add optional MemoryType to the SparseMatrix copy-ctor

Cherry-picking changes on top of our current hash.
There are things that hurt performance in laster mfem@master commits, not sure what yet. This includes: - feature/artv3/cusparse-Spmv - feature/tomstitt/temp-mem-type - patches to tmop.cpp and tmop_tools.cpp to address memory issues - patch to mem_manager/device to use mfem's default allcator instead of the host umpire one
2020-07-24 09:13:21 -07:00 · 2020-07-22 15:51:04 -07:00
231 changed files with 4488 additions and 25932 deletions
@@ -15,10 +15,8 @@ install:
 - msmpisdk.msi /passive
 - set PATH=C:\Program Files\Microsoft MPI\Bin;%PATH%

-# Install METIS, use a mirror because the original source server is not always
-# up. Original url:
-#    http://glaros.dtc.umn.edu/gkhome/fetch/sw/metis/metis-5.1.0.tar.gz
- ps: Start-FileDownload 'https://mfem.github.io/tpls/metis-5.1.0.tar.gz'
+# Install METIS
+- ps: Start-FileDownload 'http://glaros.dtc.umn.edu/gkhome/fetch/sw/metis/metis-5.1.0.tar.gz'
 - 7z x metis-5.1.0.tar.gz -so | 7z x -si -ttar > nul
 - cd metis-5.1.0
 - ps: ( get-content "GKlib\gk_arch.h") | % { If ($_.ReadCount -ge 52) {$_ -replace "#ifdef __MSC__","#ifdef DISABLE_THIS_ANCIENT_MSC_CHECK"} Else {$_} } | set-content "GKlib\gk_arch.h"
@@ -28,17 +26,17 @@ install:
 - cd ..

 # Install hypre
- ps: Start-FileDownload 'https://github.com/hypre-space/hypre/archive/v2.19.0.tar.gz'
- 7z x v2.19.0.tar.gz -so | 7z x -si -ttar > nul
- cd hypre-2.19.0/src
- cmake -H. -Bbuild -DMPI_C_INCLUDE_PATH="C:\Program Files (x86)\Microsoft SDKs\MPI\Include" -DMPI_C_LIBRARIES="C:\Program Files (x86)\Microsoft SDKs\MPI\Lib\x86\msmpi.lib" -DMPI_CXX_LIBRARIES="C:\Program Files (x86)\Microsoft SDKs\MPI\Lib\x86\msmpi.lib" -DMPI_CXX_INCLUDE_PATH="C:\Program Files (x86)\Microsoft SDKs\MPI\Include"
+- ps: Start-FileDownload 'https://github.com/hypre-space/hypre/archive/V2-10-0b.tar.gz'
+- 7z x V2-10-0b.tar.gz -so | 7z x -si -ttar > nul
+- cd hypre-2-10-0b
+- cmake -H. -Bbuild -DHYPRE_USING_FEI=OFF -DMPI_C_INCLUDE_PATH="C:\Program Files (x86)\Microsoft SDKs\MPI\Include" -DMPI_C_LIBRARIES="C:\Program Files (x86)\Microsoft SDKs\MPI\Lib\x86\msmpi.lib" -DMPI_CXX_LIBRARIES="C:\Program Files (x86)\Microsoft SDKs\MPI\Lib\x86\msmpi.lib" -DMPI_CXX_INCLUDE_PATH="C:\Program Files (x86)\Microsoft SDKs\MPI\Include"
 - cmake --build build
 - cmake --build build --target install
- cd ../..
+- cd ..

 # MFEM
 before_build:
- cmake -H. -DCMAKE_INSTALL_PREFIX=install -Bbuild_parallel -DMFEM_USE_MPI=TRUE -DMFEM_USE_METIS_5=TRUE -DMPI_CXX_LIBRARIES="C:\Program Files (x86)\Microsoft SDKs\MPI\Lib\x86\msmpi.lib" -DMPI_CXX_INCLUDE_PATH="C:\Program Files (x86)\Microsoft SDKs\MPI\Include" -DHYPRE_DIR=%cd%\hypre-2.19.0\src\hypre -DMETIS_LIBRARIES=%cd%\metis-5.1.0\build\libmetis\Debug\metis.lib -DMETIS_INCLUDE_DIRS=%cd%\metis-5.1.0\include
+- cmake -H. -DCMAKE_INSTALL_PREFIX=install -Bbuild_parallel -DMFEM_USE_MPI=TRUE -DMFEM_USE_METIS_5=TRUE -DMPI_CXX_LIBRARIES="C:\Program Files (x86)\Microsoft SDKs\MPI\Lib\x86\msmpi.lib" -DMPI_CXX_INCLUDE_PATH="C:\Program Files (x86)\Microsoft SDKs\MPI\Include" -DHYPRE_LIBRARIES=%cd%\hypre-2-10-0b\hypre\lib\HYPRE.lib -DHYPRE_INCLUDE_DIRS=%cd%\hypre-2-10-0b\hypre\include -DHYPRE_VERSION=21000 -DMETIS_LIBRARIES=%cd%\metis-5.1.0\build\libmetis\Debug\metis.lib -DMETIS_INCLUDE_DIRS=%cd%\metis-5.1.0\include
 - cmake -H. -DCMAKE_INSTALL_PREFIX=install -Bbuild_serial -DMFEM_USE_MPI=FALSE

 build_script:
@@ -122,7 +122,7 @@ examples/sundials/ex16-final.*
 examples/sundials/Example16*

 examples/petsc/ex[1-69]p
-examples/petsc/ex1[0-1]p
+examples/petsc/ex10p

 examples/petsc/mesh.*
 examples/petsc/sol.*
@@ -137,7 +137,6 @@ examples/petsc/Example9*
 examples/petsc/deformed.*
 examples/petsc/velocity.*
 examples/petsc/elastic_energy.*
-examples/petsc/mode_*

 examples/pumi/ex1
 examples/pumi/ex[126]p
@@ -244,24 +243,12 @@ miniapps/navier/navier_3dfoc
 miniapps/navier/tgv_out*.txt
 miniapps/navier/*_output

-miniapps/adjoint/cvsRoberts_ASAi_dns
-miniapps/adjoint/adjoint_advection_diffusion
-
 # Unit test binary and outputs
 tests/unit/output_meshes
 tests/unit/unit_tests
 tests/unit/punit_tests
 tests/unit/sedov_tests_*
 tests/unit/psedov_tests_*
-tests/unit/tmop_tests_*
-tests/unit/ptmop_tests_*
-tests/unit/cube.mesh
-tests/unit/star.mesh
-tests/unit/blade.mesh
-tests/unit/square01.mesh
-tests/unit/toroid-hex.mesh
-tests/unit/beam-hex-nurbs.mesh
-tests/unit/square-disc-nurbs.mesh

 # Test script output
 tests/scripts/*.err
@@ -11,20 +11,11 @@

 language: cpp

-os: linux
-dist: bionic
-
 stages:
  - checks
  - tests
  - optional

-env:
-  global:
-    - HYPRE_ARCHIVE=v2.19.0.tar.gz
-      HYPRE_URL=https://github.com/hypre-space/hypre/archive/$HYPRE_ARCHIVE
-      HYPRE_TOP_DIR=hypre-2.19.0
-
 jobs:
  include:

@@ -37,7 +28,6 @@ jobs:

    - stage: checks
      os: linux
-      dist: xenial
      name: "code-style"
      addons:
        apt:
@@ -56,6 +46,9 @@ jobs:
          packages:
            - doxygen
            - graphviz
+            - mpich
+            - libmpich-dev
+      env: MPI=YES
      script:
        - cd ${TRAVIS_BUILD_DIR}
        - cd tests/scripts
@@ -70,24 +63,13 @@ jobs:
            - mpich
            - libmpich-dev
      env: MPI=YES
-      before_script:
+      script:
        - cd ${TRAVIS_BUILD_DIR}
-        - mpicxx -v
        - make config MFEM_USE_MPI=YES MFEM_MPI_NP=2
        - make all -j3
        - make test-noclean
-      script:
        - cd tests/scripts
        - ./runtest gitignore
-      cache:
-        ccache: true
-        directories:
-          - $TRAVIS_BUILD_DIR/../$HYPRE_TOP_DIR/src/hypre
-          - $TRAVIS_BUILD_DIR/../metis-4.0
-      before_cache:
-        - cd $TRAVIS_BUILD_DIR/../metis-4.0;
-          mv libmetis.a Lib ..; rm -rf * ; mv ../libmetis.a ../Lib .;
-          rm -f Lib/*.{c,o}

    # ========================
    #  Optional Checks/Tests
@@ -96,7 +78,6 @@ jobs:

    - stage: optional
      name: "branch-history"
-      if: branch != next
      # need full git history for the binary/big files check
      git:
        depth: false
@@ -125,8 +106,6 @@ jobs:
           MPI=NO
           CODECOV=NO
           MFEM_TEST_TARGET=check
-      cache:
-        ccache: true

    - os: linux
      compiler: gcc
@@ -135,8 +114,6 @@ jobs:
           MPI=NO
           CODECOV=NO
           MFEM_TEST_TARGET=test
-      cache:
-        ccache: true

    - os: linux
      compiler: gcc
@@ -160,9 +137,9 @@ jobs:
           MFEM_TEST_TARGET=check
           NPROCS=2
      cache:
-        ccache: true
        directories:
-          - $TRAVIS_BUILD_DIR/../$HYPRE_TOP_DIR/src/hypre
+          - $TRAVIS_BUILD_DIR/../hypre-2.10.0b/src/hypre/lib
+          - $TRAVIS_BUILD_DIR/../hypre-2.10.0b/src/hypre/include
          - $TRAVIS_BUILD_DIR/../metis-4.0
      before_cache:
        - cd $TRAVIS_BUILD_DIR/../metis-4.0;
@@ -191,9 +168,9 @@ jobs:
           MFEM_TEST_TARGET=test
           NPROCS=2
      cache:
-        ccache: true
        directories:
-          - $TRAVIS_BUILD_DIR/../$HYPRE_TOP_DIR/src/hypre
+          - $TRAVIS_BUILD_DIR/../hypre-2.10.0b/src/hypre/lib
+          - $TRAVIS_BUILD_DIR/../hypre-2.10.0b/src/hypre/include
          - $TRAVIS_BUILD_DIR/../metis-4.0
      before_cache:
        - cd $TRAVIS_BUILD_DIR/../metis-4.0;
@@ -216,16 +193,16 @@ jobs:
        - cd ${TRAVIS_BUILD_DIR}/build
        - cmake ..
             -DMFEM_USE_MPI=ON
-             -DHYPRE_DIR=${TRAVIS_BUILD_DIR}/../$HYPRE_TOP_DIR/src/hypre
+             -DHYPRE_DIR=${TRAVIS_BUILD_DIR}/../hypre-2.10.0b/src/hypre
             -DMFEM_MPI_NP=$NPROCS
        - make -j3 mfem examples
        - cd ${TRAVIS_BUILD_DIR}/build/tests/unit
        - make -j3
        - ctest --output-on-failure
      cache:
-        ccache: true
        directories:
-          - $TRAVIS_BUILD_DIR/../$HYPRE_TOP_DIR/src/hypre
+          - $TRAVIS_BUILD_DIR/../hypre-2.10.0b/src/hypre/lib
+          - $TRAVIS_BUILD_DIR/../hypre-2.10.0b/src/hypre/include
          - $TRAVIS_BUILD_DIR/../metis-4.0
      before_cache:
        - cd $TRAVIS_BUILD_DIR/../metis-4.0;
@@ -241,43 +218,27 @@ jobs:
    # - parallel

    - os: osx
-      osx_image: xcode11.2
+      # osx_image: xcode7.3
      compiler: clang
      name: "Mac: Serial + Debug"
-      addons:
-        homebrew:
-          packages:
-            - ccache
      env: DEBUG=YES
           MPI=NO
           CODECOV=NO
           MFEM_TEST_TARGET=check
-      cache:
-        ccache: true

    - os: osx
-      osx_image: xcode11.2
+      # osx_image: xcode7.3
      compiler: clang
      name: "Mac: Serial"
-      addons:
-        homebrew:
-          packages:
-            - ccache
      env: DEBUG=NO
           MPI=NO
           CODECOV=NO
           MFEM_TEST_TARGET=test
-      cache:
-        ccache: true

    - os: osx
-      osx_image: xcode11.2
+      # osx_image: xcode7.3
      compiler: clang
      name: "Mac: Parallel + Debug"
-      addons:
-        homebrew:
-          packages:
-            - ccache
      env: DEBUG=YES
           MPI=YES
           CODECOV=NO
@@ -285,9 +246,9 @@ jobs:
           NPROCS=4
           TMPDIR=/tmp
      cache:
-        ccache: true
        directories:
-          - $TRAVIS_BUILD_DIR/../$HYPRE_TOP_DIR/src/hypre
+          - $TRAVIS_BUILD_DIR/../hypre-2.10.0b/src/hypre/lib
+          - $TRAVIS_BUILD_DIR/../hypre-2.10.0b/src/hypre/include
          - $TRAVIS_BUILD_DIR/../metis-4.0
          - $HOME/local-cached
      before_cache:
@@ -296,13 +257,9 @@ jobs:
          rm -f Lib/*.{c,o}

    - os: osx
-      osx_image: xcode11.2
+      # osx_image: xcode7.3
      compiler: clang
      name: "Mac: Parallel"
-      addons:
-        homebrew:
-          packages:
-            - ccache
      env: DEBUG=NO
           MPI=YES
           CODECOV=YES
@@ -310,9 +267,9 @@ jobs:
           NPROCS=4
           TMPDIR=/tmp
      cache:
-        ccache: true
        directories:
-          - $TRAVIS_BUILD_DIR/../$HYPRE_TOP_DIR/src/hypre
+          - $TRAVIS_BUILD_DIR/../hypre-2.10.0b/src/hypre/lib
+          - $TRAVIS_BUILD_DIR/../hypre-2.10.0b/src/hypre/include
          - $TRAVIS_BUILD_DIR/../metis-4.0
          - $HOME/local-cached
      before_cache:
@@ -326,19 +283,14 @@ before_install:
   #      brew install open-mpi;
   #   fi

-   # Disable ccache while building dependencies that are cached:
-   - echo "before \$PATH = $PATH";
-     export PATH=${PATH//\/usr\/lib\/ccache:/};
-     echo "after  \$PATH = $PATH"
-
-   # On Mac OS X, build and cache OpenMPI 2.1.6:
+   # On Mac OS X, build and cache OpenMPI 2.1.1:
   - if [ $TRAVIS_OS_NAME == "osx" ] && [ $MPI == "YES" ]; then
        if [ ! -e $HOME/local-cached/bin/mpicc ]; then
           mkdir -p $HOME/builds && cd $HOME/builds &&
-           wget https://download.open-mpi.org/release/open-mpi/v2.1/openmpi-2.1.6.tar.bz2 &&
-           tar jxf openmpi-2.1.6.tar.bz2 &&
+           wget https://www.open-mpi.org/software/ompi/v2.1/downloads/openmpi-2.1.1.tar.bz2 &&
+           tar jxf openmpi-2.1.1.tar.bz2 &&
           mkdir openmpi-build && cd openmpi-build &&
-           ../openmpi-2.1.6/configure --prefix=$HOME/local-cached &&
+           ../openmpi-2.1.1/configure --prefix=$HOME/local-cached &&
           make -j3 all && make install;
        fi;
        PATH=$HOME/local-cached/bin:$PATH;
@@ -383,28 +335,26 @@ install:

   # hypre
   - if [ $MPI == "YES" ]; then
-        if [ ! -e $HYPRE_TOP_DIR/src/hypre/lib/libHYPRE.a ]; then
-           wget $HYPRE_URL;
-           rm -rf $HYPRE_TOP_DIR;
-           tar xvzf $HYPRE_ARCHIVE;
-           cd $HYPRE_TOP_DIR/src;
-           ./configure --disable-fortran CC=mpicc CXX=mpic++;
+        if [ ! -e hypre-2.10.0b/src/hypre/lib/libHYPRE.a ]; then
+           wget https://computation.llnl.gov/project/linear_solvers/download/hypre-2.10.0b.tar.gz --no-check-certificate;
+           rm -rf hypre-2.10.0b;
+           tar xvzf hypre-2.10.0b.tar.gz;
+           cd hypre-2.10.0b/src;
+           ./configure --disable-fortran --without-fei CC=mpicc CXX=mpic++;
           make -j3;
           cd ../..;
        else
-            echo "Reusing cached $HYPRE_TOP_DIR/";
+            echo "Reusing cached hypre-2.10.0b/";
        fi;
-        ln -s $HYPRE_TOP_DIR hypre;
+        ln -s hypre-2.10.0b hypre;
     else
        echo "Serial build, not using hypre";
     fi

-   # METIS, use a mirror because the original source server is not always up.
-   # Original url:
-   #    http://glaros.dtc.umn.edu/gkhome/fetch/sw/metis/OLD/metis-4.0.3.tar.gz
+   # METIS
   - if [ $MPI == "YES" ]; then
        if [ ! -e metis-4.0/libmetis.a ]; then
-           wget https://mfem.github.io/tpls/metis-4.0.3.tar.gz;
+           wget http://glaros.dtc.umn.edu/gkhome/fetch/sw/metis/OLD/metis-4.0.3.tar.gz;
           tar xvzf metis-4.0.3.tar.gz;
           make -j3 -C metis-4.0.3/Lib CC="$CC" OPTFLAGS="-O2";
           rm -rf metis-4.0;
@@ -414,18 +364,6 @@ install:
        fi;
     fi

-   # Re-enable ccache on linux; enable ccache on mac os:
-   - if [ $TRAVIS_OS_NAME == "linux" ]; then
-        export PATH="/usr/lib/ccache:$PATH";
-     else
-        if [ $TRAVIS_OS_NAME == "osx" ]; then
-           export PATH="/usr/local/opt/ccache/libexec:$PATH";
-        fi;
-     fi
-
-   - printf "which \$CC  = "; which $CC;
-     printf "which \$CXX = "; which $CXX
-
 script:
   # Compiler
   - if [ $MPI == "YES" ]; then
@@ -446,9 +384,6 @@ script:
     if [ "$CODECOV" == "YES" ]; then
        CPPFLAGS="--coverage -g";
     fi;
-     if [ "$TRAVIS_OS_NAME" != "linux" ] || [ "$DEBUG" == "YES" ]; then
-        CPPFLAGS+=" -pedantic -Wall -Werror";
-     fi

   # Configure the library
   - make config MFEM_USE_MPI=$MPI MFEM_DEBUG=$DEBUG $MAKE_CXX_FLAG
@@ -33,9 +33,6 @@ Meshing improvements
  for example the periodic-annulus-sector and periodic-torus-sector files in
  the data directory.

- Added complete action of the TMOP Integrator to account for the spatial
-  derivatives of discrete and analytic targets.
-
 Performance improvements
 ------------------------
 - Added support for explicit vectorization in the high-performance templated
@@ -44,26 +41,12 @@ Performance improvements
    - x86 (SSE/AVX/AVX2/AVX512),
    - Power8 & Power9 (VSX),
    - BG/Q (QPX).
-  These are disabled by default, and can be enabled with MFEM_USE_SIMD=YES.
+  These are now enabled by default, and can be disabled with MFEM_USE_SIMD=NO.
  See the new file linalg/simd.hpp and the new directory linalg/simd.

 Improved GPU capabilities
 -------------------------
 - Added support for Chebyshev accelerated polynomial smoother on GPU.
- The TMOP mesh optimization algorithms were extended to GPU:
-   - QualityMetric #1, #2 and #7 are available in 2D, #302, #303 and #321 in 3D
-   - Both AnalyticAdaptTC and DiscreteAdaptTC TargetConstructor are available
-   - Kernels for normalization and limiting have been added
-   - The AdvectorCG now also support AssemblyLevel::PARTIAL
-
- Optimized AMD/HIP kernel support.
-
- Added a Full Assembly mode compatible with Device kernel execution. This
-  assembly level builds on top of the current Element Assembly kernels to
-  compute a global sparse matrix. All integrators supported by element assembly
-  are also supported by full assembly. See the '-fa' option in Example 9.
-
- Added support for BlockOperator on GPU. See the updated Example 5.

 Discretization improvements
 ---------------------------
@@ -88,7 +71,7 @@ Discretization improvements
  and, in the continuous field case, arbitrary mesh edges and faces.

 - Added new coefficient and vector coefficient classes for QuadratureFunctions.
-  Additionally, new LinearForm integrators were also added which make use of
+  Additionaly, new LinearForm integrators were also added which make use of
  these new QuadratureFunction coefficient classes.

 - Added support face integrals on the boundaries of NURBS meshes.
@@ -105,10 +88,6 @@ Linear and nonlinear solvers
  and solution during the solving process of an IterativeSolver after every
  iteration.

- Added support for the CVODES package in SUNDIALS which provides ODE
-  solvers with sensitivity analysis capabilities. See the CVODESSolver
-  class and the new adjoint miniapps below.
-
 - Block arrays of parallel matrices can now be merged into a single parallel
  matrix with the function HypreParMatrixFromBlocks. This could be useful for
  solving block systems with parallel direct solvers such as STRUMPACK.
@@ -116,8 +95,6 @@ Linear and nonlinear solvers
 - In SLISolver, changed the residual inner product from (Br,r) to (Br,Br) so the
  solver can work with non-SPD preconditioner B.

- Added support for the SLEPc eigensolver package.
-
 New and updated examples and miniapps
 -------------------------------------
 - Added a new example, Example 25/25p, to demonstrate the use of a Perfectly
@@ -132,21 +109,6 @@ New and updated examples and miniapps
  for applying Dirichlet, Neumann (both homogeneous and inhomogeneous), Robin,
  and periodic boundary conditions with either H1 or DG discretizations.

- Added a new miniapp, Navier, that solves the time-dependent Navier-Stokes
-  equations of incompressible fluid dynamics. See the miniapps/navier directory
-  for more details.
-
- Added a new miniapps/adjoint directory with two miniapps demonstrating how to
-  solve adjoint problems in MFEM using the CVODES package in SUNDIALS. Both of
-  these miniapps require the MFEM_USE_SUNDIALS configuration option.
-  * The cvsRoberts_ASAi_dns miniapp solves a backward adjoint problem for a
-    system of ODEs, evaluating both forward and adjoint quadratures in serial.
-  * The adjoint_advection_diffusion miniapp solves a backward adjoint problem
-    for an advection diffusion PDE, evaluating adjoint quadratures in parallel.
-
- Ported Example 11p to SLEPc, to demonstrate solving the Laplace eigenvalue
-  equation with the shift-and-invert spectral transformation method.
-
 - Added a simple meshing miniapp, Twist, which demonstrates MFEM's strategy of
  stitching together opposite surfaces of a mesh to create a topologically
  periodic mesh.
@@ -154,13 +116,11 @@ New and updated examples and miniapps
 - Added a new meshing miniapp, Minimal Surface, which solves Plateau's problem:
  the Dirichlet problem for the minimal surface equation.

- Added partial assembly support to Example 4/4p and Example 5/5p, with diagonal
+- Added partial assembly support to examples 4/4p and 5/5p, with diagonal
  preconditioning.

- Added full assembly support in Example 9/9p.
-
- Added a new test problem in Example 24/24p, demonstrating a mixed bilinear
-  form for H1, H(curl), H(div) and L_2, with partial assembly support.
+- Added a new test problem in example 24/24p, demonstrating a mixed bilinear
+  form for H(div) and L_2, with partial assembly support.

 - Added weak Dirichlet boundary conditions (Nitsche) to the NURBS miniapp.

@@ -168,8 +128,6 @@ New and updated examples and miniapps
  mesh based on element attributes. Any newly exposed boundary elements are
  assigned attribute numbers related to the trimmed element attributes.

- Added device support in Example 5/5p.
-
 Improved testing
 ----------------
 - Added a GitLab pipeline that automates PR testing on supercomputing systems
@@ -149,13 +149,9 @@ if (MFEM_USE_MPI)
      message(FATAL_ERROR "PETSc version >= 3.8.0 is required")
    endif()
    set(PETSC_INCLUDE_DIRS ${PETSC_INCLUDES})
-    if (MFEM_USE_SLEPC)
-      find_package(SLEPc REQUIRED config)
-      message(STATUS "Found SLEPc version ${SLEPC_VERSION}")
-    endif()
  endif()
 else()
-  set(PKGS_NEED_MPI SUPERLU PETSC SLEPC STRUMPACK PUMI)
+  set(PKGS_NEED_MPI SUPERLU PETSC STRUMPACK PUMI)
  foreach(PKG IN LISTS PKGS_NEED_MPI)
    if (MFEM_USE_${PKG})
      message(STATUS "Disabling package ${PKG} - requires MPI")
@@ -211,10 +207,10 @@ endif()
 # SUNDIALS
 if (MFEM_USE_SUNDIALS)
  if (NOT MFEM_USE_MPI)
-    find_package(SUNDIALS REQUIRED NVector_Serial CVODES ARKODE KINSOL)
+    find_package(SUNDIALS REQUIRED NVector_Serial CVODE ARKODE KINSOL)
  else()
    find_package(SUNDIALS REQUIRED
-      NVector_Serial NVector_Parallel NVector_ParHyp CVODES ARKODE KINSOL)
+      NVector_Serial NVector_Parallel NVector_ParHyp CVODE ARKODE KINSOL)
  endif()
 endif()

@@ -356,7 +352,7 @@ endif()
 #    integers, the METIS header (with 32-bit indices, as used by mfem) needs to
 #    be before SuiteSparse.
 set(MFEM_TPLS MPI_CXX OPENMP BLAS LAPACK METIS HYPRE SuiteSparse SUNDIALS PETSC
-    SLEPC MESQUITE SuperLUDist STRUMPACK AXOM CONDUIT Ginkgo GNUTLS GSLIB NETCDF
+    MESQUITE SuperLUDist STRUMPACK AXOM CONDUIT Ginkgo GNUTLS GSLIB NETCDF
    MPFR PUMI HIOP POSIXCLOCKS MFEMBacktrace ZLIB OCCA CEED RAJA UMPIRE ADIOS2)
 # Add all *_FOUND libraries in the variable TPL_LIBRARIES.
 set(TPL_LIBRARIES "")
@@ -109,7 +109,6 @@ The MFEM source code has the following structure:
  ├── linalg
  ├── mesh
  ├── miniapps
-  │   ├── adjoint
  │   ├── common
  │   ├── electromagnetics
  │   ├── gslib
@@ -383,10 +383,6 @@ MFEM_USE_PETSC = YES/NO
   and other features based on the PETSc package. When enabled, this option uses
   the PETSC_* library options, see below.

-MFEM_USE_SLEPC = YES/NO
-   Enable MFEM eigensolvers based on the SLEPc package. When enabled, this
-   option uses the SLEPC_* library options, see below.
-
 MFEM_USE_MPFR = YES/NO
   MPFR is a library for multiple-precision floating-point computations. This
   option enables the use of MPFR in MFEM, e.g. for precise computation of 1D
@@ -601,12 +597,6 @@ The specific libraries and their options are:
  Options: PETSC_OPT, PETSC_LIB.
  Versions: PETSc >= 3.8.0.

- SLEPc (optional), used when MFEM_USE_SLEPC = YES. SLEPc depends on PETSc and
-  uses some of the PETSc options when compiled.
-  URL: https://slepc.upv.es/
-  Options: SLEPC_OPT, SLEPC_LIB.
-  Versions: SLEPc >= 3.8.0.
-
 - Sidre (optional), part of LLNL's axom project, used when MFEM_USE_SIDRE = YES.
  Starting with MFEM v4.1, Axom version 0.3.1 or later is required.
  URL: https://github.com/LLNL/axom
@@ -659,11 +649,12 @@ The specific libraries and their options are:
  Options: OCCA_DIR, OCCA_OPT, OCCA_LIB.
  Versions: OCCA >= 1.0.9.

- libCEED (optional), used when MFEM_USE_CEED = YES.
+- libCEED (optional), used when MFEM_USE_CEED = YES. Requires libCEED v0.6
+  or later version, specifically, git-hash 3d05795 or later.
  URL: https://github.com/CEED/libCEED
       https://ceed.exascaleproject.org/libceed
  Options: CEED_DIR, CEED_OPT, CEED_LIB.
-  Versions: libCEED >= 0.6, git-hash a970f63.
+  Versions: libCEED >= 0.6.

 - RAJA (optional), used when MFEM_USE_RAJA = YES.
  Beginning with MFEM v4.1, only RAJA v0.10.0+ is supported.
@@ -244,10 +244,6 @@ IF (DEFINED TPL_ENABLE_PETSC)
  SET(MFEM_USE_PETSC ${TPL_ENABLE_PETSC} CACHE BOOL "Enable PETSc support." FORCE)
 ENDIF()

-IF (DEFINED TPL_ENABLE_SLEPC)
-  SET(MFEM_USE_SLEPC ${TPL_ENABLE_SLEPC} CACHE BOOL "Enable SLEPc support." FORCE)
-ENDIF()
-
 IF (DEFINED TPL_ENABLE_MPFR)
  SET(MFEM_USE_MPFR ${TPL_ENABLE_MPFR} CACHE BOOL "Enable MPFR usage." FORCE)
 ENDIF()
@@ -38,7 +38,6 @@ set(MFEM_USE_GNUTLS @MFEM_USE_GNUTLS@)
 set(MFEM_USE_GSLIB @MFEM_USE_GSLIB@)
 set(MFEM_USE_NETCDF @MFEM_USE_NETCDF@)
 set(MFEM_USE_PETSC @MFEM_USE_PETSC@)
-set(MFEM_USE_SLEPC @MFEM_USE_SLEPC@)
 set(MFEM_USE_MPFR @MFEM_USE_MPFR@)
 set(MFEM_USE_SIDRE @MFEM_USE_SIDRE@)
 set(MFEM_USE_CONDUIT @MFEM_USE_CONDUIT@)
@@ -104,9 +104,6 @@
 // Enable MFEM functionality based on the PETSc library
 #cmakedefine MFEM_USE_PETSC

-// Enable MFEM functionality based on the SLEPc library
-#cmakedefine MFEM_USE_SLEPC
-
 // Enable MFEM functionality based on the Sidre library
 #cmakedefine MFEM_USE_SIDRE

@@ -1,44 +0,0 @@
-# Copyright (c) 2010-2020, 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.
-
-# Sets the following variables:
-#   - SLEPC_FOUND
-#   - SLEPC_INCLUDE_DIRS
-#   - SLEPC_LIBRARIES
-
-set(SLEPc_REQUIRED_PACKAGES "PETSC" CACHE STRING
-  "Additional packages required by SLEPc")
-
-include(MfemCmakeUtilities)
-mfem_find_package(SLEPc SLEPC SLEPC_DIR
-  "include" "slepceps.h"
-  "${PETSC_ARCH}/lib" "slepc" # add NAMES_PER_DIR?
-  "Paths to headers required by SLEPc."
-  "Libraries required by SLEPc."
-  ADD_COMPONENT "config" "${PETSC_ARCH}/include" "slepcconf.h" "" ""
-  CHECK_BUILD SLEPC_VERSION_OK TRUE
-"
-#include \"petsc.h\"
-#include \"slepceps.h\"
-int main()
-{
-  PetscErrorCode ierr;
-  int argc = 0;
-  char** argv = NULL;
-  ierr = SlepcInitialize(&argc, &argv, PETSC_NULL, PETSC_NULL);
-  EPS eps;
-  ierr = EPSCreate(PETSC_COMM_SELF, &eps); CHKERRQ(ierr);
-  ierr = EPSDestroy(&eps); CHKERRQ(ierr);
-  ierr = SlepcFinalize(); CHKERRQ(ierr);
-  return 0;
-}
-"
-  )
@@ -25,6 +25,5 @@ mfem_find_package(SUNDIALS SUNDIALS SUNDIALS_DIR
  ADD_COMPONENT NVector_ParHyp
    "include" nvector/nvector_parhyp.h "lib" sundials_nvecparhyp
  ADD_COMPONENT CVODE "include" cvode/cvode.h "lib" sundials_cvode
-  ADD_COMPONENT CVODES "include" cvodes/cvodes.h "lib" sundials_cvodes
  ADD_COMPONENT ARKODE "include" arkode/arkode.h "lib" sundials_arkode
  ADD_COMPONENT KINSOL "include" kinsol/kinsol.h "lib" sundials_kinsol)
@@ -731,7 +731,7 @@ function(mfem_export_mk_files)
      MFEM_USE_LAPACK MFEM_THREAD_SAFE MFEM_USE_OPENMP MFEM_USE_LEGACY_OPENMP
      MFEM_USE_MEMALLOC MFEM_USE_SUNDIALS MFEM_USE_MESQUITE MFEM_USE_SUITESPARSE
      MFEM_USE_SUPERLU MFEM_USE_STRUMPACK MFEM_USE_GNUTLS
-      MFEM_USE_GSLIB MFEM_USE_NETCDF MFEM_USE_PETSC MFEM_USE_SLEPC MFEM_USE_MPFR MFEM_USE_SIDRE
+      MFEM_USE_GSLIB MFEM_USE_NETCDF MFEM_USE_PETSC MFEM_USE_MPFR MFEM_USE_SIDRE
      MFEM_USE_CONDUIT MFEM_USE_PUMI MFEM_USE_CUDA MFEM_USE_OCCA MFEM_USE_RAJA
      MFEM_USE_UMPIRE MFEM_USE_SIMD MFEM_USE_ADIOS2)
  foreach(var ${CONFIG_MK_BOOL_VARS})
@@ -48,9 +48,6 @@
 #ifdef MFEM_USE_PETSC
 #error Building with PETSc (MFEM_USE_PETSC=YES) requires MPI (MFEM_USE_MPI=YES)
 #endif
-#ifdef MFEM_USE_SLEPC
-#error Building with SLEPc (MFEM_USE_SLEPC=YES) requires MPI (MFEM_USE_MPI=YES)
-#endif
 #ifdef MFEM_USE_PUMI
 #error Building with PUMI (MFEM_USE_PUMI=YES) requires MPI (MFEM_USE_MPI=YES)
 #endif
@@ -118,9 +118,6 @@
 // Enable functionality based on the PETSc library
 // #define MFEM_USE_PETSC

-// Enable functionality based on the SLEPc library
-// #define MFEM_USE_SLEPC
-
 // Enable functionality based on the MPFR library.
 // #define MFEM_USE_MPFR

@@ -37,7 +37,6 @@ MFEM_USE_GINKGO        = @MFEM_USE_GINKGO@
 MFEM_USE_GNUTLS        = @MFEM_USE_GNUTLS@
 MFEM_USE_NETCDF        = @MFEM_USE_NETCDF@
 MFEM_USE_PETSC         = @MFEM_USE_PETSC@
-MFEM_USE_SLEPC         = @MFEM_USE_SLEPC@
 MFEM_USE_MPFR          = @MFEM_USE_MPFR@
 MFEM_USE_SIDRE         = @MFEM_USE_SIDRE@
 MFEM_USE_CONDUIT       = @MFEM_USE_CONDUIT@
@@ -39,7 +39,6 @@ option(MFEM_USE_GNUTLS "Enable GNUTLS usage" OFF)
 option(MFEM_USE_GSLIB "Enable GSLIB usage" OFF)
 option(MFEM_USE_NETCDF "Enable NETCDF usage" OFF)
 option(MFEM_USE_PETSC "Enable PETSc support." OFF)
-option(MFEM_USE_SLEPC "Enable SLEPc support." OFF)
 option(MFEM_USE_MPFR "Enable MPFR usage." OFF)
 option(MFEM_USE_SIDRE "Enable Axom/Sidre usage" OFF)
 option(MFEM_USE_CONDUIT "Enable Conduit usage" OFF)
@@ -50,7 +49,7 @@ option(MFEM_USE_OCCA "Enable OCCA" OFF)
 option(MFEM_USE_RAJA "Enable RAJA" OFF)
 option(MFEM_USE_CEED "Enable CEED" OFF)
 option(MFEM_USE_UMPIRE "Enable Umpire" OFF)
-option(MFEM_USE_SIMD "Enable use of SIMD intrinsics" OFF)
+option(MFEM_USE_SIMD "Enable use of SIMD intrinsics" ON)
 option(MFEM_USE_ADIOS2 "Enable ADIOS2" OFF)

 set(MFEM_MPI_NP 4 CACHE STRING "Number of processes used for MPI tests")
@@ -88,8 +87,6 @@ set(METIS_DIR "${MFEM_DIR}/../metis-4.0" CACHE PATH "Path to the METIS library."

 set(LIBUNWIND_DIR "" CACHE PATH "Path to Libunwind.")

-# For sundials_nvecparhyp and nvecparallel remember to build with MPI_ENABLED=ON
-# and modify cmake variables for hypre for sundials
 set(SUNDIALS_DIR "${MFEM_DIR}/../sundials-5.0.0/instdir" CACHE PATH
    "Path to the SUNDIALS library.")
 # The following may be necessary, if SUNDIALS was built with KLU:
@@ -158,10 +155,6 @@ set(PETSC_DIR "${MFEM_DIR}/../petsc" CACHE PATH
    "Path to the PETSc main directory.")
 set(PETSC_ARCH "arch-linux2-c-debug" CACHE STRING "PETSc build architecture.")

-set(SLEPC_DIR "${MFEM_DIR}/../slepc" CACHE PATH
-    "Path to the SLEPc main directory.")
-set(SLEPC_ARCH "arch-linux2-c-debug" CACHE STRING "SLEPC build architecture.")
-
 set(MPFR_DIR "" CACHE PATH "Path to the MPFR library.")

 set(CONDUIT_DIR "${MFEM_DIR}/../conduit" CACHE PATH
@@ -138,8 +138,7 @@ MFEM_USE_RAJA          = NO
 MFEM_USE_OCCA          = NO
 MFEM_USE_CEED          = NO
 MFEM_USE_UMPIRE        = NO
-MFEM_USE_CAMP          = NO
-MFEM_USE_SIMD          = NO
+MFEM_USE_SIMD          = YES
 MFEM_USE_ADIOS2        = NO

 # Compile and link options for zlib.
@@ -190,12 +189,10 @@ OPENMP_LIB =
 POSIX_CLOCKS_LIB = -lrt

 # SUNDIALS library configuration
-# For sundials_nvecparhyp and nvecparallel remember to build with MPI_ENABLED=ON
-# and modify cmake variables for hypre for sundials
 SUNDIALS_DIR = @MFEM_DIR@/../sundials-5.0.0/instdir
 SUNDIALS_OPT = -I$(SUNDIALS_DIR)/include
 SUNDIALS_LIB = -Wl,-rpath,$(SUNDIALS_DIR)/lib64 -L$(SUNDIALS_DIR)/lib64\
- -lsundials_arkode -lsundials_cvodes -lsundials_nvecserial -lsundials_kinsol
+ -lsundials_arkode -lsundials_cvode -lsundials_nvecserial -lsundials_kinsol

 ifeq ($(MFEM_USE_MPI),YES)
   SUNDIALS_LIB += -lsundials_nvecparhyp -lsundials_nvecparallel
@@ -373,11 +370,6 @@ UMPIRE_DIR = @MFEM_DIR@/../umpire
 UMPIRE_OPT = -I$(UMPIRE_DIR)/include
 UMPIRE_LIB = -L$(UMPIRE_DIR)/lib -lumpire

-# CAMP library configuration
-CAMP_DIR = @MFEM_DIR@/../camp
-CAMP_OPT = -I$(CAMP_DIR)/include
-CAMP_LIB = -L$(CAMP_DIR)/lib
-
 # If YES, enable some informational messages
 VERBOSE = NO

@@ -770,7 +770,6 @@ INPUT                  = @MFEM_SOURCE_DIR@/doc/CodeDocumentation.dox \
                         @MFEM_SOURCE_DIR@/examples/pumi \
                         @MFEM_SOURCE_DIR@/examples/hiop \
                         @MFEM_SOURCE_DIR@/examples/sundials \
-                         @MFEM_SOURCE_DIR@/miniapps/adjoint \			 
                         @MFEM_SOURCE_DIR@/miniapps/common \
                         @MFEM_SOURCE_DIR@/miniapps/electromagnetics \
                         @MFEM_SOURCE_DIR@/miniapps/gslib \
@@ -88,8 +88,8 @@ namespace mfem {
 * - <a class="el" href="ex24p_8cpp_source.html">Example 24p</a>: parallel mixed finite element spaces and interpolators
 * - <a class="el" href="ex25_8cpp_source.html">Example 25</a>: simulation of electromagnetic wave propagation using a Perfectly Matched Layer (PML)
 * - <a class="el" href="ex25p_8cpp_source.html">Example 25p</a>: parallel simulation of electromagnetic wave propagation using a Perfectly Matched Layer (PML)
- * - <a class="el" href="ex26_8cpp_source.html">Example 26</a>: multigrid preconditioner for the Laplace problem using nodal H1 FEM
- * - <a class="el" href="ex26p_8cpp_source.html">Example 26p</a>: parallel multigrid preconditioner for the Laplace problem using nodal H1 FEM
+ * - <a class="el" href="ex26_8cpp_source.html">Example 26</a>: multigrid preconditioner for the Laplace problem using nodal H1 FEM 
+ * - <a class="el" href="ex26p_8cpp_source.html">Example 26p</a>: parallel multigrid preconditioner for the Laplace problem using nodal H1 FEM 
 *
 * <H4>SUNDIALS Examples</H4>
 * - Variants of Examples
@@ -101,9 +101,6 @@ namespace mfem {
 *   and
 *   <a class="el" href="sundials_2ex16p_8cpp_source.html">16p</a>
 *   demonstrating the use of MFEM's \link sundials.hpp SUNDIALS classes\endlink
- * - CVODES adjoint examples:
- *   <a class="el" href="cvsRoberts__ASAi__dns_8cpp_source.html">serial ODE system</a>,
- *   <a class="el" href="adjoint__advection__diffusion_8cpp_source.html">parallel advection-diffusion</a>
 *
 * <H4>PETSc Examples</H4>
 * - Variants of Examples
@@ -143,9 +140,7 @@ 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="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="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
 * - <a class="el" href="toroid_8cpp_source.html">Toroid</a>: generate simple toroidal meshes
 * - <a class="el" href="twist_8cpp_source.html">Twist</a>: generate simple periodic meshes
@@ -162,6 +157,7 @@ namespace mfem {
 * - <a class="el" href="lor-transfer_8cpp_source.html">LOR Transfer</a>: map functions between high-order and low-order refined spaces
 * - <a class="el" href="findpts_8cpp_source.html">Find Points</a>: evaluate grid function in physical space, <a class="el" href="findpts_8cpp_source.html">serial</a> and <a class="el" href="pfindpts_8cpp_source.html">parallel</a> versions
 * - <a class="el" href="field-diff_8cpp_source.html">Field Diff</a>: compare grid functions on different meshes
+ * - <a class="el" href="classmfem_1_1navier_1_1NavierSolver.html">Navier</a>: solve the transient incompressible Navier-Stokes equations
 * - <a class="el" href="miniapps_2performance_2ex1_8cpp_source.html">HPC Example 1</a>: high-performance nodal H1 FEM for the Laplace problem
 * - <a class="el" href="miniapps_2performance_2ex1p_8cpp_source.html">HPC Example 1p</a>: high-performance parallel nodal H1 FEM for the Laplace problem
 *
@@ -91,7 +91,7 @@ foreach(SRC_FILE ${ALL_EXE_SRCS})
    add_test(NAME ${TEST_NAME}_ser
      COMMAND ${TEST_NAME} ${THIS_TEST_OPTIONS})
  else()
-    add_test(NAME ${TEST_NAME}_np=${MFEM_MPI_NP}
+    add_test(NAME ${TEST_NAME}_np=4
      COMMAND ${MPIEXEC} ${MPIEXEC_NUMPROC_FLAG} ${MFEM_MPI_NP}
      ${MPIEXEC_PREFLAGS}
      $<TARGET_FILE:${TEST_NAME}> ${THIS_TEST_OPTIONS}
@@ -101,22 +101,13 @@ endforeach()

 # If STRUMPACK is enabled, add a test run that uses it.
 if (MFEM_USE_STRUMPACK)
-  add_test(NAME ex11p_strumpack_np=${MFEM_MPI_NP}
+  add_test(NAME ex11p_strumpack_np=4
    COMMAND ${MPIEXEC} ${MPIEXEC_NUMPROC_FLAG} ${MFEM_MPI_NP}
    ${MPIEXEC_PREFLAGS}
    $<TARGET_FILE:ex11p> "-no-vis" "--strumpack"
    ${MPIEXEC_POSTFLAGS})
 endif()

-# If SuperLU_DIST is enabled, add a test run that uses it.
-if (MFEM_USE_SUPERLU)
-  add_test(NAME ex11p_superlu_np=${MFEM_MPI_NP}
-    COMMAND ${MPIEXEC} ${MPIEXEC_NUMPROC_FLAG} ${MFEM_MPI_NP}
-    ${MPIEXEC_PREFLAGS}
-    $<TARGET_FILE:ex11p> "-no-vis" "--superlu"
-    ${MPIEXEC_POSTFLAGS})
-endif()
-
 # Include the examples/sundials directory if SUNDIALS is enabled.
 if (MFEM_USE_SUNDIALS)
  add_subdirectory(sundials)
@@ -34,8 +34,7 @@
 //               ex1 -pa -d raja-omp
 //               ex1 -pa -d occa-omp
 //               ex1 -pa -d ceed-cpu
-//             * ex1 -pa -d ceed-cuda
-//               ex1 -pa -d ceed-cuda:/gpu/cuda/shared
+//               ex1 -pa -d ceed-cuda
 //               ex1 -m ../data/beam-hex.mesh -pa -d cuda
 //               ex1 -m ../data/beam-tet.mesh -pa -d ceed-cpu
 //               ex1 -m ../data/beam-tet.mesh -pa -d ceed-cuda:/gpu/cuda/ref
@@ -103,8 +102,8 @@ int main(int argc, char *argv[])
   // 3. Read the mesh from the given mesh file. We can handle triangular,
   //    quadrilateral, tetrahedral, hexahedral, surface and volume meshes with
   //    the same code.
-   Mesh mesh(mesh_file, 1, 1);
-   int dim = mesh.Dimension();
+   Mesh *mesh = new Mesh(mesh_file, 1, 1);
+   int dim = mesh->Dimension();

   // 4. Refine the mesh to increase the resolution. In this example we do
   //    'ref_levels' of uniform refinement. We choose 'ref_levels' to be the
@@ -112,10 +111,10 @@ int main(int argc, char *argv[])
   //    elements.
   {
      int ref_levels =
-         (int)floor(log(50000./mesh.GetNE())/log(2.)/dim);
+         (int)floor(log(50000./mesh->GetNE())/log(2.)/dim);
      for (int l = 0; l < ref_levels; l++)
      {
-         mesh.UniformRefinement();
+         mesh->UniformRefinement();
      }
   }

@@ -123,70 +122,66 @@ int main(int argc, char *argv[])
   //    Lagrange finite elements of the specified order. If order < 1, we
   //    instead use an isoparametric/isogeometric space.
   FiniteElementCollection *fec;
-   bool delete_fec;
   if (order > 0)
   {
      fec = new H1_FECollection(order, dim);
-      delete_fec = true;
   }
-   else if (mesh.GetNodes())
+   else if (mesh->GetNodes())
   {
-      fec = mesh.GetNodes()->OwnFEC();
-      delete_fec = false;
+      fec = mesh->GetNodes()->OwnFEC();
      cout << "Using isoparametric FEs: " << fec->Name() << endl;
   }
   else
   {
      fec = new H1_FECollection(order = 1, dim);
-      delete_fec = true;
   }
-   FiniteElementSpace fespace(&mesh, fec);
+   FiniteElementSpace *fespace = new FiniteElementSpace(mesh, fec);
   cout << "Number of finite element unknowns: "
-        << fespace.GetTrueVSize() << endl;
+        << fespace->GetTrueVSize() << endl;

   // 6. Determine the list of true (i.e. conforming) essential boundary dofs.
   //    In this example, the boundary conditions are defined by marking all
   //    the boundary attributes from the mesh as essential (Dirichlet) and
   //    converting them to a list of true dofs.
   Array<int> ess_tdof_list;
-   if (mesh.bdr_attributes.Size())
+   if (mesh->bdr_attributes.Size())
   {
-      Array<int> ess_bdr(mesh.bdr_attributes.Max());
+      Array<int> ess_bdr(mesh->bdr_attributes.Max());
      ess_bdr = 1;
-      fespace.GetEssentialTrueDofs(ess_bdr, ess_tdof_list);
+      fespace->GetEssentialTrueDofs(ess_bdr, ess_tdof_list);
   }

   // 7. Set up the linear form b(.) which corresponds to the right-hand side of
   //    the FEM linear system, which in this case is (1,phi_i) where phi_i are
   //    the basis functions in the finite element fespace.
-   LinearForm b(&fespace);
+   LinearForm *b = new LinearForm(fespace);
   ConstantCoefficient one(1.0);
-   b.AddDomainIntegrator(new DomainLFIntegrator(one));
-   b.Assemble();
+   b->AddDomainIntegrator(new DomainLFIntegrator(one));
+   b->Assemble();

   // 8. Define the solution vector x as a finite element grid function
   //    corresponding to fespace. Initialize x with initial guess of zero,
   //    which satisfies the boundary conditions.
-   GridFunction x(&fespace);
+   GridFunction x(fespace);
   x = 0.0;

   // 9. Set up the bilinear form a(.,.) on the finite element space
   //    corresponding to the Laplacian operator -Delta, by adding the Diffusion
   //    domain integrator.
-   BilinearForm a(&fespace);
-   if (pa) { a.SetAssemblyLevel(AssemblyLevel::PARTIAL); }
-   a.AddDomainIntegrator(new DiffusionIntegrator(one));
+   BilinearForm *a = new BilinearForm(fespace);
+   if (pa) { a->SetAssemblyLevel(AssemblyLevel::PARTIAL); }
+   a->AddDomainIntegrator(new DiffusionIntegrator(one));

   // 10. Assemble the bilinear form and the corresponding linear system,
   //     applying any necessary transformations such as: eliminating boundary
   //     conditions, applying conforming constraints for non-conforming AMR,
   //     static condensation, etc.
-   if (static_cond) { a.EnableStaticCondensation(); }
-   a.Assemble();
+   if (static_cond) { a->EnableStaticCondensation(); }
+   a->Assemble();

   OperatorPtr A;
   Vector B, X;
-   a.FormLinearSystem(ess_tdof_list, x, b, A, X, B);
+   a->FormLinearSystem(ess_tdof_list, x, *b, A, X, B);

   cout << "Size of linear system: " << A->Height() << endl;

@@ -207,9 +202,9 @@ int main(int argc, char *argv[])
   }
   else // Jacobi preconditioning in partial assembly mode
   {
-      if (UsesTensorBasis(fespace))
+      if (UsesTensorBasis(*fespace))
      {
-         OperatorJacobiSmoother M(a, ess_tdof_list);
+         OperatorJacobiSmoother M(*a, ess_tdof_list);
         PCG(*A, M, B, X, 1, 400, 1e-12, 0.0);
      }
      else
@@ -219,13 +214,13 @@ int main(int argc, char *argv[])
   }

   // 12. Recover the solution as a finite element grid function.
-   a.RecoverFEMSolution(X, b, x);
+   a->RecoverFEMSolution(X, *b, x);

   // 13. Save the refined mesh and the solution. This output can be viewed later
   //     using GLVis: "glvis -m refined.mesh -g sol.gf".
   ofstream mesh_ofs("refined.mesh");
   mesh_ofs.precision(8);
-   mesh.Print(mesh_ofs);
+   mesh->Print(mesh_ofs);
   ofstream sol_ofs("sol.gf");
   sol_ofs.precision(8);
   x.Save(sol_ofs);
@@ -237,14 +232,15 @@ int main(int argc, char *argv[])
      int  visport   = 19916;
      socketstream sol_sock(vishost, visport);
      sol_sock.precision(8);
-      sol_sock << "solution\n" << mesh << x << flush;
+      sol_sock << "solution\n" << *mesh << x << flush;
   }

   // 15. Free the used memory.
-   if (delete_fec)
-   {
-      delete fec;
-   }
+   delete a;
+   delete b;
+   delete fespace;
+   if (order > 0) { delete fec; }
+   delete mesh;

   return 0;
 }
@@ -88,6 +88,8 @@ private:
   Vector funval2;
   Vector nor;
   Vector fluxN;
+   IntegrationPoint eip1;
+   IntegrationPoint eip2;

 public:
   FaceIntegrator(RiemannSolver &rsolver_, const int dim);
@@ -422,16 +424,19 @@ void FaceIntegrator::AssembleFaceVector(const FiniteElement &el1,
   {
      const IntegrationPoint &ip = ir->IntPoint(i);

-      Tr.SetAllIntPoints(&ip); // set face and element int. points
+      Tr.Loc1.Transform(ip, eip1);
+      Tr.Loc2.Transform(ip, eip2);

      // Calculate basis functions on both elements at the face
-      el1.CalcShape(Tr.GetElement1IntPoint(), shape1);
-      el2.CalcShape(Tr.GetElement2IntPoint(), shape2);
+      el1.CalcShape(eip1, shape1);
+      el2.CalcShape(eip2, shape2);

      // Interpolate elfun at the point
      elfun1_mat.MultTranspose(shape1, funval1);
      elfun2_mat.MultTranspose(shape2, funval2);

+      Tr.SetIntPoint(&ip);
+
      // Get the normal vector and the flux on the face
      CalcOrtho(Tr.Jacobian(), nor);
      const double mcs = rsolver.Eval(funval1, funval2, nor, fluxN);
@@ -32,8 +32,7 @@
 //               mpirun -np 4 ex1p -pa -d occa-cuda
 //               mpirun -np 4 ex1p -pa -d raja-omp
 //               mpirun -np 4 ex1p -pa -d ceed-cpu
-//             * mpirun -np 4 ex1p -pa -d ceed-cuda
-//               mpirun -np 4 ex1p -pa -d ceed-cuda:/gpu/cuda/shared
+//               mpirun -np 4 ex1p -pa -d ceed-cuda
 //               mpirun -np 4 ex1p -m ../data/beam-tet.mesh -pa -d ceed-cpu
 //
 // Description:  This example code demonstrates the use of MFEM to define a
@@ -112,8 +111,8 @@ int main(int argc, char *argv[])
   // 4. Read the (serial) mesh from the given mesh file on all processors.  We
   //    can handle triangular, quadrilateral, tetrahedral, hexahedral, surface
   //    and volume meshes with the same code.
-   Mesh mesh(mesh_file, 1, 1);
-   int dim = mesh.Dimension();
+   Mesh *mesh = new Mesh(mesh_file, 1, 1);
+   int dim = mesh->Dimension();

   // 5. Refine the serial mesh on all processors to increase the resolution. In
   //    this example we do 'ref_levels' of uniform refinement. We choose
@@ -121,23 +120,23 @@ int main(int argc, char *argv[])
   //    more than 10,000 elements.
   {
      int ref_levels =
-         (int)floor(log(10000./mesh.GetNE())/log(2.)/dim);
+         (int)floor(log(10000./mesh->GetNE())/log(2.)/dim);
      for (int l = 0; l < ref_levels; l++)
      {
-         mesh.UniformRefinement();
+         mesh->UniformRefinement();
      }
   }

   // 6. Define a parallel mesh by a partitioning of the serial mesh. Refine
   //    this mesh further in parallel to increase the resolution. Once the
   //    parallel mesh is defined, the serial mesh can be deleted.
-   ParMesh pmesh(MPI_COMM_WORLD, mesh);
-   mesh.Clear();
+   ParMesh *pmesh = new ParMesh(MPI_COMM_WORLD, *mesh);
+   delete mesh;
   {
      int par_ref_levels = 2;
      for (int l = 0; l < par_ref_levels; l++)
      {
-         pmesh.UniformRefinement();
+         pmesh->UniformRefinement();
      }
   }

@@ -145,16 +144,13 @@ int main(int argc, char *argv[])
   //    use continuous Lagrange finite elements of the specified order. If
   //    order < 1, we instead use an isoparametric/isogeometric space.
   FiniteElementCollection *fec;
-   bool delete_fec;
   if (order > 0)
   {
      fec = new H1_FECollection(order, dim);
-      delete_fec = true;
   }
-   else if (pmesh.GetNodes())
+   else if (pmesh->GetNodes())
   {
-      fec = pmesh.GetNodes()->OwnFEC();
-      delete_fec = false;
+      fec = pmesh->GetNodes()->OwnFEC();
      if (myid == 0)
      {
         cout << "Using isoparametric FEs: " << fec->Name() << endl;
@@ -163,10 +159,9 @@ int main(int argc, char *argv[])
   else
   {
      fec = new H1_FECollection(order = 1, dim);
-      delete_fec = true;
   }
-   ParFiniteElementSpace fespace(&pmesh, fec);
-   HYPRE_Int size = fespace.GlobalTrueVSize();
+   ParFiniteElementSpace *fespace = new ParFiniteElementSpace(pmesh, fec);
+   HYPRE_Int size = fespace->GlobalTrueVSize();
   if (myid == 0)
   {
      cout << "Number of finite element unknowns: " << size << endl;
@@ -177,44 +172,44 @@ int main(int argc, char *argv[])
   //    by marking all the boundary attributes from the mesh as essential
   //    (Dirichlet) and converting them to a list of true dofs.
   Array<int> ess_tdof_list;
-   if (pmesh.bdr_attributes.Size())
+   if (pmesh->bdr_attributes.Size())
   {
-      Array<int> ess_bdr(pmesh.bdr_attributes.Max());
+      Array<int> ess_bdr(pmesh->bdr_attributes.Max());
      ess_bdr = 1;
-      fespace.GetEssentialTrueDofs(ess_bdr, ess_tdof_list);
+      fespace->GetEssentialTrueDofs(ess_bdr, ess_tdof_list);
   }

   // 9. Set up the parallel linear form b(.) which corresponds to the
   //    right-hand side of the FEM linear system, which in this case is
   //    (1,phi_i) where phi_i are the basis functions in fespace.
-   ParLinearForm b(&fespace);
+   ParLinearForm *b = new ParLinearForm(fespace);
   ConstantCoefficient one(1.0);
-   b.AddDomainIntegrator(new DomainLFIntegrator(one));
-   b.Assemble();
+   b->AddDomainIntegrator(new DomainLFIntegrator(one));
+   b->Assemble();

   // 10. Define the solution vector x as a parallel finite element grid function
   //     corresponding to fespace. Initialize x with initial guess of zero,
   //     which satisfies the boundary conditions.
-   ParGridFunction x(&fespace);
+   ParGridFunction x(fespace);
   x = 0.0;

   // 11. Set up the parallel bilinear form a(.,.) on the finite element space
   //     corresponding to the Laplacian operator -Delta, by adding the Diffusion
   //     domain integrator.
-   ParBilinearForm a(&fespace);
-   if (pa) { a.SetAssemblyLevel(AssemblyLevel::PARTIAL); }
-   a.AddDomainIntegrator(new DiffusionIntegrator(one));
+   ParBilinearForm *a = new ParBilinearForm(fespace);
+   if (pa) { a->SetAssemblyLevel(AssemblyLevel::PARTIAL); }
+   a->AddDomainIntegrator(new DiffusionIntegrator(one));

   // 12. Assemble the parallel bilinear form and the corresponding linear
   //     system, applying any necessary transformations such as: parallel
   //     assembly, eliminating boundary conditions, applying conforming
   //     constraints for non-conforming AMR, static condensation, etc.
-   if (static_cond) { a.EnableStaticCondensation(); }
-   a.Assemble();
+   if (static_cond) { a->EnableStaticCondensation(); }
+   a->Assemble();

   OperatorPtr A;
   Vector B, X;
-   a.FormLinearSystem(ess_tdof_list, x, b, A, X, B);
+   a->FormLinearSystem(ess_tdof_list, x, *b, A, X, B);

   // 13. Solve the linear system A X = B.
   //     * With full assembly, use the BoomerAMG preconditioner from hypre.
@@ -222,9 +217,9 @@ int main(int argc, char *argv[])
   Solver *prec = NULL;
   if (pa)
   {
-      if (UsesTensorBasis(fespace))
+      if (UsesTensorBasis(*fespace))
      {
-         prec = new OperatorJacobiSmoother(a, ess_tdof_list);
+         prec = new OperatorJacobiSmoother(*a, ess_tdof_list);
      }
   }
   else
@@ -242,7 +237,7 @@ int main(int argc, char *argv[])

   // 14. Recover the parallel grid function corresponding to X. This is the
   //     local finite element solution on each processor.
-   a.RecoverFEMSolution(X, b, x);
+   a->RecoverFEMSolution(X, *b, x);

   // 15. Save the refined mesh and the solution in parallel. This output can
   //     be viewed later using GLVis: "glvis -np <np> -m mesh -g sol".
@@ -253,7 +248,7 @@ int main(int argc, char *argv[])

      ofstream mesh_ofs(mesh_name.str().c_str());
      mesh_ofs.precision(8);
-      pmesh.Print(mesh_ofs);
+      pmesh->Print(mesh_ofs);

      ofstream sol_ofs(sol_name.str().c_str());
      sol_ofs.precision(8);
@@ -268,14 +263,16 @@ int main(int argc, char *argv[])
      socketstream sol_sock(vishost, visport);
      sol_sock << "parallel " << num_procs << " " << myid << "\n";
      sol_sock.precision(8);
-      sol_sock << "solution\n" << pmesh << x << flush;
+      sol_sock << "solution\n" << *pmesh << x << flush;
   }

   // 17. Free the used memory.
-   if (delete_fec)
-   {
-      delete fec;
-   }
+   delete a;
+   delete b;
+   delete fespace;
+   if (order > 0) { delete fec; }
+   delete pmesh;
+
   MPI_Finalize();

   return 0;
@@ -13,11 +13,6 @@
 //               ex22 -m ../data/inline-hex.mesh -o 2 -p 2
 //               ex22 -m ../data/star.mesh -r 1 -o 2 -sigma 10.0
 //
-// With partial assembly:
-//               ex22 -m ../data/inline-quad.mesh -o 3 -p 1 -pa
-//               ex22 -m ../data/inline-hex.mesh -o 2 -p 2 -pa
-//               ex22 -m ../data/star.mesh -r 1 -o 2 -sigma 10.0 -pa
-//
 // Description:  This example code demonstrates the use of MFEM to define and
 //               solve simple complex-valued linear systems. It implements three
 //               variants of a damped harmonic oscillator:
@@ -81,7 +76,6 @@ int main(int argc, char *argv[])
   bool visualization = 1;
   bool herm_conv = true;
   bool exact_sol = true;
-   bool pa = false;

   OptionsParser args(argc, argv);
   args.AddOption(&mesh_file, "-m", "--mesh",
@@ -112,8 +106,6 @@ int main(int argc, char *argv[])
   args.AddOption(&visualization, "-vis", "--visualization", "-no-vis",
                  "--no-visualization",
                  "Enable or disable GLVis visualization.");
-   args.AddOption(&pa, "-pa", "--partial-assembly", "-no-pa",
-                  "--no-partial-assembly", "Enable Partial Assembly.");
   args.Parse();
   if (!args.Good())
   {
@@ -290,7 +282,6 @@ int main(int argc, char *argv[])
   ConstantCoefficient negMassCoef(omega_ * omega_ * epsilon_);

   SesquilinearForm *a = new SesquilinearForm(fespace, conv);
-   if (pa) { a->SetAssemblyLevel(AssemblyLevel::PARTIAL); }
   switch (prob)
   {
      case 0:
@@ -327,8 +318,6 @@ int main(int argc, char *argv[])
   //         -Grad(a Div) - omega^2 b + omega c
   //
   BilinearForm *pcOp = new BilinearForm(fespace);
-   if (pa) { pcOp->SetAssemblyLevel(AssemblyLevel::PARTIAL); }
-
   switch (prob)
   {
      case 0:
@@ -359,8 +348,19 @@ int main(int argc, char *argv[])
   Vector B, U;

   a->FormLinearSystem(ess_tdof_list, u, b, A, U, B);
+   u = 0.0;
+   U = 0.0;

-   cout << "Size of linear system: " << A->Width() << endl << endl;
+   OperatorHandle PCOp;
+   pcOp->FormSystemMatrix(ess_tdof_list, PCOp);
+
+   {
+      ComplexSparseMatrix * Asp =
+         dynamic_cast<ComplexSparseMatrix*>(A.Ptr());
+
+      cout << "Size of linear system: "
+           << 2 * Asp->real().Width() << endl << endl;
+   }

   // 10. Define and apply a GMRES solver for AU=B with a block diagonal
   //     preconditioner based on the appropriate sparse smoother.
@@ -368,8 +368,8 @@ int main(int argc, char *argv[])
      Array<int> blockOffsets;
      blockOffsets.SetSize(3);
      blockOffsets[0] = 0;
-      blockOffsets[1] = A->Height() / 2;
-      blockOffsets[2] = A->Height() / 2;
+      blockOffsets[1] = PCOp.Ptr()->Height();
+      blockOffsets[2] = PCOp.Ptr()->Height();
      blockOffsets.PartialSum();

      BlockDiagonalPreconditioner BDP(blockOffsets);
@@ -377,31 +377,22 @@ int main(int argc, char *argv[])
      Operator * pc_r = NULL;
      Operator * pc_i = NULL;

-      if (pa)
+      double s = 1.0;
+      switch (prob)
      {
-         pc_r = new OperatorJacobiSmoother(*pcOp, ess_tdof_list);
+         case 0:
+            pc_r = new DSmoother(*PCOp.As<SparseMatrix>());
+            break;
+         case 1:
+            pc_r = new GSSmoother(*PCOp.As<SparseMatrix>());
+            s = -1.0;
+            break;
+         case 2:
+            pc_r = new DSmoother(*PCOp.As<SparseMatrix>());
+            break;
+
+         default: break; // This should be unreachable
      }
-      else
-      {
-         OperatorHandle PCOp;
-         pcOp->SetDiagonalPolicy(mfem::Operator::DIAG_ONE);
-         pcOp->FormSystemMatrix(ess_tdof_list, PCOp);
-         switch (prob)
-         {
-            case 0:
-               pc_r = new DSmoother(*PCOp.As<SparseMatrix>());
-               break;
-            case 1:
-               pc_r = new GSSmoother(*PCOp.As<SparseMatrix>());
-               break;
-            case 2:
-               pc_r = new DSmoother(*PCOp.As<SparseMatrix>());
-               break;
-            default:
-               break; // This should be unreachable
-         }
-      }
-      double s = (prob != 1) ? 1.0 : -1.0;
      pc_i = new ScaledOperator(pc_r,
                                (conv == ComplexOperator::HERMITIAN) ?
                                s:-s);
@@ -13,11 +13,6 @@
 //               mpirun -np 4 ex22p -m ../data/inline-hex.mesh -o 2 -p 2
 //               mpirun -np 4 ex22p -m ../data/star.mesh -o 2 -sigma 10.0
 //
-// With partial assembly:
-//               mpirun -np 4 ex22p -m ../data/inline-quad.mesh -o 1 -p 1 -pa
-//               mpirun -np 4 ex22p -m ../data/inline-hex.mesh -o 1 -p 2 -pa
-//               mpirun -np 4 ex22p -m ../data/star.mesh -o 2 -sigma 10.0 -pa
-//
 // Description:  This example code demonstrates the use of MFEM to define and
 //               solve simple complex-valued linear systems. It implements three
 //               variants of a damped harmonic oscillator:
@@ -89,7 +84,6 @@ int main(int argc, char *argv[])
   bool visualization = 1;
   bool herm_conv = true;
   bool exact_sol = true;
-   bool pa = false;

   OptionsParser args(argc, argv);
   args.AddOption(&mesh_file, "-m", "--mesh",
@@ -122,8 +116,6 @@ int main(int argc, char *argv[])
   args.AddOption(&visualization, "-vis", "--visualization", "-no-vis",
                  "--no-visualization",
                  "Enable or disable GLVis visualization.");
-   args.AddOption(&pa, "-pa", "--partial-assembly", "-no-pa",
-                  "--no-partial-assembly", "Enable Partial Assembly.");
   args.Parse();
   if (!args.Good())
   {
@@ -323,7 +315,6 @@ int main(int argc, char *argv[])
   ConstantCoefficient negMassCoef(omega_ * omega_ * epsilon_);

   ParSesquilinearForm *a = new ParSesquilinearForm(fespace, conv);
-   if (pa) { a->SetAssemblyLevel(AssemblyLevel::PARTIAL); }
   switch (prob)
   {
      case 0:
@@ -360,7 +351,6 @@ int main(int argc, char *argv[])
   //         -Grad(a Div) - omega^2 b + omega c
   //
   ParBilinearForm *pcOp = new ParBilinearForm(fespace);
-   if (pa) { pcOp->SetAssemblyLevel(AssemblyLevel::PARTIAL); }
   switch (prob)
   {
      case 0:
@@ -392,11 +382,19 @@ int main(int argc, char *argv[])
   Vector B, U;

   a->FormLinearSystem(ess_tdof_list, u, b, A, U, B);
+   u = 0.0;
+   U = 0.0;
+
+   OperatorHandle PCOp;
+   pcOp->FormSystemMatrix(ess_tdof_list, PCOp);

   if (myid == 0)
   {
+      ComplexHypreParMatrix * Ahyp =
+         dynamic_cast<ComplexHypreParMatrix*>(A.Ptr());
+
      cout << "Size of linear system: "
-           << 2 * fespace->GlobalTrueVSize() << endl << endl;
+           << 2 * Ahyp->real().GetGlobalNumRows() << endl << endl;
   }

   // 12. Define and apply a parallel FGMRES solver for AU=B with a block
@@ -406,8 +404,8 @@ int main(int argc, char *argv[])
      Array<int> blockTrueOffsets;
      blockTrueOffsets.SetSize(3);
      blockTrueOffsets[0] = 0;
-      blockTrueOffsets[1] = A->Height() / 2;
-      blockTrueOffsets[2] = A->Height() / 2;
+      blockTrueOffsets[1] = PCOp.Ptr()->Height();
+      blockTrueOffsets[2] = PCOp.Ptr()->Height();
      blockTrueOffsets.PartialSum();

      BlockDiagonalPreconditioner BDP(blockTrueOffsets);
@@ -415,34 +413,25 @@ int main(int argc, char *argv[])
      Operator * pc_r = NULL;
      Operator * pc_i = NULL;

-      if (pa)
+      switch (prob)
      {
-         pc_r = new OperatorJacobiSmoother(*pcOp, ess_tdof_list);
-      }
-      else
-      {
-         OperatorHandle PCOp;
-         pcOp->FormSystemMatrix(ess_tdof_list, PCOp);
-         switch (prob)
-         {
-            case 0:
-               pc_r = new HypreBoomerAMG(*PCOp.As<HypreParMatrix>());
-               break;
-            case 1:
+         case 0:
+            pc_r = new HypreBoomerAMG(*PCOp.As<HypreParMatrix>());
+            break;
+         case 1:
+            pc_r = new HypreAMS(*PCOp.As<HypreParMatrix>(), fespace);
+            break;
+         case 2:
+            if (dim == 2 )
+            {
               pc_r = new HypreAMS(*PCOp.As<HypreParMatrix>(), fespace);
-               break;
-            case 2:
-               if (dim == 2 )
-               {
-                  pc_r = new HypreAMS(*PCOp.As<HypreParMatrix>(), fespace);
-               }
-               else
-               {
-                  pc_r = new HypreADS(*PCOp.As<HypreParMatrix>(), fespace);
-               }
-               break;
-            default: break; // This should be unreachable
-         }
+            }
+            else
+            {
+               pc_r = new HypreADS(*PCOp.As<HypreParMatrix>(), fespace);
+            }
+            break;
+         default: break; // This should be unreachable
      }
      pc_i = new ScaledOperator(pc_r,
                                (conv == ComplexOperator::HERMITIAN) ?
@@ -7,7 +7,6 @@
 //               ex24 -m ../data/beam-tet.mesh
 //               ex24 -m ../data/beam-hex.mesh -o 2 -pa
 //               ex24 -m ../data/beam-hex.mesh -o 2 -pa -p 1
-//               ex24 -m ../data/beam-hex.mesh -o 2 -pa -p 2
 //               ex24 -m ../data/escher.mesh
 //               ex24 -m ../data/escher.mesh -o 2
 //               ex24 -m ../data/fichera.mesh
@@ -25,13 +24,12 @@
 //               ex24 -m ../data/beam-hex.mesh -pa -d cuda
 //
 // Description:  This example code illustrates usage of mixed finite element
-//               spaces, with three variants:
+//               spaces, with two variants:
 //
 //               1) (grad p, u) for p in H^1 tested against u in H(curl)
-//               2) (curl v, u) for v in H(curl) tested against u in H(div), 3D
-//               3) (div v, q) for v in H(div) tested against q in L_2
+//               2) (div v, q) for v in H(div) tested against q in L_2
 //
-//               Using different approaches, we project the gradient, curl, or
+//               Using different approaches, we project the gradient or
 //               divergence to the appropriate space.
 //
 //               We recommend viewing examples 1, 3, and 5 before viewing this
@@ -47,11 +45,8 @@ using namespace mfem;
 double p_exact(const Vector &x);
 void gradp_exact(const Vector &, Vector &);
 double div_gradp_exact(const Vector &x);
-void v_exact(const Vector &x, Vector &v);
-void curlv_exact(const Vector &x, Vector &cv);

 int dim;
-double freq = 1.0, kappa;

 int main(int argc, char *argv[])
 {
@@ -88,7 +83,6 @@ int main(int argc, char *argv[])
      return 1;
   }
   args.PrintOptions(cout);
-   kappa = freq * M_PI;

   // 2. Enable hardware devices such as GPUs, and programming models such as
   //    CUDA, OCCA, RAJA and OpenMP based on command line options.
@@ -125,15 +119,10 @@ int main(int argc, char *argv[])
      trial_fec = new H1_FECollection(order, dim);
      test_fec = new ND_FECollection(order, dim);
   }
-   else if (prob == 1)
-   {
-      trial_fec = new ND_FECollection(order, dim);
-      test_fec = new RT_FECollection(order-1, dim);
-   }
   else
   {
-      trial_fec = new RT_FECollection(order-1, dim);
-      test_fec = new L2_FECollection(order-1, dim);
+      trial_fec = new RT_FECollection(order - 1, dim);
+      test_fec = new L2_FECollection(order - 1, dim);
   }

   FiniteElementSpace trial_fes(mesh, trial_fec);
@@ -147,12 +136,6 @@ int main(int argc, char *argv[])
      cout << "Number of Nedelec finite element unknowns: " << test_size << endl;
      cout << "Number of H1 finite element unknowns: " << trial_size << endl;
   }
-   else if (prob == 1)
-   {
-      cout << "Number of Nedelec finite element unknowns: " << trial_size << endl;
-      cout << "Number of Raviart-Thomas finite element unknowns: " << test_size <<
-           endl;
-   }
   else
   {
      cout << "Number of Raviart-Thomas finite element unknowns: "
@@ -167,18 +150,12 @@ int main(int argc, char *argv[])
   GridFunction x(&test_fes);
   FunctionCoefficient p_coef(p_exact);
   VectorFunctionCoefficient gradp_coef(sdim, gradp_exact);
-   VectorFunctionCoefficient v_coef(sdim, v_exact);
-   VectorFunctionCoefficient curlv_coef(sdim, curlv_exact);
   FunctionCoefficient divgradp_coef(div_gradp_exact);

   if (prob == 0)
   {
      gftrial.ProjectCoefficient(p_coef);
   }
-   else if (prob == 1)
-   {
-      gftrial.ProjectCoefficient(v_coef);
-   }
   else
   {
      gftrial.ProjectCoefficient(gradp_coef);
@@ -202,11 +179,6 @@ int main(int argc, char *argv[])
      a.AddDomainIntegrator(new VectorFEMassIntegrator(one));
      a_mixed.AddDomainIntegrator(new MixedVectorGradientIntegrator(one));
   }
-   else if (prob == 1)
-   {
-      a.AddDomainIntegrator(new VectorFEMassIntegrator(one));
-      a_mixed.AddDomainIntegrator(new MixedVectorCurlIntegrator(one));
-   }
   else
   {
      a.AddDomainIntegrator(new MassIntegrator(one));
@@ -272,10 +244,6 @@ int main(int argc, char *argv[])
   {
      dlo.AddDomainInterpolator(new GradientInterpolator());
   }
-   else if (prob == 1)
-   {
-      dlo.AddDomainInterpolator(new CurlInterpolator());
-   }
   else
   {
      dlo.AddDomainInterpolator(new DivergenceInterpolator());
@@ -290,10 +258,6 @@ int main(int argc, char *argv[])
   {
      exact_proj.ProjectCoefficient(gradp_coef);
   }
-   else if (prob == 1)
-   {
-      exact_proj.ProjectCoefficient(curlv_coef);
-   }
   else
   {
      exact_proj.ProjectCoefficient(divgradp_coef);
@@ -312,21 +276,8 @@ int main(int argc, char *argv[])
      cout << "\n Solution of (E_h,v) = (grad p_h,v) for E_h and v in H(curl): "
           "|| E_h - grad p ||_{L_2} = " << errSol << '\n' << endl;
      cout << " Gradient interpolant E_h = grad p_h in H(curl): || E_h - grad p"
-           " ||_{L_2} = " << errInterp << '\n' << endl;
-      cout << " Projection E_h of exact grad p in H(curl): || E_h - grad p "
-           "||_{L_2} = " << errProj << '\n' << endl;
-   }
-   else if (prob == 1)
-   {
-      double errSol = x.ComputeL2Error(curlv_coef);
-      double errInterp = discreteInterpolant.ComputeL2Error(curlv_coef);
-      double errProj = exact_proj.ComputeL2Error(curlv_coef);
-
-      cout << "\n Solution of (E_h,w) = (curl v_h,w) for E_h and w in H(div): "
-           "|| E_h - curl v ||_{L_2} = " << errSol << '\n' << endl;
-      cout << " Curl interpolant E_h = curl v_h in H(div): || E_h - curl v "
           "||_{L_2} = " << errInterp << '\n' << endl;
-      cout << " Projection E_h of exact curl v in H(div): || E_h - curl v "
+      cout << " Projection E_h of exact grad p in H(curl): || E_h - grad p "
           "||_{L_2} = " << errProj << '\n' << endl;
   }
   else
@@ -344,7 +295,7 @@ int main(int argc, char *argv[])

      cout << "\n Solution of (f_h,q) = (div v_h,q) for f_h and q in L_2: "
           "|| f_h - div v ||_{L_2} = " << errSol << '\n' << endl;
-      cout << " Divergence interpolant f_h = div v_h in L_2: || f_h - div v "
+      cout << " Divergence interpolant f_h = div v_h in L_2: || f_h - div v"
           "||_{L_2} = " << errInterp << '\n' << endl;
      cout << " Projection f_h of exact div v in L_2: || f_h - div v "
           "||_{L_2} = " << errProj << '\n' << endl;
@@ -420,33 +371,3 @@ double div_gradp_exact(const Vector &x)

   return 0.0;
 }
-
-void v_exact(const Vector &x, Vector &v)
-{
-   if (dim == 3)
-   {
-      v(0) = sin(kappa * x(1));
-      v(1) = sin(kappa * x(2));
-      v(2) = sin(kappa * x(0));
-   }
-   else
-   {
-      v(0) = sin(kappa * x(1));
-      v(1) = sin(kappa * x(0));
-      if (x.Size() == 3) { v(2) = 0.0; }
-   }
-}
-
-void curlv_exact(const Vector &x, Vector &cv)
-{
-   if (dim == 3)
-   {
-      cv(0) = -kappa * cos(kappa * x(2));
-      cv(1) = -kappa * cos(kappa * x(0));
-      cv(2) = -kappa * cos(kappa * x(1));
-   }
-   else
-   {
-      cv = 0.0;
-   }
-}
@@ -6,8 +6,7 @@
 //               mpirun -np 4 ex24p -m ../data/square-disc.mesh -o 2
 //               mpirun -np 4 ex24p -m ../data/beam-tet.mesh
 //               mpirun -np 4 ex24p -m ../data/beam-hex.mesh -o 2 -pa
-//               mpirun -np 4 ex24p -m ../data/beam-hex.mesh -o 2 -pa -p 1
-//               mpirun -np 4 ex24p -m ../data/beam-hex.mesh -o 2 -pa -p 2
+//               mpirun -np 4 ex24p -m ../data/beam-hex.mesh -o 2 -p 1 -pa
 //               mpirun -np 4 ex24p -m ../data/escher.mesh
 //               mpirun -np 4 ex24p -m ../data/escher.mesh -o 2
 //               mpirun -np 4 ex24p -m ../data/fichera.mesh
@@ -25,13 +24,12 @@
 //               mpirun -np 4 ex24p -m ../data/beam-hex.mesh -pa -d cuda
 //
 // Description:  This example code illustrates usage of mixed finite element
-//               spaces, with three variants:
+//               spaces, with two variants:
 //
 //               1) (grad p, u) for p in H^1 tested against u in H(curl)
-//               2) (curl v, u) for v in H(curl) tested against u in H(div), 3D
-//               3) (div v, q) for v in H(div) tested against q in L_2
+//               2) (div v, q) for v in H(div) tested against q in L_2
 //
-//               Using different approaches, we project the gradient, curl, or
+//               Using different approaches, we project the gradient or
 //               divergence to the appropriate space.
 //
 //               We recommend viewing examples 1, 3, and 5 before viewing this
@@ -47,11 +45,8 @@ using namespace mfem;
 double p_exact(const Vector &x);
 void gradp_exact(const Vector &, Vector &);
 double div_gradp_exact(const Vector &x);
-void v_exact(const Vector &x, Vector &v);
-void curlv_exact(const Vector &x, Vector &cv);

 int dim;
-double freq = 1.0, kappa;

 int main(int argc, char *argv[])
 {
@@ -101,7 +96,6 @@ int main(int argc, char *argv[])
   {
      args.PrintOptions(cout);
   }
-   kappa = freq * M_PI;

   // 3. Enable hardware devices such as GPUs, and programming models such as
   //    CUDA, OCCA, RAJA and OpenMP based on command line options.
@@ -153,15 +147,10 @@ int main(int argc, char *argv[])
      trial_fec = new H1_FECollection(order, dim);
      test_fec = new ND_FECollection(order, dim);
   }
-   else if (prob == 1)
-   {
-      trial_fec = new ND_FECollection(order, dim);
-      test_fec = new RT_FECollection(order-1, dim);
-   }
   else
   {
-      trial_fec = new RT_FECollection(order-1, dim);
-      test_fec = new L2_FECollection(order-1, dim);
+      trial_fec = new RT_FECollection(order - 1, dim);
+      test_fec = new L2_FECollection(order - 1, dim);
   }

   ParFiniteElementSpace trial_fes(pmesh, trial_fec);
@@ -177,12 +166,6 @@ int main(int argc, char *argv[])
         cout << "Number of Nedelec finite element unknowns: " << test_size << endl;
         cout << "Number of H1 finite element unknowns: " << trial_size << endl;
      }
-      else if (prob == 1)
-      {
-         cout << "Number of Nedelec finite element unknowns: " << trial_size << endl;
-         cout << "Number of Raviart-Thomas finite element unknowns: " << test_size <<
-              endl;
-      }
      else
      {
         cout << "Number of Raviart-Thomas finite element unknowns: "
@@ -198,18 +181,12 @@ int main(int argc, char *argv[])
   ParGridFunction x(&test_fes);
   FunctionCoefficient p_coef(p_exact);
   VectorFunctionCoefficient gradp_coef(sdim, gradp_exact);
-   VectorFunctionCoefficient v_coef(sdim, v_exact);
-   VectorFunctionCoefficient curlv_coef(sdim, curlv_exact);
   FunctionCoefficient divgradp_coef(div_gradp_exact);

   if (prob == 0)
   {
      gftrial.ProjectCoefficient(p_coef);
   }
-   else if (prob == 1)
-   {
-      gftrial.ProjectCoefficient(v_coef);
-   }
   else
   {
      gftrial.ProjectCoefficient(gradp_coef);
@@ -233,11 +210,6 @@ int main(int argc, char *argv[])
      a.AddDomainIntegrator(new VectorFEMassIntegrator(one));
      a_mixed.AddDomainIntegrator(new MixedVectorGradientIntegrator(one));
   }
-   else if (prob == 1)
-   {
-      a.AddDomainIntegrator(new VectorFEMassIntegrator(one));
-      a_mixed.AddDomainIntegrator(new MixedVectorCurlIntegrator(one));
-   }
   else
   {
      a.AddDomainIntegrator(new MassIntegrator(one));
@@ -321,10 +293,6 @@ int main(int argc, char *argv[])
   {
      dlo.AddDomainInterpolator(new GradientInterpolator());
   }
-   else if (prob == 1)
-   {
-      dlo.AddDomainInterpolator(new CurlInterpolator());
-   }
   else
   {
      dlo.AddDomainInterpolator(new DivergenceInterpolator());
@@ -339,10 +307,6 @@ int main(int argc, char *argv[])
   {
      exact_proj.ProjectCoefficient(gradp_coef);
   }
-   else if (prob == 1)
-   {
-      exact_proj.ProjectCoefficient(curlv_coef);
-   }
   else
   {
      exact_proj.ProjectCoefficient(divgradp_coef);
@@ -360,27 +324,11 @@ int main(int argc, char *argv[])

      if (myid == 0)
      {
-         cout << "\n Solution of (E_h,v) = (grad p_h,v) for E_h and v in H(curl)"
-              ": || E_h - grad p ||_{L_2} = " << errSol << '\n' << endl;
-         cout << " Gradient interpolant E_h = grad p_h in H(curl): || E_h - grad"
-              " p ||_{L_2} = " << errInterp << '\n' << endl;
-         cout << " Projection E_h of exact grad p in H(curl): || E_h - grad p "
-              "||_{L_2} = " << errProj << '\n' << endl;
-      }
-   }
-   else if (prob == 1)
-   {
-      double errSol = x.ComputeL2Error(curlv_coef);
-      double errInterp = discreteInterpolant.ComputeL2Error(curlv_coef);
-      double errProj = exact_proj.ComputeL2Error(curlv_coef);
-
-      if (myid == 0)
-      {
-         cout << "\n Solution of (E_h,w) = (curl v_h,w) for E_h and w in "
-              "H(div): || E_h - curl v ||_{L_2} = " << errSol << '\n' << endl;
-         cout << " Curl interpolant E_h = curl v_h in H(div): || E_h - curl v "
+         cout << "\n Solution of (E_h,v) = (grad p_h,v) for E_h and v in H(curl): "
+              "|| E_h - grad p ||_{L_2} = " << errSol << '\n' << endl;
+         cout << " Gradient interpolant E_h = grad p_h in H(curl): || E_h - grad p"
              "||_{L_2} = " << errInterp << '\n' << endl;
-         cout << " Projection E_h of exact curl v in H(div): || E_h - curl v "
+         cout << " Projection E_h of exact grad p in H(curl): || E_h - grad p "
              "||_{L_2} = " << errProj << '\n' << endl;
      }
   }
@@ -402,7 +350,7 @@ int main(int argc, char *argv[])
         cout << "\n Solution of (f_h,q) = (div v_h,q) for f_h and q in L_2: "
              "|| f_h - div v ||_{L_2} = " << errSol << '\n' << endl;
         cout << " Divergence interpolant f_h = div v_h in L_2: || f_h - div v"
-              " ||_{L_2} = " << errInterp << '\n' << endl;
+              "||_{L_2} = " << errInterp << '\n' << endl;
         cout << " Projection f_h of exact div v in L_2: || f_h - div v "
              "||_{L_2} = " << errProj << '\n' << endl;
      }
@@ -488,33 +436,3 @@ double div_gradp_exact(const Vector &x)

   return 0.0;
 }
-
-void v_exact(const Vector &x, Vector &v)
-{
-   if (dim == 3)
-   {
-      v(0) = sin(kappa * x(1));
-      v(1) = sin(kappa * x(2));
-      v(2) = sin(kappa * x(0));
-   }
-   else
-   {
-      v(0) = sin(kappa * x(1));
-      v(1) = sin(kappa * x(0));
-      if (x.Size() == 3) { v(2) = 0.0; }
-   }
-}
-
-void curlv_exact(const Vector &x, Vector &cv)
-{
-   if (dim == 3)
-   {
-      cv(0) = -kappa * cos(kappa * x(2));
-      cv(1) = -kappa * cos(kappa * x(0));
-      cv(2) = -kappa * cos(kappa * x(1));
-   }
-   else
-   {
-      cv = 0.0;
-   }
-}
@@ -389,22 +389,27 @@ int main(int argc, char *argv[])
   //     applying any necessary transformations such as: assembly, eliminating
   //     boundary conditions, applying conforming constraints for
   //     non-conforming AMR, etc.
-   a.Assemble(0);
+   a.Assemble();

-   OperatorPtr A;
+   OperatorHandle Ah;
   Vector B, X;
-   a.FormLinearSystem(ess_tdof_list, x, b, A, X, B);
+   a.FormLinearSystem(ess_tdof_list, x, b, Ah, X, B);

-   // 13. Solve using a direct or an iterative solver
+   // 13. Transform to monolithic SparseMatrix
+   SparseMatrix *A = Ah.As<ComplexSparseMatrix>()->GetSystemMatrix();
+
+   cout << "Size of linear system: " << A->Height() << endl;
+
+   // 14. Solve using a direct or an iterative solver
 #ifdef MFEM_USE_SUITESPARSE
   {
-      ComplexUMFPackSolver csolver(*A.As<ComplexSparseMatrix>());
-      csolver.Control[UMFPACK_ORDERING] = UMFPACK_ORDERING_METIS;
-      csolver.SetPrintLevel(1);
-      csolver.Mult(B, X);
+      UMFPackSolver  solver(*A);
+      solver.Control[UMFPACK_ORDERING] = UMFPACK_ORDERING_METIS;
+      solver.Mult(B, X);
   }
 #else
-   // 13a. Set up the Bilinear form a(.,.) for the preconditioner
+
+   // 14a. Set up the Bilinear form a(.,.) for the preconditioner
   //
   //    In Comp
   //    Domain:   1/mu (Curl E, Curl F) + omega^2 * epsilon (E,F)
@@ -432,10 +437,10 @@ int main(int argc, char *argv[])

      prec.Assemble();

-      OperatorPtr PCOpAh;
+      OperatorHandle PCOpAh;
      prec.FormSystemMatrix(ess_tdof_list, PCOpAh);

-      // 13b. Define and apply a GMRES solver for AU=B with a block diagonal
+      // 14b. Define and apply a GMRES solver for AU=B with a block diagonal
      //      preconditioner based on the Gauss-Seidel sparse smoother.
      Array<int> offsets(3);
      offsets[0] = 0;
@@ -462,15 +467,17 @@ int main(int argc, char *argv[])
   }
 #endif

-   // 14. Recover the solution as a finite element grid function and compute the
+   // 15. Recover the solution as a finite element grid function and compute the
   //     errors if the exact solution is known.
   a.RecoverFEMSolution(X, b, x);

   // If exact is known compute the error
   if (exact_known)
   {
+      ComplexGridFunction x_gf(fespace);
      VectorFunctionCoefficient E_ex_Re(dim, E_exact_Re);
      VectorFunctionCoefficient E_ex_Im(dim, E_exact_Im);
+      x_gf.ProjectCoefficient(E_ex_Re, E_ex_Im);
      int order_quad = max(2, 2 * order + 1);
      const IntegrationRule *irs[Geometry::NumGeom];
      for (int i = 0; i < Geometry::NumGeom; ++i)
@@ -499,7 +506,7 @@ int main(int argc, char *argv[])
           << sqrt(L2Error_Re*L2Error_Re + L2Error_Im*L2Error_Im) << "\n\n";
   }

-   // 15. Save the refined mesh and the solution. This output can be viewed
+   // 16. Save the refined mesh and the solution. This output can be viewed
   //     later using GLVis: "glvis -m mesh -g sol".
   {
      ofstream mesh_ofs("ex25.mesh");
@@ -514,7 +521,7 @@ int main(int argc, char *argv[])
      x.imag().Save(sol_i_ofs);
   }

-   // 16. Send the solution by socket to a GLVis server.
+   // 17. Send the solution by socket to a GLVis server.
   if (visualization)
   {
      // Define visualization keys for GLVis (see GLVis documentation)
@@ -565,7 +572,8 @@ int main(int argc, char *argv[])
      }
   }

-   // 17. Free the used memory.
+   // 18. Free the used memory.
+   delete A;
   delete pml;
   delete fespace;
   delete fec;
@@ -419,15 +419,21 @@ int main(int argc, char *argv[])
   //     constraints for non-conforming AMR, etc.
   a.Assemble();

-   OperatorPtr Ah;
+   OperatorHandle Ah;
   Vector B, X;
   a.FormLinearSystem(ess_tdof_list, x, b, Ah, X, B);

-   // 15. Solve using a direct or an iterative solver
+   // 15. Transform to monolithic HypreParMatrix
+   HypreParMatrix *A = Ah.As<ComplexHypreParMatrix>()->GetSystemMatrix();
+
+   if (myid == 0)
+   {
+      cout << "Size of linear system: " << A->GetGlobalNumRows() << endl;
+   }
+
+   // 16. Solve using a direct or an iterative solver
 #ifdef MFEM_USE_SUPERLU
   {
-      // Transform to monolithic HypreParMatrix
-      HypreParMatrix *A = Ah.As<ComplexHypreParMatrix>()->GetSystemMatrix();
      SuperLURowLocMatrix SA(*A);
      SuperLUSolver superlu(MPI_COMM_WORLD);
      superlu.SetPrintStatistics(false);
@@ -435,9 +441,9 @@ int main(int argc, char *argv[])
      superlu.SetColumnPermutation(superlu::PARMETIS);
      superlu.SetOperator(SA);
      superlu.Mult(B, X);
-      delete A;
   }
 #else
+
   // 16a. Set up the parallel Bilinear form a(.,.) for the preconditioner
   //
   //    In Comp
@@ -466,7 +472,7 @@ int main(int argc, char *argv[])

      prec.Assemble();

-      OperatorPtr PCOpAh;
+      OperatorHandle PCOpAh;
      prec.FormSystemMatrix(ess_tdof_list, PCOpAh);

      // 16b. Define and apply a parallel GMRES solver for AU=B with a block
@@ -490,7 +496,7 @@ int main(int argc, char *argv[])
      gmres.SetMaxIter(2000);
      gmres.SetRelTol(1e-5);
      gmres.SetAbsTol(0.0);
-      gmres.SetOperator(*Ah);
+      gmres.SetOperator(*A);
      gmres.SetPreconditioner(BlockAMS);
      gmres.Mult(B, X);
   }
@@ -503,8 +509,10 @@ int main(int argc, char *argv[])
   // If exact is known compute the error
   if (exact_known)
   {
+      ParComplexGridFunction x_gf(fespace);
      VectorFunctionCoefficient E_ex_Re(dim, E_exact_Re);
      VectorFunctionCoefficient E_ex_Im(dim, E_exact_Im);
+      x_gf.ProjectCoefficient(E_ex_Re, E_ex_Im);
      int order_quad = max(2, 2 * order + 1);
      const IntegrationRule *irs[Geometry::NumGeom];
      for (int i = 0; i < Geometry::NumGeom; ++i)
@@ -621,6 +629,7 @@ int main(int argc, char *argv[])
   }

   // 20. Free the used memory.
+   delete A;
   delete pml;
   delete fespace;
   delete fec;
@@ -11,12 +11,6 @@
 //               ex5 -m ../data/escher.mesh
 //               ex5 -m ../data/fichera.mesh
 //
-// Device sample runs:
-//               ex5 -m ../data/star.mesh -pa -d cuda
-//               ex5 -m ../data/star.mesh -pa -d raja-cuda
-//               ex5 -m ../data/star.mesh -pa -d raja-omp
-//               ex5 -m ../data/beam-hex.mesh -pa -d cuda
-//
 // Description:  This example code solves a simple 2D/3D mixed Darcy problem
 //               corresponding to the saddle point system
 //                                 k*u + grad p = f
@@ -56,7 +50,6 @@ int main(int argc, char *argv[])
   const char *mesh_file = "../data/star.mesh";
   int order = 1;
   bool pa = false;
-   const char *device_config = "cpu";
   bool visualization = 1;

   OptionsParser args(argc, argv);
@@ -66,8 +59,6 @@ int main(int argc, char *argv[])
                  "Finite element order (polynomial degree).");
   args.AddOption(&pa, "-pa", "--partial-assembly", "-no-pa",
                  "--no-partial-assembly", "Enable Partial Assembly.");
-   args.AddOption(&device_config, "-d", "--device",
-                  "Device configuration string, see Device::Configure().");
   args.AddOption(&visualization, "-vis", "--visualization", "-no-vis",
                  "--no-visualization",
                  "Enable or disable GLVis visualization.");
@@ -79,18 +70,13 @@ int main(int argc, char *argv[])
   }
   args.PrintOptions(cout);

-   // 2. Enable hardware devices such as GPUs, and programming models such as
-   //    CUDA, OCCA, RAJA and OpenMP based on command line options.
-   Device device(device_config);
-   device.Print();
-
-   // 3. Read the mesh from the given mesh file. We can handle triangular,
+   // 2. Read the mesh from the given mesh file. We can handle triangular,
   //    quadrilateral, tetrahedral, hexahedral, surface and volume meshes with
   //    the same code.
   Mesh *mesh = new Mesh(mesh_file, 1, 1);
   int dim = mesh->Dimension();

-   // 4. Refine the mesh to increase the resolution. In this example we do
+   // 3. Refine the mesh to increase the resolution. In this example we do
   //    'ref_levels' of uniform refinement. We choose 'ref_levels' to be the
   //    largest number that gives a final mesh with no more than 10,000
   //    elements.
@@ -103,7 +89,7 @@ int main(int argc, char *argv[])
      }
   }

-   // 5. Define a finite element space on the mesh. Here we use the
+   // 4. Define a finite element space on the mesh. Here we use the
   //    Raviart-Thomas finite elements of the specified order.
   FiniteElementCollection *hdiv_coll(new RT_FECollection(order, dim));
   FiniteElementCollection *l2_coll(new L2_FECollection(order, dim));
@@ -111,7 +97,7 @@ int main(int argc, char *argv[])
   FiniteElementSpace *R_space = new FiniteElementSpace(mesh, hdiv_coll);
   FiniteElementSpace *W_space = new FiniteElementSpace(mesh, l2_coll);

-   // 6. Define the BlockStructure of the problem, i.e. define the array of
+   // 5. Define the BlockStructure of the problem, i.e. define the array of
   //    offsets for each variable. The last component of the Array is the sum
   //    of the dimensions of each block.
   Array<int> block_offsets(3); // number of variables + 1
@@ -126,7 +112,7 @@ int main(int argc, char *argv[])
   std::cout << "dim(R+W) = " << block_offsets.Last() << "\n";
   std::cout << "***********************************************************\n";

-   // 7. Define the coefficients, analytical solution, and rhs of the PDE.
+   // 6. Define the coefficients, analytical solution, and rhs of the PDE.
   ConstantCoefficient k(1.0);

   VectorFunctionCoefficient fcoeff(dim, fFun);
@@ -136,28 +122,25 @@ int main(int argc, char *argv[])
   VectorFunctionCoefficient ucoeff(dim, uFun_ex);
   FunctionCoefficient pcoeff(pFun_ex);

-   // 8. Allocate memory (x, rhs) for the analytical solution and the right hand
+   // 7. Allocate memory (x, rhs) for the analytical solution and the right hand
   //    side.  Define the GridFunction u,p for the finite element solution and
   //    linear forms fform and gform for the right hand side.  The data
   //    allocated by x and rhs are passed as a reference to the grid functions
   //    (u,p) and the linear forms (fform, gform).
-   MemoryType mt = device.GetMemoryType();
-   BlockVector x(block_offsets, mt), rhs(block_offsets, mt);
+   BlockVector x(block_offsets), rhs(block_offsets);

   LinearForm *fform(new LinearForm);
   fform->Update(R_space, rhs.GetBlock(0), 0);
   fform->AddDomainIntegrator(new VectorFEDomainLFIntegrator(fcoeff));
   fform->AddBoundaryIntegrator(new VectorFEBoundaryFluxLFIntegrator(fnatcoeff));
   fform->Assemble();
-   fform->SyncAliasMemory(rhs);

   LinearForm *gform(new LinearForm);
   gform->Update(W_space, rhs.GetBlock(1), 0);
   gform->AddDomainIntegrator(new DomainLFIntegrator(gcoeff));
   gform->Assemble();
-   gform->SyncAliasMemory(rhs);

-   // 9. Assemble the finite element matrices for the Darcy operator
+   // 8. Assemble the finite element matrices for the Darcy operator
   //
   //                            D = [ M  B^T ]
   //                                [ B   0  ]
@@ -202,7 +185,7 @@ int main(int argc, char *argv[])
      darcyOp.SetBlock(1,0, &B);
   }

-   // 10. Construct the operators for preconditioner
+   // 9. Construct the operators for preconditioner
   //
   //                 P = [ diag(M)         0         ]
   //                     [  0       B diag(M)^-1 B^T ]
@@ -219,11 +202,10 @@ int main(int argc, char *argv[])
   if (pa)
   {
      mVarf->AssembleDiagonal(Md);
-      auto Md_host = Md.HostRead();
      Vector invMd(mVarf->Height());
      for (int i=0; i<mVarf->Height(); ++i)
      {
-         invMd(i) = 1.0 / Md_host[i];
+         invMd(i) = 1.0 / Md(i);
      }

      Vector BMBt_diag(bVarf->Height());
@@ -264,7 +246,7 @@ int main(int argc, char *argv[])
   darcyPrec.SetDiagonalBlock(0, invM);
   darcyPrec.SetDiagonalBlock(1, invS);

-   // 11. Solve the linear system with MINRES.
+   // 10. Solve the linear system with MINRES.
   //     Check the norm of the unpreconditioned residual.
   int maxIter(1000);
   double rtol(1.e-6);
@@ -281,7 +263,6 @@ int main(int argc, char *argv[])
   solver.SetPrintLevel(1);
   x = 0.0;
   solver.Mult(rhs, x);
-   if (device.IsEnabled()) { x.HostRead(); }
   chrono.Stop();

   if (solver.GetConverged())
@@ -292,7 +273,7 @@ int main(int argc, char *argv[])
                << " iterations. Residual norm is " << solver.GetFinalNorm() << ".\n";
   std::cout << "MINRES solver took " << chrono.RealTime() << "s. \n";

-   // 12. Create the grid functions u and p. Compute the L2 error norms.
+   // 11. Create the grid functions u and p. Compute the L2 error norms.
   GridFunction u, p;
   u.MakeRef(R_space, x.GetBlock(0), 0);
   p.MakeRef(W_space, x.GetBlock(1), 0);
@@ -312,7 +293,7 @@ int main(int argc, char *argv[])
   std::cout << "|| u_h - u_ex || / || u_ex || = " << err_u / norm_u << "\n";
   std::cout << "|| p_h - p_ex || / || p_ex || = " << err_p / norm_p << "\n";

-   // 13. Save the mesh and the solution. This output can be viewed later using
+   // 12. Save the mesh and the solution. This output can be viewed later using
   //     GLVis: "glvis -m ex5.mesh -g sol_u.gf" or "glvis -m ex5.mesh -g
   //     sol_p.gf".
   {
@@ -329,13 +310,13 @@ int main(int argc, char *argv[])
      p.Save(p_ofs);
   }

-   // 14. Save data in the VisIt format
+   // 13. Save data in the VisIt format
   VisItDataCollection visit_dc("Example5", mesh);
   visit_dc.RegisterField("velocity", &u);
   visit_dc.RegisterField("pressure", &p);
   visit_dc.Save();

-   // 15. Save data in the ParaView format
+   // 14. Save data in the ParaView format
   ParaViewDataCollection paraview_dc("Example5", mesh);
   paraview_dc.SetPrefixPath("ParaView");
   paraview_dc.SetLevelsOfDetail(order);
@@ -347,7 +328,7 @@ int main(int argc, char *argv[])
   paraview_dc.RegisterField("pressure",&p);
   paraview_dc.Save();

-   // 16. Send the solution by socket to a GLVis server.
+   // 15. Send the solution by socket to a GLVis server.
   if (visualization)
   {
      char vishost[] = "localhost";
@@ -360,7 +341,7 @@ int main(int argc, char *argv[])
      p_sock << "solution\n" << *mesh << p << "window_title 'Pressure'" << endl;
   }

-   // 17. Free the used memory.
+   // 16. Free the used memory.
   delete fform;
   delete gform;
   delete invM;
@@ -11,12 +11,6 @@
 //               mpirun -np 4 ex5p -m ../data/escher.mesh
 //               mpirun -np 4 ex5p -m ../data/fichera.mesh
 //
-// Device sample runs:
-//               mpirun -np 4 ex5p -m ../data/star.mesh -r 2 -pa -d cuda
-//               mpirun -np 4 ex5p -m ../data/star.mesh -r 2 -pa -d raja-cuda
-//               mpirun -np 4 ex5p -m ../data/star.mesh -r 2 -pa -d raja-omp
-//               mpirun -np 4 ex5p -m ../data/beam-hex.mesh -pa -d cuda
-//
 // Description:  This example code solves a simple 2D/3D mixed Darcy problem
 //               corresponding to the saddle point system
 //                                 k*u + grad p = f
@@ -66,7 +60,6 @@ int main(int argc, char *argv[])
   int order = 1;
   bool par_format = false;
   bool pa = false;
-   const char *device_config = "cpu";
   bool visualization = 1;
   bool adios2 = false;

@@ -82,8 +75,6 @@ int main(int argc, char *argv[])
                  "Format to use when saving the results for VisIt.");
   args.AddOption(&pa, "-pa", "--partial-assembly", "-no-pa",
                  "--no-partial-assembly", "Enable Partial Assembly.");
-   args.AddOption(&device_config, "-d", "--device",
-                  "Device configuration string, see Device::Configure().");
   args.AddOption(&visualization, "-vis", "--visualization", "-no-vis",
                  "--no-visualization",
                  "Enable or disable GLVis visualization.");
@@ -105,18 +96,13 @@ int main(int argc, char *argv[])
      args.PrintOptions(cout);
   }

-   // 3. Enable hardware devices such as GPUs, and programming models such as
-   //    CUDA, OCCA, RAJA and OpenMP based on command line options.
-   Device device(device_config);
-   if (myid == 0) { device.Print(); }
-
-   // 4. Read the (serial) mesh from the given mesh file on all processors.  We
+   // 3. Read the (serial) mesh from the given mesh file on all processors.  We
   //    can handle triangular, quadrilateral, tetrahedral, hexahedral, surface
   //    and volume meshes with the same code.
   Mesh *mesh = new Mesh(mesh_file, 1, 1);
   int dim = mesh->Dimension();

-   // 5. Refine the serial mesh on all processors to increase the resolution. In
+   // 4. Refine the serial mesh on all processors to increase the resolution. In
   //    this example we do 'ref_levels' of uniform refinement. We choose
   //    'ref_levels' to be the largest number that gives a final mesh with no
   //    more than 10,000 elements, unless the user specifies it as input.
@@ -132,7 +118,7 @@ int main(int argc, char *argv[])
      }
   }

-   // 6. Define a parallel mesh by a partitioning of the serial mesh. Refine
+   // 5. Define a parallel mesh by a partitioning of the serial mesh. Refine
   //    this mesh further in parallel to increase the resolution. Once the
   //    parallel mesh is defined, the serial mesh can be deleted.
   ParMesh *pmesh = new ParMesh(MPI_COMM_WORLD, *mesh);
@@ -145,7 +131,7 @@ int main(int argc, char *argv[])
      }
   }

-   // 7. Define a parallel finite element space on the parallel mesh. Here we
+   // 6. Define a parallel finite element space on the parallel mesh. Here we
   //    use the Raviart-Thomas finite elements of the specified order.
   FiniteElementCollection *hdiv_coll(new RT_FECollection(order, dim));
   FiniteElementCollection *l2_coll(new L2_FECollection(order, dim));
@@ -165,7 +151,7 @@ int main(int argc, char *argv[])
      std::cout << "***********************************************************\n";
   }

-   // 8. Define the two BlockStructure of the problem.  block_offsets is used
+   // 7. Define the two BlockStructure of the problem.  block_offsets is used
   //    for Vector based on dof (like ParGridFunction or ParLinearForm),
   //    block_trueOffstes is used for Vector based on trueDof (HypreParVector
   //    for the rhs and solution of the linear system).  The offsets computed
@@ -182,7 +168,7 @@ int main(int argc, char *argv[])
   block_trueOffsets[2] = W_space->TrueVSize();
   block_trueOffsets.PartialSum();

-   // 9. Define the coefficients, analytical solution, and rhs of the PDE.
+   // 8. Define the coefficients, analytical solution, and rhs of the PDE.
   ConstantCoefficient k(1.0);

   VectorFunctionCoefficient fcoeff(dim, fFun);
@@ -192,30 +178,25 @@ int main(int argc, char *argv[])
   VectorFunctionCoefficient ucoeff(dim, uFun_ex);
   FunctionCoefficient pcoeff(pFun_ex);

-   // 10. Define the parallel grid function and parallel linear forms, solution
-   //     vector and rhs.
-   MemoryType mt = device.GetMemoryType();
-   BlockVector x(block_offsets, mt), rhs(block_offsets, mt);
-   BlockVector trueX(block_trueOffsets, mt), trueRhs(block_trueOffsets, mt);
+   // 9. Define the parallel grid function and parallel linear forms, solution
+   //    vector and rhs.
+   BlockVector x(block_offsets), rhs(block_offsets);
+   BlockVector trueX(block_trueOffsets), trueRhs(block_trueOffsets);

   ParLinearForm *fform(new ParLinearForm);
   fform->Update(R_space, rhs.GetBlock(0), 0);
   fform->AddDomainIntegrator(new VectorFEDomainLFIntegrator(fcoeff));
   fform->AddBoundaryIntegrator(new VectorFEBoundaryFluxLFIntegrator(fnatcoeff));
   fform->Assemble();
-   fform->SyncAliasMemory(rhs);
   fform->ParallelAssemble(trueRhs.GetBlock(0));
-   trueRhs.GetBlock(0).SyncAliasMemory(trueRhs);

   ParLinearForm *gform(new ParLinearForm);
   gform->Update(W_space, rhs.GetBlock(1), 0);
   gform->AddDomainIntegrator(new DomainLFIntegrator(gcoeff));
   gform->Assemble();
-   gform->SyncAliasMemory(rhs);
   gform->ParallelAssemble(trueRhs.GetBlock(1));
-   trueRhs.GetBlock(1).SyncAliasMemory(trueRhs);

-   // 11. Assemble the finite element matrices for the Darcy operator
+   // 10. Assemble the finite element matrices for the Darcy operator
   //
   //                            D = [ M  B^T ]
   //                                [ B   0  ]
@@ -268,7 +249,7 @@ int main(int argc, char *argv[])
      darcyOp->SetBlock(1,0, B);
   }

-   // 12. Construct the operators for preconditioner
+   // 11. Construct the operators for preconditioner
   //
   //                 P = [ diag(M)         0         ]
   //                     [  0       B diag(M)^-1 B^T ]
@@ -285,11 +266,10 @@ int main(int argc, char *argv[])
   {
      Md_PA.SetSize(R_space->GetTrueVSize());
      mVarf->AssembleDiagonal(Md_PA);
-      auto Md_host = Md_PA.HostRead();
      Vector invMd(Md_PA.Size());
      for (int i=0; i<Md_PA.Size(); ++i)
      {
-         invMd(i) = 1.0 / Md_host[i];
+         invMd(i) = 1.0 / Md_PA(i);
      }

      Vector BMBt_diag(W_space->GetTrueVSize());
@@ -322,7 +302,7 @@ int main(int argc, char *argv[])
   darcyPr->SetDiagonalBlock(0, invM);
   darcyPr->SetDiagonalBlock(1, invS);

-   // 13. Solve the linear system with MINRES.
+   // 12. Solve the linear system with MINRES.
   //     Check the norm of the unpreconditioned residual.
   int maxIter(pa ? 1000 : 500);
   double rtol(1.e-6);
@@ -339,7 +319,6 @@ int main(int argc, char *argv[])
   solver.SetPrintLevel(verbose);
   trueX = 0.0;
   solver.Mult(trueRhs, trueX);
-   if (device.IsEnabled()) { trueX.HostRead(); }
   chrono.Stop();

   if (verbose)
@@ -353,7 +332,7 @@ int main(int argc, char *argv[])
      std::cout << "MINRES solver took " << chrono.RealTime() << "s. \n";
   }

-   // 14. Extract the parallel grid function corresponding to the finite element
+   // 13. Extract the parallel grid function corresponding to the finite element
   //     approximation X. This is the local solution on each processor. Compute
   //     L2 error norms.
   ParGridFunction *u(new ParGridFunction);
@@ -381,7 +360,7 @@ int main(int argc, char *argv[])
      std::cout << "|| p_h - p_ex || / || p_ex || = " << err_p / norm_p << "\n";
   }

-   // 15. Save the refined mesh and the solution in parallel. This output can be
+   // 14. Save the refined mesh and the solution in parallel. This output can be
   //     viewed later using GLVis: "glvis -np <np> -m mesh -g sol_*".
   {
      ostringstream mesh_name, u_name, p_name;
@@ -402,7 +381,7 @@ int main(int argc, char *argv[])
      p->Save(p_ofs);
   }

-   // 16. Save data in the VisIt format
+   // 15. Save data in the VisIt format
   VisItDataCollection visit_dc("Example5-Parallel", pmesh);
   visit_dc.RegisterField("velocity", u);
   visit_dc.RegisterField("pressure", p);
@@ -411,7 +390,7 @@ int main(int argc, char *argv[])
                      DataCollection::PARALLEL_FORMAT);
   visit_dc.Save();

-   // 17. Save data in the ParaView format
+   // 16. Save data in the ParaView format
   ParaViewDataCollection paraview_dc("Example5P", pmesh);
   paraview_dc.SetPrefixPath("ParaView");
   paraview_dc.SetLevelsOfDetail(order);
@@ -423,7 +402,7 @@ int main(int argc, char *argv[])
   paraview_dc.RegisterField("pressure",p);
   paraview_dc.Save();

-   // 18. Optionally output a BP (binary pack) file using ADIOS2. This can be
+   // 17. Optionally output a BP (binary pack) file using ADIOS2. This can be
   //     visualized with the ParaView VTX reader.
 #ifdef MFEM_USE_ADIOS2
   if (adios2)
@@ -443,7 +422,7 @@ int main(int argc, char *argv[])
   }
 #endif

-   // 19. Send the solution by socket to a GLVis server.
+   // 18. Send the solution by socket to a GLVis server.
   if (visualization)
   {
      char vishost[] = "localhost";
@@ -463,7 +442,7 @@ int main(int argc, char *argv[])
             << endl;
   }

-   // 20. Free the used memory.
+   // 19. Free the used memory.
   delete fform;
   delete gform;
   delete u;
@@ -20,7 +20,7 @@
 //               ex6 -pa -d occa-cuda
 //               ex6 -pa -d raja-omp
 //               ex6 -pa -d ceed-cpu
-//             * ex6 -pa -d ceed-cuda
+//               * ex6 -pa -d ceed-cuda
 //               ex6 -pa -d ceed-cuda:/gpu/cuda/shared
 //
 // Description:  This is a version of Example 1 with a simple adaptive mesh
@@ -20,7 +20,7 @@
 //               mpirun -np 4 ex6p -pa -d occa-cuda
 //               mpirun -np 4 ex6p -pa -d raja-omp
 //               mpirun -np 4 ex6p -pa -d ceed-cpu
-//             * mpirun -np 4 ex6p -pa -d ceed-cuda
+//               * mpirun -np 4 ex6p -pa -d ceed-cuda
 //               mpirun -np 4 ex6p -pa -d ceed-cuda:/gpu/cuda/shared
 //
 // Description:  This is a version of Example 1 with a simple adaptive mesh
@@ -20,11 +20,8 @@
 // Device sample runs:
 //    ex9 -pa
 //    ex9 -ea
-//    ex9 -fa
 //    ex9 -pa -m ../data/periodic-cube.mesh
 //    ex9 -pa -m ../data/periodic-cube.mesh -d cuda
-//    ex9 -ea -m ../data/periodic-cube.mesh -d cuda
-//    ex9 -fa -m ../data/periodic-cube.mesh -d cuda
 //
 // Description:  This example code solves the time-dependent advection equation
 //               du/dt + v.grad(u) = 0, where v is a given fluid velocity, and
@@ -147,7 +144,6 @@ int main(int argc, char *argv[])
   int order = 3;
   bool pa = false;
   bool ea = false;
-   bool fa = false;
   const char *device_config = "cpu";
   int ode_solver_type = 4;
   double t_final = 10.0;
@@ -174,8 +170,6 @@ int main(int argc, char *argv[])
                  "--no-partial-assembly", "Enable Partial Assembly.");
   args.AddOption(&ea, "-ea", "--element-assembly", "-no-ea",
                  "--no-element-assembly", "Enable Element Assembly.");
-   args.AddOption(&fa, "-fa", "--full-assembly", "-no-fa",
-                  "--no-full-assembly", "Enable Full Assembly.");
   args.AddOption(&device_config, "-d", "--device",
                  "Device configuration string, see Device::Configure().");
   args.AddOption(&ode_solver_type, "-s", "--ode-solver",
@@ -284,11 +278,6 @@ int main(int argc, char *argv[])
      m.SetAssemblyLevel(AssemblyLevel::ELEMENT);
      k.SetAssemblyLevel(AssemblyLevel::ELEMENT);
   }
-   else if (fa)
-   {
-      m.SetAssemblyLevel(AssemblyLevel::FULL);
-      k.SetAssemblyLevel(AssemblyLevel::FULL);
-   }
   m.AddDomainIntegrator(new MassIntegrator);
   k.AddDomainIntegrator(new ConvectionIntegrator(velocity, -1.0));
   k.AddInteriorFaceIntegrator(
@@ -448,19 +437,21 @@ int main(int argc, char *argv[])
 FE_Evolution::FE_Evolution(BilinearForm &_M, BilinearForm &_K, const Vector &_b)
   : TimeDependentOperator(_M.Height()), M(_M), K(_K), b(_b), z(_M.Height())
 {
+   bool pa = M.GetAssemblyLevel() == AssemblyLevel::PARTIAL;
+   bool ea = M.GetAssemblyLevel() == AssemblyLevel::ELEMENT;
   Array<int> ess_tdof_list;
-   if (M.GetAssemblyLevel() == AssemblyLevel::LEGACYFULL)
-   {
-      M_prec = new DSmoother(M.SpMat());
-      M_solver.SetOperator(M.SpMat());
-      dg_solver = new DG_Solver(M.SpMat(), K.SpMat(), *M.FESpace());
-   }
-   else
+   if (pa || ea)
   {
      M_prec = new OperatorJacobiSmoother(M, ess_tdof_list);
      M_solver.SetOperator(M);
      dg_solver = NULL;
   }
+   else
+   {
+      M_prec = new DSmoother(M.SpMat());
+      dg_solver = new DG_Solver(M.SpMat(), K.SpMat(), *M.FESpace());
+      M_solver.SetOperator(M.SpMat());
+   }
   M_solver.SetPreconditioner(*M_prec);
   M_solver.iterative_mode = false;
   M_solver.SetRelTol(1e-9);
@@ -21,11 +21,8 @@
 // Device sample runs:
 //    mpirun -np 4 ex9p -pa
 //    mpirun -np 4 ex9p -ea
-//    mpirun -np 4 ex9p -fa
 //    mpirun -np 4 ex9p -pa -m ../data/periodic-cube.mesh
 //    mpirun -np 4 ex9p -pa -m ../data/periodic-cube.mesh -d cuda
-//    mpirun -np 4 ex9p -ea -m ../data/periodic-cube.mesh -d cuda
-//    mpirun -np 4 ex9p -fa -m ../data/periodic-cube.mesh -d cuda
 //
 // Description:  This example code solves the time-dependent advection equation
 //               du/dt + v.grad(u) = 0, where v is a given fluid velocity, and
@@ -167,7 +164,6 @@ int main(int argc, char *argv[])
   int order = 3;
   bool pa = false;
   bool ea = false;
-   bool fa = false;
   const char *device_config = "cpu";
   int ode_solver_type = 4;
   double t_final = 10.0;
@@ -197,8 +193,6 @@ int main(int argc, char *argv[])
                  "--no-partial-assembly", "Enable Partial Assembly.");
   args.AddOption(&ea, "-ea", "--element-assembly", "-no-ea",
                  "--no-element-assembly", "Enable Element Assembly.");
-   args.AddOption(&fa, "-fa", "--full-assembly", "-no-fa",
-                  "--no-full-assembly", "Enable Full Assembly.");
   args.AddOption(&device_config, "-d", "--device",
                  "Device configuration string, see Device::Configure().");
   args.AddOption(&ode_solver_type, "-s", "--ode-solver",
@@ -335,12 +329,6 @@ int main(int argc, char *argv[])
      m->SetAssemblyLevel(AssemblyLevel::ELEMENT);
      k->SetAssemblyLevel(AssemblyLevel::ELEMENT);
   }
-   else if (fa)
-   {
-      m->SetAssemblyLevel(AssemblyLevel::FULL);
-      k->SetAssemblyLevel(AssemblyLevel::FULL);
-   }
-
   m->AddDomainIntegrator(new MassIntegrator);
   k->AddDomainIntegrator(new ConvectionIntegrator(velocity, -1.0));
   k->AddInteriorFaceIntegrator(
@@ -577,21 +565,29 @@ FE_Evolution::FE_Evolution(ParBilinearForm &_M, ParBilinearForm &_K,
     M_solver(_M.ParFESpace()->GetComm()),
     z(_M.Height())
 {
-   if (_M.GetAssemblyLevel()==AssemblyLevel::LEGACYFULL)
-   {
-      M.Reset(_M.ParallelAssemble(), true);
-      K.Reset(_K.ParallelAssemble(), true);
-   }
-   else
+   bool pa = _M.GetAssemblyLevel()==AssemblyLevel::PARTIAL;
+   bool ea = _M.GetAssemblyLevel()==AssemblyLevel::ELEMENT;
+
+   if (pa || ea)
   {
      M.Reset(&_M, false);
      K.Reset(&_K, false);
   }
+   else
+   {
+      M.Reset(_M.ParallelAssemble(), true);
+      K.Reset(_K.ParallelAssemble(), true);
+   }

   M_solver.SetOperator(*M);

   Array<int> ess_tdof_list;
-   if (_M.GetAssemblyLevel()==AssemblyLevel::LEGACYFULL)
+   if (pa || ea)
+   {
+      M_prec = new OperatorJacobiSmoother(_M, ess_tdof_list);
+      dg_solver = NULL;
+   }
+   else
   {
      HypreParMatrix &M_mat = *M.As<HypreParMatrix>();
      HypreParMatrix &K_mat = *K.As<HypreParMatrix>();
@@ -600,11 +596,6 @@ FE_Evolution::FE_Evolution(ParBilinearForm &_M, ParBilinearForm &_K,

      dg_solver = new DG_Solver(M_mat, K_mat, *_M.FESpace());
   }
-   else
-   {
-      M_prec = new OperatorJacobiSmoother(_M, ess_tdof_list);
-      dg_solver = NULL;
-   }

   M_solver.SetPreconditioner(*M_prec);
   M_solver.iterative_mode = false;
@@ -114,11 +114,6 @@ ex11p-test-strumpack: ex11p
 	@$(call mfem-test,$<, $(RUN_MPI), STRUMPACK example,--strumpack)
 test-par-YES: ex11p-test-strumpack
 endif
-ifeq ($(MFEM_USE_SUPERLU),YES)
-ex11p-test-superlu: ex11p
-	@$(call mfem-test,$<, $(RUN_MPI), SuperLU_DIST example,--superlu)
-test-par-YES: ex11p-test-superlu
-endif

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

@@ -34,15 +34,6 @@ if (MFEM_USE_MPI)
    )
 endif()

-if (MFEM_USE_SLEPC)
-  list(APPEND PETSC_EXAMPLES_SRCS
-    ex11p.cpp
-    )
-  list(APPEND PETSC_RC_FILES
-    rc_ex11p_lobpcg rc_ex11p_gd
-    )
-endif()
-
 # Include the source directory where mfem.hpp and mfem-performance.hpp are.
 include_directories(BEFORE ${PROJECT_BINARY_DIR})

@@ -87,22 +78,12 @@ set(EX9_E_ARGS       -m ../../data/periodic-hexagon.mesh --usepetsc --petscopts
 set(EX9_ES_ARGS      -m ../../data/periodic-hexagon.mesh --usepetsc --petscopts rc_ex9p_expl --no-step)
 set(EX9_IS_ARGS      -m ../../data/periodic-hexagon.mesh --usepetsc --petscopts rc_ex9p_impl --implicit -tf 0.5)
 set(EX10_ARGS        -m ../../data/beam-quad.mesh --usepetsc --petscopts rc_ex10p -tf 30 -s 3 -rs 2 -dt 3)
-if (MFEM_USE_SLEPC)
-  set(EX11_ARGS_SINV   -m ../../data/star.mesh --useslepc)
-  set(EX11_ARGS_LOBPCG -m ../../data/star.mesh --useslepc --slepcopts rc_ex11p_lobpcg)
-  set(EX11_ARGS_GD -m ../../data/star.mesh --useslepc --slepcopts rc_ex11p_gd)
-endif()

 # Add the tests: one test per command-line-variable.
-set(TEST_OPTIONS_VARS
-    EX1_ARGS_W EX1_ARGS_P EX2_ARGS EX3_ARGS EX4_ARGS EX4_HYB_ARGS
-    EX5_BDDC_LB_ARGS EX5_BDDC_GB_ARGS EX5_FSPL_ARGS EX6_ARGS EX6_NONOVL_ARGS
-    EX9_E_ARGS EX9_ES_ARGS EX9_IS_ARGS EX10_ARGS)
-if (MFEM_USE_SLEPC)
-    list(APPEND TEST_OPTIONS_VARS EX11_ARGS_SINV EX11_ARGS_LOBPCG EX11_ARGS_GD)
-endif()
-
-foreach(TEST_OPTIONS_VAR ${TEST_OPTIONS_VARS})
+foreach(TEST_OPTIONS_VAR
+        EX1_ARGS_W EX1_ARGS_P EX2_ARGS EX3_ARGS EX4_ARGS EX4_HYB_ARGS
+        EX5_BDDC_LB_ARGS EX5_BDDC_GB_ARGS EX5_FSPL_ARGS EX6_ARGS EX6_NONOVL_ARGS
+        EX9_E_ARGS EX9_ES_ARGS EX9_IS_ARGS EX10_ARGS)
  string(REGEX REPLACE "^(.+)_ARGS" "\\1" TEST_NAME_UC ${TEST_OPTIONS_VAR})
  string(REGEX REPLACE "^([^_]+)" "\\1P" TEST_NAME_UC ${TEST_NAME_UC})
  string(TOLOWER ${TEST_NAME_UC} TEST_NAME_FULL)
@@ -1,440 +0,0 @@
-//                       MFEM Example 11 - Parallel Version
-//                              PETSc Modification
-//
-// Compile with: make ex11p
-//
-// Sample runs:  mpirun -np 4 ex11p -m ../../data/star.mesh
-//               mpirun -np 4 ex11p -m ../../data/star.mesh --slepcopts rc_ex11p_lobpcg
-//               mpirun -np 4 ex11p -m ../../data/star.mesh --slepcopts rc_ex11p_gd
-//
-// Description:  This example code demonstrates the use of MFEM to solve the
-//               eigenvalue problem -Delta u = lambda u with homogeneous
-//               Dirichlet boundary conditions.
-//
-//               We compute a number of the lowest eigenmodes by discretizing
-//               the Laplacian and Mass operators using a FE space of the
-//               specified order, or an isoparametric/isogeometric space if
-//               order < 1 (quadratic for quadratic curvilinear mesh, NURBS for
-//               NURBS mesh, etc.)
-//
-//               The example demonstrates the use of the SLEPc eigensolver as an
-//               alternative to the LOBPCG eigenvalue solver. The shift and
-//               invert spectral transformation is used to help the convergence
-//               to the smaller eigenvalues. Alternative solver parameters can
-//               be passed in a file with "-slepcopts".
-//
-//               Reusing a single GLVis visualization window for multiple
-//               eigenfunctions is also illustrated.
-//
-//               We recommend viewing Example 1 before viewing this example.
-
-#include "mfem.hpp"
-#include <fstream>
-#include <iostream>
-
-#ifndef MFEM_USE_SLEPC
-#error This examples requires that MFEM is build with MFEM_USE_SLEPC=YES
-#endif
-
-using namespace std;
-using namespace mfem;
-
-int main(int argc, char *argv[])
-{
-   // 1. Initialize MPI.
-   int num_procs, myid;
-   MPI_Init(&argc, &argv);
-   MPI_Comm_size(MPI_COMM_WORLD, &num_procs);
-   MPI_Comm_rank(MPI_COMM_WORLD, &myid);
-
-   // 2. Parse command-line options.
-   const char *mesh_file = "../../data/star.mesh";
-   int ser_ref_levels = 2;
-   int par_ref_levels = 1;
-   int order = 1;
-   int nev = 5;
-   int seed = 75;
-   bool slu_solver  = false;
-   bool sp_solver = false;
-   bool visualization = 1;
-   bool use_slepc = true;
-   const char *slepcrc_file = "";
-
-   OptionsParser args(argc, argv);
-   args.AddOption(&mesh_file, "-m", "--mesh",
-                  "Mesh file to use.");
-   args.AddOption(&ser_ref_levels, "-rs", "--refine-serial",
-                  "Number of times to refine the mesh uniformly in serial.");
-   args.AddOption(&par_ref_levels, "-rp", "--refine-parallel",
-                  "Number of times to refine the mesh uniformly in parallel.");
-   args.AddOption(&order, "-o", "--order",
-                  "Finite element order (polynomial degree) or -1 for"
-                  " isoparametric space.");
-   args.AddOption(&nev, "-n", "--num-eigs",
-                  "Number of desired eigenmodes.");
-   args.AddOption(&seed, "-s", "--seed",
-                  "Random seed used to initialize LOBPCG.");
-#ifdef MFEM_USE_SUPERLU
-   args.AddOption(&slu_solver, "-slu", "--superlu", "-no-slu",
-                  "--no-superlu", "Use the SuperLU Solver.");
-#endif
-#ifdef MFEM_USE_STRUMPACK
-   args.AddOption(&sp_solver, "-sp", "--strumpack", "-no-sp",
-                  "--no-strumpack", "Use the STRUMPACK Solver.");
-#endif
-   args.AddOption(&visualization, "-vis", "--visualization", "-no-vis",
-                  "--no-visualization",
-                  "Enable or disable GLVis visualization.");
-   args.AddOption(&use_slepc, "-useslepc","--useslepc","-no-slepc",
-                  "--no-slepc","Use or not SLEPc to solve the eigenvalue problem");
-   args.AddOption(&slepcrc_file, "-slepcopts", "--slepcopts",
-                  "SlepcOptions file to use.");
-   args.Parse();
-   if (slu_solver && sp_solver)
-   {
-      if (myid == 0)
-         cout << "WARNING: Both SuperLU and STRUMPACK have been selected,"
-              << " please choose either one." << endl
-              << "         Defaulting to SuperLU." << endl;
-      sp_solver = false;
-   }
-   // The command line options are also passed to the STRUMPACK
-   // solver. So do not exit if some options are not recognized.
-   if (!sp_solver)
-   {
-      if (!args.Good())
-      {
-         if (myid == 0)
-         {
-            args.PrintUsage(cout);
-         }
-         MPI_Finalize();
-         return 1;
-      }
-   }
-   if (myid == 0)
-   {
-      args.PrintOptions(cout);
-   }
-
-   // 2b. We initialize SLEPc. This internally initializes PETSc as well.
-   MFEMInitializeSlepc(NULL,NULL,slepcrc_file,NULL);
-
-   // 3. Read the (serial) mesh from the given mesh file on all processors. We
-   //    can handle triangular, quadrilateral, tetrahedral, hexahedral, surface
-   //    and volume meshes with the same code.
-   Mesh *mesh = new Mesh(mesh_file, 1, 1);
-   int dim = mesh->Dimension();
-
-   // 4. Refine the serial mesh on all processors to increase the resolution. In
-   //    this example we do 'ref_levels' of uniform refinement (2 by default, or
-   //    specified on the command line with -rs).
-   for (int lev = 0; lev < ser_ref_levels; lev++)
-   {
-      mesh->UniformRefinement();
-   }
-
-   // 5. Define a parallel mesh by a partitioning of the serial mesh. Refine
-   //    this mesh further in parallel to increase the resolution (1 time by
-   //    default, or specified on the command line with -rp). Once the parallel
-   //    mesh is defined, the serial mesh can be deleted.
-   ParMesh *pmesh = new ParMesh(MPI_COMM_WORLD, *mesh);
-   delete mesh;
-   for (int lev = 0; lev < par_ref_levels; lev++)
-   {
-      pmesh->UniformRefinement();
-   }
-
-   // 6. Define a parallel finite element space on the parallel mesh. Here we
-   //    use continuous Lagrange finite elements of the specified order. If
-   //    order < 1, we instead use an isoparametric/isogeometric space.
-   FiniteElementCollection *fec;
-   if (order > 0)
-   {
-      fec = new H1_FECollection(order, dim);
-   }
-   else if (pmesh->GetNodes())
-   {
-      fec = pmesh->GetNodes()->OwnFEC();
-   }
-   else
-   {
-      fec = new H1_FECollection(order = 1, dim);
-   }
-   ParFiniteElementSpace *fespace = new ParFiniteElementSpace(pmesh, fec);
-   HYPRE_Int size = fespace->GlobalTrueVSize();
-   if (myid == 0)
-   {
-      cout << "Number of unknowns: " << size << endl;
-   }
-
-   // 7. Set up the parallel bilinear forms a(.,.) and m(.,.) on the finite
-   //    element space. The first corresponds to the Laplacian operator -Delta,
-   //    while the second is a simple mass matrix needed on the right hand side
-   //    of the generalized eigenvalue problem below. The boundary conditions
-   //    are implemented by elimination with special values on the diagonal to
-   //    shift the Dirichlet eigenvalues out of the computational range. After
-   //    serial and parallel assembly we extract the corresponding parallel
-   //    matrices A and M.
-   ConstantCoefficient one(1.0);
-   Array<int> ess_bdr;
-   if (pmesh->bdr_attributes.Size())
-   {
-      ess_bdr.SetSize(pmesh->bdr_attributes.Max());
-      ess_bdr = 1;
-   }
-
-   ParBilinearForm *a = new ParBilinearForm(fespace);
-   a->AddDomainIntegrator(new DiffusionIntegrator(one));
-   if (pmesh->bdr_attributes.Size() == 0)
-   {
-      // Add a mass term if the mesh has no boundary, e.g. periodic mesh or
-      // closed surface.
-      a->AddDomainIntegrator(new MassIntegrator(one));
-   }
-   a->Assemble();
-   a->EliminateEssentialBCDiag(ess_bdr, 1.0);
-   a->Finalize();
-
-   ParBilinearForm *m = new ParBilinearForm(fespace);
-   m->AddDomainIntegrator(new MassIntegrator(one));
-   m->Assemble();
-   // shift the eigenvalue corresponding to eliminated dofs to a large value
-   m->EliminateEssentialBCDiag(ess_bdr, numeric_limits<double>::min());
-   m->Finalize();
-
-   PetscParMatrix *pA = NULL, *pM = NULL;
-   HypreParMatrix *A = NULL, *M = NULL;
-   Operator::Type tid =
-      !use_slepc ? Operator::Hypre_ParCSR : Operator::PETSC_MATAIJ;
-   OperatorHandle Ah(tid), Mh(tid);
-
-   a->ParallelAssemble(Ah);
-   if (!use_slepc) { Ah.Get(A); }
-   else { Ah.Get(pA); }
-   Ah.SetOperatorOwner(false);
-
-   m->ParallelAssemble(Mh);
-   if (!use_slepc) {Mh.Get(M); }
-   else {Mh.Get(pM); }
-   Mh.SetOperatorOwner(false);
-
-#if defined(MFEM_USE_SUPERLU) || defined(MFEM_USE_STRUMPACK)
-   Operator * Arow = NULL;
-#ifdef MFEM_USE_SUPERLU
-   if (slu_solver)
-   {
-      Arow = new SuperLURowLocMatrix(*A);
-   }
-#endif
-#ifdef MFEM_USE_STRUMPACK
-   if (sp_solver)
-   {
-      Arow = new STRUMPACKRowLocMatrix(*A);
-   }
-#endif
-#endif
-
-   delete a;
-   delete m;
-
-   // 8. Define and configure the LOBPCG eigensolver and the BoomerAMG
-   //    preconditioner for A to be used within the solver. Set the matrices
-   //    which define the generalized eigenproblem A x = lambda M x.
-   Solver * precond = NULL;
-   if (!use_slepc)
-   {
-      if (!slu_solver && !sp_solver)
-      {
-         HypreBoomerAMG * amg = new HypreBoomerAMG(*A);
-         amg->SetPrintLevel(0);
-         precond = amg;
-      }
-      else
-      {
-#ifdef MFEM_USE_SUPERLU
-         if (slu_solver)
-         {
-            SuperLUSolver * superlu = new SuperLUSolver(MPI_COMM_WORLD);
-            superlu->SetPrintStatistics(false);
-            superlu->SetSymmetricPattern(true);
-            superlu->SetColumnPermutation(superlu::PARMETIS);
-            superlu->SetOperator(*Arow);
-            precond = superlu;
-         }
-#endif
-#ifdef MFEM_USE_STRUMPACK
-         if (sp_solver)
-         {
-            STRUMPACKSolver * strumpack = new STRUMPACKSolver(argc, argv, MPI_COMM_WORLD);
-            strumpack->SetPrintFactorStatistics(true);
-            strumpack->SetPrintSolveStatistics(false);
-            strumpack->SetKrylovSolver(strumpack::KrylovSolver::DIRECT);
-            strumpack->SetReorderingStrategy(strumpack::ReorderingStrategy::METIS);
-            strumpack->DisableMatching();
-            strumpack->SetOperator(*Arow);
-            strumpack->SetFromCommandLine();
-            precond = strumpack;
-         }
-#endif
-      }
-   }
-
-   HypreLOBPCG * lobpcg = NULL;
-   SlepcEigenSolver * slepc = NULL;
-   if (!use_slepc)
-   {
-
-      lobpcg = new HypreLOBPCG(MPI_COMM_WORLD);
-      lobpcg->SetNumModes(nev);
-      lobpcg->SetRandomSeed(seed);
-      lobpcg->SetPreconditioner(*precond);
-      lobpcg->SetMaxIter(200);
-      lobpcg->SetTol(1e-8);
-      lobpcg->SetPrecondUsageMode(1);
-      lobpcg->SetPrintLevel(1);
-      lobpcg->SetMassMatrix(*M);
-      lobpcg->SetOperator(*A);
-   }
-   else
-   {
-      slepc = new SlepcEigenSolver(MPI_COMM_WORLD);
-      slepc->SetNumModes(nev);
-      slepc->SetWhichEigenpairs(SlepcEigenSolver::TARGET_REAL);
-      slepc->SetTarget(0.0);
-      slepc->SetSpectralTransformation(SlepcEigenSolver::SHIFT_INVERT);
-      slepc->SetOperators(*pA,*pM);
-   }
-
-   // 9. Compute the eigenmodes and extract the array of eigenvalues. Define a
-   //    parallel grid function to represent each of the eigenmodes returned by
-   //    the solver.
-   Array<double> eigenvalues;
-   if (!use_slepc)
-   {
-      lobpcg->Solve();
-      lobpcg->GetEigenvalues(eigenvalues);
-   }
-   else
-   {
-      slepc->Solve();
-      eigenvalues.SetSize(nev);
-      for (int i=0; i<nev; i++)
-      {
-         slepc->GetEigenvalue(i,eigenvalues[i]);
-      }
-   }
-   Vector temp(fespace->GetTrueVSize());
-   ParGridFunction x(fespace);
-
-   // 10. Save the refined mesh and the modes in parallel. This output can be
-   //     viewed later using GLVis: "glvis -np <np> -m mesh -g mode".
-   {
-      ostringstream mesh_name, mode_name;
-      mesh_name << "mesh." << setfill('0') << setw(6) << myid;
-
-      ofstream mesh_ofs(mesh_name.str().c_str());
-      mesh_ofs.precision(8);
-      pmesh->Print(mesh_ofs);
-
-      for (int i=0; i<nev; i++)
-      {
-         // convert eigenvector from HypreParVector to ParGridFunction
-         if (!use_slepc)
-         {
-            x = lobpcg->GetEigenvector(i);
-         }
-         else
-         {
-            slepc->GetEigenvector(i,temp);
-            x.Distribute(temp);
-
-         }
-
-         mode_name << "mode_" << setfill('0') << setw(2) << i << "."
-                   << setfill('0') << setw(6) << myid;
-
-         ofstream mode_ofs(mode_name.str().c_str());
-         mode_ofs.precision(8);
-         x.Save(mode_ofs);
-         mode_name.str("");
-      }
-   }
-
-   // 11. Send the solution by socket to a GLVis server.
-   if (visualization)
-   {
-      char vishost[] = "localhost";
-      int  visport   = 19916;
-      socketstream mode_sock(vishost, visport);
-      mode_sock.precision(8);
-
-      for (int i=0; i<nev; i++)
-      {
-         if ( myid == 0 )
-         {
-            cout << "Eigenmode " << i+1 << '/' << nev
-                 << ", Lambda = " << eigenvalues[i] << endl;
-         }
-
-         // convert eigenvector from HypreParVector to ParGridFunction
-         if (!use_slepc)
-         {
-            x = lobpcg->GetEigenvector(i);
-         }
-         else
-         {
-            slepc->GetEigenvector(i,temp);
-            x.Distribute(temp);
-         }
-
-         mode_sock << "parallel " << num_procs << " " << myid << "\n"
-                   << "solution\n" << *pmesh << x << flush
-                   << "window_title 'Eigenmode " << i+1 << '/' << nev
-                   << ", Lambda = " << eigenvalues[i] << "'" << endl;
-
-         char c;
-         if (myid == 0)
-         {
-            cout << "press (q)uit or (c)ontinue --> " << flush;
-            cin >> c;
-         }
-         MPI_Bcast(&c, 1, MPI_CHAR, 0, MPI_COMM_WORLD);
-
-         if (c != 'c')
-         {
-            break;
-         }
-      }
-      mode_sock.close();
-   }
-
-   // 12. Free the used memory.
-   if (!use_slepc)
-   {
-      delete lobpcg;
-   }
-   else
-   {
-      delete slepc;
-   }
-   delete precond;
-   delete M;
-   delete A;
-#if defined(MFEM_USE_SUPERLU) || defined(MFEM_USE_STRUMPACK)
-   delete Arow;
-#endif
-
-   delete fespace;
-   if (order > 0)
-   {
-      delete fec;
-   }
-   delete pmesh;
-
-   // We finalize SLEPc
-   MFEMFinalizeSlepc();
-   MPI_Finalize();
-
-   return 0;
-}
@@ -23,9 +23,6 @@ MFEM_LIB_FILE = mfem_is_not_built

 SEQ_EXAMPLES =
 PAR_EXAMPLES = ex1p ex2p ex3p ex4p ex5p ex6p ex9p ex10p
-ifeq ($(MFEM_USE_SLEPC),YES)
-   PAR_EXAMPLES += ex11p
-endif
 ifeq ($(MFEM_USE_MPI),NO)
   EXAMPLES = $(SEQ_EXAMPLES)
 else
@@ -90,9 +87,6 @@ EX10_ARGS        := -m ../../data/beam-quad.mesh --usepetsc --petscopts rc_ex10p
 EX10_MF_ARGS     := -m ../../data/beam-quad.mesh --usepetsc --petscopts rc_ex10p_mf -tf 6 -s 3 -rs 0 -dt 3
 EX10_MFOP_ARGS   := -m ../../data/beam-quad.mesh --usepetsc --petscopts rc_ex10p_mfop -tf 6 -s 3 -rs 0 -dt 3
 EX10_JFNK_ARGS   := -m ../../data/beam-quad.mesh --usepetsc --petscopts rc_ex10p_jfnk --jfnk -tf 6 -s 3 -rs 0 -dt 3
-EX11_ARGS_SINV   := -m ../../data/star.mesh --useslepc
-EX11_ARGS_LOBPCG := -m ../../data/star.mesh --useslepc --slepcopts rc_ex11p_lobpcg
-EX11_ARGS_GD     := -m ../../data/star.mesh --useslepc --slepcopts rc_ex11p_gd

 ex1p-test-par: ex1p
 	@$(call mfem-test,$<, $(RUN_MPI), $(TESTNAME),$(EX1_ARGS_W))
@@ -120,12 +114,6 @@ ex10p-test-par: ex10p
 	@$(call mfem-test,$<, $(RUN_MPI), $(TESTNAME),$(EX10_MF_ARGS))
 	@$(call mfem-test,$<, $(RUN_MPI), $(TESTNAME),$(EX10_MFOP_ARGS))
 	@$(call mfem-test,$<, $(RUN_MPI), $(TESTNAME),$(EX10_JFNK_ARGS))
-ifeq ($(MFEM_USE_SLEPC),YES)
-ex11p-test-par: ex11p
-	@$(call mfem-test,$<, $(RUN_MPI), $(TESTNAME),$(EX11_ARGS_SINV))
-	@$(call mfem-test,$<, $(RUN_MPI), $(TESTNAME),$(EX11_ARGS_LOBPCG))
-	@$(call mfem-test,$<, $(RUN_MPI), $(TESTNAME),$(EX11_ARGS_GD))
-endif

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

@@ -1,6 +0,0 @@
-# Options for the eigenvalue solver
-eps_view
-eps_converged_reason
-eps_type gd
-# Options for the spectral transform
-st_type precond
@@ -1,11 +0,0 @@
-# Options for the eigenvalue solver
-eps_monitor
-eps_converged_reason
-eps_view_values
-eps_type lobpcg
-eps_gen_hermitian
-eps_smallest_real
-eps_lobpcg_blocksize 5
-# Options for the spectral transform
-st_type precond
-st_pc_type gamg
@@ -50,43 +50,10 @@ set(SRCS
  fespacehierarchy.cpp
  nonlininteg_vectorconvection.cpp
  quadinterpolator.cpp
-  quadinterpolator_det.cpp
-  quadinterpolator_eval_by_nodes.cpp
-  quadinterpolator_eval_by_vdim.cpp
-  quadinterpolator_grad_by_nodes.cpp
-  quadinterpolator_grad_by_vdim.cpp
-  quadinterpolator_grad_phys_by_nodes.cpp
-  quadinterpolator_grad_phys_by_vdim.cpp
  quadinterpolator_face.cpp
  restriction.cpp
  staticcond.cpp
  tmop.cpp
-  tmop_pa.cpp
-  tmop_pa_h2d.cpp
-  tmop_pa_h2d_c0.cpp
-  tmop_pa_h2m.cpp
-  tmop_pa_h2m_c0.cpp
-  tmop_pa_h2s.cpp
-  tmop_pa_h2s_c0.cpp
-  tmop_pa_h3d.cpp
-  tmop_pa_h3d_c0.cpp
-  tmop_pa_h3m.cpp
-  tmop_pa_h3m_c0.cpp
-  tmop_pa_h3s.cpp
-  tmop_pa_h3s_c0.cpp
-  tmop_pa_jp2.cpp
-  tmop_pa_jp3.cpp
-  tmop_pa_jt2_tc.cpp
-  tmop_pa_jt3_datc.cpp
-  tmop_pa_jt3_tc.cpp
-  tmop_pa_p2.cpp
-  tmop_pa_p2_c0.cpp
-  tmop_pa_p3.cpp
-  tmop_pa_p3_c0.cpp
-  tmop_pa_w2.cpp
-  tmop_pa_w2_c0.cpp
-  tmop_pa_w3.cpp
-  tmop_pa_w3_c0.cpp
  tmop_tools.cpp
  gslib.cpp
  transfer.cpp
@@ -116,10 +83,7 @@ set(HDRS
  nonlinearform_ext.hpp
  nonlininteg.hpp
  quadinterpolator.hpp
-  quadinterpolator_eval.hpp
  quadinterpolator_face.hpp
-  quadinterpolator_grad.hpp
-  quadinterpolator_grad_phys.hpp
  restriction.hpp
  fespacehierarchy.hpp
  staticcond.hpp
@@ -132,7 +96,6 @@ set(HDRS
  tfespace.hpp
  tintrules.hpp
  tmop.hpp
-  tmop_pa.hpp
  tmop_tools.hpp
  gslib.hpp
  transfer.hpp
@@ -76,7 +76,7 @@ BilinearForm::BilinearForm(FiniteElementSpace * f)
   precompute_sparsity = 0;
   diag_policy = DIAG_KEEP;

-   assembly = AssemblyLevel::LEGACYFULL;
+   assembly = AssemblyLevel::FULL;
   batch = 1;
   ext = NULL;
 }
@@ -94,7 +94,7 @@ BilinearForm::BilinearForm (FiniteElementSpace * f, BilinearForm * bf, int ps)
   precompute_sparsity = ps;
   diag_policy = DIAG_KEEP;

-   assembly = AssemblyLevel::LEGACYFULL;
+   assembly = AssemblyLevel::FULL;
   batch = 1;
   ext = NULL;

@@ -121,10 +121,9 @@ void BilinearForm::SetAssemblyLevel(AssemblyLevel assembly_level)
   assembly = assembly_level;
   switch (assembly)
   {
-      case AssemblyLevel::LEGACYFULL:
-         break;
      case AssemblyLevel::FULL:
-         ext = new FABilinearFormExtension(this);
+         // ext = new FABilinearFormExtension(this);
+         // Use the original BilinearForm implementation for now
         break;
      case AssemblyLevel::ELEMENT:
         ext = new EABilinearFormExtension(this);
@@ -144,7 +143,7 @@ void BilinearForm::SetAssemblyLevel(AssemblyLevel assembly_level)
 void BilinearForm::EnableStaticCondensation()
 {
   delete static_cond;
-   if (assembly != AssemblyLevel::LEGACYFULL)
+   if (assembly != AssemblyLevel::FULL)
   {
      static_cond = NULL;
      MFEM_WARNING("Static condensation not supported for this assembly level");
@@ -169,7 +168,7 @@ void BilinearForm::EnableHybridization(FiniteElementSpace *constr_space,
                                       const Array<int> &ess_tdof_list)
 {
   delete hybridization;
-   if (assembly != AssemblyLevel::LEGACYFULL)
+   if (assembly != AssemblyLevel::FULL)
   {
      delete constr_integ;
      hybridization = NULL;
@@ -224,7 +223,7 @@ MatrixInverse * BilinearForm::Inverse() const

 void BilinearForm::Finalize (int skip_zeros)
 {
-   if (assembly == AssemblyLevel::LEGACYFULL)
+   if (assembly == AssemblyLevel::FULL)
   {
      if (!static_cond) { mat->Finalize(skip_zeros); }
      if (mat_e) { mat_e->Finalize(skip_zeros); }
@@ -640,7 +639,8 @@ void BilinearForm::AssembleDiagonal(Vector &diag) const
   }
   else
   {
-      mat->GetDiag(diag);
+      MFEM_ABORT("Not implemented. Maybe assemble your bilinear form into a "
+                 "matrix and use SparseMatrix::GetDiag?");
   }
 }

@@ -1083,7 +1083,7 @@ MixedBilinearForm::MixedBilinearForm (FiniteElementSpace *tr_fes,
   mat = NULL;
   mat_e = NULL;
   extern_bfs = 0;
-   assembly = AssemblyLevel::LEGACYFULL;
+   assembly = AssemblyLevel::FULL;
   ext = NULL;
 }

@@ -1108,7 +1108,7 @@ MixedBilinearForm::MixedBilinearForm (FiniteElementSpace *tr_fes,
   bbfi_marker = mbf->bbfi_marker;
   btfbfi_marker = mbf->btfbfi_marker;

-   assembly = AssemblyLevel::LEGACYFULL;
+   assembly = AssemblyLevel::FULL;
   ext = NULL;
 }

@@ -1121,8 +1121,6 @@ void MixedBilinearForm::SetAssemblyLevel(AssemblyLevel assembly_level)
   assembly = assembly_level;
   switch (assembly)
   {
-      case AssemblyLevel::LEGACYFULL:
-         break;
      case AssemblyLevel::FULL:
         // ext = new FAMixedBilinearFormExtension(this);
         // Use the original BilinearForm implementation for now
@@ -1193,7 +1191,7 @@ void MixedBilinearForm::AddMultTranspose(const Vector & x, Vector & y,

 MatrixInverse * MixedBilinearForm::Inverse() const
 {
-   if (assembly != AssemblyLevel::LEGACYFULL)
+   if (assembly != AssemblyLevel::FULL)
   {
      MFEM_WARNING("MixedBilinearForm::Inverse not possible with this assembly level!");
      return NULL;
@@ -1206,7 +1204,7 @@ MatrixInverse * MixedBilinearForm::Inverse() const

 void MixedBilinearForm::Finalize (int skip_zeros)
 {
-   if (assembly == AssemblyLevel::LEGACYFULL)
+   if (assembly == AssemblyLevel::FULL)
   {
      mat -> Finalize (skip_zeros);
   }
@@ -1483,7 +1481,7 @@ void MixedBilinearForm::AssembleDiagonal_ADAt(const Vector &D,

 void MixedBilinearForm::ConformingAssemble()
 {
-   if (assembly != AssemblyLevel::LEGACYFULL)
+   if (assembly != AssemblyLevel::FULL)
   {
      MFEM_WARNING("Conforming assemble not supported for this assembly level!");
      return;
@@ -29,11 +29,8 @@ namespace mfem
    form classes derived from Operator. */
 enum class AssemblyLevel
 {
-   /// Legacy fully assembled form, i.e. a global sparse matrix in MFEM, Hypre
-   /// or PETSC format. This assembly is ALWAYS performed on the host.
-   LEGACYFULL = 0,
-   /// Fully assembled form, i.e. a global sparse matrix in MFEM format. This
-   /// assembly is compatible with device execution.
+   /// Fully assembled form, i.e. a global sparse matrix in MFEM, Hypre or PETSC
+   /// format.
   FULL,
   /// Form assembled at element level, which computes and stores dense element
   /// matrices.
@@ -122,7 +119,7 @@ protected:
      static_cond = NULL; hybridization = NULL;
      precompute_sparsity = 0;
      diag_policy = DIAG_KEEP;
-      assembly = AssemblyLevel::LEGACYFULL;
+      assembly = AssemblyLevel::FULL;
      batch = 1;
      ext = NULL;
   }
@@ -15,7 +15,6 @@
 #include "../general/forall.hpp"
 #include "bilinearform.hpp"
 #include "libceed/ceed.hpp"
-#include "pgridfunc.hpp"

 namespace mfem
 {
@@ -116,7 +115,7 @@ void PABilinearFormExtension::AssembleDiagonal(Vector &y) const
   Array<BilinearFormIntegrator*> &integrators = *a->GetDBFI();

   const int iSz = integrators.Size();
-   if (elem_restrict && !DeviceCanUseCeed())
+   if (elem_restrict)
   {
      localY = 0.0;
      for (int i = 0; i < iSz; ++i)
@@ -293,8 +292,7 @@ void PABilinearFormExtension::MultTranspose(const Vector &x, Vector &y) const

 // Data and methods for element-assembled bilinear forms
 EABilinearFormExtension::EABilinearFormExtension(BilinearForm *form)
-   : PABilinearFormExtension(form),
-     factorize_face_terms(form->FESpace()->IsDGSpace())
+   : PABilinearFormExtension(form)
 {
 }

@@ -349,17 +347,6 @@ void EABilinearFormExtension::Assemble()
   {
      bdrFaceIntegrators[i]->AssembleEABoundaryFaces(*a->FESpace(),ea_data_bdr);
   }
-
-   if (factorize_face_terms && int_face_restrict_lex)
-   {
-      auto restFint = dynamic_cast<const L2FaceRestriction&>(*int_face_restrict_lex);
-      restFint.AddFaceMatricesToElementMatrices(ea_data_int, ea_data);
-   }
-   if (factorize_face_terms && bdr_face_restrict_lex)
-   {
-      auto restFbdr = dynamic_cast<const L2FaceRestriction&>(*bdr_face_restrict_lex);
-      restFbdr.AddFaceMatricesToElementMatrices(ea_data_bdr, ea_data);
-   }
 }

 void EABilinearFormExtension::Mult(const Vector &x, Vector &y) const
@@ -412,27 +399,24 @@ void EABilinearFormExtension::Mult(const Vector &x, Vector &y) const
         const int NDOFS = faceDofs;
         auto X = Reshape(faceIntX.Read(), NDOFS, 2, nf_int);
         auto Y = Reshape(faceIntY.ReadWrite(), NDOFS, 2, nf_int);
-         if (!factorize_face_terms)
+         auto A_int = Reshape(ea_data_int.Read(), NDOFS, NDOFS, 2, nf_int);
+         MFEM_FORALL(glob_j, nf_int*NDOFS,
         {
-            auto A_int = Reshape(ea_data_int.Read(), NDOFS, NDOFS, 2, nf_int);
-            MFEM_FORALL(glob_j, nf_int*NDOFS,
+            const int f = glob_j/NDOFS;
+            const int j = glob_j%NDOFS;
+            double res = 0.0;
+            for (int i = 0; i < NDOFS; i++)
            {
-               const int f = glob_j/NDOFS;
-               const int j = glob_j%NDOFS;
-               double res = 0.0;
-               for (int i = 0; i < NDOFS; i++)
-               {
-                  res += A_int(i, j, 0, f)*X(i, 0, f);
-               }
-               Y(j, 0, f) += res;
-               res = 0.0;
-               for (int i = 0; i < NDOFS; i++)
-               {
-                  res += A_int(i, j, 1, f)*X(i, 1, f);
-               }
-               Y(j, 1, f) += res;
-            });
-         }
+               res += A_int(i, j, 0, f)*X(i, 0, f);
+            }
+            Y(j, 0, f) += res;
+            res = 0.0;
+            for (int i = 0; i < NDOFS; i++)
+            {
+               res += A_int(i, j, 1, f)*X(i, 1, f);
+            }
+            Y(j, 1, f) += res;
+         });
         auto A_ext = Reshape(ea_data_ext.Read(), NDOFS, NDOFS, 2, nf_int);
         MFEM_FORALL(glob_j, nf_int*NDOFS,
         {
@@ -459,7 +443,7 @@ void EABilinearFormExtension::Mult(const Vector &x, Vector &y) const
   // Treatment of boundary faces
   Array<BilinearFormIntegrator*> &bdrFaceIntegrators = *a->GetBFBFI();
   const int bFISz = bdrFaceIntegrators.Size();
-   if (!factorize_face_terms && bdr_face_restrict_lex && bFISz>0)
+   if (bdr_face_restrict_lex && bFISz>0)
   {
      // Apply the Boundary Face Restriction
      bdr_face_restrict_lex->Mult(x, faceBdrX);
@@ -538,27 +522,24 @@ void EABilinearFormExtension::MultTranspose(const Vector &x, Vector &y) const
         const int NDOFS = faceDofs;
         auto X = Reshape(faceIntX.Read(), NDOFS, 2, nf_int);
         auto Y = Reshape(faceIntY.ReadWrite(), NDOFS, 2, nf_int);
-         if (!factorize_face_terms)
+         auto A_int = Reshape(ea_data_int.Read(), NDOFS, NDOFS, 2, nf_int);
+         MFEM_FORALL(glob_j, nf_int*NDOFS,
         {
-            auto A_int = Reshape(ea_data_int.Read(), NDOFS, NDOFS, 2, nf_int);
-            MFEM_FORALL(glob_j, nf_int*NDOFS,
+            const int f = glob_j/NDOFS;
+            const int j = glob_j%NDOFS;
+            double res = 0.0;
+            for (int i = 0; i < NDOFS; i++)
            {
-               const int f = glob_j/NDOFS;
-               const int j = glob_j%NDOFS;
-               double res = 0.0;
-               for (int i = 0; i < NDOFS; i++)
-               {
-                  res += A_int(j, i, 0, f)*X(i, 0, f);
-               }
-               Y(j, 0, f) += res;
-               res = 0.0;
-               for (int i = 0; i < NDOFS; i++)
-               {
-                  res += A_int(j, i, 1, f)*X(i, 1, f);
-               }
-               Y(j, 1, f) += res;
-            });
-         }
+               res += A_int(j, i, 0, f)*X(i, 0, f);
+            }
+            Y(j, 0, f) += res;
+            res = 0.0;
+            for (int i = 0; i < NDOFS; i++)
+            {
+               res += A_int(j, i, 1, f)*X(i, 1, f);
+            }
+            Y(j, 1, f) += res;
+         });
         auto A_ext = Reshape(ea_data_ext.Read(), NDOFS, NDOFS, 2, nf_int);
         MFEM_FORALL(glob_j, nf_int*NDOFS,
         {
@@ -585,7 +566,7 @@ void EABilinearFormExtension::MultTranspose(const Vector &x, Vector &y) const
   // Treatment of boundary faces
   Array<BilinearFormIntegrator*> &bdrFaceIntegrators = *a->GetBFBFI();
   const int bFISz = bdrFaceIntegrators.Size();
-   if (!factorize_face_terms && bdr_face_restrict_lex && bFISz>0)
+   if (bdr_face_restrict_lex && bFISz>0)
   {
      // Apply the Boundary Face Restriction
      bdr_face_restrict_lex->Mult(x, faceBdrX);
@@ -614,139 +595,6 @@ void EABilinearFormExtension::MultTranspose(const Vector &x, Vector &y) const
   }
 }

-// Data and methods for fully-assembled bilinear forms
-FABilinearFormExtension::FABilinearFormExtension(BilinearForm *form)
-   : EABilinearFormExtension(form),
-     mat(form->FESpace()->GetVSize(),form->FESpace()->GetVSize(),0),
-     face_mat(form->FESpace()->GetVSize(),0,0),
-     use_face_mat(false)
-{
-#ifdef MFEM_USE_MPI
-   if ( ParFiniteElementSpace* pfes =
-           dynamic_cast<ParFiniteElementSpace*>(form->FESpace()) )
-   {
-      if (pfes->IsDGSpace())
-      {
-         use_face_mat = true;
-         pfes->ExchangeFaceNbrData();
-         face_mat.SetWidth(pfes->GetFaceNbrVSize());
-      }
-   }
-#endif
-}
-
-void FABilinearFormExtension::Assemble()
-{
-   EABilinearFormExtension::Assemble();
-   FiniteElementSpace &fes = *a->FESpace();
-   if (fes.IsDGSpace())
-   {
-      const L2ElementRestriction *restE =
-         static_cast<const L2ElementRestriction*>(elem_restrict);
-      const L2FaceRestriction *restF =
-         static_cast<const L2FaceRestriction*>(int_face_restrict_lex);
-      // 1. Fill I
-      //  1.1 Increment with restE
-      restE->FillI(mat);
-      //  1.2 Increment with restF
-      if (restF) { restF->FillI(mat, face_mat); }
-      //  1.3 Sum the non-zeros in I
-      auto h_I = mat.HostReadWriteI();
-      int cpt = 0;
-      const int vd = fes.GetVDim();
-      const int ndofs = ne*elemDofs*vd;
-      for (int i = 0; i < ndofs; i++)
-      {
-         const int nnz = h_I[i];
-         h_I[i] = cpt;
-         cpt += nnz;
-      }
-      const int nnz = cpt;
-      h_I[ndofs] = nnz;
-      mat.GetMemoryJ().New(nnz, mat.GetMemoryJ().GetMemoryType());
-      mat.GetMemoryData().New(nnz, mat.GetMemoryData().GetMemoryType());
-      if (use_face_mat && restF)
-      {
-         auto h_I_face = face_mat.HostReadWriteI();
-         int cpt = 0;
-         for (int i = 0; i < ndofs; i++)
-         {
-            const int nnz = h_I_face[i];
-            h_I_face[i] = cpt;
-            cpt += nnz;
-         }
-         const int nnz_face = cpt;
-         h_I_face[ndofs] = nnz_face;
-         face_mat.GetMemoryJ().New(nnz_face,
-                                   face_mat.GetMemoryJ().GetMemoryType());
-         face_mat.GetMemoryData().New(nnz_face,
-                                      face_mat.GetMemoryData().GetMemoryType());
-      }
-      // 2. Fill J and Data
-      // 2.1 Fill J and Data with Elem ea_data
-      restE->FillJAndData(ea_data, mat);
-      // 2.2 Fill J and Data with Face ea_data_ext
-      if (restF) { restF->FillJAndData(ea_data_ext, mat, face_mat); }
-      // 2.3 Shift indirections in I back to original
-      auto I = mat.HostReadWriteI();
-      for (int i = ndofs; i > 0; i--)
-      {
-         I[i] = I[i-1];
-      }
-      I[0] = 0;
-      if (use_face_mat && restF)
-      {
-         auto I_face = face_mat.HostReadWriteI();
-         for (int i = ndofs; i > 0; i--)
-         {
-            I_face[i] = I_face[i-1];
-         }
-         I_face[0] = 0;
-      }
-   }
-   else // continuous Galerkin case
-   {
-      const ElementRestriction &rest =
-         static_cast<const ElementRestriction&>(*elem_restrict);
-      rest.FillSparseMatrix(ea_data, mat);
-   }
-}
-
-void FABilinearFormExtension::Mult(const Vector &x, Vector &y) const
-{
-   mat.Mult(x, y);
-#ifdef MFEM_USE_MPI
-   if (const ParFiniteElementSpace *pfes =
-          dynamic_cast<const ParFiniteElementSpace*>(testFes))
-   {
-      ParGridFunction x_gf;
-      x_gf.MakeRef(const_cast<ParFiniteElementSpace*>(pfes),
-                   const_cast<Vector&>(x),0);
-      x_gf.ExchangeFaceNbrData();
-      Vector &shared_x = x_gf.FaceNbrData();
-      if (shared_x.Size()) { face_mat.AddMult(shared_x, y); }
-   }
-#endif
-}
-
-void FABilinearFormExtension::MultTranspose(const Vector &x, Vector &y) const
-{
-   mat.MultTranspose(x, y);
-#ifdef MFEM_USE_MPI
-   if (const ParFiniteElementSpace *pfes =
-          dynamic_cast<const ParFiniteElementSpace*>(testFes))
-   {
-      ParGridFunction x_gf;
-      x_gf.MakeRef(const_cast<ParFiniteElementSpace*>(pfes),
-                   const_cast<Vector&>(x),0);
-      x_gf.ExchangeFaceNbrData();
-      Vector &shared_x = x_gf.FaceNbrData();
-      if (shared_x.Size()) { face_mat.AddMultTranspose(shared_x, y); }
-   }
-#endif
-}
-
-
 MixedBilinearFormExtension::MixedBilinearFormExtension(MixedBilinearForm *form)
   : Operator(form->Height(), form->Width()), a(form)
 {
@@ -62,6 +62,27 @@ public:
   virtual void Update() = 0;
 };

+/** @brief Data and methods for fully-assembled bilinear forms.
+    Not yet implemented!  Use the BilinearForm Class instead. */
+class FABilinearFormExtension : public BilinearFormExtension
+{
+public:
+   FABilinearFormExtension(BilinearForm *form)
+      : BilinearFormExtension(form) { }
+
+   /// TODO
+   void Assemble() {}
+   void FormSystemMatrix(const Array<int> &ess_tdof_list, OperatorHandle &A) {}
+   void FormLinearSystem(const Array<int> &ess_tdof_list,
+                         Vector &x, Vector &b,
+                         OperatorHandle &A, Vector &X, Vector &B,
+                         int copy_interior = 0) {}
+   void Mult(const Vector &x, Vector &y) const {}
+   void MultTranspose(const Vector &x, Vector &y) const {}
+   void Update() {}
+   ~FABilinearFormExtension() {}
+};
+
 /// Data and methods for partially-assembled bilinear forms
 class PABilinearFormExtension : public BilinearFormExtension
 {
@@ -98,12 +119,10 @@ class EABilinearFormExtension : public PABilinearFormExtension
 protected:
   int ne;
   int elemDofs;
-   // The element matrices are stored row major
   Vector ea_data;
   int nf_int, nf_bdr;
   int faceDofs;
   Vector ea_data_int, ea_data_ext, ea_data_bdr;
-   bool factorize_face_terms;

 public:
   EABilinearFormExtension(BilinearForm *form);
@@ -113,23 +132,6 @@ public:
   void MultTranspose(const Vector &x, Vector &y) const;
 };

-/// Data and methods for fully-assembled bilinear forms
-class FABilinearFormExtension : public EABilinearFormExtension
-{
-private:
-   SparseMatrix mat;
-   /// face_mat handles parallelism for DG face terms.
-   SparseMatrix face_mat;
-   bool use_face_mat;
-
-public:
-   FABilinearFormExtension(BilinearForm *form);
-
-   void Assemble();
-   void Mult(const Vector &x, Vector &y) const;
-   void MultTranspose(const Vector &x, Vector &y) const;
-};
-
 /// Data and methods for matrix-free bilinear forms NOT YET IMPLEMENTED.
 class MFBilinearFormExtension : public BilinearFormExtension
 {
@@ -926,14 +926,11 @@ void BoundaryMassIntegrator::AssembleFaceMatrix(
   for (int i = 0; i < ir->GetNPoints(); i++)
   {
      const IntegrationPoint &ip = ir->IntPoint(i);
-
-      // Set the integration point in the face and the neighboring element
-      Trans.SetAllIntPoints(&ip);
-
-      // Access the neighboring element's integration point
-      const IntegrationPoint &eip = Trans.GetElement1IntPoint();
+      IntegrationPoint eip;
+      Trans.Loc1.Transform(ip, eip);
      el1.CalcShape(eip, shape);

+      Trans.SetIntPoint(&ip);
      w = Trans.Weight() * ip.weight;
      if (Q)
      {
@@ -2574,17 +2571,16 @@ void DGTraceIntegrator::AssembleFaceMatrix(const FiniteElement &el1,
   for (int p = 0; p < ir->GetNPoints(); p++)
   {
      const IntegrationPoint &ip = ir->IntPoint(p);
-
-      // Set the integration point in the face and the neighboring elements
-      Trans.SetAllIntPoints(&ip);
-
-      // Access the neighboring elements' integration points
-      // Note: eip2 will only contain valid data if Elem2 exists
-      const IntegrationPoint &eip1 = Trans.GetElement1IntPoint();
-      const IntegrationPoint &eip2 = Trans.GetElement2IntPoint();
-
+      IntegrationPoint eip1, eip2;
+      Trans.Loc1.Transform(ip, eip1);
+      if (ndof2)
+      {
+         Trans.Loc2.Transform(ip, eip2);
+      }
      el1.CalcShape(eip1, shape1);

+      Trans.SetIntPoint(&ip);
+
      u->Eval(vu, *Trans.Elem1, eip1);

      if (dim == 1)
@@ -2731,15 +2727,10 @@ void DGDiffusionIntegrator::AssembleFaceMatrix(
   for (int p = 0; p < ir->GetNPoints(); p++)
   {
      const IntegrationPoint &ip = ir->IntPoint(p);
+      IntegrationPoint eip1, eip2;

-      // Set the integration point in the face and the neighboring elements
-      Trans.SetAllIntPoints(&ip);
-
-      // Access the neighboring elements' integration points
-      // Note: eip2 will only contain valid data if Elem2 exists
-      const IntegrationPoint &eip1 = Trans.GetElement1IntPoint();
-      const IntegrationPoint &eip2 = Trans.GetElement2IntPoint();
-
+      Trans.Loc1.Transform(ip, eip1);
+      Trans.SetIntPoint(&ip);
      if (dim == 1)
      {
         nor(0) = 2*eip1.x - 1.0;
@@ -2796,6 +2787,7 @@ void DGDiffusionIntegrator::AssembleFaceMatrix(

      if (ndof2)
      {
+         Trans.Loc2.Transform(ip, eip2);
         el2.CalcShape(eip2, shape2);
         el2.CalcDShape(eip2, dshape2);
         w = ip.weight/2/Trans.Elem2->Weight();
@@ -3013,14 +3005,9 @@ void DGElasticityIntegrator::AssembleFaceMatrix(
   for (int pind = 0; pind < ir->GetNPoints(); ++pind)
   {
      const IntegrationPoint &ip = ir->IntPoint(pind);
-
-      // Set the integration point in the face and the neighboring elements
-      Trans.SetAllIntPoints(&ip);
-
-      // Access the neighboring elements' integration points
-      // Note: eip2 will only contain valid data if Elem2 exists
-      const IntegrationPoint &eip1 = Trans.GetElement1IntPoint();
-      const IntegrationPoint &eip2 = Trans.GetElement2IntPoint();
+      IntegrationPoint eip1, eip2; // integration point in the reference space
+      Trans.Loc1.Transform(ip, eip1);
+      Trans.SetIntPoint(&ip);

      el1.CalcShape(eip1, shape1);
      el1.CalcDShape(eip1, dshape1);
@@ -3040,6 +3027,7 @@ void DGElasticityIntegrator::AssembleFaceMatrix(
      double w, wLM;
      if (ndofs2)
      {
+         Trans.Loc2.Transform(ip, eip2);
         el2.CalcShape(eip2, shape2);
         el2.CalcDShape(eip2, dshape2);
         CalcAdjugate(Trans.Elem2->Jacobian(), adjJ);
@@ -3177,22 +3165,17 @@ void TraceJumpIntegrator::AssembleFaceMatrix(
   for (int p = 0; p < ir->GetNPoints(); p++)
   {
      const IntegrationPoint &ip = ir->IntPoint(p);
-
-      // Set the integration point in the face and the neighboring elements
-      Trans.SetAllIntPoints(&ip);
-
-      // Access the neighboring elements' integration points
-      // Note: eip2 will only contain valid data if Elem2 exists
-      const IntegrationPoint &eip1 = Trans.GetElement1IntPoint();
-      const IntegrationPoint &eip2 = Trans.GetElement2IntPoint();
-
+      IntegrationPoint eip1, eip2;
      // Trace finite element shape function
+      Trans.SetIntPoint(&ip);
      trial_face_fe.CalcShape(ip, face_shape);
      // Side 1 finite element shape function
+      Trans.Loc1.Transform(ip, eip1);
      test_fe1.CalcShape(eip1, shape1);
      if (ndof2)
      {
         // Side 2 finite element shape function
+         Trans.Loc2.Transform(ip, eip2);
         test_fe2.CalcShape(eip2, shape2);
      }
      w = ip.weight;
@@ -1685,22 +1685,6 @@ protected:
   {
      trial_fe.CalcPhysCurlShape(Trans, shape);
   }
-
-   using BilinearFormIntegrator::AssemblePA;
-   virtual void AssemblePA(const FiniteElementSpace &trial_fes,
-                           const FiniteElementSpace &test_fes);
-
-   virtual void AddMultPA(const Vector&, Vector&) const;
-
-private:
-   // PA extension
-   Vector pa_data;
-   const DofToQuad *mapsO;         ///< Not owned. DOF-to-quad map, open.
-   const DofToQuad *mapsC;         ///< Not owned. DOF-to-quad map, closed.
-   const DofToQuad *mapsOtest;     ///< Not owned. DOF-to-quad map, open.
-   const DofToQuad *mapsCtest;     ///< Not owned. DOF-to-quad map, closed.
-   const GeometricFactors *geom;   ///< Not owned
-   int dim, ne, dofs1D, dofs1Dtest,quad1D, testType, trialType, coeffDim;
 };

 /** Class for integrating the bilinear form a(u,v) := (Q u, curl v) in 3D and
@@ -1740,20 +1724,6 @@ protected:
   {
      test_fe.CalcPhysCurlShape(Trans, shape);
   }
-
-   using BilinearFormIntegrator::AssemblePA;
-   virtual void AssemblePA(const FiniteElementSpace &trial_fes,
-                           const FiniteElementSpace &test_fes);
-
-   virtual void AddMultPA(const Vector&, Vector&) const;
-
-private:
-   // PA extension
-   Vector pa_data;
-   const DofToQuad *mapsO;         ///< Not owned. DOF-to-quad map, open.
-   const DofToQuad *mapsC;         ///< Not owned. DOF-to-quad map, closed.
-   const GeometricFactors *geom;   ///< Not owned
-   int dim, ne, dofs1D, quad1D, testType, trialType, coeffDim;
 };

 /** Class for integrating the bilinear form a(u,v) := - (Q u, grad v) in either
@@ -1954,7 +1924,7 @@ public:
   static const IntegrationRule &GetRule(const FiniteElement &trial_fe,
                                         const FiniteElement &test_fe);

-   void SetupPA(const FiniteElementSpace &fes);
+   void SetupPA(const FiniteElementSpace &fes, const bool force = false);
 };

 /** Class for local mass matrix assembling a(u,v) := (Q u, v) */
@@ -2030,7 +2000,7 @@ public:
                                         const FiniteElement &test_fe,
                                         ElementTransformation &Trans);

-   void SetupPA(const FiniteElementSpace &fes);
+   void SetupPA(const FiniteElementSpace &fes, const bool force = false);
 };

 /** Mass integrator (u, v) restricted to the boundary of a domain */
@@ -32,7 +32,7 @@ static void EAConvectionAssemble1D(const int NE,
   auto B = Reshape(b.Read(), Q1D, D1D);
   auto G = Reshape(g.Read(), Q1D, D1D);
   auto D = Reshape(padata.Read(), Q1D, NE);
-   auto A = Reshape(eadata.ReadWrite(), D1D, D1D, NE);
+   auto A = Reshape(eadata.Write(), D1D, D1D, NE);
   MFEM_FORALL_3D(e, NE, D1D, D1D, 1,
   {
      const int D1D = T_D1D ? T_D1D : d1d;
@@ -54,7 +54,7 @@ static void EAConvectionAssemble1D(const int NE,
            {
               val += r_Bj[k1] * D(k1, e) * r_Gi[k1];
            }
-            A(i1, j1, e) += val;
+            A(i1, j1, e) = val;
         }
      }
   });
@@ -76,7 +76,7 @@ static void EAConvectionAssemble2D(const int NE,
   auto B = Reshape(b.Read(), Q1D, D1D);
   auto G = Reshape(g.Read(), Q1D, D1D);
   auto D = Reshape(padata.Read(), Q1D, Q1D, 2, NE);
-   auto A = Reshape(eadata.ReadWrite(), D1D, D1D, D1D, D1D, NE);
+   auto A = Reshape(eadata.Write(), D1D, D1D, D1D, D1D, NE);
   MFEM_FORALL_3D(e, NE, D1D, D1D, 1,
   {
      const int D1D = T_D1D ? T_D1D : d1d;
@@ -121,7 +121,7 @@ static void EAConvectionAssemble2D(const int NE,
                               * r_B[k1][j1]* r_B[k2][j2];
                     }
                  }
-                  A(i1, i2, j1, j2, e) += val;
+                  A(i1, i2, j1, j2, e) = val;
               }
            }
         }
@@ -145,7 +145,7 @@ static void EAConvectionAssemble3D(const int NE,
   auto B = Reshape(b.Read(), Q1D, D1D);
   auto G = Reshape(g.Read(), Q1D, D1D);
   auto D = Reshape(padata.Read(), Q1D, Q1D, Q1D, 3, NE);
-   auto A = Reshape(eadata.ReadWrite(), D1D, D1D, D1D, D1D, D1D, D1D, NE);
+   auto A = Reshape(eadata.Write(), D1D, D1D, D1D, D1D, D1D, D1D, NE);
   MFEM_FORALL_3D(e, NE, D1D, D1D, D1D,
   {
      const int D1D = T_D1D ? T_D1D : d1d;
@@ -191,7 +191,7 @@ static void EAConvectionAssemble3D(const int NE,
                              }
                           }
                        }
-                        A(i1, i2, i3, j1, j2, j3, e) += val;
+                        A(i1, i2, i3, j1, j2, j3, e) = val;
                     }
                  }
               }
@@ -13,10 +13,6 @@
 #include "bilininteg.hpp"
 #include "gridfunc.hpp"

-#include "restriction.hpp"
-#include "tmop_pa.hpp"
-#include "../linalg/kernels.hpp"
-
 using namespace std;

 namespace mfem
@@ -72,53 +68,47 @@ static void PAConvectionSetup3D(const int Q1D,
                                const double alpha,
                                Vector &op)
 {
-   const auto W = Reshape(w.Read(), Q1D,Q1D,Q1D);
-   const auto J = Reshape(j.Read(), Q1D,Q1D,Q1D,3,3,NE);
+   const int NQ = Q1D*Q1D*Q1D;
+   auto W = w.Read();
+   auto J = Reshape(j.Read(), NQ, 3, 3, NE);
   const bool const_v = vel.Size() == 3;
-   const auto V = const_v ?
-                  Reshape(vel.Read(), 3,1,1,1,1) :
-                  Reshape(vel.Read(), 3,Q1D,Q1D,Q1D,NE);
-   auto y = Reshape(op.Write(), Q1D,Q1D,Q1D,3,NE);
-   MFEM_FORALL_3D(e, NE, Q1D, Q1D, Q1D,
+   auto V =
+      const_v ? Reshape(vel.Read(), 3,1,1) : Reshape(vel.Read(), 3,NQ,NE);
+   auto y = Reshape(op.Write(), NQ, 3, NE);
+   MFEM_FORALL(e, NE,
   {
-      MFEM_FOREACH_THREAD(qx,x,Q1D)
+      for (int q = 0; q < NQ; ++q)
      {
-         MFEM_FOREACH_THREAD(qy,y,Q1D)
-         {
-            MFEM_FOREACH_THREAD(qz,z,Q1D)
-            {
-               const double J11 = J(qx,qy,qz,0,0,e);
-               const double J12 = J(qx,qy,qz,0,1,e);
-               const double J13 = J(qx,qy,qz,0,2,e);
-               const double J21 = J(qx,qy,qz,1,0,e);
-               const double J22 = J(qx,qy,qz,1,1,e);
-               const double J23 = J(qx,qy,qz,1,2,e);
-               const double J31 = J(qx,qy,qz,2,0,e);
-               const double J32 = J(qx,qy,qz,2,1,e);
-               const double J33 = J(qx,qy,qz,2,2,e);
-               const double w = alpha * W(qx,qy,qz);
-               const double v0 = const_v ? V(0,0,0,0,0) : V(0,qx,qy,qz,e);
-               const double v1 = const_v ? V(1,0,0,0,0) : V(1,qx,qy,qz,e);
-               const double v2 = const_v ? V(2,0,0,0,0) : V(2,qx,qy,qz,e);
-               const double wx = w * v0;
-               const double wy = w * v1;
-               const double wz = w * v2;
-               // adj(J)
-               const double A11 = (J22 * J33) - (J23 * J32);
-               const double A12 = (J32 * J13) - (J12 * J33);
-               const double A13 = (J12 * J23) - (J22 * J13);
-               const double A21 = (J31 * J23) - (J21 * J33);
-               const double A22 = (J11 * J33) - (J13 * J31);
-               const double A23 = (J21 * J13) - (J11 * J23);
-               const double A31 = (J21 * J32) - (J31 * J22);
-               const double A32 = (J31 * J12) - (J11 * J32);
-               const double A33 = (J11 * J22) - (J12 * J21);
-               // q . J^{-1} = q . adj(J)
-               y(qx,qy,qz,0,e) = wx * A11 + wy * A12 + wz * A13;
-               y(qx,qy,qz,1,e) = wx * A21 + wy * A22 + wz * A23;
-               y(qx,qy,qz,2,e) = wx * A31 + wy * A32 + wz * A33;
-            }
-         }
+         const double J11 = J(q,0,0,e);
+         const double J21 = J(q,1,0,e);
+         const double J31 = J(q,2,0,e);
+         const double J12 = J(q,0,1,e);
+         const double J22 = J(q,1,1,e);
+         const double J32 = J(q,2,1,e);
+         const double J13 = J(q,0,2,e);
+         const double J23 = J(q,1,2,e);
+         const double J33 = J(q,2,2,e);
+         const double w = alpha * W[q];
+         const double v0 = const_v ? V(0,0,0) : V(0,q,e);
+         const double v1 = const_v ? V(1,0,0) : V(1,q,e);
+         const double v2 = const_v ? V(2,0,0) : V(2,q,e);
+         const double wx = w * v0;
+         const double wy = w * v1;
+         const double wz = w * v2;
+         // adj(J)
+         const double A11 = (J22 * J33) - (J23 * J32);
+         const double A12 = (J32 * J13) - (J12 * J33);
+         const double A13 = (J12 * J23) - (J22 * J13);
+         const double A21 = (J31 * J23) - (J21 * J33);
+         const double A22 = (J11 * J33) - (J13 * J31);
+         const double A23 = (J21 * J13) - (J11 * J23);
+         const double A31 = (J21 * J32) - (J31 * J22);
+         const double A32 = (J31 * J12) - (J11 * J32);
+         const double A33 = (J11 * J22) - (J12 * J21);
+         // q . J^{-1} = q . adj(J)
+         y(q,0,e) =  wx * A11 + wy * A12 + wz * A13;
+         y(q,1,e) =  wx * A21 + wy * A22 + wz * A23;
+         y(q,2,e) =  wx * A31 + wy * A32 + wz * A33;
      }
   });
 }
@@ -194,8 +184,8 @@ void PAConvectionApply2D(const int ne,
            Gu[dy][qx] = 0.0;
            for (int dx = 0; dx < D1D; ++dx)
            {
-               const double bx = B(qx,dx);
-               const double gx = G(qx,dx);
+               const double bx  = B(qx,dx);
+               const double gx  = G(qx,dx);
               const double x = u[dy][dx];
               Bu[dy][qx] += bx * x;
               Gu[dy][qx] += gx * x;
@@ -212,8 +202,8 @@ void PAConvectionApply2D(const int ne,
            BGu[qy][qx] = 0.0;
            for (int dy = 0; dy < D1D; ++dy)
            {
-               const double bx = B(qy,dy);
-               const double gx = G(qy,dy);
+               const double bx  = B(qy,dy);
+               const double gx  = G(qy,dy);
               GBu[qy][qx] += gx * Bu[dy][qx];
               BGu[qy][qx] += bx * Gu[dy][qx];
            }
@@ -242,7 +232,7 @@ void PAConvectionApply2D(const int ne,
            BDGu[dy][qx] = 0.0;
            for (int qy = 0; qy < Q1D; ++qy)
            {
-               const double w = Bt(dy,qy);
+               const double w  = Bt(dy,qy);
               BDGu[dy][qx] += w * DGu[qy][qx];
            }
         }
@@ -254,7 +244,7 @@ void PAConvectionApply2D(const int ne,
            double BBDGu = 0.0;
            for (int qx = 0; qx < Q1D; ++qx)
            {
-               const double w = Bt(dx,qx);
+               const double w  = Bt(dx,qx);
               BBDGu += w * BDGu[dy][qx];
            }
            y(dx,dy,e) += BBDGu;
@@ -320,7 +310,7 @@ void SmemPAConvectionApply2D(const int ne,
            {
               const double bx = B(qx,dx);
               const double gx = G(qx,dx);
-               const double x = u[tidz][dy][dx];
+               const double x  = u[tidz][dy][dx];
               Bu[tidz][dy][qx] += bx * x;
               Gu[tidz][dy][qx] += gx * x;
            }
@@ -337,8 +327,8 @@ void SmemPAConvectionApply2D(const int ne,
            BGu[tidz][qy][qx] = 0.0;
            for (int dy = 0; dy < D1D; ++dy)
            {
-               const double bx = B(qy,dy);
-               const double gx = G(qy,dy);
+               const double bx  = B(qy,dy);
+               const double gx  = G(qy,dy);
               GBu[tidz][qy][qx] += gx * Bu[tidz][dy][qx];
               BGu[tidz][qy][qx] += bx * Gu[tidz][dy][qx];
            }
@@ -369,7 +359,7 @@ void SmemPAConvectionApply2D(const int ne,
            BDGu[tidz][dy][qx] = 0.0;
            for (int qy = 0; qy < Q1D; ++qy)
            {
-               const double w = Bt(dy,qy);
+               const double w  = Bt(dy,qy);
               BDGu[tidz][dy][qx] += w * DGu[tidz][qy][qx];
            }
         }
@@ -382,7 +372,7 @@ void SmemPAConvectionApply2D(const int ne,
            double BBDGu = 0.0;
            for (int qx = 0; qx < Q1D; ++qx)
            {
-               const double w = Bt(dx,qx);
+               const double w  = Bt(dx,qx);
               BBDGu += w * BDGu[tidz][dy][qx];
            }
            y(dx,dy,e) += BBDGu;
@@ -446,8 +436,8 @@ void PAConvectionApply3D(const int ne,
               Gu[dz][dy][qx] = 0.0;
               for (int dx = 0; dx < D1D; ++dx)
               {
-                  const double bx = B(qx,dx);
-                  const double gx = G(qx,dx);
+                  const double bx  = B(qx,dx);
+                  const double gx  = G(qx,dx);
                  const double x = u[dz][dy][dx];
                  Bu[dz][dy][qx] += bx * x;
                  Gu[dz][dy][qx] += gx * x;
@@ -469,8 +459,8 @@ void PAConvectionApply3D(const int ne,
               BGu[dz][qy][qx] = 0.0;
               for (int dy = 0; dy < D1D; ++dy)
               {
-                  const double bx = B(qy,dy);
-                  const double gx = G(qy,dy);
+                  const double bx  = B(qy,dy);
+                  const double gx  = G(qy,dy);
                  BBu[dz][qy][qx] += bx * Bu[dz][dy][qx];
                  GBu[dz][qy][qx] += gx * Bu[dz][dy][qx];
                  BGu[dz][qy][qx] += bx * Gu[dz][dy][qx];
@@ -492,8 +482,8 @@ void PAConvectionApply3D(const int ne,
               BBGu[qz][qy][qx] = 0.0;
               for (int dz = 0; dz < D1D; ++dz)
               {
-                  const double bx = B(qz,dz);
-                  const double gx = G(qz,dz);
+                  const double bx  = B(qz,dz);
+                  const double gx  = G(qz,dz);
                  GBBu[qz][qy][qx] += gx * BBu[dz][qy][qx];
                  BGBu[qz][qy][qx] += bx * GBu[dz][qy][qx];
                  BBGu[qz][qy][qx] += bx * BGu[dz][qy][qx];
@@ -531,7 +521,7 @@ void PAConvectionApply3D(const int ne,
               BDGu[dz][qy][qx] = 0.0;
               for (int qz = 0; qz < Q1D; ++qz)
               {
-                  const double w = Bt(dz,qz);
+                  const double w  = Bt(dz,qz);
                  BDGu[dz][qy][qx] += w * DGu[qz][qy][qx];
               }
            }
@@ -547,7 +537,7 @@ void PAConvectionApply3D(const int ne,
               BBDGu[dz][dy][qx] = 0.0;
               for (int qy = 0; qy < Q1D; ++qy)
               {
-                  const double w = Bt(dy,qy);
+                  const double w  = Bt(dy,qy);
                  BBDGu[dz][dy][qx] += w * BDGu[dz][qy][qx];
               }
            }
@@ -562,7 +552,7 @@ void PAConvectionApply3D(const int ne,
               double BBBDGu = 0.0;
               for (int qx = 0; qx < Q1D; ++qx)
               {
-                  const double w = Bt(dx,qx);
+                  const double w  = Bt(dx,qx);
                  BBBDGu += w * BBDGu[dz][dy][qx];
               }
               y(dx,dy,dz,e) += BBBDGu;
@@ -635,8 +625,8 @@ void SmemPAConvectionApply3D(const int ne,
               double Gu_ = 0.0;
               for (int dx = 0; dx < D1D; ++dx)
               {
-                  const double bx = B(qx,dx);
-                  const double gx = G(qx,dx);
+                  const double bx  = B(qx,dx);
+                  const double gx  = G(qx,dx);
                  const double x = u[dz][dy][dx];
                  Bu_ += bx * x;
                  Gu_ += gx * x;
@@ -661,8 +651,8 @@ void SmemPAConvectionApply3D(const int ne,
               double BGu_ = 0.0;
               for (int dy = 0; dy < D1D; ++dy)
               {
-                  const double bx = B(qy,dy);
-                  const double gx = G(qy,dy);
+                  const double bx  = B(qy,dy);
+                  const double gx  = G(qy,dy);
                  BBu_ += bx * Bu[dz][dy][qx];
                  GBu_ += gx * Bu[dz][dy][qx];
                  BGu_ += bx * Gu[dz][dy][qx];
@@ -688,8 +678,8 @@ void SmemPAConvectionApply3D(const int ne,
               double BBGu_ = 0.0;
               for (int dz = 0; dz < D1D; ++dz)
               {
-                  const double bx = B(qz,dz);
-                  const double gx = G(qz,dz);
+                  const double bx  = B(qz,dz);
+                  const double gx  = G(qz,dz);
                  GBBu_ += gx * BBu[dz][qy][qx];
                  BGBu_ += bx * GBu[dz][qy][qx];
                  BBGu_ += bx * BGu[dz][qy][qx];
@@ -731,7 +721,7 @@ void SmemPAConvectionApply3D(const int ne,
               double BDGu_ = 0.0;
               for (int qz = 0; qz < Q1D; ++qz)
               {
-                  const double w = Bt(dz,qz);
+                  const double w  = Bt(dz,qz);
                  BDGu_ += w * DGu[qz][qy][qx];
               }
               BDGu[dz][qy][qx] = BDGu_;
@@ -749,7 +739,7 @@ void SmemPAConvectionApply3D(const int ne,
               double BBDGu_ = 0.0;
               for (int qy = 0; qy < Q1D; ++qy)
               {
-                  const double w = Bt(dy,qy);
+                  const double w  = Bt(dy,qy);
                  BBDGu_ += w * BDGu[dz][qy][qx];
               }
               BBDGu[dz][dy][qx] = BBDGu_;
@@ -766,7 +756,7 @@ void SmemPAConvectionApply3D(const int ne,
               double BBBDGu = 0.0;
               for (int qx = 0; qx < Q1D; ++qx)
               {
-                  const double w = Bt(dx,qx);
+                  const double w  = Bt(dx,qx);
                  BBBDGu += w * BBDGu[dz][dy][qx];
               }
               y(dx,dy,dz,e) = BBBDGu;
@@ -776,117 +766,6 @@ void SmemPAConvectionApply3D(const int ne,
   });
 }

-template<int T_VDIM = 0, int T_D1D = 0, int T_Q1D = 0, int T_MAX = 0>
-static void QEvalVGF2D(const int NE,
-                       const double *b_,
-                       const double *x_,
-                       double *y_,
-                       const int vdim = 1,
-                       const int d1d = 0,
-                       const int q1d = 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;
-
-   const auto b = Reshape(b_, Q1D, D1D);
-   const auto X = Reshape(x_, D1D, D1D, VDIM, NE);
-   auto C = Reshape(y_, VDIM, Q1D, Q1D, NE);
-
-   MFEM_FORALL_2D(e, NE, Q1D, Q1D, 1,
-   {
-      constexpr int NBZ = 1;
-      constexpr int MQ1 = T_Q1D ? T_Q1D : T_MAX;
-      constexpr int MD1 = T_D1D ? T_D1D : T_MAX;
-
-      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_SHARED double B[MQ1*MD1];
-      mfem::kernels::LoadB<MD1,MQ1>(D1D,Q1D,b,B);
-
-      MFEM_SHARED double DD[NBZ][MD1*MD1];
-      MFEM_SHARED double DQ[NBZ][MD1*MQ1];
-      MFEM_SHARED double QQ[NBZ][MQ1*MQ1];
-
-      for (int c = 0; c < VDIM; c++)
-      {
-         mfem::kernels::LoadX<MD1,NBZ>(e,D1D,c,X,DD);
-         mfem::kernels::EvalX<MD1,MQ1,NBZ>(D1D,Q1D,B,DD,DQ);
-         mfem::kernels::EvalY<MD1,MQ1,NBZ>(D1D,Q1D,B,DQ,QQ);
-
-         MFEM_FOREACH_THREAD(qx,x,Q1D)
-         {
-            MFEM_FOREACH_THREAD(qy,y,Q1D)
-            {
-               double G;
-               mfem::kernels::PullEval<MQ1,NBZ>(qx,qy,QQ,G);
-               C(c,qx,qy,e) = G;
-            }
-         }
-         MFEM_SYNC_THREAD;
-      }
-   });
-}
-
-template<int T_VDIM = 0, int T_D1D = 0, int T_Q1D = 0, int T_MAX = 0>
-static void QEvalVGF3D(const int NE,
-                       const double *b_,
-                       const double *x_,
-                       double *y_,
-                       const int vdim = 1,
-                       const int d1d = 0,
-                       const int q1d = 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;
-
-   const auto b = Reshape(b_, Q1D, D1D);
-   const auto X = Reshape(x_, D1D, D1D, D1D, VDIM, NE);
-   auto C = Reshape(y_, VDIM, Q1D, Q1D, Q1D, NE);
-
-   MFEM_FORALL_3D(e, NE, Q1D, Q1D, Q1D,
-   {
-      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 MQ1 = T_Q1D ? T_Q1D : T_MAX;
-      constexpr int MD1 = T_D1D ? T_D1D : T_MAX;
-
-      MFEM_SHARED double B[MQ1*MD1];
-      mfem::kernels::LoadB<MD1,MQ1>(D1D,Q1D,b,B);
-
-      MFEM_SHARED double DDD[MD1*MD1*MD1];
-      MFEM_SHARED double DDQ[MD1*MD1*MQ1];
-      MFEM_SHARED double DQQ[MD1*MQ1*MQ1];
-      MFEM_SHARED double QQQ[MQ1*MQ1*MQ1];
-
-      for (int c = 0; c < VDIM; c++)
-      {
-         mfem::kernels::LoadX<MD1>(e,D1D,c,X,DDD);
-         mfem::kernels::EvalX<MD1,MQ1>(D1D,Q1D,B,DDD,DDQ);
-         mfem::kernels::EvalY<MD1,MQ1>(D1D,Q1D,B,DDQ,DQQ);
-         mfem::kernels::EvalZ<MD1,MQ1>(D1D,Q1D,B,DQQ,QQQ);
-
-         MFEM_FOREACH_THREAD(qx,x,Q1D)
-         {
-            MFEM_FOREACH_THREAD(qy,y,Q1D)
-            {
-               MFEM_FOREACH_THREAD(qz,z,Q1D)
-               {
-                  double G;
-                  mfem::kernels::PullEval<MQ1>(qx,qy,qz,QQQ,G);
-                  C(c,qx,qy,qz,e) = G;
-               }
-            }
-         }
-         MFEM_SYNC_THREAD;
-      }
-   });
-}
-
 void ConvectionIntegrator::AssemblePA(const FiniteElementSpace &fes)
 {
   // Assumes tensor-product elements
@@ -899,104 +778,16 @@ void ConvectionIntegrator::AssemblePA(const FiniteElementSpace &fes)
   const int nq = ir->GetNPoints();
   dim = mesh->Dimension();
   ne = fes.GetNE();
-   const DofToQuad::Mode mode = DofToQuad::TENSOR;
-#ifdef MFEM_USE_UMPIRE
-   const MemoryType temp_type = Device::GetDeviceMemoryType() == MemoryType::DEVICE_UMPIRE
-      ? MemoryType::DEVICE_UMPIRE_2 : Device::GetDeviceMemoryType();
-#else
-   const MemoryType temp_type = Device::GetDeviceMemoryType();
-#endif
-
-   geom = mesh->GetGeometricFactors(*ir, GeometricFactors::JACOBIANS, mode, temp_type);
-   maps = &el.GetDofToQuad(*ir, mode);
+   geom = mesh->GetGeometricFactors(*ir, GeometricFactors::JACOBIANS);
+   maps = &el.GetDofToQuad(*ir, DofToQuad::TENSOR);
   dofs1D = maps->ndof;
   quad1D = maps->nqpt;
-   pa_data.SetSize(symmDims * nq * ne, temp_type);
+   pa_data.SetSize(symmDims * nq * ne, Device::GetMemoryType());
   Vector vel;
-   if (VectorConstantCoefficient *cQ =
-          dynamic_cast<VectorConstantCoefficient*>(Q))
+   if (VectorConstantCoefficient *cQ = dynamic_cast<VectorConstantCoefficient*>(Q))
   {
      vel = cQ->GetVec();
   }
-   else if (VectorGridFunctionCoefficient *vgfQ =
-               dynamic_cast<VectorGridFunctionCoefficient*>(Q))
-   {
-      Vector xe;
-      vel.SetSize(dim * nq * ne, temp_type);
-
-      const GridFunction *gf = vgfQ->GetGridFunction();
-      const ElementDofOrdering ordering = ElementDofOrdering::LEXICOGRAPHIC;
-      const FiniteElementSpace &gf_fes = *gf->FESpace();
-
-      const int vdim = gf_fes.GetVDim();
-      const Operator *R = gf_fes.GetElementRestriction(ordering);
-      const FiniteElement &el_gf = *gf_fes.GetFE(0);
-      const DofToQuad *maps_gf = &el_gf.GetDofToQuad(*ir, mode);
-      const int D1D = maps_gf->ndof;
-      const int Q1D = maps_gf->nqpt;
-
-      MFEM_VERIFY(R,"");
-      MFEM_VERIFY(vdim == dim, "");
-      MFEM_VERIFY(dim==2 || dim==3,"");
-
-      xe.SetSize(R->Height(), Device::GetMemoryType());
-      xe.UseDevice(true);
-      R->Mult(*gf, xe);
-
-      const auto B = maps_gf->B.Read();
-      const auto x = xe.Read();
-      auto y = vel.Write();
-
-      const int id = (D1D << 4 ) | Q1D;
-      if (dim == 2)
-      {
-         switch (id)
-         {
-            case 0x22: QEvalVGF2D<2,2,2>(ne,B,x,y); break;
-            case 0x33: QEvalVGF2D<2,3,3>(ne,B,x,y); break;
-            case 0x34: QEvalVGF2D<2,3,4>(ne,B,x,y); break;
-            default:
-            {
-               constexpr int MAX_DQ = 8;
-               MFEM_VERIFY(D1D <= MAX_DQ, "");
-               MFEM_VERIFY(Q1D <= MAX_DQ, "");
-               QEvalVGF2D<0,0,0,MAX_DQ>(ne,B,x,y,vdim,D1D,Q1D);
-            }
-         }
-      }
-      if (dim == 3)
-      {
-         switch (id)
-         {
-            case 0x23: QEvalVGF3D<3,2,3>(ne,B,x,y); break;
-            case 0x34: QEvalVGF3D<3,3,4>(ne,B,x,y); break;
-            case 0x35: QEvalVGF3D<3,3,5>(ne,B,x,y); break;
-            case 0x46: QEvalVGF3D<3,4,6>(ne,B,x,y); break;
-            case 0x48: QEvalVGF3D<3,4,8>(ne,B,x,y); break;
-            default:
-            {
-               constexpr int MAX_DQ = 6;
-               MFEM_VERIFY(D1D <= MAX_DQ, "");
-               MFEM_VERIFY(Q1D <= MAX_DQ, "");
-               QEvalVGF3D<0,0,0,MAX_DQ>(ne,B,x,y,vdim,D1D,Q1D);
-            }
-         }
-      }
-   }
-   else if (VectorQuadratureFunctionCoefficient* cQ =
-               dynamic_cast<VectorQuadratureFunctionCoefficient*>(Q))
-   {
-      const QuadratureFunction &qFun = cQ->GetQuadFunction();
-      MFEM_VERIFY(qFun.Size() == dim * nq * ne,
-                  "Incompatible QuadratureFunction dimension \n");
-
-      MFEM_VERIFY(ir == &qFun.GetSpace()->GetElementIntRule(0),
-                  "IntegrationRule used within integrator and in"
-                  " QuadratureFunction appear to be different");
-
-      qFun.Read();
-      vel.MakeRef(const_cast<QuadratureFunction &>(qFun),0);
-   }
   else
   {
      vel.SetSize(dim * nq * ne);
@@ -1036,12 +827,9 @@ static void PAConvectionApply(const int dim,
      switch ((D1D << 4 ) | Q1D)
      {
         case 0x22: return SmemPAConvectionApply2D<2,2,8>(NE,B,G,Bt,Gt,op,x,y);
-         case 0x33: return SmemPAConvectionApply2D<3,3,4>(NE,B,G,Bt,Gt,op,x,y);
-         case 0x34: return SmemPAConvectionApply2D<3,4,4>(NE,B,G,Bt,Gt,op,x,y);
-         case 0x44: return SmemPAConvectionApply2D<4,4,4>(NE,B,G,Bt,Gt,op,x,y);
-         case 0x46: return SmemPAConvectionApply2D<4,6,4>(NE,B,G,Bt,Gt,op,x,y);
+         case 0x33: return SmemPAConvectionApply2D<3,3,3>(NE,B,G,Bt,Gt,op,x,y);
+         case 0x44: return SmemPAConvectionApply2D<4,4,2>(NE,B,G,Bt,Gt,op,x,y);
         case 0x55: return SmemPAConvectionApply2D<5,5,2>(NE,B,G,Bt,Gt,op,x,y);
-         case 0x58: return SmemPAConvectionApply2D<5,8,2>(NE,B,G,Bt,Gt,op,x,y);
         case 0x66: return SmemPAConvectionApply2D<6,6,1>(NE,B,G,Bt,Gt,op,x,y);
         case 0x77: return SmemPAConvectionApply2D<7,7,1>(NE,B,G,Bt,Gt,op,x,y);
         case 0x88: return SmemPAConvectionApply2D<8,8,1>(NE,B,G,Bt,Gt,op,x,y);
@@ -1054,12 +842,8 @@ static void PAConvectionApply(const int dim,
      switch ((D1D << 4 ) | Q1D)
      {
         case 0x23: return SmemPAConvectionApply3D<2,3>(NE,B,G,Bt,Gt,op,x,y);
-         case 0x24: return SmemPAConvectionApply3D<2,4>(NE,B,G,Bt,Gt,op,x,y);
-         case 0x26: return SmemPAConvectionApply3D<2,6>(NE,B,G,Bt,Gt,op,x,y);
         case 0x34: return SmemPAConvectionApply3D<3,4>(NE,B,G,Bt,Gt,op,x,y);
-         case 0x35: return SmemPAConvectionApply3D<3,5>(NE,B,G,Bt,Gt,op,x,y);
         case 0x45: return SmemPAConvectionApply3D<4,5>(NE,B,G,Bt,Gt,op,x,y);
-         case 0x48: return SmemPAConvectionApply3D<4,8>(NE,B,G,Bt,Gt,op,x,y);
         case 0x56: return SmemPAConvectionApply3D<5,6>(NE,B,G,Bt,Gt,op,x,y);
         case 0x67: return SmemPAConvectionApply3D<6,7>(NE,B,G,Bt,Gt,op,x,y);
         case 0x78: return SmemPAConvectionApply3D<7,8>(NE,B,G,Bt,Gt,op,x,y);
@@ -167,19 +167,6 @@ void DGTraceIntegrator::SetupPA(const FiniteElementSpace &fes, FaceType type)
      r.SetSize(1);
      r(0) = c_rho->constant;
   }
-   else if (QuadratureFunctionCoefficient* c_rho =
-               dynamic_cast<QuadratureFunctionCoefficient*>(rho))
-   {
-      const QuadratureFunction &qFun = c_rho->GetQuadFunction();
-      MFEM_VERIFY(qFun.Size() == nq * nf,
-                  "Incompatible QuadratureFunction dimension \n");
-
-      MFEM_VERIFY(ir == &qFun.GetSpace()->GetElementIntRule(0),
-                  "IntegrationRule used within integrator and in"
-                  " QuadratureFunction appear to be different");
-      qFun.Read();
-      r.MakeRef(const_cast<QuadratureFunction &>(qFun),0);
-   }
   else
   {
      r.SetSize(nq * nf);
@@ -213,20 +200,6 @@ void DGTraceIntegrator::SetupPA(const FiniteElementSpace &fes, FaceType type)
   {
      vel = c_u->GetVec();
   }
-   else if (VectorQuadratureFunctionCoefficient* c_u =
-               dynamic_cast<VectorQuadratureFunctionCoefficient*>(u))
-   {
-      // Assumed to be in lexicographical ordering
-      const QuadratureFunction &qFun = c_u->GetQuadFunction();
-      MFEM_VERIFY(qFun.Size() == dim * nq * nf,
-                  "Incompatible QuadratureFunction dimension \n");
-
-      MFEM_VERIFY(ir == &qFun.GetSpace()->GetElementIntRule(0),
-                  "IntegrationRule used within integrator and in"
-                  " QuadratureFunction appear to be different");
-      qFun.Read();
-      vel.MakeRef(const_cast<QuadratureFunction &>(qFun),0);
-   }
   else
   {
      vel.SetSize(dim * nq * nf);
@@ -31,7 +31,7 @@ static void EADiffusionAssemble1D(const int NE,
   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
   auto G = Reshape(g.Read(), Q1D, D1D);
   auto D = Reshape(padata.Read(), Q1D, NE);
-   auto A = Reshape(eadata.ReadWrite(), D1D, D1D, NE);
+   auto A = Reshape(eadata.Write(), D1D, D1D, NE);
   MFEM_FORALL_3D(e, NE, D1D, D1D, 1,
   {
      const int D1D = T_D1D ? T_D1D : d1d;
@@ -53,7 +53,7 @@ static void EADiffusionAssemble1D(const int NE,
            {
               val += r_Gj[k1] * D(k1, e) * r_Gi[k1];
            }
-            A(i1, j1, e) += val;
+            A(i1, j1, e) = val;
         }
      }
   });
@@ -75,7 +75,7 @@ static void EADiffusionAssemble2D(const int NE,
   auto B = Reshape(b.Read(), Q1D, D1D);
   auto G = Reshape(g.Read(), Q1D, D1D);
   auto D = Reshape(padata.Read(), Q1D, Q1D, 3, NE);
-   auto A = Reshape(eadata.ReadWrite(), D1D, D1D, D1D, D1D, NE);
+   auto A = Reshape(eadata.Write(), D1D, D1D, D1D, D1D, NE);
   MFEM_FORALL_3D(e, NE, D1D, D1D, 1,
   {
      const int D1D = T_D1D ? T_D1D : d1d;
@@ -120,7 +120,7 @@ static void EADiffusionAssemble2D(const int NE,
                               + gbi * D11 * gbj;
                     }
                  }
-                  A(i1, i2, j1, j2, e) += val;
+                  A(i1, i2, j1, j2, e) = val;
               }
            }
         }
@@ -144,7 +144,7 @@ static void EADiffusionAssemble3D(const int NE,
   auto B = Reshape(b.Read(), Q1D, D1D);
   auto G = Reshape(g.Read(), Q1D, D1D);
   auto D = Reshape(padata.Read(), Q1D, Q1D, Q1D, 6, NE);
-   auto A = Reshape(eadata.ReadWrite(), D1D, D1D, D1D, D1D, D1D, D1D, NE);
+   auto A = Reshape(eadata.Write(), D1D, D1D, D1D, D1D, D1D, D1D, NE);
   MFEM_FORALL_3D(e, NE, D1D, D1D, D1D,
   {
      const int D1D = T_D1D ? T_D1D : d1d;
@@ -208,7 +208,7 @@ static void EADiffusionAssemble3D(const int NE,
                              }
                           }
                        }
-                        A(i1, i2, i3, j1, j2, j3, e) += val;
+                        A(i1, i2, i3, j1, j2, j3, e) = val;
                     }
                  }
               }
@@ -170,54 +170,47 @@ static void PADiffusionSetup3D(const int Q1D,
                               const Vector &c,
                               Vector &d)
 {
+   const int NQ = Q1D*Q1D*Q1D;
   const bool const_c = c.Size() == 1;
-   const auto W = Reshape(w.Read(), Q1D,Q1D,Q1D);
-   const auto J = Reshape(j.Read(), Q1D,Q1D,Q1D,3,3,NE);
-   const auto C = const_c ? Reshape(c.Read(), 1,1,1,1) :
-                  Reshape(c.Read(), Q1D,Q1D,Q1D,NE);
-   auto D = Reshape(d.Write(), Q1D,Q1D,Q1D, 6, NE);
-   MFEM_FORALL_3D(e, NE, Q1D, Q1D, Q1D,
+   auto W = w.Read();
+   auto J = Reshape(j.Read(), NQ, 3, 3, NE);
+   auto C = const_c ? Reshape(c.Read(), 1, 1) : Reshape(c.Read(), NQ, NE);
+   auto D = Reshape(d.Write(), NQ, 6, NE);
+   MFEM_FORALL(eq, NE*NQ,
   {
-      MFEM_FOREACH_THREAD(qx,x,Q1D)
-      {
-         MFEM_FOREACH_THREAD(qy,y,Q1D)
-         {
-            MFEM_FOREACH_THREAD(qz,z,Q1D)
-            {
-               const double J11 = J(qx,qy,qz,0,0,e);
-               const double J21 = J(qx,qy,qz,1,0,e);
-               const double J31 = J(qx,qy,qz,2,0,e);
-               const double J12 = J(qx,qy,qz,0,1,e);
-               const double J22 = J(qx,qy,qz,1,1,e);
-               const double J32 = J(qx,qy,qz,2,1,e);
-               const double J13 = J(qx,qy,qz,0,2,e);
-               const double J23 = J(qx,qy,qz,1,2,e);
-               const double J33 = J(qx,qy,qz,2,2,e);
-               const double detJ = J11 * (J22 * J33 - J32 * J23) -
-               /* */               J21 * (J12 * J33 - J32 * J13) +
-               /* */               J31 * (J12 * J23 - J22 * J13);
-               const double coeff = const_c ? C(0,0,0,0) : C(qx,qy,qz,e);
-               const double c_detJ = W(qx,qy,qz) * coeff / detJ;
-               // adj(J)
-               const double A11 = (J22 * J33) - (J23 * J32);
-               const double A12 = (J32 * J13) - (J12 * J33);
-               const double A13 = (J12 * J23) - (J22 * J13);
-               const double A21 = (J31 * J23) - (J21 * J33);
-               const double A22 = (J11 * J33) - (J13 * J31);
-               const double A23 = (J21 * J13) - (J11 * J23);
-               const double A31 = (J21 * J32) - (J31 * J22);
-               const double A32 = (J31 * J12) - (J11 * J32);
-               const double A33 = (J11 * J22) - (J12 * J21);
-               // detJ J^{-1} J^{-T} = (1/detJ) adj(J) adj(J)^T
-               D(qx,qy,qz,0,e) = c_detJ * (A11*A11 + A12*A12 + A13*A13); // 1,1
-               D(qx,qy,qz,1,e) = c_detJ * (A11*A21 + A12*A22 + A13*A23); // 2,1
-               D(qx,qy,qz,2,e) = c_detJ * (A11*A31 + A12*A32 + A13*A33); // 3,1
-               D(qx,qy,qz,3,e) = c_detJ * (A21*A21 + A22*A22 + A23*A23); // 2,2
-               D(qx,qy,qz,4,e) = c_detJ * (A21*A31 + A22*A32 + A23*A33); // 3,2
-               D(qx,qy,qz,5,e) = c_detJ * (A31*A31 + A32*A32 + A33*A33); // 3,3
-            }
-         }
-      }
+      const int e = eq / NQ;
+      const int q = eq % NQ;
+      const double J11 = J(q,0,0,e);
+      const double J21 = J(q,1,0,e);
+      const double J31 = J(q,2,0,e);
+      const double J12 = J(q,0,1,e);
+      const double J22 = J(q,1,1,e);
+      const double J32 = J(q,2,1,e);
+      const double J13 = J(q,0,2,e);
+      const double J23 = J(q,1,2,e);
+      const double J33 = J(q,2,2,e);
+      const double detJ = J11 * (J22 * J33 - J32 * J23) -
+      /* */               J21 * (J12 * J33 - J32 * J13) +
+      /* */               J31 * (J12 * J23 - J22 * J13);
+      const double coeff = const_c ? C(0,0) : C(q,e);
+      const double c_detJ = W[q] * coeff / detJ;
+      // adj(J)
+      const double A11 = (J22 * J33) - (J23 * J32);
+      const double A12 = (J32 * J13) - (J12 * J33);
+      const double A13 = (J12 * J23) - (J22 * J13);
+      const double A21 = (J31 * J23) - (J21 * J33);
+      const double A22 = (J11 * J33) - (J13 * J31);
+      const double A23 = (J21 * J13) - (J11 * J23);
+      const double A31 = (J21 * J32) - (J31 * J22);
+      const double A32 = (J31 * J12) - (J11 * J32);
+      const double A33 = (J11 * J22) - (J12 * J21);
+      // detJ J^{-1} J^{-T} = (1/detJ) adj(J) adj(J)^T
+      D(q,0,e) = c_detJ * (A11*A11 + A12*A12 + A13*A13); // 1,1
+      D(q,1,e) = c_detJ * (A11*A21 + A12*A22 + A13*A23); // 2,1
+      D(q,2,e) = c_detJ * (A11*A31 + A12*A32 + A13*A33); // 3,1
+      D(q,3,e) = c_detJ * (A21*A21 + A22*A22 + A23*A23); // 2,2
+      D(q,4,e) = c_detJ * (A21*A31 + A22*A32 + A23*A33); // 3,2
+      D(q,5,e) = c_detJ * (A31*A31 + A32*A32 + A33*A33); // 3,3
   });
 }

@@ -259,7 +252,8 @@ static void PADiffusionSetup(const int dim,
   }
 }

-void DiffusionIntegrator::SetupPA(const FiniteElementSpace &fes)
+void DiffusionIntegrator::SetupPA(const FiniteElementSpace &fes,
+                                  const bool force)
 {
   // Assuming the same element type
   fespace = &fes;
@@ -268,7 +262,7 @@ void DiffusionIntegrator::SetupPA(const FiniteElementSpace &fes)
   const FiniteElement &el = *fes.GetFE(0);
   const IntegrationRule *ir = IntRule ? IntRule : &GetRule(el, el);
 #ifdef MFEM_USE_CEED
-   if (DeviceCanUseCeed())
+   if (DeviceCanUseCeed() && !force)
   {
      if (ceedDataPtr) { delete ceedDataPtr; }
      CeedData* ptr = new CeedData();
@@ -276,16 +270,17 @@ void DiffusionIntegrator::SetupPA(const FiniteElementSpace &fes)
      InitCeedCoeff(Q, ptr);
      return CeedPADiffusionAssemble(fes, *ir, *ptr);
   }
+#else
+   MFEM_CONTRACT_VAR(force);
 #endif
   const int dims = el.GetDim();
   const int symmDims = (dims * (dims + 1)) / 2; // 1x1: 1, 2x2: 3, 3x3: 6
   const int nq = ir->GetNPoints();
   dim = mesh->Dimension();
   ne = fes.GetNE();
-   const DofToQuad::Mode mode = DofToQuad::TENSOR;
-   geom = mesh->GetGeometricFactors(*ir, GeometricFactors::JACOBIANS, mode);
+   geom = mesh->GetGeometricFactors(*ir, GeometricFactors::JACOBIANS);
   const int sdim = mesh->SpaceDimension();
-   maps = &el.GetDofToQuad(*ir, mode);
+   maps = &el.GetDofToQuad(*ir, DofToQuad::TENSOR);
   dofs1D = maps->ndof;
   quad1D = maps->nqpt;
   pa_data.SetSize(symmDims * nq * ne, Device::GetDeviceMemoryType());
@@ -300,19 +295,6 @@ void DiffusionIntegrator::SetupPA(const FiniteElementSpace &fes)
      coeff.SetSize(1);
      coeff(0) = cQ->constant;
   }
-   else if (QuadratureFunctionCoefficient* cQ =
-               dynamic_cast<QuadratureFunctionCoefficient*>(Q))
-   {
-      const QuadratureFunction &qFun = cQ->GetQuadFunction();
-      MFEM_VERIFY(qFun.Size() == ne*nq,
-                  "Incompatible QuadratureFunction dimension \n");
-
-      MFEM_VERIFY(ir == &qFun.GetSpace()->GetElementIntRule(0),
-                  "IntegrationRule used within integrator and in"
-                  " QuadratureFunction appear to be different");
-      qFun.Read();
-      coeff.MakeRef(const_cast<QuadratureFunction &>(qFun),0);
-   }
   else
   {
      coeff.SetSize(nq * ne);
@@ -740,7 +722,6 @@ static void PADiffusionAssembleDiagonal(const int dim,
         case 0x23: return SmemPADiffusionDiagonal3D<2,3>(NE,B,G,D,Y);
         case 0x34: return SmemPADiffusionDiagonal3D<3,4>(NE,B,G,D,Y);
         case 0x45: return SmemPADiffusionDiagonal3D<4,5>(NE,B,G,D,Y);
-         case 0x46: return SmemPADiffusionDiagonal3D<4,6>(NE,B,G,D,Y);
         case 0x56: return SmemPADiffusionDiagonal3D<5,6>(NE,B,G,D,Y);
         case 0x67: return SmemPADiffusionDiagonal3D<6,7>(NE,B,G,D,Y);
         case 0x78: return SmemPADiffusionDiagonal3D<7,8>(NE,B,G,D,Y);
@@ -754,17 +735,9 @@ static void PADiffusionAssembleDiagonal(const int dim,

 void DiffusionIntegrator::AssembleDiagonalPA(Vector &diag)
 {
-#ifdef MFEM_USE_CEED
-   if (DeviceCanUseCeed())
-   {
-      CeedAssembleDiagonalPA(ceedDataPtr, diag);
-   }
-   else
-#endif
-   {
-      PADiffusionAssembleDiagonal(dim, dofs1D, quad1D, ne,
-                                  maps->B, maps->G, pa_data, diag);
-   }
+   if (pa_data.Size()==0) { SetupPA(*fespace, true); }
+   PADiffusionAssembleDiagonal(dim, dofs1D, quad1D, ne,
+                               maps->B, maps->G, pa_data, diag);
 }


@@ -1333,33 +1306,7 @@ static void PADiffusionApply3D(const int NE,
   });
 }

-// Half of B and G are stored in shared to get B, Bt, G and Gt.
-// Indices computation for SmemPADiffusionApply3D.
-static MFEM_HOST_DEVICE inline int qi(const int q, const int d, const int Q)
-{
-   return (q<=d) ? q : Q-1-q;
-}
-
-static MFEM_HOST_DEVICE inline int dj(const int q, const int d, const int D)
-{
-   return (q<=d) ? d : D-1-d;
-}
-
-static MFEM_HOST_DEVICE inline int qk(const int q, const int d, const int Q)
-{
-   return (q<=d) ? Q-1-q : q;
-}
-
-static MFEM_HOST_DEVICE inline int dl(const int q, const int d, const int D)
-{
-   return (q<=d) ? D-1-d : d;
-}
-
-static MFEM_HOST_DEVICE inline double sign(const int q, const int d)
-{
-   return (q<=d) ? -1.0 : 1.0;
-}
-
+// Shared memory PA Diffusion Apply 3D kernel
 template<int T_D1D = 0, int T_Q1D = 0>
 static void SmemPADiffusionApply3D(const int NE,
                                   const Array<double> &b_,
@@ -1372,27 +1319,28 @@ static void SmemPADiffusionApply3D(const int NE,
 {
   const int D1D = T_D1D ? T_D1D : d1d;
   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   constexpr int M1Q = T_Q1D ? T_Q1D : MAX_Q1D;
-   constexpr int M1D = T_D1D ? T_D1D : MAX_D1D;
-   MFEM_VERIFY(D1D <= M1D, "");
-   MFEM_VERIFY(Q1D <= M1Q, "");
+   constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
+   constexpr int MD1 = T_D1D ? T_D1D : MAX_D1D;
+   MFEM_VERIFY(D1D <= MD1, "");
+   MFEM_VERIFY(Q1D <= MQ1, "");
   auto b = Reshape(b_.Read(), Q1D, D1D);
   auto g = Reshape(g_.Read(), Q1D, D1D);
-   auto d = Reshape(d_.Read(), Q1D, Q1D, Q1D, 6, NE);
+   auto d = Reshape(d_.Read(), Q1D*Q1D*Q1D, 6, NE);
   auto x = Reshape(x_.Read(), D1D, D1D, D1D, NE);
   auto y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, NE);
-   MFEM_FORALL_3D(e, NE, Q1D, Q1D, 1,
+   MFEM_FORALL_3D(e, NE, Q1D, Q1D, Q1D,
   {
+      const int tidz = MFEM_THREAD_ID(z);
      const int D1D = T_D1D ? T_D1D : d1d;
      const int Q1D = T_Q1D ? T_Q1D : q1d;
      constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
      constexpr int MD1 = T_D1D ? T_D1D : MAX_D1D;
-      constexpr int MDQ = (MQ1 > MD1) ? MQ1 : MD1;
-      MFEM_SHARED double sBG[MQ1*MD1];
-      double (*B)[MD1] = (double (*)[MD1]) sBG;
-      double (*G)[MD1] = (double (*)[MD1]) sBG;
-      double (*Bt)[MQ1] = (double (*)[MQ1]) sBG;
-      double (*Gt)[MQ1] = (double (*)[MQ1]) sBG;
+      constexpr int MDQ = MQ1 > MD1 ? MQ1 : MD1;
+      MFEM_SHARED double sBG[2][MQ1*MD1];
+      double (*B)[MD1] = (double (*)[MD1]) (sBG+0);
+      double (*G)[MD1] = (double (*)[MD1]) (sBG+1);
+      double (*Bt)[MQ1] = (double (*)[MQ1]) (sBG+0);
+      double (*Gt)[MQ1] = (double (*)[MQ1]) (sBG+1);
      MFEM_SHARED double sm0[3][MDQ*MDQ*MDQ];
      MFEM_SHARED double sm1[3][MDQ*MDQ*MDQ];
      double (*X)[MD1][MD1]    = (double (*)[MD1][MD1]) (sm0+2);
@@ -1410,127 +1358,108 @@ static void SmemPADiffusionApply3D(const int NE,
      double (*QDD0)[MD1][MD1] = (double (*)[MD1][MD1]) (sm0+0);
      double (*QDD1)[MD1][MD1] = (double (*)[MD1][MD1]) (sm0+1);
      double (*QDD2)[MD1][MD1] = (double (*)[MD1][MD1]) (sm0+2);
-      MFEM_FOREACH_THREAD(dy,y,D1D)
+      MFEM_FOREACH_THREAD(dz,z,D1D)
      {
-         MFEM_FOREACH_THREAD(dx,x,D1D)
+         MFEM_FOREACH_THREAD(dy,y,D1D)
         {
-            MFEM_UNROLL(MD1)
-            for (int dz = 0; dz < D1D; ++dz)
+            MFEM_FOREACH_THREAD(dx,x,D1D)
            {
               X[dz][dy][dx] = x(dx,dy,dz,e);
            }
         }
-         MFEM_FOREACH_THREAD(qx,x,Q1D)
+      }
+      if (tidz == 0)
+      {
+         MFEM_FOREACH_THREAD(d,y,D1D)
         {
-            const int i = qi(qx,dy,Q1D);
-            const int j = dj(qx,dy,D1D);
-            const int k = qk(qx,dy,Q1D);
-            const int l = dl(qx,dy,D1D);
-            B[i][j] = b(qx,dy);
-            G[k][l] = g(qx,dy) * sign(qx,dy);
+            MFEM_FOREACH_THREAD(q,x,Q1D)
+            {
+               B[q][d] = b(q,d);
+               G[q][d] = g(q,d);
+            }
         }
      }
      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(dy,y,D1D)
+      MFEM_FOREACH_THREAD(dz,z,D1D)
      {
-         MFEM_FOREACH_THREAD(qx,x,Q1D)
+         MFEM_FOREACH_THREAD(dy,y,D1D)
         {
-            double u[D1D], v[D1D];
-            MFEM_UNROLL(MD1)
-            for (int dz = 0; dz < D1D; dz++) { u[dz] = v[dz] = 0.0; }
-            MFEM_UNROLL(MD1)
-            for (int dx = 0; dx < D1D; ++dx)
+            MFEM_FOREACH_THREAD(qx,x,Q1D)
            {
-               const int i = qi(qx,dx,Q1D);
-               const int j = dj(qx,dx,D1D);
-               const int k = qk(qx,dx,Q1D);
-               const int l = dl(qx,dx,D1D);
-               const double s = sign(qx,dx);
-               MFEM_UNROLL(MD1)
-               for (int dz = 0; dz < D1D; ++dz)
+               double u = 0.0;
+               double v = 0.0;
+               for (int dx = 0; dx < D1D; ++dx)
               {
                  const double coords = X[dz][dy][dx];
-                  u[dz] += coords * B[i][j];
-                  v[dz] += coords * G[k][l] * s;
+                  u += coords * B[qx][dx];
+                  v += coords * G[qx][dx];
               }
-            }
-            MFEM_UNROLL(MD1)
-            for (int dz = 0; dz < D1D; ++dz)
-            {
-               DDQ0[dz][dy][qx] = u[dz];
-               DDQ1[dz][dy][qx] = v[dz];
+               DDQ0[dz][dy][qx] = u;
+               DDQ1[dz][dy][qx] = v;
            }
         }
      }
      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(qy,y,Q1D)
+      MFEM_FOREACH_THREAD(dz,z,D1D)
      {
-         MFEM_FOREACH_THREAD(qx,x,Q1D)
+         MFEM_FOREACH_THREAD(qy,y,Q1D)
         {
-            double u[D1D], v[D1D], w[D1D];
-            MFEM_UNROLL(MD1)
-            for (int dz = 0; dz < D1D; dz++) { u[dz] = v[dz] = w[dz] = 0.0; }
-            MFEM_UNROLL(MD1)
-            for (int dy = 0; dy < D1D; ++dy)
+            MFEM_FOREACH_THREAD(qx,x,Q1D)
            {
-               const int i = qi(qy,dy,Q1D);
-               const int j = dj(qy,dy,D1D);
-               const int k = qk(qy,dy,Q1D);
-               const int l = dl(qy,dy,D1D);
-               const double s = sign(qy,dy);
-               MFEM_UNROLL(MD1)
-               for (int dz = 0; dz < D1D; dz++)
+               double u = 0.0;
+               double v = 0.0;
+               double w = 0.0;
+               for (int dy = 0; dy < D1D; ++dy)
               {
-                  u[dz] += DDQ1[dz][dy][qx] * B[i][j];
-                  v[dz] += DDQ0[dz][dy][qx] * G[k][l] * s;
-                  w[dz] += DDQ0[dz][dy][qx] * B[i][j];
+                  u += DDQ1[dz][dy][qx] * B[qy][dy];
+                  v += DDQ0[dz][dy][qx] * G[qy][dy];
+                  w += DDQ0[dz][dy][qx] * B[qy][dy];
               }
-            }
-            MFEM_UNROLL(MD1)
-            for (int dz = 0; dz < D1D; dz++)
-            {
-               DQQ0[dz][qy][qx] = u[dz];
-               DQQ1[dz][qy][qx] = v[dz];
-               DQQ2[dz][qy][qx] = w[dz];
+               DQQ0[dz][qy][qx] = u;
+               DQQ1[dz][qy][qx] = v;
+               DQQ2[dz][qy][qx] = w;
            }
         }
      }
      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(qy,y,Q1D)
+      MFEM_FOREACH_THREAD(qz,z,Q1D)
      {
-         MFEM_FOREACH_THREAD(qx,x,Q1D)
+         MFEM_FOREACH_THREAD(qy,y,Q1D)
         {
-            double u[Q1D], v[Q1D], w[Q1D];
-            MFEM_UNROLL(MQ1)
-            for (int qz = 0; qz < Q1D; qz++) { u[qz] = v[qz] = w[qz] = 0.0; }
-            MFEM_UNROLL(MD1)
-            for (int dz = 0; dz < D1D; ++dz)
+            MFEM_FOREACH_THREAD(qx,x,Q1D)
            {
-               MFEM_UNROLL(MQ1)
-               for (int qz = 0; qz < Q1D; qz++)
+               double u = 0.0;
+               double v = 0.0;
+               double w = 0.0;
+               for (int dz = 0; dz < D1D; ++dz)
               {
-                  const int i = qi(qz,dz,Q1D);
-                  const int j = dj(qz,dz,D1D);
-                  const int k = qk(qz,dz,Q1D);
-                  const int l = dl(qz,dz,D1D);
-                  const double s = sign(qz,dz);
-                  u[qz] += DQQ0[dz][qy][qx] * B[i][j];
-                  v[qz] += DQQ1[dz][qy][qx] * B[i][j];
-                  w[qz] += DQQ2[dz][qy][qx] * G[k][l] * s;
+                  u += DQQ0[dz][qy][qx] * B[qz][dz];
+                  v += DQQ1[dz][qy][qx] * B[qz][dz];
+                  w += DQQ2[dz][qy][qx] * G[qz][dz];
               }
+               QQQ0[qz][qy][qx] = u;
+               QQQ1[qz][qy][qx] = v;
+               QQQ2[qz][qy][qx] = w;
            }
-            MFEM_UNROLL(MQ1)
-            for (int qz = 0; qz < Q1D; qz++)
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(qz,z,Q1D)
+      {
+         MFEM_FOREACH_THREAD(qy,y,Q1D)
+         {
+            MFEM_FOREACH_THREAD(qx,x,Q1D)
            {
-               const double O11 = d(qx,qy,qz,0,e);
-               const double O12 = d(qx,qy,qz,1,e);
-               const double O13 = d(qx,qy,qz,2,e);
-               const double O22 = d(qx,qy,qz,3,e);
-               const double O23 = d(qx,qy,qz,4,e);
-               const double O33 = d(qx,qy,qz,5,e);
-               const double gX = u[qz];
-               const double gY = v[qz];
-               const double gZ = w[qz];
+               const int q = qx + ((qy*Q1D) + (qz*Q1D*Q1D));
+               const double O11 = d(q,0,e);
+               const double O12 = d(q,1,e);
+               const double O13 = d(q,2,e);
+               const double O22 = d(q,3,e);
+               const double O23 = d(q,4,e);
+               const double O33 = d(q,5,e);
+               const double gX = QQQ0[qz][qy][qx];
+               const double gY = QQQ1[qz][qy][qx];
+               const double gZ = QQQ2[qz][qy][qx];
               QQQ0[qz][qy][qx] = (O11*gX) + (O12*gY) + (O13*gZ);
               QQQ1[qz][qy][qx] = (O12*gX) + (O22*gY) + (O23*gZ);
               QQQ2[qz][qy][qx] = (O13*gX) + (O23*gY) + (O33*gZ);
@@ -1538,112 +1467,78 @@ static void SmemPADiffusionApply3D(const int NE,
         }
      }
      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(d,y,D1D)
+      if (tidz == 0)
      {
-         MFEM_FOREACH_THREAD(q,x,Q1D)
+         MFEM_FOREACH_THREAD(d,y,D1D)
         {
-            const int i = qi(q,d,Q1D);
-            const int j = dj(q,d,D1D);
-            const int k = qk(q,d,Q1D);
-            const int l = dl(q,d,D1D);
-            Bt[j][i] = b(q,d);
-            Gt[l][k] = g(q,d) * sign(q,d);
+            MFEM_FOREACH_THREAD(q,x,Q1D)
+            {
+               Bt[d][q] = b(q,d);
+               Gt[d][q] = g(q,d);
+            }
         }
      }
      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(qy,y,Q1D)
+      MFEM_FOREACH_THREAD(qz,z,Q1D)
      {
-         MFEM_FOREACH_THREAD(dx,x,D1D)
+         MFEM_FOREACH_THREAD(qy,y,Q1D)
         {
-            double u[Q1D], v[Q1D], w[Q1D];
-            MFEM_UNROLL(MQ1)
-            for (int qz = 0; qz < Q1D; ++qz) { u[qz] = v[qz] = w[qz] = 0.0; }
-            MFEM_UNROLL(MQ1)
-            for (int qx = 0; qx < Q1D; ++qx)
+            MFEM_FOREACH_THREAD(dx,x,D1D)
            {
-               const int i = qi(qx,dx,Q1D);
-               const int j = dj(qx,dx,D1D);
-               const int k = qk(qx,dx,Q1D);
-               const int l = dl(qx,dx,D1D);
-               const double s = sign(qx,dx);
-               MFEM_UNROLL(MQ1)
+               double u = 0.0;
+               double v = 0.0;
+               double w = 0.0;
+               for (int qx = 0; qx < Q1D; ++qx)
+               {
+                  u += QQQ0[qz][qy][qx] * Gt[dx][qx];
+                  v += QQQ1[qz][qy][qx] * Bt[dx][qx];
+                  w += QQQ2[qz][qy][qx] * Bt[dx][qx];
+               }
+               QQD0[qz][qy][dx] = u;
+               QQD1[qz][qy][dx] = v;
+               QQD2[qz][qy][dx] = w;
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(qz,z,Q1D)
+      {
+         MFEM_FOREACH_THREAD(dy,y,D1D)
+         {
+            MFEM_FOREACH_THREAD(dx,x,D1D)
+            {
+               double u = 0.0;
+               double v = 0.0;
+               double w = 0.0;
+               for (int qy = 0; qy < Q1D; ++qy)
+               {
+                  u += QQD0[qz][qy][dx] * Bt[dy][qy];
+                  v += QQD1[qz][qy][dx] * Gt[dy][qy];
+                  w += QQD2[qz][qy][dx] * Bt[dy][qy];
+               }
+               QDD0[qz][dy][dx] = u;
+               QDD1[qz][dy][dx] = v;
+               QDD2[qz][dy][dx] = w;
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(dz,z,D1D)
+      {
+         MFEM_FOREACH_THREAD(dy,y,D1D)
+         {
+            MFEM_FOREACH_THREAD(dx,x,D1D)
+            {
+               double u = 0.0;
+               double v = 0.0;
+               double w = 0.0;
               for (int qz = 0; qz < Q1D; ++qz)
               {
-                  u[qz] += QQQ0[qz][qy][qx] * Gt[l][k] * s;
-                  v[qz] += QQQ1[qz][qy][qx] * Bt[j][i];
-                  w[qz] += QQQ2[qz][qy][qx] * Bt[j][i];
+                  u += QDD0[qz][dy][dx] * Bt[dz][qz];
+                  v += QDD1[qz][dy][dx] * Bt[dz][qz];
+                  w += QDD2[qz][dy][dx] * Gt[dz][qz];
               }
-            }
-            MFEM_UNROLL(MQ1)
-            for (int qz = 0; qz < Q1D; ++qz)
-            {
-               QQD0[qz][qy][dx] = u[qz];
-               QQD1[qz][qy][dx] = v[qz];
-               QQD2[qz][qy][dx] = w[qz];
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(dy,y,D1D)
-      {
-         MFEM_FOREACH_THREAD(dx,x,D1D)
-         {
-            double u[Q1D], v[Q1D], w[Q1D];
-            MFEM_UNROLL(MQ1)
-            for (int qz = 0; qz < Q1D; ++qz) { u[qz] = v[qz] = w[qz] = 0.0; }
-            MFEM_UNROLL(MQ1)
-            for (int qy = 0; qy < Q1D; ++qy)
-            {
-               const int i = qi(qy,dy,Q1D);
-               const int j = dj(qy,dy,D1D);
-               const int k = qk(qy,dy,Q1D);
-               const int l = dl(qy,dy,D1D);
-               const double s = sign(qy,dy);
-               MFEM_UNROLL(MQ1)
-               for (int qz = 0; qz < Q1D; ++qz)
-               {
-                  u[qz] += QQD0[qz][qy][dx] * Bt[j][i];
-                  v[qz] += QQD1[qz][qy][dx] * Gt[l][k] * s;
-                  w[qz] += QQD2[qz][qy][dx] * Bt[j][i];
-               }
-            }
-            MFEM_UNROLL(MQ1)
-            for (int qz = 0; qz < Q1D; ++qz)
-            {
-               QDD0[qz][dy][dx] = u[qz];
-               QDD1[qz][dy][dx] = v[qz];
-               QDD2[qz][dy][dx] = w[qz];
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(dy,y,D1D)
-      {
-         MFEM_FOREACH_THREAD(dx,x,D1D)
-         {
-            double u[D1D], v[D1D], w[D1D];
-            MFEM_UNROLL(MD1)
-            for (int dz = 0; dz < D1D; ++dz) { u[dz] = v[dz] = w[dz] = 0.0; }
-            MFEM_UNROLL(MQ1)
-            for (int qz = 0; qz < Q1D; ++qz)
-            {
-               MFEM_UNROLL(MD1)
-               for (int dz = 0; dz < D1D; ++dz)
-               {
-                  const int i = qi(qz,dz,Q1D);
-                  const int j = dj(qz,dz,D1D);
-                  const int k = qk(qz,dz,Q1D);
-                  const int l = dl(qz,dz,D1D);
-                  const double s = sign(qz,dz);
-                  u[dz] += QDD0[qz][dy][dx] * Bt[j][i];
-                  v[dz] += QDD1[qz][dy][dx] * Bt[j][i];
-                  w[dz] += QDD2[qz][dy][dx] * Gt[l][k] * s;
-               }
-            }
-            MFEM_UNROLL(MD1)
-            for (int dz = 0; dz < D1D; ++dz)
-            {
-               y(dx,dy,dz,e) += (u[dz] + v[dz] + w[dz]);
+               y(dx,dy,dz,e) += (u + v + w);
            }
         }
      }
@@ -1678,11 +1573,9 @@ static void PADiffusionApply(const int dim,
      MFEM_ABORT("OCCA PADiffusionApply unknown kernel!");
   }
 #endif // MFEM_USE_OCCA
-   const int ID = (D1D << 4 ) | Q1D;
-
   if (dim == 2)
   {
-      switch (ID)
+      switch ((D1D << 4 ) | Q1D)
      {
         case 0x22: return SmemPADiffusionApply2D<2,2,16>(NE,B,G,D,X,Y);
         case 0x33: return SmemPADiffusionApply2D<3,3,16>(NE,B,G,D,X,Y);
@@ -1695,13 +1588,11 @@ static void PADiffusionApply(const int dim,
         default:   return PADiffusionApply2D(NE,B,G,Bt,Gt,D,X,Y,D1D,Q1D);
      }
   }
-
-   if (dim == 3)
+   else if (dim == 3)
   {
-      switch (ID)
+      switch ((D1D << 4 ) | Q1D)
      {
         case 0x23: return SmemPADiffusionApply3D<2,3>(NE,B,G,D,X,Y);
-         case 0x24: return SmemPADiffusionApply3D<2,4>(NE,B,G,D,X,Y);
         case 0x34: return SmemPADiffusionApply3D<3,4>(NE,B,G,D,X,Y);
         case 0x45: return SmemPADiffusionApply3D<4,5>(NE,B,G,D,X,Y);
         case 0x46: return SmemPADiffusionApply3D<4,6>(NE,B,G,D,X,Y);
@@ -1722,7 +1613,29 @@ void DiffusionIntegrator::AddMultPA(const Vector &x, Vector &y) const
 #ifdef MFEM_USE_CEED
   if (DeviceCanUseCeed())
   {
-      CeedAddMultPA(ceedDataPtr, x, y);
+      const CeedScalar *x_ptr;
+      CeedScalar *y_ptr;
+      CeedMemType mem;
+      CeedGetPreferredMemType(internal::ceed, &mem);
+      if ( Device::Allows(Backend::CUDA) && mem==CEED_MEM_DEVICE )
+      {
+         x_ptr = x.Read();
+         y_ptr = y.ReadWrite();
+      }
+      else
+      {
+         x_ptr = x.HostRead();
+         y_ptr = y.HostReadWrite();
+         mem = CEED_MEM_HOST;
+      }
+      CeedVectorSetArray(ceedDataPtr->u, mem, CEED_USE_POINTER,
+                         const_cast<CeedScalar*>(x_ptr));
+      CeedVectorSetArray(ceedDataPtr->v, mem, CEED_USE_POINTER, y_ptr);
+
+      CeedOperatorApplyAdd(ceedDataPtr->oper, ceedDataPtr->u, ceedDataPtr->v,
+                           CEED_REQUEST_IMMEDIATE);
+
+      CeedVectorSyncArray(ceedDataPtr->v, mem);
   }
   else
 #endif
@@ -114,8 +114,6 @@ void PAHdivMassApply2D(const int D1D,
                       Vector &_y)
 {
   constexpr static int VDIM = 2;
-   constexpr static int MAX_D1D = HDIV_MAX_D1D;
-   constexpr static int MAX_Q1D = HDIV_MAX_Q1D;

   auto Bo = Reshape(_Bo.Read(), Q1D, D1D-1);
   auto Bc = Reshape(_Bc.Read(), Q1D, D1D);
@@ -240,7 +238,6 @@ void PAHdivMassAssembleDiagonal2D(const int D1D,
                                  Vector &_diag)
 {
   constexpr static int VDIM = 2;
-   constexpr static int MAX_Q1D = HDIV_MAX_Q1D;

   auto Bo = Reshape(_Bo.Read(), Q1D, D1D-1);
   auto Bc = Reshape(_Bc.Read(), Q1D, D1D);
@@ -617,8 +614,6 @@ static void PADivDivApply2D(const int D1D,
                            Vector &_y)
 {
   constexpr static int VDIM = 2;
-   constexpr static int MAX_D1D = HDIV_MAX_D1D;
-   constexpr static int MAX_Q1D = HDIV_MAX_Q1D;

   auto Bo = Reshape(_Bo.Read(), Q1D, D1D-1);
   auto Bot = Reshape(_Bot.Read(), D1D-1, Q1D);
@@ -982,7 +977,6 @@ static void PADivDivAssembleDiagonal2D(const int D1D,
                                       Vector &_diag)
 {
   constexpr static int VDIM = 2;
-   constexpr static int MAX_Q1D = HDIV_MAX_Q1D;

   auto Bo = Reshape(_Bo.Read(), Q1D, D1D-1);
   auto Gc = Reshape(_Gc.Read(), Q1D, D1D);
@@ -1406,8 +1400,6 @@ static void PAHdivL2Apply2D(const int D1D,
                            Vector &_y)
 {
   constexpr static int VDIM = 2;
-   constexpr static int MAX_D1D = HDIV_MAX_D1D;
-   constexpr static int MAX_Q1D = HDIV_MAX_Q1D;

   auto Bo = Reshape(_Bo.Read(), Q1D, D1D-1);
   auto Gc = Reshape(_Gc.Read(), Q1D, D1D);
@@ -1674,8 +1666,6 @@ static void PAHdivL2ApplyTranspose2D(const int D1D,
                                     Vector &_y)
 {
   constexpr static int VDIM = 2;
-   constexpr static int MAX_D1D = HDIV_MAX_D1D;
-   constexpr static int MAX_Q1D = HDIV_MAX_Q1D;

   auto L2Bo = Reshape(_L2Bo.Read(), Q1D, L2D1D);
   auto Gct = Reshape(_Gct.Read(), D1D, Q1D);
@@ -1734,7 +1724,7 @@ static void PAHdivL2ApplyTranspose2D(const int D1D,

      for (int qy = 0; qy < Q1D; ++qy)
      {
-         double aX[MAX_D1D];
+         double aX[HDIV_MAX_D1D];

         int osc = 0;
         for (int c = 0; c < VDIM; ++c)  // loop over x, y components
@@ -30,7 +30,7 @@ static void EAMassAssemble1D(const int NE,
   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
   auto B = Reshape(basis.Read(), Q1D, D1D);
   auto D = Reshape(padata.Read(), Q1D, NE);
-   auto M = Reshape(eadata.ReadWrite(), D1D, D1D, NE);
+   auto M = Reshape(eadata.Write(), D1D, D1D, NE);
   MFEM_FORALL_3D(e, NE, D1D, D1D, 1,
   {
      const int D1D = T_D1D ? T_D1D : d1d;
@@ -52,7 +52,7 @@ static void EAMassAssemble1D(const int NE,
            {
               val += r_Bi[k1] * r_Bj[k1] * D(k1, e);
            }
-            M(i1, j1, e) += val;
+            M(i1, j1, e) = val;
         }
      }
   });
@@ -72,7 +72,7 @@ static void EAMassAssemble2D(const int NE,
   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
   auto B = Reshape(basis.Read(), Q1D, D1D);
   auto D = Reshape(padata.Read(), Q1D, Q1D, NE);
-   auto M = Reshape(eadata.ReadWrite(), D1D, D1D, D1D, D1D, NE);
+   auto M = Reshape(eadata.Write(), D1D, D1D, D1D, D1D, NE);
   MFEM_FORALL_3D(e, NE, D1D, D1D, 1,
   {
      const int D1D = T_D1D ? T_D1D : d1d;
@@ -114,7 +114,7 @@ static void EAMassAssemble2D(const int NE,
                               * s_D[k1][k2];
                     }
                  }
-                  M(i1, i2, j1, j2, e) += val;
+                  M(i1, i2, j1, j2, e) = val;
               }
            }
         }
@@ -136,7 +136,7 @@ static void EAMassAssemble3D(const int NE,
   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
   auto B = Reshape(basis.Read(), Q1D, D1D);
   auto D = Reshape(padata.Read(), Q1D, Q1D, Q1D, NE);
-   auto M = Reshape(eadata.ReadWrite(), D1D, D1D, D1D, D1D, D1D, D1D, NE);
+   auto M = Reshape(eadata.Write(), D1D, D1D, D1D, D1D, D1D, D1D, NE);
   MFEM_FORALL_3D(e, NE, D1D, D1D, D1D,
   {
      const int D1D = T_D1D ? T_D1D : d1d;
@@ -189,7 +189,7 @@ static void EAMassAssemble3D(const int NE,
                              }
                           }
                        }
-                        M(i1, i2, i3, j1, j2, j3, e) += val;
+                        M(i1, i2, i3, j1, j2, j3, e) = val;
                     }
                  }
               }
@@ -23,7 +23,7 @@ namespace mfem

 // PA Mass Assemble kernel

-void MassIntegrator::SetupPA(const FiniteElementSpace &fes)
+void MassIntegrator::SetupPA(const FiniteElementSpace &fes, const bool force)
 {

   // Assuming the same element type
@@ -34,7 +34,7 @@ void MassIntegrator::SetupPA(const FiniteElementSpace &fes)
   ElementTransformation *T = mesh->GetElementTransformation(0);
   const IntegrationRule *ir = IntRule ? IntRule : &GetRule(el, el, *T);
 #ifdef MFEM_USE_CEED
-   if (DeviceCanUseCeed())
+   if (DeviceCanUseCeed() && !force)
   {
      if (ceedDataPtr) { delete ceedDataPtr; }
      CeedData* ptr = new CeedData();
@@ -46,17 +46,9 @@ void MassIntegrator::SetupPA(const FiniteElementSpace &fes)
   dim = mesh->Dimension();
   ne = fes.GetMesh()->GetNE();
   nq = ir->GetNPoints();
-   const DofToQuad::Mode mode = DofToQuad::TENSOR;
-   const int flags = GeometricFactors::JACOBIANS |
-                     GeometricFactors::COORDINATES;
-#ifdef MFEM_USE_UMPIRE
-   const MemoryType temp_type = Device::GetDeviceMemoryType() == MemoryType::DEVICE_UMPIRE
-      ? MemoryType::DEVICE_UMPIRE_2 : Device::GetDeviceMemoryType();
-#else
-   const MemoryType temp_type = Device::GetDeviceMemoryType();
-#endif
-   geom = mesh->GetGeometricFactors(*ir, flags, mode, temp_type);
-   maps = &el.GetDofToQuad(*ir, mode);
+   geom = mesh->GetGeometricFactors(*ir, GeometricFactors::COORDINATES |
+                                    GeometricFactors::JACOBIANS);
+   maps = &el.GetDofToQuad(*ir, DofToQuad::TENSOR);
   dofs1D = maps->ndof;
   quad1D = maps->nqpt;
   pa_data.SetSize(ne*nq, Device::GetDeviceMemoryType());
@@ -72,26 +64,13 @@ void MassIntegrator::SetupPA(const FiniteElementSpace &fes)
   {
      coeff = new Vector;
      coeff->SetSize(1);
-      (*coeff)(0) = cQ->constant;
+      (*coeff)(0) = 1.0;
   }
   else if (QuadratureCoefficient* cQ = dynamic_cast<QuadratureCoefficient*>(Q))
   {
      coeff = cQ->Data();
      own_coeff = false;
   }
-   else if (QuadratureFunctionCoefficient* cQ =
-               dynamic_cast<QuadratureFunctionCoefficient*>(Q))
-   {
-      const QuadratureFunction &qFun = cQ->GetQuadFunction();
-      MFEM_VERIFY(qFun.Size() == nq * ne,
-                  "Incompatible QuadratureFunction dimension \n");
-
-      MFEM_VERIFY(ir == &qFun.GetSpace()->GetElementIntRule(0),
-                  "IntegrationRule used within integrator and in"
-                  " QuadratureFunction appear to be different");
-      qFun.Read();
-      coeff->MakeRef(const_cast<QuadratureFunction &>(qFun),0);
-   }
   else
   {
      coeff = new Vector;
@@ -134,38 +113,25 @@ void MassIntegrator::SetupPA(const FiniteElementSpace &fes)
   if (dim==3)
   {
      const int NE = ne;
-      const int Q1D = quad1D;
+      const int NQ = nq;
      const bool const_c = coeff->Size() == 1;
-      const auto W = Reshape(ir->GetWeights().Read(),Q1D,Q1D,Q1D);
-      const auto J = Reshape(geom->J.Read(), Q1D,Q1D,Q1D,3,3,NE);
-      const auto C = const_c ?
-                     Reshape(coeff->Read(), 1,1,1,1) :
-                     Reshape(coeff->Read(), Q1D,Q1D,Q1D,NE);
-      auto V = Reshape(pa_data.Write(), Q1D,Q1D,Q1D,NE);
-      MFEM_FORALL_3D(e, NE, Q1D, Q1D, Q1D,
+      auto W = ir->GetWeights().Read();
+      auto J = Reshape(geom->J.Read(), NQ,3,3,NE);
+      auto C =
+         const_c ? Reshape(coeff->Read(), 1,1) : Reshape(coeff->Read(), NQ,NE);
+      auto v = Reshape(pa_data.Write(), NQ,NE);
+      MFEM_FORALL(e, NE,
      {
-         MFEM_FOREACH_THREAD(qx,x,Q1D)
+         for (int q = 0; q < NQ; ++q)
         {
-            MFEM_FOREACH_THREAD(qy,y,Q1D)
-            {
-               MFEM_FOREACH_THREAD(qz,z,Q1D)
-               {
-                  const double J11 = J(qx,qy,qz,0,0,e);
-                  const double J12 = J(qx,qy,qz,0,1,e);
-                  const double J13 = J(qx,qy,qz,0,2,e);
-                  const double J21 = J(qx,qy,qz,1,0,e);
-                  const double J22 = J(qx,qy,qz,1,1,e);
-                  const double J23 = J(qx,qy,qz,1,2,e);
-                  const double J31 = J(qx,qy,qz,2,0,e);
-                  const double J32 = J(qx,qy,qz,2,1,e);
-                  const double J33 = J(qx,qy,qz,2,2,e);
-                  const double detJ = J11 * (J22 * J33 - J32 * J23) -
-                  /* */               J21 * (J12 * J33 - J32 * J13) +
-                  /* */               J31 * (J12 * J23 - J22 * J13);
-                  const double coeff = const_c ? C(0,0,0,0) : C(qx,qy,qz,e);
-                  V(qx,qy,qz,e) = W(qx,qy,qz) * coeff * detJ;
-               }
-            }
+            const double J11 = J(q,0,0,e), J12 = J(q,0,1,e), J13 = J(q,0,2,e);
+            const double J21 = J(q,1,0,e), J22 = J(q,1,1,e), J23 = J(q,1,2,e);
+            const double J31 = J(q,2,0,e), J32 = J(q,2,1,e), J33 = J(q,2,2,e);
+            const double detJ = J11 * (J22 * J33 - J32 * J23) -
+            /* */               J21 * (J12 * J33 - J32 * J13) +
+            /* */               J31 * (J12 * J23 - J22 * J13);
+            const double coeff = const_c ? C(0,0) : C(q,e);
+            v(q,e) = W[q] * coeff * detJ;
         }
      });
   }
@@ -472,12 +438,8 @@ static void PAMassAssembleDiagonal(const int dim, const int D1D,
      switch ((D1D << 4 ) | Q1D)
      {
         case 0x23: return SmemPAMassAssembleDiagonal3D<2,3>(NE,B,D,Y);
-         case 0x24: return SmemPAMassAssembleDiagonal3D<2,4>(NE,B,D,Y);
-         case 0x26: return SmemPAMassAssembleDiagonal3D<2,6>(NE,B,D,Y);
         case 0x34: return SmemPAMassAssembleDiagonal3D<3,4>(NE,B,D,Y);
-         case 0x35: return SmemPAMassAssembleDiagonal3D<3,5>(NE,B,D,Y);
         case 0x45: return SmemPAMassAssembleDiagonal3D<4,5>(NE,B,D,Y);
-         case 0x48: return SmemPAMassAssembleDiagonal3D<4,8>(NE,B,D,Y);
         case 0x56: return SmemPAMassAssembleDiagonal3D<5,6>(NE,B,D,Y);
         case 0x67: return SmemPAMassAssembleDiagonal3D<6,7>(NE,B,D,Y);
         case 0x78: return SmemPAMassAssembleDiagonal3D<7,8>(NE,B,D,Y);
@@ -490,16 +452,8 @@ static void PAMassAssembleDiagonal(const int dim, const int D1D,

 void MassIntegrator::AssembleDiagonalPA(Vector &diag)
 {
-#ifdef MFEM_USE_CEED
-   if (DeviceCanUseCeed())
-   {
-      CeedAssembleDiagonalPA(ceedDataPtr, diag);
-   }
-   else
-#endif
-   {
-      PAMassAssembleDiagonal(dim, dofs1D, quad1D, ne, maps->B, pa_data, diag);
-   }
+   if (pa_data.Size()==0) { SetupPA(*fespace, true); }
+   PAMassAssembleDiagonal(dim, dofs1D, quad1D, ne, maps->B, pa_data, diag);
 }


@@ -697,7 +651,6 @@ static void SmemPAMassApply2D(const int NE,
                              const int d1d = 0,
                              const int q1d = 0)
 {
-   MFEM_CONTRACT_VAR(bt_);
   const int D1D = T_D1D ? T_D1D : d1d;
   const int Q1D = T_Q1D ? T_Q1D : q1d;
   constexpr int NBZ = T_NBZ ? T_NBZ : 1;
@@ -961,7 +914,6 @@ static void SmemPAMassApply3D(const int NE,
                              const int d1d = 0,
                              const int q1d = 0)
 {
-   MFEM_CONTRACT_VAR(bt_);
   const int D1D = T_D1D ? T_D1D : d1d;
   const int Q1D = T_Q1D ? T_Q1D : q1d;
   constexpr int M1Q = T_Q1D ? T_Q1D : MAX_Q1D;
@@ -1211,13 +1163,10 @@ static void PAMassApply(const int dim,
         case 0x24: return SmemPAMassApply2D<2,4,16>(NE,B,Bt,D,X,Y);
         case 0x33: return SmemPAMassApply2D<3,3,16>(NE,B,Bt,D,X,Y);
         case 0x34: return SmemPAMassApply2D<3,4,16>(NE,B,Bt,D,X,Y);
-         case 0x35: return SmemPAMassApply2D<3,5,16>(NE,B,Bt,D,X,Y);
         case 0x36: return SmemPAMassApply2D<3,6,16>(NE,B,Bt,D,X,Y);
         case 0x44: return SmemPAMassApply2D<4,4,8>(NE,B,Bt,D,X,Y);
-         case 0x46: return SmemPAMassApply2D<4,6,8>(NE,B,Bt,D,X,Y);
         case 0x48: return SmemPAMassApply2D<4,8,4>(NE,B,Bt,D,X,Y);
         case 0x55: return SmemPAMassApply2D<5,5,8>(NE,B,Bt,D,X,Y);
-         case 0x57: return SmemPAMassApply2D<5,7,8>(NE,B,Bt,D,X,Y);
         case 0x58: return SmemPAMassApply2D<5,8,2>(NE,B,Bt,D,X,Y);
         case 0x66: return SmemPAMassApply2D<6,6,4>(NE,B,Bt,D,X,Y);
         case 0x77: return SmemPAMassApply2D<7,7,4>(NE,B,Bt,D,X,Y);
@@ -1225,7 +1174,6 @@ static void PAMassApply(const int dim,
         case 0x99: return SmemPAMassApply2D<9,9,2>(NE,B,Bt,D,X,Y);
         default:   return PAMassApply2D(NE,B,Bt,D,X,Y,D1D,Q1D);
      }
-      mfem::out << "Unknown 2D kernel 0x" << std::hex << id << std::endl;
   }
   else if (dim == 3)
   {
@@ -1234,9 +1182,7 @@ static void PAMassApply(const int dim,
         case 0x23: return SmemPAMassApply3D<2,3>(NE,B,Bt,D,X,Y);
         case 0x24: return SmemPAMassApply3D<2,4>(NE,B,Bt,D,X,Y);
         case 0x34: return SmemPAMassApply3D<3,4>(NE,B,Bt,D,X,Y);
-         case 0x35: return SmemPAMassApply3D<3,5>(NE,B,Bt,D,X,Y);
         case 0x36: return SmemPAMassApply3D<3,6>(NE,B,Bt,D,X,Y);
-         case 0x37: return SmemPAMassApply3D<3,7>(NE,B,Bt,D,X,Y);
         case 0x45: return SmemPAMassApply3D<4,5>(NE,B,Bt,D,X,Y);
         case 0x46: return SmemPAMassApply3D<4,6>(NE,B,Bt,D,X,Y);
         case 0x48: return SmemPAMassApply3D<4,8>(NE,B,Bt,D,X,Y);
@@ -1248,8 +1194,8 @@ static void PAMassApply(const int dim,
         case 0x9A: return SmemPAMassApply3D<9,10>(NE,B,Bt,D,X,Y);
         default:   return PAMassApply3D(NE,B,Bt,D,X,Y,D1D,Q1D);
      }
-      mfem::out << "Unknown 3D kernel 0x" << std::hex << id << std::endl;
   }
+   mfem::out << "Unknown kernel 0x" << std::hex << id << std::endl;
   MFEM_ABORT("Unknown kernel.");
 }

@@ -1258,7 +1204,29 @@ void MassIntegrator::AddMultPA(const Vector &x, Vector &y) const
 #ifdef MFEM_USE_CEED
   if (DeviceCanUseCeed())
   {
-      CeedAddMultPA(ceedDataPtr, x, y);
+      const CeedScalar *x_ptr;
+      CeedScalar *y_ptr;
+      CeedMemType mem;
+      CeedGetPreferredMemType(internal::ceed, &mem);
+      if ( Device::Allows(Backend::CUDA) && mem==CEED_MEM_DEVICE )
+      {
+         x_ptr = x.Read();
+         y_ptr = y.ReadWrite();
+      }
+      else
+      {
+         x_ptr = x.HostRead();
+         y_ptr = y.HostReadWrite();
+         mem = CEED_MEM_HOST;
+      }
+      CeedVectorSetArray(ceedDataPtr->u, mem, CEED_USE_POINTER,
+                         const_cast<CeedScalar*>(x_ptr));
+      CeedVectorSetArray(ceedDataPtr->v, mem, CEED_USE_POINTER, y_ptr);
+
+      CeedOperatorApplyAdd(ceedDataPtr->oper, ceedDataPtr->u, ceedDataPtr->v,
+                           CEED_REQUEST_IMMEDIATE);
+
+      CeedVectorSyncArray(ceedDataPtr->v, mem);
   }
   else
 #endif
@@ -25,7 +25,7 @@ void TransposeIntegrator::AssembleEA(const FiniteElementSpace &fes,
   if (ne == 0) { return; }
   const int dofs = fes.GetFE(0)->GetDof();
   auto A = Reshape(ea_data_tmp.Write(), dofs, dofs, ne);
-   auto AT = Reshape(ea_data.ReadWrite(), dofs, dofs, ne);
+   auto AT = Reshape(ea_data.Write(), dofs, dofs, ne);
   MFEM_FORALL(e, ne,
   {
      for (int i = 0; i < dofs; i++)
@@ -9,14 +9,12 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.

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

 namespace mfem
 {

 void PAHcurlSetup2D(const int Q1D,
-                    const int coeffDim,
                    const int NE,
                    const Array<double> &w,
                    const Vector &j,
@@ -24,7 +22,6 @@ void PAHcurlSetup2D(const int Q1D,
                    Vector &op);

 void PAHcurlSetup3D(const int Q1D,
-                    const int coeffDim,
                    const int NE,
                    const Array<double> &w,
                    const Vector &j,
@@ -175,36 +172,16 @@ void VectorFEMassIntegrator::AssemblePA(const FiniteElementSpace &fes)

   pa_data.SetSize(symmDims * nq * ne, Device::GetMemoryType());

-   const int coeffDim = VQ ? VQ->GetVDim() : 1;
-
-   Vector coeff(coeffDim * ne * nq);
+   Vector coeff(ne * nq);
   coeff = 1.0;
-   auto coeffh = Reshape(coeff.HostWrite(), coeffDim, nq, ne);
-   if (Q || VQ)
+   if (Q)
   {
-      Vector D(VQ ? coeffDim : 0);
-      if (VQ)
-      {
-         MFEM_VERIFY(coeffDim == dim, "");
-      }
-
      for (int e=0; e<ne; ++e)
      {
         ElementTransformation *tr = mesh->GetElementTransformation(e);
         for (int p=0; p<nq; ++p)
         {
-            if (VQ)
-            {
-               VQ->Eval(D, *tr, ir->IntPoint(p));
-               for (int i=0; i<coeffDim; ++i)
-               {
-                  coeffh(i, p, e) = D[i];
-               }
-            }
-            else
-            {
-               coeffh(0, p, e) = Q->Eval(*tr, ir->IntPoint(p));
-            }
+            coeff[p + (e * nq)] = Q->Eval(*tr, ir->IntPoint(p));
         }
      }
   }
@@ -213,12 +190,12 @@ void VectorFEMassIntegrator::AssemblePA(const FiniteElementSpace &fes)

   if (el->GetDerivType() == mfem::FiniteElement::CURL && dim == 3)
   {
-      PAHcurlSetup3D(quad1D, coeffDim, ne, ir->GetWeights(), geom->J,
+      PAHcurlSetup3D(quad1D, ne, ir->GetWeights(), geom->J,
                     coeff, pa_data);
   }
   else if (el->GetDerivType() == mfem::FiniteElement::CURL && dim == 2)
   {
-      PAHcurlSetup2D(quad1D, coeffDim, ne, ir->GetWeights(), geom->J,
+      PAHcurlSetup2D(quad1D, ne, ir->GetWeights(), geom->J,
                     coeff, pa_data);
   }
   else if (el->GetDerivType() == mfem::FiniteElement::DIV && dim == 3)
@@ -371,12 +348,12 @@ void MixedVectorGradientIntegrator::AssemblePA(const FiniteElementSpace
   // Use the same setup functions as VectorFEMassIntegrator.
   if (test_el->GetDerivType() == mfem::FiniteElement::CURL && dim == 3)
   {
-      PAHcurlSetup3D(quad1D, 1, ne, ir->GetWeights(), geom->J,
+      PAHcurlSetup3D(quad1D, ne, ir->GetWeights(), geom->J,
                     coeff, pa_data);
   }
   else if (test_el->GetDerivType() == mfem::FiniteElement::CURL && dim == 2)
   {
-      PAHcurlSetup2D(quad1D, 1, ne, ir->GetWeights(), geom->J,
+      PAHcurlSetup2D(quad1D, ne, ir->GetWeights(), geom->J,
                     coeff, pa_data);
   }
   else
@@ -217,24 +217,18 @@ void GradientGridFunctionCoefficient::Eval(
   GridFunc->GetGradients(T, ir, M);
 }

-CurlGridFunctionCoefficient::CurlGridFunctionCoefficient(
+CurlGridFunctionCoefficient::CurlGridFunctionCoefficient (
   const GridFunction *gf)
-   : VectorCoefficient(0)
+   : VectorCoefficient ((gf) ?
+                        gf -> FESpace() -> GetMesh() -> SpaceDimension() : 0)
 {
-   SetGridFunction(gf);
+   GridFunc = gf;
 }

 void CurlGridFunctionCoefficient::SetGridFunction(const GridFunction *gf)
 {
-   if (gf)
-   {
-      int sdim = gf -> FESpace() -> GetMesh() -> SpaceDimension();
-      MFEM_VERIFY(sdim == 2 || sdim == 3,
-                  "CurlGridFunctionCoefficient "
-                  "only defind for spaces of dimension 2 or 3.");
-   }
-   GridFunc = gf;
-   vdim = (gf) ? (2 * gf -> FESpace() -> GetMesh() -> SpaceDimension() - 3) : 0;
+   GridFunc = gf; vdim = (gf) ?
+                         gf -> FESpace() -> GetMesh() -> SpaceDimension() : 0;
 }

 void CurlGridFunctionCoefficient::Eval(Vector &V, ElementTransformation &T,
@@ -30,10 +30,7 @@ class ParMesh;
 /** @brief Base class Coefficients that optionally depend on space and time.
    These are used by the BilinearFormIntegrator, LinearFormIntegrator, and
    NonlinearFormIntegrator classes to represent the physical coefficients in
-    the PDEs that are being discretized. This class can also be used in a more
-    general way to represent functions that don't necessarily belong to a FE
-    space, e.g., to project onto GridFunctions to use as initial conditions,
-    exact solutions, etc. See, e.g., ex4 or ex22 for these uses. */
+    the PDEs that are being discretized. */
 class Coefficient
 {
 protected:
@@ -342,10 +342,11 @@ SesquilinearForm::FormLinearSystem(const Array<int> &ess_tdof_list,
                                   int ci)
 {
   FiniteElementSpace * fes = blfr->FESpace();
+
   int vsize  = fes->GetVSize();

   // Allocate temporary vectors
-   Vector b_0(vsize); b_0 = 0.0;
+   Vector b_0(vsize);  b_0 = 0.0;

   // Extract the real and imaginary parts of the input vectors
   MFEM_ASSERT(x.Size() == 2 * vsize, "Input GridFunction of incorrect size!");
@@ -359,7 +360,8 @@ SesquilinearForm::FormLinearSystem(const Array<int> &ess_tdof_list,
   if (conv == ComplexOperator::BLOCK_SYMMETRIC) { b_i *= -1.0; }

   int tvsize = fes->GetTrueVSize();
-   OperatorHandle A_r, A_i;
+   SparseMatrix * A_r = nullptr;
+   SparseMatrix * A_i = nullptr;

   X.SetSize(2 * tvsize);
   B.SetSize(2 * tvsize);
@@ -372,39 +374,42 @@ SesquilinearForm::FormLinearSystem(const Array<int> &ess_tdof_list,

   if (RealInteg())
   {
+      A_r = new SparseMatrix;
      blfr->SetDiagonalPolicy(diag_policy);

      b_0 = b_r;
-      blfr->FormLinearSystem(ess_tdof_list, x_r, b_0, A_r, X_0, B_0, ci);
+      blfr->FormLinearSystem(ess_tdof_list, x_r, b_0, *A_r, X_0, B_0, ci);
      X_r = X_0; B_r = B_0;

      b_0 = b_i;
-      blfr->FormLinearSystem(ess_tdof_list, x_i, b_0, A_r, X_0, B_0, ci);
+      blfr->FormLinearSystem(ess_tdof_list, x_i, b_0, *A_r, X_0, B_0, ci);
      X_i = X_0; B_i = B_0;

      if (ImagInteg())
      {
+         A_i = new SparseMatrix;
         blfi->SetDiagonalPolicy(mfem::Matrix::DiagonalPolicy::DIAG_ZERO);

         b_0 = 0.0;
-         blfi->FormLinearSystem(ess_tdof_list, x_i, b_0, A_i, X_0, B_0, false);
+         blfi->FormLinearSystem(ess_tdof_list, x_i, b_0, *A_i, X_0, B_0, false);
         B_r -= B_0;

         b_0 = 0.0;
-         blfi->FormLinearSystem(ess_tdof_list, x_r, b_0, A_i, X_0, B_0, false);
+         blfi->FormLinearSystem(ess_tdof_list, x_r, b_0, *A_i, X_0, B_0, false);
         B_i += B_0;
      }
   }
   else if (ImagInteg())
   {
+      A_i = new SparseMatrix;
      blfi->SetDiagonalPolicy(diag_policy);

      b_0 = b_i;
-      blfi->FormLinearSystem(ess_tdof_list, x_r, b_0, A_i, X_0, B_0, ci);
+      blfi->FormLinearSystem(ess_tdof_list, x_r, b_0, *A_i, X_0, B_0, ci);
      X_r = X_0; B_i = B_0;

      b_0 = b_r; b_0 *= -1.0;
-      blfi->FormLinearSystem(ess_tdof_list, x_i, b_0, A_i, X_0, B_0, ci);
+      blfi->FormLinearSystem(ess_tdof_list, x_i, b_0, *A_i, X_0, B_0, ci);
      X_i = X_0; B_r = B_0; B_r *= -1.0;
   }
   else
@@ -412,55 +417,16 @@ SesquilinearForm::FormLinearSystem(const Array<int> &ess_tdof_list,
      MFEM_ABORT("Real and Imaginary part of the Sesquilinear form are empty");
   }

-   if (RealInteg() && ImagInteg())
-   {
-      // Modify RHS and offdiagonal blocks (imaginary parts of the matrix) to
-      // conform with standard essential BC treatment
-      if (A_i.Is<ConstrainedOperator>())
-      {
-         int n = ess_tdof_list.Size();
-         for (int k = 0; k < n; k++)
-         {
-            int j = ess_tdof_list[k];
-            B_r(j) = X_r(j);
-            B_i(j) = X_i(j);
-         }
-         A_i.As<ConstrainedOperator>()->SetDiagonalPolicy
-         (mfem::Operator::DiagonalPolicy::DIAG_ZERO);
-      }
-   }
-
   if (conv == ComplexOperator::BLOCK_SYMMETRIC)
   {
      B_i *= -1.0;
      b_i *= -1.0;
   }
-
   // A = A_r + i A_i
   A.Clear();
-   if ( A_r.Type() == Operator::MFEM_SPARSEMAT ||
-        A_i.Type() == Operator::MFEM_SPARSEMAT )
-   {
-      ComplexSparseMatrix * A_sp =
-         new ComplexSparseMatrix(A_r.As<SparseMatrix>(),
-                                 A_i.As<SparseMatrix>(),
-                                 A_r.OwnsOperator(),
-                                 A_i.OwnsOperator(),
-                                 conv);
-      A.Reset<ComplexSparseMatrix>(A_sp, true);
-   }
-   else
-   {
-      ComplexOperator * A_op =
-         new ComplexOperator(A_r.Ptr(),
-                             A_i.Ptr(),
-                             A_r.OwnsOperator(),
-                             A_i.OwnsOperator(),
-                             conv);
-      A.Reset<ComplexOperator>(A_op, true);
-   }
-   A_r.SetOperatorOwner(false);
-   A_i.SetOperatorOwner(false);
+   ComplexSparseMatrix * A_sp;
+   A_sp = new ComplexSparseMatrix(A_r, A_i, true, true, conv);
+   A.Reset<ComplexSparseMatrix>(A_sp, true);
 }

 void
@@ -468,60 +434,31 @@ SesquilinearForm::FormSystemMatrix(const Array<int> &ess_tdof_list,
                                   OperatorHandle &A)

 {
-   OperatorHandle A_r, A_i;
+   SparseMatrix * A_r = nullptr;
+   SparseMatrix * A_i = nullptr;
+
   if (RealInteg())
   {
+      A_r = new SparseMatrix;
      blfr->SetDiagonalPolicy(diag_policy);
-      blfr->FormSystemMatrix(ess_tdof_list, A_r);
+      blfr->FormSystemMatrix(ess_tdof_list, *A_r);
   }
   if (ImagInteg())
   {
-      blfi->SetDiagonalPolicy(RealInteg() ?
-                              mfem::Matrix::DiagonalPolicy::DIAG_ZERO :
-                              diag_policy);
-      blfi->FormSystemMatrix(ess_tdof_list, A_i);
+      A_i = new SparseMatrix;
+      blfr->SetDiagonalPolicy(diag_policy);
+      blfi->FormSystemMatrix(ess_tdof_list, *A_i);
   }
   if (!RealInteg() && !ImagInteg())
   {
      MFEM_ABORT("Both Real and Imaginary part of the Sesquilinear form are empty");
   }

-   if (RealInteg() && ImagInteg())
-   {
-      // Modify offdiagonal blocks (imaginary parts of the matrix) to conform
-      // with standard essential BC treatment
-      if (A_i.Is<ConstrainedOperator>())
-      {
-         A_i.As<ConstrainedOperator>()->SetDiagonalPolicy
-         (mfem::Operator::DiagonalPolicy::DIAG_ZERO);
-      }
-   }
-
   // A = A_r + i A_i
   A.Clear();
-   if ( A_r.Type() == Operator::MFEM_SPARSEMAT ||
-        A_i.Type() == Operator::MFEM_SPARSEMAT )
-   {
-      ComplexSparseMatrix * A_sp =
-         new ComplexSparseMatrix(A_r.As<SparseMatrix>(),
-                                 A_i.As<SparseMatrix>(),
-                                 A_r.OwnsOperator(),
-                                 A_i.OwnsOperator(),
-                                 conv);
-      A.Reset<ComplexSparseMatrix>(A_sp, true);
-   }
-   else
-   {
-      ComplexOperator * A_op =
-         new ComplexOperator(A_r.Ptr(),
-                             A_i.Ptr(),
-                             A_r.OwnsOperator(),
-                             A_i.OwnsOperator(),
-                             conv);
-      A.Reset<ComplexOperator>(A_op, true);
-   }
-   A_r.SetOperatorOwner(false);
-   A_i.SetOperatorOwner(false);
+   ComplexSparseMatrix * A_sp =
+      new ComplexSparseMatrix(A_r, A_i, true, true, conv);
+   A.Reset<ComplexSparseMatrix>(A_sp, true);
 }

 void
@@ -709,7 +646,7 @@ ParComplexLinearForm::ParComplexLinearForm(ParFiniteElementSpace *pfes,
   int n = (HYPRE_AssumedPartitionCheck()) ? 2 : pfes->GetNRanks();
   tdof_offsets = new HYPRE_Int[n+1];

-   for (int i = 0; i <= n; i++)
+   for (int i=0; i<=n; i++)
   {
      tdof_offsets[i] = 2 * tdof_offsets_fes[i];
   }
@@ -717,8 +654,7 @@ ParComplexLinearForm::ParComplexLinearForm(ParFiniteElementSpace *pfes,


 ParComplexLinearForm::ParComplexLinearForm(ParFiniteElementSpace *pfes,
-                                           ParLinearForm *plf_r,
-                                           ParLinearForm *plf_i,
+                                           ParLinearForm *plf_r, ParLinearForm *plf_i,
                                           ComplexOperator::Convention
                                           convention)
   : Vector(2*(pfes->GetVSize())),
@@ -734,7 +670,7 @@ ParComplexLinearForm::ParComplexLinearForm(ParFiniteElementSpace *pfes,
   int n = (HYPRE_AssumedPartitionCheck()) ? 2 : pfes->GetNRanks();
   tdof_offsets = new HYPRE_Int[n+1];

-   for (int i = 0; i <= n; i++)
+   for (int i=0; i<=n; i++)
   {
      tdof_offsets[i] = 2 * tdof_offsets_fes[i];
   }
@@ -881,8 +817,7 @@ ParSesquilinearForm::ParSesquilinearForm(ParFiniteElementSpace *pf,
 {}

 ParSesquilinearForm::ParSesquilinearForm(ParFiniteElementSpace *pf,
-                                         ParBilinearForm *pbfr,
-                                         ParBilinearForm *pbfi,
+                                         ParBilinearForm *pbfr, ParBilinearForm *pbfi,
                                         ComplexOperator::Convention convention)
   : conv(convention),
     pblfr(new ParBilinearForm(pf,pbfr)),
@@ -978,10 +913,9 @@ ParSesquilinearForm::FormLinearSystem(const Array<int> &ess_tdof_list,
   int vsize = pfes->GetVSize();

   // Allocate temporary vectors
-   Vector b_0(vsize); b_0 = 0.0;
+   Vector b_0(vsize);  b_0 = 0.0;

   // Extract the real and imaginary parts of the input vectors
-   MFEM_ASSERT(x.Size() == 2 * vsize, "Input GridFunction of incorrect size!");
   Vector x_r(x.GetData(), vsize);
   Vector x_i(&(x.GetData())[vsize], vsize);

@@ -1040,34 +974,25 @@ ParSesquilinearForm::FormLinearSystem(const Array<int> &ess_tdof_list,
      MFEM_ABORT("Real and Imaginary part of the Sesquilinear form are empty");
   }

+   // Modify RHS and offdiagonal blocks (Imaginary parts of the matrix) to
+   // conform with standard essential BC treatment i.e. zero out rows and
+   // columns and place ones on the diagonal.
   if (RealInteg() && ImagInteg())
   {
-      int n = ess_tdof_list.Size();
-      // Modify RHS to conform with standard essential BC treatment
-      for (int k = 0; k < n; k++)
-      {
-         int j=ess_tdof_list[k];
-         B_r(j) = X_r(j);
-         B_i(j) = X_i(j);
-      }
-      // Modify offdiagonal blocks (imaginary parts of the matrix) to conform
-      // with standard essential BC treatment
      if ( A_i.Type() == Operator::Hypre_ParCSR )
      {
-         HypreParMatrix * Ah;
-         A_i.Get(Ah);
-         hypre_ParCSRMatrix *Aih = *Ah;
-         for (int k = 0; k < n; k++)
+         HypreParMatrix * Ah;  A_i.Get(Ah);
+         int n = ess_tdof_list.Size();
+         hypre_ParCSRMatrix * Aih =
+            (hypre_ParCSRMatrix *)const_cast<HypreParMatrix&>(*Ah);
+         for (int k=0; k<n; k++)
         {
-            int j = ess_tdof_list[k];
+            int j=ess_tdof_list[k];
            Aih->diag->data[Aih->diag->i[j]] = 0.0;
+            B_r(j) = X_r(j);
+            B_i(j) = X_i(j);
         }
      }
-      else
-      {
-         A_i.As<ConstrainedOperator>()->SetDiagonalPolicy
-         (mfem::Operator::DiagonalPolicy::DIAG_ZERO);
-      }
   }

   if (conv == ComplexOperator::BLOCK_SYMMETRIC)
@@ -1075,7 +1000,6 @@ ParSesquilinearForm::FormLinearSystem(const Array<int> &ess_tdof_list,
      B_i *= -1.0;
      b_i *= -1.0;
   }
-
   // A = A_r + i A_i
   A.Clear();
   if ( A_r.Type() == Operator::Hypre_ParCSR ||
@@ -1099,8 +1023,6 @@ ParSesquilinearForm::FormLinearSystem(const Array<int> &ess_tdof_list,
                             conv);
      A.Reset<ComplexOperator>(A_op, true);
   }
-   A_r.SetOperatorOwner(false);
-   A_i.SetOperatorOwner(false);
 }

 void
@@ -1121,27 +1043,25 @@ ParSesquilinearForm::FormSystemMatrix(const Array<int> &ess_tdof_list,
      MFEM_ABORT("Both Real and Imaginary part of the Sesquilinear form are empty");
   }

+   // Modify offdiagonal blocks (Imaginary parts of the matrix) to conform with
+   // standard essential BC treatment i.e. zero out rows and columns and place
+   // ones on the diagonal.
   if (RealInteg() && ImagInteg())
   {
-      // Modify offdiagonal blocks (imaginary parts of the matrix) to conform
-      // with standard essential BC treatment
      if ( A_i.Type() == Operator::Hypre_ParCSR )
      {
         int n = ess_tdof_list.Size();
-         HypreParMatrix * Ah;
-         A_i.Get(Ah);
-         hypre_ParCSRMatrix * Aih = *Ah;
-         for (int k = 0; k < n; k++)
+         int j;
+
+         HypreParMatrix * Ah;  A_i.Get(Ah);
+         hypre_ParCSRMatrix * Aih =
+            (hypre_ParCSRMatrix *)const_cast<HypreParMatrix&>(*Ah);
+         for (int k=0; k<n; k++)
         {
-            int j = ess_tdof_list[k];
+            j=ess_tdof_list[k];
            Aih->diag->data[Aih->diag->i[j]] = 0.0;
         }
      }
-      else
-      {
-         A_i.As<ConstrainedOperator>()->SetDiagonalPolicy
-         (mfem::Operator::DiagonalPolicy::DIAG_ZERO);
-      }
   }

   // A = A_r + i A_i
@@ -1167,8 +1087,6 @@ ParSesquilinearForm::FormSystemMatrix(const Array<int> &ess_tdof_list,
                             conv);
      A.Reset<ComplexOperator>(A_op, true);
   }
-   A_r.SetOperatorOwner(false);
-   A_i.SetOperatorOwner(false);
 }

 void
@@ -219,21 +219,6 @@ public:
   void SetConvention(const ComplexOperator::Convention &
                      convention) { conv = convention; }

-   /// Set the desired assembly level.
-   /** Valid choices are:
-
-       - AssemblyLevel::FULL  (default)
-       - AssemblyLevel::PARTIAL
-       - AssemblyLevel::ELEMENT
-       - AssemblyLevel::NONE
-
-       This method must be called before assembly. */
-   void SetAssemblyLevel(AssemblyLevel assembly_level)
-   {
-      blfr->SetAssemblyLevel(assembly_level);
-      blfi->SetAssemblyLevel(assembly_level);
-   }
-
   BilinearForm & real() { return *blfr; }
   BilinearForm & imag() { return *blfi; }
   const BilinearForm & real() const { return *blfr; }
@@ -493,7 +478,7 @@ public:
 /** Class for a parallel sesquilinear form

    A sesquilinear form is a generalization of a bilinear form to complex-valued
-    fields. Sesquilinear forms are linear in the second argument but the
+    fields. Sesquilinear forms are linear in the second argument but but the
    first argument involves a complex conjugate in the sense that:

                a(alpha u, beta v) = conj(alpha) beta a(u, v)
@@ -539,21 +524,6 @@ public:
   void SetConvention(const ComplexOperator::Convention &
                      convention) { conv = convention; }

-   /// Set the desired assembly level.
-   /** Valid choices are:
-
-       - AssemblyLevel::FULL  (default)
-       - AssemblyLevel::PARTIAL
-       - AssemblyLevel::ELEMENT
-       - AssemblyLevel::NONE
-
-       This method must be called before assembly. */
-   void SetAssemblyLevel(AssemblyLevel assembly_level)
-   {
-      pblfr->SetAssemblyLevel(assembly_level);
-      pblfi->SetAssemblyLevel(assembly_level);
-   }
-
   ParBilinearForm & real() { return *pblfr; }
   ParBilinearForm & imag() { return *pblfi; }
   const ParBilinearForm & real() const { return *pblfr; }
@@ -415,6 +415,9 @@ void VisItDataCollection::SetMesh(MPI_Comm comm, Mesh *new_mesh)
 void VisItDataCollection::RegisterField(const std::string& name,
                                        GridFunction *gf)
 {
+   DataCollection::RegisterField(name, gf);
+   field_info_map[name] = VisItFieldInfo("nodes", gf->VectorDim());
+
   int LOD = 1;
   if (gf->FESpace()->GetNURBSext())
   {
@@ -428,27 +431,6 @@ void VisItDataCollection::RegisterField(const std::string& name,
      }
   }

-   DataCollection::RegisterField(name, gf);
-   field_info_map[name] = VisItFieldInfo("nodes", gf->VectorDim(), LOD);
-   visit_levels_of_detail = std::max(visit_levels_of_detail, LOD);
-}
-
-void VisItDataCollection::RegisterQField(const std::string& name,
-                                         QuadratureFunction *qf)
-{
-   int LOD = -1;
-   Mesh *mesh = qf->GetSpace()->GetMesh();
-   for (int e=0; e<qf->GetSpace()->GetNE(); e++)
-   {
-      int locLOD = GlobGeometryRefiner.GetRefinementLevelFromElems(
-                      mesh->GetElementBaseGeometry(e),
-                      qf->GetElementIntRule(e).GetNPoints());
-
-      LOD = std::max(LOD,locLOD);
-   }
-
-   DataCollection::RegisterQField(name, qf);
-   field_info_map[name] = VisItFieldInfo("elements", 1, LOD);
   visit_levels_of_detail = std::max(visit_levels_of_detail, LOD);
 }

@@ -616,28 +598,14 @@ void VisItDataCollection::LoadFields()
      // TODO: 1) load parallel GridFunction on one processor
      if (serial)
      {
-         if ((it->second).association == "nodes")
-         {
-            field_map.Register(it->first, new GridFunction(mesh, file), own_data);
-         }
-         else if ((it->second).association == "elements")
-         {
-            q_field_map.Register(it->first, new QuadratureFunction(mesh, file), own_data);
-         }
+         field_map.Register(it->first, new GridFunction(mesh, file), own_data);
      }
      else
      {
 #ifdef MFEM_USE_MPI
-         if ((it->second).association == "nodes")
-         {
-            field_map.Register(
-               it->first,
-               new ParGridFunction(dynamic_cast<ParMesh*>(mesh), file), own_data);
-         }
-         else if ((it->second).association == "elements")
-         {
-            q_field_map.Register(it->first, new QuadratureFunction(mesh, file), own_data);
-         }
+         field_map.Register(
+            it->first,
+            new ParGridFunction(dynamic_cast<ParMesh*>(mesh), file), own_data);
 #else
         error = READ_ERROR;
         MFEM_WARNING("Reading parallel format in serial is not supported");
@@ -672,7 +640,7 @@ std::string VisItDataCollection::GetVisItRootString()
   {
      ftags["assoc"] = picojson::value((it->second).association);
      ftags["comps"] = picojson::value(to_string((it->second).num_components));
-      ftags["lod"] = picojson::value(to_string((it->second).lod));
+      ftags["lod"] = picojson::value(to_string(visit_levels_of_detail));
      field["path"] = picojson::value(path_str + it->first + file_ext_format);
      field["tags"] = picojson::value(ftags);
      fields[it->first] = picojson::value(field);
@@ -391,10 +391,9 @@ class VisItFieldInfo
 public:
   std::string association;
   int num_components;
-   int lod;
-   VisItFieldInfo() { association = ""; num_components = 0; lod = 1;}
-   VisItFieldInfo(std::string _association, int _num_components, int _lod = 1)
-   { association = _association; num_components = _num_components; lod =_lod;}
+   VisItFieldInfo() { association = ""; num_components = 0; }
+   VisItFieldInfo(std::string _association, int _num_components)
+   { association = _association; num_components = _num_components; }
 };

 /// Data collection with VisIt I/O routines
@@ -446,12 +445,6 @@ public:
   /// Add a grid function to the collection and update the root file
   virtual void RegisterField(const std::string& field_name, GridFunction *gf);

-   /// Add a quadrature function to the collection and update the root file.
-   /** Visualization of quadrature function is not supported in VisIt(3.12).
-       A patch has been sent to VisIt developers in June 2020. */
-   virtual void RegisterQField(const std::string& q_field_name,
-                               QuadratureFunction *qf);
-
   /// Set VisIt parameter: default levels of detail for the MultiresControl
   void SetLevelsOfDetail(int levels_of_detail);

@@ -552,32 +552,26 @@ void IntegrationPointTransformation::Transform (const IntegrationRule &ir1,
   }
 }

-void FaceElementTransformations::SetIntPoint(const IntegrationPoint *face_ip)
+void FaceElementTransformations::SetIntPoint(const IntegrationPoint *ip)
 {
-   IsoparametricTransformation::SetIntPoint(face_ip);
+   IsoparametricTransformation::SetIntPoint(ip);

-   if (mask & 4)
+   if (Elem1)
   {
-      Loc1.Transform(*face_ip, eip1);
-      if (Elem1)
-      {
-         Elem1->SetIntPoint(&eip1);
-      }
+      Loc1.Transform(*ip, eip1);
+      Elem1->SetIntPoint(&eip1);
   }
-   if (mask & 8)
+   if (Elem2)
   {
-      Loc2.Transform(*face_ip, eip2);
-      if (Elem2)
-      {
-         Elem2->SetIntPoint(&eip2);
-      }
+      Loc2.Transform(*ip, eip2);
+      Elem2->SetIntPoint(&eip2);
   }
 }

 ElementTransformation &
 FaceElementTransformations::GetElement1Transformation()
 {
-   MFEM_VERIFY(mask & HAVE_ELEM1 && Elem1 != NULL, "The ElementTransformation "
+   MFEM_VERIFY(mask & 1 && Elem1 != NULL, "The ElementTransformation "
               "for the element has not been configured for side 1.");
   return *Elem1;
 }
@@ -585,7 +579,7 @@ FaceElementTransformations::GetElement1Transformation()
 ElementTransformation &
 FaceElementTransformations::GetElement2Transformation()
 {
-   MFEM_VERIFY(mask & HAVE_ELEM2 && Elem2 != NULL, "The ElementTransformation "
+   MFEM_VERIFY(mask & 2 && Elem2 != NULL, "The ElementTransformation "
               "for the element has not been configured for side 2.");
   return *Elem2;
 }
@@ -593,7 +587,7 @@ FaceElementTransformations::GetElement2Transformation()
 IntegrationPointTransformation &
 FaceElementTransformations::GetIntPoint1Transformation()
 {
-   MFEM_VERIFY(mask & HAVE_LOC1, "The IntegrationPointTransformation "
+   MFEM_VERIFY(mask & 4, "The IntegrationPointTransformation "
               "for the element has not been configured for side 1.");
   return Loc1;
 }
@@ -601,7 +595,7 @@ FaceElementTransformations::GetIntPoint1Transformation()
 IntegrationPointTransformation &
 FaceElementTransformations::GetIntPoint2Transformation()
 {
-   MFEM_VERIFY(mask & HAVE_LOC2, "The IntegrationPointTransformation "
+   MFEM_VERIFY(mask & 8, "The IntegrationPointTransformation "
               "for the element has not been configured for side 2.");
   return Loc2;
 }
@@ -609,7 +603,7 @@ FaceElementTransformations::GetIntPoint2Transformation()
 void FaceElementTransformations::Transform(const IntegrationPoint &ip,
                                           Vector &trans)
 {
-   MFEM_VERIFY(mask & HAVE_FACE, "The ElementTransformation "
+   MFEM_VERIFY(mask & 16, "The ElementTransformation "
               "for the face has not been configured.");
   IsoparametricTransformation::Transform(ip, trans);
 }
@@ -617,7 +611,7 @@ void FaceElementTransformations::Transform(const IntegrationPoint &ip,
 void FaceElementTransformations::Transform(const IntegrationRule &ir,
                                           DenseMatrix &tr)
 {
-   MFEM_VERIFY(mask & HAVE_FACE, "The ElementTransformation "
+   MFEM_VERIFY(mask & 16, "The ElementTransformation "
               "for the face has not been configured.");
   IsoparametricTransformation::Transform(ir, tr);
 }
@@ -625,7 +619,7 @@ void FaceElementTransformations::Transform(const IntegrationRule &ir,
 void FaceElementTransformations::Transform(const DenseMatrix &matrix,
                                           DenseMatrix &result)
 {
-   MFEM_VERIFY(mask & HAVE_FACE, "The ElementTransformation "
+   MFEM_VERIFY(mask & 16, "The ElementTransformation "
               "for the face has not been configured.");
   IsoparametricTransformation::Transform(matrix, result);
 }
@@ -439,57 +439,15 @@ public:
   void Transform (const IntegrationRule  &, IntegrationRule  &);
 };

-/** @brief A specialized ElementTransformation class representing a face and
-    its two neighboring elements.

-    This class can be used as a container for the element transformation data
-    needed for integrating discontinuous fields on element interfaces in a
-    Discontinuous Galerkin (DG) context.
-
-    The secondary purpose of this class is to enable the
-    GridFunction::GetValue function, and various related functions, to properly
-    evaluate fields with limited continuity on boundary elements.
-*/
 class FaceElementTransformations : public IsoparametricTransformation
 {
 private:
-
-   // Bitwise OR of ConfigMasks
   int mask;

   IntegrationPoint eip1, eip2;

-protected: // interface for Mesh to be able to configure this object.
-
-   friend class Mesh;
-#ifdef MFEM_USE_MPI
-   friend class ParMesh;
-#endif
-
-   /// Set the mask indicating which portions of the object have been setup
-   /** The argument @a m is a bitmask used in
-       Mesh::GetFaceElementTransformations to indicate which portions of the
-       FaceElementTransformations object have been configured.
-
-       mask &  1: Elem1 is configured
-       mask &  2: Elem2 is configured
-       mask &  4: Loc1 is configured
-       mask &  8: Loc2 is configured
-       mask & 16: The Face transformation itself is configured
-   */
-   void SetConfigurationMask(int m) { mask = m; }
-
 public:
-
-   enum ConfigMasks
-   {
-      HAVE_ELEM1 =  1, ///< Element on side 1 is configured
-      HAVE_ELEM2 =  2, ///< Element on side 2 is configured
-      HAVE_LOC1  =  4, ///< Point transformation for side 1 is configured
-      HAVE_LOC2  =  8, ///< Point transformation for side 2 is configured
-      HAVE_FACE  = 16  ///< Face transformation is configured
-   };
-
   int Elem1No, Elem2No;
   Geometry::Type &FaceGeom; ///< @deprecated Use GetGeometryType instead
   ElementTransformation *Elem1, *Elem2;
@@ -508,10 +466,10 @@ public:
   */
   void SetGeometryType(Geometry::Type g) { geom = g; }

-   /** @brief Return the mask defining the configuration state.
-
-       The mask value indicates which portions of FaceElementTransformations
-       object have been configured.
+   /// Set the mask indicating which portions of the object have been setup
+   /** The argument @a m is a bitmask used in
+       Mesh::GetFaceElementTransformations to indicate which portions of the
+       FaceElement Transformations object have been configured.

       mask &  1: Elem1 is configured
       mask &  2: Elem2 is configured
@@ -519,45 +477,12 @@ public:
       mask &  8: Loc2 is configured
       mask & 16: The Face transformation itself is configured
   */
-   int GetConfigurationMask() const { return mask; }
+   void SetConfigurationMask(int m) { mask = m; }
+   int  GetConfigurationMask() const { return mask; }

   /** @brief Set the integration point in the Face and the two neighboring
-       elements, if present.
-
-       The point @a face_ip must be in the reference coordinate system of the
-       face.
-   */
-   void SetIntPoint(const IntegrationPoint *face_ip);
-
-   /** @brief Set the integration point in the Face and the two neighboring
-       elements, if present.
-
-       This is a more expressive member function name than SetIntPoint, which
-       in this special case, does the same thing. This function can be used for
-       greater code clarity.
-   */
-   inline void SetAllIntPoints(const IntegrationPoint *face_ip)
-   { FaceElementTransformations::SetIntPoint(face_ip); }
-
-   /** @brief Get a const reference to the integration point in neighboring
-       element 1 corresponding to the currently set integration point on the
-       face.
-
-       This IntegrationPoint object will only contain up-to-date data if
-       SetIntPoint or SetAllIntPoints has been called with the latest
-       integration point for the face and the appropriate point transformation
-       has been configured. */
-   const IntegrationPoint &GetElement1IntPoint() { return eip1; }
-
-   /** @brief Get a const reference to the integration point in neighboring
-       element 2 corresponding to the currently set integration point on the
-       face.
-
-       This IntegrationPoint object will only contain up-to-date data if
-       SetIntPoint or SetAllIntPoints has been called with the latest
-       integration point for the face and the appropriate point transformation
-       has been configured. */
-   const IntegrationPoint &GetElement2IntPoint() { return eip2; }
+       elements, if present. */
+   void SetIntPoint(const IntegrationPoint *ip);

   virtual void Transform(const IntegrationPoint &, Vector &);
   virtual void Transform(const IntegrationRule &, DenseMatrix &);
@@ -37,8 +37,7 @@ public:
      ClosedUniform   = 4,  ///< Nodes: x_i = i/(n-1),     i=0,...,n-1
      OpenHalfUniform = 5,  ///< Nodes: x_i = (i+1/2)/n,   i=0,...,n-1
      Serendipity     = 6,  ///< Serendipity basis (squares / cubes)
-      ClosedGL        = 7,  ///< Closed GaussLegendre
-      NumBasisTypes   = 8   /**< Keep track of maximum types to prevent
+      NumBasisTypes   = 7   /**< Keep track of maximum types to prevent
                                 hard-coding */
   };
   /** @brief If the input does not represents a valid BasisType, abort with an
@@ -70,7 +69,6 @@ public:
         case ClosedUniform:   return Quadrature1D::ClosedUniform;
         case OpenHalfUniform: return Quadrature1D::OpenHalfUniform;
         case Serendipity:     return Quadrature1D::GaussLobatto;
-         case ClosedGL:        return Quadrature1D::ClosedGL;
      }
      return Quadrature1D::Invalid;
   }
@@ -84,7 +82,6 @@ public:
         case Quadrature1D::OpenUniform:     return OpenUniform;
         case Quadrature1D::ClosedUniform:   return ClosedUniform;
         case Quadrature1D::OpenHalfUniform: return OpenHalfUniform;
-         case Quadrature1D::ClosedGL:        return ClosedGL;
      }
      return Invalid;
   }
@@ -944,7 +944,7 @@ const Operator *FiniteElementSpace::GetFaceRestriction(
 }

 const QuadratureInterpolator *FiniteElementSpace::GetQuadratureInterpolator(
-   const IntegrationRule &ir, const DofToQuad::Mode mode) const
+   const IntegrationRule &ir) const
 {
   for (int i = 0; i < E2Q_array.Size(); i++)
   {
@@ -952,13 +952,13 @@ const QuadratureInterpolator *FiniteElementSpace::GetQuadratureInterpolator(
      if (qi->IntRule == &ir) { return qi; }
   }

-   QuadratureInterpolator *qi = new QuadratureInterpolator(*this, ir, mode);
+   QuadratureInterpolator *qi = new QuadratureInterpolator(*this, ir);
   E2Q_array.Append(qi);
   return qi;
 }

 const QuadratureInterpolator *FiniteElementSpace::GetQuadratureInterpolator(
-   const QuadratureSpace &qs, const DofToQuad::Mode mode) const
+   const QuadratureSpace &qs) const
 {
   for (int i = 0; i < E2Q_array.Size(); i++)
   {
@@ -966,7 +966,7 @@ const QuadratureInterpolator *FiniteElementSpace::GetQuadratureInterpolator(
      if (qi->qspace == &qs) { return qi; }
   }

-   QuadratureInterpolator *qi = new QuadratureInterpolator(*this, qs, mode);
+   QuadratureInterpolator *qi = new QuadratureInterpolator(*this, qs);
   E2Q_array.Append(qi);
   return qi;
 }
@@ -983,8 +983,8 @@ const FaceQuadratureInterpolator
         if (qi->IntRule == &ir) { return qi; }
      }

-      FaceQuadratureInterpolator *qi =
-         new FaceQuadratureInterpolator(*this, ir, type);
+      FaceQuadratureInterpolator *qi = new FaceQuadratureInterpolator(*this, ir,
+                                                                      type);
      E2IFQ_array.Append(qi);
      return qi;
   }
@@ -367,7 +367,7 @@ public:
       All elements will use the same IntegrationRule, @a ir as the target
       quadrature points. */
   const QuadratureInterpolator *GetQuadratureInterpolator(
-      const IntegrationRule &ir, const DofToQuad::Mode = DofToQuad::FULL) const;
+      const IntegrationRule &ir) const;

   /** @brief Return a QuadratureInterpolator that interpolates E-vectors to
       quadrature point values and/or derivatives (Q-vectors). */
@@ -378,7 +378,7 @@ public:
       The target quadrature points in the elements are described by the given
       QuadratureSpace, @a qs. */
   const QuadratureInterpolator *GetQuadratureInterpolator(
-      const QuadratureSpace &qs, const DofToQuad::Mode = DofToQuad::FULL) const;
+      const QuadratureSpace &qs) const;

   /** @brief Return a FaceQuadratureInterpolator that interpolates E-vectors to
       quadrature point values and/or derivatives (Q-vectors). */
@@ -756,12 +756,6 @@ public:
   /// Return the total number of quadrature points.
   int GetSize() const { return size; }

-   /// Returns the mesh
-   inline Mesh *GetMesh() const { return mesh; }
-
-   /// Returns number of elements in the mesh.
-   inline int GetNE() const { return mesh->GetNE(); }
-
   /// Get the IntegrationRule associated with mesh element @a idx.
   const IntegrationRule &GetElementIntRule(int idx) const
   { return *int_rule[mesh->GetElementBaseGeometry(idx)]; }
@@ -1344,14 +1344,15 @@ const IntegrationRule *GeometryRefiner::RefineInterior(Geometry::Type Geom,
            return NULL;
         }
         ir = FindInIntPts(Geom, Times-1);
-         if (ir) { return ir; }
-
-         ir = new IntegrationRule(Times-1);
-         for (int i = 1; i < Times; i++)
+         if (ir == NULL)
         {
-            IntegrationPoint &ip = ir->IntPoint(i-1);
-            ip.x = double(i) / Times;
-            ip.y = ip.z = 0.0;
+            ir = new IntegrationRule(Times-1);
+            for (int i = 1; i < Times; i++)
+            {
+               IntegrationPoint &ip = ir->IntPoint(i-1);
+               ip.x = double(i) / Times;
+               ip.y = ip.z = 0.0;
+            }
         }
      }
      break;
@@ -1363,17 +1364,18 @@ const IntegrationRule *GeometryRefiner::RefineInterior(Geometry::Type Geom,
            return NULL;
         }
         ir = FindInIntPts(Geom, ((Times-1)*(Times-2))/2);
-         if (ir) { return ir; }
-
-         ir = new IntegrationRule(((Times-1)*(Times-2))/2);
-         for (int k = 0, j = 1; j < Times-1; j++)
-            for (int i = 1; i < Times-j; i++, k++)
-            {
-               IntegrationPoint &ip = ir->IntPoint(k);
-               ip.x = double(i) / Times;
-               ip.y = double(j) / Times;
-               ip.z = 0.0;
-            }
+         if (ir == NULL)
+         {
+            ir = new IntegrationRule(((Times-1)*(Times-2))/2);
+            for (int k = 0, j = 1; j < Times-1; j++)
+               for (int i = 1; i < Times-j; i++, k++)
+               {
+                  IntegrationPoint &ip = ir->IntPoint(k);
+                  ip.x = double(i) / Times;
+                  ip.y = double(j) / Times;
+                  ip.z = 0.0;
+               }
+         }
      }
      break;

@@ -1384,17 +1386,18 @@ const IntegrationRule *GeometryRefiner::RefineInterior(Geometry::Type Geom,
            return NULL;
         }
         ir = FindInIntPts(Geom, (Times-1)*(Times-1));
-         if (ir) { return ir; }
-
-         ir = new IntegrationRule((Times-1)*(Times-1));
-         for (int k = 0, j = 1; j < Times; j++)
-            for (int i = 1; i < Times; i++, k++)
-            {
-               IntegrationPoint &ip = ir->IntPoint(k);
-               ip.x = double(i) / Times;
-               ip.y = double(j) / Times;
-               ip.z = 0.0;
-            }
+         if (ir == NULL)
+         {
+            ir = new IntegrationRule((Times-1)*(Times-1));
+            for (int k = 0, j = 1; j < Times; j++)
+               for (int i = 1; i < Times; i++, k++)
+               {
+                  IntegrationPoint &ip = ir->IntPoint(k);
+                  ip.x = double(i) / Times;
+                  ip.y = double(j) / Times;
+                  ip.z = 0.0;
+               }
+         }
      }
      break;

@@ -1402,121 +1405,10 @@ const IntegrationRule *GeometryRefiner::RefineInterior(Geometry::Type Geom,
         mfem_error("GeometryRefiner::RefineInterior(...)");
   }

-   MFEM_ASSERT(ir != NULL, "Failed to construct the refined IntegrationRule.");
-   IntPts[Geom].Append(ir);
-
+   if (ir) { IntPts[Geom].Append(ir); }
   return ir;
 }

-
-int GeometryRefiner::GetRefinementLevelFromPoints(Geometry::Type geom, int Npts)
-{
-   switch (geom)
-   {
-      case Geometry::POINT:
-      {
-         return -1;
-      }
-      case Geometry::SEGMENT:
-      {
-         return Npts -1;
-      }
-      case Geometry::TRIANGLE:
-      {
-         for (int n = 0, np = 0; (n < 15) && (np < Npts) ; n++)
-         {
-            np = (n+1)*(n+2)/2;
-            if (np == Npts) { return n; }
-         }
-         return -1;
-      }
-      case Geometry::SQUARE:
-      {
-         for (int n = 0, np = 0; (n < 15) && (np < Npts) ; n++)
-         {
-            np = (n+1)*(n+1);
-            if (np == Npts) { return n; }
-         }
-         return -1;
-      }
-      case Geometry::CUBE:
-      {
-         for (int n = 0, np = 0; (n < 15) && (np < Npts) ; n++)
-         {
-            np = (n+1)*(n+1)*(n+1);
-            if (np == Npts) { return n; }
-         }
-         return -1;
-      }
-      case Geometry::TETRAHEDRON:
-      {
-         for (int n = 0, np = 0; (n < 15) && (np < Npts) ; n++)
-         {
-            np = (n+3)*(n+2)*(n+1)/6;
-            if (np == Npts) { return n; }
-         }
-         return -1;
-      }
-      case Geometry::PRISM:
-      {
-         for (int n = 0, np = 0; (n < 15) && (np < Npts) ; n++)
-         {
-            np = (n+1)*(n+1)*(n+2)/2;
-            if (np == Npts) { return n; }
-         }
-         return -1;
-      }
-      default:
-      {
-         mfem_error("Non existing Geometry.");
-      }
-   }
-
-   return -1;
-}
-
-
-int GeometryRefiner::GetRefinementLevelFromElems(Geometry::Type geom, int Nels)
-{
-   switch (geom)
-   {
-      case Geometry::POINT:
-      {
-         return -1;
-      }
-      case Geometry::SEGMENT:
-      {
-         return Nels;
-      }
-      case Geometry::TRIANGLE:
-      case Geometry::SQUARE:
-      {
-         for (int n = 0; (n < 15) && (n*n < Nels+1) ; n++)
-         {
-            if (n*n == Nels) { return n-1; }
-         }
-         return -1;
-      }
-      case Geometry::CUBE:
-      case Geometry::TETRAHEDRON:
-      case Geometry::PRISM:
-      {
-         for (int n = 0; (n < 15) && (n*n*n < Nels+1) ; n++)
-         {
-            if (n*n*n == Nels) { return n-1; }
-         }
-         return -1;
-      }
-      default:
-      {
-         mfem_error("Non existing Geometry.");
-      }
-   }
-
-   return -1;
-}
-
-
 GeometryRefiner GlobGeometryRefiner;

 }
@@ -273,12 +273,6 @@ public:
   /// @note This method always uses Quadrature1D::OpenUniform points.
   const IntegrationRule *RefineInterior(Geometry::Type Geom, int Times);

-   /// Get the Refinement level based on number of points
-   virtual int GetRefinementLevelFromPoints(Geometry::Type Geom, int Npts);
-
-   /// Get the Refinement level based on number of elements
-   virtual int GetRefinementLevelFromElems(Geometry::Type geom, int Npts);
-
   ~GeometryRefiner();
 };

@@ -397,16 +397,8 @@ const
   fes->DofsToVDofs(vdim-1, dofs);
   Vector DofVal(dofs.Size()), LocVec;
   const FiniteElement *fe = fes->GetFE(i);
-   if (fe->GetMapType() == FiniteElement::VALUE)
-   {
-      fe->CalcShape(ip, DofVal);
-   }
-   else
-   {
-      ElementTransformation *Tr = fes->GetElementTransformation(i);
-      Tr->SetIntPoint(&ip);
-      fe->CalcPhysShape(*Tr, DofVal);
-   }
+   MFEM_ASSERT(fe->GetMapType() == FiniteElement::VALUE, "invalid FE map type");
+   fe->CalcShape(ip, DofVal);
   GetSubVector(dofs, LocVec);

   return (DofVal * LocVec);
@@ -423,17 +415,10 @@ void GridFunction::GetVectorValue(int i, const IntegrationPoint &ip,
   GetSubVector(vdofs, loc_data);
   if (FElem->GetRangeType() == FiniteElement::SCALAR)
   {
+      MFEM_ASSERT(FElem->GetMapType() == FiniteElement::VALUE,
+                  "invalid FE map type");
      Vector shape(dof);
-      if (FElem->GetMapType() == FiniteElement::VALUE)
-      {
-         FElem->CalcShape(ip, shape);
-      }
-      else
-      {
-         ElementTransformation *Tr = fes->GetElementTransformation(i);
-         Tr->SetIntPoint(&ip);
-         FElem->CalcPhysShape(*Tr, shape);
-      }
+      FElem->CalcShape(ip, shape);
      int vdim = fes->GetVDim();
      val.SetSize(vdim);
      for (int k = 0; k < vdim; k++)
@@ -767,21 +752,19 @@ double GridFunction::GetValue(ElementTransformation &T,
            IntegrationPoint fip;
            be_to_bfe(FET->GetGeometryType(), o, ip, fip);

-            // Compute and set the point in element 1 from fip
-            FET->SetAllIntPoints(&fip);
+            FET->SetIntPoint(&fip);
            ElementTransformation & T1 = FET->GetElement1Transformation();
            return GetValue(T1, T1.GetIntPoint(), comp);
         }
+         break;
      }
-      break;
      case ElementTransformation::BDR_FACE:
      {
         FaceElementTransformations * FET =
            dynamic_cast<FaceElementTransformations *>(&T);

         // Evaluate in neighboring element for both continuous and
-         // discontinuous fields (the integration point in T1 should have
-         // already been set).
+         // discontinuous fields.
         ElementTransformation & T1 = FET->GetElement1Transformation();
         return GetValue(T1, T1.GetIntPoint(), comp);
      }
@@ -905,21 +888,19 @@ void GridFunction::GetVectorValue(ElementTransformation &T,
            IntegrationPoint fip;
            be_to_bfe(FET->GetGeometryType(), o, ip, fip);

-            // Compute and set the point in element 1 from fip
-            FET->SetAllIntPoints(&fip);
+            FET->SetIntPoint(&fip);
            ElementTransformation & T1 = FET->GetElement1Transformation();
            return GetVectorValue(T1, T1.GetIntPoint(), val);
         }
+         break;
      }
-      break;
      case ElementTransformation::BDR_FACE:
      {
         FaceElementTransformations * FET =
            dynamic_cast<FaceElementTransformations *>(&T);

         // Evaluate in neighboring element for both continuous and
-         // discontinuous fields (the integration point in T1 should have
-         // already been set).
+         // discontinuous fields.
         ElementTransformation & T1 = FET->GetElement1Transformation();
         return GetVectorValue(T1, T1.GetIntPoint(), val);
      }
@@ -1357,262 +1338,107 @@ void GridFunction::GetVectorGradientHat(
   MultAtB(loc_data_mat, dshape, gh);
 }

-double GridFunction::GetDivergence(ElementTransformation &T) const
+double GridFunction::GetDivergence(ElementTransformation &tr) const
 {
-   switch (T.ElementType)
+   double div_v;
+   int elNo = tr.ElementNo;
+   const FiniteElement *FElem = fes->GetFE(elNo);
+   if (FElem->GetRangeType() == FiniteElement::SCALAR)
   {
-      case ElementTransformation::ELEMENT:
+      MFEM_ASSERT(FElem->GetMapType() == FiniteElement::VALUE,
+                  "invalid FE map type");
+      DenseMatrix grad_hat;
+      GetVectorGradientHat(tr, grad_hat);
+      const DenseMatrix &Jinv = tr.InverseJacobian();
+      div_v = 0.0;
+      for (int i = 0; i < Jinv.Width(); i++)
      {
-         int elNo = T.ElementNo;
-         const FiniteElement *fe = fes->GetFE(elNo);
-         if (fe->GetRangeType() == FiniteElement::SCALAR)
+         for (int j = 0; j < Jinv.Height(); j++)
         {
-            MFEM_ASSERT(fe->GetMapType() == FiniteElement::VALUE,
-                        "invalid FE map type");
-            DenseMatrix grad_hat;
-            GetVectorGradientHat(T, grad_hat);
-            const DenseMatrix &Jinv = T.InverseJacobian();
-            double div_v = 0.0;
-            for (int i = 0; i < Jinv.Width(); i++)
-            {
-               for (int j = 0; j < Jinv.Height(); j++)
-               {
-                  div_v += grad_hat(i, j) * Jinv(j, i);
-               }
-            }
-            return div_v;
+            div_v += grad_hat(i, j) * Jinv(j, i);
         }
-         else
-         {
-            // Assuming RT-type space
-            Array<int> dofs;
-            fes->GetElementDofs(elNo, dofs);
-            Vector loc_data, divshape(fe->GetDof());
-            GetSubVector(dofs, loc_data);
-            fe->CalcDivShape(T.GetIntPoint(), divshape);
-            return (loc_data * divshape) / T.Weight();
-         }
-      }
-      break;
-      case ElementTransformation::BDR_ELEMENT:
-      {
-         // In order to properly capture the derivative of the normal component
-         // of the field (as well as the transverse divergence of the
-         // tangential compoents) we must evaluate it in the neighboring
-         // element.
-         FaceElementTransformations * FET =
-            fes->GetMesh()->GetBdrFaceTransformations(T.ElementNo);
-
-         // Boundary elements and Boundary Faces may have different
-         // orientations so adjust the integration point if necessary.
-         int o = 0;
-         if (fes->GetMesh()->Dimension() == 3)
-         {
-            int f;
-            fes->GetMesh()->GetBdrElementFace(T.ElementNo, &f, &o);
-         }
-
-         IntegrationPoint fip;
-         be_to_bfe(FET->GetGeometryType(), o, T.GetIntPoint(), fip);
-
-         // Compute and set the point in element 1 from fip
-         FET->SetAllIntPoints(&fip);
-         ElementTransformation & T1 = FET->GetElement1Transformation();
-
-         return GetDivergence(T1);
-      }
-      break;
-      case ElementTransformation::BDR_FACE:
-      {
-         // This must be a DG context so this dynamic cast must succeed.
-         FaceElementTransformations * FET =
-            dynamic_cast<FaceElementTransformations *>(&T);
-
-         // Evaluate in neighboring element (the integration point in T1 should
-         // have already been set).
-         ElementTransformation & T1 = FET->GetElement1Transformation();
-         return GetDivergence(T1);
-      }
-      break;
-      default:
-      {
-         MFEM_ABORT("GridFunction::GetDivergence: Unsupported element type \""
-                    << T.ElementType << "\"");
      }
   }
-   return 0.0; // never reached
+   else
+   {
+      // Assuming RT-type space
+      Array<int> dofs;
+      fes->GetElementDofs(elNo, dofs);
+      Vector loc_data, divshape(FElem->GetDof());
+      GetSubVector(dofs, loc_data);
+      FElem->CalcDivShape(tr.GetIntPoint(), divshape);
+      div_v = (loc_data * divshape) / tr.Weight();
+   }
+   return div_v;
 }

-void GridFunction::GetCurl(ElementTransformation &T, Vector &curl) const
+void GridFunction::GetCurl(ElementTransformation &tr, Vector &curl) const
 {
-   switch (T.ElementType)
+   int elNo = tr.ElementNo;
+   const FiniteElement *FElem = fes->GetFE(elNo);
+   if (FElem->GetRangeType() == FiniteElement::SCALAR)
   {
-      case ElementTransformation::ELEMENT:
+      MFEM_ASSERT(FElem->GetMapType() == FiniteElement::VALUE,
+                  "invalid FE map type");
+      DenseMatrix grad_hat;
+      GetVectorGradientHat(tr, grad_hat);
+      const DenseMatrix &Jinv = tr.InverseJacobian();
+      DenseMatrix grad(grad_hat.Height(), Jinv.Width()); // vdim x FElem->Dim
+      Mult(grad_hat, Jinv, grad);
+      MFEM_ASSERT(grad.Height() == grad.Width(), "");
+      if (grad.Height() == 3)
      {
-         int elNo = T.ElementNo;
-         const FiniteElement *fe = fes->GetFE(elNo);
-         if (fe->GetRangeType() == FiniteElement::SCALAR)
-         {
-            MFEM_ASSERT(fe->GetMapType() == FiniteElement::VALUE,
-                        "invalid FE map type");
-            DenseMatrix grad_hat;
-            GetVectorGradientHat(T, grad_hat);
-            const DenseMatrix &Jinv = T.InverseJacobian();
-            // Dimensions of grad are vdim x FElem->Dim
-            DenseMatrix grad(grad_hat.Height(), Jinv.Width());
-            Mult(grad_hat, Jinv, grad);
-            MFEM_ASSERT(grad.Height() == grad.Width(), "");
-            if (grad.Height() == 3)
-            {
-               curl.SetSize(3);
-               curl(0) = grad(2,1) - grad(1,2);
-               curl(1) = grad(0,2) - grad(2,0);
-               curl(2) = grad(1,0) - grad(0,1);
-            }
-            else if (grad.Height() == 2)
-            {
-               curl.SetSize(1);
-               curl(0) = grad(1,0) - grad(0,1);
-            }
-         }
-         else
-         {
-            // Assuming ND-type space
-            Array<int> dofs;
-            fes->GetElementDofs(elNo, dofs);
-            Vector loc_data;
-            GetSubVector(dofs, loc_data);
-            DenseMatrix curl_shape(fe->GetDof(), fe->GetDim() == 3 ? 3 : 1);
-            fe->CalcCurlShape(T.GetIntPoint(), curl_shape);
-            curl.SetSize(curl_shape.Width());
-            if (curl_shape.Width() == 3)
-            {
-               double curl_hat[3];
-               curl_shape.MultTranspose(loc_data, curl_hat);
-               T.Jacobian().Mult(curl_hat, curl);
-            }
-            else
-            {
-               curl_shape.MultTranspose(loc_data, curl);
-            }
-            curl /= T.Weight();
-         }
+         curl.SetSize(3);
+         curl(0) = grad(2,1) - grad(1,2);
+         curl(1) = grad(0,2) - grad(2,0);
+         curl(2) = grad(1,0) - grad(0,1);
      }
-      break;
-      case ElementTransformation::BDR_ELEMENT:
+      else if (grad.Height() == 2)
      {
-         // In order to capture the tangential components of the curl we
-         // must evaluate it in the neighboring element.
-         FaceElementTransformations * FET =
-            fes->GetMesh()->GetBdrFaceTransformations(T.ElementNo);
-
-         // Boundary elements and Boundary Faces may have different
-         // orientations so adjust the integration point if necessary.
-         int o = 0;
-         if (fes->GetMesh()->Dimension() == 3)
-         {
-            int f;
-            fes->GetMesh()->GetBdrElementFace(T.ElementNo, &f, &o);
-         }
-
-         IntegrationPoint fip;
-         be_to_bfe(FET->GetGeometryType(), o, T.GetIntPoint(), fip);
-
-         // Compute and set the point in element 1 from fip
-         FET->SetAllIntPoints(&fip);
-         ElementTransformation & T1 = FET->GetElement1Transformation();
-
-         GetCurl(T1, curl);
+         curl.SetSize(1);
+         curl(0) = grad(1,0) - grad(0,1);
      }
-      break;
-      case ElementTransformation::BDR_FACE:
+   }
+   else
+   {
+      // Assuming ND-type space
+      Array<int> dofs;
+      fes->GetElementDofs(elNo, dofs);
+      Vector loc_data;
+      GetSubVector(dofs, loc_data);
+      DenseMatrix curl_shape(FElem->GetDof(), FElem->GetDim() == 3 ? 3 : 1);
+      FElem->CalcCurlShape(tr.GetIntPoint(), curl_shape);
+      curl.SetSize(curl_shape.Width());
+      if (curl_shape.Width() == 3)
      {
-         // This must be a DG context so this dynamic cast must succeed.
-         FaceElementTransformations * FET =
-            dynamic_cast<FaceElementTransformations *>(&T);
-
-         // Evaluate in neighboring element (the integration point in T1 should
-         // have already been set).
-         ElementTransformation & T1 = FET->GetElement1Transformation();
-         GetCurl(T1, curl);
+         double curl_hat[3];
+         curl_shape.MultTranspose(loc_data, curl_hat);
+         tr.Jacobian().Mult(curl_hat, curl);
      }
-      break;
-      default:
+      else
      {
-         MFEM_ABORT("GridFunction::GetCurl: Unsupported element type \""
-                    << T.ElementType << "\"");
+         curl_shape.MultTranspose(loc_data, curl);
      }
+      curl /= tr.Weight();
   }
 }

-void GridFunction::GetGradient(ElementTransformation &T, Vector &grad) const
+void GridFunction::GetGradient(ElementTransformation &tr, Vector &grad) const
 {
-   switch (T.ElementType)
-   {
-      case ElementTransformation::ELEMENT:
-      {
-         const FiniteElement * fe = fes->GetFE(T.ElementNo);
-         MFEM_ASSERT(fe->GetMapType() == FiniteElement::VALUE,
-                     "invalid FE map type");
-         int spaceDim = fes->GetMesh()->SpaceDimension();
-         int dim = fe->GetDim(), dof = fe->GetDof();
-         DenseMatrix dshape(dof, dim);
-         Vector lval, gh(dim);
-         Array<int> dofs;
+   int elNo = tr.ElementNo;
+   const FiniteElement *fe = fes->GetFE(elNo);
+   MFEM_ASSERT(fe->GetMapType() == FiniteElement::VALUE, "invalid FE map type");
+   int dim = fe->GetDim(), dof = fe->GetDof();
+   DenseMatrix dshape(dof, dim);
+   Vector lval, gh(dim);
+   Array<int> dofs;

-         grad.SetSize(spaceDim);
-         fes->GetElementDofs(T.ElementNo, dofs);
-         GetSubVector(dofs, lval);
-         fe->CalcDShape(T.GetIntPoint(), dshape);
-         dshape.MultTranspose(lval, gh);
-         T.InverseJacobian().MultTranspose(gh, grad);
-      }
-      break;
-      case ElementTransformation::BDR_ELEMENT:
-      {
-         // In order to properly capture the normal component of the gradient
-         // as well as its tangential components we must evaluate it in the
-         // neighboring element.
-         FaceElementTransformations * FET =
-            fes->GetMesh()->GetBdrFaceTransformations(T.ElementNo);
-
-         // Boundary elements and Boundary Faces may have different
-         // orientations so adjust the integration point if necessary.
-         int o = 0;
-         if (fes->GetMesh()->Dimension() == 3)
-         {
-            int f;
-            fes->GetMesh()->GetBdrElementFace(T.ElementNo, &f, &o);
-         }
-
-         IntegrationPoint fip;
-         be_to_bfe(FET->GetGeometryType(), o, T.GetIntPoint(), fip);
-
-         // Compute and set the point in element 1 from fip
-         FET->SetAllIntPoints(&fip);
-         ElementTransformation & T1 = FET->GetElement1Transformation();
-
-         GetGradient(T1, grad);
-      }
-      break;
-      case ElementTransformation::BDR_FACE:
-      {
-         // This must be a DG context so this dynamic cast must succeed.
-         FaceElementTransformations * FET =
-            dynamic_cast<FaceElementTransformations *>(&T);
-
-         // Evaluate in neighboring element (the integration point in T1 should
-         // have already been set).
-         ElementTransformation & T1 = FET->GetElement1Transformation();
-         GetGradient(T1, grad);
-      }
-      break;
-      default:
-      {
-         MFEM_ABORT("GridFunction::GetGradient: Unsupported element type \""
-                    << T.ElementType << "\"");
-      }
-   }
+   grad.SetSize(dim);
+   fes->GetElementDofs(elNo, dofs);
+   GetSubVector(dofs, lval);
+   fe->CalcDShape(tr.GetIntPoint(), dshape);
+   dshape.MultTranspose(lval, gh);
+   tr.InverseJacobian().MultTranspose(gh, grad);
 }

 void GridFunction::GetGradients(ElementTransformation &tr,
@@ -1641,65 +1467,15 @@ void GridFunction::GetGradients(ElementTransformation &tr,
 }

 void GridFunction::GetVectorGradient(
-   ElementTransformation &T, DenseMatrix &grad) const
+   ElementTransformation &tr, DenseMatrix &grad) const
 {
-   switch (T.ElementType)
-   {
-      case ElementTransformation::ELEMENT:
-      {
-         MFEM_ASSERT(fes->GetFE(T.ElementNo)->GetMapType() ==
-                     FiniteElement::VALUE, "invalid FE map type");
-         DenseMatrix grad_hat;
-         GetVectorGradientHat(T, grad_hat);
-         const DenseMatrix &Jinv = T.InverseJacobian();
-         grad.SetSize(grad_hat.Height(), Jinv.Width());
-         Mult(grad_hat, Jinv, grad);
-      }
-      break;
-      case ElementTransformation::BDR_ELEMENT:
-      {
-         // In order to capture the normal component of the gradient we
-         // must evaluate it in the neighboring element.
-         FaceElementTransformations * FET =
-            fes->GetMesh()->GetBdrFaceTransformations(T.ElementNo);
-
-         // Boundary elements and Boundary Faces may have different
-         // orientations so adjust the integration point if necessary.
-         int o = 0;
-         if (fes->GetMesh()->Dimension() == 3)
-         {
-            int f;
-            fes->GetMesh()->GetBdrElementFace(T.ElementNo, &f, &o);
-         }
-
-         IntegrationPoint fip;
-         be_to_bfe(FET->GetGeometryType(), o, T.GetIntPoint(), fip);
-
-         // Compute and set the point in element 1 from fip
-         FET->SetAllIntPoints(&fip);
-         ElementTransformation & T1 = FET->GetElement1Transformation();
-
-         GetVectorGradient(T1, grad);
-      }
-      break;
-      case ElementTransformation::BDR_FACE:
-      {
-         // This must be a DG context so this dynamic cast must succeed.
-         FaceElementTransformations * FET =
-            dynamic_cast<FaceElementTransformations *>(&T);
-
-         // Evaluate in neighboring element (the integration point in T1 should
-         // have already been set).
-         ElementTransformation & T1 = FET->GetElement1Transformation();
-         GetVectorGradient(T1, grad);
-      }
-      break;
-      default:
-      {
-         MFEM_ABORT("GridFunction::GetVectorGradient: "
-                    "Unsupported element type \"" << T.ElementType << "\"");
-      }
-   }
+   MFEM_ASSERT(fes->GetFE(tr.ElementNo)->GetMapType() == FiniteElement::VALUE,
+               "invalid FE map type");
+   DenseMatrix grad_hat;
+   GetVectorGradientHat(tr, grad_hat);
+   const DenseMatrix &Jinv = tr.InverseJacobian();
+   grad.SetSize(grad_hat.Height(), Jinv.Width());
+   Mult(grad_hat, Jinv, grad);
 }

 void GridFunction::GetElementAverages(GridFunction &avgs) const
@@ -105,7 +105,7 @@ public:
       have the same size.

       @note Defining this method overwrites the implicitly defined copy
-       assignment operator. */
+       assignemnt operator. */
   GridFunction &operator=(const GridFunction &rhs)
   { return operator=((const Vector &)rhs); }

@@ -162,8 +162,7 @@ public:
                           int vdim = 1) const;

   /** Return a vector value from within the given element. */
-   virtual void GetVectorValue(int i, const IntegrationPoint &ip,
-                               Vector &val) const;
+   void GetVectorValue(int i, const IntegrationPoint &ip, Vector &val) const;
   ///@}

   /** @name Element Index Get Values Methods
@@ -209,14 +208,13 @@ public:
   ///@{
   /** Return a scalar value from within the element indicated by the
       ElementTransformation Object. */
-   virtual double GetValue(ElementTransformation &T, const IntegrationPoint &ip,
-                           int comp = 0, Vector *tr = NULL) const;
+   double GetValue(ElementTransformation &T, const IntegrationPoint &ip,
+                   int comp = 0, Vector *tr = NULL) const;

   /** Return a vector value from within the element indicated by the
       ElementTransformation Object. */
-   virtual void GetVectorValue(ElementTransformation &T,
-                               const IntegrationPoint &ip,
-                               Vector &val, Vector *tr = NULL) const;
+   void GetVectorValue(ElementTransformation &T, const IntegrationPoint &ip,
+                       Vector &val, Vector *tr = NULL) const;
   ///@}

   /** @name ElementTransformation Get Values Methods
@@ -714,7 +712,7 @@ public:
       the same size.

       @note Defining this method overwrites the implicitly defined copy
-       assignment operator. */
+       assignemnt operator. */
   QuadratureFunction &operator=(const QuadratureFunction &v);

   /// Get the IntegrationRule associated with mesh element @a idx.
@@ -192,7 +192,6 @@ void FindPointsGSLIB::Interpolate(Array<unsigned int> &codes,
   const int ncomp      = field_in.FESpace()->GetVDim(),
             points_fld = field_in.Size() / ncomp,
             points_cnt = codes.Size();
-   field_out.SetSize(points_cnt*ncomp);

   for (int i = 0; i < ncomp; i++)
   {
@@ -618,26 +618,6 @@ void QuadratureFunctions1D::OpenHalfUniform(const int np, IntegrationRule* ir)
   CalculateUniformWeights(ir, Quadrature1D::OpenHalfUniform);
 }

-void QuadratureFunctions1D::ClosedGL(const int np, IntegrationRule* ir)
-{
-   ir->SetSize(np);
-   ir->IntPoint(0).x = 0.0;
-   ir->IntPoint(np-1).x = 1.0;
-
-   if ( np > 2 )
-   {
-      IntegrationRule gl_ir;
-      GaussLegendre(np-1, &gl_ir);
-
-      for (int i = 1; i < np-1; ++i)
-      {
-         ir->IntPoint(i).x = (gl_ir.IntPoint(i-1).x + gl_ir.IntPoint(i).x)/2;
-      }
-   }
-
-   CalculateUniformWeights(ir, Quadrature1D::ClosedGL);
-}
-
 void QuadratureFunctions1D::GivePolyPoints(const int np, double *pts,
                                           const int type)
 {
@@ -670,11 +650,6 @@ void QuadratureFunctions1D::GivePolyPoints(const int np, double *pts,
         OpenHalfUniform(np, &ir);
         break;
      }
-      case Quadrature1D::ClosedGL:
-      {
-         ClosedGL(np, &ir);
-         break;
-      }
      default:
      {
         MFEM_ABORT("Asking for an unknown type of 1D Quadrature points, "
@@ -272,7 +272,6 @@ public:
   void OpenUniform(const int np, IntegrationRule *ir);
   void ClosedUniform(const int np, IntegrationRule *ir);
   void OpenHalfUniform(const int np, IntegrationRule *ir);
-   void ClosedGL(const int np, IntegrationRule *ir);
   ///@}

   /// A helper function that will play nice with Poly_1D::OpenPoints and
@@ -294,8 +293,7 @@ public:
      GaussLobatto    = 1,
      OpenUniform     = 2,  ///< aka open Newton-Cotes
      ClosedUniform   = 3,  ///< aka closed Newton-Cotes
-      OpenHalfUniform = 4,  ///< aka "open half" Newton-Cotes
-      ClosedGL        = 5   ///< aka closed Gauss Legendre
+      OpenHalfUniform = 4   ///< aka "open half" Newton-Cotes
   };
   /** @brief If the Quadrature1D type is not closed return Invalid; otherwise
       return type. */
@@ -97,8 +97,6 @@ static void InitCeedNonTensorBasisAndRestriction(const FiniteElementSpace &fes,
   Vector qweight(Q);
   Vector shape_i(P);
   DenseMatrix grad_i(P, dim);
-
-   CeedInt compstride = fes.GetOrdering()==Ordering::byVDIM ? 1 : fes.GetNDofs();
   const Table &el_dof = fes.GetElementToDofTable();
   Array<int> tp_el_dof(el_dof.Size_of_connections());
   const TensorBasisElement * tfe =
@@ -130,15 +128,7 @@ static void InitCeedNonTensorBasisAndRestriction(const FiniteElementSpace &fes,
         const int el_offset = fe->GetDof() * i;
         for (int j = 0; j < fe->GetDof(); j++)
         {
-            if (compstride == 1)
-            {
-               tp_el_dof[j + el_offset] = fes.GetVDim()*
-                                          el_dof.GetJ()[dof_map[j] + el_offset];
-            }
-            else
-            {
-               tp_el_dof[j + el_offset] = el_dof.GetJ()[dof_map[j] + el_offset];
-            }
+            tp_el_dof[j + el_offset] = el_dof.GetJ()[dof_map[j] + el_offset];
         }
      }
   }
@@ -167,23 +157,20 @@ static void InitCeedNonTensorBasisAndRestriction(const FiniteElementSpace &fes,
      {
         for (int i = 0; i < P; i++)
         {
-            if (compstride == 1)
-            {
-               tp_el_dof[i + e*P] = fes.GetVDim()*el_dof.GetJ()[i + e*P];
-            }
-            else
-            {
-               tp_el_dof[i + e*P] = el_dof.GetJ()[i + e*P];
-            }
+            tp_el_dof[i + e*P] = el_dof.GetJ()[i + e*P];
         }
      }
   }
   CeedBasisCreateH1(ceed, GetCeedTopology(fe->GetGeomType()), fes.GetVDim(),
                     fe->GetDof(), ir.GetNPoints(), shape.GetData(),
                     grad.GetData(), qref.GetData(), qweight.GetData(), basis);
-   CeedElemRestrictionCreate(ceed, mesh->GetNE(), fe->GetDof(), fes.GetVDim(),
-                             compstride, (fes.GetVDim())*(fes.GetNDofs()),
-                             CEED_MEM_HOST, CEED_COPY_VALUES,
+   CeedInterlaceMode imode = CEED_NONINTERLACED;
+   if (fes.GetOrdering()==Ordering::byVDIM)
+   {
+      imode = CEED_INTERLACED;
+   }
+   CeedElemRestrictionCreate(ceed, imode, mesh->GetNE(), fe->GetDof(),
+                             fes.GetNDofs(), fes.GetVDim(), CEED_MEM_HOST, CEED_COPY_VALUES,
                             tp_el_dof.GetData(), restr);
 }

@@ -228,7 +215,6 @@ static void InitCeedTensorBasisAndRestriction(const FiniteElementSpace &fes,
                           grad1d.GetData(), qref1d.GetData(),
                           qweight1d.GetData(), basis);

-   CeedInt compstride = fes.GetOrdering()==Ordering::byVDIM ? 1 : fes.GetNDofs();
   const Table &el_dof = fes.GetElementToDofTable();
   Array<int> tp_el_dof(el_dof.Size_of_connections());
   for (int i = 0; i < mesh->GetNE(); i++)
@@ -236,20 +222,16 @@ static void InitCeedTensorBasisAndRestriction(const FiniteElementSpace &fes,
      const int el_offset = fe->GetDof() * i;
      for (int j = 0; j < fe->GetDof(); j++)
      {
-         if (compstride == 1)
-         {
-            tp_el_dof[j + el_offset] = fes.GetVDim()*
-                                       el_dof.GetJ()[dof_map[j] + el_offset];
-         }
-         else
-         {
-            tp_el_dof[j + el_offset] = el_dof.GetJ()[dof_map[j] + el_offset];
-         }
+         tp_el_dof[j + el_offset] = el_dof.GetJ()[dof_map[j] + el_offset];
      }
   }
-   CeedElemRestrictionCreate(ceed, mesh->GetNE(), fe->GetDof(), fes.GetVDim(),
-                             compstride, (fes.GetVDim())*(fes.GetNDofs()),
-                             CEED_MEM_HOST, CEED_COPY_VALUES,
+   CeedInterlaceMode imode = CEED_NONINTERLACED;
+   if (fes.GetOrdering()==Ordering::byVDIM)
+   {
+      imode = CEED_INTERLACED;
+   }
+   CeedElemRestrictionCreate(ceed, imode, mesh->GetNE(), fe->GetDof(),
+                             fes.GetNDofs(), fes.GetVDim(), CEED_MEM_HOST, CEED_COPY_VALUES,
                             tp_el_dof.GetData(), restr);
 }

@@ -316,9 +298,8 @@ void CeedPAAssemble(const CeedPAOperator& op,
   CeedBasisGetNumQuadraturePoints(ceedData.basis, &nqpts);

   const int qdatasize = op.qdatasize;
-   CeedElemRestrictionCreateStrided(ceed, nelem, nqpts, qdatasize,
-                                    nelem*nqpts*qdatasize, CEED_STRIDES_BACKEND,
-                                    &ceedData.restr_i);
+   CeedElemRestrictionCreateStrided(ceed, nelem, nqpts, nelem*nqpts, qdatasize,
+                                    CEED_STRIDES_BACKEND, &ceedData.restr_i);

   CeedVectorCreate(ceed, mesh->GetNodes()->Size(), &ceedData.node_coords);
   CeedVectorSetArray(ceedData.node_coords, CEED_MEM_HOST, CEED_USE_POINTER,
@@ -415,59 +396,6 @@ void CeedPAAssemble(const CeedPAOperator& op,
   CeedVectorCreate(ceed, fes.GetNDofs(), &ceedData.v);
 }

-void CeedAddMultPA(const CeedData *ceedDataPtr,
-                   const Vector &x,
-                   Vector &y)
-{
-   const CeedScalar *x_ptr;
-   CeedScalar *y_ptr;
-   CeedMemType mem;
-   CeedGetPreferredMemType(internal::ceed, &mem);
-   if ( Device::Allows(Backend::CUDA) && mem==CEED_MEM_DEVICE )
-   {
-      x_ptr = x.Read();
-      y_ptr = y.ReadWrite();
-   }
-   else
-   {
-      x_ptr = x.HostRead();
-      y_ptr = y.HostReadWrite();
-      mem = CEED_MEM_HOST;
-   }
-   CeedVectorSetArray(ceedDataPtr->u, mem, CEED_USE_POINTER,
-                      const_cast<CeedScalar*>(x_ptr));
-   CeedVectorSetArray(ceedDataPtr->v, mem, CEED_USE_POINTER, y_ptr);
-
-   CeedOperatorApplyAdd(ceedDataPtr->oper, ceedDataPtr->u, ceedDataPtr->v,
-                        CEED_REQUEST_IMMEDIATE);
-
-   CeedVectorTakeArray(ceedDataPtr->u, mem, const_cast<CeedScalar**>(&x_ptr));
-   CeedVectorTakeArray(ceedDataPtr->v, mem, &y_ptr);
-}
-
-void CeedAssembleDiagonalPA(const CeedData *ceedDataPtr,
-                            Vector &diag)
-{
-   CeedScalar *d_ptr;
-   CeedMemType mem;
-   CeedGetPreferredMemType(internal::ceed, &mem);
-   if ( Device::Allows(Backend::CUDA) && mem==CEED_MEM_DEVICE )
-   {
-      d_ptr = diag.ReadWrite();
-   }
-   else
-   {
-      d_ptr = diag.HostReadWrite();
-      mem = CEED_MEM_HOST;
-   }
-   CeedVectorSetArray(ceedDataPtr->v, mem, CEED_USE_POINTER, d_ptr);
-
-   CeedOperatorLinearAssembleAddDiagonal(ceedDataPtr->oper, ceedDataPtr->v,
-                                         CEED_REQUEST_IMMEDIATE);
-
-   CeedVectorTakeArray(ceedDataPtr->v, mem, &d_ptr);
-}
-
 } // namespace mfem

 #endif // MFEM_USE_CEED
@@ -16,7 +16,6 @@

 #ifdef MFEM_USE_CEED
 #include "../../general/device.hpp"
-#include "../../linalg/vector.hpp"
 #include <ceed.h>

 namespace mfem
@@ -145,15 +144,6 @@ const std::string &GetCeedPath();
 void CeedPAAssemble(const CeedPAOperator& op,
                    CeedData& ceedData);

-/** @brief Function that applies a libCEED PA operator. */
-void CeedAddMultPA(const CeedData *ceedDataPtr,
-                   const Vector &x,
-                   Vector &y);
-
-/** @brief Function that assembles a libCEED PA operator diagonal. */
-void CeedAssembleDiagonalPA(const CeedData *ceedDataPtr,
-                            Vector &diag);
-
 /** @brief Function that determines if a CEED kernel should be used, based on
    the current mfem::Device configuration. */
 inline bool DeviceCanUseCeed()
@@ -199,8 +199,7 @@ void LinearForm::Assemble()
 void LinearForm::Update(FiniteElementSpace *f, Vector &v, int v_offset)
 {
   fes = f;
-   NewMemoryAndSize(Memory<double>(v.GetMemory(), v_offset, f->GetVSize()),
-                    f->GetVSize(), false);
+   NewDataAndSize((double *)v + v_offset, fes->GetVSize());
   ResetDeltaLocations();
 }

@@ -159,13 +159,10 @@ void BoundaryLFIntegrator::AssembleRHSElementVect(
   for (int i = 0; i < ir->GetNPoints(); i++)
   {
      const IntegrationPoint &ip = ir->IntPoint(i);
+      IntegrationPoint eip;
+      Tr.Loc1.Transform(ip, eip);

-      // Set the integration point in the face and the neighboring element
-      Tr.SetAllIntPoints(&ip);
-
-      // Access the neighboring element's integration point
-      const IntegrationPoint &eip = Tr.GetElement1IntPoint();
-
+      Tr.Face->SetIntPoint (&ip);
      double val = Tr.Face->Weight() * ip.weight * Q.Eval(*Tr.Face, ip);

      el.CalcShape(eip, shape);
@@ -362,12 +359,10 @@ void VectorBoundaryLFIntegrator::AssembleRHSElementVect(
   for (int i = 0; i < ir->GetNPoints(); i++)
   {
      const IntegrationPoint &ip = ir->IntPoint(i);
+      IntegrationPoint eip;
+      Tr.Loc1.Transform(ip, eip);

-      // Set the integration point in the face and the neighboring element
-      Tr.SetAllIntPoints(&ip);
-
-      // Access the neighboring element's integration point
-      const IntegrationPoint &eip = Tr.GetElement1IntPoint();
+      Tr.SetIntPoint(&ip);

      // Use Tr transformation in case Q depends on boundary attribute
      Q.Eval(vec, Tr, ip);
@@ -688,14 +683,12 @@ void BoundaryFlowIntegrator::AssembleRHSElementVect(
   for (int p = 0; p < ir->GetNPoints(); p++)
   {
      const IntegrationPoint &ip = ir->IntPoint(p);
-
-      // Set the integration point in the face and the neighboring element
-      Tr.SetAllIntPoints(&ip);
-
-      // Access the neighboring element's integration point
-      const IntegrationPoint &eip = Tr.GetElement1IntPoint();
+      IntegrationPoint eip;
+      Tr.Loc1.Transform(ip, eip);
      el.CalcShape(eip, shape);

+      Tr.SetIntPoint(&ip);
+
      // Use Tr.Elem1 transformation for u so that it matches the coefficient
      // used with the ConvectionIntegrator and/or the DGTraceIntegrator.
      u->Eval(vu, *Tr.Elem1, eip);
@@ -759,13 +752,10 @@ void DGDirichletLFIntegrator::AssembleRHSElementVect(
   for (int p = 0; p < ir->GetNPoints(); p++)
   {
      const IntegrationPoint &ip = ir->IntPoint(p);
+      IntegrationPoint eip;

-      // Set the integration point in the face and the neighboring element
-      Tr.SetAllIntPoints(&ip);
-
-      // Access the neighboring element's integration point
-      const IntegrationPoint &eip = Tr.GetElement1IntPoint();
-
+      Tr.Loc1.Transform(ip, eip);
+      Tr.SetIntPoint(&ip);
      if (dim == 1)
      {
         nor(0) = 2*eip.x - 1.0;
@@ -784,14 +774,14 @@ void DGDirichletLFIntegrator::AssembleRHSElementVect(
      {
         if (Q)
         {
-            w *= Q->Eval(*Tr.Elem1, eip);
+            w *= Q->Eval(Tr, ip);
         }
         ni.Set(w, nor);
      }
      else
      {
         nh.Set(w, nor);
-         MQ->Eval(mq, *Tr.Elem1, eip);
+         MQ->Eval(mq, Tr, ip);
         mq.MultTranspose(nh, ni);
      }
      CalcAdjugate(Tr.Elem1->Jacobian(), adjJ);
@@ -855,12 +845,9 @@ void DGElasticityDirichletLFIntegrator::AssembleRHSElementVect(
   for (int pi = 0; pi < ir->GetNPoints(); ++pi)
   {
      const IntegrationPoint &ip = ir->IntPoint(pi);
-
-      // Set the integration point in the face and the neighboring element
-      Tr.SetAllIntPoints(&ip);
-
-      // Access the neighboring element's integration point
-      const IntegrationPoint &eip = Tr.GetElement1IntPoint();
+      IntegrationPoint eip;
+      Tr.Loc1.Transform(ip, eip);
+      Tr.SetIntPoint(&ip);

      // Evaluate the Dirichlet b.c. using the face transformation.
      uD.Eval(u_dir, Tr, ip);
@@ -135,18 +135,11 @@ void Multigrid::SetOperator(const Operator& op)
   MFEM_ABORT("SetOperator not supported in Multigrid");
 }

-void Multigrid::SmoothingStep(int level, bool transpose) const
+void Multigrid::SmoothingStep(int level) const
 {
   GetOperatorAtLevel(level)->Mult(*Y[level], *R[level]); // r = A x
   subtract(*X[level], *R[level], *R[level]);             // r = b - A x
-   if (transpose)
-   {
-      GetSmootherAtLevel(level)->MultTranspose(*R[level], *Z[level]); // z = S r
-   }
-   else
-   {
-      GetSmootherAtLevel(level)->Mult(*R[level], *Z[level]); // z = S r
-   }
+   GetSmootherAtLevel(level)->Mult(*R[level], *Z[level]); // z = S r
   add(*Y[level], 1.0, *Z[level], *Y[level]);             // x = x + S (b - A x)
 }

@@ -160,7 +153,7 @@ void Multigrid::Cycle(int level) const

   for (int i = 0; i < preSmoothingSteps; i++)
   {
-      SmoothingStep(level, false);
+      SmoothingStep(level);
   }

   // Compute residual
@@ -194,7 +187,7 @@ void Multigrid::Cycle(int level) const
   // Post-smooth
   for (int i = 0; i < postSmoothingSteps; i++)
   {
-      SmoothingStep(level, true);
+      SmoothingStep(level);
   }
 }

@@ -108,7 +108,7 @@ public:

 private:
   /// Application of a smoothing step at particular level
-   void SmoothingStep(int level, bool transpose) const;
+   void SmoothingStep(int level) const;

   /// Application of a cycle at particular level
   void Cycle(int level) const;
@@ -10,7 +10,6 @@
 // CONTRIBUTING.md for details.

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

 namespace mfem
 {
@@ -28,7 +27,7 @@ void NonlinearForm::SetAssemblyLevel(AssemblyLevel assembly_level)
         // This is the default behavior.
         break;
      case AssemblyLevel::PARTIAL:
-         ext = new PANonlinearForm(this);
+         ext = new PANonlinearFormExtension(this);
         break;
      default:
         mfem_error("Unknown assembly level for this form.");
@@ -81,13 +80,6 @@ void NonlinearForm::SetEssentialVDofs(const Array<int> &ess_vdofs_list)

 double NonlinearForm::GetGridFunctionEnergy(const Vector &x) const
 {
-   if (ext)
-   {
-      MFEM_VERIFY(!fnfi.Size(), "Interior faces terms not yet implemented!");
-      MFEM_VERIFY(!bfnfi.Size(), "Boundary face terms not yet implemented!");
-      return ext->GetGridFunctionEnergy(x);
-   }
-
   Array<int> vdofs;
   Vector el_x;
   const FiniteElement *fe;
@@ -146,14 +138,6 @@ void NonlinearForm::Mult(const Vector &x, Vector &y) const
   if (ext)
   {
      ext->Mult(px, py);
-      if (Serial())
-      {
-         if (cP) { cP->MultTranspose(py, y); }
-         const int N = ess_tdof_list.Size();
-         const auto tdof = ess_tdof_list.Read();
-         auto Y = y.ReadWrite();
-         MFEM_FORALL(i, N, Y[tdof[i]] = 0.0; );
-      }
      return;
   }

@@ -280,16 +264,7 @@ Operator &NonlinearForm::GetGradient(const Vector &x) const
 {
   if (ext)
   {
-      Operator &grad = ext->GetGradient(Prolongate(x));
-      hGrad.Reset(&grad, false);
-      if (Serial())
-      {
-         Operator *Gop;
-         if (cP) { hGrad.Reset(new RAPOperator(*cP, grad, *cP)); }
-         hGrad.Ptr()->Operator::FormSystemOperator(ess_tdof_list, Gop);
-         hGrad.Reset(Gop);
-      }
-      return *hGrad.Ptr();
+      MFEM_ABORT("Not yet implemented!");
   }

   const int skip_zeros = 0;
@@ -451,31 +426,7 @@ void NonlinearForm::Update()

 void NonlinearForm::Setup()
 {
-   if (ext) { return ext->Setup(); }
-}
-
-void NonlinearForm::AssembleGradientDiagonal(Vector &diag) const
-{
-   if (ext)
-   {
-      MFEM_ASSERT(diag.Size() == fes->GetTrueVSize(),
-                  "Vector for holding diagonal has wrong size!");
-      const Operator *P = fes->GetProlongationMatrix();
-      if (!IsIdentityProlongation(P))
-      {
-         Vector local_diag(P->Height());
-         ext->AssembleGradientDiagonal(local_diag);
-         P->MultTranspose(local_diag, diag);
-      }
-      else
-      {
-         ext->AssembleGradientDiagonal(diag);
-      }
-   }
-   else
-   {
-      MFEM_ABORT("Not implemented. Can be obtained through GetGradient().");
-   }
+   if (ext) { return ext->AssemblePA(); }
 }

 NonlinearForm::~NonlinearForm()
@@ -982,17 +933,6 @@ Operator &BlockNonlinearForm::GetGradientBlocked(const BlockVector &bx) const
      }
   }

-   if (!Grads(0,0)->Finalized())
-   {
-      for (int i=0; i<fes.Size(); ++i)
-      {
-         for (int j=0; j<fes.Size(); ++j)
-         {
-            Grads(i,j)->Finalize(skip_zeros);
-         }
-      }
-   }
-
   for (int s=0; s<fes.Size(); ++s)
   {
      for (int i = 0; i < ess_vdofs[s]->Size(); ++i)
@@ -1012,6 +952,17 @@ Operator &BlockNonlinearForm::GetGradientBlocked(const BlockVector &bx) const
      }
   }

+   if (!Grads(0,0)->Finalized())
+   {
+      for (int i=0; i<fes.Size(); ++i)
+      {
+         for (int j=0; j<fes.Size(); ++j)
+         {
+            Grads(i,j)->Finalize(skip_zeros);
+         }
+      }
+   }
+
   for (int i=0; i<fes.Size(); ++i)
   {
      for (int j=0; j<fes.Size(); ++j)
@@ -45,7 +45,6 @@ protected:
   Array<Array<int>*>              bfnfi_marker; // not owned

   mutable SparseMatrix *Grad, *cGrad; // owned
-   mutable OperatorHandle hGrad;

   /// A list of all essential true dofs
   Array<int> ess_tdof_list;
@@ -166,15 +165,6 @@ public:
   /// Setup the NonlinearForm
   virtual void Setup();

-   /** @brief Assemble the diagonal of the gradient into diag
-
-       For adaptively refined meshes, this returns P^T d_e, where d_e is the
-       locally assembled diagonal on each element and P^T is the transpose of
-       the conforming prolongation. In general this is not the correct diagonal
-       for an AMR mesh. */
-   void AssembleGradientDiagonal(Vector &diag) const;
-
-
   /// Get the finite element space prolongation matrix
   virtual const Operator *GetProlongation() const { return P; }
   /// Get the finite element space restriction matrix
@@ -13,101 +13,62 @@
 // PABilinearFormExtension and MFBilinearFormExtension.

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

 namespace mfem
 {

-NonlinearFormExtension::NonlinearFormExtension(const NonlinearForm *nlf)
-   : Operator(nlf->FESpace()->GetTrueVSize()), nlf(nlf) { }
-
-PANonlinearForm::PANonlinearForm(NonlinearForm *nlf):
-   NonlinearFormExtension(nlf),
-   x_grad(NULL),
-   fes(*nlf->FESpace()),
-   dnfi(*nlf->GetDNFI()),
-   R(fes.GetElementRestriction(ElementDofOrdering::LEXICOGRAPHIC))
+NonlinearFormExtension::NonlinearFormExtension(NonlinearForm *form)
+   : Operator(form->FESpace()->GetTrueVSize()), n(form)
 {
-   MFEM_VERIFY(R, "Not yet implemented!");
-   xe.SetSize(R->Height(), Device::GetMemoryType());
-   ye.SetSize(R->Height(), Device::GetMemoryType());
-   ye.UseDevice(true);
+   // empty
 }

-double PANonlinearForm::GetGridFunctionEnergy(const Vector &x) const
+PANonlinearFormExtension::PANonlinearFormExtension(NonlinearForm *form):
+   NonlinearFormExtension(form), fes(*form->FESpace())
 {
-   double energy = 0.0;
-
-   R->Mult(x, xe);
-   for (int i = 0; i < dnfi.Size(); i++)
+   const ElementDofOrdering ordering = ElementDofOrdering::LEXICOGRAPHIC;
+   elem_restrict_lex = fes.GetElementRestriction(ordering);
+   if (elem_restrict_lex)
   {
-      energy += dnfi[i]->GetGridFunctionEnergyPA(xe);
+      localX.SetSize(elem_restrict_lex->Height(), Device::GetMemoryType());
+      localY.SetSize(elem_restrict_lex->Height(), Device::GetMemoryType());
+      localY.UseDevice(true); // ensure 'localY = 0.0' is done on device
   }
-   return energy;
 }

-void PANonlinearForm::Setup()
+void PANonlinearFormExtension::AssemblePA()
 {
-   for (int i = 0; i < dnfi.Size(); ++i) { dnfi[i]->AssemblePA(fes); }
-}
-
-void PANonlinearForm::Mult(const Vector &x, Vector &y) const
-{
-   ye = 0.0;
-   R->Mult(x, xe);
-   for (int i = 0; i < dnfi.Size(); ++i) { dnfi[i]->AddMultPA(xe, ye); }
-   R->MultTranspose(ye, y);
-}
-
-void PANonlinearForm::AssembleGradientDiagonal(Vector &diag) const
-{
-   MFEM_VERIFY(x_grad, "GetGradient() has not been called");
-   R->Mult(*x_grad, xe);
-
-   ye = 0.0;
-   for (int i = 0; i < dnfi.Size(); ++i)
+   Array<NonlinearFormIntegrator*> &integrators = *n->GetDNFI();
+   const int Ni = integrators.Size();
+   for (int i = 0; i < Ni; ++i)
   {
-      dnfi[i]->AssembleGradientDiagonalPA(xe, ye);
+      integrators[i]->AssemblePA(*n->FESpace());
   }
-   R->MultTranspose(ye, diag);
 }

-Operator &PANonlinearForm::GetGradient(const Vector &x) const
+void PANonlinearFormExtension::Mult(const Vector &x, Vector &y) const
 {
-   // Store the last x that was used to compute the gradient.
-   x_grad = &x;
-
-   Grad.Reset(new PANonlinearForm::Gradient(x, *this));
-   return *Grad.Ptr();
+   Array<NonlinearFormIntegrator*> &integrators = *n->GetDNFI();
+   const int iSz = integrators.Size();
+   if (elem_restrict_lex)
+   {
+      elem_restrict_lex->Mult(x, localX);
+      localY = 0.0;
+      for (int i = 0; i < iSz; ++i)
+      {
+         integrators[i]->AddMultPA(localX, localY);
+      }
+      elem_restrict_lex->MultTranspose(localY, y);
+   }
+   else
+   {
+      y.UseDevice(true); // typically this is a large vector, so store on device
+      y = 0.0;
+      for (int i = 0; i < iSz; ++i)
+      {
+         integrators[i]->AddMultPA(x, y);
+      }
+   }
 }

-PANonlinearForm::Gradient::Gradient(const Vector &x, const PANonlinearForm &e):
-   Operator(e.fes.GetVSize()), R(e.R), dnfi(e.dnfi)
-{
-   ge.UseDevice(true);
-   ge.SetSize(R->Height(), Device::GetMemoryType());
-   R->Mult(x, ge);
-
-   xe.UseDevice(true);
-   xe.SetSize(R->Height(), Device::GetMemoryType());
-
-   ye.UseDevice(true);
-   ye.SetSize(R->Height(), Device::GetMemoryType());
-
-   ze.UseDevice(true);
-   ze.SetSize(R->Height(), Device::GetMemoryType());
-
-   // Do we still need to do this?
-   for (int i = 0; i < dnfi.Size(); ++i) { dnfi[i]->AssemblePA(e.fes); }
 }
-
-void PANonlinearForm::Gradient::Mult(const Vector &x, Vector &y) const
-{
-   ze = x;
-   ye = 0.0;
-   R->Mult(ze, xe);
-   for (int i = 0; i < dnfi.Size(); ++i) { dnfi[i]->AddMultGradPA(ge, xe, ye); }
-   R->MultTranspose(ye, y);
-}
-
-} // namespace mfem
@@ -17,60 +17,28 @@

 namespace mfem
 {
-
 class NonlinearForm;
-class NonlinearFormIntegrator;

-/** @brief Class extending the NonlinearForm class to support the different
-    AssemblyLevel%s. */
 class NonlinearFormExtension : public Operator
 {
 protected:
-   const NonlinearForm *nlf;
+   NonlinearForm *n; ///< Not owned
 public:
-   NonlinearFormExtension(const NonlinearForm*);
-   virtual void Setup() = 0;
-   virtual Operator &GetGradient(const Vector&) const = 0;
-   virtual double GetGridFunctionEnergy(const Vector &x) const = 0;
-   virtual void AssembleGradientDiagonal(Vector &diag) const
-   {
-      MFEM_ABORT("Not implemented for this assembly level!");
-   }
+   NonlinearFormExtension(NonlinearForm *form);
+   virtual void AssemblePA() = 0;
 };

-class PANonlinearForm;
-
-
 /// Data and methods for partially-assembled nonlinear forms
-class PANonlinearForm : public NonlinearFormExtension
+class PANonlinearFormExtension : public NonlinearFormExtension
 {
-private:
-   class Gradient : public Operator
-   {
-   protected:
-      const Operator *R;
-      mutable Vector ge, xe, ye, ze;
-      const Array<NonlinearFormIntegrator*> &dnfi;
-   public:
-      Gradient(const Vector &x, const PANonlinearForm &ext);
-      virtual void Mult(const Vector &x, Vector &y) const;
-   };
-
 protected:
-   mutable Vector xe, ye;
-   mutable const Vector *x_grad;
-   mutable OperatorHandle Grad;
-   const FiniteElementSpace &fes;
-   const Array<NonlinearFormIntegrator*> &dnfi;
-   const Operator *R;
-
+   const FiniteElementSpace &fes; // Not owned
+   mutable Vector localX, localY;
+   const Operator *elem_restrict_lex; // Not owned
 public:
-   PANonlinearForm(NonlinearForm *nlf);
-   void Setup();
+   PANonlinearFormExtension(NonlinearForm*);
+   void AssemblePA();
   void Mult(const Vector &x, Vector &y) const;
-   Operator &GetGradient(const Vector &x) const;
-   double GetGridFunctionEnergy(const Vector &x) const;
-   void AssembleGradientDiagonal(Vector &diag) const;
 };
 }
 #endif // NONLINEARFORM_EXT_HPP
@@ -15,13 +15,6 @@
 namespace mfem
 {

-double NonlinearFormIntegrator::GetGridFunctionEnergyPA(const Vector &x) const
-{
-   mfem_error ("NonlinearFormIntegrator::GetGridFunctionEnergyPA(...)\n"
-               "   is not implemented for this class.");
-   return 0.0;
-}
-
 void NonlinearFormIntegrator::AssemblePA(const FiniteElementSpace&)
 {
   mfem_error ("NonlinearFormIntegrator::AssemblePA(...)\n"
@@ -41,20 +34,6 @@ void NonlinearFormIntegrator::AddMultPA(const Vector &, Vector &) const
               "   is not implemented for this class.");
 }

-void NonlinearFormIntegrator::AddMultGradPA(const Vector&,
-                                            const Vector&, Vector&) const
-{
-   mfem_error ("NonlinearFormIntegrator::AddMultGradPA(...)\n"
-               "   is not implemented for this class.");
-}
-
-void NonlinearFormIntegrator::AssembleGradientDiagonalPA(const mfem::Vector &x,
-                                                         mfem::Vector &diag) const
-{
-   mfem_error ("NonlinearFormIntegrator::AssembleDiagonalPA(...)\n"
-               "   is not implemented for this class.");
-}
-
 void NonlinearFormIntegrator::AssembleElementVector(
   const FiniteElement &el, ElementTransformation &Tr,
   const Vector &elfun, Vector &elvect)
@@ -68,9 +68,6 @@ public:
                                   ElementTransformation &Tr,
                                   const Vector &elfun);

-   /// Compute the local energy with partial assembly.
-   virtual double GetGridFunctionEnergyPA(const Vector &x) const;
-
   /// Method defining partial assembly.
   /** The result of the partial assembly is stored internally so that it can be
       used later in the methods AddMultPA(). */
@@ -91,12 +88,6 @@ public:
       called. */
   virtual void AddMultPA(const Vector &x, Vector &y) const;

-   /// Method for partially assembled gradient action.
-   virtual void AddMultGradPA(const Vector &g,
-                              const Vector &x, Vector &y) const;
-
-   virtual void AssembleGradientDiagonalPA(const Vector &x, Vector &diag) const;
-
   virtual ~NonlinearFormIntegrator() { }
 };

@@ -198,9 +198,8 @@ void ParBilinearForm::AssembleSharedFaces(int skip_zeros)
   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);
+      pfes->GetFaceNbrElementVDofs(T->Elem2No, vdofs2);
      vdofs1.Copy(vdofs_all);
      for (int j = 0; j < vdofs2.Size(); j++)
      {
@@ -217,7 +216,7 @@ void ParBilinearForm::AssembleSharedFaces(int skip_zeros)
      for (int k = 0; k < fbfi.Size(); k++)
      {
         fbfi[k]->AssembleFaceMatrix(*pfes->GetFE(T->Elem1No),
-                                     *pfes->GetFaceNbrFE(Elem2NbrNo),
+                                     *pfes->GetFaceNbrFE(T->Elem2No),
                                     *T, elemmat);
         if (keep_nbr_block)
         {
@@ -241,7 +240,7 @@ void ParBilinearForm::Assemble(int skip_zeros)

   BilinearForm::Assemble(skip_zeros);

-   if (!ext && fbfi.Size() > 0)
+   if (fbfi.Size() > 0)
   {
      AssembleSharedFaces(skip_zeros);
   }
@@ -3147,7 +3147,7 @@ static void SetSubVector(const int N,
                         const Array<int> &indices,
                         const Vector &in, Vector &out)
 {
-   auto y = out.ReadWrite();
+   auto y = out.Write();
   const auto x = in.Read();
   const auto I = indices.Read();
   MFEM_FORALL(i, N, y[I[i]] = x[i];);
@@ -3234,7 +3234,7 @@ static void AddSubVector(const int num_unique_dst_indices,
                         const Vector &src,
                         Vector &dst)
 {
-   auto y = dst.ReadWrite();
+   auto y = dst.Write();
   const auto x = src.Read();
   const auto DST_I = unique_dst_indices.Read();
   const auto SRC_O = unique_to_src_offsets.Read();
@@ -347,8 +347,6 @@ public:
   const FiniteElement *GetFaceNbrFE(int i) const;
   const FiniteElement *GetFaceNbrFaceFE(int i) const;
   const HYPRE_Int *GetFaceNbrGlobalDofMap() { return face_nbr_glob_dof_map; }
-   ElementTransformation *GetFaceNbrElementTransformation(int i) const
-   { return pmesh->GetFaceNbrElementTransformation(i); }

   void Lose_Dof_TrueDof_Matrix();
   void LoseDofOffsets() { dof_offsets.LoseData(); }
@@ -271,7 +271,6 @@ const
   {
      int fes_vdim = pfes->GetVDim();
      pfes->GetFaceNbrElementVDofs(nbr_el_no, dofs);
-      const FiniteElement *fe = pfes->GetFaceNbrFE(nbr_el_no);
      if (fes_vdim > 1)
      {
         int s = dofs.Size()/fes_vdim;
@@ -284,17 +283,7 @@ const
         face_nbr_data.GetSubVector(dofs, LocVec);
         DofVal.SetSize(dofs.Size());
      }
-      if (fe->GetMapType() == FiniteElement::VALUE)
-      {
-         fe->CalcShape(ip, DofVal);
-      }
-      else
-      {
-         ElementTransformation *Tr =
-            pfes->GetFaceNbrElementTransformation(nbr_el_no);
-         Tr->SetIntPoint(&ip);
-         fe->CalcPhysShape(*Tr, DofVal);
-      }
+      pfes->GetFaceNbrFE(nbr_el_no)->CalcShape(ip, DofVal);
   }
   else
   {
@@ -302,175 +291,14 @@ const
      fes->DofsToVDofs(vdim-1, dofs);
      DofVal.SetSize(dofs.Size());
      const FiniteElement *fe = fes->GetFE(i);
-      if (fe->GetMapType() == FiniteElement::VALUE)
-      {
-         fe->CalcShape(ip, DofVal);
-      }
-      else
-      {
-         ElementTransformation *Tr = fes->GetElementTransformation(i);
-         Tr->SetIntPoint(&ip);
-         fe->CalcPhysShape(*Tr, DofVal);
-      }
+      MFEM_ASSERT(fe->GetMapType() == FiniteElement::VALUE, "invalid FE map type");
+      fe->CalcShape(ip, DofVal);
      GetSubVector(dofs, LocVec);
   }

   return (DofVal * LocVec);
 }

-void ParGridFunction::GetVectorValue(int i, const IntegrationPoint &ip,
-                                     Vector &val) const
-{
-   int nbr_el_no = i - pfes->GetParMesh()->GetNE();
-   if (nbr_el_no >= 0)
-   {
-      Array<int> dofs;
-      pfes->GetFaceNbrElementVDofs(nbr_el_no, dofs);
-      Vector loc_data;
-      face_nbr_data.GetSubVector(dofs, loc_data);
-      const FiniteElement *FElem = pfes->GetFaceNbrFE(nbr_el_no);
-      int dof = FElem->GetDof();
-      if (FElem->GetRangeType() == FiniteElement::SCALAR)
-      {
-         Vector shape(dof);
-         if (FElem->GetMapType() == FiniteElement::VALUE)
-         {
-            FElem->CalcShape(ip, shape);
-         }
-         else
-         {
-            ElementTransformation *Tr =
-               pfes->GetParMesh()->GetFaceNbrElementTransformation(nbr_el_no);
-            Tr->SetIntPoint(&ip);
-            FElem->CalcPhysShape(*Tr, shape);
-         }
-         int vdim = fes->GetVDim();
-         val.SetSize(vdim);
-         for (int k = 0; k < vdim; k++)
-         {
-            val(k) = shape * ((const double *)loc_data + dof * k);
-         }
-      }
-      else
-      {
-         int spaceDim = fes->GetMesh()->SpaceDimension();
-         DenseMatrix vshape(dof, spaceDim);
-         ElementTransformation *Tr =
-            pfes->GetParMesh()->GetFaceNbrElementTransformation(nbr_el_no);
-         Tr->SetIntPoint(&ip);
-         FElem->CalcVShape(*Tr, vshape);
-         val.SetSize(spaceDim);
-         vshape.MultTranspose(loc_data, val);
-      }
-   }
-   else
-   {
-      GridFunction::GetVectorValue(i, ip, val);
-   }
-}
-
-double ParGridFunction::GetValue(ElementTransformation &T,
-                                 const IntegrationPoint &ip,
-                                 int comp, Vector *tr) const
-{
-   // We can assume faces and edges are local
-   if (T.ElementType != ElementTransformation::ELEMENT)
-   {
-      return GridFunction::GetValue(T, ip, comp, tr);
-   }
-
-   // Check for evaluation in a local element
-   int nbr_el_no = T.ElementNo - pfes->GetParMesh()->GetNE();
-   if (nbr_el_no < 0)
-   {
-      return GridFunction::GetValue(T, ip, comp, tr);
-   }
-
-   // Evaluate using DoFs from a neighboring element
-   if (tr)
-   {
-      T.SetIntPoint(&ip);
-      T.Transform(ip, *tr);
-   }
-
-   Array<int> dofs;
-   const FiniteElement * fe = pfes->GetFaceNbrFE(nbr_el_no);
-   pfes->GetFaceNbrElementVDofs(nbr_el_no, dofs);
-
-   pfes->DofsToVDofs(comp-1, dofs);
-   Vector DofVal(dofs.Size()), LocVec;
-   if (fe->GetMapType() == FiniteElement::VALUE)
-   {
-      fe->CalcShape(ip, DofVal);
-   }
-   else
-   {
-      fe->CalcPhysShape(T, DofVal);
-   }
-   face_nbr_data.GetSubVector(dofs, LocVec);
-
-   return (DofVal * LocVec);
-}
-
-void ParGridFunction::GetVectorValue(ElementTransformation &T,
-                                     const IntegrationPoint &ip,
-                                     Vector &val, Vector *tr) const
-{
-   // We can assume faces and edges are local
-   if (T.ElementType != ElementTransformation::ELEMENT)
-   {
-      return GridFunction::GetVectorValue(T, ip, val, tr);
-   }
-
-   // Check for evaluation in a local element
-   int nbr_el_no = T.ElementNo - pfes->GetParMesh()->GetNE();
-   if (nbr_el_no < 0)
-   {
-      return GridFunction::GetVectorValue(T, ip, val, tr);
-   }
-
-   // Evaluate using DoFs from a neighboring element
-   if (tr)
-   {
-      T.SetIntPoint(&ip);
-      T.Transform(ip, *tr);
-   }
-
-   Array<int> vdofs;
-   pfes->GetFaceNbrElementVDofs(nbr_el_no, vdofs);
-   const FiniteElement *fe = pfes->GetFaceNbrFE(nbr_el_no);
-
-   int dof = fe->GetDof();
-   Vector loc_data;
-   face_nbr_data.GetSubVector(vdofs, loc_data);
-   if (fe->GetRangeType() == FiniteElement::SCALAR)
-   {
-      Vector shape(dof);
-      if (fe->GetMapType() == FiniteElement::VALUE)
-      {
-         fe->CalcShape(ip, shape);
-      }
-      else
-      {
-         fe->CalcPhysShape(T, shape);
-      }
-      int vdim = pfes->GetVDim();
-      val.SetSize(vdim);
-      for (int k = 0; k < vdim; k++)
-      {
-         val(k) = shape * ((const double *)loc_data + dof * k);
-      }
-   }
-   else
-   {
-      int spaceDim = pfes->GetMesh()->SpaceDimension();
-      DenseMatrix vshape(dof, spaceDim);
-      fe->CalcVShape(T, vshape);
-      val.SetSize(spaceDim);
-      vshape.MultTranspose(loc_data, val);
-   }
-}
-
 void ParGridFunction::ProjectCoefficient(Coefficient &coeff)
 {
   DeltaCoefficient *delta_c = dynamic_cast<DeltaCoefficient *>(&coeff);
@@ -711,9 +539,7 @@ void ParGridFunction::SaveAsOne(std::ostream &out)
   int *nfdofs = new int[NRanks];
   int *nrdofs = new int[NRanks];

-   HostReadWrite();
   values[0] = data;
-
   nv[0]     = pfes -> GetVSize();
   nvdofs[0] = pfes -> GetNVDofs();
   nedofs[0] = pfes -> GetNEDofs();
@@ -39,7 +39,7 @@ protected:
   Vector face_nbr_data;

   /** @brief Vector used as an MPI buffer to send face-neighbor data
-       in ExchangeFaceNbrData() to neighboring processors. */
+        in ExchangeFaceNbrData() to neighboring processors. */
   //TODO: Use temporary memory to avoid CUDA malloc allocation cost.
   Vector send_data;

@@ -209,18 +209,6 @@ public:
   double GetValue(ElementTransformation &T)
   { return GetValue(T.ElementNo, T.GetIntPoint()); }

-   // Redefine to handle the case when T describes a face-neighbor element
-   virtual double GetValue(ElementTransformation &T, const IntegrationPoint &ip,
-                           int comp = 0, Vector *tr = NULL) const;
-
-   virtual void GetVectorValue(int i, const IntegrationPoint &ip,
-                               Vector &val) const;
-
-   // Redefine to handle the case when T describes a face-neighbor element
-   virtual void GetVectorValue(ElementTransformation &T,
-                               const IntegrationPoint &ip,
-                               Vector &val, Vector *tr = NULL) const;
-
   using GridFunction::ProjectCoefficient;
   virtual void ProjectCoefficient(Coefficient &coeff);

@@ -14,7 +14,6 @@
 #ifdef MFEM_USE_MPI

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

 namespace mfem
 {
@@ -50,7 +49,6 @@ void ParNonlinearForm::Mult(const Vector &x, Vector &y) const

   if (fnfi.Size())
   {
-      MFEM_VERIFY(!NonlinearForm::ext,"");
      // Terms over shared interior faces in parallel.
      ParFiniteElementSpace *pfes = ParFESpace();
      ParMesh *pmesh = pfes->GetParMesh();
@@ -66,13 +64,12 @@ void ParNonlinearForm::Mult(const Vector &x, Vector &y) const
      for (int i = 0; i < n_shared_faces; i++)
      {
         tr = pmesh->GetSharedFaceTransformations(i, true);
-         int Elem2NbrNo = tr->Elem2No - pmesh->GetNE();

         fe1 = pfes->GetFE(tr->Elem1No);
-         fe2 = pfes->GetFaceNbrFE(Elem2NbrNo);
+         fe2 = pfes->GetFaceNbrFE(tr->Elem2No);

         pfes->GetElementVDofs(tr->Elem1No, vdofs1);
-         pfes->GetFaceNbrElementVDofs(Elem2NbrNo, vdofs2);
+         pfes->GetFaceNbrElementVDofs(tr->Elem2No, vdofs2);

         el_x.SetSize(vdofs1.Size() + vdofs2.Size());
         X.GetSubVector(vdofs1, el_x.GetData());
@@ -88,16 +85,15 @@ void ParNonlinearForm::Mult(const Vector &x, Vector &y) const

   P->MultTranspose(aux2, y);

-   const int N = ess_tdof_list.Size();
-   const auto idx = ess_tdof_list.Read();
-   auto Y = y.ReadWrite();
-   MFEM_FORALL(i, N, Y[idx[i]] = 0.0; );
+   y.HostReadWrite();
+   for (int i = 0; i < ess_tdof_list.Size(); i++)
+   {
+      y(ess_tdof_list[i]) = 0.0;
+   }
 }

 const SparseMatrix &ParNonlinearForm::GetLocalGradient(const Vector &x) const
 {
-   if (NonlinearForm::ext) { MFEM_ABORT("Not yet implemented!"); }
-
   NonlinearForm::GetGradient(x); // (re)assemble Grad, no b.c.

   return *Grad;
@@ -107,20 +103,16 @@ Operator &ParNonlinearForm::GetGradient(const Vector &x) const
 {
   ParFiniteElementSpace *pfes = ParFESpace();

-   Operator &grad = NonlinearForm::GetGradient(x); // (re)assemble Grad, no b.c.
-
   pGrad.Clear();

+   NonlinearForm::GetGradient(x); // (re)assemble Grad, no b.c.
+
   OperatorHandle dA(pGrad.Type()), Ph(pGrad.Type());

   if (fnfi.Size() == 0)
   {
-      if (NonlinearForm::ext) { dA.Reset(&grad, false); }
-      else
-      {
-         dA.MakeSquareBlockDiag(pfes->GetComm(), pfes->GlobalVSize(),
-                                pfes->GetDofOffsets(), Grad);
-      }
+      dA.MakeSquareBlockDiag(pfes->GetComm(), pfes->GlobalVSize(),
+                             pfes->GetDofOffsets(), Grad);
   }
   else
   {
@@ -27,24 +27,35 @@ ParL2FaceRestriction::ParL2FaceRestriction(const ParFiniteElementSpace &fes,
                                           ElementDofOrdering e_ordering,
                                           FaceType type,
                                           L2FaceValues m)
-   : L2FaceRestriction(fes, type, m)
+   : fes(fes),
+     nf(fes.GetNFbyType(type)),
+     vdim(fes.GetVDim()),
+     byvdim(fes.GetOrdering() == Ordering::byVDIM),
+     ndofs(fes.GetNDofs()),
+     dof(nf>0 ?
+         fes.GetTraceElement(0, fes.GetMesh()->GetFaceBaseGeometry(0))->GetDof()
+         : 0),
+     m(m),
+     nfdofs(nf*dof),
+     scatter_indices1(nf*dof),
+     scatter_indices2(m==L2FaceValues::DoubleValued?nf*dof:0),
+     offsets(ndofs+1),
+     gather_indices((m==L2FaceValues::DoubleValued? 2 : 1)*nf*dof)
 {
   if (nf==0) { return; }
   // If fespace == L2
-   const ParFiniteElementSpace &pfes =
-      static_cast<const ParFiniteElementSpace&>(this->fes);
-   const FiniteElement *fe = pfes.GetFE(0);
+   const FiniteElement *fe = fes.GetFE(0);
   const TensorBasisElement *tfe = dynamic_cast<const TensorBasisElement*>(fe);
   MFEM_VERIFY(tfe != NULL &&
               (tfe->GetBasisType()==BasisType::GaussLobatto ||
                tfe->GetBasisType()==BasisType::Positive),
               "Only Gauss-Lobatto and Bernstein basis are supported in "
               "ParL2FaceRestriction.");
-   MFEM_VERIFY(pfes.GetMesh()->Conforming(),
+   MFEM_VERIFY(fes.GetMesh()->Conforming(),
               "Non-conforming meshes not yet supported with partial assembly.");
   // Assuming all finite elements are using Gauss-Lobatto dofs
   height = (m==L2FaceValues::DoubleValued? 2 : 1)*vdim*nf*dof;
-   width = pfes.GetVSize();
+   width = fes.GetVSize();
   const bool dof_reorder = (e_ordering == ElementDofOrdering::LEXICOGRAPHIC);
   if (!dof_reorder)
   {
@@ -52,32 +63,32 @@ ParL2FaceRestriction::ParL2FaceRestriction(const ParFiniteElementSpace &fes,
   }
   if (dof_reorder && nf > 0)
   {
-      for (int f = 0; f < pfes.GetNF(); ++f)
+      for (int f = 0; f < fes.GetNF(); ++f)
      {
         const FiniteElement *fe =
-            pfes.GetTraceElement(f, pfes.GetMesh()->GetFaceBaseGeometry(f));
+            fes.GetTraceElement(f, fes.GetMesh()->GetFaceBaseGeometry(f));
         const TensorBasisElement* el =
            dynamic_cast<const TensorBasisElement*>(fe);
         if (el) { continue; }
         mfem_error("Finite element not suitable for lexicographic ordering");
      }
   }
-   const Table& e2dTable = pfes.GetElementToDofTable();
+   const Table& e2dTable = fes.GetElementToDofTable();
   const int* elementMap = e2dTable.GetJ();
   Array<int> faceMap1(dof), faceMap2(dof);
   int e1, e2;
   int inf1, inf2;
   int face_id1, face_id2;
   int orientation;
-   const int dof1d = pfes.GetFE(0)->GetOrder()+1;
-   const int elem_dofs = pfes.GetFE(0)->GetDof();
-   const int dim = pfes.GetMesh()->SpaceDimension();
+   const int dof1d = fes.GetFE(0)->GetOrder()+1;
+   const int elem_dofs = fes.GetFE(0)->GetDof();
+   const int dim = fes.GetMesh()->SpaceDimension();
   // Computation of scatter indices
   int f_ind=0;
-   for (int f = 0; f < pfes.GetNF(); ++f)
+   for (int f = 0; f < fes.GetNF(); ++f)
   {
-      pfes.GetMesh()->GetFaceElements(f, &e1, &e2);
-      pfes.GetMesh()->GetFaceInfos(f, &inf1, &inf2);
+      fes.GetMesh()->GetFaceElements(f, &e1, &e2);
+      fes.GetMesh()->GetFaceInfos(f, &inf1, &inf2);
      if (dof_reorder)
      {
         orientation = inf1 % 64;
@@ -125,7 +136,7 @@ ParL2FaceRestriction::ParL2FaceRestriction(const ParFiniteElementSpace &fes,
            {
               const int se2 = -1 - e2;
               Array<int> sharedDofs;
-               pfes.GetFaceNbrElementVDofs(se2, sharedDofs);
+               fes.GetFaceNbrElementVDofs(se2, sharedDofs);
               for (int d = 0; d < dof; ++d)
               {
                  const int pd = PermuteFaceL2(dim, face_id1, face_id2,
@@ -169,10 +180,10 @@ ParL2FaceRestriction::ParL2FaceRestriction(const ParFiniteElementSpace &fes,
      offsets[i] = 0;
   }
   f_ind = 0;
-   for (int f = 0; f < pfes.GetNF(); ++f)
+   for (int f = 0; f < fes.GetNF(); ++f)
   {
-      pfes.GetMesh()->GetFaceElements(f, &e1, &e2);
-      pfes.GetMesh()->GetFaceInfos(f, &inf1, &inf2);
+      fes.GetMesh()->GetFaceElements(f, &e1, &e2);
+      fes.GetMesh()->GetFaceInfos(f, &inf1, &inf2);
      if ((type==FaceType::Interior && (e2>=0 || (e2<0 && inf2>=0))) ||
          (type==FaceType::Boundary && e2<0 && inf2<0) )
      {
@@ -211,10 +222,10 @@ ParL2FaceRestriction::ParL2FaceRestriction(const ParFiniteElementSpace &fes,
      offsets[i] += offsets[i - 1];
   }
   f_ind = 0;
-   for (int f = 0; f < pfes.GetNF(); ++f)
+   for (int f = 0; f < fes.GetNF(); ++f)
   {
-      pfes.GetMesh()->GetFaceElements(f, &e1, &e2);
-      pfes.GetMesh()->GetFaceInfos(f, &inf1, &inf2);
+      fes.GetMesh()->GetFaceElements(f, &e1, &e2);
+      fes.GetMesh()->GetFaceInfos(f, &inf1, &inf2);
      if ((type==FaceType::Interior && (e2>=0 || (e2<0 && inf2>=0))) ||
          (type==FaceType::Boundary && e2<0 && inf2<0) )
      {
@@ -261,10 +272,8 @@ ParL2FaceRestriction::ParL2FaceRestriction(const ParFiniteElementSpace &fes,

 void ParL2FaceRestriction::Mult(const Vector& x, Vector& y) const
 {
-   const ParFiniteElementSpace &pfes =
-      static_cast<const ParFiniteElementSpace&>(this->fes);
   ParGridFunction x_gf;
-   x_gf.MakeRef(const_cast<ParFiniteElementSpace*>(&pfes),
+   x_gf.MakeRef(const_cast<ParFiniteElementSpace*>(&fes),
                const_cast<Vector&>(x), 0);
   x_gf.ExchangeFaceNbrData();

@@ -328,122 +337,34 @@ void ParL2FaceRestriction::Mult(const Vector& x, Vector& y) const
   }
 }

-static MFEM_HOST_DEVICE int AddNnz(const int iE, int *I, const int dofs)
+void ParL2FaceRestriction::MultTranspose(const Vector& x, Vector& y) const
 {
-   int val = AtomicAdd(I[iE],dofs);
-   return val;
-}
-
-void ParL2FaceRestriction::FillI(SparseMatrix &mat,
-                                 SparseMatrix &face_mat) const
-{
-   const int face_dofs = dof;
-   const int Ndofs = ndofs;
-   auto d_indices1 = scatter_indices1.Read();
-   auto d_indices2 = scatter_indices2.Read();
-   auto I = mat.ReadWriteI();
-   auto I_face = face_mat.ReadWriteI();
-   MFEM_FORALL(i, ne*elemDofs*vdim+1,
+   // Assumes all elements have the same number of dofs
+   const int nd = dof;
+   const int vd = vdim;
+   const bool t = byvdim;
+   const int dofs = nfdofs;
+   auto d_offsets = offsets.Read();
+   auto d_indices = gather_indices.Read();
+   auto d_x = Reshape(x.Read(), nd, vd, 2, nf);
+   auto d_y = Reshape(y.Write(), t?vd:ndofs, t?ndofs:vd);
+   MFEM_FORALL(i, ndofs,
   {
-      I_face[i] = 0;
-   });
-   MFEM_FORALL(fdof, nf*face_dofs,
-   {
-      const int f  = fdof/face_dofs;
-      const int iF = fdof%face_dofs;
-      const int iE1 = d_indices1[f*face_dofs+iF];
-      if (iE1 < Ndofs)
+      const int offset = d_offsets[i];
+      const int nextOffset = d_offsets[i + 1];
+      for (int c = 0; c < vd; ++c)
      {
-         for (int jF = 0; jF < face_dofs; jF++)
+         double dofValue = 0;
+         for (int j = offset; j < nextOffset; ++j)
         {
-            const int jE2 = d_indices2[f*face_dofs+jF];
-            if (jE2 < Ndofs)
-            {
-               AddNnz(iE1,I,1);
-            }
-            else
-            {
-               AddNnz(iE1,I_face,1);
-            }
-         }
-      }
-      const int iE2 = d_indices2[f*face_dofs+iF];
-      if (iE2 < Ndofs)
-      {
-         for (int jF = 0; jF < face_dofs; jF++)
-         {
-            const int jE1 = d_indices1[f*face_dofs+jF];
-            if (jE1 < Ndofs)
-            {
-               AddNnz(iE2,I,1);
-            }
-            else
-            {
-               AddNnz(iE2,I_face,1);
-            }
-         }
-      }
-   });
-}
-
-void ParL2FaceRestriction::FillJAndData(const Vector &ea_data,
-                                        SparseMatrix &mat,
-                                        SparseMatrix &face_mat) const
-{
-   const int face_dofs = dof;
-   const int Ndofs = ndofs;
-   auto d_indices1 = scatter_indices1.Read();
-   auto d_indices2 = scatter_indices2.Read();
-   auto mat_fea = Reshape(ea_data.Read(), face_dofs, face_dofs, 2, nf);
-   auto I = mat.ReadWriteI();
-   auto I_face = face_mat.ReadWriteI();
-   auto J = mat.WriteJ();
-   auto J_face = face_mat.WriteJ();
-   auto Data = mat.WriteData();
-   auto Data_face = face_mat.WriteData();
-   MFEM_FORALL(fdof, nf*face_dofs,
-   {
-      const int f  = fdof/face_dofs;
-      const int iF = fdof%face_dofs;
-      const int iE1 = d_indices1[f*face_dofs+iF];
-      if (iE1 < Ndofs)
-      {
-         for (int jF = 0; jF < face_dofs; jF++)
-         {
-            const int jE2 = d_indices2[f*face_dofs+jF];
-            if (jE2 < Ndofs)
-            {
-               const int offset = AddNnz(iE1,I,1);
-               J[offset] = jE2;
-               Data[offset] = mat_fea(jF,iF,1,f);
-            }
-            else
-            {
-               const int offset = AddNnz(iE1,I_face,1);
-               J_face[offset] = jE2-Ndofs;
-               Data_face[offset] = mat_fea(jF,iF,1,f);
-            }
-         }
-      }
-      const int iE2 = d_indices2[f*face_dofs+iF];
-      if (iE2 < Ndofs)
-      {
-         for (int jF = 0; jF < face_dofs; jF++)
-         {
-            const int jE1 = d_indices1[f*face_dofs+jF];
-            if (jE1 < Ndofs)
-            {
-               const int offset = AddNnz(iE2,I,1);
-               J[offset] = jE1;
-               Data[offset] = mat_fea(jF,iF,0,f);
-            }
-            else
-            {
-               const int offset = AddNnz(iE2,I_face,1);
-               J_face[offset] = jE1-Ndofs;
-               Data_face[offset] = mat_fea(jF,iF,0,f);
-            }
+            int idx_j = d_indices[j];
+            bool isE1 = idx_j < dofs;
+            idx_j = isE1 ? idx_j : idx_j - dofs;
+            dofValue +=  isE1 ?
+            d_x(idx_j % nd, c, 0, idx_j / nd)
+            :d_x(idx_j % nd, c, 1, idx_j / nd);
         }
+         d_y(t?c:i,t?i:c) += dofValue;
      }
   });
 }
@@ -26,21 +26,28 @@ class ParFiniteElementSpace;
 /// Operator that extracts Face degrees of freedom in parallel.
 /** Objects of this type are typically created and owned by FiniteElementSpace
    objects, see FiniteElementSpace::GetFaceRestriction(). */
-class ParL2FaceRestriction : public L2FaceRestriction
+class ParL2FaceRestriction : public Operator
 {
+protected:
+   const ParFiniteElementSpace &fes;
+   const int nf;
+   const int vdim;
+   const bool byvdim;
+   const int ndofs;
+   const int dof;
+   const L2FaceValues m;
+   const int nfdofs;
+   Array<int> scatter_indices1;
+   Array<int> scatter_indices2;
+   Array<int> offsets;
+   Array<int> gather_indices;
+
 public:
   ParL2FaceRestriction(const ParFiniteElementSpace&, ElementDofOrdering,
                        FaceType type,
                        L2FaceValues m = L2FaceValues::DoubleValued);
   void Mult(const Vector &x, Vector &y) const;
-   /** Fill the I array of SparseMatrix corresponding to the sparsity pattern
-       given by this ParL2FaceRestriction. */
-   void FillI(SparseMatrix &mat, SparseMatrix &face_mat) const;
-   /** Fill the J and Data arrays of SparseMatrix corresponding to the sparsity
-       pattern given by this ParL2FaceRestriction, and the values of ea_data. */
-   void FillJAndData(const Vector &ea_data,
-                     SparseMatrix &mat,
-                     SparseMatrix &face_mat) const;
+   void MultTranspose(const Vector &x, Vector &y) const;
 };

 }
--- a/Show More
+++ b/Show More