Add GPU scan support for Array<T>::PartialSum

Fix comments in scan.hpp
Add more device support to Array<T>
2025-10-21 09:57:37 -07:00 · 2025-10-21 09:57:37 -07:00 · 2025-10-21 09:57:37 -07:00 · 2025-10-21 09:57:37 -07:00 · 2025-10-20 15:38:54 -07:00 · 2025-10-18 10:48:44 -07:00
167 changed files with 17685 additions and 5253 deletions
@@ -132,12 +132,14 @@ jobs:
            hypre-target: int32
            precision: fp64
            enzyme: true
-            config-opts: MFEM_USE_ENZYME=YES ENZYME_DIR=$(brew --prefix enzyme)
+            config-opts: MFEM_USE_ENZYME=YES ENZYME_DIR=$(brew --prefix enzyme) LDFLAGS=-L$LLVM_PREFIX/lib/c++

    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: ${{ matrix.enzyme && true || false }}
+
    steps:
      # Fix 'No space left on device' errors for Ubuntu builds.
      - name: Run Actions Cleaner
@@ -292,10 +294,12 @@ jobs:
        run: |
          export HOMEBREW_NO_INSTALL_CLEANUP=1
          brew update
-          brew install llvm@20 enzyme
-          echo "LLVM_PREFIX=$(brew --prefix llvm@20)" >> $GITHUB_ENV
-          echo "OMPI_CC=$(brew --prefix llvm@20)/bin/clang" >> $GITHUB_ENV
-          echo "OMPI_CXX=$(brew --prefix llvm@20)/bin/clang++" >> $GITHUB_ENV
+          brew install enzyme
+          ENZYME_LLVM=$(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

      # MFEM build and test
      - name: build
@@ -29,3 +29,47 @@ 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
@@ -208,10 +208,13 @@ 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/gslib/field-diff
 miniapps/gslib/field-interp
@@ -27,9 +27,27 @@ 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.
+
 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.

@@ -44,13 +62,19 @@ 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
 -------------------------------------
@@ -60,10 +84,22 @@ New and updated examples and miniapps
  operators as smoothers.
  These miniapps can be found in `miniapps/diag-smoothers`.

-API changes
+- 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:
 -----------
 - 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`.

@@ -71,16 +107,25 @@ 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
 ====================================

@@ -278,6 +278,11 @@ 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)
@@ -495,14 +500,13 @@ 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)
@@ -596,7 +600,7 @@ 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)
+    ALGOIM ENZYME CUDA::cudart)

 # Add all created targets and *_FOUND libraries in the variables TPL_TARGETS and
 # TPL_LIBRARIES, respectively.
@@ -614,6 +618,7 @@ foreach(TPL IN LISTS MFEM_TPLS)
    endif()
  endif()
 endforeach(TPL)
+
 list(REVERSE TPL_LIBRARIES)
 list(REMOVE_DUPLICATES TPL_LIBRARIES)
 list(REVERSE TPL_LIBRARIES)
@@ -859,7 +859,7 @@ 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.
+  Versions: libCEED >= 0.12.0.

 - RAJA (optional), used when MFEM_USE_RAJA = YES.
  Beginning with MFEM v4.5.1, only RAJA v2022.10.3+ is supported.
@@ -84,6 +84,31 @@ 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)
@@ -27,6 +27,7 @@ if (HYPRE_FOUND OR TARGET HYPRE)
  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
@@ -37,51 +38,91 @@ if (HYPRE_FOUND OR TARGET HYPRE)
 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${HYPRE_FETCH_VERSION}" CACHE STRING "Tag, branch, or commit for HYPRE")
  add_library(HYPRE STATIC IMPORTED)
  # set options and associated dependencies
-  set(CMAKE_OPTIONS)
-  list(APPEND CMAKE_OPTIONS -DCMAKE_BUILD_TYPE:STRING=${CMAKE_BUILD_TYPE})
+  list(APPEND HYPRE_CMAKE_OPTIONS -DCMAKE_BUILD_TYPE:STRING=${CMAKE_BUILD_TYPE})
  if (MFEM_USE_CUDA)
-    list(APPEND CMAKE_OPTIONS -DHYPRE_WITH_CUDA:BOOL=ON)
+    list(APPEND HYPRE_CMAKE_OPTIONS -DHYPRE_ENABLE_CUDA:BOOL=ON -DCMAKE_CUDA_ARCHITECTURES:STRING=${CMAKE_CUDA_ARCHITECTURES})
    find_package(CUDAToolkit REQUIRED)
    target_link_libraries(HYPRE INTERFACE CUDA::cusparse CUDA::curand CUDA::cublas)
  elseif (MFEM_USE_HIP)
-    list(APPEND CMAKE_OPTIONS -DHYPRE_WITH_HIP:BOOL=ON)
+    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 CMAKE_OPTIONS -DHYPRE_ENABLE_SINGLE:BOOL=ON)
+    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_VERSION} to be built with ${CMAKE_OPTIONS}")
-  set(PREFIX ${CMAKE_BINARY_DIR}/fetch/hypre)
+  message(STATUS "Will fetch HYPRE ${HYPRE_FETCH_TAG} to be built with ${HYPRE_CMAKE_OPTIONS}")
+  set(HYPRE_INSTALL ${CMAKE_BINARY_DIR}/fetch/hypre)
  include(ExternalProject)
  ExternalProject_Add(hypre
    GIT_REPOSITORY https://github.com/hypre-space/hypre.git
-    GIT_TAG v${HYPRE_FETCH_VERSION}
+    GIT_TAG ${HYPRE_FETCH_TAG}
    GIT_SHALLOW TRUE
+    GIT_PROGRESS TRUE
    UPDATE_DISCONNECTED TRUE
    SOURCE_SUBDIR src
-    PREFIX ${PREFIX}
-    CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${PREFIX} -DCMAKE_INSTALL_LIBDIR:PATH=lib ${CMAKE_OPTIONS})
-  file(MAKE_DIRECTORY ${PREFIX}/include)
+    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 ${PREFIX}/lib/libHYPRE.a
-    INTERFACE_INCLUDE_DIRECTORIES ${PREFIX}/include)
+    IMPORTED_LOCATION ${HYPRE_INSTALL}/lib/libHYPRE.a
+    INTERFACE_INCLUDE_DIRECTORIES ${HYPRE_INSTALL}/include)
  # convert HYPRE version to integer
-  string(REGEX MATCHALL "[0-9]+" HYPRE_SPLIT_VERSION ${HYPRE_FETCH_VERSION})
-  list(GET HYPRE_SPLIT_VERSION 0 HYPRE_MAJOR_VERSION)
-  list(GET HYPRE_SPLIT_VERSION 1 HYPRE_MINOR_VERSION)
-  list(GET HYPRE_SPLIT_VERSION 2 HYPRE_PATCH_VERSION)
-  math(EXPR HYPRE_VERSION "10000*${HYPRE_MAJOR_VERSION} + 100*${HYPRE_MINOR_VERSION} + ${HYPRE_PATCH_VERSION}")
-  # set cache variables that would otherwise be set after mfem_find_package call
-  set(HYPRE_VERSION ${HYPRE_VERSION} CACHE STRING "HYPRE version." FORCE)
+  if (HYPRE_FETCH_TAG MATCHES "^v?([0-9]+)\\.([0-9]+)\\.([0-9]+)$")
+    # 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()

@@ -149,7 +190,8 @@ endif()
 if (HYPRE_FOUND AND HYPRE_USING_HIP)
  find_package(rocsparse REQUIRED)
  find_package(rocrand REQUIRED)
-  list(APPEND HYPRE_LIBRARIES ${rocsparse_LIBRARIES} ${rocrand_LIBRARIES})
+  find_package(rocsolver REQUIRED)
+  list(APPEND HYPRE_LIBRARIES ${rocsparse_LIBRARIES} ${rocrand_LIBRARIES} roc::rocsolver roc::rocblas)
  set(HYPRE_LIBRARIES ${HYPRE_LIBRARIES} CACHE STRING
      "HYPRE libraries + dependencies." FORCE)
  message(STATUS "Updated HYPRE_LIBRARIES: ${HYPRE_LIBRARIES}")
@@ -32,6 +32,7 @@ if (METIS_FETCH OR FETCH_TPLS)
    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)
@@ -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, EXECUTABLE.
+  # SHARED_LIBRARY, INTERFACE_LIBRARY, UNKNOWN_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 ("${type}" STREQUAL "SHARED_LIBRARY")
+  elseif ("${type}" STREQUAL "SHARED_LIBRARY" OR "${type}" STREQUAL "UNKNOWN_LIBRARY")
    get_target_property(Location ${tgt} LOCATION)
    if (Location)
      get_filename_component(Dir ${Location} DIRECTORY)
@@ -932,12 +932,14 @@ function(mfem_export_mk_files)
  endif()
  set(MFEM_BUILD_TAG "${CMAKE_SYSTEM}")
  set(MFEM_PREFIX "${CMAKE_INSTALL_PREFIX}")
-  # For the next 4 variable, these are the values for the build-tree version of
+  # For the next 4 variables, these are the values for the build-tree version of
  # 'config.mk'
  set(MFEM_INC_DIR "${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")
@@ -88,6 +88,7 @@ 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@
@@ -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)
+ -arch=$(CUDA_ARCH) -isystem "$(CUDA_DIR)/include"
 # Prefixes for passing flags to the host compiler and linker when using
 # CUDA_CXX=nvcc
 CUDA_XCOMPILER = -Xcompiler=
@@ -0,0 +1,593 @@
+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
@@ -0,0 +1,342 @@
+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
@@ -0,0 +1,96 @@
+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
@@ -202,6 +202,7 @@ 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
@@ -205,6 +205,15 @@ 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
@@ -27,6 +27,7 @@
 //               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
@@ -173,6 +173,12 @@ 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

@@ -171,6 +171,11 @@ 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
@@ -241,8 +246,13 @@ 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
@@ -23,6 +23,8 @@

 namespace mfem
 {
+class QuadratureSpace;
+class FaceQuadratureSpace;

 /// Abstract base class BilinearFormIntegrator
 class BilinearFormIntegrator : public NonlinearFormIntegrator
@@ -2494,8 +2496,7 @@ private:
 #endif

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

   void AssembleElementMatrix(const FiniteElement &,
                              ElementTransformation &,
@@ -2528,6 +2529,28 @@ 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,
@@ -2798,15 +2821,13 @@ protected:
   bool symmetric = true; ///< False if using a nonsymmetric matrix coefficient

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

   /* Given a particular Finite Element, compute the
      element curl-curl matrix elmat */
@@ -2836,6 +2857,34 @@ 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
@@ -3089,21 +3138,18 @@ private:
   Vector vcoeff;

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

   /** \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)
-      : vdim(vector_dimension) { }
+   VectorDiffusionIntegrator(int vector_dimension);

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

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

   /** \brief Integrator with scalar coefficient for caller-specified vector
       dimension.
@@ -3113,8 +3159,7 @@ 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)
-      : Q(&q), vdim(vector_dimension) { }
+   VectorDiffusionIntegrator(Coefficient &q, int 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
@@ -3125,8 +3170,7 @@ 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)
-      : VQ(&vq), vdim(vq.GetVDim()) { }
+   VectorDiffusionIntegrator(VectorCoefficient &vq);

   /** \brief Integrator with \c MatrixCoefficient. The vector dimension of the
       \c FiniteElementSpace is assumed to be the same as the dimension of the
@@ -3137,8 +3181,7 @@ 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)
-      : MQ(&mq), vdim(mq.GetVDim()) { }
+   VectorDiffusionIntegrator(MatrixCoefficient& mq);

   void AssembleElementMatrix(const FiniteElement &el,
                              ElementTransformation &Trans,
@@ -3154,6 +3197,28 @@ 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:
@@ -3307,8 +3372,8 @@ public:
 class DGTraceIntegrator : public BilinearFormIntegrator
 {
 protected:
-   Coefficient *rho;
-   VectorCoefficient *u;
+   Coefficient *rho = nullptr;
+   VectorCoefficient *u = nullptr;
   real_t alpha, beta;
   // PA extension
   Vector pa_data;
@@ -3321,17 +3386,16 @@ 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)
-   { rho = NULL; u = &u_; alpha = a; beta = 0.5*a; }
+   DGTraceIntegrator(VectorCoefficient &u_, real_t a);

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

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

   using BilinearFormIntegrator::AssembleFaceMatrix;
   void AssembleFaceMatrix(const FiniteElement &el1,
@@ -3370,6 +3434,26 @@ 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);
 };
@@ -3416,8 +3500,8 @@ public:
 class DGDiffusionIntegrator : public BilinearFormIntegrator
 {
 protected:
-   Coefficient *Q;
-   MatrixCoefficient *MQ;
+   Coefficient *Q = nullptr;
+   MatrixCoefficient *MQ = nullptr;
   real_t sigma, kappa;

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

 public:
-   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) { }
+   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);
   using BilinearFormIntegrator::AssembleFaceMatrix;
-   void AssembleFaceMatrix(const FiniteElement &el1,
-                           const FiniteElement &el2,
+   void AssembleFaceMatrix(const FiniteElement &el1, const FiniteElement &el2,
                           FaceElementTransformations &Trans,
                           DenseMatrix &elmat) override;

@@ -3459,6 +3539,28 @@ 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);
 };
@@ -8,6 +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_conv and f_apply_conv
 struct ConvectionContext {
@@ -91,7 +92,7 @@ CEED_QFUNCTION(f_build_conv_const)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return 0;
+   return CEED_ERROR_SUCCESS;
 }

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

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

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

 CEED_QFUNCTION(f_apply_conv_mf_quad)(void *ctx, CeedInt Q,
@@ -525,5 +526,5 @@ CEED_QFUNCTION(f_apply_conv_mf_quad)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return 0;
+   return CEED_ERROR_SUCCESS;
 }
@@ -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 0;
+   return CEED_ERROR_SUCCESS;
 }

 /// 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 0;
+   return CEED_ERROR_SUCCESS;
 }

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

 /// 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 0;
+   return CEED_ERROR_SUCCESS;
 }

 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 0;
+   return CEED_ERROR_SUCCESS;
 }
@@ -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 0;
+   return CEED_ERROR_SUCCESS;
 }

 /// 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 0;
+   return CEED_ERROR_SUCCESS;
 }

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

 /// 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 0;
+   return CEED_ERROR_SUCCESS;
 }

 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 0;
+   return CEED_ERROR_SUCCESS;
 }
@@ -8,6 +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_conv and f_apply_conv
 struct NLConvectionContext { CeedInt dim, space_dim, vdim; CeedScalar coeff; };
@@ -87,7 +88,7 @@ CEED_QFUNCTION(f_build_conv_const)(void *ctx, CeedInt Q,
         }
         break;
   }
-   return 0;
+   return CEED_ERROR_SUCCESS;
 }

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

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

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

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

 #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
-      CeedCompositeOperatorCreate(internal::ceed, &oper);
+      CeedOperatorCreateComposite(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);
-         CeedCompositeOperatorAddSub(oper, sub_op->GetCeedOperator());
+         CeedOperatorCompositeAddSub(oper, sub_op->GetCeedOperator());
      }

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

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

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

   int nsub;
   CeedOperator *subops;
-#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
+   ierr = CeedOperatorCompositeGetNumSub(op, &nsub); PCeedChk(ierr);
+   ierr = CeedOperatorCompositeGetSubList(op, &subops); PCeedChk(ierr);
   for (int isub=0; isub<nsub; ++isub)
   {
      CeedOperator subop = subops[isub];
@@ -294,7 +289,7 @@ CeedOperator CoarsenCeedCompositeOperator(
      // refcounted by existing objects
      ierr = CeedBasisDestroy(&basis_coarse); PCeedChk(ierr);
      ierr = CeedBasisDestroy(&basis_c2f); PCeedChk(ierr);
-      ierr = CeedCompositeOperatorAddSub(op_coarse, subop_coarse);
+      ierr = CeedOperatorCompositeAddSub(op_coarse, subop_coarse);
      PCeedChk(ierr);
      ierr = CeedOperatorDestroy(&subop_coarse); PCeedChk(ierr);
   }
@@ -81,12 +81,27 @@ int CeedSingleOperatorFullAssemble(CeedOperator op, SparseMatrix *out)
      ierr = CeedOperatorFieldGetVector(input_fields[i], &vec); PCeedChk(ierr);
      if (vec == CEED_VECTOR_ACTIVE)
      {
-         ierr = CeedOperatorFieldGetBasis(input_fields[i], &basisin);
-         PCeedChk(ierr);
+         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 = CeedBasisGetNumComponents(basisin, &ncomp); PCeedChk(ierr);
         ierr = CeedBasisGetDimension(basisin, &dim); PCeedChk(ierr);
-         ierr = CeedOperatorFieldGetElemRestriction(input_fields[i], &rstrin);
+         CeedElemRestriction rstr;
+         ierr = CeedOperatorFieldGetElemRestriction(input_fields[i], &rstr);
         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);
@@ -112,6 +127,9 @@ 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
@@ -127,11 +145,25 @@ int CeedSingleOperatorFullAssemble(CeedOperator op, SparseMatrix *out)
      ierr = CeedOperatorFieldGetVector(output_fields[i], &vec); PCeedChk(ierr);
      if (vec == CEED_VECTOR_ACTIVE)
      {
-         ierr = CeedOperatorFieldGetBasis(output_fields[i], &basisout);
-         PCeedChk(ierr);
-         ierr = CeedOperatorFieldGetElemRestriction(output_fields[i], &rstrout);
-         PCeedChk(ierr);
+         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);
         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);
@@ -157,6 +189,9 @@ 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;
@@ -200,7 +235,11 @@ 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
@@ -285,6 +324,10 @@ 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);

@@ -310,13 +353,8 @@ int CeedOperatorFullAssemble(CeedOperator op, SparseMatrix **mat)
   {
      CeedInt numsub;
      CeedOperator *subops;
-#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
+      ierr = CeedOperatorCompositeGetNumSub(op, &numsub); PCeedChk(ierr);
+      ierr = CeedOperatorCompositeGetSubList(op, &subops); PCeedChk(ierr);
      for (int i = 0; i < numsub; ++i)
      {
         ierr = CeedSingleOperatorFullAssemble(subops[i], out); PCeedChk(ierr);
@@ -120,7 +120,11 @@ 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"
@@ -664,7 +668,11 @@ 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)
      {
@@ -676,10 +684,19 @@ 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)
      {
@@ -691,6 +708,11 @@ 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;
@@ -741,7 +763,9 @@ 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;
 }

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

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

 #include <cmath>
 #include <limits>
@@ -80,6 +81,49 @@ 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)
 {
@@ -519,6 +563,26 @@ 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)
@@ -1065,6 +1129,41 @@ 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())
@@ -132,6 +132,9 @@ 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
@@ -894,6 +897,9 @@ public:
   void Eval(DenseMatrix &M, ElementTransformation &T,
             const IntegrationRule &ir) override;

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

@@ -1771,6 +1777,9 @@ 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.
@@ -259,6 +259,30 @@ 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
@@ -401,9 +401,6 @@ 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;
 }
@@ -2459,8 +2456,7 @@ 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 &&
-       ob_type != BasisType::IntegratedGLL)
+   if (Quadrature1D::CheckOpen(op_type) == Quadrature1D::Invalid)
   {
      const char *ob_name = BasisType::Name(ob_type); // this may abort
      MFEM_ABORT("unknown open BasisType: " << ob_name);
@@ -2518,6 +2514,7 @@ 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)) ==
@@ -2786,8 +2783,7 @@ 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 &&
-       ob_type != BasisType::IntegratedGLL)
+   if (Quadrature1D::CheckOpen(op_type) == Quadrature1D::Invalid)
   {
      const char *ob_name = BasisType::Name(ob_type);
      MFEM_ABORT("Invalid open basis point type: " << ob_name);
@@ -464,6 +464,13 @@ 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
@@ -475,6 +482,13 @@ 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.
@@ -224,7 +224,7 @@ struct DerefineMatrixOpMultFunctor
               sum += sign * bsptr[boptr[k] + i + j * block_height] *
                      xptr[this->IndexX(col, vdim, k)];
            }
-#if defined(__CUDA_ARCH__) or defined(__HIP_DEVICE_COMPILE__)
+#if defined(__CUDA_ARCH__) || defined(__HIP_DEVICE_COMPILE__)
            if (Atomic)
            {
               atomicAdd(yptr + this->IndexY(row, vdim), sum);
@@ -683,8 +683,12 @@ public:
   NURBSExtension *GetNURBSext() { return NURBSext; }
   NURBSExtension *StealNURBSext();

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

   /** Set the prolongation operator of the space to an arbitrary sparse matrix,
       creating a copy of the argument. */
@@ -922,6 +926,9 @@ 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() &&
+      if (fes->Nonconforming() && fes->GetMesh()->ncmesh &&
          fes->GetMesh()->ncmesh->IsLegacyLoaded())
      {
         LegacyNCReorder();
@@ -1374,6 +1374,50 @@ 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;
@@ -2624,7 +2668,7 @@ void GridFunction::ProjectBdrCoefficient(Coefficient *coeff[],
   }
   for (int i = 0; i < values_counter.Size(); i++)
   {
-      MFEM_ASSERT(bool(values_counter[i]) == ess_vdofs_marker[i],
+      MFEM_ASSERT(bool(values_counter[i]) == bool(ess_vdofs_marker[i]),
                  "internal error");
   }
 #endif
@@ -153,7 +153,8 @@ public:
   /// Shortcut for calling SetFromTrueDofs() with GetTrueVector() as argument.
   void SetFromTrueVector() { SetFromTrueDofs(GetTrueVector()); }

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

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

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

   void GetVectorFieldNodalValues(Vector &val, int comp) const;
@@ -359,6 +361,33 @@ 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(NULL),
-     fec_map_lin(NULL),
-     fdataD(NULL), cr(NULL), gsl_comm(NULL),
+   : mesh(nullptr),
+     fec_map_lin(nullptr),
+     fdataD(nullptr), cr(nullptr), gsl_comm(nullptr),
     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] = NULL;
-      ir_split[i] = NULL;
-      fes_rst_map[i] = NULL;
-      gf_rst_map[i] = NULL;
+      mesh_split[i] = nullptr;
+      ir_split[i] = nullptr;
+      fes_rst_map[i] = nullptr;
+      gf_rst_map[i] = nullptr;
   }

   gsl_comm = new gslib::comm;
@@ -117,27 +117,40 @@ 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()
 {
-   crystal_free(cr);
-   comm_free(gsl_comm);
-   delete gsl_comm;
-   delete cr;
-   for (int i = 0; i < 4; i++)
+   FreeData();
+#ifdef MFEM_USE_MPI
+   if (!Mpi::IsFinalized())  // currently segfaults inside gslib otherwise
+#endif
   {
-      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; }
+      crystal_free(cr);
+      comm_free(gsl_comm);
+      delete gsl_comm;
+      delete cr;
   }
-   if (fec_map_lin) { delete fec_map_lin; fec_map_lin = 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; }
 }

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

   gsl_comm = new gslib::comm;
@@ -158,12 +171,21 @@ 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
@@ -171,37 +193,9 @@ void FindPointsGSLIB::Setup(Mesh &m, const double bb_t, const double newt_tol,

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

-   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);
-   }
+   SetupSplitMeshesAndIntegrationRules(meshOrder);

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

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

 void FindPointsGSLIB::SetupSplitMeshes()
 {
-   fec_map_lin = new H1_FECollection(1, dim);
-   if (mesh->Dimension() == 2)
+   if (fec_map_lin == nullptr) { fec_map_lin = new H1_FECollection(1, dim); }
+   if (dim == 2)
   {
      int Nvert = 7;
      int NEsplit = 3;
@@ -1201,7 +1200,7 @@ void FindPointsGSLIB::SetupSplitMeshes()
      mesh_split[1] = new Mesh(Mesh::MakeCartesian2D(1, 1,
                                                     Element::QUADRILATERAL));
   }
-   else if (mesh->Dimension() == 3)
+   else if (dim == 3)
   {
      mesh_split[0] = new Mesh(Mesh::MakeCartesian3D(1, 1, 1,
                                                     Element::HEXAHEDRON));
@@ -1346,41 +1345,6 @@ 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,
@@ -1431,6 +1395,79 @@ 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)
 {
@@ -2081,6 +2118,19 @@ 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)
@@ -2386,6 +2436,10 @@ 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(); }
@@ -2395,35 +2449,7 @@ void OversetFindPointsGSLIB::Setup(Mesh &m, const int meshid,
   const FiniteElement *fe = mesh->GetNodalFESpace()->GetTypicalFE();
   unsigned dof1D = fe->GetOrder() + 1;

-   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);
-   }
+   SetupSplitMeshesAndIntegrationRules(meshOrder);

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

@@ -2480,7 +2506,7 @@ void OversetFindPointsGSLIB::FindPoints(const Vector &point_pos,
 {
   MFEM_VERIFY(setupflag, "Use OversetFindPointsGSLIB::Setup before "
               "finding points.");
-   MFEM_VERIFY(overset, "Please setup FindPoints for overlapping grids.");
+   MFEM_VERIFY(overset, "Please use OversetFindPoints 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

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

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

 #ifdef MFEM_USE_GSLIB

@@ -131,6 +135,10 @@ 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);

@@ -190,14 +198,23 @@ 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.
@@ -212,8 +229,8 @@ 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
@@ -289,7 +306,12 @@ 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. */
+       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.
+   */
   virtual void FreeData();

   /// Return code for each point searched by FindPoints: inside element (0), on
@@ -312,6 +334,9 @@ 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 = el2.GetDof();
+   const int dof2 = (Tr.Elem2No >= 0)?(el2.GetDof()):(0);

 #ifdef MFEM_THREAD_SAFE
   // Local storage for element integration
@@ -219,7 +219,9 @@ void HyperbolicFormIntegrator::AssembleFaceVector(
   const IntegrationRule *ir = IntRule;
   if (!ir)
   {
-      const int order = 2*std::max(el1.GetOrder(), el2.GetOrder()) + IntOrderOffset;
+      const int max_el_order = dof2 ? std::max(el1.GetOrder(),
+                                               el2.GetOrder()) : el1.GetOrder();
+      const int order = 2*max_el_order + IntOrderOffset;
      ir = &IntRules.Get(Tr.GetGeometryType(), order);
   }
   // loop over integration points
@@ -231,18 +233,22 @@ 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);
-      elfun2_mat.MultTranspose(shape2, state2);
+
+      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);
+      }

      // Get the normal vector and the flux on the face
      if (nor.Size() == 1)  // if 1D, use 1 or -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;
+         nor(0) = 2*Tr.GetElement1IntPoint().x - 1.;
      }
      else
      {
@@ -250,14 +256,18 @@ 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 = numFlux.Eval(state1, state2, nor, Tr, fluxN);
+      const real_t speed = (dof2) ? numFlux.Eval(state1, state2, nor, Tr, fluxN):
+                           fluxFunction.ComputeFluxDotN(state1, 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);
-      AddMult_a_VWt(+ip.weight*sign, shape2, fluxN, elvect2_mat);
+      if (dof2)
+      {
+         AddMult_a_VWt(+ip.weight*sign, shape2, fluxN, elvect2_mat);
+      }
   }
 }

@@ -268,7 +278,7 @@ 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 = el2.GetDof();
+   const int dof2 = (Tr.Elem2No >= 0)?(el2.GetDof()):(0);

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

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

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

      // Interpolate elfun at the point
      elfun1_mat.MultTranspose(shape1, state1);
-      elfun2_mat.MultTranspose(shape2, state2);
+
+      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);
+      }

      // Get the normal vector and the flux on the face
      if (nor.Size() == 1)  // if 1D, use 1 or -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;
+         nor(0) = 2*Tr.GetElement1IntPoint().x - 1.;
      }
      else
      {
@@ -335,7 +352,14 @@ void HyperbolicFormIntegrator::AssembleFaceGrad(
      // Trial side 1

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

      const int ioff = fluxFunction.num_equations * dof1;

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

-      // Trial side 2
+      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);
+      }

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

      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++)
+            const real_t w = -ip.weight * sign * JDotN(di,dj);
+            for (int j = 0; j < dof; j++)
+               for (int i = 0; i < dof; i++)
               {
-                  elmat(i+dof1*di, joff+j+dof2*dj) += w * shape1(i) * shape2(j);
+                  elmat(i+dof*di, j+dof*dj) += w * shape(i) * shape(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,
@@ -306,12 +306,14 @@ 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
@@ -331,8 +333,9 @@ private:

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

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

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

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

   /**
    * @brief Implements (F(u), ∇v) with abstract F computed by
@@ -416,6 +412,182 @@ 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,
@@ -15,6 +15,86 @@
 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
@@ -77,129 +157,16 @@ void CurlCurlIntegrator::AssemblePA(const FiniteElementSpace &fes)

 void CurlCurlIntegrator::AssembleDiagonalPA(Vector& 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!");
-   }
+   DiagonalPAKernels::Run(dim, dofs1D, quad1D, dofs1D, quad1D, symmetric, ne,
+                          mapsO->B, mapsC->B, mapsO->G, mapsC->G, pa_data,
+                          diag);
 }

 void CurlCurlIntegrator::AddMultPA(const Vector &x, Vector &y) const
 {
-   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!");
-   }
+   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);
 }

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

-   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!");
-   }
+   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);
 }

 } // namespace mfem
@@ -0,0 +1,500 @@
+// 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
@@ -11,42 +11,39 @@

 #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"

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

 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();

@@ -56,7 +53,7 @@ static void PADGDiffusionSetup2D(const int Q1D,
   // (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)};
@@ -74,7 +71,7 @@ static void PADGDiffusionSetup2D(const int Q1D,

      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;
@@ -85,17 +82,19 @@ static void PADGDiffusionSetup2D(const int Q1D,

            // 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;
@@ -104,9 +103,9 @@ static void PADGDiffusionSetup2D(const int Q1D,
            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;
         }
@@ -122,47 +121,43 @@ static void PADGDiffusionSetup2D(const int Q1D,
   });
 }

-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];
@@ -172,10 +167,7 @@ static void PADGDiffusionSetup3D(const int Q1D,

      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)
         {
@@ -200,16 +192,16 @@ static void PADGDiffusionSetup3D(const int Q1D,
      {
         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;
@@ -217,27 +209,45 @@ static void PADGDiffusionSetup3D(const int Q1D,

               // *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;
@@ -257,7 +267,8 @@ static void PADGDiffusionSetup3D(const int Q1D,
 }

 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();

@@ -326,8 +337,7 @@ 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)
 {
@@ -386,17 +396,19 @@ inline void SignedFaceNormalPermutation(int perm[3],
 }

 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)
@@ -408,9 +420,9 @@ 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);

@@ -421,16 +433,17 @@ 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
         {
@@ -448,8 +461,8 @@ 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);
@@ -458,16 +471,17 @@ 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;
@@ -476,8 +490,8 @@ 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;
@@ -489,9 +503,18 @@ 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)
@@ -501,14 +524,16 @@ 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);
   }
 }

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

-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)
+void DGDiffusionIntegrator::AddMultPAFaceNormalDerivatives(const Vector &x,
+                                                           const Vector &dxdn,
+                                                           Vector &y,
+                                                           Vector &dydn) const
 {
-   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
+   ApplyPAKernels::Run(dim, dofs1D, quad1D, nf, maps->B, maps->Bt, maps->G,
+                       maps->Gt, sigma, pa_data, x, dxdn, y, dydn, dofs1D,
+                       quad1D);
 }

-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)
+DGDiffusionIntegrator::DGDiffusionIntegrator(const real_t s, const real_t k)
+   : sigma(s), kappa(k)
 {
-   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];
-            }
-         }
-      }
-   });
+   static Kernels kernels;
 }

-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)
+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)
 {
   if (dim == 2)
   {
-      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);
+      return internal::PADGDiffusionApply2D;
   }
   else if (dim == 3)
   {
-      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);
+      return internal::PADGDiffusionApply3D;
   }
   else
   {
-      MFEM_ABORT("Unsupported dimension");
+      MFEM_ABORT("");
   }
 }

-void DGDiffusionIntegrator::AddMultPAFaceNormalDerivatives(
-   const Vector &x, const Vector &dxdn, Vector &y, Vector &dydn) const
+DGDiffusionIntegrator::Kernels::Kernels()
 {
-   PADGDiffusionApply(dim, dofs1D, quad1D, nf,
-                      maps->B, maps->Bt, maps->G, maps->Gt,
-                      sigma, pa_data, x, dxdn, y, dydn);
+   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>();
 }

+/// \endcond DO_NOT_DOCUMENT
+
 } // namespace mfem
@@ -0,0 +1,793 @@
+// 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
@@ -19,6 +19,8 @@ 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>();
@@ -27,17 +29,39 @@ 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][MD1][MQ1];
+      MFEM_SHARED real_t GQ[2][NBZ][MQ1][MQ1];
      real_t (*X)[MD1] = (real_t (*)[MD1])(Xz + 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);
+      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);
      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 (DIM == 2) { return internal::SmemPADiffusionApply2D<T_D1D,T_Q1D>; }
-   else if (DIM == 3) { return internal::SmemPADiffusionApply3D<T_D1D, T_Q1D>; }
-   else { MFEM_ABORT(""); }
+   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("");
 }

 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 (DIM == 2) { return internal::SmemPADiffusionDiagonal2D<D1D,Q1D>; }
-   else if (DIM == 3) { return internal::SmemPADiffusionDiagonal3D<D1D, Q1D>; }
-   else { MFEM_ABORT(""); }
+   if constexpr (DIM == 2) { return internal::SmemPADiffusionDiagonal2D<D1D,Q1D>; }
+   else if constexpr (DIM == 3) { return internal::SmemPADiffusionDiagonal3D<D1D, Q1D>; }
+   MFEM_ABORT("");
 }

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

-void PACurlCurlAssembleDiagonal2D(const int D1D,
-                                  const int Q1D,
-                                  const int NE,
-                                  const Array<real_t> &bo,
+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> &,
                                  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);
@@ -653,16 +651,11 @@ void PACurlCurlAssembleDiagonal2D(const int D1D,
   }); // end of element loop
 }

-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,
+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,
                       const bool useAbs)
 {

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

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

@@ -426,8 +426,11 @@ 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);
@@ -831,9 +834,12 @@ 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,
@@ -3158,6 +3164,49 @@ 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
@@ -19,6 +19,7 @@
 #include "../../linalg/vector.hpp"
 #include "../bilininteg.hpp"

+/// \cond DO_NOT_DOCUMENT
 namespace mfem
 {

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

 } // namespace mfem

+/// \endcond DO_NOT_DOCUMENT
+
 #endif
@@ -17,6 +17,8 @@ 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>();
@@ -25,17 +27,45 @@ 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 (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(""); }
+   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("");
 }

 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 (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(""); }
+   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("");
 }

 inline DiagonalKernelType MassIntegrator::DiagonalPAKernels::Fallback(
@@ -0,0 +1,355 @@
+// 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
@@ -15,9 +15,58 @@
 #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,
@@ -425,322 +474,6 @@ 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
 {
@@ -757,27 +490,29 @@ 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;
-
-      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.");
+      ApplyPAKernels::Run(dim, sdim, D1D, Q1D, ne, B, G, Bt, Gt, D, x, y, D1D,
+                          Q1D, sdim);
   }
 }

+/// \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
@@ -9,183 +9,19 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.

-#include "../../general/forall.hpp"
 #include "../../fem/kernels.hpp"
+#include "../../general/forall.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();
@@ -197,50 +33,20 @@ 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 int *M = markers.Read();
   const real_t *B = maps.B.Read();
+   const int *M = markers.Read();
   const real_t *detJ = geom->detJ.Read();
   const real_t *W = ir->GetWeights().Read();
   real_t *Y = y.ReadWrite();
-   ker(vdim, ne, d, q, map_type, M, B, detJ, W, coeff, Y);
+   DomainLFIntegrator::AssembleKernels::Run(dim, d, q, 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;
@@ -266,4 +72,49 @@ 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
@@ -0,0 +1,318 @@
+// 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
@@ -947,6 +947,7 @@ int Quadrature1D::CheckOpen(int type)
      case OpenUniform:
      case ClosedUniform:
      case OpenHalfUniform:
+      case ClosedGL:
         return type; // all types can work as open
      default:
         return Invalid;
@@ -78,7 +78,7 @@ namespace mfem
  public:                                                                      \
    const char *kernel_name = MFEM_KERNEL_NAME(KernelName);                    \
    using KernelSignature = KernelType;                                        \
-    template <MFEM_PARAM_LIST P3> static MFEM_EXPORT KernelSignature Kernel(); \
+    template <MFEM_PARAM_LIST P3> static KernelSignature Kernel();             \
    static MFEM_EXPORT KernelSignature Fallback(MFEM_PARAM_LIST P1);           \
    static MFEM_EXPORT KernelName &Get() {                                     \
      static KernelName table;                                                 \
@@ -35,6 +35,19 @@ void LinearFormIntegrator::AssembleRHSElementVect(
   mfem_error("LinearFormIntegrator::AssembleRHSElementVect(...)");
 }

+DomainLFIntegrator::DomainLFIntegrator(Coefficient &QF, int a, int b)
+   : DeltaLFIntegrator(QF), Q(QF), oa(a), ob(b)
+{
+   static Kernels kernels;
+}
+
+DomainLFIntegrator::DomainLFIntegrator(Coefficient &QF,
+                                       const IntegrationRule *ir)
+   : DeltaLFIntegrator(QF, ir), Q(QF), oa(1), ob(1)
+{
+   static Kernels kernels;
+}
+
 void DomainLFIntegrator::AssembleRHSElementVect(const FiniteElement &el,
                                                ElementTransformation &Tr,
                                                Vector &elvect)
@@ -266,6 +279,13 @@ void BoundaryTangentialLFIntegrator::AssembleRHSElementVect(
   }
 }

+VectorDomainLFIntegrator::VectorDomainLFIntegrator(VectorCoefficient &QF,
+                                                   const IntegrationRule *ir)
+   : DeltaLFIntegrator(QF, ir), Q(QF)
+{
+   static DomainLFIntegrator::Kernels kernels;
+}
+
 void VectorDomainLFIntegrator::AssembleRHSElementVect(
   const FiniteElement &el, ElementTransformation &Tr, Vector &elvect)
 {
@@ -18,6 +18,8 @@
 #include <random>
 #include "integrator.hpp"

+#include "kernel_dispatch.hpp"
+
 namespace mfem
 {

@@ -109,14 +111,12 @@ class DomainLFIntegrator : public DeltaLFIntegrator
   int oa, ob;
 public:
   /// Constructs a domain integrator with a given Coefficient
-   DomainLFIntegrator(Coefficient &QF, int a = 2, int b = 0)
-   // the old default was a = 1, b = 1
-   // for simple elliptic problems a = 2, b = -2 is OK
-      : DeltaLFIntegrator(QF), Q(QF), oa(a), ob(b) { }
+   /// the old default was a = 1, b = 1
+   /// for simple elliptic problems a = 2, b = -2 is OK
+   DomainLFIntegrator(Coefficient &QF, int a = 2, int b = 0);

   /// Constructs a domain integrator with a given Coefficient
-   DomainLFIntegrator(Coefficient &QF, const IntegrationRule *ir)
-      : DeltaLFIntegrator(QF, ir), Q(QF), oa(1), ob(1) { }
+   DomainLFIntegrator(Coefficient &QF, const IntegrationRule *ir);

   bool SupportsDevice() const override { return true; }

@@ -136,6 +136,22 @@ public:
                                 Vector &elvect) override;

   using LinearFormIntegrator::AssembleRHSElementVect;
+
+   /// args: vdim, ne, d1d, q1d, map_type, markers, B, detJ, W, coeff, y
+   using AssembleKernelType = void (*)(const int, const int, const int,
+                                       const int, const int, const int *,
+                                       const real_t *, const real_t *,
+                                       const real_t *, const Vector &coeff,
+                                       real_t *y);
+
+   /// parameters: use DIM, T_D1D, T_Q1D
+   MFEM_REGISTER_KERNELS(AssembleKernels, AssembleKernelType, (int, int, int));
+   struct Kernels { Kernels(); };
+
+   template <int DIM, int D1D, int Q1D> static void AddSpecialization()
+   {
+      AssembleKernels::Specialization<DIM, D1D, Q1D>::Add();
+   }
 };

 /// Class for domain integrator $ L(v) := (f, \nabla v) $
@@ -256,14 +272,13 @@ private:

 public:
   /// Constructs a domain integrator with a given VectorCoefficient
-   VectorDomainLFIntegrator(VectorCoefficient &QF)
-      : DeltaLFIntegrator(QF), Q(QF) { }
+   VectorDomainLFIntegrator(VectorCoefficient &QF,
+                            const IntegrationRule *ir = nullptr);

   bool SupportsDevice() const override { return true; }

   /// Method defining assembly on device
-   void AssembleDevice(const FiniteElementSpace &fes,
-                       const Array<int> &markers,
+   void AssembleDevice(const FiniteElementSpace &fes, const Array<int> &markers,
                       Vector &b) override;

   /** Given a particular Finite Element and a transformation (Tr)
@@ -277,6 +292,12 @@ public:
                                 Vector &elvect) override;

   using LinearFormIntegrator::AssembleRHSElementVect;
+
+   template <int DIM, int D1D, int Q1D> static void AddSpecialization()
+   {
+      // uses the same kernels for assembly
+      DomainLFIntegrator::AssembleKernels::Specialization<DIM, D1D, Q1D>::Add();
+   }
 };

 /** Class for domain integrator $ L(v) := (f, \nabla v) $, where
@@ -544,7 +565,7 @@ public:
    Specifically, given the Dirichlet data $u_D$, the linear form assembles the
    following integrals on the boundary:
   $$
-    \sigma \langle u_D, (Q \nabla v)) \cdot n \rangle + \kappa \langle {h^{-1} Q} u_D, v \rangle,
+    \sigma \langle u_D, (Q \nabla v) \cdot n \rangle + \kappa \langle {h^{-1} Q} u_D, v \rangle,
   $$
    where Q is a scalar or matrix diffusion coefficient and v is the test
    function. The parameters $\sigma$ and $\kappa$ should be the same as the ones
@@ -14,9 +14,11 @@
 #include "../../general/forall.hpp"
 #include <climits>
 #include "../pbilinearform.hpp"
+#include "../../fem/fe/face_map_utils.hpp"

 // Specializations
 #include "lor_h1.hpp"
+#include "lor_dg.hpp"
 #include "lor_nd.hpp"
 #include "lor_rt.hpp"

@@ -54,17 +56,18 @@ bool BatchedLORAssembly::FormIsSupported(BilinearForm &a)
   // Batched LOR requires all tensor elements
   if (!UsesTensorBasis(*a.FESpace())) { return false; }

-   if (dynamic_cast<const H1_FECollection*>(fec))
+   if (dynamic_cast<const H1_FECollection*>(fec) ||
+       dynamic_cast<const DG_FECollection*>(fec))
   {
-      if (HasIntegrators<DiffusionIntegrator, MassIntegrator>(a)) { return true; }
+      return HasIntegrators<DiffusionIntegrator, MassIntegrator>(a);
   }
   else if (dynamic_cast<const ND_FECollection*>(fec))
   {
-      if (HasIntegrators<CurlCurlIntegrator, VectorFEMassIntegrator>(a)) { return true; }
+      return HasIntegrators<CurlCurlIntegrator, VectorFEMassIntegrator>(a);
   }
   else if (dynamic_cast<const RT_FECollection*>(fec))
   {
-      if (HasIntegrators<DivDivIntegrator, VectorFEMassIntegrator>(a)) { return true; }
+      return HasIntegrators<DivDivIntegrator, VectorFEMassIntegrator>(a);
   }
   return false;
 }
@@ -75,12 +78,14 @@ void BatchedLORAssembly::FormLORVertexCoordinates(FiniteElementSpace &fes_ho,
   Mesh &mesh_ho = *fes_ho.GetMesh();
   mesh_ho.EnsureNodes();

+   const bool dg = fes_ho.IsDGSpace();
+
   // Get nodal points at the LOR vertices
   const int dim = mesh_ho.Dimension();
   const int sdim = mesh_ho.SpaceDimension();
   const int nel_ho = mesh_ho.GetNE();
   const int order = fes_ho.GetMaxElementOrder();
-   const int nd1d = order + 1;
+   const int nd1d = dg ? order + 2 : order + 1;
   const int ndof_per_el = static_cast<int>(pow(nd1d, dim));

   const GridFunction *nodal_gf = mesh_ho.GetNodes();
@@ -92,7 +97,8 @@ void BatchedLORAssembly::FormLORVertexCoordinates(FiniteElementSpace &fes_ho,
   Vector nodal_evec(nodal_restriction->Height());
   nodal_restriction->Mult(*nodal_gf, nodal_evec);

-   IntegrationRule ir = GetCollocatedIntRule(fes_ho);
+   const IntegrationRule ir = GetLobattoIntRule(
+                                 mesh_ho.GetTypicalElementGeometry(), nd1d);

   // Map from nodal E-vector to Q-vector at the LOR vertex points
   X_vert.SetSize(sdim*ndof_per_el*nel_ho);
@@ -159,6 +165,7 @@ int BatchedLORAssembly::FillI(SparseMatrix &A) const
   const auto K = dof_glob2loc_offsets_.Read();
   const auto map = Reshape(sparse_mapping.Read(), nnz_per_row, ndof_per_el);

+
   auto I = A.WriteI();

   mfem::forall(nvdof + 1, [=] MFEM_HOST_DEVICE (int ii) { I[ii] = 0; });
@@ -358,6 +365,177 @@ void BatchedLORAssembly::FillJAndData(SparseMatrix &A) const
   });
 }

+void BatchedLORAssembly::SparseIJToCSR_DG(OperatorHandle &A) const
+{
+   const int ndof_per_el = fes_ho.GetFE(0)->GetDof();
+   const int nel_ho = fes_ho.GetNE();
+   const int nnz_per_row = sparse_ij.Size()/ndof_per_el/nel_ho;
+   const int dim = fes_ho.GetMesh()->Dimension();
+   const int nrows = nel_ho*ndof_per_el;
+   const int p = fes_ho.GetMaxElementOrder();
+   const int pp1 = p + 1;
+   const int nnz = nrows*nnz_per_row;
+
+   const int face_nbr_vsize = [&]()
+   {
+#ifdef MFEM_USE_MPI
+      if (auto *par_fes = dynamic_cast<ParFiniteElementSpace*>(&fes_ho))
+      {
+         return par_fes->GetFaceNbrVSize();
+      }
+#endif
+      return 0;
+   }();
+
+   // If A contains an existing SparseMatrix, reuse it (and try to reuse its
+   // I, J, A arrays if they are big enough)
+   SparseMatrix *A_mat = A.Is<SparseMatrix>();
+   if (!A_mat)
+   {
+      A_mat = new SparseMatrix;
+      A.Reset(A_mat);
+   }
+
+   // The second argument (nrows + face_nbr_vsize) accounts for additional
+   // columns contributed by DG face neighbors in parallel finite element
+   // spaces. In serial, face_nbr_vsize is set to 0.
+   A_mat->OverrideSize(nrows, nrows + face_nbr_vsize);
+
+   EnsureCapacity(A_mat->GetMemoryI(), nrows + 1);
+   EnsureCapacity(A_mat->GetMemoryJ(), nnz);
+   EnsureCapacity(A_mat->GetMemoryData(), nnz);
+
+   Array<int> nbr_info(nel_ho*3*2*dim);
+   auto h_nbr_info = Reshape(nbr_info.HostWrite(), nel_ho, 2*dim, 3);
+   const int num_faces = fes_ho.GetMesh()->GetNumFaces();
+   for (int f = 0; f < num_faces; f++)
+   {
+      Mesh::FaceInformation finfo = fes_ho.GetMesh()->GetFaceInformation(f);
+      int e0 = finfo.element[0].index;
+      int f0 = finfo.element[0].local_face_id;
+      if (finfo.IsBoundary())
+      {
+         h_nbr_info(e0,f0,0) = -1;
+         h_nbr_info(e0,f0,1)= -1;
+         h_nbr_info(e0,f0,2)= -1;
+      }
+      else if (finfo.IsShared())
+      {
+         // Face neighbors elements are indexed after the last local element
+         h_nbr_info(e0,f0,0) = nel_ho + finfo.element[1].index;
+         h_nbr_info(e0,f0,1)= finfo.element[1].orientation;
+         h_nbr_info(e0,f0,2)= finfo.element[1].local_face_id;
+      }
+      else if (finfo.IsInterior())
+      {
+         int e1 = finfo.element[1].index;
+         int f1 = finfo.element[1].local_face_id;
+         h_nbr_info(e0,f0,0) = e1;
+         h_nbr_info(e0,f0,1)= finfo.element[1].orientation;
+         h_nbr_info(e0,f0,2)= f1;
+         h_nbr_info(e1,f1,0) = e0;
+         h_nbr_info(e1,f1,1) = finfo.element[1].orientation;
+         h_nbr_info(e1,f1,2) = f0;
+      }
+   };
+
+   auto h_I = A_mat->HostWriteI();
+   h_I[0] = 0;
+   for (int i = 0; i < nrows; ++i)
+   {
+      const int iel_ho = i / ndof_per_el;
+      const int iloc = i % ndof_per_el;
+      static const int lex_map_2[4] = {3, 1, 0, 2};
+      static const int lex_map_3[6] = {4, 2, 1, 3, 0, 5};
+      const int local_i[3] = {iloc % pp1, (iloc/pp1)%pp1, iloc/pp1/pp1};
+      int bdr_count = 0;
+      for (int n_idx = 0; n_idx < dim; ++n_idx)
+      {
+         for (int e_i = 0; e_i < 2; ++e_i)
+         {
+            const int j_lex = e_i + n_idx*2;
+            const int f = (dim == 3) ? lex_map_3[j_lex]:lex_map_2[j_lex];
+            const bool boundary = (local_i[n_idx] == e_i * p);
+            if (boundary)
+            {
+               int neighbor_idx = h_nbr_info(iel_ho, f, 0);
+               if (neighbor_idx == -1)
+               {
+                  ++bdr_count;
+               }
+            }
+         }
+      }
+      h_I[i+1] = h_I[i] + (nnz_per_row - bdr_count);
+   }
+
+   const auto V = Reshape(sparse_ij.Read(), nnz_per_row, ndof_per_el, nel_ho);
+   auto J = A_mat->WriteJ();
+   auto AV = A_mat->WriteData();
+   auto I = A_mat->ReadI();
+
+   auto d_nbr_info = Reshape(nbr_info.Read(), nel_ho, 2*dim, 3);
+   mfem::forall(nrows, [=] MFEM_HOST_DEVICE (int i)
+   {
+      const int e = i / ndof_per_el;
+      const int iloc = i % ndof_per_el;
+      const int local_x = iloc % pp1;
+      const int local_y = (iloc/pp1)%pp1;
+      const int local_z = iloc/pp1/pp1;
+      const int local_i[3] = {local_x, local_y, local_z};
+      int offset = I[i];
+      static const int lex_map_2[4] = {3, 1, 0, 2};
+      static const int lex_map_3[6] = {4,2,1,3,0,5};
+      const int *lex_map = (dim == 2) ? lex_map_2 : lex_map_3;
+      AV[offset] = V(0, iloc, e);
+      J[offset] = i;
+      ++offset;
+      for (int n_idx = 0; n_idx < dim; ++n_idx)
+      {
+         // qi is the face lexicographic index, obtained by taking the
+         // lexicographic index of the coordinates ommiting n_idx.
+         int qi = 0;
+         int stride = 1;
+         for (int d = 0; d < dim; ++d)
+         {
+            if (d != n_idx)
+            {
+               qi += local_i[d]*stride;
+               stride *= pp1;
+            }
+         }
+         for (int e_i = 0; e_i < 2; ++e_i)
+         {
+            const int j_lex = e_i + n_idx*2;
+            const int f = lex_map[j_lex];
+            const bool bdr = (local_i[n_idx] == e_i * p);
+            if (bdr)
+            {
+               const int nbr_e = d_nbr_info(e, f, 0);
+               const int nbr_ori = d_nbr_info(e, f, 1);
+               const int nbr_f = d_nbr_info(e, f, 2);
+               if (nbr_e != -1)
+               {
+                  const int nbr_loc_idx = internal::FaceIdxToVolIdx(
+                                             dim, qi, pp1, f, nbr_f, 1, nbr_ori);
+                  J[offset] = nbr_e*ndof_per_el + nbr_loc_idx;
+                  AV[offset] = V(f+1, iloc, e);
+                  ++offset;
+               }
+            }
+            else
+            {
+               int shift = (e_i == 0) ? -1 : 1;
+               for (int n = 0; n < n_idx; ++n) { shift *= pp1; }
+               J[offset] = i + shift;
+               AV[offset] = V(f+1, iloc, e);
+               ++offset;
+            }
+         }
+      }
+   });
+}
+
 void BatchedLORAssembly::SparseIJToCSR(OperatorHandle &A) const
 {
   const int nvdof = fes_ho.GetVSize();
@@ -372,12 +550,11 @@ void BatchedLORAssembly::SparseIJToCSR(OperatorHandle &A) const
   }

   A_mat->OverrideSize(nvdof, nvdof);
+   EnsureCapacity(A_mat->GetMemoryI(), nvdof + 1);

-   A_mat->GetMemoryI().New(nvdof+1, Device::GetDeviceMemoryType());
-   int nnz = FillI(*A_mat);
-
-   A_mat->GetMemoryJ().New(nnz, Device::GetDeviceMemoryType());
-   A_mat->GetMemoryData().New(nnz, Device::GetDeviceMemoryType());
+   const int nnz = FillI(*A_mat);
+   EnsureCapacity(A_mat->GetMemoryJ(), nnz);
+   EnsureCapacity(A_mat->GetMemoryData(), nnz);
   FillJAndData(*A_mat);
 }

@@ -431,6 +608,19 @@ void BatchedLORAssembly::AssembleWithoutBC(BilinearForm &a, OperatorHandle &A)
   // Assemble the matrix, depending on what the form is.
   // This fills in the arrays sparse_ij and sparse_mapping.
   const FiniteElementCollection *fec = fes_ho.FEColl();
+
+   // Handle DG case separately, because assembly of CSR matrix requires
+   // handling face terms.
+   if (dynamic_cast<const DG_FECollection*>(fec))
+   {
+      if (HasIntegrators<DiffusionIntegrator, MassIntegrator>(a))
+      {
+         AssemblyKernel<BatchedLOR_DG>(a);
+      }
+      SparseIJToCSR_DG(A);
+      return;
+   }
+
   if (dynamic_cast<const H1_FECollection*>(fec))
   {
      if (HasIntegrators<DiffusionIntegrator, MassIntegrator>(a))
@@ -453,10 +643,47 @@ void BatchedLORAssembly::AssembleWithoutBC(BilinearForm &a, OperatorHandle &A)
      }
   }

-   return SparseIJToCSR(A);
+   SparseIJToCSR(A);
 }

 #ifdef MFEM_USE_MPI
+void BatchedLORAssembly::ParAssemble_DG(SparseMatrix &A_local,
+                                        OperatorHandle &A)
+{
+   auto &par_fes = static_cast<ParFiniteElementSpace&>(fes_ho);
+
+   // handle the case when 'a' contains off-diagonal
+   const int lvsize = par_fes.GetVSize();
+   const Array<HYPRE_BigInt> &face_nbr_glob_ldof =
+      par_fes.GetFaceNbrGlobalDofMapArray();
+   const HYPRE_BigInt ldof_offset = par_fes.GetMyDofOffset();
+
+   const int nnz_local = A_local.NumNonZeroElems();
+   Array<HYPRE_BigInt> glob_J(nnz_local);
+
+   const HYPRE_BigInt *d_face_nbr_glob_ldof = face_nbr_glob_ldof.Read();
+   const int *d_J = A_local.ReadJ();
+   HYPRE_BigInt *d_glob_J = glob_J.Write();
+
+   mfem::forall(nnz_local, [=] MFEM_HOST_DEVICE (int i)
+   {
+      if (d_J[i] < lvsize)
+      {
+         d_glob_J[i] = d_J[i] + ldof_offset;
+      }
+      else
+      {
+         d_glob_J[i] = d_face_nbr_glob_ldof[d_J[i] - lvsize];
+      }
+   });
+
+   A.Reset(new HypreParMatrix(
+              par_fes.GetComm(), lvsize, par_fes.GlobalVSize(),
+              par_fes.GlobalVSize(), A_local.HostReadWriteI(),
+              glob_J.HostReadWrite(), A_local.HostReadWriteData(),
+              par_fes.GetDofOffsets(), par_fes.GetDofOffsets()));
+}
+
 void BatchedLORAssembly::ParAssemble(
   BilinearForm &a, const Array<int> &ess_dofs, OperatorHandle &A)
 {
@@ -464,13 +691,18 @@ void BatchedLORAssembly::ParAssemble(
   OperatorHandle A_local;
   AssembleWithoutBC(a, A_local);

-   ParBilinearForm *pa =
-      dynamic_cast<ParBilinearForm*>(&a);
-
-   pa->ParallelRAP(*A_local.As<SparseMatrix>(), A, true);
-
-   A.As<HypreParMatrix>()->EliminateBC(ess_dofs,
-                                       Operator::DiagonalPolicy::DIAG_ONE);
+   if (dynamic_cast<const DG_FECollection*>(fes_ho.FEColl()))
+   {
+      ParAssemble_DG(*A_local.As<SparseMatrix>(), A);
+   }
+   else
+   {
+      ParBilinearForm *pa =
+         dynamic_cast<ParBilinearForm*>(&a);
+      pa->ParallelRAP(*A_local.As<SparseMatrix>(), A, true);
+      A.As<HypreParMatrix>()->EliminateBC(ess_dofs,
+                                          Operator::DiagonalPolicy::DIAG_ONE);
+   }
 }
 #endif

@@ -504,12 +736,22 @@ BatchedLORAssembly::BatchedLORAssembly(FiniteElementSpace &fes_ho_)
   FormLORVertexCoordinates(fes_ho, X_vert);
 }

-IntegrationRule GetCollocatedIntRule(FiniteElementSpace &fes)
+IntegrationRule GetLobattoIntRule(Geometry::Type geom, int nd1d)
 {
   IntegrationRules irs(0, Quadrature1D::GaussLobatto);
-   const Geometry::Type geom = fes.GetMesh()->GetTypicalElementGeometry();
-   const int nd1d = fes.GetMaxElementOrder() + 1;
   return irs.Get(geom, 2*nd1d - 3);
 }

+IntegrationRule GetCollocatedIntRule(FiniteElementSpace &fes)
+{
+   const Geometry::Type geom = fes.GetMesh()->GetTypicalElementGeometry();
+   return GetLobattoIntRule(geom, fes.GetMaxElementOrder() + 1);
+}
+
+IntegrationRule GetCollocatedFaceIntRule(FiniteElementSpace &fes)
+{
+   const Geometry::Type geom = fes.GetMesh()->GetTypicalFaceGeometry();
+   return GetLobattoIntRule(geom, fes.GetMaxElementOrder() + 1);
+}
+
 } // namespace mfem
@@ -25,6 +25,7 @@ namespace mfem
 /// supported, currently:
 ///
 ///  - H1 diffusion + mass
+///  - DG diffusion + mass (in progress)
 ///  - ND curl-curl + mass
 ///  - RT div-div + mass
 ///
@@ -73,6 +74,9 @@ public:
   /// Return the vertices of the LOR mesh in E-vector format
   const Vector &GetLORVertexCoordinates() { return X_vert; }

+   /// Specialized implementation of SparseIJToCSR for DG spaces.
+   void SparseIJToCSR_DG(OperatorHandle &A) const;
+
 protected:
   /// After assembling the "sparse IJ" format, convert it to CSR.
   void SparseIJToCSR(OperatorHandle &A) const;
@@ -105,6 +109,9 @@ public:
   void FillJAndData(SparseMatrix &A) const;

 #ifdef MFEM_USE_MPI
+   /// Assemble the parallel DG matrix (with shared faces).
+   void ParAssemble_DG(SparseMatrix &A_local, OperatorHandle &A);
+
   /// Assemble the system in parallel and place the result in @a A.
   void ParAssemble(BilinearForm &a, const Array<int> &ess_dofs,
                    OperatorHandle &A);
@@ -128,9 +135,8 @@ void EnsureCapacity(Memory<T> &mem, int capacity)

 /// Return the first domain integrator in the form @a i of type @a T.
 template <typename T>
-static T *GetIntegrator(BilinearForm &a)
+static T *GetIntegrator(Array<BilinearFormIntegrator*> *integs)
 {
-   Array<BilinearFormIntegrator*> *integs = a.GetDBFI();
   if (integs != NULL)
   {
      for (auto *i : *integs)
@@ -144,8 +150,32 @@ static T *GetIntegrator(BilinearForm &a)
   return nullptr;
 }

+template <typename T>
+static T *GetIntegrator(BilinearForm &a)
+{
+   return GetIntegrator<T>(a.GetDBFI());
+}
+
+template <typename T>
+static T *GetInteriorFaceIntegrator(BilinearForm &a)
+{
+   return GetIntegrator<T>(a.GetFBFI());
+}
+
+/// @brief Return the Gauss-Lobatto rule for geometry @a geom with @a nd1d
+/// points per dimension.
+IntegrationRule GetLobattoIntRule(Geometry::Type geom, int nd1d);
+
+/// @brief Return the Gauss-Lobatto rule collocated with the element nodes.
+///
+/// Assumes @a fes uses Gauss-Lobatto basis.
 IntegrationRule GetCollocatedIntRule(FiniteElementSpace &fes);

+/// @brief Return the Gauss-Lobatto rule collocated with face nodes.
+///
+/// Assumes @a fes uses Gauss-Lobatto basis.
+IntegrationRule GetCollocatedFaceIntRule(FiniteElementSpace &fes);
+
 template <typename INTEGRATOR>
 void ProjectLORCoefficient(BilinearForm &a, CoefficientVector &coeff_vector)
 {
@@ -0,0 +1,79 @@
+// 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_LOR_DG
+#define MFEM_LOR_DG
+
+#include "lor_batched.hpp"
+
+namespace mfem
+{
+
+// BatchedLORKernel specialization for DG spaces. Not user facing. See the
+// classes BatchedLORAssembly and BatchedLORKernel .
+class BatchedLOR_DG : BatchedLORKernel
+{
+   IntegrationRule ir_face; ///< Collocated Gauss-Lobatto face quadrature rule.
+   real_t kappa; ///< DG penalty parameter.
+public:
+   template <int ORDER, int SDIM> void Assemble2D();
+   template <int ORDER> void Assemble3D();
+   BatchedLOR_DG(BilinearForm &a,
+                 FiniteElementSpace &fes_ho_,
+                 Vector &X_vert_,
+                 Vector &sparse_ij_,
+                 Array<int> &sparse_mapping_)
+      : BatchedLORKernel(fes_ho_, X_vert_, sparse_ij_, sparse_mapping_),
+        ir_face(GetLobattoIntRule(fes_ho_.GetMesh()->GetTypicalFaceGeometry(),
+                                  fes_ho_.GetMaxElementOrder() + 1))
+   {
+      ProjectLORCoefficient<MassIntegrator>(a, c1);
+      ProjectLORCoefficient<DiffusionIntegrator>(a, c2);
+
+      auto *integ = GetInteriorFaceIntegrator<DGDiffusionIntegrator>(a);
+      if (integ)
+      {
+         kappa = integ->GetPenaltyParameter();
+      }
+      else
+      {
+         kappa = 0.0;
+      }
+   }
+
+   /// @brief Compute and return the face info array.
+   ///
+   /// The face info array has shape (6, nf), where @a nf is the number of
+   /// faces. For each face @a i, the column (:,i) has entries (e0, f0, o0, e1,
+   /// f1, o1), where @a e is adjacent element, @a f is the local face index,
+   /// and @a o is the orientation. For boundary and shared faces, (e1, f1, o1)
+   /// are all set to -1.
+   Array<int> GetFaceInfo() const;
+
+   /// @brief Compute and return the boundary penalty factor.
+   ///
+   /// The returned vector has shape (nq, nf), where @a nq is the number of
+   /// nodes per face, and @a nf is the number of faces.
+   ///
+   /// The boundary penalty factor is $J_f / h = J_f^2 / J_e$ (since $h = J_e /
+   /// J_f$), where $J_f$ is the face Jacobian determinant, and $J_e$ is the
+   /// element Jacobian determinant.
+   Vector GetBdrPenaltyFactor() const;
+
+   /// Assemble the face penalty terms in the matrix @a sparse_ij.
+   void AssembleFaceTerms();
+};
+
+}
+
+#include "lor_dg_impl.hpp"
+
+#endif
@@ -0,0 +1,392 @@
+// 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.
+
+#pragma once
+
+#include "lor_dg.hpp"
+#include "../fe/face_map_utils.hpp"
+#include "../../linalg/dtensor.hpp"
+#include "../../general/forall.hpp"
+
+namespace mfem
+{
+
+Array<int> BatchedLOR_DG::GetFaceInfo() const
+{
+   Mesh &mesh = *fes_ho.GetMesh();
+   const int nf = mesh.GetNumFaces();
+   Array<int> face_info(nf * 6); // (e0, f0, o0, e1, f1, o1)
+   auto h_face_info = Reshape(face_info.HostWrite(), 6, nf);
+   for (int f = 0; f < nf; ++f)
+   {
+      auto finfo = mesh.GetFaceInformation(f);
+      h_face_info(0, f) = finfo.element[0].index;
+      h_face_info(1, f) = finfo.element[0].local_face_id;
+      h_face_info(2, f) = finfo.element[0].orientation;
+      if (finfo.IsLocal()) // Interior, non-shared face
+      {
+         h_face_info(3, f) = finfo.element[1].index;
+         h_face_info(4, f) = finfo.element[1].local_face_id;
+         h_face_info(5, f) = finfo.element[1].orientation;
+      }
+      else
+      {
+         h_face_info(3, f) = -1;
+         h_face_info(4, f) = -1;
+         h_face_info(5, f) = -1;
+      }
+   }
+   return face_info;
+}
+
+Vector BatchedLOR_DG::GetBdrPenaltyFactor() const
+{
+   Mesh &mesh = *fes_ho.GetMesh();
+
+   const int nf = mesh.GetNumFaces();
+   Array<int> f_int(mesh.GetNFbyType(FaceType::Interior));
+   Array<int> f_bdr(mesh.GetNFbyType(FaceType::Boundary));
+   {
+      int i_int = 0;
+      int i_bdr = 0;
+      for (int i = 0; i < nf; ++i)
+      {
+         const auto f = mesh.GetFaceInformation(i);
+         if (f.IsBoundary())
+         {
+            f_bdr[i_bdr] = i;
+            ++i_bdr;
+         }
+         else if (f.IsInterior())
+         {
+            f_int[i_int] = i;
+            ++i_int;
+         }
+      }
+   }
+
+   const auto geom = fes_ho.GetMesh()->GetGeometricFactors(
+                        ir, GeometricFactors::DETERMINANTS);
+
+   const int nq = ir_face.Size();
+   Vector face_Jh(nq * nf);
+   for (const FaceType ft : {FaceType::Interior, FaceType::Boundary})
+   {
+      const int nft = mesh.GetNFbyType(ft);
+      auto *geom_face = mesh.GetFaceGeometricFactors(
+                           ir_face, FaceGeometricFactors::DETERMINANTS, ft);
+
+      const L2FaceValues fv = (ft == FaceType::Interior)
+                              ? L2FaceValues::DoubleValued
+                              : L2FaceValues::SingleValued;
+      const int m = (fv == L2FaceValues::DoubleValued) ? 2 : 1;
+
+      auto *r = fes_ho.GetFaceRestriction(ElementDofOrdering::LEXICOGRAPHIC, ft, fv);
+      Vector detJ_r(nq * m * nft);
+      r->Mult(geom->detJ, detJ_r);
+
+      const auto *d_i = (ft == FaceType::Interior) ? f_int.Read() : f_bdr.Read();
+      const auto d_detJ_face = Reshape(geom_face->detJ.Read(), nq, nft);
+      const auto d_detJ_r = Reshape(detJ_r.Read(), nq, m, nft);
+      auto d_face_Jh = Reshape(face_Jh.Write(), nq, nf);
+
+      mfem::forall(nft * nq, [=] MFEM_HOST_DEVICE (int ii)
+      {
+         const int i = ii % nq;
+         const int f = ii / nq;
+         const real_t J_el = 0.5*(d_detJ_r(i, 0, f) + d_detJ_r(i, m==2?1:0, f));
+         const real_t J_f = d_detJ_face(i, f);
+         d_face_Jh(i, d_i[f]) = J_f * J_f / J_el;
+      });
+   }
+   return face_Jh;
+}
+
+void BatchedLOR_DG::AssembleFaceTerms()
+{
+   Mesh &mesh = *fes_ho.GetMesh();
+
+   const int nnz_per_row = 1 + mesh.Dimension()*2;
+   const int pp1 = fes_ho.GetMaxElementOrder() + 1;
+   const int nel_ho = mesh.GetNE();
+   const int nf = mesh.GetNumFaces();
+   const int nd_face = ir_face.Size();
+   const int nd = ir.Size();
+   const int dim = mesh.Dimension();
+
+   Array<int> face_info = GetFaceInfo();
+   const auto d_face_info = Reshape(face_info.Read(), 6, nf);
+
+   Vector face_Jh = GetBdrPenaltyFactor();
+   const auto d_face_Jh = Reshape(face_Jh.Read(), nd_face, nf);
+
+   const auto *w_face = ir_face.GetWeights().Read();
+
+   // Penalty parameter (avoid capturing *this in lambda)
+   const real_t d_kappa = kappa;
+
+   // Get diffusion coefficient
+   const bool const_dq = c2.Size() == 1;
+   const auto DQ = const_dq?Reshape(c2.Read(),1,1):Reshape(c2.Read(),nd,nel_ho);
+
+   // Sparse matrix entries
+   auto V = Reshape(sparse_ij.ReadWrite(), nnz_per_row, nd, nel_ho);
+
+   mfem::forall(nf, [=] MFEM_HOST_DEVICE (int f)
+   {
+      const int f_0 = d_face_info(1, f);
+      const int f_1 = d_face_info(4, f);
+      const int nsides = (f_1 >= 0) ? 2 : 1;
+      for (int el_i = 0; el_i < nsides; ++el_i)
+      {
+         const int e = d_face_info(3*el_i, f);
+         const int o = d_face_info(3*el_i + 2, f);
+         const int v_idx = 1 + ((el_i == 0) ? f_0 : f_1);
+         for (int i = 0; i < nd_face; ++i)
+         {
+            const int ii = internal::FaceIdxToVolIdx(dim, i, pp1, f_0, f_1, el_i, o);
+            const real_t Jh = d_face_Jh(i, f);
+            const real_t dq = const_dq ? DQ(0,0) : DQ(ii, e);
+            V(v_idx, ii, e) = -dq*d_kappa*Jh*w_face[i];
+         }
+      }
+   });
+}
+
+template <int ORDER, int SDIM>
+void BatchedLOR_DG::Assemble2D()
+{
+   MFEM_VERIFY(SDIM == 2, "Surface meshes not currently supported for LOR-DG.")
+
+   static constexpr int pp1 = ORDER + 1;
+   static constexpr int ndof_per_el = pp1*pp1;
+   static constexpr int nnz_per_row = 5;
+   const int nel_ho = fes_ho.GetNE();
+
+   // Get element geometric factors; calling before AssembleFaceTerms, since
+   // in AssembleFaceTerms, element Jacobian determinants are used, potentially
+   // saving recomputation.
+   const auto factors = GeometricFactors::DETERMINANTS |
+                        GeometricFactors::JACOBIANS;
+   const auto *geom = fes_ho.GetMesh()->GetGeometricFactors(ir, factors);
+
+   // Sparse matrix entries
+   sparse_ij.SetSize(nnz_per_row*ndof_per_el*nel_ho);
+   sparse_ij.UseDevice(true);
+   sparse_ij = 0.0;
+   auto V = Reshape(sparse_ij.ReadWrite(), nnz_per_row, pp1, pp1, nel_ho);
+
+   AssembleFaceTerms();
+
+   // Populate Gauss-Lobatto quadrature rule of size (p+1)
+   IntegrationRule ir_pp1;
+   QuadratureFunctions1D::GaussLobatto(pp1, &ir_pp1);
+   Vector glx_pp1(pp1), glw_pp1(pp1);
+   for (int i = 0; i < pp1; ++i)
+   {
+      glx_pp1[i] = ir_pp1[i].x;
+      glw_pp1[i] = ir_pp1[i].weight;
+   }
+   const auto *x_pp1 = glx_pp1.Read();
+   const auto *w_1d = glw_pp1.Read();
+
+   // Get coefficients for mass and diffusion
+   const bool const_mq = c1.Size() == 1;
+   const auto MQ = const_mq
+                   ? Reshape(c1.Read(), 1, 1, 1)
+                   : Reshape(c1.Read(), pp1, pp1, nel_ho);
+   const bool const_dq = c2.Size() == 1;
+   const auto DQ = const_dq
+                   ? Reshape(c2.Read(), 1, 1, 1)
+                   : Reshape(c2.Read(), pp1, pp1, nel_ho);
+
+   const auto detJ = Reshape(geom->detJ.Read(), pp1, pp1, nel_ho);
+   const auto J = Reshape(geom->J.Read(), pp1, pp1, 2, 2, nel_ho);
+   const auto W = Reshape(ir.GetWeights().Read(), pp1, pp1);
+
+   mfem::forall(nel_ho, [=] MFEM_HOST_DEVICE (int iel_ho)
+   {
+      for (int iy = 0; iy < pp1; ++iy)
+      {
+         for (int ix = 0; ix < pp1; ++ix)
+         {
+            const real_t mq = const_mq ? MQ(0,0,0) : MQ(ix, iy, iel_ho);
+            const real_t dq = const_dq ? DQ(0,0,0) : DQ(ix, iy, iel_ho);
+
+            for (int n_idx = 0; n_idx < 2; ++n_idx)
+            {
+               for (int e_i = 0; e_i < 2; ++e_i)
+               {
+                  const int i_0 = (n_idx == 0) ? ix + e_i : ix;
+                  const int j_0 = (n_idx == 1) ? iy + e_i : iy;
+
+                  const bool bdr = (n_idx == 0 && (i_0 == 0 || i_0 == pp1)) ||
+                                   (n_idx == 1 && (j_0 == 0 || j_0 == pp1));
+
+                  if (bdr) { continue; }
+
+                  static constexpr int lex_map[] = {4, 2, 1, 3};
+                  const int v_idx_lex = e_i + n_idx*2;
+                  const int v_idx = lex_map[v_idx_lex];
+
+                  const int w_idx = (n_idx == 0) ? iy : ix;
+                  const int x_idx = (n_idx == 0) ? i_0 : j_0;
+
+                  const real_t J1 = J(ix, iy, n_idx, !n_idx, iel_ho);
+                  const real_t J2 = J(ix, iy, !n_idx, !n_idx, iel_ho);
+                  const real_t Jh = (J1*J1 + J2*J2) / detJ(ix, iy, iel_ho);
+
+                  V(v_idx, ix, iy, iel_ho) =
+                     -dq * Jh * w_1d[w_idx] / (x_pp1[x_idx] - x_pp1[x_idx -1]);
+               }
+            }
+            V(0, ix, iy, iel_ho) = mq * detJ(ix, iy, iel_ho) * W(ix, iy);
+            for (int i = 1; i < nnz_per_row; ++i)
+            {
+               V(0, ix, iy, iel_ho) -= V(i, ix, iy, iel_ho);
+            }
+         }
+      }
+   });
+}
+
+template <int ORDER>
+void BatchedLOR_DG::Assemble3D()
+{
+   static constexpr int pp1 = ORDER + 1;
+   static constexpr int ndof_per_el = pp1*pp1*pp1;
+   static constexpr int nnz_per_row = 7;
+   const int nel_ho = fes_ho.GetNE();
+
+   // Get element geometric factors; calling before AssembleFaceTerms, since
+   // in AssembleFaceTerms, element Jacobian determinants are used, potentially
+   // saving recomputation.
+   const auto factors = GeometricFactors::DETERMINANTS |
+                        GeometricFactors::JACOBIANS;
+   const auto geom = fes_ho.GetMesh()->GetGeometricFactors(ir, factors);
+
+   sparse_ij.SetSize(nnz_per_row*ndof_per_el*nel_ho);
+   sparse_ij.UseDevice(true);
+   sparse_ij = 0.0;
+   auto V = Reshape(sparse_ij.Write(), nnz_per_row, pp1, pp1, pp1, nel_ho);
+
+   AssembleFaceTerms();
+
+   // Populate Gauss-Lobatto quadrature rule of size (p+1)
+   IntegrationRule ir_pp1;
+   QuadratureFunctions1D::GaussLobatto(pp1, &ir_pp1);
+   Vector glx_pp1(pp1), glw_pp1(pp1);
+   for (int i = 0; i < pp1; ++i)
+   {
+      glx_pp1[i] = ir_pp1[i].x;
+      glw_pp1[i] = ir_pp1[i].weight;
+   }
+   const auto *x_pp1 = glx_pp1.Read();
+   const auto *w_1d = glw_pp1.Read();
+
+   const bool const_mq = c1.Size() == 1;
+   const auto MQ = const_mq
+                   ? Reshape(c1.Read(), 1, 1, 1, 1)
+                   : Reshape(c1.Read(), pp1, pp1, pp1, nel_ho);
+   const bool const_dq = c2.Size() == 1;
+   const auto DQ = const_dq
+                   ? Reshape(c2.Read(), 1, 1, 1, 1)
+                   : Reshape(c2.Read(), pp1, pp1, pp1, nel_ho);
+   const auto W = Reshape(ir.GetWeights().Read(), pp1, pp1, pp1);
+
+   const auto detJ = Reshape(geom->detJ.Read(), pp1, pp1, pp1, nel_ho);
+   const auto J = Reshape(geom->J.Read(), pp1, pp1, pp1, 3, 3, nel_ho);
+
+   mfem::forall(nel_ho, [=] MFEM_HOST_DEVICE (int iel_ho)
+   {
+      for (int iz = 0; iz < pp1; ++iz)
+      {
+         for (int iy = 0; iy < pp1; ++iy)
+         {
+            for (int ix = 0; ix < pp1; ++ix)
+            {
+               const real_t mq = const_mq ? MQ(0,0,0,0) : MQ(ix, iy, iz, iel_ho);
+               const real_t dq = const_dq ? DQ(0,0,0,0) : DQ(ix, iy, iz, iel_ho);
+
+               const real_t DETJ = detJ(ix, iy, iz, iel_ho);
+
+               for (int n_idx = 0; n_idx < 3; ++n_idx)
+               {
+                  for (int e_i = 0; e_i < 2; ++e_i)
+                  {
+                     static constexpr int lex_map[] = {5,3,2,4,1,6};
+                     const int v_idx_lex = e_i + n_idx*2;
+                     const int v_idx = lex_map[v_idx_lex];
+
+                     const int i_0 = (n_idx == 0) ? ix + e_i : ix;
+                     const int j_0 = (n_idx == 1) ? iy + e_i : iy;
+                     const int k_0 = (n_idx == 2) ? iz + e_i : iz;
+
+                     const bool bdr =
+                        (n_idx == 0 && (i_0 == 0 || i_0 == pp1)) ||
+                        (n_idx == 1 && (j_0 == 0 || j_0 == pp1)) ||
+                        (n_idx == 2 && (k_0 == 0 || k_0 == pp1));
+
+                     if (bdr) { continue; }
+
+                     int x_idx = (n_idx == 0) ? i_0 : (n_idx == 1) ? j_0 : k_0;
+                     int w_idx_1 = (n_idx == 0) ? iy : (n_idx == 1) ? iz : ix;
+                     int w_idx_2 = (n_idx == 0) ? iz : (n_idx == 1) ? ix : iy;
+
+                     const real_t J00 = J(ix, iy, iz, 0, 0, iel_ho);
+                     const real_t J01 = J(ix, iy, iz, 0, 1, iel_ho);
+                     const real_t J02 = J(ix, iy, iz, 0, 2, iel_ho);
+                     const real_t J10 = J(ix, iy, iz, 1, 0, iel_ho);
+                     const real_t J11 = J(ix, iy, iz, 1, 1, iel_ho);
+                     const real_t J12 = J(ix, iy, iz, 1, 2, iel_ho);
+                     const real_t J20 = J(ix, iy, iz, 2, 0, iel_ho);
+                     const real_t J21 = J(ix, iy, iz, 2, 1, iel_ho);
+                     const real_t J22 = J(ix, iy, iz, 2, 2, iel_ho);
+
+                     real_t JinvJinvT_diag = 0.0;
+                     if (n_idx == 0)
+                     {
+                        JinvJinvT_diag = J02*J02*(J11*J11 + J21*J21) + (J12*J21 - J11*J22)*
+                                         (J12*J21 - J11*J22) - 2*J01*J02*(J11*J12 + J21*J22) + J01*J01*
+                                         (J12*J12 + J22*J22);
+                     }
+                     else if (n_idx == 1)
+                     {
+                        JinvJinvT_diag = J02*J02*(J10*J10 + J20*J20) + (J12*J20 - J10*J22)*
+                                         (J12*J20 - J10*J22) - 2*J00*J02*(J10*J12 + J20*J22) + J00*J00*
+                                         (J12*J12 + J22*J22);
+                     }
+                     else if (n_idx == 2)
+                     {
+                        JinvJinvT_diag = J01*J01*(J10*J10 + J20*J20) + (J11*J20 - J10*J21)*
+                                         (J11*J20 - J10*J21) - 2*J00*J01*(J10*J11 + J20*J21) + J00*J00*
+                                         (J11*J11 + J21*J21);
+                     }
+
+                     const real_t Jh = JinvJinvT_diag / DETJ;
+
+                     V(v_idx, ix, iy, iz, iel_ho) = -dq * Jh * w_1d[w_idx_1] * w_1d[w_idx_2] /
+                                                    (x_pp1[x_idx] - x_pp1[x_idx -1]);
+                  }
+               }
+               V(0, ix, iy, iz, iel_ho) = mq * DETJ * W(ix, iy, iz);
+               for (int i = 1; i < 7; ++i)
+               {
+                  V(0, ix, iy, iz, iel_ho) -= V(i, ix, iy, iz, iel_ho);
+               }
+            }
+         }
+      }
+
+   });
+}
+
+} // namespace mfem
@@ -436,7 +436,7 @@ void NonlinearForm::Mult(const Vector &x, Vector &y) const
   // In parallel, the result is in 'py' which is an alias for 'aux2'.
 }

-Operator &NonlinearForm::GetGradient(const Vector &x) const
+Operator &NonlinearForm::GetGradient(const Vector &x, bool finalize) const
 {
   if (ext)
   {
@@ -644,6 +644,8 @@ Operator &NonlinearForm::GetGradient(const Vector &x) const
      }
   }

+   if (!finalize) { return *Grad; }
+
   if (!Grad->Finalized())
   {
      Grad->Finalize(skip_zeros);
@@ -788,12 +790,10 @@ BlockNonlinearForm::BlockNonlinearForm(Array<FiniteElementSpace *> &f) :
 }

 void BlockNonlinearForm::SetEssentialBC(
-   const Array<Array<int> *> &bdr_attr_is_ess, Array<Vector *> &rhs)
+   const Array<Array<int>*> &bdr_attr_is_ess, Array<Vector*> &rhs)
 {
   for (int s = 0; s < fes.Size(); ++s)
   {
-      ess_tdofs[s]->SetSize(ess_tdofs.Size());
-
      fes[s]->GetEssentialTrueDofs(*bdr_attr_is_ess[s], *ess_tdofs[s]);

      if (rhs[s])
@@ -803,6 +803,19 @@ void BlockNonlinearForm::SetEssentialBC(
   }
 }

+void BlockNonlinearForm::SetEssentialTrueDofs(
+   const Array<Array<int>*> &ess_tdof_list, Array<Vector*> &rhs)
+{
+   for (int s = 0; s < fes.Size(); ++s)
+   {
+      *ess_tdofs[s] = *ess_tdof_list[s];
+      if (rhs[s])
+      {
+         rhs[s]->SetSubVector(*ess_tdofs[s], 0.0);
+      }
+   }
+}
+
 real_t BlockNonlinearForm::GetEnergyBlocked(const BlockVector &bx) const
 {
   Array<Array<int> *> vdofs(fes.Size());
@@ -1192,7 +1205,14 @@ const BlockVector &BlockNonlinearForm::Prolongate(const BlockVector &bx) const
      aux1.Update(block_offsets);
      for (int s = 0; s < fes.Size(); s++)
      {
-         P[s]->Mult(bx.GetBlock(s), aux1.GetBlock(s));
+         if (P[s])
+         {
+            P[s]->Mult(bx.GetBlock(s), aux1.GetBlock(s));
+         }
+         else
+         {
+            aux1.GetBlock(s) = bx.GetBlock(s);
+         }
      }
      return aux1;
   }
@@ -1221,11 +1241,16 @@ void BlockNonlinearForm::Mult(const Vector &x, Vector &y) const
      {
         cP[s]->MultTranspose(pby.GetBlock(s), by.GetBlock(s));
      }
+      else if (needs_prolongation)
+      {
+         by.GetBlock(s) = pby.GetBlock(s);
+      }
      by.GetBlock(s).SetSubVector(*ess_tdofs[s], 0.0);
   }
 }

-void BlockNonlinearForm::ComputeGradientBlocked(const BlockVector &bx) const
+void BlockNonlinearForm::ComputeGradientBlocked(const BlockVector &bx,
+                                                bool finalize) const
 {
   const int skip_zeros = 0;
   Array<Array<int> *> vdofs(fes.Size());
@@ -1479,7 +1504,7 @@ void BlockNonlinearForm::ComputeGradientBlocked(const BlockVector &bx) const
      }
   }

-   if (!Grads(0,0)->Finalized())
+   if (finalize && !Grads(0,0)->Finalized())
   {
      for (int i=0; i<fes.Size(); ++i)
      {
@@ -1518,7 +1543,23 @@ Operator &BlockNonlinearForm::GetGradient(const Vector &x) const
         for (int s2 = 0; s2 < fes.Size(); ++s2)
         {
            delete cGrads(s1, s2);
-            cGrads(s1, s2) = RAP(*cP[s1], *Grads(s1, s2), *cP[s2]);
+            if (cP[s1] && cP[s2])
+            {
+               cGrads(s1, s2) = RAP(*cP[s1], *Grads(s1, s2), *cP[s2]);
+            }
+            else if (cP[s1])
+            {
+               cGrads(s1, s2) = TransposeMult(*cP[s1], *Grads(s1, s2));
+            }
+            else if (cP[s2])
+            {
+               cGrads(s1, s2) = mfem::Mult(*Grads(s1, s2), *cP[s2]);
+            }
+            else
+            {
+               cGrads(s1, s2) = NULL;
+               continue;
+            }
            mGrads(s1, s2) = cGrads(s1, s2);
         }
      }
@@ -217,7 +217,12 @@ public:
       In general, @a x may have non-homogeneous essential boundary values.

       The state @a x must be a true-dof vector. */
-   Operator &GetGradient(const Vector &x) const override;
+   Operator &GetGradient(const Vector &x) const override { return GetGradient(x, true); }
+
+   /** @brief Compute the gradient Operator of the NonlinearForm corresponding
+       to the state @a x with optional finalization and elimintaion. */
+   /** @see GetGradient(const Vector &) */
+   Operator &GetGradient(const Vector &x, bool finalize) const;

   /// Update the NonlinearForm to propagate updates of the associated FE space.
   /** After calling this method, the essential boundary conditions need to be
@@ -308,7 +313,7 @@ protected:
   void MultBlocked(const BlockVector &bx, BlockVector &by) const;

   /// Specialized version of GetGradient() for BlockVector
-   void ComputeGradientBlocked(const BlockVector &bx) const;
+   void ComputeGradientBlocked(const BlockVector &bx, bool finalize = true) const;

 public:
   /// Construct an empty BlockNonlinearForm. Initialize with SetSpaces().
@@ -363,8 +368,39 @@ public:
                             Array<int> &bdr_marker)
   { bfnfi.Append(nlfi); bfnfi_marker.Append(&bdr_marker); }

-   virtual void SetEssentialBC(const Array<Array<int> *>&bdr_attr_is_ess,
-                               Array<Vector *> &rhs);
+   /** @brief Set essential boundary conditions to each finite element space
+       using boundary attribute markers.
+
+       This method calls `FiniteElementSpace::GetEssentialTrueDofs()` for each
+       space and stores ess_tdof_lists internally.
+
+       If `rhs` vectors are non-null, the entries corresponding to these
+       essential DoFs are set to zero. This ensures compatibility with the
+       output of the `Mult()` method, which also zeroes out these entries.
+
+       @param[in] bdr_attr_is_ess A list of boundary attribute markers for each
+       space.
+       @param[in,out] rhs         An array of optional right-hand side vectors.
+       If a vector at `rhs[i]` is non-null, its essential DoFs will be set
+       to zero. */
+   virtual void SetEssentialBC(const Array<Array<int>*> &bdr_attr_is_ess,
+                               Array<Vector*> &rhs);
+
+   /** @brief Set essential boundary conditions to each finite element space
+       using essential true dof lists.
+
+       This method stores a copy of the provided essential true dof lists.
+
+       If `rhs` vectors are non-null, the entries corresponding to these
+       essential DoFs are set to zero. This ensures compatibility with the
+       output of the `Mult()` method, which also zeroes out these entries.
+
+       @param[in] ess_tdof_list A list of essential true dofs for each space.
+       @param[in,out] rhs       An array of optional right-hand side vectors.
+       If a vector at `rhs[i]` is non-null, its essential DoFs will be set
+       to zero. */
+   virtual void SetEssentialTrueDofs(const Array<Array<int>*> &ess_tdof_list,
+                                     Array<Vector*> &rhs);

   virtual real_t GetEnergy(const Vector &x) const;

@@ -151,6 +151,15 @@ void ParBilinearForm::ParallelRAP(SparseMatrix &loc_A, OperatorHandle &A,
   }
 }

+HypreParMatrix *ParBilinearForm::ParallelAssembleInternalMatrix()
+{
+   if (p_mat.Ptr() == NULL)
+   {
+      ParallelAssemble(p_mat, mat);
+   }
+   return p_mat.As<HypreParMatrix>();
+}
+
 void ParBilinearForm::ParallelAssemble(OperatorHandle &A, SparseMatrix *A_local)
 {
   A.Clear();
@@ -333,6 +342,15 @@ void ParBilinearForm
   A.EliminateRowsCols(dof_list, X, B);
 }

+void ParBilinearForm::ParallelEliminateEssentialBC(
+   const Array<int> &bdr_attr_is_ess, const HypreParVector &X, HypreParVector &B)
+{
+   Array<int> dof_list;
+   pfes->GetEssentialTrueDofs(bdr_attr_is_ess, dof_list);
+
+   p_mat.As<HypreParMatrix>()->EliminateRowsCols(dof_list, X, B);
+}
+
 HypreParMatrix *ParBilinearForm::
 ParallelEliminateEssentialBC(const Array<int> &bdr_attr_is_ess,
                             HypreParMatrix &A) const
@@ -344,6 +362,26 @@ ParallelEliminateEssentialBC(const Array<int> &bdr_attr_is_ess,
   return A.EliminateRowsCols(dof_list);
 }

+void ParBilinearForm::ParallelEliminateEssentialBC(const Array<int>
+                                                   &bdr_attr_is_ess)
+{
+   Array<int> tdofs_list;
+   pfes->GetEssentialTrueDofs(bdr_attr_is_ess, tdofs_list);
+
+   ParallelEliminateTDofs(tdofs_list);
+}
+
+void ParBilinearForm::ParallelEliminateTDofs(const Array<int> &tdofs_list)
+{
+   p_mat_e.EliminateRowsCols(p_mat, tdofs_list);
+}
+
+void ParBilinearForm::ParallelEliminateTDofsInRHS(
+   const Array<int> &tdofs_list, const Vector &x, Vector &b)
+{
+   p_mat.EliminateBC(p_mat_e, tdofs_list, x, b);
+}
+
 void ParBilinearForm::TrueAddMult(const Vector &x, Vector &y, const real_t a)
 const
 {
@@ -485,7 +523,7 @@ void ParBilinearForm::FormLinearSystem(
      HypreParVector true_X(pfes), true_B(pfes);
      P.MultTranspose(b, true_B);
      R.Mult(x, true_X);
-      p_mat.EliminateBC(p_mat_e, ess_tdof_list, true_X, true_B);
+      ParallelEliminateTDofsInRHS(ess_tdof_list, true_X, true_B);
      R.MultTranspose(true_B, b);
      hybridization->ReduceRHS(true_B, B);
      X.SetSize(B.Size());
@@ -498,17 +536,11 @@ void ParBilinearForm::FormLinearSystem(
      B.SetSize(X.Size());
      P.MultTranspose(b, B);
      R.Mult(x, X);
-      p_mat.EliminateBC(p_mat_e, ess_tdof_list, X, B);
+      ParallelEliminateTDofsInRHS(ess_tdof_list, X, B);
      if (!copy_interior) { X.SetSubVectorComplement(ess_tdof_list, 0.0); }
   }
 }

-void ParBilinearForm::EliminateVDofsInRHS(
-   const Array<int> &vdofs, const Vector &x, Vector &b)
-{
-   p_mat.EliminateBC(p_mat_e, vdofs, x, b);
-}
-
 void ParBilinearForm::FormSystemMatrix(const Array<int> &ess_tdof_list,
                                       OperatorHandle &A)
 {
@@ -553,7 +585,7 @@ void ParBilinearForm::FormSystemMatrix(const Array<int> &ess_tdof_list,
         mat = NULL;
         delete mat_e;
         mat_e = NULL;
-         p_mat_e.EliminateRowsCols(p_mat, ess_tdof_list);
+         ParallelEliminateTDofs(ess_tdof_list);
      }
      if (hybridization)
      {
@@ -615,36 +647,180 @@ void ParBilinearForm::Update(FiniteElementSpace *nfes)
   p_mat_e.Clear();
 }

-
-HypreParMatrix *ParMixedBilinearForm::ParallelAssemble()
+void ParMixedBilinearForm::pAllocMat()
 {
-   // construct the block-diagonal matrix A
-   HypreParMatrix *A =
-      new HypreParMatrix(trial_pfes->GetComm(),
-                         test_pfes->GlobalVSize(),
-                         trial_pfes->GlobalVSize(),
-                         test_pfes->GetDofOffsets(),
-                         trial_pfes->GetDofOffsets(),
-                         mat);
+   const int trial_nbr_size = trial_pfes->GetFaceNbrVSize();
+   const int test_nbr_size = test_pfes->GetFaceNbrVSize();

-   HypreParMatrix *rap = RAP(test_pfes->Dof_TrueDof_Matrix(), A,
-                             trial_pfes->Dof_TrueDof_Matrix());
-
-   delete A;
-
-   return rap;
+   if (keep_nbr_block)
+   {
+      mat = new SparseMatrix(height + test_nbr_size, width + trial_nbr_size);
+   }
+   else
+   {
+      mat = new SparseMatrix(height, width + trial_nbr_size);
+   }
 }

-void ParMixedBilinearForm::ParallelAssemble(OperatorHandle &A)
+void ParMixedBilinearForm::AssembleSharedFaces(int skip_zeros)
 {
-   // construct the rectangular block-diagonal matrix dA
-   OperatorHandle dA(A.Type());
-   dA.MakeRectangularBlockDiag(trial_pfes->GetComm(),
-                               test_pfes->GlobalVSize(),
-                               trial_pfes->GlobalVSize(),
-                               test_pfes->GetDofOffsets(),
-                               trial_pfes->GetDofOffsets(),
-                               mat);
+   ParMesh *pmesh = trial_pfes->GetParMesh();
+   FaceElementTransformations *T;
+   Array<int> tr_vdofs1, tr_vdofs2, tr_vdofs_all;
+   Array<int> te_vdofs1, te_vdofs2, te_vdofs_all;
+   DenseMatrix elemmat;
+
+   int nfaces = pmesh->GetNSharedFaces();
+   for (int i = 0; i < nfaces; i++)
+   {
+      T = pmesh->GetSharedFaceTransformations(i);
+      int Elem2NbrNo = T->Elem2No - pmesh->GetNE();
+      trial_pfes->GetElementVDofs(T->Elem1No, tr_vdofs1);
+      test_pfes->GetElementVDofs(T->Elem1No, te_vdofs1);
+      trial_pfes->GetFaceNbrElementVDofs(Elem2NbrNo, tr_vdofs2);
+      test_pfes->GetFaceNbrElementVDofs(Elem2NbrNo, te_vdofs2);
+
+      tr_vdofs1.Copy(tr_vdofs_all);
+      for (int j = 0; j < tr_vdofs2.Size(); j++)
+      {
+         if (tr_vdofs2[j] >= 0)
+         {
+            tr_vdofs2[j] += width;
+         }
+         else
+         {
+            tr_vdofs2[j] -= width;
+         }
+      }
+      tr_vdofs_all.Append(tr_vdofs2);
+
+      if (keep_nbr_block)
+      {
+         te_vdofs1.Copy(te_vdofs_all);
+         for (int j = 0; j < te_vdofs2.Size(); j++)
+         {
+            if (te_vdofs2[j] >= 0)
+            {
+               te_vdofs2[j] += height;
+            }
+            else
+            {
+               te_vdofs2[j] -= height;
+            }
+         }
+         te_vdofs_all.Append(te_vdofs2);
+      }
+
+      for (int k = 0; k < interior_face_integs.Size(); k++)
+      {
+         interior_face_integs[k]->
+         AssembleFaceMatrix(*trial_pfes->GetFE(T->Elem1No),
+                            *test_pfes->GetFE(T->Elem1No),
+                            *trial_pfes->GetFaceNbrFE(Elem2NbrNo),
+                            *test_pfes->GetFaceNbrFE(Elem2NbrNo),
+                            *T, elemmat);
+         if (keep_nbr_block)
+         {
+            mat->AddSubMatrix(te_vdofs_all, tr_vdofs_all, elemmat, skip_zeros);
+         }
+         else
+         {
+            mat->AddSubMatrix(te_vdofs1, tr_vdofs_all, elemmat, skip_zeros);
+         }
+      }
+   }
+}
+
+void ParMixedBilinearForm::Assemble(int skip_zeros)
+{
+   if (interior_face_integs.Size())
+   {
+      trial_pfes->ExchangeFaceNbrData();
+      test_pfes->ExchangeFaceNbrData();
+      if (!ext && mat == NULL)
+      {
+         pAllocMat();
+      }
+   }
+
+   MixedBilinearForm::Assemble(skip_zeros);
+
+   if (!ext && interior_face_integs.Size() > 0)
+   {
+      AssembleSharedFaces(skip_zeros);
+   }
+}
+
+HypreParMatrix *ParMixedBilinearForm::ParallelAssembleInternalMatrix()
+{
+   if (p_mat.Ptr() == NULL)
+   {
+      ParallelAssemble(p_mat, mat);
+   }
+   return p_mat.As<HypreParMatrix>();
+}
+
+HypreParMatrix *ParMixedBilinearForm::ParallelAssemble(SparseMatrix *m)
+{
+   OperatorHandle Mh(Operator::Hypre_ParCSR);
+   ParallelAssemble(Mh, m);
+   Mh.SetOperatorOwner(false);
+   return Mh.As<HypreParMatrix>();
+}
+
+void ParMixedBilinearForm::ParallelAssemble(OperatorHandle &A,
+                                            SparseMatrix *A_local)
+{
+   A.Clear();
+
+   if (A_local == NULL) { return; }
+   MFEM_VERIFY(A_local->Finalized(), "the local matrix must be finalized");
+
+   OperatorHandle dA(A.Type()), hdA;
+
+   if (interior_face_integs.Size() == 0)
+   {
+      // construct the rectangular block-diagonal matrix dA
+      dA.MakeRectangularBlockDiag(trial_pfes->GetComm(),
+                                  test_pfes->GlobalVSize(),
+                                  trial_pfes->GlobalVSize(),
+                                  test_pfes->GetDofOffsets(),
+                                  trial_pfes->GetDofOffsets(),
+                                  A_local);
+   }
+   else
+   {
+      // handle the case when 'a' contains off-diagonal
+      const int lvrows = test_pfes->GetVSize();
+      const int lvcols = trial_pfes->GetVSize();
+      const HYPRE_BigInt *face_nbr_glob_lcol = trial_pfes->GetFaceNbrGlobalDofMap();
+      const HYPRE_BigInt lcol_offset = trial_pfes->GetMyDofOffset();
+
+      Array<HYPRE_BigInt> glob_J(A_local->NumNonZeroElems());
+      const int *J = A_local->GetJ();
+      for (int i = 0; i < glob_J.Size(); i++)
+      {
+         if (J[i] < lvcols)
+         {
+            glob_J[i] = J[i] + lcol_offset;
+         }
+         else
+         {
+            glob_J[i] = face_nbr_glob_lcol[J[i] - lvcols];
+         }
+      }
+
+      // TODO - construct dA directly in the A format
+      hdA.Reset(
+         new HypreParMatrix(trial_pfes->GetComm(), lvrows, test_pfes->GlobalVSize(),
+                            trial_pfes->GlobalVSize(), A_local->GetI(), glob_J,
+                            A_local->GetData(), test_pfes->GetDofOffsets(),
+                            trial_pfes->GetDofOffsets()));
+      // - hdA owns the new HypreParMatrix
+      // - the above constructor copies all input arrays
+      glob_J.DeleteAll();
+      dA.ConvertFrom(hdA);
+   }

   OperatorHandle P_test(A.Type()), P_trial(A.Type());

@@ -670,6 +846,44 @@ void ParMixedBilinearForm::TrueAddMult(const Vector &x, Vector &y,
   test_pfes->Dof_TrueDof_Matrix()->MultTranspose(a, Yaux, 1.0, y);
 }

+void ParMixedBilinearForm::ParallelEliminateTrialEssentialBC(
+   const Array<int> &bdr_attr_is_ess)
+{
+   Array<int> trial_tdof_list;
+   trial_pfes->GetEssentialTrueDofs(bdr_attr_is_ess, trial_tdof_list);
+
+   ParallelEliminateTrialTDofs(trial_tdof_list);
+}
+
+void ParMixedBilinearForm::ParallelEliminateTrialTDofs(
+   const Array<int> &trial_tdof_list)
+{
+   HypreParMatrix *temp = p_mat.As<HypreParMatrix>()->EliminateCols(
+                             trial_tdof_list);
+   p_mat_e.Reset(temp, true);
+}
+
+void ParMixedBilinearForm::ParallelEliminateTrialTDofsInRHS(
+   const Array<int> &trial_tdof_list, const Vector &x, Vector &b)
+{
+   p_mat_e.As<HypreParMatrix>()->Mult(-1.0, x, 1.0, b);
+}
+
+void ParMixedBilinearForm::ParallelEliminateTestEssentialBC(
+   const Array<int> &bdr_attr_is_ess)
+{
+   Array<int> test_tdof_list;
+   test_pfes->GetEssentialTrueDofs(bdr_attr_is_ess, test_tdof_list);
+
+   ParallelEliminateTestTDofs(test_tdof_list);
+}
+
+void ParMixedBilinearForm::ParallelEliminateTestTDofs(
+   const Array<int> &test_tdof_list)
+{
+   p_mat.As<HypreParMatrix>()->EliminateRows(test_tdof_list);
+}
+
 void ParMixedBilinearForm::FormRectangularSystemMatrix(
   const Array<int>
   &trial_tdof_list,
@@ -690,10 +904,8 @@ void ParMixedBilinearForm::FormRectangularSystemMatrix(
      mat = NULL;
      delete mat_e;
      mat_e = NULL;
-      HypreParMatrix *temp =
-         p_mat.As<HypreParMatrix>()->EliminateCols(trial_tdof_list);
-      p_mat.As<HypreParMatrix>()->EliminateRows(test_tdof_list);
-      p_mat_e.Reset(temp, true);
+      ParallelEliminateTrialTDofs(trial_tdof_list);
+      ParallelEliminateTestTDofs(test_tdof_list);
   }

   A = p_mat;
@@ -723,7 +935,7 @@ void ParMixedBilinearForm::FormRectangularLinearSystem(
   test_P->MultTranspose(b, B);
   trial_R->Mult(x, X);

-   p_mat_e.As<HypreParMatrix>()->Mult(-1.0, X, 1.0, B);
+   ParallelEliminateTrialTDofsInRHS(trial_tdof_list, X, B);
   B.SetSubVector(test_tdof_list, 0.0);
 }

@@ -73,7 +73,7 @@ public:
   /** When set to true and the ParBilinearForm has interior face integrators,
       the local SparseMatrix will include the rows (in addition to the columns)
       corresponding to face-neighbor dofs. The default behavior is to disregard
-       those rows. Must be called before the first Assemble call. */
+       those rows. Must be called before the first Assemble() call. */
   void KeepNbrBlock(bool knb = true) { keep_nbr_block = knb; }

   /** @brief Set the operator type id for the parallel matrix/operator when
@@ -101,6 +101,14 @@ public:
       diagonal for this case. */
   void AssembleDiagonal(Vector &diag) const override;

+   /// Returns the matrix assembled on the true dofs, i.e. P^t A P.
+   /** The returned matrix is the internal one, owned by the form. It is not
+       reassembled if it has been already constructed. If FormSystemMatrix()
+       has been called before, it is the system matrix with eliminated
+       essential DOFs, otherwise the parallel matrix is assembled here without
+       the elimination process. */
+   HypreParMatrix *ParallelAssembleInternalMatrix();
+
   /// Returns the matrix assembled on the true dofs, i.e. P^t A P.
   /** The returned matrix has to be deleted by the caller. */
   HypreParMatrix *ParallelAssemble() { return ParallelAssemble(mat); }
@@ -146,6 +154,13 @@ public:
                                     const HypreParVector &X,
                                     HypreParVector &B) const;

+   /// Eliminate essential boundary DOFs from the parallel system matrix.
+   /** The array @a bdr_attr_is_ess marks boundary attributes that constitute
+       the essential part of the boundary. */
+   void ParallelEliminateEssentialBC(const Array<int> &bdr_attr_is_ess,
+                                     const HypreParVector &X,
+                                     HypreParVector &B);
+
   /// Eliminate essential boundary DOFs from a parallel assembled matrix @a A.
   /** The array @a bdr_attr_is_ess marks boundary attributes that constitute
       the essential part of the boundary. The eliminated part is stored in a
@@ -157,6 +172,12 @@ public:
   HypreParMatrix *ParallelEliminateEssentialBC(const Array<int> &bdr_attr_is_ess,
                                                HypreParMatrix &A) const;

+   /// Eliminate essential boundary DOFs from the parallel system matrix.
+   /** The array @a bdr_attr_is_ess marks boundary attributes that constitute
+       the essential part of the boundary. This method relies on
+       ParallelEliminateTDofs(const Array<int> &), see it for details. */
+   void ParallelEliminateEssentialBC(const Array<int> &bdr_attr_is_ess);
+
   /// Eliminate essential true DOFs from a parallel assembled matrix @a A.
   /** Given a list of essential true dofs and the parallel assembled matrix
       @a A, eliminate the true dofs from the matrix, storing the eliminated
@@ -169,6 +190,28 @@ public:
                                          HypreParMatrix &A) const
   { return A.EliminateRowsCols(tdofs_list); }

+   /// Eliminate essential true DOFs from the parallel system matrix.
+   /** Given a list of essential true dofs, eliminate the true dofs from
+       the parallel assembled system matrix, storing the eliminated part
+       internally. This method works in conjunction with
+       ParallelEliminateTDofsInRHS() and allows elimination of boundary
+       conditions in multiple right-hand sides. */
+   void ParallelEliminateTDofs(const Array<int> &tdofs_list);
+
+   /** @brief Use the stored eliminated part of the parallel system matrix for
+       elimination of boundary conditions in the r.h.s. */
+   /** Given a list of essential true dofs, eliminate the true dofs from the
+       right-hand side @a b using the solution vector @a x and the previously
+       stored eliminated part of the parallel assembled system matrix produced
+       by ParallelEliminateTDofs(const Array<int> &). */
+   void ParallelEliminateTDofsInRHS(const Array<int> &tdofs, const Vector &x,
+                                    Vector &b);
+
+   /// @deprecated Use ParallelEliminateTDofsInRHS() instead.
+   MFEM_DEPRECATED void EliminateVDofsInRHS(const Array<int> &vdofs,
+                                            const Vector &x, Vector &b)
+   { ParallelEliminateTDofsInRHS(vdofs, x, b); }
+
   /** @brief Compute @a y += @a a (P^t A P) @a x, where @a x and @a y are
       vectors on the true dofs. */
   void TrueAddMult(const Vector &x, Vector &y, const real_t a = 1.0) const;
@@ -238,8 +281,6 @@ public:

   void Update(FiniteElementSpace *nfes = NULL) override;

-   void EliminateVDofsInRHS(const Array<int> &vdofs, const Vector &x, Vector &b);
-
   virtual ~ParBilinearForm() { }
 };

@@ -257,6 +298,13 @@ protected:
   /// Matrix and eliminated matrix
   OperatorHandle p_mat, p_mat_e;

+   bool keep_nbr_block;
+
+   // Allocate mat - called when (mat == NULL && fbfi.Size() > 0)
+   void pAllocMat();
+
+   void AssembleSharedFaces(int skip_zeros = 1);
+
 private:
   /// Copy construction is not supported; body is undefined.
   ParMixedBilinearForm(const ParMixedBilinearForm &);
@@ -276,6 +324,7 @@ public:
   {
      trial_pfes = trial_fes;
      test_pfes  = test_fes;
+      keep_nbr_block = false;
   }

   /** @brief Create a ParMixedBilinearForm on the given FiniteElementSpace%s
@@ -295,15 +344,89 @@ public:
   {
      trial_pfes = trial_fes;
      test_pfes  = test_fes;
+      keep_nbr_block = false;
   }

+   /** When set to true and the ParMixedBilinearForm has interior face
+       integrators, the local SparseMatrix will include the rows (in addition
+       to the columns) corresponding to face-neighbor dofs. The default
+       behavior is to disregard those rows. Must be called before the first
+       Assemble() call. */
+   void KeepNbrBlock(bool knb = true) { keep_nbr_block = knb; }
+
+   /// Assemble the local matrix
+   void Assemble(int skip_zeros = 1);
+
   /// Returns the matrix assembled on the true dofs, i.e. P_test^t A P_trial.
-   HypreParMatrix *ParallelAssemble();
+   /** The returned matrix is the internal one, owned by the form. It is not
+       reassembled if it has been already constructed. If
+       FormRectangularSystemMatrix() has been called before, it is the system
+       matrix with eliminated essential DOFs, otherwise the parallel matrix is
+       assembled here without the elimination process. */
+   HypreParMatrix *ParallelAssembleInternalMatrix();
+
+   /// Returns the matrix assembled on the true dofs, i.e. P_test^t A P_trial.
+   /** The returned matrix has to be deleted by the caller. */
+   HypreParMatrix *ParallelAssemble() { return ParallelAssemble(mat); }
+
+   /** @brief Returns the eliminated matrix assembled on the true dofs, i.e.
+       P_test^t A_local P_trial. */
+   /** The returned matrix has to be deleted by the caller. */
+   HypreParMatrix *ParallelAssembleElim() { return ParallelAssemble(mat_e); }
+
+   /** @brief Return the matrix @a m assembled on the true dofs, i.e. P_test^t
+       A_local P_trial. */
+   /** The returned matrix has to be deleted by the caller. */
+   HypreParMatrix *ParallelAssemble(SparseMatrix *m);

   /** @brief Returns the matrix assembled on the true dofs, i.e.
       @a A = P_test^t A_local P_trial, in the format (type id) specified by
       @a A. */
-   void ParallelAssemble(OperatorHandle &A);
+   void ParallelAssemble(OperatorHandle &A) { ParallelAssemble(A, mat); }
+
+   /** Returns the eliminated matrix assembled on the true dofs, i.e.
+       @a A_elim = P^t A_elim_local P in the format (type id) specified by @a A.
+    */
+   void ParallelAssembleElim(OperatorHandle &A_elim)
+   { ParallelAssemble(A_elim, mat_e); }
+
+   /** Returns the matrix @a A_local assembled on the true dofs, i.e.
+       @a A = P_test^t A_local P_trial in the format (type id) specified by
+       @a A. */
+   void ParallelAssemble(OperatorHandle &A, SparseMatrix *A_local);
+
+   /// Eliminate essential boundary trial DOFs from the parallel system matrix.
+   /** The array @a bdr_attr_is_ess marks boundary attributes that constitute
+       the essential part of the boundary. This method relies on
+       ParallelEliminateTrialTDofs(const Array<int> &), see it for details. */
+   void ParallelEliminateTrialEssentialBC(const Array<int> &bdr_attr_is_ess);
+
+   /// Eliminate essential trial true DOFs from the parallel system matrix.
+   /** Given a list of essential trial true dofs, eliminate the trial true dofs
+       from the parallel assembled system matrix, storing the eliminated part
+       internally. This method works in conjunction with
+       ParallelEliminateTrialTDofsInRHS() and allows elimination of boundary
+       conditions in multiple right-hand sides. */
+   void ParallelEliminateTrialTDofs(const Array<int> &trial_tdof_list);
+
+   /** @brief Use the stored eliminated part of the parallel system matrix for
+       elimination of boundary conditions in the r.h.s. */
+   /** Given a list of essential trial true dofs, eliminate the trial true dofs
+       from the right-hand side @a B using the solution vector @a X and the
+       previously stored eliminated part of the parallel assembled system
+       matrix produced by ParallelEliminateTrialTDofs(const Array<int> &). */
+   void ParallelEliminateTrialTDofsInRHS(const Array<int> &trial_tdof_list,
+                                         const Vector &X, Vector &B);
+
+   /// Eliminate essential boundary test DOFs from the parallel system matrix.
+   /** The array @a bdr_attr_is_ess marks boundary attributes that constitute
+       the essential part of the boundary. */
+   void ParallelEliminateTestEssentialBC(const Array<int> &bdr_attr_is_ess);
+
+   /// Eliminate essential test true DOFs from the parallel system matrix.
+   /** Given a list of essential test true dofs, eliminate the test true dofs
+       from the parallel assembled system matrix. */
+   void ParallelEliminateTestTDofs(const Array<int> &test_tdof_list);

   using MixedBilinearForm::FormRectangularSystemMatrix;
   using MixedBilinearForm::FormRectangularLinearSystem;
@@ -332,7 +332,7 @@ ParDerefineMatrixOp::ParDerefineMatrixOp(ParFiniteElementSpace &fespace_,
   pack_col_idcs.SetSize(send_len);
   // memory manager doesn't appear to have a graceful fallback for
   // HOST_PINNED if not built with CUDA or HIP
-#if defined(MFEM_USE_CUDA) or defined(MFEM_USE_HIP)
+#if defined(MFEM_USE_CUDA) || defined(MFEM_USE_HIP)
   xghost_send.SetSize(send_len * fespace->GetVDim(),
                       Device::GetGPUAwareMPI() ? MemoryType::DEFAULT
                       : MemoryType::HOST_PINNED);
@@ -481,6 +481,7 @@ public:
       that the number of DOFs is @a ndofs. */
   const FiniteElement *GetFaceNbrFE(int i, int ndofs = 0) const;
   const FiniteElement *GetFaceNbrFaceFE(int i) const;
+   const Array<HYPRE_BigInt> &GetFaceNbrGlobalDofMapArray() { return face_nbr_glob_dof_map; }
   const HYPRE_BigInt *GetFaceNbrGlobalDofMap() { return face_nbr_glob_dof_map; }
   ElementTransformation *GetFaceNbrElementTransformation(int i) const
   { return pmesh->GetFaceNbrElementTransformation(i); }
@@ -105,6 +105,59 @@ const SparseMatrix &ParNonlinearForm::GetLocalGradient(const Vector &x) const
   return *Grad;
 }

+void ParNonlinearForm::GradientSharedFaces(const Vector &x,
+                                           int skip_zeros) const
+{
+   ParFiniteElementSpace *pfes = ParFESpace();
+   ParMesh *pmesh = pfes->GetParMesh();
+   FaceElementTransformations *T;
+   Array<int> vdofs1, vdofs2, vdofs_all;
+   DenseMatrix elemmat;
+   Vector el_x, nbr_x, face_x;
+   const Vector &px = Prolongate(x);
+
+   ParGridFunction pgf(pfes, const_cast<Vector&>(px), 0);
+   pgf.ExchangeFaceNbrData();
+
+   int nfaces = pmesh->GetNSharedFaces();
+   for (int i = 0; i < nfaces; i++)
+   {
+      T = pmesh->GetSharedFaceTransformations(i);
+      int Elem2NbrNo = T->Elem2No - pmesh->GetNE();
+
+      pfes->GetElementVDofs(T->Elem1No, vdofs1);
+      pfes->GetFaceNbrElementVDofs(Elem2NbrNo, vdofs2);
+      face_x.SetSize(vdofs1.Size() + vdofs2.Size());
+
+      el_x.MakeRef(face_x, 0, vdofs1.Size());
+      pgf.GetSubVector(vdofs1, el_x);
+
+      nbr_x.MakeRef(face_x, vdofs1.Size(), vdofs2.Size());
+      pgf.FaceNbrData().GetSubVector(vdofs2, nbr_x);
+
+      vdofs1.Copy(vdofs_all);
+      for (int j = 0; j < vdofs2.Size(); j++)
+      {
+         if (vdofs2[j] >= 0)
+         {
+            vdofs2[j] += height;
+         }
+         else
+         {
+            vdofs2[j] -= height;
+         }
+      }
+      vdofs_all.Append(vdofs2);
+      for (int k = 0; k < fnfi.Size(); k++)
+      {
+         fnfi[k]->AssembleFaceGrad(*pfes->GetFE(T->Elem1No),
+                                   *pfes->GetFaceNbrFE(Elem2NbrNo),
+                                   *T, face_x, elemmat);
+         Grad->AddSubMatrix(vdofs1, vdofs_all, elemmat, skip_zeros);
+      }
+   }
+}
+
 Operator &ParNonlinearForm::GetGradient(const Vector &x) const
 {
   if (NonlinearForm::ext) { return NonlinearForm::GetGradient(x); }
@@ -112,19 +165,61 @@ Operator &ParNonlinearForm::GetGradient(const Vector &x) const
   ParFiniteElementSpace *pfes = ParFESpace();

   pGrad.Clear();
+   OperatorHandle dA(pGrad.Type()), Ph(pGrad.Type()), hdA;

-   NonlinearForm::GetGradient(x); // (re)assemble Grad, no b.c.
-
-   OperatorHandle dA(pGrad.Type()), Ph(pGrad.Type());
-
-   if (fnfi.Size() == 0)
+   if (fnfi.Size())
   {
-      dA.MakeSquareBlockDiag(pfes->GetComm(), pfes->GlobalVSize(),
-                             pfes->GetDofOffsets(), Grad);
+      const int skip_zeros = 0;
+
+      pfes->ExchangeFaceNbrData();
+      if (Grad == NULL)
+      {
+         int nbr_size = pfes->GetFaceNbrVSize();
+         Grad = new SparseMatrix(pfes->GetVSize(), pfes->GetVSize() + nbr_size);
+      }
+
+      NonlinearForm::GetGradient(x, false); // (re)assemble Grad, no b.c.
+
+      GradientSharedFaces(x, skip_zeros);
+
+      Grad->Finalize(skip_zeros);
+
+      // handle the case when 'a' contains off-diagonal
+      int lvsize = pfes->GetVSize();
+      const HYPRE_BigInt *face_nbr_glob_ldof = pfes->GetFaceNbrGlobalDofMap();
+      HYPRE_BigInt ldof_offset = pfes->GetMyDofOffset();
+
+      Array<HYPRE_BigInt> glob_J(Grad->NumNonZeroElems());
+      int *J = Grad->GetJ();
+      for (int i = 0; i < glob_J.Size(); i++)
+      {
+         if (J[i] < lvsize)
+         {
+            glob_J[i] = J[i] + ldof_offset;
+         }
+         else
+         {
+            glob_J[i] = face_nbr_glob_ldof[J[i] - lvsize];
+         }
+      }
+
+      // TODO - construct dA directly in the A format
+      hdA.Reset(
+         new HypreParMatrix(pfes->GetComm(), lvsize, pfes->GlobalVSize(),
+                            pfes->GlobalVSize(), Grad->GetI(), glob_J,
+                            Grad->GetData(), pfes->GetDofOffsets(),
+                            pfes->GetDofOffsets()));
+      // - hdA owns the new HypreParMatrix
+      // - the above constructor copies all input arrays
+      glob_J.DeleteAll();
+      dA.ConvertFrom(hdA);
   }
   else
   {
-      MFEM_ABORT("TODO: assemble contributions from shared face terms");
+      NonlinearForm::GetGradient(x); // (re)assemble Grad, no b.c.
+
+      dA.MakeSquareBlockDiag(pfes->GetComm(), pfes->GlobalVSize(),
+                             pfes->GetDofOffsets(), Grad);
   }

   // RAP the local gradient dA.
@@ -199,9 +294,8 @@ const ParFiniteElementSpace *ParBlockNonlinearForm::ParFESpace(int k) const
 }

 // Here, rhs is a true dof vector
-void ParBlockNonlinearForm::SetEssentialBC(const
-                                           Array<Array<int> *>&bdr_attr_is_ess,
-                                           Array<Vector *> &rhs)
+void ParBlockNonlinearForm::SetEssentialBC(
+   const Array<Array<int>*> &bdr_attr_is_ess, Array<Vector*> &rhs)
 {
   Array<Vector *> nullarray(fes.Size());
   nullarray = NULL;
@@ -217,6 +311,23 @@ void ParBlockNonlinearForm::SetEssentialBC(const
   }
 }

+void ParBlockNonlinearForm::SetEssentialTrueDofs(
+   const Array<Array<int>*> &ess_tdof_list, Array<Vector*> &rhs)
+{
+   Array<Vector *> nullarray(fes.Size());
+   nullarray = nullptr;
+
+   BlockNonlinearForm::SetEssentialTrueDofs(ess_tdof_list, nullarray);
+
+   for (int s = 0; s < fes.Size(); ++s)
+   {
+      if (rhs[s])
+      {
+         rhs[s]->SetSubVector(*ess_tdofs[s], 0.0);
+      }
+   }
+}
+
 real_t ParBlockNonlinearForm::GetEnergy(const Vector &x) const
 {
   // xs_true is not modified, so const_cast is okay
@@ -255,7 +366,70 @@ void ParBlockNonlinearForm::Mult(const Vector &x, Vector &y) const

   if (fnfi.Size() > 0)
   {
-      MFEM_ABORT("TODO: assemble contributions from shared face terms");
+      // Terms over shared interior faces in parallel.
+      ParMesh *pmesh = ParFESpace(0)->GetParMesh();
+      FaceElementTransformations *tr;
+
+      Array<Array<int> *>vdofs(fes.Size());
+      Array<Array<int> *>vdofs2(fes.Size());
+      Array<Vector *> el_x(fes.Size());
+      Array<const Vector *> el_x_const(fes.Size());
+      Array<Vector *> el_y(fes.Size());
+      Array<const FiniteElement *> fe(fes.Size());
+      Array<const FiniteElement *> fe2(fes.Size());
+      Array<ParGridFunction *> pgfs(fes.Size());
+      for (int s=0; s<fes.Size(); ++s)
+      {
+         el_x_const[s] = el_x[s] = new Vector();
+         el_y[s] = new Vector();
+         vdofs[s] = new Array<int>;
+         vdofs2[s] = new Array<int>;
+         pgfs[s] = new ParGridFunction(const_cast<ParFiniteElementSpace*>(ParFESpace(s)),
+                                       xs.GetBlock(s));
+         pgfs[s]->ExchangeFaceNbrData();
+      }
+
+      const int n_shared_faces = pmesh->GetNSharedFaces();
+      for (int i = 0; i < n_shared_faces; i++)
+      {
+         tr = pmesh->GetSharedFaceTransformations(i, true);
+         int Elem2NbrNo = tr->Elem2No - pmesh->GetNE();
+
+         for (int s=0; s<fes.Size(); ++s)
+         {
+            const ParFiniteElementSpace *pfes = ParFESpace(s);
+            fe[s] = pfes->GetFE(tr->Elem1No);
+            fe2[s] = pfes->GetFaceNbrFE(Elem2NbrNo);
+
+            pfes->GetElementVDofs(tr->Elem1No, *(vdofs[s]));
+            pfes->GetFaceNbrElementVDofs(Elem2NbrNo, *(vdofs2[s]));
+
+            el_x[s]->SetSize(vdofs[s]->Size() + vdofs2[s]->Size());
+            xs.GetBlock(s).GetSubVector(*(vdofs[s]), el_x[s]->GetData());
+            pgfs[s]->FaceNbrData().GetSubVector(*(vdofs2[s]),
+                                                el_x[s]->GetData() + vdofs[s]->Size());
+         }
+
+         for (int k = 0; k < fnfi.Size(); ++k)
+         {
+            fnfi[k]->AssembleFaceVector(fe, fe2, *tr, el_x_const, el_y);
+
+            for (int s=0; s<fes.Size(); ++s)
+            {
+               if (el_y[s]->Size() == 0) { continue; }
+               ys.GetBlock(s).AddElementVector(*(vdofs[s]), *el_y[s]);
+            }
+         }
+      }
+
+      for (int s=0; s<fes.Size(); ++s)
+      {
+         delete pgfs[s];
+         delete vdofs2[s];
+         delete vdofs[s];
+         delete el_y[s];
+         delete el_x[s];
+      }
   }

   for (int s=0; s<fes.Size(); ++s)
@@ -312,6 +486,106 @@ void ParBlockNonlinearForm::SetGradientType(Operator::Type tid)
   }
 }

+void ParBlockNonlinearForm::GradientSharedFaces(const BlockVector &xs,
+                                                int skip_zeros) const
+{
+   // Terms over shared interior faces in parallel.
+   ParMesh *pmesh = ParFESpace(0)->GetParMesh();
+   FaceElementTransformations *tr;
+
+   Array<Array<int> *>vdofs(fes.Size());
+   Array<Array<int> *>vdofs2(fes.Size());
+   Array<Array<int> *>vdofs_all(fes.Size());
+   Array<Vector *> el_x(fes.Size());
+   Array<const Vector *> el_x_const(fes.Size());
+   Array2D<DenseMatrix *> elmats(fes.Size(), fes.Size());
+   Array<const FiniteElement *> fe(fes.Size());
+   Array<const FiniteElement *> fe2(fes.Size());
+   Array<ParGridFunction *> pgfs(fes.Size());
+
+   for (int s1=0; s1<fes.Size(); ++s1)
+   {
+      el_x_const[s1] = el_x[s1] = new Vector();
+      vdofs[s1] = new Array<int>;
+      vdofs2[s1] = new Array<int>;
+      vdofs_all[s1] = new Array<int>;
+      pgfs[s1] = new ParGridFunction(
+         const_cast<ParFiniteElementSpace*>(ParFESpace(s1)),
+         const_cast<Vector&>(xs.GetBlock(s1)));
+      pgfs[s1]->ExchangeFaceNbrData();
+      for (int s2=0; s2<fes.Size(); ++s2)
+      {
+         elmats(s1,s2) = new DenseMatrix();
+      }
+   }
+
+   const int n_shared_faces = pmesh->GetNSharedFaces();
+   for (int i = 0; i < n_shared_faces; i++)
+   {
+      tr = pmesh->GetSharedFaceTransformations(i, true);
+      int Elem2NbrNo = tr->Elem2No - pmesh->GetNE();
+
+      for (int s=0; s<fes.Size(); ++s)
+      {
+         const ParFiniteElementSpace *pfes = ParFESpace(s);
+         fe[s] = pfes->GetFE(tr->Elem1No);
+         fe2[s] = pfes->GetFaceNbrFE(Elem2NbrNo);
+
+         pfes->GetElementVDofs(tr->Elem1No, *(vdofs[s]));
+         pfes->GetFaceNbrElementVDofs(Elem2NbrNo, *(vdofs2[s]));
+
+         el_x[s]->SetSize(vdofs[s]->Size() + vdofs2[s]->Size());
+         xs.GetBlock(s).GetSubVector(*(vdofs[s]), el_x[s]->GetData());
+         pgfs[s]->FaceNbrData().GetSubVector(*(vdofs2[s]),
+                                             el_x[s]->GetData() + vdofs[s]->Size());
+
+         vdofs[s]->Copy(*vdofs_all[s]);
+
+         const int lvsize = pfes->GetVSize();
+         for (int j = 0; j < vdofs2[s]->Size(); j++)
+         {
+            if ((*vdofs2[s])[j] >= 0)
+            {
+               (*vdofs2[s])[j] += lvsize;
+            }
+            else
+            {
+               (*vdofs2[s])[j] -= lvsize;
+            }
+         }
+         vdofs_all[s]->Append(*(vdofs2[s]));
+      }
+
+      for (int k = 0; k < fnfi.Size(); ++k)
+      {
+         fnfi[k]->AssembleFaceGrad(fe, fe2, *tr, el_x_const, elmats);
+
+         for (int s1=0; s1<fes.Size(); ++s1)
+         {
+            for (int s2=0; s2<fes.Size(); ++s2)
+            {
+               if (elmats(s1,s2)->Height() == 0) { continue; }
+               Grads(s1,s2)->AddSubMatrix(*vdofs[s1], *vdofs_all[s2],
+                                          *elmats(s1,s2), skip_zeros);
+            }
+         }
+      }
+   }
+
+   for (int s1=0; s1<fes.Size(); ++s1)
+   {
+      delete pgfs[s1];
+      delete vdofs_all[s1];
+      delete vdofs2[s1];
+      delete vdofs[s1];
+      delete el_x[s1];
+      for (int s2=0; s2<fes.Size(); ++s2)
+      {
+         delete elmats(s1,s2);
+      }
+   }
+}
+
 BlockOperator & ParBlockNonlinearForm::GetGradient(const Vector &x) const
 {
   if (pBlockGrad == NULL)
@@ -331,49 +605,155 @@ BlockOperator & ParBlockNonlinearForm::GetGradient(const Vector &x) const
      }
   }

-   GetLocalGradient(x); // gradients are stored in 'Grads'
+   // xs_true is not modified, so const_cast is okay
+   xs_true.Update(const_cast<Vector &>(x), block_trueOffsets);
+   xs.Update(block_offsets);
+
+   for (int s=0; s<fes.Size(); ++s)
+   {
+      fes[s]->GetProlongationMatrix()->Mult(
+         xs_true.GetBlock(s), xs.GetBlock(s));
+   }

   if (fnfi.Size() > 0)
   {
-      MFEM_ABORT("TODO: assemble contributions from shared face terms");
-   }
+      const int skip_zeros = 0;

-   for (int s1=0; s1<fes.Size(); ++s1)
-   {
-      for (int s2=0; s2<fes.Size(); ++s2)
+      for (int s=0; s<fes.Size(); ++s)
      {
-         OperatorHandle dA(phBlockGrad(s1,s2)->Type()),
-                        Ph(phBlockGrad(s1,s2)->Type()),
-                        Rh(phBlockGrad(s1,s2)->Type());
+         const_cast<ParFiniteElementSpace*>(pfes[s])->ExchangeFaceNbrData();
+      }

-         if (s1 == s2)
+      for (int s1=0; s1<fes.Size(); ++s1)
+      {
+         for (int s2=0; s2<fes.Size(); ++s2)
         {
-            dA.MakeSquareBlockDiag(pfes[s1]->GetComm(), pfes[s1]->GlobalVSize(),
-                                   pfes[s1]->GetDofOffsets(), Grads(s1,s1));
-            Ph.ConvertFrom(pfes[s1]->Dof_TrueDof_Matrix());
-            phBlockGrad(s1,s1)->MakePtAP(dA, Ph);
-
-            OperatorHandle Ae;
-            Ae.EliminateRowsCols(*phBlockGrad(s1,s1), *ess_tdofs[s1]);
+            if (Grads(s1,s2) == NULL)
+            {
+               int nbr_size = pfes[s2]->GetFaceNbrVSize();
+               Grads(s1,s2) = new SparseMatrix(pfes[s1]->GetVSize(),
+                                               pfes[s2]->GetVSize() + nbr_size);
+            }
         }
-         else
+      }
+
+      // (re)assemble Grad without b.c. into 'Grads'
+      BlockNonlinearForm::ComputeGradientBlocked(xs, false);
+
+      GradientSharedFaces(xs, skip_zeros);
+
+      // finalize the gradients
+      for (int s1=0; s1<fes.Size(); ++s1)
+         for (int s2=0; s2<fes.Size(); ++s2)
         {
-            dA.MakeRectangularBlockDiag(pfes[s1]->GetComm(),
-                                        pfes[s1]->GlobalVSize(),
-                                        pfes[s2]->GlobalVSize(),
-                                        pfes[s1]->GetDofOffsets(),
-                                        pfes[s2]->GetDofOffsets(),
-                                        Grads(s1,s2));
-            Rh.ConvertFrom(pfes[s1]->Dof_TrueDof_Matrix());
-            Ph.ConvertFrom(pfes[s2]->Dof_TrueDof_Matrix());
-
-            phBlockGrad(s1,s2)->MakeRAP(Rh, dA, Ph);
-
-            phBlockGrad(s1,s2)->EliminateRows(*ess_tdofs[s1]);
-            phBlockGrad(s1,s2)->EliminateCols(*ess_tdofs[s2]);
+            Grads(s1,s2)->Finalize(skip_zeros);
         }

-         pBlockGrad->SetBlock(s1, s2, phBlockGrad(s1,s2)->Ptr());
+      for (int s1=0; s1<fes.Size(); ++s1)
+      {
+         for (int s2=0; s2<fes.Size(); ++s2)
+         {
+            OperatorHandle hdA;
+            OperatorHandle dA(phBlockGrad(s1,s2)->Type()),
+                           Ph(phBlockGrad(s1,s2)->Type()),
+                           Rh(phBlockGrad(s1,s2)->Type());
+
+            // handle the case when 'a' contains off-diagonal
+            int lvsize = pfes[s2]->GetVSize();
+            const HYPRE_BigInt *face_nbr_glob_ldof =
+               const_cast<ParFiniteElementSpace*>(pfes[s2])->GetFaceNbrGlobalDofMap();
+            HYPRE_BigInt ldof_offset = pfes[s2]->GetMyDofOffset();
+
+            Array<HYPRE_BigInt> glob_J(Grads(s1,s2)->NumNonZeroElems());
+            int *J = Grads(s1,s2)->GetJ();
+            for (int i = 0; i < glob_J.Size(); i++)
+            {
+               if (J[i] < lvsize)
+               {
+                  glob_J[i] = J[i] + ldof_offset;
+               }
+               else
+               {
+                  glob_J[i] = face_nbr_glob_ldof[J[i] - lvsize];
+               }
+            }
+
+            // TODO - construct dA directly in the A format
+            hdA.Reset(
+               new HypreParMatrix(pfes[s2]->GetComm(), pfes[s1]->GetVSize(),
+                                  pfes[s1]->GlobalVSize(), pfes[s2]->GlobalVSize(),
+                                  Grads(s1,s2)->GetI(), glob_J, Grads(s1,s2)->GetData(),
+                                  pfes[s1]->GetDofOffsets(), pfes[s2]->GetDofOffsets()));
+            // - hdA owns the new HypreParMatrix
+            // - the above constructor copies all input arrays
+            glob_J.DeleteAll();
+            dA.ConvertFrom(hdA);
+
+            if (s1 == s2)
+            {
+               Ph.ConvertFrom(pfes[s1]->Dof_TrueDof_Matrix());
+               phBlockGrad(s1,s1)->MakePtAP(dA, Ph);
+
+               OperatorHandle Ae;
+               Ae.EliminateRowsCols(*phBlockGrad(s1,s1), *ess_tdofs[s1]);
+            }
+            else
+            {
+               Rh.ConvertFrom(pfes[s1]->Dof_TrueDof_Matrix());
+               Ph.ConvertFrom(pfes[s2]->Dof_TrueDof_Matrix());
+
+               phBlockGrad(s1,s2)->MakeRAP(Rh, dA, Ph);
+
+               phBlockGrad(s1,s2)->EliminateRows(*ess_tdofs[s1]);
+               phBlockGrad(s1,s2)->EliminateCols(*ess_tdofs[s2]);
+            }
+
+            pBlockGrad->SetBlock(s1, s2, phBlockGrad(s1,s2)->Ptr());
+         }
+      }
+   }
+   else
+   {
+      // (re)assemble Grad without b.c. into 'Grads'
+      BlockNonlinearForm::ComputeGradientBlocked(xs);
+
+      for (int s1=0; s1<fes.Size(); ++s1)
+      {
+         for (int s2=0; s2<fes.Size(); ++s2)
+         {
+            OperatorHandle dA(phBlockGrad(s1,s2)->Type()),
+                           Ph(phBlockGrad(s1,s2)->Type()),
+                           Rh(phBlockGrad(s1,s2)->Type());
+
+            if (s1 == s2)
+            {
+               dA.MakeSquareBlockDiag(pfes[s1]->GetComm(), pfes[s1]->GlobalVSize(),
+                                      pfes[s1]->GetDofOffsets(), Grads(s1,s1));
+               Ph.ConvertFrom(pfes[s1]->Dof_TrueDof_Matrix());
+               phBlockGrad(s1,s1)->MakePtAP(dA, Ph);
+
+               OperatorHandle Ae;
+               Ae.EliminateRowsCols(*phBlockGrad(s1,s1), *ess_tdofs[s1]);
+            }
+            else
+            {
+               dA.MakeRectangularBlockDiag(pfes[s1]->GetComm(),
+                                           pfes[s1]->GlobalVSize(),
+                                           pfes[s2]->GlobalVSize(),
+                                           pfes[s1]->GetDofOffsets(),
+                                           pfes[s2]->GetDofOffsets(),
+                                           Grads(s1,s2));
+               Rh.ConvertFrom(pfes[s1]->Dof_TrueDof_Matrix());
+               Ph.ConvertFrom(pfes[s2]->Dof_TrueDof_Matrix());
+
+               phBlockGrad(s1,s2)->MakeRAP(Rh, dA, Ph);
+
+               phBlockGrad(s1,s2)->EliminateRows(*ess_tdofs[s1]);
+               phBlockGrad(s1,s2)->EliminateCols(*ess_tdofs[s2]);
+            }
+
+            pBlockGrad->SetBlock(s1, s2, phBlockGrad(s1,s2)->Ptr());
+         }
      }
   }

@@ -29,6 +29,8 @@ protected:
   mutable ParGridFunction X, Y;
   mutable OperatorHandle pGrad;

+   void GradientSharedFaces(const Vector &x, int skip_zeros = 1) const;
+
 public:
   ParNonlinearForm(ParFiniteElementSpace *pf);

@@ -81,6 +83,8 @@ protected:
   mutable Array2D<OperatorHandle *> phBlockGrad;
   mutable BlockOperator *pBlockGrad;

+   void GradientSharedFaces(const BlockVector &xs, int skip_zeros) const;
+
 public:
   /// Computes the energy of the system
   real_t GetEnergy(const Vector &x) const override;
@@ -102,9 +106,39 @@ public:
       gradient-type (if different from the default) must be set again. */
   void SetParSpaces(Array<ParFiniteElementSpace *> &pf);

-   // Here, rhs is a true dof vector
-   void SetEssentialBC(const Array<Array<int> *>&bdr_attr_is_ess,
-                       Array<Vector *> &rhs) override;
+   /** @brief Set essential boundary conditions to each finite element space
+       using boundary attribute markers.
+
+       This method calls `FiniteElementSpace::GetEssentialTrueDofs()` for each
+       space and stores ess_tdof_lists internally.
+
+       If `rhs` vectors are non-null, the entries corresponding to these
+       essential DoFs are set to zero. This ensures compatibility with the
+       output of the `Mult()` method, which also zeroes out these entries.
+
+       @param[in] bdr_attr_is_ess A list of boundary attribute markers for each
+       space.
+       @param[in,out] rhs         An array of optional right-hand side vectors.
+       If a vector at `rhs[i]` is non-null, its essential DoFs will be set
+       to zero. */
+   virtual void SetEssentialBC(const Array<Array<int>*> &bdr_attr_is_ess,
+                               Array<Vector*> &rhs) override;
+
+   /** @brief Set essential boundary conditions to each finite element space
+       using essential true dof lists.
+
+       This method stores a copy of the provided essential true dof lists.
+
+       If `rhs` vectors are non-null, the entries corresponding to these
+       essential DoFs are set to zero. This ensures compatibility with the
+       output of the `Mult()` method, which also zeroes out these entries.
+
+       @param[in] ess_tdof_list A list of essential true dofs for each space.
+       @param[in,out] rhs       An array of optional right-hand side vectors.
+       If a vector at `rhs[i]` is non-null, its essential DoFs will be set
+       to zero. */
+   virtual void SetEssentialTrueDofs(const Array<Array<int>*> &ess_tdof_list,
+                                     Array<Vector*> &rhs) override;

   /// Block T-Vector to Block T-Vector
   void Mult(const Vector &x, Vector &y) const override;
@@ -37,6 +37,8 @@ void InitDetKernels()
   k::Specialization<3,3,3,3>::Add();
   k::Specialization<3,3,3,5>::Add();
   k::Specialization<3,3,3,6>::Add();
+   k::Specialization<3,3,4,6>::Add();
+   k::Specialization<3,3,3,4>::Add();
 }

 } // namespace quadrature_interpolator
@@ -28,6 +28,7 @@ void InitEvalByNodesKernels()
   k::Specialization<2,QVectorLayout::byNODES,1,2,4>::Opt<1>::Add();
   k::Specialization<2,QVectorLayout::byNODES,1,3,2>::Opt<1>::Add();
   k::Specialization<2,QVectorLayout::byNODES,1,3,4>::Opt<1>::Add();
+   k::Specialization<2,QVectorLayout::byNODES,1,3,6>::Opt<1>::Add();
   k::Specialization<2,QVectorLayout::byNODES,1,4,3>::Opt<1>::Add();
   k::Specialization<2,QVectorLayout::byNODES,1,4,4>::Opt<1>::Add();

@@ -30,6 +30,7 @@ void InitEvalByVDimKernels()
   k::Specialization<2,QVectorLayout::byVDIM,2,2,4>::Opt<8>::Add();
   k::Specialization<2,QVectorLayout::byVDIM,2,3,4>::Opt<8>::Add();
   k::Specialization<2,QVectorLayout::byVDIM,2,3,6>::Opt<4>::Add();
+   k::Specialization<2,QVectorLayout::byVDIM,2,4,6>::Opt<2>::Add();
   k::Specialization<2,QVectorLayout::byVDIM,2,4,8>::Opt<2>::Add();
   // 3D
   k::Specialization<3,QVectorLayout::byVDIM,1,2,4>::Opt<1>::Add();
@@ -47,6 +48,9 @@ void InitEvalByVDimKernels()
   k::Specialization<3,QVectorLayout::byVDIM,3,7,7>::Opt<1>::Add();
   k::Specialization<3,QVectorLayout::byVDIM,3,8,8>::Opt<1>::Add();
   k::Specialization<3,QVectorLayout::byVDIM,3,9,9>::Opt<1>::Add();
+
+   k::Specialization<3,QVectorLayout::byVDIM,3,4,6>::Opt<1>::Add();
+   k::Specialization<3,QVectorLayout::byVDIM,3,3,4>::Opt<1>::Add();
 }

 } // namespace quadrature_interpolator
@@ -22,7 +22,7 @@ QuadratureSpaceBase::QuadratureSpaceBase(Mesh &mesh_, Geometry::Type geom,
 {
   for (int g = 0; g < Geometry::NumGeom; g++)
   {
-      int_rule[g] = NULL;
+      int_rule[g] = nullptr;
   }
   int_rule[geom] = &ir;
 }
@@ -37,6 +37,29 @@ void QuadratureSpaceBase::ConstructIntRules(int dim)
   }
 }

+const Array<int> &QuadratureSpaceBase::Offsets(
+   QSpaceOffsetStorage storage) const
+{
+   if (storage == QSpaceOffsetStorage::COMPRESSED || offsets.Size() > 1)
+   {
+      return offsets;
+   }
+   else
+   {
+      if (full_offset_cache.Size() == 0)
+      {
+         const int nq = size / ne;
+         full_offset_cache.SetSize(ne + 1);
+         int *d_full_offset_cache = full_offset_cache.Write();
+         mfem::forall(ne + 1, [=] MFEM_HOST_DEVICE (int e)
+         {
+            d_full_offset_cache[e] = nq * e;
+         });
+      }
+      return full_offset_cache;
+   }
+}
+
 namespace
 {

@@ -116,7 +139,7 @@ void QuadratureSpace::ConstructOffsets()
      {
         offsets[i] = offset;
         const Geometry::Type geom = mesh.GetElementBaseGeometry(i);
-         MFEM_ASSERT(int_rule[geom] != NULL, "Missing integration rule.");
+         MFEM_ASSERT(int_rule[geom] != nullptr, "Missing integration rule.");
         offset += int_rule[geom]->GetNPoints();
      }
      offsets[num_elem] = offset;
@@ -181,9 +204,9 @@ const Vector &QuadratureSpace::GetGeometricFactorWeights() const

 FaceQuadratureSpace::FaceQuadratureSpace(Mesh &mesh_, int order_,
                                         FaceType face_type_)
-   : QuadratureSpaceBase(mesh_, order_),
-     face_type(face_type_),
-     num_faces(mesh.GetNFbyType(face_type))
+   : QuadratureSpaceBase(mesh_, order_), face_type(face_type_),
+     face_indices(mesh.GetFaceIndices(face_type_)),
+     face_indices_inv(mesh.GetInvFaceIndices(face_type_))
 {
   Construct();
 }
@@ -192,7 +215,8 @@ FaceQuadratureSpace::FaceQuadratureSpace(Mesh &mesh_, const IntegrationRule &ir,
                                         FaceType face_type_)
   : QuadratureSpaceBase(mesh_, mesh_.GetTypicalFaceGeometry(), ir),
     face_type(face_type_),
-     num_faces(mesh.GetNFbyType(face_type))
+     face_indices(mesh.GetFaceIndices(face_type_)),
+     face_indices_inv(mesh.GetInvFaceIndices(face_type_))
 {
   MFEM_VERIFY(mesh.GetNumGeometries(mesh.Dimension() - 1) <= 1,
               "Constructor not valid for mixed meshes");
@@ -201,28 +225,30 @@ FaceQuadratureSpace::FaceQuadratureSpace(Mesh &mesh_, const IntegrationRule &ir,

 void FaceQuadratureSpace::ConstructOffsets()
 {
-   face_indices.SetSize(num_faces);
-   offsets.SetSize(num_faces + 1);
-   ne = num_faces;
-   int offset = 0;
-   int f_idx = 0;
-   for (int i = 0; i < mesh.GetNumFacesWithGhost(); i++)
-   {
-      const Mesh::FaceInformation face = mesh.GetFaceInformation(i);
-      if (face.IsNonconformingCoarse() || !face.IsOfFaceType(face_type))
-      {
-         continue;
-      }
-      face_indices[f_idx] = i;
-      face_indices_inv[i] = f_idx;
-      offsets[f_idx] = offset;
-      Geometry::Type geom = mesh.GetFaceGeometry(i);
-      MFEM_ASSERT(int_rule[geom] != NULL, "Missing integration rule");
-      offset += int_rule[geom]->GetNPoints();
+   ne = face_indices.Size();

-      f_idx++;
+   if (mesh.GetNumGeometries(mesh.Dimension() - 1) == 1)
+   {
+      Array<Geometry::Type> geoms;
+      mesh.GetGeometries(mesh.Dimension() - 1, geoms);
+      offsets.SetSize(1);
+      offsets.HostWrite();
+      offsets[0] = int_rule[geoms[0]]->GetNPoints();
+      size = ne * offsets[0];
+   }
+   else
+   {
+      offsets.SetSize(face_indices.Size() + 1);
+      int offset = 0;
+      for (int i = 0; i < mesh.GetNFbyType(face_type); ++i)
+      {
+         offsets[i] = offset;
+         Geometry::Type geom = mesh.GetFaceGeometry(face_indices[i]);
+         MFEM_ASSERT(int_rule[geom] != nullptr, "Missing integration rule");
+         offset += int_rule[geom]->GetNPoints();
+      }
+      offsets[face_indices.Size()] = size = offset;
   }
-   offsets[num_faces] = size = offset;
 }

 void FaceQuadratureSpace::Construct()
@@ -248,6 +274,16 @@ int FaceQuadratureSpace::GetPermutedIndex(int idx, int iq) const
   }
 }

+ElementTransformation *FaceQuadratureSpace::GetTransformation(int idx)
+{
+   ElementTransformation *T = mesh.GetFaceTransformation(face_indices[idx]);
+   if (face_type == FaceType::Boundary)
+   {
+      T->Attribute = mesh.GetBdrFaceAttributes()[idx];
+   }
+   return T;
+}
+
 int FaceQuadratureSpace::GetEntityIndex(const ElementTransformation &T) const
 {
   auto get_face_index = [this](const int idx)
@@ -19,6 +19,12 @@
 namespace mfem
 {

+enum class QSpaceOffsetStorage
+{
+   FULL,
+   COMPRESSED
+};
+
 /// Abstract base class for QuadratureSpace and FaceQuadratureSpace.
 /** This class represents the storage layout for QuadratureFunction%s, that may
    be defined either on mesh elements or mesh faces. */
@@ -30,7 +36,7 @@ protected:
   Mesh &mesh; ///< The underlying mesh.
   int order; ///< The order of integration rule.
   int size; ///< Total number of quadrature points.
-   int ne; ///< Actual number of entities
+   int ne; ///< Number of entities
   mutable Vector weights; ///< Integration weights.
   mutable long nodes_sequence = 0; ///< Nodes counter for cache invalidation.

@@ -41,7 +47,17 @@ protected:
   /// can be computed as i * offsets[0], where i is the entity index. Otherwise
   /// has size num_entities + 1.
   ///
+   /// In the non-compressed case, the quadrature point values for entity i are
+   /// stored in the indices between offsets[i] and offsets[i+1].
   Array<int> offsets;
+
+   /// @brief Cached version of the "full" offsets, returned by Offsets() when
+   /// QSpaceOffsetStorage::FULL is provided.
+   ///
+   /// The quadrature point values for entity i are stored in the indices
+   /// between offsets[i] and offsets[i+1].
+   mutable Array<int> full_offset_cache;
+
   /// The quadrature rules used for each geometry type.
   const IntegrationRule *int_rule[Geometry::NumGeom];

@@ -74,12 +90,18 @@ public:

   /// @brief Entity quadrature point offset array.
   ///
-   /// Supports a constant compression scheme for meshes which have a single
+   /// If @a storage is QSpaceOffsetStorage::COMPRESSED, then the returned array
+   /// supports a constant compression scheme for meshes which have a single
   /// geometry type. When compressed, will have a single value. The true offset
   /// can be computed as i * offsets[0], where i is the entity index. Otherwise
   /// has size num_entities + 1.
   ///
-   const Array<int> &Offsets() const { return offsets; }
+   /// If @a storage is QSpaceOffsetStorage::FULL, then the array will never be
+   /// compressed.
+   ///
+   /// In the non-compressed case, the quadrature point values for entity i are
+   /// stored in the indices between offsets[i] and offsets[i+1].
+   const Array<int> &Offsets(QSpaceOffsetStorage storage) const;

   /// Return the total number of quadrature points.
   int GetSize() const { return size; }
@@ -191,13 +213,12 @@ public:
 class FaceQuadratureSpace : public QuadratureSpaceBase
 {
   FaceType face_type; ///< Is the space defined on interior or boundary faces?
-   const int num_faces; ///< Number of faces.

   /// Map from boundary or interior face indices to mesh face indices.
-   Array<int> face_indices;
+   const Array<int> &face_indices;

   /// Inverse of the map @a face_indices.
-   std::unordered_map<int,int> face_indices_inv;
+   const std::unordered_map<int,int> &face_indices_inv;

   const Vector &GetGeometricFactorWeights() const override;
   void ConstructOffsets();
@@ -213,14 +234,13 @@ public:
                       FaceType face_type_);

   /// Returns number of faces in the mesh.
-   inline int GetNumFaces() const { return num_faces; }
+   inline int GetNumFaces() const { return face_indices.Size(); }

   /// Returns the face type (boundary or interior).
   FaceType GetFaceType() const { return face_type; }

   /// Returns the face transformation of face @a idx.
-   ElementTransformation *GetTransformation(int idx) override
-   { return mesh.GetFaceTransformation(face_indices[idx]); }
+   ElementTransformation *GetTransformation(int idx) override;

   /// Returns the geometry type of face @a idx.
   Geometry::Type GetGeometry(int idx) const override
@@ -140,6 +140,36 @@ void add_3D(const scalartype &scalar, const std::vector<type> &u,

 /* Metric definitions */

+// W = ||T||^2 - 2*det(T).
+template <typename type>
+type mu4_ad(const std::vector<type> &T, const std::vector<type> &W)
+{
+   auto fnorm2 = fnorm2_2D(T);
+   auto det = det_2D(T);
+   return fnorm2 - 2*det;
+};
+
+// W = ||T-I||^2.
+template <typename type>
+type mu14_ad(const std::vector<type> &T, const std::vector<type> &W)
+{
+   DenseMatrix Id(2,2); Id = 0.0;
+   Id(0,0) = 1; Id(1,1) = 1;
+
+   std::vector<type> Mat;
+   add_2D(real_t{-1.0}, T, &Id, Mat);
+
+   return fnorm2_2D(Mat);
+};
+
+// W = (det(T)-1)^2.
+template <typename type>
+type mu55_ad(const std::vector<type> &T, const std::vector<type> &W)
+{
+   auto det = det_2D(T);
+   return pow(det-1.0, 2.0);
+};
+
 // W = |T-T'|^2, where T'= |T|*I/sqrt(2).
 template <typename type>
 type mu85_ad(const std::vector<type> &T, const std::vector<type> &W)
@@ -163,6 +193,63 @@ type mu98_ad(const std::vector<type> &T, const std::vector<type> &W)
   return fnorm2_2D(Mat)/det_2D(T);
 };

+template <typename type>
+type make_one_type()
+{
+   return 1.0;
+}
+// add specialization for AD1Type
+template <>
+AD1Type make_one_type<AD1Type>()
+{
+   return AD1Type{1.0, 0.0};
+}
+// add specialization for AD2Type
+template <>
+AD2Type make_one_type<AD2Type>()
+{
+   return AD2Type{AD1Type{1.0, 0.0}, AD1Type{0.0, 0.0}};
+}
+
+using TWCUO = TMOP_WorstCaseUntangleOptimizer_Metric;
+template <typename type>
+type wcuo_ad(type mu,
+             const std::vector<type> &T, const std::vector<type> &W,
+             real_t alpha, real_t min_detT, real_t detT_ep,
+             int exponent, real_t max_muT, real_t muT_ep,
+             TWCUO::BarrierType bt,
+             TWCUO::WorstCaseType wct)
+{
+   type one = make_one_type<type>();
+   type zero = 0.0*one;
+   type denom = one;
+   if (bt == TWCUO::BarrierType::Shifted)
+   {
+      auto val1 = alpha*min_detT-detT_ep < 0.0 ?
+                  (alpha*min_detT-detT_ep)*one :
+                  zero;
+      denom = 2.0*(det_2D(T)-val1);
+   }
+   else if (bt == TWCUO::BarrierType::Pseudo)
+   {
+      auto detT = det_2D(T);
+      denom = detT + sqrt(detT*detT + detT_ep*detT_ep);
+   }
+   mu = mu/denom;
+
+   if (wct == TWCUO::WorstCaseType::PMean)
+   {
+      auto exp = exponent*one;
+      mu = pow(mu, exp);
+   }
+   else if (wct == TWCUO::WorstCaseType::Beta)
+   {
+      auto beta = (max_muT+muT_ep)*one;
+      mu = mu/(beta-mu);
+   }
+   return mu;
+}
+
 // W = 1/(tau^0.5) |T-I|^2.
 template <typename type>
 type mu342_ad(const std::vector<type> &T, const std::vector<type> &W)
@@ -421,7 +508,7 @@ void TMOP_QualityMetric::DefaultAssembleH(const DenseTensor &H,
         {
            for (int cc = 0; cc < dim; cc++)
            {
-               const double entry_rr_cc = Hrc(rr, cc);
+               const real_t entry_rr_cc = Hrc(rr, cc);

               for (int i = 0; i < dof; i++)
               {
@@ -481,6 +568,30 @@ void TMOP_Combo_QualityMetric::EvalPW(const DenseMatrix &Jpt,
   }
 }

+AD1Type TMOP_Combo_QualityMetric::EvalW_AD1(const std::vector<AD1Type> &T,
+                                            const std::vector<AD1Type> &W)
+const
+{
+   AD1Type metric = {0., 0.};
+   for (int i = 0; i < tmop_q_arr.Size(); i++)
+   {
+      metric += wt_arr[i]*tmop_q_arr[i]->EvalW_AD1(T, W);
+   }
+   return metric;
+}
+
+AD2Type TMOP_Combo_QualityMetric::EvalW_AD2(const std::vector<AD2Type> &T,
+                                            const std::vector<AD2Type> &W)
+const
+{
+   AD2Type metric = {{0., 0.},{0., 0.}};
+   for (int i = 0; i < tmop_q_arr.Size(); i++)
+   {
+      metric += wt_arr[i]*tmop_q_arr[i]->EvalW_AD2(T, W);
+   }
+   return metric;
+}
+
 void TMOP_Combo_QualityMetric::AssembleH(const DenseMatrix &Jpt,
                                         const DenseMatrix &DS,
                                         const real_t weight,
@@ -645,6 +756,64 @@ real_t TMOP_WorstCaseUntangleOptimizer_Metric::EvalWBarrier(
   return tmop_metric.EvalW(Jpt)/denominator;
 }

+AD1Type TMOP_WorstCaseUntangleOptimizer_Metric::EvalW_AD1(
+   const std::vector<AD1Type> &T,
+   const std::vector<AD1Type> &W) const
+{
+   return wcuo_ad(tmop_metric.EvalW_AD1(T,W), T, W, alpha, min_detT, detT_ep,
+                  exponent, max_muT, muT_ep, btype, wctype);
+}
+
+AD2Type TMOP_WorstCaseUntangleOptimizer_Metric::EvalW_AD2(
+   const std::vector<AD2Type> &T,
+   const std::vector<AD2Type> &W) const
+{
+   return wcuo_ad(tmop_metric.EvalW_AD2(T,W), T, W, alpha, min_detT, detT_ep,
+                  exponent, max_muT, muT_ep, btype, wctype);
+}
+
+void TMOP_WorstCaseUntangleOptimizer_Metric::EvalP(const DenseMatrix &Jpt,
+                                                   DenseMatrix &P) const
+{
+   auto mu_ad_fn = [this](std::vector<AD1Type> &T, std::vector<AD1Type> &W)
+   {
+      return EvalW_AD1(T,W);
+   };
+   if (tmop_metric.Id() == 4 || tmop_metric.Id() == 14 ||
+       tmop_metric.Id() == 66)
+   {
+      ADGrad(mu_ad_fn, P, Jpt);
+      return;
+   }
+   MFEM_ABORT("EvalW_AD1 not implemented with this metric for "
+              "TMOP_WorstCaseUntangleOptimizer_Metric. "
+              "Please use metric 4/14/66.");
+}
+
+void TMOP_WorstCaseUntangleOptimizer_Metric::AssembleH(
+   const DenseMatrix &Jpt,
+   const DenseMatrix &DS,
+   const real_t weight,
+   DenseMatrix &A) const
+{
+   DenseTensor H(Jpt.Height(), Jpt.Height(), Jpt.TotalSize());
+   H = 0.0;
+   auto mu_ad_fn = [this](std::vector<AD2Type> &T, std::vector<AD2Type> &W)
+   {
+      return EvalW_AD2(T,W);
+   };
+   if (tmop_metric.Id() == 4 || tmop_metric.Id() == 14 ||
+       tmop_metric.Id() == 66)
+   {
+      ADHessian(mu_ad_fn, H, Jpt);
+      this->DefaultAssembleH(H,DS,weight,A);
+      return;
+   }
+   MFEM_ABORT("EvalW_AD1 not implemented with this metric for "
+              "TMOP_WorstCaseUntangleOptimizer_Metric. "
+              "Please use metric 4/14/66.");
+}
+
 real_t TMOP_Metric_001::EvalW(const DenseMatrix &Jpt) const
 {
   ie.SetJacobian(Jpt.GetData());
@@ -850,6 +1019,25 @@ void TMOP_Metric_004::AssembleH(const DenseMatrix &Jpt,
   ie.Assemble_ddI2b(-2.0*weight, A.GetData());
 }

+template <typename type>
+type TMOP_Metric_004::EvalW_AD_impl(const std::vector<type> &T,
+                                    const std::vector<type> &W) const
+{
+   return mu4_ad(T, W);
+}
+
+AD1Type TMOP_Metric_004::EvalW_AD1(const std::vector<AD1Type> &T,
+                                   const std::vector<AD1Type> &W) const
+{
+   return EvalW_AD_impl<AD1Type>(T,W);
+}
+
+AD2Type TMOP_Metric_004::EvalW_AD2(const std::vector<AD2Type> &T,
+                                   const std::vector<AD2Type> &W) const
+{
+   return EvalW_AD_impl<AD2Type>(T,W);
+}
+
 real_t TMOP_Metric_007::EvalW(const DenseMatrix &Jpt) const
 {
   // mu_7 = |J-J^{-t}|^2 = |J|^2 + |J^{-1}|^2 - 4
@@ -971,6 +1159,25 @@ void TMOP_Metric_014::AssembleH(const DenseMatrix &Jpt,
   ie.Assemble_ddI1(weight, A.GetData());
 }

+template <typename type>
+type TMOP_Metric_014::EvalW_AD_impl(const std::vector<type> &T,
+                                    const std::vector<type> &W) const
+{
+   return mu14_ad(T, W);
+}
+
+AD1Type TMOP_Metric_014::EvalW_AD1(const std::vector<AD1Type> &T,
+                                   const std::vector<AD1Type> &W) const
+{
+   return EvalW_AD_impl<AD1Type>(T,W);
+}
+
+AD2Type TMOP_Metric_014::EvalW_AD2(const std::vector<AD2Type> &T,
+                                   const std::vector<AD2Type> &W) const
+{
+   return EvalW_AD_impl<AD2Type>(T,W);
+}
+
 real_t TMOP_Metric_022::EvalW(const DenseMatrix &Jpt) const
 {
   // mu_22 = (0.5*|J|^2 - det(J)) / (det(J) - tau0)
@@ -1100,6 +1307,25 @@ void TMOP_Metric_055::AssembleH(const DenseMatrix &Jpt,
   ie.Assemble_ddI2b(2*weight*(ie.Get_I2b() - 1.0), A.GetData());
 }

+template <typename type>
+type TMOP_Metric_055::EvalW_AD_impl(const std::vector<type> &T,
+                                    const std::vector<type> &W) const
+{
+   return mu55_ad(T, W);
+}
+
+AD1Type TMOP_Metric_055::EvalW_AD1(const std::vector<AD1Type> &T,
+                                   const std::vector<AD1Type> &W) const
+{
+   return EvalW_AD_impl<AD1Type>(T,W);
+}
+
+AD2Type TMOP_Metric_055::EvalW_AD2(const std::vector<AD2Type> &T,
+                                   const std::vector<AD2Type> &W) const
+{
+   return EvalW_AD_impl<AD2Type>(T,W);
+}
+
 real_t TMOP_Metric_056::EvalWMatrixForm(const DenseMatrix &Jpt) const
 {
   // mu_56 = 0.5 (det(J) + 1 / det(J)) - 1.
@@ -4096,7 +4322,7 @@ real_t TMOP_Integrator::GetElementEnergy(const FiniteElement &el,

         const IntegrationPoint &ip_s = ir_s->IntPoint(s);
         Tpr->SetIntPoint(&ip_s);
-         double w = surf_fit_coeff->Eval(*Tpr, ip_s) * surf_fit_normal *
+         real_t w = surf_fit_coeff->Eval(*Tpr, ip_s) * surf_fit_normal *
                    1.0 / surf_fit_dof_count[scalar_dof_id];

         if (surf_fit_gf)
@@ -14,10 +14,14 @@

 #include "../linalg/invariants.hpp"
 #include "nonlininteg.hpp"
+#include "../linalg/dual.hpp"

 namespace mfem
 {

+using AD1Type = future::dual<real_t, real_t>;
+using AD2Type = future::dual<AD1Type, AD1Type>;
+
 /** @brief Abstract class for local mesh quality metrics in the target-matrix
    optimization paradigm (TMOP) by P. Knupp et al. */
 class TMOP_QualityMetric : public HyperelasticModel
@@ -69,6 +73,22 @@ public:
   virtual void EvalPW(const DenseMatrix &Jpt, DenseMatrix &PW) const
   { PW = 0.0;}

+   /// @brief First-derivative hook for AD-based computations.
+   /// @warning Not for public use. Internal use for AD-based computations.
+   virtual AD1Type EvalW_AD1(const std::vector<AD1Type> &T,
+                             const std::vector<AD1Type> &W) const
+   {
+      MFEM_ABORT("EvalW_AD1 not implemented for this metric");
+   }
+
+   /// @brief Second-derivative hook for AD-based computations.
+   /// @warning Not for public use. Internal use for AD-based computations.
+   virtual AD2Type EvalW_AD2(const std::vector<AD2Type> &T,
+                             const std::vector<AD2Type> &W) const
+   {
+      MFEM_ABORT("EvalW_AD2 not implemented for this metric");
+   }
+
   /** @brief Evaluate the derivative of the 1st Piola-Kirchhoff stress tensor
       and assemble its contribution to the local gradient matrix 'A'.
       @param[in] Jpt     Represents the target->physical transformation
@@ -124,6 +144,12 @@ public:
   void AssembleH(const DenseMatrix &Jpt, const DenseMatrix &DS,
                  const real_t weight, DenseMatrix &A) const override;

+   AD1Type EvalW_AD1(const std::vector<AD1Type> &T,
+                     const std::vector<AD1Type> &W) const override;
+
+   AD2Type EvalW_AD2(const std::vector<AD2Type> &T,
+                     const std::vector<AD2Type> &W) const override;
+
   /// Computes the averages of all metrics (integral of metric / volume).
   /// Works in parallel when called with a ParGridFunction.
   void ComputeAvgMetrics(const GridFunction &nodes,
@@ -221,12 +247,16 @@ public:

   real_t EvalW(const DenseMatrix &Jpt) const override;

-   void EvalP(const DenseMatrix &Jpt, DenseMatrix &P) const override
-   { MFEM_ABORT("Not implemented"); }
+   AD1Type EvalW_AD1(const std::vector<AD1Type> &T,
+                     const std::vector<AD1Type> &W) const override;
+
+   AD2Type EvalW_AD2(const std::vector<AD2Type> &T,
+                     const std::vector<AD2Type> &W) const override;
+
+   void EvalP(const DenseMatrix &Jpt, DenseMatrix &P) const override;

   void AssembleH(const DenseMatrix &Jpt, const DenseMatrix &DS,
-                  const real_t weight, DenseMatrix &A) const override
-   { MFEM_ABORT("Not implemented"); }
+                  const real_t weight, DenseMatrix &A) const override;

   // Compute mu_hat.
   real_t EvalWBarrier(const DenseMatrix &Jpt) const;
@@ -368,6 +398,10 @@ class TMOP_Metric_004 : public TMOP_QualityMetric
 protected:
   mutable InvariantsEvaluator2D<real_t> ie;

+   template<typename type>
+   type EvalW_AD_impl(const std::vector<type> &T,
+                      const std::vector<type> &W) const;
+
 public:
   // W = |J|^2 - 2*det(J)
   real_t EvalW(const DenseMatrix &Jpt) const override;
@@ -377,6 +411,12 @@ public:
   void AssembleH(const DenseMatrix &Jpt, const DenseMatrix &DS,
                  const real_t weight, DenseMatrix &A) const override;

+   AD1Type EvalW_AD1(const std::vector<AD1Type> &T,
+                     const std::vector<AD1Type> &W) const override;
+
+   AD2Type EvalW_AD2(const std::vector<AD2Type> &T,
+                     const std::vector<AD2Type> &W) const override;
+
   int Id() const override { return 4; }
 };

@@ -420,6 +460,10 @@ class TMOP_Metric_014 : public TMOP_QualityMetric
 protected:
   mutable InvariantsEvaluator2D<real_t> ie;

+   template <typename type>
+   type EvalW_AD_impl(const std::vector<type> &T,
+                      const std::vector<type> &W) const;
+
 public:
   // W = |J - I|^2.
   real_t EvalWMatrixForm(const DenseMatrix &Jpt) const override;
@@ -431,6 +475,14 @@ public:

   void AssembleH(const DenseMatrix &Jpt, const DenseMatrix &DS,
                  const real_t weight, DenseMatrix &A) const override;
+
+   AD1Type EvalW_AD1(const std::vector<AD1Type> &T,
+                     const std::vector<AD1Type> &W) const override;
+
+   AD2Type EvalW_AD2(const std::vector<AD2Type> &T,
+                     const std::vector<AD2Type> &W) const override;
+
+   int Id() const override { return 14; }
 };

 /// 2D Shifted barrier form of shape metric (mu_2).
@@ -479,6 +531,10 @@ class TMOP_Metric_055 : public TMOP_QualityMetric
 protected:
   mutable InvariantsEvaluator2D<real_t> ie;

+   template<typename type>
+   type EvalW_AD_impl(const std::vector<type> &T,
+                      const std::vector<type> &W) const;
+
 public:
   // W = (det(J) - 1)^2.
   real_t EvalW(const DenseMatrix &Jpt) const override;
@@ -488,6 +544,14 @@ public:
   void AssembleH(const DenseMatrix &Jpt, const DenseMatrix &DS,
                  const real_t weight, DenseMatrix &A) const override;

+   AD1Type EvalW_AD1(const std::vector<AD1Type> &T,
+                     const std::vector<AD1Type> &W) const override;
+
+   AD2Type EvalW_AD2(const std::vector<AD2Type> &T,
+                     const std::vector<AD2Type> &W) const override;
+
+   int Id() const override { return 55; }
+
 };

 /// 2D barrier size (V) metric (polyconvex).
@@ -111,7 +111,7 @@ public:

   ~InterpolatorFP()
   {
-      finder->FreeData();
+      if (finder) { finder->FreeData(); }
      delete finder;
   }
 };
@@ -1,199 +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.
-
-
-// Abstract array data type
-
-#include "array.hpp"
-#include "../general/forall.hpp"
-#include <fstream>
-#include <type_traits>
-
-namespace mfem
-{
-
-template <class T>
-void Array<T>::Print(std::ostream &os, int width) const
-{
-   for (int i = 0; i < size; i++)
-   {
-      os << data[i];
-      if ( !((i+1) % width) || i+1 == size )
-      {
-         os << '\n';
-      }
-      else
-      {
-         os << " ";
-      }
-   }
-}
-
-template <class T>
-void Array<T>::Save(std::ostream &os, int fmt) const
-{
-   if (fmt == 0)
-   {
-      os << size << '\n';
-   }
-   for (int i = 0; i < size; i++)
-   {
-      os << operator[](i) << '\n';
-   }
-}
-
-template <class T>
-void Array<T>::Load(std::istream &in, int fmt)
-{
-   if (fmt == 0)
-   {
-      int new_size;
-      in >> new_size;
-      SetSize(new_size);
-   }
-   for (int i = 0; i < size; i++)
-   {
-      in >> operator[](i);
-   }
-}
-
-template <class T>
-T Array<T>::Max() const
-{
-   MFEM_ASSERT(size > 0, "Array is empty with size " << size);
-
-   T max = operator[](0);
-   for (int i = 1; i < size; i++)
-   {
-      if (max < operator[](i))
-      {
-         max = operator[](i);
-      }
-   }
-
-   return max;
-}
-
-template <class T>
-T Array<T>::Min() const
-{
-   MFEM_ASSERT(size > 0, "Array is empty with size " << size);
-
-   T min = operator[](0);
-   for (int i = 1; i < size; i++)
-   {
-      if (operator[](i) < min)
-      {
-         min = operator[](i);
-      }
-   }
-
-   return min;
-}
-
-// Partial Sum
-template <class T>
-void Array<T>::PartialSum()
-{
-   T sum = static_cast<T>(0);
-   for (int i = 0; i < size; i++)
-   {
-      sum+=operator[](i);
-      operator[](i) = sum;
-   }
-}
-
-template <class T>
-void Array<T>::Abs()
-{
-   static_assert(std::is_arithmetic<T>::value, "Use with arithmetic types!");
-   const bool useDevice = UseDevice();
-   const int N = size;
-   auto y = ReadWrite(useDevice);
-   mfem::forall_switch(useDevice, N, [=] MFEM_HOST_DEVICE (int i)
-   {
-      y[i] = std::abs(y[i]);
-   });
-}
-
-// Sum
-template <class T>
-T Array<T>::Sum() const
-{
-   T sum = static_cast<T>(0);
-   for (int i = 0; i < size; i++)
-   {
-      sum+=operator[](i);
-   }
-
-   return sum;
-}
-
-template <class T>
-int Array<T>::IsSorted() const
-{
-   T val_prev = operator[](0), val;
-   for (int i = 1; i < size; i++)
-   {
-      val=operator[](i);
-      if (val < val_prev)
-      {
-         return 0;
-      }
-      val_prev = val;
-   }
-
-   return 1;
-}
-
-
-template <class T>
-void Array2D<T>::Load(const char *filename, int fmt)
-{
-   std::ifstream in;
-   in.open(filename, std::ifstream::in);
-   MFEM_VERIFY(in.is_open(), "File " << filename << " does not exist.");
-   Load(in, fmt);
-   in.close();
-}
-
-template <class T>
-void Array2D<T>::Print(std::ostream &os, int width_)
-{
-   int height = this->NumRows();
-   int width  = this->NumCols();
-
-   for (int i = 0; i < height; i++)
-   {
-      os << "[row " << i << "]\n";
-      for (int j = 0; j < width; j++)
-      {
-         os << (*this)(i,j);
-         if ( (j+1) == width_ || (j+1) % width_ == 0 )
-         {
-            os << '\n';
-         }
-         else
-         {
-            os << ' ';
-         }
-      }
-   }
-}
-
-template class Array<char>;
-template class Array<int>;
-template class Array<long long>;
-template class Array<real_t>;
-template class Array2D<int>;
-template class Array2D<real_t>;
-
-} // namespace mfem
@@ -16,9 +16,13 @@
 #include "mem_manager.hpp"
 #include "device.hpp"
 #include "error.hpp"
+#include "forall.hpp"
 #include "globals.hpp"
+#include "reducers.hpp"
+#include "scan.hpp"

 #include <iostream>
+#include <fstream>
 #include <cstdlib>
 #include <cstring>
 #include <algorithm>
@@ -135,6 +139,8 @@ public:
   /// Return the device flag of the Memory object used by the Array
   bool UseDevice() const { return data.UseDevice(); }

+   void UseDevice(bool use_dev) { data.UseDevice(use_dev); }
+
   /// Return true if the data will be deleted by the Array
   inline bool OwnsData() const { return data.OwnsHostPtr(); }

@@ -217,7 +223,7 @@ public:
   /// Reduces the capacity of the array to exactly match the current size.
   inline void ShrinkToFit();

-   ///  Create a copy of the internal array to the provided @a copy.
+   /// Create a copy of the internal array to the provided @a copy.
   inline void Copy(Array &copy) const;

   /// Make this Array a reference to a pointer.
@@ -275,11 +281,11 @@ public:

   /** @brief Find the maximal element in the array, using the comparison
       operator `<` for class T. */
-   T Max() const;
+   inline T Max() const;

   /** @brief Find the minimal element in the array, using the comparison
       operator `<` for class T. */
-   T Min() const;
+   inline T Min() const;

   /// Sorts the array in ascending order. This requires operator< to be defined for T.
   void Sort() { std::sort((T*)data, data + size); }
@@ -297,19 +303,22 @@ public:
   }

   /// Return 1 if the array is sorted from lowest to highest.  Otherwise return 0.
-   int IsSorted() const;
+   inline int IsSorted() const;

   /// Does the Array have Size zero.
   bool IsEmpty() const { return Size() == 0; }

+   /// Return true if all entries of the array are the same.
+   inline bool IsConstant() const;
+
   /// Fill the entries of the array with the cumulative sum of the entries.
-   void PartialSum();
+   inline void PartialSum();

   /// Replace each entry of the array with its absolute value.
-   void Abs();
+   inline void Abs();

   /// Return the sum of all the array entries using the '+'' operator for class 'T'.
-   T Sum() const;
+   inline T Sum() const;

   /// Set all entries of the array to the provided constant.
   inline void operator=(const T &a);
@@ -412,10 +421,13 @@ private:

 public:
   Array2D() { M = N = 0; }
+
+   /// Construct an m x n 2D array.
   Array2D(int m, int n) : array1d(m*n) { M = m; N = n; }

   Array2D(const Array2D &) = default;

+   /// Set the 2D array size to m x n.
   void SetSize(int m, int n) { array1d.SetSize(m*n); M = m; N = n; }

   int NumRows() const { return M; }
@@ -472,9 +484,11 @@ public:
   void Load(int new_size0,int new_size1, std::istream &in)
   { SetSize(new_size0,new_size1); Load(in, 1); }

+   /// Create a copy of the internal array to the provided @a copy.
   void Copy(Array2D &copy) const
   { copy.M = M; copy.N = N; array1d.Copy(copy.array1d); }

+   /// Set all entries of the array to the provided constant.
   inline void operator=(const T &a)
   { array1d = a; }

@@ -489,6 +503,14 @@ public:

   /// Prints array to stream with width elements per row
   void Print(std::ostream &out = mfem::out, int width = 4);
+
+   /** @brief Find the maximal element in the array, using the comparison
+        operator `<` for class T. */
+   T Max() const { return array1d.Max(); }
+
+   /** @brief Find the minimal element in the array, using the comparison
+       operator `<` for class T. */
+   T Min() const { return array1d.Min(); }
 };


@@ -501,15 +523,32 @@ private:

 public:
   Array3D() { N2 = N3 = 0; }
+
+   /// Construct a 3D array of size n1 x n2 x n3.
   Array3D(int n1, int n2, int n3)
      : array1d(n1*n2*n3) { N2 = n2; N3 = n3; }

+   /// Set the 3D array size to n1 x n2 x n3.
   void SetSize(int n1, int n2, int n3)
   { array1d.SetSize(n1*n2*n3); N2 = n2; N3 = n3; }

+   /// Get the 3D array size in the first dimension.
+   int GetSize1() const
+   {
+      const int size = array1d.Size();
+      return size == 0 ? 0 : size / (N2 * N3);
+   }
+
+   /// Get the 3D array size in the second dimension.
+   int GetSize2() const { return N2; }
+
+   /// Get the 3D array size in the third dimension.
+   int GetSize3() const { return N3; }
+
   inline const T &operator()(int i, int j, int k) const;
   inline       T &operator()(int i, int j, int k);

+   /// Set all entries of the array to the provided constant.
   inline void operator=(const T &a)
   { array1d = a; }
 };
@@ -764,8 +803,14 @@ template <typename T> template <typename CT>
 inline Array<T> &Array<T>::operator=(const Array<CT> &src)
 {
   SetSize(src.Size());
-   for (int i = 0; i < size; i++) { (*this)[i] = T(src[i]); }
-   return *this;
+
+   const bool use_dev = UseDevice() || src.UseDevice();
+   const auto x = src.Read(use_dev);
+   auto y = Write(use_dev);
+   mfem::forall_switch(use_dev, size, [=] MFEM_HOST_DEVICE (int i)
+   {
+      y[i] = x[i];
+   });
 }

 template <class T>
@@ -981,19 +1026,24 @@ template <class T>
 inline void Array<T>::GetSubArray(int offset, int sa_size, Array<T> &sa) const
 {
   sa.SetSize(sa_size);
-   for (int i = 0; i < sa_size; i++)
+   const bool use_dev = UseDevice() || sa.UseDevice();
+   const auto x = Read(use_dev);
+   auto y = sa.Write(use_dev);
+   mfem::forall_switch(use_dev, sa_size, [=] MFEM_HOST_DEVICE (int i)
   {
-      sa[i] = (*this)[offset+i];
-   }
+      y[i] = x[offset + i];
+   });
 }

 template <class T>
 inline void Array<T>::operator=(const T &a)
 {
-   for (int i = 0; i < size; i++)
+   const bool use_dev = UseDevice();
+   auto x = Write(use_dev);
+   mfem::forall_switch(use_dev, size, [=] MFEM_HOST_DEVICE (int i)
   {
-      data[i] = a;
-   }
+      x[i] = a;
+   });
 }

 template <class T>
@@ -1002,6 +1052,153 @@ inline void Array<T>::Assign(const T *p)
   data.CopyFromHost(p, Size());
 }

+template <class T>
+inline void Array<T>::Print(std::ostream &os, int width) const
+{
+   for (int i = 0; i < size; i++)
+   {
+      os << data[i];
+      if ( !((i+1) % width) || i+1 == size )
+      {
+         os << '\n';
+      }
+      else
+      {
+         os << " ";
+      }
+   }
+}
+
+template <class T>
+inline void Array<T>::Save(std::ostream &os, int fmt) const
+{
+   if (fmt == 0)
+   {
+      os << size << '\n';
+   }
+   for (int i = 0; i < size; i++)
+   {
+      os << operator[](i) << '\n';
+   }
+}
+
+template <class T>
+void Array<T>::Load(std::istream &in, int fmt)
+{
+   if (fmt == 0)
+   {
+      int new_size;
+      in >> new_size;
+      SetSize(new_size);
+   }
+   for (int i = 0; i < size; i++)
+   {
+      in >> operator[](i);
+   }
+}
+
+template <class T>
+inline T Array<T>::Max() const
+{
+   MFEM_ASSERT(size > 0, "Array is empty with size " << size);
+
+   T max = operator[](0);
+   for (int i = 1; i < size; i++)
+   {
+      if (max < operator[](i))
+      {
+         max = operator[](i);
+      }
+   }
+
+   return max;
+}
+
+template <class T>
+inline T Array<T>::Min() const
+{
+   MFEM_ASSERT(size > 0, "Array is empty with size " << size);
+
+   T min = operator[](0);
+   for (int i = 1; i < size; i++)
+   {
+      if (operator[](i) < min)
+      {
+         min = operator[](i);
+      }
+   }
+
+   return min;
+}
+
+// Partial Sum
+template <class T>
+inline void Array<T>::PartialSum()
+{
+   auto data_ptr = ReadWrite(UseDevice());
+   InclusiveScan(UseDevice(), data_ptr, data_ptr, size);
+}
+
+template <class T>
+inline void Array<T>::Abs()
+{
+   static_assert(std::is_arithmetic<T>::value, "Use with arithmetic types!");
+   const bool useDevice = UseDevice();
+   const int N = size;
+   auto y = ReadWrite(useDevice);
+   mfem::forall_switch(useDevice, N, [=] MFEM_HOST_DEVICE (int i)
+   {
+      y[i] = std::abs(y[i]);
+   });
+}
+
+// Sum
+template <class T>
+inline T Array<T>::Sum() const
+{
+   T sum = static_cast<T>(0);
+   if (size > 0)
+   {
+      const auto m_data = Read(UseDevice());
+      reduce(size, sum, [=] MFEM_HOST_DEVICE(int i, T &r) { r += m_data[i]; },
+      /*  */ SumReducer<T> {}, UseDevice());
+   }
+   return sum;
+}
+
+template <class T>
+inline int Array<T>::IsSorted() const
+{
+   T val_prev = operator[](0), val;
+   for (int i = 1; i < size; i++)
+   {
+      val=operator[](i);
+      if (val < val_prev)
+      {
+         return 0;
+      }
+      val_prev = val;
+   }
+
+   return 1;
+}
+
+template <class T>
+inline bool Array<T>::IsConstant() const
+{
+   if (size < 2) { return true; }
+   const T v0 = data[0];
+   for (int i = 1; i < size; i++)
+   {
+      if (data[i] != v0)
+      {
+         return false;
+      }
+   }
+
+   return true;
+}
+

 template <class T>
 inline const T &Array2D<T>::operator()(int i, int j) const
@@ -1041,6 +1238,40 @@ inline T *Array2D<T>::operator[](int i)
   return &array1d[i*N];
 }

+template <class T>
+void Array2D<T>::Load(const char *filename, int fmt)
+{
+   std::ifstream in;
+   in.open(filename, std::ifstream::in);
+   MFEM_VERIFY(in.is_open(), "File " << filename << " does not exist.");
+   Load(in, fmt);
+   in.close();
+}
+
+template <class T>
+void Array2D<T>::Print(std::ostream &os, int width_)
+{
+   int height = this->NumRows();
+   int width  = this->NumCols();
+
+   for (int i = 0; i < height; i++)
+   {
+      os << "[row " << i << "]\n";
+      for (int j = 0; j < width; j++)
+      {
+         os << (*this)(i,j);
+         if ( (j+1) == width_ || (j+1) % width_ == 0 )
+         {
+            os << '\n';
+         }
+         else
+         {
+            os << ' ';
+         }
+      }
+   }
+}
+

 template <class T>
 inline void Swap(Array2D<T> &a, Array2D<T> &b)
@@ -57,9 +57,9 @@ struct Hashed4
 *  each time this class is invoked.
 *
 *  There are two main methods this class provides. The Get(...) methods always
- *  return an item given the two or four indices. If the item didn't previously
+ *  return an item given the two or four indices. If the item did not previously
 *  exist, the methods creates a new one. The Find(...) methods, on the other
- *  hand, just return NULL or -1 if the item doesn't exist.
+ *  hand, just return NULL or -1 if the item does not exist.
 *
 *  Each new item is automatically assigned a unique ID - the index of the item
 *  inside the BlockArray. The IDs may (but need not) be used as p1, p2, ... of
@@ -95,14 +95,14 @@ public:
       @param[in] init_hash_size The initial size of the hash table. Must be
                                 a power of 2. */
   HashTable(int block_size = 16*1024, int init_hash_size = 32*1024);
-   /// @brief Deep copy
+   /// Deep copy
   HashTable(const HashTable& other);
-   /// @brief Copy assignment not supported
+   /// Copy assignment not supported
   HashTable& operator=(const HashTable&) = delete;
   ~HashTable();

-   /** @brief Item accessor with key (or parents) the pair 'p1', 'p2'. Default
-       construct an item of type T if no value correspond to the requested key.
+   /** @brief Item accessor with key (or parents) the pair p1, p2. Default
+       construct an item of type T if no value corresponds to the requested key.

       @param[in] p1 First part of the key.
       @param[in] p2 Second part of the key.
@@ -111,9 +111,9 @@ public:
       @warning This method should only be called if T inherits from Hashed2. */
   T* Get(int p1, int p2);

-   /** @brief Item accessor with key (or parents) the quadruplet 'p1', 'p2',
-       'p3', 'p4'. The key 'p4' is optional. Default construct an item of type T
-       if no value corresponds to the requested key.
+   /** @brief Item accessor with key (or parents) the quadruplet p1, p2, p3, p4.
+       The key p4 is optional. Default construct an item of type T if no value
+       corresponds to the requested key.

       @param[in] p1 First part of the key.
       @param[in] p2 Second part of the key.
@@ -124,10 +124,10 @@ public:
      @warning This method should only be called if T inherits from Hashed4. */
   T* Get(int p1, int p2, int p3, int p4 = -1 /* p4 optional */);

-   /// Get id of item whose parents are p1, p2... Create it if it doesn't exist.
-   /** @brief Get the "id" of an item, this "id" corresponding to the index of the
-       item in the underlying BlockArray<T> object. Default construct an item
-       and id if no value corresponds to the requested key.
+   /** @brief Get the "id" of the item whose parents are p1, p2, this "id"
+       corresponding to the index of the item in the underlying BlockArray<T>
+       object. Default construct an item and "id" if no value corresponds to the
+       requested key.

       @param[in] p1 First part of the key.
       @param[in] p2 Second part of the key.
@@ -136,9 +136,9 @@ public:
       @warning This method should only be called if T inherits from Hashed2. */
   int GetId(int p1, int p2);

-   /** @brief Get the "id" of an item, this "id" corresponding to the index of the
-       item in the underlying BlockArray<T> object. Default construct an item
-       and id if no value correspond to the requested key.
+   /** @brief Get the "id" of an item, this "id" corresponding to the index of
+       the item in the underlying BlockArray<T> object. Default construct an item
+       and "id" if no value corresponds to the requested key.

       @param[in] p1 First part of the key.
       @param[in] p2 Second part of the key.
@@ -149,9 +149,8 @@ public:
       @warning This method should only be called if T inherits from Hashed4. */
   int GetId(int p1, int p2, int p3, int p4 = -1);

-   /// Find item whose parents are p1, p2... Return NULL if it doesn't exist.
-   /** @brief Item accessor with key (or parents) the pair 'p1', 'p2'. Return
-       nullptr if no value correspond to the requested key.
+   /** @brief Item accessor with key (or parents) the pair p1, p2. Return
+       NULL if no value corresponds to the requested key.

       @param[in] p1 First part of the key.
       @param[in] p2 Second part of the key.
@@ -160,9 +159,9 @@ public:
       @warning This method should only be called if T inherits from Hashed2. */
   T* Find(int p1, int p2);

-   /** @brief Item accessor with key (or parents) the quadruplet 'p1', 'p2',
-       'p3', 'p4'. The key 'p4' is optional. Return nullptr if no value
-       correspond to the requested key.
+   /** @brief Item accessor with key (or parents) the quadruplet p1, p2, p3, p4.
+       The key p4 is optional. Return NULL if no value corresponds to the
+       requested key.

       @param[in] p1 First part of the key.
       @param[in] p2 Second part of the key.
@@ -173,8 +172,8 @@ public:
       @warning This method should only be called if T inherits from Hashed4. */
   T* Find(int p1, int p2, int p3, int p4 = -1);

-   /** @brief Item const accessor with key (or parents) the pair 'p1', 'p2'.
-       Return nullptr if no value correspond to the requested key.
+   /** @brief Item const accessor with key (or parents) the pair p1, p2.
+       Return NULL if no value corresponds to the requested key.

       @param[in] p1 First part of the key.
       @param[in] p2 Second part of the key.
@@ -183,9 +182,9 @@ public:
      @warning This method should only be called if T inherits from Hashed2. */
   const T* Find(int p1, int p2) const;

-   /** @brief Item const accessor with key (or parents) the quadruplet 'p1',
-       'p2', 'p3', 'p4'. The key 'p4' is optional. Return nullptr if no value
-       correspond to the requested key.
+   /** @brief Item const accessor with key (or parents) the quadruplet p1, p2,
+       p3, p4. The key p4 is optional. Return NULL if no value corresponds to the
+       requested key.

       @param[in] p1 First part of the key.
       @param[in] p2 Second part of the key.
@@ -196,10 +195,12 @@ public:
       @warning This method should only be called if T inherits from Hashed4. */
   const T* Find(int p1, int p2, int p3, int p4 = -1) const;

-   /// Find id of item whose parents are p1, p2... Return -1 if it doesn't exist.
-   /** @brief Find the "id" of an item, this "id" corresponding to the index of
-       the item in the underlying BlockArray<T> object. Default construct an
-       item and id if no value correspond to the requested key.
+   /** @brief Find the "id" of an item whose parents are p1, p2. Return -1 if it
+       does not exist.
+
+       This "id" corresponds to the index of the item in the underlying
+       BlockArray<T> object. Default construct an item and "id" if no value
+       corresponds to the requested key.

       @param[in] p1 First part of the key.
       @param[in] p2 Second part of the key.
@@ -209,8 +210,9 @@ public:
   int FindId(int p1, int p2) const;

   /** @brief Find the "id" of an item, this "id" corresponding to the index of
-       the item in the underlying BlockArray<T> object. Default construct an
-       item and id if no value correspond to the requested key.
+       the item in the underlying BlockArray<T> object. Return -1 if it does not
+       exist. Default construct an item and "id" if no value corresponds to the
+       requested key.

       @param[in] p1 First part of the key.
       @param[in] p2 Second part of the key.
@@ -221,16 +223,16 @@ public:
       @warning This method should only be called if T inherits from Hashed4. */
   int FindId(int p1, int p2, int p3, int p4 = -1) const;

-   /// @brief Return the number of elements currently stored in the HashTable.
+   /// Return the number of elements currently stored in the HashTable.
   int Size() const { return Base::Size() - unused.Size(); }

-   /// @brief Return the total number of ids (used and unused) in the HashTable.
+   /// Return the total number of ids (used and unused) in the HashTable.
   int NumIds() const { return Base::Size(); }

-   /// @brief Return the number of free/unused ids in the HashTable.
+   /// Return the number of free/unused ids in the HashTable.
   int NumFreeIds() const { return unused.Size(); }

-   /** @brief Return true if item 'id' exists in (is used by) the container.
+   /** @brief Return true if item @a id exists in (is used by) the container.

       @param[in] id Index of the item in the underlying BlockArray<T>.

@@ -241,13 +243,13 @@ public:

       @param[in] id Index of the item in the underlying BlockArray<T>.

-       @warning Its id will be reused by newly added items. */
+       @warning Its @a id will be reused by newly added items. */
   void Delete(int id);

-   /// @brief Remove all items.
+   /// Remove all items.
   void DeleteAll();

-   /** @brief Allocate an item at 'id'. Enlarge the underlying BlockArray if
+   /** @brief Allocate an item at @a id. Enlarge the underlying BlockArray if
       necessary.

       @param[in] id Index of the item in the underlying BlockArray<T>.
@@ -255,7 +257,7 @@ public:
       @param[in] p2 Second part of the key.

       @warning This is a special purpose method used when loading data from a
-       file. Does nothing if the slot 'id' has already been allocated. */
+       file. Does nothing if the slot @a id has already been allocated. */
   void Alloc(int id, int p1, int p2);

   /** @brief Reinitialize the internal list of unallocated items.
@@ -287,13 +289,13 @@ public:
       @warning This method should only be called if T inherits from Hashed4. */
   void Reparent(int id, int new_p1, int new_p2, int new_p3, int new_p4 = -1);

-   /// @brief Return total size of allocated memory (tables plus items), in bytes.
+   /// Return total size of allocated memory (tables plus items), in bytes.
   std::size_t MemoryUsage() const;

-   /// @brief Write details of the memory usage to the mfem output stream.
+   /// Write details of the memory usage to the mfem output stream.
   void PrintMemoryDetail() const;

-   /// @brief Print a histogram of bin sizes for debugging purposes.
+   /// Print a histogram of bin sizes for debugging purposes.
   void PrintStats() const;

   class iterator : public Base::iterator
@@ -346,7 +348,7 @@ public:

 protected:
   /** The hash table: each bin is a linked list of items. For each non-empty
-       bin, this arrays stores the 'id' of the first item in the list, or -1
+       bin, this arrays stores the "id" of the first item in the list, or -1
       if the bin is empty. */
   int* table;

@@ -384,11 +386,11 @@ protected:
   { return (984120265ul*p1 + 125965121ul*p2 + 495698413ul*p3) & mask; }

   // Delete() and Reparent() use one of these:
-   /// @brief Hash function for items of type T that inherit from Hashed2.
+   /// Hash function for items of type T that inherit from Hashed2.
   inline int Hash(const Hashed2& item) const
   { return Hash(item.p1, item.p2); }

-   /// @brief Hash function for items of type T that inherit from Hashed4.
+   /// Hash function for items of type T that inherit from Hashed4.
   inline int Hash(const Hashed4& item) const
   { return Hash(item.p1, item.p2, item.p3); }

@@ -415,15 +417,15 @@ protected:
       @warning This method should only be called if T inherits from Hashed4. */
   int SearchList(int id, int p1, int p2, int p3) const;

-   /** @brief Insert the item 'id' into bin 'idx'.
+   /** @brief Insert the item @a id into bin @a idx.

       @param[in] idx The bin/bucket index.
       @param[in] id The index of the item in the BlockArray<T>.
       @param[in] item The item to insert at the beginning of the linked list.

-       @warning The method only works with bin 'idx' and does not check the
-                overall fill factor of the hash table. If appropriate,
-                use CheckRehash() for that. */
+       @warning The method only works with bin @a idx and does not check the
+                overall fill factor of the hash table. If appropriate, use
+                CheckRehash() for that. */
   inline void Insert(int idx, int id, T &item);

   /** @brief Unlink an item @a id from the linked list of bin @a idx.
@@ -444,11 +446,11 @@ protected:
       and reinsert all items into the new bins.

       NOTE: Rehashing is computationally expensive (O(N) in the number of items),
-       but since it is only done rarely (when the number of items doubles),
-       the amortized complexity of inserting an item is still O(1). */
+       but since it is only done rarely (when the number of items doubles), the
+       amortized complexity of inserting an item is still O(1). */
   void DoRehash();

-   /** @brief Return the size of the bin "idx".
+   /** @brief Return the size of the bin @a idx.

       @param[in] idx The index of the bin.
       @return The size of the bin. */
@@ -1384,8 +1384,11 @@ void MemoryManager::Insert(void *h_ptr, size_t bytes,
   {
      auto &m = res.first->second;
      MFEM_VERIFY(m.bytes >= bytes && m.h_mt == h_mt &&
-                  (m.d_mt == d_mt || (d_mt == MemoryType::DEFAULT &&
-                                      m.d_mt == GetDualMemoryType(h_mt))),
+                  (m.d_mt == d_mt ||
+                   (d_mt == MemoryType::DEFAULT &&
+                    m.d_mt == GetDualMemoryType(h_mt)) ||
+                   (m.d_mt == MemoryType::DEFAULT &&
+                    d_mt == GetDualMemoryType(m.h_mt))),
                  "Address already present with different attributes!");
 #ifdef MFEM_TRACK_MEM_MANAGER
      mfem::out << "[mfem memory manager]: repeated registration of h_ptr: "
@@ -12,7 +12,6 @@
 #ifndef MFEM_REDUCERS_HPP
 #define MFEM_REDUCERS_HPP

-#include "array.hpp"
 #include "forall.hpp"

 #include <cmath>
@@ -466,7 +465,7 @@ template<class B, class R> struct reduction_kernel
   /// helper for computing the reduction block size
   static int block_log2(unsigned N)
   {
-#if defined(__GNUC__) or defined(__clang__)
+#if defined(__GNUC__) || defined(__clang__)
      return N ? (sizeof(unsigned) * 8 - __builtin_clz(N)) : 0;
 #elif defined(_MSC_VER)
      return sizeof(unsigned) * 8 - __lzclz(N);
@@ -514,6 +513,33 @@ template<class B, class R> struct reduction_kernel
      }
   }
 };
+
+template <class T>
+class ReductionWorkspace
+{
+   Memory<T> workspace;
+
+   static ReductionWorkspace &Instance()
+   {
+      static ReductionWorkspace instance;
+      return instance;
+   }
+
+   ~ReductionWorkspace() { workspace.Delete(); }
+
+public:
+   static T *Get(int num_blocks)
+   {
+      ReductionWorkspace &instance = Instance();
+      if (instance.workspace.Capacity() < num_blocks)
+      {
+         instance.workspace.Delete();
+         instance.workspace.New(num_blocks, MemoryType::HOST_PINNED);
+      }
+      return instance.workspace;
+   }
+};
+
 }

 /**
@@ -529,8 +555,7 @@ template<class B, class R> struct reduction_kernel
 @tparam T value_type to operate on
 */
 template <class T, class B, class R>
-void reduce(int N, T &res, B &&body, const R &reducer, bool use_dev,
-            Array<T> &workspace)
+void reduce(int N, T &res, B &&body, const R &reducer, bool use_dev)
 {
   if (N == 0)
   {
@@ -567,13 +592,7 @@ void reduce(int N, T &res, B &&body, const R &reducer, bool use_dev,

      red_type red{nullptr, std::forward<B>(body), reducer, N, items_per_thread};
      // allocate res to fit block_size entries
-      auto mt = workspace.GetMemory().GetMemoryType();
-      if (mt != MemoryType::HOST_PINNED && mt != MemoryType::MANAGED)
-      {
-         mt = MemoryType::HOST_PINNED;
-      }
-      workspace.SetSize(nblocks, mt);
-      auto work = workspace.HostWrite();
+      auto work = internal::ReductionWorkspace<T>::Get(nblocks);
      red.work = work;
      forall_2D(nblocks, block_size, 1, std::move(red));
      // wait for results
@@ -0,0 +1,258 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#ifndef MFEM_SCAN_HPP
+#define MFEM_SCAN_HPP
+
+#include "backends.hpp"
+
+#ifdef MFEM_USE_CUDA
+#include <cub/device/device_scan.cuh>
+#define MFEM_CUB_NAMESPACE cub
+#elif defined(MFEM_USE_HIP)
+#include <hipcub/device/device_scan.hpp>
+#define MFEM_CUB_NAMESPACE hipcub
+#endif
+
+#include <functional>
+#include <numeric>
+#include <cstddef>
+
+namespace mfem
+{
+
+namespace internal
+{
+class ScanWorkspace
+{
+   Memory<std::byte> workspace;
+   static ScanWorkspace &Instance()
+   {
+      static ScanWorkspace instance;
+      return instance;
+   }
+   ~ScanWorkspace() { workspace.Delete(); }
+public:
+   static std::byte *Get(int num_bytes)
+   {
+      ScanWorkspace &instance = Instance();
+      if (Size() < num_bytes)
+      {
+         instance.workspace.Delete();
+         instance.workspace.New(num_bytes);
+      }
+      return instance.workspace.Write(MemoryClass::DEVICE, Size());
+   }
+   static int Size()
+   {
+      return Instance().workspace.Capacity();
+   }
+};
+}
+
+/// Equivalent to InclusiveScan(use_dev, d_in, d_out, num_items, std::plus<>{})
+template <class InputIt, class OutputIt>
+void InclusiveScan(bool use_dev, InputIt d_in, OutputIt d_out, size_t num_items)
+{
+   // forward to InclusiveSum for potentially faster kernels
+#if defined(MFEM_USE_CUDA) || defined(MFEM_USE_HIP)
+   if (use_dev && mfem::Device::Allows(Backend::CUDA_MASK | Backend::HIP_MASK))
+   {
+      using internal::ScanWorkspace;
+      size_t bytes = ScanWorkspace::Size();
+      if (bytes > 0)
+      {
+         auto err = MFEM_CUB_NAMESPACE::DeviceScan::InclusiveSum(
+                       ScanWorkspace::Get(bytes), bytes, d_in, d_out, num_items);
+#if defined(MFEM_USE_CUDA)
+         if (err == cudaSuccess)
+         {
+            return;
+         }
+#elif defined(MFEM_USE_HIP)
+         if (err == hipSuccess)
+         {
+            return;
+         }
+#endif
+      }
+      // try allocating a larger buffer
+      bytes = 0;
+      // get size of buffer
+      MFEM_GPU_CHECK(MFEM_CUB_NAMESPACE::DeviceScan::InclusiveSum(
+                        nullptr, bytes, d_in, d_out, num_items));
+      // resize buffer (in ScanWorkspace::Get) and try again
+      MFEM_GPU_CHECK(MFEM_CUB_NAMESPACE::DeviceScan::InclusiveSum(
+                        ScanWorkspace::Get(bytes), bytes, d_in, d_out, num_items));
+      return;
+   }
+#endif
+#if 0
+   std::inclusive_scan(d_in, d_in + num_items, d_out);
+#else
+   // work-around to some compilers not fully supporting C++17
+   if (num_items)
+   {
+      *d_out = *d_in;
+      auto prev = d_out;
+      ++d_in;
+      ++d_out;
+      for (size_t i = 1; i < num_items; ++i)
+      {
+         *d_out = (*prev) + (*d_in);
+         prev = d_out;
+         ++d_in;
+         ++d_out;
+      }
+   }
+#endif
+}
+
+/// @brief Performs an inclusive scan of [d_in, d_in+num_items) -> [d_out,
+/// d_out+num_items). This call is potentially asynchronous on the device.
+///
+/// @a d_in input start.
+/// @a d_out output start. Can perform in-place scans with d_out = d_in
+/// @a scan_op binary scan functor. Must be associative. If only weakly
+/// associative (i.e. floating point addition) results are not deterministic. On
+/// device this must also be commutative.
+template <class InputIt, class OutputIt, class ScanOp>
+void InclusiveScan(bool use_dev, InputIt d_in, OutputIt d_out, size_t num_items,
+                   ScanOp scan_op)
+{
+#if defined(MFEM_USE_CUDA) || defined(MFEM_USE_HIP)
+   if (use_dev && mfem::Device::Allows(Backend::CUDA_MASK | Backend::HIP_MASK))
+   {
+      using internal::ScanWorkspace;
+      size_t bytes = ScanWorkspace::Size();
+      if (bytes > 0)
+      {
+         auto err = MFEM_CUB_NAMESPACE::DeviceScan::InclusiveScan(
+                       ScanWorkspace::Get(bytes), bytes, d_in, d_out, scan_op,
+                       num_items);
+#if defined(MFEM_USE_CUDA)
+         if (err == cudaSuccess)
+         {
+            return;
+         }
+#elif defined(MFEM_USE_HIP)
+         if (err == hipSuccess)
+         {
+            return;
+         }
+#endif
+      }
+      // try allocating a larger buffer
+      bytes = 0;
+      MFEM_GPU_CHECK(MFEM_CUB_NAMESPACE::DeviceScan::InclusiveScan(
+                        nullptr, bytes, d_in, d_out, scan_op, num_items));
+      MFEM_GPU_CHECK(MFEM_CUB_NAMESPACE::DeviceScan::InclusiveScan(
+                        ScanWorkspace::Get(bytes), bytes, d_in, d_out, scan_op,
+                        num_items));
+      return;
+   }
+#endif
+#if 0
+   std::inclusive_scan(d_in, d_in + num_items, d_out, scan_op);
+#else
+   // work-around to some compilers not fully supporting C++17
+   if (num_items)
+   {
+      *d_out = *d_in;
+      auto prev = d_out;
+      ++d_in;
+      ++d_out;
+      for (size_t i = 1; i < num_items; ++i)
+      {
+         *d_out = scan_op(*prev, *d_in);
+         prev = d_out;
+         ++d_in;
+         ++d_out;
+      }
+   }
+#endif
+}
+
+/// Performs an exclusive scan of [d_in, d_in+num_items) -> [d_out,
+/// d_out+num_items). This call is potentially asynchronous on the device.
+/// @a d_in input start.
+/// @a d_out output start. Can perform in-place scans with d_out = d_in
+/// @a scan_op binary scan functor. Must be associative. If only weakly
+/// associative (i.e. floating point addition) results are not deterministic. On
+/// device this must also be commutative.
+template <class InputIt, class OutputIt, class T, class ScanOp>
+void ExclusiveScan(bool use_dev, InputIt d_in, OutputIt d_out, size_t num_items,
+                   T init_value, ScanOp scan_op)
+{
+#if defined(MFEM_USE_CUDA) || defined(MFEM_USE_HIP)
+   if (use_dev && mfem::Device::Allows(Backend::CUDA_MASK | Backend::HIP_MASK))
+   {
+      using internal::ScanWorkspace;
+      size_t bytes = ScanWorkspace::Size();
+      if (bytes)
+      {
+         auto err = MFEM_CUB_NAMESPACE::DeviceScan::ExclusiveScan(
+                       ScanWorkspace::Get(bytes), bytes, d_in, d_out, scan_op,
+                       init_value, num_items);
+#if defined(MFEM_USE_CUDA)
+         if (err == cudaSuccess)
+         {
+            return;
+         }
+#elif defined(MFEM_USE_HIP)
+         if (err == hipSuccess)
+         {
+            return;
+         }
+#endif
+      }
+      // try allocating a larger buffer
+      bytes = 0;
+      MFEM_GPU_CHECK(MFEM_CUB_NAMESPACE::DeviceScan::ExclusiveScan(
+                        nullptr, bytes, d_in, d_out, scan_op, init_value, num_items));
+      MFEM_GPU_CHECK(MFEM_CUB_NAMESPACE::DeviceScan::ExclusiveScan(
+                        ScanWorkspace::Get(bytes), bytes, d_in, d_out, scan_op,
+                        init_value, num_items));
+      return;
+   }
+#endif
+#if 0
+   std::exclusive_scan(d_in, d_in + num_items, d_out, init_value, scan_op);
+#else
+   // work-around to some compilers not fully supporting C++17
+   if (num_items)
+   {
+      for (size_t i = 0; i < num_items; ++i)
+      {
+         auto next = scan_op(init_value, *d_in);
+         *d_out = init_value;
+         init_value = next;
+         ++d_out;
+         ++d_in;
+      }
+   }
+#endif
+}
+
+/// Equivalent to ExclusiveScan(use_dev, d_in, d_out, num_items, init_value,
+/// std::plus<>{})
+template <class InputIt, class OutputIt, class T>
+void ExclusiveScan(bool use_dev, InputIt d_in, OutputIt d_out, size_t num_items,
+                   T init_value)
+{
+   ExclusiveScan(use_dev, d_in, d_out, num_items, init_value, std::plus<> {});
+}
+
+} // namespace mfem
+
+#undef MFEM_CUB_NAMESPACE
+
+#endif
@@ -68,7 +68,7 @@ void MagmaBatchedLinAlg::AddMult(const DenseTensor &A, const Vector &x,
   auto d_x = x.Read(); // Shape (n, k, n_mat);
   auto d_y = beta == 0.0 ? y.Write() : y.ReadWrite(); // Shape (m, k, n_mat);

-   magma_trans_t magma_op = tr ? MagmaNoTrans : MagmaTrans;
+   magma_trans_t magma_op = tr ? MagmaTrans : MagmaNoTrans;

   MFEM_MAGMABLAS_PREFIX(gemm_batched_strided)(
      magma_op, MagmaNoTrans, m, k, n, alpha, d_A, m, m*n, d_x, n, n*k,
@@ -167,7 +167,7 @@ void MagmaBatchedLinAlg::Invert(DenseTensor &A) const
   magma_int_t status;

   status = MFEM_MAGMA_PREFIX(getrf_batched)(
-               n, n, d_A_ptrs, n, d_P_ptrs, info_array.Write(), n_mat,
+               n, n, d_LU_ptrs, n, d_P_ptrs, info_array.Write(), n_mat,
               Magma::Queue());
   MFEM_VERIFY(status == MAGMA_SUCCESS, "");

@@ -561,7 +561,8 @@ void CopyMemory(Memory<T> &src, Memory<T> &dst, MemoryClass dst_mc,
    this function. In particular, @a dst should be empty or deleted before
    calling this function. */
 template <typename SrcT, typename DstT>
-void CopyConvertMemory(Memory<SrcT> &src, MemoryClass dst_mc, Memory<DstT> &dst)
+void CopyConvertMemory(const Memory<SrcT> &src, MemoryClass dst_mc,
+                       Memory<DstT> &dst)
 {
   auto capacity = src.Capacity();
   dst.New(capacity, GetMemoryType(dst_mc));
@@ -842,8 +843,8 @@ static int GetPartitioningArraySize(MPI_Comm comm)
 ///
 /// Both @a row and @a col are partitioning arrays, whose length is returned by
 /// GetPartitioningArraySize(), see @ref hypre_partitioning_descr.
-static bool RowAndColStartsAreEqual(MPI_Comm comm, HYPRE_BigInt *rows,
-                                    HYPRE_BigInt *cols)
+static bool RowAndColStartsAreEqual(MPI_Comm comm, const HYPRE_BigInt *rows,
+                                    const HYPRE_BigInt *cols)
 {
   const int part_size = GetPartitioningArraySize(comm);
   bool are_equal = true;
@@ -1131,7 +1132,7 @@ HypreParMatrix::HypreParMatrix(
 HypreParMatrix::HypreParMatrix(MPI_Comm comm,
                               HYPRE_BigInt *row_starts,
                               HYPRE_BigInt *col_starts,
-                               SparseMatrix *sm_a)
+                               const SparseMatrix *sm_a)
 {
   MFEM_ASSERT(sm_a != NULL, "invalid input");
   MFEM_VERIFY(!HYPRE_AssumedPartitionCheck(),
@@ -1145,7 +1146,7 @@ HypreParMatrix::HypreParMatrix(MPI_Comm comm,

   hypre_CSRMatrixSetDataOwner(csr_a,0);
   MemoryIJData mem_a;
-   CopyCSR(sm_a, mem_a, csr_a, false);
+   CopyCSR(const_cast<SparseMatrix*>(sm_a), mem_a, csr_a, false);
   hypre_CSRMatrixSetRownnz(csr_a);

   // NOTE: this call creates a matrix on host even when device support is
@@ -1307,10 +1308,11 @@ HypreParMatrix::HypreParMatrix(MPI_Comm comm, int id, int np,
 HypreParMatrix::HypreParMatrix(MPI_Comm comm, int nrows,
                               HYPRE_BigInt glob_nrows,
                               HYPRE_BigInt glob_ncols,
-                               int *I, HYPRE_BigInt *J,
-                               real_t *data,
-                               HYPRE_BigInt *rows,
-                               HYPRE_BigInt *cols)
+                               const int *I,
+                               const HYPRE_BigInt *J,
+                               const real_t *data,
+                               const HYPRE_BigInt *rows,
+                               const HYPRE_BigInt *cols)
 {
   Init();

@@ -2327,8 +2329,8 @@ void HypreParMatrix::Threshold(real_t threshold)
   /* TODO: GenerateDiagAndOffd() uses an int array of size equal to the number
      of columns in csr_A_wo_z which is the global number of columns in A. This
      does not scale well. */
-   ierr += GenerateDiagAndOffd(csr_A_wo_z,parcsr_A_ptr,
-                               col_start,col_end);
+   ierr += hypre_GenerateDiagAndOffd(csr_A_wo_z,parcsr_A_ptr,
+                                     col_start,col_end);

   ierr += hypre_CSRMatrixDestroy(csr_A_wo_z);

@@ -25,11 +25,18 @@
 #define HYPRE_TIMING

 // hypre header files
+#if MFEM_HYPRE_VERSION < 30000
 #include <seq_mv.h>
 #include <temp_multivector.h>
+#else
+#include <_hypre_seq_mv.h>
+#include <_hypre_lobpcg_temp_multivector.h>
+#endif
 #include <_hypre_parcsr_mv.h>
 #include <_hypre_parcsr_ls.h>

+#include <HYPRE_parcsr_ls.h>
+
 #ifdef HYPRE_COMPLEX
 #error "MFEM does not work with HYPRE's complex numbers support"
 #endif
@@ -53,6 +60,10 @@
 #error "MFEM_USE_HIP=YES is required when HYPRE is built with HIP!"
 #endif

+#if MFEM_HYPRE_VERSION > 21500
+#define HYPRE_AssumedPartitionCheck() 1
+#endif
+
 namespace mfem
 {

@@ -554,7 +565,7 @@ public:
       partitioning arrays @a row_starts and @a col_starts. */
   HypreParMatrix(MPI_Comm comm, HYPRE_BigInt *row_starts,
                  HYPRE_BigInt *col_starts,
-                  SparseMatrix *a); // constructor with 4 arguments, v2
+                  const SparseMatrix *a); // constructor with 4 arguments, v2

   /// Creates boolean block-diagonal rectangular parallel matrix.
   /** The new HypreParMatrix does not take ownership of any of the input
@@ -583,9 +594,9 @@ public:
       arrays (so they can be deleted). See @ref hypre_partitioning_descr "here"
       for a description of the partitioning arrays @a rows and @a cols. */
   HypreParMatrix(MPI_Comm comm, int nrows, HYPRE_BigInt glob_nrows,
-                  HYPRE_BigInt glob_ncols, int *I, HYPRE_BigInt *J,
-                  real_t *data, HYPRE_BigInt *rows,
-                  HYPRE_BigInt *cols); // constructor with 9 arguments
+                  HYPRE_BigInt glob_ncols, const int *I, const HYPRE_BigInt *J,
+                  const real_t *data, const HYPRE_BigInt *rows,
+                  const HYPRE_BigInt *cols); // constructor with 9 arguments

   /** @brief Copy constructor for a ParCSR matrix which creates a deep copy of
       structure and data from @a P. */
@@ -1916,9 +1916,9 @@ hypre_ParCSRMatrixAdd(hypre_ParCSRMatrix *A,
      /* FIXME: GenerateDiagAndOffd() uses an int array of size equal to the
         number of columns in csr_C_temp which is the global number of columns
         in A and B. This does not scale well. */
-      ierr += GenerateDiagAndOffd(csr_C_temp, C,
-                                  hypre_ParCSRMatrixFirstColDiag(A),
-                                  hypre_ParCSRMatrixLastColDiag(A));
+      ierr += hypre_GenerateDiagAndOffd(csr_C_temp, C,
+                                        hypre_ParCSRMatrixFirstColDiag(A),
+                                        hypre_ParCSRMatrixLastColDiag(A));

      /* delete CSR version of C */
      ierr += hypre_CSRMatrixDestroy(csr_C_temp);
@@ -21,6 +21,10 @@
 // hypre header files
 #include <_hypre_parcsr_mv.h>

+#if MFEM_HYPRE_VERSION < 30000
+#define hypre_GenerateDiagAndOffd GenerateDiagAndOffd
+#endif
+
 // Older hypre versions do not define HYPRE_BigInt and HYPRE_MPI_BIG_INT, so we
 // define them here for backward compatibility.
 #if MFEM_HYPRE_VERSION < 21600
@@ -1019,7 +1019,7 @@ MMA::MMA(MPI_Comm comm_, int nVar, int nCon, real_t *xval, int iter)
   mSubProblem.reset(new MMA::MMASubSvanberg(*this, nVar, nCon));
 }

-MMA::MMA(MPI_Comm comm_, const int & nVar, const int & nCon,
+MMA::MMA(MPI_Comm comm_, const int nVar, const int nCon,
         const Vector & xval, int iter) : MMA(comm_, nVar, nCon, xval.GetData(), iter)
 {}
 #endif
@@ -25,25 +25,42 @@ namespace mfem
 // forward declaration
 class Vector;

-/** \brief MMA (Method of Moving Asymptotes) solves an optimization problem
- *         of the form:
+/** \brief MMA (Method of Moving Asymptotes) solves a nonlinear optimization
+ *         problem involving an objective function, inequality constraints,
+ *         and variable bounds.
 *
- *    Find       x that minimizes the objective function F(x),
- *    subject to C(x)_i <= 0,         for all i = 1, ... m
- *               x_lo <= x <= x_hi.
+ *  \details
+ *  This class finds ${\bf x} \in R^n$ that solves the following nonlinear
+ *  program:
+ *   $$
+ *    \begin{array}{ll}
+ *    \min_{{\bf x} \in R^n} & F({\bf x})\\
+ *     \textrm{subject to}   & C({\bf x})_i \leq 0,\quad
+ *                             \textrm{for all}\quad i = 1,\ldots m\\
+ *                           & {\bf x}_{\textrm{lo}} \leq {\bf x} \leq
+ *                             {\bf x}_{\textrm{hi}}.
+ *    \end{array}
+ *   $$
+ *   Here $F : R^n \to R$ is the objective function, and
+ *   $C : R^n \to R^m$ is a set of $m$ inequality constraints. The
+ *   variable bounds are sometimes called box constraints. By
+ *   convention, the routine seeks ${\bf x}$ that minimizes the
+ *   objective function, $F$. Maximization problems should be
+ *   reformulated as a minimization of $-F$.
 *
 *    The objective functions are replaced by convex functions
 *    chosen based on gradient information, and solved using a dual method.
 *    The unique optimal solution of this subproblem is returned as the next
 *    iteration point. Optimality is determined by the KKT conditions.
 *
- *  The "Update" function in MMA advances the optimization and must be called
- *  in every optimization iteration. Current and previous iteration points
- *  construct the "moving asymptotes". The design variables, objective function,
- *  constraints are passed to an approximating subproblem. The design variables
- *  are updated and returned. Its implementation closely follows the original
- *  formulation of 'Svanberg, K. (2007). MMA and GCMMA-two methods
- *  for nonlinear optimization. vol, 1, 1-15.'
+ *  The "Update" function in MMA advances the optimization and must be
+ *  called in every optimization iteration. Current and previous iteration
+ *  points construct the "moving asymptotes". The design variables,
+ *  objective function, constraints are passed to an approximating
+ *  subproblem. The design variables are updated and returned. Its
+ *  implementation closely follows the original formulation of <a
+ *  href="https://people.kth.se/~krille/mmagcmma.pdf">'Svanberg, K. (2007).
+ *  MMA and GCMMA-two methods for nonlinear optimization. vol, 1, 1-15.'</a>
 *
 *  When used in parallel, all Vectors are assumed to be true dof vectors,
 *  and the operators are expected to be defined for tdof vectors.
@@ -52,46 +69,164 @@ class Vector;
 class MMA
 {
 public:
-   /// Serial constructor:
-   /// nVar - number of design parameters;
-   /// nCon - number of constraints;
-   /// xval[nVar] - initial parameter values
+   /**
+    * \brief Serial constructor
+    * \param nVar   total number of design parameters
+    * \param nCon   number of inequality constraints (i.e., $C$)
+    * \param xval   initial values for design parameters (a pointer
+    *               to \p nVar doubles). Caller retains ownership of
+    *               this pointer/data.
+    * \param iterationNumber the starting iteration number
+    */
   MMA(int nVar, int nCon, real_t *xval, int iterationNumber = 0);
+
+   /**
+    * \brief Serial constructor
+    * \param nVar   total number of design parameters
+    * \param nCon   number of inequality constraints (i.e., $C$)
+    * \param xval   initial values for design parameters (size should
+    *               be \p nVar). Caller retains ownership of
+    *               this Vector.
+    * \param iterationNumber the starting iteration number
+    */
   MMA(const int nVar, int nCon, Vector & xval, int iterationNumber = 0);

 #ifdef MFEM_USE_MPI
-   /// Parallel constructor:
-   /// comm_ - communicator
+   /**
+    * \brief Parallel constructor
+    * \param comm_  the MPI communicator participating in the NLP solve
+    * \param nVar   number of design parameters on this MPI rank
+    * \param nCon   total number of inequality constraints (i.e., $C$).
+    *               Every MPI rank provides the same value here.
+    * \param xval   initial values for design parameters on this MPI rank
+    *               (a pointer to \p nVar doubles). Caller retains ownership
+    *               of this pointer/data.
+    * \param iterationNumber the starting iteration number. All MPI ranks
+    *               should pass in the same value here.
+    *
+    * \details
+    * Each MPI rank has a subset of the total design variable vector, and
+    * calls for that MPI rank always address its subset of the design
+    * variable vector and gradients with respect to its subset of the design
+    * variable vector.
+    *
+    * If you wanted to determine the global number of design variables, it
+    * would be determined as follows:
+    * \code{.cpp}
+    * int globalDesignVars;
+    * MPI_Allreduce(&nVar, &globalDesignVars, 1, MPI_INT, MPI_SUM, comm_);
+    * \endcode
+    */
   MMA(MPI_Comm comm_, int nVar, int nCon, real_t *xval,
       int iterationNumber = 0);
-   MMA(MPI_Comm comm_, const int & nVar, const int & nCon, const Vector & xval,
+   /**
+    * \brief Parallel constructor
+    * \param comm_  the MPI communicator participating in the NLP solve
+    * \param nVar   number of design parameters on this MPI rank
+    * \param nCon   total number of inequality constraints (i.e., $C$).
+    *               Every MPI rank provides the same value here.
+    * \param xval   initial values for design parameters (size should
+    *               be \p nVar). Caller retains ownership of
+    *               this Vector.
+    * \param iterationNumber the starting iteration number. All MPI ranks
+    *               should pass in the same value here.
+    *
+    * \details
+    * Each MPI rank has a subset of the total design variable vector, and
+    * calls for that MPI rank always address its subset of the design
+    * variable vector and gradients with respect to its subset of the design
+    * variable vector.
+    *
+    * If you wanted to determine the global number of design variables, it
+    * would be determined as follows:
+    * \code{.cpp}
+    * int globalDesignVars;
+    * MPI_Allreduce(&nVar, &globalDesignVars, 1, MPI_INT, MPI_SUM, comm_);
+    * \endcode
+    */
+   MMA(MPI_Comm comm_, const int nVar, const int nCon, const Vector & xval,
       int iterationNumber = 0);
 #endif

   /// Destructor
   ~MMA();

-   /// Update the optimization parameters
-   /// dfdx[nVar] - gradients of the objective
-   /// gx[nCon] - values of the constraints
-   /// dgdx[nCon*nVar] - gradients of the constraints ordered
-   ///                   constraint by constraint, e.g. {dg0dx0, dg0dx1, ... ,}
-   ///                                                  {dg1dx0, dg1dx1, ... ,}
-   /// xmin[nVar] - lower bounds
-   /// xmax[nVar] - upper bounds
-   /// xval[nVar] - input/output for optimization parameters
+   /**
+    * \brief Update the optimization parameters for a constrained
+    *        nonlinear program
+    * \param dfdx    vector of size nVar holding the gradients of the
+    *                objective function with respect to
+    *                the design variables,
+    *                $\frac{\partial F}{\partial {\bf x}_i}$
+    *                for each variable on this rank.
+    * \param gx      vector of size nCon holding the values of the
+    *                inequality constraints. Every MPI rank should
+    *                pass in the same values here.
+    * \param dgdx    vector of size $\textrm{nCon}\cdot\textrm{nVar}$
+    *                holding the gradients of the constraints in
+    *                row-major order. For example, {dg0dx0, dg0dx1, ...,}
+    *                {dg1dx0, dg1dx1, ..., }, ...
+    * \param xmin    vector of size nVar holding the lower bounds on
+    *                the design values. \p xmin and \p xmax are
+    *                the box constraints.
+    * \param xmax    vector of size nVar holding the upper bounds on
+    *                the design values.  \p xmin and \p xmax are
+    *                the box constraints.
+    * \param xval    vector of size nVar. On entry, this holds the
+    *                value of the design variables where the objective,
+    *                constraints, and their gradients were evaluated.
+    *                On exit, this holds the result of the MMA iteration,
+    *                the next design variable value to use.
+    *
+    * \details
+    * The caller retains ownership of all Vectors passed into this method.
+    */
   void Update(const Vector& dfdx,
               const Vector& gx, const Vector& dgdx,
               const Vector& xmin, const Vector& xmax,
               Vector& xval);
-   /// Unconstrained
+
+   /**
+    * \brief Update the optimization parameters for an unconstrained
+    *        nonlinear program
+    * \param dfdx    vector of size nVar holding the gradients of the
+    *                objective function with respect to
+    *                the design variables,
+    *                $\frac{\partial F}{\partial {\bf x}_i}$
+    *                for each variable on this rank.
+    * \param xmin    vector of size nVar holding the lower bounds on
+    *                the design values. \p xmin and \p xmax are
+    *                the box constraints.
+    * \param xmax    vector of size nVar holding the upper bounds on
+    *                the design values.  \p xmin and \p xmax are
+    *                the box constraints.
+    * \param xval    vector of size nVar. On entry, this holds the
+    *                value of the design variables where the objective,
+    *                constraints, and their gradients were evaluated.
+    *                On exit, this holds the result of the MMA iteration,
+    *                the next design variable value to use.
+    *
+    * \details
+    * The caller retains ownership of all Vectors passed into this method.
+    * This should be used when the number of inequality constraints is zero.
+    */
   void Update( const Vector& dfdx,
                const Vector& xmin, const Vector& xmax,
                Vector& xval);

+   /**
+    * \brief Change the iteration number
+    * \param iterationNumber    the new iteration number
+    */
   void SetIteration( int iterationNumber ) { iter = iterationNumber; };
-   int GetIteration() { return iter; };

+   /// Return the current iteration number
+   int GetIteration() const { return iter; };
+
+   /**
+    * \brief change the print level
+    * \param print_lvl    the new print level
+    */
   void SetPrintLevel(int print_lvl) { print_level = print_lvl; }

 protected:
@@ -123,7 +258,7 @@ private:
   /// KKT norm
   real_t kktnorm;

-   /// intialization state
+   /// initialization state
   bool isInitialized = false;

 #ifdef MFEM_USE_MPI
@@ -1356,6 +1356,7 @@ void PetscParMatrix::MakeWrapper(MPI_Comm comm, const Operator* op, Mat *A)
                      PETSC_DECIDE,PETSC_DECIDE); PCHKERRQ(A,ierr);
   ierr = MatSetType(*A,MATSHELL); PCHKERRQ(A,ierr);
   ierr = MatShellSetContext(*A,(void *)op); PCHKERRQ(A,ierr);
+#if PETSC_VERSION_LT(3,24,0)
   ierr = MatShellSetOperation(*A,MATOP_MULT,
                               (void (*)())__mfem_mat_shell_apply);
   PCHKERRQ(A,ierr);
@@ -1367,6 +1368,19 @@ void PetscParMatrix::MakeWrapper(MPI_Comm comm, const Operator* op, Mat *A)
   PCHKERRQ(A,ierr);
   ierr = MatShellSetOperation(*A,MATOP_DESTROY,
                               (void (*)())__mfem_mat_shell_destroy);
+#else
+   ierr = MatShellSetOperation(*A,MATOP_MULT,
+                               (PetscErrorCodeFn*)__mfem_mat_shell_apply);
+   PCHKERRQ(A,ierr);
+   ierr = MatShellSetOperation(*A,MATOP_MULT_TRANSPOSE,
+                               (PetscErrorCodeFn*)__mfem_mat_shell_apply_transpose);
+   PCHKERRQ(A,ierr);
+   ierr = MatShellSetOperation(*A,MATOP_COPY,
+                               (PetscErrorCodeFn*)__mfem_mat_shell_copy);
+   PCHKERRQ(A,ierr);
+   ierr = MatShellSetOperation(*A,MATOP_DESTROY,
+                               (PetscErrorCodeFn*)__mfem_mat_shell_destroy);
+#endif
 #if defined(_USE_DEVICE)
   MemoryType mt = GetMemoryType(op->GetMemoryClass());
   if (mt == MemoryType::DEVICE || mt == MemoryType::MANAGED)
--- a/Show More
+++ b/Show More