriction::ElementRestriction(const FiniteElementSpace &f,
       offsets[i] = offsets[i - 1];
    }
    offsets[0] = 0;
+   MFEM_PERF_END(_MFEM_FUNC_NAME);
 }
 
 void ElementRestriction::Mult(const Vector& x, Vector& y) const
 {
+   MFEM_PERF_FUNCTION;
+
    // Assumes all elements have the same number of dofs
    const int nd = dof;
    const int vd = vdim;
@@ -152,6 +155,8 @@ void ElementRestriction::AbsMult(const Vector& x, Vector& y) const
 template <bool ADD>
 void ElementRestriction::TAddMultTranspose(const Vector& x, Vector& y) const
 {
+   MFEM_PERF_FUNCTION;
+
    // Assumes all elements have the same number of dofs
    const int nd = dof;
    const int vd = vdim;

fem/transfer.cpp

@@ -13,6 +13,7 @@
 #include "bilinearform.hpp"
 #include "pbilinearform.hpp"
 #include "../general/forall.hpp"
+#include "kernels.hpp"
 
 namespace mfem
 {
@@ -2322,6 +2323,76 @@ void Prolongation2D(const int NE, const int D1D, const int Q1D,
    });
 }
 
+template <int DLO, int DHI>
+static void SmemProlongation3D(const int NE,
+                               const Vector& localL, Vector& localH,
+                               const Array<real_t> &b, const Vector& mask)
+{
+   MFEM_PERF_FUNCTION;
+
+   auto u_lo = Reshape(localL.Read(), DLO, DLO, DLO, NE);
+   auto u_hi = Reshape(localH.Write(), DHI, DHI, DHI, NE);
+   auto d_b = Reshape(b.Read(), DHI, DLO);
+   auto m_ = Reshape(mask.Read(), DHI, DHI, DHI, NE);
+
+   mfem::forall_2D(NE, DHI, DHI, [=] MFEM_HOST_DEVICE (int e)
+   {
+      // Load B into shared memory
+      MFEM_SHARED real_t s_B[DHI*DLO];
+      kernels::internal::LoadBt<DLO,DHI>(DLO,DHI,d_b,s_B);
+      const DeviceMatrix B(s_B, DHI, DLO);
+
+      MFEM_SHARED real_t s_u[DHI*DHI*DLO];
+      const DeviceCube u(s_u, DHI, DHI, DLO);
+      real_t v[DHI];
+
+      MFEM_FOREACH_THREAD(lx,x,DLO)
+      {
+         MFEM_FOREACH_THREAD(ly,y,DLO)
+         {
+            for (int hz = 0; hz < DHI; ++hz) { v[hz] = 0.0; }
+            for (int lz = 0; lz < DLO; ++lz)
+            {
+               const real_t XYZ = u_lo(lx,ly,lz,e);
+               for (int hz = 0; hz < DHI; ++hz) { v[hz] += XYZ * B(hz,lz); }
+            }
+            for (int hz = 0; hz < DHI; ++hz) { u(hz,ly,lx) = v[hz]; }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(hz,y,DHI)
+      {
+         MFEM_FOREACH_THREAD(lx,x,DLO)
+         {
+            for (int hy = 0; hy < DHI; ++hy) { v[hy] = 0.0; }
+            for (int ly = 0; ly < DLO; ++ly)
+            {
+               const real_t zYX = u(hz,ly,lx);
+               for (int hy = 0; hy < DHI; ++hy) { v[hy] += zYX * B(hy,ly); }
+            }
+            for (int hy = 0; hy < DHI; ++hy) { u(hz,hy,lx) = v[hy]; }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(hz,y,DHI)
+      {
+         MFEM_FOREACH_THREAD(hy,x,DHI)
+         {
+            for (int hx = 0; hx < DHI; ++hx) { v[hx] = 0.0; }
+            for (int lx = 0; lx < DLO; ++lx)
+            {
+               const real_t zyX = u(hz,hy,lx);
+               for (int hx = 0; hx < DHI; ++hx) { v[hx] += zyX * B(hx,lx); }
+            }
+            for (int hx = 0; hx < DHI; ++hx)
+            {
+               u_hi(hx,hy,hz,e) = m_(hx,hy,hz,e)*v[hx];
+            }
+         }
+      }
+   });
+}
+
 void Prolongation3D(const int NE, const int D1D, const int Q1D,
                     const Vector& localL, Vector& localH,
                     const Array<real_t>& B, const Vector& mask)
@@ -2403,9 +2474,9 @@ void Prolongation3D(const int NE, const int D1D, const int Q1D,
    });
 }
 
-void Restriction2D(const int NE, const int D1D, const int Q1D,
-                   const Vector& localH, Vector& localL,
-                   const Array<real_t>& Bt, const Vector& mask)
+void ProlongationTranspose2D(const int NE, const int D1D, const int Q1D,
+                             const Vector& localH, Vector& localL,
+                             const Array<real_t>& Bt, const Vector& mask)
 {
    auto x_ = Reshape(localH.Read(), Q1D, Q1D, NE);
    auto y_ = Reshape(localL.Write(), D1D, D1D, NE);
@@ -2448,9 +2519,80 @@ void Restriction2D(const int NE, const int D1D, const int Q1D,
       }
    });
 }
-void Restriction3D(const int NE, const int D1D, const int Q1D,
-                   const Vector& localH, Vector& localL,
-                   const Array<real_t>& Bt, const Vector& mask)
+
+template <int DLO, int DHI>
+static void SmemProlongationTranspose3D(
+   const int NE, const Vector& localH, Vector& localL,
+   const Array<real_t>& bt, const Vector& mask)
+{
+   MFEM_PERF_FUNCTION;
+
+   auto u_h = Reshape(localH.Read(), DHI, DHI, DHI, NE);
+   auto u_l = Reshape(localL.Write(), DLO, DLO, DLO, NE);
+   auto d_bt = Reshape(bt.Read(), DLO, DHI);
+   auto m_ = Reshape(mask.Read(), DHI, DHI, DHI, NE);
+
+   mfem::forall_2D(NE, DHI, DHI, [=] MFEM_HOST_DEVICE (int e)
+   {
+      // Load Bt into shared memory
+      MFEM_SHARED real_t s_Bt[DHI*DLO];
+      kernels::internal::LoadBt<DHI,DLO>(DHI,DLO,d_bt,s_Bt);
+      const DeviceMatrix Bt(s_Bt, DLO, DHI);
+
+      MFEM_SHARED real_t s_u[DLO*DHI*DHI];
+      const DeviceCube u(s_u, DLO, DHI, DHI);
+      real_t v[DLO];
+
+      MFEM_FOREACH_THREAD(hx,x,DHI)
+      {
+         MFEM_FOREACH_THREAD(hy,y,DHI)
+         {
+            for (int lz = 0; lz < DLO; ++lz) { v[lz] = 0.0; }
+            for (int hz = 0; hz < DHI; ++hz)
+            {
+               const real_t XYZ = m_(hx,hy,hz,e)*u_h(hx,hy,hz,e);
+               for (int lz = 0; lz < DLO; ++lz) { v[lz] += XYZ * Bt(lz,hz); }
+            }
+            for (int lz = 0; lz < DLO; ++lz) { u(lz,hy,hx) = v[lz]; }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(lz,y,DLO)
+      {
+         MFEM_FOREACH_THREAD(hx,x,DHI)
+         {
+            for (int ly = 0; ly < DLO; ++ly) { v[ly] = 0.0; }
+            for (int hy = 0; hy < DHI; ++hy)
+            {
+               const real_t zYX = u(lz,hy,hx);
+               for (int ly = 0; ly < DLO; ++ly) { v[ly] += zYX * Bt(ly,hy); }
+            }
+            for (int ly = 0; ly < DLO; ++ly) { u(lz,ly,hx) = v[ly]; }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(lz,y,DLO)
+      {
+         MFEM_FOREACH_THREAD(ly,x,DLO)
+         {
+            for (int lx = 0; lx < DLO; ++lx) { v[lx] = 0.0; }
+            for (int hx = 0; hx < DHI; ++hx)
+            {
+               const real_t zyX = u(lz,ly,hx);
+               for (int lx = 0; lx < DLO; ++lx) { v[lx] += zyX * Bt(lx,hx); }
+            }
+            for (int lx = 0; lx < DLO; ++lx)
+            {
+               u_l(lx,ly,lz,e) = v[lx];
+            }
+         }
+      }
+   });
+}
+
+void ProlongationTranspose3D(const int NE, const int D1D, const int Q1D,
+                             const Vector& localH, Vector& localL,
+                             const Array<real_t>& Bt, const Vector& mask)
 {
    auto x_ = Reshape(localH.Read(), Q1D, Q1D, Q1D, NE);
    auto y_ = Reshape(localL.Write(), D1D, D1D, D1D, NE);
@@ -2518,11 +2660,15 @@ void Restriction3D(const int NE, const int D1D, const int Q1D,
       }
    });
 }
+
 } // namespace TransferKernels
 
 void TensorProductPRefinementTransferOperator::Mult(const Vector& x,
                                                     Vector& y) const
 {
+   MFEM_PERF_FUNCTION;
+   using namespace TransferKernels;
+
    if (lFESpace.GetMesh()->GetNE() == 0)
    {
       return;
@@ -2531,11 +2677,25 @@ void TensorProductPRefinementTransferOperator::Mult(const Vector& x,
    elem_restrict_lex_l->Mult(x, localL);
    if (dim == 2)
    {
-      TransferKernels::Prolongation2D(NE, D1D, Q1D, localL, localH, B, mask);
+      Prolongation2D(NE, D1D, Q1D, localL, localH, B, mask);
    }
    else if (dim == 3)
    {
-      TransferKernels::Prolongation3D(NE, D1D, Q1D, localL, localH, B, mask);
+      switch ((D1D << 4 ) | Q1D)
+      {
+         case 0x23:
+            SmemProlongation3D<2,3>(NE, localL, localH, B, mask); break;
+         case 0x24:
+            SmemProlongation3D<2,4>(NE, localL, localH, B, mask); break;
+         case 0x35:
+            SmemProlongation3D<3,5>(NE, localL, localH, B, mask); break;
+         case 0x46:
+            SmemProlongation3D<4,6>(NE, localL, localH, B, mask); break;
+         case 0x47:
+            SmemProlongation3D<4,7>(NE, localL, localH, B, mask); break;
+         default:
+            Prolongation3D(NE, D1D, Q1D, localL, localH, B, mask); break;
+      }
    }
    else
    {
@@ -2549,6 +2709,9 @@ void TensorProductPRefinementTransferOperator::Mult(const Vector& x,
 void TensorProductPRefinementTransferOperator::MultTranspose(const Vector& x,
                                                              Vector& y) const
 {
+   MFEM_PERF_FUNCTION;
+   using namespace TransferKernels;
+
    if (lFESpace.GetMesh()->GetNE() == 0)
    {
       return;
@@ -2557,11 +2720,25 @@ void TensorProductPRefinementTransferOperator::MultTranspose(const Vector& x,
    elem_restrict_lex_h->Mult(x, localH);
    if (dim == 2)
    {
-      TransferKernels::Restriction2D(NE, D1D, Q1D, localH, localL, Bt, mask);
+      ProlongationTranspose2D(NE, D1D, Q1D, localH, localL, Bt, mask);
    }
    else if (dim == 3)
    {
-      TransferKernels::Restriction3D(NE, D1D, Q1D, localH, localL, Bt, mask);
+      switch ((D1D << 4 ) | Q1D)
+      {
+         case 0x23:
+            SmemProlongationTranspose3D<2,3>(NE, localH, localL, Bt, mask); break;
+         case 0x24:
+            SmemProlongationTranspose3D<2,4>(NE, localH, localL, Bt, mask); break;
+         case 0x35:
+            SmemProlongationTranspose3D<3,5>(NE, localH, localL, Bt, mask); break;
+         case 0x46:
+            SmemProlongationTranspose3D<4,6>(NE, localH, localL, Bt, mask); break;
+         case 0x47:
+            SmemProlongationTranspose3D<4,7>(NE, localH, localL, Bt, mask); break;
+         default:
+            ProlongationTranspose3D(NE, D1D, Q1D, localH, localL, Bt, mask); break;
+      }
    }
    else
    {
@@ -2583,20 +2760,20 @@ TrueTransferOperator::TrueTransferOperator(const FiniteElementSpace& lFESpace_,
 
    P = lFESpace.GetProlongationMatrix();
    R = hFESpace.IsVariableOrder() ? hFESpace.GetHpRestrictionMatrix() :
-       hFESpace.GetRestrictionMatrix();
+       hFESpace.GetRestrictionOperator();
 
    // P and R can be both null
    // P can be null and R not null
    // If P is not null it is assumed that R is not null as well
    if (P) { MFEM_VERIFY(R, "Both P and R have to be not NULL") }
 
-   if (P)
+   if (!IsIdentityProlongation(P))
    {
       tmpL.SetSize(lFESpace_.GetVSize());
       tmpH.SetSize(hFESpace_.GetVSize());
    }
    // P can be null and R not null
-   else if (R)
+   else if (!IsIdentityProlongation(R))
    {
       tmpH.SetSize(hFESpace_.GetVSize());
    }
@@ -2609,13 +2786,14 @@ TrueTransferOperator::~TrueTransferOperator()
 
 void TrueTransferOperator::Mult(const Vector& x, Vector& y) const
 {
-   if (P)
+   MFEM_PERF_FUNCTION;
+   if (!IsIdentityProlongation(P))
    {
       P->Mult(x, tmpL);
       localTransferOperator->Mult(tmpL, tmpH);
       R->Mult(tmpH, y);
    }
-   else if (R)
+   else if (!IsIdentityProlongation(R))
    {
       localTransferOperator->Mult(x, tmpH);
       R->Mult(tmpH, y);
@@ -2628,13 +2806,14 @@ void TrueTransferOperator::Mult(const Vector& x, Vector& y) const
 
 void TrueTransferOperator::MultTranspose(const Vector& x, Vector& y) const
 {
-   if (P)
+   MFEM_PERF_FUNCTION;
+   if (!IsIdentityProlongation(P))
    {
       R->MultTranspose(x, tmpH);
       localTransferOperator->MultTranspose(tmpH, tmpL);
       P->MultTranspose(tmpL, y);
    }
-   else if (R)
+   else if (!IsIdentityProlongation(R))
    {
       R->MultTranspose(x, tmpH);
       localTransferOperator->MultTranspose(tmpH, y);

fem/transfer.hpp

@@ -621,9 +621,6 @@ public:
       const FiniteElementSpace& lFESpace_,
       const FiniteElementSpace& hFESpace_);
 
-   /// Destructor
-   virtual ~TensorProductPRefinementTransferOperator() { }
-
    /// @brief Interpolation or prolongation of a vector \p x corresponding to
    /// the coarse space to the vector \p y corresponding to the fine space.
    void Mult(const Vector& x, Vector& y) const override;
@@ -642,7 +639,7 @@ private:
    const FiniteElementSpace& lFESpace;
    const FiniteElementSpace& hFESpace;
    const Operator * P = nullptr;
-   const SparseMatrix * R = nullptr;
+   const Operator * R = nullptr;
    TransferOperator* localTransferOperator;
    mutable Vector tmpL;
    mutable Vector tmpH;

general/annotation.hpp

@@ -14,24 +14,98 @@
 
 #include "../config/config.hpp"
 
-#ifdef MFEM_USE_CALIPER
+#define MFEM_CONCAT_(X,Y)  X##Y
+#define MFEM_CONCAT(X,Y)  MFEM_CONCAT_(X,Y)
 
+#ifdef MFEM_USE_CALIPER
+#include "device.hpp"
+#include "backends.hpp"
+#ifdef MFEM_USE_MPI
+#include "communication.hpp"
+#endif
 #include <caliper/cali.h>
 #include <caliper/cali-manager.h>
-#define MFEM_PERF_FUNCTION CALI_CXX_MARK_FUNCTION
-#define MFEM_PERF_BEGIN(s) CALI_MARK_BEGIN(s)
-#define MFEM_PERF_END(s) CALI_MARK_END(s)
+#endif
+
+namespace mfem
+{
+
+namespace internal
+{
+
+extern int annotation_sync_stream; // defined in globals.cpp
+extern int annotation_sync_mpi; // defined in globals.cpp
+
+#ifdef MFEM_USE_CALIPER
+
+inline void AnnotationSync()
+{
+   if (annotation_sync_stream && Device::Allows(Backend::DEVICE_MASK))
+   {
+      MFEM_STREAM_SYNC;
+   }
+#ifdef MFEM_USE_MPI
+   if (annotation_sync_mpi && Mpi::IsInitialized() && !Mpi::IsFinalized())
+   {
+      MPI_Barrier(GetGlobalMPI_Comm());
+   }
+#endif
+}
+
+struct FunctionAnnotation
+{
+   ::cali::Function cali_func;
+
+   FunctionAnnotation(const char *fname)
+      : cali_func((AnnotationSync(), fname)) { }
+
+   ~FunctionAnnotation() { AnnotationSync(); }
+};
+
+struct ScopeAnnotation
+{
+   ::cali::ScopeAnnotation cali_scope;
+
+   ScopeAnnotation(const char *name)
+      : cali_scope((AnnotationSync(), name)) { }
+
+   ~ScopeAnnotation() { AnnotationSync(); }
+};
+
+#endif // #ifdef MFEM_USE_CALIPER
+
+} // namespace internal
+
+} // namespace mfem
+
+
+#ifdef MFEM_USE_CALIPER
+
+#define MFEM_PERF_FUNCTION \
+   mfem::internal::FunctionAnnotation mfem_func_annotation_(_MFEM_FUNC_NAME)
+#define MFEM_PERF_BEGIN(s) \
+   (mfem::internal::AnnotationSync(), CALI_MARK_BEGIN(s))
+#define MFEM_PERF_END(s) \
+   (mfem::internal::AnnotationSync(), CALI_MARK_END(s))
 #define MFEM_PERF_SCOPE(name) \
-  cali::Annotation::Guard cali_autogenerated_guard_name(cali::Annotation("function").begin(std::string(name).c_str()))
+   mfem::internal::ScopeAnnotation \
+      MFEM_CONCAT(mfem_scope_annotation_,__LINE__)(name)
 
+#define MFEM_PERF_SYNC_STREAM(b) (mfem::internal::annotation_sync_stream = (b))
+#define MFEM_PERF_SYNC_MPI(b) (mfem::internal::annotation_sync_mpi = (b))
+#define MFEM_PERF_SYNC(b) (MFEM_PERF_SYNC_STREAM(b), MFEM_PERF_SYNC_MPI(b))
 
-#else
+#else // #ifdef MFEM_USE_CALIPER
 
 #define MFEM_PERF_FUNCTION
-#define MFEM_PERF_BEGIN(s)
+#define MFEM_PERF_BEGIN(s) ((void)(0))
 #define MFEM_PERF_END(s)
 #define MFEM_PERF_SCOPE(name)
 
-#endif
+#define MFEM_PERF_SYNC_STREAM(b)
+#define MFEM_PERF_SYNC_MPI(b)
+#define MFEM_PERF_SYNC(b)
 
-#endif
+#endif // #ifdef MFEM_USE_CALIPER
+
+#endif // MFEM_ANNOTATION_HPP

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

@@ -23,13 +23,9 @@
 #include <mpi.h>
 #include <cstdint>
 
-// can't directly use MPI_CXX_BOOL because Microsoft's MPI implementation
-// doesn't include MPI_CXX_BOOL. Fallback to MPI_C_BOOL if unavailable.
-#ifdef MPI_CXX_BOOL
-#define MFEM_MPI_CXX_BOOL MPI_CXX_BOOL
-#else
-#define MFEM_MPI_CXX_BOOL MPI_C_BOOL
-#endif
+// Some MPI implementations do not have MPI_CXX_BOOL or do not handle it
+// correctly, so we use MPI_UNSIGNED_CHAR as the MPI type for 'bool':
+#define MFEM_MPI_CXX_BOOL MPI_UNSIGNED_CHAR
 
 namespace mfem
 {
@@ -425,16 +421,24 @@ 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 +603,50 @@ 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 +655,6 @@ 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

@@ -151,6 +151,22 @@ Device::Device()
    {
       SetGPUAwareMPI(true);
    }
+
+   if (const char *mfem_perf_sync = GetEnv("MFEM_PERF_SYNC"))
+   {
+      MFEM_PERF_SYNC(std::atoi(mfem_perf_sync));
+      MFEM_CONTRACT_VAR(mfem_perf_sync);
+   }
+   if (const char *mfem_perf_sync_stream = GetEnv("MFEM_PERF_SYNC_STREAM"))
+   {
+      MFEM_PERF_SYNC_STREAM(std::atoi(mfem_perf_sync_stream));
+      MFEM_CONTRACT_VAR(mfem_perf_sync_stream);
+   }
+   if (const char *mfem_perf_sync_mpi = GetEnv("MFEM_PERF_SYNC_MPI"))
+   {
+      MFEM_PERF_SYNC_MPI(std::atoi(mfem_perf_sync_mpi));
+      MFEM_CONTRACT_VAR(mfem_perf_sync_mpi);
+   }
 }
 
 Device::~Device()
@@ -615,7 +631,7 @@ 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.cpp

@@ -193,4 +193,4 @@ void mfem_warning(const char *msg)
    }
 }
 
-}
+} // namespace mfem

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__);     \
@@ -208,4 +208,4 @@ __device__ void abort_msg(T & msg)
 #define MFEM_ASSERT_KERNEL(x,...)
 #endif
 
-#endif
+#endif // MFEM_ERROR_HPP

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/globals.cpp

@@ -31,6 +31,9 @@ namespace internal
 {
 bool mfem_out_initialized = false;
 bool mfem_err_initialized = false;
+
+int annotation_sync_stream = 0; // declared in annotation.hpp
+int annotation_sync_mpi = 0; // declared in annotation.hpp
 }
 
 void OutStream::Init()

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

@@ -657,7 +657,8 @@ private: // Static methods used by the Memory<T> class
    /// Return the host pointer.
    MFEM_ENZYME_INACTIVE static void *Register_(void *ptr, void *h_ptr,
                                                size_t bytes, MemoryType mt,
-                                               bool own, bool alias, unsigned &flags);
+                                               bool own, bool alias,
+                                               unsigned &flags);
 
    /// Register a pair of external host and device pointers
    static void Register2_(void *h_ptr, void *d_ptr, size_t bytes,
@@ -741,7 +742,7 @@ private:
 
    /// Insert a host address @a h_ptr and size *a bytes in the memory map to be
    /// managed.
-   void Insert(void *h_ptr, size_t bytes, MemoryType h_mt,  MemoryType d_mt);
+   void Insert(void *h_ptr, size_t bytes, MemoryType h_mt, MemoryType d_mt);
 
    /// Insert a device and the host addresses in the memory map
    void InsertDevice(void *d_ptr, void *h_ptr, size_t bytes,
@@ -896,6 +897,7 @@ 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;
@@ -980,7 +982,7 @@ inline void Memory<T>::Wrap(T *ptr, int size, bool own)
 #ifdef MFEM_DEBUG
    if (own && MemoryManager::Exists())
    {
-      MemoryType h_ptr_mt = MemoryManager::GetHostMemoryType_(h_ptr);
+      MemoryType h_ptr_mt = MemoryManager::GetHostMemoryType_((void*)h_ptr);
       MFEM_VERIFY(h_mt == h_ptr_mt,
                   "h_mt = " << (int)h_mt << ", h_ptr_mt = " << (int)h_ptr_mt);
    }
@@ -988,7 +990,8 @@ inline void Memory<T>::Wrap(T *ptr, int size, bool own)
    if (own && h_mt != MemoryType::HOST)
    {
       const size_t bytes = size*sizeof(T);
-      MemoryManager::Register_(ptr, ptr, bytes, h_mt, own, false, flags);
+      MemoryManager::Register_((void*)ptr, (void*)ptr, bytes, h_mt, own, false,
+                               flags);
    }
 }
 
@@ -1057,7 +1060,8 @@ 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))

linalg/batched/gpu_blas.cpp

@@ -20,13 +20,13 @@
 #define MFEM_CU_or_HIP(stub) HIP##stub
 #endif
 
-#define MFEM_CONCAT(x, y, z) MFEM_CONCAT_(x, y, z)
-#define MFEM_CONCAT_(x, y, z) x ## y ## z
+#define MFEM_CONCAT3(x, y, z) MFEM_CONCAT3_(x, y, z)
+#define MFEM_CONCAT3_(x, y, z) x ## y ## z
 
 #ifdef MFEM_USE_SINGLE
-#define MFEM_GPUBLAS_PREFIX(stub) MFEM_CONCAT(MFEM_cu_or_hip(blas), S, stub)
+#define MFEM_GPUBLAS_PREFIX(stub) MFEM_CONCAT3(MFEM_cu_or_hip(blas), S, stub)
 #elif defined(MFEM_USE_DOUBLE)
-#define MFEM_GPUBLAS_PREFIX(stub) MFEM_CONCAT(MFEM_cu_or_hip(blas), D, stub)
+#define MFEM_GPUBLAS_PREFIX(stub) MFEM_CONCAT3(MFEM_cu_or_hip(blas), D, stub)
 #endif
 
 #define MFEM_BLAS_SUCCESS MFEM_CU_or_HIP(BLAS_STATUS_SUCCESS)

linalg/hypre.cpp

@@ -1868,6 +1868,8 @@ HYPRE_Int HypreParMatrix::Mult(HypreParVector &x, HypreParVector &y,
 
 void HypreParMatrix::Mult(real_t a, const Vector &x, real_t b, Vector &y) const
 {
+   MFEM_PERF_FUNCTION;
+
    MFEM_ASSERT(x.Size() == Width(), "invalid x.Size() = " << x.Size()
                << ", expected size = " << Width());
    MFEM_ASSERT(y.Size() == Height(), "invalid y.Size() = " << y.Size()
@@ -1926,6 +1928,8 @@ void HypreParMatrix::Mult(real_t a, const Vector &x, real_t b, Vector &y) const
 void HypreParMatrix::MultTranspose(real_t a, const Vector &x,
                                    real_t b, Vector &y) const
 {
+   MFEM_PERF_FUNCTION;
+
    MFEM_ASSERT(x.Size() == Height(), "invalid x.Size() = " << x.Size()
                << ", expected size = " << Height());
    MFEM_ASSERT(y.Size() == Width(), "invalid y.Size() = " << y.Size()
@@ -4091,6 +4095,8 @@ void HypreSolver::Setup(const HypreParVector &b, HypreParVector &x) const
 {
    if (setup_called) { return; }
 
+   MFEM_PERF_FUNCTION;
+
    MFEM_VERIFY(A != NULL, "HypreParMatrix A is missing");
 
    HYPRE_Int err_flag = SetupFcn()(*this, *A, b, x);
@@ -4116,6 +4122,8 @@ void HypreSolver::Setup(const Vector &b, Vector &x) const
 
 void HypreSolver::Mult(const HypreParVector &b, HypreParVector &x) const
 {
+   MFEM_PERF_FUNCTION;
+
    HYPRE_Int err_flag;
    if (A == NULL)
    {

linalg/operator.cpp

@@ -50,17 +50,19 @@ void Operator::InitTVectors(const Operator *Po, const Operator *Ri,
 
 void Operator::AddMult(const Vector &x, Vector &y, const real_t a) const
 {
-   mfem::Vector z(y.Size());
-   Mult(x, z);
-   y.Add(a, z);
+   z_am.SetSize(y.Size());
+   z_am.UseDevice(true);
+   Mult(x, z_am);
+   y.Add(a, z_am);
 }
 
 void Operator::AddMultTranspose(const Vector &x, Vector &y,
                                 const real_t a) const
 {
-   mfem::Vector z(y.Size());
-   MultTranspose(x, z);
-   y.Add(a, z);
+   z_am.SetSize(y.Size());
+   z_am.UseDevice(true);
+   MultTranspose(x, z_am);
+   y.Add(a, z_am);
 }
 
 void Operator::ArrayMult(const Array<const Vector *> &X,
@@ -586,6 +588,8 @@ void ConstrainedOperator::EliminateRHS(const Vector &x, Vector &b) const
 void ConstrainedOperator::ConstrainedMult(const Vector &x, Vector &y,
                                           const bool transpose) const
 {
+   MFEM_PERF_FUNCTION;
+
    const int csz = constraint_list.Size();
    if (csz == 0)
    {
@@ -785,6 +789,8 @@ void RectangularConstrainedOperator::EliminateRHS(const Vector &x,
 
 void RectangularConstrainedOperator::Mult(const Vector &x, Vector &y) const
 {
+   MFEM_PERF_FUNCTION;
+
    const int trial_csz = trial_constraints.Size();
    const int test_csz = test_constraints.Size();
    if (trial_csz == 0)
@@ -820,6 +826,8 @@ void RectangularConstrainedOperator::Mult(const Vector &x, Vector &y) const
 void RectangularConstrainedOperator::MultTranspose(const Vector &x,
                                                    Vector &y) const
 {
+   MFEM_PERF_FUNCTION;
+
    const int trial_csz = trial_constraints.Size();
    const int test_csz = test_constraints.Size();
    if (test_csz == 0)

linalg/operator.hpp

@@ -13,6 +13,7 @@
 #define MFEM_OPERATOR
 
 #include "vector.hpp"
+#include "../general/annotation.hpp"
 
 namespace mfem
 {
@@ -23,6 +24,13 @@ class RectangularConstrainedOperator;
 /// Abstract operator
 class Operator
 {
+private:
+   /// Auxiliary Vector used by the methods AddMult() and AddMultTranspose().
+   /** @note This Vector is private to prevent derived classes from accidentaly
+       using it in their implementation of Mult() or MultTranspose() which may
+       lead to hard-to-find bugs. */
+   mutable Vector z_am;
+
 protected:
    int height; ///< Dimension of the output / number of rows in the matrix.
    int width;  ///< Dimension of the input / number of columns in the matrix.
@@ -818,10 +826,12 @@ public:
    explicit IdentityOperator(int n) : Operator(n) { }
 
    /// Operator application
-   void Mult(const Vector &x, Vector &y) const override { y = x; }
+   void Mult(const Vector &x, Vector &y) const override
+   { MFEM_PERF_FUNCTION; y = x; }
 
    /// Application of the transpose
-   void MultTranspose(const Vector &x, Vector &y) const override { y = x; }
+   void MultTranspose(const Vector &x, Vector &y) const override
+   { MFEM_PERF_FUNCTION; y = x; }
 };
 
 /// Returns true if P is the identity prolongation, i.e. if it is either NULL or

linalg/solvers.cpp

@@ -55,6 +55,8 @@ IterativeSolver::IterativeSolver(MPI_Comm comm_)
 
 real_t IterativeSolver::Dot(const Vector &x, const Vector &y) const
 {
+   MFEM_PERF_FUNCTION;
+
 #ifndef MFEM_USE_MPI
    return (x * y);
 #else
@@ -314,25 +316,29 @@ void OperatorJacobiSmoother::Mult(const Vector &x, Vector &y) const
    MFEM_VERIFY(x.Size() == Width(), "invalid input vector");
    MFEM_VERIFY(y.Size() == Height(), "invalid output vector");
 
+   auto DI = dinv.Read();
+   auto X = x.Read();
    if (iterative_mode)
    {
       MFEM_VERIFY(oper, "iterative_mode == true requires the forward operator");
       oper->Mult(y, residual);  // r = A y
-      subtract(x, residual, residual); // r = x - A y
+      auto R = residual.Read();
+      auto Y = y.ReadWrite();
+      // y += D^{-1} (x - A y)
+      mfem::forall(height, [=] MFEM_HOST_DEVICE (int i)
+      {
+         Y[i] += DI[i] * (X[i] - R[i]);
+      });
    }
    else
    {
-      residual = x;
-      y.UseDevice(true);
-      y = 0.0;
+      auto Y = y.Write();
+      // y = D^{-1} x
+      mfem::forall(height, [=] MFEM_HOST_DEVICE (int i)
+      {
+         Y[i] = DI[i] * X[i];
+      });
    }
-   auto DI = dinv.Read();
-   auto R = residual.Read();
-   auto Y = y.ReadWrite();
-   mfem::forall(height, [=] MFEM_HOST_DEVICE (int i)
-   {
-      Y[i] += DI[i] * R[i];
-   });
 }
 
 OperatorChebyshevSmoother::OperatorChebyshevSmoother(const Operator &oper_,
@@ -348,7 +354,8 @@ OperatorChebyshevSmoother::OperatorChebyshevSmoother(const Operator &oper_,
    diag(d),
    coeffs(order),
    ess_tdof_list(ess_tdofs),
-   residual(N),
+   residual(order > 1 ? N : 0),
+   z(order > 1 ? N : 0),
    oper(&oper_) { Setup(); }
 
 #ifdef MFEM_USE_MPI
@@ -368,14 +375,15 @@ OperatorChebyshevSmoother::OperatorChebyshevSmoother(const Operator &oper_,
                                                      real_t power_tolerance,
                                                      int power_seed)
 #endif
-   : Solver(d.Size()),
+   : Solver((MFEM_PERF_BEGIN(_MFEM_FUNC_NAME), d.Size())),
      order(order_),
      N(d.Size()),
      dinv(N),
      diag(d),
      coeffs(order),
      ess_tdof_list(ess_tdofs),
-     residual(N),
+     residual(order > 1 ? N : 0),
+     z(order > 1 ? N : 0),
      oper(&oper_)
 {
    OperatorJacobiSmoother invDiagOperator(diag, ess_tdofs, 1.0);
@@ -394,6 +402,7 @@ OperatorChebyshevSmoother::OperatorChebyshevSmoother(const Operator &oper_,
                                                             power_seed);
 
    Setup();
+   MFEM_PERF_END(_MFEM_FUNC_NAME);
 }
 
 OperatorChebyshevSmoother::OperatorChebyshevSmoother(const Operator* oper_,
@@ -422,7 +431,7 @@ void OperatorChebyshevSmoother::Setup()
 {
    // Invert diagonal
    residual.UseDevice(true);
-   helperVector.UseDevice(true);
+   z.UseDevice(true);
    auto D = diag.Read();
    auto X = dinv.Write();
    mfem::forall(N, [=] MFEM_HOST_DEVICE (int i) { X[i] = 1.0 / D[i]; });
@@ -432,6 +441,20 @@ void OperatorChebyshevSmoother::Setup()
       X[I[i]] = 1.0;
    });
 
+   const int order_save = order;
+   order = -1; // avoid early exit in SetOrder() when 'new_order' == 'order'
+   SetOrder(order_save);
+}
+
+void OperatorChebyshevSmoother::SetOrder(int new_order)
+{
+   if (new_order == order) { return; }
+
+   order = new_order;
+   coeffs.SetSize(order);
+   residual.SetSize(order > 1 ? N : 0);
+   z.SetSize(order > 1 ? N : 0);
+
    // Set up Chebyshev coefficients
    // For reference, see e.g., Parallel multigrid smoothing: polynomial versus
    // Gauss-Seidel by Adams et al.
@@ -501,6 +524,8 @@ void OperatorChebyshevSmoother::Setup()
 
 void OperatorChebyshevSmoother::Mult(const Vector& x, Vector &y) const
 {
+   MFEM_PERF_FUNCTION;
+
    if (iterative_mode)
    {
       MFEM_ABORT("Chebyshev smoother not implemented for iterative mode");
@@ -511,32 +536,55 @@ void OperatorChebyshevSmoother::Mult(const Vector& x, Vector &y) const
       MFEM_ABORT("Chebyshev smoother requires operator");
    }
 
-   residual = x;
-   helperVector.SetSize(x.Size());
-   helperVector.UseDevice(true);
-
-   y.UseDevice(true);
-   y = 0.0;
-
-   for (int k = 0; k < order; ++k)
+   // for k = 0, perform:
+   //    r = D^{-1} x
+   //    y = C_0 r
+   const real_t C_0 = coeffs[0];
+   auto Dinv = dinv.Read();
+   auto X = x.Read();
+   auto Y0 = y.Write();
+   if (order == 1)
    {
-      // Apply
-      if (k > 0)
+      mfem::forall(N, [=] MFEM_HOST_DEVICE (int i)
       {
-         oper->Mult(residual, helperVector);
-         residual = helperVector;
-      }
+         Y0[i] = C_0 * Dinv[i] * X[i];
+      });
+   }
+   else
+   {
+      auto R0 = residual.Write();
+      mfem::forall(N, [=] MFEM_HOST_DEVICE (int i)
+      {
+         Y0[i] = C_0 * (R0[i] = Dinv[i] * X[i]);
+      });
+   }
 
-      // Scale residual by inverse diagonal
-      const int n = N;
-      auto Dinv = dinv.Read();
-      auto R = residual.ReadWrite();
-      mfem::forall(n, [=] MFEM_HOST_DEVICE (int i) { R[i] *= Dinv[i]; });
+   for (int k = 1; k < order; ++k)
+   {
+      // Apply: z = A r
+      oper->Mult(residual, z);
 
-      // Add weighted contribution to y
+      // Scale residual by inverse diagonal and add weighted contribution to y:
+      //   r = D^{-1} z
+      //   y += C_k r
+      const real_t C_k = coeffs[k];
+      auto Z = z.Read();
       auto Y = y.ReadWrite();
-      auto C = coeffs.Read();
-      mfem::forall(n, [=] MFEM_HOST_DEVICE (int i) { Y[i] += C[k] * R[i]; });
+      if (k < order-1)
+      {
+         auto R = residual.Write();
+         mfem::forall(N, [=] MFEM_HOST_DEVICE (int i)
+         {
+            Y[i] += C_k * (R[i] = Dinv[i] * Z[i]);
+         });
+      }
+      else
+      {
+         mfem::forall(N, [=] MFEM_HOST_DEVICE (int i)
+         {
+            Y[i] += C_k * Dinv[i] * Z[i];
+         });
+      }
    }
 }
 
@@ -3213,7 +3261,7 @@ void ResidualBCMonitor::MonitorResidual(
    MPI_Comm comm = iter_solver->GetComm();
    if (comm != MPI_COMM_NULL)
    {
-      double glob_bc_norm_squared = 0.0;
+      real_t glob_bc_norm_squared = 0.0;
       MPI_Reduce(&bc_norm_squared, &glob_bc_norm_squared, 1,
                  MPITypeMap<real_t>::mpi_type,
                  MPI_SUM, 0, comm);

linalg/solvers.hpp

@@ -380,11 +380,11 @@ public:
    void SetPositiveDiagonal(bool pos_diag = true) { use_abs_diag = pos_diag; }
 
    /// Approach the solution of the linear system by applying Jacobi smoothing.
-   void Mult(const Vector &x, Vector &y) const;
+   void Mult(const Vector &x, Vector &y) const override;
 
    /** @brief Approach the solution of the transposed linear system by applying
        Jacobi smoothing. */
-   void MultTranspose(const Vector &x, Vector &y) const { Mult(x, y); }
+   void MultTranspose(const Vector &x, Vector &y) const override { Mult(x, y); }
 
    /** @brief Recompute the diagonal using the method AssembleDiagonal of the
        given new Operator, @a op. */
@@ -397,7 +397,7 @@ public:
        When the new Operator, @a op, is not a (Par)BilinearForm, any previously
        set array of essential true-dofs will be thrown away because in this case
        any essential b.c. will be handled by the AssembleDiagonal method. */
-   void SetOperator(const Operator &op);
+   void SetOperator(const Operator &op) override;
 
 private:
    Vector dinv;
@@ -481,21 +481,22 @@ public:
 
    /** @brief Approach the solution of the linear system by applying Chebyshev
        smoothing. */
-   void Mult(const Vector &x, Vector &y) const;
+   void Mult(const Vector &x, Vector &y) const override;
 
    /** @brief Approach the solution of the transposed linear system by applying
        Chebyshev smoothing. */
-   void MultTranspose(const Vector &x, Vector &y) const { Mult(x, y); }
+   void MultTranspose(const Vector &x, Vector &y) const override { Mult(x, y); }
 
-   void SetOperator(const Operator &op_)
+   void SetOperator(const Operator &op_) override
    {
       oper = &op_;
    }
 
    void Setup();
+   void SetOrder(int new_order);
 
 private:
-   const int order;
+   int order;
    real_t max_eig_estimate;
    const int N;
    Vector dinv;
@@ -503,7 +504,7 @@ private:
    Array<real_t> coeffs;
    const Array<int>& ess_tdof_list;
    mutable Vector residual;
-   mutable Vector helperVector;
+   mutable Vector z;
    const Operator* oper;
 };
 

linalg/sparsemat.cpp

@@ -764,6 +764,8 @@ void SparseMatrix::Mult(const Vector &x, Vector &y) const
 
 void SparseMatrix::AddMult(const Vector &x, Vector &y, const real_t a) const
 {
+   MFEM_PERF_FUNCTION;
+
    MFEM_ASSERT(width == x.Size(), "Input vector size (" << x.Size()
                << ") must match matrix width (" << width << ")");
    MFEM_ASSERT(height == y.Size(), "Output vector size (" << y.Size()
@@ -964,6 +966,8 @@ void SparseMatrix::MultTranspose(const Vector &x, Vector &y) const
 void SparseMatrix::AddMultTranspose(const Vector &x, Vector &y,
                                     const real_t a) const
 {
+   MFEM_PERF_FUNCTION;
+
    MFEM_ASSERT(height == x.Size(), "Input vector size (" << x.Size()
                << ") must match matrix height (" << height << ")");
    MFEM_ASSERT(width == y.Size(), "Output vector size (" << y.Size()

linalg/vector.cpp

@@ -117,7 +117,7 @@ Vector::Vector(const Vector &v)
    UseDevice(v.UseDevice());
 }
 
-Vector::Vector(Vector &&v)
+Vector::Vector(Vector &&v) : Vector()
 {
    *this = std::move(v);
 }
@@ -205,14 +205,16 @@ Vector &Vector::operator=(const Vector &v)
    data.CopyFrom(v.data, v.Size());
    UseDevice(v.UseDevice());
 #else
-   SetSize(v.Size());
    const bool vuse = v.UseDevice();
    const bool use_dev = UseDevice() || vuse;
+   if (use_dev) { MFEM_PERF_BEGIN(_MFEM_FUNC_NAME); }
+   SetSize(v.Size());
    v.UseDevice(use_dev);
    // keep 'data' where it is, unless 'use_dev' is true
    if (use_dev) { Write(); }
    data.CopyFrom(v.data, v.Size());
    v.UseDevice(vuse);
+   if (use_dev) { MFEM_PERF_END(_MFEM_FUNC_NAME); }
 #endif
    return *this;
 }
@@ -227,9 +229,11 @@ Vector &Vector::operator=(Vector &&v)
 Vector &Vector::operator=(real_t value)
 {
    const bool use_dev = UseDevice();
+   if (use_dev) { MFEM_PERF_BEGIN(_MFEM_FUNC_NAME); }
    const int N = size;
    auto y = Write(use_dev);
    mfem::forall_switch(use_dev, N, [=] MFEM_HOST_DEVICE (int i) { y[i] = value; });
+   if (use_dev) { MFEM_PERF_END(_MFEM_FUNC_NAME); }
    return *this;
 }
 
@@ -290,10 +294,12 @@ Vector &Vector::operator-=(const Vector &v)
    MFEM_ASSERT(size == v.size, "incompatible Vectors!");
 
    const bool use_dev = UseDevice() || v.UseDevice();
+   if (use_dev) { MFEM_PERF_BEGIN(_MFEM_FUNC_NAME); }
    const int N = size;
    const auto x = v.Read(use_dev);
    auto y = ReadWrite(use_dev);
    mfem::forall_switch(use_dev, N, [=] MFEM_HOST_DEVICE (int i) { y[i] -= x[i]; });
+   if (use_dev) { MFEM_PERF_END(_MFEM_FUNC_NAME); }
    return *this;
 }
 
@@ -311,10 +317,12 @@ Vector &Vector::operator+=(const Vector &v)
    MFEM_ASSERT(size == v.size, "incompatible Vectors!");
 
    const bool use_dev = UseDevice() || v.UseDevice();
+   if (use_dev) { MFEM_PERF_BEGIN(_MFEM_FUNC_NAME); }
    const int N = size;
    const auto x = v.Read(use_dev);
    auto y = ReadWrite(use_dev);
    mfem::forall_switch(use_dev, N, [=] MFEM_HOST_DEVICE (int i) { y[i] += x[i]; });
+   if (use_dev) { MFEM_PERF_END(_MFEM_FUNC_NAME); }
    return *this;
 }
 
@@ -326,9 +334,11 @@ Vector &Vector::Add(const real_t a, const Vector &Va)
    {
       const int N = size;
       const bool use_dev = UseDevice() || Va.UseDevice();
+      if (use_dev) { MFEM_PERF_BEGIN(_MFEM_FUNC_NAME); }
       const auto x = Va.Read(use_dev);
       auto y = ReadWrite(use_dev);
       mfem::forall_switch(use_dev, N, [=] MFEM_HOST_DEVICE (int i) { y[i] += a * x[i]; });
+      if (use_dev) { MFEM_PERF_END(_MFEM_FUNC_NAME); }
    }
    return *this;
 }
@@ -445,6 +455,7 @@ void add(const Vector &v1, real_t alpha, const Vector &v2, Vector &v)
    {
 #if !defined(MFEM_USE_LEGACY_OPENMP)
       const bool use_dev = v1.UseDevice() || v2.UseDevice() || v.UseDevice();
+      if (use_dev) { MFEM_PERF_BEGIN(_MFEM_FUNC_NAME); }
       const int N = v.size;
       // Note: get read access first, in case v is the same as v1/v2.
       const auto d_x = v1.Read(use_dev);
@@ -454,6 +465,7 @@ void add(const Vector &v1, real_t alpha, const Vector &v2, Vector &v)
       {
          d_z[i] = d_x[i] + alpha * d_y[i];
       });
+      if (use_dev) { MFEM_PERF_END(_MFEM_FUNC_NAME); }
 #else
       const real_t *v1p = v1.data, *v2p = v2.data;
       real_t *vp = v.data;
@@ -569,6 +581,7 @@ void subtract(const Vector &x, const Vector &y, Vector &z)
 
 #if !defined(MFEM_USE_LEGACY_OPENMP)
    const bool use_dev = x.UseDevice() || y.UseDevice() || z.UseDevice();
+   if (use_dev) { MFEM_PERF_BEGIN(_MFEM_FUNC_NAME); }
    const int N = x.size;
    // Note: get read access first, in case z is the same as x/y.
    const auto xd = x.Read(use_dev);
@@ -578,6 +591,7 @@ void subtract(const Vector &x, const Vector &y, Vector &z)
    {
       zd[i] = xd[i] - yd[i];
    });
+   if (use_dev) { MFEM_PERF_END(_MFEM_FUNC_NAME); }
 #else
    const real_t *xp = x.data;
    const real_t *yp = y.data;

makefile

@@ -125,7 +125,8 @@ EXAMPLE_TEST_DIRS := examples
 
 MINIAPP_SUBDIRS = common electromagnetics meshing navier performance tools \
  toys nurbs gslib adjoint solvers shifted mtop parelag tribol autodiff dfem \
- hooke multidomain dpg hdiv-linear-solver spde diag-smoothers
+ hooke multidomain dpg hdiv-linear-solver spde diag-smoothers \
+ benchmarks/ceed-solver-bps
 MINIAPP_DIRS := $(addprefix miniapps/,$(MINIAPP_SUBDIRS))
 MINIAPP_TEST_DIRS := $(filter-out %/common,$(MINIAPP_DIRS))
 MINIAPP_USE_COMMON := $(addprefix miniapps/,electromagnetics meshing tools \

mesh/mesh.cpp

@@ -32,6 +32,7 @@
 #include <cstring>
 #include <ctime>
 #include <functional>
+#include <set>
 #include <numeric>
 #include <unordered_map>
 #include <unordered_set>
@@ -921,8 +922,66 @@ 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::DeleteGeometricFactors()
 {
+   MFEM_PERF_FUNCTION;
    for (int i = 0; i < geom_factors.Size(); i++)
    {
       delete geom_factors[i];
@@ -1804,6 +1863,8 @@ void Mesh::Destroy()
    TetMemory.Clear();
 #endif
 
+   elem_attrs_cache.DeleteAll();
+   bdr_face_attrs_cache.DeleteAll();
    attributes.DeleteAll();
    bdr_attributes.DeleteAll();
 }
@@ -1817,36 +1878,48 @@ 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 +4527,10 @@ 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()
@@ -6503,6 +6580,7 @@ void XYZ_VectorFunction(const Vector &p, Vector &v)
 
 void Mesh::GetNodes(GridFunction &nodes) const
 {
+   MFEM_PERF_FUNCTION;
    if (Nodes == NULL || Nodes->FESpace() != nodes.FESpace())
    {
       const int newSpaceDim = nodes.FESpace()->GetVDim();
@@ -6523,6 +6601,7 @@ void Mesh::SetNodalFESpace(FiniteElementSpace *nfes)
 
 void Mesh::EnsureNodes()
 {
+   MFEM_PERF_FUNCTION;
    if (Nodes)
    {
       const FiniteElementCollection *fec = GetNodalFESpace()->FEColl();
@@ -6575,6 +6654,7 @@ const FiniteElementSpace *Mesh::GetNodalFESpace() const
 
 void Mesh::SetCurvature(int order, bool discont, int space_dim, int ordering)
 {
+   MFEM_PERF_FUNCTION;
    if (order <= 0)
    {
       delete Nodes;
@@ -7721,6 +7801,12 @@ 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);
 }
 
@@ -10850,6 +10936,10 @@ 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);
 }
 
 void Mesh::GetElementData(const Array<Element*> &elem_array, int geom,
@@ -13257,7 +13347,7 @@ 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 +13360,7 @@ void Mesh::Transform(void (*f)(const Vector&, Vector&))
             vold(j) = vertices[i](j);
          }
          vnew.SetData(vertices[i]());
-         (*f)(vold, vnew);
+         f(vold, vnew);
       }
    }
    else
@@ -14649,7 +14739,7 @@ GeometricFactors::GeometricFactors(const GridFunction &nodes,
 void GeometricFactors::Compute(const GridFunction &nodes,
                                MemoryType d_mt)
 {
-
+   MFEM_PERF_FUNCTION;
    const FiniteElementSpace *fespace = nodes.FESpace();
    const FiniteElement *fe = fespace->GetTypicalFE();
    const int dim  = fe->GetDim();

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(),
@@ -1122,13 +1127,14 @@ 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 +2260,7 @@ 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 +2273,35 @@ 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!");
+      }
    }
 }
 
@@ -2010,6 +2016,7 @@ void ParMesh::DeleteFaceNbrData()
 
 void ParMesh::SetCurvature(int order, bool discont, int space_dim, int ordering)
 {
+   MFEM_PERF_FUNCTION;
    DeleteFaceNbrData();
    space_dim = (space_dim == -1) ? spaceDim : space_dim;
    FiniteElementCollection* nfec;

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/benchmarks/ceed-solver-bps/kershaw.hpp

@@ -0,0 +1,156 @@
+// 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 __KERSHAW_HPP__
+#define __KERSHAW_HPP__
+
+#include "mfem.hpp"
+
+namespace mfem
+{
+
+// 1D transformation at the right boundary.
+real_t right(const real_t eps, const real_t x)
+{
+   return (x <= 0.5) ? (2-eps) * x : 1 + eps*(x-1);
+}
+
+// 1D transformation at the left boundary
+real_t left(const real_t eps, const real_t x)
+{
+   return 1-right(eps,1-x);
+}
+
+// Transition from a value of "a" for x=0, to a value of "b" for x=1. Smoothness
+// is controlled by the parameter "s", taking values 0, 1, or 2.
+real_t step(const real_t a, const real_t b, real_t x, int s)
+{
+   if (x <= 0) { return a; }
+   if (x >= 1) { return b; }
+   switch (s)
+   {
+      case 0:
+      default:
+         return a + (b-a) * (x);
+      case 1: return a + (b-a) * (x*x*(3-2*x));
+      case 2: return a + (b-a) * (x*x*x*(x*(6*x-15)+10));
+   }
+}
+
+// 3D version of a generalized Kershaw mesh transformation, see D. Kershaw,
+// "Differencing of the diffusion equation in Lagrangian hydrodynamic codes",
+// JCP, 39:375–395, 1981.
+//
+// The input mesh should be Cartesian nx x ny x nz with nx divisible by 6 and
+// ny, nz divisible by 2.
+//
+// The eps parameters are in (0, 1]. Uniform mesh is recovered for epsy=epsz=1.
+void kershaw(const real_t epsy, const real_t epsz, const int smoothness,
+             const real_t x, const real_t y, const real_t z,
+             real_t &X, real_t &Y, real_t &Z)
+{
+   X = x;
+
+   int layer = x*6.0;
+   real_t lambda = (x-layer/6.0)*6;
+
+   // The x-range is split in 6 layers going from left-to-left, left-to-right,
+   // right-to-left (2 layers), left-to-right and right-to-right yz-faces.
+   switch (layer)
+   {
+      case 0:
+         Y = left(epsy, y);
+         Z = left(epsz, z);
+         break;
+      case 1:
+      case 4:
+         Y = step(left(epsy, y), right(epsy, y), lambda, smoothness);
+         Z = step(left(epsz, z), right(epsz, z), lambda, smoothness);
+         break;
+      case 2:
+         Y = step(right(epsy, y), left(epsy, y), lambda/2, smoothness);
+         Z = step(right(epsz, z), left(epsz, z), lambda/2, smoothness);
+         break;
+      case 3:
+         Y = step(right(epsy, y), left(epsy, y), (1+lambda)/2, smoothness);
+         Z = step(right(epsz, z), left(epsz, z), (1+lambda)/2, smoothness);
+         break;
+      default:
+         Y = right(epsy, y);
+         Z = right(epsz, z);
+         break;
+   }
+}
+
+struct KershawTransformation : VectorCoefficient
+{
+   real_t epsy, epsz;
+   int dim, s;
+   KershawTransformation(int dim_, real_t epsy_, real_t epsz_, int s_=0)
+      : VectorCoefficient(dim_), epsy(epsy_), epsz(epsz_), dim(dim_), s(s_) { }
+   using VectorCoefficient::Eval;
+   void Eval(Vector &V, ElementTransformation &T,
+             const IntegrationPoint &ip) override
+   {
+      real_t xyz[3];
+      Vector transip(xyz, 3);
+      T.Transform(ip, transip);
+      if (dim == 1)
+      {
+         V[0] = xyz[0]; // no transformation in 1D
+      }
+      else if (dim == 2)
+      {
+         real_t z=0, zt;
+         kershaw(epsy, epsz, s, xyz[0], xyz[1], z, V[0], V[1], zt);
+      }
+      else // dim == 3
+      {
+         kershaw(epsy, epsz, s, xyz[0], xyz[1], xyz[2], V[0], V[1], V[2]);
+      }
+   }
+};
+
+ParMesh CreateKershawMesh(int nx, int ny, int nz, real_t epsy, real_t epsz)
+{
+   const bool sfc_order = true;
+   Mesh serial_mesh;
+   if (nx > 0 && ny == 0 && nz == 0)
+   {
+      serial_mesh = Mesh::MakeCartesian1D(nx, 1.0);
+   }
+   else if (nx > 0 && ny > 0 && nz == 0)
+   {
+      serial_mesh = Mesh::MakeCartesian2D(nx, ny, Element::QUADRILATERAL,
+                                          false, 1, 1, sfc_order);
+   }
+   else if (nx > 0 && ny > 0 && nz > 0)
+   {
+      serial_mesh = Mesh::MakeCartesian3D(nx, ny, nz, Element::HEXAHEDRON,
+                                          1, 1, 1, sfc_order);
+   }
+   else
+   {
+      MFEM_ABORT("Bad grid size");
+   }
+   KershawTransformation kt(serial_mesh.Dimension(), epsy, epsz);
+   serial_mesh.Transform(kt);
+   return ParMesh(MPI_COMM_WORLD, serial_mesh);
+}
+
+ParMesh CreateKershawMesh(int n, real_t eps)
+{
+   return CreateKershawMesh(n, n, n, eps, eps);
+}
+
+}
+
+#endif

miniapps/benchmarks/ceed-solver-bps/makefile

@@ -0,0 +1,77 @@
+# 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.
+
+# Use the MFEM build directory
+MFEM_DIR ?= ../../..
+MFEM_BUILD_DIR ?= ../../..
+MFEM_INSTALL_DIR ?= ../../../mfem
+SRC = $(if $(MFEM_DIR:../../..=),$(MFEM_DIR)/miniapps/benchmarks/ceed-solver-bps/,)
+CONFIG_MK = $(or $(wildcard $(MFEM_BUILD_DIR)/config/config.mk),\
+   $(wildcard $(MFEM_INSTALL_DIR)/share/mfem/config.mk))
+
+MFEM_LIB_FILE = mfem_is_not_built
+-include $(CONFIG_MK)
+
+SEQ_MINIAPPS =
+PAR_MINIAPPS = solver-bp
+ifeq ($(MFEM_USE_MPI),NO)
+   MINIAPPS = $(SEQ_MINIAPPS)
+else
+   MINIAPPS = $(PAR_MINIAPPS) $(SEQ_MINIAPPS)
+endif
+
+EXTRA_SOURCES = preconditioners.cpp
+EXTRA_HEADERS = kershaw.hpp rhs.hpp preconditioners.hpp
+EXTRA_OBJECTS = $(EXTRA_SOURCES:.cpp=.o)
+
+.SUFFIXES:
+.SUFFIXES: .o .cpp .mk
+.PHONY: all clean clean-build clean-exec
+.PRECIOUS: %.o
+
+# Remove built-in rules
+%: %.cpp
+%.o: %.cpp
+
+all: $(MINIAPPS)
+
+# Rule for building solver-bp
+solver-bp: solver-bp.o $(addprefix $(SRC),$(EXTRA_HEADERS)) \
+ $(EXTRA_OBJECTS) $(MFEM_LIB_FILE) $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_LINK_FLAGS) $< -o $@ $(EXTRA_OBJECTS) $(MFEM_LIBS)
+
+# Rules for compiling *.o files
+# -I$(MFEM_DIR) is needed for "general/forall.hpp" for out-of-source builds
+%.o: $(SRC)%.cpp $(wildcard $(SRC)%.hpp) $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -I$(MFEM_DIR) -c $(<) -o $(@)
+
+MFEM_TESTS = MINIAPPS
+include $(MFEM_TEST_MK)
+
+# Testing: Specific execution options
+RUN_MPI = $(MFEM_MPIEXEC) $(MFEM_MPIEXEC_NP) $(MFEM_MPI_NP)
+solver-bp-test-par: solver-bp
+	@$(call mfem-test,$<, $(RUN_MPI), CEED Solver BP,,SKIP-NO-VIS)
+
+# Testing: "test" target and mfem-test* variables are defined in config/test.mk
+
+# Generate an error message if the MFEM library is not built and exit
+$(MFEM_LIB_FILE):
+	$(error The MFEM library is not built)
+
+clean: clean-build clean-exec
+
+clean-build:
+	rm -f *.o *~ $(SEQ_MINIAPPS) $(PAR_MINIAPPS) $(EXTRA_OBJECTS)
+	rm -rf *.dSYM *.TVD.*breakpoints
+
+clean-exec:
+	@true

miniapps/benchmarks/ceed-solver-bps/plot_csv.py

@@ -0,0 +1,129 @@
+import csv
+from pylab import *
+
+fields=[
+    ['code ID', 'str'],
+    ['preconditioner ID', 'str'],
+    ['machine ID', 'str'],
+    ['number of nodes', 'int'],
+    ['number of MPI ranks', 'int'],
+    ['n_x', 'int'], ['n_y', 'int'], ['n_z', 'int'],
+    ['solution polynomial degree', 'int'],
+    ['number of 1D quadrature points', 'float'],
+    ['eps_y', 'float'], ['eps_z', 'float'],
+    ['ndofs (including Dirichlet boundary)', 'int'],
+    ['niter', 'int'],
+    ['initial residual', 'float'], ['final residual', 'float'],
+    ['error', 'float'],
+    ['t_setup (preconditioner setup)', 'float'],
+    ['t_solve (total iter time)', 'float']]
+fields_dict=dict(fields)
+
+def convert(obj, type_str):
+    ctor=getattr(__builtins__, type_str)
+    return ctor(obj)
+
+input_csv='run-001.csv'
+print('reading %s ...' % input_csv)
+runs = []
+with open(input_csv) as csvfile:
+    csvreader = csv.DictReader(csvfile, fieldnames=[f[0] for f in fields],
+                               restkey='additional notes')
+    for row in csvreader:
+        for i in fields_dict:
+            row[i]=convert(row[i], fields_dict[i])
+        runs.append(row)
+
+orders=[r['solution polynomial degree'] for r in runs]
+orders=unique(orders) # numpy function
+# orders=[1]
+
+nps=[r['number of MPI ranks'] for r in runs]
+nps=unique(nps)
+if len(nps) > 1:
+    print('multiple num-ranks present: %s' % nps)
+    quit()
+np=nps[0]
+
+# plot fx (or fx/fn) vs fy, (or fx/fn/fy, etc) for all orders
+fn='number of MPI ranks'
+fx='ndofs (including Dirichlet boundary)'
+fy='t_solve (total iter time)'
+# fy='niter'
+# fy='error'
+fz='niter'
+
+figure()
+for p in orders:
+    rr=[r for r in runs if (r['solution polynomial degree']==p and
+                            r['niter']>0)]
+    if len(rr)==0:
+        continue
+
+    # pl_data=asarray([[r[fx],r[fx]/r[fy]] for r in rr])
+    # pl_data=asarray([[r[fx],r[fy]] for r in rr])
+    # pl_data=asarray([[r[fx],r[fx]/(r[fy]/r[fz])] for r in rr])
+
+    pl_data=asarray([[r[fx]/r[fn],r[fx]/r[fn]/r[fy]] for r in rr])
+    # pl_data=asarray([[r[fx]/r[fn],r[fy]] for r in rr])
+
+    plot(pl_data[:,0],pl_data[:,1], 'o-', label='p=%i'%p)
+    rnx=asarray([r['n_x'] for r in rr])
+    rerr=asarray([r['error'] for r in rr])
+    rate=arange(1.0,len(rnx))
+    for l in range(1,len(rnx)):
+        rate[l-1]=log(rerr[l-1]/rerr[l])/log(rnx[l]/rnx[l-1])
+    set_printoptions(formatter={'float':"{:6.2f}".format},linewidth=120)
+    print(f"p={p} rate:{rate}")
+
+# xscale('log', basex=10) # older matplotlib
+xscale('log', base=10)
+# xlim(4e4,3.1e7)
+xlim(4e4,5e6)
+# yscale('log', basey=10) # older matplotlib
+# yscale('log', base=10)
+# ylim(1e5,2e7)
+# ylim(0,2.55e7)
+# ylim(0,3.25e7)
+# ylim(0,5e6)
+ymin,ymax=ylim()
+ylim(0,ymax)
+# ylim(1e-2,2e1)
+# ylim(3e-3,6e-2)
+# xlabel(fx)
+# xlabel('# DOFs')
+xlabel('# DOFs / # Ranks')
+# ylabel(fx + ' / ' + fy)
+# ylabel(fy)
+# ylabel('# DOFs / t_solve')
+ylabel('# DOFs / # Ranks / t_solve')
+# ylabel('t_solve')
+# ylabel('# DOFs / (t_solve / # Iter)')
+# ylabel('# Iter')
+# ylabel('L2 error')
+# ylabel('Grad L2 error')
+grid('on', color='gray', ls='dotted')
+grid('on', axis='both', which='minor', color='gray', ls='dotted')
+legend(ncol=2, loc='best')
+ranks='1 MPI rank'
+if np > 1:
+   ranks='%s MPI ranks' % (np,np)
+hypre='hypre CPU'
+# hypre='hypre HIP'
+# prec=hypre+', p-MG(1,1)'
+prec=hypre+', LOR'
+# prec='Jacobi'
+# eps='1'
+eps='0.3'
+mfem='MFEM CPU'
+# mfem='MFEM HIP'
+title(mfem + ', ' + prec + ', $\\varepsilon = ' + eps + '$, ' + ranks)
+
+if 1: # write .pdf file?
+    pdf_file='plot.pdf'
+    print('saving figure --> %s'%pdf_file)
+    savefig(pdf_file, format='pdf', bbox_inches='tight')
+
+if 0: # show the figures?
+    print('\nshowing figures ...')
+    show()

miniapps/benchmarks/ceed-solver-bps/preconditioners.cpp

@@ -0,0 +1,241 @@
+// 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 "preconditioners.hpp"
+
+namespace mfem
+{
+
+AssemblyLevel GetCoarseAssemblyLevel(SolverConfig config)
+{
+   switch (config.type)
+   {
+      case SolverConfig::JACOBI:
+      case SolverConfig::LOR_HYPRE:
+      case SolverConfig::LOR_AMGX:
+         return AssemblyLevel::PARTIAL;
+      default:
+         return AssemblyLevel::FULL;
+         // return AssemblyLevel::LEGACYFULL;
+   }
+}
+
+bool NeedsLOR(SolverConfig config)
+{
+   switch (config.type)
+   {
+      case SolverConfig::LOR_HYPRE:
+      case SolverConfig::LOR_AMGX:
+         return true;
+      default:
+         return false;
+   }
+}
+
+DiffusionMultigrid::DiffusionMultigrid(
+   ParFiniteElementSpaceHierarchy& hierarchy,
+   Coefficient &coeff_,
+   Array<int>& ess_bdr,
+   SolverConfig coarse_solver_config,
+   int q1d_inc_,
+   int smoothers_cheby_order_)
+   : GeometricMultigrid(hierarchy, ess_bdr),
+     coeff(coeff_),
+     q1d_inc(q1d_inc_),
+     irs(0, Quadrature1D::GaussLegendre),
+     smoothers_cheby_order(smoothers_cheby_order_)
+{
+   ConstructCoarseOperatorAndSolver(
+      coarse_solver_config, hierarchy.GetFESpaceAtLevel(0), ess_bdr);
+   int nlevels = hierarchy.GetNumLevels();
+   for (int i=1; i<nlevels; ++i)
+   {
+      ConstructOperatorAndSmoother(hierarchy.GetFESpaceAtLevel(i), ess_bdr);
+   }
+}
+
+void DiffusionMultigrid::ConstructBilinearForm(
+   ParFiniteElementSpace &fespace, Array<int> &ess_bdr, AssemblyLevel asm_lvl)
+{
+   ParBilinearForm *form = new ParBilinearForm(&fespace);
+   form->SetAssemblyLevel(asm_lvl);
+
+   DiffusionIntegrator *integ = new DiffusionIntegrator(coeff);
+
+   int p = fespace.GetOrder(0);
+   int dim = fespace.GetMesh()->Dimension();
+   // Integration rule for high-order problem: (p+1+q1d_inc)^d Gauss-Legendre
+   // points
+   int int_order = 2*(p+1+q1d_inc) - 1;
+   Geometry::Type geom = fespace.GetMesh()->GetElementBaseGeometry(0);
+   const IntegrationRule &ir = irs.Get(geom, int_order);
+   MFEM_VERIFY(ir.Size() == pow(p+1+q1d_inc,dim), "Wrong quadrature");
+   integ->SetIntegrationRule(ir);
+
+   form->AddDomainIntegrator(integ);
+   form->Assemble();
+   bfs.Append(form);
+
+   essentialTrueDofs.Append(new Array<int>());
+   fespace.GetEssentialTrueDofs(ess_bdr, *essentialTrueDofs.Last());
+}
+
+void DiffusionMultigrid::ConstructOperatorAndSmoother(
+   ParFiniteElementSpace& fespace, Array<int>& ess_bdr)
+{
+   ConstructBilinearForm(fespace, ess_bdr, AssemblyLevel::PARTIAL);
+
+   OperatorPtr opr;
+   bfs.Last()->FormSystemMatrix(*essentialTrueDofs.Last(), opr);
+   opr.SetOperatorOwner(false);
+
+   Vector diag(fespace.GetTrueVSize());
+   bfs.Last()->AssembleDiagonal(diag);
+
+   Solver* smoother = new OperatorChebyshevSmoother(
+      *opr, diag, *essentialTrueDofs.Last(), smoothers_cheby_order,
+      fespace.GetParMesh()->GetComm());
+
+   AddLevel(opr.Ptr(), smoother, true, true);
+}
+
+void DiffusionMultigrid::ConstructCoarseOperatorAndSolver(
+   SolverConfig config, ParFiniteElementSpace& fespace, Array<int>& ess_bdr)
+{
+   ConstructBilinearForm(fespace, ess_bdr, GetCoarseAssemblyLevel(config));
+   ParBilinearForm &a = static_cast<ParBilinearForm&>(*bfs.Last());
+   Array<int> &ess_dofs = *essentialTrueDofs.Last();
+
+   a.FormSystemMatrix(ess_dofs, A_coarse);
+
+   OperatorPtr A_prec;
+   if (NeedsLOR(config))
+   {
+      if (Mpi::Root())
+      {
+         std::cout << "Forming LOR discretization..." << std::endl;
+      }
+      lor.reset(new ParLORDiscretization(a, ess_dofs));
+      A_prec = lor->GetAssembledSystem();
+      if (Mpi::Root())
+      {
+         std::cout << "Forming LOR discretization... Done." << std::endl;
+      }
+   }
+   else
+   {
+      A_prec = A_coarse;
+   }
+
+   if (Mpi::Root()) { std::cout << "Forming preconditioner... " << std::endl; }
+   switch (config.type)
+   {
+      case SolverConfig::JACOBI:
+         coarse_precond.reset(new OperatorJacobiSmoother(a, ess_dofs));
+         break;
+      case SolverConfig::FA_HYPRE:
+      case SolverConfig::LOR_HYPRE:
+      {
+         HypreBoomerAMG *amg = new HypreBoomerAMG(*A_prec.As<HypreParMatrix>());
+         amg->SetPrintLevel(1);
+         Vector b(amg->Height());
+         Vector x(amg->Height());
+         b = 0.0;
+         x = 0.0;
+         amg->Setup(b, x); // Force setup;
+         coarse_precond.reset(amg);
+         break;
+      }
+#ifdef MFEM_USE_AMGX
+      case SolverConfig::FA_AMGX:
+      case SolverConfig::LOR_AMGX:
+      {
+         AmgXSolver *amg = new AmgXSolver;
+         amg->ReadParameters(config.amgx_config_file, AmgXSolver::EXTERNAL);
+         amg->InitExclusiveGPU(MPI_COMM_WORLD);
+         amg->SetOperator(*A_prec.As<HypreParMatrix>());
+         coarse_precond.reset(amg);
+         break;
+      }
+#endif
+      default:
+         MFEM_ABORT("Not available.")
+   }
+
+   if (config.inner_sli) // coarse_solver = SLI
+   {
+      SLISolver *sli = new SLISolver(fespace.GetComm());
+      sli->SetPrintLevel(0);
+      sli->SetAbsTol(0.0);
+      sli->SetRelTol(0.0);
+      sli->SetMaxIter(config.inner_sli_iter);
+      sli->SetOperator(*A_coarse);
+      sli->SetPreconditioner(*coarse_precond);
+      coarse_solver.reset(sli);
+   }
+   else if (config.inner_cg)
+   {
+      CGSolver *cg = new CGSolver(MPI_COMM_WORLD);
+      cg->SetPrintLevel(2);
+      cg->SetMaxIter(100);
+      cg->SetRelTol(1e-8);
+      cg->SetAbsTol(0.0);
+      cg->SetOperator(*A_coarse);
+      cg->SetPreconditioner(*coarse_precond);
+      cg->iterative_mode = false;
+      coarse_solver.reset(cg);
+   }
+   else
+   {
+      coarse_solver = coarse_precond;
+   }
+   if (Mpi::Root())
+   {
+      std::cout << "Forming preconditioner... Done.\n" << std::endl;
+   }
+
+   if (config.coarse_smooth)
+   {
+      Vector diag(fespace.GetTrueVSize());
+      a.AssembleDiagonal(diag);
+
+      Solver *smoother = new OperatorChebyshevSmoother(
+         *A_coarse, diag, ess_dofs, smoothers_cheby_order,
+         fespace.GetParMesh()->GetComm());
+
+      AddLevel(A_coarse.Ptr(), smoother, false, true);
+      AddCoarseSolver(coarse_solver.get(), false);
+   }
+   else
+   {
+      AddLevel(A_coarse.Ptr(), coarse_solver.get(), false, false);
+   }
+}
+
+void DiffusionMultigrid::SetSmoothersChebyshevOrder(int new_cheby_order)
+{
+   for (int level = MultigridBase::coarse_solver ? 0 : 1;
+        level < NumLevels(); level++)
+   {
+      OperatorChebyshevSmoother *cheby =
+         dynamic_cast<OperatorChebyshevSmoother*>(GetSmootherAtLevel(level));
+      if (cheby) { cheby->SetOrder(new_cheby_order); }
+   }
+   smoothers_cheby_order = new_cheby_order;
+}
+
+void DiffusionMultigrid::SetInnerSLINumIter(int inner_sli_iter)
+{
+   SLISolver *sli = dynamic_cast<SLISolver*>(coarse_solver.get());
+   if (sli) { sli->SetMaxIter(inner_sli_iter); }
+}
+
+} // namespace mfem

miniapps/benchmarks/ceed-solver-bps/preconditioners.hpp

@@ -0,0 +1,100 @@
+// 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 __SOLVER_BP_HPP__
+#define __SOLVER_BP_HPP__
+
+#include "mfem.hpp"
+#include <memory>
+
+namespace mfem
+{
+
+struct SolverConfig
+{
+   enum SolverType
+   {
+      JACOBI = 0,
+      FA_HYPRE = 1,
+      LOR_HYPRE = 2,
+      FA_AMGX = 3,
+      LOR_AMGX = 4
+   };
+   SolverType type;
+   const char *amgx_config_file = "amgx/amgx.json";
+   bool inner_cg = false; //<-- use inner CG iteration for coarse solver
+   bool inner_sli = false; //<-- use inner SLI iteration for coarse solver
+   int inner_sli_iter = 1; //<- number of iterations for the inner SLI solver
+   bool coarse_smooth = false; //<- enable level 0 smoothing
+   SolverConfig(SolverType type_) : type(type_) { }
+
+   void Print()
+   {
+      mfem::out << "Coarse solver: ";
+      switch (type)
+      {
+         case JACOBI: mfem::out << "Jacobi"; break;
+         case FA_HYPRE: mfem::out << "Hypre (full)"; break;
+         case LOR_HYPRE: mfem::out << "Hypre (LOR)"; break;
+         case FA_AMGX: mfem::out << "AmgX (full)"; break;
+         case LOR_AMGX: mfem::out << "AmgX (LOR)"; break;
+      }
+      mfem::out << std::endl;
+      // If inner_sli is true inner_cg is not used, see
+      // DiffusionMultigrid::ConstructCoarseOperatorAndSolver():
+      if (inner_sli) { inner_cg = false; }
+      mfem::out << "Inner CG:      "
+                << (inner_cg ? "On" : "Off")
+                << std::endl;
+      mfem::out << "Inner SLI:     " << (inner_sli ? "On" : "Off") << '\n';
+      mfem::out << "Coarse smooth: " << (coarse_smooth ? "On" : "Off") << '\n';
+   }
+};
+
+struct DiffusionMultigrid : GeometricMultigrid
+{
+   Coefficient &coeff;
+   int q1d_inc;
+   IntegrationRules irs;
+   std::unique_ptr<ParLORDiscretization> lor;
+   OperatorPtr A_coarse;
+   std::shared_ptr<Solver> coarse_solver, coarse_precond;
+   int smoothers_cheby_order;
+
+   DiffusionMultigrid(
+      ParFiniteElementSpaceHierarchy& hierarchy,
+      Coefficient &coeff_,
+      Array<int>& ess_bdr,
+      SolverConfig coarse_solver_config,
+      int q1d_inc_ = 0,
+      int smoothers_cheby_order_ = 1);
+
+   void ConstructBilinearForm(
+      ParFiniteElementSpace &fespace,
+      Array<int> &ess_bdr,
+      AssemblyLevel asm_lvl);
+
+   void ConstructOperatorAndSmoother(
+      ParFiniteElementSpace &fespace,
+      Array<int> &ess_bdr);
+
+   void ConstructCoarseOperatorAndSolver(
+      SolverConfig config,
+      ParFiniteElementSpace &fespace,
+      Array<int> &ess_bdr);
+
+   void SetSmoothersChebyshevOrder(int new_cheby_order);
+   void SetInnerSLINumIter(int inner_sli_iter);
+};
+
+} // namespace mfem
+
+#endif

miniapps/benchmarks/ceed-solver-bps/rhs.hpp

@@ -0,0 +1,364 @@
+// 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 __RHS_HPP__
+#define __RHS_HPP__
+
+#include "mfem.hpp"
+#include "general/forall.hpp"
+
+// 0 - Solution described in the CEED MS 36 report
+// 1 - Solution from the "ecp_special_2023" paper (option with cosine):
+//       w(n,x) = \sum_{k=0}^n a^k \cos(b^k \pi (x - 1/2)),   x \in [0,1]
+//     with a = 1/2, b = 3.
+// 2 - Solution from the "ecp_special_2023" paper (option with sine):
+//       w(n,x) = \sum_{k=0}^n a^k \sin(b^k \pi x),   x \in [0,1]
+//     with a = 1/2, b = 3.
+#define CEED_SOLVER_BP_SOLUTION_OPTION 1
+
+namespace mfem
+{
+
+constexpr real_t pi = real_t(M_PI);
+
+#if (CEED_SOLVER_BP_SOLUTION_OPTION == 0)
+
+MFEM_HOST_DEVICE inline
+real_t s(int k, real_t x)
+{
+   return sin(2*pi*k*x);
+}
+
+MFEM_HOST_DEVICE inline
+real_t u(int k, real_t x)
+{
+   real_t skx = s(k,x);
+   real_t sgn = skx < 0 ? -1.0 : 1.0;
+   return exp(-1/skx/skx)*sgn;
+}
+
+MFEM_HOST_DEVICE inline
+real_t u_xx(int k, real_t x)
+{
+   real_t kpix = k*pi*x;
+   real_t csc_2kpix = 1.0/sin(2*kpix);
+   real_t sgn = sin(2*kpix) < 0 ? -1.0 : 1.0;
+   return 2*exp(-csc_2kpix*csc_2kpix)*k*k*pi*pi
+          *(1 + 6*cos(4*kpix) + cos(8*kpix))
+          *pow(csc_2kpix,6)
+          *sgn;
+}
+
+MFEM_HOST_DEVICE inline
+real_t w(int n, real_t x)
+{
+   real_t wkx = 0.0;
+   real_t xx = 2*x - 1; // transform from [0,1] to [-1,1]
+   for (int j=0; j<n; ++j)
+   {
+      int k = pow(3, j);
+      wkx += u(k, xx);
+   }
+   return wkx;
+}
+
+MFEM_HOST_DEVICE inline
+real_t w_xx(int n, real_t x)
+{
+   real_t wkx = 0.0;
+   real_t xx = 2*x - 1; // transform from [0,1] to [-1,1]
+   if (xx == 0.0) { return 0.0; }
+   for (int j=0; j<n; ++j)
+   {
+      int k = pow(3, j);
+      wkx += 4*u_xx(k, xx); // factor of four from reference interval transf.
+   }
+   return wkx;
+}
+
+#elif (CEED_SOLVER_BP_SOLUTION_OPTION == 1)
+
+MFEM_HOST_DEVICE inline
+real_t w(int n, real_t x)
+{
+   //  w(n,x) = \sum_{k=0}^n a^k \cos(b^k \pi (x - 1/2))
+   const real_t a = 0.5, b = 3.;
+   real_t ak = 1.0;
+   real_t xk = pi * (x - 0.5);
+   real_t w_ = ak * cos(xk);
+   for (int k = 1; k <= n; k++)
+   {
+      ak *= a;
+      xk *= b;
+      w_ += ak * cos(xk);
+   }
+   return w_;
+}
+
+MFEM_HOST_DEVICE inline
+real_t w_x(int n, real_t x)
+{
+   //  w'(n,x) = -\pi \sum_{k=0}^n a^k b^k \sin(b^k \pi (x - 1/2))
+   const real_t a = 0.5, b = 3.;
+   real_t ck = -pi;
+   real_t xk = pi * (x - 0.5);
+   real_t w_x_ = ck * sin(xk);
+   for (int k = 1; k <= n; k++)
+   {
+      ck *= a * b;
+      xk *= b;
+      w_x_ += ck * sin(xk);
+   }
+   return w_x_;
+}
+
+MFEM_HOST_DEVICE inline
+real_t w_xx(int n, real_t x)
+{
+   // w''(n,x) = -\pi^2 \sum_{k=0}^n a^k b^{2 k} \cos(b^k \pi (x - 1/2))
+   const real_t a = 0.5, b = 3.;
+   real_t ck = -(pi * pi);
+   real_t xk = pi * (x - 0.5);
+   real_t w_xx_ = ck * cos(xk);
+   for (int k = 1; k <= n; k++)
+   {
+      ck *= a * b*b;
+      xk *= b;
+      w_xx_ += ck * cos(xk);
+   }
+   return w_xx_;
+}
+
+#elif (CEED_SOLVER_BP_SOLUTION_OPTION == 2)
+
+MFEM_HOST_DEVICE inline
+real_t w(int n, real_t x)
+{
+   //  w(n,x) = \sum_{k=0}^n a^k \sin(b^k \pi x)
+   const real_t a = 0.5, b = 3.;
+   real_t ak = 1.0;
+   real_t xk = pi * x;
+   real_t w_ = ak * sin(xk);
+   for (int k = 1; k <= n; k++)
+   {
+      ak *= a;
+      xk *= b;
+      w_ += ak * sin(xk);
+   }
+   return w_;
+}
+
+MFEM_HOST_DEVICE inline
+real_t w_xx(int n, real_t x)
+{
+   // w''(n,x) = -\pi^2 \sum_{k=0}^n a^k b^{2 k} \sin(b^k \pi x)
+   const real_t a = 0.5, b = 3.;
+   real_t ck = -(pi * pi);
+   real_t xk = pi * x;
+   real_t w_xx_ = ck * sin(xk);
+   for (int k = 1; k <= n; k++)
+   {
+      ck *= a * b*b;
+      xk *= b;
+      w_xx_ += ck * sin(xk);
+   }
+   return w_xx_;
+}
+
+#endif // CEED_SOLVER_BP_SOLUTION_OPTION
+
+using BPSFunctionType = real_t(*)(int n, const real_t *xyz);
+
+template <BPSFunctionType F>
+void ProjectBPSFunction(int n, QuadratureFunction &qf)
+{
+   MFEM_PERF_FUNCTION;
+   QuadratureSpaceBase &qs = *qf.GetSpace();
+   Mesh &mesh = *qs.GetMesh();
+   const IntegrationRule &ir = qs.GetIntRule(0);
+
+   auto *geom = mesh.GetGeometricFactors(ir, GeometricFactors::COORDINATES);
+
+   const int dim = qs.GetMesh()->Dimension();
+   const int nq = ir.Size();
+   const int N = qf.Size();
+
+   const real_t *d_x = geom->X.Read();
+   real_t *d_q = qf.Write();
+
+   mfem::forall(N, [=] MFEM_HOST_DEVICE (int ii)
+   {
+      const int i = ii / nq;
+      const int j = ii % nq;
+      real_t xvec[3];
+      for (int d = 0; d < dim; ++d)
+      {
+         xvec[d] = d_x[j + d*nq + i*dim*nq];
+      }
+      d_q[ii] = F(n, xvec);
+   });
+}
+
+MFEM_HOST_DEVICE inline
+real_t sol_1d(const int n, const real_t *xyz)
+{
+   return w(n, xyz[0]);
+}
+
+MFEM_HOST_DEVICE inline
+real_t sol_2d(const int n, const real_t *xyz)
+{
+   return w(n, xyz[0])*w(n, xyz[1]);
+}
+
+MFEM_HOST_DEVICE inline
+real_t sol_3d(const int n, const real_t *xyz)
+{
+   return w(n, xyz[0])*w(n, xyz[1])*w(n, xyz[2]);
+}
+
+struct ExactSolution : Coefficient
+{
+   int dim, n;
+   ExactSolution(int dim_, int n_=0) : dim(dim_), n(n_) { }
+   using Coefficient::Eval;
+   real_t Eval(ElementTransformation &T, const IntegrationPoint &ip) override
+   {
+      real_t xyz[3];
+      Vector transip(xyz, 3);
+      T.Transform(ip, transip);
+      if (dim == 1)
+      {
+         return w(n, xyz[0]);
+      }
+      if (dim == 2)
+      {
+         return w(n, xyz[0])*w(n, xyz[1]);
+      }
+      else // dim == 3
+      {
+         return w(n, xyz[0])*w(n, xyz[1])*w(n, xyz[2]);
+      }
+   }
+
+   void Project(QuadratureFunction &qf) override
+   {
+      switch (dim)
+      {
+         case 1: ProjectBPSFunction<sol_1d>(n, qf); break;
+         case 2: ProjectBPSFunction<sol_2d>(n, qf); break;
+         case 3: ProjectBPSFunction<sol_3d>(n, qf); break;
+         default: MFEM_ABORT("Unsupported dimension.");
+      }
+   }
+};
+
+struct ExactGrad : VectorCoefficient
+{
+   int dim, n;
+   ExactGrad(int dim_, int n_)
+      : VectorCoefficient(dim_), dim(dim_), n(n_) { }
+   using VectorCoefficient::Eval;
+   void Eval(Vector &V, ElementTransformation &T,
+             const IntegrationPoint &ip) override
+   {
+      real_t xyz[3];
+      Vector transip(xyz, 3);
+      T.Transform(ip, transip);
+      V.SetSize(dim);
+      if (dim == 1)
+      {
+         V(0) = w_x(n, xyz[0]);
+      }
+      if (dim == 2)
+      {
+         V(0) = w_x(n, xyz[0])*  w(n, xyz[1]);
+         V(1) =   w(n, xyz[0])*w_x(n, xyz[1]);
+      }
+      else // dim == 3
+      {
+         const real_t wnx = w(n, xyz[0]);
+         const real_t wny = w(n, xyz[1]);
+         const real_t wnz = w(n, xyz[2]);
+         V(0) = w_x(n, xyz[0])*wny           *wnz;
+         V(1) = wnx           *w_x(n, xyz[1])*wnz;
+         V(2) = wnx           *wny           *w_x(n, xyz[2]);
+      }
+   }
+};
+
+MFEM_HOST_DEVICE inline
+real_t rhs_1d(const int n, const real_t *xyz)
+{
+   return -w_xx(n, xyz[0]);
+}
+
+MFEM_HOST_DEVICE inline
+real_t rhs_2d(const int n, const real_t *xyz)
+{
+   return -w_xx(n, xyz[0])*w(n, xyz[1]) - w(n, xyz[0])*w_xx(n, xyz[1]);
+}
+
+MFEM_HOST_DEVICE inline
+real_t rhs_3d(const int n, const real_t *xyz)
+{
+   return -w_xx(n, xyz[0])*w(n, xyz[1])*w(n, xyz[2])
+          - w(n, xyz[0])*w_xx(n, xyz[1])*w(n, xyz[2])
+          - w(n, xyz[0])*w(n, xyz[1])*w_xx(n, xyz[2]);
+}
+
+void ProjectRHS(int n, QuadratureFunction &qf)
+{
+   const int dim = qf.GetSpace()->GetMesh()->Dimension();
+   switch (dim)
+   {
+      case 1: ProjectBPSFunction<rhs_1d>(n, qf); break;
+      case 2: ProjectBPSFunction<rhs_2d>(n, qf); break;
+      case 3: ProjectBPSFunction<rhs_3d>(n, qf); break;
+      default: MFEM_ABORT("Unsupported dimension.");
+   }
+}
+
+struct RHS : Coefficient
+{
+   int dim, n;
+   RHS(int dim_, int n_=0) : dim(dim_), n(n_) { }
+   using Coefficient::Eval;
+   real_t Eval(ElementTransformation &T, const IntegrationPoint &ip) override
+   {
+      real_t xyz[3];
+      Vector transip(xyz, 3);
+      T.Transform(ip, transip);
+      if (dim == 1)
+      {
+         return -w_xx(n, xyz[0]);
+      }
+      if (dim == 2)
+      {
+         return -w_xx(n, xyz[0])*w(n, xyz[1]) - w(n, xyz[0])*w_xx(n, xyz[1]);
+      }
+      else // dim == 3
+      {
+         return -w_xx(n, xyz[0])*w(n, xyz[1])*w(n, xyz[2])
+                - w(n, xyz[0])*w_xx(n, xyz[1])*w(n, xyz[2])
+                - w(n, xyz[0])*w(n, xyz[1])*w_xx(n, xyz[2]);
+      }
+   }
+
+   void Project(QuadratureFunction &qf) override
+   {
+      ProjectRHS(n,qf);
+   }
+};
+
+}
+
+#endif

miniapps/benchmarks/ceed-solver-bps/run.sh

@@ -0,0 +1,125 @@
+bsep="============================================================"
+ssep="----------------------------------------"
+# Enable GPU-aware MPI:
+# gpu_aware_mpi_env_cmd="env MPICH_GPU_SUPPORT_ENABLED=1"
+# gpu_aware_mpi="-g"
+# number of MPI ranks, number of ranks per node, number of nodes:
+np=1
+nrnode=4
+((nnodes = (np+nrnode-1)/nrnode))
+# dev="-d gpu ${gpu_aware_mpi}"
+eps="0.3"
+# mpirun_np="mpirun -np"
+mpirun_np="env MFEM_REPORT_KERNELS=1 mpirun -np"
+# mpirun_np="${gpu_aware_mpi_env_cmd} flux run -x -N ${nnodes} -n"
+# dry run:
+# mpirun_np="echo ${mpirun_np}"
+#    p-MG/LOR + FA-hypre, or diagonal (Jacobi smoother)
+#    prec_type: "p-mg", "lor", or "diag"
+prec_type="lor"
+p_mg_opts="-cb 1"
+# p_mg_opts="-cb 5 -sli -sli-it 6"
+# lor_opts="-cls -cb 5 -sli -sli-it 6"
+# lor_opts="-cls -cb 2 -sli -sli-it 2"
+lor_opts="-cb 2 -sli -sli-it 2"
+mg_set=("1" "1 2" "1 3" "1 2 4" "1 3 5" "1 3 6")
+# mg_set=("1 2")
+#    p=7 and p=8 fail at the moment: "1 3 5 7" "1 3 5 8"
+# per-rank limits on the number of LOR elements for different p, in 2^20 units:
+# (bigger sizes run out of GPU memory, at least with LOR prec.)
+lor_ne_max_all=(4 4 4 4 4 4 4 4)
+# lor_ne_max_all=(18 22 24 24 27 24 8 8) # MI250X
+((lor_ne_min = 40*2**10))
+((np_ = np))
+((mm = 1))
+while ((np_ > 8)); do
+   ((mm++))
+   ((np_ = (np_-1)/8+1))
+done
+((mf = 2**mm))
+((mff = 3*mf))
+echo " *** np = ${np}, mf = ${mf}, mff = ${mff}"
+for mg in "${mg_set[@]}"; do
+   echo "${bsep}"
+   p=(${mg})
+   # p=${p[-1]}
+   p="${p[$((${#p[@]}-1))]}"
+   lor_ne_max="${lor_ne_max_all[$((p-1))]}"
+   ((lor_ne_max *= 2**20))
+   # n_max = floor(lor_ne_max^(1/3))
+   n_max=$(echo "a=e((1/3)*l(${np}*${lor_ne_max}));scale=0;a/1" | bc -l)
+   # for np*lor_ne_max=256^3, the above gives 255, so we adjust the result:
+   while (( (n_max+1)**3 <= np*lor_ne_max )); do
+      ((n_max++))
+   done
+   echo " *** p = ${p}, n_max = ${n_max}"
+   if (( n_max**3 > np*lor_ne_max )); then
+      echo "error: n_max^3 > np*lor_ne_max"
+      exit 1
+   fi
+   echo "${bsep}"
+   nx_set=()
+   for ((nx = (n_max/p/mff)*mff, last_nx = 2*nx; nx >= 6; nx -= mff)); do
+      ((last_ne = last_nx**3))
+      ((ne = nx**3))
+      ((lor_ne = (p*nx)**3))
+      if ((np*lor_ne_min > lor_ne)); then break; fi
+      if ((last_ne < ne*4/3)); then continue; fi
+      nx_set=("${nx}" "${nx_set[@]}")
+      ((ndofs = (p*nx+1)**3))
+      ((rhs_n=0))
+      while ((2*3**(rhs_n+1) <= p*nx)); do
+         ((rhs_n++))
+      done
+      # 2*3**rhs_n <= p*nx < 2*3**(rhs_n+1)
+      printf "np = ${np}, p = ${p}, nx = ${nx}, ndofs = ${ndofs}"
+      # rhs_n for eps = 1:
+      # printf ", rhs_n = ${rhs_n}"
+      printf "\n"
+      ((last_nx = nx))
+   done
+   for nx in "${nx_set[@]}"; do
+      # break;
+      if ((nx % mf != 0)); then
+         echo " *** internal error!"
+         exit 1
+      fi
+      ((rp = mm))
+      ((nx /= mf))
+      if false; then
+         # 0, 1, or 2 additional parallel refinements for 1, 8, or 64 ranks
+         ((np_=np))
+         while ((np_%8 == 0)); do
+            ((np_=np_/8))
+            ((rp++))
+         done
+      fi
+      ((ndofs = (p*nx*2**rp+1)**3))
+      echo "${bsep}"
+      echo "np = ${np}, p = ${p}, ndofs = ${ndofs}"
+      if [[ "$prec_type" == "p-mg" ]]; then
+         # p-MG
+         printf "$mpirun_np ${np} ./solver-bp ${dev} -nrn ${nrnode}"
+         printf " -ey ${eps} -mg \"${mg}\" -cs 1 ${p_mg_opts}"
+         printf " -nx ${nx} -rp ${rp}\n"
+         echo "${ssep}"
+         $mpirun_np "${np}" ./solver-bp ${dev} -nrn ${nrnode} \
+            -ey ${eps} -mg "${mg}" -cs 1 ${p_mg_opts} -nx "${nx}" -rp "${rp}"
+      elif [[ "$prec_type" == "lor" ]]; then
+         # LOR
+         printf "$mpirun_np ${np} ./solver-bp ${dev} -nrn ${nrnode}"
+         printf " -ey ${eps} -mg \"${p}\" -cs 2 ${lor_opts}"
+         printf " -nx ${nx} -rp ${rp}\n"
+         echo "${ssep}"
+         $mpirun_np "${np}" ./solver-bp ${dev} -nrn ${nrnode} \
+            -ey ${eps} -mg "${p}" -cs 2 ${lor_opts} -nx "${nx}" -rp "${rp}"
+      elif [[ "$prec_type" == "diag" ]]; then
+         # Diag
+         printf "$mpirun_np ${np} ./solver-bp ${dev} -nrn ${nrnode}"
+         printf " -ey ${eps} -mg \"${p}\" -cs 0 -nx ${nx} -rp ${rp}\n"
+         echo "${ssep}"
+         $mpirun_np "${np}" ./solver-bp ${dev} -nrn ${nrnode} \
+            -ey ${eps} -mg "${p}" -cs 0 -nx "${nx}" -rp "${rp}"
+      fi
+   done
+done

miniapps/benchmarks/ceed-solver-bps/solver-bp.cpp

@@ -0,0 +1,845 @@
+// 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.
+
+// --------------------------------------------------------------
+//    MFEM Implementation of the CEED Solver Bake-off Problems
+// --------------------------------------------------------------
+//
+// Run a suite of benchmarks and view the results:
+//
+//   1. Edit 'run.sh' to adjust machine and size parameters.
+//   2. Run 'run.sh' redirecting output to a file, e.g.:
+//        bash run.sh > run-001.out
+//   3. Extract the CSV output:
+//        sed -n -e 's/^= CSV:\(.*\)$/\1/p' run-001.out > run-001.csv
+//   3. Edit the script 'plot_csv.py' set the name of your CSV file and,
+//      optionally, customize the plot it generates.
+//   4. Process the CSV file:
+//        python3 plot_csv.py
+//
+// Sample runs:
+//
+//   solver-bp -nx 6
+//   solver-bp -nx 6 -mg "1 2 3"
+//   solver-bp -nx 6 -mg "1 r r 2 3"
+//   solver-bp -nx 6 -rp 2 -mg 3 -cs 1
+//   solver-bp -nx 6 -rp 2 -mg 3 -cs 2
+//
+// Device sample runs:
+//
+//   solver-bp -d cuda -nx 6 -mg "1 r r 2 3" -cs 0
+//   solver-bp -d cuda -nx 6 -rp 2 -mg 3 -cs 3
+//   solver-bp -d cuda -nx 6 -rp 2 -mg 3 -cs 4
+//
+
+#include "mfem.hpp"
+#include "kershaw.hpp"
+#include "rhs.hpp"
+#include "preconditioners.hpp"
+#include <regex>
+#include <fem/integ/bilininteg_diffusion_kernels.hpp>
+
+using namespace std;
+using namespace mfem;
+
+struct MGRefinement
+{
+   enum Type { P_MG, H_MG };
+   const Type type;
+   const int order;
+   MGRefinement(Type type_, int order_) : type(type_), order(order_) { }
+   static MGRefinement p(int order_) { return MGRefinement(P_MG, order_); }
+   static MGRefinement h() { return MGRefinement(H_MG, 0); }
+};
+
+struct CGMonitor : IterativeSolverMonitor
+{
+   const real_t tol;
+   real_t initial_nrm, final_nrm, saved_nrm;
+   int final_it, saved_it;
+
+   CGMonitor(real_t tol_) : tol(tol_) { }
+
+   void MonitorResidual(int it, real_t norm, const Vector &r, bool final)
+   override
+   {
+      MFEM_PERF_FUNCTION;
+
+      MFEM_CONTRACT_VAR(norm);
+      // Avoid recomputing the norm if it was already computed -- this method
+      // is called two times for the final iteration: once with final = false
+      // (possibly triggering the monitor convergence criterion) and a second
+      // time with final = true.
+      bool init_call = (it == 0 && !final);
+      const real_t nrm =
+         (!init_call && it == saved_it) ?
+         saved_nrm :
+         sqrt(InnerProduct(iter_solver->GetComm(), r, r));
+      if ((it == 0 || final) && Mpi::Root())
+      {
+         mfem::out << (final ? "Final" : "   Initial")
+                   << " l2 norm of residual: " << nrm << '\n';
+      }
+      if (init_call)
+      {
+         initial_nrm = nrm;
+         converged = false;
+         final_nrm = -1.0;
+         final_it = -1;
+      }
+      saved_nrm = nrm;
+      saved_it = it;
+      // Check for monitor-triggered convergence
+      converged = (nrm <= tol*initial_nrm);
+      if (final)
+      {
+         final_nrm = nrm;
+         final_it = it;
+      }
+      if (final && Mpi::Root())
+      {
+         mfem::out << "Final relative l2 residual: ";
+         if (initial_nrm == 0.0)
+         {
+            mfem::out << "N/A (initial norm is 0)" << endl;
+         }
+         else
+         {
+            const real_t rel_nrm = nrm/initial_nrm;
+            mfem::out << rel_nrm << '\n';
+            mfem::out << "Average l2 reduction factor: ";
+            if (it == 0) { mfem::out << "N/A"; }
+            else { mfem::out << pow(rel_nrm, 1.0/it); }
+            mfem::out << " [" << it << " iterations]" << endl;
+         }
+      }
+   }
+};
+
+void report_hypre_gpu_status(bool gpu_aware_mpi_requested);
+void report_env_vars();
+real_t verify_ess_bdr(const Vector &b, const Vector &x,
+                      const Array<int> &ess_tdof_list);
+
+template <typename T> void PrintPair(const string &name, T val)
+{
+   cout << setw(14) << left << name << val << '\n';
+}
+
+
+int main(int argc, char *argv[])
+{
+   DiffusionIntegrator::AddSpecialization<3,3,3>();
+   DiffusionIntegrator::AddSpecialization<3,4,4>();
+   DiffusionIntegrator::AddSpecialization<3,5,5>();
+   DiffusionIntegrator::AddSpecialization<3,6,6>();
+   DiffusionIntegrator::AddSpecialization<3,7,7>();
+
+   Mpi::Init(argc, argv);
+   Hypre::Init();
+
+   const char *device_config = "cpu";
+   int nrnode = 4; // number of ranks per node, used for CSV output only
+   bool gpu_aware_mpi = false;
+   int nx = 6, ny = -1, nz = -1;
+   int rhs_n = -1;
+   const char *mg_spec = "1";
+   int q1d_inc = 0; // num 1D qpts = p + 1 + q1d_inc
+   int smoothers_cheby_order = 1;
+   real_t epsy = 1.0, epsz = -1;
+   int ref_par = 0;
+   bool glvis = false;
+   bool paraview = false;
+   SolverConfig coarse_solver(SolverConfig::JACOBI);
+
+   OptionsParser args(argc, argv);
+   args.AddOption(&device_config, "-d", "--device",
+                  "Device configuration string, see Device::Configure().");
+   args.AddOption(&nrnode, "-nrn", "--num-ranks-per-node",
+                  "Number of ranks per compute node. Used to compute the number"
+                  " of nodes written in CSV output.");
+   args.AddOption(&gpu_aware_mpi, "-g", "--gpu-aware-mpi", "-no-g",
+                  "--no-gpu-aware-mpi", "Enable GPU-aware MPI.");
+   args.AddOption(&mg_spec, "-mg", "--multigrid-spec",
+                  "Multigrid specification. See README for description.");
+   args.AddOption(&q1d_inc, "-qi", "--quadrature-points-increment",
+                  "Increment for the 1D quadrature points relative to p + 1");
+   args.AddOption(&smoothers_cheby_order, "-cb",
+                  "--smoothers-chebyshev-order",
+                  "Order of the Chebyshev smoothers for the multigrid.");
+   args.AddOption((int*)&coarse_solver.type, "-cs", "--coarse-solver-config",
+                  "Coarse solver configuration. 0: Jacobi, 1: FA-HYPRE, "
+                  "2: LOR-HYPRE, 3: FA-AMGX, 4: LOR-AMGX.");
+   args.AddOption(&coarse_solver.inner_cg, "-cg", "--inner-cg",
+                  "-no-cg", "--no-inner-cg",
+                  "Use inner CG iteration for the coarse solver.");
+   args.AddOption(&coarse_solver.inner_sli, "-sli", "--inner-sli",
+                  "-no-sli", "--no-inner-sli",
+                  "Use inner SLI iteration for the coarse solver.");
+   args.AddOption(&coarse_solver.inner_sli_iter, "-sli-it",
+                  "--inner-sli-iterations",
+                  "Number of iterations for the inner SLI solver.");
+   args.AddOption(&coarse_solver.coarse_smooth, "-cls", "--coarse-level-smooth",
+                  "-no-cls", "--no-coarse-level-smooth",
+                  "Use coarse smoothing in addition to the coarse solver.");
+   args.AddOption(&coarse_solver.amgx_config_file, "-amgx", "--amgx-config",
+                  "AmgX config JSON file.");
+   args.AddOption(&nx, "-nx", "--nx", "Number of elements in x direction.");
+   args.AddOption(&ny, "-ny", "--ny", "Number of elements in y direction.");
+   args.AddOption(&nz, "-nz", "--nz", "Number of elements in z direction.");
+   args.AddOption(&epsy, "-ey", "--epsy", "Kershaw parameter epsilon y.");
+   args.AddOption(&epsz, "-ez", "--epsz", "Kershaw parameter epsilon z.");
+   args.AddOption(&rhs_n, "-rn", "--rhs-n",
+                  "Parameter n in the RHS function; -1 for default.");
+   args.AddOption(&ref_par, "-rp", "--ref-par",
+                  "Number of uniform parallel refinements to perform.");
+   args.AddOption(&glvis, "-gv", "--glvis", "-no-gv", "--no-glvis",
+                  "Save the mesh and solution for GLVis visualization.");
+   args.AddOption(&paraview, "-pv", "--paraview", "-no-pv", "--no-paraview",
+                  "Save data files for ParaView visualization.");
+   args.ParseCheck();
+
+   if (ny < 0) { ny = nx; }
+   if (nz < 0) { nz = nx; }
+   if (epsz < 0) { epsz = epsy; }
+   // rhs_n default is handled later
+
+   Device device(device_config);
+   device.SetGPUAwareMPI(gpu_aware_mpi);
+   if (Mpi::Root()) { device.Print(); }
+   // Report HYPRE's GPU config and GPU-aware MPI config. Terminates if
+   // GPU-aware MPI is requested but HYPRE's GPU-aware MPI support is disabled.
+   report_hypre_gpu_status(gpu_aware_mpi);
+   // Report environment variables like {CUDA,ROCR}_VISIBLE_DEVICES:
+   report_env_vars();
+
+   // Generate mesh
+   MFEM_PERF_BEGIN("CreateKershawMesh");
+   ParMesh mesh_coarse = CreateKershawMesh(nx, ny, nz, epsy, epsz);
+   MFEM_PERF_END("CreateKershawMesh");
+   const int dim = mesh_coarse.Dimension();
+   for (int i=0; i<ref_par; ++i)
+   {
+      MFEM_PERF_SCOPE("Mesh UniformRefinement");
+      mesh_coarse.UniformRefinement();
+   }
+
+   int coarse_order = 0, order = 0, h_ref = ref_par;
+   // Parse order specification
+   vector<MGRefinement> mg_refinements;
+   {
+      istringstream mg_stream(mg_spec);
+      string ref;
+      mg_stream >> coarse_order;
+      int prev_order = order = coarse_order;
+      if (Mpi::Root()) { cout << "\nCoarse order " << coarse_order << '\n'; }
+      while (mg_stream >> ref)
+      {
+         if (ref == "r")
+         {
+            if (Mpi::Root()) { cout << "h-MG uniform refinement\n"; }
+            mg_refinements.push_back(MGRefinement::h());
+            ++h_ref;
+         }
+         else
+         {
+            try { order = stoi(ref); }
+            catch (...)
+            {
+               MFEM_ABORT("Multigrid refinement must either be an integer or "
+                          "the character `r`");
+            }
+            if (Mpi::Root()) { cout << "p-MG order   " << order << '\n'; }
+            MFEM_VERIFY(order > 0, "Orders must be positive");
+            MFEM_VERIFY(order > prev_order, "Orders must be increasing");
+            mg_refinements.push_back(MGRefinement::p(order));
+            prev_order = order;
+         }
+      }
+   }
+
+   if (order == 1 && coarse_solver.type == SolverConfig::LOR_HYPRE)
+   {
+      // Using ~10^7 elements with p=1 overflows a Vector in the LOR setup.
+      // The Vector has size (3D): (p+1)^3 * 27 * num_elem_ho.
+      // In 3D, for p > 1, the overflow will happen around:
+      // - p=2: ~23.6 million dofs or 2,945,794 elements
+      // - p=3: ~33.6 million dofs or 1,242,757 elements
+      // - p=4: ~40.7 million dofs or   636,292 elements
+      // - p=5: ~46.0 million dofs or   368,225 elements
+      // - p=6: ~50.1 million dofs or   231,885 elements
+      //
+      // Note: the size of the Jacobians at quadrature points (with q1d=p+1) in
+      // 3D is: (p+1)^3 * 9 * num_elem, so 3x smaller than the above Vector.
+      //
+      // For q1d=p+2, the overflow happens around:
+      // - p=1: 8,837,382 elements or ~8.8 million dofs
+      // - p=2: 3,728,271 elements or ~29.8 million dofs
+      // - p=3: 1,908,875 elements or ~51.5 million dofs
+      // - p=4: 1,104,673 elements or ~70.7 million dofs
+      // - p=5:   695,654 elements or ~87.0 million dofs
+      // - p=6:   466,034 elements or ~100.7 million dofs
+      coarse_solver.type = SolverConfig::FA_HYPRE;
+      if (Mpi::Root())
+      {
+         cout << "\nOrder is 1: switching from LOR-HYPRE to FA-HYPRE.\n";
+      }
+   }
+#if 0
+   if (order == 1 && coarse_solver.type == SolverConfig::FA_HYPRE &&
+       coarse_solver.inner_sli)
+   {
+      coarse_solver.inner_sli = false;
+      if (Mpi::Root())
+      {
+         cout << "\nOrder is 1: turning off the inner SLI.\n";
+      }
+   }
+#endif
+
+   MFEM_PERF_BEGIN("Setup [hierarchy]");
+   vector<unique_ptr<FiniteElementCollection>> fe_collections;
+   fe_collections.emplace_back(new H1_FECollection(coarse_order, dim));
+   ParFiniteElementSpace fes_coarse(&mesh_coarse, fe_collections.back().get());
+   ParFiniteElementSpaceHierarchy hierarchy(&mesh_coarse, &fes_coarse,
+                                            false, false);
+
+   for (MGRefinement ref : mg_refinements)
+   {
+      if (ref.type == MGRefinement::H_MG)
+      {
+         hierarchy.AddUniformlyRefinedLevel();
+      }
+      else // P_MG
+      {
+         fe_collections.emplace_back(new H1_FECollection(ref.order, dim));
+         hierarchy.AddOrderRefinedLevel(fe_collections.back().get());
+      }
+   }
+   MFEM_PERF_END("Setup [hierarchy]");
+
+   const int nlevels = hierarchy.GetNumLevels();
+   if (Mpi::Root())
+   {
+      if (nlevels == 1)
+      {
+         cout << "1 level in MG hierarchy. Using coarse solver only." << endl;
+      }
+      else
+      {
+         cout << nlevels << " levels in MG hierarchy." << endl;
+      }
+      coarse_solver.Print();
+      cout << endl;
+   }
+
+   // Determine final nx, ny, nz and use them to determine the default rhs_n.
+   const int ref_factor = pow(2, h_ref);
+   nx *= ref_factor;
+   ny *= ref_factor;
+   nz *= ref_factor;
+   if (rhs_n < 0)
+   {
+      int n_min = min(nx, ny);
+      if (nz > 0) { n_min = min(n_min, nz); }
+      // Find rhs_n such that 2*3^rhs_n <= (order*n_min) < 2*3^{rhs_n+1}
+      rhs_n = 0;
+      for (int l = 2*3; l <= order*n_min; l *= 3) { rhs_n++; }
+      if (epsy < 0.8) { rhs_n--; }
+      if (Mpi::Root()) { cout << "Using rhs_n = " << rhs_n << '\n' << endl; }
+   }
+
+   ParFiniteElementSpace &fes = hierarchy.GetFinestFESpace();
+   ParMesh &mesh = *fes.GetParMesh();
+   MFEM_PERF_BEGIN("ParMesh PrintInfo");
+   mesh.PrintInfo(cout);
+   MFEM_PERF_END("ParMesh PrintInfo");
+   HYPRE_Int ndof = fes.GlobalTrueVSize();
+   if (Mpi::Root())
+   {
+      cout << "\nTotal number of DOFs: " << ndof << endl << endl;
+   }
+
+   // All Dirichlet boundaries
+   Array<int> ess_bdr;
+   if (mesh.bdr_attributes.Size())
+   {
+      ess_bdr.SetSize(mesh.bdr_attributes.Max());
+      ess_bdr = 1;
+   }
+
+   ConstantCoefficient one(1.0);
+   ConstantCoefficient coeff(1.0); // Diffusion coefficient
+   // Set up RHS
+   if (Mpi::Root()) { cout << "Assembling right-hand side..." << endl; }
+   MFEM_PERF_BEGIN("Setup [RHS]");
+   RHS rhs_coeff(dim, rhs_n);
+   ParLinearForm b(&fes);
+   const int rhs_ir_inc = 2*q1d_inc+1;
+   // --> ir_order = 2*(p+1+q1d_inc)-1 --> q1d = p+1+q1d_inc
+   b.AddDomainIntegrator(new DomainLFIntegrator(rhs_coeff, 2, rhs_ir_inc));
+   b.UseFastAssembly(true);
+   b.Assemble();
+   MFEM_PERF_END("Setup [RHS]");
+   if (Mpi::Root()) { cout << "Assembling right-hand side... Done." << endl; }
+
+   // Free device memory: the geometric facros computed so far are:
+   // * the coordinates, for the rhs coefficient evaluation, and
+   // * the detJ, for the DomainLFIntegrator.
+   // These are no-longer needed (?), so we can free the memory.
+   mesh.DeleteGeometricFactors();
+
+   // make sure the GPU is done with any previous tasks:
+   if (Device::Allows(Backend::DEVICE_MASK)) { MFEM_STREAM_SYNC; }
+   // make sure all ranks are done with any previous tasks:
+   MPI_Barrier(MPI_COMM_WORLD);
+   MFEM_PERF_BEGIN("Setup [DiffusionMultigrid]");
+   tic();
+   // Set up operators in the multigrid hierarchy
+   DiffusionMultigrid MG(hierarchy, coeff, ess_bdr, coarse_solver, q1d_inc,
+                         smoothers_cheby_order);
+   MG.SetCycleType(Multigrid::CycleType::VCYCLE, 1, 1);
+   // make sure the GPU is done with all setup tasks:
+   if (Device::Allows(Backend::DEVICE_MASK)) { MFEM_STREAM_SYNC; }
+   // make sure all ranks are done with all setup tasks:
+   MPI_Barrier(MPI_COMM_WORLD);
+   const double t_setup = tic_toc.RealTime();
+   MFEM_PERF_END("Setup [DiffusionMultigrid]");
+
+   ParGridFunction x(&fes);
+   x = 0.0;
+
+   OperatorPtr A;
+   Vector X, B;
+   MFEM_PERF_BEGIN("Setup [MG.FormFineLinearSystem]");
+   MG.FormFineLinearSystem(x, b, A, X, B);
+   MFEM_PERF_END("Setup [MG.FormFineLinearSystem]");
+
+   const real_t l2_tol = 1e-8;
+   CGMonitor monitor(l2_tol);
+
+   CGSolver cg(MPI_COMM_WORLD);
+   cg.SetRelTol(0.0); // use the 'monitor' for convergence
+   cg.SetPrintLevel(3);
+   cg.SetOperator(*A);
+   cg.SetPreconditioner(MG);
+   cg.SetMonitor(monitor);
+   // Run 2 CG iterations to ensure everything is allocated and initialized for
+   // the full CG solve:
+   if (Mpi::Root()) { cout << "Running 2 warm-up CG iterations ...\n"; }
+   MFEM_PERF_BEGIN("Warm-up");
+   cg.SetMaxIter(2);
+   {
+      Vector X_save(X);
+      cg.Mult(B, X);
+      X = X_save;
+   }
+   MFEM_PERF_END("Warm-up");
+   if (coarse_solver.inner_sli &&
+       ((coarse_solver.type == SolverConfig::FA_HYPRE /* && order > 1 */) ||
+        coarse_solver.type == SolverConfig::LOR_HYPRE))
+   {
+      MFEM_PERF_SCOPE("Auto-tuning");
+      // timing data: (t-solve,sli-iter,cheby-order,pcg-iter)
+      std::vector<std::tuple<double,int,int,int>> timings;
+      Vector X_save(X);
+      if (Mpi::Root()) { cout << "\nFinding optimal MG parameters ...\n"; }
+      cg.SetMaxIter(500);
+      for (int sli_it = 1; sli_it <= coarse_solver.inner_sli_iter; sli_it++)
+      {
+         MG.SetInnerSLINumIter(sli_it);
+         for (int cheby_order = 1; cheby_order <= smoothers_cheby_order;
+              cheby_order++)
+         {
+            MFEM_PERF_SCOPE(("Timing [" + to_string(sli_it) + "," +
+                             to_string(cheby_order) + "]").c_str());
+            MG.SetSmoothersChebyshevOrder(cheby_order);
+
+            if (Mpi::Root())
+            {
+               cout << "\nRunning and timing parameters (sli iter, cheby order)"
+                    << " = (" << sli_it << ',' << cheby_order << ") ...\n";
+            }
+            // make sure the GPU is done with any previous tasks:
+            if (Device::Allows(Backend::DEVICE_MASK)) { MFEM_STREAM_SYNC; }
+            // make sure all ranks are done with any previous tasks:
+            MPI_Barrier(MPI_COMM_WORLD);
+            tic();
+            cg.Mult(B, X);
+            // make sure the GPU is done with all solve tasks:
+            if (Device::Allows(Backend::DEVICE_MASK)) { MFEM_STREAM_SYNC; }
+            // make sure all ranks are done with all solve tasks:
+            MPI_Barrier(MPI_COMM_WORLD);
+            const double t_solve = tic_toc.RealTime();
+            if (cg.GetConverged())
+            {
+               timings.emplace_back(t_solve, sli_it, cheby_order,
+                                    cg.GetNumIterations());
+            }
+            X = X_save;
+         }
+      }
+      std::sort(timings.begin(), timings.end());
+      if (Mpi::Root())
+      {
+         cout << "\nSorted timings from rank 0:\n";
+         const auto old_prec = cout.precision(6);
+         const auto old_fmtflags = cout.flags();
+         cout << std::fixed;
+         for (size_t i = 0; i < timings.size(); i++)
+         {
+            cout << setw(2) << i << ": "
+                 << 1e3*std::get<0>(timings[i]) << " ms: ("
+                 << std::get<1>(timings[i]) << ','
+                 << std::get<2>(timings[i]) << "): "
+                 << setw(3) << std::get<3>(timings[i]) << " iter\n";
+         }
+         cout.flags(old_fmtflags);
+         cout.precision(old_prec);
+      }
+      if (timings.size() > 0)
+      {
+         // Use the fastest parameters (as timed on rank 0) for the full solve:
+         int si = std::get<1>(timings[0]);
+         int co = std::get<2>(timings[0]);
+         MPI_Bcast(&si, 1, MPI_INT, 0, MPI_COMM_WORLD);
+         MPI_Bcast(&co, 1, MPI_INT, 0, MPI_COMM_WORLD);
+         MG.SetInnerSLINumIter(si);
+         MG.SetSmoothersChebyshevOrder(co);
+         coarse_solver.inner_sli_iter = si;
+         smoothers_cheby_order = co;
+         if (Mpi::Root())
+         {
+            cout << "\nUsing the fastest option (sli iter, cheby order) = ("
+                 << si << ',' << co << ")\n";
+         }
+      }
+      else
+      {
+         MG.SetInnerSLINumIter(1);
+         MG.SetSmoothersChebyshevOrder(1);
+         coarse_solver.inner_sli_iter = 1;
+         smoothers_cheby_order = 1;
+         if (Mpi::Root())
+         {
+            cout << "\nAll options failed to converge!"
+                 << " Using (sli iter, cheby order) = (1,1)\n";
+         }
+      }
+   }
+   if (Mpi::Root()) { cout << "\nRunning and timing the full CG solve ...\n"; }
+   cg.SetMaxIter(500);
+   // make sure the GPU is done with any previous tasks:
+   if (Device::Allows(Backend::DEVICE_MASK)) { MFEM_STREAM_SYNC; }
+   // make sure all ranks are done with any previous tasks:
+   MPI_Barrier(MPI_COMM_WORLD);
+   MFEM_PERF_BEGIN("Final CG Solve");
+   tic();
+   cg.Mult(B, X);
+   // make sure the GPU is done with all solve tasks:
+   if (Device::Allows(Backend::DEVICE_MASK)) { MFEM_STREAM_SYNC; }
+   // make sure all ranks are done with all solve tasks:
+   MPI_Barrier(MPI_COMM_WORLD);
+   const double t_solve = tic_toc.RealTime();
+   MFEM_PERF_END("Final CG Solve");
+
+   const int niter = cg.GetConverged() ? cg.GetNumIterations() : -1;
+
+   const real_t bdr_err = verify_ess_bdr(B, X, MG.GetFineEssentialTrueDofs());
+   if (Mpi::Root())
+   {
+      MFEM_VERIFY(bdr_err == 0.0, "Incorrect boundary values in solution!"
+                  " bdr_err = " << bdr_err);
+   }
+
+   MG.RecoverFineFEMSolution(X, b, x);
+
+   MFEM_PERF_BEGIN("Compute L2 Error");
+   ExactSolution exact_coeff(dim, rhs_n);
+   // ExactGrad exact_grad_coeff(dim, rhs_n);
+   real_t L2_err = x.ComputeL2Error(exact_coeff);
+   // real_t grad_err = x.ComputeGradError(&exact_grad_coeff);
+   MFEM_PERF_END("Compute L2 Error");
+   if (Mpi::Root())
+   {
+      cout << "\nL2 Error: " << setprecision(10) << scientific
+           << L2_err << '\n';
+      // cout << "\nGrad Error: " << setprecision(10) << scientific
+      //      << grad_err << '\n';
+   }
+
+   if (glvis)
+   {
+      ofstream mesh_ofs(MakeParFilename("mesh.", Mpi::WorldRank()));
+      mesh_ofs.precision(8);
+      mesh.Print(mesh_ofs);
+
+      ofstream sol_ofs(MakeParFilename("sol.", Mpi::WorldRank()));
+      sol_ofs.precision(8);
+      x.Save(sol_ofs);
+   }
+
+   if (paraview)
+   {
+      ParGridFunction rhs_gf(&fes), exact_gf(&fes), error_gf(&fes);
+      rhs_gf.ProjectCoefficient(rhs_coeff);
+      exact_gf.ProjectCoefficient(exact_coeff);
+
+      subtract(exact_gf, x, error_gf);
+
+      ParaViewDataCollection dc("SolverBP", &mesh);
+      dc.RegisterField("u", &x);
+      dc.RegisterField("rhs", &rhs_gf);
+      dc.RegisterField("exact", &exact_gf);
+      dc.RegisterField("error", &error_gf);
+      dc.SetPrefixPath("ParaView");
+      dc.SetLevelsOfDetail(order);
+      dc.SetHighOrderOutput(true);
+      dc.SetCycle(0);
+      dc.SetTime(0.0);
+      dc.Save();
+   }
+
+   const long long nel = mesh.GetGlobalNE();
+   if (nz == 0) { MFEM_VERIFY(nel == nx*ny, "Wrong number of elements"); }
+   else { MFEM_VERIFY(nel == nx*ny*nz, "Wrong number of elements"); }
+
+   if (Mpi::Root())
+   {
+      cout << "\n= Results\n";
+      PrintPair("nranks", Mpi::WorldSize());
+      PrintPair("nx", nx);
+      PrintPair("ny", ny);
+      PrintPair("nz", nz);
+      PrintPair("degree", order);
+      PrintPair("rhs_n", rhs_n);
+      PrintPair("epsy", epsy);
+      PrintPair("epsz", epsz);
+      PrintPair("ndof", ndof);
+      PrintPair("niter", niter);
+
+      // Should also output:
+      // code id
+      // prec id
+      // machine id
+      // number of supercomputer nodes
+      // number of 1d quadrature points
+      // initial and final residuals
+      // error
+
+      // Timings
+      PrintPair("t_setup", t_setup);
+      PrintPair("t_solve", t_solve);
+
+      cout << "\nSolve MDOFs/rank/sec:   "
+           << ndof/1e6/Mpi::WorldSize()/t_solve << '\n';
+
+      // CSV fields:
+      // 1. code ID
+      // 2. preconditioner ID
+      // 3. machine ID
+      // 4. number of nodes
+      // 5. number of MPI ranks
+      // 6,7,8. n_x, n_y, n_z
+      // 9. solution polynomial degree
+      // 10. number of 1D quadrature points
+      // 11,12. eps_y, eps_z
+      // 13. ndofs (including Dirichlet boundary)
+      // 14. niter
+      // 15,16. initial and final residuals
+      // 17. error
+      // 18. t_setup (preconditioner setup)
+      // 19. t_solve (total iter time)
+      //
+      // extract the CSV lines from the output with:
+      //    grep "= CSV:" out.txt | sed -e 's/^= CSV://' > out.csv
+      cout << "\n= CSV:"
+           << "MFEM-" + string(device_config); // 1
+      string hypre_str =
+#if defined(HYPRE_USING_HIP)
+         "hypre-hip"
+#elif defined(HYPRE_USING_CUDA)
+         "hypre-cuda"
+#else
+         "hypre-cpu"
+#endif
+         ;
+      auto cs = coarse_solver.type;
+      string prec_id;
+      if (cs == SolverConfig::FA_HYPRE) // p-MG, add (sli-iter,cheby-order)
+      {
+         prec_id = hypre_str + "-pMG(";
+      }
+      else if (cs == SolverConfig::LOR_HYPRE) // LOR, add (sli-iter,cheby-order)
+      {
+         prec_id = hypre_str + "-LOR(";
+      }
+      else if (cs == SolverConfig::JACOBI)
+      {
+         prec_id = "diag(";
+      }
+      else
+      {
+         prec_id = "(unknown)(";
+      }
+      if (coarse_solver.inner_cg)
+      {
+         prec_id += "cg;";
+      }
+      if (coarse_solver.inner_sli)
+      {
+         prec_id += to_string(coarse_solver.inner_sli_iter) + ";";
+      }
+      prec_id += to_string(smoothers_cheby_order) +
+                 (coarse_solver.coarse_smooth ? "c" : "") + ")";
+      prec_id += "-" + regex_replace(mg_spec, regex(" "), "-");
+      cout << ',' << prec_id; // 2
+      const char *hostname = getenv("HOSTNAME");
+      if (!hostname) { hostname = getenv("HOST"); }
+      string host_id = regex_replace(hostname ? hostname : "(unknown)",
+                                     regex("[0-9]*$"), "");
+      cout << ',' << host_id; // 3
+      cout << ',' << (fes.GetNRanks() + (nrnode-1))/nrnode; // 4
+      cout << ',' << fes.GetNRanks(); // 5
+      cout << ',' << nx << ',' << ny << ',' << nz; // 6,7,8
+      cout << ',' << order; // 9
+      // DiffusionMultigrid::ConstructBilinearForm p+1+q1d_inc 1D points
+      real_t Q1D = order + 1 + q1d_inc;
+      cout << ',' << defaultfloat << Q1D; // 10 (note: written as real_t)
+      cout << ',' << scientific << epsy << ',' << epsz; // 11,12
+      cout << ',' << ndof; // 13
+      cout << ',' << niter; // 14
+      cout << ',' << monitor.initial_nrm << ',' << monitor.final_nrm; // 15,16
+      cout << ',' << L2_err; // 17
+      // cout << ',' << grad_err; // 17 *** for testing ***
+      cout << ',' << t_setup << ',' << t_solve; // 18,19
+      cout << endl;
+   }
+
+   return 0;
+}
+
+void report_hypre_gpu_status(bool gpu_aware_mpi_requested)
+{
+#if defined(HYPRE_WITH_GPU_AWARE_MPI) || defined(HYPRE_USING_GPU_AWARE_MPI)
+   bool hypre_gpu_aware_mpi = true;
+#else
+   bool hypre_gpu_aware_mpi = false;
+#endif
+#if (MFEM_HYPRE_VERSION > 23000)
+   hypre_gpu_aware_mpi = hypre_gpu_aware_mpi && hypre_GetGpuAwareMPI();
+#endif
+   if (Mpi::Root())
+   {
+      MFEM_VERIFY(!gpu_aware_mpi_requested || hypre_gpu_aware_mpi,
+                  "GPU-aware MPI requested but HYPRE's GPU-aware MPI support"
+                  " is not enabled");
+      cout << "\nHYPRE GPU support: "
+           << (HypreUsingGPU() ? "enabled" : "disabled");
+      cout << "\nHYPRE GPU-aware MPI support: "
+           << (hypre_gpu_aware_mpi ? "enabled" : "disabled") << endl;
+   }
+}
+
+void report_env_vars()
+{
+   const int myid = Mpi::WorldRank();
+   // const int lastid = min(Mpi::WorldSize(),4)-1; // show up to 4 ranks
+   const int lastid = Mpi::WorldSize()-1;
+   if (myid > lastid) { return; }
+   Array<char> recv_buf;
+   int buflen = -1, tag = 42;
+   const char *env_vars[] =
+   {
+      "HOST", "HOSTNAME", "MPICH_GPU_SUPPORT_ENABLED", "CUDA_VISIBLE_DEVICES",
+      "ROCR_VISIBLE_DEVICES"
+   };
+   const int num_env_vars = sizeof(env_vars)/sizeof(env_vars[0]);
+   // Send strings to rank 0, so that they can be printed in order, guaranteed.
+   // Every rank > 0 sends to rank 0:
+   if (myid > 0)
+   {
+      for (int ev = 0; ev < num_env_vars; ev++)
+      {
+         const char *env_var_val = getenv(env_vars[ev]);
+         buflen = env_var_val ? int(strlen(env_var_val)+1) : -1;
+         MPI_Send(&buflen, 1, MPI_INT, 0, tag, MPI_COMM_WORLD);
+         if (env_var_val)
+         {
+            MPI_Send(env_var_val, buflen, MPI_CHAR, 0, tag, MPI_COMM_WORLD);
+         }
+      }
+   }
+   else // myid == 0
+   {
+      cout << "\nDefined environment variables:\n";
+      for (int id = 0; id <= lastid; id++)
+      {
+         cout << "[rank " << id << "]:";
+         for (int ev = 0, vars_shown = 0; ev < num_env_vars; ev++)
+         {
+            const char *env_var_val = nullptr;
+            if (id == 0)
+            {
+               env_var_val = getenv(env_vars[ev]);
+               buflen = env_var_val ? 0 : -1;
+            }
+            else
+            {
+               MPI_Recv(&buflen, 1, MPI_INT, id, tag, MPI_COMM_WORLD,
+                        MPI_STATUS_IGNORE);
+            }
+            if (buflen != -1)
+            {
+               if (id > 0)
+               {
+                  recv_buf.SetSize(buflen);
+                  MPI_Recv(recv_buf.begin(), buflen, MPI_CHAR, id, tag,
+                           MPI_COMM_WORLD, MPI_STATUS_IGNORE);
+                  env_var_val = recv_buf.begin();
+               }
+               if (vars_shown)
+               {
+                  cout << "\n[rank " << id << "]:";
+               }
+               cout << ' ' << env_vars[ev] << '=' << env_var_val;
+               vars_shown++;
+            }
+         }
+         cout << '\n';
+      }
+      if (lastid < Mpi::WorldSize()-1)
+      {
+         cout << "... [only " << lastid+1 << '/' << Mpi::WorldSize()
+              << " ranks shown]\n";
+      }
+      cout << flush;
+   }
+}
+
+real_t verify_ess_bdr(const Vector &b, const Vector &x,
+                      const Array<int> &ess_tdof_list)
+{
+   Vector d(ess_tdof_list.Size());
+   auto d_b = b.Read();
+   auto d_x = x.Read();
+   auto d_d = d.Write();
+   auto d_ess_ind = ess_tdof_list.Read();
+   mfem::forall(ess_tdof_list.Size(), [=] MFEM_HOST_DEVICE (int i)
+   {
+      const int ind = d_ess_ind[i];
+      d_d[i] = -fabs(d_b[ind] - d_x[ind]);
+   });
+   real_t d_max = -d.Min(); // max is not implemented on device
+   MPI_Allreduce(MPI_IN_PLACE, &d_max, 1, MFEM_MPI_REAL_T, MPI_MAX,
+                 MPI_COMM_WORLD);
+   return d_max;
+}

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

miniapps/shifted/distance.cpp

@@ -384,8 +384,10 @@ int main(int argc, char *argv[])
    dacol.Save();
 
    ConstantCoefficient zero(0.0);
+   Vector zero_vec(dim); zero_vec = 0_r;
+   VectorConstantCoefficient vzero(zero_vec);
    const real_t s_norm  = distance_s.ComputeL2Error(zero),
-                v_norm  = distance_v.ComputeL2Error(zero);
+                v_norm  = distance_v.ComputeL2Error(vzero);
    if (myid == 0)
    {
       cout << fixed << setprecision(10) << "Norms: "