Merge branch 'complex-coef-dev' of github.com:mfem/mfem into complex-coef-dev

Switching to complex_t
Adding complex_t type using code by @camierjs
2025-07-22 10:46:29 -07:00 · 2025-07-22 10:45:56 -07:00 · 2025-07-22 10:43:59 -07:00 · 2025-07-15 09:22:34 +08:00 · 2025-07-14 13:44:17 -07:00 · 2025-07-14 13:18:58 -07:00
230 changed files with 9508 additions and 19119 deletions
@@ -12,7 +12,7 @@ coverage:
        threshold: 0%
        base: auto
        branches:
-          - main
+          - master
        if_ci_failed: error
        informational: true
        only_pulls: true
@@ -22,7 +22,7 @@ coverage:
        threshold: 1%  # allows variations around the target
        base: auto
        branches:
-          - main
+          - master
        if_ci_failed: error
        only_pulls: true

@@ -132,14 +132,12 @@ jobs:
            hypre-target: int32
            precision: fp64
            enzyme: true
-            config-opts: MFEM_USE_ENZYME=YES ENZYME_DIR=$(brew --prefix enzyme) LDFLAGS=-L$LLVM_PREFIX/lib/c++
+            config-opts: MFEM_USE_ENZYME=YES ENZYME_DIR=$(brew --prefix enzyme)

    name: ${{ matrix.os }}-${{ matrix.build-system }}-${{ matrix.target }}-${{ matrix.mpi }}-${{ matrix.hypre-target }}-${{ matrix.precision }}${{ matrix.enzyme && '-enzyme' || '' }}

    runs-on: ${{ matrix.os }}

-    continue-on-error: ${{ matrix.enzyme && true || false }}
-
    steps:
      # Fix 'No space left on device' errors for Ubuntu builds.
      - name: Run Actions Cleaner
@@ -170,13 +168,10 @@ jobs:
          env
        shell: bash

-      # For info on Xcode see:
-      #  - https://github.com/actions/runner-images/issues/12541
-      #  - https://github.com/actions/runner-images/blob/releases/macos-15-arm64/20250811/images/macos/macos-15-arm64-Readme.md#xcode
      - name: Xcode version setup (MacOS)
        if: matrix.os == 'macos-latest'
        run: |
-          XCODE_PATH="/Applications/Xcode_16.4.app"
+          XCODE_PATH="/Applications/Xcode_15.3.app"
          echo "> sudo xcode-select -s ${XCODE_PATH}"
          sudo xcode-select -s ${XCODE_PATH}
          echo "> g++ -v"
@@ -294,12 +289,10 @@ jobs:
        run: |
          export HOMEBREW_NO_INSTALL_CLEANUP=1
          brew update
-          brew install enzyme
-          ENZYME_LLVM=$(brew info enzyme | sed -n 's/^Required:.*\(llvm[^ ]*\).*/\1/p')
-          LLVM_PREFIX=$(brew --prefix $ENZYME_LLVM)
-          echo "LLVM_PREFIX=$LLVM_PREFIX" >> $GITHUB_ENV
-          echo "OMPI_CC=$LLVM_PREFIX/bin/clang" >> $GITHUB_ENV
-          echo "OMPI_CXX=$LLVM_PREFIX/bin/clang++" >> $GITHUB_ENV
+          brew install llvm@19 enzyme
+          echo "LLVM_PREFIX=$(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

      # MFEM build and test
      - name: build
@@ -29,47 +29,3 @@ jobs:
        operations-per-run: 500
        exempt-issue-labels: "bug,WIP,ready-for-review,in-review,in-next"
        exempt-pr-labels: "bug,WIP,ready-for-review,in-review,in-next"
-
-  # Stale action for PRs with "in-review" label.
-  stale-in-review-pr:
-
-    runs-on: ubuntu-latest
-    permissions:
-      issues: write
-      pull-requests: write
-      actions: write
-
-    steps:
-    - uses: actions/stale@v9
-      with:
-        repo-token: ${{ secrets.GITHUB_TOKEN }}
-        stale-pr-message: ':warning: This PR has been automatically marked as stale because it has not had any activity in the last 150 days. *If no activity occurs in the next 30 days, it will be automatically closed.* Thank you for your contributions.'
-        only-pr-labels: "in-review"
-        days-before-pr-stale: 150
-        days-before-pr-close: 30
-        days-before-issue-stale: -1
-        days-before-issue-close: -1
-        stale-pr-label: 'stale'
-        operations-per-run: 500
-
-  # Stale action for PRs with "WIP" label.
-  stale-wip-pr:
-
-    runs-on: ubuntu-latest
-    permissions:
-      issues: write
-      pull-requests: write
-      actions: write
-
-    steps:
-    - uses: actions/stale@v9
-      with:
-        repo-token: ${{ secrets.GITHUB_TOKEN }}
-        stale-pr-message: ':warning: This PR has been automatically marked as stale because it has not had any activity in the last 300 days. *If no activity occurs in the next 30 days, it will be automatically closed.* Thank you for your contributions.'
-        only-pr-labels: "WIP"
-        days-before-pr-stale: 300
-        days-before-pr-close: 30
-        days-before-issue-stale: -1
-        days-before-issue-close: -1
-        stale-pr-label: 'stale'
-        operations-per-run: 500
@@ -208,13 +208,10 @@ miniapps/electromagnetics/volta
 miniapps/electromagnetics/tesla
 miniapps/electromagnetics/maxwell
 miniapps/electromagnetics/joule
-miniapps/electromagnetics/lorentz
 miniapps/electromagnetics/Volta-AMR*
 miniapps/electromagnetics/Tesla-AMR*
 miniapps/electromagnetics/Maxwell-Parallel*
-miniapps/electromagnetics/Joule_[0-9]*
-miniapps/electromagnetics/Lorentz_[0-9]*
-miniapps/electromagnetics/Lorentz.dat
+miniapps/electromagnetics/Joule_*

 miniapps/gslib/field-diff
 miniapps/gslib/field-interp
@@ -270,9 +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
@@ -280,7 +277,7 @@ miniapps/autodiff/seqtest
 miniapps/autodiff/par_example
 miniapps/autodiff/seq_example
 miniapps/autodiff/seq_test
-miniapps/autodiff/Example/*
+miniapps/autodiff/Exampl*

 miniapps/navier/navier_mms
 miniapps/navier/navier_kovasznay
@@ -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 Dane resource are allocated/released once for all.
+# - Allocate/Release is where ruby 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:
-dane-build-and-test:
+ruby-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 @@ dane-build-and-test:
    AUTOTEST: "${AUTOTEST}"
    AUTOTEST_COMMIT: "${AUTOTEST_COMMIT}"
  trigger:
-    include: .gitlab/dane-build-and-test.yml
+    include: .gitlab/ruby-build-and-test.yml
    strategy: depend

-dane-baseline:
+ruby-baseline:
  stage: sub-pipelines
  variables:
    # Explicitly pass down values that we want to be able to set when triggering
@@ -73,7 +73,7 @@ dane-baseline:
    AUTOTEST: "${AUTOTEST}"
    AUTOTEST_COMMIT: "${AUTOTEST_COMMIT}"
  trigger:
-    include: .gitlab/dane-baseline.yml
+    include: .gitlab/ruby-baseline.yml
    strategy: depend

 lassen-build-and-test:
@@ -24,7 +24,7 @@ and `test type`.

 Machines typically include:

-* Dane: Intel Sapphire Rapids
+* Ruby: 2nd Gen Intel Xeon (Cascade Lake)
 * 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 Dane for example resumes to:
+spack spec to use. Adding a job on ruby for example resumes to:

 ```yaml
 <job_name>:
  variables:
    SPEC: "<spack_spec>"
-  extends: .build_and_test_on_dane
+  extends: .build_and_test_on_ruby
 ```

 The remaining and non trivial work is to make sure this spec is working. To
@@ -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 Dane, there is only one allocation shared among jobs in order to
+# On LLNL's ruby, 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}
@@ -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 Dane machine at LLNL
+# GitLab pipelines configurations for the Ruby machine at LLNL
 variables:
-  MACHINE_NAME: dane
+  MACHINE_NAME: ruby

-.on_dane:
+.on_ruby:
  tags:
    - shell
-    - dane
+    - ruby
  rules:
-    # Don't run dane jobs if...
-    - if: '$CI_COMMIT_BRANCH =~ /_qnone/ || $ON_DANE == "OFF"'
+    # Don't run ruby jobs if...
+    - if: '$CI_COMMIT_BRANCH =~ /_qnone/ || $ON_RUBY == "OFF"'
      when: never
    # Don't run autotest update if...
    - if: '$CI_JOB_NAME =~ /report/ && $AUTOTEST != "YES"'
@@ -40,17 +40,16 @@ variables:
    - when: on_success

 # Spack helped builds
-# Generic dane build job, extending build script
-.build_and_test_on_dane:
-  extends: [.on_dane]
+# Generic ruby build job, extending build script
+.build_and_test_on_ruby:
+  extends: [.on_ruby]
  stage: build_and_test
  script:
    # THREADS is used by 'tests/gitlab/build_and_test', run below
-    # Dane has 224 threads/node and we run 7 separate jobs: 224=7*32
-    - export THREADS=28
+    - export THREADS=16
    - 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=${JOBID}" ) --reservation=ci -t 60 -N 1 tests/gitlab/build_and_test --spec "${SPEC}" --data-dir "${MFEM_DATA_DIR}" --data
+    - 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
@@ -18,7 +18,7 @@
 setup_baseline:
  tags:
    - shell
-    - dane
+    - ruby
  stage: setup
  variables:
    GIT_STRATEGY: none
@@ -16,7 +16,7 @@
 setup:
  tags:
    - shell
-    - dane
+    - ruby
  stage: setup
  variables:
    GIT_STRATEGY: none
@@ -19,8 +19,8 @@ stages:
  - cleanup
  - baseline_publish

-baselinecheck_mfem_intel_dane:
-  extends: [.on_dane]
+baselinecheck_mfem_intel_ruby:
+  extends: [.on_ruby]
  stage: baseline_check
  variables:
    # TPLS_DIR is used in .gitlab/scripts/baseline to provide the tpls location
@@ -31,8 +31,8 @@ baselinecheck_mfem_intel_dane:
  script:
    - echo ${BUILD_ROOT}
    - echo ${TPLS_DIR}
-    # Used by the tests in MFEM/tests, dane has 224 threads/node:
-    - export MFEM_TEST_NP=192
+    # Used by the tests in MFEM/tests:
+    - export MFEM_TEST_NP=48
    # 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_dane:
  allow_failure: true

 cleanup:
-  extends: .on_dane
+  extends: .on_ruby
  stage: cleanup
  variables:
    GIT_STRATEGY: none
@@ -53,7 +53,7 @@ cleanup:
    - rm -rf "${BUILD_ROOT}" || true

 report_baseline:
-  extends: [.on_dane]
+  extends: [.on_ruby]
  stage: baseline_report
  script:
    - echo ${MACHINE_NAME}
@@ -113,8 +113,8 @@ report_baseline:
        exit $err
      ) 9> autotest.lock

-baselinepublish_mfem_dane:
-  extends: [.on_dane]
+baselinepublish_mfem_ruby:
+  extends: [.on_ruby]
  stage: baseline_publish
  rules:
    # - if: '$CI_COMMIT_BRANCH == "master" || $REBASELINE == "YES"'
@@ -129,5 +129,5 @@ baselinepublish_mfem_dane:

 include:
  - local: .gitlab/configs/common.yml
-  - local: .gitlab/configs/dane-config.yml
+  - local: .gitlab/configs/ruby-config.yml
  - local: .gitlab/configs/setup-baseline.yml
@@ -19,54 +19,54 @@ stages:
 allocate_resource:
  variables:
    GIT_STRATEGY: none
-  extends: .on_dane
+  extends: .on_ruby
  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 Dane machine at LLNL
+# GitLab jobs for the Ruby machine at LLNL
 debug_ser_gcc_10:
  variables:
    SPEC: "%gcc@10.3.1 +debug~mpi"
-  extends: .build_and_test_on_dane
+  extends: .build_and_test_on_ruby

 debug_par_gcc_10:
  variables:
    SPEC: "%gcc@10.3.1 +debug+mpi"
-  extends: .build_and_test_on_dane
+  extends: .build_and_test_on_ruby

 opt_ser_gcc_10:
  variables:
    SPEC: "%gcc@10.3.1 ~mpi"
-  extends: .build_and_test_on_dane
+  extends: .build_and_test_on_ruby

 opt_par_gcc_10:
  variables:
    SPEC: "%gcc@10.3.1"
-  extends: .build_and_test_on_dane
+  extends: .build_and_test_on_ruby

 opt_par_gcc_10_sundials:
  variables:
    SPEC: "%gcc@10.3.1 +sundials"
-  extends: .build_and_test_on_dane
+  extends: .build_and_test_on_ruby

 opt_par_gcc_10_petsc:
  variables:
    SPEC: "%gcc@10.3.1 +petsc ^petsc+mumps~superlu-dist"
-  extends: .build_and_test_on_dane
+  extends: .build_and_test_on_ruby

 opt_par_gcc_10_pumi:
  variables:
    SPEC: "%gcc@10.3.1 +pumi"
-  extends: .build_and_test_on_dane
+  extends: .build_and_test_on_ruby

 # Release
 release_resource:
  variables:
    GIT_STRATEGY: none
-  extends: .on_dane
+  extends: .on_ruby
  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_dane
+    - .on_ruby
    - .report_job_success

 report_job_failure:
  stage: release_resource_and_report
  extends:
-    - .on_dane
+    - .on_ruby
    - .report_job_failure

 include:
  - local: .gitlab/configs/common.yml
-  - local: .gitlab/configs/dane-config.yml
+  - local: .gitlab/configs/ruby-config.yml
  - local: .gitlab/configs/setup-build-and-test.yml
  - local: .gitlab/configs/report-build-and-test.yml
@@ -14,7 +14,7 @@
 # locals
 glob_err=${BASELINE_TEST}.err
 base=${BASELINE_TEST}-${SYS_TYPE}
-if [[ "${MACHINE_NAME}" == "dane" ]]; then
+if [[ "${MACHINE_NAME}" == "ruby" ]]; then
  base="${BASELINE_TEST}-${MACHINE_NAME}"
 fi
 base_diff=${base}.diff
@@ -31,7 +31,7 @@ cd tests
 mkdir _${BASELINE_TEST} && cd _${BASELINE_TEST}

 # run
-if [[ "${MACHINE_NAME}" == "dane" ]]; then
+if [[ "${MACHINE_NAME}" == "ruby" ]]; 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}"
@@ -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 dane baselines.
+# There will be collision between corona and ruby 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 dane baselines.
+# There will be collision between corona and ruby baselines.
 # Once the corresponding files have been generated, we can switch to machine
 # specific ref.
 SAVED_NAME=baseline-${SYS_TYPE}.saved
@@ -27,29 +27,9 @@ Discretization improvements
 - In the ParMoonolith integration, added support for variational resampling of
  H1 vector fields.

- Added support for boundary integration to the hyperbolic framework. In this
-  regard, new classes `BdrHyperbolicDirichletIntegrator` and
-  `BoundaryHyperbolicFlowIntegrator` have been introduced for implementation
-  of weak Dirichlet boundary conditions with a general flux or for the linear
-  case respectively.
-
- Added method to compute piecewise linear bounds on high-order functions on
-  tensor-product elements.
-
- Parallel anisotropic refinement of hexahedral meshes is now supported,
-  provided that neighboring hexahedra are not refined in conflicting directions.
-  A new ParMesh method is added to check for such conflicts, before refinement.
-
- Renamed the default GitHub branch from "master" to "main".
-
 Meshing improvements
 --------------------

- Introduced NC-patch NURBS meshes, which are conforming element-wise but allow
-  for nonconforming patch topology. This new mesh format supports element
-  spacing formulas for refinement, as well as local refinement factors for a
-  subset of knot vectors.
-
 - Added support for higher order meshes in Mesh::MakeSimplicial and
  ParMesh::MakeSimplicial.

@@ -64,20 +44,9 @@ GPU computing
  set. This is most often used for setting constant essential boundary
  conditions. A new function Vector::SetSubVectorHost has been added in cases
  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`.
- Added new method: GridFunction::GetGradients, with GPU support, for computing
-  the gradients of a GridFunction on all elements.
- Added GPU support in GradientGridFunctionCoefficient and
-  InnerProductCoefficient by implementing their Project methods.
-
 New and updated examples and miniapps
 -------------------------------------
 - Added miniapps to demonstrate an implementation of the absolute-value
@@ -86,22 +55,10 @@ New and updated examples and miniapps
  operators as smoothers.
  These miniapps can be found in `miniapps/diag-smoothers`.

- Added a new miniapp (meshing/mesh-bounding-boxes) that computes the bounding
-  boxes for each element of a given mesh, and the bounds on the determinant of
-  the Jacobian of the transformation.
-
- Added a new miniapp (tools/gridfunction-bounds) to compute piecewise linear
-  bounds on a given high-order grid function.
-
- Added a new miniapp (electromagnetics/lorentz) which computes the trajectory
-  of a charged particle, subject to Lorentz forces, in electrostatic and/or
-  magnetostatic fields as computed by the volta or tesla miniapps.
-
-API changes:
+API changes
 -----------
 - mfem::internal::tensor and mfem::internal::dual have been moved to
  mfem::future::tensor and mfem::future::dual.
-
 - API addition: in class `Operator`, added virtual functions: `AbsMult`, and
  `AbsMultTranspose`; in class `Vector`, added `Abs` and `Pow`.

@@ -109,25 +66,16 @@ Miscellaneous
 -------------
 - Added the "gpu", "raja-gpu", and "ceed-gpu" backend aliases/shortcuts which
  automatically select between CUDA or HIP.
-
 - The CUDA-specific names used by some of the unit tests like 'cunit_tests' and
  'pcunit_tests' were replaced by names using 'gpu' instead of 'c' (short for
  CUDA) or 'cuda'. These tests automatically run the CUDA/HIP tests based on the
  MFEM build configuration.
-
 - Added the option to enable GPU-aware MPI in MFEM using the environment
  variable 'MFEM_GPU_AWARE_MPI' set to any value. Setting this environment
  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.

- FindPointsGSLIB has a new constructor that accepts the mesh object and
-  internally calls the Setup() method so that the user does not have to.
-  The FreeData() method has also been moved to the destructor so the user does
-  not need to manually free-up the memory if the destructor is called before
-  MPI_Finalize().
-
 Version 4.8, released on Apr 9, 2025
 ====================================

@@ -278,11 +278,6 @@ if (MFEM_USE_OPENMP OR MFEM_USE_LEGACY_OPENMP)
  endif()
 endif()

-# Umpire (must be included before hypre, so hypre can use it if needed)
-if (MFEM_USE_UMPIRE)
-   find_package(UMPIRE REQUIRED)
-endif()
-
 # MPI -> hypre; PETSc (optional)
 if (MFEM_USE_MPI)
  find_package(MPI REQUIRED)
@@ -500,13 +495,14 @@ endif()

 # RAJA
 if (MFEM_USE_RAJA)
-   # RAJA uses FindCUDA, which needs CMP0146=OLD in CMake >= 3.27
-   if(CMAKE_VERSION VERSION_GREATER_EQUAL 3.27.0)
-      cmake_policy(SET CMP0146 OLD)
-   endif()
   find_package(RAJA REQUIRED)
 endif()

+# UMPIRE
+if (MFEM_USE_UMPIRE)
+   find_package(UMPIRE REQUIRED)
+endif()
+
 # GOOGLE-BENCHMARK
 if (MFEM_USE_BENCHMARK)
   find_package(Benchmark REQUIRED)
@@ -600,25 +596,18 @@ set(MFEM_TPLS OPENMP HYPRE LAPACK BLAS SuperLUDist STRUMPACK METIS SuiteSparse
    NETCDF MPFR PUMI HIOP POSIXCLOCKS MFEMBacktrace ZLIB OCCA CEED RAJA UMPIRE
    ADIOS2 MKL_CPARDISO MKL_PARDISO AMGX MAGMA CUSPARSE CUBLAS CALIPER CODIPACK
    BENCHMARK PARELAG TRIBOL MPI_CXX HIP HIPBLAS HIPSPARSE MOONOLITH BLITZ
-    ALGOIM ENZYME CUDA::cudart)
+    ALGOIM ENZYME)

-# Add all created targets and *_FOUND libraries in the variables TPL_TARGETS and
-# TPL_LIBRARIES, respectively.
-set(TPL_TARGETS)
+# Add all *_FOUND libraries in the variable TPL_LIBRARIES.
 set(TPL_LIBRARIES "")
 set(TPL_INCLUDE_DIRS "")
 foreach(TPL IN LISTS MFEM_TPLS)
-  if (${TPL}_FOUND OR TARGET ${TPL})
+  if (${TPL}_FOUND)
    message(STATUS "MFEM: using package ${TPL}")
-    if (TARGET ${TPL})
-      list(APPEND TPL_TARGETS ${TPL})
-    else()
-      list(APPEND TPL_LIBRARIES ${${TPL}_LIBRARIES})
-      list(APPEND TPL_INCLUDE_DIRS ${${TPL}_INCLUDE_DIRS})
-    endif()
+    list(APPEND TPL_LIBRARIES ${${TPL}_LIBRARIES})
+    list(APPEND TPL_INCLUDE_DIRS ${${TPL}_INCLUDE_DIRS})
  endif()
 endforeach(TPL)
-
 list(REVERSE TPL_LIBRARIES)
 list(REMOVE_DUPLICATES TPL_LIBRARIES)
 list(REVERSE TPL_LIBRARIES)
@@ -652,7 +641,7 @@ if (MFEM_USE_CUDA)
 endif()

 add_subdirectory(config)
-set(MAIN_HEADERS
+set(MASTER_HEADERS
  ${PROJECT_SOURCE_DIR}/mfem.hpp
  ${PROJECT_SOURCE_DIR}/mfem-performance.hpp)

@@ -689,12 +678,9 @@ set(MFEM_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
 set(MFEM_INSTALL_DIR ${CMAKE_INSTALL_PREFIX})

 # Declaring the library
-mfem_add_library(mfem ${SOURCES} ${HEADERS} ${MAIN_HEADERS})
+mfem_add_library(mfem ${SOURCES} ${HEADERS} ${MASTER_HEADERS})
 # message(STATUS "TPL_LIBRARIES = ${TPL_LIBRARIES}")
-target_link_libraries(mfem PUBLIC ${TPL_LIBRARIES} ${TPL_TARGETS})
-if (TPL_TARGETS)
-  add_dependencies(mfem ${TPL_TARGETS})
-endif()
+target_link_libraries(mfem PUBLIC ${TPL_LIBRARIES})
 if (MINGW)
  target_link_libraries(mfem PRIVATE ws2_32)
 endif()
@@ -888,12 +874,12 @@ install(TARGETS ${PROJECT_NAME}
  LIBRARY DESTINATION ${INSTALL_LIB_DIR}
  ARCHIVE DESTINATION ${INSTALL_LIB_DIR})

-# Install the main headers
+# Install the master headers
 foreach(Header mfem.hpp mfem-performance.hpp)
  install(FILES ${PROJECT_BINARY_DIR}/InstallHeaders/${Header}
    DESTINATION ${INSTALL_INCLUDE_DIR})
 endforeach()
-install(FILES ${MAIN_HEADERS} DESTINATION ${INSTALL_INCLUDE_DIR}/mfem)
+install(FILES ${MASTER_HEADERS} DESTINATION ${INSTALL_INCLUDE_DIR}/mfem)

 # Install the headers (except common miniapp which is installed from its subdir)
 install(DIRECTORY ${MFEM_SOURCE_DIRS}
@@ -3,10 +3,10 @@
 </p>

 <p align="center">
-<a href="https://github.com/mfem/mfem/blob/main/LICENSE"><img alt="License" src="https://img.shields.io/badge/License-BSD-brightgreen.svg"></a>
-<a href="https://github.com/mfem/mfem/actions?query=workflow%3Arepo-check+branch%3Amain"><img alt="Repo check" src="https://github.com/mfem/mfem/actions/workflows/repo-check.yml/badge.svg?branch=main"></a>
-<a href="https://github.com/mfem/mfem/actions?query=workflow%3Abuild-analysis+branch%3Amain"><img alt="Build Analysis" src="https://github.com/mfem/mfem/actions/workflows/mfem-analysis.yml/badge.svg?branch=main"></a>
-<a href="https://github.com/mfem/mfem/actions?query=workflow%3Abuilds-and-tests+branch%3Amain"><img alt="Builds and Tests" src="https://github.com/mfem/mfem/actions/workflows/builds-and-tests.yml/badge.svg?branch=main"></a>
+<a href="https://github.com/mfem/mfem/blob/master/LICENSE"><img alt="License" src="https://img.shields.io/badge/License-BSD-brightgreen.svg"></a>
+<a href="https://github.com/mfem/mfem/actions?query=workflow%3Arepo-check+branch%3Amaster"><img alt="Repo check" src="https://github.com/mfem/mfem/actions/workflows/repo-check.yml/badge.svg?branch=master"></a>
+<a href="https://github.com/mfem/mfem/actions?query=workflow%3Abuild-analysis+branch%3Amaster"><img alt="Build Analysis" src="https://github.com/mfem/mfem/actions/workflows/mfem-analysis.yml/badge.svg?branch=master"></a>
+<a href="https://github.com/mfem/mfem/actions?query=workflow%3Abuilds-and-tests+branch%3Amaster"><img alt="Builds and Tests" src="https://github.com/mfem/mfem/actions/workflows/builds-and-tests.yml/badge.svg?branch=master"></a>
 <a href="https://ci.appveyor.com/project/mfem/mfem"><img alt="Build Status" src="https://ci.appveyor.com/api/projects/status/19non9sqm6msi2wy?svg=true"></a>
 <a href="https://docs.mfem.org/html/index.html"><img alt="Doxygen" src="https://img.shields.io/badge/code-documented-brightgreen.svg"></a>
 </p>
@@ -27,7 +27,7 @@ in the MFEM community, you agree to abide by its rules.
 If you plan on contributing to MFEM, consider reviewing the
 [issue tracker](https://github.com/mfem/mfem/issues) first to check if a thread
 already exists for your desired feature or the bug you ran into. Use a pull
-request (PR) toward the `mfem:main` branch to propose your contribution. If
+request (PR) toward the `mfem:master` branch to propose your contribution. If
 you are planning significant code changes or have questions, you may want to
 open an [issue](https://github.com/mfem/mfem/issues) before issuing a PR. In
 addition to technical contributions, we are also interested in your results and
@@ -47,7 +47,7 @@ back to them before issuing pull requests:
  - [Pull Requests](#pull-requests)
  - [MFEM PR Rules](#mfem-pr-rules)
  - [Pull Request Checklist](#pull-request-checklist)
-  - [Main/Next Workflow](#mainnext-workflow)
+  - [Master/Next Workflow](#masternext-workflow)
  - [Releases](#releases)
  - [Release Checklist](#release-checklist)
 - [LLNL Workflow](#llnl-workflow)
@@ -66,12 +66,12 @@ Origin](#developers-certificate-of-origin-11) at the end of this file.*
 ## Quick Summary

 - We encourage you to [join the MFEM organization](#mfem-organization) and create
-  development branches off `mfem:main`.
+  development branches off `mfem:master`.
 - Please follow the [developer guidelines](#developer-guidelines), in particular
  with regards to documentation and code styling.
 - Please do not commit large/binary files to the central repository (use a fork
  instead).
- Pull requests should be issued toward `mfem:main`. Make sure
+- Pull requests should be issued toward `mfem:master`. Make sure
  to check the items off the [Pull Request Checklist](#pull-request-checklist) and
  follow the [MFEM PR Rules](#mfem-pr-rules).
 - When your contribution is fully working and ready to be reviewed, add
@@ -81,8 +81,8 @@ Origin](#developers-certificate-of-origin-11) at the end of this file.*
 - The reviewers have 3 weeks to evaluate the PR and work with the author to
  fix issues and implement improvements.
 - During review there should be no force pushes/rewriting history in the branch.
- After approval, MFEM developers merge the PR manually in the [mfem:next branch](#mainnext-workflow).
- After a week of testing in `mfem:next`, the original PR is merged in `mfem:main`.
+- After approval, MFEM developers merge the PR manually in the [mfem:next branch](#masternext-workflow).
+- After a week of testing in `mfem:next`, the original PR is merged in `mfem:master`.
 - We use [milestones](https://github.com/mfem/mfem/milestones) to coordinate the
  work on different PRs toward a release.
 - Don't hesitate to [contact us](#contact-information) if you have any questions.
@@ -289,7 +289,7 @@ Before you can start, you need a GitHub account, here are a few suggestions:
  author, is willing to work on it and be its champion.

 - The author creates a branch for the new feature (with suffix `-dev`), off
-  the `main` branch, or another existing feature branch, for example:
+  the `master` branch, or another existing feature branch, for example:

  ```
  # Clone assuming you have setup your ssh keys on GitHub:
@@ -298,8 +298,8 @@ Before you can start, you need a GitHub account, here are a few suggestions:
  # Alternatively, clone using the "https" protocol:
  git clone https://github.com/mfem/mfem.git

-  # Create a new feature branch starting from "main":
-  git checkout main
+  # Create a new feature branch starting from "master":
+  git checkout master
  git pull
  git checkout -b feature-dev

@@ -375,7 +375,7 @@ Before you can start, you need a GitHub account, here are a few suggestions:
 ### Pull Requests

 - When your branch is ready for other developers to review / comment on
-  the code, create a pull request towards `mfem:main`.
+  the code, create a pull request towards `mfem:master`.

 - Pull request typically have titles like:

@@ -411,7 +411,7 @@ Before you can start, you need a GitHub account, here are a few suggestions:
  assigned, the PR is considered under review. To help with the review process
  there should be no force pushes/rewriting history in the branch.

- After approval, the PR is [tested](#mainnext-workflow) for a week with
+- After approval, the PR is [tested](#masternext-workflow) for a week with
  other approved PRs in the `mfem:next` branch.

 - Consider manually running the tests in `tests/scripts` before merging in
@@ -447,7 +447,7 @@ The Pull Request (PR) approval process in MFEM is similar to the approval of pap

 3. A PR can be (manually) merged in the *next* branch only if 2 of the assigned reviewers have approved it and it has passed internal testing. This merge can be performed by any of the assigned reviewers or by any of the editors.

-4. A PR can be merged in the *main* branch only if it has been tested successfully for a week in *next* and an editor has (optionally) taken a final look. This merge can be performed only by one of the editors.
+4. A PR can be merged in the *master* branch only if it has been tested successfully for a week in *next* and an editor has (optionally) taken a final look. This merge can be performed only by one of the editors.

 #### Responsibilities of Editors

@@ -470,7 +470,7 @@ The current list of MFEM editors is:

 5. To remind the reviewers about timely completion of their review.

-6. To take a final look and complete the PR merge in *main*. The final look step is optional and shouldn't take more than 3 days.
+6. To take a final look and complete the PR merge in *master*. The final look step is optional and shouldn't take more than 3 days.

 7. The assignment of bugfixes should be expedited proportional to their importance, e.g. in some cases the editor can assign much shorter review window.

@@ -492,7 +492,7 @@ Everyone on the MFEM team can be asked to serve as a reviewer on a PR in their a

 5. To test the PR thoroughly before merging in *next*. The PR author is also encouraged to perform testing and inform the reviewers about the results.

-6. To monitor the PR impact on the testing in the *next* branch and alert the editors that the PR is ready for merging in *main*.
+6. To monitor the PR impact on the testing in the *next* branch and alert the editors that the PR is ready for merging in *master*.

 7. The review of bugfixes should be expedited proportional to their importance. The review window can be much less than three weeks in such cases.

@@ -582,11 +582,11 @@ Before a PR can be merged, it should satisfy the following:
   - [ ] Update internal tests to include the new features.


-### Main/Next Workflow
+### Master/Next Workflow

-MFEM uses a `main`/`next`-branch workflow as described below:
+MFEM uses a `master`/`next`-branch workflow as described below:

- The `main` branch should always be of release quality and changes should not
+- The `master` branch should always be of release quality and changes should not
  be merged until they have been fully tested. This branch is protected, and
  changes can only be made through pull requests.

@@ -613,20 +613,20 @@ MFEM uses a `main`/`next`-branch workflow as described below:

 - After a week of testing in `next` (excluding bugfixes), both on GitHub, as
  well as [internally](#tests-at-llnl) at LLNL, the original PR is merged into
-  `main` (provided there are no issues).
+  `master` (provided there are no issues).

 - After the merge, the feature branch is deleted (unless it is a long-term
  project with periodic PRs).

 - The `next` branch is used just for integrated testing of all PRs approved for
-  merging into `main` to verify that each works individually and that all of
+  merging into `master` to verify that each works individually and that all of
  them work as a group. This branch can be discarded at any time, though we
  typically do that only at the end of a [release cycle](#releases).


 ### Releases

- Releases are just tags in the `main` branch, e.g. https://github.com/mfem/mfem/releases/tag/v3.3.2,
+- Releases are just tags in the `master` branch, e.g. https://github.com/mfem/mfem/releases/tag/v3.3.2,
  and have a version that ends in an even "patch" number, e.g. `v3.2.2` or
  `v3.4` (by convention `v3.4` is the same as `v3.4.0`.)  Between releases, the
  version ends in an odd "patch" number, e.g. `v3.3.3`.
@@ -679,19 +679,19 @@ MFEM uses a `main`/`next`-branch workflow as described below:

 ### Mirroring on Bitbucket

- The GitHub `main` and `next` branches are mirrored to the LLNL institutional
-  Bitbucket repository as `gh-main` and `gh-next`.
+- The GitHub `master` and `next` branches are mirrored to the LLNL institutional
+  Bitbucket repository as `gh-master` and `gh-next`.

- `gh-main` is merged into LLNL's internal `main` through pull requests; write
-  permissions to `main` are restricted to ensure this is the only way in which it
+- `gh-master` is merged into LLNL's internal `master` through pull requests; write
+  permissions to `master` are restricted to ensure this is the only way in which it
  gets updated.

 - We never push directly from LLNL to GitHub.

 - Versions of the code on LLNL's internal server, from most to least stable:
  - MFEM official release on mfem.org -- Most stable, tested in many apps.
-  - `mfem:main` -- Recent development version, guaranteed to work.
-  - `mfem:gh-main` -- Stable development version, passed testing, you can use
+  - `mfem:master` -- Recent development version, guaranteed to work.
+  - `mfem:gh-master` -- Stable development version, passed testing, you can use
     it to build your code between releases.
  - `mfem:gh-next` -- Bleeding-edge development version, may be broken, use at
     your own risk.
@@ -721,14 +721,14 @@ directory.


 ### Linux and Mac smoke tests
-We use GitHub Actions to drive the default tests on the `main` and `next`
+We use GitHub Actions to drive the default tests on the `master` and `next`
 branches. See the `.github/workflows` files and the logs at
 [https://github.com/mfem/mfem/actions](https://github.com/mfem/mfem/actions).

 Testing using GitHub Actions should be kept lightweight, as there is a time
 constraint on jobs. Two virtual machines are configured - Mac (OS X) and Linux.

- Tests on the `main` branch are triggered whenever a PR is issued on this branch.
+- Tests on the `master` branch are triggered whenever a PR is issued on this branch.
 - Tests on the `next` branch are currently scheduled to run each night.


@@ -744,7 +744,7 @@ and debug build is performed with a simple run of `ex1` to verify the executable

 ### Tests at LLNL

- We mirror the `main` and `next` branches internally (to `gh-main` and
+- We mirror the `master` and `next` branches internally (to `gh-master` and
  `gh-next`) and run longer nightly tests via cron. On the weekends, a more
  extensive test is run which extracts and executes all the different sample
  runs from each example and most miniapps.
@@ -121,11 +121,6 @@ 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>
@@ -668,7 +663,6 @@ 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).
@@ -848,7 +842,6 @@ 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
@@ -859,7 +852,7 @@ The specific libraries and their options are:
  URL: https://github.com/CEED/libCEED
       https://ceed.exascaleproject.org/libceed
  Options: CEED_DIR, CEED_OPT, CEED_LIB.
-  Versions: libCEED >= 0.12.0.
+  Versions: libCEED >= 0.12.

 - RAJA (optional), used when MFEM_USE_RAJA = YES.
  Beginning with MFEM v4.5.1, only RAJA v2022.10.3+ is supported.
@@ -1081,9 +1074,6 @@ 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):
 ---------------------------
@@ -1145,12 +1135,6 @@ 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
 ==================================
@@ -84,31 +84,6 @@ set_and_check(MFEM_LIBRARY_DIR "@PACKAGE_LIB_INSTALL_DIR@")

 check_required_components(MFEM)

-include(CMakeFindDependencyMacro)
-
-if (MFEM_USE_CUDA)
-  # required for projects linking to MFEM+CUDA, even if they don't use CUDA directly
-  find_dependency(CUDAToolkit)
-endif (MFEM_USE_CUDA)
-
-if (MFEM_USE_HIP)
-  # hip/rocm uses the modern MFEM way of linking to targets, need to find dependencies
-  find_dependency(HIP)
-  find_dependency(HIPBLAS)
-  find_dependency(HIPSPARSE)
-  if (MFEM_USE_MPI)
-    # assume HYPRE uses HIP
-    # alternatively could check HYPRE_USING_HIP
-    find_dependency(rocsparse)
-    find_dependency(rocrand)
-    find_dependency(rocsolver)
-  endif (MFEM_USE_MPI)
-endif (MFEM_USE_HIP)
-
-if (MFEM_USE_RAJA)
-   find_dependency(RAJA)
-endif()
-
 if (NOT TARGET mfem)
  include(${CMAKE_CURRENT_LIST_DIR}/MFEMTargets.cmake)
 endif (NOT TARGET mfem)
@@ -9,25 +9,21 @@
 # terms of the BSD-3 license. We welcome feedback and contributions, see file
 # CONTRIBUTING.md for details.

-# Defines the following variables if fetching of TPLs is disabled (default):
+# Defines the following variables:
 #   - 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 OR TARGET HYPRE)
+if (HYPRE_FOUND)
  if (HYPRE_USING_CUDA)
    find_package(CUDAToolkit REQUIRED)
  endif()
  if (HYPRE_USING_HIP)
    find_package(rocsparse REQUIRED)
    find_package(rocrand REQUIRED)
-    find_package(rocsolver REQUIRED)
  endif()
  if (HYPRE_LIBRARIES AND HYPRE_INCLUDE_DIRS AND HYPRE_VERSION)
    find_package_handle_standard_args(HYPRE
@@ -37,95 +33,6 @@ if (HYPRE_FOUND OR TARGET HYPRE)
  endif()
 endif()

-if (HYPRE_FETCH OR FETCH_TPLS)
-  # Collect all HYPRE_ENABLE variables and pass them to hypre, assuming they are BOOL.
-  set(HYPRE_CMAKE_OPTIONS "")
-  get_cmake_property(all_vars VARIABLES)
-  foreach(var ${all_vars})
-    if(var MATCHES "^HYPRE_ENABLE")
-      list(APPEND HYPRE_CMAKE_OPTIONS "-D${var}:BOOL=${${var}}")
-    endif()
-  endforeach()
-
-  set(HYPRE_FETCH_VERSION 2.33.0)
-  set(HYPRE_FETCH_TAG "v${HYPRE_FETCH_VERSION}" CACHE STRING "Tag, branch, or commit for HYPRE")
-  add_library(HYPRE STATIC IMPORTED)
-  # set options and associated dependencies
-  list(APPEND HYPRE_CMAKE_OPTIONS -DCMAKE_BUILD_TYPE:STRING=${CMAKE_BUILD_TYPE})
-  if (MFEM_USE_CUDA)
-    list(APPEND HYPRE_CMAKE_OPTIONS -DHYPRE_ENABLE_CUDA:BOOL=ON -DCMAKE_CUDA_ARCHITECTURES:STRING=${CMAKE_CUDA_ARCHITECTURES})
-    find_package(CUDAToolkit REQUIRED)
-    target_link_libraries(HYPRE INTERFACE CUDA::cusparse CUDA::curand CUDA::cublas)
-  elseif (MFEM_USE_HIP)
-    list(APPEND HYPRE_CMAKE_OPTIONS -DHYPRE_ENABLE_HIP:BOOL=ON)
-    find_package(rocsparse REQUIRED)
-    find_package(rocrand REQUIRED)
-    target_link_libraries(HYPRE INTERFACE rocsparse rocrand)
-  endif()
-  if (MFEM_USE_CUDA OR MFEM_USE_HIP)
-    if (MFEM_USE_UMPIRE)
-      if (EXISTS ${umpire_DIR})
-        list(APPEND HYPRE_CMAKE_OPTIONS -DHYPRE_ENABLE_UMPIRE:BOOL=ON -Dumpire_DIR:PATH=${umpire_DIR})
-      else()
-        message(FATAL_ERROR "MFEM_USE_UMPIRE=ON, however umpire_DIR isn't visible to HYPRE")
-      endif()
-    else()
-      list(APPEND HYPRE_CMAKE_OPTIONS -DHYPRE_ENABLE_UMPIRE:BOOL=OFF)
-      message(WARNING
-"================================================================================
- Umpire is disabled while building HYPRE with GPU support.
- This is not recommended for performance reasons!
- Consider enabling Umpire with -DMFEM_USE_UMPIRE=ON and providing -DUMPIRE_DIR.
-================================================================================")
-    endif()
-  endif()
-  if (MFEM_USE_SINGLE)
-    list(APPEND HYPRE_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_TAG} to be built with ${HYPRE_CMAKE_OPTIONS}")
-  set(HYPRE_INSTALL ${CMAKE_BINARY_DIR}/fetch/hypre)
-  include(ExternalProject)
-  ExternalProject_Add(hypre
-    GIT_REPOSITORY https://github.com/hypre-space/hypre.git
-    GIT_TAG ${HYPRE_FETCH_TAG}
-    GIT_SHALLOW TRUE
-    GIT_PROGRESS TRUE
-    UPDATE_DISCONNECTED TRUE
-    SOURCE_SUBDIR src
-    PREFIX ${HYPRE_INSTALL}
-    BUILD_COMMAND ${CMAKE_COMMAND} --build . -- -j${CMAKE_BUILD_PARALLEL_LEVEL}
-    CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${HYPRE_INSTALL} -DCMAKE_INSTALL_LIBDIR:PATH=lib ${HYPRE_CMAKE_OPTIONS})
-  file(MAKE_DIRECTORY ${HYPRE_INSTALL}/include)
-  # set imported library target properties
-  add_dependencies(HYPRE hypre)
-  set_target_properties(HYPRE PROPERTIES
-    IMPORTED_LOCATION ${HYPRE_INSTALL}/lib/libHYPRE.a
-    INTERFACE_INCLUDE_DIRECTORIES ${HYPRE_INSTALL}/include)
-  # convert HYPRE version to integer
-  if (HYPRE_FETCH_TAG MATCHES "^v?([0-9]+)\\.([0-9]+)\\.([0-9]+)$")
-    # Exact release tag X.Y.Z
-    string(REGEX MATCHALL "[0-9]+" HYPRE_SPLIT_VERSION "${HYPRE_FETCH_TAG}")
-  elseif (HYPRE_FETCH_VERSION MATCHES "([0-9]+)\\.([0-9]+)(\\.([0-9]+))?")
-    string(REGEX MATCHALL "[0-9]+" HYPRE_SPLIT_VERSION "${HYPRE_FETCH_VERSION}")
-  else (NOT DEFINED HYPRE_VERSION)
-    message(FATAL_ERROR "Unable to find HYPRE release version. Please provide it via -DHYPRE_VERSION")
-  endif()
-  if (HYPRE_SPLIT_VERSION AND NOT DEFINED HYPRE_VERSION)
-    list(GET HYPRE_SPLIT_VERSION 0 HYPRE_MAJOR_VERSION)
-    list(GET HYPRE_SPLIT_VERSION 1 HYPRE_MINOR_VERSION)
-    if (HYPRE_SPLIT_VERSION GREATER 2)
-      list(GET HYPRE_SPLIT_VERSION 2 HYPRE_PATCH_VERSION)
-    else()
-      set(HYPRE_PATCH_VERSION 0)
-    endif()
-    math(EXPR HYPRE_VERSION "10000*${HYPRE_MAJOR_VERSION} + 100*${HYPRE_MINOR_VERSION} + ${HYPRE_PATCH_VERSION}")
-    set(HYPRE_VERSION ${HYPRE_VERSION} CACHE STRING "HYPRE version." FORCE)
-  endif()
-  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."
@@ -190,8 +97,7 @@ endif()
 if (HYPRE_FOUND AND HYPRE_USING_HIP)
  find_package(rocsparse REQUIRED)
  find_package(rocrand REQUIRED)
-  find_package(rocsolver REQUIRED)
-  list(APPEND HYPRE_LIBRARIES ${rocsparse_LIBRARIES} ${rocrand_LIBRARIES} roc::rocsolver roc::rocblas)
+  list(APPEND HYPRE_LIBRARIES ${rocsparse_LIBRARIES} ${rocrand_LIBRARIES})
  set(HYPRE_LIBRARIES ${HYPRE_LIBRARIES} CACHE STRING
      "HYPRE libraries + dependencies." FORCE)
  message(STATUS "Updated HYPRE_LIBRARIES: ${HYPRE_LIBRARIES}")
@@ -9,39 +9,10 @@
 # terms of the BSD-3 license. We welcome feedback and contributions, see file
 # CONTRIBUTING.md for details.

-# Defines the following variables if fetching of TPLs is disabled (default):
+# Defines the following variables:
 #   - 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
-    BUILD_COMMAND $(MAKE) COPTIONS=-Wno-incompatible-pointer-types
-    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"
@@ -718,7 +718,7 @@ function(mfem_get_target_options Target CompileOptsVar LinkOptsVar)
  get_target_property(IsImported ${tgt} IMPORTED)
  # message(STATUS "${tgt}[IMPORTED]: ${IsImported}")
  # Generally, the possible target types are: STATIC_LIBRARY, MODULE_LIBRARY,
-  # SHARED_LIBRARY, INTERFACE_LIBRARY, UNKNOWN_LIBRARY, EXECUTABLE.
+  # SHARED_LIBRARY, INTERFACE_LIBRARY, EXECUTABLE.
  get_target_property(type ${tgt} TYPE)
  # message(STATUS "${tgt}[TYPE]: ${type}")
  unset(ImportConfig)
@@ -766,7 +766,7 @@ function(mfem_get_target_options Target CompileOptsVar LinkOptsVar)
    else()
      message(STATUS " *** Warning: [${tgt}] LOCATION not defined!")
    endif()
-  elseif ("${type}" STREQUAL "SHARED_LIBRARY" OR "${type}" STREQUAL "UNKNOWN_LIBRARY")
+  elseif ("${type}" STREQUAL "SHARED_LIBRARY")
    get_target_property(Location ${tgt} LOCATION)
    if (Location)
      get_filename_component(Dir ${Location} DIRECTORY)
@@ -932,14 +932,12 @@ function(mfem_export_mk_files)
  endif()
  set(MFEM_BUILD_TAG "${CMAKE_SYSTEM}")
  set(MFEM_PREFIX "${CMAKE_INSTALL_PREFIX}")
-  # For the next 4 variables, these are the values for the build-tree version of
+  # For the next 4 variable, these are the values for the build-tree version of
  # 'config.mk'
  set(MFEM_INC_DIR "${PROJECT_BINARY_DIR}")
  set(MFEM_LIB_DIR "${PROJECT_BINARY_DIR}")
  set(MFEM_TEST_MK "${PROJECT_SOURCE_DIR}/config/test.mk")
  set(MFEM_CONFIG_EXTRA "MFEM_BUILD_DIR ?= ${PROJECT_BINARY_DIR}")
-  # TODO: CUDA/HIP support:
-  set(MFEM_XLINKER "${CMAKE_CXX_LINKER_WRAPPER_FLAG}")
  set(MFEM_MPIEXEC ${MPIEXEC})
  if (NOT MFEM_MPIEXEC)
    set(MFEM_MPIEXEC "mpirun")
@@ -23,14 +23,11 @@
 #include "_config.hpp"
 #endif

-#include <cstdint>
-#include <climits>
-
 namespace mfem
 {

 #if (defined(MFEM_USE_CUDA) && defined(__CUDACC__)) || \
-    (defined(MFEM_USE_HIP) && defined(__HIP__))
+    (defined(MFEM_USE_HIP) && defined(__HIPCC__))
 #define MFEM_HOST_DEVICE __host__ __device__
 #else
 #define MFEM_HOST_DEVICE
@@ -88,7 +88,6 @@ MFEM_BUILD_TAG = @MFEM_BUILD_TAG@
 MFEM_PREFIX    = @MFEM_PREFIX@
 MFEM_INC_DIR   = @MFEM_INC_DIR@
 MFEM_LIB_DIR   = @MFEM_LIB_DIR@
-MFEM_XLINKER   = @MFEM_XLINKER@

 # Location of test.mk
 MFEM_TEST_MK = @MFEM_TEST_MK@
@@ -89,12 +89,6 @@ 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++)
@@ -57,7 +57,7 @@ CUDA_DIR = $(or $(CUDA_HOME),$(patsubst %/,%,$(dir \
 CLANG_CUDA_FLAGS = -xcuda --cuda-path=$(CUDA_DIR) --cuda-gpu-arch=$(CUDA_ARCH)
 # flags for nvcc
 NVCC_FLAGS = -x=cu --expt-extended-lambda --expt-relaxed-constexpr \
- -arch=$(CUDA_ARCH) -isystem "$(CUDA_DIR)/include"
+ -arch=$(CUDA_ARCH)
 # Prefixes for passing flags to the host compiler and linker when using
 # CUDA_CXX=nvcc
 CUDA_XCOMPILER = -Xcompiler=
@@ -33,8 +33,8 @@ RUN mkdir -p /opt/mfem-env \
 RUN cd /opt/mfem-env && \
    . /opt/spack/share/spack/setup-env.sh && \
    spack env activate . && \
-    spack develop --path /code mfem@main+examples+miniapps && \
-    spack add mfem@main+examples+miniapps # && \
+    spack develop --path /code mfem@master+examples+miniapps && \
+    spack add mfem@master+examples+miniapps && \
    spack install

 # ensure mfem always on various paths
@@ -1,11 +1,11 @@
 spack:
-  specs: [mfem@main+examples+miniapps]
-  view:
+  specs: [mfem@master+examples+miniapps]
+  view: 
    mfem:
-      root: /opt/mfem-view
+      root: /opt/mfem-view 
      link_type: copy
  concretization: together
  develop:
    mfem:
      path: /code
-      spec: mfem@main+examples+miniapps
+      spec: mfem@master+examples+miniapps
@@ -115,7 +115,7 @@ fi

 # branch-history
 if [[ "${option}" == "--history" || "${option}" == "" ]]; then
-  git fetch origin main:main
+  git fetch origin master:master
  cd tests/scripts
  if ! ./runtest branch-history; then code=1; fi
  cd -
@@ -1,593 +0,0 @@
-MFEM mesh v1.0
-
-# Created by: Pointwise
-
-# MFEM Geometry Types:
-#
-# POINT       = 0
-# SEGMENT     = 1
-# TRIANGLE    = 2
-# SQUARE      = 3
-# TETRAHEDRON = 4
-# CUBE        = 5
-# PRISM       = 6
-
-dimension
-2
-
-elements
-160
-1 3 1 164 163 0
-1 3 164 165 162 163
-1 3 2 166 164 1
-1 3 166 132 165 164
-1 3 3 167 166 2
-1 3 167 131 132 166
-1 3 4 168 167 3
-1 3 168 130 131 167
-1 3 5 169 168 4
-1 3 169 129 130 168
-1 3 6 170 169 5
-1 3 170 128 129 169
-1 3 171 172 170 6
-1 3 172 127 128 170
-1 3 124 125 172 171
-1 3 125 126 127 172
-1 3 162 165 173 161
-1 3 165 132 133 173
-1 3 161 173 174 160
-1 3 173 133 134 174
-1 3 160 174 175 159
-1 3 174 134 135 175
-1 3 6 7 176 171
-1 3 7 8 177 176
-1 3 171 176 123 124
-1 3 176 177 122 123
-1 3 159 175 178 158
-1 3 175 135 136 178
-1 3 158 178 179 157
-1 3 178 136 137 179
-1 3 157 179 180 156
-1 3 179 137 138 180
-1 3 122 177 181 121
-1 3 177 8 182 181
-1 3 8 9 183 182
-1 3 9 10 184 183
-1 3 10 11 185 184
-1 3 11 12 186 185
-1 3 12 13 187 186
-1 3 13 14 15 187
-1 3 121 181 119 120
-1 3 181 182 118 119
-1 3 182 183 117 118
-1 3 183 184 188 117
-1 3 184 185 109 188
-1 3 185 186 108 109
-1 3 186 187 189 108
-1 3 187 15 16 189
-1 3 109 110 190 188
-1 3 110 111 191 190
-1 3 111 112 113 191
-1 3 188 190 116 117
-1 3 190 191 115 116
-1 3 191 113 114 115
-1 3 189 192 107 108
-1 3 192 193 106 107
-1 3 193 194 105 106
-1 3 194 195 104 105
-1 3 195 196 103 104
-1 3 16 17 192 189
-1 3 17 18 193 192
-1 3 18 19 194 193
-1 3 19 20 195 194
-1 3 20 21 196 195
-1 3 97 98 197 96
-1 3 98 99 198 197
-1 3 99 100 199 198
-1 3 100 101 200 199
-1 3 101 102 201 200
-1 3 102 103 202 201
-1 3 103 196 203 202
-1 3 196 21 22 203
-1 3 96 197 204 95
-1 3 197 198 39 204
-1 3 198 199 38 39
-1 3 199 200 205 38
-1 3 200 201 32 205
-1 3 201 202 31 32
-1 3 202 203 206 31
-1 3 203 22 23 206
-1 3 32 33 207 205
-1 3 33 34 35 207
-1 3 205 207 37 38
-1 3 207 35 36 37
-1 3 39 40 208 204
-1 3 40 41 209 208
-1 3 41 42 210 209
-1 3 42 43 211 210
-1 3 43 44 212 211
-1 3 204 208 94 95
-1 3 208 209 93 94
-1 3 209 210 92 93
-1 3 210 211 91 92
-1 3 211 212 90 91
-1 3 90 212 213 89
-1 3 212 44 214 213
-1 3 44 45 215 214
-1 3 45 46 216 215
-1 3 46 47 217 216
-1 3 47 48 218 217
-1 3 48 49 219 218
-1 3 49 50 51 219
-1 3 89 213 87 88
-1 3 213 214 86 87
-1 3 214 215 85 86
-1 3 215 216 84 85
-1 3 216 217 83 84
-1 3 217 218 82 83
-1 3 218 219 220 82
-1 3 219 51 52 220
-1 3 53 221 220 52
-1 3 221 81 82 220
-1 3 54 222 221 53
-1 3 222 80 81 221
-1 3 55 223 222 54
-1 3 223 79 80 222
-1 3 26 27 224 25
-1 3 27 28 29 224
-1 3 25 224 225 24
-1 3 224 29 30 225
-1 3 24 225 206 23
-1 3 225 30 31 206
-1 3 154 155 226 153
-1 3 155 156 180 226
-1 3 153 226 227 152
-1 3 226 180 138 227
-1 3 152 227 228 151
-1 3 227 138 139 228
-1 3 151 228 229 150
-1 3 228 139 140 229
-1 3 150 229 230 149
-1 3 229 140 141 230
-1 3 149 230 231 148
-1 3 230 141 142 231
-1 3 148 231 232 147
-1 3 231 142 143 232
-1 3 147 232 145 146
-1 3 232 143 144 145
-1 3 56 233 223 55
-1 3 233 78 79 223
-1 3 57 234 233 56
-1 3 234 77 78 233
-1 3 58 235 234 57
-1 3 235 76 77 234
-1 3 61 236 59 60
-1 3 236 235 58 59
-1 3 62 237 236 61
-1 3 237 76 235 236
-1 3 63 238 237 62
-1 3 238 75 76 237
-1 3 64 239 238 63
-1 3 239 74 75 238
-1 3 65 240 239 64
-1 3 240 73 74 239
-1 3 66 241 240 65
-1 3 241 72 73 240
-1 3 67 242 241 66
-1 3 242 71 72 241
-1 3 68 69 242 67
-1 3 69 70 71 242
-
-boundary
-164
-3 1 0 1
-3 1 1 2
-3 1 2 3
-3 1 3 4
-3 1 4 5
-3 1 5 6
-3 1 6 7
-3 1 7 8
-3 1 8 9
-3 1 9 10
-3 1 10 11
-3 1 11 12
-3 1 12 13
-3 1 13 14
-3 1 16 17
-3 1 17 18
-3 1 18 19
-3 1 19 20
-3 1 20 21
-3 1 21 22
-3 1 22 23
-3 1 23 24
-3 1 24 25
-3 1 25 26
-3 1 26 27
-3 1 27 28
-3 1 28 29
-3 1 29 30
-3 1 30 31
-3 1 31 32
-3 1 32 33
-3 1 33 34
-3 1 34 35
-3 1 35 36
-3 1 36 37
-3 1 37 38
-3 1 38 39
-3 1 39 40
-3 1 40 41
-3 1 41 42
-3 1 42 43
-3 1 43 44
-3 1 49 50
-3 1 48 49
-3 1 47 48
-3 1 46 47
-3 1 45 46
-3 1 44 45
-3 1 52 53
-3 1 53 54
-3 1 54 55
-3 1 57 58
-3 1 56 57
-3 1 55 56
-3 1 60 61
-3 1 61 62
-3 1 62 63
-3 1 63 64
-3 1 64 65
-3 1 65 66
-3 1 66 67
-3 1 67 68
-3 1 75 76
-3 1 74 75
-3 1 73 74
-3 1 72 73
-3 1 71 72
-3 1 70 71
-3 1 76 77
-3 1 77 78
-3 1 78 79
-3 1 81 82
-3 1 80 81
-3 1 79 80
-3 1 82 83
-3 1 83 84
-3 1 84 85
-3 1 85 86
-3 1 86 87
-3 1 87 88
-3 1 94 95
-3 1 93 94
-3 1 92 93
-3 1 91 92
-3 1 90 91
-3 1 96 97
-3 1 95 96
-3 1 97 98
-3 1 98 99
-3 1 99 100
-3 1 100 101
-3 1 101 102
-3 1 102 103
-3 1 107 108
-3 1 106 107
-3 1 105 106
-3 1 104 105
-3 1 103 104
-3 1 108 109
-3 1 109 110
-3 1 110 111
-3 1 111 112
-3 1 112 113
-3 1 113 114
-3 1 114 115
-3 1 115 116
-3 1 116 117
-3 1 119 120
-3 1 118 119
-3 1 117 118
-3 1 131 132
-3 1 130 131
-3 1 129 130
-3 1 128 129
-3 1 127 128
-3 1 126 127
-3 1 132 133
-3 1 133 134
-3 1 134 135
-3 1 137 138
-3 1 136 137
-3 1 135 136
-3 1 138 139
-3 1 139 140
-3 1 140 141
-3 1 141 142
-3 1 142 143
-3 1 143 144
-3 1 147 148
-3 1 146 147
-3 1 153 154
-3 1 152 153
-3 1 151 152
-3 1 150 151
-3 1 149 150
-3 1 148 149
-3 1 156 157
-3 1 157 158
-3 1 158 159
-3 1 161 162
-3 1 160 161
-3 1 159 160
-2 1 69 70
-2 1 68 69
-3 1 88 89
-3 1 89 90
-3 1 121 122
-3 1 120 121
-3 1 123 124
-3 1 122 123
-3 1 125 126
-3 1 124 125
-1 1 144 145
-1 1 145 146
-3 1 15 16
-3 1 14 15
-3 1 50 51
-3 1 51 52
-3 1 59 60
-3 1 58 59
-3 1 154 155
-3 1 155 156
-3 1 163 0
-3 1 162 163
-
-vertices
-243
-2
-4 4
-4 3.5
-4 3
-4 2.5
-4 2
-4 1.5
-4 1
-4.5 1
-5 1
-5 1.5
-5 2
-5 2.5
-5 3
-5 3.5
-5 4
-5.500 4
-6 4
-6.500 4
-7 4
-7.5 4
-8 4
-8.5 4
-9 4
-9.5 4
-10 4
-10.5 4
-11 4
-11 3.5
-11 3
-10.5 3
-10 3
-9.5 3
-9.5 2.5
-10 2.5
-10.5 2.5
-10.5 2
-10.5 1.5
-10 1.5
-9.5 1.5
-9.5 1
-10 1
-10.5 1
-11 1
-11.5 1
-12 1
-12 1.5
-12 2
-12 2.5
-12 3
-12 3.5
-12 4
-12.5 4
-13 4
-13.333 3.75
-13.666 3.5
-14.000 3.25
-14.333 3.5
-14.666 3.75
-15.000 4
-15.500 4
-16.000 4
-16.000 3.5
-16.000 3
-16.000 2.5
-16.000 2
-16.000 1.5
-16.000 1
-16.000 0.5
-16.000 0
-15.500 0
-15.000 0
-15.000 0.5000000000000002
-15.000 1
-15.000 1.5
-15.000 2
-15.000 2.5
-15.000 3
-14.666 2.75
-14.333 2.5
-14.000 2.25
-13.666 2.5
-13.333 2.75
-13 3
-13 2.5
-13 2
-13 1.5
-13 1
-13 0.500
-13 0
-12.5 0
-12 0
-11.5 0
-11 0
-10.5 0
-10 0
-9.5 0
-9 0
-8.5 0
-8.5 0.5
-8.5 1
-8.5 1.5
-8.5 2
-8.5 2.5
-8.5 3
-8 3
-7.5 3
-7 3
-6.500 3
-6 3
-6 2.5
-6.5 2.5
-7 2.5
-7.5 2.5
-7.5 2
-7.5 1.5
-7.000 1.5
-6.5 1.5
-6 1.5
-6 1
-6 0.5
-6 0
-5.5 0
-5 0
-4.5 0
-4 0
-3.5 0
-3 0
-3 0.500
-3 1
-3 1.5
-3 2
-3 2.5
-3 3
-2.666 2.75
-2.333 2.5
-2.000 2.25
-1.666 2.5
-1.333 2.75
-1.000 3
-1.000 2.5
-1.000 2
-1.000 1.5
-1.000 1
-1.000 0.5000
-1.000 0
-0.5000 0
-0.0000 0
-0.0000 0.5
-0.0000 1
-0.0000 1.5
-0.0000 2
-0.0000 2.5
-0.0000 3
-0.0000 3.5
-0.0000 4
-0.5000 4
-1.000 4
-1.333 3.75
-1.666 3.5
-2.000 3.25
-2.333 3.5
-2.666 3.75
-3 4
-3.5 4
-3.5 3.5
-3 3.5
-3.5 3
-3.5 2.5
-3.5 2
-3.5 1.5
-3.5 1
-4 0.5
-3.5 0.5
-2.666 3.25
-2.333 3
-2.000 2.75
-4.5 0.5
-5 0.5
-1.666 3
-1.333 3.25
-1.000 3.5
-5.5 0.5
-5.500 1
-5.500 1.5
-5.500 2
-5.500 2.5
-5.500 3
-5.500 3.5
-6 2
-6 3.5
-6.5 2
-7 2
-6.5 3.5
-7 3.5
-7.5 3.5
-8 3.5
-8.5 3.5
-9 0.5
-9 1
-9 1.5
-9 2
-9 2.5
-9 3
-9 3.5
-9.5 0.5
-9.5 2
-9.5 3.5
-10 2
-10 0.5
-10.5 0.5
-11 0.5
-11.5 0.5
-12 0.5
-12.5 0.500
-12.5 1
-12.5 1.5
-12.5 2
-12.5 2.5
-12.5 3
-12.5 3.5
-13 3.5
-13.333 3.250
-13.666 3
-14.000 2.75
-10.5 3.5
-10 3.5
-0.500 3.5
-0.500 3
-0.500 2.5
-0.500 2
-0.500 1.5
-0.500 1
-0.500 0.5
-14.333 3
-14.666 3.25
-15.000 3.5
-15.500 3.5
-15.500 3
-15.500 2.5
-15.500 2
-15.500 1.5
-15.500 1
-15.500 0.5
@@ -1,342 +0,0 @@
-MFEM NURBS NC-patch mesh v1.0
-dimension
-3
-
-elements
-13
-0 1 5 0 8 10 11 9 4 6 7 5
-0 1 5 0 18 8 24 32 30 23 36 38
-0 1 5 0 0 18 32 14 12 30 38 29
-0 1 5 0 32 24 10 20 38 36 26 35
-0 1 5 0 14 32 20 2 29 38 35 16
-0 1 5 0 30 23 36 38 31 22 37 39
-0 1 5 0 12 30 38 29 13 31 39 28
-0 1 5 0 38 36 26 35 39 37 27 34
-0 1 5 0 29 38 35 16 28 39 34 17
-0 1 5 0 31 22 37 39 19 9 25 33
-0 1 5 0 13 31 39 28 1 19 33 15
-0 1 5 0 39 37 27 34 33 25 11 21
-0 1 5 0 28 39 34 17 15 33 21 3
-
-boundary
-31
-9999 3 8 10 6 4
-9999 3 10 11 7 6
-9999 3 11 9 5 7
-9999 3 9 8 4 5
-9999 3 4 6 7 5
-9999 3 32 24 8 18
-9999 3 18 8 23 30
-9999 3 14 32 18 0
-9999 3 0 18 30 12
-9999 3 14 0 12 29
-9999 3 20 10 24 32
-9999 3 10 20 35 26
-9999 3 2 20 32 14
-9999 3 20 2 16 35
-9999 3 2 14 29 16
-9999 3 30 23 22 31
-9999 3 12 30 31 13
-9999 3 29 12 13 28
-9999 3 26 35 34 27
-9999 3 35 16 17 34
-9999 3 16 29 28 17
-9999 3 31 22 9 19
-9999 3 19 9 25 33
-9999 3 13 31 19 1
-9999 3 28 13 1 15
-9999 3 1 19 33 15
-9999 3 27 34 21 11
-9999 3 33 25 11 21
-9999 3 34 17 3 21
-9999 3 17 28 15 3
-9999 3 15 33 21 3
-
-vertex_to_knotspan
-8
-23 0 1 8 10 11 9
-22 0 2 8 10 11 9
-24 1 0 8 10 11 9
-36 1 1 8 10 11 9
-37 1 2 8 10 11 9
-25 1 3 8 10 11 9
-26 2 1 8 10 11 9
-27 2 2 8 10 11 9
-
-coordinates
-40
-3
-0 0 0
-0 1 0
-4 0 0
-4 1 0
-0 0 4
-0 1 4
-4 0 4
-4 1 4
-0 0 2
-0 1 2
-4 0 2
-4 1 2
-0 0.333333333333333 0
-0 0.666666666666667 0
-2 0 0
-2 1 0
-4 0.333333333333334 0
-4 0.666666666666667 0
-0 0 1
-0 1 1
-4 0 1
-4 1 1
-0 0.666666666666667 2
-0 0.333333333333333 2
-2 0 2
-2 1 2
-4 0.333333333333333 2
-4 0.666666666666667 2
-2 0.666666666666667 0
-2 0.333333333333333 0
-0 0.333333333333333 1
-0 0.666666666666667 1
-1.81325211007895 0 1
-1.81325211007895 1 1
-4 0.666666666666667 1
-4 0.333333333333333 1
-2 0.333333333333333 2
-2 0.666666666666667 2
-1.81325211007895 0.333333333333333 1
-1.81325211007895 0.666666666666667 1
-
-edges
-87
-0 8 10
-1 10 11
-0 9 11
-1 8 9
-0 4 6
-1 6 7
-0 5 7
-1 4 5
-2 4 8
-2 6 10
-2 7 11
-2 5 9
-9 18 8
-7 8 24
-9 32 24
-7 18 32
-9 30 23
-7 23 36
-9 38 36
-7 30 38
-3 18 30
-3 8 23
-3 24 36
-3 32 38
-8 0 18
-8 14 32
-7 0 14
-8 12 30
-8 29 38
-7 12 29
-3 0 12
-3 14 29
-6 24 10
-9 20 10
-6 32 20
-6 36 26
-9 35 26
-6 38 35
-3 10 26
-3 20 35
-8 2 20
-6 14 2
-8 16 35
-6 29 16
-3 2 16
-9 31 22
-7 22 37
-9 39 37
-7 31 39
-4 30 31
-4 23 22
-4 36 37
-4 38 39
-8 13 31
-8 28 39
-7 13 28
-4 12 13
-4 29 28
-6 37 27
-9 34 27
-6 39 34
-4 26 27
-4 35 34
-8 17 34
-6 28 17
-4 16 17
-9 19 9
-7 9 25
-9 33 25
-7 19 33
-5 31 19
-5 22 9
-5 37 25
-5 39 33
-8 1 19
-8 15 33
-7 1 15
-5 13 1
-5 28 15
-6 25 11
-9 21 11
-6 33 21
-5 27 11
-5 34 21
-8 3 21
-6 15 3
-5 17 3
-
-knotvectors
-10
-1 3 0 0 0.5 1 1
-1 4 0 0 0.333333333333333 0.666666666666667 1 1
-1 3 0 0 0.5 1 1
-1 2 0 0 1 1
-1 2 0 0 1 1
-1 2 0 0 1 1
-1 2 0 0 1 1
-1 2 0 0 1 1
-1 2 0 0 1 1
-1 2 0 0 1 1
-
-spacing
-0
-
-weights
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-
-FiniteElementSpace
-FiniteElementCollection: NURBS1
-VDim: 3
-Ordering: 1
-
-0 0 0
-0 1 0
-4 0 0
-4 1 0
-0 0 4
-0 1 4
-4 0 4
-4 1 4
-0 0 2
-0 1 2
-4 0 2
-4 1 2
-0 0.333333333333333 0
-0 0.666666666666667 0
-2 0 0
-2 1 0
-4 0.333333333333334 0
-4 0.666666666666667 0
-0 0 1
-0 1 1
-4 0 1
-4 1 1
-0 0.666666666666667 2
-0 0.333333333333333 2
-2 0 2
-2 1 2
-4 0.333333333333333 2
-4 0.666666666666667 2
-2 0.666666666666667 0
-2 0.333333333333333 0
-0 0.333333333333333 1
-0 0.666666666666667 1
-1.81325211007895 0 1
-1.81325211007895 1 1
-4 0.666666666666667 1
-4 0.333333333333333 1
-2 0.333333333333333 2
-2 0.666666666666667 2
-1.81325211007895 0.333333333333333 1
-1.81325211007895 0.666666666666667 1
-2 0 4
-4 0.333333333333333 4
-4 0.666666666666667 4
-2 1 4
-0 0.333333333333333 4
-0 0.666666666666667 4
-0 0 3
-4 0 3
-4 1 3
-0 1 3
-2 0 3
-4 0.333333333333333 3
-4 0.666666666666667 3
-2 1 3
-0 0.666666666666667 3
-0 0.333333333333333 3
-2 0.333333333333333 4
-2 0.666666666666667 4
-2 0.333333333333333 3
-2 0.666666666666667 3
@@ -1,96 +0,0 @@
-MFEM NURBS NC-patch mesh v1.0
-dimension
-2
-
-# rank attr geom ref_type nodes/children
-elements
-3
-0 1 3 0 0 4 5 1
-0 1 3 0 6 7 4 2
-0 1 3 0 6 3 5 7
-
-# attr geom nodes
-boundary
-7
-1 1 0 4
-1 1 5 1
-1 1 1 0
-1 1 2 6
-1 1 6 3
-1 1 4 2
-1 1 5 3
-
-vertex_to_knotspan
-1
-7 1 4 5
-
-# top-level node coordinates
-coordinates
-8
-2
-0 0
-0 1
-2 0
-2 1
-1 0
-1 1
-2 0.5
-1 0.5
-
-edges
-11
-0 0 4
-1 4 5
-0 1 5
-1 0 1
-2 6 7
-4 7 4
-2 2 4
-4 6 2
-3 6 3
-2 3 5
-3 7 5
-
-knotvectors
-5
-1 3 0 0 0.5 1 1
-1 3 0 0 0.5 1 1
-1 2 0 0 1 1
-1 2 0 0 1 1
-1 2 0 0 1 1
-
-spacing
-0
-
-weights
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-1.0
-
-FiniteElementSpace
-FiniteElementCollection: NURBS1
-VDim: 2
-Ordering: 1
-
-0 0
-0 1
-2 0
-2 1
-1 0
-1 1
-2 0.5
-1 0.5
-0.5 0
-0.5 1
-0 0.5
-0.5 0.5
-mfem_mesh_end
@@ -202,7 +202,6 @@ namespace mfem {
 * - <a class="el" href="tesla_8cpp_source.html">Tesla</a>: simple magnetostatics simulation code
 * - <a class="el" href="maxwell_8cpp_source.html">Maxwell</a>: simple transient full-wave electromagnetics simulation code
 * - <a class="el" href="joule_8cpp_source.html">Joule</a>: transient magnetics and Joule heating miniapp
- * - <a class="el" href="lorentz_8cpp_source.html">Lorentz</a>: simple particle tracking code based on the Lorentz force
 * - <a class="el" href="classmfem_1_1navier_1_1NavierSolver.html">Navier</a>: solve the transient incompressible Navier-Stokes equations
 * - <a class="el" href="mobius-strip_8cpp_source.html">Mobius Strip</a>: generate various Mobius strip-like meshes
 * - <a class="el" href="klein-bottle_8cpp_source.html">Klein Bottle</a>: generate three types of Klein bottle surfaces
@@ -205,15 +205,6 @@ if (MFEM_ENABLE_TESTING)
      $<TARGET_FILE:ex25p> "-no-vis" "--mumps-solver"
      ${MPIEXEC_POSTFLAGS})
  endif()
-
-  # Parallel libCEED example
-  if (MFEM_USE_CEED AND MFEM_USE_MPI)
-    add_test(NAME ex1p_ceed_np=${MFEM_MPI_NP}
-      COMMAND ${MPIEXEC} ${MPIEXEC_NUMPROC_FLAG} ${MFEM_MPI_NP}
-      ${MPIEXEC_PREFLAGS}
-      $<TARGET_FILE:ex1p> "-no-vis" "-d ceed-cpu" "-pa" "-a"
-      ${MPIEXEC_POSTFLAGS})
-  endif()
 endif()

 # Include the examples/amgx directory if AmgX is enabled
@@ -27,7 +27,6 @@
 //               ex1 -m ../data/fichera-amr.mesh
 //               ex1 -m ../data/mobius-strip.mesh
 //               ex1 -m ../data/mobius-strip.mesh -o -1 -sc
-//               ex1 -m ../data/nc3-nurbs.mesh -o -1
 //
 // Device sample runs:
 //               ex1 -pa -d cuda
@@ -62,14 +62,9 @@ static real_t epsilon_ = 1.0;
 static real_t sigma_ = 20.0;
 static real_t omega_ = 10.0;

-real_t u0_real_exact(const Vector &);
-real_t u0_imag_exact(const Vector &);
-
-void u1_real_exact(const Vector &, Vector &);
-void u1_imag_exact(const Vector &, Vector &);
-
-void u2_real_exact(const Vector &, Vector &);
-void u2_imag_exact(const Vector &, Vector &);
+complex<real_t> u0_exact(const Vector &x);
+void u1_exact(const Vector &, ComplexVector &);
+void u2_exact(const Vector &, ComplexVector &);

 bool check_for_inline_mesh(const char * mesh_file);

@@ -215,54 +210,48 @@ int main(int argc, char *argv[])
   ComplexGridFunction * u_exact = NULL;
   if (exact_sol) { u_exact = new ComplexGridFunction(fespace); }

-   FunctionCoefficient u0_r(u0_real_exact);
-   FunctionCoefficient u0_i(u0_imag_exact);
-   VectorFunctionCoefficient u1_r(dim, u1_real_exact);
-   VectorFunctionCoefficient u1_i(dim, u1_imag_exact);
-   VectorFunctionCoefficient u2_r(dim, u2_real_exact);
-   VectorFunctionCoefficient u2_i(dim, u2_imag_exact);
+   ComplexFunctionCoefficient u0(u0_exact);
+   ComplexVectorFunctionCoefficient u1(dim, u1_exact);
+   ComplexVectorFunctionCoefficient u2(dim, u2_exact);

-   ConstantCoefficient zeroCoef(0.0);
-   ConstantCoefficient oneCoef(1.0);
+   ComplexConstantCoefficient oneCoef(1.0);

-   Vector zeroVec(dim); zeroVec = 0.0;
   Vector  oneVec(dim);  oneVec = 0.0; oneVec[(prob==2)?(dim-1):0] = 1.0;
-   VectorConstantCoefficient zeroVecCoef(zeroVec);
-   VectorConstantCoefficient oneVecCoef(oneVec);
+   ComplexVectorConstantCoefficient oneVecCoef(oneVec);

   switch (prob)
   {
      case 0:
         if (exact_sol)
         {
-            u.ProjectBdrCoefficient(u0_r, u0_i, ess_bdr);
-            u_exact->ProjectCoefficient(u0_r, u0_i);
+            u.ProjectBdrCoefficient(u0, ess_bdr);
+            u_exact->ProjectCoefficient(u0);
         }
         else
         {
-            u.ProjectBdrCoefficient(oneCoef, zeroCoef, ess_bdr);
+            u.ProjectBdrCoefficient(oneCoef, ess_bdr);
         }
         break;
      case 1:
         if (exact_sol)
         {
-            u.ProjectBdrCoefficientTangent(u1_r, u1_i, ess_bdr);
-            u_exact->ProjectCoefficient(u1_r, u1_i);
+            u.ProjectBdrCoefficientTangent(u1, ess_bdr);
+            u_exact->ProjectCoefficient(u1);
         }
         else
         {
-            u.ProjectBdrCoefficientTangent(oneVecCoef, zeroVecCoef, ess_bdr);
+            u.ProjectBdrCoefficientTangent(oneVecCoef, ess_bdr);
         }
         break;
      case 2:
         if (exact_sol)
         {
-            u.ProjectBdrCoefficientNormal(u2_r, u2_i, ess_bdr);
-            u_exact->ProjectCoefficient(u2_r, u2_i);
+            u.ProjectBdrCoefficientNormal(u2, ess_bdr);
+            u_exact->ProjectCoefficient(u2);
         }
         else
         {
-            u.ProjectBdrCoefficientNormal(oneVecCoef, zeroVecCoef, ess_bdr);
+            u.ProjectBdrCoefficientNormal(oneVecCoef, ess_bdr);
         }
         break;
      default: break; // This should be unreachable
@@ -300,27 +289,24 @@ int main(int argc, char *argv[])
   ConstantCoefficient lossCoef(omega_ * sigma_);
   ConstantCoefficient negMassCoef(omega_ * omega_ * epsilon_);

+   ComplexConstantCoefficient complexMassCoef(-omega_ * omega_ * epsilon_,
+                                              omega_ * sigma_);
+
   SesquilinearForm *a = new SesquilinearForm(fespace, conv);
   if (pa) { a->SetAssemblyLevel(AssemblyLevel::PARTIAL); }
   switch (prob)
   {
      case 0:
-         a->AddDomainIntegrator(new DiffusionIntegrator(stiffnessCoef),
-                                NULL);
-         a->AddDomainIntegrator(new MassIntegrator(massCoef),
-                                new MassIntegrator(lossCoef));
+         a->AddDomainIntegrator<DiffusionIntegrator>(stiffnessCoef);
+         a->AddDomainIntegrator<MassIntegrator>(complexMassCoef);
         break;
      case 1:
-         a->AddDomainIntegrator(new CurlCurlIntegrator(stiffnessCoef),
-                                NULL);
-         a->AddDomainIntegrator(new VectorFEMassIntegrator(massCoef),
-                                new VectorFEMassIntegrator(lossCoef));
+         a->AddDomainIntegrator<CurlCurlIntegrator>(stiffnessCoef);
+         a->AddDomainIntegrator<VectorFEMassIntegrator>(complexMassCoef);
         break;
      case 2:
-         a->AddDomainIntegrator(new DivDivIntegrator(stiffnessCoef),
-                                NULL);
-         a->AddDomainIntegrator(new VectorFEMassIntegrator(massCoef),
-                                new VectorFEMassIntegrator(lossCoef));
+         a->AddDomainIntegrator<DivDivIntegrator>(stiffnessCoef);
+         a->AddDomainIntegrator<VectorFEMassIntegrator>(complexMassCoef);
         break;
      default: break; // This should be unreachable
   }
@@ -436,29 +422,24 @@ int main(int argc, char *argv[])

   if (exact_sol)
   {
-      real_t err_r = -1.0;
-      real_t err_i = -1.0;
+      real_t err_u = -1.0;

      switch (prob)
      {
         case 0:
-            err_r = u.real().ComputeL2Error(u0_r);
-            err_i = u.imag().ComputeL2Error(u0_i);
+            err_u = u.ComputeL2Error(u0);
            break;
         case 1:
-            err_r = u.real().ComputeL2Error(u1_r);
-            err_i = u.imag().ComputeL2Error(u1_i);
+            err_u = u.ComputeL2Error(u1);
            break;
         case 2:
-            err_r = u.real().ComputeL2Error(u2_r);
-            err_i = u.imag().ComputeL2Error(u2_i);
+            err_u = u.ComputeL2Error(u2);
            break;
         default: break; // This should be unreachable
      }

      cout << endl;
-      cout << "|| Re (u_h - u) ||_{L^2} = " << err_r << endl;
-      cout << "|| Im (u_h - u) ||_{L^2} = " << err_i << endl;
+      cout << "|| u_h - u ||_{L^2} = " << err_u << endl;
      cout << endl;
   }

@@ -564,36 +545,14 @@ complex<real_t> u0_exact(const Vector &x)
   return std::exp(-i * kappa * x[dim - 1]);
 }

-real_t u0_real_exact(const Vector &x)
-{
-   return u0_exact(x).real();
-}
-
-real_t u0_imag_exact(const Vector &x)
-{
-   return u0_exact(x).imag();
-}
-
-void u1_real_exact(const Vector &x, Vector &v)
+void u1_exact(const Vector &x, ComplexVector &v)
 {
   int dim = x.Size();
-   v.SetSize(dim); v = 0.0; v[0] = u0_real_exact(x);
+   v.SetSize(dim); v = 0.0; v[0] = u0_exact(x);
 }

-void u1_imag_exact(const Vector &x, Vector &v)
+void u2_exact(const Vector &x, ComplexVector &v)
 {
   int dim = x.Size();
-   v.SetSize(dim); v = 0.0; v[0] = u0_imag_exact(x);
-}
-
-void u2_real_exact(const Vector &x, Vector &v)
-{
-   int dim = x.Size();
-   v.SetSize(dim); v = 0.0; v[dim-1] = u0_real_exact(x);
-}
-
-void u2_imag_exact(const Vector &x, Vector &v)
-{
-   int dim = x.Size();
-   v.SetSize(dim); v = 0.0; v[dim-1] = u0_imag_exact(x);
+   v.SetSize(dim); v = 0.0; v[dim-1] = u0_exact(x);
 }
@@ -62,6 +62,10 @@ static real_t epsilon_ = 1.0;
 static real_t sigma_ = 20.0;
 static real_t omega_ = 10.0;

+complex<real_t> u0_exact(const Vector &x);
+void u1_exact(const Vector &, ComplexVector &);
+void u2_exact(const Vector &, ComplexVector &);
+
 real_t u0_real_exact(const Vector &);
 real_t u0_imag_exact(const Vector &);

@@ -244,13 +248,22 @@ int main(int argc, char *argv[])
   ParComplexGridFunction * u_exact = NULL;
   if (exact_sol) { u_exact = new ParComplexGridFunction(fespace); }

+   ComplexFunctionCoefficient u0(u0_exact);
+   ComplexVectorFunctionCoefficient u1(dim, u1_exact);
+   ComplexVectorFunctionCoefficient u2(dim, u2_exact);
+
+   ComplexConstantCoefficient oneCoef(1.0);
+
+   Vector  oneVec(dim);  oneVec = 0.0; oneVec[(prob==2)?(dim-1):0] = 1.0;
+   ComplexVectorConstantCoefficient oneVecCoef(oneVec);
+
   FunctionCoefficient u0_r(u0_real_exact);
   FunctionCoefficient u0_i(u0_imag_exact);
   VectorFunctionCoefficient u1_r(dim, u1_real_exact);
   VectorFunctionCoefficient u1_i(dim, u1_imag_exact);
   VectorFunctionCoefficient u2_r(dim, u2_real_exact);
   VectorFunctionCoefficient u2_i(dim, u2_imag_exact);
-
+   /*
   ConstantCoefficient zeroCoef(0.0);
   ConstantCoefficient oneCoef(1.0);

@@ -258,40 +271,40 @@ int main(int argc, char *argv[])
   Vector  oneVec(dim);  oneVec = 0.0; oneVec[(prob==2)?(dim-1):0] = 1.0;
   VectorConstantCoefficient zeroVecCoef(zeroVec);
   VectorConstantCoefficient oneVecCoef(oneVec);
-
+   */
   switch (prob)
   {
      case 0:
         if (exact_sol)
         {
-            u.ProjectBdrCoefficient(u0_r, u0_i, ess_bdr);
-            u_exact->ProjectCoefficient(u0_r, u0_i);
+            u.ProjectBdrCoefficient(u0, ess_bdr);
+            u_exact->ProjectCoefficient(u0);
         }
         else
         {
-            u.ProjectBdrCoefficient(oneCoef, zeroCoef, ess_bdr);
+            u.ProjectBdrCoefficient(oneCoef, ess_bdr);
         }
         break;
      case 1:
         if (exact_sol)
         {
-            u.ProjectBdrCoefficientTangent(u1_r, u1_i, ess_bdr);
-            u_exact->ProjectCoefficient(u1_r, u1_i);
+            u.ProjectBdrCoefficientTangent(u1, ess_bdr);
+            u_exact->ProjectCoefficient(u1);
         }
         else
         {
-            u.ProjectBdrCoefficientTangent(oneVecCoef, zeroVecCoef, ess_bdr);
+            u.ProjectBdrCoefficientTangent(oneVecCoef, ess_bdr);
         }
         break;
      case 2:
         if (exact_sol)
         {
-            u.ProjectBdrCoefficientNormal(u2_r, u2_i, ess_bdr);
-            u_exact->ProjectCoefficient(u2_r, u2_i);
+            u.ProjectBdrCoefficientNormal(u2, ess_bdr);
+            u_exact->ProjectCoefficient(u2);
         }
         else
         {
-            u.ProjectBdrCoefficientNormal(oneVecCoef, zeroVecCoef, ess_bdr);
+            u.ProjectBdrCoefficientNormal(oneVecCoef, ess_bdr);
         }
         break;
      default: break; // This should be unreachable
@@ -331,27 +344,24 @@ int main(int argc, char *argv[])
   ConstantCoefficient lossCoef(omega_ * sigma_);
   ConstantCoefficient negMassCoef(omega_ * omega_ * epsilon_);

+   ComplexConstantCoefficient complexMassCoef(-omega_ * omega_ * epsilon_,
+                                              omega_ * sigma_);
+
   ParSesquilinearForm *a = new ParSesquilinearForm(fespace, conv);
   if (pa) { a->SetAssemblyLevel(AssemblyLevel::PARTIAL); }
   switch (prob)
   {
      case 0:
-         a->AddDomainIntegrator(new DiffusionIntegrator(stiffnessCoef),
-                                NULL);
-         a->AddDomainIntegrator(new MassIntegrator(massCoef),
-                                new MassIntegrator(lossCoef));
+         a->AddDomainIntegrator<DiffusionIntegrator>(stiffnessCoef);
+         a->AddDomainIntegrator<MassIntegrator>(complexMassCoef);
         break;
      case 1:
-         a->AddDomainIntegrator(new CurlCurlIntegrator(stiffnessCoef),
-                                NULL);
-         a->AddDomainIntegrator(new VectorFEMassIntegrator(massCoef),
-                                new VectorFEMassIntegrator(lossCoef));
+         a->AddDomainIntegrator<CurlCurlIntegrator>(stiffnessCoef);
+         a->AddDomainIntegrator<VectorFEMassIntegrator>(complexMassCoef);
         break;
      case 2:
-         a->AddDomainIntegrator(new DivDivIntegrator(stiffnessCoef),
-                                NULL);
-         a->AddDomainIntegrator(new VectorFEMassIntegrator(massCoef),
-                                new VectorFEMassIntegrator(lossCoef));
+         a->AddDomainIntegrator<DivDivIntegrator>(stiffnessCoef);
+         a->AddDomainIntegrator<VectorFEMassIntegrator>(complexMassCoef);
         break;
      default: break; // This should be unreachable
   }
@@ -475,22 +485,18 @@ int main(int argc, char *argv[])

   if (exact_sol)
   {
-      real_t err_r = -1.0;
-      real_t err_i = -1.0;
+      real_t err_u = -1.0;

      switch (prob)
      {
         case 0:
-            err_r = u.real().ComputeL2Error(u0_r);
-            err_i = u.imag().ComputeL2Error(u0_i);
+            err_u = u.ComputeL2Error(u0);
            break;
         case 1:
-            err_r = u.real().ComputeL2Error(u1_r);
-            err_i = u.imag().ComputeL2Error(u1_i);
+            err_u = u.ComputeL2Error(u1);
            break;
         case 2:
-            err_r = u.real().ComputeL2Error(u2_r);
-            err_i = u.imag().ComputeL2Error(u2_i);
+            err_u = u.ComputeL2Error(u2);
            break;
         default: break; // This should be unreachable
      }
@@ -498,8 +504,7 @@ int main(int argc, char *argv[])
      if ( myid == 0 )
      {
         cout << endl;
-         cout << "|| Re (u_h - u) ||_{L^2} = " << err_r << endl;
-         cout << "|| Im (u_h - u) ||_{L^2} = " << err_i << endl;
+         cout << "|| u_h - u ||_{L^2} = " << err_u << endl;
         cout << endl;
      }
   }
@@ -627,6 +632,12 @@ real_t u0_imag_exact(const Vector &x)
   return u0_exact(x).imag();
 }

+void u1_exact(const Vector &x, ComplexVector &v)
+{
+   int dim = x.Size();
+   v.SetSize(dim); v = 0.0; v[0] = u0_exact(x);
+}
+
 void u1_real_exact(const Vector &x, Vector &v)
 {
   int dim = x.Size();
@@ -639,6 +650,12 @@ void u1_imag_exact(const Vector &x, Vector &v)
   v.SetSize(dim); v = 0.0; v[0] = u0_imag_exact(x);
 }

+void u2_exact(const Vector &x, ComplexVector &v)
+{
+   int dim = x.Size();
+   v.SetSize(dim); v = 0.0; v[dim-1] = u0_exact(x);
+}
+
 void u2_real_exact(const Vector &x, Vector &v)
 {
   int dim = x.Size();
@@ -1,6 +1,6 @@
 # Jupyter Notebooks using xeus-cling

-[![Binder](https://mybinder.org/badge_logo.svg)](https://mybinder.org/v2/gh/mfem/mfem/main?filepath=examples%2Fjupyter%2Fex.ipynb)
+[![Binder](https://mybinder.org/badge_logo.svg)](https://mybinder.org/v2/gh/mfem/mfem/master?filepath=examples%2Fjupyter%2Fex.ipynb)

 [xeus-cling](https://github.com/jupyter-xeus/xeus-cling) is a C++ Jupyter Kernel based on [cling](https://github.com/root-project/cling),
 which can be used to create interactive C++ MFEM and GLVis notebooks.
@@ -173,12 +173,6 @@ ex11p-test-cpardiso: ex11p
 	@$(call mfem-test,$<, $(RUN_MPI), MKL_CPARDISO example,--cpardiso)
 test-par-YES: ex11p-test-cpardiso
 endif
-ifeq ($(MFEM_USE_CEED),YES)
-ex1p-test-ceed: ex1p
-	@$(call mfem-test,$<, $(RUN_MPI),\
-	   Parallel libCEED example,-d ceed-cpu -pa -a)
-test-par-YES: ex1p-test-ceed
-endif

 # Testing: "test" target and mfem-test* variables are defined in config/test.mk

@@ -59,6 +59,7 @@ set(SRCS
  integ/nonlininteg_vecconvection_pa.cpp
  integ/nonlininteg_vecconvection_mf.cpp
  coefficient.cpp
+  complex_coefficient.cpp
  complex_fem.cpp
  convergence.cpp
  datacollection.cpp
@@ -82,8 +83,6 @@ set(SRCS
  fe/fe_ser.cpp
  fe_coll.cpp
  fespace.cpp
-  derefmat_op.cpp
-  pderefmat_op.cpp
  geom.cpp
  gridfunc.cpp
  hybridization.cpp
@@ -171,11 +170,6 @@ set(HDRS
  bilinearform.hpp
  bilinearform_ext.hpp
  bilininteg.hpp
-  integ/lininteg_domain_kernels.hpp
-  integ/bilininteg_dgdiffusion_kernels.hpp
-  integ/bilininteg_dgtrace_kernels.hpp
-  integ/bilininteg_vecdiffusion_kernels.hpp
-  integ/bilininteg_convection_kernels.hpp
  integ/bilininteg_diffusion_kernels.hpp
  integ/bilininteg_elasticity_kernels.hpp
  integ/bilininteg_hcurl_kernels.hpp
@@ -183,6 +177,7 @@ set(HDRS
  integ/bilininteg_hcurlhdiv_kernels.hpp
  integ/bilininteg_mass_kernels.hpp
  coefficient.hpp
+  complex_coefficient.hpp
  complex_fem.hpp
  convergence.hpp
  datacollection.hpp
@@ -246,13 +241,8 @@ set(HDRS
  lor/lor_ams.hpp
  lor/lor_batched.hpp
  lor/lor_h1.hpp
-  lor/lor_dg.hpp
  lor/lor_nd.hpp
  lor/lor_rt.hpp
-  lor/lor_h1_impl.hpp
-  lor/lor_dg_impl.hpp
-  lor/lor_nd_impl.hpp
-  lor/lor_rt_impl.hpp
  lor/lor_util.hpp
  multigrid.hpp
  nonlinearform.hpp
@@ -266,7 +266,11 @@ void PABilinearFormExtension::SetupRestrictionOperators(const L2FaceValues m)

      // Gather the attributes on the host from all the elements
      const Mesh &mesh = *trial_fes->GetMesh();
-      elem_attributes = &mesh.GetElementAttributes();
+      elem_attributes.SetSize(mesh.GetNE());
+      for (int i = 0; i < mesh.GetNE(); ++i)
+      {
+         elem_attributes[i] = mesh.GetAttribute(i);
+      }
   }

   // Construct face restriction operators only if the bilinear form has
@@ -325,7 +329,45 @@ void PABilinearFormExtension::SetupRestrictionOperators(const L2FaceValues m)
         bdr_face_dYdn.SetSize(bdr_face_restrict_lex->Height());
      }

-      bdr_face_attributes = &trial_fes->GetMesh()->GetBdrFaceAttributes();
+      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.");
+      }
   }
 }

@@ -387,7 +429,7 @@ void PABilinearFormExtension::AssembleDiagonal(Vector &y) const
         mfem::forall(ne, [=] MFEM_HOST_DEVICE (int e)
         {
            const int attr = d_attr[e];
-            if (attr <= 0 || d_m[attr - 1] == 0)
+            if (d_m[attr - 1] == 0)
            {
               for (int i = 0; i < nd; ++i)
               {
@@ -408,7 +450,7 @@ 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);
@@ -434,7 +476,7 @@ 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);
      }
   }

@@ -447,7 +489,7 @@ 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_face_attributes, bdr_face_Y);
+                                        bdr_attributes, bdr_face_Y);
      }
      bdr_face_restrict_lex->AddAbsMultTranspose(bdr_face_Y, y);
   }
@@ -546,7 +588,7 @@ 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)
         {
@@ -648,8 +690,8 @@ 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_face_attributes, false, bdr_face_Y);
+            AddMultWithMarkers(*bdr_integs[i], bdr_face_X, bdr_markers[i], bdr_attributes,
+                               false, bdr_face_Y);
         }
         for (int i = 0; i < n_bdr_face_integs; ++i)
         {
@@ -657,14 +699,12 @@ void PABilinearFormExtension::MultInternal(const Vector &x, Vector &y,
            {
               AddMultNormalDerivativesWithMarkers(
                  *bdr_face_integs[i], bdr_face_X, bdr_face_dXdn,
-                  bdr_face_markers[i], *bdr_face_attributes, bdr_face_Y,
-                  bdr_face_dYdn);
+                  bdr_face_markers[i], bdr_attributes, bdr_face_Y, bdr_face_dYdn);
            }
            else
            {
-               AddMultWithMarkers(*bdr_face_integs[i], bdr_face_X,
-                                  bdr_face_markers[i], *bdr_face_attributes, false,
-                                  bdr_face_Y);
+               AddMultWithMarkers(*bdr_face_integs[i], bdr_face_X, bdr_face_markers[i],
+                                  bdr_attributes, false, bdr_face_Y);
            }
         }
         bdr_face_restrict_lex->AddMultTransposeInPlace(bdr_face_Y, y);
@@ -687,7 +727,7 @@ 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);
@@ -734,14 +774,13 @@ 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_face_attributes, true, bdr_face_Y);
+            AddMultWithMarkers(*bdr_integs[i], bdr_face_X, bdr_markers[i], bdr_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_face_attributes, true,
-                               bdr_face_Y);
+            AddMultWithMarkers(*bdr_face_integs[i], bdr_face_X, bdr_face_markers[i],
+                               bdr_attributes, true, bdr_face_Y);
         }
         bdr_face_restrict_lex->AddMultTransposeInPlace(bdr_face_Y, y);
      }
@@ -765,7 +804,7 @@ static void AddWithMarkers_(
   mfem::forall(ne, [=] MFEM_HOST_DEVICE (int e)
   {
      const int attr = d_attr[e];
-      if (attr <= 0 || d_m[attr - 1] == 0) { return; }
+      if (d_m[attr - 1] == 0) { return; }
      for (int i = 0; i < nd; ++i)
      {
         d_y(i, e) += d_x(i, e);
@@ -881,8 +920,7 @@ void EABilinearFormExtension::Assemble()
      {
         const int i = idx % sz;
         const int e = idx / sz;
-         const real_t val =
-            d_a[e] > 0 ? (d_m[d_a[e] - 1] ? d_ea_1(i, e) : 0) : 0;
+         const real_t val = d_m[d_a[e] - 1] ? d_ea_1(i, e) : 0.0;
         if (add)
         {
            d_ea_2(i, e) += val;
@@ -915,7 +953,7 @@ 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);
         }
      }
   }
@@ -944,7 +982,7 @@ 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_face_attributes, add);
+                             bdr_attributes, add);
         }
      }
   }
@@ -993,7 +1031,7 @@ void EABilinearFormExtension::Assemble()
                                                        ea_data_tmp,
                                                        add);
            add_with_markers(ea_data_tmp, ea_data_bdr, nf_bdr, *markers,
-                             *bdr_face_attributes, add);
+                             bdr_attributes, add);
         }
      }
   }
@@ -69,8 +69,7 @@ class PABilinearFormExtension : public BilinearFormExtension
 protected:
   const FiniteElementSpace *trial_fes, *test_fes; // Not owned
   /// Attributes of all mesh elements.
-   const Array<int> *elem_attributes; // Not owned
-   const Array<int> *bdr_face_attributes; // Not owned
+   Array<int> elem_attributes, bdr_attributes;
   mutable Vector tmp_evec; // Work array
   mutable Vector localX, localY;
   mutable Vector int_face_X, int_face_Y;
@@ -23,8 +23,6 @@

 namespace mfem
 {
-class QuadratureSpace;
-class FaceQuadratureSpace;

 /// Abstract base class BilinearFormIntegrator
 class BilinearFormIntegrator : public NonlinearFormIntegrator
@@ -814,7 +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 );
   }

@@ -886,7 +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 );
   }

@@ -921,7 +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 );
   }

@@ -1602,7 +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 );
   }

@@ -1637,7 +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 );
   }

@@ -1671,7 +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 );
   }

@@ -1762,7 +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 );
   }

@@ -1795,7 +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   );
   }
@@ -1834,7 +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
             );
@@ -1975,7 +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 );
   }

@@ -2496,7 +2494,8 @@ private:
 #endif

 public:
-   ConvectionIntegrator(VectorCoefficient &q, real_t a = 1.0);
+   ConvectionIntegrator(VectorCoefficient &q, real_t a = 1.0)
+      : Q(&q) { alpha = a; }

   void AssembleElementMatrix(const FiniteElement &,
                              ElementTransformation &,
@@ -2529,28 +2528,6 @@ public:

   bool SupportsCeed() const override { return DeviceCanUseCeed(); }

-   /// arguments: NE, B, G, Bt, Gt, pa_data, x, y, D1D, Q1D
-   using ApplyKernelType = void (*)(const int, const Array<real_t> &,
-                                    const Array<real_t> &,
-                                    const Array<real_t> &,
-                                    const Array<real_t> &, const Vector &,
-                                    const Vector &, Vector &, const int,
-                                    const int);
-
-   /// arguments: DIMS, D1D, Q1D
-   MFEM_REGISTER_KERNELS(ApplyPAKernels, ApplyKernelType, (int, int, int));
-   /// arguments: DIMS, D1D, Q1D
-   MFEM_REGISTER_KERNELS(ApplyPATKernels, ApplyKernelType, (int, int, int));
-
-   template <int DIM, int D1D, int Q1D>
-   static void AddSpecialization()
-   {
-      ApplyPAKernels::Specialization<DIM, D1D, Q1D>::Add();
-      ApplyPATKernels::Specialization<DIM, D1D, Q1D>::Add();
-   }
-
-   struct Kernels { Kernels(); };
-
 protected:
   const IntegrationRule* GetDefaultIntegrationRule(
      const FiniteElement& trial_fe,
@@ -2821,13 +2798,15 @@ protected:
   bool symmetric = true; ///< False if using a nonsymmetric matrix coefficient

 public:
-   CurlCurlIntegrator();
+   CurlCurlIntegrator() { Q = NULL; DQ = NULL; MQ = NULL; }
   /// Construct a bilinear form integrator for Nedelec elements
-   CurlCurlIntegrator(Coefficient &q, const IntegrationRule *ir = nullptr);
+   CurlCurlIntegrator(Coefficient &q, const IntegrationRule *ir = NULL) :
+      BilinearFormIntegrator(ir), Q(&q), DQ(NULL), MQ(NULL) { }
   CurlCurlIntegrator(DiagonalMatrixCoefficient &dq,
-                      const IntegrationRule *ir = nullptr);
-   CurlCurlIntegrator(MatrixCoefficient &mq,
-                      const IntegrationRule *ir = nullptr);
+                      const IntegrationRule *ir = NULL) :
+      BilinearFormIntegrator(ir), Q(NULL), DQ(&dq), MQ(NULL) { }
+   CurlCurlIntegrator(MatrixCoefficient &mq, const IntegrationRule *ir = NULL) :
+      BilinearFormIntegrator(ir), Q(NULL), DQ(NULL), MQ(&mq) { }

   /* Given a particular Finite Element, compute the
      element curl-curl matrix elmat */
@@ -2857,34 +2836,6 @@ public:
   void AssembleDiagonalPA(Vector& diag) override;

   const Coefficient *GetCoefficient() const { return Q; }
-
-   /// arguments: d1d, q1d, symmetric, NE, bo, bc, bot, bct, gc, gct, pa_data,
-   /// x, y, useAbs
-   using ApplyKernelType = void (*)(
-                              const int, const int, const bool, const int, const Array<real_t> &,
-                              const Array<real_t> &, const Array<real_t> &, const Array<real_t> &,
-                              const Array<real_t> &, const Array<real_t> &, const Vector &,
-                              const Vector &, Vector &, const bool);
-
-   /// arguments: d1d, q1d, symmetric, ne, Bo, Bc, Go, Gc, pa_data, diag
-   using DiagonalKernelType = void (*)(const int, const int, const bool,
-                                       const int, const Array<real_t> &,
-                                       const Array<real_t> &,
-                                       const Array<real_t> &,
-                                       const Array<real_t> &, const Vector &,
-                                       Vector &);
-
-   /// parameters: dim, d1d, q1d
-   MFEM_REGISTER_KERNELS(ApplyPAKernels, ApplyKernelType, (int, int, int));
-   /// parameters: dim, d1d, q1d
-   MFEM_REGISTER_KERNELS(DiagonalPAKernels, DiagonalKernelType, (int, int, int));
-   struct Kernels { Kernels(); };
-
-   template <int DIM, int D1D, int Q1D> static void AddSpecialization()
-   {
-      ApplyPAKernels::Specialization<DIM, D1D, Q1D>::Add();
-      DiagonalPAKernels::Specialization<DIM, D1D, Q1D>::Add();
-   }
 };

 /** Integrator for $(\mathrm{curl}(u), \mathrm{curl}(v))$ for FE spaces defined by 'dim' copies of a
@@ -3138,18 +3089,21 @@ private:
   Vector vcoeff;

 public:
-   VectorDiffusionIntegrator(const IntegrationRule *ir = nullptr);
+   VectorDiffusionIntegrator() { }

   /** \brief Integrator with unit coefficient for caller-specified vector
       dimension.

       If the vector dimension does not match the true dimension of the space,
       the resulting element matrix will be mathematically invalid. */
-   VectorDiffusionIntegrator(int vector_dimension);
+   VectorDiffusionIntegrator(int vector_dimension)
+      : vdim(vector_dimension) { }

-   VectorDiffusionIntegrator(Coefficient &q);
+   VectorDiffusionIntegrator(Coefficient &q)
+      : Q(&q) { }

-   VectorDiffusionIntegrator(Coefficient &q, const IntegrationRule *ir);
+   VectorDiffusionIntegrator(Coefficient &q, const IntegrationRule *ir)
+      : BilinearFormIntegrator(ir), Q(&q) { }

   /** \brief Integrator with scalar coefficient for caller-specified vector
       dimension.
@@ -3159,7 +3113,8 @@ public:

       If the vector dimension does not match the true dimension of the space,
       the resulting element matrix will be mathematically invalid. */
-   VectorDiffusionIntegrator(Coefficient &q, int vector_dimension);
+   VectorDiffusionIntegrator(Coefficient &q, int vector_dimension)
+      : Q(&q), vdim(vector_dimension) { }

   /** \brief Integrator with \c VectorCoefficient. The vector dimension of the
       \c FiniteElementSpace is assumed to be the same as the dimension of the
@@ -3170,7 +3125,8 @@ public:

       If the vector dimension does not match the true dimension of the space,
       the resulting element matrix will be mathematically invalid. */
-   VectorDiffusionIntegrator(VectorCoefficient &vq);
+   VectorDiffusionIntegrator(VectorCoefficient &vq)
+      : VQ(&vq), vdim(vq.GetVDim()) { }

   /** \brief Integrator with \c MatrixCoefficient. The vector dimension of the
       \c FiniteElementSpace is assumed to be the same as the dimension of the
@@ -3181,7 +3137,8 @@ public:

       If the vector dimension does not match the true dimension of the space,
       the resulting element matrix will be mathematically invalid. */
-   VectorDiffusionIntegrator(MatrixCoefficient& mq);
+   VectorDiffusionIntegrator(MatrixCoefficient& mq)
+      : MQ(&mq), vdim(mq.GetVDim()) { }

   void AssembleElementMatrix(const FiniteElement &el,
                              ElementTransformation &Trans,
@@ -3197,28 +3154,6 @@ public:
   void AddMultPA(const Vector &x, Vector &y) const override;
   void AddMultMF(const Vector &x, Vector &y) const override;
   bool SupportsCeed() const override { return DeviceCanUseCeed(); }
-
-   /// arguments: ne, B, G, Bt, Gt, pa_data, x, y, d1d, q1d, vdim
-   using ApplyKernelType = void (*)(const int, const Array<real_t> &,
-                                    const Array<real_t> &,
-                                    const Array<real_t> &,
-                                    const Array<real_t> &, const Vector &,
-                                    const Vector &, Vector &, const int,
-                                    const int, const int);
-
-   /// arguments: dim, vdim, d1d, q1d
-   MFEM_REGISTER_KERNELS(ApplyPAKernels, ApplyKernelType, (int, int, int, int));
-
-   template <int DIM, int VDIM, int D1D, int Q1D>
-   static void AddSpecialization()
-   {
-      ApplyPAKernels::Specialization<DIM, VDIM, D1D, Q1D>::Add();
-   }
-
-   struct Kernels
-   {
-      Kernels();
-   };
 };

 /** Integrator for the linear elasticity form:
@@ -3372,8 +3307,8 @@ public:
 class DGTraceIntegrator : public BilinearFormIntegrator
 {
 protected:
-   Coefficient *rho = nullptr;
-   VectorCoefficient *u = nullptr;
+   Coefficient *rho;
+   VectorCoefficient *u;
   real_t alpha, beta;
   // PA extension
   Vector pa_data;
@@ -3386,16 +3321,17 @@ private:
   Vector tr_shape1, te_shape1, tr_shape2, te_shape2;

 public:
-   DGTraceIntegrator(real_t a, real_t b);
-
   /// Construct integrator with $\rho = 1$, $\beta = \alpha/2$.
-   DGTraceIntegrator(VectorCoefficient &u_, real_t a);
+   DGTraceIntegrator(VectorCoefficient &u_, real_t a)
+   { rho = NULL; u = &u_; alpha = a; beta = 0.5*a; }

   /// Construct integrator with $\rho = 1$.
-   DGTraceIntegrator(VectorCoefficient &u_, real_t a, real_t b);
+   DGTraceIntegrator(VectorCoefficient &u_, real_t a, real_t b)
+   { rho = NULL; u = &u_; alpha = a; beta = b; }

   DGTraceIntegrator(Coefficient &rho_, VectorCoefficient &u_,
-                     real_t a, real_t b);
+                     real_t a, real_t b)
+   { rho = &rho_; u = &u_; alpha = a; beta = b; }

   using BilinearFormIntegrator::AssembleFaceMatrix;
   void AssembleFaceMatrix(const FiniteElement &el1,
@@ -3434,26 +3370,6 @@ public:
   static const IntegrationRule &GetRule(Geometry::Type geom, int order,
                                         const ElementTransformation &T);

-   /// arguments: nf, B, Bt, pa_data, x, y, dofs1D, quad1D
-   using ApplyKernelType = void (*)(const int, const Array<real_t> &,
-                                    const Array<real_t> &, const Vector &,
-                                    const Vector &, Vector &, const int,
-                                    const int);
-
-   /// arguments: DIM, d1d, q1d
-   MFEM_REGISTER_KERNELS(ApplyPAKernels, ApplyKernelType, (int, int, int));
-   /// arguments: DIM, d1d, q1d
-   MFEM_REGISTER_KERNELS(ApplyPATKernels, ApplyKernelType, (int, int, int));
-
-   template <int DIM, int D1D, int Q1D> static void AddSpecialization()
-   {
-      ApplyPAKernels::Specialization<DIM, D1D, Q1D>::Add();
-      ApplyPATKernels::Specialization<DIM, D1D, Q1D>::Add();
-   }
-
-   struct Kernels { Kernels(); };
-
-
 private:
   void SetupPA(const FiniteElementSpace &fes, FaceType type);
 };
@@ -3500,8 +3416,8 @@ public:
 class DGDiffusionIntegrator : public BilinearFormIntegrator
 {
 protected:
-   Coefficient *Q = nullptr;
-   MatrixCoefficient *MQ = nullptr;
+   Coefficient *Q;
+   MatrixCoefficient *MQ;
   real_t sigma, kappa;

   // these are not thread-safe!
@@ -3516,11 +3432,15 @@ protected:
   IntegrationRules irs{0, Quadrature1D::GaussLobatto};

 public:
-   DGDiffusionIntegrator(const real_t s, const real_t k);
-   DGDiffusionIntegrator(Coefficient &q, const real_t s, const real_t k);
-   DGDiffusionIntegrator(MatrixCoefficient &q, const real_t s, const real_t k);
+   DGDiffusionIntegrator(const real_t s, const real_t k)
+      : Q(NULL), MQ(NULL), sigma(s), kappa(k) { }
+   DGDiffusionIntegrator(Coefficient &q, const real_t s, const real_t k)
+      : Q(&q), MQ(NULL), sigma(s), kappa(k) { }
+   DGDiffusionIntegrator(MatrixCoefficient &q, const real_t s, const real_t k)
+      : Q(NULL), MQ(&q), sigma(s), kappa(k) { }
   using BilinearFormIntegrator::AssembleFaceMatrix;
-   void AssembleFaceMatrix(const FiniteElement &el1, const FiniteElement &el2,
+   void AssembleFaceMatrix(const FiniteElement &el1,
+                           const FiniteElement &el2,
                           FaceElementTransformations &Trans,
                           DenseMatrix &elmat) override;

@@ -3539,28 +3459,6 @@ public:

   const IntegrationRule &GetRule(int order, Geometry::Type geom);

-   real_t GetPenaltyParameter() const { return kappa; }
-
-   /// arguments: nf, B, Bt, G, Gt, sigma, pa_data, x, dxdn, y, dydn, dofs1D,
-   /// quad1D
-   using ApplyKernelType = void (*)(const int, const Array<real_t> &,
-                                    const Array<real_t> &,
-                                    const Array<real_t> &,
-                                    const Array<real_t> &, const real_t,
-                                    const Vector &, const Vector &_,
-                                    const Vector &, Vector &, Vector &,
-                                    const int, const int);
-
-   /// arguments: DIM, d1d, q1d
-   MFEM_REGISTER_KERNELS(ApplyPAKernels, ApplyKernelType, (int, int, int));
-
-   template <int DIM, int D1D, int Q1D> static void AddSpecialization()
-   {
-      ApplyPAKernels::Specialization<DIM, D1D, Q1D>::Add();
-   }
-
-   struct Kernels { Kernels(); };
-
 private:
   void SetupPA(const FiniteElementSpace &fes, FaceType type);
 };
@@ -8,7 +8,6 @@
 // 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 <ceed/types.h>

 /// A structure used to pass additional data to f_build_conv and f_apply_conv
 struct ConvectionContext {
@@ -92,7 +91,7 @@ CEED_QFUNCTION(f_build_conv_const)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 /// libCEED Q-function for building quadrature data for a convection operator
@@ -168,7 +167,7 @@ CEED_QFUNCTION(f_build_conv_quad)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 /// libCEED Q-function for applying a conv operator
@@ -234,7 +233,7 @@ CEED_QFUNCTION(f_apply_conv)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 /// libCEED Q-function for applying a conv operator
@@ -382,7 +381,7 @@ CEED_QFUNCTION(f_apply_conv_mf_const)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 CEED_QFUNCTION(f_apply_conv_mf_quad)(void *ctx, CeedInt Q,
@@ -526,5 +525,5 @@ CEED_QFUNCTION(f_apply_conv_mf_quad)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }
@@ -8,7 +8,7 @@
 // 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 <ceed/types.h>
+

 /// A structure used to pass additional data to f_build_diff and f_apply_diff
 struct DiffusionContext { CeedInt dim, space_dim, vdim; CeedScalar coeff; };
@@ -85,7 +85,7 @@ CEED_QFUNCTION(f_build_diff_const)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 /// libCEED Q-function for building quadrature data for a diffusion operator
@@ -161,7 +161,7 @@ CEED_QFUNCTION(f_build_diff_quad)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 /// libCEED Q-function for applying a diff operator
@@ -241,7 +241,7 @@ CEED_QFUNCTION(f_apply_diff)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 /// libCEED Q-function for applying a diff operator
@@ -394,7 +394,7 @@ CEED_QFUNCTION(f_apply_diff_mf_const)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 CEED_QFUNCTION(f_apply_diff_mf_quad)(void *ctx, CeedInt Q,
@@ -549,5 +549,5 @@ CEED_QFUNCTION(f_apply_diff_mf_quad)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }
@@ -8,7 +8,7 @@
 // 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 <ceed/types.h>
+

 /// A structure used to pass additional data to f_build_diff and f_apply_diff
 struct MassContext { CeedInt dim, space_dim, vdim; CeedScalar coeff; };
@@ -53,7 +53,7 @@ CEED_QFUNCTION(f_build_mass_const)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 /// libCEED Q-function for building quadrature data for a mass operator with a
@@ -95,7 +95,7 @@ CEED_QFUNCTION(f_build_mass_quad)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 /// libCEED Q-function for applying a mass operator
@@ -135,7 +135,7 @@ CEED_QFUNCTION(f_apply_mass)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 /// libCEED Q-function for applying a diff operator
@@ -199,7 +199,7 @@ CEED_QFUNCTION(f_apply_mass_mf_const)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 CEED_QFUNCTION(f_apply_mass_mf_quad)(void *ctx, CeedInt Q,
@@ -266,5 +266,5 @@ CEED_QFUNCTION(f_apply_mass_mf_quad)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }
@@ -8,7 +8,6 @@
 // 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 <ceed/types.h>

 /// A structure used to pass additional data to f_build_conv and f_apply_conv
 struct NLConvectionContext { CeedInt dim, space_dim, vdim; CeedScalar coeff; };
@@ -88,7 +87,7 @@ CEED_QFUNCTION(f_build_conv_const)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 /// libCEED Q-function for building quadrature data for a convection operator
@@ -168,7 +167,7 @@ CEED_QFUNCTION(f_build_conv_quad)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 /// libCEED Q-function for applying a conv operator
@@ -248,7 +247,7 @@ CEED_QFUNCTION(f_apply_conv)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 /// libCEED Q-function for applying a conv operator
@@ -363,7 +362,7 @@ CEED_QFUNCTION(f_apply_conv_mf_const)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }

 CEED_QFUNCTION(f_apply_conv_mf_quad)(void *ctx, CeedInt Q,
@@ -476,5 +475,5 @@ CEED_QFUNCTION(f_apply_conv_mf_quad)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return CEED_ERROR_SUCCESS;
+   return 0;
 }
@@ -18,21 +18,10 @@

 #include <ceed.h>

-#if !CEED_VERSION_GE(0, 12, 0)
+#if !CEED_VERSION_GE(0,12,0)
 #error MFEM requires a libCEED version >= 0.12.0
 #endif

-#if !CEED_VERSION_GE(0, 13, 0)
-#define CeedOperatorCreateComposite(ceed, op)                                  \
-    CeedCompositeOperatorCreate((ceed), (op))
-#define CeedOperatorCompositeAddSub(op, sub)                                   \
-    CeedCompositeOperatorAddSub((op), (sub))
-#define CeedOperatorCompositeGetNumSub(op, num)                                \
-    CeedCompositeOperatorGetNumSub((op), (num))
-#define CeedOperatorCompositeGetSubList(op, list)                              \
-    CeedCompositeOperatorGetSubList((op), (list))
-#endif
-
 namespace mfem
 {

@@ -83,7 +83,7 @@ public:
      }

      // Create composite CeedOperator
-      CeedOperatorCreateComposite(internal::ceed, &oper);
+      CeedCompositeOperatorCreate(internal::ceed, &oper);

      // Create each sub-CeedOperator
      sub_ops.reserve(element_indices.size());
@@ -101,7 +101,7 @@ public:
         int nelem = *count[value.first];
         sub_op->Assemble(info, fes, ir, nelem, indices, Q);
         sub_ops.push_back(sub_op);
-         CeedOperatorCompositeAddSub(oper, sub_op->GetCeedOperator());
+         CeedCompositeOperatorAddSub(oper, sub_op->GetCeedOperator());
      }

      const int ndofs = fes.GetVDim() * fes.GetNDofs();
@@ -140,7 +140,11 @@ int CeedOperatorGetActiveField(CeedOperator oper, CeedOperatorField *field)
   CeedOperator *subops;
   if (isComposite)
   {
-      ierr = CeedOperatorCompositeGetSubList(oper, &subops); PCeedChk(ierr);
+#if CEED_VERSION_GE(0, 10, 2)
+      ierr = CeedCompositeOperatorGetSubList(oper, &subops); PCeedChk(ierr);
+#else
+      ierr = CeedOperatorGetSubList(oper, &subops); PCeedChk(ierr);
+#endif
      ierr = CeedOperatorGetQFunction(subops[0], &qf); PCeedChk(ierr);
   }
   else
@@ -167,11 +171,7 @@ int CeedOperatorGetActiveField(CeedOperator oper, CeedOperatorField *field)
   for (int i = 0; i < numinputfields; ++i)
   {
      ierr = CeedOperatorFieldGetVector(inputfields[i], &if_vector); PCeedChk(ierr);
-      bool is_active = if_vector == CEED_VECTOR_ACTIVE;
-#if CEED_VERSION_GE(0, 13, 0)
-      ierr = CeedVectorDestroy(&if_vector); PCeedChk(ierr);
-#endif
-      if (is_active)
+      if (if_vector == CEED_VECTOR_ACTIVE)
      {
         if (found)
         {
@@ -228,7 +228,7 @@ void AddToCompositeOperator(BilinearFormIntegrator *integ, CeedOperator op)
 {
   if (integ->SupportsCeed())
   {
-      CeedOperatorCompositeAddSub(op, integ->GetCeedOp().GetCeedOperator());
+      CeedCompositeOperatorAddSub(op, integ->GetCeedOp().GetCeedOperator());
   }
   else
   {
@@ -240,7 +240,7 @@ CeedOperator CreateCeedCompositeOperatorFromBilinearForm(BilinearForm &form)
 {
   int ierr;
   CeedOperator op;
-   ierr = CeedOperatorCreateComposite(internal::ceed, &op); PCeedChk(ierr);
+   ierr = CeedCompositeOperatorCreate(internal::ceed, &op); PCeedChk(ierr);

   MFEM_VERIFY(form.GetBBFI()->Size() == 0,
               "Not implemented for this integrator!");
@@ -271,13 +271,18 @@ CeedOperator CoarsenCeedCompositeOperator(
   MFEM_ASSERT(isComposite, "");

   CeedOperator op_coarse;
-   ierr = CeedOperatorCreateComposite(internal::ceed,
+   ierr = CeedCompositeOperatorCreate(internal::ceed,
                                      &op_coarse); PCeedChk(ierr);

   int nsub;
   CeedOperator *subops;
-   ierr = CeedOperatorCompositeGetNumSub(op, &nsub); PCeedChk(ierr);
-   ierr = CeedOperatorCompositeGetSubList(op, &subops); PCeedChk(ierr);
+#if CEED_VERSION_GE(0, 10, 2)
+   ierr = CeedCompositeOperatorGetNumSub(op, &nsub); PCeedChk(ierr);
+   ierr = CeedCompositeOperatorGetSubList(op, &subops); PCeedChk(ierr);
+#else
+   ierr = CeedOperatorGetNumSub(op, &nsub); PCeedChk(ierr);
+   ierr = CeedOperatorGetSubList(op, &subops); PCeedChk(ierr);
+#endif
   for (int isub=0; isub<nsub; ++isub)
   {
      CeedOperator subop = subops[isub];
@@ -289,7 +294,7 @@ CeedOperator CoarsenCeedCompositeOperator(
      // refcounted by existing objects
      ierr = CeedBasisDestroy(&basis_coarse); PCeedChk(ierr);
      ierr = CeedBasisDestroy(&basis_c2f); PCeedChk(ierr);
-      ierr = CeedOperatorCompositeAddSub(op_coarse, subop_coarse);
+      ierr = CeedCompositeOperatorAddSub(op_coarse, subop_coarse);
      PCeedChk(ierr);
      ierr = CeedOperatorDestroy(&subop_coarse); PCeedChk(ierr);
   }
@@ -81,27 +81,12 @@ int CeedSingleOperatorFullAssemble(CeedOperator op, SparseMatrix *out)
      ierr = CeedOperatorFieldGetVector(input_fields[i], &vec); PCeedChk(ierr);
      if (vec == CEED_VECTOR_ACTIVE)
      {
-         CeedBasis basis;
-         ierr = CeedOperatorFieldGetBasis(input_fields[i], &basis); PCeedChk(ierr);
-         if (!basisin)
-         {
-            ierr = CeedBasisReferenceCopy(basis, &basisin); PCeedChk(ierr);
-         }
-#if CEED_VERSION_GE(0, 13, 0)
-         ierr = CeedBasisDestroy(&basis); PCeedChk(ierr);
-#endif
+         ierr = CeedOperatorFieldGetBasis(input_fields[i], &basisin);
+         PCeedChk(ierr);
         ierr = CeedBasisGetNumComponents(basisin, &ncomp); PCeedChk(ierr);
         ierr = CeedBasisGetDimension(basisin, &dim); PCeedChk(ierr);
-         CeedElemRestriction rstr;
-         ierr = CeedOperatorFieldGetElemRestriction(input_fields[i], &rstr);
+         ierr = CeedOperatorFieldGetElemRestriction(input_fields[i], &rstrin);
         PCeedChk(ierr);
-         if (!rstrin)
-         {
-            ierr = CeedElemRestrictionReferenceCopy(rstr, &rstrin); PCeedChk(ierr);
-         }
-#if CEED_VERSION_GE(0, 13, 0)
-         ierr = CeedElemRestrictionDestroy(&rstr); PCeedChk(ierr);
-#endif
         CeedEvalMode emode;
         ierr = CeedQFunctionFieldGetEvalMode(qffields[i], &emode);
         PCeedChk(ierr);
@@ -127,9 +112,6 @@ int CeedSingleOperatorFullAssemble(CeedOperator op, SparseMatrix *out)
               break; // Caught by QF Assembly
         }
      }
-#if CEED_VERSION_GE(0, 13, 0)
-      ierr = CeedVectorDestroy(&vec); PCeedChk(ierr);
-#endif
   }

   // Determine active output basis
@@ -145,25 +127,11 @@ int CeedSingleOperatorFullAssemble(CeedOperator op, SparseMatrix *out)
      ierr = CeedOperatorFieldGetVector(output_fields[i], &vec); PCeedChk(ierr);
      if (vec == CEED_VECTOR_ACTIVE)
      {
-         CeedBasis basis;
-         ierr = CeedOperatorFieldGetBasis(output_fields[i], &basis); PCeedChk(ierr);
-         if (!basisout)
-         {
-            ierr = CeedBasisReferenceCopy(basis, &basisout); PCeedChk(ierr);
-         }
-#if CEED_VERSION_GE(0, 13, 0)
-         ierr = CeedBasisDestroy(&basis); PCeedChk(ierr);
-#endif
-         CeedElemRestriction rstr;
-         ierr = CeedOperatorFieldGetElemRestriction(output_fields[i], &rstr);
+         ierr = CeedOperatorFieldGetBasis(output_fields[i], &basisout);
+         PCeedChk(ierr);
+         ierr = CeedOperatorFieldGetElemRestriction(output_fields[i], &rstrout);
+         PCeedChk(ierr);
         PCeedChk(ierr);
-         if (!rstrout)
-         {
-            ierr = CeedElemRestrictionReferenceCopy(rstr, &rstrout); PCeedChk(ierr);
-         }
-#if CEED_VERSION_GE(0, 13, 0)
-         ierr = CeedElemRestrictionDestroy(&rstr); PCeedChk(ierr);
-#endif
         CeedEvalMode emode;
         ierr = CeedQFunctionFieldGetEvalMode(qffields[i], &emode);
         PCeedChk(ierr);
@@ -189,9 +157,6 @@ int CeedSingleOperatorFullAssemble(CeedOperator op, SparseMatrix *out)
               break; // Caught by QF Assembly
         }
      }
-#if CEED_VERSION_GE(0, 13, 0)
-      ierr = CeedVectorDestroy(&vec); PCeedChk(ierr);
-#endif
   }

   CeedInt nelem, elemsize, nqpts;
@@ -235,11 +200,7 @@ int CeedSingleOperatorFullAssemble(CeedOperator op, SparseMatrix *out)
   PCeedChk(ierr);

   CeedInt layout[3];
-#if CEED_VERSION_GE(0, 13, 0)
-   ierr = CeedElemRestrictionGetELayout(rstr_q, layout); PCeedChk(ierr);
-#else
   ierr = CeedElemRestrictionGetELayout(rstr_q, &layout); PCeedChk(ierr);
-#endif
   ierr = CeedElemRestrictionDestroy(&rstr_q); PCeedChk(ierr);

   // enforce structurally symmetric for later elimination
@@ -324,10 +285,6 @@ int CeedSingleOperatorFullAssemble(CeedOperator op, SparseMatrix *out)
   ierr = CeedVectorRestoreArrayRead(assembledqf, &assembledqfarray);
   PCeedChk(ierr);
   ierr = CeedVectorDestroy(&assembledqf); PCeedChk(ierr);
-   ierr = CeedElemRestrictionDestroy(&rstrin); PCeedChk(ierr);
-   ierr = CeedElemRestrictionDestroy(&rstrout); PCeedChk(ierr);
-   ierr = CeedBasisDestroy(&basisin); PCeedChk(ierr);
-   ierr = CeedBasisDestroy(&basisout); PCeedChk(ierr);
   ierr = CeedHackFree(&emodein); PCeedChk(ierr);
   ierr = CeedHackFree(&emodeout); PCeedChk(ierr);

@@ -353,8 +310,13 @@ int CeedOperatorFullAssemble(CeedOperator op, SparseMatrix **mat)
   {
      CeedInt numsub;
      CeedOperator *subops;
-      ierr = CeedOperatorCompositeGetNumSub(op, &numsub); PCeedChk(ierr);
-      ierr = CeedOperatorCompositeGetSubList(op, &subops); PCeedChk(ierr);
+#if CEED_VERSION_GE(0, 10, 2)
+      CeedCompositeOperatorGetNumSub(op, &numsub);
+      ierr = CeedCompositeOperatorGetSubList(op, &subops); PCeedChk(ierr);
+#else
+      CeedOperatorGetNumSub(op, &numsub);
+      ierr = CeedOperatorGetSubList(op, &subops); PCeedChk(ierr);
+#endif
      for (int i = 0; i < numsub; ++i)
      {
         ierr = CeedSingleOperatorFullAssemble(subops[i], out); PCeedChk(ierr);
@@ -120,11 +120,7 @@ int CeedATPMGElemRestriction(int order,
   }
   ierr = CeedVectorRestoreArray(in_lvec, &lvec_data); PCeedChk(ierr);
   CeedInt in_layout[3];
-#if CEED_VERSION_GE(0, 13, 0)
-   ierr = CeedElemRestrictionGetELayout(er_in, in_layout); PCeedChk(ierr);
-#else
   ierr = CeedElemRestrictionGetELayout(er_in, &in_layout); PCeedChk(ierr);
-#endif
   if (in_layout[0] == 0 && in_layout[1] == 0 && in_layout[2] == 0)
   {
      return CeedError(ceed, 1, "Cannot interpret e-vector ordering of given"
@@ -668,11 +664,7 @@ int CeedATPMGOperator(CeedOperator oper, int order_reduction,

   for (int i = 0; i < numinputfields; ++i)
   {
-#if CEED_VERSION_GE(0, 13, 0)
-      const char * fieldname;
-#else
      char * fieldname;
-#endif
      ierr = CeedQFunctionFieldGetName(inputqfields[i], &fieldname); PCeedChk(ierr);
      if (if_vector[i] == CEED_VECTOR_ACTIVE)
      {
@@ -684,19 +676,10 @@ int CeedATPMGOperator(CeedOperator oper, int order_reduction,
         ierr = CeedOperatorSetField(coper, fieldname, er_input[i], basis_input[i],
                                     if_vector[i]); PCeedChk(ierr);
      }
-#if CEED_VERSION_GE(0, 13, 0)
-      ierr = CeedVectorDestroy(&if_vector[i]); PCeedChk(ierr);
-      ierr = CeedElemRestrictionDestroy(&er_input[i]); PCeedChk(ierr);
-      ierr = CeedBasisDestroy(&basis_input[i]); PCeedChk(ierr);
-#endif
   }
   for (int i = 0; i < numoutputfields; ++i)
   {
-#if CEED_VERSION_GE(0, 13, 0)
-      const char * fieldname;
-#else
      char * fieldname;
-#endif
      ierr = CeedQFunctionFieldGetName(outputqfields[i], &fieldname); PCeedChk(ierr);
      if (of_vector[i] == CEED_VECTOR_ACTIVE)
      {
@@ -708,11 +691,6 @@ int CeedATPMGOperator(CeedOperator oper, int order_reduction,
         ierr = CeedOperatorSetField(coper, fieldname, er_output[i], basis_output[i],
                                     of_vector[i]); PCeedChk(ierr);
      }
-#if CEED_VERSION_GE(0, 13, 0)
-      ierr = CeedVectorDestroy(&of_vector[i]); PCeedChk(ierr);
-      ierr = CeedElemRestrictionDestroy(&er_output[i]); PCeedChk(ierr);
-      ierr = CeedBasisDestroy(&basis_output[i]); PCeedChk(ierr);
-#endif
   }
   delete [] er_input;
   delete [] er_output;
@@ -763,9 +741,7 @@ int CeedOperatorGetOrder(CeedOperator oper, CeedInt * order)
   int P1d;
   ierr = CeedBasisGetNumNodes1D(basis, &P1d); PCeedChk(ierr);
   *order = P1d - 1;
-#if CEED_VERSION_GE(0, 13, 0)
-   ierr = CeedBasisDestroy(&basis); PCeedChk(ierr);
-#endif
+
   return 0;
 }

@@ -12,7 +12,6 @@
 // Implementation of Coefficient class

 #include "fem.hpp"
-#include "../general/forall.hpp"

 #include <cmath>
 #include <limits>
@@ -81,49 +80,6 @@ real_t PWConstCoefficient::Eval(ElementTransformation & T,
   return (constants(att-1));
 }

-void PWConstCoefficient::Project(QuadratureFunction &qf)
-{
-   auto &qs = *qf.GetSpace();
-
-   const bool compressed =
-      qs.Offsets(QSpaceOffsetStorage::COMPRESSED).Size() == 1;
-   const int *offsets = qs.Offsets(QSpaceOffsetStorage::COMPRESSED).Read();
-   const int ne = qs.GetNE();
-
-   const int *attributes = [&]()
-   {
-      if (dynamic_cast<QuadratureSpace*>(&qs) != nullptr)
-      {
-         return qs.GetMesh()->GetElementAttributes().Read();
-      }
-      else if (auto *qs_f = dynamic_cast<FaceQuadratureSpace*>(&qs))
-      {
-         MFEM_VERIFY(qs_f->GetFaceType() == FaceType::Boundary,
-                     "Interior faces do not have attributes.");
-         return qs.GetMesh()->GetBdrFaceAttributes().Read();
-      }
-      else
-      {
-         MFEM_ABORT("Unsupported case.");
-      }
-   }();
-
-   const real_t *d_c = constants.Read();
-   real_t *d_qf = qf.Write();
-
-   mfem::forall(ne, [=] MFEM_HOST_DEVICE (int e)
-   {
-      const int a = attributes[e];
-      const real_t elementConstant = d_c[a - 1];
-      const int begin = compressed ? e*offsets[0] : offsets[e];
-      const int end = compressed ? (e+1)*offsets[0] : offsets[e+1];
-      for (int i = begin; i < end; ++i)
-      {
-         d_qf[i] = elementConstant;
-      }
-   });
-}
-
 void PWCoefficient::InitMap(const Array<int> & attr,
                            const Array<Coefficient*> & coefs)
 {
@@ -563,26 +519,6 @@ void GradientGridFunctionCoefficient::Eval(
   }
 }

-void GradientGridFunctionCoefficient::Project(QuadratureFunction &qf)
-{
-   const FiniteElementSpace &fes = *GridFunc->FESpace();
-   const Mesh &mesh = *fes.GetMesh();
-   const int sdim = mesh.SpaceDimension();
-   const int gf_vdim = fes.GetVDim(); // assumed to be 1 in this class
-   qf.SetVDim(sdim*gf_vdim);
-   if (mesh.GetNE() == 0) { return; }
-   // All mesh element must be the same type:
-   MFEM_VERIFY(mesh.GetNumGeometries(mesh.Dimension()) == 1,
-               "All mesh elements must be the same type!");
-   const IntegrationRule &ir = qf.GetIntRule(0);
-   // All elements must use the same quadrature rule:
-   MFEM_VERIFY(qf.Size() == sdim*gf_vdim*ir.GetNPoints()*mesh.GetNE(),
-               "All mesh elements must use the same quadrature rule!");
-   // QuadratureFunction uses the layout qf_vdim x nq x ne, i.e.
-   // gf_vdim x sdim x nq x nq, so we need to request QVectorLayout::byVDIM:
-   GridFunc->GetGradients(ir, qf, QVectorLayout::byVDIM);
-}
-
 CurlGridFunctionCoefficient::CurlGridFunctionCoefficient(
   const GridFunction *gf)
   : VectorCoefficient(0)
@@ -1129,41 +1065,6 @@ real_t InnerProductCoefficient::Eval(ElementTransformation &T,
   return va * vb;
 }

-void InnerProductCoefficient::Project(QuadratureFunction &qf)
-{
-   MFEM_VERIFY(a->GetVDim() == b->GetVDim(),
-               "Incompatible vector coefficients: a->GetVDim(): "
-               << a->GetVDim() << ", b->GetVDim(): " << b->GetVDim());
-
-   const int vdim = a->GetVDim();
-   MFEM_VERIFY(vdim >= 1, "invalid vdim: " << vdim);
-
-   // When running on device, make sure the output data is allocated before any
-   // local temporary data to reduce potential heap fragmentation:
-   auto dot_d = qf.Write();
-
-   QuadratureFunction qf_a(qf.GetSpace(), vdim);
-   QuadratureFunction qf_b(qf.GetSpace(), vdim);
-
-   a->Project(qf_a);
-   b->Project(qf_b);
-
-   auto a_d = qf_a.Read();
-   auto b_d = qf_b.Read();
-
-   mfem::forall(qf.GetSpace()->GetSize(), [=] MFEM_HOST_DEVICE (int i)
-   {
-      const real_t *ai = a_d + i*vdim;
-      const real_t *bi = b_d + i*vdim;
-      real_t dot = ai[0]*bi[0];
-      for (int d = 1; d < vdim; d++)
-      {
-         dot += ai[d]*bi[d];
-      }
-      dot_d[i] = dot;
-   });
-}
-
 VectorRotProductCoefficient::VectorRotProductCoefficient(VectorCoefficient &A,
                                                         VectorCoefficient &B)
   : a(&A), b(&B), va(A.GetVDim()), vb(B.GetVDim())
@@ -132,9 +132,6 @@ public:
   /// Evaluate the coefficient.
   real_t Eval(ElementTransformation &T,
               const IntegrationPoint &ip) override;
-
-   /// Fill the QuadratureFunction @a qf with the piecewise constant values.
-   void Project(QuadratureFunction &qf) override;
 };

 /** @brief A piecewise coefficient with the pieces keyed off the element
@@ -897,9 +894,6 @@ public:
   void Eval(DenseMatrix &M, ElementTransformation &T,
             const IntegrationRule &ir) override;

-   /// @copydoc VectorCoefficient::Project(QuadratureFunction &)
-   void Project(QuadratureFunction &qf) override;
-
   virtual ~GradientGridFunctionCoefficient() { }
 };

@@ -1777,9 +1771,6 @@ public:
   /// Evaluate the coefficient at @a ip.
   real_t Eval(ElementTransformation &T,
               const IntegrationPoint &ip) override;
-
-   /// @copydoc Coefficient::Project(QuadratureFunction &)
-   void Project(QuadratureFunction &qf) override;
 };

 /// Scalar coefficient defined as a cross product of two vectors in the xy-plane.
@@ -0,0 +1,217 @@
+// 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 "complex_fem.hpp"
+#include "../general/forall.hpp"
+
+using namespace std;
+
+namespace mfem
+{
+
+real_t
+RealPartCoefficient::Eval(ElementTransformation &T,
+                          const IntegrationPoint &ip)
+{
+   complex_t val = complex_coef_.Eval(T, ip);
+   return val.real();
+}
+
+real_t
+ImagPartCoefficient::Eval(ElementTransformation &T,
+                          const IntegrationPoint &ip)
+{
+   complex_t val = complex_coef_.Eval(T, ip);
+   return val.imag();
+}
+
+RealPartVectorCoefficient::RealPartVectorCoefficient(ComplexVectorCoefficient &
+                                                     complex_vcoef)
+   : VectorCoefficient(complex_vcoef.GetVDim()),
+     complex_vcoef_(complex_vcoef),
+     val_(vdim)
+{}
+
+void
+RealPartVectorCoefficient::Eval(Vector &V, ElementTransformation &T,
+                                const IntegrationPoint &ip)
+{
+   complex_vcoef_.Eval(val_, T, ip);
+   V = val_.real();
+}
+
+ImagPartVectorCoefficient::ImagPartVectorCoefficient(ComplexVectorCoefficient &
+                                                     complex_vcoef)
+   : VectorCoefficient(complex_vcoef.GetVDim()),
+     complex_vcoef_(complex_vcoef),
+     val_(vdim)
+{}
+
+void
+ImagPartVectorCoefficient::Eval(Vector &V, ElementTransformation &T,
+                                const IntegrationPoint &ip)
+{
+   complex_vcoef_.Eval(val_, T, ip);
+   V = val_.imag();
+}
+
+RealPartMatrixCoefficient::RealPartMatrixCoefficient(ComplexMatrixCoefficient &
+                                                     complex_mcoef)
+   : MatrixCoefficient(complex_mcoef.GetHeight(), complex_mcoef.GetWidth()),
+     complex_mcoef_(complex_mcoef),
+     val_(height, width)
+{}
+
+void
+RealPartMatrixCoefficient::Eval(DenseMatrix &M, ElementTransformation &T,
+                                const IntegrationPoint &ip)
+{
+   complex_mcoef_.Eval(val_, T, ip);
+   M = val_.real();
+}
+
+ImagPartMatrixCoefficient::ImagPartMatrixCoefficient(ComplexMatrixCoefficient &
+                                                     complex_mcoef)
+   : MatrixCoefficient(complex_mcoef.GetHeight(), complex_mcoef.GetWidth()),
+     complex_mcoef_(complex_mcoef),
+     val_(height, width)
+{}
+
+void
+ImagPartMatrixCoefficient::Eval(DenseMatrix &M, ElementTransformation &T,
+                                const IntegrationPoint &ip)
+{
+   complex_mcoef_.Eval(val_, T, ip);
+   M = val_.imag();
+}
+
+ComplexCoefficient::ComplexCoefficient()
+   : time(0.),
+     re_part_coef_(*this), im_part_coef_(*this),
+     real_coef_(re_part_coef_), imag_coef_(im_part_coef_)
+{ }
+
+ComplexCoefficient::ComplexCoefficient(Coefficient &c_r,
+                                       Coefficient &c_i)
+   : time(c_r.GetTime()),
+     re_part_coef_(*this), im_part_coef_(*this),
+     real_coef_(c_r), imag_coef_(c_i)
+{
+   c_i.SetTime(time);
+}
+
+complex_t
+ComplexCoefficient::Eval(ElementTransformation &T,
+                         const IntegrationPoint &ip)
+{
+   // Avoid circular dependency
+   MFEM_VERIFY(std::addressof(real_coef_) != std::addressof(re_part_coef_) &&
+               std::addressof(imag_coef_) != std::addressof(im_part_coef_),
+               "Classes dervied from ComplexCoefficient must either "
+               "implement an Eval method or supply Coefficients "
+               "for both the real and imaginary parts of the field.");
+
+   return complex_t(real_coef_.Eval(T, ip), imag_coef_.Eval(T, ip));
+}
+
+ComplexVectorCoefficient::ComplexVectorCoefficient(VectorCoefficient &v_r,
+                                                   VectorCoefficient &v_i)
+   : vdim(v_r.GetVDim()), time(v_r.GetTime()),
+     re_part_vcoef_(*this), im_part_vcoef_(*this),
+     real_vcoef_(v_r), imag_vcoef_(v_i)
+{
+   MFEM_ASSERT(v_r.GetVDim() == v_i.GetVDim(), "ComplexVectorCoefficient"
+               " - incompatible vector dimensions of real and imaginary parts.");
+
+   v_i.SetTime(time);
+}
+
+void ComplexVectorCoefficient::Eval(ComplexVector &V, ElementTransformation &T,
+                                    const IntegrationPoint &ip)
+{
+   // Avoid circular dependency
+   MFEM_VERIFY(std::addressof(real_vcoef_) != std::addressof(re_part_vcoef_) &&
+               std::addressof(imag_vcoef_) != std::addressof(im_part_vcoef_),
+               "Classes dervied from ComplexVectorCoefficient must either "
+               "implement an Eval method or supply VectorCoefficients "
+               "for both the real and imaginary parts of the field.");
+
+   V_r_.SetSize(vdim);
+   V_i_.SetSize(vdim);
+
+   real_vcoef_.Eval(V_r_, T, ip);
+   imag_vcoef_.Eval(V_i_, T, ip);
+
+   V.Set(V_r_, V_i_);
+}
+
+ComplexConstantCoefficient::ComplexConstantCoefficient(
+   const complex_t z)
+   : val(z), real_coef(z.real()), imag_coef(z.imag())
+{
+   real_coef_ = real_coef;
+   imag_coef_ = imag_coef;
+}
+
+ComplexConstantCoefficient::ComplexConstantCoefficient(
+   real_t z_r, real_t z_i)
+   : real_coef(z_r), imag_coef(z_i)
+{
+   val = complex_t(z_r, z_i);
+
+   real_coef_ = real_coef;
+   imag_coef_ = imag_coef;
+}
+
+complex_t ComplexFunctionCoefficient::Eval(ElementTransformation & T,
+                                           const IntegrationPoint & ip)
+{
+   real_t x[3];
+   Vector transip(x, 3);
+
+   T.Transform(ip, transip);
+
+   if (Function)
+   {
+      return Function(transip);
+   }
+   else
+   {
+      return TDFunction(transip, GetTime());
+   }
+}
+
+void ComplexVectorFunctionCoefficient::Eval(ComplexVector &V,
+                                            ElementTransformation &T,
+                                            const IntegrationPoint &ip)
+{
+   real_t x[3];
+   Vector transip(x, 3);
+
+   T.Transform(ip, transip);
+
+   V.SetSize(vdim);
+   if (Function)
+   {
+      Function(transip, V);
+   }
+   else
+   {
+      TDFunction(transip, GetTime(), V);
+   }
+   if (Q)
+   {
+      V *= Q->Eval(T, ip, GetTime());
+   }
+}
+
+} // end namespace mfem
+
@@ -0,0 +1,523 @@
+// 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_COMPLEX_COEFFICIENT
+#define MFEM_COMPLEX_COEFFICIENT
+
+#include "../config/config.hpp"
+#include "../linalg/linalg.hpp"
+#include "coefficient.hpp"
+#include "intrules.hpp"
+#include "eltrans.hpp"
+
+namespace mfem
+{
+
+class ComplexCoefficient;
+class ComplexVectorCoefficient;
+class ComplexMatrixCoefficient;
+
+/// Standard Coefficient which returns the real part of a ComplexCoefficient
+class RealPartCoefficient : public Coefficient
+{
+private:
+   ComplexCoefficient &complex_coef_;
+
+public:
+   RealPartCoefficient(ComplexCoefficient & complex_coef)
+      : complex_coef_(complex_coef) {}
+
+   real_t Eval(ElementTransformation &T,
+               const IntegrationPoint &ip);
+};
+
+/// Standard Coefficient which returns the imaginary part of a
+/// ComplexCoefficient
+class ImagPartCoefficient : public Coefficient
+{
+private:
+   ComplexCoefficient &complex_coef_;
+
+public:
+   ImagPartCoefficient(ComplexCoefficient & complex_coef)
+      : complex_coef_(complex_coef) {}
+
+   real_t Eval(ElementTransformation &T,
+               const IntegrationPoint &ip);
+};
+
+typedef ImagPartCoefficient ImaginaryPartCoefficient;
+
+class RealPartVectorCoefficient : public VectorCoefficient
+{
+private:
+   ComplexVectorCoefficient &complex_vcoef_;
+   mutable ComplexVector val_;
+
+public:
+   RealPartVectorCoefficient(ComplexVectorCoefficient & complex_vcoef);
+
+   void Eval(Vector &V, ElementTransformation &T,
+             const IntegrationPoint &ip);
+};
+
+class ImagPartVectorCoefficient : public VectorCoefficient
+{
+private:
+   ComplexVectorCoefficient &complex_vcoef_;
+   mutable ComplexVector val_;
+
+public:
+   ImagPartVectorCoefficient(ComplexVectorCoefficient & complex_vcoef);
+
+   void Eval(Vector &V, ElementTransformation &T,
+             const IntegrationPoint &ip);
+};
+
+typedef ImagPartVectorCoefficient ImaginaryPartVectorCoefficient;
+
+class RealPartMatrixCoefficient : public MatrixCoefficient
+{
+private:
+   ComplexMatrixCoefficient &complex_mcoef_;
+   mutable ComplexTypeDenseMatrix val_;
+
+public:
+   RealPartMatrixCoefficient(ComplexMatrixCoefficient & complex_mcoef);
+
+   void Eval(DenseMatrix &M, ElementTransformation &T,
+             const IntegrationPoint &ip);
+};
+
+class ImagPartMatrixCoefficient : public MatrixCoefficient
+{
+private:
+   ComplexMatrixCoefficient &complex_mcoef_;
+   mutable ComplexTypeDenseMatrix val_;
+
+public:
+   ImagPartMatrixCoefficient(ComplexMatrixCoefficient & complex_mcoef);
+
+   void Eval(DenseMatrix &V, ElementTransformation &T,
+             const IntegrationPoint &ip);
+};
+
+typedef ImagPartMatrixCoefficient ImaginaryPartMatrixCoefficient;
+
+/** @brief Base class ComplexCoefficients that optionally depend on space and
+    time. These are used by the SesquilinearForm, ComplexLinearForm, and
+    ComplexGridFunction classes to represent the physical coefficients in
+    the PDEs that are being discretized. This class can also be used in a more
+    general way to represent functions that don't necessarily belong to a FE
+    space, e.g., to project onto ComplexGridFunctions to use as initial
+    conditions, exact solutions, etc. See, e.g., ex22 for these uses. */
+class ComplexCoefficient
+{
+protected:
+   real_t time;
+
+private:
+   RealPartCoefficient re_part_coef_;
+   ImagPartCoefficient im_part_coef_;
+
+protected:
+   Coefficient &real_coef_;
+   Coefficient &imag_coef_;
+
+public:
+
+   ComplexCoefficient();
+   ComplexCoefficient(Coefficient &c_r, Coefficient &c_i);
+
+   /// Set the time for time dependent coefficients
+   virtual void SetTime(real_t t)
+   { time = t; real_coef_.SetTime(t); imag_coef_.SetTime(t); }
+
+   /// Get the time for time dependent coefficients
+   real_t GetTime() { return time; }
+
+   /** @brief Evaluate the coefficient in the element described by @a T at the
+       point @a ip. */
+   /** @note When this method is called, the caller must make sure that the
+       IntegrationPoint associated with @a T is the same as @a ip. This can be
+       achieved by calling T.SetIntPoint(&ip). */
+   virtual complex_t Eval(ElementTransformation &T,
+                          const IntegrationPoint &ip);
+
+   /** @brief Evaluate the coefficient in the element described by @a T at the
+       point @a ip at time @a t. */
+   /** @note When this method is called, the caller must make sure that the
+       IntegrationPoint associated with @a T is the same as @a ip. This can be
+       achieved by calling T.SetIntPoint(&ip). */
+   complex_t Eval(ElementTransformation &T,
+                  const IntegrationPoint &ip, real_t t)
+   {
+      SetTime(t);
+      return Eval(T, ip);
+   }
+
+   /** @brief Access a standard Coefficient object reproducing the real part of
+       the complex-valued field */
+   /** @note By default this method returns an internal object which
+       computes the complex value using the above Eval method and
+       returns its real part. Custom implementations may choose to
+       override this method with a more efficient real-valued
+       coefficient. */
+   virtual Coefficient & real() { return real_coef_; }
+
+   /** @brief Access a standard Coefficient object reproducing the imaginary
+       part of the complex-valued field */
+   /** @note By default this method returns an internal object which
+       computes the complex value using the above Eval method and
+       returns its imaginary part. Custom implementations may choose to
+       override this method with a more efficient real-valued
+       coefficient. */
+   virtual Coefficient & imag() { return imag_coef_; }
+
+   virtual ~ComplexCoefficient() { }
+};
+
+/** @brief Base class ComplexVectorCoefficients that optionally depend
+    on space and time. These are used by the SesquilinearForm,
+    ComplexLinearForm, and ComplexGridFunction classes to represent
+    the physical vector-valued coefficients in the PDEs that are being
+    discretized. This class can also be used in a more general way to
+    represent functions that don't necessarily belong to a FE space,
+    e.g., to project onto ComplexGridFunctions to use as initial
+    conditions, exact solutions, etc. See, e.g., ex22 for these
+    uses. */
+class ComplexVectorCoefficient
+{
+protected:
+   int vdim;
+   real_t time;
+
+private:
+   RealPartVectorCoefficient re_part_vcoef_;
+   ImagPartVectorCoefficient im_part_vcoef_;
+
+protected:
+   VectorCoefficient &real_vcoef_;
+   VectorCoefficient &imag_vcoef_;
+
+   mutable Vector V_r_;
+   mutable Vector V_i_;
+
+public:
+   ComplexVectorCoefficient(int vd)
+      : vdim(vd), time(0.),
+        re_part_vcoef_(*this), im_part_vcoef_(*this),
+        real_vcoef_(re_part_vcoef_), imag_vcoef_(im_part_vcoef_)
+   { }
+
+   ComplexVectorCoefficient(VectorCoefficient &v_r, VectorCoefficient &v_i);
+
+
+   /// Set the time for time dependent coefficients
+   virtual void SetTime(real_t t)
+   { time = t; real_vcoef_.SetTime(t); imag_vcoef_.SetTime(t); }
+
+   /// Get the time for time dependent coefficients
+   real_t GetTime() { return time; }
+
+   /// Returns dimension of the vector.
+   int GetVDim() { return vdim; }
+
+   /** @brief Evaluate the vector coefficient in the element described by @a T
+       at the point @a ip, storing the result in @a V. */
+   /** @note When this method is called, the caller must make sure that the
+       IntegrationPoint associated with @a T is the same as @a ip. This can be
+       achieved by calling T.SetIntPoint(&ip). */
+   virtual void Eval(ComplexVector &V, ElementTransformation &T,
+                     const IntegrationPoint &ip);
+
+   /** @brief Evaluate the vector coefficient in the element described by @a T
+       at the point @a ip at time @a t, storing the result in @a V. */
+   /** @note When this method is called, the caller must make sure that the
+       IntegrationPoint associated with @a T is the same as @a ip. This can be
+       achieved by calling T.SetIntPoint(&ip). */
+   void Eval(ComplexVector &V, ElementTransformation &T,
+             const IntegrationPoint &ip, real_t t)
+   {
+      SetTime(t);
+      Eval(V, T, ip);
+   }
+
+   /** @brief Access a standard Coefficient object reproducing the real part of
+       the complex-valued field */
+   /** @note By default this method returns an internal object which
+       computes the complex value using the above Eval method and
+       returns its real part. Custom implementations may choose to
+       override this method with a more efficient real-valued
+       coefficient. */
+   virtual VectorCoefficient & real() { return real_vcoef_; }
+
+   /** @brief Access a standard Coefficient object reproducing the imaginary
+       part of the complex-valued field */
+   /** @note By default this method returns an internal object which
+       computes the complex value using the above Eval method and
+       returns its imaginary part. Custom implementations may choose to
+       override this method with a more efficient real-valued
+       coefficient. */
+   virtual VectorCoefficient & imag() { return imag_vcoef_; }
+
+   virtual ~ComplexVectorCoefficient() { }
+};
+
+/** @brief Base class ComplexMatrixCoefficients that optionally depend
+    on space and time. These are used by the SesquilinearForm,
+    ComplexLinearForm, and ComplexGridFunction classes to represent
+    the physical matrix-valued coefficients in the PDEs that are being
+    discretized. This class can also be used in a more general way to
+    represent functions that don't necessarily belong to a FE space.
+    See, e.g., ex22 for these uses. */
+class ComplexMatrixCoefficient
+{
+protected:
+   int height, width;
+   real_t time;
+
+private:
+   RealPartMatrixCoefficient re_part_mcoef_;
+   ImagPartMatrixCoefficient im_part_mcoef_;
+
+protected:
+   MatrixCoefficient &real_mcoef_;
+   MatrixCoefficient &imag_mcoef_;
+
+   mutable DenseMatrix M_r_;
+   mutable DenseMatrix M_i_;
+
+public:
+   /// Construct a dim x dim matrix coefficient.
+   explicit ComplexMatrixCoefficient(int dim)
+      : height(dim), width(dim), time(0.),
+        re_part_mcoef_(*this), im_part_mcoef_(*this),
+        real_mcoef_(re_part_mcoef_), imag_mcoef_(im_part_mcoef_)
+   { }
+
+   /// Construct a h x w matrix coefficient.
+   ComplexMatrixCoefficient(int h, int w) :
+      height(h), width(w), time(0.),
+      re_part_mcoef_(*this), im_part_mcoef_(*this),
+      real_mcoef_(re_part_mcoef_), imag_mcoef_(im_part_mcoef_)
+   { }
+
+   /// Set the time for time dependent coefficients
+   virtual void SetTime(real_t t) { time = t; }
+
+   /// Get the time for time dependent coefficients
+   real_t GetTime() { return time; }
+
+   /// Get the height of the matrix.
+   int GetHeight() const { return height; }
+
+   /// Get the width of the matrix.
+   int GetWidth() const { return width; }
+
+   /// For backward compatibility get the width of the matrix.
+   int GetVDim() const { return width; }
+
+   /** @brief Evaluate the matrix coefficient in the element described by @a T
+       at the point @a ip, storing the result in @a K. */
+   /** @note When this method is called, the caller must make sure that the
+       IntegrationPoint associated with @a T is the same as @a ip. This can be
+       achieved by calling T.SetIntPoint(&ip). */
+   virtual void Eval(ComplexTypeDenseMatrix &K, ElementTransformation &T,
+                     const IntegrationPoint &ip) = 0;
+
+   /** @brief Access a standard Coefficient object reproducing the real part of
+       the complex-valued field */
+   /** @note By default this method returns an internal object which
+       computes the complex value using the above Eval method and
+       returns its real part. Custom implementations may choose to
+       override this method with a more efficient real-valued
+       coefficient. */
+   virtual MatrixCoefficient & real() { return real_mcoef_; }
+
+   /** @brief Access a standard Coefficient object reproducing the imaginary
+       part of the complex-valued field */
+   /** @note By default this method returns an internal object which
+       computes the complex value using the above Eval method and
+       returns its imaginary part. Custom implementations may choose to
+       override this method with a more efficient real-valued
+       coefficient. */
+   virtual MatrixCoefficient & imag() { return imag_mcoef_; }
+
+   virtual ~ComplexMatrixCoefficient() { }
+};
+
+/// A complex-valued coefficient that is constant across space and time
+class ComplexConstantCoefficient : public ComplexCoefficient
+{
+private:
+   complex_t val;
+
+   ConstantCoefficient real_coef;
+   ConstantCoefficient imag_coef;
+
+public:
+   ComplexConstantCoefficient(const complex_t z);
+
+   ComplexConstantCoefficient(real_t z_r, real_t z_i = 0.);
+
+   complex_t Eval(ElementTransformation &T,
+                  const IntegrationPoint &ip) { return val; }
+};
+
+/// Complex-valued vector coefficient that is constant in space and time.
+class ComplexVectorConstantCoefficient : public ComplexVectorCoefficient
+{
+private:
+   ComplexVector vec;
+
+public:
+   /// Construct the coefficient with constant vector @a v.
+   ComplexVectorConstantCoefficient(const ComplexVector &v)
+      : ComplexVectorCoefficient(v.Size()), vec(v) { }
+
+   /// Construct the coefficient with constant vector @a v.
+   ComplexVectorConstantCoefficient(const Vector &v)
+      : ComplexVectorCoefficient(v.Size()), vec(v) { }
+
+   using ComplexVectorCoefficient::Eval;
+
+   ///  Evaluate the vector coefficient at @a ip.
+   void Eval(ComplexVector &V, ElementTransformation &T,
+             const IntegrationPoint &ip) override { V = vec; }
+
+   /// Return a reference to the constant vector in this class.
+   const ComplexVector& GetVec() const { return vec; }
+};
+
+/// Complex-valued vector coefficient that is constant in space and time.
+class ComplexMatrixConstantCoefficient : public ComplexMatrixCoefficient
+{
+private:
+   ComplexTypeDenseMatrix mat;
+
+public:
+   /// Construct the coefficient with constant vector @a v.
+   ComplexMatrixConstantCoefficient(const ComplexTypeDenseMatrix &m)
+      : ComplexMatrixCoefficient(m.Height(), m.Width()), mat(m) { }
+
+   /// Construct the coefficient with constant vector @a v.
+   ComplexMatrixConstantCoefficient(const DenseMatrix &m)
+      : ComplexMatrixCoefficient(m.Height(), m.Width()), mat(m) { }
+
+   using ComplexMatrixCoefficient::Eval;
+
+   ///  Evaluate the matrix coefficient at @a ip.
+   void Eval(ComplexTypeDenseMatrix &M, ElementTransformation &T,
+             const IntegrationPoint &ip) override { M = mat; }
+
+   /// Return a reference to the constant matrix in this class.
+   const ComplexTypeDenseMatrix& GetMat() const { return mat; }
+};
+
+/// A general complex-valued function coefficient
+class ComplexFunctionCoefficient : public ComplexCoefficient
+{
+protected:
+   std::function<complex_t(const Vector &)> Function;
+   std::function<complex_t(const Vector &, real_t)> TDFunction;
+
+public:
+   /// Define a time-independent coefficient from a std function
+   /** \param F time-independent std::function */
+   ComplexFunctionCoefficient(std::function<complex_t
+                              (const Vector &)> F)
+      : Function(std::move(F))
+   { }
+
+   /// Define a time-dependent coefficient from a std function
+   /** \param TDF time-dependent function */
+   ComplexFunctionCoefficient(std::function<complex_t
+                              (const Vector &, real_t)> TDF)
+      : TDFunction(std::move(TDF))
+   { }
+
+   /// (DEPRECATED) Define a time-independent coefficient from a C-function
+   /** @deprecated Use the method where the C-function, @a f, uses a const
+       Vector argument instead of Vector. */
+   MFEM_DEPRECATED ComplexFunctionCoefficient(complex_t
+                                              (*f)(Vector &))
+   {
+      // Cast first to (void*) to suppress a warning from newer version of
+      // Clang when using -Wextra.
+      Function = reinterpret_cast<complex_t(*)
+                 (const Vector&)>((void*)f);
+      TDFunction = NULL;
+   }
+
+   /// (DEPRECATED) Define a time-dependent coefficient from a C-function
+   /** @deprecated Use the method where the C-function, @a tdf, uses a const
+       Vector argument instead of Vector. */
+   MFEM_DEPRECATED ComplexFunctionCoefficient(complex_t
+                                              (*tdf)(Vector &, real_t))
+   {
+      Function = NULL;
+      // Cast first to (void*) to suppress a warning from newer version of
+      // Clang when using -Wextra.
+      TDFunction =
+         reinterpret_cast<complex_t(*)(const Vector&,
+                                       real_t)>((void*)tdf);
+   }
+
+   /// Evaluate the coefficient at @a ip.
+   complex_t Eval(ElementTransformation &T,
+                  const IntegrationPoint &ip) override;
+};
+
+/// A general vector function coefficient
+class ComplexVectorFunctionCoefficient : public ComplexVectorCoefficient
+{
+private:
+   std::function<void(const Vector &, ComplexVector &)> Function;
+   std::function<void(const Vector &, real_t, ComplexVector &)> TDFunction;
+   ComplexCoefficient *Q;
+
+public:
+   /// Define a time-independent complex-valued vector coefficient
+   /// from a std function
+   /** \param dim - the size of the vector
+       \param F - time-independent function
+       \param q - optional scalar Coefficient to scale the vector coefficient */
+   ComplexVectorFunctionCoefficient(int dim,
+                                    std::function<void(const Vector &,
+                                                       ComplexVector &)> F,
+                                    ComplexCoefficient *q = nullptr)
+      : ComplexVectorCoefficient(dim), Function(std::move(F)), Q(q)
+   { }
+
+   /// Define a time-dependent complex-valued vector coefficient from
+   /// a std function
+   /** \param dim - the size of the vector
+       \param TDF - time-dependent function
+       \param q - optional scalar ComplexCoefficient to scale the vector coefficient */
+   ComplexVectorFunctionCoefficient(int dim,
+                                    std::function<void(const Vector &, real_t,
+                                                       ComplexVector &)> TDF,
+                                    ComplexCoefficient *q = nullptr)
+      : ComplexVectorCoefficient(dim), TDFunction(std::move(TDF)), Q(q)
+   { }
+
+   using ComplexVectorCoefficient::Eval;
+   /// Evaluate the vector coefficient at @a ip.
+   void Eval(ComplexVector &V, ElementTransformation &T,
+             const IntegrationPoint &ip) override;
+
+   virtual ~ComplexVectorFunctionCoefficient() { }
+};
+
+} // end namespace mfem
+
+#endif
@@ -96,6 +96,23 @@ ComplexGridFunction::ProjectCoefficient(Coefficient &real_coeff,
   gfi->SyncAliasMemory(*this);
 }

+void
+ComplexGridFunction::ProjectCoefficient(Coefficient &real_coeff)
+{
+   gfr->SyncMemory(*this);
+   gfi->SyncMemory(*this);
+   gfr->ProjectCoefficient(real_coeff);
+   *gfi = 0.0;
+   gfr->SyncAliasMemory(*this);
+   gfi->SyncAliasMemory(*this);
+}
+
+void
+ComplexGridFunction::ProjectCoefficient(ComplexCoefficient &coeff)
+{
+   this->ProjectCoefficient(coeff.real(), coeff.imag());
+}
+
 void
 ComplexGridFunction::ProjectCoefficient(VectorCoefficient &real_vcoeff,
                                        VectorCoefficient &imag_vcoeff)
@@ -108,6 +125,23 @@ ComplexGridFunction::ProjectCoefficient(VectorCoefficient &real_vcoeff,
   gfi->SyncAliasMemory(*this);
 }

+void
+ComplexGridFunction::ProjectCoefficient(VectorCoefficient &real_vcoeff)
+{
+   gfr->SyncMemory(*this);
+   gfi->SyncMemory(*this);
+   gfr->ProjectCoefficient(real_vcoeff);
+   *gfi = 0.0;
+   gfr->SyncAliasMemory(*this);
+   gfi->SyncAliasMemory(*this);
+}
+
+void
+ComplexGridFunction::ProjectCoefficient(ComplexVectorCoefficient &vcoeff)
+{
+   this->ProjectCoefficient(vcoeff.real(), vcoeff.imag());
+}
+
 void
 ComplexGridFunction::ProjectBdrCoefficient(Coefficient &real_coeff,
                                           Coefficient &imag_coeff,
@@ -121,6 +155,26 @@ ComplexGridFunction::ProjectBdrCoefficient(Coefficient &real_coeff,
   gfi->SyncAliasMemory(*this);
 }

+void
+ComplexGridFunction::ProjectBdrCoefficient(Coefficient &real_coeff,
+                                           Array<int> &attr)
+{
+   ConstantCoefficient zero_coeff(0.0);
+   gfr->SyncMemory(*this);
+   gfi->SyncMemory(*this);
+   gfr->ProjectBdrCoefficient(real_coeff, attr);
+   gfi->ProjectBdrCoefficient(zero_coeff, attr);
+   gfr->SyncAliasMemory(*this);
+   gfi->SyncAliasMemory(*this);
+}
+
+void
+ComplexGridFunction::ProjectBdrCoefficient(ComplexCoefficient &coeff,
+                                           Array<int> &attr)
+{
+   this->ProjectBdrCoefficient(coeff.real(), coeff.imag(), attr);
+}
+
 void
 ComplexGridFunction::ProjectBdrCoefficientNormal(VectorCoefficient &real_vcoeff,
                                                 VectorCoefficient &imag_vcoeff,
@@ -134,6 +188,28 @@ ComplexGridFunction::ProjectBdrCoefficientNormal(VectorCoefficient &real_vcoeff,
   gfi->SyncAliasMemory(*this);
 }

+void
+ComplexGridFunction::ProjectBdrCoefficientNormal(VectorCoefficient &real_vcoeff,
+                                                 Array<int> &attr)
+{
+   Vector zero_vec(real_vcoeff.GetVDim()); zero_vec = 0.;
+   VectorConstantCoefficient zero_vcoeff(zero_vec);
+   gfr->SyncMemory(*this);
+   gfi->SyncMemory(*this);
+   gfr->ProjectBdrCoefficientNormal(real_vcoeff, attr);
+   gfi->ProjectBdrCoefficientNormal(zero_vcoeff, attr);
+   gfr->SyncAliasMemory(*this);
+   gfi->SyncAliasMemory(*this);
+}
+
+void
+ComplexGridFunction::ProjectBdrCoefficientNormal(
+   ComplexVectorCoefficient &vcoeff,
+   Array<int> &attr)
+{
+   this->ProjectBdrCoefficientNormal(vcoeff.real(), vcoeff.imag(), attr);
+}
+
 void
 ComplexGridFunction::ProjectBdrCoefficientTangent(VectorCoefficient
                                                  &real_vcoeff,
@@ -149,6 +225,80 @@ ComplexGridFunction::ProjectBdrCoefficientTangent(VectorCoefficient
   gfi->SyncAliasMemory(*this);
 }

+void
+ComplexGridFunction::ProjectBdrCoefficientTangent(VectorCoefficient
+                                                  &real_vcoeff,
+                                                  Array<int> &attr)
+{
+   Vector zero_vec(real_vcoeff.GetVDim()); zero_vec = 0.;
+   VectorConstantCoefficient zero_vcoeff(zero_vec);
+   gfr->SyncMemory(*this);
+   gfi->SyncMemory(*this);
+   gfr->ProjectBdrCoefficientTangent(real_vcoeff, attr);
+   gfi->ProjectBdrCoefficientTangent(zero_vcoeff, attr);
+   gfr->SyncAliasMemory(*this);
+   gfi->SyncAliasMemory(*this);
+}
+
+void
+ComplexGridFunction::ProjectBdrCoefficientTangent(
+   ComplexVectorCoefficient &vcoeff,
+   Array<int> &attr)
+{
+   this->ProjectBdrCoefficientTangent(vcoeff.real(), vcoeff.imag(), attr);
+}
+
+real_t
+ComplexGridFunction::ComputeL2Error(Coefficient &re_exsol,
+                                    Coefficient &im_exsol,
+                                    const IntegrationRule *irs[],
+                                    const Array<int> *elems) const
+{
+   real_t err_r = gfr->ComputeL2Error(re_exsol, irs, elems);
+   real_t err_i = gfi->ComputeL2Error(im_exsol, irs, elems);
+
+   return sqrt(err_r * err_r + err_i * err_i);
+}
+
+real_t
+ComplexGridFunction::ComputeL2Error(Coefficient &re_exsol,
+                                    const IntegrationRule *irs[],
+                                    const Array<int> *elems) const
+{
+   ConstantCoefficient zero_coef(0.0);
+
+   real_t err_r = gfr->ComputeL2Error(re_exsol, irs, elems);
+   real_t err_i = gfi->ComputeL2Error(zero_coef, irs, elems);
+
+   return sqrt(err_r * err_r + err_i * err_i);
+}
+
+real_t
+ComplexGridFunction::ComputeL2Error(VectorCoefficient &re_exsol,
+                                    VectorCoefficient &im_exsol,
+                                    const IntegrationRule *irs[],
+                                    const Array<int> *elems) const
+{
+   real_t err_r = gfr->ComputeL2Error(re_exsol, irs, elems);
+   real_t err_i = gfi->ComputeL2Error(im_exsol, irs, elems);
+
+   return sqrt(err_r * err_r + err_i * err_i);
+}
+
+real_t
+ComplexGridFunction::ComputeL2Error(VectorCoefficient &re_exsol,
+                                    const IntegrationRule *irs[],
+                                    const Array<int> *elems) const
+{
+   Vector zero_vec(re_exsol.GetVDim()); zero_vec = 0.0;
+   VectorConstantCoefficient zero_coef(zero_vec);
+
+   real_t err_r = gfr->ComputeL2Error(re_exsol, irs, elems);
+   real_t err_i = gfi->ComputeL2Error(zero_coef, irs, elems);
+
+   return sqrt(err_r * err_r + err_i * err_i);
+}
+

 ComplexLinearForm::ComplexLinearForm(FiniteElementSpace *fes,
                                     ComplexOperator::Convention convention)
@@ -731,6 +881,17 @@ ParComplexGridFunction::ProjectCoefficient(Coefficient &real_coeff,
   pgfi->SyncAliasMemory(*this);
 }

+void
+ParComplexGridFunction::ProjectCoefficient(Coefficient &real_coeff)
+{
+   pgfr->SyncMemory(*this);
+   pgfi->SyncMemory(*this);
+   pgfr->ProjectCoefficient(real_coeff);
+   *pgfi = 0.0;
+   pgfr->SyncAliasMemory(*this);
+   pgfi->SyncAliasMemory(*this);
+}
+
 void
 ParComplexGridFunction::ProjectCoefficient(VectorCoefficient &real_vcoeff,
                                           VectorCoefficient &imag_vcoeff)
@@ -743,6 +904,17 @@ ParComplexGridFunction::ProjectCoefficient(VectorCoefficient &real_vcoeff,
   pgfi->SyncAliasMemory(*this);
 }

+void
+ParComplexGridFunction::ProjectCoefficient(VectorCoefficient &real_vcoeff)
+{
+   pgfr->SyncMemory(*this);
+   pgfi->SyncMemory(*this);
+   pgfr->ProjectCoefficient(real_vcoeff);
+   *pgfi = 0.0;
+   pgfr->SyncAliasMemory(*this);
+   pgfi->SyncAliasMemory(*this);
+}
+
 void
 ParComplexGridFunction::ProjectBdrCoefficient(Coefficient &real_coeff,
                                              Coefficient &imag_coeff,
@@ -756,6 +928,19 @@ ParComplexGridFunction::ProjectBdrCoefficient(Coefficient &real_coeff,
   pgfi->SyncAliasMemory(*this);
 }

+void
+ParComplexGridFunction::ProjectBdrCoefficient(Coefficient &real_coeff,
+                                              Array<int> &attr)
+{
+   ConstantCoefficient zero_coeff(0.0);
+   pgfr->SyncMemory(*this);
+   pgfi->SyncMemory(*this);
+   pgfr->ProjectBdrCoefficient(real_coeff, attr);
+   pgfi->ProjectBdrCoefficient(zero_coeff, attr);
+   pgfr->SyncAliasMemory(*this);
+   pgfi->SyncAliasMemory(*this);
+}
+
 void
 ParComplexGridFunction::ProjectBdrCoefficientNormal(VectorCoefficient
                                                    &real_vcoeff,
@@ -771,6 +956,21 @@ ParComplexGridFunction::ProjectBdrCoefficientNormal(VectorCoefficient
   pgfi->SyncAliasMemory(*this);
 }

+void
+ParComplexGridFunction::ProjectBdrCoefficientNormal(VectorCoefficient
+                                                    &real_vcoeff,
+                                                    Array<int> &attr)
+{
+   Vector zero_vec(real_vcoeff.GetVDim()); zero_vec = 0.;
+   VectorConstantCoefficient zero_vcoeff(zero_vec);
+   pgfr->SyncMemory(*this);
+   pgfi->SyncMemory(*this);
+   pgfr->ProjectBdrCoefficientNormal(real_vcoeff, attr);
+   pgfi->ProjectBdrCoefficientNormal(zero_vcoeff, attr);
+   pgfr->SyncAliasMemory(*this);
+   pgfi->SyncAliasMemory(*this);
+}
+
 void
 ParComplexGridFunction::ProjectBdrCoefficientTangent(VectorCoefficient
                                                     &real_vcoeff,
@@ -786,6 +986,21 @@ ParComplexGridFunction::ProjectBdrCoefficientTangent(VectorCoefficient
   pgfi->SyncAliasMemory(*this);
 }

+void
+ParComplexGridFunction::ProjectBdrCoefficientTangent(VectorCoefficient
+                                                     &real_vcoeff,
+                                                     Array<int> &attr)
+{
+   Vector zero_vec(real_vcoeff.GetVDim()); zero_vec = 0.;
+   VectorConstantCoefficient zero_vcoeff(zero_vec);
+   pgfr->SyncMemory(*this);
+   pgfi->SyncMemory(*this);
+   pgfr->ProjectBdrCoefficientTangent(real_vcoeff, attr);
+   pgfi->ProjectBdrCoefficientTangent(zero_vcoeff, attr);
+   pgfr->SyncAliasMemory(*this);
+   pgfi->SyncAliasMemory(*this);
+}
+
 void
 ParComplexGridFunction::Distribute(const Vector *tv)
 {
@@ -825,6 +1040,31 @@ ParComplexGridFunction::ParallelProject(Vector &tv) const
   tvi.SyncAliasMemory(tv);
 }

+real_t
+ParComplexGridFunction::ComputeL2Error(Coefficient &exsolr,
+                                       const IntegrationRule *irs[],
+                                       Array<int> *elems) const
+{
+   ConstantCoefficient zeroCoef(0.0);
+
+   real_t err_r = pgfr->ComputeL2Error(exsolr, irs, elems);
+   real_t err_i = pgfi->ComputeL2Error(zeroCoef, irs, elems);
+   return sqrt(err_r * err_r + err_i * err_i);
+}
+
+real_t
+ParComplexGridFunction::ComputeL2Error(VectorCoefficient &exsolr,
+                                       const IntegrationRule *irs[],
+                                       Array<int> *elems) const
+{
+   Vector zeroVec(exsolr.GetVDim()); zeroVec = 0.0;
+   VectorConstantCoefficient zeroCoef(zeroVec);
+
+   real_t err_r = pgfr->ComputeL2Error(exsolr, irs, elems);
+   real_t err_i = pgfi->ComputeL2Error(zeroCoef, irs, elems);
+   return sqrt(err_r * err_r + err_i * err_i);
+}
+

 ParComplexLinearForm::ParComplexLinearForm(ParFiniteElementSpace *pfes,
                                           ComplexOperator::Convention
@@ -912,7 +912,7 @@ ConduitDataCollection::GridFunctionToBlueprintField(mfem::GridFunction *gf,

   if (vdim == 1) // scalar case
   {
-      n_field["values"].set_external(const_cast<real_t *>(gf->HostRead()),
+      n_field["values"].set_external(gf->GetData(),
                                     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(real_t) * entry_stride;
+      index_t stride = sizeof(double) * 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(const_cast<real_t *>(gf->HostRead()),
+         n_field["values"][comp_name].set_external(gf->GetData(),
                                                   ndofs,
                                                   offset,
                                                   stride);
-         offset +=  sizeof(real_t) * vdim_stride;
+         offset +=  sizeof(double) * vdim_stride;
      }
   }

@@ -1,266 +0,0 @@
-// 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
@@ -1,65 +0,0 @@
-// 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
@@ -231,53 +231,22 @@ public:
      const std::vector<FieldDescriptor> &parameters,
      const ParMesh &mesh);

-   /// MultLevel enum to indicate if the T->L Operators are used in the
-   /// Mult method.
-   enum MultLevel
-   {
-      TVECTOR,
-      LVECTOR
-   };
-
-   /// @brief Set the MultLevel mode for the DifferentiableOperator.
-   /// The default is TVECTOR, which means that the Operator will use
-   /// T->L before Mult and L->T Operators after.
-   void SetMultLevel(MultLevel level)
-   {
-      mult_level = level;
-   }
-
   /// @brief Compute the action of the operator on a given vector.
   ///
-   /// @param solutions_in The solution vector in which to compute the action.
-   /// This has to be a T-dof vector if MultLevel is set to TVECTOR, or L-dof
-   /// Vector if MultLevel is set to LVECTOR.
-   /// @param result_in Result vector of the action of the operator on
-   /// solutions. The result is a T-dof vector or L-dof vector depending on
-   /// the MultLevel.
-   void Mult(const Vector &solutions_in, Vector &result_in) const override
+   /// @param solutions_t The solution vector in which to compute the action.
+   /// This has to be a T-dof vector.
+   /// @param result_t Result vector of the action of the operator on
+   /// solutions_t. The result is a T-dof vector.
+   void Mult(const Vector &solutions_t, Vector &result_t) const override
   {
      MFEM_ASSERT(!action_callbacks.empty(), "no integrators have been set");
-
-      if (mult_level == MultLevel::LVECTOR)
+      prolongation(solutions, solutions_t, solutions_l);
+      residual_l = 0.0;
+      for (auto &action : action_callbacks)
      {
-         get_lvectors(solutions, solutions_in, solutions_l);
-         result_in = 0.0;
-         for (auto &action : action_callbacks)
-         {
-            action(solutions_l, parameters_l, result_in);
-         }
-      }
-      else
-      {
-         prolongation(solutions, solutions_in, solutions_l);
-         residual_l = 0.0;
-         for (auto &action : action_callbacks)
-         {
-            action(solutions_l, parameters_l, residual_l);
-         }
-         prolongation_transpose(residual_l, result_in);
+         action(solutions_l, parameters_l, residual_l);
      }
+      prolongation_transpose(residual_l, result_t);
   }

   /// @brief Add a domain integrator to the operator.
@@ -376,8 +345,6 @@ public:
 private:
   const ParMesh &mesh;

-   MultLevel mult_level = TVECTOR;
-
   std::vector<action_t> action_callbacks;
   std::map<size_t,
       std::vector<derivative_action_t>> derivative_action_callbacks;
@@ -387,6 +354,7 @@ private:
       std::vector<assemble_derivative_hypreparmatrix_callback_t>>
       assemble_derivative_hypreparmatrix_callbacks;

+
   std::vector<FieldDescriptor> solutions;
   std::vector<FieldDescriptor> parameters;
   // solutions and parameters
@@ -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, MPITypeMap<size_t>::mpi_type,
-              lengths.data(), 1, MPITypeMap<size_t>::mpi_type,
+   MPI_Gather(&msg_len, 1, MPI_INT,
+              lengths.data(), 1, MPI_INT,
              0, MPI_COMM_WORLD);

   if (myrank == 0)
@@ -568,7 +568,7 @@ struct ThreadBlocks
   int z = 1;
 };

-#if defined(MFEM_USE_CUDA_OR_HIP)
+#if (defined(MFEM_USE_CUDA) || defined(MFEM_USE_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_OR_HIP)
+#if (defined(MFEM_USE_CUDA) || defined(MFEM_USE_HIP))
      // int gridsize = (N + Z - 1) / Z;
      int num_bytes = num_shmem * sizeof(decltype(shmem));
      dim3 block_size(blocks.x, blocks.y, blocks.z);
@@ -1076,24 +1076,6 @@ void prolongation(const std::vector<FieldDescriptor> fields,
   }
 }

-inline
-void get_lvectors(const std::vector<FieldDescriptor> fields,
-                  const Vector &x,
-                  std::vector<Vector> &fields_l)
-{
-   int data_offset = 0;
-   for (std::size_t i = 0; i < fields.size(); i++)
-   {
-      const int sz = GetVSize(fields[i]);
-      fields_l[i].SetSize(sz);
-
-      const Vector x_i(const_cast<Vector&>(x), data_offset, sz);
-      fields_l[i] = x_i;
-
-      data_offset += sz;
-   }
-}
-
 /// @brief Get a transpose prolongation callback for a field descriptor.
 ///
 /// In the special case of a one field operator, the transpose prolongation
@@ -259,30 +259,6 @@ inline void FaceIdxToVolIdx3D(const int index, const int size1d,
   i = yz_plane ? level : _i;
 }

-MFEM_HOST_DEVICE
-inline int FaceIdxToVolIdx(int dim, int i, int size1d, int face0, int face1,
-                           int side, int orientation)
-{
-   if (dim == 2)
-   {
-      int ix, iy;
-      internal::FaceIdxToVolIdx2D(i, size1d, face0, face1, side, ix, iy);
-      return ix + iy*size1d;
-   }
-   else if (dim == 3)
-   {
-      int ix, iy, iz;
-      internal::FaceIdxToVolIdx3D(i, size1d, face0, face1, side, orientation,
-                                  ix, iy, iz);
-      return ix + size1d*iy + size1d*size1d*iz;
-   }
-   else
-   {
-      MFEM_ABORT_KERNEL("Invalid dimension");
-      return -1;
-   }
-};
-
 } // namespace internal

 } // namespace mfem
@@ -401,6 +401,9 @@ FiniteElementCollection *FiniteElementCollection::New(const char *name)
   {
      MFEM_ABORT("unknown FiniteElementCollection: " << name);
   }
+   MFEM_VERIFY(!strcmp(fec->Name(), name), "input name: \"" << name
+               << "\" does not match the created collection name: \""
+               << fec->Name() << '"');

   return fec;
 }
@@ -2456,7 +2459,8 @@ RT_FECollection::RT_FECollection(const int order, const int dim,
      const char *cb_name = BasisType::Name(cb_type); // this may abort
      MFEM_ABORT("unknown closed BasisType: " << cb_name);
   }
-   if (Quadrature1D::CheckOpen(op_type) == Quadrature1D::Invalid)
+   if (Quadrature1D::CheckOpen(op_type) == Quadrature1D::Invalid &&
+       ob_type != BasisType::IntegratedGLL)
   {
      const char *ob_name = BasisType::Name(ob_type); // this may abort
      MFEM_ABORT("unknown open BasisType: " << ob_name);
@@ -2514,7 +2518,6 @@ RT_FECollection::RT_FECollection(const int p, const int dim,
                                 const int map_type, const bool signs,
                                 const int ob_type)
   : FiniteElementCollection(p + 1)
-   , dim(dim)
   , ob_type(ob_type)
 {
   if (Quadrature1D::CheckOpen(BasisType::GetQuadrature1D(ob_type)) ==
@@ -2783,7 +2786,8 @@ ND_FECollection::ND_FECollection(const int p, const int dim,
   int cp_type = BasisType::GetQuadrature1D(cb_type);

   // Error checking
-   if (Quadrature1D::CheckOpen(op_type) == Quadrature1D::Invalid)
+   if (Quadrature1D::CheckOpen(op_type) == Quadrature1D::Invalid &&
+       ob_type != BasisType::IntegratedGLL)
   {
      const char *ob_name = BasisType::Name(ob_type);
      MFEM_ABORT("Invalid open basis point type: " << ob_name);
@@ -464,13 +464,6 @@ public:
   RT_Trace_FECollection(const int p, const int dim,
                         const int map_type = FiniteElement::INTEGRAL,
                         const int ob_type = BasisType::GaussLegendre);
-
-   FiniteElementCollection *Clone(int p) const override
-   {
-      const int map_type = (strncmp(rt_name, "RT_Trace", 8) == 0)?
-                           (FiniteElement::INTEGRAL):(FiniteElement::VALUE);
-      return new RT_Trace_FECollection(p, dim, map_type, ob_type);
-   }
 };

 /** Arbitrary order discontinuous finite elements defined on the interface
@@ -482,13 +475,6 @@ public:
   DG_Interface_FECollection(const int p, const int dim,
                             const int map_type = FiniteElement::VALUE,
                             const int ob_type = BasisType::GaussLegendre);
-
-   FiniteElementCollection *Clone(int p) const override
-   {
-      const int map_type = (strncmp(rt_name, "DG_Iface", 8) == 0)?
-                           (FiniteElement::VALUE):(FiniteElement::INTEGRAL);
-      return new DG_Interface_FECollection(p, dim, map_type, ob_type);
-   }
 };

 /// Arbitrary order H(curl)-conforming Nedelec finite elements.
@@ -1,249 +0,0 @@
-// 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__) || 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
@@ -17,9 +17,6 @@
 #include "fem.hpp"
 #include "ceed/interface/util.hpp"

-#include "derefmat_op.hpp"
-
-#include <algorithm>
 #include <cmath>
 #include <cstdarg>

@@ -27,9 +24,9 @@ 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();
@@ -43,9 +40,8 @@ void Ordering::DofsToVDofs<Ordering::byNODES>(int ndofs, int vdim,
   }
 }

-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();
@@ -59,6 +55,7 @@ void Ordering::DofsToVDofs<Ordering::byVDIM>(int ndofs, int vdim,
   }
 }

+
 FiniteElementSpace::FiniteElementSpace()
   : mesh(NULL), fec(NULL), vdim(0), ordering(Ordering::byNODES),
     ndofs(0), nvdofs(0), nedofs(0), nfdofs(0), nbdofs(0),
@@ -4247,11 +4244,7 @@ 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)
@@ -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.
@@ -246,7 +246,6 @@ 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).
@@ -683,12 +682,8 @@ public:
   NURBSExtension *GetNURBSext() { return NURBSext; }
   NURBSExtension *StealNURBSext();

-   bool Conforming() const
-   {
-      return NURBSext != NULL ||
-             (mesh->Conforming() && cP == NULL);
-   }
-   bool Nonconforming() const { return !Conforming(); }
+   bool Conforming() const { return mesh->Conforming() && cP == NULL; }
+   bool Nonconforming() const { return mesh->Nonconforming() || cP != NULL; }

   /** Set the prolongation operator of the space to an arbitrary sparse matrix,
       creating a copy of the argument. */
@@ -926,9 +921,6 @@ public:
   { return mesh->GetBdrElementType(i); }

   /// Returns ElementTransformation for the @a i-th element.
-   /// @note The returned pointer references an object owned by the associated
-   /// @a Mesh that will be modified by other calls to `GetElementTransformation`.
-   /// As such, this pointer should @b not be deleted by the caller.
   ElementTransformation *GetElementTransformation(int i) const
   { return mesh->GetElementTransformation(i); }

@@ -68,7 +68,7 @@ GridFunction::GridFunction(Mesh *m, std::istream &input)
      Vector::Load(input, fes->GetVSize());

      // if the mesh is a legacy (v1.1) NC mesh, it has old vertex ordering
-      if (fes->Nonconforming() && fes->GetMesh()->ncmesh &&
+      if (fes->Nonconforming() &&
          fes->GetMesh()->ncmesh->IsLegacyLoaded())
      {
         LegacyNCReorder();
@@ -1374,50 +1374,6 @@ void GridFunction::GetVectorGradientHat(
   MultAtB(loc_data_mat, dshape, gh);
 }

-void GridFunction::GetGradients(const IntegrationRule &ir, Vector &grad,
-                                QVectorLayout ql, MemoryType d_mt) const
-{
-   const FiniteElement &fe = *fes->GetTypicalFE();
-   const int dim  = fe.GetDim();
-   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();
-
-   // ql == QVectorLayout::byNODES :   NQ x VDIM x  DIM x NE
-   // ql == QVectorLayout::byVDIM  : VDIM x  DIM x NQPT x NE
-   grad.SetSize(dim*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);
-
-   // Pre-compute the geometric factors in order to set the desired MemoryType
-   // they use:
-   fes->GetMesh()->GetGeometricFactors(
-      ir, GeometricFactors::JACOBIANS, my_d_mt);
-
-   if (elem_restr) // currently, always true
-   {
-      Vector f_e(vdim*ND*NE, my_d_mt);
-      elem_restr->Mult(*this, f_e);
-      qi.PhysDerivatives(f_e, grad);
-   }
-   else
-   {
-      qi.PhysDerivatives(*this, grad);
-   }
-}
-
 real_t GridFunction::GetDivergence(ElementTransformation &T) const
 {
   DofTransformation doftrans;
@@ -2668,7 +2624,7 @@ void GridFunction::ProjectBdrCoefficient(Coefficient *coeff[],
   }
   for (int i = 0; i < values_counter.Size(); i++)
   {
-      MFEM_ASSERT(bool(values_counter[i]) == bool(ess_vdofs_marker[i]),
+      MFEM_ASSERT(bool(values_counter[i]) == ess_vdofs_marker[i],
                  "internal error");
   }
 #endif
@@ -153,8 +153,7 @@ public:
   /// Shortcut for calling SetFromTrueDofs() with GetTrueVector() as argument.
   void SetFromTrueVector() { SetFromTrueDofs(GetTrueVector()); }

-   /** @brief Returns the values at the vertices of element @a i for the 1-based
-       dimension vdim. */
+   /// Returns the values in the vertices of i'th element for dimension vdim.
   void GetNodalValues(int i, Array<real_t> &nval, int vdim = 1) const;

   /** @name Element index Get Value Methods
@@ -309,8 +308,7 @@ public:
   /// For a vector grid function, makes sure that the ordering is byNODES.
   void ReorderByNodes();

-   /** @brief Returns the values as a vector at mesh vertices, for the 1-based
-       dimension vdim. */
+   /// Return the values as a vector on mesh vertices for dimension vdim.
   void GetNodalValues(Vector &nval, int vdim = 1) const;

   void GetVectorFieldNodalValues(Vector &val, int comp) const;
@@ -361,33 +359,6 @@ public:
       variable. */
   void GetVectorGradientHat(ElementTransformation &T, DenseMatrix &gh) const;

-   /** @brief Evaluate the gradients of the GridFunction at the given quadrature
-       points, @a ir, in all mesh elements. */
-   /** This method assumes that all mesh elements are the same type and that the
-       IntegrationRule @a ir is consistent with that type of element.
-
-       @param[in] ir     Quadrature points at which the gradients are to be
-                         evaluated.
-       @param[out] grad  Output vector of size `SDIM*VDIM*NQ*NE` where `SDIM` is
-                         the spatial dimention of the mesh, `VDIM` is the vector
-                         dimension of the GridFunction, `NQ` is the number of
-                         quadrature points in @a ir, and `NE` is the number of
-                         elements in the mesh. The layout of @a grad is
-                         determined by the parameter @a ql: when @a ql is
-                         QVectorLayout::byNODES, the layout is
-                         `NQ x VDIM x SDIM x NE`; when @a ql is
-                         QVectorLayout::byVDIM, the layout is
-                         `VDIM x SDIM x NQ x NE`.
-       @param[in] ql     Determines the layout of the output vector @a grad; see
-                         the description of @a grad for details.
-       @param[in] d_mt   MemoryType to use for allocating the output vector
-                         @a grad, as well the GeometricFactors and temporary
-                         vector used by the method. By default, the current
-                         device memory type is used. */
-   void GetGradients(const IntegrationRule &ir, Vector &grad,
-                     QVectorLayout ql = QVectorLayout::byNODES,
-                     MemoryType d_mt = MemoryType::DEFAULT) const;
-
   /** Compute $ (\int_{\Omega} (*this) \psi_i)/(\int_{\Omega} \psi_i) $,
       where $ \psi_i $ are the basis functions for the FE space of avgs.
       Both FE spaces should be scalar and on the same mesh. */
@@ -85,9 +85,9 @@ namespace mfem
 {

 FindPointsGSLIB::FindPointsGSLIB()
-   : mesh(nullptr),
-     fec_map_lin(nullptr),
-     fdataD(nullptr), cr(nullptr), gsl_comm(nullptr),
+   : mesh(NULL),
+     fec_map_lin(NULL),
+     fdataD(NULL), cr(NULL), gsl_comm(NULL),
     dim(-1), points_cnt(-1), setupflag(false), default_interp_value(0),
     avgtype(AvgType::ARITHMETIC), bdr_tol(1e-8)
 {
@@ -97,10 +97,10 @@ FindPointsGSLIB::FindPointsGSLIB()
   gf_rst_map.SetSize(4);
   for (int i = 0; i < mesh_split.Size(); i++)
   {
-      mesh_split[i] = nullptr;
-      ir_split[i] = nullptr;
-      fes_rst_map[i] = nullptr;
-      gf_rst_map[i] = nullptr;
+      mesh_split[i] = NULL;
+      ir_split[i] = NULL;
+      fes_rst_map[i] = NULL;
+      gf_rst_map[i] = NULL;
   }

   gsl_comm = new gslib::comm;
@@ -117,40 +117,27 @@ FindPointsGSLIB::FindPointsGSLIB()
   crystal_init(cr, gsl_comm);
 }

-FindPointsGSLIB::FindPointsGSLIB(Mesh &mesh_in, const double bb_t,
-                                 const double newt_tol, const int npt_max)
-   : FindPointsGSLIB()
-{
-   Setup(mesh_in, bb_t, newt_tol, npt_max);
-}
-
 FindPointsGSLIB::~FindPointsGSLIB()
 {
-   FreeData();
-#ifdef MFEM_USE_MPI
-   if (!Mpi::IsFinalized())  // currently segfaults inside gslib otherwise
-#endif
+   crystal_free(cr);
+   comm_free(gsl_comm);
+   delete gsl_comm;
+   delete cr;
+   for (int i = 0; i < 4; i++)
   {
-      crystal_free(cr);
-      comm_free(gsl_comm);
-      delete gsl_comm;
-      delete cr;
+      if (mesh_split[i]) { delete mesh_split[i]; mesh_split[i] = NULL; }
+      if (ir_split[i]) { delete ir_split[i]; ir_split[i] = NULL; }
+      if (fes_rst_map[i]) { delete fes_rst_map[i]; fes_rst_map[i] = NULL; }
+      if (gf_rst_map[i]) { delete gf_rst_map[i]; gf_rst_map[i] = NULL; }
   }
-   for (int i = 0; i < mesh_split.Size(); i++)
-   {
-      if (mesh_split[i]) { delete mesh_split[i]; mesh_split[i] = nullptr; }
-      if (ir_split[i]) { delete ir_split[i]; ir_split[i] = nullptr; }
-      if (fes_rst_map[i]) { delete fes_rst_map[i]; fes_rst_map[i] = nullptr; }
-      if (gf_rst_map[i]) { delete gf_rst_map[i]; gf_rst_map[i] = nullptr; }
-   }
-   if (fec_map_lin) { delete fec_map_lin; fec_map_lin = nullptr; }
+   if (fec_map_lin) { delete fec_map_lin; fec_map_lin = NULL; }
 }

 #ifdef MFEM_USE_MPI
 FindPointsGSLIB::FindPointsGSLIB(MPI_Comm comm_)
-   : mesh(nullptr),
-     fec_map_lin(nullptr),
-     fdataD(nullptr), cr(nullptr), gsl_comm(nullptr),
+   : mesh(NULL),
+     fec_map_lin(NULL),
+     fdataD(NULL), cr(NULL), gsl_comm(NULL),
     dim(-1), points_cnt(-1), setupflag(false), default_interp_value(0),
     avgtype(AvgType::ARITHMETIC), bdr_tol(1e-8)
 {
@@ -160,10 +147,10 @@ FindPointsGSLIB::FindPointsGSLIB(MPI_Comm comm_)
   gf_rst_map.SetSize(4);
   for (int i = 0; i < mesh_split.Size(); i++)
   {
-      mesh_split[i] = nullptr;
-      ir_split[i] = nullptr;
-      fes_rst_map[i] = nullptr;
-      gf_rst_map[i] = nullptr;
+      mesh_split[i] = NULL;
+      ir_split[i] = NULL;
+      fes_rst_map[i] = NULL;
+      gf_rst_map[i] = NULL;
   }

   gsl_comm = new gslib::comm;
@@ -171,21 +158,12 @@ FindPointsGSLIB::FindPointsGSLIB(MPI_Comm comm_)
   comm_init(gsl_comm, comm_);
   crystal_init(cr, gsl_comm);
 }
-
-FindPointsGSLIB::FindPointsGSLIB(ParMesh &mesh_in, const double bb_t,
-                                 const double newt_tol, const int npt_max)
-   : FindPointsGSLIB(mesh_in.GetComm())
-{
-   Setup(mesh_in, bb_t, newt_tol, npt_max);
-}
 #endif

 void FindPointsGSLIB::Setup(Mesh &m, const double bb_t, const double newt_tol,
                            const int npt_max)
 {
   MFEM_VERIFY(m.GetNodes() != NULL, "Mesh nodes are required.");
-   MFEM_VERIFY(m.SpaceDimension() == m.Dimension(),
-               "Mesh spatial dimension and reference element dimension must be the same");
   const int meshOrder = m.GetNodes()->FESpace()->GetMaxElementOrder();

   // call FreeData if FindPointsGSLIB::Setup has been called already
@@ -193,9 +171,37 @@ void FindPointsGSLIB::Setup(Mesh &m, const double bb_t, const double newt_tol,

   mesh = &m;
   dim  = mesh->Dimension();
-   const unsigned int dof1D = meshOrder+1;
+   unsigned dof1D = meshOrder + 1;

-   SetupSplitMeshesAndIntegrationRules(meshOrder);
+   SetupSplitMeshes();
+   if (dim == 2)
+   {
+      if (ir_split[0]) { delete ir_split[0]; ir_split[0] = NULL; }
+      ir_split[0] = new IntegrationRule(3*pow(dof1D, dim));
+      SetupIntegrationRuleForSplitMesh(mesh_split[0], ir_split[0], meshOrder);
+
+      if (ir_split[1]) { delete ir_split[1]; ir_split[1] = NULL; }
+      ir_split[1] = new IntegrationRule(pow(dof1D, dim));
+      SetupIntegrationRuleForSplitMesh(mesh_split[1], ir_split[1], meshOrder);
+   }
+   else if (dim == 3)
+   {
+      if (ir_split[0]) { delete ir_split[0]; ir_split[0] = NULL; }
+      ir_split[0] = new IntegrationRule(pow(dof1D, dim));
+      SetupIntegrationRuleForSplitMesh(mesh_split[0], ir_split[0], meshOrder);
+
+      if (ir_split[1]) { delete ir_split[1]; ir_split[1] = NULL; }
+      ir_split[1] = new IntegrationRule(4*pow(dof1D, dim));
+      SetupIntegrationRuleForSplitMesh(mesh_split[1], ir_split[1], meshOrder);
+
+      if (ir_split[2]) { delete ir_split[2]; ir_split[2] = NULL; }
+      ir_split[2] = new IntegrationRule(3*pow(dof1D, dim));
+      SetupIntegrationRuleForSplitMesh(mesh_split[2], ir_split[2], meshOrder);
+
+      if (ir_split[3]) { delete ir_split[3]; ir_split[3] = NULL; }
+      ir_split[3] = new IntegrationRule(8*pow(dof1D, dim));
+      SetupIntegrationRuleForSplitMesh(mesh_split[3], ir_split[3], meshOrder);
+   }

   GetNodalValues(mesh->GetNodes(), gsl_mesh);

@@ -1122,18 +1128,13 @@ void FindPointsGSLIB::Interpolate(Mesh &m, const Vector &point_pos,
 void FindPointsGSLIB::FreeData()
 {
   if (!setupflag) { return; }
-#ifdef MFEM_USE_MPI
-   if (!Mpi::IsFinalized())  // currently segfaults inside gslib otherwise
-#endif
+   if (dim == 2)
   {
-      if (dim == 2)
-      {
-         findpts_free_2((gslib::findpts_data_2 *)this->fdataD);
-      }
-      else
-      {
-         findpts_free_3((gslib::findpts_data_3 *)this->fdataD);
-      }
+      findpts_free_2((gslib::findpts_data_2 *)this->fdataD);
+   }
+   else
+   {
+      findpts_free_3((gslib::findpts_data_3 *)this->fdataD);
   }
   gsl_code.DeleteAll();
   gsl_proc.DeleteAll();
@@ -1157,8 +1158,8 @@ void FindPointsGSLIB::FreeData()

 void FindPointsGSLIB::SetupSplitMeshes()
 {
-   if (fec_map_lin == nullptr) { fec_map_lin = new H1_FECollection(1, dim); }
-   if (dim == 2)
+   fec_map_lin = new H1_FECollection(1, dim);
+   if (mesh->Dimension() == 2)
   {
      int Nvert = 7;
      int NEsplit = 3;
@@ -1200,7 +1201,7 @@ void FindPointsGSLIB::SetupSplitMeshes()
      mesh_split[1] = new Mesh(Mesh::MakeCartesian2D(1, 1,
                                                     Element::QUADRILATERAL));
   }
-   else if (dim == 3)
+   else if (mesh->Dimension() == 3)
   {
      mesh_split[0] = new Mesh(Mesh::MakeCartesian3D(1, 1, 1,
                                                     Element::HEXAHEDRON));
@@ -1345,6 +1346,41 @@ void FindPointsGSLIB::SetupSplitMeshes()
         }
      }
   }
+
+   NE_split_total = 0;
+   split_element_map.SetSize(0);
+   split_element_index.SetSize(0);
+   int NEsplit = 0;
+   for (int e = 0; e < mesh->GetNE(); e++)
+   {
+      const Geometry::Type gt   = mesh->GetElement(e)->GetGeometryType();
+      if (gt == Geometry::TRIANGLE || gt == Geometry::PRISM)
+      {
+         NEsplit = 3;
+      }
+      else if (gt == Geometry::TETRAHEDRON)
+      {
+         NEsplit = 4;
+      }
+      else if (gt == Geometry::PYRAMID)
+      {
+         NEsplit = 8;
+      }
+      else if (gt == Geometry::SQUARE || gt == Geometry::CUBE)
+      {
+         NEsplit = 1;
+      }
+      else
+      {
+         MFEM_ABORT("Unsupported geometry type.");
+      }
+      NE_split_total += NEsplit;
+      for (int i = 0; i < NEsplit; i++)
+      {
+         split_element_map.Append(e);
+         split_element_index.Append(i);
+      }
+   }
 }

 void FindPointsGSLIB::SetupIntegrationRuleForSplitMesh(Mesh *meshin,
@@ -1395,79 +1431,6 @@ void FindPointsGSLIB::SetupIntegrationRuleForSplitMesh(Mesh *meshin,
   }
 }

-void FindPointsGSLIB::SetupSplitMeshesAndIntegrationRules(const int order)
-{
-   MFEM_VERIFY(mesh, "Setup FindPointsGSLIB with mesh first.");
-   const int dof1D = order+1;
-   const int dim = mesh->Dimension();
-
-   SetupSplitMeshes();
-   if (dim == 2)
-   {
-      if (ir_split[0]) { delete ir_split[0]; ir_split[0] = NULL; }
-      ir_split[0] = new IntegrationRule(3*pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[0], ir_split[0], order);
-
-      if (ir_split[1]) { delete ir_split[1]; ir_split[1] = NULL; }
-      ir_split[1] = new IntegrationRule(pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[1], ir_split[1], order);
-   }
-   else if (dim == 3)
-   {
-      if (ir_split[0]) { delete ir_split[0]; ir_split[0] = NULL; }
-      ir_split[0] = new IntegrationRule(pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[0], ir_split[0], order);
-
-      if (ir_split[1]) { delete ir_split[1]; ir_split[1] = NULL; }
-      ir_split[1] = new IntegrationRule(4*pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[1], ir_split[1], order);
-
-      if (ir_split[2]) { delete ir_split[2]; ir_split[2] = NULL; }
-      ir_split[2] = new IntegrationRule(3*pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[2], ir_split[2], order);
-
-      if (ir_split[3]) { delete ir_split[3]; ir_split[3] = NULL; }
-      ir_split[3] = new IntegrationRule(8*pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[3], ir_split[3], order);
-   }
-
-   // Setup map for non tensor-product elements
-   NE_split_total = 0;
-   split_element_map.SetSize(0);
-   split_element_index.SetSize(0);
-   int NEsplit = 0;
-   for (int e = 0; e < mesh->GetNE(); e++)
-   {
-      const Geometry::Type gt   = mesh->GetElement(e)->GetGeometryType();
-      if (gt == Geometry::TRIANGLE || gt == Geometry::PRISM)
-      {
-         NEsplit = 3;
-      }
-      else if (gt == Geometry::TETRAHEDRON)
-      {
-         NEsplit = 4;
-      }
-      else if (gt == Geometry::PYRAMID)
-      {
-         NEsplit = 8;
-      }
-      else if (gt == Geometry::SQUARE || gt == Geometry::CUBE)
-      {
-         NEsplit = 1;
-      }
-      else
-      {
-         MFEM_ABORT("Unsupported geometry type.");
-      }
-      NE_split_total += NEsplit;
-      for (int i = 0; i < NEsplit; i++)
-      {
-         split_element_map.Append(e);
-         split_element_index.Append(i);
-      }
-   }
-}
-
 void FindPointsGSLIB::GetNodalValues(const GridFunction *gf_in,
                                     Vector &node_vals)
 {
@@ -2118,19 +2081,6 @@ void FindPointsGSLIB::InterpolateGeneral(const GridFunction &field_in,
   } // parallel
 }

-Array<unsigned int> FindPointsGSLIB::GetPointsNotFoundIndices() const
-{
-   Array<unsigned int> nf_idxs;
-   for (int i = 0; i < gsl_code.Size(); i++)
-   {
-      if (gsl_code[i] == 2)
-      {
-         nf_idxs.Append(i);
-      }
-   }
-   return nf_idxs;
-}
-
 void FindPointsGSLIB::DistributePointInfoToOwningMPIRanks(
   Array<unsigned int> &recv_elem, Vector &recv_ref,
   Array<unsigned int> &recv_code)
@@ -2436,10 +2386,6 @@ void OversetFindPointsGSLIB::Setup(Mesh &m, const int meshid,
 {
   MFEM_VERIFY(m.GetNodes() != NULL, "Mesh nodes are required.");
   const int meshOrder = m.GetNodes()->FESpace()->GetMaxElementOrder();
-   const int gfOrder = gfmax ? gfmax->FESpace()->GetMaxElementOrder() :
-                       meshOrder;
-   MFEM_VERIFY(meshOrder == gfOrder,
-               "Mesh order must match gfmax order in OversetFindPointsGSLIB.");

   // FreeData if OversetFindPointsGSLIB::Setup has been called already
   if (setupflag) { FreeData(); }
@@ -2449,7 +2395,35 @@ void OversetFindPointsGSLIB::Setup(Mesh &m, const int meshid,
   const FiniteElement *fe = mesh->GetNodalFESpace()->GetTypicalFE();
   unsigned dof1D = fe->GetOrder() + 1;

-   SetupSplitMeshesAndIntegrationRules(meshOrder);
+   SetupSplitMeshes();
+   if (dim == 2)
+   {
+      if (ir_split[0]) { delete ir_split[0]; ir_split[0] = NULL; }
+      ir_split[0] = new IntegrationRule(3*pow(dof1D, dim));
+      SetupIntegrationRuleForSplitMesh(mesh_split[0], ir_split[0], meshOrder);
+
+      if (ir_split[1]) { delete ir_split[1]; ir_split[1] = NULL; }
+      ir_split[1] = new IntegrationRule(pow(dof1D, dim));
+      SetupIntegrationRuleForSplitMesh(mesh_split[1], ir_split[1], meshOrder);
+   }
+   else if (dim == 3)
+   {
+      if (ir_split[0]) { delete ir_split[0]; ir_split[0] = NULL; }
+      ir_split[0] = new IntegrationRule(pow(dof1D, dim));
+      SetupIntegrationRuleForSplitMesh(mesh_split[0], ir_split[0], meshOrder);
+
+      if (ir_split[1]) { delete ir_split[1]; ir_split[1] = NULL; }
+      ir_split[1] = new IntegrationRule(4*pow(dof1D, dim));
+      SetupIntegrationRuleForSplitMesh(mesh_split[1], ir_split[1], meshOrder);
+
+      if (ir_split[2]) { delete ir_split[2]; ir_split[2] = NULL; }
+      ir_split[2] = new IntegrationRule(3*pow(dof1D, dim));
+      SetupIntegrationRuleForSplitMesh(mesh_split[2], ir_split[2], meshOrder);
+
+      if (ir_split[3]) { delete ir_split[3]; ir_split[3] = NULL; }
+      ir_split[3] = new IntegrationRule(8*pow(dof1D, dim));
+      SetupIntegrationRuleForSplitMesh(mesh_split[3], ir_split[3], meshOrder);
+   }

   GetNodalValues(mesh->GetNodes(), gsl_mesh);

@@ -2506,7 +2480,7 @@ void OversetFindPointsGSLIB::FindPoints(const Vector &point_pos,
 {
   MFEM_VERIFY(setupflag, "Use OversetFindPointsGSLIB::Setup before "
               "finding points.");
-   MFEM_VERIFY(overset, "Please use OversetFindPoints for overlapping grids.");
+   MFEM_VERIFY(overset, "Please setup FindPoints for overlapping grids.");
   points_cnt = point_pos.Size() / dim;
   unsigned int match = 0; // Don't find points in the mesh if point_id=mesh_id

@@ -13,11 +13,7 @@
 #define MFEM_GSLIB

 #include "../config/config.hpp"
-#ifdef MFEM_USE_MPI
-#include "pgridfunc.hpp"
-#else
 #include "gridfunc.hpp"
-#endif

 #ifdef MFEM_USE_GSLIB

@@ -135,10 +131,6 @@ protected:
                                                 IntegrationRule *irule,
                                                 int order);

-   /// Helper function that calls \ref SetupSplitMeshes and
-   /// \ref SetupIntegrationRuleForSplitMesh.
-   virtual void SetupSplitMeshesAndIntegrationRules(const int order);
-
   /// Get GridFunction value at the points expected by GSLIB.
   virtual void GetNodalValues(const GridFunction *gf_in, Vector &node_vals);

@@ -198,23 +190,14 @@ protected:
   void InterpolateOnDevice(const Vector &field_in_evec, Vector &field_out,
                            const int nel, const int ncomp,
                            const int dof1dsol, const int ordering);
-
 public:
   FindPointsGSLIB();
-   FindPointsGSLIB(Mesh &mesh_in, const double bb_t = 0.1,
-                   const double newt_tol = 1.0e-12,
-                   const int npt_max = 256);

 #ifdef MFEM_USE_MPI
   FindPointsGSLIB(MPI_Comm comm_);
-   FindPointsGSLIB(ParMesh &mesh_in, const double bb_t = 0.1,
-                   const double newt_tol = 1.0e-12,
-                   const int npt_max = 256);
 #endif

   virtual ~FindPointsGSLIB();
-   FindPointsGSLIB(const FindPointsGSLIB&) = delete;
-   FindPointsGSLIB& operator=(const FindPointsGSLIB&) = delete;

   /** Initializes the internal mesh in gslib, by sending the positions of the
       Gauss-Lobatto nodes of the input Mesh object \p m.
@@ -229,8 +212,8 @@ public:
       @param[in] npt_max   (Optional) Number of points for simultaneous
                            iteration. This alters performance and
                            memory footprint.*/
-
-   void Setup(Mesh &m, const double bb_t = 0.1, const double newt_tol = 1.0e-12,
+   void Setup(Mesh &m, const double bb_t = 0.1,
+              const double newt_tol = 1.0e-12,
              const int npt_max = 256);
   /** Searches positions given in physical space by \p point_pos.
       These positions can be ordered byNodes: (XXX...,YYY...,ZZZ) or
@@ -306,12 +289,7 @@ public:

   /** Cleans up memory allocated internally by gslib.
       Note that in parallel, this must be called before MPI_Finalize(), as it
-       calls MPI_Comm_free() for internal gslib communicators. FreeData is
-       also called by the class destructor and there are no memory leaks if the
-       destructor is called before MPI_Finalize(). If the destructor is called
-       after MPI_Finalize(), there will be an error because gslib will try to
-       invoke some MPI functions.
-   */
+       calls MPI_Comm_free() for internal gslib communicators. */
   virtual void FreeData();

   /// Return code for each point searched by FindPoints: inside element (0), on
@@ -334,9 +312,6 @@ public:
   /// point found by FindPoints.
   virtual const Vector &GetGSLIBReferencePosition() const { return gsl_ref; }

-   /// Get array of indices of not-found points.
-   Array<unsigned int> GetPointsNotFoundIndices() const;
-
   /** @name Methods to support a custom interpolation procedure.
       \brief The physical-space point that the user seeks to interpolate at
       could be located inside an element on another mpi rank.
@@ -181,7 +181,7 @@ void HyperbolicFormIntegrator::AssembleFaceVector(
   // current elements' the number of degrees of freedom
   // does not consider the number of equations
   const int dof1 = el1.GetDof();
-   const int dof2 = (Tr.Elem2No >= 0)?(el2.GetDof()):(0);
+   const int dof2 = el2.GetDof();

 #ifdef MFEM_THREAD_SAFE
   // Local storage for element integration
@@ -219,9 +219,7 @@ void HyperbolicFormIntegrator::AssembleFaceVector(
   const IntegrationRule *ir = IntRule;
   if (!ir)
   {
-      const int max_el_order = dof2 ? std::max(el1.GetOrder(),
-                                               el2.GetOrder()) : el1.GetOrder();
-      const int order = 2*max_el_order + IntOrderOffset;
+      const int order = 2*std::max(el1.GetOrder(), el2.GetOrder()) + IntOrderOffset;
      ir = &IntRules.Get(Tr.GetGeometryType(), order);
   }
   // loop over integration points
@@ -233,22 +231,18 @@ void HyperbolicFormIntegrator::AssembleFaceVector(

      // Calculate basis functions on both elements at the face
      el1.CalcShape(Tr.GetElement1IntPoint(), shape1);
+      el2.CalcShape(Tr.GetElement2IntPoint(), shape2);

      // Interpolate elfun at the point
      elfun1_mat.MultTranspose(shape1, state1);
-
-      if (dof2)
-      {
-         // Calculate basis functions on both elements at the face
-         el2.CalcShape(Tr.GetElement2IntPoint(), shape2);
-         // Interpolate elfun at the point
-         elfun2_mat.MultTranspose(shape2, state2);
-      }
+      elfun2_mat.MultTranspose(shape2, state2);

      // Get the normal vector and the flux on the face
      if (nor.Size() == 1)  // if 1D, use 1 or -1.
      {
-         nor(0) = 2*Tr.GetElement1IntPoint().x - 1.;
+         // This assume the 1D integration point is in (0,1). This may not work
+         // if this changes.
+         nor(0) = (Tr.GetElement1IntPoint().x - 0.5) * 2.0;
      }
      else
      {
@@ -256,18 +250,14 @@ void HyperbolicFormIntegrator::AssembleFaceVector(
      }
      // Compute F(u+, x) and F(u-, x) with maximum characteristic speed
      // Compute hat(F) using evaluated quantities
-      const real_t speed = (dof2) ? numFlux.Eval(state1, state2, nor, Tr, fluxN):
-                           fluxFunction.ComputeFluxDotN(state1, nor, Tr, fluxN);
+      const real_t speed = numFlux.Eval(state1, state2, nor, Tr, fluxN);

      // Update the global max char speed
      max_char_speed = std::max(speed, max_char_speed);

      // pre-multiply integration weight to flux
      AddMult_a_VWt(-ip.weight*sign, shape1, fluxN, elvect1_mat);
-      if (dof2)
-      {
-         AddMult_a_VWt(+ip.weight*sign, shape2, fluxN, elvect2_mat);
-      }
+      AddMult_a_VWt(+ip.weight*sign, shape2, fluxN, elvect2_mat);
   }
 }

@@ -278,7 +268,7 @@ void HyperbolicFormIntegrator::AssembleFaceGrad(
   // current elements' the number of degrees of freedom
   // does not consider the number of equations
   const int dof1 = el1.GetDof();
-   const int dof2 = (Tr.Elem2No >= 0)?(el2.GetDof()):(0);
+   const int dof2 = el2.GetDof();

 #ifdef MFEM_THREAD_SAFE
   // Local storage for element integration
@@ -312,9 +302,7 @@ void HyperbolicFormIntegrator::AssembleFaceGrad(
   const IntegrationRule *ir = IntRule;
   if (!ir)
   {
-      const int max_el_order = dof2 ? std::max(el1.GetOrder(),
-                                               el2.GetOrder()) : el1.GetOrder();
-      const int order = 2*max_el_order + IntOrderOffset;
+      const int order = 2*std::max(el1.GetOrder(), el2.GetOrder()) + IntOrderOffset;
      ir = &IntRules.Get(Tr.GetGeometryType(), order);
   }
   // loop over integration points
@@ -324,25 +312,20 @@ void HyperbolicFormIntegrator::AssembleFaceGrad(

      Tr.SetAllIntPoints(&ip); // set face and element int. points

-      // Calculate basis functions of the first element at the face
+      // Calculate basis functions on both elements at the face
      el1.CalcShape(Tr.GetElement1IntPoint(), shape1);
+      el2.CalcShape(Tr.GetElement2IntPoint(), shape2);

      // Interpolate elfun at the point
      elfun1_mat.MultTranspose(shape1, state1);
-
-      if (dof2)
-      {
-         // Calculate basis function of the second element at the face
-         el2.CalcShape(Tr.GetElement2IntPoint(), shape2);
-
-         // Interpolate elfun at the point
-         elfun2_mat.MultTranspose(shape2, state2);
-      }
+      elfun2_mat.MultTranspose(shape2, state2);

      // Get the normal vector and the flux on the face
      if (nor.Size() == 1)  // if 1D, use 1 or -1.
      {
-         nor(0) = 2*Tr.GetElement1IntPoint().x - 1.;
+         // This assume the 1D integration point is in (0,1). This may not work
+         // if this changes.
+         nor(0) = (Tr.GetElement1IntPoint().x - 0.5) * 2.0;
      }
      else
      {
@@ -352,14 +335,7 @@ void HyperbolicFormIntegrator::AssembleFaceGrad(
      // Trial side 1

      // Compute hat(J) using evaluated quantities
-      if (dof2)
-      {
-         numFlux.Grad(1, state1, state2, nor, Tr, JDotN);
-      }
-      else
-      {
-         fluxFunction.ComputeFluxJacobianDotN(state1, nor, Tr, JDotN);
-      }
+      numFlux.Grad(1, state1, state2, nor, Tr, JDotN);

      const int ioff = fluxFunction.num_equations * dof1;

@@ -384,325 +360,36 @@ void HyperbolicFormIntegrator::AssembleFaceGrad(
            }
         }

-      if (dof2)
-      {
-         // Trial side 2
-
-         // Compute hat(J) using evaluated quantities
-         numFlux.Grad(2, state1, state2, nor, Tr, JDotN);
-
-         const int joff = ioff;
-
-         for (int di = 0; di < fluxFunction.num_equations; di++)
-            for (int dj = 0; dj < fluxFunction.num_equations; dj++)
-            {
-               // pre-multiply integration weight to Jacobian
-               const real_t w = +ip.weight * sign * JDotN(di,dj);
-               for (int j = 0; j < dof2; j++)
-               {
-                  // Test side 1
-                  for (int i = 0; i < dof1; i++)
-                  {
-                     elmat(i+dof1*di, joff+j+dof2*dj) += w * shape1(i) * shape2(j);
-                  }
-
-                  // Test side 2
-                  for (int i = 0; i < dof2; i++)
-                  {
-                     elmat(ioff+i+dof2*di, joff+j+dof2*dj) -= w * shape2(i) * shape2(j);
-                  }
-               }
-            }
-      }
-   }
-}
-
-BdrHyperbolicDirichletIntegrator::BdrHyperbolicDirichletIntegrator(
-   const NumericalFlux &numFlux,
-   VectorCoefficient &bdrState,
-   const int IntOrderOffset,
-   real_t sign)
-   : NonlinearFormIntegrator(),
-     numFlux(numFlux),
-     fluxFunction(numFlux.GetFluxFunction()),
-     u_vcoeff(bdrState),
-     IntOrderOffset(IntOrderOffset),
-     sign(sign),
-     num_equations(fluxFunction.num_equations)
-{
-   MFEM_VERIFY(fluxFunction.num_equations == bdrState.GetVDim(),
-               "Flux function does not match the vector dimension of the coefficient!");
-#ifndef MFEM_THREAD_SAFE
-   state_in.SetSize(num_equations);
-   state_out.SetSize(num_equations);
-   fluxN.SetSize(num_equations);
-   JDotN.SetSize(num_equations);
-   nor.SetSize(fluxFunction.dim);
-#endif
-   ResetMaxCharSpeed();
-}
-
-void BdrHyperbolicDirichletIntegrator::AssembleFaceVector(
-   const FiniteElement &el, const FiniteElement &,
-   FaceElementTransformations &Tr, const Vector &elfun, Vector &elvect)
-{
-   MFEM_ASSERT(Tr.Elem2No < 0, "Not a boundary face!");
-
-   // current elements' the number of degrees of freedom
-   // does not consider the number of equations
-   const int dof = el.GetDof();
-
-#ifdef MFEM_THREAD_SAFE
-   // Local storage for element integration
-
-   // shape function value at an integration point
-   Vector shape(dof);
-   // normal vector (usually not a unit vector)
-   Vector nor(Tr.GetSpaceDim());
-   // state value at an integration point - interior
-   Vector state_in(num_equations);
-   // state value at an integration point - boundary
-   Vector state_out(num_equations);
-   // hat(F)(u,x)
-   Vector fluxN(num_equations);
-#else
-   shape.SetSize(dof);
-#endif
-
-   elvect.SetSize(dof * num_equations);
-   elvect = 0.0;
-
-   const DenseMatrix elfun_mat(elfun.GetData(), dof, num_equations);
-
-   DenseMatrix elvect_mat(elvect.GetData(), dof, num_equations);
-
-   // Obtain integration rule. If integration is rule is given, then use it.
-   // Otherwise, get (2*p + IntOrderOffset) order integration rule
-   const IntegrationRule *ir = IntRule;
-   if (!ir)
-   {
-      const int order = 2*el.GetOrder() + IntOrderOffset;
-      ir = &IntRules.Get(Tr.GetGeometryType(), order);
-   }
-   // loop over integration points
-   for (int i = 0; i < ir->GetNPoints(); i++)
-   {
-      const IntegrationPoint &ip = ir->IntPoint(i);
-
-      Tr.SetAllIntPoints(&ip); // set face and element int. points
-
-      // Calculate basis functions at the face
-      el.CalcShape(Tr.GetElement1IntPoint(), shape);
-
-      // Interpolate elfun at the point
-      elfun_mat.MultTranspose(shape, state_in);
-
-      // Evaluate boundary state at the point
-      u_vcoeff.Eval(state_out, Tr, ip);
-
-      // Get the normal vector and the flux on the face
-      if (nor.Size() == 1)  // if 1D, use 1 or -1.
-      {
-         nor(0) = 2*Tr.GetElement1IntPoint().x - 1.;
-      }
-      else
-      {
-         CalcOrtho(Tr.Jacobian(), nor);
-      }
-      // Compute F(u+, x) and F(u_b, x) with maximum characteristic speed
-      // Compute hat(F) using evaluated quantities
-      const real_t speed = numFlux.Eval(state_in, state_out, nor, Tr, fluxN);
-
-      // Update the global max char speed
-      max_char_speed = std::max(speed, max_char_speed);
-
-      // pre-multiply integration weight to flux
-      AddMult_a_VWt(-ip.weight*sign, shape, fluxN, elvect_mat);
-   }
-}
-
-void BdrHyperbolicDirichletIntegrator::AssembleFaceGrad(
-   const FiniteElement &el, const FiniteElement &,
-   FaceElementTransformations &Tr, const Vector &elfun, DenseMatrix &elmat)
-{
-   // current elements' the number of degrees of freedom
-   // does not consider the number of equations
-   const int dof = el.GetDof();
-
-#ifdef MFEM_THREAD_SAFE
-   // Local storage for element integration
-
-   // shape function value at an integration point
-   Vector shape(dof);
-   // normal vector (usually not a unit vector)
-   Vector nor(Tr.GetSpaceDim());
-   // state value at an integration point - interior
-   Vector state_in(num_equations);
-   // state value at an integration point - boundary
-   Vector state_out(num_equations);
-   // hat(J)(u,x)
-   DenseMatrix JDotN(num_equations);
-#else
-   shape.SetSize(dof);
-#endif
-
-   elmat.SetSize(dof * num_equations);
-   elmat = 0.0;
-
-   const DenseMatrix elfun_mat(elfun.GetData(), dof, num_equations);
-
-   // Obtain integration rule. If integration is rule is given, then use it.
-   // Otherwise, get (2*p + IntOrderOffset) order integration rule
-   const IntegrationRule *ir = IntRule;
-   if (!ir)
-   {
-      const int order = 2*el.GetOrder() + IntOrderOffset;
-      ir = &IntRules.Get(Tr.GetGeometryType(), order);
-   }
-   // loop over integration points
-   for (int q = 0; q < ir->GetNPoints(); q++)
-   {
-      const IntegrationPoint &ip = ir->IntPoint(q);
-
-      Tr.SetAllIntPoints(&ip); // set face and element int. points
-
-      // Calculate basis functions at the face
-      el.CalcShape(Tr.GetElement1IntPoint(), shape);
-
-      // Interpolate elfun at the point
-      elfun_mat.MultTranspose(shape, state_in);
-
-      // Evaluate boundary state at the point
-      u_vcoeff.Eval(state_out, Tr, ip);
-
-      // Get the normal vector and the flux on the face
-      if (nor.Size() == 1)  // if 1D, use 1 or -1.
-      {
-         nor(0) = 2*Tr.GetElement1IntPoint().x - 1.;
-      }
-      else
-      {
-         CalcOrtho(Tr.Jacobian(), nor);
-      }
+      // Trial side 2

      // Compute hat(J) using evaluated quantities
-      numFlux.Grad(1, state_in, state_out, nor, Tr, JDotN);
+      numFlux.Grad(2, state1, state2, nor, Tr, JDotN);
+
+      const int joff = ioff;

      for (int di = 0; di < fluxFunction.num_equations; di++)
         for (int dj = 0; dj < fluxFunction.num_equations; dj++)
         {
            // pre-multiply integration weight to Jacobian
-            const real_t w = -ip.weight * sign * JDotN(di,dj);
-            for (int j = 0; j < dof; j++)
-               for (int i = 0; i < dof; i++)
+            const real_t w = +ip.weight * sign * JDotN(di,dj);
+            for (int j = 0; j < dof2; j++)
+            {
+               // Test side 1
+               for (int i = 0; i < dof1; i++)
               {
-                  elmat(i+dof*di, j+dof*dj) += w * shape(i) * shape(j);
+                  elmat(i+dof1*di, joff+j+dof2*dj) += w * shape1(i) * shape2(j);
               }
+
+               // Test side 2
+               for (int i = 0; i < dof2; i++)
+               {
+                  elmat(ioff+i+dof2*di, joff+j+dof2*dj) -= w * shape2(i) * shape2(j);
+               }
+            }
         }
   }
 }

-BoundaryHyperbolicFlowIntegrator::BoundaryHyperbolicFlowIntegrator(
-   const FluxFunction &flux, VectorCoefficient &u, real_t alpha_, real_t beta_,
-   const int IntOrderOffset_)
-   : fluxFunction(flux), u_vcoeff(u), alpha(alpha_), beta(beta_),
-     IntOrderOffset(IntOrderOffset_)
-{
-   MFEM_VERIFY(fluxFunction.num_equations == u_vcoeff.GetVDim(),
-               "Flux function does not match the vector dimension of the coefficient!");
-#ifndef MFEM_THREAD_SAFE
-   state.SetSize(fluxFunction.num_equations);
-   nor.SetSize(fluxFunction.dim);
-   fluxN.SetSize(fluxFunction.num_equations);
-#endif
-   ResetMaxCharSpeed();
-}
-
-void BoundaryHyperbolicFlowIntegrator::AssembleRHSElementVect(
-   const FiniteElement &el, ElementTransformation &Tr, Vector &elvect)
-{
-   mfem_error("BoundaryHyperbolicFlowIntegrator::AssembleRHSElementVect\n"
-              "  is not implemented as boundary integrator!\n"
-              "  Use LinearForm::AddBdrFaceIntegrator instead of\n"
-              "  LinearForm::AddBoundaryIntegrator.");
-}
-
-void BoundaryHyperbolicFlowIntegrator::AssembleRHSElementVect(
-   const FiniteElement &el, FaceElementTransformations &Tr, Vector &elvect)
-{
-   // current elements' the number of degrees of freedom
-   // does not consider the number of equations
-   const int dof = el.GetDof();
-
-#ifdef MFEM_THREAD_SAFE
-   // Local storage for element integration
-
-   // shape function value at an integration point
-   Vector shape(dof);
-   // state value at an integration point
-   Vector state(fluxFunction.num_equations);
-   // normal vector (usually not a unit vector)
-   Vector nor(Tr.GetSpaceDim());
-   // hat(F)(u,x)
-   Vector fluxN(fluxFunction.num_equations);
-#else
-   shape.SetSize(dof);
-#endif
-
-   elvect.SetSize(dof * fluxFunction.num_equations);
-   elvect = 0.0;
-
-   DenseMatrix elvect_mat(elvect.GetData(), dof, fluxFunction.num_equations);
-
-   // Obtain integration rule. If integration is rule is given, then use it.
-   // Otherwise, get (2*p + IntOrderOffset) order integration rule
-   const IntegrationRule *ir = IntRule;
-   if (!ir)
-   {
-      const int order = 2*el.GetOrder() + IntOrderOffset;
-      ir = &IntRules.Get(Tr.GetGeometryType(), order);
-   }
-   // loop over integration points
-   for (int i = 0; i < ir->GetNPoints(); i++)
-   {
-      const IntegrationPoint &ip = ir->IntPoint(i);
-
-      Tr.SetAllIntPoints(&ip); // set face and element int. points
-
-      // Calculate basis functions on both elements at the face
-      el.CalcShape(Tr.GetElement1IntPoint(), shape);
-
-      // Evaluate the coefficient at the point
-      u_vcoeff.Eval(state, Tr, ip);
-
-      // Get the normal vector and the flux on the face
-      if (nor.Size() == 1)  // if 1D, use 1 or -1.
-      {
-         nor(0) = 2*Tr.GetElement1IntPoint().x - 1.;
-      }
-      else
-      {
-         CalcOrtho(Tr.Jacobian(), nor);
-      }
-      // Compute F(u, x) with maximum characteristic speed
-      const real_t speed = fluxFunction.ComputeFluxDotN(state, nor, Tr, fluxN);
-
-      // Update the global max char speed
-      max_char_speed = std::max(speed, max_char_speed);
-
-      // pre-multiply integration weight to flux
-      const real_t a = 0.5 * alpha  * ip.weight;
-      const real_t b = beta * ip.weight;
-
-      for (int n = 0; n < fluxFunction.num_equations; n++)
-      {
-         fluxN(n) = a * fluxN(n) - b * fabs(fluxN(n));
-      }
-
-      AddMultVWt(shape, fluxN, elvect_mat);
-   }
-}
-
 real_t FluxFunction::ComputeFluxDotN(const Vector &U,
                                     const Vector &normal,
                                     FaceElementTransformations &Tr,
@@ -306,14 +306,12 @@ MFEM_DEPRECATED typedef NumericalFlux RiemannSolver;
 class HyperbolicFormIntegrator : public NonlinearFormIntegrator
 {
 private:
+   // The maximum characteristic speed, updated during element/face vector assembly
+   real_t max_char_speed;
   const NumericalFlux &numFlux;   // Numerical flux that maps F(u±,x) to F̂
   const FluxFunction &fluxFunction;
   const int IntOrderOffset; // integration order offset, 2*p + IntOrderOffset.
   const real_t sign;
-
-   // The maximum characteristic speed, updated during element/face vector assembly
-   real_t max_char_speed;
-
 #ifndef MFEM_THREAD_SAFE
   // Local storage for element integration
   Vector shape;              // shape function value at an integration point
@@ -333,9 +331,8 @@ private:

 public:
   const int num_equations;  // the number of equations
-
   /**
-    * @brief Construct a new HyperbolicFormIntegrator object
+    * @brief Construct a new Hyperbolic Form Integrator object
    *
    * @param[in] numFlux numerical flux
    * @param[in] IntOrderOffset integration order offset
@@ -346,14 +343,21 @@ public:
      const int IntOrderOffset = 0,
      const real_t sign = 1.);

-   /// Reset the maximum characteristic speed to zero
-   void ResetMaxCharSpeed() { max_char_speed = 0.0; }
+   /**
+    * @brief Reset the Max Char Speed 0
+    *
+    */
+   void ResetMaxCharSpeed()
+   {
+      max_char_speed = 0.0;
+   }

-   /// Get the maximum characteristic speed
-   real_t GetMaxCharSpeed() const { return max_char_speed; }
+   real_t GetMaxCharSpeed()
+   {
+      return max_char_speed;
+   }

-   /// Get the associated flux function
-   const FluxFunction &GetFluxFunction() const { return fluxFunction; }
+   const FluxFunction &GetFluxFunction() { return fluxFunction; }

   /**
    * @brief Implements (F(u), ∇v) with abstract F computed by
@@ -412,182 +416,6 @@ public:
                         const Vector &elfun, DenseMatrix &elmat) override;
 };

-/**
- * @brief Abstract boundary hyperbolic form integrator, assembling
- * <F̂(u⁻,u_b,x) n, [v]> term for scalar finite elements at the boundary.
- *
- * This form integrator is coupled with a NumericalFlux that implements the
- * numerical flux F̂ at the boundary faces. The flux F is obtained from the
- * FluxFunction assigned to the aforementioned NumericalFlux with the given
- * boundary coefficient for the state u_b.
- *
- * Note the class can be used for imposing conditions on interior interfaces.
- */
-class BdrHyperbolicDirichletIntegrator : public NonlinearFormIntegrator
-{
-private:
-   const NumericalFlux &numFlux;    // Numerical flux that maps F to F̂
-   const FluxFunction &fluxFunction;
-   VectorCoefficient &u_vcoeff;     // Boundary state vector coefficient
-   const int IntOrderOffset; // integration order offset, 2*p + IntOrderOffset.
-   const real_t sign;
-
-   // The maximum characteristic speed, updated during element/face vector assembly
-   real_t max_char_speed;
-
-#ifndef MFEM_THREAD_SAFE
-   // Local storage for element integration
-   Vector shape;  // shape function value at an integration point
-   Vector state_in;  // state value at an integration point - interior
-   Vector state_out;  // state value at an integration point - boundary
-   Vector nor;     // normal vector, see mfem::CalcOrtho()
-   Vector fluxN;   // F̂(u⁻,u_b,x) n
-   DenseMatrix JDotN;   // Ĵ(u⁻,u_b,x) n
-#endif
-
-public:
-   const int num_equations;  // the number of equations
-
-   /**
-    * @brief Construct a new BdrHyperbolicDirichletIntegrator object
-    *
-    * @param[in] numFlux numerical flux
-    * @param[in] bdrState boundary state coefficient
-    * @param[in] IntOrderOffset integration order offset
-    * @param[in] sign sign of the convection term
-    */
-   BdrHyperbolicDirichletIntegrator(
-      const NumericalFlux &numFlux,
-      VectorCoefficient &bdrState,
-      const int IntOrderOffset = 0,
-      const real_t sign = 1.);
-
-   /// Reset the maximum characteristic speed to zero
-   void ResetMaxCharSpeed() { max_char_speed = 0.0; }
-
-   /// Get the maximum characteristic speed
-   real_t GetMaxCharSpeed() const { return max_char_speed; }
-
-   /// Get the associated flux function
-   const FluxFunction &GetFluxFunction() const { return fluxFunction; }
-
-   /**
-    * @brief Implements <-F̂(u⁻,u_b,x) n, [v]> with abstract F̂ computed by
-    * NumericalFlux::Eval() of the numerical flux object
-    *
-    * @param[in] el1 finite element of the interior element
-    * @param[in] el2 not used
-    * @param[in] Tr face element transformations
-    * @param[in] elfun local coefficient of basis for the interior element
-    * @param[out] elvect evaluated dual vector <-F̂(u⁻,u_b,x) n, [v]>
-    */
-   void AssembleFaceVector(const FiniteElement &el1,
-                           const FiniteElement &el2,
-                           FaceElementTransformations &Tr,
-                           const Vector &elfun, Vector &elvect) override;
-
-   /**
-    * @brief Implements <-Ĵ(u⁻,u_b,x) n, [v]> with abstract Ĵ computed by
-    * NumericalFlux::Grad() of the numerical flux object
-    *
-    * @param[in] el1 finite element of the interior element
-    * @param[in] el2 not used
-    * @param[in] Tr face element transformations
-    * @param[in] elfun local coefficient of basis for the interior element
-    * @param[out] elmat evaluated Jacobian matrix <-Ĵ(u⁻,u_b,x) n, [v]>
-    */
-   void AssembleFaceGrad(const FiniteElement &el1,
-                         const FiniteElement &el2,
-                         FaceElementTransformations &Tr,
-                         const Vector &elfun, DenseMatrix &elmat) override;
-};
-
-/**
- * @brief Abstract boundary hyperbolic linear form integrator, assembling
- * <ɑ/2 F(u,x) n - β |F(u,x) n|, v> terms for scalar finite elements.
- *
- * This form integrator is coupled with a FluxFunction that evaluates the
- * flux F at the boundary.
- *
- * Note the upwinding is performed component-wise. For general boundary
- * integration with a numerical flux, see BdrHyperbolicDirichletIntegrator.
- */
-class BoundaryHyperbolicFlowIntegrator : public LinearFormIntegrator
-{
-   const FluxFunction &fluxFunction;
-   VectorCoefficient &u_vcoeff;
-   const real_t alpha, beta;
-   const int IntOrderOffset; // integration order offset, 2*p + IntOrderOffset.
-
-   // The maximum characteristic speed, updated during face vector assembly
-   real_t max_char_speed;
-
-#ifndef MFEM_THREAD_SAFE
-   // Local storage for element integration
-   Vector shape;   // shape function value at an integration point
-   Vector state;   // state value at an integration point
-   Vector nor;     // normal vector, see mfem::CalcOrtho()
-   Vector fluxN;   // F(u,x) n
-#endif
-
-public:
-   /**
-    * @brief Construct a new BoundaryHyperbolicFlowIntegrator object
-    *
-    * @param[in] flux flux function
-    * @param[in] u vector state coefficient
-    * @param[in] alpha ɑ coefficient (β = ɑ/2)
-    * @param[in] IntOrderOffset integration order offset
-    */
-   BoundaryHyperbolicFlowIntegrator(const FluxFunction &flux, VectorCoefficient &u,
-                                    real_t alpha = -1., int IntOrderOffset = 0)
-      : BoundaryHyperbolicFlowIntegrator(flux, u, alpha, alpha/2., IntOrderOffset) { }
-
-   /**
-    * @brief Construct a new BoundaryHyperbolicFlowIntegrator object
-    *
-    * @param[in] flux flux function
-    * @param[in] u vector state coefficient
-    * @param[in] alpha ɑ coefficient
-    * @param[in] beta β coefficient
-    * @param[in] IntOrderOffset integration order offset
-    */
-   BoundaryHyperbolicFlowIntegrator(const FluxFunction &flux, VectorCoefficient &u,
-                                    real_t alpha, real_t beta, int IntOrderOffset = 0);
-
-   /// Reset the maximum characteristic speed to zero
-   void ResetMaxCharSpeed() { max_char_speed = 0.0; }
-
-   /// Get the maximum characteristic speed
-   real_t GetMaxCharSpeed() const { return max_char_speed; }
-
-   /// Get the associated flux function
-   const FluxFunction &GetFluxFunction() const { return fluxFunction; }
-
-   using LinearFormIntegrator::AssembleRHSElementVect;
-
-   /**
-    * @warning Boundary element integration not implemented, use
-    * AssembleRHSElementVect(const FiniteElement&,
-    * FaceElementTransformations &, Vector &) instead
-    */
-   void AssembleRHSElementVect(const FiniteElement &el,
-                               ElementTransformation &Tr,
-                               Vector &elvect) override;
-
-   /**
-    * @brief Implements <-F(u,x) n, v> with abstract F computed by
-    * FluxFunction::ComputeFluxDotN() of the flux function object
-    *
-    * @param[in] el finite element
-    * @param[in] Tr face element transformations
-    * @param[out] elvect evaluated dual vector <F(u,x) n, v>
-    */
-   void AssembleRHSElementVect(const FiniteElement &el,
-                               FaceElementTransformations &Tr,
-                               Vector &elvect) override;
-};
-

 /**
 * @brief Rusanov flux, also known as local Lax-Friedrichs,
@@ -15,86 +15,6 @@
 namespace mfem
 {

-CurlCurlIntegrator::CurlCurlIntegrator() : Q(nullptr), DQ(nullptr), MQ(nullptr)
-{
-   static Kernels kernels;
-}
-
-CurlCurlIntegrator::CurlCurlIntegrator(Coefficient &q,
-                                       const IntegrationRule *ir)
-   : BilinearFormIntegrator(ir), Q(&q), DQ(nullptr), MQ(nullptr)
-{
-   static Kernels kernels;
-}
-
-CurlCurlIntegrator::CurlCurlIntegrator(DiagonalMatrixCoefficient &dq,
-                                       const IntegrationRule *ir)
-   : BilinearFormIntegrator(ir), Q(nullptr), DQ(&dq), MQ(nullptr)
-{
-   static Kernels kernels;
-}
-
-CurlCurlIntegrator::CurlCurlIntegrator(MatrixCoefficient &mq,
-                                       const IntegrationRule *ir)
-   : BilinearFormIntegrator(ir), Q(nullptr), DQ(nullptr), MQ(&mq)
-{
-   static Kernels kernels;
-}
-
-/// \cond DO_NOT_DOCUMENT
-
-CurlCurlIntegrator::Kernels::Kernels()
-{
-   CurlCurlIntegrator::AddSpecialization<3, 2, 3>();
-   CurlCurlIntegrator::AddSpecialization<3, 3, 4>();
-   CurlCurlIntegrator::AddSpecialization<3, 4, 5>();
-   CurlCurlIntegrator::AddSpecialization<3, 5, 6>();
-}
-
-CurlCurlIntegrator::ApplyKernelType
-CurlCurlIntegrator::ApplyPAKernels::Fallback(int DIM, int, int)
-{
-   if (DIM == 2) { return internal::PACurlCurlApply2D; }
-   else if (DIM == 3)
-   {
-      if (Device::Allows(Backend::DEVICE_MASK))
-      {
-         return internal::SmemPACurlCurlApply3D;
-      }
-      else
-      {
-         return internal::PACurlCurlApply3D;
-      }
-   }
-   else { MFEM_ABORT(""); }
-}
-
-CurlCurlIntegrator::DiagonalKernelType
-CurlCurlIntegrator::DiagonalPAKernels::Fallback(int DIM, int, int)
-{
-   if (DIM == 2)
-   {
-      return internal::PACurlCurlAssembleDiagonal2D;
-   }
-   else if (DIM == 3)
-   {
-      if (Device::Allows(Backend::DEVICE_MASK))
-      {
-         return internal::SmemPACurlCurlAssembleDiagonal3D;
-      }
-      else
-      {
-         return internal::PACurlCurlAssembleDiagonal3D;
-      }
-   }
-   else
-   {
-      MFEM_ABORT("");
-   }
-}
-
-/// \endcond DO_NOT_DOCUMENT
-
 void CurlCurlIntegrator::AssemblePA(const FiniteElementSpace &fes)
 {
   // Assumes tensor-product elements
@@ -157,16 +77,129 @@ void CurlCurlIntegrator::AssemblePA(const FiniteElementSpace &fes)

 void CurlCurlIntegrator::AssembleDiagonalPA(Vector& diag)
 {
-   DiagonalPAKernels::Run(dim, dofs1D, quad1D, dofs1D, quad1D, symmetric, ne,
-                          mapsO->B, mapsC->B, mapsO->G, mapsC->G, pa_data,
-                          diag);
+   if (dim == 3)
+   {
+      if (Device::Allows(Backend::DEVICE_MASK))
+      {
+         const int ID = (dofs1D << 4) | quad1D;
+         switch (ID)
+         {
+            case 0x23:
+               return internal::SmemPACurlCurlAssembleDiagonal3D<2,3>(
+                         dofs1D,
+                         quad1D,
+                         symmetric, ne,
+                         mapsO->B, mapsC->B,
+                         mapsO->G, mapsC->G,
+                         pa_data, diag);
+            case 0x34:
+               return internal::SmemPACurlCurlAssembleDiagonal3D<3,4>(
+                         dofs1D,
+                         quad1D,
+                         symmetric, ne,
+                         mapsO->B, mapsC->B,
+                         mapsO->G, mapsC->G,
+                         pa_data, diag);
+            case 0x45:
+               return internal::SmemPACurlCurlAssembleDiagonal3D<4,5>(
+                         dofs1D,
+                         quad1D,
+                         symmetric, ne,
+                         mapsO->B, mapsC->B,
+                         mapsO->G, mapsC->G,
+                         pa_data, diag);
+            case 0x56:
+               return internal::SmemPACurlCurlAssembleDiagonal3D<5,6>(
+                         dofs1D,
+                         quad1D,
+                         symmetric, ne,
+                         mapsO->B, mapsC->B,
+                         mapsO->G, mapsC->G,
+                         pa_data, diag);
+            default:
+               return internal::SmemPACurlCurlAssembleDiagonal3D(
+                         dofs1D, quad1D,
+                         symmetric, ne,
+                         mapsO->B, mapsC->B,
+                         mapsO->G, mapsC->G,
+                         pa_data, diag);
+         }
+      }
+      else
+      {
+         internal::PACurlCurlAssembleDiagonal3D(dofs1D, quad1D, symmetric, ne,
+                                                mapsO->B, mapsC->B,
+                                                mapsO->G, mapsC->G,
+                                                pa_data, diag);
+      }
+   }
+   else if (dim == 2)
+   {
+      internal::PACurlCurlAssembleDiagonal2D(dofs1D, quad1D, ne,
+                                             mapsO->B, mapsC->G, pa_data, diag);
+   }
+   else
+   {
+      MFEM_ABORT("Unsupported dimension!");
+   }
 }

 void CurlCurlIntegrator::AddMultPA(const Vector &x, Vector &y) const
 {
-   ApplyPAKernels::Run(dim, dofs1D, quad1D, dofs1D, quad1D, symmetric, ne,
-                       mapsO->B, mapsC->B, mapsO->Bt, mapsC->Bt, mapsC->G,
-                       mapsC->Gt, pa_data, x, y, false);
+   if (dim == 3)
+   {
+      if (Device::Allows(Backend::DEVICE_MASK))
+      {
+         const int ID = (dofs1D << 4) | quad1D;
+         switch (ID)
+         {
+            case 0x23:
+               return internal::SmemPACurlCurlApply3D<2,3>(
+                         dofs1D, quad1D,
+                         symmetric, ne,
+                         mapsO->B, mapsC->B, mapsO->Bt, mapsC->Bt,
+                         mapsC->G, mapsC->Gt, pa_data, x, y);
+            case 0x34:
+               return internal::SmemPACurlCurlApply3D<3,4>(
+                         dofs1D, quad1D,
+                         symmetric, ne,
+                         mapsO->B, mapsC->B, mapsO->Bt, mapsC->Bt,
+                         mapsC->G, mapsC->Gt, pa_data, x, y);
+            case 0x45:
+               return internal::SmemPACurlCurlApply3D<4,5>(
+                         dofs1D, quad1D,
+                         symmetric, ne,
+                         mapsO->B, mapsC->B, mapsO->Bt, mapsC->Bt,
+                         mapsC->G, mapsC->Gt, pa_data, x, y);
+            case 0x56:
+               return internal::SmemPACurlCurlApply3D<5,6>(
+                         dofs1D, quad1D,
+                         symmetric, ne,
+                         mapsO->B, mapsC->B, mapsO->Bt, mapsC->Bt,
+                         mapsC->G, mapsC->Gt, pa_data, x, y);
+            default:
+               return internal::SmemPACurlCurlApply3D(
+                         dofs1D, quad1D, symmetric, ne,
+                         mapsO->B, mapsC->B, mapsO->Bt, mapsC->Bt,
+                         mapsC->G, mapsC->Gt, pa_data, x, y);
+         }
+      }
+      else
+      {
+         internal::PACurlCurlApply3D(dofs1D, quad1D, symmetric, ne, mapsO->B, mapsC->B,
+                                     mapsO->Bt, mapsC->Bt, mapsC->G, mapsC->Gt,
+                                     pa_data, x, y);
+      }
+   }
+   else if (dim == 2)
+   {
+      internal::PACurlCurlApply2D(dofs1D, quad1D, ne, mapsO->B, mapsO->Bt,
+                                  mapsC->G, mapsC->Gt, pa_data, x, y);
+   }
+   else
+   {
+      MFEM_ABORT("Unsupported dimension!");
+   }
 }

 void CurlCurlIntegrator::AddAbsMultPA(const Vector &x, Vector &y) const
@@ -176,9 +209,61 @@ void CurlCurlIntegrator::AddAbsMultPA(const Vector &x, Vector &y) const
   auto absO = mapsO->Abs();
   auto absC = mapsC->Abs();

-   ApplyPAKernels::Run(dim, dofs1D, quad1D, dofs1D, quad1D, symmetric, ne,
-                       absO.B, absC.B, absO.Bt, absC.Bt, absC.G, absC.Gt,
-                       abs_pa_data, x, y, true);
+   if (dim == 3)
+   {
+      if (Device::Allows(Backend::DEVICE_MASK))
+      {
+         const int ID = (dofs1D << 4) | quad1D;
+         switch (ID)
+         {
+            case 0x23:
+               return internal::SmemPACurlCurlApply3D<2,3>(
+                         dofs1D, quad1D,
+                         symmetric, ne,
+                         absO.B, absC.B, absO.Bt, absC.Bt,
+                         absC.G, absC.Gt, abs_pa_data, x, y, true);
+            case 0x34:
+               return internal::SmemPACurlCurlApply3D<3,4>(
+                         dofs1D, quad1D,
+                         symmetric, ne,
+                         absO.B, absC.B, absO.Bt, absC.Bt,
+                         absC.G, absC.Gt, abs_pa_data, x, y, true);
+            case 0x45:
+               return internal::SmemPACurlCurlApply3D<4,5>(
+                         dofs1D, quad1D,
+                         symmetric, ne,
+                         absO.B, absC.B, absO.Bt, absC.Bt,
+                         absC.G, absC.Gt, abs_pa_data, x, y, true);
+            case 0x56:
+               return internal::SmemPACurlCurlApply3D<5,6>(
+                         dofs1D, quad1D,
+                         symmetric, ne,
+                         absO.B, absC.B, absO.Bt, absC.Bt,
+                         absC.G, absC.Gt, abs_pa_data, x, y, true);
+            default:
+               return internal::SmemPACurlCurlApply3D<0,0>(
+                         dofs1D, quad1D, symmetric, ne,
+                         absO.B, absC.B, absO.Bt, absC.Bt,
+                         absC.G, absC.Gt, abs_pa_data, x, y, true);
+         }
+      }
+      else
+      {
+         internal::PACurlCurlApply3D<0,0>(
+            dofs1D, quad1D, symmetric, ne,
+            absO.B, absC.B, absO.Bt, absC.Bt, absC.G, absC.Gt,
+            abs_pa_data, x, y, true);
+      }
+   }
+   else if (dim == 2)
+   {
+      internal::PACurlCurlApply2D(dofs1D, quad1D, ne, absO.B, absO.Bt,
+                                  absC.G, absC.Gt, abs_pa_data, x, y, true);
+   }
+   else
+   {
+      MFEM_ABORT("Unsupported dimension!");
+   }
 }

 } // namespace mfem
@@ -1,500 +0,0 @@
-// 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_BILININTEG_DGDIFFUSION_KERNELS_HPP
-#define MFEM_BILININTEG_DGDIFFUSION_KERNELS_HPP
-
-#include "../../general/forall.hpp"
-#include "../../mesh/face_nbr_geom.hpp"
-#include "../fe/face_map_utils.hpp"
-#include "../gridfunc.hpp"
-#include "../qfunction.hpp"
-
-/// \cond DO_NOT_DOCUMENT
-namespace mfem
-{
-
-namespace internal
-{
-
-template <int T_D1D = 0, int T_Q1D = 0>
-static void PADGDiffusionApply2D(const int NF, const Array<real_t> &b,
-                                 const Array<real_t> &bt,
-                                 const Array<real_t> &g,
-                                 const Array<real_t> &gt, const real_t sigma,
-                                 const Vector &pa_data, const Vector &x_,
-                                 const Vector &dxdn_, Vector &y_, Vector &dydn_,
-                                 const int d1d = 0, const int q1d = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-
-   auto B_ = Reshape(b.Read(), Q1D, D1D);
-   auto G_ = Reshape(g.Read(), Q1D, D1D);
-
-   auto pa =
-      Reshape(pa_data.Read(), 6, Q1D, NF); // (q, 1/h, J00, J01, J10, J11)
-
-   auto x = Reshape(x_.Read(), D1D, 2, NF);
-   auto y = Reshape(y_.ReadWrite(), D1D, 2, NF);
-   auto dxdn = Reshape(dxdn_.Read(), D1D, 2, NF);
-   auto dydn = Reshape(dydn_.ReadWrite(), D1D, 2, NF);
-
-   const int NBX = std::max(D1D, Q1D);
-
-   mfem::forall_2D(NF, NBX, 2, [=] MFEM_HOST_DEVICE(int f) -> void
-   {
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-
-      MFEM_SHARED real_t u0[max_D1D];
-      MFEM_SHARED real_t u1[max_D1D];
-      MFEM_SHARED real_t du0[max_D1D];
-      MFEM_SHARED real_t du1[max_D1D];
-
-      MFEM_SHARED real_t Bu0[max_Q1D];
-      MFEM_SHARED real_t Bu1[max_Q1D];
-      MFEM_SHARED real_t Bdu0[max_Q1D];
-      MFEM_SHARED real_t Bdu1[max_Q1D];
-
-      MFEM_SHARED real_t r[max_Q1D];
-
-      MFEM_SHARED real_t BG[2 * max_D1D * max_Q1D];
-      DeviceMatrix B(BG, Q1D, D1D);
-      DeviceMatrix G(BG + D1D * Q1D, Q1D, D1D);
-
-      if (MFEM_THREAD_ID(y) == 0)
-      {
-         MFEM_FOREACH_THREAD(p, x, Q1D)
-         {
-            for (int d = 0; d < D1D; ++d)
-            {
-               B(p, d) = B_(p, d);
-               G(p, d) = G_(p, d);
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-
-      // copy edge values to u0, u1 and copy edge normals to du0, du1
-      MFEM_FOREACH_THREAD(side, y, 2)
-      {
-         real_t *u = (side == 0) ? u0 : u1;
-         real_t *du = (side == 0) ? du0 : du1;
-         MFEM_FOREACH_THREAD(d, x, D1D)
-         {
-            u[d] = x(d, side, f);
-            du[d] = dxdn(d, side, f);
-         }
-      }
-      MFEM_SYNC_THREAD;
-
-      // eval @ quad points
-      MFEM_FOREACH_THREAD(side, y, 2)
-      {
-         real_t *u = (side == 0) ? u0 : u1;
-         real_t *du = (side == 0) ? du0 : du1;
-         real_t *Bu = (side == 0) ? Bu0 : Bu1;
-         real_t *Bdu = (side == 0) ? Bdu0 : Bdu1;
-
-         MFEM_FOREACH_THREAD(p, x, Q1D)
-         {
-            const real_t Je_side[] = {pa(2 + 2 * side, p, f),
-                                      pa(2 + 2 * side + 1, p, f)
-                                     };
-
-            Bu[p] = 0.0;
-            Bdu[p] = 0.0;
-
-            for (int d = 0; d < D1D; ++d)
-            {
-               const real_t b = B(p, d);
-               const real_t g = G(p, d);
-
-               Bu[p] += b * u[d];
-               Bdu[p] += Je_side[0] * b * du[d] + Je_side[1] * g * u[d];
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-
-      // term - < {Q du/dn}, [v] > +  kappa * < {Q/h} [u], [v] >:
-      if (MFEM_THREAD_ID(y) == 0)
-      {
-         MFEM_FOREACH_THREAD(p, x, Q1D)
-         {
-            const real_t q = pa(0, p, f);
-            const real_t hi = pa(1, p, f);
-            const real_t jump = Bu0[p] - Bu1[p];
-            const real_t avg = Bdu0[p] + Bdu1[p]; // = {Q du/dn} * w * det(J)
-            r[p] = -avg + hi * q * jump;
-         }
-      }
-      MFEM_SYNC_THREAD;
-
-      MFEM_FOREACH_THREAD(d, x, D1D)
-      {
-         real_t Br = 0.0;
-
-         for (int p = 0; p < Q1D; ++p)
-         {
-            Br += B(p, d) * r[p];
-         }
-
-         u0[d] = Br; // overwrite u0, u1
-         u1[d] = -Br;
-      } // for d
-      MFEM_SYNC_THREAD;
-
-      MFEM_FOREACH_THREAD(side, y, 2)
-      {
-         real_t *du = (side == 0) ? du0 : du1;
-         MFEM_FOREACH_THREAD(d, x, D1D) { du[d] = 0.0; }
-      }
-      MFEM_SYNC_THREAD;
-
-      // term sigma * < [u], {Q dv/dn} >
-      MFEM_FOREACH_THREAD(side, y, 2)
-      {
-         real_t *const du = (side == 0) ? du0 : du1;
-         real_t *const u = (side == 0) ? u0 : u1;
-
-         MFEM_FOREACH_THREAD(d, x, D1D)
-         {
-            for (int p = 0; p < Q1D; ++p)
-            {
-               const real_t Je[] = {pa(2 + 2 * side, p, f),
-                                    pa(2 + 2 * side + 1, p, f)
-                                   };
-               const real_t jump = Bu0[p] - Bu1[p];
-               const real_t r_p = Je[0] * jump; // normal
-               const real_t w_p = Je[1] * jump; // tangential
-               du[d] += sigma * B(p, d) * r_p;
-               u[d] += sigma * G(p, d) * w_p;
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-
-      MFEM_FOREACH_THREAD(side, y, 2)
-      {
-         real_t *u = (side == 0) ? u0 : u1;
-         real_t *du = (side == 0) ? du0 : du1;
-         MFEM_FOREACH_THREAD(d, x, D1D)
-         {
-            y(d, side, f) += u[d];
-            dydn(d, side, f) += du[d];
-         }
-      }
-   }); // mfem::forall
-}
-
-template <int T_D1D = 0, int T_Q1D = 0>
-static void PADGDiffusionApply3D(const int NF, const Array<real_t> &b,
-                                 const Array<real_t> &bt,
-                                 const Array<real_t> &g,
-                                 const Array<real_t> &gt, const real_t sigma,
-                                 const Vector &pa_data, const Vector &x_,
-                                 const Vector &dxdn_, Vector &y_, Vector &dydn_,
-                                 const int d1d = 0, const int q1d = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-
-   auto B_ = Reshape(b.Read(), Q1D, D1D);
-   auto G_ = Reshape(g.Read(), Q1D, D1D);
-
-   // (J0[0], J0[1], J0[2], J1[0], J1[1], J1[2], q/h)
-   auto pa = Reshape(pa_data.Read(), 7, Q1D, Q1D, NF);
-
-   auto x = Reshape(x_.Read(), D1D, D1D, 2, NF);
-   auto y = Reshape(y_.ReadWrite(), D1D, D1D, 2, NF);
-   auto dxdn = Reshape(dxdn_.Read(), D1D, D1D, 2, NF);
-   auto dydn = Reshape(dydn_.ReadWrite(), D1D, D1D, 2, NF);
-
-   const int NBX = std::max(D1D, Q1D);
-
-   mfem::forall_3D(NF, NBX, NBX, 2, [=] MFEM_HOST_DEVICE(int f) -> void
-   {
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-
-      MFEM_SHARED real_t u0[max_Q1D][max_Q1D];
-      MFEM_SHARED real_t u1[max_Q1D][max_Q1D];
-
-      MFEM_SHARED real_t du0[max_Q1D][max_Q1D];
-      MFEM_SHARED real_t du1[max_Q1D][max_Q1D];
-
-      MFEM_SHARED real_t Gu0[max_Q1D][max_Q1D];
-      MFEM_SHARED real_t Gu1[max_Q1D][max_Q1D];
-
-      MFEM_SHARED real_t Bu0[max_Q1D][max_Q1D];
-      MFEM_SHARED real_t Bu1[max_Q1D][max_Q1D];
-
-      MFEM_SHARED real_t Bdu0[max_Q1D][max_Q1D];
-      MFEM_SHARED real_t Bdu1[max_Q1D][max_Q1D];
-
-      MFEM_SHARED real_t kappa_Qh[max_Q1D][max_Q1D];
-
-      MFEM_SHARED real_t nJe[2][max_Q1D][max_Q1D][3];
-      MFEM_SHARED real_t BG[2 * max_D1D * max_Q1D];
-
-      // some buffers are reused multiple times, but for clarity have new names:
-      real_t(*Bj0)[max_Q1D] = Bu0;
-      real_t(*Bj1)[max_Q1D] = Bu1;
-      real_t(*Bjn0)[max_Q1D] = Bdu0;
-      real_t(*Bjn1)[max_Q1D] = Bdu1;
-      real_t(*Gj0)[max_Q1D] = Gu0;
-      real_t(*Gj1)[max_Q1D] = Gu1;
-
-      DeviceMatrix B(BG, Q1D, D1D);
-      DeviceMatrix G(BG + D1D * Q1D, Q1D, D1D);
-
-      // copy face values to u0, u1 and copy normals to du0, du1
-      MFEM_FOREACH_THREAD(side, z, 2)
-      {
-         real_t(*u)[max_Q1D] = (side == 0) ? u0 : u1;
-         real_t(*du)[max_Q1D] = (side == 0) ? du0 : du1;
-
-         MFEM_FOREACH_THREAD(d2, x, D1D)
-         {
-            MFEM_FOREACH_THREAD(d1, y, D1D)
-            {
-               u[d2][d1] = x(d1, d2, side,
-                             f); // copy transposed for better memory access
-               du[d2][d1] = dxdn(d1, d2, side, f);
-            }
-         }
-
-         MFEM_FOREACH_THREAD(p1, x, Q1D)
-         {
-            MFEM_FOREACH_THREAD(p2, y, Q1D)
-            {
-               for (int l = 0; l < 3; ++l)
-               {
-                  nJe[side][p2][p1][l] = pa(3 * side + l, p1, p2, f);
-               }
-
-               if (side == 0)
-               {
-                  kappa_Qh[p2][p1] = pa(6, p1, p2, f);
-               }
-            }
-         }
-
-         if (side == 0)
-         {
-            MFEM_FOREACH_THREAD(p, x, Q1D)
-            {
-               MFEM_FOREACH_THREAD(d, y, D1D)
-               {
-                  B(p, d) = B_(p, d);
-                  G(p, d) = G_(p, d);
-               }
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-
-      // eval u and normal derivative @ quad points
-      MFEM_FOREACH_THREAD(side, z, 2)
-      {
-         real_t(*u)[max_Q1D] = (side == 0) ? u0 : u1;
-         real_t(*du)[max_Q1D] = (side == 0) ? du0 : du1;
-         real_t(*Bu)[max_Q1D] = (side == 0) ? Bu0 : Bu1;
-         real_t(*Bdu)[max_Q1D] = (side == 0) ? Bdu0 : Bdu1;
-         real_t(*Gu)[max_Q1D] = (side == 0) ? Gu0 : Gu1;
-
-         MFEM_FOREACH_THREAD(p1, x, Q1D)
-         {
-            MFEM_FOREACH_THREAD(d2, y, D1D)
-            {
-               real_t bu = 0.0;
-               real_t bdu = 0.0;
-               real_t gu = 0.0;
-
-               for (int d1 = 0; d1 < D1D; ++d1)
-               {
-                  const real_t b = B(p1, d1);
-                  const real_t g = G(p1, d1);
-
-                  bu += b * u[d2][d1];
-                  bdu += b * du[d2][d1];
-                  gu += g * u[d2][d1];
-               }
-
-               Bu[p1][d2] = bu;
-               Bdu[p1][d2] = bdu;
-               Gu[p1][d2] = gu;
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-
-      MFEM_FOREACH_THREAD(side, z, 2)
-      {
-         real_t(*u)[max_Q1D] = (side == 0) ? u0 : u1;
-         real_t(*du)[max_Q1D] = (side == 0) ? du0 : du1;
-         real_t(*Bu)[max_Q1D] = (side == 0) ? Bu0 : Bu1;
-         real_t(*Gu)[max_Q1D] = (side == 0) ? Gu0 : Gu1;
-         real_t(*Bdu)[max_Q1D] = (side == 0) ? Bdu0 : Bdu1;
-
-         MFEM_FOREACH_THREAD(p2, x, Q1D)
-         {
-            MFEM_FOREACH_THREAD(p1, y, Q1D)
-            {
-               const real_t *Je = nJe[side][p2][p1];
-
-               real_t bbu = 0.0;
-               real_t bgu = 0.0;
-               real_t gbu = 0.0;
-               real_t bbdu = 0.0;
-
-               for (int d2 = 0; d2 < D1D; ++d2)
-               {
-                  const real_t b = B(p2, d2);
-                  const real_t g = G(p2, d2);
-                  bbu += b * Bu[p1][d2];
-                  gbu += g * Bu[p1][d2];
-                  bgu += b * Gu[p1][d2];
-                  bbdu += b * Bdu[p1][d2];
-               }
-
-               u[p2][p1] = bbu;
-               // du <- Q du/dn * w * det(J)
-               du[p2][p1] = Je[0] * bbdu + Je[1] * bgu + Je[2] * gbu;
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-
-      MFEM_FOREACH_THREAD(side, z, 2)
-      {
-         real_t(*Bj)[max_Q1D] = (side == 0) ? Bj0 : Bj1;
-         real_t(*Bjn)[max_Q1D] = (side == 0) ? Bjn0 : Bjn1;
-         real_t(*Gj)[max_Q1D] = (side == 0) ? Gj0 : Gj1;
-
-         MFEM_FOREACH_THREAD(d1, x, D1D)
-         {
-            MFEM_FOREACH_THREAD(p2, y, Q1D)
-            {
-               real_t bj = 0.0;
-               real_t bjn = 0.0;
-               real_t gj = 0.0;
-               real_t br = 0.0;
-
-               for (int p1 = 0; p1 < Q1D; ++p1)
-               {
-                  const real_t b = B(p1, d1);
-                  const real_t g = G(p1, d1);
-
-                  const real_t *Je = nJe[side][p2][p1];
-
-                  const real_t jump = u0[p2][p1] - u1[p2][p1];
-                  const real_t avg = du0[p2][p1] + du1[p2][p1];
-
-                  // r = - < {Q du/dn}, [v] > + kappa * < {Q/h} [u], [v] >
-                  const real_t r = -avg + kappa_Qh[p2][p1] * jump;
-
-                  // bj, gj, bjn contribute to sigma term
-                  bj += b * Je[0] * jump;
-                  gj += g * Je[1] * jump;
-                  bjn += b * Je[2] * jump;
-
-                  br += b * r;
-               }
-
-               Bj[d1][p2] = sigma * bj;
-               Bjn[d1][p2] = sigma * bjn;
-
-               // group br and gj together since we will multiply them both by B
-               // and then sum
-               const real_t sgn = (side == 0) ? 1.0 : -1.0;
-               Gj[d1][p2] = sgn * br + sigma * gj;
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-
-      MFEM_FOREACH_THREAD(side, z, 2)
-      {
-         real_t(*u)[max_Q1D] = (side == 0) ? u0 : u1;
-         real_t(*du)[max_Q1D] = (side == 0) ? du0 : du1;
-         real_t(*Bj)[max_Q1D] = (side == 0) ? Bj0 : Bj1;
-         real_t(*Bjn)[max_Q1D] = (side == 0) ? Bjn0 : Bjn1;
-         real_t(*Gj)[max_Q1D] = (side == 0) ? Gj0 : Gj1;
-
-         MFEM_FOREACH_THREAD(d2, x, D1D)
-         {
-            MFEM_FOREACH_THREAD(d1, y, D1D)
-            {
-               real_t bbj = 0.0;
-               real_t gbj = 0.0;
-               real_t bgj = 0.0;
-
-               for (int p2 = 0; p2 < Q1D; ++p2)
-               {
-                  const real_t b = B(p2, d2);
-                  const real_t g = G(p2, d2);
-
-                  bbj += b * Bj[d1][p2];
-                  bgj += b * Gj[d1][p2];
-                  gbj += g * Bjn[d1][p2];
-               }
-
-               du[d2][d1] = bbj;
-               u[d2][d1] = bgj + gbj;
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-
-      // map back to y and dydn
-      MFEM_FOREACH_THREAD(side, z, 2)
-      {
-         const real_t(*u)[max_Q1D] = (side == 0) ? u0 : u1;
-         const real_t(*du)[max_Q1D] = (side == 0) ? du0 : du1;
-
-         MFEM_FOREACH_THREAD(d2, x, D1D)
-         {
-            MFEM_FOREACH_THREAD(d1, y, D1D)
-            {
-               y(d1, d2, side, f) += u[d2][d1];
-               dydn(d1, d2, side, f) += du[d2][d1];
-            }
-         }
-      }
-   });
-}
-
-} // namespace internal
-
-template <int DIM, int D1D, int Q1D>
-DGDiffusionIntegrator::ApplyKernelType
-DGDiffusionIntegrator::ApplyPAKernels::Kernel()
-{
-   if constexpr (DIM == 2)
-   {
-      return internal::PADGDiffusionApply2D<D1D, Q1D>;
-   }
-   else if constexpr (DIM == 3)
-   {
-      return internal::PADGDiffusionApply3D<D1D, Q1D>;
-   }
-   MFEM_ABORT("");
-}
-} // namespace mfem
-/// \endcond DO_NOT_DOCUMENT
-#endif
@@ -11,39 +11,42 @@

 #include "../../general/forall.hpp"
 #include "../../mesh/face_nbr_geom.hpp"
-#include "../fe/face_map_utils.hpp"
 #include "../gridfunc.hpp"
 #include "../qfunction.hpp"
+#include "../fe/face_map_utils.hpp"

-#include "bilininteg_dgdiffusion_kernels.hpp"
+using namespace std;

 namespace mfem
 {

-static void PADGDiffusionSetup2D(const int Q1D, const int NE, const int NF,
+static void PADGDiffusionSetup2D(const int Q1D,
+                                 const int NE,
+                                 const int NF,
                                 const Array<real_t> &w,
                                 const GeometricFactors &el_geom,
                                 const FaceGeometricFactors &face_geom,
                                 const FaceNeighborGeometricFactors *nbr_geom,
-                                 const Vector &q, const real_t sigma,
-                                 const real_t kappa, Vector &pa_data,
+                                 const Vector &q,
+                                 const real_t sigma,
+                                 const real_t kappa,
+                                 Vector &pa_data,
                                 const Array<int> &face_info_)
 {
   const auto J_loc = Reshape(el_geom.J.Read(), Q1D, Q1D, 2, 2, NE);
   const auto detJe_loc = Reshape(el_geom.detJ.Read(), Q1D, Q1D, NE);

   const int n_nbr = nbr_geom ? nbr_geom->num_neighbor_elems : 0;
-   const auto J_shared =
-      Reshape(nbr_geom ? nbr_geom->J.Read() : nullptr, Q1D, Q1D, 2, 2, n_nbr);
-   const auto detJ_shared =
-      Reshape(nbr_geom ? nbr_geom->detJ.Read() : nullptr, Q1D, Q1D, n_nbr);
+   const auto J_shared = Reshape(nbr_geom ? nbr_geom->J.Read() : nullptr,
+                                 Q1D, Q1D, 2, 2, n_nbr);
+   const auto detJ_shared = Reshape(nbr_geom ? nbr_geom->detJ.Read() : nullptr,
+                                    Q1D, Q1D, n_nbr);

   const auto detJf = Reshape(face_geom.detJ.Read(), Q1D, NF);
   const auto n = Reshape(face_geom.normal.Read(), Q1D, 2, NF);

   const bool const_q = (q.Size() == 1);
-   const auto Q =
-      const_q ? Reshape(q.Read(), 1, 1) : Reshape(q.Read(), Q1D, NF);
+   const auto Q = const_q ? Reshape(q.Read(), 1,1) : Reshape(q.Read(), Q1D,NF);

   const auto W = w.Read();

@@ -53,7 +56,7 @@ static void PADGDiffusionSetup2D(const int Q1D, const int NE, const int NF,
   // (q, 1/h, J0_0, J0_1, J1_0, J1_1)
   auto pa = Reshape(pa_data.Write(), 6, Q1D, NF);

-   mfem::forall(NF, [=] MFEM_HOST_DEVICE(int f) -> void
+   mfem::forall(NF, [=] MFEM_HOST_DEVICE (int f) -> void
   {
      const int normal_dir[] = {face_info(0, f), face_info(1, f)};
      const int fid[] = {face_info(4, f), face_info(5, f)};
@@ -71,7 +74,7 @@ static void PADGDiffusionSetup2D(const int Q1D, const int NE, const int NF,

      for (int p = 0; p < Q1D; ++p)
      {
-         const real_t Qp = const_q ? Q(0, 0) : Q(p, f);
+         const real_t Qp = const_q ? Q(0,0) : Q(p, f);
         pa(0, p, f) = kappa * Qp * W[p] * detJf(p, f);

         real_t hi = 0.0;
@@ -82,19 +85,17 @@ static void PADGDiffusionSetup2D(const int Q1D, const int NE, const int NF,

            // Always opposite direction in "native" ordering
            // Need to multiply the native=>lex0 with native=>lex1 and negate
-            const int sgn = (side == 1) ? -1 * sgn0 * sgn1 : 1;
+            const int sgn = (side == 1) ? -1*sgn0*sgn1 : 1;

            const int e = el[side];
            const auto &J = (side == 1 && shared) ? J_shared : J_loc;
            const auto &detJ = (side == 1 && shared) ? detJ_shared : detJe_loc;

            real_t nJi[2];
-            nJi[0] =
-               n(p, 0, f) * J(i, j, 1, 1, e) - n(p, 1, f) * J(i, j, 0, 1, e);
-            nJi[1] =
-               -n(p, 0, f) * J(i, j, 1, 0, e) + n(p, 1, f) * J(i, j, 0, 0, e);
+            nJi[0] = n(p,0,f)*J(i,j, 1,1, e) - n(p,1,f)*J(i,j,0,1,e);
+            nJi[1] = -n(p,0,f)*J(i,j,1,0, e) + n(p,1,f)*J(i,j,0,0,e);

-            const real_t dJe = detJ(i, j, e);
+            const real_t dJe = detJ(i,j,e);
            const real_t dJf = detJf(p, f);

            const real_t w = factor * Qp * W[p] * dJf / dJe;
@@ -103,9 +104,9 @@ static void PADGDiffusionSetup2D(const int Q1D, const int NE, const int NF,
            const int ti = 1 - ni;

            // Normal
-            pa(2 + 2 * side + 0, p, f) = w * nJi[ni];
+            pa(2 + 2*side + 0, p, f) = w * nJi[ni];
            // Tangential
-            pa(2 + 2 * side + 1, p, f) = sgn * w * nJi[ti];
+            pa(2 + 2*side + 1, p, f) = sgn * w * nJi[ti];

            hi += factor * dJf / dJe;
         }
@@ -121,43 +122,47 @@ static void PADGDiffusionSetup2D(const int Q1D, const int NE, const int NF,
   });
 }

-static void PADGDiffusionSetup3D(const int Q1D, const int NE, const int NF,
+static void PADGDiffusionSetup3D(const int Q1D,
+                                 const int NE,
+                                 const int NF,
                                 const Array<real_t> &w,
                                 const GeometricFactors &el_geom,
                                 const FaceGeometricFactors &face_geom,
                                 const FaceNeighborGeometricFactors *nbr_geom,
-                                 const Vector &q, const real_t sigma,
-                                 const real_t kappa, Vector &pa_data,
+                                 const Vector &q,
+                                 const real_t sigma,
+                                 const real_t kappa,
+                                 Vector &pa_data,
                                 const Array<int> &face_info_)
 {
   const auto J_loc = Reshape(el_geom.J.Read(), Q1D, Q1D, Q1D, 3, 3, NE);
   const auto detJe_loc = Reshape(el_geom.detJ.Read(), Q1D, Q1D, Q1D, NE);

   const int n_nbr = nbr_geom ? nbr_geom->num_neighbor_elems : 0;
-   const auto J_shared = Reshape(nbr_geom ? nbr_geom->J.Read() : nullptr, Q1D,
-                                 Q1D, Q1D, 3, 3, n_nbr);
-   const auto detJ_shared =
-      Reshape(nbr_geom ? nbr_geom->detJ.Read() : nullptr, Q1D, Q1D, Q1D, n_nbr);
+   const auto J_shared = Reshape(nbr_geom ? nbr_geom->J.Read() : nullptr,
+                                 Q1D, Q1D, Q1D, 3, 3, n_nbr);
+   const auto detJ_shared = Reshape(nbr_geom ? nbr_geom->detJ.Read() : nullptr,
+                                    Q1D, Q1D, Q1D, n_nbr);

   const auto detJf = Reshape(face_geom.detJ.Read(), Q1D, Q1D, NF);
   const auto n = Reshape(face_geom.normal.Read(), Q1D, Q1D, 3, NF);

   const bool const_q = (q.Size() == 1);
-   const auto Q =
-      const_q ? Reshape(q.Read(), 1, 1, 1) : Reshape(q.Read(), Q1D, Q1D, NF);
+   const auto Q = const_q ? Reshape(q.Read(), 1, 1, 1)
+                  : Reshape(q.Read(), Q1D, Q1D, NF);

   const auto W = Reshape(w.Read(), Q1D, Q1D);

   // (perm[0], perm[1], perm[2], element_index, local_face_id, orientation)
   const auto face_info = Reshape(face_info_.Read(), 6, 2, NF);
-   constexpr int _el_ = 3;  // offset in face_info for element index
+   constexpr int _el_ = 3; // offset in face_info for element index
   constexpr int _fid_ = 4; // offset in face_info for local face id
-   constexpr int _or_ = 5;  // offset in face_info for orientation
+   constexpr int _or_ = 5; // offset in face_info for orientation

   // (J00, J01, J02, J10, J11, J12, q/h)
   const auto pa = Reshape(pa_data.Write(), 7, Q1D, Q1D, NF);

-   mfem::forall_2D(NF, Q1D, Q1D, [=] MFEM_HOST_DEVICE(int f) -> void
+   mfem::forall_2D(NF, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int f) -> void
   {
      MFEM_SHARED int perm[2][3];
      MFEM_SHARED int el[2];
@@ -167,7 +172,10 @@ static void PADGDiffusionSetup3D(const int Q1D, const int NE, const int NF,

      MFEM_FOREACH_THREAD(side, x, 2)
      {
-         MFEM_FOREACH_THREAD(i, y, 3) { perm[side][i] = face_info(i, side, f); }
+         MFEM_FOREACH_THREAD(i, y, 3)
+         {
+            perm[side][i] = face_info(i, side, f);
+         }

         if (MFEM_THREAD_ID(y) == 0)
         {
@@ -192,16 +200,16 @@ static void PADGDiffusionSetup3D(const int Q1D, const int NE, const int NF,
      {
         MFEM_FOREACH_THREAD(p2, y, Q1D)
         {
-            const real_t Qp = const_q ? Q(0, 0, 0) : Q(p1, p2, f);
-            const real_t dJf = detJf(p1, p2, f);
+            const real_t Qp = const_q ? Q(0,0,0) : Q(p1, p2, f);
+            const real_t dJf = detJf(p1,p2,f);

            real_t hi = 0.0;

            for (int side = 0; side < nsides; ++side)
            {
               int i, j, k;
-               internal::FaceIdxToVolIdx3D(p1 + Q1D * p2, Q1D, fid[0], fid[1],
-                                           side, ortn[1], i, j, k);
+               internal::FaceIdxToVolIdx3D(
+                  p1 + Q1D*p2, Q1D, fid[0], fid[1], side, ortn[1], i, j, k);

               const int e = el[side];
               const auto &J = shared[side] ? J_shared : J_loc;
@@ -209,45 +217,27 @@ static void PADGDiffusionSetup3D(const int Q1D, const int NE, const int NF,

               // *INDENT-OFF*
               real_t nJi[3];
-               nJi[0] = (-J(i, j, k, 1, 2, e) * J(i, j, k, 2, 1, e) +
-                         J(i, j, k, 1, 1, e) * J(i, j, k, 2, 2, e)) *
-                           n(p1, p2, 0, f) +
-                        (J(i, j, k, 0, 2, e) * J(i, j, k, 2, 1, e) -
-                         J(i, j, k, 0, 1, e) * J(i, j, k, 2, 2, e)) *
-                           n(p1, p2, 1, f) +
-                        (-J(i, j, k, 0, 2, e) * J(i, j, k, 1, 1, e) +
-                         J(i, j, k, 0, 1, e) * J(i, j, k, 1, 2, e)) *
-                           n(p1, p2, 2, f);
+               nJi[0] = (  -J(i,j,k, 1,2, e)*J(i,j,k, 2,1, e) + J(i,j,k, 1,1, e)*J(i,j,k, 2,2, e)) * n(p1,p2, 0, f)
+                        + ( J(i,j,k, 0,2, e)*J(i,j,k, 2,1, e) - J(i,j,k, 0,1, e)*J(i,j,k, 2,2, e)) * n(p1,p2, 1, f)
+                        + (-J(i,j,k, 0,2, e)*J(i,j,k, 1,1, e) + J(i,j,k, 0,1, e)*J(i,j,k, 1,2, e)) * n(p1,p2, 2, f);

-               nJi[1] = (J(i, j, k, 1, 2, e) * J(i, j, k, 2, 0, e) -
-                         J(i, j, k, 1, 0, e) * J(i, j, k, 2, 2, e)) *
-                           n(p1, p2, 0, f) +
-                        (-J(i, j, k, 0, 2, e) * J(i, j, k, 2, 0, e) +
-                         J(i, j, k, 0, 0, e) * J(i, j, k, 2, 2, e)) *
-                           n(p1, p2, 1, f) +
-                        (J(i, j, k, 0, 2, e) * J(i, j, k, 1, 0, e) -
-                         J(i, j, k, 0, 0, e) * J(i, j, k, 1, 2, e)) *
-                           n(p1, p2, 2, f);
+               nJi[1] = (   J(i,j,k, 1,2, e)*J(i,j,k, 2,0, e) - J(i,j,k, 1,0, e)*J(i,j,k, 2,2, e)) * n(p1,p2, 0, f)
+                        + (-J(i,j,k, 0,2, e)*J(i,j,k, 2,0, e) + J(i,j,k, 0,0, e)*J(i,j,k, 2,2, e)) * n(p1,p2, 1, f)
+                        + ( J(i,j,k, 0,2, e)*J(i,j,k, 1,0, e) - J(i,j,k, 0,0, e)*J(i,j,k, 1,2, e)) * n(p1,p2, 2, f);

-               nJi[2] = (-J(i, j, k, 1, 1, e) * J(i, j, k, 2, 0, e) +
-                         J(i, j, k, 1, 0, e) * J(i, j, k, 2, 1, e)) *
-                           n(p1, p2, 0, f) +
-                        (J(i, j, k, 0, 1, e) * J(i, j, k, 2, 0, e) -
-                         J(i, j, k, 0, 0, e) * J(i, j, k, 2, 1, e)) *
-                           n(p1, p2, 1, f) +
-                        (-J(i, j, k, 0, 1, e) * J(i, j, k, 1, 0, e) +
-                         J(i, j, k, 0, 0, e) * J(i, j, k, 1, 1, e)) *
-                           n(p1, p2, 2, f);
+               nJi[2] = (  -J(i,j,k, 1,1, e)*J(i,j,k, 2,0, e) + J(i,j,k, 1,0, e)*J(i,j,k, 2,1, e)) * n(p1,p2, 0, f)
+                        + ( J(i,j,k, 0,1, e)*J(i,j,k, 2,0, e) - J(i,j,k, 0,0, e)*J(i,j,k, 2,1, e)) * n(p1,p2, 1, f)
+                        + (-J(i,j,k, 0,1, e)*J(i,j,k, 1,0, e) + J(i,j,k, 0,0, e)*J(i,j,k, 1,1, e)) * n(p1,p2, 2, f);
               // *INDENT-ON*

-               const real_t dJe = detJe(i, j, k, e);
+               const real_t dJe = detJe(i,j,k,e);
               const real_t val = factor * Qp * W(p1, p2) * dJf / dJe;

               for (int d = 0; d < 3; ++d)
               {
                  const int idx = std::abs(perm[side][d]) - 1;
                  const int sgn = (perm[side][d] < 0) ? -1 : 1;
-                  pa(3 * side + d, p1, p2, f) = sgn * val * nJi[idx];
+                  pa(3*side + d, p1, p2, f) = sgn * val * nJi[idx];
               }

               hi += factor * dJf / dJe;
@@ -267,8 +257,7 @@ static void PADGDiffusionSetup3D(const int Q1D, const int NE, const int NF,
 }

 static void PADGDiffusionSetupFaceInfo2D(const int nf, const Mesh &mesh,
-                                         const FaceType type,
-                                         Array<int> &face_info_)
+                                         const FaceType type, Array<int> &face_info_)
 {
   const int ne = mesh.GetNE();

@@ -337,7 +326,8 @@ inline void FaceNormalPermutation(int perm[3], const int face_id)

 // Assigns to perm the permutation as in FaceNormalPermutation for the second
 // element on the face but signed to indicate the sign of the normal derivative.
-inline void SignedFaceNormalPermutation(int perm[3], const int face_id1,
+inline void SignedFaceNormalPermutation(int perm[3],
+                                        const int face_id1,
                                        const int face_id2,
                                        const int orientation)
 {
@@ -396,19 +386,17 @@ inline void SignedFaceNormalPermutation(int perm[3], const int face_id1,
 }

 static void PADGDiffusionSetupFaceInfo3D(const int nf, const Mesh &mesh,
-                                         const FaceType type,
-                                         Array<int> &face_info_)
+                                         const FaceType type, Array<int> &face_info_)
 {
   const int ne = mesh.GetNE();

   int fidx = 0;
-   // face_info array has 12 entries per face, 6 for each of the adjacent
-   // elements: (perm[0], perm[1], perm[2], element_index, local_face_id,
-   // orientation)
+   // face_info array has 12 entries per face, 6 for each of the adjacent elements:
+   // (perm[0], perm[1], perm[2], element_index, local_face_id, orientation)
   face_info_.SetSize(nf * 12);
-   constexpr int _e_ = 3;   // offset for element index
+   constexpr int _e_ = 3; // offset for element index
   constexpr int _fid_ = 4; // offset for local face id
-   constexpr int _or_ = 5;  // offset for orientation
+   constexpr int _or_ = 5; // offset for orientation

   auto face_info = Reshape(face_info_.HostWrite(), 6, 2, nf);
   for (int f = 0; f < mesh.GetNumFaces(); ++f)
@@ -420,9 +408,9 @@ static void PADGDiffusionSetupFaceInfo3D(const int nf, const Mesh &mesh,
         const int fid0 = f_info.element[0].local_face_id;
         const int or0 = f_info.element[0].orientation;

-         face_info(_e_, 0, fidx) = f_info.element[0].index;
+         face_info(  _e_, 0, fidx) = f_info.element[0].index;
         face_info(_fid_, 0, fidx) = fid0;
-         face_info(_or_, 0, fidx) = or0;
+         face_info( _or_, 0, fidx) = or0;

         FaceNormalPermutation(&face_info(0, 0, fidx), fid0);

@@ -433,17 +421,16 @@ static void PADGDiffusionSetupFaceInfo3D(const int nf, const Mesh &mesh,

            if (f_info.IsShared())
            {
-               face_info(_e_, 1, fidx) = ne + f_info.element[1].index;
+               face_info(  _e_, 1, fidx) = ne + f_info.element[1].index;
            }
            else
            {
-               face_info(_e_, 1, fidx) = f_info.element[1].index;
+               face_info(  _e_, 1, fidx) = f_info.element[1].index;
            }
            face_info(_fid_, 1, fidx) = fid1;
-            face_info(_or_, 1, fidx) = or1;
+            face_info( _or_, 1, fidx) = or1;

-            SignedFaceNormalPermutation(&face_info(0, 1, fidx), fid0, fid1,
-                                        or1);
+            SignedFaceNormalPermutation(&face_info(0, 1, fidx), fid0, fid1, or1);
         }
         else
         {
@@ -461,8 +448,8 @@ static void PADGDiffusionSetupFaceInfo3D(const int nf, const Mesh &mesh,
 void DGDiffusionIntegrator::SetupPA(const FiniteElementSpace &fes,
                                    FaceType type)
 {
-   const MemoryType mt =
-      (pa_mt == MemoryType::DEFAULT) ? Device::GetDeviceMemoryType() : pa_mt;
+   const MemoryType mt = (pa_mt == MemoryType::DEFAULT) ?
+                         Device::GetDeviceMemoryType() : pa_mt;

   const int ne = fes.GetNE();
   nf = fes.GetNFbyType(type);
@@ -471,17 +458,16 @@ void DGDiffusionIntegrator::SetupPA(const FiniteElementSpace &fes,
   Mesh &mesh = *fes.GetMesh();
   const Geometry::Type face_geom_type = mesh.GetTypicalFaceGeometry();
   const FiniteElement &el = *fes.GetTypicalTraceElement();
-   const int ir_order = IntRule
-                        ? IntRule->GetOrder()
+   const int ir_order = IntRule ? IntRule->GetOrder()
                        : GetRule(el.GetOrder(), face_geom_type).GetOrder();
   const IntegrationRule &ir = irs.Get(face_geom_type, ir_order);
   dim = mesh.Dimension();
-   const int q1d = (ir.GetOrder() + 3) / 2;
-   MFEM_ASSERT(q1d == pow(real_t(ir.Size()), 1.0 / (dim - 1)), "");
+   const int q1d = (ir.GetOrder() + 3)/2;
+   MFEM_ASSERT(q1d == pow(real_t(ir.Size()), 1.0/(dim - 1)), "");

   const auto vol_ir = irs.Get(mesh.GetTypicalElementGeometry(), ir_order);
-   const auto geom_flags =
-      GeometricFactors::JACOBIANS | GeometricFactors::DETERMINANTS;
+   const auto geom_flags = GeometricFactors::JACOBIANS |
+                           GeometricFactors::DETERMINANTS;
   const auto el_geom = mesh.GetGeometricFactors(vol_ir, geom_flags, mt);

   std::unique_ptr<FaceNeighborGeometricFactors> nbr_geom;
@@ -490,8 +476,8 @@ void DGDiffusionIntegrator::SetupPA(const FiniteElementSpace &fes,
      nbr_geom.reset(new FaceNeighborGeometricFactors(*el_geom));
   }

-   const auto face_geom_flags =
-      FaceGeometricFactors::DETERMINANTS | FaceGeometricFactors::NORMALS;
+   const auto face_geom_flags = FaceGeometricFactors::DETERMINANTS |
+                                FaceGeometricFactors::NORMALS;
   auto face_geom = mesh.GetFaceGeometricFactors(ir, face_geom_flags, type, mt);
   maps = &el.GetDofToQuad(ir, DofToQuad::TENSOR);
   dofs1D = maps->ndof;
@@ -503,18 +489,9 @@ void DGDiffusionIntegrator::SetupPA(const FiniteElementSpace &fes,
   // Evaluate the coefficient at the face quadrature points.
   FaceQuadratureSpace fqs(mesh, ir, type);
   CoefficientVector q(fqs, CoefficientStorage::COMPRESSED);
-   if (Q)
-   {
-      q.Project(*Q);
-   }
-   else if (MQ)
-   {
-      MFEM_ABORT("Not yet implemented"); /* q.Project(*MQ); */
-   }
-   else
-   {
-      q.SetConstant(1.0);
-   }
+   if (Q) { q.Project(*Q); }
+   else if (MQ) { MFEM_ABORT("Not yet implemented"); /* q.Project(*MQ); */ }
+   else { q.SetConstant(1.0); }

   Array<int> face_info;
   if (dim == 1)
@@ -524,16 +501,14 @@ void DGDiffusionIntegrator::SetupPA(const FiniteElementSpace &fes,
   else if (dim == 2)
   {
      PADGDiffusionSetupFaceInfo2D(nf, mesh, type, face_info);
-      PADGDiffusionSetup2D(quad1D, ne, nf, ir.GetWeights(), *el_geom,
-                           *face_geom, nbr_geom.get(), q, sigma, kappa, pa_data,
-                           face_info);
+      PADGDiffusionSetup2D(quad1D, ne, nf, ir.GetWeights(), *el_geom, *face_geom,
+                           nbr_geom.get(), q, sigma, kappa, pa_data, face_info);
   }
   else if (dim == 3)
   {
      PADGDiffusionSetupFaceInfo3D(nf, mesh, type, face_info);
-      PADGDiffusionSetup3D(quad1D, ne, nf, ir.GetWeights(), *el_geom,
-                           *face_geom, nbr_geom.get(), q, sigma, kappa, pa_data,
-                           face_info);
+      PADGDiffusionSetup3D(quad1D, ne, nf, ir.GetWeights(), *el_geom, *face_geom,
+                           nbr_geom.get(), q, sigma, kappa, pa_data, face_info);
   }
 }

@@ -549,76 +524,529 @@ void DGDiffusionIntegrator::AssemblePABoundaryFaces(
   SetupPA(fes, FaceType::Boundary);
 }

-void DGDiffusionIntegrator::AddMultPAFaceNormalDerivatives(const Vector &x,
-                                                           const Vector &dxdn,
-                                                           Vector &y,
-                                                           Vector &dydn) const
+template<int T_D1D = 0, int T_Q1D = 0> static
+void PADGDiffusionApply2D(const int NF,
+                          const Array<real_t> &b,
+                          const Array<real_t> &bt,
+                          const Array<real_t>& g,
+                          const Array<real_t>& gt,
+                          const real_t sigma,
+                          const Vector &pa_data,
+                          const Vector &x_,
+                          const Vector &dxdn_,
+                          Vector &y_,
+                          Vector &dydn_,
+                          const int d1d = 0,
+                          const int q1d = 0)
 {
-   ApplyPAKernels::Run(dim, dofs1D, quad1D, nf, maps->B, maps->Bt, maps->G,
-                       maps->Gt, sigma, pa_data, x, dxdn, y, dydn, dofs1D,
-                       quad1D);
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+
+   auto B_ = Reshape(b.Read(), Q1D, D1D);
+   auto G_ = Reshape(g.Read(), Q1D, D1D);
+
+   auto pa = Reshape(pa_data.Read(), 6, Q1D, NF); // (q, 1/h, J00, J01, J10, J11)
+
+   auto x =    Reshape(x_.Read(),         D1D, 2, NF);
+   auto y =    Reshape(y_.ReadWrite(),    D1D, 2, NF);
+   auto dxdn = Reshape(dxdn_.Read(),      D1D, 2, NF);
+   auto dydn = Reshape(dydn_.ReadWrite(), D1D, 2, NF);
+
+   const int NBX = std::max(D1D, Q1D);
+
+   mfem::forall_2D(NF, NBX, 2, [=] MFEM_HOST_DEVICE (int f) -> void
+   {
+      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
+      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+
+      MFEM_SHARED real_t u0[max_D1D];
+      MFEM_SHARED real_t u1[max_D1D];
+      MFEM_SHARED real_t du0[max_D1D];
+      MFEM_SHARED real_t du1[max_D1D];
+
+      MFEM_SHARED real_t Bu0[max_Q1D];
+      MFEM_SHARED real_t Bu1[max_Q1D];
+      MFEM_SHARED real_t Bdu0[max_Q1D];
+      MFEM_SHARED real_t Bdu1[max_Q1D];
+
+      MFEM_SHARED real_t r[max_Q1D];
+
+      MFEM_SHARED real_t BG[2*max_D1D*max_Q1D];
+      DeviceMatrix B(BG, Q1D, D1D);
+      DeviceMatrix G(BG + D1D*Q1D, Q1D, D1D);
+
+      if (MFEM_THREAD_ID(y) == 0)
+      {
+         MFEM_FOREACH_THREAD(p,x,Q1D)
+         {
+            for (int d = 0; d < D1D; ++d)
+            {
+               B(p,d) = B_(p,d);
+               G(p,d) = G_(p,d);
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      // copy edge values to u0, u1 and copy edge normals to du0, du1
+      MFEM_FOREACH_THREAD(side,y,2)
+      {
+         real_t *u = (side == 0) ? u0 : u1;
+         real_t *du = (side == 0) ? du0 : du1;
+         MFEM_FOREACH_THREAD(d,x,D1D)
+         {
+            u[d] = x(d, side, f);
+            du[d] = dxdn(d, side, f);
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      // eval @ quad points
+      MFEM_FOREACH_THREAD(side,y,2)
+      {
+         real_t *u = (side == 0) ? u0 : u1;
+         real_t *du = (side == 0) ? du0 : du1;
+         real_t *Bu = (side == 0) ? Bu0 : Bu1;
+         real_t *Bdu = (side == 0) ? Bdu0 : Bdu1;
+
+         MFEM_FOREACH_THREAD(p,x,Q1D)
+         {
+            const real_t Je_side[] = {pa(2 + 2*side, p, f), pa(2 + 2*side + 1, p, f)};
+
+            Bu[p] = 0.0;
+            Bdu[p] = 0.0;
+
+            for (int d = 0; d < D1D; ++d)
+            {
+               const real_t b = B(p,d);
+               const real_t g = G(p,d);
+
+               Bu[p] += b*u[d];
+               Bdu[p] += Je_side[0] * b * du[d] + Je_side[1] * g * u[d];
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      // term - < {Q du/dn}, [v] > +  kappa * < {Q/h} [u], [v] >:
+      if (MFEM_THREAD_ID(y) == 0)
+      {
+         MFEM_FOREACH_THREAD(p,x,Q1D)
+         {
+            const real_t q = pa(0, p, f);
+            const real_t hi = pa(1, p, f);
+            const real_t jump = Bu0[p] - Bu1[p];
+            const real_t avg = Bdu0[p] + Bdu1[p]; // = {Q du/dn} * w * det(J)
+            r[p] = -avg + hi * q * jump;
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      MFEM_FOREACH_THREAD(d,x,D1D)
+      {
+         real_t Br = 0.0;
+
+         for (int p = 0; p < Q1D; ++p)
+         {
+            Br += B(p, d) * r[p];
+         }
+
+         u0[d] =  Br; // overwrite u0, u1
+         u1[d] = -Br;
+      } // for d
+      MFEM_SYNC_THREAD;
+
+
+      MFEM_FOREACH_THREAD(side,y,2)
+      {
+         real_t *du = (side == 0) ? du0 : du1;
+         MFEM_FOREACH_THREAD(d,x,D1D)
+         {
+            du[d] = 0.0;
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      // term sigma * < [u], {Q dv/dn} >
+      MFEM_FOREACH_THREAD(side,y,2)
+      {
+         real_t * const du = (side == 0) ? du0 : du1;
+         real_t * const u = (side == 0) ? u0 : u1;
+
+         MFEM_FOREACH_THREAD(d,x,D1D)
+         {
+            for (int p = 0; p < Q1D; ++p)
+            {
+               const real_t Je[] = {pa(2 + 2*side, p, f), pa(2 + 2*side + 1, p, f)};
+               const real_t jump = Bu0[p] - Bu1[p];
+               const real_t r_p = Je[0] * jump; // normal
+               const real_t w_p = Je[1] * jump; // tangential
+               du[d] += sigma * B(p, d) * r_p;
+               u[d] += sigma * G(p, d) * w_p;
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      MFEM_FOREACH_THREAD(side,y,2)
+      {
+         real_t *u = (side == 0) ? u0 : u1;
+         real_t *du = (side == 0) ? du0 : du1;
+         MFEM_FOREACH_THREAD(d,x,D1D)
+         {
+            y(d, side, f) += u[d];
+            dydn(d, side, f) += du[d];
+         }
+      }
+   }); // mfem::forall
 }

-DGDiffusionIntegrator::DGDiffusionIntegrator(const real_t s, const real_t k)
-   : sigma(s), kappa(k)
+template <int T_D1D = 0, int T_Q1D = 0>
+static void PADGDiffusionApply3D(const int NF,
+                                 const Array<real_t>& b,
+                                 const Array<real_t>& bt,
+                                 const Array<real_t>& g,
+                                 const Array<real_t>& gt,
+                                 const real_t sigma,
+                                 const Vector& pa_data,
+                                 const Vector& x_,
+                                 const Vector& dxdn_,
+                                 Vector& y_,
+                                 Vector& dydn_,
+                                 const int d1d = 0,
+                                 const int q1d = 0)
 {
-   static Kernels kernels;
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+
+   auto B_ = Reshape(b.Read(), Q1D, D1D);
+   auto G_ = Reshape(g.Read(), Q1D, D1D);
+
+   // (J0[0], J0[1], J0[2], J1[0], J1[1], J1[2], q/h)
+   auto pa = Reshape(pa_data.Read(), 7, Q1D, Q1D, NF);
+
+   auto x =    Reshape(x_.Read(),         D1D, D1D, 2, NF);
+   auto y =    Reshape(y_.ReadWrite(),    D1D, D1D, 2, NF);
+   auto dxdn = Reshape(dxdn_.Read(),      D1D, D1D, 2, NF);
+   auto dydn = Reshape(dydn_.ReadWrite(), D1D, D1D, 2, NF);
+
+   const int NBX = std::max(D1D, Q1D);
+
+   mfem::forall_3D(NF, NBX, NBX, 2, [=] MFEM_HOST_DEVICE (int f) -> void
+   {
+      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
+      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+
+      MFEM_SHARED real_t u0[max_Q1D][max_Q1D];
+      MFEM_SHARED real_t u1[max_Q1D][max_Q1D];
+
+      MFEM_SHARED real_t du0[max_Q1D][max_Q1D];
+      MFEM_SHARED real_t du1[max_Q1D][max_Q1D];
+
+      MFEM_SHARED real_t Gu0[max_Q1D][max_Q1D];
+      MFEM_SHARED real_t Gu1[max_Q1D][max_Q1D];
+
+      MFEM_SHARED real_t Bu0[max_Q1D][max_Q1D];
+      MFEM_SHARED real_t Bu1[max_Q1D][max_Q1D];
+
+      MFEM_SHARED real_t Bdu0[max_Q1D][max_Q1D];
+      MFEM_SHARED real_t Bdu1[max_Q1D][max_Q1D];
+
+      MFEM_SHARED real_t kappa_Qh[max_Q1D][max_Q1D];
+
+      MFEM_SHARED real_t nJe[2][max_Q1D][max_Q1D][3];
+      MFEM_SHARED real_t BG[2*max_D1D*max_Q1D];
+
+      // some buffers are reused multiple times, but for clarity have new names:
+      real_t (*Bj0)[max_Q1D] = Bu0;
+      real_t (*Bj1)[max_Q1D] = Bu1;
+      real_t (*Bjn0)[max_Q1D] = Bdu0;
+      real_t (*Bjn1)[max_Q1D] = Bdu1;
+      real_t (*Gj0)[max_Q1D] = Gu0;
+      real_t (*Gj1)[max_Q1D] = Gu1;
+
+      DeviceMatrix B(BG, Q1D, D1D);
+      DeviceMatrix G(BG + D1D*Q1D, Q1D, D1D);
+
+      // copy face values to u0, u1 and copy normals to du0, du1
+      MFEM_FOREACH_THREAD(side, z, 2)
+      {
+         real_t (*u)[max_Q1D] = (side == 0) ? u0 : u1;
+         real_t (*du)[max_Q1D] = (side == 0) ? du0 : du1;
+
+         MFEM_FOREACH_THREAD(d2, x, D1D)
+         {
+            MFEM_FOREACH_THREAD(d1, y, D1D)
+            {
+               u[d2][d1] = x(d1, d2, side, f); // copy transposed for better memory access
+               du[d2][d1] = dxdn(d1, d2, side, f);
+            }
+         }
+
+         MFEM_FOREACH_THREAD(p1, x, Q1D)
+         {
+            MFEM_FOREACH_THREAD(p2, y, Q1D)
+            {
+               for (int l=0; l < 3; ++l)
+               {
+                  nJe[side][p2][p1][l] = pa(3*side + l, p1, p2, f);
+               }
+
+               if (side == 0)
+               {
+                  kappa_Qh[p2][p1] = pa(6, p1, p2, f);
+               }
+            }
+         }
+
+         if (side == 0)
+         {
+            MFEM_FOREACH_THREAD(p, x, Q1D)
+            {
+               MFEM_FOREACH_THREAD(d, y, D1D)
+               {
+                  B(p, d) = B_(p, d);
+                  G(p, d) = G_(p, d);
+               }
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      // eval u and normal derivative @ quad points
+      MFEM_FOREACH_THREAD(side, z, 2)
+      {
+         real_t (*u)[max_Q1D] = (side == 0) ? u0 : u1;
+         real_t (*du)[max_Q1D] = (side == 0) ? du0 : du1;
+         real_t (*Bu)[max_Q1D] = (side == 0) ? Bu0 : Bu1;
+         real_t (*Bdu)[max_Q1D] = (side == 0) ? Bdu0 : Bdu1;
+         real_t (*Gu)[max_Q1D] = (side == 0) ? Gu0 : Gu1;
+
+         MFEM_FOREACH_THREAD(p1, x, Q1D)
+         {
+            MFEM_FOREACH_THREAD(d2, y, D1D)
+            {
+               real_t bu = 0.0;
+               real_t bdu = 0.0;
+               real_t gu = 0.0;
+
+               for (int d1=0; d1 < D1D; ++d1)
+               {
+                  const real_t b = B(p1, d1);
+                  const real_t g = G(p1, d1);
+
+                  bu += b * u[d2][d1];
+                  bdu += b * du[d2][d1];
+                  gu += g * u[d2][d1];
+               }
+
+               Bu[p1][d2] = bu;
+               Bdu[p1][d2] = bdu;
+               Gu[p1][d2] = gu;
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      MFEM_FOREACH_THREAD(side, z, 2)
+      {
+         real_t (*u)[max_Q1D] = (side == 0) ? u0 : u1;
+         real_t (*du)[max_Q1D] = (side == 0) ? du0 : du1;
+         real_t (*Bu)[max_Q1D] = (side == 0) ? Bu0 : Bu1;
+         real_t (*Gu)[max_Q1D] = (side == 0) ? Gu0 : Gu1;
+         real_t (*Bdu)[max_Q1D] = (side == 0) ? Bdu0 : Bdu1;
+
+         MFEM_FOREACH_THREAD(p2, x, Q1D)
+         {
+            MFEM_FOREACH_THREAD(p1, y, Q1D)
+            {
+               const real_t * Je = nJe[side][p2][p1];
+
+               real_t bbu = 0.0;
+               real_t bgu = 0.0;
+               real_t gbu = 0.0;
+               real_t bbdu = 0.0;
+
+               for (int d2 = 0; d2 < D1D; ++d2)
+               {
+                  const real_t b = B(p2, d2);
+                  const real_t g = G(p2, d2);
+                  bbu += b * Bu[p1][d2];
+                  gbu += g * Bu[p1][d2];
+                  bgu += b * Gu[p1][d2];
+                  bbdu += b * Bdu[p1][d2];
+               }
+
+               u[p2][p1] = bbu;
+               // du <- Q du/dn * w * det(J)
+               du[p2][p1] = Je[0] * bbdu + Je[1] * bgu + Je[2] * gbu;
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      MFEM_FOREACH_THREAD(side, z, 2)
+      {
+         real_t (*Bj)[max_Q1D] = (side == 0) ? Bj0 : Bj1;
+         real_t (*Bjn)[max_Q1D] = (side == 0) ? Bjn0 : Bjn1;
+         real_t (*Gj)[max_Q1D] = (side == 0) ? Gj0 : Gj1;
+
+         MFEM_FOREACH_THREAD(d1, x, D1D)
+         {
+            MFEM_FOREACH_THREAD(p2, y, Q1D)
+            {
+               real_t bj = 0.0;
+               real_t bjn = 0.0;
+               real_t gj = 0.0;
+               real_t br = 0.0;
+
+               for (int p1 = 0; p1 < Q1D; ++p1)
+               {
+                  const real_t b = B(p1, d1);
+                  const real_t g = G(p1, d1);
+
+                  const real_t * Je = nJe[side][p2][p1];
+
+                  const real_t jump = u0[p2][p1] - u1[p2][p1];
+                  const real_t avg = du0[p2][p1] + du1[p2][p1];
+
+                  // r = - < {Q du/dn}, [v] > + kappa * < {Q/h} [u], [v] >
+                  const real_t r = -avg + kappa_Qh[p2][p1] * jump;
+
+                  // bj, gj, bjn contribute to sigma term
+                  bj += b * Je[0] * jump;
+                  gj += g * Je[1] * jump;
+                  bjn += b * Je[2] * jump;
+
+                  br += b * r;
+               }
+
+               Bj[d1][p2] = sigma * bj;
+               Bjn[d1][p2] = sigma * bjn;
+
+               // group br and gj together since we will multiply them both by B
+               // and then sum
+               const real_t sgn = (side == 0) ? 1.0 : -1.0;
+               Gj[d1][p2] = sgn * br + sigma * gj;
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      MFEM_FOREACH_THREAD(side, z, 2)
+      {
+         real_t (*u)[max_Q1D] = (side == 0) ? u0 : u1;
+         real_t (*du)[max_Q1D] = (side == 0) ? du0 : du1;
+         real_t (*Bj)[max_Q1D] = (side == 0) ? Bj0 : Bj1;
+         real_t (*Bjn)[max_Q1D] = (side == 0) ? Bjn0 : Bjn1;
+         real_t (*Gj)[max_Q1D] = (side == 0) ? Gj0 : Gj1;
+
+         MFEM_FOREACH_THREAD(d2, x, D1D)
+         {
+            MFEM_FOREACH_THREAD(d1, y, D1D)
+            {
+               real_t bbj = 0.0;
+               real_t gbj = 0.0;
+               real_t bgj = 0.0;
+
+               for (int p2 = 0; p2 < Q1D; ++p2)
+               {
+                  const real_t b = B(p2, d2);
+                  const real_t g = G(p2, d2);
+
+                  bbj += b * Bj[d1][p2];
+                  bgj += b * Gj[d1][p2];
+                  gbj += g * Bjn[d1][p2];
+               }
+
+               du[d2][d1] = bbj;
+               u[d2][d1] = bgj + gbj;
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      // map back to y and dydn
+      MFEM_FOREACH_THREAD(side, z, 2)
+      {
+         const real_t (*u)[max_Q1D] = (side == 0) ? u0 : u1;
+         const real_t (*du)[max_Q1D] = (side == 0) ? du0 : du1;
+
+         MFEM_FOREACH_THREAD(d2, x, D1D)
+         {
+            MFEM_FOREACH_THREAD(d1, y, D1D)
+            {
+               y(d1, d2, side, f) += u[d2][d1];
+               dydn(d1, d2, side, f) += du[d2][d1];
+            }
+         }
+      }
+   });
 }

-DGDiffusionIntegrator::DGDiffusionIntegrator(Coefficient &q, const real_t s,
-                                             const real_t k)
-   : DGDiffusionIntegrator(s, k)
-{
-   Q = &q;
-}
-
-DGDiffusionIntegrator::DGDiffusionIntegrator(MatrixCoefficient &q,
-                                             const real_t s, const real_t k)
-   : DGDiffusionIntegrator(s, k)
-{
-   MQ = &q;
-}
-
-/// \cond DO_NOT_DOCUMENT
-
-DGDiffusionIntegrator::ApplyKernelType
-DGDiffusionIntegrator::ApplyPAKernels::Fallback(int dim, int, int)
+static void PADGDiffusionApply(const int dim,
+                               const int D1D,
+                               const int Q1D,
+                               const int NF,
+                               const Array<real_t> &B,
+                               const Array<real_t> &Bt,
+                               const Array<real_t> &G,
+                               const Array<real_t> &Gt,
+                               const real_t sigma,
+                               const Vector &pa_data,
+                               const Vector &x,
+                               const Vector &dxdn,
+                               Vector &y,
+                               Vector &dydn)
 {
   if (dim == 2)
   {
-      return internal::PADGDiffusionApply2D;
+      auto kernel = PADGDiffusionApply2D<0,0>;
+      switch ((D1D << 4 ) | Q1D)
+      {
+         case 0x23: kernel = PADGDiffusionApply2D<2,3>; break;
+         case 0x34: kernel = PADGDiffusionApply2D<3,4>; break;
+         case 0x45: kernel = PADGDiffusionApply2D<4,5>; break;
+         case 0x56: kernel = PADGDiffusionApply2D<5,6>; break;
+         case 0x67: kernel = PADGDiffusionApply2D<6,7>; break;
+         case 0x78: kernel = PADGDiffusionApply2D<7,8>; break;
+         case 0x89: kernel = PADGDiffusionApply2D<8,9>; break;
+         case 0x9A: kernel = PADGDiffusionApply2D<9,10>; break;
+      }
+      kernel(NF, B, Bt, G, Gt, sigma, pa_data, x, dxdn, y, dydn, D1D, Q1D);
   }
   else if (dim == 3)
   {
-      return internal::PADGDiffusionApply3D;
+      auto kernel = PADGDiffusionApply3D<0,0>;
+      switch ((D1D << 4) | Q1D)
+      {
+         case 0x24: kernel = PADGDiffusionApply3D<2,4>; break;
+         case 0x35: kernel = PADGDiffusionApply3D<3,5>; break;
+         case 0x46: kernel = PADGDiffusionApply3D<4,6>; break;
+         case 0x57: kernel = PADGDiffusionApply3D<5,7>; break;
+         case 0x68: kernel = PADGDiffusionApply3D<6,8>; break;
+         case 0x79: kernel = PADGDiffusionApply3D<7,9>; break;
+         case 0x8A: kernel = PADGDiffusionApply3D<8,10>; break;
+         case 0x9B: kernel = PADGDiffusionApply3D<9,11>; break;
+      }
+      kernel(NF, B, Bt, G, Gt, sigma, pa_data, x, dxdn, y, dydn, D1D, Q1D);
   }
   else
   {
-      MFEM_ABORT("");
+      MFEM_ABORT("Unsupported dimension");
   }
 }

-DGDiffusionIntegrator::Kernels::Kernels()
+void DGDiffusionIntegrator::AddMultPAFaceNormalDerivatives(
+   const Vector &x, const Vector &dxdn, Vector &y, Vector &dydn) const
 {
-   DGDiffusionIntegrator::AddSpecialization<2, 2, 3>();
-   DGDiffusionIntegrator::AddSpecialization<2, 3, 4>();
-   DGDiffusionIntegrator::AddSpecialization<2, 4, 5>();
-   DGDiffusionIntegrator::AddSpecialization<2, 5, 6>();
-   DGDiffusionIntegrator::AddSpecialization<2, 6, 7>();
-   DGDiffusionIntegrator::AddSpecialization<2, 7, 8>();
-   DGDiffusionIntegrator::AddSpecialization<2, 8, 9>();
-   DGDiffusionIntegrator::AddSpecialization<2, 9, 10>();
-
-   DGDiffusionIntegrator::AddSpecialization<3, 2, 4>();
-   DGDiffusionIntegrator::AddSpecialization<3, 3, 5>();
-   DGDiffusionIntegrator::AddSpecialization<3, 4, 6>();
-   DGDiffusionIntegrator::AddSpecialization<3, 5, 7>();
-   DGDiffusionIntegrator::AddSpecialization<3, 6, 8>();
-   DGDiffusionIntegrator::AddSpecialization<3, 7, 9>();
-   DGDiffusionIntegrator::AddSpecialization<3, 8, 10>();
-   DGDiffusionIntegrator::AddSpecialization<3, 9, 11>();
+   PADGDiffusionApply(dim, dofs1D, quad1D, nf,
+                      maps->B, maps->Bt, maps->G, maps->Gt,
+                      sigma, pa_data, x, dxdn, y, dydn);
 }

-/// \endcond DO_NOT_DOCUMENT
-
 } // namespace mfem
@@ -1,793 +0,0 @@
-// 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 BILININTEG_DGTRACE_KERNELS_HPP
-#define BILININTEG_DGTRACE_KERNELS_HPP
-
-#include "../../general/forall.hpp"
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
-#include "../qfunction.hpp"
-#include "../restriction.hpp"
-
-/// \cond DO_NOT_DOCUMENT
-namespace mfem
-{
-
-namespace internal
-{
-
-// PA DGTrace Apply 2D kernel for Gauss-Lobatto/Bernstein
-template <int T_D1D = 0, int T_Q1D = 0>
-static void PADGTraceApply2D(const int NF, const Array<real_t> &b,
-                             const Array<real_t> &bt, const Vector &op_,
-                             const Vector &x_, Vector &y_, const int d1d = 0,
-                             const int q1d = 0)
-{
-   const int VDIM = 1;
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto Bt = Reshape(bt.Read(), D1D, Q1D);
-   auto op = Reshape(op_.Read(), Q1D, 2, 2, NF);
-   auto x = Reshape(x_.Read(), D1D, VDIM, 2, NF);
-   auto y = Reshape(y_.ReadWrite(), D1D, VDIM, 2, NF);
-
-   mfem::forall(NF, [=] MFEM_HOST_DEVICE(int f)
-   {
-      const int VDIM = 1;
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      // the following variables are evaluated at compile time
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-      real_t u0[max_D1D][VDIM];
-      real_t u1[max_D1D][VDIM];
-      for (int d = 0; d < D1D; d++)
-      {
-         for (int c = 0; c < VDIM; c++)
-         {
-            u0[d][c] = x(d, c, 0, f);
-            u1[d][c] = x(d, c, 1, f);
-         }
-      }
-      real_t Bu0[max_Q1D][VDIM];
-      real_t Bu1[max_Q1D][VDIM];
-      for (int q = 0; q < Q1D; ++q)
-      {
-         for (int c = 0; c < VDIM; c++)
-         {
-            Bu0[q][c] = 0.0;
-            Bu1[q][c] = 0.0;
-         }
-         for (int d = 0; d < D1D; ++d)
-         {
-            const real_t b = B(q, d);
-            for (int c = 0; c < VDIM; c++)
-            {
-               Bu0[q][c] += b * u0[d][c];
-               Bu1[q][c] += b * u1[d][c];
-            }
-         }
-      }
-      real_t DBu[max_Q1D][VDIM];
-      for (int q = 0; q < Q1D; ++q)
-      {
-         for (int c = 0; c < VDIM; c++)
-         {
-            DBu[q][c] = op(q, 0, 0, f) * Bu0[q][c] + op(q, 1, 0, f) * Bu1[q][c];
-         }
-      }
-      real_t BDBu[max_D1D][VDIM];
-      for (int d = 0; d < D1D; ++d)
-      {
-         for (int c = 0; c < VDIM; c++)
-         {
-            BDBu[d][c] = 0.0;
-         }
-         for (int q = 0; q < Q1D; ++q)
-         {
-            const real_t b = Bt(d, q);
-            for (int c = 0; c < VDIM; c++)
-            {
-               BDBu[d][c] += b * DBu[q][c];
-            }
-         }
-         for (int c = 0; c < VDIM; c++)
-         {
-            y(d, c, 0, f) += BDBu[d][c];
-            y(d, c, 1, f) += -BDBu[d][c];
-         }
-      }
-   });
-}
-
-// PA DGTrace Apply 3D kernel for Gauss-Lobatto/Bernstein
-template <int T_D1D = 0, int T_Q1D = 0>
-static void PADGTraceApply3D(const int NF, const Array<real_t> &b,
-                             const Array<real_t> &bt, const Vector &op_,
-                             const Vector &x_, Vector &y_, const int d1d = 0,
-                             const int q1d = 0)
-{
-   const int VDIM = 1;
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto Bt = Reshape(bt.Read(), D1D, Q1D);
-   auto op = Reshape(op_.Read(), Q1D, Q1D, 2, 2, NF);
-   auto x = Reshape(x_.Read(), D1D, D1D, VDIM, 2, NF);
-   auto y = Reshape(y_.ReadWrite(), D1D, D1D, VDIM, 2, NF);
-
-   mfem::forall(NF, [=] MFEM_HOST_DEVICE(int f)
-   {
-      const int VDIM = 1;
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      // the following variables are evaluated at compile time
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-      real_t u0[max_D1D][max_D1D][VDIM];
-      real_t u1[max_D1D][max_D1D][VDIM];
-      for (int d1 = 0; d1 < D1D; d1++)
-      {
-         for (int d2 = 0; d2 < D1D; d2++)
-         {
-            for (int c = 0; c < VDIM; c++)
-            {
-               u0[d1][d2][c] = x(d1, d2, c, 0, f);
-               u1[d1][d2][c] = x(d1, d2, c, 1, f);
-            }
-         }
-      }
-      real_t Bu0[max_Q1D][max_D1D][VDIM];
-      real_t Bu1[max_Q1D][max_D1D][VDIM];
-      for (int q = 0; q < Q1D; ++q)
-      {
-         for (int d2 = 0; d2 < D1D; d2++)
-         {
-            for (int c = 0; c < VDIM; c++)
-            {
-               Bu0[q][d2][c] = 0.0;
-               Bu1[q][d2][c] = 0.0;
-            }
-            for (int d1 = 0; d1 < D1D; ++d1)
-            {
-               const real_t b = B(q, d1);
-               for (int c = 0; c < VDIM; c++)
-               {
-                  Bu0[q][d2][c] += b * u0[d1][d2][c];
-                  Bu1[q][d2][c] += b * u1[d1][d2][c];
-               }
-            }
-         }
-      }
-      real_t BBu0[max_Q1D][max_Q1D][VDIM];
-      real_t BBu1[max_Q1D][max_Q1D][VDIM];
-      for (int q1 = 0; q1 < Q1D; ++q1)
-      {
-         for (int q2 = 0; q2 < Q1D; q2++)
-         {
-            for (int c = 0; c < VDIM; c++)
-            {
-               BBu0[q1][q2][c] = 0.0;
-               BBu1[q1][q2][c] = 0.0;
-            }
-            for (int d2 = 0; d2 < D1D; ++d2)
-            {
-               const real_t b = B(q2, d2);
-               for (int c = 0; c < VDIM; c++)
-               {
-                  BBu0[q1][q2][c] += b * Bu0[q1][d2][c];
-                  BBu1[q1][q2][c] += b * Bu1[q1][d2][c];
-               }
-            }
-         }
-      }
-      real_t DBBu[max_Q1D][max_Q1D][VDIM];
-      for (int q1 = 0; q1 < Q1D; ++q1)
-      {
-         for (int q2 = 0; q2 < Q1D; q2++)
-         {
-            for (int c = 0; c < VDIM; c++)
-            {
-               DBBu[q1][q2][c] = op(q1, q2, 0, 0, f) * BBu0[q1][q2][c] +
-                                 op(q1, q2, 1, 0, f) * BBu1[q1][q2][c];
-            }
-         }
-      }
-      real_t BDBBu[max_Q1D][max_D1D][VDIM];
-      for (int q1 = 0; q1 < Q1D; ++q1)
-      {
-         for (int d2 = 0; d2 < D1D; d2++)
-         {
-            for (int c = 0; c < VDIM; c++)
-            {
-               BDBBu[q1][d2][c] = 0.0;
-            }
-            for (int q2 = 0; q2 < Q1D; ++q2)
-            {
-               const real_t b = Bt(d2, q2);
-               for (int c = 0; c < VDIM; c++)
-               {
-                  BDBBu[q1][d2][c] += b * DBBu[q1][q2][c];
-               }
-            }
-         }
-      }
-      real_t BBDBBu[max_D1D][max_D1D][VDIM];
-      for (int d1 = 0; d1 < D1D; ++d1)
-      {
-         for (int d2 = 0; d2 < D1D; d2++)
-         {
-            for (int c = 0; c < VDIM; c++)
-            {
-               BBDBBu[d1][d2][c] = 0.0;
-            }
-            for (int q1 = 0; q1 < Q1D; ++q1)
-            {
-               const real_t b = Bt(d1, q1);
-               for (int c = 0; c < VDIM; c++)
-               {
-                  BBDBBu[d1][d2][c] += b * BDBBu[q1][d2][c];
-               }
-            }
-            for (int c = 0; c < VDIM; c++)
-            {
-               y(d1, d2, c, 0, f) += BBDBBu[d1][d2][c];
-               y(d1, d2, c, 1, f) += -BBDBBu[d1][d2][c];
-            }
-         }
-      }
-   });
-}
-
-// Optimized PA DGTrace Apply 3D kernel for Gauss-Lobatto/Bernstein
-template <int T_D1D = 0, int T_Q1D = 0, int T_NBZ = 0>
-static void SmemPADGTraceApply3D(const int NF, const Array<real_t> &b,
-                                 const Array<real_t> &bt, const Vector &op_,
-                                 const Vector &x_, Vector &y_,
-                                 const int d1d = 0, const int q1d = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   constexpr int NBZ = T_NBZ ? T_NBZ : 1;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto Bt = Reshape(bt.Read(), D1D, Q1D);
-   auto op = Reshape(op_.Read(), Q1D, Q1D, 2, 2, NF);
-   auto x = Reshape(x_.Read(), D1D, D1D, 2, NF);
-   auto y = Reshape(y_.ReadWrite(), D1D, D1D, 2, NF);
-
-   mfem::forall_2D_batch(NF, Q1D, Q1D, NBZ, [=] MFEM_HOST_DEVICE(int f)
-   {
-      const int tidz = MFEM_THREAD_ID(z);
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      // the following variables are evaluated at compile time
-      constexpr int NBZ = T_NBZ ? T_NBZ : 1;
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-      MFEM_SHARED real_t u0[NBZ][max_D1D][max_D1D];
-      MFEM_SHARED real_t u1[NBZ][max_D1D][max_D1D];
-      MFEM_FOREACH_THREAD(d1, x, D1D)
-      {
-         MFEM_FOREACH_THREAD(d2, y, D1D)
-         {
-            u0[tidz][d1][d2] = x(d1, d2, 0, f);
-            u1[tidz][d1][d2] = x(d1, d2, 1, f);
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_SHARED real_t Bu0[NBZ][max_Q1D][max_D1D];
-      MFEM_SHARED real_t Bu1[NBZ][max_Q1D][max_D1D];
-      MFEM_FOREACH_THREAD(q1, x, Q1D)
-      {
-         MFEM_FOREACH_THREAD(d2, y, D1D)
-         {
-            real_t Bu0_ = 0.0;
-            real_t Bu1_ = 0.0;
-            for (int d1 = 0; d1 < D1D; ++d1)
-            {
-               const real_t b = B(q1, d1);
-               Bu0_ += b * u0[tidz][d1][d2];
-               Bu1_ += b * u1[tidz][d1][d2];
-            }
-            Bu0[tidz][q1][d2] = Bu0_;
-            Bu1[tidz][q1][d2] = Bu1_;
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_SHARED real_t BBu0[NBZ][max_Q1D][max_Q1D];
-      MFEM_SHARED real_t BBu1[NBZ][max_Q1D][max_Q1D];
-      MFEM_FOREACH_THREAD(q1, x, Q1D)
-      {
-         MFEM_FOREACH_THREAD(q2, y, Q1D)
-         {
-            real_t BBu0_ = 0.0;
-            real_t BBu1_ = 0.0;
-            for (int d2 = 0; d2 < D1D; ++d2)
-            {
-               const real_t b = B(q2, d2);
-               BBu0_ += b * Bu0[tidz][q1][d2];
-               BBu1_ += b * Bu1[tidz][q1][d2];
-            }
-            BBu0[tidz][q1][q2] = BBu0_;
-            BBu1[tidz][q1][q2] = BBu1_;
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_SHARED real_t DBBu[NBZ][max_Q1D][max_Q1D];
-      MFEM_FOREACH_THREAD(q1, x, Q1D)
-      {
-         MFEM_FOREACH_THREAD(q2, y, Q1D)
-         {
-            DBBu[tidz][q1][q2] = op(q1, q2, 0, 0, f) * BBu0[tidz][q1][q2] +
-                                 op(q1, q2, 1, 0, f) * BBu1[tidz][q1][q2];
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_SHARED real_t BDBBu[NBZ][max_Q1D][max_D1D];
-      MFEM_FOREACH_THREAD(q1, x, Q1D)
-      {
-         MFEM_FOREACH_THREAD(d2, y, D1D)
-         {
-            real_t BDBBu_ = 0.0;
-            for (int q2 = 0; q2 < Q1D; ++q2)
-            {
-               const real_t b = Bt(d2, q2);
-               BDBBu_ += b * DBBu[tidz][q1][q2];
-            }
-            BDBBu[tidz][q1][d2] = BDBBu_;
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(d1, x, D1D)
-      {
-         MFEM_FOREACH_THREAD(d2, y, D1D)
-         {
-            real_t BBDBBu_ = 0.0;
-            for (int q1 = 0; q1 < Q1D; ++q1)
-            {
-               const real_t b = Bt(d1, q1);
-               BBDBBu_ += b * BDBBu[tidz][q1][d2];
-            }
-            y(d1, d2, 0, f) += BBDBBu_;
-            y(d1, d2, 1, f) += -BBDBBu_;
-         }
-      }
-   });
-}
-
-// PA DGTrace Apply 2D kernel for Gauss-Lobatto/Bernstein
-template <int T_D1D = 0, int T_Q1D = 0>
-static void PADGTraceApplyTranspose2D(const int NF, const Array<real_t> &b,
-                                      const Array<real_t> &bt,
-                                      const Vector &op_, const Vector &x_,
-                                      Vector &y_, const int d1d = 0,
-                                      const int q1d = 0)
-{
-   const int VDIM = 1;
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto Bt = Reshape(bt.Read(), D1D, Q1D);
-   auto op = Reshape(op_.Read(), Q1D, 2, 2, NF);
-   auto x = Reshape(x_.Read(), D1D, VDIM, 2, NF);
-   auto y = Reshape(y_.ReadWrite(), D1D, VDIM, 2, NF);
-
-   mfem::forall(NF, [=] MFEM_HOST_DEVICE(int f)
-   {
-      const int VDIM = 1;
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      // the following variables are evaluated at compile time
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-      real_t u0[max_D1D][VDIM];
-      real_t u1[max_D1D][VDIM];
-      for (int d = 0; d < D1D; d++)
-      {
-         for (int c = 0; c < VDIM; c++)
-         {
-            u0[d][c] = x(d, c, 0, f);
-            u1[d][c] = x(d, c, 1, f);
-         }
-      }
-      real_t Bu0[max_Q1D][VDIM];
-      real_t Bu1[max_Q1D][VDIM];
-      for (int q = 0; q < Q1D; ++q)
-      {
-         for (int c = 0; c < VDIM; c++)
-         {
-            Bu0[q][c] = 0.0;
-            Bu1[q][c] = 0.0;
-         }
-         for (int d = 0; d < D1D; ++d)
-         {
-            const real_t b = B(q, d);
-            for (int c = 0; c < VDIM; c++)
-            {
-               Bu0[q][c] += b * u0[d][c];
-               Bu1[q][c] += b * u1[d][c];
-            }
-         }
-      }
-      real_t DBu0[max_Q1D][VDIM];
-      real_t DBu1[max_Q1D][VDIM];
-      for (int q = 0; q < Q1D; ++q)
-      {
-         for (int c = 0; c < VDIM; c++)
-         {
-            DBu0[q][c] =
-               op(q, 0, 0, f) * Bu0[q][c] + op(q, 0, 1, f) * Bu1[q][c];
-            DBu1[q][c] =
-               op(q, 1, 0, f) * Bu0[q][c] + op(q, 1, 1, f) * Bu1[q][c];
-         }
-      }
-      real_t BDBu0[max_D1D][VDIM];
-      real_t BDBu1[max_D1D][VDIM];
-      for (int d = 0; d < D1D; ++d)
-      {
-         for (int c = 0; c < VDIM; c++)
-         {
-            BDBu0[d][c] = 0.0;
-            BDBu1[d][c] = 0.0;
-         }
-         for (int q = 0; q < Q1D; ++q)
-         {
-            const real_t b = Bt(d, q);
-            for (int c = 0; c < VDIM; c++)
-            {
-               BDBu0[d][c] += b * DBu0[q][c];
-               BDBu1[d][c] += b * DBu1[q][c];
-            }
-         }
-         for (int c = 0; c < VDIM; c++)
-         {
-            y(d, c, 0, f) += BDBu0[d][c];
-            y(d, c, 1, f) += BDBu1[d][c];
-         }
-      }
-   });
-}
-
-// PA DGTrace Apply Transpose 3D kernel for Gauss-Lobatto/Bernstein
-template <int T_D1D = 0, int T_Q1D = 0>
-static void PADGTraceApplyTranspose3D(const int NF, const Array<real_t> &b,
-                                      const Array<real_t> &bt,
-                                      const Vector &op_, const Vector &x_,
-                                      Vector &y_, const int d1d = 0,
-                                      const int q1d = 0)
-{
-   const int VDIM = 1;
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto Bt = Reshape(bt.Read(), D1D, Q1D);
-   auto op = Reshape(op_.Read(), Q1D, Q1D, 2, 2, NF);
-   auto x = Reshape(x_.Read(), D1D, D1D, VDIM, 2, NF);
-   auto y = Reshape(y_.ReadWrite(), D1D, D1D, VDIM, 2, NF);
-
-   mfem::forall(NF, [=] MFEM_HOST_DEVICE(int f)
-   {
-      const int VDIM = 1;
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      // the following variables are evaluated at compile time
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-      real_t u0[max_D1D][max_D1D][VDIM];
-      real_t u1[max_D1D][max_D1D][VDIM];
-      for (int d1 = 0; d1 < D1D; d1++)
-      {
-         for (int d2 = 0; d2 < D1D; d2++)
-         {
-            for (int c = 0; c < VDIM; c++)
-            {
-               u0[d1][d2][c] = x(d1, d2, c, 0, f);
-               u1[d1][d2][c] = x(d1, d2, c, 1, f);
-            }
-         }
-      }
-      real_t Bu0[max_Q1D][max_D1D][VDIM];
-      real_t Bu1[max_Q1D][max_D1D][VDIM];
-      for (int q1 = 0; q1 < Q1D; ++q1)
-      {
-         for (int d2 = 0; d2 < D1D; ++d2)
-         {
-            for (int c = 0; c < VDIM; c++)
-            {
-               Bu0[q1][d2][c] = 0.0;
-               Bu1[q1][d2][c] = 0.0;
-            }
-            for (int d1 = 0; d1 < D1D; ++d1)
-            {
-               const real_t b = B(q1, d1);
-               for (int c = 0; c < VDIM; c++)
-               {
-                  Bu0[q1][d2][c] += b * u0[d1][d2][c];
-                  Bu1[q1][d2][c] += b * u1[d1][d2][c];
-               }
-            }
-         }
-      }
-      real_t BBu0[max_Q1D][max_Q1D][VDIM];
-      real_t BBu1[max_Q1D][max_Q1D][VDIM];
-      for (int q1 = 0; q1 < Q1D; ++q1)
-      {
-         for (int q2 = 0; q2 < Q1D; ++q2)
-         {
-            for (int c = 0; c < VDIM; c++)
-            {
-               BBu0[q1][q2][c] = 0.0;
-               BBu1[q1][q2][c] = 0.0;
-            }
-            for (int d2 = 0; d2 < D1D; ++d2)
-            {
-               const real_t b = B(q2, d2);
-               for (int c = 0; c < VDIM; c++)
-               {
-                  BBu0[q1][q2][c] += b * Bu0[q1][d2][c];
-                  BBu1[q1][q2][c] += b * Bu1[q1][d2][c];
-               }
-            }
-         }
-      }
-      real_t DBu0[max_Q1D][max_Q1D][VDIM];
-      real_t DBu1[max_Q1D][max_Q1D][VDIM];
-      for (int q1 = 0; q1 < Q1D; ++q1)
-      {
-         for (int q2 = 0; q2 < Q1D; ++q2)
-         {
-            const real_t D00 = op(q1, q2, 0, 0, f);
-            const real_t D01 = op(q1, q2, 0, 1, f);
-            const real_t D10 = op(q1, q2, 1, 0, f);
-            const real_t D11 = op(q1, q2, 1, 1, f);
-            for (int c = 0; c < VDIM; c++)
-            {
-               DBu0[q1][q2][c] = D00 * BBu0[q1][q2][c] + D01 * BBu1[q1][q2][c];
-               DBu1[q1][q2][c] = D10 * BBu0[q1][q2][c] + D11 * BBu1[q1][q2][c];
-            }
-         }
-      }
-      real_t BDBu0[max_D1D][max_Q1D][VDIM];
-      real_t BDBu1[max_D1D][max_Q1D][VDIM];
-      for (int d1 = 0; d1 < D1D; ++d1)
-      {
-         for (int q2 = 0; q2 < Q1D; ++q2)
-         {
-            for (int c = 0; c < VDIM; c++)
-            {
-               BDBu0[d1][q2][c] = 0.0;
-               BDBu1[d1][q2][c] = 0.0;
-            }
-            for (int q1 = 0; q1 < Q1D; ++q1)
-            {
-               const real_t b = Bt(d1, q1);
-               for (int c = 0; c < VDIM; c++)
-               {
-                  BDBu0[d1][q2][c] += b * DBu0[q1][q2][c];
-                  BDBu1[d1][q2][c] += b * DBu1[q1][q2][c];
-               }
-            }
-         }
-      }
-      real_t BBDBu0[max_D1D][max_D1D][VDIM];
-      real_t BBDBu1[max_D1D][max_D1D][VDIM];
-      for (int d1 = 0; d1 < D1D; ++d1)
-      {
-         for (int d2 = 0; d2 < D1D; ++d2)
-         {
-            for (int c = 0; c < VDIM; c++)
-            {
-               BBDBu0[d1][d2][c] = 0.0;
-               BBDBu1[d1][d2][c] = 0.0;
-            }
-            for (int q2 = 0; q2 < Q1D; ++q2)
-            {
-               const real_t b = Bt(d2, q2);
-               for (int c = 0; c < VDIM; c++)
-               {
-                  BBDBu0[d1][d2][c] += b * BDBu0[d1][q2][c];
-                  BBDBu1[d1][d2][c] += b * BDBu1[d1][q2][c];
-               }
-            }
-            for (int c = 0; c < VDIM; c++)
-            {
-               y(d1, d2, c, 0, f) += BBDBu0[d1][d2][c];
-               y(d1, d2, c, 1, f) += BBDBu1[d1][d2][c];
-            }
-         }
-      }
-   });
-}
-
-// Optimized PA DGTrace Apply Transpose 3D kernel for Gauss-Lobatto/Bernstein
-template <int T_D1D = 0, int T_Q1D = 0, int T_NBZ = 0>
-static void SmemPADGTraceApplyTranspose3D(const int NF, const Array<real_t> &b,
-                                          const Array<real_t> &bt,
-                                          const Vector &op_, const Vector &x_,
-                                          Vector &y_, const int d1d = 0,
-                                          const int q1d = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   constexpr int NBZ = T_NBZ ? T_NBZ : 1;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto Bt = Reshape(bt.Read(), D1D, Q1D);
-   auto op = Reshape(op_.Read(), Q1D, Q1D, 2, 2, NF);
-   auto x = Reshape(x_.Read(), D1D, D1D, 2, NF);
-   auto y = Reshape(y_.ReadWrite(), D1D, D1D, 2, NF);
-
-   mfem::forall_2D_batch(NF, Q1D, Q1D, NBZ, [=] MFEM_HOST_DEVICE(int f)
-   {
-      const int tidz = MFEM_THREAD_ID(z);
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      // the following variables are evaluated at compile time
-      constexpr int NBZ = T_NBZ ? T_NBZ : 1;
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-      MFEM_SHARED real_t u0[NBZ][max_D1D][max_D1D];
-      MFEM_SHARED real_t u1[NBZ][max_D1D][max_D1D];
-      MFEM_FOREACH_THREAD(d1, x, D1D)
-      {
-         MFEM_FOREACH_THREAD(d2, y, D1D)
-         {
-            u0[tidz][d1][d2] = x(d1, d2, 0, f);
-            u1[tidz][d1][d2] = x(d1, d2, 1, f);
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_SHARED real_t Bu0[NBZ][max_Q1D][max_D1D];
-      MFEM_SHARED real_t Bu1[NBZ][max_Q1D][max_D1D];
-      MFEM_FOREACH_THREAD(q1, x, Q1D)
-      {
-         MFEM_FOREACH_THREAD(d2, y, D1D)
-         {
-            real_t Bu0_ = 0.0;
-            real_t Bu1_ = 0.0;
-            for (int d1 = 0; d1 < D1D; ++d1)
-            {
-               const real_t b = B(q1, d1);
-               Bu0_ += b * u0[tidz][d1][d2];
-               Bu1_ += b * u1[tidz][d1][d2];
-            }
-            Bu0[tidz][q1][d2] = Bu0_;
-            Bu1[tidz][q1][d2] = Bu1_;
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_SHARED real_t BBu0[NBZ][max_Q1D][max_Q1D];
-      MFEM_SHARED real_t BBu1[NBZ][max_Q1D][max_Q1D];
-      MFEM_FOREACH_THREAD(q1, x, Q1D)
-      {
-         MFEM_FOREACH_THREAD(q2, y, Q1D)
-         {
-            real_t BBu0_ = 0.0;
-            real_t BBu1_ = 0.0;
-            for (int d2 = 0; d2 < D1D; ++d2)
-            {
-               const real_t b = B(q2, d2);
-               BBu0_ += b * Bu0[tidz][q1][d2];
-               BBu1_ += b * Bu1[tidz][q1][d2];
-            }
-            BBu0[tidz][q1][q2] = BBu0_;
-            BBu1[tidz][q1][q2] = BBu1_;
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_SHARED real_t DBBu0[NBZ][max_Q1D][max_Q1D];
-      MFEM_SHARED real_t DBBu1[NBZ][max_Q1D][max_Q1D];
-      MFEM_FOREACH_THREAD(q1, x, Q1D)
-      {
-         MFEM_FOREACH_THREAD(q2, y, Q1D)
-         {
-            const real_t D00 = op(q1, q2, 0, 0, f);
-            const real_t D01 = op(q1, q2, 0, 1, f);
-            const real_t D10 = op(q1, q2, 1, 0, f);
-            const real_t D11 = op(q1, q2, 1, 1, f);
-            const real_t u0q = BBu0[tidz][q1][q2];
-            const real_t u1q = BBu1[tidz][q1][q2];
-            DBBu0[tidz][q1][q2] = D00 * u0q + D01 * u1q;
-            DBBu1[tidz][q1][q2] = D10 * u0q + D11 * u1q;
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_SHARED real_t BDBBu0[NBZ][max_Q1D][max_D1D];
-      MFEM_SHARED real_t BDBBu1[NBZ][max_Q1D][max_D1D];
-      MFEM_FOREACH_THREAD(q1, x, Q1D)
-      {
-         MFEM_FOREACH_THREAD(d2, y, D1D)
-         {
-            real_t BDBBu0_ = 0.0;
-            real_t BDBBu1_ = 0.0;
-            for (int q2 = 0; q2 < Q1D; ++q2)
-            {
-               const real_t b = Bt(d2, q2);
-               BDBBu0_ += b * DBBu0[tidz][q1][q2];
-               BDBBu1_ += b * DBBu1[tidz][q1][q2];
-            }
-            BDBBu0[tidz][q1][d2] = BDBBu0_;
-            BDBBu1[tidz][q1][d2] = BDBBu1_;
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(d1, x, D1D)
-      {
-         MFEM_FOREACH_THREAD(d2, y, D1D)
-         {
-            real_t BBDBBu0_ = 0.0;
-            real_t BBDBBu1_ = 0.0;
-            for (int q1 = 0; q1 < Q1D; ++q1)
-            {
-               const real_t b = Bt(d1, q1);
-               BBDBBu0_ += b * BDBBu0[tidz][q1][d2];
-               BBDBBu1_ += b * BDBBu1[tidz][q1][d2];
-            }
-            y(d1, d2, 0, f) += BBDBBu0_;
-            y(d1, d2, 1, f) += BBDBBu1_;
-         }
-      }
-   });
-}
-
-} // namespace internal
-
-template <int DIM, int D1D, int Q1D>
-DGTraceIntegrator::ApplyKernelType DGTraceIntegrator::ApplyPAKernels::Kernel()
-{
-   if constexpr (DIM == 2)
-   {
-      return internal::PADGTraceApply2D<D1D, Q1D>;
-   }
-   else if constexpr (DIM == 3)
-   {
-      if constexpr (D1D == 3 || D1D == 4)
-      {
-         return internal::SmemPADGTraceApply3D<D1D, Q1D, 2>;
-      }
-      else
-      {
-         return internal::SmemPADGTraceApply3D<D1D, Q1D>;
-      }
-   }
-   MFEM_ABORT("");
-}
-
-template <int DIM, int D1D, int Q1D>
-DGTraceIntegrator::ApplyKernelType DGTraceIntegrator::ApplyPATKernels::Kernel()
-{
-   if constexpr (DIM == 2)
-   {
-      return internal::PADGTraceApplyTranspose2D<D1D, Q1D>;
-   }
-   else if constexpr (DIM == 3)
-   {
-      return internal::SmemPADGTraceApplyTranspose3D<D1D, Q1D>;
-   }
-   MFEM_ABORT("");
-}
-} // namespace mfem
-
-/// \endcond DO_NOT_DOCUMENT
-#endif
@@ -19,8 +19,6 @@ namespace mfem
 DiffusionIntegrator::Kernels::Kernels()
 {
   // 2D
-   // Q = P+1
-   DiffusionIntegrator::AddSpecialization<2,1,1>();
   DiffusionIntegrator::AddSpecialization<2,2,2>();
   DiffusionIntegrator::AddSpecialization<2,3,3>();
   DiffusionIntegrator::AddSpecialization<2,4,4>();
@@ -29,39 +27,17 @@ DiffusionIntegrator::Kernels::Kernels()
   DiffusionIntegrator::AddSpecialization<2,7,7>();
   DiffusionIntegrator::AddSpecialization<2,8,8>();
   DiffusionIntegrator::AddSpecialization<2,9,9>();
-   // Q = P+2
-   DiffusionIntegrator::AddSpecialization<2,1,2>();
-   DiffusionIntegrator::AddSpecialization<2,2,3>();
-   DiffusionIntegrator::AddSpecialization<2,3,4>();
-   DiffusionIntegrator::AddSpecialization<2,4,5>();
-   DiffusionIntegrator::AddSpecialization<2,5,6>();
-   DiffusionIntegrator::AddSpecialization<2,6,7>();
-   DiffusionIntegrator::AddSpecialization<2,7,8>();
-   DiffusionIntegrator::AddSpecialization<2,8,9>();
-   DiffusionIntegrator::AddSpecialization<2,9,10>();
-   // others
   // 3D
-   // Q = P+1
-   DiffusionIntegrator::AddSpecialization<3,1,1>();
   DiffusionIntegrator::AddSpecialization<3,2,2>();
-   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>();
-   DiffusionIntegrator::AddSpecialization<3,8,8>();
-   // Q = P+2
-   DiffusionIntegrator::AddSpecialization<3,1,2>();
   DiffusionIntegrator::AddSpecialization<3,2,3>();
   DiffusionIntegrator::AddSpecialization<3,3,4>();
   DiffusionIntegrator::AddSpecialization<3,4,5>();
+   DiffusionIntegrator::AddSpecialization<3,4,6>();
   DiffusionIntegrator::AddSpecialization<3,5,6>();
+   DiffusionIntegrator::AddSpecialization<3,5,8>();
   DiffusionIntegrator::AddSpecialization<3,6,7>();
   DiffusionIntegrator::AddSpecialization<3,7,8>();
   DiffusionIntegrator::AddSpecialization<3,8,9>();
-   // others
-   DiffusionIntegrator::AddSpecialization<3,4,6>();
-   DiffusionIntegrator::AddSpecialization<3,5,8>();
 }

 namespace internal
@@ -672,12 +672,12 @@ inline void SmemPADiffusionApply2D(const int NE,
      real_t (*Gt)[MQ1] = (real_t (*)[MQ1]) (sBG+1);
      MFEM_SHARED real_t Xz[NBZ][MD1][MD1];
      MFEM_SHARED real_t GD[2][NBZ][MD1][MQ1];
-      MFEM_SHARED real_t GQ[2][NBZ][MQ1][MQ1];
+      MFEM_SHARED real_t GQ[2][NBZ][MD1][MQ1];
      real_t (*X)[MD1] = (real_t (*)[MD1])(Xz + tidz);
-      real_t (*DQ0)[MQ1] = (real_t (*)[MQ1])(GD[0] + tidz);
-      real_t (*DQ1)[MQ1] = (real_t (*)[MQ1])(GD[1] + tidz);
-      real_t (*QQ0)[MQ1] = (real_t (*)[MQ1])(GQ[0] + tidz);
-      real_t (*QQ1)[MQ1] = (real_t (*)[MQ1])(GQ[1] + tidz);
+      real_t (*DQ0)[MD1] = (real_t (*)[MD1])(GD[0] + tidz);
+      real_t (*DQ1)[MD1] = (real_t (*)[MD1])(GD[1] + tidz);
+      real_t (*QQ0)[MD1] = (real_t (*)[MD1])(GQ[0] + tidz);
+      real_t (*QQ1)[MD1] = (real_t (*)[MD1])(GQ[1] + tidz);
      MFEM_FOREACH_THREAD(dy,y,D1D)
      {
         MFEM_FOREACH_THREAD(dx,x,D1D)
@@ -1221,9 +1221,9 @@ using DiagonalKernelType = DiffusionIntegrator::DiagonalKernelType;
 template<int DIM, int T_D1D, int T_Q1D>
 ApplyKernelType DiffusionIntegrator::ApplyPAKernels::Kernel()
 {
-   if constexpr (DIM == 2) { return internal::SmemPADiffusionApply2D<T_D1D,T_Q1D>; }
-   else if constexpr (DIM == 3) { return internal::SmemPADiffusionApply3D<T_D1D, T_Q1D>; }
-   MFEM_ABORT("");
+   if (DIM == 2) { return internal::SmemPADiffusionApply2D<T_D1D,T_Q1D>; }
+   else if (DIM == 3) { return internal::SmemPADiffusionApply3D<T_D1D, T_Q1D>; }
+   else { MFEM_ABORT(""); }
 }

 inline
@@ -1237,9 +1237,9 @@ ApplyKernelType DiffusionIntegrator::ApplyPAKernels::Fallback(int DIM, int, int)
 template<int DIM, int D1D, int Q1D>
 DiagonalKernelType DiffusionIntegrator::DiagonalPAKernels::Kernel()
 {
-   if constexpr (DIM == 2) { return internal::SmemPADiffusionDiagonal2D<D1D,Q1D>; }
-   else if constexpr (DIM == 3) { return internal::SmemPADiffusionDiagonal3D<D1D, Q1D>; }
-   MFEM_ABORT("");
+   if (DIM == 2) { return internal::SmemPADiffusionDiagonal2D<D1D,Q1D>; }
+   else if (DIM == 3) { return internal::SmemPADiffusionDiagonal3D<D1D, Q1D>; }
+   else { MFEM_ABORT(""); }
 }

 inline DiagonalKernelType
@@ -599,11 +599,13 @@ void PACurlCurlSetup3D(const int Q1D,
   });
 }

-void PACurlCurlAssembleDiagonal2D(const int D1D, const int Q1D, const bool,
-                                  const int NE, const Array<real_t> &bo,
-                                  const Array<real_t> &, const Array<real_t> &,
+void PACurlCurlAssembleDiagonal2D(const int D1D,
+                                  const int Q1D,
+                                  const int NE,
+                                  const Array<real_t> &bo,
                                  const Array<real_t> &gc,
-                                  const Vector &pa_data, Vector &diag)
+                                  const Vector &pa_data,
+                                  Vector &diag)
 {
   auto Bo = Reshape(bo.Read(), Q1D, D1D-1);
   auto Gc = Reshape(gc.Read(), Q1D, D1D);
@@ -651,11 +653,16 @@ void PACurlCurlAssembleDiagonal2D(const int D1D, const int Q1D, const bool,
   }); // end of element loop
 }

-void PACurlCurlApply2D(const int D1D, const int Q1D, const bool, const int NE,
-                       const Array<real_t> &bo, const Array<real_t> &,
-                       const Array<real_t> &bot, const Array<real_t> &,
-                       const Array<real_t> &gc, const Array<real_t> &gct,
-                       const Vector &pa_data, const Vector &x, Vector &y,
+void PACurlCurlApply2D(const int D1D,
+                       const int Q1D,
+                       const int NE,
+                       const Array<real_t> &bo,
+                       const Array<real_t> &bot,
+                       const Array<real_t> &gc,
+                       const Array<real_t> &gct,
+                       const Vector &pa_data,
+                       const Vector &x,
+                       Vector &y,
                       const bool useAbs)
 {

@@ -24,7 +24,7 @@

 namespace mfem
 {
-/// \cond DO_NOT_DOCUMENT
+
 namespace internal
 {

@@ -426,11 +426,8 @@ void PACurlCurlSetup3D(const int Q1D,
 // PA H(curl) curl-curl Diagonal 2D kernel
 void PACurlCurlAssembleDiagonal2D(const int D1D,
                                  const int Q1D,
-                                  const bool symmetric, // unused
                                  const int NE,
                                  const Array<real_t> &bo,
-                                  const Array<real_t> &bc, // unused
-                                  const Array<real_t> &go, // unused
                                  const Array<real_t> &gc,
                                  const Vector &pa_data,
                                  Vector &diag);
@@ -834,12 +831,9 @@ inline void SmemPACurlCurlAssembleDiagonal3D(const int d1d,
 // PA H(curl) curl-curl Apply/AbsApply 2D kernel
 void PACurlCurlApply2D(const int D1D,
                       const int Q1D,
-                       const bool symmetric, // unused
                       const int NE,
                       const Array<real_t> &bo,
-                       const Array<real_t> &bc, // unused
                       const Array<real_t> &bot,
-                       const Array<real_t> &bct, // unused
                       const Array<real_t> &gc,
                       const Array<real_t> &gct,
                       const Vector &pa_data,
@@ -3164,49 +3158,6 @@ inline void SmemPAHcurlL2ApplyTranspose3D(const int d1d,

 } // namespace internal

-template<int DIM, int T_D1D, int T_Q1D>
-CurlCurlIntegrator::ApplyKernelType CurlCurlIntegrator::ApplyPAKernels::Kernel()
-{
-   if constexpr (DIM == 2)
-   {
-      return internal::PACurlCurlApply2D;
-   }
-   else if constexpr (DIM == 3)
-   {
-      if (Device::Allows(Backend::DEVICE_MASK))
-      {
-         return internal::SmemPACurlCurlApply3D<T_D1D, T_Q1D>;
-      }
-      else
-      {
-         return internal::PACurlCurlApply3D;
-      }
-   }
-   MFEM_ABORT("");
-}
-
-template <int DIM, int T_D1D, int T_Q1D>
-CurlCurlIntegrator::DiagonalKernelType
-CurlCurlIntegrator::DiagonalPAKernels::Kernel()
-{
-   if constexpr (DIM == 2)
-   {
-      return internal::PACurlCurlAssembleDiagonal2D;
-   }
-   else if constexpr (DIM == 3)
-   {
-      if (Device::Allows(Backend::DEVICE_MASK))
-      {
-         return internal::SmemPACurlCurlAssembleDiagonal3D<T_D1D, T_Q1D>;
-      }
-      else
-      {
-         return internal::PACurlCurlAssembleDiagonal3D;
-      }
-   }
-   MFEM_ABORT("");
-}
-/// \endcond DO_NOT_DOCUMENT
 } // namespace mfem

 #endif
@@ -19,7 +19,6 @@
 #include "../../linalg/vector.hpp"
 #include "../bilininteg.hpp"

-/// \cond DO_NOT_DOCUMENT
 namespace mfem
 {

@@ -820,6 +819,4 @@ inline void PAHcurlHdivApplyTranspose3D(const int d1d,

 } // namespace mfem

-/// \endcond DO_NOT_DOCUMENT
-
 #endif
@@ -17,8 +17,6 @@ namespace mfem
 MassIntegrator::Kernels::Kernels()
 {
   // 2D
-   // Q=P+1
-   MassIntegrator::AddSpecialization<2,1,1>();
   MassIntegrator::AddSpecialization<2,2,2>();
   MassIntegrator::AddSpecialization<2,3,3>();
   MassIntegrator::AddSpecialization<2,4,4>();
@@ -27,45 +25,17 @@ MassIntegrator::Kernels::Kernels()
   MassIntegrator::AddSpecialization<2,7,7>();
   MassIntegrator::AddSpecialization<2,8,8>();
   MassIntegrator::AddSpecialization<2,9,9>();
-   // Q=P+2
-   MassIntegrator::AddSpecialization<2,1,2>();
-   MassIntegrator::AddSpecialization<2,2,3>();
-   MassIntegrator::AddSpecialization<2,3,4>();
-   MassIntegrator::AddSpecialization<2,4,5>();
-   MassIntegrator::AddSpecialization<2,5,6>();
-   MassIntegrator::AddSpecialization<2,6,7>();
-   MassIntegrator::AddSpecialization<2,7,8>();
-   MassIntegrator::AddSpecialization<2,8,9>();
-   MassIntegrator::AddSpecialization<2,9,10>();
-   // others
-   MassIntegrator::AddSpecialization<2,2,4>();
-   MassIntegrator::AddSpecialization<2,3,6>();
-   MassIntegrator::AddSpecialization<2,4,6>();
   // 3D
-   // Q=P+1
-   MassIntegrator::AddSpecialization<3,1,1>();
   MassIntegrator::AddSpecialization<3,2,2>();
-   MassIntegrator::AddSpecialization<3,3,3>();
-   MassIntegrator::AddSpecialization<3,4,4>();
-   MassIntegrator::AddSpecialization<3,5,5>();
-   MassIntegrator::AddSpecialization<3,6,6>();
-   MassIntegrator::AddSpecialization<3,7,7>();
-   MassIntegrator::AddSpecialization<3,8,8>();
-   MassIntegrator::AddSpecialization<3,9,9>();
-   // Q=P+2
-   MassIntegrator::AddSpecialization<3,1,2>();
   MassIntegrator::AddSpecialization<3,2,3>();
   MassIntegrator::AddSpecialization<3,3,4>();
-   MassIntegrator::AddSpecialization<3,3,6>();
   MassIntegrator::AddSpecialization<3,4,5>();
+   MassIntegrator::AddSpecialization<3,4,6>();
   MassIntegrator::AddSpecialization<3,5,6>();
+   MassIntegrator::AddSpecialization<3,5,8>();
   MassIntegrator::AddSpecialization<3,6,7>();
   MassIntegrator::AddSpecialization<3,7,8>();
   MassIntegrator::AddSpecialization<3,8,9>();
-   // others
-   MassIntegrator::AddSpecialization<3,2,4>();
-   MassIntegrator::AddSpecialization<3,4,6>();
-   MassIntegrator::AddSpecialization<3,5,8>();
 }

 namespace internal
@@ -1392,10 +1392,10 @@ using DiagonalKernelType = MassIntegrator::DiagonalKernelType;
 template<int DIM, int T_D1D, int T_Q1D>
 ApplyKernelType MassIntegrator::ApplyPAKernels::Kernel()
 {
-   if constexpr (DIM == 1) { return internal::PAMassApply1D; }
-   else if constexpr (DIM == 2) { return internal::SmemPAMassApply2D<T_D1D,T_Q1D>; }
-   else if constexpr (DIM == 3) { return internal::SmemPAMassApply3D<T_D1D, T_Q1D>; }
-   MFEM_ABORT("");
+   if (DIM == 1) { return internal::PAMassApply1D; }
+   else if (DIM == 2) { return internal::SmemPAMassApply2D<T_D1D,T_Q1D>; }
+   else if (DIM == 3) { return internal::SmemPAMassApply3D<T_D1D, T_Q1D>; }
+   else { MFEM_ABORT(""); }
 }

 inline ApplyKernelType MassIntegrator::ApplyPAKernels::Fallback(
@@ -1410,10 +1410,10 @@ inline ApplyKernelType MassIntegrator::ApplyPAKernels::Fallback(
 template<int DIM, int T_D1D, int T_Q1D>
 DiagonalKernelType MassIntegrator::DiagonalPAKernels::Kernel()
 {
-   if constexpr (DIM == 1) { return internal::PAMassAssembleDiagonal1D; }
-   else if constexpr (DIM == 2) { return internal::SmemPAMassAssembleDiagonal2D<T_D1D,T_Q1D>; }
-   else if constexpr (DIM == 3) { return internal::SmemPAMassAssembleDiagonal3D<T_D1D, T_Q1D>; }
-   MFEM_ABORT("");
+   if (DIM == 1) { return internal::PAMassAssembleDiagonal1D; }
+   else if (DIM == 2) { return internal::SmemPAMassAssembleDiagonal2D<T_D1D,T_Q1D>; }
+   else if (DIM == 3) { return internal::SmemPAMassAssembleDiagonal3D<T_D1D, T_Q1D>; }
+   else { MFEM_ABORT(""); }
 }

 inline DiagonalKernelType MassIntegrator::DiagonalPAKernels::Fallback(
@@ -59,23 +59,26 @@ 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)
   {
-      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)
+      MFEM_FOREACH_THREAD(i, x, NQ)
      {
-         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);
-      });
-   }
+         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);
+      }
+   });
 }

 void MassIntegrator::AssemblePABoundary(const FiniteElementSpace &fes)
@@ -106,22 +109,50 @@ void MassIntegrator::AssemblePABoundary(const FiniteElementSpace &fes)
   CoefficientVector coeff(Q, qs, CoefficientStorage::COMPRESSED);

   const int NE = ne;
-   const int NQ = nq;
+   const int Q1D = quad1D;
   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(), 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)
+      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 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);
+         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);
+         }
      });
   }
+   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)
@@ -1,355 +0,0 @@
-// 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_BILININTEG_VECDIFFUSION_KERNELS_HPP
-#define MFEM_BILININTEG_VECDIFFUSION_KERNELS_HPP
-
-#include "../../general/forall.hpp"
-#include "../bilininteg.hpp"
-#include "../ceed/integrators/diffusion/diffusion.hpp"
-#include "../gridfunc.hpp"
-#include "../qfunction.hpp"
-/// \cond DO_NOT_DOCUMENT
-namespace mfem
-{
-
-namespace internal
-{
-
-// PA Diffusion Apply 2D kernel
-template <int T_D1D = 0, int T_Q1D = 0, int T_VDIM = 0>
-static void
-PAVectorDiffusionApply2D(const int NE, const Array<real_t> &b,
-                         const Array<real_t> &g, const Array<real_t> &bt,
-                         const Array<real_t> &gt, const Vector &d_,
-                         const Vector &x_, Vector &y_, const int d1d = 0,
-                         const int q1d = 0, const int vdim = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   const int VDIM = T_VDIM ? T_VDIM : vdim;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto G = Reshape(g.Read(), Q1D, D1D);
-   auto Bt = Reshape(bt.Read(), D1D, Q1D);
-   auto Gt = Reshape(gt.Read(), D1D, Q1D);
-   auto D = Reshape(d_.Read(), Q1D * Q1D, 3, NE);
-   auto x = Reshape(x_.Read(), D1D, D1D, VDIM, NE);
-   auto y = Reshape(y_.ReadWrite(), D1D, D1D, VDIM, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE(int e)
-   {
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      const int VDIM = T_VDIM ? T_VDIM : vdim;
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-
-      real_t grad[max_Q1D][max_Q1D][2];
-      for (int c = 0; c < VDIM; c++)
-      {
-         for (int qy = 0; qy < Q1D; ++qy)
-         {
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               grad[qy][qx][0] = 0.0;
-               grad[qy][qx][1] = 0.0;
-            }
-         }
-         for (int dy = 0; dy < D1D; ++dy)
-         {
-            real_t gradX[max_Q1D][2];
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               gradX[qx][0] = 0.0;
-               gradX[qx][1] = 0.0;
-            }
-            for (int dx = 0; dx < D1D; ++dx)
-            {
-               const real_t s = x(dx, dy, c, e);
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  gradX[qx][0] += s * B(qx, dx);
-                  gradX[qx][1] += s * G(qx, dx);
-               }
-            }
-            for (int qy = 0; qy < Q1D; ++qy)
-            {
-               const real_t wy = B(qy, dy);
-               const real_t wDy = G(qy, dy);
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  grad[qy][qx][0] += gradX[qx][1] * wy;
-                  grad[qy][qx][1] += gradX[qx][0] * wDy;
-               }
-            }
-         }
-         // Calculate Dxy, xDy in plane
-         for (int qy = 0; qy < Q1D; ++qy)
-         {
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               const int q = qx + qy * Q1D;
-               const real_t O11 = D(q, 0, e);
-               const real_t O12 = D(q, 1, e);
-               const real_t O22 = D(q, 2, e);
-               const real_t gradX = grad[qy][qx][0];
-               const real_t gradY = grad[qy][qx][1];
-               grad[qy][qx][0] = (O11 * gradX) + (O12 * gradY);
-               grad[qy][qx][1] = (O12 * gradX) + (O22 * gradY);
-            }
-         }
-         for (int qy = 0; qy < Q1D; ++qy)
-         {
-            real_t gradX[max_D1D][2];
-            for (int dx = 0; dx < D1D; ++dx)
-            {
-               gradX[dx][0] = 0.0;
-               gradX[dx][1] = 0.0;
-            }
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               const real_t gX = grad[qy][qx][0];
-               const real_t gY = grad[qy][qx][1];
-               for (int dx = 0; dx < D1D; ++dx)
-               {
-                  const real_t wx = Bt(dx, qx);
-                  const real_t wDx = Gt(dx, qx);
-                  gradX[dx][0] += gX * wDx;
-                  gradX[dx][1] += gY * wx;
-               }
-            }
-            for (int dy = 0; dy < D1D; ++dy)
-            {
-               const real_t wy = Bt(dy, qy);
-               const real_t wDy = Gt(dy, qy);
-               for (int dx = 0; dx < D1D; ++dx)
-               {
-                  y(dx, dy, c, e) +=
-                     ((gradX[dx][0] * wy) + (gradX[dx][1] * wDy));
-               }
-            }
-         }
-      }
-   });
-}
-
-// PA Diffusion Apply 3D kernel
-template <const int T_D1D = 0, const int T_Q1D = 0>
-static void
-PAVectorDiffusionApply3D(const int NE, const Array<real_t> &b,
-                         const Array<real_t> &g, const Array<real_t> &bt,
-                         const Array<real_t> &gt, const Vector &op_,
-                         const Vector &x_, Vector &y_, const int d1d = 0,
-                         const int q1d = 0, const int sdim = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   constexpr int VDIM = 3;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto G = Reshape(g.Read(), Q1D, D1D);
-   auto Bt = Reshape(bt.Read(), D1D, Q1D);
-   auto Gt = Reshape(gt.Read(), D1D, Q1D);
-   auto op = Reshape(op_.Read(), Q1D * Q1D * Q1D, 6, NE);
-   auto x = Reshape(x_.Read(), D1D, D1D, D1D, VDIM, NE);
-   auto y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, VDIM, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE(int e)
-   {
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-      for (int c = 0; c < VDIM; ++c)
-      {
-         real_t grad[max_Q1D][max_Q1D][max_Q1D][3];
-         for (int qz = 0; qz < Q1D; ++qz)
-         {
-            for (int qy = 0; qy < Q1D; ++qy)
-            {
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  grad[qz][qy][qx][0] = 0.0;
-                  grad[qz][qy][qx][1] = 0.0;
-                  grad[qz][qy][qx][2] = 0.0;
-               }
-            }
-         }
-         for (int dz = 0; dz < D1D; ++dz)
-         {
-            real_t gradXY[max_Q1D][max_Q1D][3];
-            for (int qy = 0; qy < Q1D; ++qy)
-            {
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  gradXY[qy][qx][0] = 0.0;
-                  gradXY[qy][qx][1] = 0.0;
-                  gradXY[qy][qx][2] = 0.0;
-               }
-            }
-            for (int dy = 0; dy < D1D; ++dy)
-            {
-               real_t gradX[max_Q1D][2];
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  gradX[qx][0] = 0.0;
-                  gradX[qx][1] = 0.0;
-               }
-               for (int dx = 0; dx < D1D; ++dx)
-               {
-                  const real_t s = x(dx, dy, dz, c, e);
-                  for (int qx = 0; qx < Q1D; ++qx)
-                  {
-                     gradX[qx][0] += s * B(qx, dx);
-                     gradX[qx][1] += s * G(qx, dx);
-                  }
-               }
-               for (int qy = 0; qy < Q1D; ++qy)
-               {
-                  const real_t wy = B(qy, dy);
-                  const real_t wDy = G(qy, dy);
-                  for (int qx = 0; qx < Q1D; ++qx)
-                  {
-                     const real_t wx = gradX[qx][0];
-                     const real_t wDx = gradX[qx][1];
-                     gradXY[qy][qx][0] += wDx * wy;
-                     gradXY[qy][qx][1] += wx * wDy;
-                     gradXY[qy][qx][2] += wx * wy;
-                  }
-               }
-            }
-            for (int qz = 0; qz < Q1D; ++qz)
-            {
-               const real_t wz = B(qz, dz);
-               const real_t wDz = G(qz, dz);
-               for (int qy = 0; qy < Q1D; ++qy)
-               {
-                  for (int qx = 0; qx < Q1D; ++qx)
-                  {
-                     grad[qz][qy][qx][0] += gradXY[qy][qx][0] * wz;
-                     grad[qz][qy][qx][1] += gradXY[qy][qx][1] * wz;
-                     grad[qz][qy][qx][2] += gradXY[qy][qx][2] * wDz;
-                  }
-               }
-            }
-         }
-         // Calculate Dxyz, xDyz, xyDz in plane
-         for (int qz = 0; qz < Q1D; ++qz)
-         {
-            for (int qy = 0; qy < Q1D; ++qy)
-            {
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  const int q = qx + (qy + qz * Q1D) * Q1D;
-                  const real_t O11 = op(q, 0, e);
-                  const real_t O12 = op(q, 1, e);
-                  const real_t O13 = op(q, 2, e);
-                  const real_t O22 = op(q, 3, e);
-                  const real_t O23 = op(q, 4, e);
-                  const real_t O33 = op(q, 5, e);
-                  const real_t gradX = grad[qz][qy][qx][0];
-                  const real_t gradY = grad[qz][qy][qx][1];
-                  const real_t gradZ = grad[qz][qy][qx][2];
-                  grad[qz][qy][qx][0] =
-                     (O11 * gradX) + (O12 * gradY) + (O13 * gradZ);
-                  grad[qz][qy][qx][1] =
-                     (O12 * gradX) + (O22 * gradY) + (O23 * gradZ);
-                  grad[qz][qy][qx][2] =
-                     (O13 * gradX) + (O23 * gradY) + (O33 * gradZ);
-               }
-            }
-         }
-         for (int qz = 0; qz < Q1D; ++qz)
-         {
-            real_t gradXY[max_D1D][max_D1D][3];
-            for (int dy = 0; dy < D1D; ++dy)
-            {
-               for (int dx = 0; dx < D1D; ++dx)
-               {
-                  gradXY[dy][dx][0] = 0;
-                  gradXY[dy][dx][1] = 0;
-                  gradXY[dy][dx][2] = 0;
-               }
-            }
-            for (int qy = 0; qy < Q1D; ++qy)
-            {
-               real_t gradX[max_D1D][3];
-               for (int dx = 0; dx < D1D; ++dx)
-               {
-                  gradX[dx][0] = 0;
-                  gradX[dx][1] = 0;
-                  gradX[dx][2] = 0;
-               }
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  const real_t gX = grad[qz][qy][qx][0];
-                  const real_t gY = grad[qz][qy][qx][1];
-                  const real_t gZ = grad[qz][qy][qx][2];
-                  for (int dx = 0; dx < D1D; ++dx)
-                  {
-                     const real_t wx = Bt(dx, qx);
-                     const real_t wDx = Gt(dx, qx);
-                     gradX[dx][0] += gX * wDx;
-                     gradX[dx][1] += gY * wx;
-                     gradX[dx][2] += gZ * wx;
-                  }
-               }
-               for (int dy = 0; dy < D1D; ++dy)
-               {
-                  const real_t wy = Bt(dy, qy);
-                  const real_t wDy = Gt(dy, qy);
-                  for (int dx = 0; dx < D1D; ++dx)
-                  {
-                     gradXY[dy][dx][0] += gradX[dx][0] * wy;
-                     gradXY[dy][dx][1] += gradX[dx][1] * wDy;
-                     gradXY[dy][dx][2] += gradX[dx][2] * wy;
-                  }
-               }
-            }
-            for (int dz = 0; dz < D1D; ++dz)
-            {
-               const real_t wz = Bt(dz, qz);
-               const real_t wDz = Gt(dz, qz);
-               for (int dy = 0; dy < D1D; ++dy)
-               {
-                  for (int dx = 0; dx < D1D; ++dx)
-                  {
-                     y(dx, dy, dz, c, e) +=
-                        ((gradXY[dy][dx][0] * wz) + (gradXY[dy][dx][1] * wz) +
-                         (gradXY[dy][dx][2] * wDz));
-                  }
-               }
-            }
-         }
-      }
-   });
-}
-} // namespace internal
-
-template <int DIM, int VDIM, int T_D1D, int T_Q1D>
-VectorDiffusionIntegrator::ApplyKernelType
-VectorDiffusionIntegrator::ApplyPAKernels::Kernel()
-{
-   if constexpr (DIM == 2)
-   {
-      return internal::PAVectorDiffusionApply2D<T_D1D, T_Q1D, VDIM>;
-   }
-   else if constexpr (DIM == 3)
-   {
-      return internal::PAVectorDiffusionApply3D;
-   }
-   MFEM_ABORT("");
-}
-
-} // namespace mfem
-/// \endcond DO_NOT_DOCUMENT
-#endif
@@ -15,58 +15,9 @@
 #include "../qfunction.hpp"
 #include "../ceed/integrators/diffusion/diffusion.hpp"

-#include "bilininteg_vecdiffusion_kernels.hpp"
-
 namespace mfem
 {

-VectorDiffusionIntegrator::VectorDiffusionIntegrator(const IntegrationRule *ir)
-   : BilinearFormIntegrator(ir)
-{
-   static Kernels kernels;
-}
-
-VectorDiffusionIntegrator::VectorDiffusionIntegrator(Coefficient &q)
-   : VectorDiffusionIntegrator()
-{
-   Q = &q;
-}
-
-VectorDiffusionIntegrator::VectorDiffusionIntegrator(int vector_dimension)
-   : VectorDiffusionIntegrator()
-{
-   vdim = vector_dimension;
-}
-
-VectorDiffusionIntegrator::VectorDiffusionIntegrator(Coefficient &q,
-                                                     const IntegrationRule *ir)
-   : VectorDiffusionIntegrator(ir)
-{
-   Q = &q;
-}
-
-VectorDiffusionIntegrator::VectorDiffusionIntegrator(Coefficient &q,
-                                                     int vector_dimension)
-   : VectorDiffusionIntegrator()
-{
-   Q = &q;
-   vdim = vector_dimension;
-}
-
-VectorDiffusionIntegrator::VectorDiffusionIntegrator(VectorCoefficient &vq)
-   : VectorDiffusionIntegrator()
-{
-   VQ = &vq;
-   vdim = vq.GetVDim();
-}
-
-VectorDiffusionIntegrator::VectorDiffusionIntegrator(MatrixCoefficient &mq)
-   : VectorDiffusionIntegrator()
-{
-   MQ = &mq;
-   vdim = mq.GetVDim();
-}
-
 // PA Diffusion Assemble 2D kernel
 static void PAVectorDiffusionSetup2D(const int Q1D,
                                     const int NE,
@@ -474,6 +425,322 @@ void VectorDiffusionIntegrator::AssembleDiagonalPA(Vector &diag)
   }
 }

+// PA Diffusion Apply 2D kernel
+template<int T_D1D = 0, int T_Q1D = 0, int T_VDIM = 0> static
+void PAVectorDiffusionApply2D(const int NE,
+                              const Array<real_t> &b,
+                              const Array<real_t> &g,
+                              const Array<real_t> &bt,
+                              const Array<real_t> &gt,
+                              const Vector &d_,
+                              const Vector &x_,
+                              Vector &y_,
+                              const int d1d = 0,
+                              const int q1d = 0,
+                              const int vdim = 0)
+{
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+   const int VDIM = T_VDIM ? T_VDIM : vdim;
+   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+   auto B = Reshape(b.Read(), Q1D, D1D);
+   auto G = Reshape(g.Read(), Q1D, D1D);
+   auto Bt = Reshape(bt.Read(), D1D, Q1D);
+   auto Gt = Reshape(gt.Read(), D1D, Q1D);
+   auto D = Reshape(d_.Read(), Q1D*Q1D, 3, NE);
+   auto x = Reshape(x_.Read(), D1D, D1D, VDIM, NE);
+   auto y = Reshape(y_.ReadWrite(), D1D, D1D, VDIM, NE);
+   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   {
+      const int D1D = T_D1D ? T_D1D : d1d;
+      const int Q1D = T_Q1D ? T_Q1D : q1d;
+      const int VDIM = T_VDIM ? T_VDIM : vdim;
+      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
+      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+
+      real_t grad[max_Q1D][max_Q1D][2];
+      for (int c = 0; c < VDIM; c++)
+      {
+         for (int qy = 0; qy < Q1D; ++qy)
+         {
+            for (int qx = 0; qx < Q1D; ++qx)
+            {
+               grad[qy][qx][0] = 0.0;
+               grad[qy][qx][1] = 0.0;
+            }
+         }
+         for (int dy = 0; dy < D1D; ++dy)
+         {
+            real_t gradX[max_Q1D][2];
+            for (int qx = 0; qx < Q1D; ++qx)
+            {
+               gradX[qx][0] = 0.0;
+               gradX[qx][1] = 0.0;
+            }
+            for (int dx = 0; dx < D1D; ++dx)
+            {
+               const real_t s = x(dx,dy,c,e);
+               for (int qx = 0; qx < Q1D; ++qx)
+               {
+                  gradX[qx][0] += s * B(qx,dx);
+                  gradX[qx][1] += s * G(qx,dx);
+               }
+            }
+            for (int qy = 0; qy < Q1D; ++qy)
+            {
+               const real_t wy  = B(qy,dy);
+               const real_t wDy = G(qy,dy);
+               for (int qx = 0; qx < Q1D; ++qx)
+               {
+                  grad[qy][qx][0] += gradX[qx][1] * wy;
+                  grad[qy][qx][1] += gradX[qx][0] * wDy;
+               }
+            }
+         }
+         // Calculate Dxy, xDy in plane
+         for (int qy = 0; qy < Q1D; ++qy)
+         {
+            for (int qx = 0; qx < Q1D; ++qx)
+            {
+               const int q = qx + qy * Q1D;
+               const real_t O11 = D(q,0,e);
+               const real_t O12 = D(q,1,e);
+               const real_t O22 = D(q,2,e);
+               const real_t gradX = grad[qy][qx][0];
+               const real_t gradY = grad[qy][qx][1];
+               grad[qy][qx][0] = (O11 * gradX) + (O12 * gradY);
+               grad[qy][qx][1] = (O12 * gradX) + (O22 * gradY);
+            }
+         }
+         for (int qy = 0; qy < Q1D; ++qy)
+         {
+            real_t gradX[max_D1D][2];
+            for (int dx = 0; dx < D1D; ++dx)
+            {
+               gradX[dx][0] = 0.0;
+               gradX[dx][1] = 0.0;
+            }
+            for (int qx = 0; qx < Q1D; ++qx)
+            {
+               const real_t gX = grad[qy][qx][0];
+               const real_t gY = grad[qy][qx][1];
+               for (int dx = 0; dx < D1D; ++dx)
+               {
+                  const real_t wx  = Bt(dx,qx);
+                  const real_t wDx = Gt(dx,qx);
+                  gradX[dx][0] += gX * wDx;
+                  gradX[dx][1] += gY * wx;
+               }
+            }
+            for (int dy = 0; dy < D1D; ++dy)
+            {
+               const real_t wy  = Bt(dy,qy);
+               const real_t wDy = Gt(dy,qy);
+               for (int dx = 0; dx < D1D; ++dx)
+               {
+                  y(dx,dy,c,e) += ((gradX[dx][0] * wy) + (gradX[dx][1] * wDy));
+               }
+            }
+         }
+      }
+   });
+}
+
+// PA Diffusion Apply 3D kernel
+template<const int T_D1D = 0,
+         const int T_Q1D = 0> static
+void PAVectorDiffusionApply3D(const int NE,
+                              const Array<real_t> &b,
+                              const Array<real_t> &g,
+                              const Array<real_t> &bt,
+                              const Array<real_t> &gt,
+                              const Vector &op_,
+                              const Vector &x_,
+                              Vector &y_,
+                              int d1d = 0, int q1d = 0)
+{
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+   constexpr int VDIM = 3;
+   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+   auto B = Reshape(b.Read(), Q1D, D1D);
+   auto G = Reshape(g.Read(), Q1D, D1D);
+   auto Bt = Reshape(bt.Read(), D1D, Q1D);
+   auto Gt = Reshape(gt.Read(), D1D, Q1D);
+   auto op = Reshape(op_.Read(), Q1D*Q1D*Q1D, 6, NE);
+   auto x = Reshape(x_.Read(), D1D, D1D, D1D, VDIM, NE);
+   auto y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, VDIM, NE);
+   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   {
+      const int D1D = T_D1D ? T_D1D : d1d;
+      const int Q1D = T_Q1D ? T_Q1D : q1d;
+      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
+      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+      for (int c = 0; c < VDIM; ++ c)
+      {
+         real_t grad[max_Q1D][max_Q1D][max_Q1D][3];
+         for (int qz = 0; qz < Q1D; ++qz)
+         {
+            for (int qy = 0; qy < Q1D; ++qy)
+            {
+               for (int qx = 0; qx < Q1D; ++qx)
+               {
+                  grad[qz][qy][qx][0] = 0.0;
+                  grad[qz][qy][qx][1] = 0.0;
+                  grad[qz][qy][qx][2] = 0.0;
+               }
+            }
+         }
+         for (int dz = 0; dz < D1D; ++dz)
+         {
+            real_t gradXY[max_Q1D][max_Q1D][3];
+            for (int qy = 0; qy < Q1D; ++qy)
+            {
+               for (int qx = 0; qx < Q1D; ++qx)
+               {
+                  gradXY[qy][qx][0] = 0.0;
+                  gradXY[qy][qx][1] = 0.0;
+                  gradXY[qy][qx][2] = 0.0;
+               }
+            }
+            for (int dy = 0; dy < D1D; ++dy)
+            {
+               real_t gradX[max_Q1D][2];
+               for (int qx = 0; qx < Q1D; ++qx)
+               {
+                  gradX[qx][0] = 0.0;
+                  gradX[qx][1] = 0.0;
+               }
+               for (int dx = 0; dx < D1D; ++dx)
+               {
+                  const real_t s = x(dx,dy,dz,c,e);
+                  for (int qx = 0; qx < Q1D; ++qx)
+                  {
+                     gradX[qx][0] += s * B(qx,dx);
+                     gradX[qx][1] += s * G(qx,dx);
+                  }
+               }
+               for (int qy = 0; qy < Q1D; ++qy)
+               {
+                  const real_t wy  = B(qy,dy);
+                  const real_t wDy = G(qy,dy);
+                  for (int qx = 0; qx < Q1D; ++qx)
+                  {
+                     const real_t wx  = gradX[qx][0];
+                     const real_t wDx = gradX[qx][1];
+                     gradXY[qy][qx][0] += wDx * wy;
+                     gradXY[qy][qx][1] += wx  * wDy;
+                     gradXY[qy][qx][2] += wx  * wy;
+                  }
+               }
+            }
+            for (int qz = 0; qz < Q1D; ++qz)
+            {
+               const real_t wz  = B(qz,dz);
+               const real_t wDz = G(qz,dz);
+               for (int qy = 0; qy < Q1D; ++qy)
+               {
+                  for (int qx = 0; qx < Q1D; ++qx)
+                  {
+                     grad[qz][qy][qx][0] += gradXY[qy][qx][0] * wz;
+                     grad[qz][qy][qx][1] += gradXY[qy][qx][1] * wz;
+                     grad[qz][qy][qx][2] += gradXY[qy][qx][2] * wDz;
+                  }
+               }
+            }
+         }
+         // Calculate Dxyz, xDyz, xyDz in plane
+         for (int qz = 0; qz < Q1D; ++qz)
+         {
+            for (int qy = 0; qy < Q1D; ++qy)
+            {
+               for (int qx = 0; qx < Q1D; ++qx)
+               {
+                  const int q = qx + (qy + qz * Q1D) * Q1D;
+                  const real_t O11 = op(q,0,e);
+                  const real_t O12 = op(q,1,e);
+                  const real_t O13 = op(q,2,e);
+                  const real_t O22 = op(q,3,e);
+                  const real_t O23 = op(q,4,e);
+                  const real_t O33 = op(q,5,e);
+                  const real_t gradX = grad[qz][qy][qx][0];
+                  const real_t gradY = grad[qz][qy][qx][1];
+                  const real_t gradZ = grad[qz][qy][qx][2];
+                  grad[qz][qy][qx][0] = (O11*gradX)+(O12*gradY)+(O13*gradZ);
+                  grad[qz][qy][qx][1] = (O12*gradX)+(O22*gradY)+(O23*gradZ);
+                  grad[qz][qy][qx][2] = (O13*gradX)+(O23*gradY)+(O33*gradZ);
+               }
+            }
+         }
+         for (int qz = 0; qz < Q1D; ++qz)
+         {
+            real_t gradXY[max_D1D][max_D1D][3];
+            for (int dy = 0; dy < D1D; ++dy)
+            {
+               for (int dx = 0; dx < D1D; ++dx)
+               {
+                  gradXY[dy][dx][0] = 0;
+                  gradXY[dy][dx][1] = 0;
+                  gradXY[dy][dx][2] = 0;
+               }
+            }
+            for (int qy = 0; qy < Q1D; ++qy)
+            {
+               real_t gradX[max_D1D][3];
+               for (int dx = 0; dx < D1D; ++dx)
+               {
+                  gradX[dx][0] = 0;
+                  gradX[dx][1] = 0;
+                  gradX[dx][2] = 0;
+               }
+               for (int qx = 0; qx < Q1D; ++qx)
+               {
+                  const real_t gX = grad[qz][qy][qx][0];
+                  const real_t gY = grad[qz][qy][qx][1];
+                  const real_t gZ = grad[qz][qy][qx][2];
+                  for (int dx = 0; dx < D1D; ++dx)
+                  {
+                     const real_t wx  = Bt(dx,qx);
+                     const real_t wDx = Gt(dx,qx);
+                     gradX[dx][0] += gX * wDx;
+                     gradX[dx][1] += gY * wx;
+                     gradX[dx][2] += gZ * wx;
+                  }
+               }
+               for (int dy = 0; dy < D1D; ++dy)
+               {
+                  const real_t wy  = Bt(dy,qy);
+                  const real_t wDy = Gt(dy,qy);
+                  for (int dx = 0; dx < D1D; ++dx)
+                  {
+                     gradXY[dy][dx][0] += gradX[dx][0] * wy;
+                     gradXY[dy][dx][1] += gradX[dx][1] * wDy;
+                     gradXY[dy][dx][2] += gradX[dx][2] * wy;
+                  }
+               }
+            }
+            for (int dz = 0; dz < D1D; ++dz)
+            {
+               const real_t wz  = Bt(dz,qz);
+               const real_t wDz = Gt(dz,qz);
+               for (int dy = 0; dy < D1D; ++dy)
+               {
+                  for (int dx = 0; dx < D1D; ++dx)
+                  {
+                     y(dx,dy,dz,c,e) +=
+                        ((gradXY[dy][dx][0] * wz) +
+                         (gradXY[dy][dx][1] * wz) +
+                         (gradXY[dy][dx][2] * wDz));
+                  }
+               }
+            }
+         }
+      }
+   });
+}
+
 // PA Diffusion Apply kernel
 void VectorDiffusionIntegrator::AddMultPA(const Vector &x, Vector &y) const
 {
@@ -490,29 +757,27 @@ void VectorDiffusionIntegrator::AddMultPA(const Vector &x, Vector &y) const
      const Array<real_t> &Bt = maps->Bt;
      const Array<real_t> &Gt = maps->Gt;
      const Vector &D = pa_data;
-      ApplyPAKernels::Run(dim, sdim, D1D, Q1D, ne, B, G, Bt, Gt, D, x, y, D1D,
-                          Q1D, sdim);
+
+      if (dim == 2 && sdim == 3)
+      {
+         switch ((dofs1D << 4 ) | quad1D)
+         {
+            case 0x22: return PAVectorDiffusionApply2D<2,2,3>(ne,B,G,Bt,Gt,D,x,y);
+            case 0x33: return PAVectorDiffusionApply2D<3,3,3>(ne,B,G,Bt,Gt,D,x,y);
+            case 0x44: return PAVectorDiffusionApply2D<4,4,3>(ne,B,G,Bt,Gt,D,x,y);
+            case 0x55: return PAVectorDiffusionApply2D<5,5,3>(ne,B,G,Bt,Gt,D,x,y);
+            default:
+               return PAVectorDiffusionApply2D(ne,B,G,Bt,Gt,D,x,y,D1D,Q1D,sdim);
+         }
+      }
+      if (dim == 2 && sdim == 2)
+      { return PAVectorDiffusionApply2D(ne,B,G,Bt,Gt,D,x,y,D1D,Q1D,sdim); }
+
+      if (dim == 3 && sdim == 3)
+      { return PAVectorDiffusionApply3D(ne,B,G,Bt,Gt,D,x,y,D1D,Q1D); }
+
+      MFEM_ABORT("Unknown kernel.");
   }
 }

-/// \cond DO_NOT_DOCUMENT
-
-VectorDiffusionIntegrator::ApplyKernelType
-VectorDiffusionIntegrator::ApplyPAKernels::Fallback(int DIM, int, int, int)
-{
-   if (DIM == 2) { return internal::PAVectorDiffusionApply2D; }
-   else if (DIM == 3) { return internal::PAVectorDiffusionApply3D; }
-   else { MFEM_ABORT(""); }
-}
-
-VectorDiffusionIntegrator::Kernels::Kernels()
-{
-   VectorDiffusionIntegrator::AddSpecialization<2, 3, 2, 2>();
-   VectorDiffusionIntegrator::AddSpecialization<2, 3, 3, 3>();
-   VectorDiffusionIntegrator::AddSpecialization<2, 3, 4, 4>();
-   VectorDiffusionIntegrator::AddSpecialization<2, 3, 5, 5>();
-}
-
-/// \endcond DO_NOT_DOCUMENT
-
 } // namespace mfem
@@ -9,19 +9,183 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.

-#include "../../fem/kernels.hpp"
 #include "../../general/forall.hpp"
+#include "../../fem/kernels.hpp"
 #include "../fem.hpp"

-#include "lininteg_domain_kernels.hpp"
-
-/// \cond DO_NOT_DOCUMENT
-
 namespace mfem
 {
+
+template<int T_D1D = 0, int T_Q1D = 0>
+static void DLFEvalAssemble2D(const int vdim, const int ne, const int d,
+                              const int q,
+                              const int map_type, const int *markers, const real_t *b,
+                              const real_t *detj, const real_t *weights,
+                              const Vector &coeff, real_t *y)
+{
+   const auto F = coeff.Read();
+   const auto M = Reshape(markers, ne);
+   const auto B = Reshape(b, q, d);
+   const auto DETJ = Reshape(detj, q, q, ne);
+   const auto W = Reshape(weights, q, q);
+   const bool cst = coeff.Size() == vdim;
+   const auto C = cst ? Reshape(F,vdim,1,1,1) : Reshape(F,vdim,q,q,ne);
+   auto Y = Reshape(y, d,d, vdim, ne);
+
+   mfem::forall_2D(ne, q, q, [=] MFEM_HOST_DEVICE (int e)
+   {
+      if (M(e) == 0) { return; } // ignore
+
+      constexpr int Q = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+      constexpr int D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
+
+      MFEM_SHARED real_t sBt[Q*D];
+      MFEM_SHARED real_t sQQ[Q*Q];
+      MFEM_SHARED real_t sQD[Q*D];
+
+      const DeviceMatrix Bt(sBt, d, q);
+      kernels::internal::LoadB<D,Q>(d, q, B, sBt);
+
+      const DeviceMatrix QQ(sQQ, q, q);
+      const DeviceMatrix QD(sQD, q, d);
+
+      for (int c = 0; c < vdim; ++c)
+      {
+         const real_t cst_val = C(c,0,0,0);
+         MFEM_FOREACH_THREAD(x,x,q)
+         {
+            MFEM_FOREACH_THREAD(y,y,q)
+            {
+               const real_t detJ = (map_type == FiniteElement::VALUE) ? DETJ(x,y,e) : 1.0;
+               const real_t coeff_val = cst ? cst_val : C(c,x,y,e);
+               QQ(y,x) = W(x,y) * coeff_val * detJ;
+            }
+         }
+         MFEM_SYNC_THREAD;
+         MFEM_FOREACH_THREAD(qy,y,q)
+         {
+            MFEM_FOREACH_THREAD(dx,x,d)
+            {
+               real_t u = 0.0;
+               for (int qx = 0; qx < q; ++qx) { u += QQ(qy,qx) * Bt(dx,qx); }
+               QD(qy,dx) = u;
+            }
+         }
+         MFEM_SYNC_THREAD;
+         MFEM_FOREACH_THREAD(dy,y,d)
+         {
+            MFEM_FOREACH_THREAD(dx,x,d)
+            {
+               real_t u = 0.0;
+               for (int qy = 0; qy < q; ++qy) { u += QD(qy,dx) * Bt(dy,qy); }
+               Y(dx,dy,c,e) += u;
+            }
+         }
+         MFEM_SYNC_THREAD;
+      }
+   });
+}
+
+template<int T_D1D = 0, int T_Q1D = 0>
+static void DLFEvalAssemble3D(const int vdim, const int ne, const int d,
+                              const int q,
+                              const int map_type, const int *markers, const real_t *b,
+                              const real_t *detj, const real_t *weights,
+                              const Vector &coeff, real_t *y)
+{
+   const auto F = coeff.Read();
+   const auto M = Reshape(markers, ne);
+   const auto B = Reshape(b, q,d);
+   const auto DETJ = Reshape(detj, q, q, q, ne);
+   const auto W = Reshape(weights, q,q,q);
+   const bool cst_coeff = coeff.Size() == vdim;
+   const auto C = cst_coeff ? Reshape(F,vdim,1,1,1,1):Reshape(F,vdim,q,q,q,ne);
+
+   auto Y = Reshape(y, d,d,d, vdim, ne);
+
+   mfem::forall_2D(ne, q, q, [=] MFEM_HOST_DEVICE (int e)
+   {
+      if (M(e) == 0) { return; } // ignore
+
+      constexpr int Q = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+      constexpr int D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
+      constexpr int MQD = (Q >= D) ? Q : D;
+
+      real_t u[D];
+
+      MFEM_SHARED real_t sBt[Q*D];
+      const DeviceMatrix Bt(sBt, d,q);
+      kernels::internal::LoadB<D,Q>(d,q,B,sBt);
+
+      MFEM_SHARED real_t sQQQ[MQD*MQD*MQD];
+      const DeviceCube QQQ(sQQQ, MQD, MQD, MQD);
+
+      for (int c = 0; c < vdim; ++c)
+      {
+         const real_t cst_val = C(c,0,0,0,0);
+         MFEM_FOREACH_THREAD(x,x,q)
+         {
+            MFEM_FOREACH_THREAD(y,y,q)
+            {
+               for (int z = 0; z < q; ++z)
+               {
+                  const real_t detJ = (map_type == FiniteElement::VALUE) ? DETJ(x,y,z,e) : 1.0;
+                  const real_t coeff_val = cst_coeff ? cst_val : C(c,x,y,z,e);
+                  QQQ(z,y,x) = W(x,y,z) * coeff_val * detJ;
+               }
+            }
+         }
+         MFEM_SYNC_THREAD;
+         MFEM_FOREACH_THREAD(qx,x,q)
+         {
+            MFEM_FOREACH_THREAD(qy,y,q)
+            {
+               for (int dz = 0; dz < d; ++dz) { u[dz] = 0.0; }
+               for (int qz = 0; qz < q; ++qz)
+               {
+                  const real_t ZYX = QQQ(qz,qy,qx);
+                  for (int dz = 0; dz < d; ++dz) { u[dz] += ZYX * Bt(dz,qz); }
+               }
+               for (int dz = 0; dz < d; ++dz) { QQQ(dz,qy,qx) = u[dz]; }
+            }
+         }
+         MFEM_SYNC_THREAD;
+         MFEM_FOREACH_THREAD(dz,y,d)
+         {
+            MFEM_FOREACH_THREAD(qx,x,q)
+            {
+               for (int dy = 0; dy < d; ++dy) { u[dy] = 0.0; }
+               for (int qy = 0; qy < q; ++qy)
+               {
+                  const real_t zYX = QQQ(dz,qy,qx);
+                  for (int dy = 0; dy < d; ++dy) { u[dy] += zYX * Bt(dy,qy); }
+               }
+               for (int dy = 0; dy < d; ++dy) { QQQ(dz,dy,qx) = u[dy]; }
+            }
+         }
+         MFEM_SYNC_THREAD;
+         MFEM_FOREACH_THREAD(dz,y,d)
+         {
+            MFEM_FOREACH_THREAD(dy,x,d)
+            {
+               for (int dx = 0; dx < d; ++dx) { u[dx] = 0.0; }
+               for (int qx = 0; qx < q; ++qx)
+               {
+                  const real_t zyX = QQQ(dz,dy,qx);
+                  for (int dx = 0; dx < d; ++dx) { u[dx] += zyX * Bt(dx,qx); }
+               }
+               for (int dx = 0; dx < d; ++dx) { Y(dx,dy,dz,c,e) += u[dx]; }
+            }
+         }
+         MFEM_SYNC_THREAD;
+      }
+   });
+}
+
 static void DLFEvalAssemble(const FiniteElementSpace &fes,
                            const IntegrationRule *ir,
-                            const Array<int> &markers, const Vector &coeff,
+                            const Array<int> &markers,
+                            const Vector &coeff,
                            Vector &y)
 {
   Mesh *mesh = fes.GetMesh();
@@ -33,20 +197,50 @@ static void DLFEvalAssemble(const FiniteElementSpace &fes,
   constexpr int flags = GeometricFactors::DETERMINANTS;
   const GeometricFactors *geom = mesh->GetGeometricFactors(*ir, flags, mt);
   const int map_type = fes.GetTypicalFE()->GetMapType();
+   decltype(&DLFEvalAssemble2D<>) ker =
+      dim == 2 ? DLFEvalAssemble2D<> : DLFEvalAssemble3D<>;
+
+   if (dim==2)
+   {
+      if (d==1 && q==1) { ker=DLFEvalAssemble2D<1,1>; }
+      if (d==2 && q==2) { ker=DLFEvalAssemble2D<2,2>; }
+      if (d==3 && q==3) { ker=DLFEvalAssemble2D<3,3>; }
+      if (d==4 && q==4) { ker=DLFEvalAssemble2D<4,4>; }
+      if (d==5 && q==5) { ker=DLFEvalAssemble2D<5,5>; }
+      if (d==2 && q==3) { ker=DLFEvalAssemble2D<2,3>; }
+      if (d==3 && q==4) { ker=DLFEvalAssemble2D<3,4>; }
+      if (d==4 && q==5) { ker=DLFEvalAssemble2D<4,5>; }
+      if (d==5 && q==6) { ker=DLFEvalAssemble2D<5,6>; }
+   }
+
+   if (dim==3)
+   {
+      if (d==1 && q==1) { ker=DLFEvalAssemble3D<1,1>; }
+      if (d==2 && q==2) { ker=DLFEvalAssemble3D<2,2>; }
+      if (d==3 && q==3) { ker=DLFEvalAssemble3D<3,3>; }
+      if (d==4 && q==4) { ker=DLFEvalAssemble3D<4,4>; }
+      if (d==5 && q==5) { ker=DLFEvalAssemble3D<5,5>; }
+      if (d==2 && q==3) { ker=DLFEvalAssemble3D<2,3>; }
+      if (d==3 && q==4) { ker=DLFEvalAssemble3D<3,4>; }
+      if (d==4 && q==5) { ker=DLFEvalAssemble3D<4,5>; }
+      if (d==5 && q==6) { ker=DLFEvalAssemble3D<5,6>; }
+   }
+
+   MFEM_VERIFY(ker, "No kernel ndof " << d << " nqpt " << q);

   const int vdim = fes.GetVDim();
   const int ne = fes.GetMesh()->GetNE();
-   const real_t *B = maps.B.Read();
   const int *M = markers.Read();
+   const real_t *B = maps.B.Read();
   const real_t *detJ = geom->detJ.Read();
   const real_t *W = ir->GetWeights().Read();
   real_t *Y = y.ReadWrite();
-   DomainLFIntegrator::AssembleKernels::Run(dim, d, q, vdim, ne, d, q, map_type,
-                                            M, B, detJ, W, coeff, Y);
+   ker(vdim, ne, d, q, map_type, M, B, detJ, W, coeff, Y);
 }

 void DomainLFIntegrator::AssembleDevice(const FiniteElementSpace &fes,
-                                        const Array<int> &markers, Vector &b)
+                                        const Array<int> &markers,
+                                        Vector &b)
 {
   const FiniteElement &fe = *fes.GetTypicalFE();
   const int qorder = oa * fe.GetOrder() + ob;
@@ -72,49 +266,4 @@ void VectorDomainLFIntegrator::AssembleDevice(const FiniteElementSpace &fes,
   DLFEvalAssemble(fes, ir, markers, coeff, b);
 }

-DomainLFIntegrator::AssembleKernelType
-DomainLFIntegrator::AssembleKernels::Fallback(int DIM, int, int)
-{
-   switch (DIM)
-   {
-      case 1:
-         return DLFEvalAssemble1D<0, 0>;
-      case 2:
-         return DLFEvalAssemble2D<0, 0>;
-      case 3:
-         return DLFEvalAssemble3D<0, 0>;
-   }
-   MFEM_ABORT("");
-}
-
-DomainLFIntegrator::Kernels::Kernels()
-{
-   // 2D
-   // Q = P+1
-   DomainLFIntegrator::AddSpecialization<2, 1, 1>();
-   DomainLFIntegrator::AddSpecialization<2, 2, 2>();
-   DomainLFIntegrator::AddSpecialization<2, 3, 3>();
-   DomainLFIntegrator::AddSpecialization<2, 4, 4>();
-   DomainLFIntegrator::AddSpecialization<2, 5, 5>();
-   // Q = P+2
-   DomainLFIntegrator::AddSpecialization<2, 2, 3>();
-   DomainLFIntegrator::AddSpecialization<2, 3, 4>();
-   DomainLFIntegrator::AddSpecialization<2, 4, 5>();
-   DomainLFIntegrator::AddSpecialization<2, 5, 6>();
-   // 3D
-   // Q = P+1
-   DomainLFIntegrator::AddSpecialization<3, 1, 1>();
-   DomainLFIntegrator::AddSpecialization<3, 2, 2>();
-   DomainLFIntegrator::AddSpecialization<3, 3, 3>();
-   DomainLFIntegrator::AddSpecialization<3, 4, 4>();
-   DomainLFIntegrator::AddSpecialization<3, 5, 5>();
-   // Q = P+2
-   DomainLFIntegrator::AddSpecialization<3, 2, 3>();
-   DomainLFIntegrator::AddSpecialization<3, 3, 4>();
-   DomainLFIntegrator::AddSpecialization<3, 4, 5>();
-   DomainLFIntegrator::AddSpecialization<3, 5, 6>();
-}
-
-/// \endcond DO_NOT_DOCUMENT
-
 } // namespace mfem
@@ -1,318 +0,0 @@
-// 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_LININTEG_DOMAIN_KERNELS_HPP
-#define MFEM_LININTEG_DOMAIN_KERNELS_HPP
-
-#include "../../fem/kernels.hpp"
-#include "../../general/forall.hpp"
-#include "../fem.hpp"
-
-/// \cond DO_NOT_DOCUMENT
-
-namespace mfem
-{
-
-template <int T_D1D = 0, int T_Q1D = 0>
-static void DLFEvalAssemble1D(const int vdim, const int ne, const int d,
-                              const int q, const int map_type,
-                              const int *markers, const real_t *b,
-                              const real_t *detj, const real_t *weights,
-                              const Vector &coeff, real_t *y)
-{
-   {
-      constexpr int Q = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-      constexpr int D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      MFEM_VERIFY(q <= Q, "");
-      MFEM_VERIFY(d <= D, "");
-   }
-
-   const auto F = coeff.Read();
-   const auto B = Reshape(b, q, d);
-   const auto DETJ = Reshape(detj, q, ne);
-   const bool cst = coeff.Size() == vdim;
-   const auto C = cst ? Reshape(F, vdim, 1, 1) : Reshape(F, vdim, q, ne);
-   auto Y = Reshape(y, d, vdim, ne);
-
-   mfem::forall_2D(ne, d, 1, [=] MFEM_HOST_DEVICE(int e)
-   {
-      if (markers[e] == 0)
-      {
-         return;
-      } // ignore
-
-      constexpr int Q = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-      constexpr int D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-
-      MFEM_SHARED real_t sBt[Q * D];
-      const DeviceMatrix Bt(sBt, d, q);
-      kernels::internal::LoadB<D, Q>(d, q, B, sBt);
-
-      for (int c = 0; c < vdim; ++c)
-      {
-         const real_t cst_val = C(c, 0, 0);
-         MFEM_FOREACH_THREAD(dx, x, d)
-         {
-            real_t u = 0;
-            for (int qx = 0; qx < q; ++qx)
-            {
-               const real_t detJ =
-                  (map_type == FiniteElement::VALUE) ? DETJ(qx, e) : 1.0;
-               const real_t coeff_val = cst ? cst_val : C(c, qx, e);
-               u += weights[qx] * coeff_val * detJ * Bt(dx, qx);
-            }
-            Y(dx, c, e) += u;
-         }
-      }
-   });
-}
-
-template <int T_D1D = 0, int T_Q1D = 0>
-static void DLFEvalAssemble2D(const int vdim, const int ne, const int d,
-                              const int q, const int map_type,
-                              const int *markers, const real_t *b,
-                              const real_t *detj, const real_t *weights,
-                              const Vector &coeff, real_t *y)
-{
-   {
-      constexpr int Q = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-      constexpr int D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      MFEM_VERIFY(q <= Q, "");
-      MFEM_VERIFY(d <= D, "");
-   }
-
-   const auto F = coeff.Read();
-   const auto B = Reshape(b, q, d);
-   const auto DETJ = Reshape(detj, q, q, ne);
-   const auto W = Reshape(weights, q, q);
-   const bool cst = coeff.Size() == vdim;
-   const auto C = cst ? Reshape(F, vdim, 1, 1, 1) : Reshape(F, vdim, q, q, ne);
-   auto Y = Reshape(y, d, d, vdim, ne);
-
-   mfem::forall_2D(ne, q, q, [=] MFEM_HOST_DEVICE(int e)
-   {
-      if (markers[e] == 0)
-      {
-         return;
-      } // ignore
-
-      constexpr int Q = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-      constexpr int D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-
-      MFEM_SHARED real_t sBt[Q * D];
-      MFEM_SHARED real_t sQQ[Q * Q];
-      MFEM_SHARED real_t sQD[Q * D];
-
-      const DeviceMatrix Bt(sBt, d, q);
-      kernels::internal::LoadB<D, Q>(d, q, B, sBt);
-
-      const DeviceMatrix QQ(sQQ, q, q);
-      const DeviceMatrix QD(sQD, q, d);
-
-      for (int c = 0; c < vdim; ++c)
-      {
-         const real_t cst_val = C(c, 0, 0, 0);
-         MFEM_FOREACH_THREAD(x, x, q)
-         {
-            MFEM_FOREACH_THREAD(y, y, q)
-            {
-               const real_t detJ =
-                  (map_type == FiniteElement::VALUE) ? DETJ(x, y, e) : 1.0;
-               const real_t coeff_val = cst ? cst_val : C(c, x, y, e);
-               QQ(y, x) = W(x, y) * coeff_val * detJ;
-            }
-         }
-         MFEM_SYNC_THREAD;
-         MFEM_FOREACH_THREAD(qy, y, q)
-         {
-            MFEM_FOREACH_THREAD(dx, x, d)
-            {
-               real_t u = 0.0;
-               for (int qx = 0; qx < q; ++qx)
-               {
-                  u += QQ(qy, qx) * Bt(dx, qx);
-               }
-               QD(qy, dx) = u;
-            }
-         }
-         MFEM_SYNC_THREAD;
-         MFEM_FOREACH_THREAD(dy, y, d)
-         {
-            MFEM_FOREACH_THREAD(dx, x, d)
-            {
-               real_t u = 0.0;
-               for (int qy = 0; qy < q; ++qy)
-               {
-                  u += QD(qy, dx) * Bt(dy, qy);
-               }
-               Y(dx, dy, c, e) += u;
-            }
-         }
-         MFEM_SYNC_THREAD;
-      }
-   });
-}
-
-template <int T_D1D = 0, int T_Q1D = 0>
-static void DLFEvalAssemble3D(const int vdim, const int ne, const int d,
-                              const int q, const int map_type,
-                              const int* markers, const real_t *b,
-                              const real_t *detj, const real_t *weights,
-                              const Vector &coeff, real_t *y)
-{
-   {
-      constexpr int Q = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-      constexpr int D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      MFEM_VERIFY(q <= Q, "");
-      MFEM_VERIFY(d <= D, "");
-   }
-
-   const auto F = coeff.Read();
-   const auto B = Reshape(b, q, d);
-   const auto DETJ = Reshape(detj, q, q, q, ne);
-   const auto W = Reshape(weights, q, q, q);
-   const bool cst_coeff = coeff.Size() == vdim;
-   const auto C =
-      cst_coeff ? Reshape(F, vdim, 1, 1, 1, 1) : Reshape(F, vdim, q, q, q, ne);
-
-   auto Y = Reshape(y, d, d, d, vdim, ne);
-
-   mfem::forall_2D(ne, q, q, [=] MFEM_HOST_DEVICE(int e)
-   {
-      if (markers[e] == 0)
-      {
-         return;
-      } // ignore
-
-      constexpr int Q = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-      constexpr int D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int MQD = (Q >= D) ? Q : D;
-
-      real_t u[D];
-
-      MFEM_SHARED real_t sBt[Q * D];
-      const DeviceMatrix Bt(sBt, d, q);
-      kernels::internal::LoadB<D, Q>(d, q, B, sBt);
-
-      MFEM_SHARED real_t sQQQ[MQD * MQD * MQD];
-      const DeviceCube QQQ(sQQQ, MQD, MQD, MQD);
-
-      for (int c = 0; c < vdim; ++c)
-      {
-         const real_t cst_val = C(c, 0, 0, 0, 0);
-         MFEM_FOREACH_THREAD(x, x, q)
-         {
-            MFEM_FOREACH_THREAD(y, y, q)
-            {
-               for (int z = 0; z < q; ++z)
-               {
-                  const real_t detJ = (map_type == FiniteElement::VALUE)
-                                      ? DETJ(x, y, z, e)
-                                      : 1.0;
-                  const real_t coeff_val =
-                     cst_coeff ? cst_val : C(c, x, y, z, e);
-                  QQQ(z, y, x) = W(x, y, z) * coeff_val * detJ;
-               }
-            }
-         }
-         MFEM_SYNC_THREAD;
-         MFEM_FOREACH_THREAD(qx, x, q)
-         {
-            MFEM_FOREACH_THREAD(qy, y, q)
-            {
-               for (int dz = 0; dz < d; ++dz)
-               {
-                  u[dz] = 0.0;
-               }
-               for (int qz = 0; qz < q; ++qz)
-               {
-                  const real_t ZYX = QQQ(qz, qy, qx);
-                  for (int dz = 0; dz < d; ++dz)
-                  {
-                     u[dz] += ZYX * Bt(dz, qz);
-                  }
-               }
-               for (int dz = 0; dz < d; ++dz)
-               {
-                  QQQ(dz, qy, qx) = u[dz];
-               }
-            }
-         }
-         MFEM_SYNC_THREAD;
-         MFEM_FOREACH_THREAD(dz, y, d)
-         {
-            MFEM_FOREACH_THREAD(qx, x, q)
-            {
-               for (int dy = 0; dy < d; ++dy)
-               {
-                  u[dy] = 0.0;
-               }
-               for (int qy = 0; qy < q; ++qy)
-               {
-                  const real_t zYX = QQQ(dz, qy, qx);
-                  for (int dy = 0; dy < d; ++dy)
-                  {
-                     u[dy] += zYX * Bt(dy, qy);
-                  }
-               }
-               for (int dy = 0; dy < d; ++dy)
-               {
-                  QQQ(dz, dy, qx) = u[dy];
-               }
-            }
-         }
-         MFEM_SYNC_THREAD;
-         MFEM_FOREACH_THREAD(dz, y, d)
-         {
-            MFEM_FOREACH_THREAD(dy, x, d)
-            {
-               for (int dx = 0; dx < d; ++dx)
-               {
-                  u[dx] = 0.0;
-               }
-               for (int qx = 0; qx < q; ++qx)
-               {
-                  const real_t zyX = QQQ(dz, dy, qx);
-                  for (int dx = 0; dx < d; ++dx)
-                  {
-                     u[dx] += zyX * Bt(dx, qx);
-                  }
-               }
-               for (int dx = 0; dx < d; ++dx)
-               {
-                  Y(dx, dy, dz, c, e) += u[dx];
-               }
-            }
-         }
-         MFEM_SYNC_THREAD;
-      }
-   });
-}
-
-template <int DIM, int T_D1D, int T_Q1D>
-DomainLFIntegrator::AssembleKernelType
-DomainLFIntegrator::AssembleKernels::Kernel()
-{
-   switch (DIM)
-   {
-      case 1:
-         return DLFEvalAssemble1D<T_D1D, T_Q1D>;
-      case 2:
-         return DLFEvalAssemble2D<T_D1D, T_Q1D>;
-      case 3:
-         return DLFEvalAssemble3D<T_D1D, T_Q1D>;
-   }
-   MFEM_ABORT("");
-}
-/// \endcond DO_NOT_DOCUMENT
-
-} // namespace mfem
-#endif
@@ -947,7 +947,6 @@ int Quadrature1D::CheckOpen(int type)
      case OpenUniform:
      case ClosedUniform:
      case OpenHalfUniform:
-      case ClosedGL:
         return type; // all types can work as open
      default:
         return Invalid;
@@ -346,13 +346,13 @@ private:
      template<typename T>
      T operator() (const blitz::TinyVector<T,3>& x) const
      {
-         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());
+         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 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
      {
-         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);
+         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];

-         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());
+         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);

         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
      {
-         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());
+         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 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
      {
-         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);
+         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];

-         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[0], u1, d1);
+         TmplPoly_1D::CalcBernstein(el_order,x[1], u2, d2);

         blitz::TinyVector<T,2> res(T(0.0),T(0.0));

@@ -78,7 +78,7 @@ namespace mfem
  public:                                                                      \
    const char *kernel_name = MFEM_KERNEL_NAME(KernelName);                    \
    using KernelSignature = KernelType;                                        \
-    template <MFEM_PARAM_LIST P3> static KernelSignature Kernel();             \
+    template <MFEM_PARAM_LIST P3> static MFEM_EXPORT KernelSignature Kernel(); \
    static MFEM_EXPORT KernelSignature Fallback(MFEM_PARAM_LIST P1);           \
    static MFEM_EXPORT KernelName &Get() {                                     \
      static KernelName table;                                                 \
@@ -51,7 +51,7 @@ void LinearFormExtension::Assemble()
      {
         // scan the attributes to set the markers to 0 or 1
         const int NE = fes.GetNE();
-         const auto attr = attributes->Read();
+         const auto attr = attributes.Read();
         const auto dimk = domain_integs_marker_k->Read();
         auto markers_w = markers.Write();
         mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
@@ -93,14 +93,13 @@ void LinearFormExtension::Assemble()
      else
      {
         // scan the attributes to set the markers to 0 or 1
-         const int NBE = bdr_face_attributes->Size();
-         const auto attr = bdr_face_attributes->Read();
+         const int NBE = bdr_attributes.Size();
+         const auto attr = bdr_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[e] > 0 ? (attr_markers[attr[e] - 1] == 1) : false;
+            markers_w[e] = attr_markers[attr[e]-1] == 1;
         });
      }

@@ -126,7 +125,8 @@ void LinearFormExtension::Update()
      //markers.UseDevice(true);

      // Gather the attributes on the host from all the elements
-      attributes = &mesh.GetElementAttributes();
+      attributes.SetSize(NE);
+      for (int i = 0; i < NE; ++i) { attributes[i] = mesh.GetAttribute(i); }

      elem_restrict_lex = fes.GetElementRestriction(ordering);
      MFEM_VERIFY(elem_restrict_lex, "Element restriction not available");
@@ -136,12 +136,34 @@ void LinearFormExtension::Update()

   if (lf->boundary_integs.Size() > 0)
   {
-      bdr_face_attributes = &mesh.GetBdrFaceAttributes();
-
-      const int nf_bdr = bdr_face_attributes->Size();
+      const int nf_bdr = fes.GetNFbyType(FaceType::Boundary);
      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,
@@ -25,8 +25,7 @@ class LinearForm;
 class LinearFormExtension
 {
   /// Attributes of all mesh elements.
-   const Array<int> *attributes; // Not owned
-   const Array<int> *bdr_face_attributes; // Not owned
+   Array<int> attributes, bdr_attributes;

   /// Temporary markers for device kernels.
   Array<int> markers, bdr_markers;
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Stowell, Mark L.	c968516f36	Merge branch 'complex-coef-dev' of github.com:mfem/mfem into complex-coef-dev	2025-07-22 10:46:29 -07:00
Stowell, Mark L.	9a9d1ea967	Switching to complex_t	2025-07-22 10:45:56 -07:00
Stowell, Mark L.	46ee2ab5dc	Adding complex_t type using code by @camierjs	2025-07-22 10:43:59 -07:00
adam-sim-dev	fb672667cc	Merge branch 'master' into complex-coef-dev	2025-07-15 09:22:34 +08:00
Stowell, Mark L.	1d35fafd21	Removing complex<int> unit test due to a "Build Analysis" error	2025-07-14 13:44:17 -07:00
Stowell, Mark L.	a0ed1bfbca	Adding ComplexVector unit test	2025-07-14 13:18:58 -07:00
Stowell, Mark L.	0c08279225	Adding coefficient as template argument in SesquilinearForm	2025-07-14 10:40:53 -07:00
Stowell, Mark L.	bc84ce3b47	Adding ComplexMatrixConstantCoefficient	2025-07-14 10:40:02 -07:00
Stowell, Mark L.	36abe386e0	Adding DenseMatrix constructor to StdComplexDenseMatrix	2025-07-14 10:39:31 -07:00
Stowell, Mark L.	a4868f2a98	Adding ComplexMatrixCoefficient	2025-07-12 16:05:19 -07:00
Stowell, Mark L.	21321b3abc	Adding new type of complex dense matrix	2025-07-12 16:05:03 -07:00
Stowell, Mark L.	bb4f39c3d7	Avoiding circular dependency	2025-07-12 11:49:55 -07:00
Stowell, Mark L.	b605a29988	Making Real/Imag part coefficients publicly available	2025-07-12 11:49:35 -07:00
Stowell, Mark L.	7967e13f1d	Moving complex coefficient code into separate files	2025-07-12 10:30:00 -07:00
Stowell, Mark L.	b615f22b66	Adding complex coefficient support to ComplexLinearForm classes	2025-07-12 10:20:34 -07:00
Stowell, Mark L.	e0fe515f21	Adding ComplexCoefficient methods to ComplexLinearForm classes	2025-07-12 07:42:11 -07:00
Stowell, Mark L.	269ee766db	Updating ex22p	2025-07-11 17:14:17 -07:00
Stowell, Mark L.	da8a221097	Adding complex coefficients to parallel classes	2025-07-11 17:14:04 -07:00
Stowell, Mark L.	20d6e63df0	Ubuntu portability	2025-07-11 14:48:12 -07:00
Stowell, Mark L.	2288cdcb7f	Ubuntu portability	2025-07-11 14:32:25 -07:00
Stowell, Mark L.	233337c9d1	Adding UseDevice to extraction of real and imaginary parts of ComplexVector	2025-07-11 14:26:26 -07:00
Stowell, Mark L.	4d9cd853b7	Fix for Ubuntu portability	2025-07-11 14:18:25 -07:00
Stowell, Mark L.	27352658c3	Using the new complex coefficient classes in ex22	2025-07-11 14:00:22 -07:00
Stowell, Mark L.	b8a303a07a	Adding Complex coefficient classes	2025-07-11 14:00:02 -07:00
Stowell, Mark L.	cee9bf3bb2	Adding a ComplexVector class	2025-07-11 13:59:23 -07:00