including an obstacle problem variant wherein the essential dofs are not given to the optimizer as free variables that it must choose the values for.

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

@@ -162,7 +162,7 @@ jobs:
 
     - name: get hypre
       if: matrix.mpi == 'par' && steps.hypre-cache.outputs.cache-hit != 'true' && matrix.os != 'windows-latest'
-      uses: mfem/github-actions/build-hypre@v2.4
+      uses: mfem/github-actions/build-hypre@v2.2
       with:
         archive: ${{ env.HYPRE_ARCHIVE }}
         dir: ${{ env.HYPRE_TOP_DIR }}
@@ -171,7 +171,7 @@ jobs:
 
     - name: get hypre (Windows)
       if: matrix.mpi == 'par' && steps.hypre-cache.outputs.cache-hit != 'true' && matrix.os == 'windows-latest'
-      uses: mfem/github-actions/build-hypre@v2.4
+      uses: mfem/github-actions/build-hypre@v2.2
       with:
         archive: ${{ env.HYPRE_ARCHIVE }}
         dir: ${{ env.HYPRE_TOP_DIR }}
@@ -190,7 +190,7 @@ jobs:
 
     - name: install metis
       if: matrix.mpi == 'par' && matrix.os != 'windows-latest' && steps.metis-cache.outputs.cache-hit != 'true'
-      uses: mfem/github-actions/build-metis@v2.4
+      uses: mfem/github-actions/build-metis@v2.2
       with:
         archive: ${{ env.METIS_ARCHIVE }}
         dir: ${{ env.METIS_TOP_DIR }}
@@ -217,7 +217,7 @@ jobs:
 
     # MFEM build and test
     - name: build
-      uses: mfem/github-actions/build-mfem@v2.4
+      uses: mfem/github-actions/build-mfem@v2.3
       env:
         VCPKG_DEFAULT_BINARY_CACHE: ${{ github.workspace }}/vcpkg_cache
       with:
@@ -263,15 +263,13 @@ jobs:
       if: matrix.build-system == 'cmake' && matrix.target == 'opt' && matrix.os != 'ubuntu-latest'
       run: |
         CTEST_CONFIG="Release"
-        cd ${{ env.MFEM_TOP_DIR }}/build && \
-        ctest --output-on-failure -C ${CTEST_CONFIG} || \
-        ctest --rerun-failed --output-on-failure -C ${CTEST_CONFIG}
+        cd ${{ env.MFEM_TOP_DIR }}/build && ctest --output-on-failure -C ${CTEST_CONFIG}
       shell: bash
 
     # Code coverage (process and upload reports)
     - name: codecov
       if: matrix.codecov == 'YES'
-      uses: mfem/github-actions/upload-coverage@v2.4
+      uses: mfem/github-actions/upload-coverage@v2.2
       with:
         name: ${{ matrix.os }}-${{ matrix.build-system }}-${{ matrix.target }}-${{ matrix.mpi }}-${{ matrix.hypre-target }}
         project_dir: ${{ env.MFEM_TOP_DIR }}

.github/workflows/mfem-analysis.yml

@@ -57,7 +57,7 @@ jobs:
 
     - name: Get Hypre
       if: steps.hypre-cache.outputs.cache-hit != 'true'
-      uses: mfem/github-actions/build-hypre@v2.4
+      uses: mfem/github-actions/build-hypre@v2.2
       with:
         archive: ${{ env.HYPRE_ARCHIVE }}
         dir: ${{ env.HYPRE_TOP_DIR }}
@@ -72,14 +72,14 @@ jobs:
 
     - name: Install Metis
       if: steps.metis-cache.outputs.cache-hit != 'true'
-      uses: mfem/github-actions/build-metis@v2.4
+      uses: mfem/github-actions/build-metis@v2.2
       with:
         archive: ${{ env.METIS_ARCHIVE }}
         dir: ${{ env.METIS_TOP_DIR }}
 
     # MFEM build and test
     - name: build-mfem
-      uses: mfem/github-actions/build-mfem@v2.4
+      uses: mfem/github-actions/build-mfem@v2.2
       with:
         os: ${{ runner.os }}
         target: opt

.github/workflows/mfem-sanitizer.yml

@@ -1,70 +0,0 @@
-# Copyright (c) 2010-2023, 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.
-
-name: "Sanitizer"
-
-permissions:
-  actions: write
-
-on:
-  push:
-    branches:
-      - master
-      - next
-  pull_request:
-  workflow_dispatch:
-
-jobs:
-  Serial:
-    runs-on: ubuntu-latest
-
-    steps:
-    - name: Cancel Previous Runs
-      uses: styfle/cancel-workflow-action@0.11.0
-      with:
-        access_token: ${{ github.token }}
-
-    - name: MFEM Checkout
-      uses: actions/checkout@v3
-      with:
-        path: mfem
-
-    - name: MFEM Build
-      uses: mfem/github-actions/build-mfem@v2.4
-      with:
-        os: ${{ runner.os }}
-        target: opt
-        mpi: seq
-        hypre-dir: unused-hypre-dir
-        metis-dir: unused-metis-dir
-        mfem-dir: mfem
-        build-system: make
-        library-only: false
-        config-options:
-          CXX="clang++-14"
-          CXXFLAGS="-g -O1 -std=c++11
-                    -fsanitize=address
-                    -fno-omit-frame-pointer
-                    -fsanitize-address-use-after-scope"
-
-    - name: MFEM Info
-      working-directory: mfem
-      run: make info
-
-    - name: MFEM Sanitize
-      working-directory: mfem
-      run:
-        ASAN_OPTIONS="detect_leaks=1,
-                      strict_init_order=1,
-                      strict_string_checks=1,
-                      check_initialization_order=1,
-                      detect_stack_use_after_return=1"
-        make test

.gitignore

@@ -29,7 +29,6 @@ CMakeFiles/
 config/_config.hpp
 config/config.mk
 config/sample-runs-build.log
-config/user.cmake
 config/user.mk
 doc/CodeDocumentation.conf
 doc/CodeDocumentation.html
@@ -215,7 +214,6 @@ miniapps/meshing/pmesh-fitting
 miniapps/meshing/minimal-surface
 miniapps/meshing/pminimal-surface
 miniapps/meshing/polar-nc
-miniapps/meshing/mesh-quality
 miniapps/meshing/mobius-strip.mesh
 miniapps/meshing/klein-bottle.mesh
 miniapps/meshing/toroid-*.mesh
@@ -318,28 +316,12 @@ miniapps/solvers/ParaView
 miniapps/solvers/mesh.*
 miniapps/solvers/sol.*
 
-miniapps/hdiv-linear-solver/darcy
-miniapps/hdiv-linear-solver/grad_div
-
 miniapps/parelag/MultilevelHcurlHdivSolver
 miniapps/parelag/*.mesh
 
 miniapps/multidomain/multidomain
 miniapps/hooke/hooke
 
-miniapps/dpg/diffusion
-miniapps/dpg/pdiffusion
-miniapps/dpg/convection-diffusion
-miniapps/dpg/pconvection-diffusion
-miniapps/dpg/acoustics
-miniapps/dpg/pacoustics
-miniapps/dpg/maxwell
-miniapps/dpg/pmaxwell
-miniapps/dpg/ParaView
-
-miniapps/spde/generate_random_field
-miniapps/spde/ParaView
-
 # Unit test binary and outputs
 tests/unit/output_meshes
 tests/unit/unit_tests
@@ -352,8 +334,6 @@ tests/unit/tmop_pa_tests_*
 tests/unit/ptmop_pa_tests_*
 tests/unit/ceed_tests
 tests/unit/debug_device_tests
-tests/unit/parallel_in_serial.mesh
-tests/unit/parallel_in_serial.gf
 
 # Benchmark binaries
 tests/benchmarks/bench_ceed

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

@@ -22,10 +22,12 @@
         date
         echo "Waiting to acquire lock on '$PWD/autotest.lock' ..."
         # try to get an exclusive lock on fd 9 (autotest.lock) repeating the try
-        # every 5 seconds; we may want to add a counter for the number of
+        # every 5 seconds; simply using no timeout, i.e. 'flock 9', causes the
+        # command to hang indefinitely sometimes, so we use the timeout & retry
+        # as a workaround; we may want to add a counter for the number of
         # retries to interrupt a potential infinite loop
-        while ! flock -n 9; do
-          sleep 5
+        while ! flock -w 5 9; do
+          true
         done
         echo "Acquired lock on '$PWD/autotest.lock'"
         date
@@ -55,10 +57,12 @@
         date
         echo "Waiting to acquire lock on '$PWD/autotest.lock' ..."
         # try to get an exclusive lock on fd 9 (autotest.lock) repeating the try
-        # every 5 seconds; we may want to add a counter for the number of
+        # every 5 seconds; simply using no timeout, i.e. 'flock 9', causes the
+        # command to hang indefinitely sometimes, so we use the timeout & retry
+        # as a workaround; we may want to add a counter for the number of
         # retries to interrupt a potential infinite loop
-        while ! flock -n 9; do
-          sleep 5
+        while ! flock -w 5 9; do
+          true
         done
         echo "Acquired lock on '$PWD/autotest.lock'"
         date

.gitlab/configs/setup-baseline.yml

@@ -47,10 +47,12 @@ setup_baseline:
         date
         echo "Waiting to acquire lock on '$PWD/autotest.lock' ..."
         # try to get an exclusive lock on fd 9 (autotest.lock) repeating the try
-        # every 5 seconds; we may want to add a counter for the number of
+        # every 5 seconds; simply using no timeout, i.e. 'flock 9', causes the
+        # command to hang indefinitely sometimes, so we use the timeout & retry
+        # as a workaround; we may want to add a counter for the number of
         # retries to interrupt a potential infinite loop
-        while ! flock -n 9; do
-          sleep 5
+        while ! flock -w 5 9; do
+          true
         done
         echo "Acquired lock on '$PWD/autotest.lock'"
         date

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

@@ -35,11 +35,13 @@ setup:
       (
         date
         echo "Waiting to acquire lock on '$PWD/mfem-data.lock' ..."
-        # try to get an exclusive lock on fd 9 (mfem-data.lock) repeating the
-        # try every 5 seconds; we may want to add a counter for the number of
+        # try to get an exclusive lock on fd 9 (mfem-data.lock) repeating the try
+        # every 5 seconds; simply using no timeout, i.e. 'flock 9', causes the
+        # command to hang indefinitely sometimes, so we use the timeout & retry
+        # as a workaround; we may want to add a counter for the number of
         # retries to interrupt a potential infinite loop
-        while ! flock -n 9; do
-          sleep 5
+        while ! flock -w 5 9; do
+          true
         done
         echo "Acquired lock on '$PWD/mfem-data.lock'"
         date
@@ -67,10 +69,12 @@ setup:
         date
         echo "Waiting to acquire lock on '$PWD/autotest.lock' ..."
         # try to get an exclusive lock on fd 9 (autotest.lock) repeating the try
-        # every 5 seconds; we may want to add a counter for the number of
+        # every 5 seconds; simply using no timeout, i.e. 'flock 9', causes the
+        # command to hang indefinitely sometimes, so we use the timeout & retry
+        # as a workaround; we may want to add a counter for the number of
         # retries to interrupt a potential infinite loop
-        while ! flock -n 9; do
-          sleep 5
+        while ! flock -w 5 9; do
+          true
         done
         echo "Acquired lock on '$PWD/autotest.lock'"
         date

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

@@ -14,14 +14,14 @@ stages:
   - build_and_test
   - report
 
-opt_mpi_cuda_xl_16_1_1_12:
+opt_mpi_cuda_xl_16_1_1_8:
   variables:
-    SPEC: "%xl@16.1.1.12 +mpi +cuda cuda_arch=70"
+    SPEC: "%xl@16.1.1.8 +mpi +cuda cuda_arch=70"
   extends: .build_and_test_on_lassen
 
 opt_mpi_cuda_hypre_cuda_xl:
   variables:
-    SPEC: "%xl@16.1.1.12 +mpi +cuda cuda_arch=70 ^hypre+cuda~shared cuda_arch=70"
+    SPEC: "%xl@16.1.1.8 +mpi +cuda cuda_arch=70 ^hypre+cuda~shared cuda_arch=70"
   extends: .build_and_test_on_lassen
 
 # Jobs report

.gitlab/quartz-baseline.yml

@@ -51,8 +51,6 @@ cleanup:
   script:
     - echo "BUILD_ROOT=${BUILD_ROOT}"
     - rm -rf "${BUILD_ROOT}" || true
-    - echo "CI_PROJECT_DIR=${CI_PROJECT_DIR}"
-    - make -C "${CI_PROJECT_DIR}" distclean
 
 report_baseline:
   extends: [.on_quartz]
@@ -68,10 +66,12 @@ report_baseline:
         date
         echo "Waiting to acquire lock on '$PWD/autotest.lock' ..."
         # try to get an exclusive lock on fd 9 (autotest.lock) repeating the try
-        # every 5 seconds; we may want to add a counter for the number of
+        # every 5 seconds; simply using no timeout, i.e. 'flock 9', causes the
+        # command to hang indefinitely sometimes, so we use the timeout & retry
+        # as a workaround; we may want to add a counter for the number of
         # retries to interrupt a potential infinite loop
-        while ! flock -n 9; do
-          sleep 5
+        while ! flock -w 5 9; do
+          true
         done
         echo "Acquired lock on '$PWD/autotest.lock'"
         date
@@ -82,14 +82,12 @@ report_baseline:
         rundir="${MACHINE_NAME}/$(date +%Y-%m-%d)-gitlab-${BASELINE_TEST}-${CI_COMMIT_REF_SLUG}"
         rundir=$(${CI_PROJECT_DIR}/.gitlab/scripts/safe_create_rundir ${rundir})
         cp ${CI_PROJECT_DIR}/${ARTIFACTS_DIR}/* ${rundir}
+        # We create an autotest-email.html file, because that's how we signal that there was a diff (temporary).
+        if [[ -f ${rundir}/${BASELINE_TEST}.err ]]; then
+          cp ${rundir}/${BASELINE_TEST}.err ${rundir}/autotest-email.html
+        fi
         printf "%s\n" "" "Pipeline URL:" "$CI_PIPELINE_URL" \
           >> ${rundir}/pipeline.txt
-        # We create an autotest-email.html file, because that's how we signal
-        # that there was an error / diff (temporary).
-        if [[ -f ${rundir}/${BASELINE_TEST}.err ]] || \
-           [[ -f ${rundir}/${BASELINE_TEST}-${SYS_TYPE}.diff ]]; then
-          cp ${rundir}/pipeline.txt ${rundir}/autotest-email.html
-        fi
         msg="GitLab CI log for ${BASELINE_TEST} on ${MACHINE_NAME} ($(date +%Y-%m-%d))"
         if [[ "$AUTOTEST_COMMIT" != "NO" ]]; then
           git pull && \

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

@@ -27,39 +27,39 @@ allocate_resource:
   timeout: 6h
 
 # GitLab jobs for the Quartz machine at LLNL
-debug_ser_gcc_10:
+debug_ser_gcc_6_1_0:
   variables:
-    SPEC: "%gcc@10.3.1 +debug~mpi"
+    SPEC: "%gcc@6.1.0 +debug~mpi"
   extends: .build_and_test_on_quartz
 
-debug_par_gcc_10:
+debug_par_gcc_6_1_0:
   variables:
-    SPEC: "%gcc@10.3.1 +debug+mpi"
+    SPEC: "%gcc@6.1.0 +debug+mpi"
   extends: .build_and_test_on_quartz
 
-opt_ser_gcc_10:
+opt_ser_gcc_6_1_0:
   variables:
-    SPEC: "%gcc@10.3.1 ~mpi"
+    SPEC: "%gcc@6.1.0 ~mpi"
   extends: .build_and_test_on_quartz
 
-opt_par_gcc_10:
+opt_par_gcc_6_1_0:
   variables:
-    SPEC: "%gcc@10.3.1"
+    SPEC: "%gcc@6.1.0"
   extends: .build_and_test_on_quartz
 
-opt_par_gcc_10_sundials:
+opt_par_gcc_6_1_0_sundials:
   variables:
-    SPEC: "%gcc@10.3.1 +sundials"
+    SPEC: "%gcc@6.1.0 +sundials"
   extends: .build_and_test_on_quartz
 
-opt_par_gcc_10_petsc:
+opt_par_gcc_6_1_0_petsc:
   variables:
-    SPEC: "%gcc@10.3.1 +petsc ^petsc+mumps~superlu-dist"
+    SPEC: "%gcc@6.1.0 +petsc ^petsc+mumps~superlu-dist"
   extends: .build_and_test_on_quartz
 
-opt_par_gcc_10_pumi:
+opt_par_gcc_6_1_0_pumi:
   variables:
-    SPEC: "%gcc@10.3.1 +pumi"
+    SPEC: "%gcc@6.1.0 +pumi"
   extends: .build_and_test_on_quartz
 
 # Release

.gitlab/scripts/baseline

@@ -42,34 +42,47 @@ fi
 # post
 mkdir ${artifacts_path}
 
-status=0
-if [[ -f ${BASELINE_TEST}.out ]]; then
-  cp ${BASELINE_TEST}.out ${artifacts_path}
-fi
-if [[ -s ${glob_err} ]]; then
-  echo "ERROR during ${BASELINE_TEST} execution"
-  echo "Here is the ${glob_err} file content"
+if [[ -s ${glob_err} ]]
+then
+  echo "ERROR during ${BASELINE_TEST} execution";
+  echo "Here is the ${glob_err} file content";
   cat ${glob_err}
   cp ${glob_err} ${artifacts_path}/${glob_err}
-  status=1
-fi
-if [[ -f ${base_patch} ]]; then
+  exit 1;
+elif [[ ! -f ${base_patch} && ! -f ${base_out} ]]
+then
+  echo "Something went WRONG in ${BASELINE_TEST}:";
+  echo "Either ${base_patch} or ${base_out} should exists";
+  exit 1;
+elif [[ -f ${base_patch} ]]
+then
   echo "${BASELINE_TEST}: Differences found, patch generated"
   cp ${base_patch} ${artifacts_path}/${base_patch}
-elif [[ -f ${base_out} ]]; then
+elif [[ -f ${base_out} ]]
+then
   echo "${BASELINE_TEST}: Differences found, replacement file generated"
   cp ${base_out} ${artifacts_path}/${base_out}
 fi
+
+if [[ -f ${BASELINE_TEST}.out ]]; then
+  cp ${BASELINE_TEST}.out ${artifacts_path}
+fi
+
 # base_diff won't even exist if there is no difference.
-if [[ -f ${base_diff} ]]; then
+if [[ -f ${base_diff} ]]
+then
   echo "${BASELINE_TEST}: Relevant differences (filtered diff) ..."
   cat ${base_diff}
   cp ${base_diff} ${artifacts_path}/${base_diff}
-  status=1
+  # We create a .err file, because that's how we signal that there was a diff.
+  cp ${base_diff} ${artifacts_path}/gitlab-${BASELINE_TEST}-${MACHINE_NAME}.err
 fi
-if [[ $status -eq 0 ]]; then
+
+if [[ ! -s ${base_diff} ]]
+then
   echo "${BASELINE_TEST}: PASSED"
+  true
 else
   echo "${BASELINE_TEST}: FAILED"
+  false
 fi
-exit $status

CHANGELOG

@@ -11,39 +11,14 @@
 Version 4.5.3 (development)
 ===========================
 
-New and updated examples and miniapps
--------------------------------------
-- Added a new miniapp, Mesh Quality, for evaluating mesh quality using size,
-  skewness, and aspect-ratio computed from the Jacobian of the transformation.
-
-- Added a new miniapp for interface and boundary fitting to implicit domains
-  defined using level-set functions. See miniapps/meshing/pmesh-fitting.cpp
-
-- Added new Discontinuous Petrov-Galerkin (DPG) miniapp which includes serial
-  and parallel examples for diffusion, convection-diffusion, acoustics and
-  Maxwell equations. The miniapp includes new classes such as (Par)DPGWeakForm,
-  (Par)ComplexDPGWeakForm and (Complex)BlockStaticCondensation. Three new
-  integrators are added in support of DPG systems: TraceIntegrator,
-  NormalTraceIntegrator and TangentTraceIntegrator.
-
-- Added a new H(div) solvers miniapp in miniapps/hdiv-linear-solver,
-  demonstrating the use of a matrix-free saddle-point solver methodology,
-  suitable for high-order discretizations and for GPU acceleration. Examples
-  illustrating the solution of Darcy and grad-div problems are included.
-
-- Added a random refinement option to the mesh-explorer miniapp to assist users
-  in experimenting with nonconforming meshes.
-
-- Moved the distance solver methods from miniapps/shifted to miniapps/common.
-
-Meshing improvements
---------------------
 - Added new methods in the Mesh class to set and get attributes on NURBS patches
   and patch boundaries.
 
-- TMOP improvement: added asymptotically-balanced compound metrics 90, 94, 328,
-  338. Added the tmop-metric-magnitude tool for tracking how metrics change
-  under geometric perturbations.
+New and updated examples and miniapps
+-------------------------------------
+- Added a miniapp pmesh-fitting in miniapps/meshing for interface and boundary fitting to implicit domains defined using level-set functions.
+
+- Moved the distance solver methods from miniapps/shifted to miniapps/common. 
 
 Discretization improvements
 ---------------------------
@@ -52,33 +27,6 @@ Discretization improvements
 
 - VectorFEBoundaryFluxLFIntegrator is now supported on device/GPU.
 
-- Added support for p-refined meshes in FindPointsGSLIB.
-
-Linear and nonlinear solvers
-----------------------------
-- Updated interface to MUMPS direct solver to support multiple right-hand
-  sides, block low-rank compression, builds using 64-bit integers, and other
-  improvements.
-
-- Added an interface to the MKL Pardiso sparse direct solver developed by Intel.
-  This interface provides a serial (OpenMP shared memory) version of Pardiso for
-  use with SparseMatrix. This complements the existing parallel (MPI distributed
-  memory) version already available through the CPardiso MFEM integration.
-
-Integrations, testing and documentation
----------------------------------------
-- Added an address sanitizer GitHub action for a serial build/test on Ubuntu,
-  based on Clang/LLVM (https://clang.llvm.org/docs/AddressSanitizer.html).
-
-Miscellaneous
--------------
-- Improved lambda body debugging with the addition of mfem::forall functions.
-  These functions can take the place of the MFEM_FORALL macros, which have been
-  preserved for backwards compatibility.
-
-- Reorganized files for bilinear form, linear form, and nonlinear form integrators
-  in the fem/integ/ subdirectory.
-
 
 Version 4.5.2, released on March 23, 2023
 =========================================

CMakeLists.txt

@@ -82,7 +82,7 @@ if (MFEM_USE_CONDUIT OR
    #  * find_package(PETSc REQUIRED)
    set(XSDK_ENABLE_C ON)
 endif()
-if (MFEM_USE_STRUMPACK OR MFEM_USE_MUMPS)
+if (MFEM_USE_STRUMPACK)
   # Just needed to find the MPI_Fortran libraries to link with
   set(XSDK_ENABLE_Fortran ON)
 endif()
@@ -333,7 +333,6 @@ endif()
 if (MFEM_USE_MUMPS)
   if (MFEM_USE_MPI)
     find_package(MUMPS REQUIRED mumps_common pord)
-    set(MFEM_MUMPS_VERSION ${MUMPS_VERSION})
   else()
     message(FATAL_ERROR " *** MUMPS requires that MPI be enabled.")
   endif()
@@ -467,18 +466,12 @@ if (MFEM_USE_ADIOS2)
    find_package(ADIOS2 REQUIRED)
 endif()
 
-# MKL CPardiso
 if (MFEM_USE_MKL_CPARDISO)
   if (MFEM_USE_MPI)
    find_package(MKL_CPARDISO REQUIRED MKL_SEQUENTIAL MKL_LP64 MKL_MPI_WRAPPER)
   endif()
 endif()
 
-# MKL Pardiso
-if (MFEM_USE_MKL_PARDISO)
-   find_package(MKL_PARDISO REQUIRED MKL_SEQUENTIAL MKL_LP64)
-endif()
-
 # PARELAG
 if (MFEM_USE_PARELAG)
    find_package(PARELAG REQUIRED)
@@ -528,8 +521,8 @@ find_package(Threads REQUIRED)
 set(MFEM_TPLS OPENMP HYPRE LAPACK BLAS SuperLUDist STRUMPACK METIS SuiteSparse
     SUNDIALS PETSC SLEPC MUMPS AXOM FMS CONDUIT Ginkgo GNUTLS GSLIB
     NETCDF MPFR PUMI HIOP POSIXCLOCKS MFEMBacktrace ZLIB OCCA CEED RAJA UMPIRE
-    ADIOS2 CUBLAS CUSPARSE MKL_CPARDISO MKL_PARDISO AMGX CALIPER CODIPACK
-    BENCHMARK PARELAG MPI_CXX HIP HIPSPARSE MOONOLITH BLITZ ALGOIM ENZYME)
+    ADIOS2 CUBLAS CUSPARSE MKL_CPARDISO AMGX CALIPER CODIPACK BENCHMARK PARELAG
+    MPI_CXX HIP HIPSPARSE MOONOLITH BLITZ ALGOIM ENZYME)
 
 # Add all *_FOUND libraries in the variable TPL_LIBRARIES.
 set(TPL_LIBRARIES "")

CONTRIBUTING.md

@@ -121,7 +121,6 @@ The MFEM source code has the following structure:
   ├── fem
   │   ├── ceed
   │   ├── fe
-  │   ├── integ
   │   ├── lor
   │   ├── moonolith
   │   ├── qinterp
@@ -137,7 +136,6 @@ The MFEM source code has the following structure:
   │   ├── common
   │   ├── electromagnetics
   │   ├── gslib
-  │   ├── hdiv-linear-solver
   │   ├── hooke
   │   ├── meshing
   │   ├── mtop
@@ -211,7 +209,7 @@ device/host memory manager.
 - The main device-relevant classes and sources are:
   + [`Device`](https://docs.mfem.org/html/device_8hpp.html)
   + [`MemoryManager`](https://docs.mfem.org/html/mem_manager_8hpp.html)
-  + the [`mfem::forall`](https://docs.mfem.org/html/forall_8hpp.html) function
+  + the [`MFEM_FORALL`](https://docs.mfem.org/html/forall_8hpp.html) macro
   + the [`cuda.hpp`](https://docs.mfem.org/html/cuda_8hpp.html) and [`occa.hpp`](https://docs.mfem.org/html/occa_8hpp.html) files
 
 #### Utilities, building and documentation

config/cmake/config.hpp.in

@@ -80,102 +80,97 @@
 // Internal MFEM option: enable group/batch allocation for some small objects.
 #cmakedefine MFEM_USE_MEMALLOC
 
-// Which library functions to use in class StopWatch for measuring time.
-// For a list of the available options, see INSTALL.
-// If not defined, an option is selected automatically.
-#cmakedefine MFEM_TIMER_TYPE @MFEM_TIMER_TYPE@
-
-// Enable MFEM functionality based on the SUNDIALS libraries.
-#cmakedefine MFEM_USE_SUNDIALS
-
 // Enable MFEM functionality based on the SuiteSparse library.
 #cmakedefine MFEM_USE_SUITESPARSE
 
 // Enable MFEM functionality based on the SuperLU_DIST library.
 #cmakedefine MFEM_USE_SUPERLU
-#cmakedefine MFEM_USE_SUPERLU5
 
 // Enable MFEM functionality based on the MUMPS library.
 #cmakedefine MFEM_USE_MUMPS
-#cmakedefine MFEM_MUMPS_VERSION @MFEM_MUMPS_VERSION@
 
 // Enable MFEM functionality based on the STRUMPACK library.
 #cmakedefine MFEM_USE_STRUMPACK
 
-// Enable functionality based on the Ginkgo library.
+// Enable functionality based on the Ginkgo library
 #cmakedefine MFEM_USE_GINKGO
 
-// Enable MFEM functionality based on the AmgX library.
+// Enable MFEM functionality based on the AmgX library
 #cmakedefine MFEM_USE_AMGX
 
-// Enable secure socket streams based on the GNUTLS library.
+// Enable MFEM functionality based on the GnuTLS library
 #cmakedefine MFEM_USE_GNUTLS
 
-// Enable Sidre support.
-#cmakedefine MFEM_USE_SIDRE
-
-// Enable the use of SIMD in the high performance templated classes.
-#cmakedefine MFEM_USE_SIMD
-
-// Enable FMS support.
-#cmakedefine MFEM_USE_FMS
-
-// Enable Conduit support.
-#cmakedefine MFEM_USE_CONDUIT
-
-// Enable functionality based on the NetCDF library (reading CUBIT files).
-#cmakedefine MFEM_USE_NETCDF
-
-// Enable functionality based on the PETSc library.
-#cmakedefine MFEM_USE_PETSC
-
-// Enable functionality based on the SLEPc library.
-#cmakedefine MFEM_USE_SLEPC
-
-// Enable functionality based on the MPFR library.
-#cmakedefine MFEM_USE_MPFR
-
-// Enable MFEM functionality based on the PUMI library.
-#cmakedefine MFEM_USE_PUMI
-
-// Enable Moonolith-based general interpolation between finite element spaces.
-#cmakedefine MFEM_USE_MOONOLITH
-
-// Enable MFEM functionality based on the HIOP library.
-#cmakedefine MFEM_USE_HIOP
-
-// Enable MFEM functionality based on the GSLIB library.
+// Enable MFEM functionality based on the GSLIB library
 #cmakedefine MFEM_USE_GSLIB
 
-// Build the NVIDIA GPU/CUDA-enabled version of the MFEM library.
+// Enable MFEM functionality based on the NetCDF library
+#cmakedefine MFEM_USE_NETCDF
+
+// 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
+
+// Enable the use of SIMD in the high performance templated classes
+#cmakedefine MFEM_USE_SIMD
+
+// Enable MFEM functionality based on the FMS library
+#cmakedefine MFEM_USE_FMS
+
+// Enable MFEM functionality based on Conduit
+#cmakedefine MFEM_USE_CONDUIT
+
+// Enable MFEM functionality based on the PUMI library
+#cmakedefine MFEM_USE_PUMI
+
+// Enable MFEM functionality based on the Moonolith library
+#cmakedefine MFEM_USE_MOONOLITH
+
+// Enable MFEM functionality based on the HiOp library
+#cmakedefine MFEM_USE_HIOP
+
+// Build the GPU/CUDA-enabled version of the MFEM library.
 // Requires a CUDA compiler (nvcc).
 #cmakedefine MFEM_USE_CUDA
 
-// Build the AMD GPU/HIP-enabled version of the MFEM library.
+// Build the HIP-enabled version of the MFEM library.
 // Requires a HIP compiler (hipcc).
 #cmakedefine MFEM_USE_HIP
 
-// Enable functionality based on the RAJA library.
+// Enable MFEM functionality based on the RAJA library
 #cmakedefine MFEM_USE_RAJA
 
-// Enable functionality based on the OCCA library.
+// Enable MFEM functionality based on the OCCA library
 #cmakedefine MFEM_USE_OCCA
 
-// Enable functionality based on the libCEED library.
+// Enable MFEM functionality based on the libCEED library
 #cmakedefine MFEM_USE_CEED
 
-// Enable functionality based on the Caliper library.
-#cmakedefine MFEM_USE_CALIPER
-
-// Enable functionality based on the Algoim library.
-#cmakedefine MFEM_USE_ALGOIM
-
-// Enable functionality based on the Umpire library.
+// Enable MFEM functionality based on the Umpire library
 #cmakedefine MFEM_USE_UMPIRE
 
-// Enable IO functionality based on the ADIOS2 library.
+// Enable MFEM functionality based on the ADIOS2 library
 #cmakedefine MFEM_USE_ADIOS2
 
+// Enable MFEM functionality based on the Caliper library
+#cmakedefine MFEM_USE_CALIPER
+
+// Enable MFEM functionality based on the Algoim library
+#cmakedefine MFEM_USE_ALGOIM
+
+// Which library functions to use in class StopWatch for measuring time.
+// For a list of the available options, see INSTALL.
+// If not defined, an option is selected automatically.
+#define MFEM_TIMER_TYPE @MFEM_TIMER_TYPE@
+
+// Enable MFEM functionality based on the SUNDIALS libraries.
+#cmakedefine MFEM_USE_SUNDIALS
+
 // Version of HYPRE used for building MFEM.
 #cmakedefine MFEM_HYPRE_VERSION @MFEM_HYPRE_VERSION@
 
@@ -186,13 +181,13 @@
 // Enable interface to the MKL CPardiso library.
 #cmakedefine MFEM_USE_MKL_CPARDISO
 
-// Use forward mode for automatic differentiation.
+// Use forward mode for automatic differentiation
 #cmakedefine MFEM_USE_ADFORWARD
 
-// Enable the use of the CoDiPack library for AD.
+// Enable the use of the CoDiPack library for AD
 #cmakedefine MFEM_USE_CODIPACK
 
-// Enable functionality based on the Google Benchmark library.
+// Enable MFEM functionality based on the Google Benchmark library.
 #cmakedefine MFEM_USE_BENCHMARK
 
 // Enable Enzyme for AD

config/cmake/modules/FindMKL_PARDISO.cmake

@@ -1,27 +0,0 @@
-# Copyright (c) 2010-2023, 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.
-
-# Defines the following variables:
-#   - MKL_PARDISO_FOUND
-#   - MKL_PARDISO_LIBRARIES
-#   - MKL_PARDISO_INCLUDE_DIRS
-
-if(NOT MKL_LIBRARY_DIR)
-  message(WARNING "Using default MKL library path. Double check the variable MKL_LIBRARY_DIR")
-  set(MKL_LIBRARY_DIR "lib/intel64")
-endif()
-
-include(MfemCmakeUtilities)
-mfem_find_package(MKL_PARDISO MKL_PARDISO
-    MKL_PARDISO_DIR "include" mkl_pardiso.h ${MKL_LIBRARY_DIR} mkl_core
-  "Paths to headers required by MKL Pardiso." "Libraries required by MKL PARDISO."
-  ADD_COMPONENT MKL_LP64 "include" "" ${MKL_LIBRARY_DIR} mkl_intel_lp64
-  ADD_COMPONENT MKL_SEQUENTIAL "include" "" ${MKL_LIBRARY_DIR} mkl_sequential)

config/cmake/modules/FindMUMPS.cmake

@@ -11,9 +11,8 @@
 
 # Sets the following variables:
 #   - MUMPS_FOUND
-#   - MUMPS_LIBRARIES
 #   - MUMPS_INCLUDE_DIRS
-#   - MUMPS_VERSION
+#   - MUMPS_LIBRARIES
 
 include(MfemCmakeUtilities)
 mfem_find_package(MUMPS MUMPS MUMPS_DIR
@@ -22,18 +21,3 @@ mfem_find_package(MUMPS MUMPS MUMPS_DIR
   "Libraries required by MUMPS."
   ADD_COMPONENT mumps_common "include" dmumps_c.h "lib" mumps_common
   ADD_COMPONENT pord "include" dmumps_c.h "lib" pord)
-
-if (MUMPS_FOUND AND (NOT MUMPS_VERSION))
-  try_run(MUMPS_VERSION_RUN_RESULT MUMPS_VERSION_COMPILE_RESULT
-          ${CMAKE_CURRENT_BINARY_DIR}/config
-          ${CMAKE_CURRENT_SOURCE_DIR}/config/get_mumps_version.cpp
-          CMAKE_FLAGS -DINCLUDE_DIRECTORIES:STRING=${MUMPS_INCLUDE_DIRS}
-          RUN_OUTPUT_VARIABLE MUMPS_VERSION_OUTPUT)
-  if ((MUMPS_VERSION_RUN_RESULT EQUAL 0) AND MUMPS_VERSION_OUTPUT)
-    string(STRIP "${MUMPS_VERSION_OUTPUT}" MUMPS_VERSION)
-    set(MUMPS_VERSION ${MUMPS_VERSION} CACHE STRING "MUMPS version." FORCE)
-    message(STATUS "Found MUMPS version ${MUMPS_VERSION}")
-  else()
-    message(FATAL_ERROR "Unable to determine MUMPS version.")
-  endif()
-endif()

config/config.hpp.in

@@ -30,10 +30,10 @@
 #define MFEM_VERSION_MINOR (((MFEM_VERSION)/100)%100)
 #define MFEM_VERSION_PATCH ((MFEM_VERSION)%100)
 
-// The absolute path of the MFEM source prefix.
+// The absolute path of the MFEM source prefix
 // #define MFEM_SOURCE_DIR "@MFEM_SOURCE_DIR@"
 
-// The absolute path of the MFEM installation prefix.
+// The absolute path of the MFEM installation prefix
 // #define MFEM_INSTALL_DIR "@MFEM_INSTALL_DIR@"
 
 // Description of the git commit used to build MFEM.
@@ -91,7 +91,7 @@
 // Enable MFEM functionality based on the SuiteSparse library.
 // #define MFEM_USE_SUITESPARSE
 
-// Enable MFEM functionality based on the SuperLU_DIST library.
+// Enable MFEM functionality based on the SuperLU library.
 // #define MFEM_USE_SUPERLU
 // #define MFEM_USE_SUPERLU5
 
@@ -102,40 +102,40 @@
 // Enable MFEM functionality based on the STRUMPACK library.
 // #define MFEM_USE_STRUMPACK
 
-// Enable MFEM features based on the Ginkgo library.
+// Enable MFEM features based on the Ginkgo library
 // #define MFEM_USE_GINKGO
 
 // Enable MFEM functionality based on the AmgX library.
 // #define MFEM_USE_AMGX
 
-// Enable secure socket streams based on the GNUTLS library.
+// Enable secure socket streams based on the GNUTLS library
 // #define MFEM_USE_GNUTLS
 
-// Enable Sidre support.
+// Enable Sidre support
 // #define MFEM_USE_SIDRE
 
-// Enable the use of SIMD in the high performance templated classes.
+// Enable the use of SIMD in the high performance templated classes
 // #define MFEM_USE_SIMD
 
-// Enable FMS support.
+// Enable FMS support
 // #define MFEM_USE_FMS
 
-// Enable Conduit support.
+// Enable Conduit support
 // #define MFEM_USE_CONDUIT
 
-// Enable functionality based on the NetCDF library (reading CUBIT files).
+// Enable functionality based on the NetCDF library (reading CUBIT files)
 // #define MFEM_USE_NETCDF
 
-// Enable functionality based on the PETSc library.
+// Enable functionality based on the PETSc library
 // #define MFEM_USE_PETSC
 
-// Enable functionality based on the SLEPc library.
+// Enable functionality based on the SLEPc library
 // #define MFEM_USE_SLEPC
 
 // Enable functionality based on the MPFR library.
 // #define MFEM_USE_MPFR
 
-// Enable MFEM functionality based on the PUMI library.
+// Enable MFEM functionality based on the PUMI library
 // #define MFEM_USE_PUMI
 
 // Enable Moonolith-based general interpolation between finite element spaces.
@@ -144,7 +144,7 @@
 // Enable MFEM functionality based on the HIOP library.
 // #define MFEM_USE_HIOP
 
-// Enable MFEM functionality based on the GSLIB library.
+// Enable MFEM functionality based on the GSLIB library
 // #define MFEM_USE_GSLIB
 
 // Build the NVIDIA GPU/CUDA-enabled version of the MFEM library.
@@ -186,13 +186,10 @@
 // Enable interface to the MKL CPardiso library.
 // #define MFEM_USE_MKL_CPARDISO
 
-// Enable interface to the MKL Pardiso library.
-// #define MFEM_USE_MKL_PARDISO
-
-// Use forward mode for automatic differentiation.
+// Use forward mode for automatic differentiation
 // #define MFEM_USE_ADFORWARD
 
-// Enable the use of the CoDiPack library for AD.
+// Enable the use of the CoDiPack library for AD
 // #define MFEM_USE_CODIPACK
 
 // Enable functionality based on the Google Benchmark library.

config/config.mk.in

@@ -57,7 +57,6 @@ MFEM_USE_UMPIRE        = @MFEM_USE_UMPIRE@
 MFEM_USE_SIMD          = @MFEM_USE_SIMD@
 MFEM_USE_ADIOS2        = @MFEM_USE_ADIOS2@
 MFEM_USE_MKL_CPARDISO  = @MFEM_USE_MKL_CPARDISO@
-MFEM_USE_MKL_PARDISO   = @MFEM_USE_MKL_PARDISO@
 MFEM_USE_MOONOLITH     = @MFEM_USE_MOONOLITH@
 MFEM_USE_ADFORWARD     = @MFEM_USE_ADFORWARD@
 MFEM_USE_CODIPACK      = @MFEM_USE_CODIPACK@

config/defaults.cmake

@@ -60,7 +60,6 @@ option(MFEM_USE_ADIOS2 "Enable ADIOS2" OFF)
 option(MFEM_USE_CALIPER "Enable Caliper support" OFF)
 option(MFEM_USE_ALGOIM "Enable Algoim support" OFF)
 option(MFEM_USE_MKL_CPARDISO "Enable MKL CPardiso" OFF)
-option(MFEM_USE_MKL_PARDISO "Enable MKL Pardiso" OFF)
 option(MFEM_USE_ADFORWARD "Enable forward mode for AD" OFF)
 option(MFEM_USE_CODIPACK "Enable automatic differentiation (AD) using CoDiPack" OFF)
 option(MFEM_USE_BENCHMARK "Enable Google Benchmark" OFF)
@@ -135,18 +134,16 @@ set(ParMETIS_DIR "${MFEM_DIR}/../parmetis-4.0.3" CACHE PATH
 set(ParMETIS_REQUIRED_PACKAGES "METIS" CACHE STRING
     "Additional packages required by ParMETIS.")
 
-set(SuperLUDist_DIR "${MFEM_DIR}/../SuperLU_DIST_8.1.2" CACHE PATH
+set(SuperLUDist_DIR "${MFEM_DIR}/../SuperLU_DIST_6.3.1" CACHE PATH
     "Path to the SuperLU_DIST library.")
 # SuperLU_DIST may also depend on "OpenMP", depending on how it was compiled.
-set(SuperLUDist_REQUIRED_PACKAGES "MPI" "ParMETIS" "METIS"
-    "LAPACK" "BLAS" CACHE STRING
+set(SuperLUDist_REQUIRED_PACKAGES "MPI" "BLAS" "ParMETIS" CACHE STRING
     "Additional packages required by SuperLU_DIST.")
 
-set(MUMPS_DIR "${MFEM_DIR}/../MUMPS_5.5.0" CACHE PATH
+set(MUMPS_DIR "${MFEM_DIR}/../MUMPS_5.2.0" CACHE PATH
     "Path to the MUMPS library.")
-# MUMPS may also depend on "OpenMP", depending on how it was compiled.
-set(MUMPS_REQUIRED_PACKAGES "MPI" "MPI_Fortran" "ParMETIS" "METIS"
-    "ScaLAPACK" "LAPACK" "BLAS" CACHE STRING
+# Packages required by MUMPS, depending on how it was compiled.
+set(MUMPS_REQUIRED_PACKAGES "MPI" "BLAS" "METIS" "ScaLAPACK" CACHE STRING
     "Additional packages required by MUMPS.")
 # If the MPI package does not find all required Fortran libraries:
 # set(MUMPS_REQUIRED_LIBRARIES "gfortran" "mpi_mpifh" CACHE STRING
@@ -229,8 +226,6 @@ set(MKL_CPARDISO_DIR "" CACHE STRING "MKL installation path.")
 set(MKL_MPI_WRAPPER_LIB "mkl_blacs_mpich_lp64" CACHE STRING "MKL MPI wrapper library")
 set(MKL_LIBRARY_DIR "" CACHE STRING "Custom library subdirectory")
 
-set(MKL_PARDISO_DIR "" CACHE STRING "MKL installation path.")
-
 set(OCCA_DIR "${MFEM_DIR}/../occa" CACHE PATH "Path to OCCA")
 set(RAJA_DIR "${MFEM_DIR}/../raja" CACHE PATH "Path to RAJA")
 set(CEED_DIR "${MFEM_DIR}/../libCEED" CACHE PATH "Path to libCEED")

config/defaults.mk

@@ -160,7 +160,6 @@ MFEM_USE_UMPIRE        = NO
 MFEM_USE_SIMD          = NO
 MFEM_USE_ADIOS2        = NO
 MFEM_USE_MKL_CPARDISO  = NO
-MFEM_USE_MKL_PARDISO   = NO
 MFEM_USE_MOONOLITH     = NO
 MFEM_USE_ADFORWARD     = NO
 MFEM_USE_CODIPACK      = NO
@@ -285,10 +284,10 @@ ifeq ($(MFEM_USE_SUPERLU5),YES)
    SUPERLU_LIB = $(XLINKER)-rpath,$(SUPERLU_DIR)/lib -L$(SUPERLU_DIR)/lib\
       -lsuperlu_dist_5.1.0
 else
-   SUPERLU_DIR = @MFEM_DIR@/../SuperLU_DIST_8.1.2
+   SUPERLU_DIR = @MFEM_DIR@/../SuperLU_DIST_6.3.1
    SUPERLU_OPT = -I$(SUPERLU_DIR)/include
    SUPERLU_LIB = $(XLINKER)-rpath,$(SUPERLU_DIR)/lib64 -L$(SUPERLU_DIR)/lib64\
-      -lsuperlu_dist $(LAPACK_LIB)
+      -lsuperlu_dist -lblas
 endif
 
 # SCOTCH library configuration (required by STRUMPACK <= v2.1.0, optional in
@@ -312,7 +311,7 @@ MPI_FORTRAN_LIB = -lmpifort
 # MPI_FORTRAN_LIB += -lgfortran
 
 # MUMPS library configuration
-MUMPS_DIR = @MFEM_DIR@/../MUMPS_5.5.0
+MUMPS_DIR = @MFEM_DIR@/../MUMPS_5.2.0
 MUMPS_OPT = -I$(MUMPS_DIR)/include
 MUMPS_LIB = $(XLINKER)-rpath,$(MUMPS_DIR)/lib -L$(MUMPS_DIR)/lib -ldmumps\
  -lmumps_common -lpord $(SCALAPACK_LIB) $(LAPACK_LIB) $(MPI_FORTRAN_LIB)
@@ -485,6 +484,7 @@ ifdef GOTCHA_DIR
    CALIPER_LIB += $(XLINKER)-rpath,$(GOTCHA_DIR)/lib64 $(XLINKER)-rpath,$(GOTCHA_DIR)/lib -L$(GOTCHA_DIR)/lib64 -L$(GOTCHA_DIR)/lib -lgotcha
 endif
 
+
 # BLITZ library configuration
 BLITZ_DIR = @MFEM_DIR@/../blitz
 BLITZ_OPT = -I$(BLITZ_DIR)/include
@@ -539,14 +539,6 @@ MKL_CPARDISO_LIB = $(XLINKER)-rpath,$(MKL_CPARDISO_DIR)/$(MKL_LIBRARY_SUBDIR)\
    -L$(MKL_CPARDISO_DIR)/$(MKL_LIBRARY_SUBDIR) -l$(MKL_MPI_WRAPPER)\
    -lmkl_intel_lp64 -lmkl_sequential -lmkl_core
 
-# MKL Pardiso library configuration
-MKL_PARDISO_DIR ?=
-MKL_LIBRARY_SUBDIR ?= lib
-MKL_PARDISO_OPT = -I$(MKL_PARDISO_DIR)/include
-MKL_PARDISO_LIB = $(XLINKER)-rpath,$(MKL_PARDISO_DIR)/$(MKL_LIBRARY_SUBDIR)\
-   -L$(MKL_PARDISO_DIR)/$(MKL_LIBRARY_SUBDIR)\
-   -lmkl_intel_lp64 -lmkl_sequential -lmkl_core
-
 # PARELAG library configuration
 PARELAG_DIR = @MFEM_DIR@/../parelag
 PARELAG_OPT = -I$(PARELAG_DIR)/src -I$(PARELAG_DIR)/build/src

data/disk.mesh

@@ -0,0 +1,109 @@
+MFEM NURBS mesh v1.0
+
+#
+# MFEM Geometry Types (see mesh/geom.hpp):
+#
+# SEGMENT     = 1
+# SQUARE      = 3
+# CUBE        = 5
+#
+
+dimension
+2
+
+elements
+5
+1 3 4 5 6 7
+1 3 0 1 5 4
+1 3 1 2 6 5
+1 3 3 7 6 2
+1 3 0 4 7 3
+
+boundary
+4
+1 1 0 1
+1 1 2 3
+1 1 1 2
+1 1 3 0
+
+edges
+12
+0 0 1
+0 4 5
+0 7 6
+0 3 2
+1 1 2
+1 5 6
+1 4 7
+1 0 3
+2 0 4
+2 1 5
+2 2 6
+2 3 7
+
+vertices
+8
+
+knotvectors
+3
+2 3 0 0 0 1 1 1
+2 3 0 0 0 1 1 1
+2 3 0 0 0 1 1 1
+
+weights
+1
+1
+1
+1
+1
+1
+1
+1
+0.70710678118655
+1
+1
+0.70710678118655
+0.70710678118655
+1
+1
+0.70710678118655
+1
+1
+1
+1
+1
+0.85355339059327
+0.85355339059327
+0.85355339059327
+0.85355339059327
+
+FiniteElementSpace
+FiniteElementCollection: NURBS2
+VDim: 2
+Ordering: 1
+
+-0.70710678118 -0.70710678118
+0.70710678118 -0.70710678118
+0.70710678118 0.70710678118
+-0.70710678118 0.70710678118
+-0.35355339059 -0.35355339059
+0.35355339059 -0.35355339059
+0.35355339059 0.35355339059
+-0.35355339059 0.35355339059
+0 -1.41421356236
+0 -0.35355339059
+0 0.35355339059
+0 1.41421356236
+1.41421356236 0
+0.35355339059 0
+-0.35355339059 0
+-1.41421356236 0
+-0.530330085885 -0.530330085885
+0.530330085885 -0.530330085885
+0.530330085885 0.530330085885
+-0.530330085885 0.530330085885
+0 0
+0 -0.883883476475
+0.883883476475 0
+0 0.883883476475
+-0.883883476475 0

doc/CodeDocumentation.conf.in

@@ -795,7 +795,6 @@ INPUT                  = @MFEM_SOURCE_DIR@/doc/CodeDocumentation.dox \
                          @MFEM_SOURCE_DIR@/miniapps/common \
                          @MFEM_SOURCE_DIR@/miniapps/electromagnetics \
                          @MFEM_SOURCE_DIR@/miniapps/gslib \
-                         @MFEM_SOURCE_DIR@/miniapps/hdiv-linear-solver \
                          @MFEM_SOURCE_DIR@/miniapps/hooke \
                          @MFEM_SOURCE_DIR@/miniapps/hooke/kernels \
                          @MFEM_SOURCE_DIR@/miniapps/hooke/materials \
@@ -811,10 +810,7 @@ INPUT                  = @MFEM_SOURCE_DIR@/doc/CodeDocumentation.dox \
                          @MFEM_SOURCE_DIR@/miniapps/shifted \
                          @MFEM_SOURCE_DIR@/miniapps/solvers \
                          @MFEM_SOURCE_DIR@/miniapps/tools \
-                         @MFEM_SOURCE_DIR@/miniapps/toys \
-                         @MFEM_SOURCE_DIR@/miniapps/spde \
-                         @MFEM_SOURCE_DIR@/miniapps/dpg \
-                         @MFEM_SOURCE_DIR@/miniapps/dpg/util
+                         @MFEM_SOURCE_DIR@/miniapps/toys
 
 # This tag can be used to specify the character encoding of the source files
 # that doxygen parses. Internally doxygen uses the UTF-8 encoding. Doxygen uses

doc/CodeDocumentation.dox

@@ -39,7 +39,7 @@ namespace mfem {
  * - Device
  * - Memory
  * - MemoryManager
- * - mfem::forall functions in forall.hpp
+ * - MFEM_FORALL macro in forall.hpp
  *
  * <H3>Example codes</H3>
  * - <a class="el" href="ex0_8cpp_source.html">Example 0</a>: simplest example, nodal H1 FEM for the Laplace problem
@@ -186,7 +186,6 @@ namespace mfem {
  * - <a class="el" href="extruder_8cpp_source.html">Extruder</a>: extrude a low-dimensional mesh into a higher dimension
  * - <a class="el" href="mesh-explorer_8cpp_source.html">Mesh Explorer</a>: visualize and manipulate meshes
  * - <a class="el" href="mesh-optimizer_8cpp_source.html">Mesh Optimizer</a>: optimize high-order meshes, <a class="el" href="mesh-optimizer_8cpp_source.html">serial</a> and <a class="el" href="pmesh-optimizer_8cpp_source.html">parallel</a> versions
- * - <a class="el" href="mesh-quality_8cpp_source.html">Mesh Quality</a>: visualize and check mesh quality
  * - <a class="el" href="trimmer_8cpp_source.html">Trimmer</a>: trim elements from existing meshes
  * - <a class="el" href="display-basis_8cpp_source.html">Display Basis</a>: visualize finite element basis functions
  * - <a class="el" href="get-values_8cpp_source.html">Get Values</a>: extract field values via DataCollection classes
@@ -199,15 +198,12 @@ namespace mfem {
  * - <a class="el" href="distance_8cpp_source.html">Distance</a>: finite element distance function solver
  * - <a class="el" href="diffusion_8cpp_source.html">Shifted Diffusion</a>: shifted boundary diffusion solver
  * - <a class="el" href="extrapolate_8cpp_source.html">Extrapolation</a>: PDE-based extrapolation of finite element functions
- * - <a class="el" href="block-solvers_8cpp_source.html">Block Solvers</a>: comparison of saddle point system solvers
+ * - <a class="el" href="distance_8cpp_source.html">Block Solvers</a>: comparison of saddle point system solvers
  * - <a class="el" href="parheat_8cpp_source.html">Optimization gradients</a>: Gradients of PDE-constrained function
  * - <a class="el" href="par__example_8cpp_source.html">Parallel AD</a>: Parallel p-Laplacian example
  * - <a class="el" href="seq__example_8cpp_source.html">Serial AD</a>: Serial p-Laplacian example
  * - <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
- * - <a class="el" href="generate__random__field_8cpp_source.html">SPDE Solvers</a>: SPDE solver random field generation
- * - <a class="el" href="pdiffusion_8cpp_source.html">DPG Diffusion example</a>: DPG formulation for the diffusion problem
- * - <a class="el" href="pmaxwell_8cpp_source.html">DPG Maxwell example</a>: DPG formulation for the indefinite Maxwell problem
  *
  * See also the <a class="el" href="https://mfem.org/examples/">examples documentation</a> online.
  */

examples/CMakeLists.txt

@@ -40,8 +40,6 @@ list(APPEND ALL_EXE_SRCS
   ex30.cpp
   ex31.cpp
   ex33.cpp
-  ex34.cpp
-  ex35.cpp
   )
 
 if (MFEM_USE_MPI)
@@ -163,15 +161,6 @@ if (MFEM_ENABLE_TESTING)
       $<TARGET_FILE:ex11p> "-no-vis" "--superlu"
       ${MPIEXEC_POSTFLAGS})
   endif()
-
-  # If MUMPS is enabled, add a test run that uses it.
-  if (MFEM_USE_MUMPS)
-    add_test(NAME ex25p_mumps_np=${MFEM_MPI_NP}
-      COMMAND ${MPIEXEC} ${MPIEXEC_NUMPROC_FLAG} ${MFEM_MPI_NP}
-      ${MPIEXEC_PREFLAGS}
-      $<TARGET_FILE:ex25p> "-no-vis" "--mumps-solver"
-      ${MPIEXEC_POSTFLAGS})
-  endif()
 endif()
 
 # Include the examples/amgx directory if AmgX is enabled

examples/contact/DirichletObstacleProblem.cpp

@@ -0,0 +1,141 @@
+#include "mfem.hpp"
+#include "Problems.hpp"
+#include "IPsolver.hpp"
+#include <fstream>
+#include <iostream>
+
+using namespace std;
+using namespace mfem;
+
+
+
+double fRhs(const Vector &pt);
+double obstacle(const Vector &pt);
+double dmanufacturedFun(const Vector &pt);
+
+int main(int argc, char *argv[])
+{
+  int FEorder = 1; // order of the finite elements
+  int linSolver = 0;
+  int maxIPMiters = 30;
+  bool iAmRoot = true;
+  int ref_levels = 3; 
+  OptionsParser args(argc, argv);
+  args.AddOption(&FEorder, "-o", "--order",\
+		  "Order of the finite elements.");
+  args.AddOption(&linSolver, "-linSolver", "--linearSolver", \
+       "IP-Newton linear system solution strategy.");
+  args.AddOption(&maxIPMiters, "-IPMiters", "--IPMiters",\
+		  "Maximum number of IPM iterations");
+  args.AddOption(&ref_levels, "-r", "--mesh_refinement", \
+		  "Mesh Refinement");
+  
+  args.Parse();
+  if(!args.Good())
+  {
+    args.PrintUsage(cout);
+    return 1;
+  }
+  else
+  {
+    if( iAmRoot )
+    {
+      args.PrintOptions(cout);
+    }
+  }
+
+  const char *meshFile = "../../data/inline-quad.mesh";
+  Mesh *mesh = new Mesh(meshFile, 1, 1);
+  int dim = mesh->Dimension(); // geometric dimension of the domain
+  for (int l = 0; l < ref_levels; l++)
+  {
+      mesh->UniformRefinement();
+  }
+
+  FiniteElementCollection *fec = new H1_FECollection(FEorder, dim);
+  FiniteElementSpace      *Vh  = new FiniteElementSpace(mesh, fec);
+  Array<int> ess_tdof_list;
+  if (mesh->bdr_attributes.Size())
+  {
+     Array<int> ess_bdr(mesh->bdr_attributes.Max());
+     ess_bdr = 1;
+     Vh->GetEssentialTrueDofs(ess_bdr, ess_tdof_list);
+  }
+  
+  double DC_val = 0.0;
+  int dimD = Vh->GetTrueVSize();
+  Vector x0(dimD); x0 = DC_val;
+  Vector xf(dimD); xf = 0.0;
+
+  ObstacleProblem problem(Vh, x0, &fRhs, &obstacle, ess_tdof_list);
+  
+
+  InteriorPointSolver optimizer(&problem); 
+  optimizer.SetTol(1.e-7);
+  optimizer.SetLinearSolver(linSolver);
+  optimizer.SetMaxIter(maxIPMiters);
+  optimizer.Mult(x0, xf);
+  
+  double Einitial = problem.E(x0);
+  double Efinal = problem.E(xf);
+  cout << "Energy objective at initial point = " << Einitial << endl;
+  cout << "Energy objective at optimizer = " << Efinal << endl;
+
+  GridFunction d_gf(Vh);
+
+  d_gf = xf;
+  
+  FunctionCoefficient dm_fc(dmanufacturedFun); // pseudo-manufactured solution
+  GridFunction dm_gf(Vh);
+  dm_gf.ProjectCoefficient(dm_fc);
+
+  
+  ParaViewDataCollection paraview_dc("BarrierProblemSolution", mesh);
+  paraview_dc.SetPrefixPath("ParaView");
+  paraview_dc.SetLevelsOfDetail(FEorder);
+  paraview_dc.SetDataFormat(VTKFormat::BINARY);
+  paraview_dc.SetHighOrderOutput(true);
+  paraview_dc.SetCycle(0);
+  paraview_dc.SetTime(0.0);
+  paraview_dc.RegisterField("d(x) (numerical)", &d_gf);
+  paraview_dc.RegisterField("d(x) (pseudo-manufactured)", &dm_gf);
+  paraview_dc.Save();
+  
+
+  delete Vh;
+  delete fec;
+  delete mesh;
+  return 0;
+}
+
+
+
+double dmanufacturedFun(const Vector &pt)
+{
+  double alpha = 16.5;
+  return sin(M_PI * pt(1)) * (sin(M_PI * pt(0)) - alpha * pow(pt(0) * (1. - pt(0)), 2));
+}
+
+
+// f(x) forcing term... which enters the objective energy functional
+// E(d) = 0.5 d^T K d - f^T d, where f is a discrete vector representation
+// of f(x). f(x) is such that in the absence of bound-constraints then
+// the solution of the optimization problem satisfies the PDE
+// -div(grad(d)) + d = f + homogeneous Neumann conditions on the unit interval,
+// for d(x) = cos(2 \pi x) + a0 + a3 (x^3 - 1.5 x^2), a2 = 0.2, a3 = -2
+
+double fRhs(const Vector &pt)
+{
+  double alpha = 16.5;
+  double fx;
+  fx = pow(M_PI, 2) * sin(M_PI * pt(0));
+  fx += alpha * (2. * pow(pt(0), 2) + 2. * pow(1.-pt(0), 2) - 8. * pt(0) * (1.-pt(0)));
+  fx += pow(M_PI, 2) * sin(M_PI * pt(0)) * dmanufacturedFun(pt);
+  fx *= sin(M_PI * pt(1));
+  return fx;
+}
+
+double obstacle(const Vector &pt)
+{
+  return 0.0;
+}

examples/contact/DirichletObstacleProblemVariant.cpp

@@ -0,0 +1,156 @@
+#include "mfem.hpp"
+#include "Problems.hpp"
+#include "IPsolver.hpp"
+#include <fstream>
+#include <iostream>
+
+using namespace std;
+using namespace mfem;
+
+
+
+double fRhs(const Vector &pt);
+double obstacle(const Vector &pt);
+double dmanufacturedFun(const Vector &pt);
+
+int main(int argc, char *argv[])
+{
+  int FEorder = 1; // order of the finite elements
+  int linSolver = 0;
+  int maxIPMiters = 30;
+  bool iAmRoot = true;
+  int ref_levels = 3; 
+  OptionsParser args(argc, argv);
+  args.AddOption(&FEorder, "-o", "--order",\
+		  "Order of the finite elements.");
+  args.AddOption(&linSolver, "-linSolver", "--linearSolver", \
+       "IP-Newton linear system solution strategy.");
+  args.AddOption(&maxIPMiters, "-IPMiters", "--IPMiters",\
+		  "Maximum number of IPM iterations");
+  args.AddOption(&ref_levels, "-r", "--mesh_refinement", \
+		  "Mesh Refinement");
+  
+  args.Parse();
+  if(!args.Good())
+  {
+    args.PrintUsage(cout);
+    return 1;
+  }
+  else
+  {
+    if( iAmRoot )
+    {
+      args.PrintOptions(cout);
+    }
+  }
+
+  const char *meshFile = "../../data/inline-quad.mesh";
+  Mesh *mesh = new Mesh(meshFile, 1, 1);
+  int dim = mesh->Dimension(); // geometric dimension of the domain
+  for (int l = 0; l < ref_levels; l++)
+  {
+      mesh->UniformRefinement();
+  }
+
+  FiniteElementCollection *fec = new H1_FECollection(FEorder, dim);
+  FiniteElementSpace      *Vh  = new FiniteElementSpace(mesh, fec);
+  Array<int> ess_tdof_list;
+  if (mesh->bdr_attributes.Size())
+  {
+     Array<int> ess_bdr(mesh->bdr_attributes.Max());
+     ess_bdr = 1;
+     Vh->GetEssentialTrueDofs(ess_bdr, ess_tdof_list);
+  }
+  
+  double DC_val = 0.06;
+  Vector x0DC(Vh->GetTrueVSize()); x0DC = DC_val;
+  int dimD = Vh->GetTrueVSize() - ess_tdof_list.Size();
+  Vector x0(dimD); x0 = 0.0;
+  Vector xf(dimD); xf = 0.0;
+
+  ObstacleProblemVariant problem(Vh, x0DC, &fRhs, &obstacle, ess_tdof_list);
+
+  InteriorPointSolver optimizer(&problem); 
+  optimizer.SetTol(1.e-7);
+  optimizer.SetLinearSolver(linSolver);
+  optimizer.SetMaxIter(maxIPMiters);
+  optimizer.Mult(x0, xf);
+  
+  double Einitial = problem.E(x0);
+  double Efinal = problem.E(xf);
+  cout << "Energy objective at initial point = " << Einitial << endl;
+  cout << "Energy objective at optimizer = " << Efinal << endl;
+
+  Array<int> noness_tdof_list;
+  noness_tdof_list.SetSize(dimD);
+  int i = 0;
+  for(int j = 0; j < Vh->GetTrueVSize(); j++)
+  {
+     if(ess_tdof_list.Find(j) == -1)
+     {
+        noness_tdof_list[i] = j;
+        i += 1;
+     }
+  }
+
+
+  GridFunction d_gf(Vh);
+  d_gf.Set(1.0, x0DC);
+  d_gf.SetSubVector(noness_tdof_list, xf);
+
+
+
+  FunctionCoefficient dm_fc(dmanufacturedFun); // pseudo-manufactured solution
+  GridFunction dm_gf(Vh);
+  dm_gf.ProjectCoefficient(dm_fc);
+
+  
+  ParaViewDataCollection paraview_dc("BarrierProblemSolution", mesh);
+  paraview_dc.SetPrefixPath("ParaView");
+  paraview_dc.SetLevelsOfDetail(FEorder);
+  paraview_dc.SetDataFormat(VTKFormat::BINARY);
+  paraview_dc.SetHighOrderOutput(true);
+  paraview_dc.SetCycle(0);
+  paraview_dc.SetTime(0.0);
+  paraview_dc.RegisterField("d(x) (numerical)", &d_gf);
+  paraview_dc.RegisterField("d(x) (pseudo-manufactured)", &dm_gf);
+  paraview_dc.Save();
+ 
+
+  delete Vh;
+  delete fec;
+  delete mesh;
+  return 0;
+}
+
+
+
+double dmanufacturedFun(const Vector &pt)
+{
+  double alpha = 16.5;
+  return sin(M_PI * pt(1)) * (sin(M_PI * pt(0)) - alpha * pow(pt(0) * (1. - pt(0)), 2));
+}
+
+
+// f(x) forcing term... which enters the objective energy functional
+// E(d) = 0.5 d^T K d - f^T d, where f is a discrete vector representation
+// of f(x). f(x) is such that in the absence of bound-constraints then
+// the solution of the optimization problem satisfies the PDE
+// -div(grad(d)) + d = f + homogeneous Neumann conditions on the unit interval,
+// for d(x) = cos(2 \pi x) + a0 + a3 (x^3 - 1.5 x^2), a2 = 0.2, a3 = -2
+
+double fRhs(const Vector &pt)
+{
+  double alpha = 16.5;
+  double fx;
+  fx = pow(M_PI, 2) * sin(M_PI * pt(0));
+  fx += alpha * (2. * pow(pt(0), 2) + 2. * pow(1.-pt(0), 2) - 8. * pt(0) * (1.-pt(0)));
+  fx += pow(M_PI, 2) * sin(M_PI * pt(0)) * dmanufacturedFun(pt);
+  fx *= sin(M_PI * pt(1));
+  return fx;
+}
+
+double obstacle(const Vector &pt)
+{
+  return 0.0;
+}

examples/contact/IPsolver.cpp

@@ -0,0 +1,827 @@
+#include "mfem.hpp"
+#include "IPsolver.hpp"
+#include "Problems.hpp"
+#include <fstream>
+#include <iostream>
+#include <cstdlib>
+
+using namespace std;
+using namespace mfem;
+
+
+
+InteriorPointSolver::InteriorPointSolver(GeneralOptProblem * Problem) : optProblem(Problem), block_offsetsumlz(5), block_offsetsuml(4), block_offsetsx(3),
+Huu(nullptr), Hum(nullptr), Hmu(nullptr), Hmm(nullptr), Wmm(nullptr), D(nullptr), Ju(nullptr), Jm(nullptr), JuT(nullptr), JmT(nullptr), Huucl(nullptr), HLuu(nullptr), saveLogBarrierIterates(false)
+{
+  rel_tol  = 1.e-2;
+  max_iter = 20;
+  mu_k     = 1.0;
+
+  sMax     = 1.e2;
+  kSig     = 1.e10;   // control deviation from primal Hessian
+  tauMin   = 0.8;     // control rate at which iterates can approach the boundary
+  eta      = 1.e-4;   // backtracking constant
+  thetaMin = 1.e-4;   // allowed violation of the equality constraints
+
+  // constants in line-step A-5.4
+  delta    = 1.0;
+  sTheta   = 1.1;
+  sPhi     = 2.3;
+
+  // control the rate at which the penalty parameter is decreased
+  kMu     = 0.2;
+  thetaMu = 1.5;
+
+  // TO DO -- include the filter
+
+  thetaMax = 1.e6; // maximum constraint violation
+  // data for the second order correction
+  kSoc     = 0.99;
+
+  // equation (18)
+  gTheta = 1.e-5;
+  gPhi   = 1.e-5;
+
+  kEps   = 1.e1;
+
+  dimU = optProblem->GetDimU();
+  dimM = optProblem->GetDimM();
+  dimC = optProblem->GetDimC();
+  ckSoc.SetSize(dimC);
+  
+  block_offsetsumlz[0] = 0;
+  block_offsetsumlz[1] = dimU; // u
+  block_offsetsumlz[2] = dimM; // m
+  block_offsetsumlz[3] = dimC; // lambda
+  block_offsetsumlz[4] = dimM; // zl
+  block_offsetsumlz.PartialSum();
+
+  for(int i = 0; i < block_offsetsuml.Size(); i++)  { block_offsetsuml[i] = block_offsetsumlz[i]; }
+  for(int i = 0; i < block_offsetsx.Size(); i++)    { block_offsetsx[i]   = block_offsetsuml[i] ; }
+
+  // lower-bound for the inequality constraint m >= ml
+  ml = optProblem->Getml();
+  
+  lk.SetSize(dimC);  lk  = 0.0;
+  zlk.SetSize(dimM); zlk = 0.0;
+
+  linSolver = 0;
+  MyRank = 0;
+  iAmRoot = MyRank == 0 ? true : false;
+}
+
+double InteriorPointSolver::MaxStepSize(Vector &x, Vector &xl, Vector &xhat, double tau)
+{
+  double alphaMaxloc = 1.0;
+  double alphaTmp;
+  for(int i = 0; i < x.Size(); i++)
+  {
+    if( xhat(i) < 0. )
+    {
+      alphaTmp = -1. * tau * (x(i) - xl(i)) / xhat(i);
+      alphaMaxloc = min(alphaMaxloc, alphaTmp);
+    } 
+  }
+
+  // alphaMaxloc is the local maximum step size which is
+  // distinct on each MPI process. Need to compute
+  // the global maximum step size 
+  double alphaMaxglb;
+  alphaMaxglb = alphaMaxloc;
+  return alphaMaxglb;
+}
+
+double InteriorPointSolver::MaxStepSize(Vector &x, Vector &xhat, double tau)
+{
+  Vector zero(x.Size()); zero = 0.0;
+  return MaxStepSize(x, zero, xhat, tau);
+}
+
+
+void InteriorPointSolver::Mult(const Vector &x0, Vector &xf)
+{
+  BlockVector x0block(block_offsetsx); x0block = 0.0;
+  x0block.GetBlock(0).Set(1.0, x0);
+  // To do: give options for user specificiation of initialization m0
+  x0block.GetBlock(1) = 100.;
+  x0block.GetBlock(1).Add(1.0, ml);
+  BlockVector xfblock(block_offsetsx); xfblock = 0.0;
+  Mult(x0block, xfblock);
+  xf.Set(1.0, xfblock.GetBlock(0));
+}
+
+void InteriorPointSolver::Mult(const BlockVector &x0, BlockVector &xf)
+{
+  converged = false;
+  
+  BlockVector xk(block_offsetsx), xhat(block_offsetsx); xk = 0; xhat = 0.0;
+  BlockVector Xk(block_offsetsumlz), Xhat(block_offsetsumlz); Xk = 0.0; Xhat = 0.0;
+  BlockVector Xhatuml(block_offsetsuml); Xhatuml = 0.0;
+  Vector zlhat(dimM); zlhat = 0.0;
+
+  xk.GetBlock(0).Set(1.0, x0.GetBlock(0));
+  xk.GetBlock(1).Set(1.0, x0.GetBlock(1));
+  // running estimate of the final values of the Lagrange multipliers
+  lk  = 0.0;
+  zlk = 0.0;
+
+  for(int i = 0; i < dimM; i++)
+  {
+    zlk(i) = 1.e1 * mu_k / (xk(i+dimU) - ml(i));
+  }
+
+  Xk.GetBlock(0).Set(1.0, xk.GetBlock(0));
+  Xk.GetBlock(1).Set(1.0, xk.GetBlock(1));
+  Xk.GetBlock(2).Set(1.0, lk);
+  Xk.GetBlock(3).Set(1.0, zlk);
+
+  /* set theta0 = theta(x0)
+   *     thetaMin
+   *     thetaMax
+   * when theta(xk) < thetaMin and the switching condition holds
+   * then we ask for the Armijo sufficient decrease of the barrier
+   * objective to be satisfied, in order to accept the trial step length alphakl
+   * 
+   * thetaMax controls how the filter is initialized for each log-barrier subproblem
+   * F0 = {(th, phi) s.t. th > thetaMax}
+   * that is the filter does not allow for iterates where the constraint violation
+   * is larger than that of thetaMax
+   */
+  double theta0 = theta(xk);
+  thetaMin = 1.e-4 * max(1.0, theta0);
+  thetaMax = 1.e8  * thetaMin;
+
+  double Eeval, maxBarrierSolves, Eevalmu0;
+  bool printOptimalityError; // control optimality error print to console for log-barrier subproblems
+  
+  maxBarrierSolves = 10;
+
+  for(jOpt = 0; jOpt < max_iter; jOpt++)
+  {
+    mfem::out << "interior-point solve step " << jOpt << endl;
+    // A-2. Check convergence of overall optimization problem
+    printOptimalityError = false;
+    Eevalmu0 = E(xk, lk, zlk, printOptimalityError);
+    if(Eevalmu0 < rel_tol)
+    {
+      converged = true;
+      mfem::out << "solved optimization problem :)\n";
+      break;
+    }
+    
+    if(jOpt > 0) { maxBarrierSolves = 1; }
+    
+    for(int i = 0; i < maxBarrierSolves; i++)
+    {
+      // A-3. Check convergence of the barrier subproblem
+      printOptimalityError = true;
+      Eeval = E(xk, lk, zlk, mu_k, printOptimalityError);
+      if(Eeval < kEps * mu_k)
+      {
+        mfem::out << "solved barrier subproblem, for mu = " << mu_k << endl;
+        // A-3.1. Recompute the barrier parameter
+        mu_k  = max(rel_tol / 10., min(kMu * mu_k, pow(mu_k, thetaMu)));
+        // A-3.2. Re-initialize the filter
+        F1.DeleteAll();
+        F2.DeleteAll();
+      }
+      else
+      {
+        break;
+      }
+    }
+    
+    // A-4. Compute the search direction
+    // solve for (uhat, mhat, lhat)
+    mfem::out << "\n** A-4. IP-Newton solve **\n";
+    zlhat = 0.0; Xhatuml = 0.0;
+    // why do we have Xhatuml ....???
+    // TO DO: remove Xhatuml in favor of passing Xhat
+    IPNewtonSolve(xk, lk, zlk, zlhat, Xhatuml, mu_k, false); 
+
+
+    // assign data stack, X = (u, m, l, zl)
+    Xk = 0.0;
+    Xk.GetBlock(0).Set(1.0, xk.GetBlock(0));
+    Xk.GetBlock(1).Set(1.0, xk.GetBlock(1));
+    Xk.GetBlock(2).Set(1.0, lk);
+    Xk.GetBlock(3).Set(1.0, zlk);
+
+    // assign data stack, Xhat = (uhat, mhat, lhat, zlhat)
+    Xhat = 0.0;
+    for(int i = 0; i < 3; i++)
+    {
+      Xhat.GetBlock(i).Set(1.0, Xhatuml.GetBlock(i));
+    }
+    Xhat.GetBlock(3).Set(1.0, zlhat);
+
+
+    // A-5. Backtracking line search.
+    mfem::out << "\n** A-5. Linesearch **\n";
+    mfem::out << "mu = " << mu_k << endl;
+    
+    lineSearch(Xk, Xhat, mu_k);
+    
+    if(lineSearchSuccess)
+    {
+      if(!switchCondition || !sufficientDecrease)
+      {
+        F1.Append( (1. - gTheta) * thx0);
+        F2.Append( phx0 - gPhi * thx0);
+      }
+      // ----- A-6: Accept the trial point
+      // print info regarding zl...
+      xk.GetBlock(0).Add(alpha, Xhat.GetBlock(0));
+      xk.GetBlock(1).Add(alpha, Xhat.GetBlock(1));
+      lk.Add(alpha,   Xhat.GetBlock(2));
+      zlk.Add(alphaz, Xhat.GetBlock(3));
+      projectZ(xk, zlk, mu_k);
+    }
+    else
+    {
+      mfem::out << "lineSearch not successful :(\n";
+      mfem::out << "attempting feasibility restoration with theta = " << thx0 << endl;
+      mfem::out << "no feasibility restoration implemented, exiting now \n";
+      break;
+    }
+    //
+    if(jOpt + 1 == max_iter) 
+    { 
+      mfem::out << "maximum optimization iterations :(\n";
+    }
+  }
+  // done with optimization routine, just reassign data to xf reference so
+  // that the application code has access to the optimal point
+  xf = 0.0;
+  xf.GetBlock(0).Set(1.0, xk.GetBlock(0));
+  xf.GetBlock(1).Set(1.0, xk.GetBlock(1));
+}
+
+void InteriorPointSolver::FormIPNewtonMat(BlockVector & x, Vector & l, Vector &zl, BlockOperator &Ak)
+{
+  // WARNING: Huu, Hum, Hmu, Hmm should all be Hessian terms of the Lagrangian, currently we 
+  //          them by Hessian terms of the objective function and neglect the Hessian of l^T c
+
+  Huu = optProblem->Duuf(x); Hum = optProblem->Dumf(x);
+  Hmu = optProblem->Dmuf(x); Hmm = optProblem->Dmmf(x);
+
+  Vector DiagLogBar(dimM); DiagLogBar = 0.0;
+  for(int ii = 0; ii < dimM; ii++)
+  {
+    DiagLogBar(ii) = zl(ii) / (x(ii+dimU) - ml(ii));
+  }
+  if(saveLogBarrierIterates)
+  {
+    std::ofstream diagStream;
+    char diagString[100];
+    snprintf(diagString, 100, "logBarrierHessiandata/D%d.dat", jOpt);
+    diagStream.open(diagString, ios::out | ios::trunc);
+    for(int ii = 0; ii < dimM; ii++)
+    {
+      diagStream << setprecision(30) << DiagLogBar(ii) << endl;
+    }
+    diagStream.close();
+  } 
+
+
+  D = new SparseMatrix(DiagLogBar);
+  
+  if(Hmm != nullptr)
+  {
+    Wmm = new SparseMatrix(*Hmm);
+    Wmm->Add(1.0, *D);
+  }
+  else
+  {
+    Wmm = D;
+  }
+
+  Ju = optProblem->Duc(x); JuT = Transpose(*Ju);
+  Jm = optProblem->Dmc(x); JmT = Transpose(*Jm);
+ 
+  Huucl = optProblem->lDuuc(x, l);
+  if(Huucl != nullptr)
+  {
+    HLuu = Add(*Huucl, *Huu);
+    Ak.SetBlock(0, 0, HLuu);
+  } 
+  else
+  {
+    Ak.SetBlock(0, 0, Huu);
+  }
+
+  //         IP-Newton system matrix
+  //    Ak = [[H_(u,u)  H_(u,m)   J_u^T]
+  //          [H_(m,u)  W_(m,m)   J_m^T]
+  //          [ J_u      J_m       0  ]]
+
+  Ak.SetBlock(0, 2, JuT);
+  Ak.SetBlock(1, 1, Wmm); Ak.SetBlock(1, 2, JmT);
+  Ak.SetBlock(2, 0,  Ju); Ak.SetBlock(2, 1,  Jm);
+
+  if(Hum != nullptr) { Ak.SetBlock(0, 1, Hum); Ak.SetBlock(1, 0, Hmu); }
+}
+
+
+// perturbed KKT system solve
+// determine the search direction
+void InteriorPointSolver::IPNewtonSolve(BlockVector &x, Vector &l, Vector &zl, Vector &zlhat, BlockVector &Xhat, double mu, bool socSolve)
+{
+  // solve A x = b, where A is the IP-Newton matrix
+  BlockOperator A(block_offsetsuml, block_offsetsuml); BlockVector b(block_offsetsuml); b = 0.0;
+  FormIPNewtonMat(x, l, zl, A);
+
+  //       [grad_u phi + Ju^T l]
+  // b = - [grad_m phi + Jm^T l]
+  //       [          c        ]
+  BlockVector gradphi(block_offsetsx); gradphi = 0.0;
+  BlockVector JTl(block_offsetsx); JTl = 0.0;
+  Dxphi(x, mu, gradphi);
+  
+  (A.GetBlock(0,2)).Mult(l, JTl.GetBlock(0));
+  (A.GetBlock(1,2)).Mult(l, JTl.GetBlock(1));
+
+  for(int ii = 0; ii < 2; ii++)
+  {
+    b.GetBlock(ii).Set(1.0, gradphi.GetBlock(ii));
+    b.GetBlock(ii).Add(1.0, JTl.GetBlock(ii));
+  }
+  if(!socSolve) 
+  {
+    optProblem->c(x, b.GetBlock(2));
+  }
+  else
+  {
+    b.GetBlock(2).Set(1.0, ckSoc);
+  }
+  b *= -1.0; 
+  Xhat = 0.0;
+
+
+  #ifdef MFEM_USE_SUITESPARSE
+    // Direct solve for IP-Newton saddle-point system
+    // A = [ [ Huu   0    Ju^T]
+    //       [  0    D     -I ]
+    //       [ Ju   -I      0 ]]
+    if(linSolver == 0)
+    {
+      BlockMatrix ABlockMatrix(block_offsetsuml, block_offsetsuml);
+      for(int ii = 0; ii < 3; ii++)
+      {
+        for(int jj = 0; jj < 3; jj++)
+        {
+          if(!A.IsZeroBlock(ii, jj))
+          {
+            ABlockMatrix.SetBlock(ii, jj, dynamic_cast<SparseMatrix *>(&(A.GetBlock(ii, jj))));
+          }
+        }
+      }
+      /* direct solve of the 3x3 IP-Newton linear system */
+      UMFPackSolver ASolver;
+      SparseMatrix *ASparse = ABlockMatrix.CreateMonolithic();
+      ASolver.SetOperator(*ASparse);
+      ASolver.Mult(b, Xhat);
+
+      Vector residual(Xhat.Size());
+      ASparse->Mult(Xhat, residual);
+      residual.Add(-1.0, b);
+      delete ASparse;
+    }
+    else if(linSolver == 1)
+    {
+      // Direct solve for 0,0 Schur complement of IP-Newton system, Huu + Ju^T Wmm Ju,
+      // where Wmm = D for contact problems
+      SparseMatrix * Huuloc = dynamic_cast<SparseMatrix *>(&(A.GetBlock(0, 0)));
+      SparseMatrix * Wmmloc = dynamic_cast<SparseMatrix *>(&(A.GetBlock(1, 1)));
+      SparseMatrix * Juloc  = dynamic_cast<SparseMatrix *>(&(A.GetBlock(2, 0)));
+      SparseMatrix * JuTloc = dynamic_cast<SparseMatrix *>(&(A.GetBlock(0, 2)));
+      Vector DVec(dimM); DVec = 0.0;
+      Vector one(dimM); one = 1.0;
+      D->Mult(one, DVec);
+      SparseMatrix *JuTDJu = Mult_AtDA(*Juloc, DVec);     // Ju^T D Ju
+      SparseMatrix *Areduced = Add(*Huuloc, *JuTDJu);  // Huu + Ju^T D Ju
+
+      
+      /* prepare the reduced rhs */
+      // breduced = bu + Ju^T (bm + Wmm bl)
+      Vector breduced(dimU); breduced = 0.0;
+      Vector tempVec(dimM); tempVec = 0.0;
+      Wmmloc->Mult(b.GetBlock(2), tempVec);
+      tempVec.Add(1.0, b.GetBlock(1));
+      JuTloc->Mult(tempVec, breduced);
+      breduced.Add(1.0, b.GetBlock(0));
+
+      // solve the reduced linear system
+      UMFPackSolver AreducedSolver;
+      AreducedSolver.SetOperator(*Areduced);
+      AreducedSolver.Mult(breduced, Xhat.GetBlock(0));
+
+      // now propagate solved uhat to obtain mhat and lhat
+      // xm = Ju xu - bl
+      Juloc->Mult(Xhat.GetBlock(0), Xhat.GetBlock(1));
+      Xhat.GetBlock(1).Add(-1.0, b.GetBlock(2));
+
+      // xl = Wmm xm - bm
+      Wmmloc->Mult(Xhat.GetBlock(1), Xhat.GetBlock(2));
+      Xhat.GetBlock(2).Add(-1.0, b.GetBlock(1));
+
+      delete JuTDJu;
+      delete Areduced;
+    }
+  #else
+    MFEM_VERIFY(linSolver > 1, "linSolver = 0, 1 require MFEM_USE_SUITESPARSE=YES");
+  #endif
+  if(linSolver > 1)
+  { 
+    // Iterative solve for 0,0 Schur complement of IP-Newton system, Huu + Ju^T Wmm Ju,
+    // where Wmm = D for contact problems
+    // here the iterative solver is a Jacobi-preconditioned CG-solve
+    SparseMatrix * Huuloc = dynamic_cast<SparseMatrix *>(&(A.GetBlock(0, 0)));
+    SparseMatrix * Wmmloc = dynamic_cast<SparseMatrix *>(&(A.GetBlock(1, 1)));
+    SparseMatrix * Juloc  = dynamic_cast<SparseMatrix *>(&(A.GetBlock(2, 0)));
+    SparseMatrix * JuTloc = dynamic_cast<SparseMatrix *>(&(A.GetBlock(0, 2)));
+    Vector DVec(dimM); DVec = 0.0;
+    Vector one(dimM); one = 1.0;
+    D->Mult(one, DVec);
+    SparseMatrix *JuTDJu   = Mult_AtDA(*Juloc, DVec);     // Ju^T D Ju
+    SparseMatrix *Areduced = Add(*Huuloc, *JuTDJu);  // Huu + Ju^T D Ju
+
+    /* prepare the reduced rhs */
+    // breduced = bu + Ju^T (bm + Wmm bl)
+    Vector breduced(dimU); breduced = 0.0;
+    Vector tempVec(dimM); tempVec = 0.0;
+    Wmmloc->Mult(b.GetBlock(2), tempVec);
+    tempVec.Add(1.0, b.GetBlock(1));
+    JuTloc->Mult(tempVec, breduced);
+    breduced.Add(1.0, b.GetBlock(0));
+    
+    if (linSolver == 2)
+    {
+      /* Jacobi preconditioned conjugate-gradient solve */
+      DSmoother AreducedPrec((SparseMatrix &)(*Areduced));
+      CGSolver AreducedSolver;
+      AreducedSolver.SetOperator(*Areduced);
+      AreducedSolver.SetAbsTol(1.e-12);
+      AreducedSolver.SetRelTol(1.e-8);
+      AreducedSolver.SetMaxIter(500);
+      AreducedSolver.SetPreconditioner(AreducedPrec);
+      AreducedSolver.SetPrintLevel(1);
+      AreducedSolver.Mult(breduced, Xhat.GetBlock(0));
+    }
+    else
+    {
+      /* Gauss-Seidel preconditioned GMRES solve */
+      GSSmoother AreducedPrec((SparseMatrix &)(*Areduced));
+      GMRESSolver AreducedSolver;
+      AreducedSolver.SetOperator(*Areduced);
+      AreducedSolver.SetAbsTol(1.e-12);
+      AreducedSolver.SetRelTol(1.e-8);
+      AreducedSolver.SetMaxIter(500);
+      AreducedSolver.SetPreconditioner(AreducedPrec);
+      AreducedSolver.SetPrintLevel(1);
+      AreducedSolver.Mult(breduced, Xhat.GetBlock(0));
+    }
+    
+    // now propagate solved uhat to obtain mhat and lhat
+    // xm = Ju xu - bl
+    Juloc->Mult(Xhat.GetBlock(0), Xhat.GetBlock(1));
+    Xhat.GetBlock(1).Add(-1.0, b.GetBlock(2));
+
+    // xl = Wmm xm - bm
+    Wmmloc->Mult(Xhat.GetBlock(1), Xhat.GetBlock(2));
+    Xhat.GetBlock(2).Add(-1.0, b.GetBlock(1));
+
+    delete JuTDJu;
+    delete Areduced;
+  }
+    
+  
+  /* backsolve to determine zlhat */
+  for(int ii = 0; ii < dimM; ii++)
+  {
+    zlhat(ii) = -1.*(zl(ii) + (zl(ii) * Xhat(ii + dimU) - mu) / (x(ii + dimU) - ml(ii)) );
+  }
+
+  // free memory
+  if(Hmm != nullptr)
+  {
+    delete Wmm;
+  }
+  if( Huucl != nullptr)
+  {
+    delete HLuu; HLuu = nullptr;
+  }
+  delete D;
+  delete JuT;
+  delete JmT;
+}
+
+// here Xhat, X will be BlockVectors w.r.t. the 4 partitioning X = (u, m, l, zl)
+
+void InteriorPointSolver::lineSearch(BlockVector& X0, BlockVector& Xhat, double mu)
+{
+  double tau  = max(tauMin, 1.0 - mu);
+  Vector u0   = X0.GetBlock(0);
+  Vector m0   = X0.GetBlock(1);
+  Vector l0   = X0.GetBlock(2);
+  Vector z0   = X0.GetBlock(3);
+  Vector uhat = Xhat.GetBlock(0);
+  Vector mhat = Xhat.GetBlock(1);
+  Vector lhat = Xhat.GetBlock(2);
+  Vector zhat = Xhat.GetBlock(3);
+  double alphaMax  = MaxStepSize(m0, ml, mhat, tau);
+  double alphaMaxz = MaxStepSize(z0, zhat, tau);
+  alphaz = alphaMaxz;
+
+
+  BlockVector x0(block_offsetsx); x0 = 0.0;
+  x0.GetBlock(0).Set(1.0, u0);
+  x0.GetBlock(1).Set(1.0, m0);
+  
+  BlockVector xhat(block_offsetsx); xhat = 0.0;
+  xhat.GetBlock(0).Set(1.0, uhat);
+  xhat.GetBlock(1).Set(1.0, mhat);
+  
+  BlockVector xtrial(block_offsetsx); xtrial = 0.0;
+  BlockVector Dxphi0(block_offsetsx); Dxphi0 = 0.0;
+  int maxBacktrack = 20;
+  alpha = alphaMax;
+
+
+  Vector ck0(dimC); ck0 = 0.0;
+  Vector zhatsoc(dimM); zhatsoc = 0.0;
+  BlockVector Xhatumlsoc(block_offsetsuml); Xhatumlsoc = 0.0;
+  BlockVector xhatsoc(block_offsetsx); xhatsoc = 0.0;
+  Vector uhatsoc(dimU); uhatsoc = 0.0;
+  Vector mhatsoc(dimM); mhatsoc = 0.0;
+
+  Dxphi(x0, mu, Dxphi0);
+  Dxphi0_xhat = InnerProduct(Dxphi0, xhat);
+  descentDirection = Dxphi0_xhat < 0. ? true : false;
+  if(descentDirection)
+  {
+    mfem::out << "is a descent direction for the log-barrier objective\n";
+  }
+  else
+  {
+    mfem::out << "is not a descent direction for the log-barrier objective\n";
+  }
+  mfem::out << "Dxphi^T xhat / (|| Dxphi||_2 * || xhat ||_2) = " << Dxphi0_xhat / (xhat.Norml2() * Dxphi0.Norml2()) << endl;
+  thx0 = theta(x0);
+  phx0 = phi(x0, mu);
+
+  lineSearchSuccess = false;
+  for(int i = 0; i < maxBacktrack; i++)
+  {
+    mfem::out << "\n--------- alpha = " << alpha << " ---------\n";
+
+    // ----- A-5.2. Compute trial point: xtrial = x0 + alpha_i xhat
+    xtrial.Set(1.0, x0);
+    xtrial.Add(alpha, xhat);
+
+    // ------ A-5.3. if not in filter region go to A.5.4 otherwise go to A-5.5.
+    thxtrial = theta(xtrial);
+    phxtrial = phi(xtrial, mu);
+
+    filterCheck(thxtrial, phxtrial);    
+    if(!inFilterRegion)
+    {
+      mfem::out << "not in filter region :)\n";
+      // ------ A.5.4: Check sufficient decrease
+      if(!descentDirection)
+      {
+        switchCondition = false;
+      }
+      else
+      {
+        switchCondition = (alpha * pow(abs(Dxphi0_xhat), sPhi) > delta * pow(thx0, sTheta)) ? true : false;
+      }
+      mfem::out << "theta(x0) = "     << thx0     << ", thetaMin = "                  << thetaMin             << endl;
+      mfem::out << "theta(xtrial) = " << thxtrial << ", (1-gTheta) *theta(x0) = "     << (1. - gTheta) * thx0 << endl;
+      mfem::out << "phi(xtrial) = "   << phxtrial << ", phi(x0) - gPhi *theta(x0) = " << phx0 - gPhi * thx0   << endl;
+      
+      // Case I      
+      if(thx0 <= thetaMin && switchCondition)
+      {
+        sufficientDecrease = phxtrial <= phx0 + eta * alpha * Dxphi0_xhat ? true : false;
+        if(sufficientDecrease)
+        {
+          mfem::out << "Accepted step length -- sufficient decrease in log-barrier objective.\n"; 
+          // accept the trial step
+          lineSearchSuccess = true;
+          break;
+        }
+      }
+      else
+      {
+        if(thxtrial <= (1. - gTheta) * thx0 || phxtrial <= phx0 - gPhi * thx0)
+        {
+          mfem::out << "Accepted step length -- decrease in either constraint violation or log-barrier objective.\n"; 
+          // accept the trial step
+          lineSearchSuccess = true;
+          break;
+        }
+      }
+      // A-5.5: Initialize the second-order correction
+      if((!(thx0 < thxtrial)) && i == 0)
+      {
+        mfem::out << "second order correction\n";
+        optProblem->c(xtrial, ckSoc);
+        optProblem->c(x0, ck0);
+        ckSoc.Add(alphaMax, ck0);
+        // A-5.6 Compute the second-order correction.
+        IPNewtonSolve(x0, l0, z0, zhatsoc, Xhatumlsoc, mu, true);
+        mhatsoc.Set(1.0, Xhatumlsoc.GetBlock(1));
+        // alphasoc = MaxStepSize(m0, ml, mhatsoc, tau);
+        //WARNING: not complete but currently solver isn't entering this region
+      }
+    }
+    else
+    {
+      mfem::out << "in filter region\n";
+    }
+
+    // include more if needed
+    alpha *= 0.5;
+
+  } 
+}
+
+
+void InteriorPointSolver::projectZ(const Vector &x, Vector &z, double mu)
+{
+  double zi;
+  double mudivmml;
+  for(int i = 0; i < dimM; i++)
+  {
+    zi = z(i);
+    mudivmml = mu / (x(i + dimU) - ml(i));
+    z(i) = max(min(zi, kSig * mudivmml), mudivmml / kSig);
+  }
+}
+
+void InteriorPointSolver::filterCheck(double th, double ph)
+{
+  inFilterRegion = false;
+  if(th > thetaMax)
+  {
+    inFilterRegion = true;
+  }
+  else
+  {
+    for(int i = 0; i < F1.Size(); i++)
+    {
+      if(th >= F1[i] && ph >= F2[i])
+      {
+        inFilterRegion = true;
+        break;
+      }
+    }
+  }
+}
+
+double InteriorPointSolver::E(const BlockVector &x, const Vector &l, const Vector &zl, double mu, bool print)
+{
+  double E1, E2, E3;
+  double sc, sd;
+  BlockVector gradL(block_offsetsx); gradL = 0.0; // stationarity grad L = grad f + J^T l - z
+  Vector cx(dimC); cx = 0.0;     // feasibility c = c(x)
+  Vector comp(dimM); comp = 0.0; // complementarity M Z - mu 1
+
+  DxL(x, l, zl, gradL);
+  E1 = gradL.Normlinf();
+
+  optProblem->c(x, cx);
+  E2 = cx.Normlinf();
+
+  for(int ii = 0; ii < dimM; ii++) 
+  { 
+    comp(ii) = x(dimU + ii) * zl(ii) - mu;
+  }
+  E3 = comp.Normlinf();
+
+  double ll1, zl1;
+  zl1 = zl.Norml1() / double(dimC + dimM);
+  ll1 =  l.Norml1();
+  sc = max(sMax, zl1 / (double(dimM)) ) / sMax;
+  sd = max(sMax, (ll1 + zl1) / (double(dimC + dimM))) / sMax;
+  if(print)
+  {
+    mfem::out << "evaluating optimality error for mu = " << mu << endl;
+    mfem::out << "stationarity measure = "    << E1 / sd << endl;
+    mfem::out << "feasibility measure  = "    << E2      << endl;
+    mfem::out << "complimentarity measure = " << E3 / sc << endl;
+  }
+  return max(max(E1 / sd, E2), E3 / sc);
+}
+
+double InteriorPointSolver::E(const BlockVector &x, const Vector &l, const Vector &zl, bool print)
+{
+  return E(x, l, zl, 0.0, print);
+}
+
+double InteriorPointSolver::theta(const BlockVector &x)
+{
+  Vector cx(dimC); cx = 0.0;
+  optProblem->c(x, cx);
+  return cx.Norml2();
+}
+
+// log-barrier objective
+double InteriorPointSolver::phi(const BlockVector &x, double mu)
+{
+  double fx = optProblem->CalcObjective(x); 
+  double logBarrierLoc = 0.0;
+  for(int i = 0; i < dimM; i++) 
+  { 
+    logBarrierLoc += log(x(dimU+i)-ml(i));
+  }
+  double logBarrierGlb = 0.0;
+  logBarrierGlb = logBarrierLoc;
+  return fx - mu * logBarrierGlb;
+}
+
+
+
+
+// gradient of log-barrier objective with respect to x = (u, m)
+void InteriorPointSolver::Dxphi(const BlockVector &x, double mu, BlockVector &y)
+{
+  optProblem->CalcObjectiveGrad(x, y);
+  for(int i = 0; i < dimM; i++) 
+  { 
+    y(dimU + i) -= mu / (x(dimU + i));
+  } 
+}
+
+
+// Lagrangian function evaluation
+// L(x, l, zl) = f(x) + l^T c(x) - zl^T m
+double InteriorPointSolver::L(const BlockVector &x, const Vector &l, const Vector &zl)
+{
+  double fx = optProblem->CalcObjective(x);
+  Vector cx(dimC); optProblem->c(x, cx);
+  return (fx + InnerProduct(cx, l) - InnerProduct(x.GetBlock(1), zl));
+}
+
+void InteriorPointSolver::DxL(const BlockVector &x, const Vector &l, const Vector &zl, BlockVector &y)
+{
+  // evaluate the gradient of the objective with respect to the primal variables x = (u, m)
+  BlockVector gradxf(block_offsetsx); gradxf = 0.0;
+  optProblem->CalcObjectiveGrad(x, gradxf);
+  
+  SparseMatrix *Jacu, *Jacm, *JacuT, *JacmT;
+  Jacu = optProblem->Duc(x); Jacm = optProblem->Dmc(x);
+  JacuT = Transpose(*Jacu);
+  JacmT = Transpose(*Jacm);
+  JacuT->Mult(l, y.GetBlock(0));
+  JacmT->Mult(l, y.GetBlock(1));
+  delete JacuT; 
+  delete JacmT;
+  y.Add(1.0, gradxf);
+  (y.GetBlock(1)).Add(-1.0, zl);
+}
+
+
+bool InteriorPointSolver::GetConverged() const
+{
+  return converged;
+}
+
+void InteriorPointSolver::SetTol(double Tol)
+{
+  rel_tol = Tol;
+}
+
+void InteriorPointSolver::SetMaxIter(int max_it)
+{
+  max_iter = max_it;
+}
+
+void InteriorPointSolver::SetBarrierParameter(double mu_0)
+{
+  mu_k = mu_0;
+}
+
+void InteriorPointSolver::SaveLogBarrierHessianIterates(bool save)
+{
+  MFEM_ASSERT(MyRank == 0 || save == false, "currently can only save logbarrier hessian in serial codes");
+  saveLogBarrierIterates = save;
+}
+
+void InteriorPointSolver::SetLinearSolver(int LinSolver)
+{
+  linSolver = LinSolver;
+}
+
+
+
+InteriorPointSolver::~InteriorPointSolver() 
+{
+  F1.DeleteAll();
+  F2.DeleteAll();
+  block_offsetsx.DeleteAll();
+  block_offsetsumlz.DeleteAll();
+  block_offsetsuml.DeleteAll();
+  ml.SetSize(0);
+}

examples/contact/IPsolver.hpp

@@ -0,0 +1,80 @@
+#include "mfem.hpp"
+#include "Problems.hpp"
+#include <fstream>
+#include <iostream>
+
+using namespace std;
+using namespace mfem;
+
+
+#ifndef IPSOLVER 
+#define IPSOLVER
+
+class InteriorPointSolver
+{
+protected:
+    GeneralOptProblem* optProblem;
+    double rel_tol;
+    int  max_iter;
+    double mu_k; // \mu_k
+    Vector lk, zlk;
+
+    double sMax, kSig, tauMin, eta, thetaMin, delta, sTheta, sPhi, kMu, thetaMu;
+    double thetaMax, kSoc, gTheta, gPhi, kEps;
+	
+    // filter
+    Array<double> F1, F2;
+	
+    // quantities computed in lineSearch
+    double alpha, alphaz;
+    double thx0, thxtrial;
+    double phx0, phxtrial;
+    bool descentDirection, switchCondition, sufficientDecrease, lineSearchSuccess, inFilterRegion;
+    double Dxphi0_xhat;
+
+    int dimU, dimM, dimC;
+    Array<int> block_offsetsumlz, block_offsetsuml, block_offsetsx;
+    Vector ml;
+
+    Vector ckSoc;
+    SparseMatrix * Huu, * Hum, * Hmu, * Hmm, * Wmm, *D, * Ju, * Jm, * JuT, * JmT;
+    SparseMatrix * Huucl, *HLuu;    
+    int jOpt;
+    bool converged;
+    
+    int MyRank;
+    bool iAmRoot;
+
+    bool saveLogBarrierIterates;
+
+    int linSolver;
+
+public:
+    InteriorPointSolver(GeneralOptProblem*);
+    void Mult(const BlockVector& , BlockVector&); // used when the user wants to be aware of bound-constrained variable m >= ml
+    void Mult(const Vector&, Vector &); // useful when the user doesn't need to know about bound-constrained variable   m >= ml, e.g., when m is a slack variable
+    double MaxStepSize(Vector& , Vector& , Vector& , double);
+    double MaxStepSize(Vector& , Vector& , double);
+    void FormIPNewtonMat(BlockVector& , Vector& , Vector& , BlockOperator &);
+    void IPNewtonSolve(BlockVector& , Vector& , Vector& , Vector&, BlockVector& , double, bool);
+    void lineSearch(BlockVector& , BlockVector& , double);
+    void projectZ(const Vector & , Vector &, double);
+    void filterCheck(double, double);
+    double E(const BlockVector &, const Vector &, const Vector &, double, bool);
+    double E(const BlockVector &, const Vector &, const Vector &, bool);
+    bool GetConverged() const;
+    // TO DO: include Hessian of Lagrangian
+    double theta(const BlockVector &);
+    double phi(const BlockVector &, double);
+    void Dxphi(const BlockVector &, double, BlockVector &);
+    double L(const BlockVector &, const Vector &, const Vector &);
+    void DxL(const BlockVector &, const Vector &, const Vector &, BlockVector &);
+    void SetTol(double);
+    void SetMaxIter(int);
+    void SetBarrierParameter(double);    
+    void SaveLogBarrierHessianIterates(bool);
+    void SetLinearSolver(int);
+    virtual ~InteriorPointSolver();
+};
+
+#endif

examples/contact/ObstacleProblem.cpp

@@ -0,0 +1,125 @@
+#include "mfem.hpp"
+#include "Problems.hpp"
+#include "IPsolver.hpp"
+#include <fstream>
+#include <iostream>
+
+using namespace std;
+using namespace mfem;
+
+
+
+double dmanufacturedFun(const Vector &);
+double fRhs(const Vector &);
+double obstacle(const Vector &);
+
+int main(int argc, char *argv[])
+{
+  int FEorder = 1; // order of the finite elements
+  int linSolver = 0;
+  int maxIPMiters = 30;
+  bool iAmRoot = true;
+  
+  OptionsParser args(argc, argv);
+  args.AddOption(&FEorder, "-o", "--order",\
+		  "Order of the finite elements.");
+  args.AddOption(&linSolver, "-linSolver", "--linearSolver", \
+       "IP-Newton linear system solution strategy.");
+  args.AddOption(&maxIPMiters, "-IPMiters", "--IPMiters",\
+		  "Maximum number of IPM iterations");
+  
+  args.Parse();
+  if(!args.Good())
+  {
+    args.PrintUsage(cout);
+    return 1;
+  }
+  else
+  {
+    if( iAmRoot )
+    {
+      args.PrintOptions(cout);
+    }
+  }
+
+  const char *meshFile = "../../data/inline-quad.mesh";
+  Mesh *mesh = new Mesh(meshFile, 1, 1);
+  int dim = mesh->Dimension(); // geometric dimension of the domain
+  {
+     int ref_levels = 3;
+     for (int l = 0; l < ref_levels; l++)
+     {
+        mesh->UniformRefinement();
+     }
+  }
+
+  FiniteElementCollection *fec = new H1_FECollection(FEorder, dim);
+  FiniteElementSpace      *Vh  = new FiniteElementSpace(mesh, fec);
+  ObstacleProblem problem(Vh, &fRhs, &obstacle);
+  
+  int dimD = problem.GetDimD();
+  Vector x0(dimD); x0 = 0.0;
+  Vector xf(dimD); xf = 0.0;
+
+  InteriorPointSolver optimizer(&problem); 
+  optimizer.SetTol(1.e-7);
+  optimizer.SetLinearSolver(linSolver);
+  optimizer.SetMaxIter(maxIPMiters);
+  optimizer.Mult(x0, xf);
+
+  double Einitial = problem->E(x0);
+  double Efinal = problem->E(xf);
+  cout << "Energy objective at initial point = " << Einitial << endl;
+  cout << "Energy objective at QP optimizer = " << Efinal << endl;
+
+
+  GridFunction d_gf(Vh);
+
+  d_gf = xf;
+
+  FunctionCoefficient dm_fc(dmanufacturedFun); // pseudo-manufactured solution
+  GridFunction dm_gf(Vh);
+  dm_gf.ProjectCoefficient(dm_fc);
+  
+  ParaViewDataCollection paraview_dc("BarrierProblemSolution", mesh);
+  paraview_dc.SetPrefixPath("ParaView");
+  paraview_dc.SetLevelsOfDetail(FEorder);
+  paraview_dc.SetDataFormat(VTKFormat::BINARY);
+  paraview_dc.SetHighOrderOutput(true);
+  paraview_dc.SetCycle(0);
+  paraview_dc.SetTime(0.0);
+  paraview_dc.RegisterField("d(x) (numerical)", &d_gf);
+  paraview_dc.RegisterField("d(x) (pseudo-manufactured)", &dm_gf);
+  paraview_dc.Save();
+  
+  
+  delete Vh;
+  delete fec;
+  delete mesh;
+  return 0;
+}
+
+
+double dmanufacturedFun(const Vector &x)
+{
+  return cos(2*M_PI*x(0)) + 0.2 - 2.0*(pow(x(0),3) - 1.5*pow(x(0),2));
+}
+
+
+// f(x) forcing term... which enters the objective energy functional
+// E(d) = 0.5 d^T K d - f^T d, where f is a discrete vector representation
+// of f(x). f(x) is such that in the absence of bound-constraints then
+// the solution of the optimization problem satisfies the PDE
+// -div(grad(d)) + d = f + homogeneous Neumann conditions on the unit interval,
+// for d(x) = cos(2 \pi x) + a0 + a3 (x^3 - 1.5 x^2), a2 = 0.2, a3 = -2
+double fRhs(const Vector &x)
+{
+  double fx = 0.;
+  fx = 0.2 - 2.0 * (pow(x(0),3)- 1.5*pow(x(0),2.) - 6 * x(0) + 3.) + (1. + pow(2.*M_PI,2))*cos(2.*M_PI*x(0));
+  return fx;
+}
+
+double obstacle(const Vector &x)
+{
+  return 0.0;
+}

examples/contact/ParIPsolver.cpp

@@ -0,0 +1,834 @@
+#include "mfem.hpp"
+#include "ParIPsolver.hpp"
+#include "ParProblems.hpp"
+#include <fstream>
+#include <iostream>
+#include <cstdlib>
+
+using namespace std;
+using namespace mfem;
+
+
+ParInteriorPointSolver::ParInteriorPointSolver(ParGeneralOptProblem * problem_) 
+                     : problem(problem_), 
+                       block_offsetsumlz(5), block_offsetsuml(4), block_offsetsx(3),
+                       Huu(nullptr), Hum(nullptr), Hmu(nullptr), 
+                       Hmm(nullptr), Wmm(nullptr), D(nullptr), 
+                       Ju(nullptr), Jm(nullptr), JuT(nullptr), JmT(nullptr), 
+                       saveLogBarrierIterates(false)
+{
+   OptTol  = 1.e-2;
+   max_iter = 20;
+   mu_k     = 1.0;
+
+   sMax     = 1.e2;
+   kSig     = 1.e10;   // control deviation from primal Hessian
+   tauMin   = 0.8;     // control rate at which iterates can approach the boundary
+   eta      = 1.e-4;   // backtracking constant
+   thetaMin = 1.e-4;   // allowed violation of the equality constraints
+
+   // constants in line-step A-5.4
+   delta    = 1.0;
+   sTheta   = 1.1;
+   sPhi     = 2.3;
+
+   // control the rate at which the penalty parameter is decreased
+   kMu     = 0.2;
+   thetaMu = 1.5;
+
+   thetaMax = 1.e6; // maximum constraint violation
+   // data for the second order correction
+   kSoc     = 0.99;
+
+   // equation (18)
+   gTheta = 1.e-5;
+   gPhi   = 1.e-5;
+
+   kEps   = 1.e1;
+
+   dimU = problem->GetDimU();
+   dimM = problem->GetDimM();
+   dimC = problem->GetDimC();
+   MPI_Allreduce(&dimU, &dimUglb, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+   MPI_Allreduce(&dimM, &dimMglb, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+   MPI_Allreduce(&dimC, &dimCglb, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
+   ckSoc.SetSize(dimC);
+  
+   block_offsetsumlz[0] = 0;
+   block_offsetsumlz[1] = dimU; // u
+   block_offsetsumlz[2] = dimM; // m
+   block_offsetsumlz[3] = dimC; // lambda
+   block_offsetsumlz[4] = dimM; // zl
+   block_offsetsumlz.PartialSum();
+
+   for(int i = 0; i < block_offsetsuml.Size(); i++)  
+   { 
+      block_offsetsuml[i] = block_offsetsumlz[i]; 
+   }
+   for(int i = 0; i < block_offsetsx.Size(); i++)    
+   { 
+      block_offsetsx[i] = block_offsetsuml[i] ; 
+   }
+
+  
+   ml = problem->Getml();
+  
+   lk.SetSize(dimC);  lk  = 0.0;
+   zlk.SetSize(dimM); zlk = 0.0;
+
+   linSolver = 0;
+   linSolveTol = 1.e-8;
+   MyRank = Mpi::WorldRank();
+   iAmRoot = MyRank == 0 ? true : false;
+}
+
+double ParInteriorPointSolver::MaxStepSize(Vector &x, Vector &xl, Vector &xhat, double tau)
+{
+   double alphaMaxloc = 1.0;
+   double alphaTmp;
+   for(int i = 0; i < x.Size(); i++)
+   {   
+      if( xhat(i) < 0. )
+      {
+         alphaTmp = -1. * tau * (x(i) - xl(i)) / xhat(i);
+         alphaMaxloc = min(alphaMaxloc, alphaTmp);
+      } 
+   }
+
+   // alphaMaxloc is the local maximum step size which is
+   // distinct on each MPI process. Need to compute
+   // the global maximum step size 
+   double alphaMaxglb;
+   MPI_Allreduce(&alphaMaxloc, &alphaMaxglb, 1, MPI_DOUBLE, MPI_MIN, MPI_COMM_WORLD);
+   return alphaMaxglb;
+}
+
+double ParInteriorPointSolver::MaxStepSize(Vector &x, Vector &xhat, double tau)
+{
+   Vector zero(x.Size()); zero = 0.0;
+   return MaxStepSize(x, zero, xhat, tau);
+}
+
+
+void ParInteriorPointSolver::Mult(const Vector &x0, Vector &xf)
+{
+   BlockVector x0block(block_offsetsx); x0block = 0.0;
+   x0block.GetBlock(0).Set(1.0, x0);
+   x0block.GetBlock(1) = 100.;
+   x0block.GetBlock(1).Add(1.0, ml);
+   BlockVector xfblock(block_offsetsx); xfblock = 0.0;
+   Mult(x0block, xfblock);
+   xf.Set(1.0, xfblock.GetBlock(0));
+}
+
+
+void ParInteriorPointSolver::Mult(const BlockVector &x0, BlockVector &xf)
+{
+   converged = false;
+   
+   BlockVector xk(block_offsetsx), xhat(block_offsetsx); xk = 0; xhat = 0.0;
+   BlockVector Xk(block_offsetsumlz), Xhat(block_offsetsumlz); Xk = 0.0; Xhat = 0.0;
+   BlockVector Xhatuml(block_offsetsuml); Xhatuml = 0.0;
+   Vector zlhat(dimM); zlhat = 0.0;
+
+   xk.GetBlock(0).Set(1.0, x0.GetBlock(0));
+   xk.GetBlock(1).Set(1.0, x0.GetBlock(1));
+   // running estimate of the final values of the Lagrange multipliers
+   lk  = 0.0;
+   zlk = 0.0;
+
+   for(int i = 0; i < dimM; i++)
+   {
+      zlk(i) = 1.e1 * mu_k / (xk(i+dimU) - ml(i));
+   }
+
+   Xk.GetBlock(0).Set(1.0, xk.GetBlock(0));
+   Xk.GetBlock(1).Set(1.0, xk.GetBlock(1));
+   Xk.GetBlock(2).Set(1.0, lk);
+   Xk.GetBlock(3).Set(1.0, zlk);
+
+  /* set theta0 = theta(x0)
+   *     thetaMin
+   *     thetaMax
+   * when theta(xk) < thetaMin and the switching condition holds
+   * then we ask for the Armijo sufficient decrease of the barrier
+   * objective to be satisfied, in order to accept the trial step length alphakl
+   * 
+   * thetaMax controls how the filter is initialized for each log-barrier subproblem
+   * F0 = {(th, phi) s.t. th > thetaMax}
+   * that is the filter does not allow for iterates where the constraint violation
+   * is larger than that of thetaMax
+   */
+   double theta0 = theta(xk);
+   thetaMin = 1.e-4 * max(1.0, theta0);
+   thetaMax = 1.e8  * thetaMin; // 1.e4 * max(1.0, theta0)
+
+   double Eeval, maxBarrierSolves, Eevalmu0;
+   bool printOptimalityError; // control optimality error print to console for log-barrier subproblems
+   
+   maxBarrierSolves = 10;
+
+   for(jOpt = 0; jOpt < max_iter; jOpt++)
+   {
+      if(iAmRoot)
+      {
+         cout << "interior-point solve step " << jOpt << endl;
+      }
+      // A-2. Check convergence of overall optimization problem
+      printOptimalityError = false;
+      Eevalmu0 = E(xk, lk, zlk, printOptimalityError);
+      if(Eevalmu0 < OptTol)
+      {
+         converged = true;
+         if(iAmRoot)
+         {
+            cout << "solved optimization problem :)\n";
+         }
+         break;
+      }
+      
+      if(jOpt > 0) { maxBarrierSolves = 1; }
+      
+      for(int i = 0; i < maxBarrierSolves; i++)
+      {
+         // A-3. Check convergence of the barrier subproblem
+         printOptimalityError = true;
+         Eeval = E(xk, lk, zlk, mu_k, printOptimalityError);
+         if(iAmRoot)
+         {
+            cout << "E = " << Eeval << endl;
+         }
+         if(Eeval < kEps * mu_k)
+         {
+            if(iAmRoot)
+            {
+               cout << "solved barrier subproblem :), for mu = " << mu_k << endl;
+            }
+            // A-3.1. Recompute the barrier parameter
+            mu_k  = max(OptTol / 10., min(kMu * mu_k, pow(mu_k, thetaMu)));
+            // A-3.2. Re-initialize the filter
+            F1.DeleteAll();
+            F2.DeleteAll();
+         }
+         else
+         {
+            break;
+         }
+      }
+    
+      // A-4. Compute the search direction
+      // solve for (uhat, mhat, lhat)
+      if(iAmRoot)
+      {
+         cout << "\n** A-4. IP-Newton solve **\n";
+      }
+      zlhat = 0.0; Xhatuml = 0.0;
+      // why do we have Xhatuml ....???
+      // TO DO: remove Xhatuml in favor of passing Xhat
+      IPNewtonSolve(xk, lk, zlk, zlhat, Xhatuml, mu_k, false); 
+
+      // assign data stack, X = (u, m, l, zl)
+      Xk = 0.0;
+      Xk.GetBlock(0).Set(1.0, xk.GetBlock(0));
+      Xk.GetBlock(1).Set(1.0, xk.GetBlock(1));
+      Xk.GetBlock(2).Set(1.0, lk);
+      Xk.GetBlock(3).Set(1.0, zlk);
+
+      // assign data stack, Xhat = (uhat, mhat, lhat, zlhat)
+      Xhat = 0.0;
+      for(int i = 0; i < 3; i++)
+      {
+         Xhat.GetBlock(i).Set(1.0, Xhatuml.GetBlock(i));
+      }
+      Xhat.GetBlock(3).Set(1.0, zlhat);
+
+      // A-5. Backtracking line search.
+      if(iAmRoot)
+      {
+         cout << "\n** A-5. Linesearch **\n";
+         cout << "mu = " << mu_k << endl;
+      }
+      lineSearch(Xk, Xhat, mu_k);
+
+      if(lineSearchSuccess)
+      {
+         if(iAmRoot)
+         {
+            cout << "lineSearch successful :)\n";
+         }
+         if(!switchCondition || !sufficientDecrease)
+         {
+            F1.Append( (1. - gTheta) * thx0);
+            F2.Append( phx0 - gPhi * thx0);
+         }
+         // ----- A-6: Accept the trial point
+         // print info regarding zl...
+         xk.GetBlock(0).Add(alpha, Xhat.GetBlock(0));
+         xk.GetBlock(1).Add(alpha, Xhat.GetBlock(1));
+         lk.Add(alpha,   Xhat.GetBlock(2));
+         zlk.Add(alphaz, Xhat.GetBlock(3));
+         projectZ(xk, zlk, mu_k);
+      }
+      else
+      {
+         if(iAmRoot)
+         {
+            cout << "lineSearch not successful :(\n";
+            cout << "attempting feasibility restoration with theta = " << thx0 << endl;
+            cout << "no feasibility restoration implemented, exiting now \n";
+         }
+         break;
+      }
+      if(jOpt + 1 == max_iter && iAmRoot) 
+      {  
+         cout << "maximum optimization iterations :(\n";
+      }
+   }
+   // done with optimization routine, just reassign data to xf reference so
+   // that the application code has access to the optimal point
+   xf = 0.0;
+   xf.GetBlock(0).Set(1.0, xk.GetBlock(0));
+   xf.GetBlock(1).Set(1.0, xk.GetBlock(1));
+}
+
+void ParInteriorPointSolver::FormIPNewtonMat(BlockVector & x, Vector & l, Vector &zl, BlockOperator &Ak)
+{
+   // WARNING: Huu, Hum, Hmu, Hmm should all be Hessian terms of the Lagrangian, currently we 
+   //          them by Hessian terms of the objective function and neglect the Hessian of l^T c
+
+   Huu = problem->Duuf(x); 
+   Hum = problem->Dumf(x);
+   Hmu = problem->Dmuf(x);
+   Hmm = problem->Dmmf(x);
+
+   Vector DiagLogBar(dimM); DiagLogBar = 0.0;
+   for(int ii = 0; ii < dimM; ii++)
+   {
+      DiagLogBar(ii) = zl(ii) / (x(ii+dimU) - ml(ii));
+   }
+   if(saveLogBarrierIterates)
+   {
+      std::ofstream diagStream;
+      char diagString[100];
+      snprintf(diagString, 100, "logBarrierHessiandata/D%d.dat", jOpt);
+      diagStream.open(diagString, ios::out | ios::trunc);
+      for(int ii = 0; ii < dimM; ii++)
+      {
+         diagStream << setprecision(30) << DiagLogBar(ii) << endl;
+      }
+      diagStream.close();
+   } 
+
+   SparseMatrix * Ds = new SparseMatrix(DiagLogBar);
+   ParFiniteElementSpace * fes = problem->GetfesM();
+   D = new HypreParMatrix(fes->GetComm(), fes->GlobalTrueVSize(), fes->GetTrueDofOffsets(), Ds);
+   HypreStealOwnership(*D,*Ds);
+   delete Ds;
+  
+   if(Hmm != nullptr)
+   {
+      Wmm = Hmm;
+      Wmm->Add(1.0, *D);
+   }
+   else
+   {
+      Wmm = D;
+   }
+
+   Ju = problem->Duc(x); JuT = Ju->Transpose();
+   Jm = problem->Dmc(x); JmT = Jm->Transpose();
+  
+   //         IP-Newton system matrix
+   //    Ak = [[H_(u,u)  H_(u,m)   J_u^T]
+   //          [H_(m,u)  W_(m,m)   J_m^T]
+   //          [ J_u      J_m       0  ]]
+
+   Ak.SetBlock(0, 0, Huu);                         Ak.SetBlock(0, 2, JuT);
+                           Ak.SetBlock(1, 1, Wmm); Ak.SetBlock(1, 2, JmT);
+   Ak.SetBlock(2, 0,  Ju); Ak.SetBlock(2, 1,  Jm);
+
+   if(Hum != nullptr) { Ak.SetBlock(0, 1, Hum); Ak.SetBlock(1, 0, Hmu); }
+}
+
+// perturbed KKT system solve
+// determine the search direction
+void ParInteriorPointSolver::IPNewtonSolve(BlockVector &x, Vector &l, Vector &zl, Vector &zlhat, BlockVector &Xhat, double mu, bool socSolve)
+{
+   // solve A x = b, where A is the IP-Newton matrix
+   BlockOperator A(block_offsetsuml, block_offsetsuml); BlockVector b(block_offsetsuml); b = 0.0;
+   FormIPNewtonMat(x, l, zl, A);
+
+   //       [grad_u phi + Ju^T l]
+   // b = - [grad_m phi + Jm^T l]
+   //       [          c        ]
+   BlockVector gradphi(block_offsetsx); gradphi = 0.0;
+   BlockVector JTl(block_offsetsx); JTl = 0.0;
+   Dxphi(x, mu, gradphi);
+   
+   (A.GetBlock(0,2)).Mult(l, JTl.GetBlock(0));
+   (A.GetBlock(1,2)).Mult(l, JTl.GetBlock(1));
+
+   for(int ii = 0; ii < 2; ii++)
+   {
+      b.GetBlock(ii).Set(1.0, gradphi.GetBlock(ii));
+      b.GetBlock(ii).Add(1.0, JTl.GetBlock(ii));
+   }
+   if(!socSolve) 
+   {
+      problem->c(x, b.GetBlock(2));
+   }
+   else
+   {
+      b.GetBlock(2).Set(1.0, ckSoc);
+   }
+   b *= -1.0; 
+   Xhat = 0.0;
+
+
+   // Direct solver (default)
+   if(linSolver == 0)
+   {
+      Array2D<HypreParMatrix *> ABlockMatrix(3,3);
+      for(int ii = 0; ii < 3; ii++)
+      {
+      for(int jj = 0; jj < 3; jj++)
+      {
+         if(!A.IsZeroBlock(ii, jj))
+         {
+            ABlockMatrix(ii, jj) = dynamic_cast<HypreParMatrix *>(&(A.GetBlock(ii, jj)));
+	 }
+	 else
+	 {
+	    ABlockMatrix(ii, jj) = nullptr;
+	 }
+      }
+      }
+      
+      HypreParMatrix * Ah = HypreParMatrixFromBlocks(ABlockMatrix);   
+      
+      /* direct solve of the 3x3 IP-Newton linear system */
+      #ifdef MFEM_USE_MUMPS
+        MUMPSSolver ASolver;
+        ASolver.SetPrintLevel(0);
+        ASolver.SetMatrixSymType(MUMPSSolver::MatType::SYMMETRIC_INDEFINITE);
+        ASolver.SetOperator(*Ah);
+        ASolver.Mult(b, Xhat);
+      #else 
+        #ifdef MFEM_USE_MKL_CPARDISO
+          CPardisoSolver ASolver(MPI_COMM_WORLD);
+          ASolver.SetOperator(*Ah);
+          ASolver.Mult(b, Xhat);
+	#else
+	  MFEM_VERIFY(false, "linSolver 0 will not work unless compiled with MUMPS or MKL");
+        #endif
+      #endif
+
+      delete Ah;
+   }
+   else if(linSolver == 1 || linSolver == 2)
+   {
+      // form A = Huu + Ju^T D Ju, Wmm = D for contact
+      HypreParMatrix * Huuloc = dynamic_cast<HypreParMatrix *>(&(A.GetBlock(0, 0)));
+      HypreParMatrix * Wmmloc = dynamic_cast<HypreParMatrix *>(&(A.GetBlock(1, 1)));
+      HypreParMatrix * Juloc  = dynamic_cast<HypreParMatrix *>(&(A.GetBlock(2, 0)));
+      HypreParMatrix * JuTloc = dynamic_cast<HypreParMatrix *>(&(A.GetBlock(0, 2)));
+      
+      
+      HypreParMatrix *JuTDJu   = RAP(Wmmloc, Juloc);     // Ju^T D Ju
+      HypreParMatrix *Areduced = ParAdd(Huuloc, JuTDJu);  // Huu + Ju^T D Ju
+      /* prepare the reduced rhs */
+      // breduced = bu + Ju^T (bm + Wmm bl)
+      Vector breduced(dimU); breduced = 0.0;
+      Vector tempVec(dimM); tempVec = 0.0;
+      Wmmloc->Mult(b.GetBlock(2), tempVec);
+      tempVec.Add(1.0, b.GetBlock(1));
+      JuTloc->Mult(tempVec, breduced);
+      breduced.Add(1.0, b.GetBlock(0));
+      
+      if(linSolver == 1)
+      {
+         // setup the solver for the reduced linear system
+         #ifdef MFEM_USE_MUMPS
+	   MUMPSSolver AreducedSolver;   
+           AreducedSolver.SetPrintLevel(0);
+           AreducedSolver.SetMatrixSymType(MUMPSSolver::MatType::SYMMETRIC_INDEFINITE);
+	   AreducedSolver.SetOperator(*Areduced);
+	   AreducedSolver.Mult(breduced, Xhat.GetBlock(0));
+         #else 
+           #ifdef MFEM_USE_MKL_CPARDISO
+	     CPardisoSolver AreducedSolver(MPI_COMM_WORLD);
+	     AreducedSolver.SetOperator(*Areduced);
+	     AreducedSolver.Mult(breduced, Xhat.GetBlock(0));
+           #else
+	     MFEM_VERIFY(false, "linSolver 1 will not work unless compiled with MUMPS or MKL");
+           #endif
+         #endif
+      }
+      else
+      {
+         HyprePCG AreducedSolver(MPI_COMM_WORLD);
+	 AreducedSolver.SetOperator(*Areduced);
+	 HypreBoomerAMG AreducedPrec;
+         AreducedSolver.SetTol(linSolveTol);
+         AreducedSolver.SetMaxIter(500);
+         AreducedSolver.SetPreconditioner(AreducedPrec);
+         AreducedSolver.SetResidualConvergenceOptions(); // convergence criteria based on residual norm
+	 AreducedSolver.SetPrintLevel(2);
+         AreducedSolver.Mult(breduced, Xhat.GetBlock(0));
+      }
+
+      // now propagate solved uhat to obtain mhat and lhat
+      // xm = Ju xu - bl
+      Juloc->Mult(Xhat.GetBlock(0), Xhat.GetBlock(1));
+      Xhat.GetBlock(1).Add(-1.0, b.GetBlock(2));
+
+      // xl = Wmm xm - bm
+      Wmmloc->Mult(Xhat.GetBlock(1), Xhat.GetBlock(2));
+      Xhat.GetBlock(2).Add(-1.0, b.GetBlock(1));
+
+      delete JuTDJu;
+      delete Areduced;
+   }
+
+   /* backsolve to determine zlhat */
+   for(int ii = 0; ii < dimM; ii++)
+   {
+      zlhat(ii) = -1.*(zl(ii) + (zl(ii) * Xhat(ii + dimU) - mu) / (x(ii + dimU) - ml(ii)) );
+   }
+
+   // free memory
+   delete D;
+   delete JuT;
+   delete JmT;
+   if(Hmm != nullptr)
+   {
+      delete Wmm;
+   }
+}
+
+// here Xhat, X will be BlockVectors w.r.t. the 4 partitioning X = (u, m, l, zl)
+
+void ParInteriorPointSolver::lineSearch(BlockVector& X0, BlockVector& Xhat, double mu)
+{
+   double tau  = max(tauMin, 1.0 - mu);
+   Vector u0   = X0.GetBlock(0);
+   Vector m0   = X0.GetBlock(1);
+   Vector l0   = X0.GetBlock(2);
+   Vector z0   = X0.GetBlock(3);
+   Vector uhat = Xhat.GetBlock(0);
+   Vector mhat = Xhat.GetBlock(1);
+   Vector lhat = Xhat.GetBlock(2);
+   Vector zhat = Xhat.GetBlock(3);
+   double alphaMax  = MaxStepSize(m0, ml, mhat, tau);
+   double alphaMaxz = MaxStepSize(z0, zhat, tau);
+   alphaz = alphaMaxz;
+
+   BlockVector x0(block_offsetsx); x0 = 0.0;
+   x0.GetBlock(0).Set(1.0, u0);
+   x0.GetBlock(1).Set(1.0, m0);
+   
+   BlockVector xhat(block_offsetsx); xhat = 0.0;
+   xhat.GetBlock(0).Set(1.0, uhat);
+   xhat.GetBlock(1).Set(1.0, mhat);
+   
+   BlockVector xtrial(block_offsetsx); xtrial = 0.0;
+   BlockVector Dxphi0(block_offsetsx); Dxphi0 = 0.0;
+   int maxBacktrack = 20;
+   alpha = alphaMax;
+
+   Vector ck0(dimC); ck0 = 0.0;
+   Vector zhatsoc(dimM); zhatsoc = 0.0;
+   BlockVector Xhatumlsoc(block_offsetsuml); Xhatumlsoc = 0.0;
+   BlockVector xhatsoc(block_offsetsx); xhatsoc = 0.0;
+   Vector uhatsoc(dimU); uhatsoc = 0.0;
+   Vector mhatsoc(dimM); mhatsoc = 0.0;
+
+   Dxphi(x0, mu, Dxphi0);
+
+   Dxphi0_xhat = InnerProduct(MPI_COMM_WORLD, Dxphi0, xhat);
+   descentDirection = Dxphi0_xhat < 0. ? true : false;
+   if(descentDirection)
+   {
+      if (iAmRoot)
+      {
+         cout << "is a descent direction for the log-barrier objective\n";
+      }
+   }
+   else
+   {
+      cout << "is not a descent direction for the log-barrier objective\n";
+   }
+   thx0 = theta(x0);
+   phx0 = phi(x0, mu);
+
+   lineSearchSuccess = false;
+   for(int i = 0; i < maxBacktrack; i++)
+   {
+      if (iAmRoot)
+      {
+         cout << "\n--------- alpha = " << alpha << " ---------\n";
+      }
+      // ----- A-5.2. Compute trial point: xtrial = x0 + alpha_i xhat
+      xtrial.Set(1.0, x0);
+      xtrial.Add(alpha, xhat);
+
+      // ------ A-5.3. if not in filter region go to A.5.4 otherwise go to A-5.5.
+      thxtrial = theta(xtrial);
+      phxtrial = phi(xtrial, mu);
+      filterCheck(thxtrial, phxtrial);    
+      if(!inFilterRegion)
+      {
+         if (iAmRoot)
+         {
+            cout << "not in filter region :)\n";
+         }
+         // ------ A.5.4: Check sufficient decrease
+         if(!descentDirection)
+         {
+            switchCondition = false;
+         }
+         else
+         {
+            switchCondition = (alpha * pow(abs(Dxphi0_xhat), sPhi) > delta * pow(thx0, sTheta)) ? true : false;
+         }
+         if (iAmRoot)
+         {
+            cout << "theta(x0) = "     << thx0     << ", thetaMin = "                  << thetaMin             << endl;
+            cout << "theta(xtrial) = " << thxtrial << ", (1-gTheta) *theta(x0) = "     << (1. - gTheta) * thx0 << endl;
+            cout << "phi(xtrial) = "   << phxtrial << ", phi(x0) - gPhi *theta(x0) = " << phx0 - gPhi * thx0   << endl;
+         }      
+         // Case I      
+         if(thx0 <= thetaMin && switchCondition)
+         {
+            sufficientDecrease = (phxtrial <= phx0 + eta * alpha * Dxphi0_xhat) ? true : false;
+            if(sufficientDecrease)
+            {
+               if(iAmRoot) { cout << "Line search successful: sufficient decrease in log-barrier objective.\n"; } 
+               // accept the trial step
+               lineSearchSuccess = true;
+               break;
+            }
+         }
+         else
+         {
+            if(thxtrial <= (1. - gTheta) * thx0 || phxtrial <= phx0 - gPhi * thx0)
+            {
+               if(iAmRoot) { cout << "Line search successful: infeasibility or log-barrier objective decreased.\n"; } 
+               // accept the trial step
+               lineSearchSuccess = true;
+               break;
+            }
+         }
+         // A-5.5: Initialize the second-order correction
+         if((!(thx0 < thxtrial)) && i == 0)
+         {
+            if (iAmRoot)
+            {
+               cout << "second order correction\n";
+            }
+            problem->c(xtrial, ckSoc);
+            problem->c(x0, ck0);
+            ckSoc.Add(alphaMax, ck0);
+            // A-5.6 Compute the second-order correction.
+            IPNewtonSolve(x0, l0, z0, zhatsoc, Xhatumlsoc, mu, true);
+            mhatsoc.Set(1.0, Xhatumlsoc.GetBlock(1));
+            //WARNING: not complete but currently solver isn't entering this region
+         }
+      }
+      else
+      {
+         if (iAmRoot)
+         {
+            cout << "in filter region :(\n"; 
+         }
+      }
+      // include more if needed
+      alpha *= 0.5;
+
+   } 
+}
+
+
+void ParInteriorPointSolver::projectZ(const Vector &x, Vector &z, double mu)
+{
+   double zi;
+   double mudivmml;
+   for(int i = 0; i < dimM; i++)
+   {
+      zi = z(i);
+      mudivmml = mu / (x(i + dimU) - ml(i));
+      z(i) = max(min(zi, kSig * mudivmml), mudivmml / kSig);
+   }
+}
+
+void ParInteriorPointSolver::filterCheck(double th, double ph)
+{
+   inFilterRegion = false;
+   if(th > thetaMax)
+   {
+      inFilterRegion = true;
+   }
+   else
+   {
+      for(int i = 0; i < F1.Size(); i++)
+      {
+         if(th >= F1[i] && ph >= F2[i])
+         {
+            inFilterRegion = true;
+            break;
+         }
+      }
+   }
+}
+
+double ParInteriorPointSolver::E(const BlockVector &x, const Vector &l, const Vector &zl, double mu, bool printEeval)
+{
+   double E1, E2, E3;
+   double sc, sd;
+   BlockVector gradL(block_offsetsx); gradL = 0.0; // stationarity grad L = grad f + J^T l - z
+   Vector cx(dimC); cx = 0.0;     // feasibility c = c(x)
+   Vector comp(dimM); comp = 0.0; // complementarity M Z - mu 1
+
+   DxL(x, l, zl, gradL);
+   E1 = GlobalLpNorm(infinity(), gradL.Normlinf(), MPI_COMM_WORLD); 
+
+   problem->c(x, cx);
+   E2 = GlobalLpNorm(infinity(), cx.Normlinf(), MPI_COMM_WORLD); 
+
+
+   for(int ii = 0; ii < dimM; ii++) 
+   { 
+      comp(ii) = x(dimU + ii) * zl(ii) - mu;
+   }
+   E3 = GlobalLpNorm(infinity(), comp.Normlinf(), MPI_COMM_WORLD); 
+
+   double ll1, zl1;
+
+   zl1 = GlobalLpNorm(1, zl.Norml1(), MPI_COMM_WORLD)/ double(dimCglb + dimMglb); 
+   ll1 = GlobalLpNorm(1, l.Norml1(), MPI_COMM_WORLD);
+   sc = max(sMax, zl1 / (double(dimMglb)) ) / sMax;
+   sd = max(sMax, (ll1 + zl1) / (double(dimCglb + dimMglb))) / sMax;
+   if(iAmRoot && printEeval)
+   {
+      cout << "evaluating optimality error for mu = " << mu << endl;
+      cout << "stationarity measure = "    << E1 / sd << endl;
+      cout << "feasibility measure  = "    << E2      << endl;
+      cout << "complimentarity measure = " << E3 / sc << endl;
+   }
+   return max(max(E1 / sd, E2), E3 / sc);
+}
+
+double ParInteriorPointSolver::E(const BlockVector &x, const Vector &l, const Vector &zl, bool printEeval)
+{
+  return E(x, l, zl, 0.0, printEeval);
+}
+
+double ParInteriorPointSolver::theta(const BlockVector &x)
+{
+  Vector cx(dimC); cx = 0.0;
+  problem->c(x, cx);
+  return sqrt(InnerProduct(MPI_COMM_WORLD,cx, cx));
+}
+
+// log-barrier objective
+double ParInteriorPointSolver::phi(const BlockVector &x, double mu)
+{
+   double fx = problem->CalcObjective(x); 
+   double logBarrierLoc = 0.0;
+   for(int i = 0; i < dimM; i++) 
+   { 
+     logBarrierLoc += log(x(dimU+i)-ml(i));
+   }
+   double logBarrierGlb;
+   MPI_Allreduce(&logBarrierLoc, &logBarrierGlb, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
+   return fx - mu * logBarrierGlb;
+}
+
+// gradient of log-barrier objective with respect to x = (u, m)
+void ParInteriorPointSolver::Dxphi(const BlockVector &x, double mu, BlockVector &y)
+{
+   problem->CalcObjectiveGrad(x, y);
+   
+   for(int i = 0; i < dimM; i++) 
+   { 
+      y(dimU + i) -= mu / (x(dimU + i));
+   } 
+}
+
+// Lagrangian function evaluation
+// L(x, l, zl) = f(x) + l^T c(x) - zl^T m
+double ParInteriorPointSolver::L(const BlockVector &x, const Vector &l, const Vector &zl)
+{
+   double fx = problem->CalcObjective(x);
+   Vector cx(dimC); problem->c(x, cx);
+   return (fx + InnerProduct(MPI_COMM_WORLD,cx, l) - InnerProduct(MPI_COMM_WORLD, x.GetBlock(1), zl));
+}
+
+void ParInteriorPointSolver::DxL(const BlockVector &x, const Vector &l, const Vector &zl, BlockVector &y)
+{
+   // evaluate the gradient of the objective with respect to the primal variables x = (u, m)
+   BlockVector gradxf(block_offsetsx); gradxf = 0.0;
+   problem->CalcObjectiveGrad(x, gradxf);
+   
+   HypreParMatrix *Jacu, *Jacm, *JacuT, *JacmT;
+   Jacu = problem->Duc(x); 
+   Jacm = problem->Dmc(x);
+   JacuT = Jacu->Transpose();
+   JacmT = Jacm->Transpose();
+   
+   JacuT->Mult(l, y.GetBlock(0));
+   JacmT->Mult(l, y.GetBlock(1));
+   
+   delete JacuT;
+   delete JacmT;
+   
+   y.Add(1.0, gradxf);
+   (y.GetBlock(1)).Add(-1.0, zl);
+}
+
+bool ParInteriorPointSolver::GetConverged() const
+{
+   return converged;
+}
+
+void ParInteriorPointSolver::SetTol(double Tol)
+{
+   OptTol = Tol;
+}
+
+void ParInteriorPointSolver::SetMaxIter(int max_it)
+{
+   max_iter = max_it;
+}
+
+void ParInteriorPointSolver::SetBarrierParameter(double mu_0)
+{
+   mu_k = mu_0;
+}
+
+void ParInteriorPointSolver::SaveLogBarrierHessianIterates(bool save)
+{
+   MFEM_ASSERT(MyRank == 0 || save == false, "currently can only save logbarrier hessian in serial codes");
+   saveLogBarrierIterates = save;
+}
+
+void ParInteriorPointSolver::SetLinearSolver(int LinSolver)
+{
+   linSolver = LinSolver;
+}
+
+void ParInteriorPointSolver::SetLinearSolveTol(double Tol)
+{
+  linSolveTol = Tol;
+}
+
+
+ParInteriorPointSolver::~ParInteriorPointSolver() 
+{
+   F1.DeleteAll();
+   F2.DeleteAll();
+   block_offsetsx.DeleteAll();
+   block_offsetsumlz.DeleteAll();
+   block_offsetsuml.DeleteAll();
+   ml.SetSize(0);
+}

examples/contact/ParIPsolver.hpp

@@ -0,0 +1,82 @@
+#include "mfem.hpp"
+#include "ParProblems.hpp"
+#include <fstream>
+#include <iostream>
+
+using namespace std;
+using namespace mfem;
+
+
+#ifndef PARIPSOLVER 
+#define PARIPSOLVER
+
+class ParInteriorPointSolver
+{
+protected:
+    ParGeneralOptProblem* problem;
+    double OptTol;
+    int  max_iter;
+    double mu_k; // \mu_k
+    Vector lk, zlk;
+
+    double sMax, kSig, tauMin, eta, thetaMin, delta, sTheta, sPhi, kMu, thetaMu;
+    double thetaMax, kSoc, gTheta, gPhi, kEps;
+	
+    // filter
+    Array<double> F1, F2;
+	
+    // quantities computed in lineSearch
+    double alpha, alphaz;
+    double thx0, thxtrial;
+    double phx0, phxtrial;
+    bool descentDirection, switchCondition, sufficientDecrease, lineSearchSuccess, inFilterRegion;
+    double Dxphi0_xhat;
+
+    int dimU, dimM, dimC;
+    int dimUglb, dimMglb, dimCglb;
+    Array<int> block_offsetsumlz, block_offsetsuml, block_offsetsx;
+    Vector ml;
+
+    Vector ckSoc;
+    HypreParMatrix * Huu, * Hum, * Hmu, * Hmm, * Wmm, *D, * Ju, * Jm, * JuT, * JmT;
+    
+    int jOpt;
+    bool converged;
+    
+    int MyRank;
+    bool iAmRoot;
+
+    bool saveLogBarrierIterates;
+
+    int linSolver;
+    double linSolveTol;
+public:
+    ParInteriorPointSolver(ParGeneralOptProblem*);
+    double MaxStepSize(Vector& , Vector& , Vector& , double);
+    double MaxStepSize(Vector& , Vector& , double);
+    void Mult(const BlockVector& , BlockVector&);
+    void Mult(const Vector&, Vector &); 
+    void FormIPNewtonMat(BlockVector& , Vector& , Vector& , BlockOperator &);
+    void IPNewtonSolve(BlockVector& , Vector& , Vector& , Vector&, BlockVector& , double, bool);
+    void lineSearch(BlockVector& , BlockVector& , double);
+    void projectZ(const Vector & , Vector &, double);
+    void filterCheck(double, double);
+    double E(const BlockVector &, const Vector &, const Vector &, double, bool);
+    double E(const BlockVector &, const Vector &, const Vector &, bool);
+    bool GetConverged() const;
+    // TO DO: include Hessian of Lagrangian
+    double theta(const BlockVector &);
+    double phi(const BlockVector &, double);
+    void Dxphi(const BlockVector &, double, BlockVector &);
+    double L(const BlockVector &, const Vector &, const Vector &);
+    void DxL(const BlockVector &, const Vector &, const Vector &, BlockVector &);
+    void SetTol(double);
+    void SetMaxIter(int);
+    void SetBarrierParameter(double);    
+    void SaveLogBarrierHessianIterates(bool);
+    void SetLinearSolver(int);
+    void SetLinearSolveTol(double);
+    virtual ~ParInteriorPointSolver();
+};
+
+#endif

examples/contact/ParObstacleProblem.cpp

@@ -0,0 +1,129 @@
+//                         Obstacle Problem
+//
+//
+// Compile with: make ParObstacleProblem
+//
+// Sample runs: mpirun -np 4 ./ParObstacleProblem
+//
+//
+// Description: This example code demonstrates the use of MFEM to solve the
+//              bound-constrained energy minimization problem
+//
+//                      minimize (||∇u||² + ||u||²) subject to u ≥ ϕ in H¹.
+
+#include "mfem.hpp"
+#include "ParProblems.hpp"
+#include "ParIPsolver.hpp"
+#include <fstream>
+#include <iostream>
+
+using namespace std;
+using namespace mfem;
+
+double dmanufacturedFun(const Vector &);
+double fRhs(const Vector &);
+
+int main(int argc, char *argv[])
+{
+  // Initialize MPI
+   Mpi::Init();
+   int num_procs = Mpi::WorldSize();
+   int myid = Mpi::WorldRank();
+   Hypre::Init();
+
+   int FEorder = 1; // order of the finite elements
+   int linSolver = 2;
+   int maxIPMiters = 30;
+   int ref_levels = 3;
+   OptionsParser args(argc, argv);
+   args.AddOption(&FEorder, "-o", "--order",\
+	 	  "Order of the finite elements.");
+   args.AddOption(&linSolver, "-linSolver", "--linearSolver", \
+        "IP-Newton linear system solution strategy.");
+   args.AddOption(&maxIPMiters, "-IPMiters", "--IPMiters",\
+	 	  "Maximum number of IPM iterations");
+   args.AddOption(&ref_levels, "-r", "--mesh_refinement", \
+		  "Mesh Refinement");
+  
+   args.Parse();
+   if(!args.Good())
+   {
+     args.PrintUsage(cout);
+     return 1;
+   }
+   else
+   {
+     if(Mpi::Root())
+     {  
+       args.PrintOptions(cout);
+     }
+   }
+
+   const char *meshFile = "../../data/inline-quad.mesh";
+   Mesh mesh(meshFile, 1, 1);
+   int dim = mesh.Dimension(); // geometric dimension of the meshed domain
+   {
+      for (int l = 0; l < ref_levels; l++)
+      {
+         mesh.UniformRefinement();
+      }
+   }
+
+   ParMesh pmesh(MPI_COMM_WORLD, mesh);
+
+   FiniteElementCollection *fec = new H1_FECollection(FEorder, dim);
+   ParFiniteElementSpace   *Vh  = new ParFiniteElementSpace(&pmesh, fec);
+
+   ParObstacleProblem problem(Vh,Vh,&fRhs);
+  
+   int dimD = problem.GetDimD();
+   Vector x0(dimD); x0 = 100.0;
+   Vector xf(dimD); xf = 0.0;
+
+   ParInteriorPointSolver optimizer(&problem); 
+   optimizer.SetTol(1.e-8);
+   optimizer.SetLinearSolveTol(1.e-10);
+   optimizer.SetLinearSolver(linSolver);
+   optimizer.SetMaxIter(maxIPMiters);
+   optimizer.Mult(x0, xf);
+
+   ParGridFunction d_gf(Vh);
+
+   d_gf.SetFromTrueDofs(xf);
+
+
+   FunctionCoefficient dm_fc(dmanufacturedFun); // manufactured solution
+   ParGridFunction dm_gf(Vh);
+   dm_gf.ProjectCoefficient(dm_fc);
+
+   char vishost[] = "localhost";
+   int  visport   = 19916;
+   socketstream exact_sock(vishost, visport);
+   exact_sock.precision(8);
+   exact_sock << "parallel " << num_procs << " " << myid << "\n";
+   exact_sock << "solution\n" << pmesh << dm_gf
+              << "window_title 'Manufactured solution'" << flush;
+
+   socketstream sol_sock(vishost, visport);
+   sol_sock.precision(8);
+   sol_sock << "parallel " << num_procs << " " << myid << "\n";
+   sol_sock << "solution\n" << pmesh << d_gf 
+            << "window_title 'Numerical solution'" << flush;
+
+   delete Vh;
+   delete fec;
+   return 0;
+}
+
+
+double dmanufacturedFun(const Vector &x)
+{
+  return cos(2*M_PI*x(0)) + 0.2 - 2.0*(pow(x(0),3) - 1.5*pow(x(0),2));
+}
+
+double fRhs(const Vector &x)
+{
+  double fx = 0.;
+  fx = 0.2 - 2.0 * (pow(x(0),3)- 1.5*pow(x(0),2.) - 6 * x(0) + 3.) + (1. + pow(2.*M_PI,2))*cos(2.*M_PI*x(0));
+  return fx;
+}

examples/contact/ParProblems.cpp

@@ -0,0 +1,212 @@
+#include "mfem.hpp"
+#include "ParProblems.hpp"
+#include <fstream>
+#include <iostream>
+
+using namespace std;
+using namespace mfem;
+
+ParGeneralOptProblem::ParGeneralOptProblem(ParFiniteElementSpace * fesU_, ParFiniteElementSpace * fesM_) 
+: fesU(fesU_), fesM(fesM_) 
+{ 
+   dimU = fesU->GetTrueVSize();
+   dimM = fesM->GetTrueVSize();
+   dimC = fesM->GetTrueVSize();
+}
+
+void ParGeneralOptProblem::CalcObjectiveGrad(const BlockVector &x, BlockVector &y) const
+{
+   Duf(x, y.GetBlock(0));
+   Dmf(x, y.GetBlock(1));
+}
+
+ParGeneralOptProblem::~ParGeneralOptProblem()
+{
+   block_offsetsx.DeleteAll();
+}
+
+
+// min E(d) s.t. g(d) >= 0
+// min_(d,s) E(d) s.t. c(d,s) := g(d) - s = 0, s >= 0
+ParOptProblem::ParOptProblem(ParFiniteElementSpace * fesU_, 
+                                     ParFiniteElementSpace * fesM_)
+ : ParGeneralOptProblem(fesU_, fesM_), block_offsetsx(3)
+{
+   block_offsetsx[0] = 0;
+   block_offsetsx[1] = dimU;
+   block_offsetsx[2] = dimM;
+   block_offsetsx.PartialSum();
+   ml.SetSize(dimM); ml = 0.0;
+   Vector negIdentDiag(dimM);
+   negIdentDiag = -1.0;
+   SparseMatrix * diag = new SparseMatrix(negIdentDiag);
+   Ih = new HypreParMatrix(fesM->GetComm(), fesM->GlobalTrueVSize(),
+                                             fesM->GetTrueDofOffsets(), diag);
+   HypreStealOwnership(*Ih, *diag);
+   delete diag;
+}
+
+double ParOptProblem::CalcObjective(const BlockVector &x) const { return E(x.GetBlock(0)); }
+
+void ParOptProblem::Duf(const BlockVector &x, Vector &y) const { DdE(x.GetBlock(0), y); }
+
+void ParOptProblem::Dmf(const BlockVector &x, Vector &y) const { y = 0.0; }
+
+HypreParMatrix * ParOptProblem::Duuf(const BlockVector &x) 
+{ 
+   return DddE(x.GetBlock(0)); 
+}
+
+HypreParMatrix * ParOptProblem::Dumf(const BlockVector &x) { return nullptr; }
+
+HypreParMatrix * ParOptProblem::Dmuf(const BlockVector &x) { return nullptr; }
+
+HypreParMatrix * ParOptProblem::Dmmf(const BlockVector &x) { return nullptr; }
+
+void ParOptProblem::c(const BlockVector &x, Vector &y) const // c(u,m) = g(u) - m 
+{
+   g(x.GetBlock(0), y);
+   y.Add(-1.0, x.GetBlock(1));  
+}
+
+HypreParMatrix * ParOptProblem::Duc(const BlockVector &x) 
+{ 
+   return Ddg(x.GetBlock(0)); 
+}
+
+HypreParMatrix * ParOptProblem::Dmc(const BlockVector &x) 
+{ 
+   return Ih;
+} 
+
+ParOptProblem::~ParOptProblem() 
+{
+   delete Ih;
+}
+
+
+// Obstacle Problem, no essential boundary conditions enforced
+// Hessian of energy term is K + M (stiffness + mass)
+ParObstacleProblem::ParObstacleProblem(ParFiniteElementSpace *fesU_, 
+                                       ParFiniteElementSpace *fesM_, 
+                                       double (*fSource)(const Vector &)) : 
+                                       ParOptProblem(fesU_,fesM_), f(dimU), psi(dimU), J(nullptr)
+{
+   Kform = new ParBilinearForm(fesU);
+   Kform->AddDomainIntegrator(new MassIntegrator);
+   Kform->AddDomainIntegrator(new DiffusionIntegrator);
+   Kform->Assemble();
+   Kform->Finalize();
+   Kform->FormSystemMatrix(ess_tdof_list, K);
+
+   FunctionCoefficient fcoeff(fSource);
+   fform = new ParLinearForm(fesU);
+   fform->AddDomainIntegrator(new DomainLFIntegrator(fcoeff));
+   fform->Assemble();
+   Vector F(dimU);
+   fform->ParallelAssemble(F);
+   f.SetSize(dimU);
+   f.Set(1.0, F);
+
+   psi = 0.0;
+   
+   Vector iDiag(dimU); iDiag = 1.0;
+   SparseMatrix * Jacg = new SparseMatrix(iDiag);
+   
+   J = new HypreParMatrix(fesU->GetComm(),fesU->GlobalTrueVSize(),fesU->GetTrueDofOffsets(),Jacg);
+   HypreStealOwnership(*J, *Jacg);
+   delete Jacg;
+}
+
+// Obstacle Problem, essential boundary conditions enforced
+// Hessian of energy term is K (stiffness)
+ParObstacleProblem::ParObstacleProblem(ParFiniteElementSpace *fesU_, 
+                                       ParFiniteElementSpace *fesM_, 
+				       double (*fSource)(const Vector &),
+				       double (*obstacleSource)(const Vector &),
+				       Array<int> tdof_list, Vector &xDC) : ParOptProblem(fesU_,fesM_), f(dimU), psi(dimU), J(nullptr)
+{
+   // elastic energy functional terms	
+   ess_tdof_list = tdof_list;
+   Kform = new ParBilinearForm(fesU);
+   Kform->AddDomainIntegrator(new DiffusionIntegrator);
+   Kform->Assemble();
+   Kform->Finalize();
+   Kform->FormSystemMatrix(ess_tdof_list, K);
+
+   FunctionCoefficient fcoeff(fSource);
+   fform = new ParLinearForm(fesU);
+   fform->AddDomainIntegrator(new DomainLFIntegrator(fcoeff));
+   fform->Assemble();
+   Vector F(dimU);
+   fform->ParallelAssemble(F);
+   f.SetSize(dimU);
+   f.Set(1.0, F);
+   Kform->EliminateVDofsInRHS(ess_tdof_list, xDC, f);
+   
+   // obstacle constraints --  
+   Vector iDiag(dimU); iDiag = 1.0;
+   for(int i = 0; i < ess_tdof_list.Size(); i++)
+   {
+     iDiag(ess_tdof_list[i]) = 0.0;
+   }
+   SparseMatrix * Jacg = new SparseMatrix(iDiag);
+
+   J = new HypreParMatrix(fesU->GetComm(),fesU->GlobalTrueVSize(),fesU->GetTrueDofOffsets(),Jacg);
+   HypreStealOwnership(*J, *Jacg);
+   delete Jacg;
+
+   FunctionCoefficient psi_fc(obstacleSource);
+   ParGridFunction psi_gf(fesU);
+   psi_gf.ProjectCoefficient(psi_fc);
+   psi.Set(1.0, (*psi_gf.GetTrueDofs()));
+   for(int i = 0; i < ess_tdof_list.Size(); i++)
+   {
+     psi(ess_tdof_list[i]) -= 1.e-8;
+   }
+}
+
+
+
+double ParObstacleProblem::E(const Vector &d) const
+{
+   Vector Kd(K.Height()); Kd = 0.0;
+   MFEM_VERIFY(d.Size() == K.Width(), "ParObstacleProblem::E - Inconsistent dimensions");
+   K.Mult(d, Kd);
+   return 0.5 * InnerProduct(MPI_COMM_WORLD, d, Kd) - InnerProduct(MPI_COMM_WORLD, f, d);
+}
+
+void ParObstacleProblem::DdE(const Vector &d, Vector &gradE) const
+{
+   gradE.SetSize(K.Height());
+   MFEM_VERIFY(d.Size() == K.Width(), "ParObstacleProblem::DdE - Inconsistent dimensions");
+   K.Mult(d, gradE);
+   MFEM_VERIFY(f.Size() == K.Height(), "ParObstacleProblem::DdE - Inconsistent dimensions");
+   gradE.Add(-1.0, f);
+}
+
+HypreParMatrix * ParObstacleProblem::DddE(const Vector &d)
+{
+   return &K; 
+}
+
+// g(d) = d >= \psi
+void ParObstacleProblem::g(const Vector &d, Vector &gd) const
+{
+   MFEM_VERIFY(d.Size() == J->Width(), "ParObstacleProblem::g - Inconsistent dimensions");
+   J->Mult(d, gd);
+   MFEM_VERIFY(gd.Size() == J->Height(), "ParObstacleProblem::g - Inconsistent dimensions");
+   gd.Add(-1.0, psi);
+}
+
+HypreParMatrix * ParObstacleProblem::Ddg(const Vector &d)
+{
+   return J;
+}
+
+ParObstacleProblem::~ParObstacleProblem()
+{
+   delete Kform;
+   delete fform;
+   delete J;
+}

examples/contact/ParProblems.hpp

@@ -0,0 +1,107 @@
+#include "mfem.hpp"
+#include <fstream>
+#include <iostream>
+
+using namespace std;
+using namespace mfem;
+
+#ifndef PARPROBLEM_DEFS
+#define PARPROBLEM_DEFS
+
+// abstract ParGeneralOptProblem class
+// of the form
+// min_(u,m) f(u,m) s.t. c(u,m)=0 and m>=ml
+// the primal variable (u, m) is represented as a BlockVector
+// think about supporting general lower and upper bounds (see HiOP user manual) 
+
+class ParGeneralOptProblem
+{
+protected:
+    int dimU, dimM, dimC;
+    ParFiniteElementSpace * fesU = nullptr;
+    ParFiniteElementSpace * fesM = nullptr;
+    Array<int> block_offsetsx;
+    Vector ml;
+public:
+    ParGeneralOptProblem(ParFiniteElementSpace * fesU_, ParFiniteElementSpace * fesM_); // constructor
+    virtual double CalcObjective(const BlockVector &) const = 0;
+    virtual void Duf(const BlockVector &, Vector &) const = 0;
+    virtual void Dmf(const BlockVector &, Vector &) const = 0;
+    void CalcObjectiveGrad(const BlockVector &, BlockVector &) const;
+    virtual HypreParMatrix * Duuf(const BlockVector &) = 0;
+    virtual HypreParMatrix * Dumf(const BlockVector &) = 0;
+    virtual HypreParMatrix * Dmuf(const BlockVector &) = 0;
+    virtual HypreParMatrix * Dmmf(const BlockVector &) = 0;
+    virtual HypreParMatrix * Duc(const BlockVector &) = 0;
+    virtual HypreParMatrix * Dmc(const BlockVector &) = 0;
+    // TO DO: include Hessian terms of constraint c
+    virtual void c(const BlockVector &, Vector &) const = 0;
+    int GetDimU() const { return dimU; };
+    int GetDimM() const { return dimM; }; 
+    int GetDimC() const { return dimC; };
+    ParFiniteElementSpace * GetfesU() {return fesU;}
+    ParFiniteElementSpace * GetfesM() {return fesM;}
+    Vector Getml() const { return ml; };
+    ~ParGeneralOptProblem(); // destructor
+};
+
+
+// abstract ContactProblem class
+// of the form
+// min_d e(d) s.t. g(d) >= 0
+class ParOptProblem : public ParGeneralOptProblem
+{
+protected:
+    Array<int> block_offsetsx;
+    HypreParMatrix * Ih;
+public:
+    ParOptProblem(ParFiniteElementSpace * fesU_, ParFiniteElementSpace * fesM_);        // constructor
+    double CalcObjective(const BlockVector &) const; // objective e
+    void Duf(const BlockVector &, Vector &) const;
+    void Dmf(const BlockVector &, Vector &) const;
+
+    HypreParMatrix * Duuf(const BlockVector &);
+    HypreParMatrix * Dumf(const BlockVector &);
+    HypreParMatrix * Dmuf(const BlockVector &);
+    HypreParMatrix * Dmmf(const BlockVector &);
+    HypreParMatrix * Duc(const BlockVector &);
+    HypreParMatrix * Dmc(const BlockVector &);
+
+    void c(const BlockVector &, Vector &) const;
+    virtual double E(const Vector &) const = 0;               // objective e(d) (energy function)
+    virtual void DdE(const Vector &, Vector &) const = 0;      // gradient of objective De / Dd
+    virtual HypreParMatrix * DddE(const Vector &) = 0;
+    // Hessian of objective D^2 e / D d^2
+    virtual HypreParMatrix * Ddg(const Vector &) = 0;
+    // Jacobian of inequality constraint Dg / Dd
+    virtual void g(const Vector &, Vector &) const = 0;       // inequality constraint g(d) >= 0 (gap function)
+    int GetDimD() const { return fesU->GetTrueVSize(); };
+    int GetDimS() const { return fesM->GetTrueVSize(); };
+    virtual ~ParOptProblem();
+};
+
+class ParObstacleProblem : public ParOptProblem
+{
+protected:
+   // data to define energy objective function e(d) = 0.5 d^T K d - f^T d, g(d) = d >= \psi
+   // stiffness matrix used to define objective
+   ParBilinearForm *Kform;
+   ParLinearForm   *fform;
+   Array<int> ess_tdof_list; // needed for calls to FormSystemMatrix
+   HypreParMatrix  K;
+   HypreParMatrix  *J;
+   ParFiniteElementSpace *Vh;
+   Vector f;
+   Vector psi;
+public :
+   ParObstacleProblem(ParFiniteElementSpace*, ParFiniteElementSpace*, double (*fSource)(const Vector &));
+   ParObstacleProblem(ParFiniteElementSpace*, ParFiniteElementSpace*, double (*fSource)(const Vector &), double (*obstacleSource)(const Vector &), Array<int> tdof_list, Vector &);
+   double E(const Vector &) const;
+   void DdE(const Vector &, Vector &) const;
+   HypreParMatrix* DddE(const Vector &);
+   void g(const Vector &, Vector &) const;
+   HypreParMatrix* Ddg(const Vector &);
+   virtual ~ParObstacleProblem();
+};
+
+#endif

examples/contact/ParSphericalObstacleProblem.cpp

@@ -0,0 +1,173 @@
+//                       Spherical Obstacle Problem
+//
+//
+// Compile with: make ParSphericalObstacleProblem
+//
+// Sample runs: mpirun -np 4 ./ParSphericalObstacleProblem -linSolver 0
+//              mpirun -np 4 ./ParSphericalObstacleProblem -linSolver 1
+//              mpirun -np 4 ./ParSphericalObstacleProblem -linSolver 2
+//
+//
+// Description: This example code demonstrates the use of MFEM to solve the
+//              bound-constrained energy minimization problem
+//
+//                      minimize ||∇u||² subject to u ≥ ϕ in H¹₀.
+
+#include "mfem.hpp"
+#include "ParProblems.hpp"
+#include "ParIPsolver.hpp"
+#include <fstream>
+#include <iostream>
+
+using namespace std;
+using namespace mfem;
+
+double fRhs(const Vector &);
+double spherical_obstacle(const Vector &);
+double exact_solution_obstacle(const Vector &);
+
+
+int main(int argc, char *argv[])
+{
+  // Initialize MPI
+   Mpi::Init();
+   int num_procs = Mpi::WorldSize();
+   int myid = Mpi::WorldRank();
+   Hypre::Init();
+
+   int FEorder = 1; // order of the finite elements
+   int linSolver = 2;
+   int maxIPMiters = 30;
+   int ref_levels = 3;
+   OptionsParser args(argc, argv);
+   args.AddOption(&FEorder, "-o", "--order",\
+	 	  "Order of the finite elements.");
+   args.AddOption(&linSolver, "-linSolver", "--linearSolver", \
+        "IP-Newton linear system solution strategy.");
+   args.AddOption(&maxIPMiters, "-IPMiters", "--IPMiters",\
+	 	  "Maximum number of IPM iterations");
+   args.AddOption(&ref_levels, "-r", "--mesh_refinement", \
+		  "Mesh Refinement");
+  
+   args.Parse();
+   if(!args.Good())
+   {
+     args.PrintUsage(cout);
+     return 1;
+   }
+   else
+   {
+     if(myid == 0)
+     {  
+       args.PrintOptions(cout);
+     }
+   }
+
+   const char *meshFile = "../../data/disk.mesh";
+   Mesh mesh(meshFile, 1, 1);
+   int dim = mesh.Dimension(); // geometric dimension of the meshed domain
+   {
+      for (int l = 0; l < ref_levels; l++)
+      {
+         mesh.UniformRefinement();
+      }
+   }
+
+   ParMesh pmesh(MPI_COMM_WORLD, mesh);
+   
+   FiniteElementCollection *fec = new H1_FECollection(FEorder, dim);
+   ParFiniteElementSpace   *Vh  = new ParFiniteElementSpace(&pmesh, fec);
+   Array<int> boundary_dofs;
+   Vh->GetBoundaryTrueDofs(boundary_dofs);
+   int dimD = Vh->GetTrueVSize();
+   Vector xDC(dimD); xDC = 0.0;
+
+   ParObstacleProblem problem(Vh, Vh, &fRhs, &spherical_obstacle, boundary_dofs, xDC);
+   Vector x0(dimD); x0.Set(1.0, xDC);
+   Vector xf(dimD); xf = 0.0;
+
+   ParInteriorPointSolver optimizer(&problem); 
+   optimizer.SetTol(1.e-7);
+   optimizer.SetLinearSolveTol(1.e-10);
+   optimizer.SetLinearSolver(linSolver);
+   optimizer.SetMaxIter(maxIPMiters);
+   optimizer.Mult(x0, xf);
+
+   ParGridFunction d_gf(Vh);
+
+   d_gf.SetFromTrueDofs(xf);
+
+
+   FunctionCoefficient dtrue_fc(exact_solution_obstacle); // analytic solution
+   ParGridFunction dtrue_gf(Vh);
+   dtrue_gf.ProjectCoefficient(dtrue_fc);
+
+   double L2error = d_gf.ComputeL2Error(dtrue_fc);
+   if (myid == 0)
+   {
+     cout << "\n|| u_h - u ||_{L^2} = " << L2error << '\n' << endl;
+   } 
+   
+   ParaViewDataCollection paraview_dc("SphericalObstacleProblem", &pmesh);
+   paraview_dc.SetPrefixPath("ParaView");
+   paraview_dc.SetLevelsOfDetail(FEorder);
+   paraview_dc.SetDataFormat(VTKFormat::BINARY);
+   paraview_dc.SetHighOrderOutput(true);
+   paraview_dc.SetCycle(0);
+   paraview_dc.SetTime(0.0);
+   paraview_dc.RegisterField("u(x,y) (analytic)", &dtrue_gf);
+   paraview_dc.RegisterField("u(x,y) (numerical)", &d_gf);
+   paraview_dc.Save();
+
+   delete Vh;
+   delete fec;
+   return 0;
+}
+
+
+double fRhs(const Vector &x)
+{
+  return 0.;
+}
+
+
+double spherical_obstacle(const Vector &pt)
+{
+   double x = pt(0), y = pt(1);
+   double r = sqrt(x*x + y*y);
+   double r0 = 0.5;
+   double beta = 0.9;
+
+   double b = r0*beta;
+   double tmp = sqrt(r0*r0 - b*b);
+   double B = tmp + b*b/tmp;
+   double C = -b/tmp;
+
+   if (r > b)
+   {
+      return B + r * C;
+   }
+   else
+   {
+      return sqrt(r0*r0 - r*r);
+   }
+}
+
+double exact_solution_obstacle(const Vector &pt)
+{
+   double x = pt(0), y = pt(1);
+   double r = sqrt(x*x + y*y);
+   double r0 = 0.5;
+   double a =  0.348982574111686;
+   double A = -0.340129705945858;
+
+   if (r > a)
+   {
+      return A * log(r);
+   }
+   else
+   {
+      return sqrt(r0*r0-r*r);
+   }
+} 
+

examples/contact/Problems.hpp

@@ -0,0 +1,300 @@
+#include "mfem.hpp"
+#include <fstream>
+#include <iostream>
+#include <set>
+
+using namespace std;
+using namespace mfem;
+
+#ifndef PROBLEM_DEFS
+#define PROBLEM_DEFS
+
+
+
+// abstract GeneralOptProblem class
+// for the problem
+// min_(u,m) f(u,m) 
+// such that c(u,m)=0 and m >= ml
+class GeneralOptProblem
+{
+protected:
+    int dimU, dimM, dimC;
+    Array<int> block_offsetsx;
+    Vector ml;
+public:
+    GeneralOptProblem();
+    virtual double CalcObjective(const BlockVector &) const = 0;
+    virtual void Duf(const BlockVector &, Vector &) const = 0;
+    virtual void Dmf(const BlockVector &, Vector &) const = 0;
+    void CalcObjectiveGrad(const BlockVector &, BlockVector &) const;
+    virtual SparseMatrix* Duuf(const BlockVector &) = 0;
+    virtual SparseMatrix* Dumf(const BlockVector &) = 0;
+    virtual SparseMatrix* Dmuf(const BlockVector &) = 0;
+    virtual SparseMatrix* Dmmf(const BlockVector &) = 0;
+    virtual void c(const BlockVector &, Vector &) const = 0;
+    virtual SparseMatrix* Duc(const BlockVector &) = 0;
+    virtual SparseMatrix* Dmc(const BlockVector &) = 0;
+    virtual SparseMatrix* lDuuc(const BlockVector &, const Vector &) = 0;
+    virtual SparseMatrix* lDumc(const BlockVector &, const Vector &) = 0;
+    virtual SparseMatrix* lDmuc(const BlockVector &, const Vector &) = 0;
+    virtual SparseMatrix* lDmmc(const BlockVector &, const Vector &) = 0;
+    // TO DO: include log-barrier lumped-mass and pass that
+    // to the optimizer
+    //virtual SparseMatrix* GetLogBarrierLumpedMass() = 0;
+    int GetDimU() const { return dimU; };
+    int GetDimM() const { return dimM; }; 
+    int GetDimC() const { return dimC; };
+    Vector Getml() const { return ml; };
+    ~GeneralOptProblem();
+};
+
+
+// abstract OptProblem class
+// of the form
+// min_d e(d) s.t. g(d) >= 0
+class OptProblem : public GeneralOptProblem
+{
+protected:
+    int dimD;
+    int dimS;
+    Array<int> block_offsetsx;
+    SparseMatrix * negIdentity;
+    SparseMatrix * zeroMatum;
+    SparseMatrix * zeroMatmu;
+    SparseMatrix * zeroMatmm;
+public:
+    //OptProblem(int, int);        // constructor
+    OptProblem();
+    void InitializeParentData(int, int);
+    double CalcObjective(const BlockVector &) const; // objective e
+    void Duf(const BlockVector &, Vector &) const;
+    void Dmf(const BlockVector &, Vector &) const;
+    SparseMatrix* Duuf(const BlockVector &);
+    SparseMatrix* Dumf(const BlockVector &);
+    SparseMatrix* Dmuf(const BlockVector &);
+    SparseMatrix* Dmmf(const BlockVector &);
+    void c(const BlockVector &, Vector &) const;
+    SparseMatrix* Duc(const BlockVector &);
+    SparseMatrix* Dmc(const BlockVector &);
+    SparseMatrix* lDuuc(const BlockVector &, const Vector &);
+    SparseMatrix* lDumc(const BlockVector &, const Vector &);
+    SparseMatrix* lDmuc(const BlockVector &, const Vector &);
+    SparseMatrix* lDmmc(const BlockVector &, const Vector &);
+    virtual double E(const Vector &) const = 0;           // objective e(d) (energy function)
+    virtual void DdE(const Vector &, Vector &) const = 0; // gradient of objective De / Dd
+    virtual SparseMatrix* DddE(const Vector &) = 0;       // Hessian of objective D^2 e / D d^2
+    virtual void g(const Vector &, Vector &) const = 0;   // inequality constraint g(d) >= 0 (gap function)
+    virtual SparseMatrix* Ddg(const Vector &) = 0;        // Jacobian of inequality constraint Dg / Dd
+    virtual SparseMatrix* lDddg(const Vector &, const Vector &) = 0;
+    int GetDimD() const { return dimD; };
+    int GetDimS() const { return dimS; };
+    virtual ~OptProblem();
+};
+
+
+class ObstacleProblem : public OptProblem
+{
+protected:
+   // data to define energy objective function e(d) = 0.5 d^T K d - f^T d, g(d) = d + \psi >= 0
+   // stiffness matrix used to define objective
+   BilinearForm *Kform;
+   LinearForm   *fform;
+   Array<int> ess_tdof_list;
+   SparseMatrix  *K;
+   SparseMatrix  *J;
+   SparseMatrix  *Hcl;
+   FiniteElementSpace *Vh;
+   Vector f;
+   Vector psil;
+   Vector psiu;
+   bool twoBounds;
+   Vector xDC;
+   double Ce;
+public : 
+   ObstacleProblem(FiniteElementSpace*, double (*fSource)(const Vector &), double (*obstacleSource)(const Vector &));
+   ObstacleProblem(FiniteElementSpace*, Vector&,  double (*fSource)(const Vector &), double (*obstacleSource)(const Vector &), Array<int> tdof_list);
+   ObstacleProblem(FiniteElementSpace*, Vector &, double (*fSource)(const Vector &), double (*obstacleSourcel)(const Vector &), double (*obstacleSourceu)(const Vector &), Array<int> tdof_list);
+   double E(const Vector &) const;
+   void DdE(const Vector &, Vector &) const;
+   SparseMatrix* DddE(const Vector &);
+   void g(const Vector &, Vector &) const;
+   SparseMatrix* Ddg(const Vector &);
+   SparseMatrix * lDddg(const Vector &, const Vector &);
+   virtual ~ObstacleProblem();
+};
+
+
+SparseMatrix * GenerateProjector(int n, Array<int> ess_tdof_list);
+
+
+class ObstacleProblemVariant : public OptProblem
+{
+protected:
+   // data to define energy objective function e(d) = 0.5 d^T K d - f^T d, g(d) = d + \psi >= 0
+   // stiffness matrix used to define objective
+   BilinearForm *Kform;
+   LinearForm   *fform;
+   Array<int> ess_tdof_list;
+   Array<int> noness_tdof_list;
+   SparseMatrix  *K;
+   SparseMatrix  *RKP; // R K P = R K R^T
+   SparseMatrix  *J;
+   SparseMatrix  *Hcl;
+   SparseMatrix *R;
+   FiniteElementSpace *Vh;
+   Vector f;
+   Vector psil;
+   Vector xDC;
+   double Ce;
+public : 
+   ObstacleProblemVariant(FiniteElementSpace*, Vector&,  double (*fSource)(const Vector &), double (*obstacleSource)(const Vector &), Array<int> tdof_list);
+   double E(const Vector &) const;
+   void DdE(const Vector &, Vector &) const;
+   SparseMatrix* DddE(const Vector &);
+   void g(const Vector &, Vector &) const;
+   SparseMatrix* Ddg(const Vector &);
+   SparseMatrix * lDddg(const Vector &, const Vector &);
+   virtual ~ObstacleProblemVariant();
+};
+
+
+
+
+class QPOptProblem : public OptProblem
+{
+protected:
+  SparseMatrix *K;
+  SparseMatrix *J;
+  SparseMatrix *zeroMatdd;
+  Vector f;
+  Vector g0;
+public:
+  QPOptProblem(const SparseMatrix, const SparseMatrix, const Vector, const Vector);
+  double E(const Vector &) const;
+  void DdE(const Vector &, Vector &) const;
+  SparseMatrix* DddE(const Vector &);
+  void g(const Vector &, Vector &) const;
+  SparseMatrix* Ddg(const Vector &);
+  SparseMatrix * lDddg(const Vector &, const Vector &);
+  virtual ~QPOptProblem();
+};
+
+
+class ExContactBlockTL : public OptProblem
+{
+public:
+   double E(const Vector &) const;
+   void DdE(const Vector &, Vector &) const;
+   SparseMatrix* DddE(const Vector &);
+   void g(const Vector &, Vector &) const;
+   SparseMatrix* Ddg(const Vector &);
+   SparseMatrix * lDddg(const Vector &, const Vector &);
+   FiniteElementSpace GetVh1();
+   FiniteElementSpace GetVh2();
+   SparseMatrix *zeroMatdd;
+public:
+   /** default constructor */
+   ExContactBlockTL(Mesh *, Mesh *, int);
+   
+
+   /** default destructor */
+   virtual ~ExContactBlockTL();
+
+private:
+   void update_g() const;
+
+private:
+   /**@name Methods to block default compiler methods.
+    *
+    * The compiler automatically generates the following three methods.
+    *  Since the default compiler implementation is generally not what
+    *  you want (for all but the most simple classes), we usually
+    *  put the declarations of these methods in the private section
+    *  and never implement them. This prevents the compiler from
+    *  implementing an incorrect "default" behavior without us
+    *  knowing. (See Scott Meyers book, "Effective C++")
+    */
+   ExContactBlockTL(
+      const ExContactBlockTL&
+   );
+
+   ExContactBlockTL& operator=(
+      const ExContactBlockTL&
+   );
+
+   Array<int> attr;
+   Array<int> m_attr;
+   Array<int> s_conn; // connectivity of the second/slave mesh
+   std::string mesh_file1;
+   std::string mesh_file2;
+   Mesh* mesh1;
+   Mesh* mesh2;
+   FiniteElementCollection* fec1;
+   FiniteElementCollection* fec2;
+   FiniteElementSpace* fespace1;
+   FiniteElementSpace* fespace2;
+   Array<int> ess_tdof_list1;
+   Array<int> ess_tdof_list2;
+   GridFunction nodes0;
+   GridFunction* nodes1;
+   GridFunction* nodes2;
+   mutable GridFunction* x1;
+   mutable GridFunction* x2;
+   PWConstCoefficient* lambda1_func;
+   PWConstCoefficient* lambda2_func;
+   PWConstCoefficient* mu1_func;
+   PWConstCoefficient* mu2_func;
+   BilinearForm* a1;
+   BilinearForm* a2;
+
+   mfem::Vector lambda1;
+   mfem::Vector lambda2;
+   mfem::Vector mu1;
+   mfem::Vector mu2;
+   mutable mfem::Vector xyz;
+
+   std::set<int> bdryVerts2;
+
+   int dim;
+   // degrees of freedom of both meshes
+   int ndof_1;
+   int ndof_2;
+   int ndofs;
+   // number of nodes for each mesh
+   int nnd_1;
+   int nnd_2;
+   int nnd;
+
+   int npoints;
+
+   SparseMatrix A1;
+   mfem::Vector B1, X1;
+   SparseMatrix A2;
+   mfem::Vector B2, X2;
+   BlockVector *B;
+   SparseMatrix* K;
+   mutable mfem::Vector gapv;
+   mutable mfem::Vector m_xi;
+   mutable mfem::Vector xs;
+
+   mutable Array<int> m_conn; // only works for linear elements that have 4 vertices!
+   mutable DenseMatrix* coordsm;
+   mutable SparseMatrix* M;
+
+   mutable std::vector<SparseMatrix>* dM;
+
+   Array<int> Dirichlet_dof;
+   Array<double> Dirichlet_val;
+   Array<int> block_offsets;
+public: 
+   Mesh * GetMesh1() {return mesh1;}
+   Mesh * GetMesh2() {return mesh2;}
+   GridFunction & GetMesh1GridFunction() {return *x1;}
+   GridFunction & GetMesh2GridFunction() {return *x2;}
+   Array<int> & GetMesh1DirichletDofs() {return ess_tdof_list1;}
+   Array<int> & GetMesh2DirichletDofs() {return ess_tdof_list2;}
+};
+
+
+
+#endif

examples/contact/SphericalObstacleProblem.cpp

@@ -0,0 +1,175 @@
+//                          Spherical Obstacle Problem
+//
+//
+// Compile with: make SphericalobstacleProblem
+//
+// Sample runs: ./SphericalobstacleProblem
+//
+//
+// Description: This example code demonstrates the use of MFEM to solve the
+//              bound-constrained energy minimization problem
+//
+//                      minimize ||∇u||² subject to u ≥ ϕ in H¹₀.
+
+
+#include "mfem.hpp"
+#include "Problems.hpp"
+#include "IPsolver.hpp"
+#include <fstream>
+#include <iostream>
+
+using namespace std;
+using namespace mfem;
+
+
+double fRhs(const Vector &);
+double spherical_obstacle(const Vector &);
+double exact_solution_obstacle(const Vector &);
+
+int main(int argc, char *argv[])
+{
+  int FEorder = 1;     // finite element order
+  int linSolver = 0;   // linear solver 0 (direct), 1 (iterative) or 2 (iterative)
+  int maxIPMiters = 30;
+  bool iAmRoot = true;
+  int ref_levels = 3;
+  
+  OptionsParser args(argc, argv);
+  args.AddOption(&FEorder, "-o", "--order",\
+		  "Order of the finite elements.");
+  args.AddOption(&linSolver, "-linSolver", "--linearSolver", \
+       "IP-Newton linear system solution strategy.");
+  args.AddOption(&maxIPMiters, "-IPMiters", "--IPMiters",\
+		  "Maximum number of IPM iterations");
+  args.AddOption(&ref_levels, "-r", "--mesh_refinement", \
+		  "Mesh Refinement");
+  
+  args.Parse();
+  if(!args.Good())
+  {
+    args.PrintUsage(cout);
+    return 1;
+  }
+  else
+  {
+    if( iAmRoot )
+    {
+      args.PrintOptions(cout);
+    }
+  }
+
+  const char *meshFile = "../../data/disk.mesh";
+  Mesh *mesh = new Mesh(meshFile, 1, 1);
+  int dim = mesh->Dimension(); // geometric dimension of the domain
+  {
+     for (int l = 0; l < ref_levels; l++)
+     {
+        mesh->UniformRefinement();
+     }
+  }
+  double h_min, h_max, kappa_min, kappa_max;
+  mesh->GetCharacteristics(h_min, h_max, kappa_min, kappa_max);
+
+  FiniteElementCollection *fec = new H1_FECollection(FEorder, dim);
+  FiniteElementSpace      *Vh  = new FiniteElementSpace(mesh, fec);
+  Array<int> ess_tdof_list;
+  if (mesh->bdr_attributes.Size())
+  {
+     Array<int> ess_bdr(mesh->bdr_attributes.Max());
+     ess_bdr = 1;
+     Vh->GetEssentialTrueDofs(ess_bdr, ess_tdof_list);
+  }
+  
+  int dimD = Vh->GetTrueVSize();
+  Vector x0(dimD); x0 = 0.0;
+  Vector xf(dimD); xf = 0.0;
+  
+  ObstacleProblem problem(Vh, x0, &fRhs, &spherical_obstacle, ess_tdof_list);
+  InteriorPointSolver optimizer(&problem); 
+  
+  optimizer.SetTol(1.e-7);
+  optimizer.SetLinearSolver(linSolver);
+  optimizer.SetMaxIter(maxIPMiters);
+  optimizer.Mult(x0, xf);
+  
+  
+  double Einitial = problem.E(x0);
+  double Efinal = problem.E(xf);
+  cout << "Energy objective at initial point = " << Einitial << endl;
+  cout << "Energy objective at optimizer = " << Efinal << endl;
+
+  GridFunction d_gf(Vh);
+  d_gf = xf;
+
+  FunctionCoefficient dtrue_fc(exact_solution_obstacle); // exact solution
+  GridFunction dtrue_gf(Vh);
+  dtrue_gf.ProjectCoefficient(dtrue_fc);
+  
+  ParaViewDataCollection paraview_dc("BarrierProblemSolution", mesh);
+  paraview_dc.SetPrefixPath("ParaView");
+  paraview_dc.SetLevelsOfDetail(FEorder);
+  paraview_dc.SetDataFormat(VTKFormat::BINARY);
+  paraview_dc.SetHighOrderOutput(true);
+  paraview_dc.SetCycle(0);
+  paraview_dc.SetTime(0.0);
+  paraview_dc.RegisterField("d(x) (numerical)", &d_gf);
+  paraview_dc.RegisterField("d(x) (true)", &dtrue_gf);
+  paraview_dc.Save();
+  
+  FunctionCoefficient exact_coef(exact_solution_obstacle);
+  double L2_error = d_gf.ComputeL2Error(exact_coef);
+  cout << "||u - u_true||_L^2(Omega) = " << L2_error << ", hmax = " << h_max << ", hmin = " << h_min << endl;
+  
+  delete Vh;
+  delete fec;
+  delete mesh;
+  return 0;
+}
+
+
+double fRhs(const Vector &x)
+{
+  return 0.;
+}
+
+
+double spherical_obstacle(const Vector &pt)
+{
+   double x = pt(0), y = pt(1);
+   double r = sqrt(x*x + y*y);
+   double r0 = 0.5;
+   double beta = 0.9;
+
+   double b = r0*beta;
+   double tmp = sqrt(r0*r0 - b*b);
+   double B = tmp + b*b/tmp;
+   double C = -b/tmp;
+
+   if (r > b)
+   {
+      return B + r * C;
+   }
+   else
+   {
+      return sqrt(r0*r0 - r*r);
+   }
+}
+
+double exact_solution_obstacle(const Vector &pt)
+{
+   double x = pt(0), y = pt(1);
+   double r = sqrt(x*x + y*y);
+   double r0 = 0.5;
+   double a =  0.348982574111686;
+   double A = -0.340129705945858;
+
+   if (r > a)
+   {
+      return A * log(r);
+   }
+   else
+   {
+      return sqrt(r0*r0-r*r);
+   }
+} 
+

examples/contact/TwoSidedDirichletObstacleProblem.cpp

@@ -0,0 +1,143 @@
+#include "mfem.hpp"
+#include "Problems.hpp"
+#include "IPsolver.hpp"
+#include <fstream>
+#include <iostream>
+
+using namespace std;
+using namespace mfem;
+
+
+
+double fRhs(const Vector &pt);
+double obstaclel(const Vector &pt);
+double obstacleu(const Vector &pt);
+double dmanufacturedFun(const Vector &pt);
+
+int main(int argc, char *argv[])
+{
+  int FEorder = 1; // order of the finite elements
+  int linSolver = 0;
+  int maxIPMiters = 30;
+  bool iAmRoot = true;
+  int ref_levels = 1; 
+  OptionsParser args(argc, argv);
+  args.AddOption(&FEorder, "-o", "--order",\
+		  "Order of the finite elements.");
+  args.AddOption(&linSolver, "-linSolver", "--linearSolver", \
+       "IP-Newton linear system solution strategy.");
+  args.AddOption(&maxIPMiters, "-IPMiters", "--IPMiters",\
+		  "Maximum number of IPM iterations");
+  args.AddOption(&ref_levels, "-r", "--mesh_refinement", \
+		  "Mesh Refinement");
+  
+  args.Parse();
+  if(!args.Good())
+  {
+    args.PrintUsage(cout);
+    return 1;
+  }
+  else
+  {
+    if( iAmRoot )
+    {
+      args.PrintOptions(cout);
+    }
+  }
+
+  const char *meshFile = "../../data/inline-quad.mesh";
+  Mesh *mesh = new Mesh(meshFile, 1, 1);
+  int dim = mesh->Dimension(); // geometric dimension of the domain
+  for (int l = 0; l < ref_levels; l++)
+  {
+      mesh->UniformRefinement();
+  }
+
+  FiniteElementCollection *fec = new H1_FECollection(FEorder, dim);
+  FiniteElementSpace      *Vh  = new FiniteElementSpace(mesh, fec);
+  Array<int> ess_tdof_list;
+  if (mesh->bdr_attributes.Size())
+  {
+     Array<int> ess_bdr(mesh->bdr_attributes.Max());
+     ess_bdr = 1;
+     Vh->GetEssentialTrueDofs(ess_bdr, ess_tdof_list);
+  }
+  
+  double DC_val = 0.0;
+  int dimD = Vh->GetTrueVSize();
+
+  Vector x0(dimD); x0 = DC_val;
+  Vector xf(dimD); xf = 0.0;
+
+  ObstacleProblem problem(Vh, x0, &fRhs, &obstaclel, &obstacleu, ess_tdof_list);
+
+  InteriorPointSolver optimizer(&problem); 
+  optimizer.SetTol(1.e-7);
+  optimizer.SetLinearSolver(linSolver);
+  optimizer.SetMaxIter(maxIPMiters);
+  optimizer.Mult(x0, xf);
+  
+
+  GridFunction d_gf(Vh);
+
+  d_gf = xf;
+  
+  FunctionCoefficient dm_fc(dmanufacturedFun); // pseudo-manufactured solution
+  GridFunction dm_gf(Vh);
+  dm_gf.ProjectCoefficient(dm_fc);
+
+  
+  ParaViewDataCollection paraview_dc("BarrierProblemSolution", mesh);
+  paraview_dc.SetPrefixPath("ParaView");
+  paraview_dc.SetLevelsOfDetail(FEorder);
+  paraview_dc.SetDataFormat(VTKFormat::BINARY);
+  paraview_dc.SetHighOrderOutput(true);
+  paraview_dc.SetCycle(0);
+  paraview_dc.SetTime(0.0);
+  paraview_dc.RegisterField("d(x) (numerical)", &d_gf);
+  paraview_dc.RegisterField("d(x) (pseudo-manufactured)", &dm_gf);
+  paraview_dc.Save();
+  
+
+  delete Vh;
+  delete fec;
+  delete mesh;
+  return 0;
+}
+
+
+
+double dmanufacturedFun(const Vector &pt)
+{
+  double alpha = 16.5;
+  return sin(M_PI * pt(1)) * (sin(M_PI * pt(0)) - alpha * pow(pt(0) * (1. - pt(0)), 2));
+}
+
+
+// f(x) forcing term... which enters the objective energy functional
+// E(d) = 0.5 d^T K d - f^T d, where f is a discrete vector representation
+// of f(x). f(x) is such that in the absence of bound-constraints then
+// the solution of the optimization problem satisfies the PDE
+// -div(grad(d)) + d = f + homogeneous Neumann conditions on the unit interval,
+// for d(x) = cos(2 \pi x) + a0 + a3 (x^3 - 1.5 x^2), a2 = 0.2, a3 = -2
+
+double fRhs(const Vector &pt)
+{
+  double alpha = 16.5;
+  double fx;
+  fx = pow(M_PI, 2) * sin(M_PI * pt(0));
+  fx += alpha * (2. * pow(pt(0), 2) + 2. * pow(1.-pt(0), 2) - 8. * pt(0) * (1.-pt(0)));
+  fx += pow(M_PI, 2) * sin(M_PI * pt(0)) * dmanufacturedFun(pt);
+  fx *= sin(M_PI * pt(1));
+  return fx;
+}
+
+double obstaclel(const Vector &pt)
+{
+  return 0.0;
+}
+
+double obstacleu(const Vector &pt)
+{
+  return 0.08;
+}

examples/contact/block1.mesh

@@ -0,0 +1,103 @@
+MFEM mesh v1.0
+
+#
+# MFEM Geometry Types (see mesh/geom.hpp):
+#
+# POINT       = 0
+# SEGMENT     = 1
+# TRIANGLE    = 2
+# SQUARE      = 3
+# TETRAHEDRON = 4
+# CUBE        = 5
+# PRISM       = 6
+#
+
+dimension
+3
+
+elements
+9
+1   5     0      1     3     2     8    9     11    10
+1   5     2      3     5     4    10    11    13    12
+1   5     4      5     7     6    12    13    15    14
+1   5     8      9    11    10    16    17    19    18
+1   5     10    11    13    12    18    19    21    20
+1   5     12    13    15    14    20    21    23    22
+1   5     16    17    19    18    24    25    27    26
+1   5     18    19    21    20    26    27    29    28
+1   5     20    21    23    22    28    29    31    30
+
+
+
+# 0 nothing
+# 1 dirichlet bc
+# 2 contact
+boundary
+30
+1   3    1     0     2     3
+1   3    3     2     4     5
+1   3    5     4     6     7
+1   3    24    25    27    26
+1   3    26    27    29    28
+1   3    28    29    31    30
+2   3    2     0     8    10
+2   3    4     2    10    12
+2   3    6     4    12    14
+2  3    10    8    16    18
+2  3    12    10    18    20
+2  3    14    12    20    22
+2  3    18    16    24    26
+2  3    20    18    26    28
+2  3    22    20    28    30
+3  3    1     3    11    9
+3  3     3     5    13    11
+3  3     5     7    15    13
+3  3    9    11    19    17
+3  3    11    13    21    19
+3  3    13    15    23    21
+3  3    17    19    27    25
+3  3    19    21    29    27
+3  3    21    23    31    29
+1  3     8     0     1    9
+1  3    16     8    9    17
+1  3    24    16    17    25
+1  3     6    14    15     7
+1  3    14    22    23    15
+1  3    22    30    31    23
+
+
+vertices
+32
+3
+   -1.0000         0         0
+         0         0         0
+   -1.0000    0.3000         0
+         0    0.3000         0
+   -1.0000    0.6500         0
+         0    0.6500         0
+   -1.0000    1.0000         0
+         0    1.0000         0
+   -1.0000         0    0.3000
+         0         0    0.3000
+   -1.0000    0.3000    0.3500
+         0    0.3000    0.3500
+   -1.0000    0.6500    0.3000
+         0    0.6500    0.3000
+   -1.0000    1.0000    0.3000
+         0    1.0000    0.3000
+   -1.0000         0    0.6500
+         0         0    0.6500
+   -1.0000    0.3000    0.6500
+         0    0.3000    0.6500
+   -1.0000    0.6500    0.6500
+         0    0.6500    0.6500
+   -1.0000    1.0000    0.6500
+         0    1.0000    0.6500
+   -1.0000         0    1.0000
+         0         0    1.0000
+   -1.0000    0.3000    1.0000
+         0    0.3000    1.0000
+   -1.0000    0.6500    1.0000
+         0    0.6500    1.0000
+   -1.0000    1.0000    1.0000
+         0    1.0000    1.0000

examples/contact/exQPContactBlockTL.cpp

@@ -0,0 +1,246 @@
+//                                Quadratic-Programming (QP) Contact example
+//
+// Compile with: make exQPContactBlockTL
+//
+// Sample runs:  ./exQPContactBlockTL
+
+
+#include <fstream>
+#include <iostream>
+#include <array>
+
+#include "mfem.hpp"
+#include "Problems.hpp"
+#include "IPsolver.hpp"
+
+using namespace std;
+using namespace mfem;
+
+
+int main(int argc, char *argv[])
+{
+  int linSolver = 0;
+  int maxIPMiters = 30;
+  bool iAmRoot = true;
+  int ref_levels = 0;
+
+  OptionsParser args(argc, argv);
+  args.AddOption(&linSolver, "-linSolver", "--linearSolver", \
+       "IP-Newton linear system solution strategy.");
+  args.AddOption(&maxIPMiters, "-IPMiters", "--IPMiters",\
+		  "Maximum number of IPM iterations");
+  args.AddOption(&ref_levels, "-r", "--mesh_refinement", \
+		  "Mesh Refinement");
+
+   
+  args.Parse();
+  if(!args.Good())
+  {
+    args.PrintUsage(cout);
+    return 1;
+  }
+  else
+  {
+    if( iAmRoot )
+    {
+      args.PrintOptions(cout);
+    }
+  }
+	
+  Mesh * mesh1 = new Mesh("block1.mesh", 1, 1);
+  Mesh * mesh2 = new Mesh("rotatedblock2.mesh", 1, 1);
+  for(int i = 0; i < ref_levels; i++)
+  {
+     mesh1->UniformRefinement();
+     mesh2->UniformRefinement(); 
+  }
+
+  // Create an instance of the nlp
+  ExContactBlockTL * contact = new ExContactBlockTL(mesh1, mesh2, 1);
+  int ndofs = contact->GetDimD();
+  int nconstraints = contact->GetDimS();
+  
+  // set up a QP-problem
+  // E(d) = 1 / 2 d^T K d + f^T d
+  // g(d) = J d + g0
+  // where K, J, f and g0 are evaluated at d0 (a valid configuration)
+  
+  // to do: seems more appropriate to evaluate at a valid configuration...
+  //        that is one where the Dirichlet conditions hold... need to pull
+  //        this data from contactBlockTL...
+  Vector d0(ndofs); d0 = 0.0;
+  Array<int> ess_tdofs1 = contact->GetMesh1DirichletDofs();
+  Array<int> ess_tdofs2 = contact->GetMesh2DirichletDofs();
+  int sz1 = ess_tdofs1.Size();
+  int sz2 = ess_tdofs2.Size();
+  Array<int> DirichletDofs(sz1+sz2);
+  for (int i = 0; i<sz1; i++)
+  {
+    DirichletDofs[i] = ess_tdofs1[i];
+  }
+  for (int i = 0; i<sz2; i++)
+  {
+    DirichletDofs[i+sz1] = ess_tdofs2[i]+contact->GetVh1().GetTrueVSize();
+  }
+  GridFunction x1 = contact->GetMesh1GridFunction();
+  GridFunction x2 = contact->GetMesh2GridFunction();
+
+  SparseMatrix *K;
+  Vector f(ndofs); f = 0.0;
+  contact->DdE(d0, f); K = contact->DddE(d0);
+  d0.SetVector(x1,0);
+  d0.SetVector(x2,x1.Size());
+  SparseMatrix *J;
+  Vector g0(nconstraints); g0 = 0.0;
+  contact->g(d0, g0); J = contact->Ddg(d0);
+  Vector temp(nconstraints);
+  J->Mult(d0, temp);
+  g0.Add(-1.0, temp);
+
+  // check which rows of the Jacobian are zero!
+  Vector ei(nconstraints); ei = 0.0;
+  Vector JTei(ndofs); JTei = 0.0;
+
+  double normJTei;
+
+  Array<int> nonZeroRows;
+  for(int i = 0; i < nconstraints; i++)
+  {
+    Array<int> col_tmp;
+    Vector v_tmp; v_tmp = 0.0;
+    J->GetRow(i, col_tmp, v_tmp);
+    normJTei = v_tmp.Norml2();
+    if (normJTei > 1.e-12)
+    {
+      nonZeroRows.Append(i);
+    }
+  }
+  mfem::out << J->Height() << " linearized constraints\n";
+  mfem::out << nonZeroRows.Size() << " (reduced) linearized constraints\n"; 
+  
+  // remove zero rows of the gap function Jacobian and corresponding gap function entries
+  SparseMatrix * Jreduced = new SparseMatrix(nonZeroRows.Size(), ndofs);
+  Vector g0reduced(nonZeroRows.Size()); g0reduced = 0.0; 
+  
+  
+  for(int i = 0; i < nonZeroRows.Size(); i++)
+  {
+    Array<int> col_tmp;
+    Vector v_tmp; v_tmp = 0.0;
+    J->GetRow(nonZeroRows[i], col_tmp, v_tmp);
+    
+    /* obtain subset of columns of the given nonZero Jacobian row that are not Dirichlet constrained */
+    bool freeDof;
+    Array<int> free_col_indicies;
+    for(int j = 0; j < col_tmp.Size(); j++)
+    {
+      freeDof = true;
+      for(int k = 0; k < DirichletDofs.Size(); k++)
+      {
+        if(col_tmp[j] == DirichletDofs[k])
+        {
+          freeDof = false;
+        }
+      }
+      if(freeDof)
+      {
+        free_col_indicies.Append(j);
+      }
+    }
+
+    Array<int> col_tmp_reduced(free_col_indicies.Size());
+    Vector v_tmp_reduced(free_col_indicies.Size());
+    for(int j = 0; j < free_col_indicies.Size(); j++)
+    {
+      col_tmp_reduced[j] = col_tmp[free_col_indicies[j]];
+      v_tmp_reduced(j)   = v_tmp(free_col_indicies[j]);
+    }
+
+    Jreduced->SetRow(i, col_tmp_reduced, v_tmp_reduced);
+    g0reduced(i) = g0(nonZeroRows[i]);
+  }
+
+  QPOptProblem *QPContact = new QPOptProblem(*K, *Jreduced, f, g0reduced);
+  
+  InteriorPointSolver * QPContactOptimizer = new InteriorPointSolver(QPContact);
+  QPContactOptimizer->SetTol(1.e-6);
+  QPContactOptimizer->SetLinearSolver(linSolver);
+  Vector x0(ndofs); x0 = 0.0;
+  x0.SetVector(x1,0);
+  x0.SetVector(x2,x1.Size());
+
+  Vector xf(ndofs); xf = 0.0;
+  QPContactOptimizer->Mult(x0, xf);
+  
+  MFEM_VERIFY(QPContactOptimizer->GetConverged(), "Interior point solver did not converge.");
+  double Einitial = QPContact->E(x0);
+  double Efinal = QPContact->E(xf);
+  cout << "Energy objective at initial point = " << Einitial << endl;
+  cout << "Energy objective at QP optimizer = " << Efinal << endl;
+ 
+  
+  
+  int gdim = mesh1->Dimension();
+  FiniteElementCollection * fec = new H1_FECollection(1, gdim);
+  FiniteElementSpace * fespace1 = new FiniteElementSpace(mesh1, fec, gdim, Ordering::byVDIM);
+  FiniteElementSpace * fespace2 = new FiniteElementSpace(mesh2, fec, gdim, Ordering::byVDIM);
+   
+  GridFunction x1_gf(fespace1);
+  GridFunction x2_gf(fespace2);
+
+  int ndof1 = fespace1->GetTrueVSize();
+  int ndof2 = fespace2->GetTrueVSize();
+  int ndof  = ndof1 + ndof2;
+  for(int i = 0; i < ndof1; i++)
+  {
+    x1_gf(i) = xf(i);
+  }
+  for(int i = ndof1; i < ndof; i++)
+  {
+    x2_gf(i - ndof1) = xf(i);
+  }
+
+  mesh1->SetNodalFESpace(fespace1);
+  mesh2->SetNodalFESpace(fespace2);
+  GridFunction *nodes1 = mesh1->GetNodes();
+  GridFunction *nodes2 = mesh2->GetNodes();
+
+  {
+    *nodes1 += x1_gf;
+    *nodes2 += x2_gf;
+  }
+  
+
+  ParaViewDataCollection paraview_dc1("QPContactBody1", mesh1);
+  paraview_dc1.SetPrefixPath("ParaView");
+  paraview_dc1.SetLevelsOfDetail(1);
+  paraview_dc1.SetDataFormat(VTKFormat::BINARY);
+  paraview_dc1.SetHighOrderOutput(true);
+  paraview_dc1.SetCycle(0);
+  paraview_dc1.SetTime(0.0);
+  paraview_dc1.RegisterField("Body1", &x1_gf);
+  paraview_dc1.Save();
+  
+  ParaViewDataCollection paraview_dc2("QPContactBody2", mesh2);
+  paraview_dc2.SetPrefixPath("ParaView");
+  paraview_dc2.SetLevelsOfDetail(1);
+  paraview_dc2.SetDataFormat(VTKFormat::BINARY);
+  paraview_dc2.SetHighOrderOutput(true);
+  paraview_dc2.SetCycle(0);
+  paraview_dc2.SetTime(0.0);
+  paraview_dc2.RegisterField("Body2", &x2_gf);
+  paraview_dc2.Save();
+
+  delete fespace1;
+  delete fespace2;
+  delete fec;
+  delete mesh1;
+  delete mesh2;
+  
+  delete QPContact;
+  delete QPContactOptimizer;
+  
+  delete Jreduced;
+  delete contact;
+  return 0;
+}

examples/contact/makefile

@@ -0,0 +1,125 @@
+# Copyright (c) 2010-2023, Lawrence Livermore National Security, LLC. Produced
+# at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+# LICENSE and NOTICE for details. LLNL-CODE-806117.
+#
+# This file is part of the MFEM library. For more information and source code
+# availability visit https://mfem.org.
+#
+# MFEM is free software; you can redistribute it and/or modify it under the
+# terms of the BSD-3 license. We welcome feedback and contributions, see file
+# CONTRIBUTING.md for details.
+
+# Use the MFEM build directory
+MFEM_DIR ?= ../..
+MFEM_BUILD_DIR ?= ../../
+SRC = $(if $(MFEM_DIR:../..=),$(MFEM_DIR)/examples/contact/,)
+CONFIG_MK = $(MFEM_BUILD_DIR)/config/config.mk
+
+MFEM_LIB_FILE = mfem_is_not_built
+-include $(CONFIG_MK)
+
+SEQ_EXAMPLES = ObstacleProblem SphericalObstacleProblem DirichletObstacleProblem exQPContactBlockTL
+PAR_EXAMPLES = ParObstacleProblem
+EXAMPLES = $(SEQ_EXAMPLES) $(PAR_EXAMPLES)
+
+ifeq ($(MFEM_USE_SUITESPARSE),NO)
+$(SEQ_EXAMPLES):
+	$(error MFEM is not configured with SUITESPARSE)
+endif
+
+ifeq ($(MFEM_USE_MUMPS),NO)
+ifeq ($(MFEM_USE_MKL_CPARDISO), NO)
+$(PAR_EXAMPLES):
+	$(error MFEM is not configured with MUMPS or CPARDISO)
+endif
+endif
+
+all: $(EXAMPLES)
+
+ObstacleProblem: ObstacleProblem.o Problems.o IPsolver.o $(MFEM_LIB_FILE)
+	$(MFEM_CXX) $(MFEM_FLAGS) ObstacleProblem.o Problems.o IPsolver.o -o $@ $(MFEM_LIBS)
+
+SphericalObstacleProblem: SphericalObstacleProblem.o Problems.o IPsolver.o $(MFEM_LIB_FILE)
+	$(MFEM_CXX) $(MFEM_FLAGS) SphericalObstacleProblem.o Problems.o IPsolver.o -o $@ $(MFEM_LIBS)
+
+DirichletObstacleProblem: DirichletObstacleProblem.o Problems.o IPsolver.o $(MFEM_LIB_FILE)
+	$(MFEM_CXX) $(MFEM_FLAGS) DirichletObstacleProblem.o Problems.o IPsolver.o -o $@ $(MFEM_LIBS)
+
+DirichletObstacleProblemVariant: DirichletObstacleProblemVariant.o Problems.o IPsolver.o $(MFEM_LIB_FILE)
+	$(MFEM_CXX) $(MFEM_FLAGS) DirichletObstacleProblemVariant.o Problems.o IPsolver.o -o $@ $(MFEM_LIBS)
+
+TwoSidedDirichletObstacleProblem: TwoSidedDirichletObstacleProblem.o Problems.o IPsolver.o $(MFEM_LIB_FILE)
+	$(MFEM_CXX) $(MFEM_FLAGS) TwoSidedDirichletObstacleProblem.o Problems.o IPsolver.o -o $@ $(MFEM_LIBS)
+
+
+exQPContactBlockTL: exQPContactBlockTL.o Problems.o IPsolver.o $(MFEM_LIB_FILE)
+	$(MFEM_CXX) $(MFEM_FLAGS) exQPContactBlockTL.o Problems.o IPsolver.o -o $@ $(MFEM_LIBS)
+
+ParTest: ParTest.o $(MFEM_LIB_FILE)
+	$(MFEM_CXX) $(MFEM_FLAGS) ParTest.o -o $@ $(MFEM_LIBS)
+
+
+ObstacleProblem.o: $(SRC)ObstacleProblem.cpp $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -c $< 
+
+SphericalObstacleProblem.o: $(SRC)SphericalObstacleProblem.cpp $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -c $< 
+
+DirichletObstacleProblem.o: $(SRC)DirichletObstacleProblem.cpp $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -c $<
+
+DirichletObstacleProblemVariant.o: $(SRC)DirichletObstacleProblemVariant.cpp $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -c $<
+
+TwoSidedDirichletObstacleProblem.o: $(SRC)TwoSidedDirichletObstacleProblem.cpp $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -c $<
+
+exQPContactBlockTL.o: $(SRC)exQPContactBlockTL.cpp $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -c $<
+
+Problems.o: $(SRC)Problems.cpp $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -c $< 
+
+IPsolver.o: $(SRC)IPsolver.cpp $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -c $<
+
+ParObstacleProblem: ParObstacleProblem.o ParProblems.o ParIPsolver.o $(MFEM_LIB_FILE)
+	$(MFEM_CXX) $(MFEM_FLAGS) ParObstacleProblem.o ParProblems.o ParIPsolver.o -o $@ $(MFEM_LIBS)
+
+ParSphericalObstacleProblem: ParSphericalObstacleProblem.o ParProblems.o ParIPsolver.o $(MFEM_LIB_FILE)
+	$(MFEM_CXX) $(MFEM_FLAGS) ParSphericalObstacleProblem.o ParProblems.o ParIPsolver.o -o $@ $(MFEM_LIBS)
+
+ParObstacleProblem.o: $(SRC)ParObstacleProblem.cpp $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -c $< 
+
+ParSphericalObstacleProblem.o: $(SRC)ParSphericalObstacleProblem.cpp $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -c $< 
+
+ParProblems.o: $(SRC)ParProblems.cpp $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -c $< 
+
+ParIPsolver.o: $(SRC)ParIPsolver.cpp $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -c $<
+
+ParTest.o: $(SRC)ParTest.cpp $(CONFIG_MK)
+	$(MFEM_CXX) $(MFEM_FLAGS) -c $<
+
+clean: clean-build 
+
+clean-build:
+	rm -f *.o *~ $(SEQ_EXAMPLES) $(PAR_EXAMPLES)
+	rm -rf *.dSYM *.TVD.*breakpoints
+
+# For out-of-source builds, link the data files from the source tree:
+ifneq ($(SRC),)
+DATA_FILES = block1.mesh rotatedblock2.mesh
+$(DATA_FILES): %: $(SRC)%
+	ln -sf $(<) .
+copy-data: | $(DATA_FILES)
+# For out-of-source builds, the test and sample runs for 'field-interp' need
+# data from the meshing miniapps directory:
+exQPContactBlockTL: | mesh-data
+.PHONY: mesh-data
+mesh-data:
+	$(MAKE) -C ./ copy-data
+endif

examples/contact/nodepair.hpp

@@ -0,0 +1,896 @@
+
+using namespace std;
+using namespace mfem;
+
+void BasisEval(const Vector xi, Vector &N, DenseMatrix &dNdxi) // dNdxi is 2*4
+{
+   N[0] = 0.25*(1-xi[0])*(1-xi[1]);
+   N[1] = 0.25*(1+xi[0])*(1-xi[1]);
+   N[2] = 0.25*(1+xi[0])*(1+xi[1]);
+   N[3] = 0.25*(1-xi[0])*(1+xi[1]);
+
+   dNdxi(0,0) = 0.25*(-1+xi[1]);
+   dNdxi(0,1) = 0.25*(1-xi[1]);
+   dNdxi(0,2) = 0.25*(1+xi[1]);
+   dNdxi(0,3) = 0.25*(-1-xi[1]);
+   dNdxi(1,0) = 0.25*(-1+xi[0]);
+   dNdxi(1,1) = 0.25*(-1-xi[0]);
+   dNdxi(1,2) = 0.25*(1+xi[0]);
+   dNdxi(1,3) = 0.25*(1-xi[0]);
+}
+
+
+void BasisEvalDerivs(const Vector xi, Vector& N, DenseMatrix& dNdxi,
+                     DenseMatrix& dN2dxi)
+{
+   N[0] = 0.25*(1-xi[0])*(1-xi[1]);
+   N[1] = 0.25*(1+xi[0])*(1-xi[1]);
+   N[2] = 0.25*(1+xi[0])*(1+xi[1]);
+   N[3] = 0.25*(1-xi[0])*(1+xi[1]);
+
+   dNdxi.SetSize(2,4); dNdxi = 0.0;
+   dN2dxi.SetSize(3,4);
+   dN2dxi = 0.0;  // first row dxi2, second detadxi, third deta2
+
+   dNdxi(0,0) = 0.25*(-1+xi[1]); dNdxi(0,1) = 0.25*(1-xi[1]);
+   dNdxi(0,2) = 0.25*(1+xi[1]);  dNdxi(0,3) = 0.25*(-1-xi[1]);
+   dNdxi(1,0) = 0.25*(-1+xi[0]); dNdxi(1,1) = 0.25*(-1-xi[0]);
+   dNdxi(1,2) = 0.25*(1+xi[0]);  dNdxi(1,3) = 0.25*(1-xi[0]);
+
+   dN2dxi(1,0) = 0.25; dN2dxi(1,1) = -0.25; dN2dxi(1,2) = 0.25;
+   dN2dxi(1,3) = -0.25;
+}
+
+// returns the vector and matrix form of the shape functions and its derivative
+void BasisVectorDerivs(const Vector xi, DenseMatrix& N, DenseMatrix& dNdxi,
+                       DenseMatrix& ddNdxi)
+{
+   N.SetSize(3,12); N = 0.0;
+   N(0,0) = 0.25*(1-xi[0])*(1-xi[1]); N(0,3) = 0.25*(1+xi[0])*(1-xi[1]);
+   N(0,6) = 0.25*(1+xi[0])*(1+xi[1]); N(0,9) = 0.25*(1-xi[0])*(1+xi[1]);
+
+   N(1,1) = 0.25*(1-xi[0])*(1-xi[1]); N(1,4) = 0.25*(1+xi[0])*(1-xi[1]);
+   N(1,7) = 0.25*(1+xi[0])*(1+xi[1]); N(1,10) = 0.25*(1-xi[0])*(1+xi[1]);
+
+   N(2,2) = 0.25*(1-xi[0])*(1-xi[1]); N(2,5) = 0.25*(1+xi[0])*(1-xi[1]);
+   N(2,8) = 0.25*(1+xi[0])*(1+xi[1]); N(2,11) = 0.25*(1-xi[0])*(1+xi[1]);
+
+   dNdxi.SetSize(3*2, 3*4);  dNdxi = 0.0;
+   dNdxi(0,0) = 0.25*(-1+xi[1]);  dNdxi(0,3) = 0.25*(1-xi[1]);
+   dNdxi(0,6) = 0.25*(1+xi[1]);  dNdxi(0,9) = 0.25*(-1-xi[1]);
+   dNdxi(1,1) = 0.25*(-1+xi[1]);  dNdxi(1,4) = 0.25*(1-xi[1]);
+   dNdxi(1,7) = 0.25*(1+xi[1]);  dNdxi(1,10) = 0.25*(-1-xi[1]);
+   dNdxi(2,2) = 0.25*(-1+xi[1]);  dNdxi(2,5) = 0.25*(1-xi[1]);
+   dNdxi(2,8) = 0.25*(1+xi[1]);  dNdxi(2,11) = 0.25*(-1-xi[1]);
+
+   dNdxi(3,0) = 0.25*(-1+xi[0]);  dNdxi(3,3) = 0.25*(-1-xi[0]);
+   dNdxi(3,6) = 0.25*(1+xi[0]);  dNdxi(3,9) = 0.25*(1-xi[0]);
+   dNdxi(4,1) = 0.25*(-1+xi[0]);  dNdxi(4,4) = 0.25*(-1-xi[0]);
+   dNdxi(4,7) = 0.25*(1+xi[0]);  dNdxi(4,10) = 0.25*(1-xi[0]);
+   dNdxi(5,2) = 0.25*(-1+xi[0]);  dNdxi(5,5) = 0.25*(-1-xi[0]);
+   dNdxi(5,8) = 0.25*(1+xi[0]);  dNdxi(5,11) = 0.25*(1-xi[0]);
+
+   ddNdxi.SetSize(3*4, 3*4); ddNdxi = 0.0;
+   ddNdxi(3,0) = 0.25; ddNdxi(3,3) = -0.25;
+   ddNdxi(3,6) = 0.25; ddNdxi(3,9) = -0.25;
+   ddNdxi(4,1) = 0.25; ddNdxi(4,4) = -0.25;
+   ddNdxi(4,7) = 0.25; ddNdxi(4,10) = -0.25;
+   ddNdxi(5,2) = 0.25; ddNdxi(5,5) = -0.25;
+   ddNdxi(5,8) = 0.25; ddNdxi(5,11) = -0.25;
+
+   ddNdxi(6,0) = 0.25; ddNdxi(6,3) = -0.25;
+   ddNdxi(6,6) = 0.25; ddNdxi(6,9) = -0.25;
+   ddNdxi(7,1) = 0.25; ddNdxi(7,4) = -0.25;
+   ddNdxi(7,7) = 0.25; ddNdxi(7,10) = -0.25;
+   ddNdxi(8,2) = 0.25; ddNdxi(8,5) = -0.25;
+   ddNdxi(8,8) = 0.25; ddNdxi(8,11) = -0.25;
+}
+
+
+void cross(const Vector a, const Vector b, Vector& c)
+{
+   assert(a.Size()==3);
+   c.SetSize(3);
+   c[0] = a[1]*b[2] - a[2]*b[1];
+   c[1] = -a[0]*b[2] + b[0]*a[2];
+   c[2] = a[0]*b[1] - a[1]*b[0];
+
+}
+// a outer b
+void outer(const Vector a, const Vector b, DenseMatrix& c)
+{
+   int m = a.Size();
+   int n = b.Size();
+   assert(c.Height()==m);
+   assert(c.Width() ==n);
+   for (int i=0; i<m; i++)
+   {
+      for (int j=0; j<n; j++)
+      {
+         c(i,j) = a[i]*b[j];
+      }
+   }
+}
+// dphidxi 2*4
+// coords 4*3
+void ComputeNormal(const DenseMatrix& dphidxi, const DenseMatrix& coords,
+                   Vector& normal, double& nnorm)
+{
+
+   DenseMatrix dxdxi(2,3);
+   Mult(dphidxi, coords, dxdxi);
+   Vector dxdxi1(3);
+   Vector dxdxi2(3);
+
+   dxdxi.GetRow(0,dxdxi1);
+   dxdxi.GetRow(1,dxdxi2);
+
+   cross(dxdxi1, dxdxi2, normal);  // is there a cross product?  no
+   // VectorCrossProductCoefficient::Eval has hard-coded cross product
+   nnorm = normal.Norml2( );
+   normal /=  nnorm;
+}
+
+void SlaveToMaster(const DenseMatrix& m_coords, const Vector& s_x, Vector& xi)
+{
+   bool converged = false;
+   bool pt_on_elem = false;
+   int dim = 3;
+   xi.SetSize(dim-1);
+   xi = 0.0;
+   int max_iter = 15;
+   double off_el_xi = 1e-2;
+   double proj_newton_tol = 1e-13;
+   double proj_max_gap = 0.5;
+   Vector gap_v(dim);
+   // warm start from linear solution
+
+   for (int it=0; it<max_iter; it++)
+   {
+      //cout<<it<<endl;
+      Vector m_N(4);
+      m_N = 0.;
+      DenseMatrix m_dN(2,4);
+      m_dN = 0.;
+      DenseMatrix m_dN2(3,4);
+      m_dN2 = 0.;
+      BasisEvalDerivs(xi, m_N, m_dN, m_dN2);
+
+      Vector x_c(dim);
+      m_coords.MultTranspose(m_N, x_c);
+
+      gap_v = s_x;
+      gap_v -= x_c;
+
+      DenseMatrix m_dx(2,3);
+      m_dx = 0.;
+      Mult(m_dN, m_coords, m_dx);
+
+      Vector r(dim-1);
+      r = 0.0;
+      m_dx.Mult(gap_v, r);
+
+      if (r.Normlinf() < proj_newton_tol)
+      {
+         converged = true;
+         break;
+      }
+
+      DenseMatrix drdxi(dim-1,dim-1);
+      drdxi = 0.;
+      MultABt(m_dx, m_dx, drdxi); // m_dx * m_dx.T
+      drdxi *= -1.0;
+
+      DenseMatrix m_dx2(3,3); m_dx2 = 0.0;
+      Mult(m_dN2,m_coords, m_dx2);
+
+      //m_d2x = m_dN(:,:,2) * m_elem_coords(1:4,:);  //m_dN(:,:,2) is 3*4
+      for (int d=0; d<3; d++)
+      {
+         DenseMatrix Mtemp(2,2); Mtemp = 0.0;
+         Mtemp(0,0) = m_dx2(0,d); Mtemp(0,1) = m_dx2(1,d);
+         Mtemp(1,0) = m_dx2(1,d); Mtemp(1,1) = m_dx2(2,d);
+
+         drdxi.Add(gap_v[d], Mtemp);
+      }
+
+      //cond_num = rcond(drdxi);  condition number?
+      //drdxi.TestInversion();
+      DenseMatrixInverse drdxi_inv(drdxi);
+      Vector xi_tmp(dim-1);
+
+      drdxi_inv.Mult(r,xi_tmp);
+      xi -= xi_tmp;
+   }
+   if (!converged)
+   {
+      xi = 0.0;
+   }
+   off_el_xi += 1 ; // tolerance of offset of xi outside [-1,1]
+
+   //cout<<gap_v.Norml2()<<" " <<xi.Normlinf()<<endl;
+   //
+   // Discuss with Frank... what is happening here
+   if (gap_v.Norml2() < proj_max_gap && xi.Normlinf() <= off_el_xi)
+   {
+      pt_on_elem = true;
+   }
+
+   if (pt_on_elem)
+   {
+     //cout << "convergence of node to segment projection? " << converged << endl;
+     //for(int i = 0; i < 2; i++)
+     //{
+     //  cout << "xi_" << i << " = " << xi(i) << endl;
+     //}
+   }
+   MFEM_VERIFY(pt_on_elem == true, "xi went out of bounds");
+   MFEM_VERIFY(converged == true, "projection didn't converge");
+}
+
+
+
+// m_coords is expected to be 4 * 3
+void  ComputeGapJacobian(const Vector x_s, const Vector xi,
+                         const DenseMatrix m_coords,
+                         double& gap, Vector& normal, Vector& dgdxm, Vector& dgdxs)
+{
+   Vector m_N(4);
+   DenseMatrix m_dN(2,4);
+   DenseMatrix m_dN2(3,4);
+   BasisEvalDerivs(xi, m_N, m_dN, m_dN2);
+
+   Vector x_c(3);
+   m_coords.MultTranspose(m_N, x_c);
+
+   Vector gap_v(3); gap_v = 0.0;
+   gap_v  = x_s;
+   gap_v -= x_c;
+
+   DenseMatrix m_dx(2,3);
+   Mult(m_dN, m_coords, m_dx);
+
+   double nnorm = 0;
+   ComputeNormal(m_dN, m_coords, normal, nnorm);
+
+   gap = gap_v * normal; // gap function value, dot product between vectors
+
+   //dr_dx = zeros(2,4,3); % nsegment, nodes in quad, ndim
+
+   DenseMatrix dr_dx_res1(4,3); dr_dx_res1 = 0.;
+   DenseMatrix dr_dx_res2(4,3); dr_dx_res2 = 0.;
+
+   Vector m_dxrow1(3);
+   m_dx.GetRow(0, m_dxrow1);
+   outer(m_N, m_dxrow1, dr_dx_res1);// 4*1 times 1*3
+   dr_dx_res1 *= -1.0;
+
+   Vector m_dxrow2(3);
+   m_dx.GetRow(1, m_dxrow2);
+   outer(m_N, m_dxrow2, dr_dx_res2);// 4*1 times 1*3
+   dr_dx_res2 *= -1.0;
+
+   Vector m_dNrow1(4); m_dN.GetRow(0, m_dNrow1);
+   Vector m_dNrow2(4); m_dN.GetRow(1, m_dNrow2);
+
+   DenseMatrix dr_dx_res1_tmp(4,3); dr_dx_res1_tmp = 0.;
+   DenseMatrix dr_dx_res2_tmp(4,3); dr_dx_res2_tmp = 0.;
+   outer(m_dNrow1, gap_v, dr_dx_res1_tmp);// 4*1 times 1*3
+   outer(m_dNrow2, gap_v, dr_dx_res2_tmp);// 4*1 times 1*3
+
+   dr_dx_res1 += dr_dx_res1_tmp; // outer product in vector?
+   dr_dx_res2 += dr_dx_res2_tmp;
+
+
+   DenseMatrix K_dxidx1(2,2); // 2*2
+   K_dxidx1 = 0.;
+   MultABt(m_dx, m_dx, K_dxidx1); // m_dx * m_dx.T
+
+   Vector v_dxidx2(4);
+   m_coords.Mult(gap_v, v_dxidx2); // m_coords * gap_v;  // 4*3 * 3 = 4
+
+   DenseMatrix K_dxidx2(2,2); K_dxidx2 = 0.0;
+
+   Vector m_dN2row1(4); m_dN2.GetRow(0, m_dN2row1);
+   Vector m_dN2row2(4); m_dN2.GetRow(1, m_dN2row2);
+   Vector m_dN2row3(4); m_dN2.GetRow(2, m_dN2row3);
+   // how to get 2nd order? multidimensional matrix?
+   K_dxidx2(0,0) = m_dN2row1 * v_dxidx2; // how would 4*1 * 1*4 be computed?
+   K_dxidx2(0,1) = m_dN2row2 * v_dxidx2;
+   K_dxidx2(1,0) = m_dN2row2 * v_dxidx2;
+   K_dxidx2(1,1) = m_dN2row3 * v_dxidx2;
+
+   DenseMatrix K_dxidx(2,2);
+   K_dxidx -= K_dxidx1;
+   K_dxidx += K_dxidx2;
+
+   // resize the vectors and matrices
+   Vector dxidx(24); dxidx = 0.0;
+   Vector drdx_r(24); drdx_r = 0.0;
+
+   for (int i=0; i<4; i++)
+   {
+      for (int j=0; j<3; j++)
+      {
+         drdx_r[4*j+i] = dr_dx_res1(i,j);
+         drdx_r[4*j+i+12] = dr_dx_res2(i,j);
+
+      }
+   }
+   //drdx_r(1:4*3,1) = reshape(dr_dx_res(:,:,1),4*3,1);
+   //drdx_r(4*3+1:2*4*3,1) = reshape(dr_dx_res(:,:,2),4*3,1);
+   DenseMatrix drdx_K(24,24); drdx_K = 0.;
+   for (int i =0; i<12; i++)
+   {
+      drdx_K(i,i) = K_dxidx(0,0);
+      drdx_K(i,12+i) = K_dxidx(0,1);
+      drdx_K(12+i,i) = K_dxidx(1,0);
+      drdx_K(12+i,12+i) = K_dxidx(1,1);
+   }
+
+   DenseMatrixInverse drdxK_inv(drdx_K);
+   drdxK_inv.Mult(drdx_r,dxidx);
+   // LinearSolve (drdx_K,drdx_r, dxidx) ; //???
+   dxidx *= -1.0;
+
+
+
+   Vector drdxs_r(6);
+   drdxs_r[0] = m_dx(0,0); drdxs_r[1] = m_dx(0,1); drdxs_r[2] = m_dx(0,2);
+   drdxs_r[3] = m_dx(1,0); drdxs_r[4] = m_dx(1,1); drdxs_r[5] = m_dx(1,2);
+
+   DenseMatrix drdxs_K(6,6); drdxs_K = 0.;
+   for (int i=0; i<3; i++)
+   {
+      drdxs_K(i,i) = K_dxidx(0,0);
+      drdxs_K(i,3+i) = K_dxidx(0,1);
+      drdxs_K(i+3,i) = K_dxidx(1,0);
+      drdxs_K(i+3,i+3) = K_dxidx(1,1);
+   }
+
+   Vector dxidxs(6); dxidxs = 0.0;
+   DenseMatrixInverse drdxsK_inv(drdxs_K);
+   drdxsK_inv.Mult(drdxs_r,dxidxs);
+   dxidxs *= -1.0;
+   //dxidxs = -drdxs_K\drdxs_r;
+
+   //dxidx = reshape(dxidx, 4,3,2); dxidxs = reshape(dxidxs, 1,3,2);
+
+   dgdxm.SetSize(12); dgdxm = 0.;
+   DenseMatrix dgdxm_tmp(4,3);
+   outer(m_N, normal,dgdxm_tmp);
+   for (int i=0; i<4; i++)
+   {
+      for (int j=0; j<3; j++)
+      {
+         dgdxm[3*i+j] = -dgdxm_tmp(i,j);
+      }
+   }
+   //dxidx_M = -m_dN(1:2,:,1) * (m_coords(1:4,:)*normal'); % this turns out to be 0
+
+   dgdxs.SetSize(3);
+   dgdxs += normal;
+   //dgdxs = dgdxs + dxidx_M(1) * dxidxs(:,:,1) + dxidx_M(2) * dxidxs(:,:,2);
+};
+
+void ComputeGapHessian(const Vector x_s, const Vector xi,
+                       const DenseMatrix m_coords,
+                       DenseMatrix& dg2dx)
+{
+   Vector m_N(4);
+   DenseMatrix m_dN(2,4);
+   DenseMatrix m_dN2(3,4);
+   BasisEvalDerivs(xi, m_N, m_dN, m_dN2);
+
+   int dim = 3;
+   int num_dofs1 = dim;
+   int num_dofs2 = 4*dim;
+   int num_dofs = num_dofs1 + num_dofs2;
+   dg2dx.SetSize(num_dofs,num_dofs); dg2dx = 0.0;
+
+   Vector x_c(3);
+   m_coords.MultTranspose(m_N,x_c);
+
+   Vector gap_v(3); gap_v = 0.0;
+   gap_v  = x_s;
+   gap_v -= x_c;
+
+   DenseMatrix m_dx(2,3);
+   Mult(m_dN, m_coords, m_dx);
+
+   DenseMatrix m_dx2(3,3); m_dx2 = 0.0;
+   Mult(m_dN2,m_coords, m_dx2);
+   double nnorm = 0.0;
+   Vector normal(3); normal = 0.0;
+   ComputeNormal(m_dN, m_coords, normal, nnorm);
+
+   double gap = gap_v * normal; // gap function value, dot product between vectors
+
+   DenseMatrix M(2,2); M = 0.0;
+   MultABt(m_dx, m_dx, M);
+
+   DenseMatrix f(2, num_dofs2); f = 0.0;
+
+   for (int d=0; d<3; d++)
+   {
+      DenseMatrix Mtemp(2,2); Mtemp = 0.0;
+      Mtemp(0,0) = m_dx2(0,d); Mtemp(0,1) = m_dx2(1,d);
+      Mtemp(1,0) = m_dx2(1,d); Mtemp(1,1) = m_dx2(2,d);
+
+      M.Add(-gap_v[d], Mtemp);
+
+      Vector m_dxcol(2); m_dx.GetColumn(d, m_dxcol);
+      DenseMatrix ftmp(2,4);
+      outer(m_dxcol, m_N, ftmp);
+      ftmp *= -1;
+      ftmp.Add( gap_v[d], m_dN);   // 2*4
+
+      for (int j=0; j<4; j++)
+      {
+         assert(d+3*j<num_dofs2);
+         f(0,d+j*3) = ftmp(0,j);
+         f(1,d+j*3) = ftmp(1,j);
+      }
+   }
+   //fprintf('hess dxidxm\n');
+   DenseMatrixInverse Minv(M);
+   DenseMatrix dxidxm(2,num_dofs2); dxidxm = 0.0;
+   Minv.Mult(f, dxidxm);
+   //LinearSolve??
+   //dxidxm = M\f;
+
+   DenseMatrix nde2(2,2); nde2 = 0.0;
+   DenseMatrix Nndx2(2,num_dofs2); Nndx2 = 0.0;
+
+   for (int d=0; d<3; d++)
+   {
+      DenseMatrix ndetmp(2,2); ndetmp = 0.0;
+      ndetmp(0,0) = normal(d)*m_dx2(0,d); ndetmp(0,1) = normal(d)*m_dx2(1,d);
+      ndetmp(1,0) = normal(d)*m_dx2(1,d); ndetmp(1,1) = normal(d)*m_dx2(2,d);
+
+      nde2 += ndetmp;
+
+      for (int j=0; j<4; j++)
+      {
+         assert(d+3*j<num_dofs2);
+         Nndx2(0,d+j*3) = normal[d]*m_dN(0,j);
+         Nndx2(1,d+j*3) = normal[d]*m_dN(1,j);
+      }
+
+   }
+
+   DenseMatrix Ndn(2,num_dofs2); Ndn = 0.0;
+   Ndn += Nndx2;
+   AddMult(nde2, dxidxm, Ndn);
+
+
+   DenseMatrix M2(2,2); M2 = 0.0;
+   MultABt(m_dx, m_dx, M2);
+   DenseMatrixInverse M2inv(M2);
+   DenseMatrix diag2(2,2); diag2(0,0) = 1.0; diag2(1,1) = 1.0;
+   DenseMatrix m_con(2,2); m_con = 0.0;
+
+   M2inv.Mult(diag2, m_con);
+
+   DenseMatrix dg2dxm(num_dofs2, num_dofs2); dg2dxm = 0.0;
+
+   DenseMatrix dg2dxm_tmp(num_dofs2,2); dg2dxm_tmp = 0.0;
+   MultAtB(Ndn, m_con, dg2dxm_tmp);
+   Mult(dg2dxm_tmp, Ndn, dg2dxm);
+   dg2dxm *= gap;
+
+   DenseMatrix dg2dxm_tmp2(num_dofs2,num_dofs2); dg2dxm_tmp2 = 0.0;
+   MultAtB(Nndx2, dxidxm, dg2dxm_tmp2);
+   dg2dxm.Add(-1.0, dg2dxm_tmp2);
+
+   dg2dxm_tmp = 0.0;
+   MultAtB(dxidxm, nde2, dg2dxm_tmp);
+
+   AddMult_a(-1.0, dg2dxm_tmp, dxidxm, dg2dxm);
+
+   dg2dxm_tmp2 = 0.0;
+   MultAtB(dxidxm, Nndx2, dg2dxm_tmp2);
+   dg2dxm.Add(-1.0, dg2dxm_tmp2);
+
+   Vector v_dxidx2(4);
+   m_coords.Mult(gap_v, v_dxidx2); // m_coords * gap_v;  // 4*3 * 3 = 4
+
+   DenseMatrix K_dxidx2(2,2); K_dxidx2 = 0.0;
+
+   Vector m_dN2row1(4); m_dN2.GetRow(0, m_dN2row1);
+   Vector m_dN2row2(4); m_dN2.GetRow(1, m_dN2row2);
+   Vector m_dN2row3(4); m_dN2.GetRow(2, m_dN2row3);
+   K_dxidx2(0,0) = m_dN2row1 * v_dxidx2; // how would 4*1 * 1*4 be computed?
+   K_dxidx2(0,1) = m_dN2row2 * v_dxidx2;
+   K_dxidx2(1,0) = m_dN2row2 * v_dxidx2;
+   K_dxidx2(1,1) = m_dN2row3 * v_dxidx2;
+
+   DenseMatrix K_dxidx(2,2);
+   K_dxidx -= M2;
+   K_dxidx += K_dxidx2;
+
+   Vector drdxs_r(6);
+   drdxs_r[0] = m_dx(0,0); drdxs_r[1] = m_dx(0,1); drdxs_r[2] = m_dx(0,2);
+   drdxs_r[3] = m_dx(1,0); drdxs_r[4] = m_dx(1,1); drdxs_r[5] = m_dx(1,2);
+
+   DenseMatrix drdxs_K(6,6); drdxs_K = 0.;
+   for (int i=0; i<3; i++)
+   {
+      drdxs_K(i,i) = K_dxidx(0,0);
+      drdxs_K(i,3+i) = K_dxidx(0,1);
+      drdxs_K(i+3,i) = K_dxidx(1,0);
+      drdxs_K(i+3,i+3) = K_dxidx(1,1);
+   }
+   Vector dxidxs(6);
+
+   DenseMatrixInverse drdxsK_inv(drdxs_K);
+   drdxsK_inv.Mult(drdxs_r,dxidxs);
+   dxidxs *= -1.0;
+   //dxidxs = -drdxs_K\drdxs_r;
+
+   DenseMatrix dxidxs_m(2,3); dxidxs_m = 0.0;
+   dxidxs_m(0,0) = dxidxs[0]; dxidxs_m(0,1) = dxidxs[1]; dxidxs_m(0,2) = dxidxs[2];
+   dxidxs_m(1,0) = dxidxs[3]; dxidxs_m(1,1) = dxidxs[4]; dxidxs_m(1,2) = dxidxs[5];
+
+   DenseMatrix dtao1dxs(3,3); dtao1dxs = 0.0;
+   DenseMatrix dtao2dxs(3,3); dtao2dxs = 0.0;
+
+   Vector dxidxs_row1(3); dxidxs_row1 = 0.0; Vector dxidxs_row2(3);
+   dxidxs_row2 = 0.0;
+   Vector mdx2_row1(3); mdx2_row1 = 0.0; Vector mdx2_row2(3); mdx2_row2 = 0.0;
+   Vector mdx2_row3(3); mdx2_row3 = 0.0;
+   dxidxs_m.GetRow(0,dxidxs_row1);
+   dxidxs_m.GetRow(1,dxidxs_row2);
+   m_dx2.GetRow(0,mdx2_row1);
+   m_dx2.GetRow(1,mdx2_row2);
+   m_dx2.GetRow(2,mdx2_row3);
+
+   DenseMatrix dtaotmp(3,3); dtaotmp = 0.0;
+   outer(mdx2_row1, dxidxs_row1,dtaotmp);
+   dtao1dxs += dtaotmp; dtaotmp = 0.0;
+   outer(mdx2_row2, dxidxs_row1,dtaotmp);
+   dtao1dxs += dtaotmp; dtaotmp = 0.0;
+
+   outer(mdx2_row2, dxidxs_row2, dtaotmp);
+   dtao2dxs += dtaotmp; dtaotmp = 0.0;
+   outer(mdx2_row3, dxidxs_row2, dtaotmp);
+   dtao2dxs += dtaotmp; dtaotmp = 0.0;
+
+   DenseMatrix dtaodxs(3,3); dtaodxs = 0.0; //tao = tao1 cross tao2
+
+   for (int d=0; d<3; d++)
+   {
+      Vector dtao1dxs_tmp(3); dtao1dxs_tmp = 0.0;
+      dtao1dxs.GetColumn(d,dtao1dxs_tmp);
+      Vector m_dxrow(3); m_dx.GetRow(1, m_dxrow);
+
+      Vector dtaodxs_tmp(3); dtaodxs_tmp = 0.0;
+      cross(dtao1dxs_tmp, m_dxrow, dtaodxs_tmp);
+
+      Vector dtaodxs_tmp2(3); dtaodxs_tmp2 = 0.0;
+      m_dx.GetRow(0, m_dxrow);
+      dtao1dxs_tmp = 0.0;  // reuse the same vector for dtao2
+      dtao2dxs.GetColumn(d,dtao1dxs_tmp);
+      cross(m_dxrow, dtao1dxs_tmp, dtaodxs_tmp2);
+
+      dtaodxs_tmp2 += dtaodxs_tmp;
+      dtaodxs.SetCol(d, dtaodxs_tmp2);
+   }
+
+   DenseMatrix dndxs(3,3); dndxs = 0.0; dndxs += dtaodxs; dndxs *= 1.0/nnorm;
+   DenseMatrix dndxs_tmp(3,3); dndxs_tmp = 0.0;
+   outer(normal, normal, dndxs_tmp);
+   AddMult_a(-1/nnorm, dndxs_tmp, dtaodxs, dndxs);
+
+   DenseMatrix dgvdxs(3,3); dgvdxs = 0.0;
+   MultAtB(m_dx, dxidxs_m, dgvdxs);
+   dgvdxs *= -1;
+   for (int d=0; d<3; d++)
+   {
+      dgvdxs(d,d) += 1.0;
+   }
+   //dxidxs: 2*3
+
+   DenseMatrix dg2dxs(3,3); dg2dxs = 0.0;
+   DenseMatrix dg2dxs_tmp(3,2); dg2dxs_tmp = 0.0;
+   MultAtB(dxidxs_m, nde2, dg2dxs_tmp);
+   AddMult_a(-1.0, dg2dxs_tmp, dxidxs_m, dg2dxs);
+   DenseMatrix dg2dxs_tmp2(3,3); dg2dxs_tmp2 = 0.0;
+   MultAtB(dgvdxs, dndxs, dg2dxs_tmp2);
+   dg2dxs += dg2dxs_tmp2;
+   dg2dxs_tmp2 = 0.0;
+   MultAtB(dndxs, dndxs_tmp, dg2dxs_tmp2);
+   AddMult(dg2dxs_tmp2, dgvdxs, dg2dxs);
+
+   DenseMatrix Ne(3,12), Be(6,12), dBe(12,12);
+   BasisVectorDerivs(xi, Ne, Be, dBe);
+
+   DenseMatrix dtao1dxm(3,12); dtao1dxm.CopyRows(Be, 0, 2);
+   DenseMatrix dtao2dxm(3,12); dtao2dxm.CopyRows(Be, 3, 5);
+
+   Vector m_coords_v(12);
+   for (int i=0; i<4; i++)
+   {
+      for (int j=0; j<3; j++)
+      {
+         m_coords_v[i*3+j] = m_coords(i,j);
+      }
+   }
+
+   for (int i=0; i<2; i++)
+   {
+      Vector dxidxm_tmp(num_dofs2); dxidxm_tmp = 0.0;
+      dxidxm.GetRow(i,dxidxm_tmp);
+
+      DenseMatrix dBe_tmp(3,12);
+      dBe_tmp.CopyRows(dBe,i*3,(i+1)*3-1);
+
+      DenseMatrix dtaodxm_tmp(12,12); dtaodxm_tmp = 0.0;
+      outer(m_coords_v, dxidxm_tmp, dtaodxm_tmp);
+      AddMult(dBe_tmp, dtaodxm_tmp, dtao1dxm);
+
+      //dtao1dxm += dBe(:,:,i)*reshape(m_coords(1:4,:)',12,1)*reshape(dxidxm(i,:),1,12); % 3*12
+      dBe_tmp = 0.0;
+      dBe_tmp.CopyRows(dBe,(i+2)*3,(i+3)*3-1);
+      AddMult(dBe_tmp, dtaodxm_tmp, dtao2dxm);
+
+   }
+
+   DenseMatrix dtaodxm(3,12); dtaodxm = 0.0;//tao = tao1 cross tao2
+
+   for (int d=0; d<12; d++)
+   {
+      Vector dtaodxm_tmp(3); dtaodxm_tmp = 0.0;
+      Vector dtaodxm_tmp2(3); dtaodxm_tmp2 = 0.0;
+      Vector tmp1(3); tmp1 = 0.0;  dtao1dxm.GetColumn(d,tmp1);
+      Vector m_dxrow2(3); m_dx.GetRow(1, m_dxrow2);
+      Vector m_dxrow1(3); m_dx.GetRow(0, m_dxrow1);
+      Vector tmp2(3); tmp2 = 0.0;  dtao2dxm.GetColumn(d,tmp2);
+
+      cross(tmp1, m_dxrow2, dtaodxm_tmp);
+      cross(m_dxrow1,tmp2, dtaodxm_tmp2);
+      dtaodxm_tmp += dtaodxm_tmp2;
+
+      dtaodxm.SetCol(d, dtaodxm_tmp);
+   }
+
+   DenseMatrix dndxm(3,12); dndxm = 0.0;
+   dndxm += dtaodxm;
+   dndxm *= 1.0/nnorm;
+   AddMult_a(-1/nnorm, dndxs_tmp, dtaodxm, dndxm); //dndxs_tmp = normal'*normal
+
+   DenseMatrix dgvdxm(3,12); dgvdxm = 0.0;
+   dgvdxm -= Ne;
+
+   for (int i=0; i<2; i++)
+   {
+      Vector dxidxm_tmp(num_dofs2); dxidxm_tmp = 0.0;
+      dxidxm.GetRow(i,dxidxm_tmp);
+
+      DenseMatrix Be_tmp(3,12);
+      Be_tmp.CopyRows(Be,i*3,(i+1)*3-1);
+
+      DenseMatrix dgvdxm_tmp(12,12); dgvdxm_tmp = 0.0;
+      outer(m_coords_v, dxidxm_tmp, dgvdxm_tmp);
+      AddMult_a(-1.0, Be_tmp, dgvdxm_tmp, dgvdxm);
+
+   }
+
+   DenseMatrix dg2dxsxm(3,12); dg2dxsxm = 0.0;
+   DenseMatrix dg2dxsxm_tmp(3,3); dg2dxsxm_tmp = 0.0;
+   MultAtB(dgvdxs, dndxm, dg2dxsxm);
+
+   MultAtB(dndxs, dndxs_tmp, dg2dxsxm_tmp);
+   AddMult(dg2dxsxm_tmp, dgvdxm, dg2dxsxm); // += dndxs'*normal'*normal*dgvdxm;
+
+   DenseMatrix dgvdxsxmn(3,12); dgvdxsxmn = 0.0;
+   DenseMatrix dgvdxsxmn_tmp(3,2); dgvdxsxmn_tmp = 0.0;
+   MultAtB(dxidxs_m, nde2, dgvdxsxmn_tmp);   //dxidxs_m: 2*3
+
+   AddMult_a(-1.0, dgvdxsxmn_tmp, dxidxm, dgvdxsxmn);
+
+
+   for (int i =0; i<2; i++)
+   {
+      DenseMatrix Be_tmp(3,12);
+      Be_tmp.CopyRows(Be,i*3,(i+1)*3-1);
+
+      Vector dxidxs_row(3); dxidxs_row = 0.0; dxidxs_m.GetRow(i,dxidxs_row);
+      DenseMatrix dgvdxsxmn_tmp2(3,3); dgvdxsxmn_tmp2 = 0.0;
+      outer(dxidxs_row, normal, dgvdxsxmn_tmp2);
+      AddMult_a(-1.0, dgvdxsxmn_tmp2, Be_tmp, dgvdxsxmn);
+   }
+
+   dg2dxsxm += dgvdxsxmn;
+
+   DenseMatrix dg2dxmxs(12,3); dg2dxmxs = 0.0;
+   DenseMatrix dg2dxmxs_tmp(12,3); dg2dxmxs_tmp = 0.0;
+   MultAtB(dgvdxm, dndxs, dg2dxmxs);
+   MultAtB(dndxm, dndxs_tmp, dg2dxmxs_tmp);
+   AddMult(dg2dxmxs_tmp, dgvdxs, dg2dxmxs);
+
+   DenseMatrix dgvdxmxsn(12,3); dgvdxmxsn = 0.0;
+   DenseMatrix dgvdxmxsn_tmp(12,2); dgvdxmxsn_tmp = 0.0;
+
+   MultAtB(dxidxm, nde2, dgvdxmxsn_tmp);
+   dgvdxmxsn_tmp *= -1.0;
+   AddMult(dgvdxmxsn_tmp, dxidxs_m, dgvdxmxsn);
+
+   for (int i =0; i<2; i++)
+   {
+      DenseMatrix Be_tmp(3,12);
+      Be_tmp.CopyRows(Be,i*3,(i+1)*3-1);
+      Be_tmp.Transpose(); // Be is now 12*3
+
+      Vector dxidxs_row(3); dxidxs_row = 0.0; dxidxs_m.GetRow(i,dxidxs_row);
+      DenseMatrix dgvdxmxsn_tmp2(3,3); dgvdxmxsn_tmp2 = 0.0;
+      outer(normal, dxidxs_row, dgvdxmxsn_tmp2);
+      AddMult_a(-1.0, Be_tmp, dgvdxmxsn_tmp2, dgvdxmxsn);
+
+   }
+
+   dg2dxmxs += dgvdxmxsn;
+
+   dg2dx.CopyMN(dg2dxs, 0, 0);
+   dg2dx.CopyMN(dg2dxm, 3, 3);
+   dg2dx.CopyMN(dg2dxsxm, 0, 3);
+   dg2dx.CopyMN(dg2dxmxs, 3, 0);
+
+};
+
+
+
+void NodeSegConPairs(const Vector x1, const Vector xi2,
+                     const DenseMatrix coords2,
+                     double& node_g, Vector& node_dg, DenseMatrix& node_dg2)
+{
+   double gap = 0.0;
+   Vector normal(3); normal = 0.0;
+   Vector dgdxm(12); dgdxm = 0.0;
+   Vector dgdxs(3); dgdxs = 0.0;
+
+   ComputeGapJacobian(x1, xi2, coords2, gap, normal, dgdxm, dgdxs);
+   node_g = gap;
+
+   node_dg.SetSize(12+3);
+   for (int i=0; i<3; i++) { node_dg[i] = dgdxs[i]; }
+   for (int i=0; i<12; i++) { node_dg[i+3] = dgdxm[i]; }
+
+   DenseMatrix dg2dx(15,15); dg2dx = 0.0;
+   DenseMatrix dgvdxmxsn(12,3); dgvdxmxsn = 0.0;
+   ComputeGapHessian(x1, xi2, coords2, dg2dx);
+
+   node_dg2.SetSize(15,15);
+   node_dg2 = dg2dx;
+
+   /*
+   if(obj.space1.conns{e1}(i)==150)  % for debugging purpose
+
+   v1 = 1:3;
+   v2 = 1:12;
+   %v1 = ones(1,3)
+   %v2 = ones(1,12)
+   v2 = reshape(v2,4,3);
+   x1n1 = x1 + 0.01*v1;
+   coords2n1 = coords2 + 0.001*v2;
+      [xi2n1, gapv1, ~, ~] = SlaveToMaster(obj, coords2n1, x1n1);
+      [gapn1, n1,dgdxmn1, dgdxsn1] = ComputeGapJacobian(obj, x1n1, xi2n1, coords2n1);
+   x1n2 = x1 - 0.01*v1;
+   coords2n2 = coords2 - 0.001*v2;
+      [xi2n2, gapv2, ~, ~] = SlaveToMaster(obj, coords2n2, x1n2);
+      [gapn2, n2,dgdxmn2, dgdxsn2] = ComputeGapJacobian(obj, x1n2, xi2n2, coords2n2);
+   fprintf('fd\n');
+      %gapv1-gapv2
+   [dgdxsn1(:)',dgdxmn1(:)'] - [dgdxsn2(:)',dgdxmn2(:)']
+
+   %dgdxsn1-dgdxsn2
+   fprintf('code\n');
+   v2n = v2';
+   %dg2dx(1:3,1:3)*0.04*ones(3,1)
+   temp = zeros(12,3);
+   for i = 1:4
+   temp1 = dg2dx(3+(i-1)*3+1:3+i*3,1:3);
+   temp((i-1)*3+1:i*3,:) = temp1';
+   end
+   temp2 = zeros(3,12);
+   for i = 1:4
+   temp3 = dg2dx(1:3,3+(i-1)*3+1:3+i*3);
+   temp2(:,(i-1)*3+1:i*3) = temp3';
+   end
+   %dg2dx
+   %dg2dx(4:end,1:3) = temp;
+   %dg2dx(1:3,4:end) = temp2;
+   %dgvdxm * 0.002*v2n(:)
+   (dg2dx*[0.02*v1(:)',0.002*v2n(:)']')'
+   %dg2dx(4:end,1:3)
+   end*/
+
+};
+
+
+// coordsm : (npoints*4, 3) use what class?
+// m_conn: (npoints*4)
+void Assemble_Contact(const int m, const int npoints, const int ndofs,
+                      const Vector x_s,
+                      const Vector xi, const DenseMatrix coordsm, const Array<int> s_conn,
+                      const Array<int> m_conn, Vector& g, SparseMatrix& M,
+                      std::vector<SparseMatrix>& dM)
+{
+   int ndim = 3;
+
+   g.SetSize(m);
+   g = 0.0;
+
+   //SparseMatrix M(m, n);  // M needs to be the correct size
+
+   //dM.resize(m); // needs to clear?
+
+   double g_tmp = 0.;
+   Vector dg(4*ndim+ndim);
+   dg = 0.;
+   DenseMatrix dg2(4*ndim+ndim,4*ndim+ndim);
+   dg2 = 0.;
+
+   for (int i=0; i<npoints; i++)
+   {
+      Vector x1(ndim);
+      x1[0] = x_s[i*ndim];
+      x1[1] = x_s[i*ndim+1];
+      x1[2] = x_s[i*ndim+2];
+
+      Vector xi2(ndim-1);
+      xi2[0] = xi[i*(ndim-1)];
+      xi2[1] = xi[i*(ndim-1)+1];
+
+      DenseMatrix coords2(4,3);
+      coords2.CopyRows(coordsm, i*4,(i+1)*4-1);
+
+      //how to get coords2?
+      dg = 0.0;
+      dg2 = 0.;
+      NodeSegConPairs(x1, xi2, coords2, g_tmp, dg, dg2);
+      //x1.Print();
+      //xi2.Print();
+      //coords2.Print();
+      g[s_conn[i]] = g_tmp; // should be unique
+      Array<int> m_conn_i(4);
+      m_conn.GetSubArray(4*i, 4, m_conn_i);
+
+      Array<int> node_conn(5);
+      node_conn[0] = s_conn[i];
+      for (int j=0; j<4; j++)
+      {
+         node_conn[j+1] = m_conn_i[j];
+      }
+
+      Array<int>  M_i_tmp(1);
+      M_i_tmp[0] = s_conn[i];
+
+      //j_idx = (node_conn-1)*obj.disp_field.num_components +repmat((1:obj.disp_field.num_components)', 1, length(node_conn{i}));
+      Array<int> j_idx(5*ndim); j_idx = 0;
+      for (int j=0; j< 5; j++)
+      {
+         for (int k=0; k<ndim; k++)
+         {
+            j_idx[j*ndim+k] = node_conn[j]*ndim+k;
+         }
+      }
+      DenseMatrix M_v_tmp(1, ndim*(4+1));   // SetData now?
+      M_v_tmp.SetRow(0, dg);
+
+      M.AddSubMatrix(M_i_tmp, j_idx, M_v_tmp);
+
+      Array<int> dM_i(ndim*(4+1));
+      Array<int> dM_j(ndim*(4+1));
+
+      for (int j=0; j< ndim*(4+1); j++)
+      {
+         dM_i[j] = j_idx[j];
+         dM_j[j] = j_idx[j];
+      }
+      //dg2.Print();
+      //dM[s_conn[i]].Print();
+      dM[s_conn[i]].AddSubMatrix(dM_i,dM_j, dg2);
+   }
+};
+

examples/contact/obstacleProblem.cpp

@@ -0,0 +1,119 @@
+#include "mfem.hpp"
+#include "Problems.hpp"
+#include "IPsolver.hpp"
+#include <fstream>
+#include <iostream>
+
+using namespace std;
+using namespace mfem;
+
+
+
+double dmanufacturedFun(const Vector &);
+double fRhs(const Vector &);
+double obstacle(const Vector &);
+
+int main(int argc, char *argv[])
+{
+  int FEorder = 1; // order of the finite elements
+  int linSolver = 0;
+  int maxIPMiters = 30;
+  bool iAmRoot = true;
+  
+  OptionsParser args(argc, argv);
+  args.AddOption(&FEorder, "-o", "--order",\
+		  "Order of the finite elements.");
+  args.AddOption(&linSolver, "-linSolver", "--linearSolver", \
+       "IP-Newton linear system solution strategy.");
+  args.AddOption(&maxIPMiters, "-IPMiters", "--IPMiters",\
+		  "Maximum number of IPM iterations");
+  
+  args.Parse();
+  if(!args.Good())
+  {
+    args.PrintUsage(cout);
+    return 1;
+  }
+  else
+  {
+    if( iAmRoot )
+    {
+      args.PrintOptions(cout);
+    }
+  }
+
+  const char *meshFile = "../../data/inline-quad.mesh";
+  Mesh *mesh = new Mesh(meshFile, 1, 1);
+  int dim = mesh->Dimension(); // geometric dimension of the domain
+  {
+     int ref_levels = 3;
+     for (int l = 0; l < ref_levels; l++)
+     {
+        mesh->UniformRefinement();
+     }
+  }
+
+  FiniteElementCollection *fec = new H1_FECollection(FEorder, dim);
+  FiniteElementSpace      *Vh  = new FiniteElementSpace(mesh, fec);
+  ObstacleProblem problem(Vh, &fRhs, &obstacle);
+  
+  int dimD = problem.GetDimD();
+  Vector x0(dimD); x0 = 0.0;
+  Vector xf(dimD); xf = 0.0;
+
+  InteriorPointSolver optimizer(&problem); 
+  optimizer.SetTol(1.e-7);
+  optimizer.SetLinearSolver(linSolver);
+  optimizer.SetMaxIter(maxIPMiters);
+  optimizer.Mult(x0, xf);
+
+  GridFunction d_gf(Vh);
+
+  d_gf = xf;
+
+  FunctionCoefficient dm_fc(dmanufacturedFun); // pseudo-manufactured solution
+  GridFunction dm_gf(Vh);
+  dm_gf.ProjectCoefficient(dm_fc);
+  
+  ParaViewDataCollection paraview_dc("BarrierProblemSolution", mesh);
+  paraview_dc.SetPrefixPath("ParaView");
+  paraview_dc.SetLevelsOfDetail(FEorder);
+  paraview_dc.SetDataFormat(VTKFormat::BINARY);
+  paraview_dc.SetHighOrderOutput(true);
+  paraview_dc.SetCycle(0);
+  paraview_dc.SetTime(0.0);
+  paraview_dc.RegisterField("d(x) (numerical)", &d_gf);
+  paraview_dc.RegisterField("d(x) (pseudo-manufactured)", &dm_gf);
+  paraview_dc.Save();
+  
+  
+  delete Vh;
+  delete fec;
+  delete mesh;
+  return 0;
+}
+
+
+double dmanufacturedFun(const Vector &x)
+{
+  return cos(2*M_PI*x(0)) + 0.2 - 2.0*(pow(x(0),3) - 1.5*pow(x(0),2));
+}
+
+
+// f(x) forcing term... which enters the objective energy functional
+// E(d) = 0.5 d^T K d - f^T d, where f is a discrete vector representation
+// of f(x). f(x) is such that in the absence of bound-constraints then
+// the solution of the optimization problem satisfies the PDE
+// -div(grad(d)) + d = f + homogeneous Neumann conditions on the unit interval,
+// for d(x) = cos(2 \pi x) + a0 + a3 (x^3 - 1.5 x^2), a2 = 0.2, a3 = -2
+double fRhs(const Vector &x)
+{
+  double fx = 0.;
+  fx = 0.2 - 2.0 * (pow(x(0),3)- 1.5*pow(x(0),2.) - 6 * x(0) + 3.) + (1. + pow(2.*M_PI,2))*cos(2.*M_PI*x(0));
+  return fx;
+}
+
+double obstacle(const Vector &x)
+{
+  return 0.0;
+}

examples/contact/rotatedblock2.mesh

@@ -0,0 +1,70 @@
+
+MFEM mesh v1.0
+
+#
+# MFEM Geometry Types (see mesh/geom.hpp):
+#
+# POINT       = 0
+# SEGMENT     = 1
+# TRIANGLE    = 2
+# SQUARE      = 3
+# TETRAHEDRON = 4
+# CUBE        = 5
+# PRISM       = 6
+#
+
+dimension
+3
+
+# 1 nothing
+elements
+4
+1   5     0     1     3     2     6     7    9     8
+1   5     2     3     5     4     8     9    11    10
+1   5     6     7     9     8    12    13    15    14
+1   5     8     9    11    10    14    15    17    16
+
+# 0 nothing
+# 1 dirichlet bc
+# 2 contact
+boundary
+16
+1   3     1     0     2     3
+1   3     3     2     4     5
+1   3    12    13    15    14
+1   3    14    15    17    16
+3   3     2     0     6     8
+3   3     4     2     8    10
+3   3     8     6    12    14
+3   3    10     8    14    16
+2   3     1     3     9     7
+2   3     3     5    11    9
+2   3     7    9    15    13
+2   3     9    11    17    15
+1   3     6     0     1     7
+1   3    12     6     7    13
+1   3     4    10    11     5
+1   3    10    16    17    11
+
+vertices
+18
+3
+
+0.000000000000 0.145770950245 0.443895630208
+0.507100000000 0.145770950245 0.443895630208
+0.000000000000 0.350937660019 0.294833290227
+0.507100000000 0.350937660019 0.294833290227
+0.000000000000 0.556104369792 0.145770950245
+0.507100000000 0.556104369792 0.145770950245
+0.000000000000 0.294833290227 0.649062339981
+0.507100000000 0.294833290227 0.649062339981
+0.000000000000 0.500000000000 0.500000000000
+0.507100000000 0.500000000000 0.500000000000
+0.000000000000 0.705166709773 0.350937660019
+0.507100000000 0.705166709773 0.350937660019
+0.000000000000 0.443895630208 0.854229049755
+0.507100000000 0.443895630208 0.854229049755
+0.000000000000 0.649062339981 0.705166709773
+0.507100000000 0.649062339981 0.705166709773
+0.000000000000 0.854229049755 0.556104369792
+0.507100000000 0.854229049755 0.556104369792

examples/ex25.cpp

@@ -32,7 +32,6 @@
 //               We recommend viewing Example 22 before viewing this example.
 
 #include "mfem.hpp"
-#include <memory>
 #include <fstream>
 #include <iostream>
 
@@ -45,7 +44,7 @@ using namespace std;
 using namespace mfem;
 
 // Class for setting up a simple Cartesian PML region
-class PML
+class CartesianPML
 {
 private:
    Mesh *mesh;
@@ -70,7 +69,7 @@ private:
 
 public:
    // Constructor
-   PML(Mesh *mesh_,Array2D<double> length_);
+   CartesianPML(Mesh *mesh_,Array2D<double> length_);
 
    // Return Computational Domain Boundary
    Array2D<double> GetCompDomainBdr() {return comp_dom_bdr;}
@@ -92,12 +91,12 @@ public:
 class PMLDiagMatrixCoefficient : public VectorCoefficient
 {
 private:
-   PML * pml = nullptr;
-   void (*Function)(const Vector &, PML *, Vector &);
+   CartesianPML * pml = nullptr;
+   void (*Function)(const Vector &, CartesianPML *, Vector &);
 public:
-   PMLDiagMatrixCoefficient(int dim, void(*F)(const Vector &, PML *,
+   PMLDiagMatrixCoefficient(int dim, void(*F)(const Vector &, CartesianPML *,
                                               Vector &),
-                            PML * pml_)
+                            CartesianPML * pml_)
       : VectorCoefficient(dim), pml(pml_), Function(F)
    {}
 
@@ -126,13 +125,13 @@ void source(const Vector &x, Vector & f);
 
 // Functions for computing the necessary coefficients after PML stretching.
 // J is the Jacobian matrix of the stretching function
-void detJ_JT_J_inv_Re(const Vector &x, PML * pml, Vector &D);
-void detJ_JT_J_inv_Im(const Vector &x, PML * pml, Vector &D);
-void detJ_JT_J_inv_abs(const Vector &x, PML * pml, Vector &D);
+void detJ_JT_J_inv_Re(const Vector &x, CartesianPML * pml, Vector &D);
+void detJ_JT_J_inv_Im(const Vector &x, CartesianPML * pml, Vector &D);
+void detJ_JT_J_inv_abs(const Vector &x, CartesianPML * pml, Vector &D);
 
-void detJ_inv_JT_J_Re(const Vector &x, PML * pml, Vector &D);
-void detJ_inv_JT_J_Im(const Vector &x, PML * pml, Vector &D);
-void detJ_inv_JT_J_abs(const Vector &x, PML * pml, Vector &D);
+void detJ_inv_JT_J_Re(const Vector &x, CartesianPML * pml, Vector &D);
+void detJ_inv_JT_J_Im(const Vector &x, CartesianPML * pml, Vector &D);
+void detJ_inv_JT_J_abs(const Vector &x, CartesianPML * pml, Vector &D);
 
 Array2D<double> comp_domain_bdr;
 Array2D<double> domain_bdr;
@@ -268,7 +267,7 @@ int main(int argc, char *argv[])
          length = 0.25;
          break;
    }
-   PML * pml = new PML(mesh,length);
+   CartesianPML * pml = new CartesianPML(mesh,length);
    comp_domain_bdr = pml->GetCompDomainBdr();
    domain_bdr = pml->GetDomainBdr();
 
@@ -468,14 +467,16 @@ int main(int argc, char *argv[])
       offsets[2] = fespace->GetTrueVSize();
       offsets.PartialSum();
 
-      std::unique_ptr<Operator> pc_r;
-      std::unique_ptr<Operator> pc_i;
+      Operator *pc_r = nullptr;
+      Operator *pc_i = nullptr;
       double s = (conv == ComplexOperator::HERMITIAN) ? -1.0 : 1.0;
       if (pa)
       {
          // Jacobi Smoother
-         pc_r.reset(new OperatorJacobiSmoother(prec, ess_tdof_list));
-         pc_i.reset(new ScaledOperator(pc_r.get(), s));
+         OperatorJacobiSmoother *d00 = new OperatorJacobiSmoother(prec, ess_tdof_list);
+         ScaledOperator *d11 = new ScaledOperator(d00, s);
+         pc_r = d00;
+         pc_i = d11;
       }
       else
       {
@@ -484,13 +485,15 @@ int main(int argc, char *argv[])
          prec.FormSystemMatrix(ess_tdof_list, PCOpAh);
 
          // Gauss-Seidel Smoother
-         pc_r.reset(new GSSmoother(*PCOpAh.As<SparseMatrix>()));
-         pc_i.reset(new ScaledOperator(pc_r.get(), s));
+         GSSmoother *gs00 = new GSSmoother(*PCOpAh.As<SparseMatrix>());
+         ScaledOperator *gs11 = new ScaledOperator(gs00, s);
+         pc_r = gs00;
+         pc_i = gs11;
       }
 
       BlockDiagonalPreconditioner BlockDP(offsets);
-      BlockDP.SetDiagonalBlock(0, pc_r.get());
-      BlockDP.SetDiagonalBlock(1, pc_i.get());
+      BlockDP.SetDiagonalBlock(0, pc_r);
+      BlockDP.SetDiagonalBlock(1, pc_i);
 
       GMRESSolver gmres;
       gmres.SetPrintLevel(1);
@@ -804,7 +807,7 @@ void E_bdr_data_Im(const Vector &x, Vector &E)
    }
 }
 
-void detJ_JT_J_inv_Re(const Vector &x, PML * pml, Vector &D)
+void detJ_JT_J_inv_Re(const Vector &x, CartesianPML * pml, Vector &D)
 {
    vector<complex<double>> dxs(dim);
    complex<double> det(1.0, 0.0);
@@ -821,7 +824,7 @@ void detJ_JT_J_inv_Re(const Vector &x, PML * pml, Vector &D)
    }
 }
 
-void detJ_JT_J_inv_Im(const Vector &x, PML * pml, Vector &D)
+void detJ_JT_J_inv_Im(const Vector &x, CartesianPML * pml, Vector &D)
 {
    vector<complex<double>> dxs(dim);
    complex<double> det = 1.0;
@@ -838,7 +841,7 @@ void detJ_JT_J_inv_Im(const Vector &x, PML * pml, Vector &D)
    }
 }
 
-void detJ_JT_J_inv_abs(const Vector &x, PML * pml, Vector &D)
+void detJ_JT_J_inv_abs(const Vector &x, CartesianPML * pml, Vector &D)
 {
    vector<complex<double>> dxs(dim);
    complex<double> det = 1.0;
@@ -855,7 +858,7 @@ void detJ_JT_J_inv_abs(const Vector &x, PML * pml, Vector &D)
    }
 }
 
-void detJ_inv_JT_J_Re(const Vector &x, PML * pml, Vector &D)
+void detJ_inv_JT_J_Re(const Vector &x, CartesianPML * pml, Vector &D)
 {
    vector<complex<double>> dxs(dim);
    complex<double> det(1.0, 0.0);
@@ -880,7 +883,7 @@ void detJ_inv_JT_J_Re(const Vector &x, PML * pml, Vector &D)
    }
 }
 
-void detJ_inv_JT_J_Im(const Vector &x, PML * pml, Vector &D)
+void detJ_inv_JT_J_Im(const Vector &x, CartesianPML * pml, Vector &D)
 {
    vector<complex<double>> dxs(dim);
    complex<double> det = 1.0;
@@ -904,7 +907,7 @@ void detJ_inv_JT_J_Im(const Vector &x, PML * pml, Vector &D)
    }
 }
 
-void detJ_inv_JT_J_abs(const Vector &x, PML * pml, Vector &D)
+void detJ_inv_JT_J_abs(const Vector &x, CartesianPML * pml, Vector &D)
 {
    vector<complex<double>> dxs(dim);
    complex<double> det = 1.0;
@@ -928,14 +931,14 @@ void detJ_inv_JT_J_abs(const Vector &x, PML * pml, Vector &D)
    }
 }
 
-PML::PML(Mesh *mesh_, Array2D<double> length_)
+CartesianPML::CartesianPML(Mesh *mesh_, Array2D<double> length_)
    : mesh(mesh_), length(length_)
 {
    dim = mesh->Dimension();
    SetBoundaries();
 }
 
-void PML::SetBoundaries()
+void CartesianPML::SetBoundaries()
 {
    comp_dom_bdr.SetSize(dim, 2);
    dom_bdr.SetSize(dim, 2);
@@ -950,7 +953,7 @@ void PML::SetBoundaries()
    }
 }
 
-void PML::SetAttributes(Mesh *mesh_)
+void CartesianPML::SetAttributes(Mesh *mesh_)
 {
    // Initialize bdr attributes
    for (int i = 0; i < mesh_->GetNBE(); ++i)
@@ -999,8 +1002,8 @@ void PML::SetAttributes(Mesh *mesh_)
    mesh_->SetAttributes();
 }
 
-void PML::StretchFunction(const Vector &x,
-                          vector<complex<double>> &dxs)
+void CartesianPML::StretchFunction(const Vector &x,
+                                   vector<complex<double>> &dxs)
 {
    complex<double> zi = complex<double>(0., 1.);
 

examples/ex25p.cpp

@@ -44,7 +44,7 @@ using namespace std;
 using namespace mfem;
 
 // Class for setting up a simple Cartesian PML region
-class PML
+class CartesianPML
 {
 private:
    Mesh *mesh;
@@ -69,7 +69,7 @@ private:
 
 public:
    // Constructor
-   PML(Mesh *mesh_,Array2D<double> length_);
+   CartesianPML(Mesh *mesh_,Array2D<double> length_);
 
    // Return Computational Domain Boundary
    Array2D<double> GetCompDomainBdr() {return comp_dom_bdr;}
@@ -91,12 +91,12 @@ public:
 class PMLDiagMatrixCoefficient : public VectorCoefficient
 {
 private:
-   PML * pml = nullptr;
-   void (*Function)(const Vector &, PML *, Vector &);
+   CartesianPML * pml = nullptr;
+   void (*Function)(const Vector &, CartesianPML *, Vector &);
 public:
-   PMLDiagMatrixCoefficient(int dim, void(*F)(const Vector &, PML *,
+   PMLDiagMatrixCoefficient(int dim, void(*F)(const Vector &, CartesianPML *,
                                               Vector &),
-                            PML * pml_)
+                            CartesianPML * pml_)
       : VectorCoefficient(dim), pml(pml_), Function(F)
    {}
 
@@ -125,13 +125,13 @@ void source(const Vector &x, Vector & f);
 
 // Functions for computing the necessary coefficients after PML stretching.
 // J is the Jacobian matrix of the stretching function
-void detJ_JT_J_inv_Re(const Vector &x, PML * pml, Vector & D);
-void detJ_JT_J_inv_Im(const Vector &x, PML * pml, Vector & D);
-void detJ_JT_J_inv_abs(const Vector &x, PML * pml, Vector & D);
+void detJ_JT_J_inv_Re(const Vector &x, CartesianPML * pml, Vector & D);
+void detJ_JT_J_inv_Im(const Vector &x, CartesianPML * pml, Vector & D);
+void detJ_JT_J_inv_abs(const Vector &x, CartesianPML * pml, Vector & D);
 
-void detJ_inv_JT_J_Re(const Vector &x, PML * pml, Vector & D);
-void detJ_inv_JT_J_Im(const Vector &x, PML * pml, Vector & D);
-void detJ_inv_JT_J_abs(const Vector &x, PML * pml, Vector & D);
+void detJ_inv_JT_J_Re(const Vector &x, CartesianPML * pml, Vector & D);
+void detJ_inv_JT_J_Im(const Vector &x, CartesianPML * pml, Vector & D);
+void detJ_inv_JT_J_abs(const Vector &x, CartesianPML * pml, Vector & D);
 
 Array2D<double> comp_domain_bdr;
 Array2D<double> domain_bdr;
@@ -295,7 +295,7 @@ int main(int argc, char *argv[])
          length = 0.25;
          break;
    }
-   PML * pml = new PML(mesh,length);
+   CartesianPML * pml = new CartesianPML(mesh,length);
    comp_domain_bdr = pml->GetCompDomainBdr();
    domain_bdr = pml->GetDomainBdr();
 
@@ -478,11 +478,11 @@ int main(int argc, char *argv[])
    if (!pa && mumps_solver)
    {
       HypreParMatrix *A = Ah.As<ComplexHypreParMatrix>()->GetSystemMatrix();
-      MUMPSSolver mumps(A->GetComm());
+      MUMPSSolver mumps;
       mumps.SetPrintLevel(0);
       mumps.SetMatrixSymType(MUMPSSolver::MatType::UNSYMMETRIC);
       mumps.SetOperator(*A);
-      mumps.Mult(B, X);
+      mumps.Mult(B,X);
       delete A;
    }
 #endif
@@ -524,14 +524,16 @@ int main(int argc, char *argv[])
       offsets[2] = fespace->GetTrueVSize();
       offsets.PartialSum();
 
-      std::unique_ptr<Operator> pc_r;
-      std::unique_ptr<Operator> pc_i;
+      Operator *pc_r = nullptr;
+      Operator *pc_i = nullptr;
       int s = (conv == ComplexOperator::HERMITIAN) ? -1.0 : 1.0;
       if (pa)
       {
          // Jacobi Smoother
-         pc_r.reset(new OperatorJacobiSmoother(prec, ess_tdof_list));
-         pc_i.reset(new ScaledOperator(pc_r.get(), s));
+         OperatorJacobiSmoother *d00 = new OperatorJacobiSmoother(prec, ess_tdof_list);
+         ScaledOperator *d11 = new ScaledOperator(d00, s);
+         pc_r = d00;
+         pc_i = d11;
       }
       else
       {
@@ -539,13 +541,15 @@ int main(int argc, char *argv[])
          prec.FormSystemMatrix(ess_tdof_list, PCOpAh);
 
          // Hypre AMS
-         pc_r.reset(new HypreAMS(*PCOpAh.As<HypreParMatrix>(), fespace));
-         pc_i.reset(new ScaledOperator(pc_r.get(), s));
+         HypreAMS *ams00 = new HypreAMS(*PCOpAh.As<HypreParMatrix>(), fespace);
+         ScaledOperator *ams11 = new ScaledOperator(ams00, s);
+         pc_r = ams00;
+         pc_i = ams11;
       }
 
       BlockDiagonalPreconditioner BlockDP(offsets);
-      BlockDP.SetDiagonalBlock(0, pc_r.get());
-      BlockDP.SetDiagonalBlock(1, pc_i.get());
+      BlockDP.SetDiagonalBlock(0, pc_r);
+      BlockDP.SetDiagonalBlock(1, pc_i);
 
       GMRESSolver gmres(MPI_COMM_WORLD);
       gmres.SetPrintLevel(1);
@@ -880,7 +884,7 @@ void E_bdr_data_Im(const Vector &x, Vector &E)
    }
 }
 
-void detJ_JT_J_inv_Re(const Vector &x, PML * pml, Vector & D)
+void detJ_JT_J_inv_Re(const Vector &x, CartesianPML * pml, Vector & D)
 {
    vector<complex<double>> dxs(dim);
    complex<double> det(1.0, 0.0);
@@ -897,7 +901,7 @@ void detJ_JT_J_inv_Re(const Vector &x, PML * pml, Vector & D)
    }
 }
 
-void detJ_JT_J_inv_Im(const Vector &x, PML * pml, Vector & D)
+void detJ_JT_J_inv_Im(const Vector &x, CartesianPML * pml, Vector & D)
 {
    vector<complex<double>> dxs(dim);
    complex<double> det = 1.0;
@@ -914,7 +918,7 @@ void detJ_JT_J_inv_Im(const Vector &x, PML * pml, Vector & D)
    }
 }
 
-void detJ_JT_J_inv_abs(const Vector &x, PML * pml, Vector & D)
+void detJ_JT_J_inv_abs(const Vector &x, CartesianPML * pml, Vector & D)
 {
    vector<complex<double>> dxs(dim);
    complex<double> det = 1.0;
@@ -931,7 +935,7 @@ void detJ_JT_J_inv_abs(const Vector &x, PML * pml, Vector & D)
    }
 }
 
-void detJ_inv_JT_J_Re(const Vector &x, PML * pml, Vector & D)
+void detJ_inv_JT_J_Re(const Vector &x, CartesianPML * pml, Vector & D)
 {
    vector<complex<double>> dxs(dim);
    complex<double> det(1.0, 0.0);
@@ -956,7 +960,7 @@ void detJ_inv_JT_J_Re(const Vector &x, PML * pml, Vector & D)
    }
 }
 
-void detJ_inv_JT_J_Im(const Vector &x, PML * pml, Vector & D)
+void detJ_inv_JT_J_Im(const Vector &x, CartesianPML * pml, Vector & D)
 {
    vector<complex<double>> dxs(dim);
    complex<double> det = 1.0;
@@ -980,7 +984,7 @@ void detJ_inv_JT_J_Im(const Vector &x, PML * pml, Vector & D)
    }
 }
 
-void detJ_inv_JT_J_abs(const Vector &x, PML * pml, Vector & D)
+void detJ_inv_JT_J_abs(const Vector &x, CartesianPML * pml, Vector & D)
 {
    vector<complex<double>> dxs(dim);
    complex<double> det = 1.0;
@@ -1004,14 +1008,14 @@ void detJ_inv_JT_J_abs(const Vector &x, PML * pml, Vector & D)
    }
 }
 
-PML::PML(Mesh *mesh_, Array2D<double> length_)
+CartesianPML::CartesianPML(Mesh *mesh_, Array2D<double> length_)
    : mesh(mesh_), length(length_)
 {
    dim = mesh->Dimension();
    SetBoundaries();
 }
 
-void PML::SetBoundaries()
+void CartesianPML::SetBoundaries()
 {
    comp_dom_bdr.SetSize(dim, 2);
    dom_bdr.SetSize(dim, 2);
@@ -1026,7 +1030,7 @@ void PML::SetBoundaries()
    }
 }
 
-void PML::SetAttributes(ParMesh *pmesh)
+void CartesianPML::SetAttributes(ParMesh *pmesh)
 {
    // Initialize bdr attributes
    for (int i = 0; i < pmesh->GetNBE(); ++i)
@@ -1076,8 +1080,8 @@ void PML::SetAttributes(ParMesh *pmesh)
    pmesh->SetAttributes();
 }
 
-void PML::StretchFunction(const Vector &x,
-                          vector<complex<double>> &dxs)
+void CartesianPML::StretchFunction(const Vector &x,
+                                   vector<complex<double>> &dxs)
 {
    complex<double> zi = complex<double>(0., 1.);
 

examples/ex34.cpp

@@ -1,412 +0,0 @@
-//                                MFEM Example 34
-//
-// Compile with: make ex34
-//
-// Sample runs:  ex34
-//
-// Description:  This example code demonstrates the use of MFEM to define a
-//               discontinuous Galerkin (DG) finite element discretization of
-//               the Laplace problem -Delta u = f with Dirichlet boundary
-//               conditions. Finite element spaces of any order, including zero
-//               on regular grids, are supported. The example highlights the
-//               use of coupling solution domains though custom physics defined
-//               on internal boundaries.
-//
-//               We recommend viewing examples 1 and 14 before viewing this
-//               example.
-
-#include "mfem.hpp"
-#include <fstream>
-#include <iostream>
-
-using namespace std;
-using namespace mfem;
-
-class InteriorLFIntegrator : public LinearFormIntegrator
-{
-public:
-   InteriorLFIntegrator(Coefficient &Q)
-      : Q(Q)
-   {}
-
-   void AssembleRHSElementVect(const FiniteElement &el1,
-                               const FiniteElement &el2,
-                               FaceElementTransformations &trans,
-                               Vector &mesh_coords_bar) override;
-
-   void AssembleRHSElementVect(const FiniteElement &el,
-                               ElementTransformation &trans,
-                               Vector &elvect) override
-   {
-      mfem_error("AssembleRHSElementVect(...)");
-   }
-
-private:
-   Coefficient &Q;
-
-#ifndef MFEM_THREAD_SAFE
-   Vector shape1;
-   Vector shape2;
-#endif
-};
-
-Mesh generate_mesh(int ref, int internal_bdr_attr = 5);
-
-int main(int argc, char *argv[])
-{
-   // 1. Parse command-line options.
-   int ref_levels = 0;
-   int order = 1;
-   int sol_order = 3;
-   double jump = -2;
-   double sigma = -1.0;
-   double kappa = -1.0;
-   double eta = 0.0;
-   bool visualization = 1;
-
-   OptionsParser args(argc, argv);
-   args.AddOption(&ref_levels, "-r", "--refine",
-                  "Number of times to refine the mesh uniformly, -1 for auto.");
-   args.AddOption(&order, "-o", "--order",
-                  "Finite element order (polynomial degree) >= 0.");
-   args.AddOption(&sigma, "-s", "--sigma",
-                  "One of the three DG penalty parameters, typically +1/-1."
-                  " See the documentation of class DGDiffusionIntegrator.");
-   args.AddOption(&kappa, "-k", "--kappa",
-                  "One of the three DG penalty parameters, should be positive."
-                  " Negative values are replaced with (order+1)^2.");
-   args.AddOption(&eta, "-e", "--eta", "BR2 penalty parameter.");
-   args.AddOption(&sol_order, "-so", "--solution_order",
-                  "Polynomial order of the exact solution >= 0.");
-   args.AddOption(&jump, "-j", "--jump",
-                  "Value of the discontinuity between the material regions.");
-   args.AddOption(&visualization, "-vis", "--visualization", "-no-vis",
-                  "--no-visualization",
-                  "Enable or disable GLVis visualization.");
-   args.Parse();
-   if (!args.Good())
-   {
-      args.PrintUsage(cout);
-      return 1;
-   }
-   if (kappa < 0)
-   {
-      kappa = (order+1)*(order+1);
-   }
-   if (sol_order < 0)
-   {
-      sol_order = 1;
-   }
-   args.PrintOptions(cout);
-
-   // 2. Construct the (serial) mesh and refine it if requested.
-   auto mesh = generate_mesh(ref_levels);
-
-   int dim = mesh.Dimension();
-
-   if (mesh.NURBSext)
-   {
-      mesh.SetCurvature(max(order, 1));
-   }
-
-   // 3. Define a finite element space on the mesh. Here we use discontinuous
-   //    finite elements of the specified order >= 0.
-   DG_FECollection fec(order, dim);
-   FiniteElementSpace fespace(&mesh, &fec);
-   cout << "Number of unknowns: " << fespace.GetVSize() << endl;
-
-   // 4. Set up the linear form b(.) which corresponds to the right-hand side of
-   //    the FEM linear system.
-   LinearForm b(&fespace);
-
-   Array<int> p1_attr_marker(mesh.attributes.Max());
-   p1_attr_marker = 0;
-   p1_attr_marker[0] = 1;
-
-   FunctionCoefficient p1_source([sol_order](const Vector &p)
-   {
-      const double x = p(0);
-      const double val = -(sol_order - 1)*sol_order*pow(x, sol_order-2);
-      return val;
-   });
-   b.AddDomainIntegrator(new DomainLFIntegrator(p1_source), p1_attr_marker);
-
-   Array<int> p2_attr_marker(mesh.attributes.Max());
-   p2_attr_marker = 0;
-   p2_attr_marker[1] = 1;
-
-   FunctionCoefficient p2_source([sol_order](const Vector &p)
-   {
-      const double x = p(0);
-      double val = -(sol_order - 1)*sol_order*pow(x - 2, sol_order-2);
-      if (sol_order % 2 == 0)
-      {
-         val *= -1.0;
-      }
-      return val;
-   });
-   b.AddDomainIntegrator(new DomainLFIntegrator(p2_source), p2_attr_marker);
-
-   ConstantCoefficient one(1.0);
-
-   Array<int> p1_bdr_attr_marker(mesh.bdr_attributes.Max());
-   p1_bdr_attr_marker = 0;
-   p1_bdr_attr_marker[0] = 1;
-
-   ConstantCoefficient left_bc_val(0.0);
-   b.AddBdrFaceIntegrator(
-      new DGDirichletLFIntegrator(left_bc_val, one, sigma, kappa),
-      p1_bdr_attr_marker);
-
-   Array<int> p2_bdr_attr_marker(mesh.bdr_attributes.Max());
-   p2_bdr_attr_marker = 0;
-   p2_bdr_attr_marker[1] = 1;
-
-   ConstantCoefficient right_bc_val(2.0 + jump);
-   b.AddBdrFaceIntegrator(
-      new DGDirichletLFIntegrator(right_bc_val, one, sigma, kappa),
-      p2_bdr_attr_marker);
-
-   Array<int> internal_bdr_attr_marker(mesh.bdr_attributes.Max());
-   internal_bdr_attr_marker = 0;
-   internal_bdr_attr_marker[4] = 1;
-
-   ConstantCoefficient interface_flux(sol_order);
-   b.AddInternalBoundaryFaceIntegrator(
-      new InteriorLFIntegrator(interface_flux),
-      internal_bdr_attr_marker);
-
-   b.Assemble();
-
-   // 5. Define the solution vector x as a finite element grid function
-   //    corresponding to fespace. Initialize x with initial guess of zero.
-   GridFunction x(&fespace);
-   x = 0.0;
-
-   // 6. Set up the bilinear form a(.,.) on the finite element space
-   //    corresponding to the Laplacian operator -Delta, by adding the Diffusion
-   //    domain integrator and the interior and boundary DG face integrators.
-   //    Note that boundary conditions are imposed weakly in the form, so there
-   //    is no need for dof elimination. After assembly and finalizing we
-   //    extract the corresponding sparse matrix A.
-   BilinearForm a(&fespace);
-   a.AddDomainIntegrator(new DiffusionIntegrator(one));
-
-   a.AddInteriorFaceIntegrator(new DGDiffusionIntegrator(one, sigma, kappa));
-   a.AddBdrFaceIntegrator(new DGDiffusionIntegrator(one, sigma, kappa),
-                          p1_bdr_attr_marker);
-   a.AddBdrFaceIntegrator(new DGDiffusionIntegrator(one, sigma, kappa),
-                          p2_bdr_attr_marker);
-   if (eta > 0)
-   {
-      a.AddInteriorFaceIntegrator(new DGDiffusionBR2Integrator(fespace, eta));
-      a.AddBdrFaceIntegrator(new DGDiffusionBR2Integrator(fespace, eta));
-   }
-
-   // 7. Negate the DG interface terms along the internal boundary so that the
-   //    only coupling between domains is from the chosen model (constant flux
-   //    in this case).
-   ProductCoefficient neg_one(-1.0, one);
-   a.AddInternalBoundaryFaceIntegrator(new DGDiffusionIntegrator(neg_one, sigma,
-                                                                 kappa),
-                                       internal_bdr_attr_marker);
-   if (eta > 0)
-   {
-      a.AddInternalBoundaryFaceIntegrator(new DGDiffusionBR2Integrator(fespace,
-                                                                       neg_one, eta),
-                                          internal_bdr_attr_marker);
-   }
-
-   a.Assemble();
-   a.Finalize();
-   const SparseMatrix &A = a.SpMat();
-
-#ifndef MFEM_USE_SUITESPARSE
-   // 8. Define a simple symmetric Gauss-Seidel preconditioner and use it to
-   //    solve the system Ax=b with PCG in the symmetric case, and GMRES in the
-   //    non-symmetric one.
-   GSSmoother M(A);
-   if (sigma == -1.0)
-   {
-      PCG(A, M, b, x, 1, 500, 1e-12, 0.0);
-   }
-   else
-   {
-      GMRES(A, M, b, x, 1, 500, 500, 1e-24, 0.0);
-   }
-#else
-   // 8. If MFEM was compiled with SuiteSparse, use UMFPACK to solve the system.
-   UMFPackSolver umf_solver;
-   umf_solver.Control[UMFPACK_ORDERING] = UMFPACK_ORDERING_METIS;
-   umf_solver.SetOperator(A);
-   umf_solver.Mult(b, x);
-#endif
-
-   // 9. 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);
-   ofstream sol_ofs("sol.gf");
-   sol_ofs.precision(8);
-   x.Save(sol_ofs);
-
-   // 10. Send the solution by socket to a GLVis server.
-   if (visualization)
-   {
-      char vishost[] = "localhost";
-      int  visport   = 19916;
-      socketstream sol_sock(vishost, visport);
-      sol_sock.precision(8);
-      sol_sock << "solution\n" << mesh << x << flush;
-   }
-
-   return 0;
-}
-
-void InteriorLFIntegrator::AssembleRHSElementVect(
-   const FiniteElement &el1,
-   const FiniteElement &el2,
-   FaceElementTransformations &trans,
-   Vector &elvect)
-{
-   int ndof1 = el1.GetDof();
-   int ndof2 = el2.GetDof();
-   int ndof = ndof1 + ndof2;
-
-#ifdef MFEM_THREAD_SAFE
-   Vector shape1;
-   Vector shape2;
-#endif
-   shape1.SetSize(ndof1);
-   shape2.SetSize(ndof2);
-
-   const auto *ir = IntRule;
-   if (ir == NULL)
-   {
-      int order = 2 * max(el1.GetOrder(), el2.GetOrder());
-      ir = &IntRules.Get(trans.GetGeometryType(), order);
-   }
-
-   elvect.SetSize(ndof);
-   Vector elvect1(elvect.GetData(), ndof1);
-   Vector elvect2(elvect.GetData() + ndof1, ndof2);
-   elvect = 0.0;
-   for (int i = 0; i < ir->GetNPoints(); i++)
-   {
-      const auto &ip = ir->IntPoint(i);
-
-      // Set the integration point in the face and the neighboring element
-      trans.SetAllIntPoints(&ip);
-
-      const double w = ip.weight * trans.Weight();
-
-      // Access the neighboring element's integration point
-      const auto &eip1 = trans.GetElement1IntPoint();
-      const auto &eip2 = trans.GetElement2IntPoint();
-
-      double Q_val = Q.Eval(trans, ip);
-
-      el1.CalcShape(eip1, shape1);
-      el2.CalcShape(eip2, shape2);
-
-      elvect1.Add(Q_val * w, shape1);
-      elvect2.Add(-Q_val * w, shape2);
-   }
-}
-
-Mesh generate_mesh(int ref, int internal_bdr_attr)
-{
-   int nxy = 4 * (ref+1);
-   auto mesh = Mesh::MakeCartesian2D(nxy, nxy, Element::TRIANGLE, true, 2.0, 1.0);
-   // auto mesh = Mesh::MakeCartesian2D(nxy, nxy, Element::QUADRILATERAL, true, 2.0, 1.0);
-
-   // assign element attributes to left and right sides
-   for (int i = 0; i < mesh.GetNE(); ++i)
-   {
-      auto *elem = mesh.GetElement(i);
-
-      Array<int> verts;
-      elem->GetVertices(verts);
-
-      bool left = true;
-      for (int j = 0; j < verts.Size(); ++j)
-      {
-         auto *vtx = mesh.GetVertex(verts[j]);
-         if (vtx[0] <= 1.0)
-         {
-            continue;
-         }
-         else
-         {
-            left = false;
-         }
-      }
-      if (left)
-      {
-         elem->SetAttribute(1);
-      }
-      else
-      {
-         elem->SetAttribute(2);
-      }
-   }
-
-   // assign boundary element attributes to left and right sides
-   for (int i = 0; i < mesh.GetNBE(); ++i)
-   {
-      auto *elem = mesh.GetBdrElement(i);
-
-      Array<int> verts;
-      elem->GetVertices(verts);
-
-      bool left = true;
-      bool right = true;
-      bool top = true;
-      bool bottom = true;
-      for (int j = 0; j < verts.Size(); ++j)
-      {
-         auto *vtx = mesh.GetVertex(verts[j]);
-         left = left && abs(vtx[0] - 0.0) < 1e-12;
-         right = right && abs(vtx[0] - 2.0) < 1e-12;
-         top = top && abs(vtx[1] - 1.0) < 1e-12;
-         bottom = bottom && abs(vtx[1] - 0.0) < 1e-12;
-      }
-      if (left)
-      {
-         elem->SetAttribute(1);
-      }
-      else if (right)
-      {
-         elem->SetAttribute(2);
-      }
-      else if (top)
-      {
-         elem->SetAttribute(3);
-      }
-      else if (bottom)
-      {
-         elem->SetAttribute(4);
-      }
-   }
-
-   // add internal boundary elements
-   for (int i = 0; i < mesh.GetNumFaces(); ++i)
-   {
-      int e1, e2;
-      mesh.GetFaceElements(i, &e1, &e2);
-      if (e1 >= 0 && e2 >= 0 && mesh.GetAttribute(e1) != mesh.GetAttribute(e2))
-      {
-         // This is the internal face between attributes.
-         auto *new_elem = mesh.GetFace(i)->Duplicate(&mesh);
-         new_elem->SetAttribute(internal_bdr_attr);
-         mesh.AddBdrElement(new_elem);
-      }
-   }
-
-   mesh.FinalizeTopology(); // Finalize to build relevant tables
-   mesh.Finalize();
-   mesh.SetAttributes();
-
-   return mesh;
-}

examples/ex35.cpp

@@ -1,451 +0,0 @@
-//                                MFEM Example 36
-//
-// Compile with: make ex36
-//
-// Sample runs:  ex36
-//
-// Description:  This example code demonstrates the use of MFEM to define a
-//               discontinuous Galerkin (DG) finite element discretization of
-//               the Laplace problem -Delta u = f with Dirichlet boundary
-//               conditions. Finite element spaces of any order, including zero
-//               on regular grids, are supported. The example highlights the
-//               use of coupling solution domains though custom physics defined
-//               on internal boundaries.
-//
-//               We recommend viewing examples 1, 14, and 34 before viewing this
-//               example.
-
-#include "mfem.hpp"
-#include <fstream>
-#include <iostream>
-
-using namespace std;
-using namespace mfem;
-
-class InteriorMassIntegrator : public BilinearFormIntegrator
-{
-public:
-   InteriorMassIntegrator(Coefficient &Q)
-      : Q(Q)
-   {}
-
-   void AssembleFaceMatrix(const FiniteElement &el1,
-                           const FiniteElement &el2,
-                           FaceElementTransformations &trans,
-                           DenseMatrix &elmat) override;
-
-   using BilinearFormIntegrator::AssembleFaceMatrix;
-
-private:
-   Coefficient &Q;
-
-#ifndef MFEM_THREAD_SAFE
-   Vector shape1;
-   Vector shape2;
-   DenseMatrix elmat11;
-   DenseMatrix elmat12;
-   DenseMatrix elmat21;
-   DenseMatrix elmat22;
-#endif
-};
-
-Mesh generate_mesh(int ref, int internal_bdr_attr = 5);
-
-int main(int argc, char *argv[])
-{
-   // 1. Parse command-line options.
-   int ref_levels = 0;
-   int order = 1;
-   int sol_order = 3;
-   double jump = -2;
-   double sigma = -1.0;
-   double kappa = -1.0;
-   double eta = 0.0;
-   bool visualization = 1;
-
-   OptionsParser args(argc, argv);
-   args.AddOption(&ref_levels, "-r", "--refine",
-                  "Number of times to refine the mesh uniformly, -1 for auto.");
-   args.AddOption(&order, "-o", "--order",
-                  "Finite element order (polynomial degree) >= 0.");
-   args.AddOption(&sigma, "-s", "--sigma",
-                  "One of the three DG penalty parameters, typically +1/-1."
-                  " See the documentation of class DGDiffusionIntegrator.");
-   args.AddOption(&kappa, "-k", "--kappa",
-                  "One of the three DG penalty parameters, should be positive."
-                  " Negative values are replaced with (order+1)^2.");
-   args.AddOption(&eta, "-e", "--eta", "BR2 penalty parameter.");
-   args.AddOption(&sol_order, "-so", "--solution_order",
-                  "Polynomial order of the exact solution >= 0.");
-   args.AddOption(&jump, "-j", "--jump",
-                  "Value of the discontinuity between the material regions.");
-   args.AddOption(&visualization, "-vis", "--visualization", "-no-vis",
-                  "--no-visualization",
-                  "Enable or disable GLVis visualization.");
-   args.Parse();
-   if (!args.Good())
-   {
-      args.PrintUsage(cout);
-      return 1;
-   }
-   if (kappa < 0)
-   {
-      kappa = (order+1)*(order+1);
-   }
-   if (sol_order < 0)
-   {
-      sol_order = 1;
-   }
-   args.PrintOptions(cout);
-
-   // 2. Construct the (serial) mesh and refine it if requested.
-   auto mesh = generate_mesh(ref_levels);
-
-   int dim = mesh.Dimension();
-
-   if (mesh.NURBSext)
-   {
-      mesh.SetCurvature(max(order, 1));
-   }
-
-   // 3. Define a finite element space on the mesh. Here we use discontinuous
-   //    finite elements of the specified order >= 0.
-   DG_FECollection fec(order, dim);
-   FiniteElementSpace fespace(&mesh, &fec);
-   cout << "Number of unknowns: " << fespace.GetVSize() << endl;
-
-   // 4. Set up the linear form b(.) which corresponds to the right-hand side of
-   //    the FEM linear system.
-   LinearForm b(&fespace);
-
-   Array<int> p1_attr_marker(mesh.attributes.Max());
-   p1_attr_marker = 0;
-   p1_attr_marker[0] = 1;
-
-   FunctionCoefficient p1_source([sol_order](const Vector &p)
-   {
-      const double x = p(0);
-      const double val = -(sol_order - 1)*sol_order*pow(x, sol_order-2);
-      return val;
-   });
-   b.AddDomainIntegrator(new DomainLFIntegrator(p1_source), p1_attr_marker);
-
-   Array<int> p2_attr_marker(mesh.attributes.Max());
-   p2_attr_marker = 0;
-   p2_attr_marker[1] = 1;
-
-   FunctionCoefficient p2_source([sol_order](const Vector &p)
-   {
-      const double x = p(0);
-      double val = -(sol_order - 1)*sol_order*pow(x - 2, sol_order-2);
-      if (sol_order % 2 == 0)
-      {
-         val *= -1.0;
-      }
-      return val;
-   });
-   b.AddDomainIntegrator(new DomainLFIntegrator(p2_source), p2_attr_marker);
-
-   ConstantCoefficient one(1.0);
-
-   Array<int> p1_bdr_attr_marker(mesh.bdr_attributes.Max());
-   p1_bdr_attr_marker = 0;
-   p1_bdr_attr_marker[0] = 1;
-
-   ConstantCoefficient left_bc_val(0.0);
-   b.AddBdrFaceIntegrator(
-      new DGDirichletLFIntegrator(left_bc_val, one, sigma, kappa),
-      p1_bdr_attr_marker);
-
-   Array<int> p2_bdr_attr_marker(mesh.bdr_attributes.Max());
-   p2_bdr_attr_marker = 0;
-   p2_bdr_attr_marker[1] = 1;
-
-   ConstantCoefficient right_bc_val(2.0 + jump);
-   b.AddBdrFaceIntegrator(
-      new DGDirichletLFIntegrator(right_bc_val, one, sigma, kappa),
-      p2_bdr_attr_marker);
-
-   b.Assemble();
-
-   // 5. Define the solution vector x as a finite element grid function
-   //    corresponding to fespace. Initialize x with initial guess of zero.
-   GridFunction x(&fespace);
-   x = 0.0;
-
-   // 6. Set up the bilinear form a(.,.) on the finite element space
-   //    corresponding to the Laplacian operator -Delta, by adding the Diffusion
-   //    domain integrator and the interior and boundary DG face integrators.
-   //    Note that boundary conditions are imposed weakly in the form, so there
-   //    is no need for dof elimination. After assembly and finalizing we
-   //    extract the corresponding sparse matrix A.
-   BilinearForm a(&fespace);
-   a.AddDomainIntegrator(new DiffusionIntegrator(one));
-
-   a.AddInteriorFaceIntegrator(new DGDiffusionIntegrator(one, sigma, kappa));
-   a.AddBdrFaceIntegrator(new DGDiffusionIntegrator(one, sigma, kappa),
-                          p1_bdr_attr_marker);
-   a.AddBdrFaceIntegrator(new DGDiffusionIntegrator(one, sigma, kappa),
-                          p2_bdr_attr_marker);
-   if (eta > 0)
-   {
-      a.AddInteriorFaceIntegrator(new DGDiffusionBR2Integrator(fespace, eta));
-      a.AddBdrFaceIntegrator(new DGDiffusionBR2Integrator(fespace, eta));
-   }
-
-   // 7. Negate the DG interface terms along the internal boundary so that the
-   //    only coupling between domains is from the chosen model (constant flux
-   //    in this case).
-   Array<int> internal_bdr_attr_marker(mesh.bdr_attributes.Max());
-   internal_bdr_attr_marker = 0;
-   internal_bdr_attr_marker[4] = 1;
-
-   ProductCoefficient neg_one(-1.0, one);
-   a.AddInternalBoundaryFaceIntegrator(new DGDiffusionIntegrator(neg_one, sigma,
-                                                                 kappa),
-                                       internal_bdr_attr_marker);
-   if (eta > 0)
-   {
-      a.AddInternalBoundaryFaceIntegrator(new DGDiffusionBR2Integrator(fespace,
-                                                                       neg_one, eta),
-                                          internal_bdr_attr_marker);
-   }
-
-   ConstantCoefficient mass_coeff(sol_order / jump);
-   a.AddInternalBoundaryFaceIntegrator(new InteriorMassIntegrator(mass_coeff),
-                                       internal_bdr_attr_marker);
-
-   a.Assemble();
-   a.Finalize();
-   const SparseMatrix &A = a.SpMat();
-
-#ifndef MFEM_USE_SUITESPARSE
-   // 8. Define a simple symmetric Gauss-Seidel preconditioner and use it to
-   //    solve the system Ax=b with PCG in the symmetric case, and GMRES in the
-   //    non-symmetric one.
-   GSSmoother M(A);
-   if (sigma == -1.0 && !(jump < 0))
-   {
-      PCG(A, M, b, x, 1, 500, 1e-12, 0.0);
-   }
-   else
-   {
-      GMRES(A, M, b, x, 1, 500, 500, 1e-24, 0.0);
-   }
-#else
-   // 8. If MFEM was compiled with SuiteSparse, use UMFPACK to solve the system.
-   UMFPackSolver umf_solver;
-   umf_solver.Control[UMFPACK_ORDERING] = UMFPACK_ORDERING_METIS;
-   umf_solver.SetOperator(A);
-   umf_solver.Mult(b, x);
-#endif
-
-   // 9. 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);
-   ofstream sol_ofs("sol.gf");
-   sol_ofs.precision(8);
-   x.Save(sol_ofs);
-
-   // 10. Send the solution by socket to a GLVis server.
-   if (visualization)
-   {
-      char vishost[] = "localhost";
-      int  visport   = 19916;
-      socketstream sol_sock(vishost, visport);
-      sol_sock.precision(8);
-      sol_sock << "solution\n" << mesh << x << flush;
-   }
-
-   return 0;
-}
-
-void InteriorMassIntegrator::AssembleFaceMatrix(
-   const FiniteElement &el1,
-   const FiniteElement &el2,
-   FaceElementTransformations &trans,
-   DenseMatrix &elmat)
-{
-   int ndof1 = el1.GetDof();
-   int ndof2 = el2.GetDof();
-   int ndof = ndof1 + ndof2;
-
-#ifdef MFEM_THREAD_SAFE
-   Vector shape1;
-   Vector shape2;
-   DenseMatrix elmat11;
-   DenseMatrix elmat12;
-   DenseMatrix elmat21;
-   DenseMatrix elmat22;
-#endif
-   shape1.SetSize(ndof1);
-   shape2.SetSize(ndof2);
-
-   elmat11.SetSize(ndof1);
-   elmat12.SetSize(ndof1, ndof2);
-   elmat21.SetSize(ndof2, ndof1);
-   elmat22.SetSize(ndof2);
-
-   const auto *ir = IntRule;
-   if (ir == NULL)
-   {
-      int order = 2 * max(el1.GetOrder(), el2.GetOrder());
-      ir = &IntRules.Get(trans.GetGeometryType(), order);
-   }
-
-   elmat.SetSize(ndof);
-   elmat = 0.0;
-   for (int i = 0; i < ir->GetNPoints(); i++)
-   {
-      const auto &ip = ir->IntPoint(i);
-
-      // Set the integration point in the face and the neighboring element
-      trans.SetAllIntPoints(&ip);
-
-      const double w = ip.weight * trans.Weight();
-
-      // Access the neighboring element's integration point
-      const auto &eip1 = trans.GetElement1IntPoint();
-      const auto &eip2 = trans.GetElement2IntPoint();
-
-      el1.CalcShape(eip1, shape1);
-      el2.CalcShape(eip2, shape2);
-
-      const double Q_val = Q.Eval(trans, ip);
-
-      elmat11 = 0.0;
-      AddMult_a_VVt(Q_val * w, shape1, elmat11);
-
-      elmat12 = 0.0;
-      AddMult_a_VWt(-Q_val * w, shape2, shape1, elmat12);
-
-      elmat21 = 0.0;
-      AddMult_a_VWt(-Q_val * w, shape1, shape2, elmat21);
-
-      elmat22 = 0.0;
-      AddMult_a_VVt(Q_val * w, shape2, elmat22);
-
-      for (int j = 0; j < ndof1; ++j)
-      {
-         for (int k = 0; k < ndof1; ++k)
-         {
-            elmat(j, k) += elmat11(j, k);
-         }
-      }
-
-      for (int j = 0; j < ndof1; ++j)
-      {
-         for (int k = 0; k < ndof2; ++k)
-         {
-            elmat(j, k + ndof1) += elmat12(j, k);
-            elmat(k + ndof1, j) += elmat21(k, j);
-         }
-      }
-
-      for (int j = 0; j < ndof2; ++j)
-      {
-         for (int k = 0; k < ndof2; ++k)
-         {
-            elmat(j + ndof1, k + ndof1) += elmat22(j, k);
-         }
-      }
-   }
-}
-
-Mesh generate_mesh(int ref, int internal_bdr_attr)
-{
-   int nxy = 4 * (ref+1);
-   auto mesh = Mesh::MakeCartesian2D(nxy, nxy, Element::TRIANGLE, true, 2.0, 1.0);
-   // auto mesh = Mesh::MakeCartesian2D(nxy, nxy, Element::QUADRILATERAL, true, 2.0, 1.0);
-
-   // assign element attributes to left and right sides
-   for (int i = 0; i < mesh.GetNE(); ++i)
-   {
-      auto *elem = mesh.GetElement(i);
-
-      Array<int> verts;
-      elem->GetVertices(verts);
-
-      bool left = true;
-      for (int j = 0; j < verts.Size(); ++j)
-      {
-         auto *vtx = mesh.GetVertex(verts[j]);
-         if (vtx[0] <= 1.0)
-         {
-            continue;
-         }
-         else
-         {
-            left = false;
-         }
-      }
-      if (left)
-      {
-         elem->SetAttribute(1);
-      }
-      else
-      {
-         elem->SetAttribute(2);
-      }
-   }
-
-   // assign boundary element attributes to left and right sides
-   for (int i = 0; i < mesh.GetNBE(); ++i)
-   {
-      auto *elem = mesh.GetBdrElement(i);
-
-      Array<int> verts;
-      elem->GetVertices(verts);
-
-      bool left = true;
-      bool right = true;
-      bool top = true;
-      bool bottom = true;
-      for (int j = 0; j < verts.Size(); ++j)
-      {
-         auto *vtx = mesh.GetVertex(verts[j]);
-         left = left && abs(vtx[0] - 0.0) < 1e-12;
-         right = right && abs(vtx[0] - 2.0) < 1e-12;
-         top = top && abs(vtx[1] - 1.0) < 1e-12;
-         bottom = bottom && abs(vtx[1] - 0.0) < 1e-12;
-      }
-      if (left)
-      {
-         elem->SetAttribute(1);
-      }
-      else if (right)
-      {
-         elem->SetAttribute(2);
-      }
-      else if (top)
-      {
-         elem->SetAttribute(3);
-      }
-      else if (bottom)
-      {
-         elem->SetAttribute(4);
-      }
-   }
-
-   // add internal boundary elements
-   for (int i = 0; i < mesh.GetNumFaces(); ++i)
-   {
-      int e1, e2;
-      mesh.GetFaceElements(i, &e1, &e2);
-      if (e1 >= 0 && e2 >= 0 && mesh.GetAttribute(e1) != mesh.GetAttribute(e2))
-      {
-         // This is the internal face between attributes.
-         auto *new_elem = mesh.GetFace(i)->Duplicate(&mesh);
-         new_elem->SetAttribute(internal_bdr_attr);
-         mesh.AddBdrElement(new_elem);
-      }
-   }
-
-   mesh.FinalizeTopology(); // Finalize to build relevant tables
-   mesh.Finalize();
-   mesh.SetAttributes();
-
-   return mesh;
-}

examples/makefile

@@ -23,7 +23,7 @@ MFEM_LIB_FILE = mfem_is_not_built
 
 SEQ_EXAMPLES = ex0 ex1 ex2 ex3 ex4 ex5 ex6 ex7 ex8 ex9 ex10 ex14 ex15 ex16 \
   ex17 ex18 ex19 ex20 ex21 ex22 ex23 ex24 ex25 ex26 ex27 ex28 ex29 ex30 \
-  ex31 ex33 ex34 ex35
+  ex31 ex33
 PAR_EXAMPLES = ex0p ex1p ex2p ex3p ex4p ex5p ex6p ex7p ex8p ex9p ex10p ex11p \
   ex12p ex13p ex14p ex15p ex16p ex17p ex18p ex19p ex20p ex21p ex22p ex24p \
   ex25p ex26p ex27p ex28p ex29p ex30p ex31p ex32p ex33p
@@ -36,7 +36,7 @@ ifeq ($(MFEM_USE_MPI),NO)
 else
    EXAMPLES = $(PAR_EXAMPLES) $(SEQ_EXAMPLES)
 endif
-SUBDIRS =
+SUBDIRS = contact 
 ifeq ($(MFEM_USE_AMGX),YES)
    SUBDIRS += amgx
 endif

examples/superlu/ex1p.cpp

@@ -67,7 +67,6 @@ int main(int argc, char *argv[])
    int slu_colperm = 4;
    int slu_rowperm = 1;
    int slu_iterref = 2;
-   int slu_npdep = 1;
 
    OptionsParser args(argc, argv);
    args.AddOption(&mesh_file, "-m", "--mesh",
@@ -86,11 +85,9 @@ int main(int argc, char *argv[])
                   "6-ZOLTAN");
    args.AddOption(&slu_rowperm, "-rp", "--rowperm",
                   "SuperLU Row Permutation Method:  0-NOROWPERM, 1-LargeDiag");
-   args.AddOption(&slu_iterref, "-ir", "--iterref",
+   args.AddOption(&slu_iterref, "-rp", "--rowperm",
                   "SuperLU Iterative Refinement:  0-NOREFINE, 1-Single, "
                   "2-Double, 3-Extra");
-   args.AddOption(&slu_npdep, "-npdep", "--npdepth",
-                  "Depth of 3D parition for SuperLU (>= 7.2.0)");
 
    args.Parse();
    if (!args.Good())
@@ -217,7 +214,7 @@ int main(int argc, char *argv[])
    a.FormLinearSystem(ess_tdof_list, x, b, A, X, B);
 
    // 13. Solve the linear system A X = B utilizing SuperLU.
-   SuperLUSolver *superlu = new SuperLUSolver(MPI_COMM_WORLD, slu_npdep);
+   SuperLUSolver *superlu = new SuperLUSolver(MPI_COMM_WORLD);
    Operator *SLU_A = new SuperLURowLocMatrix(*A.As<HypreParMatrix>());
    superlu->SetPrintStatistics(true);
    superlu->SetSymmetricPattern(false);
@@ -284,9 +281,10 @@ int main(int argc, char *argv[])
    superlu->SetOperator(*SLU_A);
    superlu->SetPrintStatistics(true);
    superlu->Mult(B, X);
+   superlu->DismantleGrid();
 
-   delete superlu;
    delete SLU_A;
+   delete superlu;
 
    // 14. Recover the parallel grid function corresponding to X. This is the
    //     local finite element solution on each processor.

fem/CMakeLists.txt

@@ -13,44 +13,28 @@ set(SRCS
   bilinearform.cpp
   bilinearform_ext.cpp
   bilininteg.cpp
-  integ/bilininteg_br2.cpp
-  integ/bilininteg_convection_mf.cpp
-  integ/bilininteg_convection_pa.cpp
-  integ/bilininteg_convection_ea.cpp
-  integ/bilininteg_curlcurl_pa.cpp
-  integ/bilininteg_dgtrace_pa.cpp
-  integ/bilininteg_dgtrace_ea.cpp
-  integ/bilininteg_diffusion_mf.cpp
-  integ/bilininteg_diffusion_pa.cpp
-  integ/bilininteg_diffusion_ea.cpp
-  integ/bilininteg_divdiv_pa.cpp
-  integ/bilininteg_gradient_pa.cpp
-  integ/bilininteg_interp_pa.cpp
-  integ/bilininteg_mass_mf.cpp
-  integ/bilininteg_mass_pa.cpp
-  integ/bilininteg_mass_ea.cpp
-  integ/bilininteg_mixedcurl_pa.cpp
-  integ/bilininteg_mixedvecgrad_pa.cpp
-  integ/bilininteg_transpose_ea.cpp
-  integ/bilininteg_vecdiffusion_mf.cpp
-  integ/bilininteg_vecdiffusion_pa.cpp
-  integ/bilininteg_vecdiv_pa.cpp
-  integ/bilininteg_vecmass_mf.cpp
-  integ/bilininteg_vecmass_pa.cpp
-  integ/bilininteg_vectorfediv_pa.cpp
-  integ/bilininteg_vectorfemass_pa.cpp
-  integ/bilininteg_diffusion_kernels.cpp
-  integ/bilininteg_hcurl_kernels.cpp
-  integ/bilininteg_hdiv_kernels.cpp
-  integ/bilininteg_hcurlhdiv_kernels.cpp
-  integ/bilininteg_mass_kernels.cpp
-  integ/lininteg_boundary.cpp
-  integ/lininteg_boundary_flux.cpp
-  integ/lininteg_domain.cpp
-  integ/lininteg_domain_grad.cpp
-  integ/lininteg_domain_vectorfe.cpp
-  integ/nonlininteg_vecconvection_pa.cpp
-  integ/nonlininteg_vecconvection_mf.cpp
+  bilininteg_br2.cpp
+  bilininteg_convection_mf.cpp
+  bilininteg_convection_pa.cpp
+  bilininteg_convection_ea.cpp
+  bilininteg_dgtrace_pa.cpp
+  bilininteg_dgtrace_ea.cpp
+  bilininteg_diffusion_mf.cpp
+  bilininteg_diffusion_pa.cpp
+  bilininteg_diffusion_ea.cpp
+  bilininteg_divergence.cpp
+  bilininteg_hcurl.cpp
+  bilininteg_hdiv.cpp
+  bilininteg_vectorfe.cpp
+  bilininteg_gradient.cpp
+  bilininteg_mass_mf.cpp
+  bilininteg_mass_pa.cpp
+  bilininteg_mass_ea.cpp
+  bilininteg_transpose_ea.cpp
+  bilininteg_vecdiffusion.cpp
+  bilininteg_vecdiffusion_mf.cpp
+  bilininteg_vecmass.cpp
+  bilininteg_vecmass_mf.cpp
   coefficient.cpp
   complex_fem.cpp
   convergence.cpp
@@ -90,6 +74,11 @@ set(SRCS
   linearform.cpp
   linearform_ext.cpp
   lininteg.cpp
+  lininteg_boundary.cpp
+  lininteg_boundary_flux.cpp
+  lininteg_domain.cpp
+  lininteg_domain_grad.cpp
+  lininteg_vectorfe_domain.cpp
   lor/lor.cpp
   lor/lor_ads.cpp
   lor/lor_ams.cpp
@@ -102,6 +91,8 @@ set(SRCS
   nonlinearform_ext.cpp
   nonlininteg.cpp
   fespacehierarchy.cpp
+  nonlininteg_vectorconvection.cpp
+  nonlininteg_vectorconvection_mf.cpp
   qfunction.cpp
   qinterp/det.cpp
   qinterp/eval_by_nodes.cpp
@@ -152,11 +143,7 @@ set(HDRS
   bilinearform.hpp
   bilinearform_ext.hpp
   bilininteg.hpp
-  integ/bilininteg_diffusion_kernels.hpp
-  integ/bilininteg_hcurl_kernels.hpp
-  integ/bilininteg_hdiv_kernels.hpp
-  integ/bilininteg_hcurlhdiv_kernels.hpp
-  integ/bilininteg_mass_kernels.hpp
+  bilininteg_mass_pa.hpp
   coefficient.hpp
   complex_fem.hpp
   convergence.hpp

fem/bilinearform.cpp

@@ -14,7 +14,6 @@
 #include "fem.hpp"
 #include "../general/device.hpp"
 #include <cmath>
-#include <cstddef>
 
 namespace mfem
 {
@@ -110,9 +109,6 @@ BilinearForm::BilinearForm (FiniteElementSpace * f, BilinearForm * bf, int ps)
    boundary_face_integs = bf->boundary_face_integs;
    boundary_face_integs_marker = bf->boundary_face_integs_marker;
 
-   internal_boundary_face_integs = bf->internal_boundary_face_integs;
-   internal_boundary_face_integs_marker = bf->internal_boundary_face_integs_marker;
-
    AllocMat();
 }
 
@@ -282,22 +278,6 @@ void BilinearForm::AddBdrFaceIntegrator(BilinearFormIntegrator *bfi,
    boundary_face_integs_marker.Append(&bdr_marker);
 }
 
-void BilinearForm::AddInternalBoundaryFaceIntegrator(BilinearFormIntegrator
-                                                     *bfi)
-{
-   internal_boundary_face_integs.Append(bfi);
-   // nullptr -> all attributes are active
-   internal_boundary_face_integs_marker.Append(nullptr);
-}
-
-void BilinearForm::AddInternalBoundaryFaceIntegrator(BilinearFormIntegrator
-                                                     *bfi,
-                                                     Array<int> &internal_bdr_attr_marker)
-{
-   internal_boundary_face_integs.Append(bfi);
-   internal_boundary_face_integs_marker.Append(&internal_bdr_attr_marker);
-}
-
 void BilinearForm::ComputeElementMatrix(int i, DenseMatrix &elmat)
 {
    if (element_matrices)
@@ -650,59 +630,6 @@ void BilinearForm::Assemble(int skip_zeros)
       }
    }
 
-   if (internal_boundary_face_integs.Size())
-   {
-      // Which internal boundary attributes need to be processed?
-      Array<int> bdr_attr_marker(mesh->bdr_attributes.Size() ?
-                                 mesh->bdr_attributes.Max() : 0);
-      bdr_attr_marker = 0;
-      for (int k = 0; k < internal_boundary_face_integs.Size(); k++)
-      {
-         if (internal_boundary_face_integs_marker[k] == NULL)
-         {
-            bdr_attr_marker = 1;
-            break;
-         }
-         auto &bdr_marker = *internal_boundary_face_integs_marker[k];
-         MFEM_ASSERT(bdr_marker.Size() == bdr_attr_marker.Size(),
-                     "invalid boundary marker for internal boundary face "
-                     "integrator #" << k << ", counting from zero");
-         for (int i = 0; i < bdr_attr_marker.Size(); i++)
-         {
-            bdr_attr_marker[i] |= bdr_marker[i];
-         }
-      }
-
-      Array<int> vdofs2;
-      for (int i = 0; i < mesh->GetNBE(); i++)
-      {
-         const int bdr_attr = mesh->GetBdrAttribute(i);
-         if (bdr_attr_marker[bdr_attr-1] == 0) { continue; }
-
-         auto *tr = mesh->GetInternalBdrFaceTransformations(i);
-         if (tr != nullptr)
-         {
-            fes->GetElementVDofs(tr->Elem1No, vdofs);
-            fes->GetElementVDofs(tr->Elem2No, vdofs2);
-            vdofs.Append(vdofs2);
-            const auto *fe1 = fes->GetFE(tr->Elem1No);
-            const auto *fe2 = fes->GetFE(tr->Elem2No);
-            for (int k = 0; k < internal_boundary_face_integs.Size(); k++)
-            {
-               if (internal_boundary_face_integs_marker[k] &&
-                   (*internal_boundary_face_integs_marker[k])[bdr_attr - 1] == 0)
-               {
-                  continue;
-               }
-
-               internal_boundary_face_integs[k]->AssembleFaceMatrix(
-                  *fe1, *fe2, *tr, elemmat);
-               mat->AddSubMatrix(vdofs, vdofs, elemmat, skip_zeros);
-            }
-         }
-      }
-   }
-
 #ifdef MFEM_USE_LEGACY_OPENMP
    if (free_element_matrices)
    {
@@ -1216,10 +1143,6 @@ BilinearForm::~BilinearForm()
       { delete interior_face_integs[k]; }
       for (k=0; k < boundary_face_integs.Size(); k++)
       { delete boundary_face_integs[k]; }
-      for (int i = 0; i < internal_boundary_face_integs.Size(); i++)
-      {
-         delete internal_boundary_face_integs[i];
-      }
    }
 
    delete ext;

fem/bilinearform.hpp

@@ -113,10 +113,6 @@ protected:
    Array<BilinearFormIntegrator*> boundary_face_integs;
    Array<Array<int>*> boundary_face_integs_marker; ///< Entries are not owned.
 
-   /// Set of internal boundary face integrators to be applied.
-   Array<BilinearFormIntegrator*> internal_boundary_face_integs;
-   Array<Array<int>*> internal_boundary_face_integs_marker; ///< Entries not owned.
-
    DenseMatrix elemmat;
    Array<int>  vdofs;
 
@@ -420,18 +416,6 @@ public:
    void AddBdrFaceIntegrator(BilinearFormIntegrator *bfi,
                              Array<int> &bdr_marker);
 
-   /// @brief Add new internal boundary face integrator. Assumes ownership of
-   /// @a bfi.
-   void AddInternalBoundaryFaceIntegrator(BilinearFormIntegrator *bfi);
-
-   /** @brief Add new internal boundary face integrator, restricted to the given
-       boundary attributes.
-
-       Assumes ownership of @a bfi. The array @a internal_bdr_attr_marker is
-       stored internally as a pointer to the given Array<int> object. */
-   void AddInternalBoundaryFaceIntegrator(BilinearFormIntegrator *bfi,
-                                          Array<int> &internal_bdr_attr_marker);
-
    /// Sets all sparse values of \f$ M \f$ and \f$ M_e \f$ to 'a'.
    void operator=(const double a)
    {

fem/bilinearform_ext.cpp

@@ -56,9 +56,6 @@ void MFBilinearFormExtension::Assemble()
    {
       integrators[i]->AssembleMF(*a->FESpace());
    }
-
-   MFEM_VERIFY(a->GetBBFI()->Size() == 0, "AddBoundaryIntegrator is not "
-               "currently supported in MFBilinearFormExtension");
 }
 
 void MFBilinearFormExtension::AssembleDiagonal(Vector &y) const
@@ -278,9 +275,7 @@ void PABilinearFormExtension::SetupRestrictionOperators(const L2FaceValues m)
       int_face_Y.UseDevice(true); // ensure 'int_face_Y = 0.0' is done on device
    }
 
-   const bool has_bdr_integs = (a->GetBFBFI()->Size() > 0 ||
-                                a->GetBBFI()->Size() > 0);
-   if (bdr_face_restrict_lex == NULL && has_bdr_integs)
+   if (bdr_face_restrict_lex == NULL && a->GetBFBFI()->Size() > 0)
    {
       bdr_face_restrict_lex = trial_fes->GetFaceRestriction(
                                  ElementDofOrdering::LEXICOGRAPHIC,
@@ -297,27 +292,27 @@ void PABilinearFormExtension::Assemble()
    SetupRestrictionOperators(L2FaceValues::DoubleValued);
 
    Array<BilinearFormIntegrator*> &integrators = *a->GetDBFI();
-   for (BilinearFormIntegrator *integ : integrators)
+   const int integratorCount = integrators.Size();
+   for (int i = 0; i < integratorCount; ++i)
    {
-      integ->AssemblePA(*a->FESpace());
+      integrators[i]->AssemblePA(*a->FESpace());
    }
 
-   Array<BilinearFormIntegrator*> &bdr_integrators = *a->GetBBFI();
-   for (BilinearFormIntegrator *integ : bdr_integrators)
-   {
-      integ->AssemblePABoundary(*a->FESpace());
-   }
+   MFEM_VERIFY(a->GetBBFI()->Size() == 0,
+               "Partial assembly does not support AddBoundaryIntegrator yet.");
 
    Array<BilinearFormIntegrator*> &intFaceIntegrators = *a->GetFBFI();
-   for (BilinearFormIntegrator *integ : intFaceIntegrators)
+   const int intFaceIntegratorCount = intFaceIntegrators.Size();
+   for (int i = 0; i < intFaceIntegratorCount; ++i)
    {
-      integ->AssemblePAInteriorFaces(*a->FESpace());
+      intFaceIntegrators[i]->AssemblePAInteriorFaces(*a->FESpace());
    }
 
    Array<BilinearFormIntegrator*> &bdrFaceIntegrators = *a->GetBFBFI();
-   for (BilinearFormIntegrator *integ : bdrFaceIntegrators)
+   const int boundFaceIntegratorCount = bdrFaceIntegrators.Size();
+   for (int i = 0; i < boundFaceIntegratorCount; ++i)
    {
-      integ->AssemblePABoundaryFaces(*a->FESpace());
+      bdrFaceIntegrators[i]->AssemblePABoundaryFaces(*a->FESpace());
    }
 }
 
@@ -328,27 +323,20 @@ void PABilinearFormExtension::AssembleDiagonal(Vector &y) const
    const int iSz = integrators.Size();
    if (elem_restrict && !DeviceCanUseCeed())
    {
-      if (iSz > 0)
+      localY = 0.0;
+      for (int i = 0; i < iSz; ++i)
       {
-         localY = 0.0;
-         for (int i = 0; i < iSz; ++i)
-         {
-            integrators[i]->AssembleDiagonalPA(localY);
-         }
-         const ElementRestriction* H1elem_restrict =
-            dynamic_cast<const ElementRestriction*>(elem_restrict);
-         if (H1elem_restrict)
-         {
-            H1elem_restrict->MultTransposeUnsigned(localY, y);
-         }
-         else
-         {
-            elem_restrict->MultTranspose(localY, y);
-         }
+         integrators[i]->AssembleDiagonalPA(localY);
+      }
+      const ElementRestriction* H1elem_restrict =
+         dynamic_cast<const ElementRestriction*>(elem_restrict);
+      if (H1elem_restrict)
+      {
+         H1elem_restrict->MultTransposeUnsigned(localY, y);
       }
       else
       {
-         y = 0.0;
+         elem_restrict->MultTranspose(localY, y);
       }
    }
    else
@@ -360,18 +348,6 @@ void PABilinearFormExtension::AssembleDiagonal(Vector &y) const
          integrators[i]->AssembleDiagonalPA(y);
       }
    }
-
-   Array<BilinearFormIntegrator*> &bdr_integs = *a->GetBBFI();
-   const int n_bdr_integs = bdr_integs.Size();
-   if (bdr_face_restrict_lex && n_bdr_integs > 0)
-   {
-      bdr_face_Y = 0.0;
-      for (int i = 0; i < n_bdr_integs; ++i)
-      {
-         bdr_integs[i]->AssembleDiagonalPA(bdr_face_Y);
-      }
-      bdr_face_restrict_lex->AddMultTransposeUnsigned(bdr_face_Y, y);
-   }
 }
 
 void PABilinearFormExtension::Update()
@@ -421,20 +397,13 @@ void PABilinearFormExtension::Mult(const Vector &x, Vector &y) const
    }
    else
    {
-      if (iSz)
+      elem_restrict->Mult(x, localX);
+      localY = 0.0;
+      for (int i = 0; i < iSz; ++i)
       {
-         elem_restrict->Mult(x, localX);
-         localY = 0.0;
-         for (int i = 0; i < iSz; ++i)
-         {
-            integrators[i]->AddMultPA(localX, localY);
-         }
-         elem_restrict->MultTranspose(localY, y);
-      }
-      else
-      {
-         y = 0.0;
+         integrators[i]->AddMultPA(localX, localY);
       }
+      elem_restrict->MultTranspose(localY, y);
    }
 
    Array<BilinearFormIntegrator*> &intFaceIntegrators = *a->GetFBFI();
@@ -453,24 +422,17 @@ void PABilinearFormExtension::Mult(const Vector &x, Vector &y) const
       }
    }
 
-   Array<BilinearFormIntegrator*> &bdr_integs = *a->GetBBFI();
-   Array<BilinearFormIntegrator*> &bdr_face_integs = *a->GetBFBFI();
-   const int n_bdr_integs = bdr_integs.Size();
-   const int n_bdr_face_integs = bdr_face_integs.Size();
-   const bool has_bdr_integs = (n_bdr_face_integs > 0 || n_bdr_integs > 0);
-   if (bdr_face_restrict_lex && has_bdr_integs)
+   Array<BilinearFormIntegrator*> &bdrFaceIntegrators = *a->GetBFBFI();
+   const int bFISz = bdrFaceIntegrators.Size();
+   if (bdr_face_restrict_lex && bFISz>0)
    {
       bdr_face_restrict_lex->Mult(x, bdr_face_X);
       if (bdr_face_X.Size()>0)
       {
          bdr_face_Y = 0.0;
-         for (int i = 0; i < n_bdr_integs; ++i)
+         for (int i = 0; i < bFISz; ++i)
          {
-            bdr_integs[i]->AddMultPA(bdr_face_X, bdr_face_Y);
-         }
-         for (int i = 0; i < n_bdr_face_integs; ++i)
-         {
-            bdr_face_integs[i]->AddMultPA(bdr_face_X, bdr_face_Y);
+            bdrFaceIntegrators[i]->AddMultPA(bdr_face_X, bdr_face_Y);
          }
          bdr_face_restrict_lex->AddMultTransposeInPlace(bdr_face_Y, y);
       }
@@ -634,7 +596,7 @@ void EABilinearFormExtension::Mult(const Vector &x, Vector &y) const
       auto X = Reshape(useRestrict?localX.Read():x.Read(), NDOFS, ne);
       auto Y = Reshape(useRestrict?localY.ReadWrite():y.ReadWrite(), NDOFS, ne);
       auto A = Reshape(ea_data.Read(), NDOFS, NDOFS, ne);
-      mfem::forall(ne*NDOFS, [=] MFEM_HOST_DEVICE (int glob_j)
+      MFEM_FORALL(glob_j, ne*NDOFS,
       {
          const int e = glob_j/NDOFS;
          const int j = glob_j%NDOFS;
@@ -669,7 +631,7 @@ void EABilinearFormExtension::Mult(const Vector &x, Vector &y) const
          if (!factorize_face_terms)
          {
             auto A_int = Reshape(ea_data_int.Read(), NDOFS, NDOFS, 2, nf_int);
-            mfem::forall(nf_int*NDOFS, [=] MFEM_HOST_DEVICE (int glob_j)
+            MFEM_FORALL(glob_j, nf_int*NDOFS,
             {
                const int f = glob_j/NDOFS;
                const int j = glob_j%NDOFS;
@@ -688,7 +650,7 @@ void EABilinearFormExtension::Mult(const Vector &x, Vector &y) const
             });
          }
          auto A_ext = Reshape(ea_data_ext.Read(), NDOFS, NDOFS, 2, nf_int);
-         mfem::forall(nf_int*NDOFS, [=] MFEM_HOST_DEVICE (int glob_j)
+         MFEM_FORALL(glob_j, nf_int*NDOFS,
          {
             const int f = glob_j/NDOFS;
             const int j = glob_j%NDOFS;
@@ -725,7 +687,7 @@ void EABilinearFormExtension::Mult(const Vector &x, Vector &y) const
          auto X = Reshape(bdr_face_X.Read(), NDOFS, nf_bdr);
          auto Y = Reshape(bdr_face_Y.ReadWrite(), NDOFS, nf_bdr);
          auto A = Reshape(ea_data_bdr.Read(), NDOFS, NDOFS, nf_bdr);
-         mfem::forall(nf_bdr*NDOFS, [=] MFEM_HOST_DEVICE (int glob_j)
+         MFEM_FORALL(glob_j, nf_bdr*NDOFS,
          {
             const int f = glob_j/NDOFS;
             const int j = glob_j%NDOFS;
@@ -762,7 +724,7 @@ void EABilinearFormExtension::MultTranspose(const Vector &x, Vector &y) const
       auto X = Reshape(useRestrict?localX.Read():x.Read(), NDOFS, ne);
       auto Y = Reshape(useRestrict?localY.ReadWrite():y.ReadWrite(), NDOFS, ne);
       auto A = Reshape(ea_data.Read(), NDOFS, NDOFS, ne);
-      mfem::forall(ne*NDOFS, [=] MFEM_HOST_DEVICE (int glob_j)
+      MFEM_FORALL(glob_j, ne*NDOFS,
       {
          const int e = glob_j/NDOFS;
          const int j = glob_j%NDOFS;
@@ -797,7 +759,7 @@ void EABilinearFormExtension::MultTranspose(const Vector &x, Vector &y) const
          if (!factorize_face_terms)
          {
             auto A_int = Reshape(ea_data_int.Read(), NDOFS, NDOFS, 2, nf_int);
-            mfem::forall(nf_int*NDOFS, [=] MFEM_HOST_DEVICE (int glob_j)
+            MFEM_FORALL(glob_j, nf_int*NDOFS,
             {
                const int f = glob_j/NDOFS;
                const int j = glob_j%NDOFS;
@@ -816,7 +778,7 @@ void EABilinearFormExtension::MultTranspose(const Vector &x, Vector &y) const
             });
          }
          auto A_ext = Reshape(ea_data_ext.Read(), NDOFS, NDOFS, 2, nf_int);
-         mfem::forall(nf_int*NDOFS, [=] MFEM_HOST_DEVICE (int glob_j)
+         MFEM_FORALL(glob_j, nf_int*NDOFS,
          {
             const int f = glob_j/NDOFS;
             const int j = glob_j%NDOFS;
@@ -853,7 +815,7 @@ void EABilinearFormExtension::MultTranspose(const Vector &x, Vector &y) const
          auto X = Reshape(bdr_face_X.Read(), NDOFS, nf_bdr);
          auto Y = Reshape(bdr_face_Y.ReadWrite(), NDOFS, nf_bdr);
          auto A = Reshape(ea_data_bdr.Read(), NDOFS, NDOFS, nf_bdr);
-         mfem::forall(nf_bdr*NDOFS, [=] MFEM_HOST_DEVICE (int glob_j)
+         MFEM_FORALL(glob_j, nf_bdr*NDOFS,
          {
             const int f = glob_j/NDOFS;
             const int j = glob_j%NDOFS;
@@ -1068,13 +1030,13 @@ void FABilinearFormExtension::DGMult(const Vector &x, Vector &y) const
       const int local_size = a->FESpace()->GetVSize();
       auto dg_x_ptr = dg_x.Write();
       auto x_ptr = x.Read();
-      mfem::forall(local_size, [=] MFEM_HOST_DEVICE (int i)
+      MFEM_FORALL(i,local_size,
       {
          dg_x_ptr[i] = x_ptr[i];
       });
       const int shared_size = shared_x.Size();
       auto shared_x_ptr = shared_x.Read();
-      mfem::forall(shared_size, [=] MFEM_HOST_DEVICE (int i)
+      MFEM_FORALL(i,shared_size,
       {
          dg_x_ptr[local_size+i] = shared_x_ptr[i];
       });
@@ -1085,7 +1047,7 @@ void FABilinearFormExtension::DGMult(const Vector &x, Vector &y) const
          // DG Restriction
          auto dg_y_ptr = dg_y.Read();
          auto y_ptr = y.ReadWrite();
-         mfem::forall(local_size, [=] MFEM_HOST_DEVICE (int i)
+         MFEM_FORALL(i,local_size,
          {
             y_ptr[i] += dg_y_ptr[i];
          });
@@ -1129,13 +1091,13 @@ void FABilinearFormExtension::DGMultTranspose(const Vector &x, Vector &y) const
       const int local_size = a->FESpace()->GetVSize();
       auto dg_x_ptr = dg_x.Write();
       auto x_ptr = x.Read();
-      mfem::forall(local_size, [=] MFEM_HOST_DEVICE (int i)
+      MFEM_FORALL(i,local_size,
       {
          dg_x_ptr[i] = x_ptr[i];
       });
       const int shared_size = shared_x.Size();
       auto shared_x_ptr = shared_x.Read();
-      mfem::forall(shared_size, [=] MFEM_HOST_DEVICE (int i)
+      MFEM_FORALL(i,shared_size,
       {
          dg_x_ptr[local_size+i] = shared_x_ptr[i];
       });
@@ -1146,7 +1108,7 @@ void FABilinearFormExtension::DGMultTranspose(const Vector &x, Vector &y) const
          // DG Restriction
          auto dg_y_ptr = dg_y.Read();
          auto y_ptr = y.ReadWrite();
-         mfem::forall(local_size, [=] MFEM_HOST_DEVICE (int i)
+         MFEM_FORALL(i,local_size,
          {
             y_ptr[i] += dg_y_ptr[i];
          });
@@ -1484,7 +1446,7 @@ void PADiscreteLinearOperatorExtension::Assemble()
    }
 
    auto tm = test_multiplicity.ReadWrite();
-   mfem::forall(test_multiplicity.Size(), [=] MFEM_HOST_DEVICE (int i)
+   MFEM_FORALL(i, test_multiplicity.Size(),
    {
       tm[i] = 1.0 / tm[i];
    });
@@ -1536,7 +1498,7 @@ void PADiscreteLinearOperatorExtension::AddMultTranspose(
    MFEM_VERIFY(x.Size() == test_multiplicity.Size(), "Input vector of wrong size");
    auto xs = xscaled.ReadWrite();
    auto tm = test_multiplicity.Read();
-   mfem::forall(x.Size(), [=] MFEM_HOST_DEVICE (int i)
+   MFEM_FORALL(i, x.Size(),
    {
       xs[i] *= tm[i];
    });

fem/bilininteg.cpp

@@ -22,47 +22,41 @@ namespace mfem
 
 void BilinearFormIntegrator::AssemblePA(const FiniteElementSpace&)
 {
-   MFEM_ABORT("BilinearFormIntegrator::AssemblePA(fes)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AssemblePA(fes)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssemblePA(const FiniteElementSpace&,
                                         const FiniteElementSpace&)
 {
-   MFEM_ABORT("BilinearFormIntegrator::AssemblePA(fes, fes)\n"
-              "   is not implemented for this class.");
-}
-
-void BilinearFormIntegrator::AssemblePABoundary(const FiniteElementSpace&)
-{
-   MFEM_ABORT("BilinearFormIntegrator::AssemblePABoundary(fes)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AssemblePA(fes, fes)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssemblePAInteriorFaces(const FiniteElementSpace&)
 {
-   MFEM_ABORT("BilinearFormIntegrator::AssemblePAInteriorFaces(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AssemblePAInteriorFaces(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssemblePABoundaryFaces(const FiniteElementSpace&)
 {
-   MFEM_ABORT("BilinearFormIntegrator::AssemblePABoundaryFaces(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AssemblePABoundaryFaces(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssembleDiagonalPA(Vector &)
 {
-   MFEM_ABORT("BilinearFormIntegrator::AssembleDiagonalPA(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AssembleDiagonalPA(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssembleEA(const FiniteElementSpace &fes,
                                         Vector &emat,
                                         const bool add)
 {
-   MFEM_ABORT("BilinearFormIntegrator::AssembleEA(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AssembleEA(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssembleEAInteriorFaces(const FiniteElementSpace
@@ -71,8 +65,8 @@ void BilinearFormIntegrator::AssembleEAInteriorFaces(const FiniteElementSpace
                                                      Vector &ea_data_ext,
                                                      const bool add)
 {
-   MFEM_ABORT("BilinearFormIntegrator::AssembleEAInteriorFaces(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AssembleEAInteriorFaces(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssembleEABoundaryFaces(const FiniteElementSpace
@@ -80,8 +74,8 @@ void BilinearFormIntegrator::AssembleEABoundaryFaces(const FiniteElementSpace
                                                      Vector &ea_data_bdr,
                                                      const bool add)
 {
-   MFEM_ABORT("BilinearFormIntegrator::AssembleEABoundaryFaces(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AssembleEABoundaryFaces(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssembleDiagonalPA_ADAt(const Vector &, Vector &)
@@ -92,62 +86,62 @@ void BilinearFormIntegrator::AssembleDiagonalPA_ADAt(const Vector &, Vector &)
 
 void BilinearFormIntegrator::AddMultPA(const Vector &, Vector &) const
 {
-   MFEM_ABORT("BilinearFormIntegrator::MultAssembled(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::MultAssembled(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AddMultTransposePA(const Vector &, Vector &) const
 {
-   MFEM_ABORT("BilinearFormIntegrator::AddMultTransposePA(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AddMultTransposePA(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssembleMF(const FiniteElementSpace &fes)
 {
-   MFEM_ABORT("BilinearFormIntegrator::AssembleMF(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AssembleMF(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AddMultMF(const Vector &, Vector &) const
 {
-   MFEM_ABORT("BilinearFormIntegrator::AddMultMF(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AddMultMF(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AddMultTransposeMF(const Vector &, Vector &) const
 {
-   MFEM_ABORT("BilinearFormIntegrator::AddMultTransposeMF(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AddMultTransposeMF(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssembleDiagonalMF(Vector &)
 {
-   MFEM_ABORT("BilinearFormIntegrator::AssembleDiagonalMF(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AssembleDiagonalMF(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssembleElementMatrix (
    const FiniteElement &el, ElementTransformation &Trans,
    DenseMatrix &elmat )
 {
-   MFEM_ABORT("BilinearFormIntegrator::AssembleElementMatrix(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AssembleElementMatrix(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssembleElementMatrix2 (
    const FiniteElement &el1, const FiniteElement &el2,
    ElementTransformation &Trans, DenseMatrix &elmat )
 {
-   MFEM_ABORT("BilinearFormIntegrator::AssembleElementMatrix2(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AssembleElementMatrix2(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssembleFaceMatrix (
    const FiniteElement &el1, const FiniteElement &el2,
    FaceElementTransformations &Trans, DenseMatrix &elmat)
 {
-   MFEM_ABORT("BilinearFormIntegrator::AssembleFaceMatrix(...)\n"
-              "   is not implemented for this class.");
+   mfem_error ("BilinearFormIntegrator::AssembleFaceMatrix(...)\n"
+               "   is not implemented for this class.");
 }
 
 void BilinearFormIntegrator::AssembleFaceMatrix(
@@ -159,16 +153,6 @@ void BilinearFormIntegrator::AssembleFaceMatrix(
               " Integrator class.");
 }
 
-void BilinearFormIntegrator::AssembleTraceFaceMatrix (int elem,
-                                                      const FiniteElement &trial_face_fe,
-                                                      const FiniteElement &test_fe1,
-                                                      FaceElementTransformations &Trans,
-                                                      DenseMatrix &elmat)
-{
-   MFEM_ABORT("AssembleTraceFaceMatrix (DPG form) is not implemented for this"
-              " Integrator class.");
-}
-
 void BilinearFormIntegrator::AssembleElementVector(
    const FiniteElement &el, ElementTransformation &Tr, const Vector &elfun,
    Vector &elvect)
@@ -2649,7 +2633,7 @@ void VectorFEMassIntegrator::AssembleElementMatrix2(
    }
    else
    {
-      MFEM_ABORT("VectorFEMassIntegrator::AssembleElementMatrix2(...)\n"
+      mfem_error("VectorFEMassIntegrator::AssembleElementMatrix2(...)\n"
                  "   is not implemented for given trial and test bases.");
    }
 }
@@ -4013,216 +3997,6 @@ void NormalTraceJumpIntegrator::AssembleFaceMatrix(
    }
 }
 
-void TraceIntegrator::AssembleTraceFaceMatrix(int elem,
-                                              const FiniteElement &trial_face_fe,
-                                              const FiniteElement &test_fe,
-                                              FaceElementTransformations & Trans,
-                                              DenseMatrix &elmat)
-{
-   MFEM_VERIFY(test_fe.GetMapType() == FiniteElement::VALUE,
-               "TraceIntegrator::AssembleTraceFaceMatrix: Test space should be H1");
-   MFEM_VERIFY(trial_face_fe.GetMapType() == FiniteElement::INTEGRAL,
-               "TraceIntegrator::AssembleTraceFaceMatrix: Trial space should be RT trace");
-
-   int i, j, face_ndof, ndof;
-   int order;
-
-   face_ndof = trial_face_fe.GetDof();
-   ndof = test_fe.GetDof();
-
-   face_shape.SetSize(face_ndof);
-   shape.SetSize(ndof);
-
-   elmat.SetSize(ndof, face_ndof);
-   elmat = 0.0;
-
-   const IntegrationRule *ir = IntRule;
-   if (ir == NULL)
-   {
-      order = test_fe.GetOrder();
-      order += trial_face_fe.GetOrder();
-      ir = &IntRules.Get(Trans.GetGeometryType(), order);
-   }
-
-   int iel = Trans.Elem1->ElementNo;
-   if (iel != elem)
-   {
-      MFEM_VERIFY(elem == Trans.Elem2->ElementNo, "Elem != Trans.Elem2->ElementNo");
-   }
-
-   double scale = 1.0;
-   if (iel != elem) { scale = -1.; }
-   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);
-      // Trace finite element shape function
-      trial_face_fe.CalcPhysShape(Trans,face_shape);
-
-      // Finite element shape function
-      ElementTransformation * eltrans = (iel == elem) ? Trans.Elem1 : Trans.Elem2;
-      test_fe.CalcPhysShape(*eltrans, shape);
-
-      face_shape *= Trans.Weight()*ip.weight*scale;
-      for (i = 0; i < ndof; i++)
-      {
-         for (j = 0; j < face_ndof; j++)
-         {
-            elmat(i, j) += shape(i) * face_shape(j);
-         }
-      }
-   }
-}
-
-void NormalTraceIntegrator::AssembleTraceFaceMatrix(int elem,
-                                                    const FiniteElement &trial_face_fe,
-                                                    const FiniteElement &test_fe,
-                                                    FaceElementTransformations &Trans,
-                                                    DenseMatrix &elmat)
-{
-   int i, j, face_ndof, ndof, dim;
-   int order;
-
-   MFEM_VERIFY(test_fe.GetMapType() == FiniteElement::H_DIV,
-               "NormalTraceIntegrator::AssembleTraceFaceMatrix: Test space should be RT");
-   MFEM_VERIFY(trial_face_fe.GetMapType() == FiniteElement::VALUE,
-               "NormalTraceIntegrator::AssembleTraceFaceMatrix: Trial space should be H1 (trace)");
-
-   face_ndof = trial_face_fe.GetDof();
-   ndof = test_fe.GetDof();
-   dim = test_fe.GetDim();
-
-   face_shape.SetSize(face_ndof);
-   normal.SetSize(dim);
-   shape.SetSize(ndof,dim);
-   shape_n.SetSize(ndof);
-
-   elmat.SetSize(ndof, face_ndof);
-   elmat = 0.0;
-
-   const IntegrationRule *ir = IntRule;
-   if (ir == NULL)
-   {
-      order = test_fe.GetOrder();
-      order += trial_face_fe.GetOrder();
-      ir = &IntRules.Get(Trans.GetGeometryType(), order);
-   }
-
-   int iel = Trans.Elem1->ElementNo;
-   if (iel != elem)
-   {
-      MFEM_VERIFY(elem == Trans.Elem2->ElementNo, "Elem != Trans.Elem2->ElementNo");
-   }
-
-   double scale = 1.0;
-   if (iel != elem) { scale = -1.; }
-
-   for (int p = 0; p < ir->GetNPoints(); p++)
-   {
-      const IntegrationPoint &ip = ir->IntPoint(p);
-      Trans.SetAllIntPoints(&ip);
-      trial_face_fe.CalcPhysShape(Trans, face_shape);
-      CalcOrtho(Trans.Jacobian(),normal);
-      ElementTransformation * etrans = (iel == elem) ? Trans.Elem1 : Trans.Elem2;
-      test_fe.CalcVShape(*etrans, shape);
-      shape.Mult(normal, shape_n);
-      face_shape *= ip.weight*scale;
-
-      for (i = 0; i < ndof; i++)
-      {
-         for (j = 0; j < face_ndof; j++)
-         {
-            elmat(i, j) += shape_n(i) * face_shape(j);
-         }
-      }
-   }
-}
-
-void TangentTraceIntegrator::AssembleTraceFaceMatrix(int elem,
-                                                     const FiniteElement &trial_face_fe,
-                                                     const FiniteElement &test_fe,
-                                                     FaceElementTransformations & Trans,
-                                                     DenseMatrix &elmat)
-{
-
-   MFEM_VERIFY(test_fe.GetMapType() == FiniteElement::H_CURL,
-               "TangentTraceIntegrator::AssembleTraceFaceMatrix: Test space should be ND");
-
-   int face_ndof, ndof, dim;
-   int order;
-   dim = test_fe.GetDim();
-   if (dim == 3)
-   {
-      std::string msg =
-         "Trial space should be ND face trace and test space should be a ND vector field in 3D ";
-      MFEM_VERIFY(trial_face_fe.GetMapType() == FiniteElement::H_CURL &&
-                  trial_face_fe.GetDim() == 2 && test_fe.GetDim() == 3, msg);
-   }
-   else
-   {
-      std::string msg =
-         "Trial space should be H1 edge trace and test space should be a ND vector field in 2D";
-      MFEM_VERIFY(trial_face_fe.GetMapType() == FiniteElement::VALUE &&
-                  trial_face_fe.GetDim() == 1 && test_fe.GetDim() == 2, msg);
-   }
-   face_ndof = trial_face_fe.GetDof();
-   ndof = test_fe.GetDof();
-
-   int dimc = (dim == 3) ? 3 : 1;
-
-   face_shape.SetSize(face_ndof,dimc);
-   shape_n.SetSize(ndof,dimc);
-   shape.SetSize(ndof,dim);
-   normal.SetSize(dim);
-   DenseMatrix face_shape_n(face_ndof,dimc);
-
-   elmat.SetSize(ndof, face_ndof);
-   elmat = 0.0;
-
-   const IntegrationRule *ir = IntRule;
-   if (ir == NULL)
-   {
-      order = test_fe.GetOrder();
-      order += trial_face_fe.GetOrder();
-      ir = &IntRules.Get(Trans.GetGeometryType(), order);
-   }
-
-   int iel = Trans.Elem1->ElementNo;
-   if (iel != elem)
-   {
-      MFEM_VERIFY(elem == Trans.Elem2->ElementNo, "Elem != Trans.Elem2->ElementNo");
-   }
-
-   double scale = 1.0;
-   if (iel != elem) { scale = -1.; }
-   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);
-      // Trace finite element shape function
-      if (dim == 3)
-      {
-         trial_face_fe.CalcVShape(Trans,face_shape);
-      }
-      else
-      {
-         face_shape.GetColumnReference(0,temp);
-         trial_face_fe.CalcPhysShape(Trans,temp);
-      }
-      CalcOrtho(Trans.Jacobian(),normal);
-      ElementTransformation * eltrans = (iel == elem) ? Trans.Elem1 : Trans.Elem2;
-      test_fe.CalcVShape(*eltrans, shape);
-
-      // rotate
-      cross_product(normal, shape, shape_n);
-
-      const double w = scale*ip.weight;
-      AddMult_a_ABt(w,shape_n, face_shape, elmat);
-   }
-}
 
 void NormalInterpolator::AssembleElementMatrix2(
    const FiniteElement &dom_fe, const FiniteElement &ran_fe,

fem/bilininteg.hpp

@@ -61,8 +61,6 @@ public:
    virtual void AssemblePA(const FiniteElementSpace &trial_fes,
                            const FiniteElementSpace &test_fes);
 
-   virtual void AssemblePABoundary(const FiniteElementSpace &fes);
-
    virtual void AssemblePAInteriorFaces(const FiniteElementSpace &fes);
 
    virtual void AssemblePABoundaryFaces(const FiniteElementSpace &fes);
@@ -161,15 +159,6 @@ public:
                                    FaceElementTransformations &Trans,
                                    DenseMatrix &elmat);
 
-   /** Abstract method used for assembling TraceFaceIntegrators for
-       DPG weak formulations. */
-   virtual void AssembleTraceFaceMatrix(int elem,
-                                        const FiniteElement &trial_face_fe,
-                                        const FiniteElement &test_fe,
-                                        FaceElementTransformations &Trans,
-                                        DenseMatrix &elmat);
-
-
    /// @brief Perform the local action of the BilinearFormIntegrator.
    /// Note that the default implementation in the base class is general but not
    /// efficient.
@@ -303,12 +292,6 @@ public:
       bfi->AssemblePA(fes);
    }
 
-   virtual void AssemblePA(const FiniteElementSpace &trial_fes,
-                           const FiniteElementSpace &test_fes)
-   {
-      bfi->AssemblePA(test_fes, trial_fes); // Reverse test and trial
-   }
-
    virtual void AssemblePAInteriorFaces(const FiniteElementSpace &fes)
    {
       bfi->AssemblePAInteriorFaces(fes);
@@ -2200,9 +2183,8 @@ protected:
    // PA extension
    const FiniteElementSpace *fespace;
    Vector pa_data;
-   const DofToQuad *maps;                 ///< Not owned
-   const GeometricFactors *geom;          ///< Not owned
-   const FaceGeometricFactors *face_geom; ///< Not owned
+   const DofToQuad *maps;         ///< Not owned
+   const GeometricFactors *geom;  ///< Not owned
    int dim, ne, nq, dofs1D, quad1D;
 
 public:
@@ -2229,8 +2211,6 @@ public:
 
    virtual void AssemblePA(const FiniteElementSpace &fes);
 
-   virtual void AssemblePABoundary(const FiniteElementSpace &fes);
-
    virtual void AssembleEA(const FiniteElementSpace &fes, Vector &emat,
                            const bool add);
 
@@ -3321,87 +3301,6 @@ public:
                                    DenseMatrix &elmat);
 };
 
-/** Integrator for the DPG form: < v, w > over a face (the interface) where
-    the trial variable v is defined on the interface
-    (H^-1/2 i.e., v:=u⋅n normal trace of H(div))
-    and the test variable w is in an H1-conforming space. */
-class TraceIntegrator : public BilinearFormIntegrator
-{
-private:
-   Vector face_shape, shape;
-public:
-   TraceIntegrator() { }
-   void AssembleTraceFaceMatrix(int elem,
-                                const FiniteElement &trial_face_fe,
-                                const FiniteElement &test_fe,
-                                FaceElementTransformations &Trans,
-                                DenseMatrix &elmat);
-};
-
-/** Integrator for the form: < v, w.n > over a face (the interface) where
-    the trial variable v is defined on the interface (H^1/2, i.e., trace of H1)
-    and the test variable w is in an H(div)-conforming space. */
-class NormalTraceIntegrator : public BilinearFormIntegrator
-{
-private:
-   Vector face_shape, normal, shape_n;
-   DenseMatrix shape;
-
-public:
-   NormalTraceIntegrator() { }
-   virtual void AssembleTraceFaceMatrix(int ielem,
-                                        const FiniteElement &trial_face_fe,
-                                        const FiniteElement &test_fe,
-                                        FaceElementTransformations &Trans,
-                                        DenseMatrix &elmat);
-};
-
-
-/** Integrator for the form: < v, w × n > over a face (the interface)
- *  In 3D the trial variable v is defined on the interface (H^-1/2(curl), trace of H(curl))
- *  In 2D it's defined on the interface (H^1/2, trace of H1)
- *  The test variable w is in an H(curl)-conforming space. */
-class TangentTraceIntegrator : public BilinearFormIntegrator
-{
-private:
-   DenseMatrix face_shape, shape, shape_n;
-   Vector normal;
-   Vector temp;
-
-   void cross_product(const Vector & x, const DenseMatrix & Y, DenseMatrix & Z)
-   {
-      int dim = x.Size();
-      MFEM_VERIFY(Y.Width() == dim, "Size missmatch");
-      int dimc = dim == 3 ? dim : 1;
-      int h = Y.Height();
-      Z.SetSize(h,dimc);
-      if (dim == 3)
-      {
-         for (int i = 0; i<h; i++)
-         {
-            Z(i,0) = x(2) * Y(i,1) - x(1) * Y(i,2);
-            Z(i,1) = x(0) * Y(i,2) - x(2) * Y(i,0);
-            Z(i,2) = x(1) * Y(i,0) - x(0) * Y(i,1);
-         }
-      }
-      else
-      {
-         for (int i = 0; i<h; i++)
-         {
-            Z(i,0) = x(1) * Y(i,0) - x(0) * Y(i,1);
-         }
-      }
-   }
-
-public:
-   TangentTraceIntegrator() { }
-   void AssembleTraceFaceMatrix(int elem,
-                                const FiniteElement &trial_face_fe,
-                                const FiniteElement &test_fe,
-                                FaceElementTransformations &Trans,
-                                DenseMatrix &elmat);
-};
-
 /** Abstract class to serve as a base for local interpolators to be used in the
     DiscreteLinearOperator class. */
 class DiscreteInterpolator : public BilinearFormIntegrator { };
@@ -3437,7 +3336,7 @@ public:
 
 private:
    /// 1D finite element that generates and owns the 1D DofToQuad maps below
-   FiniteElement *dofquad_fe;
+   FiniteElement * dofquad_fe;
 
    bool B_id; // is the B basis operator (maps_C_C) the identity?
    const DofToQuad *maps_C_C; // one-d map with Lobatto rows, Lobatto columns
@@ -3452,8 +3351,6 @@ private:
 class IdentityInterpolator : public DiscreteInterpolator
 {
 public:
-   IdentityInterpolator(): dofquad_fe(NULL) { }
-
    virtual void AssembleElementMatrix2(const FiniteElement &dom_fe,
                                        const FiniteElement &ran_fe,
                                        ElementTransformation &Trans,
@@ -3468,11 +3365,9 @@ public:
    virtual void AddMultPA(const Vector &x, Vector &y) const;
    virtual void AddMultTransposePA(const Vector &x, Vector &y) const;
 
-   virtual ~IdentityInterpolator() { delete dofquad_fe; }
-
 private:
    /// 1D finite element that generates and owns the 1D DofToQuad maps below
-   FiniteElement *dofquad_fe;
+   FiniteElement * dofquad_fe;
 
    const DofToQuad *maps_C_C; // one-d map with Lobatto rows, Lobatto columns
    const DofToQuad *maps_O_C; // one-d map with Legendre rows, Lobatto columns
@@ -3627,5 +3522,17 @@ protected:
    VectorCoefficient *VQ;
 };
 
+
+
+// PA Diffusion Assemble 2D kernel
+template<const int T_SDIM>
+void PADiffusionSetup2D(const int Q1D,
+                        const int coeffDim,
+                        const int NE,
+                        const Array<double> &w,
+                        const Vector &j,
+                        const Vector &c,
+                        Vector &d);
+
 }
 #endif

fem/bilininteg_br2.cpp

@@ -9,8 +9,8 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../bilininteg.hpp"
-#include "../pfespace.hpp"
+#include "bilininteg.hpp"
+#include "pfespace.hpp"
 #include <algorithm>
 
 namespace mfem

fem/bilininteg_convection_ea.cpp

@@ -9,9 +9,9 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../../general/forall.hpp"
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
 
 namespace mfem
 {
@@ -34,7 +34,7 @@ static void EAConvectionAssemble1D(const int NE,
    auto G = Reshape(g.Read(), Q1D, D1D);
    auto D = Reshape(padata.Read(), Q1D, NE);
    auto A = Reshape(eadata.ReadWrite(), D1D, D1D, NE);
-   mfem::forall_2D(NE, D1D, D1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, D1D, D1D, 1,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -86,7 +86,7 @@ static void EAConvectionAssemble2D(const int NE,
    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);
-   mfem::forall_2D(NE, D1D, D1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, D1D, D1D, 1,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -163,7 +163,7 @@ static void EAConvectionAssemble3D(const int NE,
    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);
-   mfem::forall_3D(NE, D1D, D1D, D1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, D1D, D1D, D1D,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;

fem/bilininteg_convection_mf.cpp

@@ -9,9 +9,12 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
-#include "../ceed/integrators/convection/convection.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
+#include "ceed/integrators/convection/convection.hpp"
+
+using namespace std;
 
 namespace mfem
 {

fem/bilininteg_convection_pa.cpp

@@ -9,15 +9,18 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../../general/forall.hpp"
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
-#include "../qfunction.hpp"
-#include "../ceed/integrators/convection/convection.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
+#include "qfunction.hpp"
+#include "ceed/integrators/convection/convection.hpp"
+#include "quadinterpolator.hpp"
 
 namespace mfem
 {
 
+// PA Convection Integrator
+
 // PA Convection Assemble 2D kernel
 static void PAConvectionSetup2D(const int NQ,
                                 const int NE,
@@ -38,7 +41,7 @@ static void PAConvectionSetup2D(const int NQ,
                   Reshape(vel.Read(), DIM,NQ,NE);
    auto y = Reshape(op.Write(), NQ,DIM,NE);
 
-   mfem::forall(NE*NQ, [=] MFEM_HOST_DEVICE (int q_global)
+   MFEM_FORALL(q_global, NE*NQ,
    {
       const int e = q_global / NQ;
       const int q = q_global % NQ;
@@ -75,7 +78,7 @@ static void PAConvectionSetup3D(const int NQ,
                   Reshape(vel.Read(), 3,1,1) :
                   Reshape(vel.Read(), 3,NQ,NE);
    auto y = Reshape(op.Write(), NQ,3,NE);
-   mfem::forall(NE*NQ, [=] MFEM_HOST_DEVICE (int q_global)
+   MFEM_FORALL(q_global, NE*NQ,
    {
       const int e = q_global / NQ;
       const int q = q_global % NQ;
@@ -132,61 +135,6 @@ static void PAConvectionSetup(const int dim,
    }
 }
 
-void ConvectionIntegrator::AssemblePA(const FiniteElementSpace &fes)
-{
-   const MemoryType mt = (pa_mt == MemoryType::DEFAULT) ?
-                         Device::GetDeviceMemoryType() : pa_mt;
-   // Assumes tensor-product elements
-   Mesh *mesh = fes.GetMesh();
-   const FiniteElement &el = *fes.GetFE(0);
-   ElementTransformation &Trans = *fes.GetElementTransformation(0);
-   const IntegrationRule *ir = IntRule ? IntRule : &GetRule(el, Trans);
-   if (DeviceCanUseCeed())
-   {
-      delete ceedOp;
-      const bool mixed = mesh->GetNumGeometries(mesh->Dimension()) > 1 ||
-                         fes.IsVariableOrder();
-      if (mixed)
-      {
-         ceedOp = new ceed::MixedPAConvectionIntegrator(*this, fes, Q, alpha);
-      }
-      else
-      {
-         ceedOp = new ceed::PAConvectionIntegrator(fes, *ir, Q, alpha);
-      }
-      return;
-   }
-   const int dims = el.GetDim();
-   const int symmDims = dims;
-   nq = ir->GetNPoints();
-   dim = mesh->Dimension();
-   ne = fes.GetNE();
-   geom = mesh->GetGeometricFactors(*ir, GeometricFactors::JACOBIANS, mt);
-   maps = &el.GetDofToQuad(*ir, DofToQuad::TENSOR);
-   dofs1D = maps->ndof;
-   quad1D = maps->nqpt;
-   pa_data.SetSize(symmDims * nq * ne, mt);
-
-   QuadratureSpace qs(*mesh, *ir);
-   CoefficientVector vel(*Q, qs, CoefficientStorage::COMPRESSED);
-
-   PAConvectionSetup(dim, nq, ne, ir->GetWeights(), geom->J,
-                     vel, alpha, pa_data);
-}
-
-void ConvectionIntegrator::AssembleDiagonalPA(Vector &diag)
-{
-   if (DeviceCanUseCeed())
-   {
-      ceedOp->GetDiagonal(diag);
-   }
-   else
-   {
-      MFEM_ABORT("AssembleDiagonalPA not yet implemented for"
-                 " ConvectionIntegrator.");
-   }
-}
-
 // PA Convection Apply 2D kernel
 template<int T_D1D = 0, int T_Q1D = 0> static
 void PAConvectionApply2D(const int ne,
@@ -211,7 +159,7 @@ void PAConvectionApply2D(const int ne,
    auto op = Reshape(op_.Read(), Q1D, Q1D, 2, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, NE);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -331,7 +279,7 @@ void SmemPAConvectionApply2D(const int ne,
    auto op = Reshape(op_.Read(), Q1D, Q1D, 2, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, NE);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, NE);
-   mfem::forall_2D_batch(NE, Q1D, Q1D, NBZ, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_2D(e, NE, Q1D, Q1D, NBZ,
    {
       const int tidz = MFEM_THREAD_ID(z);
       const int D1D = T_D1D ? T_D1D : d1d;
@@ -458,7 +406,7 @@ void PAConvectionApply3D(const int ne,
    auto op = Reshape(op_.Read(), Q1D, Q1D, Q1D, 3, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, D1D, NE);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -639,7 +587,7 @@ void SmemPAConvectionApply3D(const int ne,
    auto op = Reshape(op_.Read(), Q1D, Q1D, Q1D, 3, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, D1D, NE);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, NE);
-   mfem::forall_3D(NE, Q1D, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, Q1D, Q1D, Q1D,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -843,7 +791,7 @@ void PAConvectionApplyT2D(const int ne,
    auto op = Reshape(op_.Read(), Q1D, Q1D, 2, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, NE);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -959,7 +907,7 @@ void SmemPAConvectionApplyT2D(const int ne,
    auto op = Reshape(op_.Read(), Q1D, Q1D, 2, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, NE);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, NE);
-   mfem::forall_2D_batch(NE, Q1D, Q1D, NBZ, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_2D(e, NE, Q1D, Q1D, NBZ,
    {
       const int tidz = MFEM_THREAD_ID(z);
       const int D1D = T_D1D ? T_D1D : d1d;
@@ -1081,7 +1029,7 @@ void PAConvectionApplyT3D(const int ne,
    auto op = Reshape(op_.Read(), Q1D, Q1D, Q1D, 3, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, D1D, NE);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -1257,7 +1205,7 @@ void SmemPAConvectionApplyT3D(const int ne,
    auto op = Reshape(op_.Read(), Q1D, Q1D, Q1D, 3, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, D1D, NE);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, NE);
-   mfem::forall_3D(NE, Q1D, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, Q1D, Q1D, Q1D,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -1427,6 +1375,48 @@ void SmemPAConvectionApplyT3D(const int ne,
    });
 }
 
+void ConvectionIntegrator::AssemblePA(const FiniteElementSpace &fes)
+{
+   const MemoryType mt = (pa_mt == MemoryType::DEFAULT) ?
+                         Device::GetDeviceMemoryType() : pa_mt;
+   // Assumes tensor-product elements
+   Mesh *mesh = fes.GetMesh();
+   const FiniteElement &el = *fes.GetFE(0);
+   ElementTransformation &Trans = *fes.GetElementTransformation(0);
+   const IntegrationRule *ir = IntRule ? IntRule : &GetRule(el, Trans);
+   if (DeviceCanUseCeed())
+   {
+      delete ceedOp;
+      const bool mixed = mesh->GetNumGeometries(mesh->Dimension()) > 1 ||
+                         fes.IsVariableOrder();
+      if (mixed)
+      {
+         ceedOp = new ceed::MixedPAConvectionIntegrator(*this, fes, Q, alpha);
+      }
+      else
+      {
+         ceedOp = new ceed::PAConvectionIntegrator(fes, *ir, Q, alpha);
+      }
+      return;
+   }
+   const int dims = el.GetDim();
+   const int symmDims = dims;
+   nq = ir->GetNPoints();
+   dim = mesh->Dimension();
+   ne = fes.GetNE();
+   geom = mesh->GetGeometricFactors(*ir, GeometricFactors::JACOBIANS, mt);
+   maps = &el.GetDofToQuad(*ir, DofToQuad::TENSOR);
+   dofs1D = maps->ndof;
+   quad1D = maps->nqpt;
+   pa_data.SetSize(symmDims * nq * ne, mt);
+
+   QuadratureSpace qs(*mesh, *ir);
+   CoefficientVector vel(*Q, qs, CoefficientStorage::COMPRESSED);
+
+   PAConvectionSetup(dim, nq, ne, ir->GetWeights(), geom->J,
+                     vel, alpha, pa_data);
+}
+
 static void PAConvectionApply(const int dim,
                               const int D1D,
                               const int Q1D,
@@ -1531,6 +1521,7 @@ static void PAConvectionApplyT(const int dim,
    MFEM_ABORT("Unknown kernel.");
 }
 
+// PA Convection Apply kernel
 void ConvectionIntegrator::AddMultPA(const Vector &x, Vector &y) const
 {
    if (DeviceCanUseCeed())
@@ -1545,6 +1536,7 @@ void ConvectionIntegrator::AddMultPA(const Vector &x, Vector &y) const
    }
 }
 
+// PA Convection Apply transpose kernel
 void ConvectionIntegrator::AddMultTransposePA(const Vector &x, Vector &y) const
 {
    if (DeviceCanUseCeed())
@@ -1560,4 +1552,17 @@ void ConvectionIntegrator::AddMultTransposePA(const Vector &x, Vector &y) const
    }
 }
 
+void ConvectionIntegrator::AssembleDiagonalPA(Vector &diag)
+{
+   if (DeviceCanUseCeed())
+   {
+      ceedOp->GetDiagonal(diag);
+   }
+   else
+   {
+      MFEM_ABORT("AssembleDiagonalPA not yet implemented for"
+                 " ConvectionIntegrator.");
+   }
+}
+
 } // namespace mfem

fem/bilininteg_dgtrace_ea.cpp

@@ -9,9 +9,9 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../../general/forall.hpp"
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
 
 namespace mfem
 {
@@ -26,7 +26,7 @@ static void EADGTraceAssemble1DInt(const int NF,
    auto D = Reshape(padata.Read(), 2, 2, NF);
    auto A_int = Reshape(eadata_int.ReadWrite(), 2, NF);
    auto A_ext = Reshape(eadata_ext.ReadWrite(), 2, NF);
-   mfem::forall(NF, [=] MFEM_HOST_DEVICE (int f)
+   MFEM_FORALL(f, NF,
    {
       double val_int0, val_int1, val_ext01, val_ext10;
       val_int0  = D(0, 0, f);
@@ -58,7 +58,7 @@ static void EADGTraceAssemble1DBdr(const int NF,
 {
    auto D = Reshape(padata.Read(), 2, 2, NF);
    auto A_bdr = Reshape(eadata_bdr.ReadWrite(), NF);
-   mfem::forall(NF, [=] MFEM_HOST_DEVICE (int f)
+   MFEM_FORALL(f, NF,
    {
       if (add)
       {
@@ -89,7 +89,7 @@ static void EADGTraceAssemble2DInt(const int NF,
    auto D = Reshape(padata.Read(), Q1D, 2, 2, NF);
    auto A_int = Reshape(eadata_int.ReadWrite(), D1D, D1D, 2, NF);
    auto A_ext = Reshape(eadata_ext.ReadWrite(), D1D, D1D, 2, NF);
-   mfem::forall_2D(NF, D1D, D1D, [=] MFEM_HOST_DEVICE (int f)
+   MFEM_FORALL_3D(f, NF, D1D, D1D, 1,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -143,7 +143,7 @@ static void EADGTraceAssemble2DBdr(const int NF,
    auto B = Reshape(basis.Read(), Q1D, D1D);
    auto D = Reshape(padata.Read(), Q1D, 2, 2, NF);
    auto A_bdr = Reshape(eadata_bdr.ReadWrite(), D1D, D1D, NF);
-   mfem::forall_2D(NF, D1D, D1D, [=] MFEM_HOST_DEVICE (int f)
+   MFEM_FORALL_3D(f, NF, D1D, D1D, 1,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -187,7 +187,7 @@ static void EADGTraceAssemble3DInt(const int NF,
    auto D = Reshape(padata.Read(), Q1D, Q1D, 2, 2, NF);
    auto A_int = Reshape(eadata_int.ReadWrite(), D1D, D1D, D1D, D1D, 2, NF);
    auto A_ext = Reshape(eadata_ext.ReadWrite(), D1D, D1D, D1D, D1D, 2, NF);
-   mfem::forall_2D(NF, D1D, D1D, [=] MFEM_HOST_DEVICE (int f)
+   MFEM_FORALL_3D(f, NF, D1D, D1D, 1,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -283,7 +283,7 @@ static void EADGTraceAssemble3DBdr(const int NF,
    auto B = Reshape(basis.Read(), Q1D, D1D);
    auto D = Reshape(padata.Read(), Q1D, Q1D, 2, 2, NF);
    auto A_bdr = Reshape(eadata_bdr.ReadWrite(), D1D, D1D, D1D, D1D, NF);
-   mfem::forall_2D(NF, D1D, D1D, [=] MFEM_HOST_DEVICE (int f)
+   MFEM_FORALL_3D(f, NF, D1D, D1D, 1,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;

fem/bilininteg_dgtrace_pa.cpp

@@ -9,15 +9,16 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../../general/forall.hpp"
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
-#include "../qfunction.hpp"
-#include "../restriction.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
+#include "qfunction.hpp"
+#include "restriction.hpp"
+
+using namespace std;
 
 namespace mfem
 {
-
 // PA DG Trace Integrator
 static void PADGTraceSetup2D(const int Q1D,
                              const int NF,
@@ -43,7 +44,7 @@ static void PADGTraceSetup2D(const int Q1D,
    auto W = w.Read();
    auto qd = Reshape(op.Write(), Q1D, 2, 2, NF);
 
-   mfem::forall(Q1D*NF, [=] MFEM_HOST_DEVICE (int tid)
+   MFEM_FORALL(tid, Q1D*NF,
    {
       const int f = tid / Q1D;
       const int q = tid % Q1D;
@@ -86,7 +87,7 @@ static void PADGTraceSetup3D(const int Q1D,
    auto W = w.Read();
    auto qd = Reshape(op.Write(), Q1D, Q1D, 2, 2, NF);
 
-   mfem::forall(Q1D*Q1D*NF, [=] MFEM_HOST_DEVICE (int tid)
+   MFEM_FORALL(tid, Q1D*Q1D*NF,
    {
       int f = tid / (Q1D * Q1D);
       int q2 = (tid / Q1D) % Q1D;
@@ -98,7 +99,7 @@ static void PADGTraceSetup3D(const int Q1D,
             const double v1 = const_v ? V(1,0,0,0) : V(1,q1,q2,f);
             const double v2 = const_v ? V(2,0,0,0) : V(2,q1,q2,f);
             const double dot = n(q1,q2,0,f) * v0 + n(q1,q2,1,f) * v1 +
-                               n(q1,q2,2,f) * v2;
+            /* */              n(q1,q2,2,f) * v2;
             const double abs = dot > 0.0 ? dot : -dot;
             const double w = W[q1+q2*Q1D]*r*d(q1,q2,f);
             qd(q1,q2,0,0,f) = w*( alpha/2 * dot + beta * abs );
@@ -266,7 +267,7 @@ void PADGTraceApply2D(const int NF,
    auto x = Reshape(x_.Read(), D1D, VDIM, 2, NF);
    auto y = Reshape(y_.ReadWrite(), D1D, VDIM, 2, NF);
 
-   mfem::forall(NF, [=] MFEM_HOST_DEVICE (int f)
+   MFEM_FORALL(f, NF,
    {
       const int VDIM = 1;
       const int D1D = T_D1D ? T_D1D : d1d;
@@ -357,7 +358,7 @@ void PADGTraceApply3D(const int NF,
    auto x = Reshape(x_.Read(), D1D, D1D, VDIM, 2, NF);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, VDIM, 2, NF);
 
-   mfem::forall(NF, [=] MFEM_HOST_DEVICE (int f)
+   MFEM_FORALL(f, NF,
    {
       const int VDIM = 1;
       const int D1D = T_D1D ? T_D1D : d1d;
@@ -502,7 +503,7 @@ void SmemPADGTraceApply3D(const int NF,
    auto x = Reshape(x_.Read(), D1D, D1D, 2, NF);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, 2, NF);
 
-   mfem::forall_2D_batch(NF, Q1D, Q1D, NBZ, [=] MFEM_HOST_DEVICE (int f)
+   MFEM_FORALL_2D(f, NF, Q1D, Q1D, NBZ,
    {
       const int tidz = MFEM_THREAD_ID(z);
       const int D1D = T_D1D ? T_D1D : d1d;
@@ -667,7 +668,7 @@ void PADGTraceApplyTranspose2D(const int NF,
    auto x = Reshape(x_.Read(), D1D, VDIM, 2, NF);
    auto y = Reshape(y_.ReadWrite(), D1D, VDIM, 2, NF);
 
-   mfem::forall(NF, [=] MFEM_HOST_DEVICE (int f)
+   MFEM_FORALL(f, NF,
    {
       const int VDIM = 1;
       const int D1D = T_D1D ? T_D1D : d1d;
@@ -763,7 +764,7 @@ void PADGTraceApplyTranspose3D(const int NF,
    auto x = Reshape(x_.Read(), D1D, D1D, VDIM, 2, NF);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, VDIM, 2, NF);
 
-   mfem::forall(NF, [=] MFEM_HOST_DEVICE (int f)
+   MFEM_FORALL(f, NF,
    {
       const int VDIM = 1;
       const int D1D = T_D1D ? T_D1D : d1d;
@@ -919,7 +920,7 @@ void SmemPADGTraceApplyTranspose3D(const int NF,
    auto x = Reshape(x_.Read(), D1D, D1D, 2, NF);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, 2, NF);
 
-   mfem::forall_2D_batch(NF, Q1D, Q1D, NBZ, [=] MFEM_HOST_DEVICE (int f)
+   MFEM_FORALL_2D(f, NF, Q1D, Q1D, NBZ,
    {
       const int tidz = MFEM_THREAD_ID(z);
       const int D1D = T_D1D ? T_D1D : d1d;

fem/bilininteg_diffusion_ea.cpp

@@ -9,9 +9,9 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../../general/forall.hpp"
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
 
 namespace mfem
 {
@@ -33,7 +33,7 @@ static void EADiffusionAssemble1D(const int NE,
    auto G = Reshape(g.Read(), Q1D, D1D);
    auto D = Reshape(padata.Read(), Q1D, NE);
    auto A = Reshape(eadata.ReadWrite(), D1D, D1D, NE);
-   mfem::forall_2D(NE, D1D, D1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, D1D, D1D, 1,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -85,7 +85,7 @@ static void EADiffusionAssemble2D(const int NE,
    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);
-   mfem::forall_2D(NE, D1D, D1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, D1D, D1D, 1,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -162,7 +162,7 @@ static void EADiffusionAssemble3D(const int NE,
    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);
-   mfem::forall_3D(NE, D1D, D1D, D1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, D1D, D1D, D1D,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;

fem/bilininteg_diffusion_mf.cpp

@@ -9,9 +9,12 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
-#include "../ceed/integrators/diffusion/diffusion.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
+#include "ceed/integrators/diffusion/diffusion.hpp"
+
+using namespace std;
 
 namespace mfem
 {

fem/bilininteg_diffusion_pa.cpp

@@ -9,42 +9,189 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#ifndef MFEM_BILININTEG_DIFFUSION_KERNELS_HPP
-#define MFEM_BILININTEG_DIFFUSION_KERNELS_HPP
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
+#include "qfunction.hpp"
+#include "ceed/integrators/diffusion/diffusion.hpp"
 
-#include "../../config/config.hpp"
-#include "../../general/array.hpp"
-#include "../../general/forall.hpp"
-#include "../../linalg/dtensor.hpp"
-#include "../../linalg/vector.hpp"
-#include "../bilininteg.hpp"
+using namespace std;
 
 namespace mfem
 {
 
-namespace internal
+// PA Diffusion Integrator
+
+// OCCA 2D Assemble kernel
+#ifdef MFEM_USE_OCCA
+static void OccaPADiffusionSetup2D(const int D1D,
+                                   const int Q1D,
+                                   const int NE,
+                                   const Array<double> &W,
+                                   const Vector &J,
+                                   const Vector &C,
+                                   Vector &op)
 {
+   occa::properties props;
+   props["defines/D1D"] = D1D;
+   props["defines/Q1D"] = Q1D;
+   const occa::memory o_W = OccaMemoryRead(W.GetMemory(), W.Size());
+   const occa::memory o_J = OccaMemoryRead(J.GetMemory(), J.Size());
+   const occa::memory o_C = OccaMemoryRead(C.GetMemory(), C.Size());
+   occa::memory o_op = OccaMemoryWrite(op.GetMemory(), op.Size());
+   const bool const_c = C.Size() == 1;
+   const occa_id_t id = std::make_pair(D1D,Q1D);
+   static occa_kernel_t OccaDiffSetup2D_ker;
+   if (OccaDiffSetup2D_ker.find(id) == OccaDiffSetup2D_ker.end())
+   {
+      const occa::kernel DiffusionSetup2D =
+         mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
+                                     "DiffusionSetup2D", props);
+      OccaDiffSetup2D_ker.emplace(id, DiffusionSetup2D);
+   }
+   OccaDiffSetup2D_ker.at(id)(NE, o_W, o_J, o_C, o_op, const_c);
+}
 
-void PADiffusionSetup(const int dim,
-                      const int sdim,
-                      const int D1D,
-                      const int Q1D,
-                      const int coeffDim,
-                      const int NE,
-                      const Array<double> &W,
-                      const Vector &J,
-                      const Vector &C,
-                      Vector &D);
+static void OccaPADiffusionSetup3D(const int D1D,
+                                   const int Q1D,
+                                   const int NE,
+                                   const Array<double> &W,
+                                   const Vector &J,
+                                   const Vector &C,
+                                   Vector &op)
+{
+   occa::properties props;
+   props["defines/D1D"] = D1D;
+   props["defines/Q1D"] = Q1D;
+   const occa::memory o_W = OccaMemoryRead(W.GetMemory(), W.Size());
+   const occa::memory o_J = OccaMemoryRead(J.GetMemory(), J.Size());
+   const occa::memory o_C = OccaMemoryRead(C.GetMemory(), C.Size());
+   occa::memory o_op = OccaMemoryWrite(op.GetMemory(), op.Size());
+   const bool const_c = C.Size() == 1;
+   const occa_id_t id = std::make_pair(D1D,Q1D);
+   static occa_kernel_t OccaDiffSetup3D_ker;
+   if (OccaDiffSetup3D_ker.find(id) == OccaDiffSetup3D_ker.end())
+   {
+      const occa::kernel DiffusionSetup3D =
+         mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
+                                     "DiffusionSetup3D", props);
+      OccaDiffSetup3D_ker.emplace(id, DiffusionSetup3D);
+   }
+   OccaDiffSetup3D_ker.at(id)(NE, o_W, o_J, o_C, o_op, const_c);
+}
+#endif // MFEM_USE_OCCA
 
-// PA Diffusion Assemble 2D kernel
-template<int T_SDIM>
-void PADiffusionSetup2D(const int Q1D,
-                        const int coeffDim,
-                        const int NE,
-                        const Array<double> &w,
-                        const Vector &j,
-                        const Vector &c,
-                        Vector &d);
+template<>
+void PADiffusionSetup2D<2>(const int Q1D,
+                           const int coeffDim,
+                           const int NE,
+                           const Array<double> &w,
+                           const Vector &j,
+                           const Vector &c,
+                           Vector &d)
+{
+   const bool symmetric = (coeffDim != 4);
+   const bool const_c = c.Size() == 1;
+   MFEM_VERIFY(coeffDim < 3 ||
+               !const_c, "Constant matrix coefficient not supported");
+   const auto W = Reshape(w.Read(), Q1D,Q1D);
+   const auto J = Reshape(j.Read(), Q1D,Q1D,2,2,NE);
+   const auto C = const_c ? Reshape(c.Read(), 1,1,1,1) :
+                  Reshape(c.Read(), coeffDim,Q1D,Q1D,NE);
+   auto D = Reshape(d.Write(), Q1D,Q1D, symmetric ? 3 : 4, NE);
+   MFEM_FORALL_2D(e, NE, Q1D,Q1D,1,
+   {
+      MFEM_FOREACH_THREAD(qx,x,Q1D)
+      {
+         MFEM_FOREACH_THREAD(qy,y,Q1D)
+         {
+            const double J11 = J(qx,qy,0,0,e);
+            const double J21 = J(qx,qy,1,0,e);
+            const double J12 = J(qx,qy,0,1,e);
+            const double J22 = J(qx,qy,1,1,e);
+            const double w_detJ = W(qx,qy) / ((J11*J22)-(J21*J12));
+            if (coeffDim == 3 || coeffDim == 4) // Matrix coefficient
+            {
+               // First compute entries of R = MJ^{-T}, without det J factor.
+               const double M11 = C(0,qx,qy,e);
+               const double M12 = C(1,qx,qy,e);
+               const double M21 = symmetric ? M12 : C(2,qx,qy,e);
+               const double M22 = symmetric ? C(2,qx,qy,e) : C(3,qx,qy,e);
+               const double R11 = M11*J22 - M12*J12;
+               const double R21 = M21*J22 - M22*J12;
+               const double R12 = -M11*J21 + M12*J11;
+               const double R22 = -M21*J21 + M22*J11;
+
+               // Now set y to J^{-1}R.
+               D(qx,qy,0,e) = w_detJ * ( J22*R11 - J12*R21); // 1,1
+               D(qx,qy,1,e) = w_detJ * (-J21*R11 + J11*R21); // 2,1
+               D(qx,qy,2,e) = w_detJ * (symmetric ? (-J21*R12 + J11*R22) :
+                                        (J22*R12 - J12*R22)); // 2,2 or 1,2
+               if (!symmetric)
+               {
+                  D(qx,qy,3,e) = w_detJ * (-J21*R12 + J11*R22); // 2,2
+               }
+            }
+            else // Vector or scalar coefficient
+            {
+               const double C1 = const_c ? C(0,0,0,0) : C(0,qx,qy,e);
+               const double C2 = const_c ? C(0,0,0,0) :
+                                 (coeffDim == 2 ? C(1,qx,qy,e) : C(0,qx,qy,e));
+
+               D(qx,qy,0,e) =  w_detJ * (C2*J12*J12 + C1*J22*J22); // 1,1
+               D(qx,qy,1,e) = -w_detJ * (C2*J12*J11 + C1*J22*J21); // 1,2
+               D(qx,qy,2,e) =  w_detJ * (C2*J11*J11 + C1*J21*J21); // 2,2
+            }
+         }
+      }
+   });
+}
+
+// PA Diffusion Assemble 2D kernel with 3D node coords
+template<>
+void PADiffusionSetup2D<3>(const int Q1D,
+                           const int coeffDim,
+                           const int NE,
+                           const Array<double> &w,
+                           const Vector &j,
+                           const Vector &c,
+                           Vector &d)
+{
+   MFEM_VERIFY(coeffDim == 1, "Matrix and vector coefficients not supported");
+   constexpr int DIM = 2;
+   constexpr int SDIM = 3;
+   const bool const_c = c.Size() == 1;
+   const auto W = Reshape(w.Read(), Q1D,Q1D);
+   const auto J = Reshape(j.Read(), Q1D,Q1D,SDIM,DIM,NE);
+   const auto C = const_c ? Reshape(c.Read(), 1,1,1) :
+                  Reshape(c.Read(), Q1D,Q1D,NE);
+   auto D = Reshape(d.Write(), Q1D,Q1D, 3, NE);
+   MFEM_FORALL_2D(e, NE, Q1D,Q1D,1,
+   {
+      MFEM_FOREACH_THREAD(qx,x,Q1D)
+      {
+         MFEM_FOREACH_THREAD(qy,y,Q1D)
+         {
+            const double wq = W(qx,qy);
+            const double J11 = J(qx,qy,0,0,e);
+            const double J21 = J(qx,qy,1,0,e);
+            const double J31 = J(qx,qy,2,0,e);
+            const double J12 = J(qx,qy,0,1,e);
+            const double J22 = J(qx,qy,1,1,e);
+            const double J32 = J(qx,qy,2,1,e);
+            const double E = J11*J11 + J21*J21 + J31*J31;
+            const double G = J12*J12 + J22*J22 + J32*J32;
+            const double F = J11*J12 + J21*J22 + J31*J32;
+            const double iw = 1.0 / sqrt(E*G - F*F);
+            const double coeff = const_c ? C(0,0,0) : C(qx,qy,e);
+            const double alpha = wq * coeff * iw;
+            D(qx,qy,0,e) =  alpha * G; // 1,1
+            D(qx,qy,1,e) = -alpha * F; // 1,2
+            D(qx,qy,2,e) =  alpha * E; // 2,2
+         }
+      }
+   });
+}
 
 // PA Diffusion Assemble 3D kernel
 void PADiffusionSetup3D(const int Q1D,
@@ -53,41 +200,217 @@ void PADiffusionSetup3D(const int Q1D,
                         const Array<double> &w,
                         const Vector &j,
                         const Vector &c,
-                        Vector &d);
+                        Vector &d)
+{
+   const bool symmetric = (coeffDim != 9);
+   const bool const_c = c.Size() == 1;
+   MFEM_VERIFY(coeffDim < 6 ||
+               !const_c, "Constant matrix coefficient not supported");
+   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(), 1,1,1,1,1) :
+                  Reshape(c.Read(), coeffDim,Q1D,Q1D,Q1D,NE);
+   auto D = Reshape(d.Write(), Q1D,Q1D,Q1D, symmetric ? 6 : 9, NE);
+   MFEM_FORALL_3D(e, NE, Q1D, Q1D, Q1D,
+   {
+      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 w_detJ = W(qx,qy,qz) / 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);
 
+               if (coeffDim == 6 || coeffDim == 9) // Matrix coefficient version
+               {
+                  // Compute entries of R = MJ^{-T} = M adj(J)^T, without det J.
+                  const double M11 = C(0, qx,qy,qz, e);
+                  const double M12 = C(1, qx,qy,qz, e);
+                  const double M13 = C(2, qx,qy,qz, e);
+                  const double M21 = (!symmetric) ? C(3, qx,qy,qz, e) : M12;
+                  const double M22 = (!symmetric) ? C(4, qx,qy,qz, e) : C(3, qx,qy,qz, e);
+                  const double M23 = (!symmetric) ? C(5, qx,qy,qz, e) : C(4, qx,qy,qz, e);
+                  const double M31 = (!symmetric) ? C(6, qx,qy,qz, e) : M13;
+                  const double M32 = (!symmetric) ? C(7, qx,qy,qz, e) : M23;
+                  const double M33 = (!symmetric) ? C(8, qx,qy,qz, e) : C(5, qx,qy,qz, e);
+
+                  const double R11 = M11*A11 + M12*A12 + M13*A13;
+                  const double R12 = M11*A21 + M12*A22 + M13*A23;
+                  const double R13 = M11*A31 + M12*A32 + M13*A33;
+                  const double R21 = M21*A11 + M22*A12 + M23*A13;
+                  const double R22 = M21*A21 + M22*A22 + M23*A23;
+                  const double R23 = M21*A31 + M22*A32 + M23*A33;
+                  const double R31 = M31*A11 + M32*A12 + M33*A13;
+                  const double R32 = M31*A21 + M32*A22 + M33*A23;
+                  const double R33 = M31*A31 + M32*A32 + M33*A33;
+
+                  // Now set D to J^{-1} R = adj(J) R
+                  D(qx,qy,qz,0,e) = w_detJ * (A11*R11 + A12*R21 + A13*R31); // 1,1
+                  const double D12 = w_detJ * (A11*R12 + A12*R22 + A13*R32);
+                  D(qx,qy,qz,1,e) = D12; // 1,2
+                  D(qx,qy,qz,2,e) = w_detJ * (A11*R13 + A12*R23 + A13*R33); // 1,3
+
+                  const double D22 = w_detJ * (A21*R12 + A22*R22 + A23*R32);
+                  const double D23 = w_detJ * (A21*R13 + A22*R23 + A23*R33);
+
+                  const double D33 = w_detJ * (A31*R13 + A32*R23 + A33*R33);
+
+                  D(qx,qy,qz,4,e) = symmetric ? D23 : D22; // 2,3 or 2,2
+                  D(qx,qy,qz,5,e) = symmetric ? D33 : D23; // 3,3 or 2,3
+
+                  if (symmetric)
+                  {
+                     D(qx,qy,qz,3,e) = D22; // 2,2
+                  }
+                  else
+                  {
+                     D(qx,qy,qz,3,e) = w_detJ * (A21*R11 + A22*R21 + A23*R31); // 2,1
+                     D(qx,qy,qz,6,e) = w_detJ * (A31*R11 + A32*R21 + A33*R31); // 3,1
+                     D(qx,qy,qz,7,e) = w_detJ * (A31*R12 + A32*R22 + A33*R32); // 3,2
+                     D(qx,qy,qz,8,e) = D33; // 3,3
+                  }
+               }
+               else  // Vector or scalar coefficient version
+               {
+                  const double C1 = const_c ? C(0,0,0,0,0) : C(0,qx,qy,qz,e);
+                  const double C2 = const_c ? C(0,0,0,0,0) :
+                                    (coeffDim == 3 ? C(1,qx,qy,qz,e) : C(0,qx,qy,qz,e));
+                  const double C3 = const_c ? C(0,0,0,0,0) :
+                                    (coeffDim == 3 ? C(2,qx,qy,qz,e) : C(0,qx,qy,qz,e));
+
+                  // detJ J^{-1} J^{-T} = (1/detJ) adj(J) adj(J)^T
+                  D(qx,qy,qz,0,e) = w_detJ * (C1*A11*A11 + C2*A12*A12 + C3*A13*A13); // 1,1
+                  D(qx,qy,qz,1,e) = w_detJ * (C1*A11*A21 + C2*A12*A22 + C3*A13*A23); // 2,1
+                  D(qx,qy,qz,2,e) = w_detJ * (C1*A11*A31 + C2*A12*A32 + C3*A13*A33); // 3,1
+                  D(qx,qy,qz,3,e) = w_detJ * (C1*A21*A21 + C2*A22*A22 + C3*A23*A23); // 2,2
+                  D(qx,qy,qz,4,e) = w_detJ * (C1*A21*A31 + C2*A22*A32 + C3*A23*A33); // 3,2
+                  D(qx,qy,qz,5,e) = w_detJ * (C1*A31*A31 + C2*A32*A32 + C3*A33*A33); // 3,3
+               }
+            }
+         }
+      }
+   });
+}
+
+static void PADiffusionSetup(const int dim,
+                             const int sdim,
+                             const int D1D,
+                             const int Q1D,
+                             const int coeffDim,
+                             const int NE,
+                             const Array<double> &W,
+                             const Vector &J,
+                             const Vector &C,
+                             Vector &D)
+{
+   if (dim == 1) { MFEM_ABORT("dim==1 not supported in PADiffusionSetup"); }
+   if (dim == 2)
+   {
 #ifdef MFEM_USE_OCCA
-// OCCA 2D Assemble kernel
-void OccaPADiffusionSetup2D(const int D1D,
-                            const int Q1D,
-                            const int NE,
-                            const Array<double> &W,
-                            const Vector &J,
-                            const Vector &C,
-                            Vector &op);
-
-// OCCA 3D Assemble kernel
-void OccaPADiffusionSetup3D(const int D1D,
-                            const int Q1D,
-                            const int NE,
-                            const Array<double> &W,
-                            const Vector &J,
-                            const Vector &C,
-                            Vector &op);
+      if (DeviceCanUseOcca())
+      {
+         OccaPADiffusionSetup2D(D1D, Q1D, NE, W, J, C, D);
+         return;
+      }
+#else
+      MFEM_CONTRACT_VAR(D1D);
 #endif // MFEM_USE_OCCA
+      if (sdim == 2) { PADiffusionSetup2D<2>(Q1D, coeffDim, NE, W, J, C, D); }
+      if (sdim == 3) { PADiffusionSetup2D<3>(Q1D, coeffDim, NE, W, J, C, D); }
+   }
+   if (dim == 3)
+   {
+#ifdef MFEM_USE_OCCA
+      if (DeviceCanUseOcca())
+      {
+         OccaPADiffusionSetup3D(D1D, Q1D, NE, W, J, C, D);
+         return;
+      }
+#endif // MFEM_USE_OCCA
+      PADiffusionSetup3D(Q1D, coeffDim, NE, W, J, C, D);
+   }
+}
 
-void PADiffusionAssembleDiagonal(const int dim,
-                                 const int D1D,
-                                 const int Q1D,
-                                 const int NE,
-                                 const bool symm,
-                                 const Array<double> &B,
-                                 const Array<double> &G,
-                                 const Vector &D,
-                                 Vector &Y);
+void DiffusionIntegrator::AssemblePA(const FiniteElementSpace &fes)
+{
+   const MemoryType mt = (pa_mt == MemoryType::DEFAULT) ?
+                         Device::GetDeviceMemoryType() : pa_mt;
+   // Assuming the same element type
+   fespace = &fes;
+   Mesh *mesh = fes.GetMesh();
+   if (mesh->GetNE() == 0) { return; }
+   const FiniteElement &el = *fes.GetFE(0);
+   const IntegrationRule *ir = IntRule ? IntRule : &GetRule(el, el);
+   if (DeviceCanUseCeed())
+   {
+      delete ceedOp;
+      MFEM_VERIFY(!VQ && !MQ,
+                  "Only scalar coefficient supported for DiffusionIntegrator"
+                  " with libCEED");
+      const bool mixed = mesh->GetNumGeometries(mesh->Dimension()) > 1 ||
+                         fes.IsVariableOrder();
+      if (mixed)
+      {
+         ceedOp = new ceed::MixedPADiffusionIntegrator(*this, fes, Q);
+      }
+      else
+      {
+         ceedOp = new ceed::PADiffusionIntegrator(fes, *ir, Q);
+      }
+      return;
+   }
+   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();
+   geom = mesh->GetGeometricFactors(*ir, GeometricFactors::JACOBIANS, mt);
+   const int sdim = mesh->SpaceDimension();
+   maps = &el.GetDofToQuad(*ir, DofToQuad::TENSOR);
+   dofs1D = maps->ndof;
+   quad1D = maps->nqpt;
+
+   QuadratureSpace qs(*mesh, *ir);
+   CoefficientVector coeff(qs, CoefficientStorage::COMPRESSED);
+
+   if (MQ) { coeff.ProjectTranspose(*MQ); }
+   else if (VQ) { coeff.Project(*VQ); }
+   else if (Q) { coeff.Project(*Q); }
+   else { coeff.SetConstant(1.0); }
+
+   const int coeff_dim = coeff.GetVDim();
+   symmetric = (coeff_dim != dims*dims);
+   const int pa_size = symmetric ? symmDims : dims*dims;
+
+   pa_data.SetSize(pa_size * nq * ne, mt);
+   PADiffusionSetup(dim, sdim, dofs1D, quad1D, coeff_dim, ne, ir->GetWeights(),
+                    geom->J, coeff, pa_data);
+}
 
-// PA Diffusion Diagonal 2D kernel
 template<int T_D1D = 0, int T_Q1D = 0>
-inline void PADiffusionDiagonal2D(const int NE,
+static void PADiffusionDiagonal2D(const int NE,
                                   const bool symmetric,
                                   const Array<double> &b,
                                   const Array<double> &g,
@@ -106,7 +429,7 @@ inline void PADiffusionDiagonal2D(const int NE,
    // store necessary entries
    auto D = Reshape(d.Read(), Q1D*Q1D, symmetric ? 3 : 4, NE);
    auto Y = Reshape(y.ReadWrite(), D1D, D1D, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -153,7 +476,7 @@ inline void PADiffusionDiagonal2D(const int NE,
 
 // Shared memory PA Diffusion Diagonal 2D kernel
 template<int T_D1D = 0, int T_Q1D = 0, int T_NBZ = 0>
-inline void SmemPADiffusionDiagonal2D(const int NE,
+static void SmemPADiffusionDiagonal2D(const int NE,
                                       const bool symmetric,
                                       const Array<double> &b_,
                                       const Array<double> &g_,
@@ -173,7 +496,7 @@ inline void SmemPADiffusionDiagonal2D(const int NE,
    auto g = Reshape(g_.Read(), Q1D, D1D);
    auto D = Reshape(d_.Read(), Q1D*Q1D, symmetric ? 3 : 4, NE);
    auto Y = Reshape(y_.ReadWrite(), D1D, D1D, NE);
-   mfem::forall_2D_batch(NE, Q1D, Q1D, NBZ, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_2D(e, NE, Q1D, Q1D, NBZ,
    {
       const int tidz = MFEM_THREAD_ID(z);
       const int D1D = T_D1D ? T_D1D : d1d;
@@ -246,9 +569,8 @@ inline void SmemPADiffusionDiagonal2D(const int NE,
    });
 }
 
-// PA Diffusion Diagonal 3D kernel
 template<int T_D1D = 0, int T_Q1D = 0>
-inline void PADiffusionDiagonal3D(const int NE,
+static void PADiffusionDiagonal3D(const int NE,
                                   const bool symmetric,
                                   const Array<double> &b,
                                   const Array<double> &g,
@@ -268,7 +590,7 @@ inline void PADiffusionDiagonal3D(const int NE,
    auto G = Reshape(g.Read(), Q1D, D1D);
    auto Q = Reshape(d.Read(), Q1D*Q1D*Q1D, symmetric ? 6 : 9, NE);
    auto Y = Reshape(y.ReadWrite(), D1D, D1D, D1D, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -292,8 +614,8 @@ inline void PADiffusionDiagonal3D(const int NE,
                      {
                         const int q = qx + (qy + qz * Q1D) * Q1D;
                         const int ksym = j >= i ?
-                                         3 - (3-i)*(2-i)/2 + j:
-                                         3 - (3-j)*(2-j)/2 + i;
+                        3 - (3-i)*(2-i)/2 + j:
+                        3 - (3-j)*(2-j)/2 + i;
                         const int k = symmetric ? ksym : (i*DIM) + j;
                         const double O = Q(q,k,e);
                         const double Bz = B(qz,dz);
@@ -349,7 +671,7 @@ inline void PADiffusionDiagonal3D(const int NE,
 
 // Shared memory PA Diffusion Diagonal 3D kernel
 template<int T_D1D = 0, int T_Q1D = 0>
-inline void SmemPADiffusionDiagonal3D(const int NE,
+static void SmemPADiffusionDiagonal3D(const int NE,
                                       const bool symmetric,
                                       const Array<double> &b_,
                                       const Array<double> &g_,
@@ -369,7 +691,7 @@ inline void SmemPADiffusionDiagonal3D(const int NE,
    auto g = Reshape(g_.Read(), Q1D, D1D);
    auto D = Reshape(d_.Read(), Q1D*Q1D*Q1D, symmetric ? 6 : 9, NE);
    auto Y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, NE);
-   mfem::forall_3D(NE, Q1D, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, Q1D, Q1D, Q1D,
    {
       const int tidz = MFEM_THREAD_ID(z);
       const int D1D = T_D1D ? T_D1D : d1d;
@@ -466,48 +788,169 @@ inline void SmemPADiffusionDiagonal3D(const int NE,
    });
 }
 
-void PADiffusionApply(const int dim,
-                      const int D1D,
-                      const int Q1D,
-                      const int NE,
-                      const bool symm,
-                      const Array<double> &B,
-                      const Array<double> &G,
-                      const Array<double> &Bt,
-                      const Array<double> &Gt,
-                      const Vector &D,
-                      const Vector &X,
-                      Vector &Y);
+static void PADiffusionAssembleDiagonal(const int dim,
+                                        const int D1D,
+                                        const int Q1D,
+                                        const int NE,
+                                        const bool symm,
+                                        const Array<double> &B,
+                                        const Array<double> &G,
+                                        const Vector &D,
+                                        Vector &Y)
+{
+   if (dim == 2)
+   {
+      switch ((D1D << 4 ) | Q1D)
+      {
+         case 0x22: return SmemPADiffusionDiagonal2D<2,2,8>(NE,symm,B,G,D,Y);
+         case 0x33: return SmemPADiffusionDiagonal2D<3,3,8>(NE,symm,B,G,D,Y);
+         case 0x44: return SmemPADiffusionDiagonal2D<4,4,4>(NE,symm,B,G,D,Y);
+         case 0x55: return SmemPADiffusionDiagonal2D<5,5,4>(NE,symm,B,G,D,Y);
+         case 0x66: return SmemPADiffusionDiagonal2D<6,6,2>(NE,symm,B,G,D,Y);
+         case 0x77: return SmemPADiffusionDiagonal2D<7,7,2>(NE,symm,B,G,D,Y);
+         case 0x88: return SmemPADiffusionDiagonal2D<8,8,1>(NE,symm,B,G,D,Y);
+         case 0x99: return SmemPADiffusionDiagonal2D<9,9,1>(NE,symm,B,G,D,Y);
+         default: return PADiffusionDiagonal2D(NE,symm,B,G,D,Y,D1D,Q1D);
+      }
+   }
+   else if (dim == 3)
+   {
+      switch ((D1D << 4 ) | Q1D)
+      {
+         case 0x22: return SmemPADiffusionDiagonal3D<2,2>(NE,symm,B,G,D,Y);
+         case 0x23: return SmemPADiffusionDiagonal3D<2,3>(NE,symm,B,G,D,Y);
+         case 0x34: return SmemPADiffusionDiagonal3D<3,4>(NE,symm,B,G,D,Y);
+         case 0x45: return SmemPADiffusionDiagonal3D<4,5>(NE,symm,B,G,D,Y);
+         case 0x46: return SmemPADiffusionDiagonal3D<4,6>(NE,symm,B,G,D,Y);
+         case 0x56: return SmemPADiffusionDiagonal3D<5,6>(NE,symm,B,G,D,Y);
+         case 0x67: return SmemPADiffusionDiagonal3D<6,7>(NE,symm,B,G,D,Y);
+         case 0x78: return SmemPADiffusionDiagonal3D<7,8>(NE,symm,B,G,D,Y);
+         case 0x89: return SmemPADiffusionDiagonal3D<8,9>(NE,symm,B,G,D,Y);
+         case 0x9A: return SmemPADiffusionDiagonal3D<9,10>(NE,symm,B,G,D,Y);
+         default: return PADiffusionDiagonal3D(NE,symm,B,G,D,Y,D1D,Q1D);
+      }
+   }
+   MFEM_ABORT("Unknown kernel.");
+}
+
+void DiffusionIntegrator::AssembleDiagonalPA(Vector &diag)
+{
+   if (DeviceCanUseCeed())
+   {
+      ceedOp->GetDiagonal(diag);
+   }
+   else
+   {
+      if (pa_data.Size()==0) { AssemblePA(*fespace); }
+      PADiffusionAssembleDiagonal(dim, dofs1D, quad1D, ne, symmetric,
+                                  maps->B, maps->G, pa_data, diag);
+   }
+}
+
 
 #ifdef MFEM_USE_OCCA
 // OCCA PA Diffusion Apply 2D kernel
-void OccaPADiffusionApply2D(const int D1D,
-                            const int Q1D,
-                            const int NE,
-                            const Array<double> &B,
-                            const Array<double> &G,
-                            const Array<double> &Bt,
-                            const Array<double> &Gt,
-                            const Vector &D,
-                            const Vector &X,
-                            Vector &Y);
+static void OccaPADiffusionApply2D(const int D1D,
+                                   const int Q1D,
+                                   const int NE,
+                                   const Array<double> &B,
+                                   const Array<double> &G,
+                                   const Array<double> &Bt,
+                                   const Array<double> &Gt,
+                                   const Vector &D,
+                                   const Vector &X,
+                                   Vector &Y)
+{
+   occa::properties props;
+   props["defines/D1D"] = D1D;
+   props["defines/Q1D"] = Q1D;
+   const occa::memory o_B = OccaMemoryRead(B.GetMemory(), B.Size());
+   const occa::memory o_G = OccaMemoryRead(G.GetMemory(), G.Size());
+   const occa::memory o_Bt = OccaMemoryRead(Bt.GetMemory(), Bt.Size());
+   const occa::memory o_Gt = OccaMemoryRead(Gt.GetMemory(), Gt.Size());
+   const occa::memory o_D = OccaMemoryRead(D.GetMemory(), D.Size());
+   const occa::memory o_X = OccaMemoryRead(X.GetMemory(), X.Size());
+   occa::memory o_Y = OccaMemoryReadWrite(Y.GetMemory(), Y.Size());
+   const occa_id_t id = std::make_pair(D1D,Q1D);
+   if (!Device::Allows(Backend::OCCA_CUDA))
+   {
+      static occa_kernel_t OccaDiffApply2D_cpu;
+      if (OccaDiffApply2D_cpu.find(id) == OccaDiffApply2D_cpu.end())
+      {
+         const occa::kernel DiffusionApply2D_CPU =
+            mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
+                                        "DiffusionApply2D_CPU", props);
+         OccaDiffApply2D_cpu.emplace(id, DiffusionApply2D_CPU);
+      }
+      OccaDiffApply2D_cpu.at(id)(NE, o_B, o_G, o_Bt, o_Gt, o_D, o_X, o_Y);
+   }
+   else
+   {
+      static occa_kernel_t OccaDiffApply2D_gpu;
+      if (OccaDiffApply2D_gpu.find(id) == OccaDiffApply2D_gpu.end())
+      {
+         const occa::kernel DiffusionApply2D_GPU =
+            mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
+                                        "DiffusionApply2D_GPU", props);
+         OccaDiffApply2D_gpu.emplace(id, DiffusionApply2D_GPU);
+      }
+      OccaDiffApply2D_gpu.at(id)(NE, o_B, o_G, o_Bt, o_Gt, o_D, o_X, o_Y);
+   }
+}
 
 // OCCA PA Diffusion Apply 3D kernel
-void OccaPADiffusionApply3D(const int D1D,
-                            const int Q1D,
-                            const int NE,
-                            const Array<double> &B,
-                            const Array<double> &G,
-                            const Array<double> &Bt,
-                            const Array<double> &Gt,
-                            const Vector &D,
-                            const Vector &X,
-                            Vector &Y);
+static void OccaPADiffusionApply3D(const int D1D,
+                                   const int Q1D,
+                                   const int NE,
+                                   const Array<double> &B,
+                                   const Array<double> &G,
+                                   const Array<double> &Bt,
+                                   const Array<double> &Gt,
+                                   const Vector &D,
+                                   const Vector &X,
+                                   Vector &Y)
+{
+   occa::properties props;
+   props["defines/D1D"] = D1D;
+   props["defines/Q1D"] = Q1D;
+   const occa::memory o_B = OccaMemoryRead(B.GetMemory(), B.Size());
+   const occa::memory o_G = OccaMemoryRead(G.GetMemory(), G.Size());
+   const occa::memory o_Bt = OccaMemoryRead(Bt.GetMemory(), Bt.Size());
+   const occa::memory o_Gt = OccaMemoryRead(Gt.GetMemory(), Gt.Size());
+   const occa::memory o_D = OccaMemoryRead(D.GetMemory(), D.Size());
+   const occa::memory o_X = OccaMemoryRead(X.GetMemory(), X.Size());
+   occa::memory o_Y = OccaMemoryReadWrite(Y.GetMemory(), Y.Size());
+   const occa_id_t id = std::make_pair(D1D,Q1D);
+   if (!Device::Allows(Backend::OCCA_CUDA))
+   {
+      static occa_kernel_t OccaDiffApply3D_cpu;
+      if (OccaDiffApply3D_cpu.find(id) == OccaDiffApply3D_cpu.end())
+      {
+         const occa::kernel DiffusionApply3D_CPU =
+            mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
+                                        "DiffusionApply3D_CPU", props);
+         OccaDiffApply3D_cpu.emplace(id, DiffusionApply3D_CPU);
+      }
+      OccaDiffApply3D_cpu.at(id)(NE, o_B, o_G, o_Bt, o_Gt, o_D, o_X, o_Y);
+   }
+   else
+   {
+      static occa_kernel_t OccaDiffApply3D_gpu;
+      if (OccaDiffApply3D_gpu.find(id) == OccaDiffApply3D_gpu.end())
+      {
+         const occa::kernel DiffusionApply3D_GPU =
+            mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
+                                        "DiffusionApply3D_GPU", props);
+         OccaDiffApply3D_gpu.emplace(id, DiffusionApply3D_GPU);
+      }
+      OccaDiffApply3D_gpu.at(id)(NE, o_B, o_G, o_Bt, o_Gt, o_D, o_X, o_Y);
+   }
+}
 #endif // MFEM_USE_OCCA
 
 // PA Diffusion Apply 2D kernel
 template<int T_D1D = 0, int T_Q1D = 0>
-inline void PADiffusionApply2D(const int NE,
+static void PADiffusionApply2D(const int NE,
                                const bool symmetric,
                                const Array<double> &b_,
                                const Array<double> &g_,
@@ -530,7 +973,7 @@ inline void PADiffusionApply2D(const int NE,
    auto D = Reshape(d_.Read(), Q1D*Q1D, symmetric ? 3 : 4, NE);
    auto X = Reshape(x_.Read(), D1D, D1D, NE);
    auto Y = Reshape(y_.ReadWrite(), D1D, D1D, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -629,7 +1072,7 @@ inline void PADiffusionApply2D(const int NE,
 
 // Shared memory PA Diffusion Apply 2D kernel
 template<int T_D1D = 0, int T_Q1D = 0, int T_NBZ = 0>
-inline void SmemPADiffusionApply2D(const int NE,
+static void SmemPADiffusionApply2D(const int NE,
                                    const bool symmetric,
                                    const Array<double> &b_,
                                    const Array<double> &g_,
@@ -651,7 +1094,7 @@ inline void SmemPADiffusionApply2D(const int NE,
    auto D = Reshape(d_.Read(), Q1D*Q1D, symmetric ? 3 : 4, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, NE);
    auto Y = Reshape(y_.ReadWrite(), D1D, D1D, NE);
-   mfem::forall_2D_batch(NE, Q1D, Q1D, NBZ, [=] MFEM_HOST_DEVICE(int e)
+   MFEM_FORALL_2D(e, NE, Q1D, Q1D, NBZ,
    {
       const int tidz = MFEM_THREAD_ID(z);
       const int D1D = T_D1D ? T_D1D : d1d;
@@ -787,7 +1230,7 @@ inline void SmemPADiffusionApply2D(const int NE,
 
 // PA Diffusion Apply 3D kernel
 template<int T_D1D = 0, int T_Q1D = 0>
-inline void PADiffusionApply3D(const int NE,
+static void PADiffusionApply3D(const int NE,
                                const bool symmetric,
                                const Array<double> &b,
                                const Array<double> &g,
@@ -809,7 +1252,7 @@ inline void PADiffusionApply3D(const int NE,
    auto D = Reshape(d_.Read(), Q1D*Q1D*Q1D, symmetric ? 6 : 9, NE);
    auto X = Reshape(x_.Read(), D1D, D1D, D1D, NE);
    auto Y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -978,9 +1421,8 @@ inline void PADiffusionApply3D(const int NE,
    });
 }
 
-// Shared memory PA Diffusion Apply 3D kernel
 template<int T_D1D = 0, int T_Q1D = 0>
-inline void SmemPADiffusionApply3D(const int NE,
+static void SmemPADiffusionApply3D(const int NE,
                                    const bool symmetric,
                                    const Array<double> &b_,
                                    const Array<double> &g_,
@@ -1001,7 +1443,7 @@ inline void SmemPADiffusionApply3D(const int NE,
    auto d = Reshape(d_.Read(), Q1D, Q1D, Q1D, symmetric ? 6 : 9, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, D1D, NE);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, NE);
-   mfem::forall_3D(NE, Q1D, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, Q1D, Q1D, Q1D,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -1201,8 +1643,99 @@ inline void SmemPADiffusionApply3D(const int NE,
    });
 }
 
-} // namespace internal
+static void PADiffusionApply(const int dim,
+                             const int D1D,
+                             const int Q1D,
+                             const int NE,
+                             const bool symm,
+                             const Array<double> &B,
+                             const Array<double> &G,
+                             const Array<double> &Bt,
+                             const Array<double> &Gt,
+                             const Vector &D,
+                             const Vector &X,
+                             Vector &Y)
+{
+#ifdef MFEM_USE_OCCA
+   if (DeviceCanUseOcca())
+   {
+      if (dim == 2)
+      {
+         OccaPADiffusionApply2D(D1D,Q1D,NE,B,G,Bt,Gt,D,X,Y);
+         return;
+      }
+      if (dim == 3)
+      {
+         OccaPADiffusionApply3D(D1D,Q1D,NE,B,G,Bt,Gt,D,X,Y);
+         return;
+      }
+      MFEM_ABORT("OCCA PADiffusionApply unknown kernel!");
+   }
+#endif // MFEM_USE_OCCA
+   const int id = (D1D << 4) | Q1D;
+
+   if (dim == 2)
+   {
+      switch (id)
+      {
+         case 0x22: return SmemPADiffusionApply2D<2,2,16>(NE,symm,B,G,D,X,Y);
+         case 0x33: return SmemPADiffusionApply2D<3,3,16>(NE,symm,B,G,D,X,Y);
+         case 0x44: return SmemPADiffusionApply2D<4,4,8>(NE,symm,B,G,D,X,Y);
+         case 0x55: return SmemPADiffusionApply2D<5,5,8>(NE,symm,B,G,D,X,Y);
+         case 0x66: return SmemPADiffusionApply2D<6,6,4>(NE,symm,B,G,D,X,Y);
+         case 0x77: return SmemPADiffusionApply2D<7,7,4>(NE,symm,B,G,D,X,Y);
+         case 0x88: return SmemPADiffusionApply2D<8,8,2>(NE,symm,B,G,D,X,Y);
+         case 0x99: return SmemPADiffusionApply2D<9,9,2>(NE,symm,B,G,D,X,Y);
+         default:   return PADiffusionApply2D(NE,symm,B,G,Bt,Gt,D,X,Y,D1D,Q1D);
+      }
+   }
+
+   if (dim == 3)
+   {
+      switch (id)
+      {
+         case 0x22: return SmemPADiffusionApply3D<2,2>(NE,symm,B,G,D,X,Y);
+         case 0x23: return SmemPADiffusionApply3D<2,3>(NE,symm,B,G,D,X,Y);
+         case 0x34: return SmemPADiffusionApply3D<3,4>(NE,symm,B,G,D,X,Y);
+         case 0x45: return SmemPADiffusionApply3D<4,5>(NE,symm,B,G,D,X,Y);
+         case 0x46: return SmemPADiffusionApply3D<4,6>(NE,symm,B,G,D,X,Y);
+         case 0x56: return SmemPADiffusionApply3D<5,6>(NE,symm,B,G,D,X,Y);
+         case 0x58: return SmemPADiffusionApply3D<5,8>(NE,symm,B,G,D,X,Y);
+         case 0x67: return SmemPADiffusionApply3D<6,7>(NE,symm,B,G,D,X,Y);
+         case 0x78: return SmemPADiffusionApply3D<7,8>(NE,symm,B,G,D,X,Y);
+         case 0x89: return SmemPADiffusionApply3D<8,9>(NE,symm,B,G,D,X,Y);
+         default:   return PADiffusionApply3D(NE,symm,B,G,Bt,Gt,D,X,Y,D1D,Q1D);
+      }
+   }
+   MFEM_ABORT("Unknown kernel: 0x"<<std::hex << id << std::dec);
+}
+
+// PA Diffusion Apply kernel
+void DiffusionIntegrator::AddMultPA(const Vector &x, Vector &y) const
+{
+   if (DeviceCanUseCeed())
+   {
+      ceedOp->AddMult(x, y);
+   }
+   else
+   {
+      PADiffusionApply(dim, dofs1D, quad1D, ne, symmetric,
+                       maps->B, maps->G, maps->Bt, maps->Gt,
+                       pa_data, x, y);
+   }
+}
+
+void DiffusionIntegrator::AddMultTransposePA(const Vector &x, Vector &y) const
+{
+   if (symmetric)
+   {
+      AddMultPA(x, y);
+   }
+   else
+   {
+      MFEM_ABORT("DiffusionIntegrator::AddMultTransposePA only implemented in "
+                 "the symmetric case.")
+   }
+}
 
 } // namespace mfem
-
-#endif

fem/bilininteg_divergence.cpp

@@ -9,13 +9,17 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../../general/forall.hpp"
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
+
+using namespace std;
 
 namespace mfem
 {
 
+// PA Divergence Integrator
+
 // PA Divergence Assemble 2D kernel
 static void PADivergenceSetup2D(const int Q1D,
                                 const int NE,
@@ -29,7 +33,7 @@ static void PADivergenceSetup2D(const int Q1D,
    auto J = Reshape(j.Read(), NQ, 2, 2, NE);
    auto y = Reshape(op.Write(), NQ, 2, 2, NE);
 
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       for (int q = 0; q < NQ; ++q)
       {
@@ -58,7 +62,7 @@ static void PADivergenceSetup3D(const int Q1D,
    auto W = w.Read();
    auto J = Reshape(j.Read(), NQ, 3, 3, NE);
    auto y = Reshape(op.Write(), NQ, 3, 3, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       for (int q = 0; q < NQ; ++q)
       {
@@ -179,7 +183,7 @@ static void PADivergenceApply2D(const int NE,
    auto op = Reshape(op_.Read(), Q1D*Q1D, 2,2, NE);
    auto x = Reshape(x_.Read(), TR_D1D, TR_D1D, 2, NE);
    auto y = Reshape(y_.ReadWrite(), TE_D1D, TE_D1D, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int TR_D1D = T_TR_D1D ? T_TR_D1D : tr_d1d;
       const int TE_D1D = T_TE_D1D ? T_TE_D1D : te_d1d;
@@ -317,7 +321,7 @@ static void PADivergenceApplyTranspose2D(const int NE,
    auto op = Reshape(op_.Read(), Q1D*Q1D, 2,2, NE);
    auto x  = Reshape(x_.Read(), TE_D1D, TE_D1D, NE);
    auto y  = Reshape(y_.ReadWrite(), TR_D1D, TR_D1D, 2, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int TR_D1D = T_TR_D1D ? T_TR_D1D : tr_d1d;
       const int TE_D1D = T_TE_D1D ? T_TE_D1D : te_d1d;
@@ -433,7 +437,7 @@ static void PADivergenceApply3D(const int NE,
    auto op = Reshape(op_.Read(), Q1D*Q1D*Q1D, 3,3, NE);
    auto x = Reshape(x_.Read(), TR_D1D, TR_D1D, TR_D1D, 3, NE);
    auto y = Reshape(y_.ReadWrite(), TE_D1D, TE_D1D, TE_D1D, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int TR_D1D = T_TR_D1D ? T_TR_D1D : tr_d1d;
       const int TE_D1D = T_TE_D1D ? T_TE_D1D : te_d1d;
@@ -616,7 +620,7 @@ static void PADivergenceApplyTranspose3D(const int NE,
    auto op = Reshape(op_.Read(), Q1D*Q1D*Q1D, 3,3, NE);
    auto x  = Reshape(x_.Read(), TE_D1D, TE_D1D, TE_D1D, NE);
    auto y  = Reshape(y_.ReadWrite(), TR_D1D, TR_D1D, TR_D1D, 3, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int TR_D1D = T_TR_D1D ? T_TR_D1D : tr_d1d;
       const int TE_D1D = T_TE_D1D ? T_TE_D1D : te_d1d;
@@ -797,7 +801,7 @@ static void SmemPADivergenceApply3D(const int NE,
    auto x = Reshape(x_.Read(), TR_D1D, TR_D1D, TR_D1D, 3, NE);
    auto y = Reshape(y_.ReadWrite(), TE_D1D, TE_D1D, TE_D1D, NE);
 
-   mfem::forall_3D(NE, Q1D, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, Q1D, Q1D, Q1D,
    {
       constexpr int VDIM = 3;
       const int tidz = MFEM_THREAD_ID(z);
@@ -1036,25 +1040,11 @@ static void PADivergenceApply(const int dim,
 {
    if (dim == 2)
    {
-      if (transpose)
-      {
-         return PADivergenceApplyTranspose2D(NE,B,G,Bt,op,x,y,TR_D1D,TE_D1D,Q1D);
-      }
-      else
-      {
-         return PADivergenceApply2D(NE,B,G,Bt,op,x,y,TR_D1D,TE_D1D,Q1D);
-      }
+      return PADivergenceApply2D(NE,B,G,Bt,op,x,y,TR_D1D,TE_D1D,Q1D);
    }
    if (dim == 3)
    {
-      if (transpose)
-      {
-         return PADivergenceApplyTranspose3D(NE,B,G,Bt,op,x,y,TR_D1D,TE_D1D,Q1D);
-      }
-      else
-      {
-         return PADivergenceApply3D(NE,B,G,Bt,op,x,y,TR_D1D,TE_D1D,Q1D);
-      }
+      return PADivergenceApply3D(NE,B,G,Bt,op,x,y,TR_D1D,TE_D1D,Q1D);
    }
    MFEM_ABORT("Unknown kernel.");
 }

fem/bilininteg_gradient.cpp

@@ -9,14 +9,18 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../../general/forall.hpp"
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
-#include "../qfunction.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
+#include "qfunction.hpp"
+
+using namespace std;
 
 namespace mfem
 {
 
+// PA Gradient Integrator
+
 /* Description of the *SetupND functions
    Inputs are as follows
    \b Q1D number of quadrature points in one dimension.
@@ -58,8 +62,8 @@ namespace mfem
    The shared memory (Smem) versions of the kernels differ from the regular
    versions in the following properties.
 
-   \b mfem::forall is using only one level of parallelism.
-   \b mfem::forall_ND uses an additional level of parallelism
+   \b MFEM_FORALL is using only one level of parallelism.
+   \b MFEM_FORALL_ND uses an additional level of parallelism
    \b MFEM_FOREACH_THREAD
 
    These macros allow automatic mapping of manually defined blocks to
@@ -83,7 +87,7 @@ static void PAGradientSetup2D(const int Q1D,
    const auto C = const_c ? Reshape(c.Read(), 1,1) :
                   Reshape(c.Read(), NQ, NE);
 
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       for (int q = 0; q < NQ; ++q)
       {
@@ -118,7 +122,7 @@ static void PAGradientSetup3D(const int Q1D,
    const auto C = const_c ? Reshape(c.Read(), 1,1) :
                   Reshape(c.Read(), NQ,NE);
 
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       for (int q = 0; q < NQ; ++q)
       {
@@ -238,7 +242,7 @@ static void PAGradientApply2D(const int NE,
    auto op = Reshape(op_.Read(), Q1D*Q1D, 2,2, NE);
    auto x = Reshape(x_.Read(), TR_D1D, TR_D1D, NE);
    auto y = Reshape(y_.ReadWrite(), TE_D1D, TE_D1D, 2, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int TR_D1D = T_TR_D1D ? T_TR_D1D : tr_d1d;
       const int TE_D1D = T_TE_D1D ? T_TE_D1D : te_d1d;
@@ -368,7 +372,7 @@ static void PAGradientApply3D(const int NE,
    auto op = Reshape(op_.Read(), Q1D*Q1D*Q1D, 3,3, NE);
    auto x = Reshape(x_.Read(), TR_D1D, TR_D1D, TR_D1D, NE);
    auto y = Reshape(y_.ReadWrite(), TE_D1D, TE_D1D, TE_D1D, 3, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int TR_D1D = T_TR_D1D ? T_TR_D1D : tr_d1d;
       const int TE_D1D = T_TE_D1D ? T_TE_D1D : te_d1d;
@@ -568,8 +572,7 @@ static void SmemPAGradientApply3D(const int NE,
    auto x = Reshape(x_.Read(), TR_D1D, TR_D1D, TR_D1D, NE);
    auto y = Reshape(y_.ReadWrite(), TE_D1D, TE_D1D, TE_D1D, 3, NE);
 
-   mfem::forall_3D(NE, (Q1D>8)?8:Q1D, (Q1D>8)?8:Q1D, (Q1D>8)?8:Q1D,
-                   [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, (Q1D>8)?8:Q1D, (Q1D>8)?8:Q1D, (Q1D>8)?8:Q1D,
    {
       const int tidz = MFEM_THREAD_ID(z);
       const int D1DR = T_TR_D1D ? T_TR_D1D : tr_d1d;

fem/bilininteg_mass_ea.cpp

@@ -9,9 +9,9 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../../general/forall.hpp"
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
 
 namespace mfem
 {
@@ -32,7 +32,7 @@ static void EAMassAssemble1D(const int NE,
    auto B = Reshape(basis.Read(), Q1D, D1D);
    auto D = Reshape(padata.Read(), Q1D, NE);
    auto M = Reshape(eadata.ReadWrite(), D1D, D1D, NE);
-   mfem::forall_2D(NE, D1D, D1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, D1D, D1D, 1,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -82,7 +82,7 @@ static void EAMassAssemble2D(const int NE,
    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);
-   mfem::forall_2D(NE, D1D, D1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, D1D, D1D, 1,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -154,51 +154,20 @@ static void EAMassAssemble3D(const int NE,
    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);
-   mfem::forall_3D(NE, D1D, D1D, D1D, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_3D(e, NE, D1D, D1D, D1D,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
       constexpr int MD1 = T_D1D ? T_D1D : MAX_D1D;
       constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
-      constexpr int DQ = T_D1D * T_Q1D;
-
-      // For quadratic and lower it's better to use registers but for higher-order you start to
-      // spill and it's better to use shared memory
-      constexpr bool USE_REG = DQ != 0 && DQ <= 12;
-      constexpr int MD1r = USE_REG ? MD1 : 1;
-      constexpr int MQ1r = USE_REG ? MQ1 : 1;
-      constexpr int MD1s = USE_REG ? 1 : MD1;
-      constexpr int MQ1s = USE_REG ? 1 : MQ1;
-
-      MFEM_SHARED double s_B[MQ1s][MD1s];
-      double r_B[MQ1r][MD1r];
-      double (*l_B)[MD1] = nullptr;
-      if (USE_REG)
+      double r_B[MQ1][MD1];
+      for (int d = 0; d < D1D; d++)
       {
-         for (int d = 0; d < D1D; d++)
+         for (int q = 0; q < Q1D; q++)
          {
-            for (int q = 0; q < Q1D; q++)
-            {
-               r_B[q][d] = B(q,d);
-            }
+            r_B[q][d] = B(q,d);
          }
-         l_B = (double (*)[MD1])r_B;
       }
-      else
-      {
-         if (MFEM_THREAD_ID(z) == 0)
-         {
-            MFEM_FOREACH_THREAD(d,x,D1D)
-            {
-               MFEM_FOREACH_THREAD(q,y,Q1D)
-               {
-                  s_B[q][d] = B(q,d);
-               }
-            }
-         }
-         l_B = (double (*)[MD1])s_B;
-      }
-
       MFEM_SHARED double s_D[MQ1][MQ1][MQ1];
       MFEM_FOREACH_THREAD(k1,x,Q1D)
       {
@@ -230,9 +199,9 @@ static void EAMassAssemble3D(const int NE,
                            {
                               for (int k3 = 0; k3 < Q1D; ++k3)
                               {
-                                 val += l_B[k1][i1] * l_B[k1][j1]
-                                        * l_B[k2][i2] * l_B[k2][j2]
-                                        * l_B[k3][i3] * l_B[k3][j3]
+                                 val += r_B[k1][i1] * r_B[k1][j1]
+                                        * r_B[k2][i2] * r_B[k2][j2]
+                                        * r_B[k3][i3] * r_B[k3][j3]
                                         * s_D[k1][k2][k3];
                               }
                            }

fem/bilininteg_mass_mf.cpp

@@ -9,9 +9,12 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
-#include "../ceed/integrators/mass/mass.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
+#include "ceed/integrators/mass/mass.hpp"
+
+using namespace std;
 
 namespace mfem
 {

fem/bilininteg_mass_pa.cpp

@@ -0,0 +1,736 @@
+// Copyright (c) 2010-2023, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
+#include "qfunction.hpp"
+#include "ceed/integrators/mass/mass.hpp"
+#include "bilininteg_mass_pa.hpp"
+
+using namespace std;
+
+namespace mfem
+{
+
+// PA Mass Integrator
+
+// PA Mass Assemble kernel
+
+void MassIntegrator::AssemblePA(const FiniteElementSpace &fes)
+{
+   const MemoryType mt = (pa_mt == MemoryType::DEFAULT) ?
+                         Device::GetDeviceMemoryType() : pa_mt;
+
+   // Assuming the same element type
+   fespace = &fes;
+   Mesh *mesh = fes.GetMesh();
+   if (mesh->GetNE() == 0) { return; }
+   const FiniteElement &el = *fes.GetFE(0);
+   ElementTransformation *T0 = mesh->GetElementTransformation(0);
+   const IntegrationRule *ir = IntRule ? IntRule : &GetRule(el, el, *T0);
+   if (DeviceCanUseCeed())
+   {
+      delete ceedOp;
+      const bool mixed = mesh->GetNumGeometries(mesh->Dimension()) > 1 ||
+                         fes.IsVariableOrder();
+      if (mixed)
+      {
+         ceedOp = new ceed::MixedPAMassIntegrator(*this, fes, Q);
+      }
+      else
+      {
+         ceedOp = new ceed::PAMassIntegrator(fes, *ir, Q);
+      }
+      return;
+   }
+   int map_type = el.GetMapType();
+   dim = mesh->Dimension();
+   ne = fes.GetMesh()->GetNE();
+   nq = ir->GetNPoints();
+   geom = mesh->GetGeometricFactors(*ir, GeometricFactors::DETERMINANTS, mt);
+   maps = &el.GetDofToQuad(*ir, DofToQuad::TENSOR);
+   dofs1D = maps->ndof;
+   quad1D = maps->nqpt;
+   pa_data.SetSize(ne*nq, mt);
+
+   QuadratureSpace qs(*mesh, *ir);
+   CoefficientVector coeff(Q, qs, CoefficientStorage::COMPRESSED);
+
+   if (dim==1) { MFEM_ABORT("Not supported yet... stay tuned!"); }
+   if (dim==2)
+   {
+      const int NE = ne;
+      const int Q1D = quad1D;
+      const bool const_c = coeff.Size() == 1;
+      const bool by_val = map_type == FiniteElement::VALUE;
+      const auto W = Reshape(ir->GetWeights().Read(), Q1D,Q1D);
+      const auto J = Reshape(geom->detJ.Read(), Q1D,Q1D,NE);
+      const auto C = const_c ? Reshape(coeff.Read(), 1,1,1) :
+                     Reshape(coeff.Read(), Q1D,Q1D,NE);
+      auto v = Reshape(pa_data.Write(), Q1D,Q1D, NE);
+      MFEM_FORALL_2D(e, NE, Q1D,Q1D,1,
+      {
+         MFEM_FOREACH_THREAD(qx,x,Q1D)
+         {
+            MFEM_FOREACH_THREAD(qy,y,Q1D)
+            {
+               const double detJ = J(qx,qy,e);
+               const double coeff = const_c ? C(0,0,0) : C(qx,qy,e);
+               v(qx,qy,e) =  W(qx,qy) * coeff * (by_val ? detJ : 1.0/detJ);
+            }
+         }
+      });
+   }
+   if (dim==3)
+   {
+      const int NE = ne;
+      const int Q1D = quad1D;
+      const bool const_c = coeff.Size() == 1;
+      const bool by_val = map_type == FiniteElement::VALUE;
+      const auto W = Reshape(ir->GetWeights().Read(), Q1D,Q1D,Q1D);
+      const auto J = Reshape(geom->detJ.Read(), Q1D,Q1D,Q1D,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,
+      {
+         MFEM_FOREACH_THREAD(qx,x,Q1D)
+         {
+            MFEM_FOREACH_THREAD(qy,y,Q1D)
+            {
+               MFEM_FOREACH_THREAD(qz,z,Q1D)
+               {
+                  const double detJ = J(qx,qy,qz,e);
+                  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 * (by_val ? detJ : 1.0/detJ);
+               }
+            }
+         }
+      });
+   }
+}
+
+template<int T_D1D = 0, int T_Q1D = 0>
+static void PAMassAssembleDiagonal2D(const int NE,
+                                     const Array<double> &b,
+                                     const Vector &d,
+                                     Vector &y,
+                                     const int d1d = 0,
+                                     const int q1d = 0)
+{
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+   MFEM_VERIFY(D1D <= MAX_D1D, "");
+   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
+   auto B = Reshape(b.Read(), Q1D, D1D);
+   auto D = Reshape(d.Read(), Q1D, Q1D, NE);
+   auto Y = Reshape(y.ReadWrite(), D1D, D1D, NE);
+   MFEM_FORALL(e, NE,
+   {
+      const int D1D = T_D1D ? T_D1D : d1d;
+      const int Q1D = T_Q1D ? T_Q1D : q1d;
+      constexpr int MD1 = T_D1D ? T_D1D : MAX_D1D;
+      constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
+      double QD[MQ1][MD1];
+      for (int qx = 0; qx < Q1D; ++qx)
+      {
+         for (int dy = 0; dy < D1D; ++dy)
+         {
+            QD[qx][dy] = 0.0;
+            for (int qy = 0; qy < Q1D; ++qy)
+            {
+               QD[qx][dy] += B(qy, dy) * B(qy, dy) * D(qx, qy, e);
+            }
+         }
+      }
+      for (int dy = 0; dy < D1D; ++dy)
+      {
+         for (int dx = 0; dx < D1D; ++dx)
+         {
+            for (int qx = 0; qx < Q1D; ++qx)
+            {
+               Y(dx,dy,e) += B(qx, dx) * B(qx, dx) * QD[qx][dy];
+            }
+         }
+      }
+   });
+}
+
+template<int T_D1D = 0, int T_Q1D = 0, int T_NBZ = 0>
+static void SmemPAMassAssembleDiagonal2D(const int NE,
+                                         const Array<double> &b_,
+                                         const Vector &d_,
+                                         Vector &y_,
+                                         const int d1d = 0,
+                                         const int q1d = 0)
+{
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+   constexpr int NBZ = T_NBZ ? T_NBZ : 1;
+   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 D = Reshape(d_.Read(), Q1D, Q1D, NE);
+   auto Y = Reshape(y_.ReadWrite(), D1D, D1D, NE);
+   MFEM_FORALL_2D(e, NE, Q1D, Q1D, NBZ,
+   {
+      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 NBZ = T_NBZ ? T_NBZ : 1;
+      constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
+      constexpr int MD1 = T_D1D ? T_D1D : MAX_D1D;
+      MFEM_SHARED double B[MQ1][MD1];
+      MFEM_SHARED double QDZ[NBZ][MQ1][MD1];
+      double (*QD)[MD1] = (double (*)[MD1])(QDZ + tidz);
+      if (tidz == 0)
+      {
+         MFEM_FOREACH_THREAD(d,y,D1D)
+         {
+            MFEM_FOREACH_THREAD(q,x,Q1D)
+            {
+               B[q][d] = b(q,d);
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(qx,x,Q1D)
+      {
+         MFEM_FOREACH_THREAD(dy,y,D1D)
+         {
+            QD[qx][dy] = 0.0;
+            for (int qy = 0; qy < Q1D; ++qy)
+            {
+               QD[qx][dy] += B[qy][dy] * B[qy][dy] * D(qx, qy, e);
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(dy,y,D1D)
+      {
+         MFEM_FOREACH_THREAD(dx,x,D1D)
+         {
+            for (int qx = 0; qx < Q1D; ++qx)
+            {
+               // might need absolute values on next line
+               Y(dx,dy,e) += B[qx][dx] * B[qx][dx] * QD[qx][dy];
+            }
+         }
+      }
+   });
+}
+
+template<int T_D1D = 0, int T_Q1D = 0>
+static void PAMassAssembleDiagonal3D(const int NE,
+                                     const Array<double> &b,
+                                     const Vector &d,
+                                     Vector &y,
+                                     const int d1d = 0,
+                                     const int q1d = 0)
+{
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+   MFEM_VERIFY(D1D <= MAX_D1D, "");
+   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
+   auto B = Reshape(b.Read(), Q1D, D1D);
+   auto D = Reshape(d.Read(), Q1D, Q1D, Q1D, NE);
+   auto Y = Reshape(y.ReadWrite(), D1D, D1D, D1D, NE);
+   MFEM_FORALL(e, NE,
+   {
+      const int D1D = T_D1D ? T_D1D : d1d;
+      const int Q1D = T_Q1D ? T_Q1D : q1d;
+      constexpr int MD1 = T_D1D ? T_D1D : MAX_D1D;
+      constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
+      double QQD[MQ1][MQ1][MD1];
+      double QDD[MQ1][MD1][MD1];
+      for (int qx = 0; qx < Q1D; ++qx)
+      {
+         for (int qy = 0; qy < Q1D; ++qy)
+         {
+            for (int dz = 0; dz < D1D; ++dz)
+            {
+               QQD[qx][qy][dz] = 0.0;
+               for (int qz = 0; qz < Q1D; ++qz)
+               {
+                  QQD[qx][qy][dz] += B(qz, dz) * B(qz, dz) * D(qx, qy, qz, e);
+               }
+            }
+         }
+      }
+      for (int qx = 0; qx < Q1D; ++qx)
+      {
+         for (int dz = 0; dz < D1D; ++dz)
+         {
+            for (int dy = 0; dy < D1D; ++dy)
+            {
+               QDD[qx][dy][dz] = 0.0;
+               for (int qy = 0; qy < Q1D; ++qy)
+               {
+                  QDD[qx][dy][dz] += B(qy, dy) * B(qy, dy) * QQD[qx][qy][dz];
+               }
+            }
+         }
+      }
+      for (int dz = 0; dz < D1D; ++dz)
+      {
+         for (int dy = 0; dy < D1D; ++dy)
+         {
+            for (int dx = 0; dx < D1D; ++dx)
+            {
+               double t = 0.0;
+               for (int qx = 0; qx < Q1D; ++qx)
+               {
+                  t += B(qx, dx) * B(qx, dx) * QDD[qx][dy][dz];
+               }
+               Y(dx, dy, dz, e) += t;
+            }
+         }
+      }
+   });
+}
+
+template<int T_D1D = 0, int T_Q1D = 0>
+static void SmemPAMassAssembleDiagonal3D(const int NE,
+                                         const Array<double> &b_,
+                                         const Vector &d_,
+                                         Vector &y_,
+                                         const int d1d = 0,
+                                         const int q1d = 0)
+{
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+   constexpr int 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 D = Reshape(d_.Read(), Q1D, Q1D, Q1D, NE);
+   auto Y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, NE);
+   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 MD1 = T_D1D ? T_D1D : MAX_D1D;
+      constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
+      MFEM_SHARED double B[MQ1][MD1];
+      MFEM_SHARED double QQD[MQ1][MQ1][MD1];
+      MFEM_SHARED double QDD[MQ1][MD1][MD1];
+      if (tidz == 0)
+      {
+         MFEM_FOREACH_THREAD(d,y,D1D)
+         {
+            MFEM_FOREACH_THREAD(q,x,Q1D)
+            {
+               B[q][d] = b(q,d);
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(qx,x,Q1D)
+      {
+         MFEM_FOREACH_THREAD(qy,y,Q1D)
+         {
+            MFEM_FOREACH_THREAD(dz,z,D1D)
+            {
+               QQD[qx][qy][dz] = 0.0;
+               for (int qz = 0; qz < Q1D; ++qz)
+               {
+                  QQD[qx][qy][dz] += B[qz][dz] * B[qz][dz] * D(qx, qy, qz, e);
+               }
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(qx,x,Q1D)
+      {
+         MFEM_FOREACH_THREAD(dz,z,D1D)
+         {
+            MFEM_FOREACH_THREAD(dy,y,D1D)
+            {
+               QDD[qx][dy][dz] = 0.0;
+               for (int qy = 0; qy < Q1D; ++qy)
+               {
+                  QDD[qx][dy][dz] += B[qy][dy] * B[qy][dy] * QQD[qx][qy][dz];
+               }
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(dz,z,D1D)
+      {
+         MFEM_FOREACH_THREAD(dy,y,D1D)
+         {
+            MFEM_FOREACH_THREAD(dx,x,D1D)
+            {
+               double t = 0.0;
+               for (int qx = 0; qx < Q1D; ++qx)
+               {
+                  t += B[qx][dx] * B[qx][dx] * QDD[qx][dy][dz];
+               }
+               Y(dx, dy, dz, e) += t;
+            }
+         }
+      }
+   });
+}
+
+static void PAMassAssembleDiagonal(const int dim, const int D1D,
+                                   const int Q1D, const int NE,
+                                   const Array<double> &B,
+                                   const Vector &D,
+                                   Vector &Y)
+{
+   if (dim == 2)
+   {
+      switch ((D1D << 4 ) | Q1D)
+      {
+         case 0x22: return SmemPAMassAssembleDiagonal2D<2,2,16>(NE,B,D,Y);
+         case 0x33: return SmemPAMassAssembleDiagonal2D<3,3,16>(NE,B,D,Y);
+         case 0x44: return SmemPAMassAssembleDiagonal2D<4,4,8>(NE,B,D,Y);
+         case 0x55: return SmemPAMassAssembleDiagonal2D<5,5,8>(NE,B,D,Y);
+         case 0x66: return SmemPAMassAssembleDiagonal2D<6,6,4>(NE,B,D,Y);
+         case 0x77: return SmemPAMassAssembleDiagonal2D<7,7,4>(NE,B,D,Y);
+         case 0x88: return SmemPAMassAssembleDiagonal2D<8,8,2>(NE,B,D,Y);
+         case 0x99: return SmemPAMassAssembleDiagonal2D<9,9,2>(NE,B,D,Y);
+         default:   return PAMassAssembleDiagonal2D(NE,B,D,Y,D1D,Q1D);
+      }
+   }
+   else if (dim == 3)
+   {
+      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);
+         case 0x89: return SmemPAMassAssembleDiagonal3D<8,9>(NE,B,D,Y);
+         default:   return PAMassAssembleDiagonal3D(NE,B,D,Y,D1D,Q1D);
+      }
+   }
+   MFEM_ABORT("Unknown kernel.");
+}
+
+void MassIntegrator::AssembleDiagonalPA(Vector &diag)
+{
+   if (DeviceCanUseCeed())
+   {
+      ceedOp->GetDiagonal(diag);
+   }
+   else
+   {
+      PAMassAssembleDiagonal(dim, dofs1D, quad1D, ne, maps->B, pa_data, diag);
+   }
+}
+
+
+#ifdef MFEM_USE_OCCA
+// OCCA PA Mass Apply 2D kernel
+static void OccaPAMassApply2D(const int D1D,
+                              const int Q1D,
+                              const int NE,
+                              const Array<double> &B,
+                              const Array<double> &Bt,
+                              const Vector &D,
+                              const Vector &X,
+                              Vector &Y)
+{
+   occa::properties props;
+   props["defines/D1D"] = D1D;
+   props["defines/Q1D"] = Q1D;
+   const occa::memory o_B = OccaMemoryRead(B.GetMemory(), B.Size());
+   const occa::memory o_Bt = OccaMemoryRead(Bt.GetMemory(), Bt.Size());
+   const occa::memory o_D = OccaMemoryRead(D.GetMemory(), D.Size());
+   const occa::memory o_X = OccaMemoryRead(X.GetMemory(), X.Size());
+   occa::memory o_Y = OccaMemoryReadWrite(Y.GetMemory(), Y.Size());
+   const occa_id_t id = std::make_pair(D1D,Q1D);
+   if (!Device::Allows(Backend::OCCA_CUDA))
+   {
+      static occa_kernel_t OccaMassApply2D_cpu;
+      if (OccaMassApply2D_cpu.find(id) == OccaMassApply2D_cpu.end())
+      {
+         const occa::kernel MassApply2D_CPU =
+            mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
+                                        "MassApply2D_CPU", props);
+         OccaMassApply2D_cpu.emplace(id, MassApply2D_CPU);
+      }
+      OccaMassApply2D_cpu.at(id)(NE, o_B, o_Bt, o_D, o_X, o_Y);
+   }
+   else
+   {
+      static occa_kernel_t OccaMassApply2D_gpu;
+      if (OccaMassApply2D_gpu.find(id) == OccaMassApply2D_gpu.end())
+      {
+         const occa::kernel MassApply2D_GPU =
+            mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
+                                        "MassApply2D_GPU", props);
+         OccaMassApply2D_gpu.emplace(id, MassApply2D_GPU);
+      }
+      OccaMassApply2D_gpu.at(id)(NE, o_B, o_Bt, o_D, o_X, o_Y);
+   }
+}
+
+// OCCA PA Mass Apply 3D kernel
+static void OccaPAMassApply3D(const int D1D,
+                              const int Q1D,
+                              const int NE,
+                              const Array<double> &B,
+                              const Array<double> &Bt,
+                              const Vector &D,
+                              const Vector &X,
+                              Vector &Y)
+{
+   occa::properties props;
+   props["defines/D1D"] = D1D;
+   props["defines/Q1D"] = Q1D;
+   const occa::memory o_B = OccaMemoryRead(B.GetMemory(), B.Size());
+   const occa::memory o_Bt = OccaMemoryRead(Bt.GetMemory(), Bt.Size());
+   const occa::memory o_D = OccaMemoryRead(D.GetMemory(), D.Size());
+   const occa::memory o_X = OccaMemoryRead(X.GetMemory(), X.Size());
+   occa::memory o_Y = OccaMemoryReadWrite(Y.GetMemory(), Y.Size());
+   const occa_id_t id = std::make_pair(D1D,Q1D);
+   if (!Device::Allows(Backend::OCCA_CUDA))
+   {
+      static occa_kernel_t OccaMassApply3D_cpu;
+      if (OccaMassApply3D_cpu.find(id) == OccaMassApply3D_cpu.end())
+      {
+         const occa::kernel MassApply3D_CPU =
+            mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
+                                        "MassApply3D_CPU", props);
+         OccaMassApply3D_cpu.emplace(id, MassApply3D_CPU);
+      }
+      OccaMassApply3D_cpu.at(id)(NE, o_B, o_Bt, o_D, o_X, o_Y);
+   }
+   else
+   {
+      static occa_kernel_t OccaMassApply3D_gpu;
+      if (OccaMassApply3D_gpu.find(id) == OccaMassApply3D_gpu.end())
+      {
+         const occa::kernel MassApply3D_GPU =
+            mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
+                                        "MassApply3D_GPU", props);
+         OccaMassApply3D_gpu.emplace(id, MassApply3D_GPU);
+      }
+      OccaMassApply3D_gpu.at(id)(NE, o_B, o_Bt, o_D, o_X, o_Y);
+   }
+}
+#endif // MFEM_USE_OCCA
+
+template<int T_D1D = 0, int T_Q1D = 0>
+static void PAMassApply2D(const int NE,
+                          const Array<double> &b_,
+                          const Array<double> &bt_,
+                          const Vector &d_,
+                          const Vector &x_,
+                          Vector &y_,
+                          const int d1d = 0,
+                          const int q1d = 0)
+{
+   MFEM_VERIFY(T_D1D ? T_D1D : d1d <= MAX_D1D, "");
+   MFEM_VERIFY(T_Q1D ? T_Q1D : q1d <= MAX_Q1D, "");
+
+   const auto B = b_.Read();
+   const auto Bt = bt_.Read();
+   const auto D = d_.Read();
+   const auto X = x_.Read();
+   auto Y = y_.ReadWrite();
+
+   MFEM_FORALL(e, NE,
+   {
+      internal::PAMassApply2D_Element(e, NE, B, Bt, D, X, Y, d1d, q1d);
+   });
+}
+
+template<int T_D1D = 0, int T_Q1D = 0, int T_NBZ = 0>
+static void SmemPAMassApply2D(const int NE,
+                              const Array<double> &b_,
+                              const Array<double> &bt_,
+                              const Vector &d_,
+                              const Vector &x_,
+                              Vector &y_,
+                              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;
+   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, "");
+   const auto b = b_.Read();
+   const auto D = d_.Read();
+   const auto x = x_.Read();
+   auto Y = y_.ReadWrite();
+   MFEM_FORALL_2D(e, NE, Q1D, Q1D, NBZ,
+   {
+      internal::SmemPAMassApply2D_Element<T_D1D,T_Q1D,T_NBZ>(e, NE, b, D, x, Y, d1d, q1d);
+   });
+}
+
+template<int T_D1D = 0, int T_Q1D = 0>
+static void PAMassApply3D(const int NE,
+                          const Array<double> &b_,
+                          const Array<double> &bt_,
+                          const Vector &d_,
+                          const Vector &x_,
+                          Vector &y_,
+                          const int d1d = 0,
+                          const int q1d = 0)
+{
+   MFEM_VERIFY(T_D1D ? T_D1D : d1d <= MAX_D1D, "");
+   MFEM_VERIFY(T_Q1D ? T_Q1D : q1d <= MAX_Q1D, "");
+
+   const auto B = b_.Read();
+   const auto Bt = bt_.Read();
+   const auto D = d_.Read();
+   const auto X = x_.Read();
+   auto Y = y_.ReadWrite();
+
+   MFEM_FORALL(e, NE,
+   {
+      internal::PAMassApply3D_Element(e, NE, B, Bt, D, X, Y, d1d, q1d);
+   });
+}
+
+template<int T_D1D = 0, int T_Q1D = 0>
+static void SmemPAMassApply3D(const int NE,
+                              const Array<double> &b_,
+                              const Array<double> &bt_,
+                              const Vector &d_,
+                              const Vector &x_,
+                              Vector &y_,
+                              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;
+   constexpr int M1D = T_D1D ? T_D1D : MAX_D1D;
+   MFEM_VERIFY(D1D <= M1D, "");
+   MFEM_VERIFY(Q1D <= M1Q, "");
+   auto b = b_.Read();
+   auto d = d_.Read();
+   auto x = x_.Read();
+   auto y = y_.ReadWrite();
+   MFEM_FORALL_3D(e, NE, Q1D, Q1D, 1,
+   {
+      internal::SmemPAMassApply3D_Element<T_D1D,T_Q1D>(e, NE, b, d, x, y, d1d, q1d);
+   });
+}
+
+static void PAMassApply(const int dim,
+                        const int D1D,
+                        const int Q1D,
+                        const int NE,
+                        const Array<double> &B,
+                        const Array<double> &Bt,
+                        const Vector &D,
+                        const Vector &X,
+                        Vector &Y)
+{
+#ifdef MFEM_USE_OCCA
+   if (DeviceCanUseOcca())
+   {
+      if (dim == 2)
+      {
+         return OccaPAMassApply2D(D1D,Q1D,NE,B,Bt,D,X,Y);
+      }
+      if (dim == 3)
+      {
+         return OccaPAMassApply3D(D1D,Q1D,NE,B,Bt,D,X,Y);
+      }
+      MFEM_ABORT("OCCA PA Mass Apply unknown kernel!");
+   }
+#endif // MFEM_USE_OCCA
+   const int id = (D1D << 4) | Q1D;
+
+   if (dim == 2)
+   {
+      switch (id)
+      {
+         case 0x22: return SmemPAMassApply2D<2,2,16>(NE,B,Bt,D,X,Y);
+         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);
+         case 0x88: return SmemPAMassApply2D<8,8,2>(NE,B,Bt,D,X,Y);
+         case 0x99: return SmemPAMassApply2D<9,9,2>(NE,B,Bt,D,X,Y);
+         default:   return PAMassApply2D(NE,B,Bt,D,X,Y,D1D,Q1D);
+      }
+   }
+   else if (dim == 3)
+   {
+      switch (id)
+      {
+         case 0x22: return SmemPAMassApply3D<2,2>(NE,B,Bt,D,X,Y);
+         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 0x26: return SmemPAMassApply3D<2,6>(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);
+         case 0x56: return SmemPAMassApply3D<5,6>(NE,B,Bt,D,X,Y);
+         case 0x58: return SmemPAMassApply3D<5,8>(NE,B,Bt,D,X,Y);
+         case 0x67: return SmemPAMassApply3D<6,7>(NE,B,Bt,D,X,Y);
+         case 0x78: return SmemPAMassApply3D<7,8>(NE,B,Bt,D,X,Y);
+         case 0x89: return SmemPAMassApply3D<8,9>(NE,B,Bt,D,X,Y);
+         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 kernel 0x" << std::hex << id << std::endl;
+   MFEM_ABORT("Unknown kernel.");
+}
+
+void MassIntegrator::AddMultPA(const Vector &x, Vector &y) const
+{
+   if (DeviceCanUseCeed())
+   {
+      ceedOp->AddMult(x, y);
+   }
+   else
+   {
+      PAMassApply(dim, dofs1D, quad1D, ne, maps->B, maps->Bt, pa_data, x, y);
+   }
+}
+
+void MassIntegrator::AddMultTransposePA(const Vector &x, Vector &y) const
+{
+   // Mass integrator is symmetric
+   AddMultPA(x, y);
+}
+
+} // namespace mfem

fem/bilininteg_mass_pa.hpp

@@ -9,15 +9,12 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#ifndef MFEM_BILININTEG_MASS_KERNELS_HPP
-#define MFEM_BILININTEG_MASS_KERNELS_HPP
+#ifndef MFEM_BILININTEG_MASS_PA_HPP
+#define MFEM_BILININTEG_MASS_PA_HPP
 
-#include "../../config/config.hpp"
-#include "../../general/array.hpp"
-#include "../../general/forall.hpp"
-#include "../../linalg/dtensor.hpp"
-#include "../../linalg/vector.hpp"
-#include "../bilininteg.hpp"
+#include "../config/config.hpp"
+#include "../general/forall.hpp"
+#include "../linalg/dtensor.hpp"
 
 namespace mfem
 {
@@ -25,315 +22,6 @@ namespace mfem
 namespace internal
 {
 
-void PAMassAssembleDiagonal(const int dim, const int D1D,
-                            const int Q1D, const int NE,
-                            const Array<double> &B,
-                            const Vector &D,
-                            Vector &Y);
-
-// PA Mass Diagonal 2D kernel
-template<int T_D1D = 0, int T_Q1D = 0>
-inline void PAMassAssembleDiagonal2D(const int NE,
-                                     const Array<double> &b,
-                                     const Vector &d,
-                                     Vector &y,
-                                     const int d1d = 0,
-                                     const int q1d = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   MFEM_VERIFY(D1D <= MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto D = Reshape(d.Read(), Q1D, Q1D, NE);
-   auto Y = Reshape(y.ReadWrite(), D1D, D1D, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
-   {
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      constexpr int MD1 = T_D1D ? T_D1D : MAX_D1D;
-      constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
-      double QD[MQ1][MD1];
-      for (int qx = 0; qx < Q1D; ++qx)
-      {
-         for (int dy = 0; dy < D1D; ++dy)
-         {
-            QD[qx][dy] = 0.0;
-            for (int qy = 0; qy < Q1D; ++qy)
-            {
-               QD[qx][dy] += B(qy, dy) * B(qy, dy) * D(qx, qy, e);
-            }
-         }
-      }
-      for (int dy = 0; dy < D1D; ++dy)
-      {
-         for (int dx = 0; dx < D1D; ++dx)
-         {
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               Y(dx,dy,e) += B(qx, dx) * B(qx, dx) * QD[qx][dy];
-            }
-         }
-      }
-   });
-}
-
-// Shared memory PA Mass Diagonal 2D kernel
-template<int T_D1D = 0, int T_Q1D = 0, int T_NBZ = 0>
-inline void SmemPAMassAssembleDiagonal2D(const int NE,
-                                         const Array<double> &b_,
-                                         const Vector &d_,
-                                         Vector &y_,
-                                         const int d1d = 0,
-                                         const int q1d = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   constexpr int NBZ = T_NBZ ? T_NBZ : 1;
-   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 D = Reshape(d_.Read(), Q1D, Q1D, NE);
-   auto Y = Reshape(y_.ReadWrite(), D1D, D1D, NE);
-   mfem::forall_2D_batch(NE, Q1D, Q1D, NBZ, [=] MFEM_HOST_DEVICE (int e)
-   {
-      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 NBZ = T_NBZ ? T_NBZ : 1;
-      constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
-      constexpr int MD1 = T_D1D ? T_D1D : MAX_D1D;
-      MFEM_SHARED double B[MQ1][MD1];
-      MFEM_SHARED double QDZ[NBZ][MQ1][MD1];
-      double (*QD)[MD1] = (double (*)[MD1])(QDZ + tidz);
-      if (tidz == 0)
-      {
-         MFEM_FOREACH_THREAD(d,y,D1D)
-         {
-            MFEM_FOREACH_THREAD(q,x,Q1D)
-            {
-               B[q][d] = b(q,d);
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(qx,x,Q1D)
-      {
-         MFEM_FOREACH_THREAD(dy,y,D1D)
-         {
-            QD[qx][dy] = 0.0;
-            for (int qy = 0; qy < Q1D; ++qy)
-            {
-               QD[qx][dy] += B[qy][dy] * B[qy][dy] * D(qx, qy, e);
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(dy,y,D1D)
-      {
-         MFEM_FOREACH_THREAD(dx,x,D1D)
-         {
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               // might need absolute values on next line
-               Y(dx,dy,e) += B[qx][dx] * B[qx][dx] * QD[qx][dy];
-            }
-         }
-      }
-   });
-}
-
-// PA Mass Diagonal 3D kernel
-template<int T_D1D = 0, int T_Q1D = 0>
-inline void PAMassAssembleDiagonal3D(const int NE,
-                                     const Array<double> &b,
-                                     const Vector &d,
-                                     Vector &y,
-                                     const int d1d = 0,
-                                     const int q1d = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   MFEM_VERIFY(D1D <= MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto D = Reshape(d.Read(), Q1D, Q1D, Q1D, NE);
-   auto Y = Reshape(y.ReadWrite(), D1D, D1D, D1D, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
-   {
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      constexpr int MD1 = T_D1D ? T_D1D : MAX_D1D;
-      constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
-      double QQD[MQ1][MQ1][MD1];
-      double QDD[MQ1][MD1][MD1];
-      for (int qx = 0; qx < Q1D; ++qx)
-      {
-         for (int qy = 0; qy < Q1D; ++qy)
-         {
-            for (int dz = 0; dz < D1D; ++dz)
-            {
-               QQD[qx][qy][dz] = 0.0;
-               for (int qz = 0; qz < Q1D; ++qz)
-               {
-                  QQD[qx][qy][dz] += B(qz, dz) * B(qz, dz) * D(qx, qy, qz, e);
-               }
-            }
-         }
-      }
-      for (int qx = 0; qx < Q1D; ++qx)
-      {
-         for (int dz = 0; dz < D1D; ++dz)
-         {
-            for (int dy = 0; dy < D1D; ++dy)
-            {
-               QDD[qx][dy][dz] = 0.0;
-               for (int qy = 0; qy < Q1D; ++qy)
-               {
-                  QDD[qx][dy][dz] += B(qy, dy) * B(qy, dy) * QQD[qx][qy][dz];
-               }
-            }
-         }
-      }
-      for (int dz = 0; dz < D1D; ++dz)
-      {
-         for (int dy = 0; dy < D1D; ++dy)
-         {
-            for (int dx = 0; dx < D1D; ++dx)
-            {
-               double t = 0.0;
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  t += B(qx, dx) * B(qx, dx) * QDD[qx][dy][dz];
-               }
-               Y(dx, dy, dz, e) += t;
-            }
-         }
-      }
-   });
-}
-
-// Shared memory PA Mass Diagonal 3D kernel
-template<int T_D1D = 0, int T_Q1D = 0>
-inline void SmemPAMassAssembleDiagonal3D(const int NE,
-                                         const Array<double> &b_,
-                                         const Vector &d_,
-                                         Vector &y_,
-                                         const int d1d = 0,
-                                         const int q1d = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   constexpr int 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 D = Reshape(d_.Read(), Q1D, Q1D, Q1D, NE);
-   auto Y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, NE);
-   mfem::forall_3D(NE, Q1D, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
-   {
-      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 MD1 = T_D1D ? T_D1D : MAX_D1D;
-      constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
-      MFEM_SHARED double B[MQ1][MD1];
-      MFEM_SHARED double QQD[MQ1][MQ1][MD1];
-      MFEM_SHARED double QDD[MQ1][MD1][MD1];
-      if (tidz == 0)
-      {
-         MFEM_FOREACH_THREAD(d,y,D1D)
-         {
-            MFEM_FOREACH_THREAD(q,x,Q1D)
-            {
-               B[q][d] = b(q,d);
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(qx,x,Q1D)
-      {
-         MFEM_FOREACH_THREAD(qy,y,Q1D)
-         {
-            MFEM_FOREACH_THREAD(dz,z,D1D)
-            {
-               QQD[qx][qy][dz] = 0.0;
-               for (int qz = 0; qz < Q1D; ++qz)
-               {
-                  QQD[qx][qy][dz] += B[qz][dz] * B[qz][dz] * D(qx, qy, qz, e);
-               }
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(qx,x,Q1D)
-      {
-         MFEM_FOREACH_THREAD(dz,z,D1D)
-         {
-            MFEM_FOREACH_THREAD(dy,y,D1D)
-            {
-               QDD[qx][dy][dz] = 0.0;
-               for (int qy = 0; qy < Q1D; ++qy)
-               {
-                  QDD[qx][dy][dz] += B[qy][dy] * B[qy][dy] * QQD[qx][qy][dz];
-               }
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(dz,z,D1D)
-      {
-         MFEM_FOREACH_THREAD(dy,y,D1D)
-         {
-            MFEM_FOREACH_THREAD(dx,x,D1D)
-            {
-               double t = 0.0;
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  t += B[qx][dx] * B[qx][dx] * QDD[qx][dy][dz];
-               }
-               Y(dx, dy, dz, e) += t;
-            }
-         }
-      }
-   });
-}
-
-void PAMassApply(const int dim,
-                 const int D1D,
-                 const int Q1D,
-                 const int NE,
-                 const Array<double> &B,
-                 const Array<double> &Bt,
-                 const Vector &D,
-                 const Vector &X,
-                 Vector &Y);
-
-#ifdef MFEM_USE_OCCA
-// OCCA PA Mass Apply 2D kernel
-void OccaPAMassApply2D(const int D1D,
-                       const int Q1D,
-                       const int NE,
-                       const Array<double> &B,
-                       const Array<double> &Bt,
-                       const Vector &D,
-                       const Vector &X,
-                       Vector &Y);
-
-// OCCA PA Mass Apply 3D kernel
-void OccaPAMassApply3D(const int D1D,
-                       const int Q1D,
-                       const int NE,
-                       const Array<double> &B,
-                       const Array<double> &Bt,
-                       const Vector &D,
-                       const Vector &X,
-                       Vector &Y);
-#endif // MFEM_USE_OCCA
-
 template <bool ACCUMULATE = true>
 MFEM_HOST_DEVICE inline
 void PAMassApply2D_Element(const int e,
@@ -937,116 +625,6 @@ void SmemPAMassApply3D_Element(const int e,
    MFEM_SYNC_THREAD;
 }
 
-// PA Mass Apply 2D kernel
-template<int T_D1D = 0, int T_Q1D = 0>
-inline void PAMassApply2D(const int NE,
-                          const Array<double> &b_,
-                          const Array<double> &bt_,
-                          const Vector &d_,
-                          const Vector &x_,
-                          Vector &y_,
-                          const int d1d = 0,
-                          const int q1d = 0)
-{
-   MFEM_VERIFY(T_D1D ? T_D1D : d1d <= MAX_D1D, "");
-   MFEM_VERIFY(T_Q1D ? T_Q1D : q1d <= MAX_Q1D, "");
-
-   const auto B = b_.Read();
-   const auto Bt = bt_.Read();
-   const auto D = d_.Read();
-   const auto X = x_.Read();
-   auto Y = y_.ReadWrite();
-
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
-   {
-      internal::PAMassApply2D_Element(e, NE, B, Bt, D, X, Y, d1d, q1d);
-   });
-}
-
-// Shared memory PA Mass Apply 2D kernel
-template<int T_D1D = 0, int T_Q1D = 0, int T_NBZ = 0>
-inline void SmemPAMassApply2D(const int NE,
-                              const Array<double> &b_,
-                              const Array<double> &bt_,
-                              const Vector &d_,
-                              const Vector &x_,
-                              Vector &y_,
-                              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;
-   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, "");
-   const auto b = b_.Read();
-   const auto D = d_.Read();
-   const auto x = x_.Read();
-   auto Y = y_.ReadWrite();
-   mfem::forall_2D_batch(NE, Q1D, Q1D, NBZ, [=] MFEM_HOST_DEVICE (int e)
-   {
-      internal::SmemPAMassApply2D_Element<T_D1D,T_Q1D,T_NBZ>(e, NE, b, D, x, Y, d1d,
-                                                             q1d);
-   });
-}
-
-// PA Mass Apply 3D kernel
-template<int T_D1D = 0, int T_Q1D = 0>
-inline void PAMassApply3D(const int NE,
-                          const Array<double> &b_,
-                          const Array<double> &bt_,
-                          const Vector &d_,
-                          const Vector &x_,
-                          Vector &y_,
-                          const int d1d = 0,
-                          const int q1d = 0)
-{
-   MFEM_VERIFY(T_D1D ? T_D1D : d1d <= MAX_D1D, "");
-   MFEM_VERIFY(T_Q1D ? T_Q1D : q1d <= MAX_Q1D, "");
-
-   const auto B = b_.Read();
-   const auto Bt = bt_.Read();
-   const auto D = d_.Read();
-   const auto X = x_.Read();
-   auto Y = y_.ReadWrite();
-
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
-   {
-      internal::PAMassApply3D_Element(e, NE, B, Bt, D, X, Y, d1d, q1d);
-   });
-}
-
-// Shared memory PA Mass Apply 2D kernel
-template<int T_D1D = 0, int T_Q1D = 0>
-inline void SmemPAMassApply3D(const int NE,
-                              const Array<double> &b_,
-                              const Array<double> &bt_,
-                              const Vector &d_,
-                              const Vector &x_,
-                              Vector &y_,
-                              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;
-   constexpr int M1D = T_D1D ? T_D1D : MAX_D1D;
-   MFEM_VERIFY(D1D <= M1D, "");
-   MFEM_VERIFY(Q1D <= M1Q, "");
-   auto b = b_.Read();
-   auto d = d_.Read();
-   auto x = x_.Read();
-   auto y = y_.ReadWrite();
-   mfem::forall_2D(NE, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
-   {
-      internal::SmemPAMassApply3D_Element<T_D1D,T_Q1D>(e, NE, b, d, x, y, d1d, q1d);
-   });
-}
-
 } // namespace internal
 
 } // namespace mfem

fem/bilininteg_transpose_ea.cpp

@@ -9,8 +9,8 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../../general/forall.hpp"
-#include "../bilininteg.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
 
 namespace mfem
 {
@@ -27,7 +27,7 @@ void TransposeIntegrator::AssembleEA(const FiniteElementSpace &fes,
       const int dofs = fes.GetFE(0)->GetDof();
       auto A = Reshape(ea_data_tmp.Read(), dofs, dofs, ne);
       auto AT = Reshape(ea_data.ReadWrite(), dofs, dofs, ne);
-      mfem::forall(ne, [=] MFEM_HOST_DEVICE (int e)
+      MFEM_FORALL(e, ne,
       {
          for (int i = 0; i < dofs; i++)
          {
@@ -46,7 +46,7 @@ void TransposeIntegrator::AssembleEA(const FiniteElementSpace &fes,
       if (ne == 0) { return; }
       const int dofs = fes.GetFE(0)->GetDof();
       auto A = Reshape(ea_data.ReadWrite(), dofs, dofs, ne);
-      mfem::forall(ne, [=] MFEM_HOST_DEVICE (int e)
+      MFEM_FORALL(e, ne,
       {
          for (int i = 0; i < dofs; i++)
          {
@@ -80,7 +80,7 @@ void TransposeIntegrator::AssembleEAInteriorFaces(const FiniteElementSpace& fes,
       auto A_ext = Reshape(ea_data_ext_tmp.Read(), faceDofs, faceDofs, 2, nf);
       auto AT_int = Reshape(ea_data_int.ReadWrite(), faceDofs, faceDofs, 2, nf);
       auto AT_ext = Reshape(ea_data_ext.ReadWrite(), faceDofs, faceDofs, 2, nf);
-      mfem::forall(nf, [=] MFEM_HOST_DEVICE (int f)
+      MFEM_FORALL(f, nf,
       {
          for (int i = 0; i < faceDofs; i++)
          {
@@ -105,7 +105,7 @@ void TransposeIntegrator::AssembleEAInteriorFaces(const FiniteElementSpace& fes,
                                                fes.GetMesh()->GetFaceGeometry(0))->GetDof();
       auto A_int = Reshape(ea_data_int.ReadWrite(), faceDofs, faceDofs, 2, nf);
       auto A_ext = Reshape(ea_data_ext.ReadWrite(), faceDofs, faceDofs, 2, nf);
-      mfem::forall(nf, [=] MFEM_HOST_DEVICE (int f)
+      MFEM_FORALL(f, nf,
       {
          for (int i = 0; i < faceDofs; i++)
          {
@@ -149,7 +149,7 @@ void TransposeIntegrator::AssembleEABoundaryFaces(const FiniteElementSpace& fes,
                                                fes.GetMesh()->GetFaceGeometry(0))->GetDof();
       auto A_bdr = Reshape(ea_data_bdr_tmp.Read(), faceDofs, faceDofs, nf);
       auto AT_bdr = Reshape(ea_data_bdr.ReadWrite(), faceDofs, faceDofs, nf);
-      mfem::forall(nf, [=] MFEM_HOST_DEVICE (int f)
+      MFEM_FORALL(f, nf,
       {
          for (int i = 0; i < faceDofs; i++)
          {
@@ -167,7 +167,7 @@ void TransposeIntegrator::AssembleEABoundaryFaces(const FiniteElementSpace& fes,
       const int faceDofs = fes.GetTraceElement(0,
                                                fes.GetMesh()->GetFaceGeometry(0))->GetDof();
       auto A_bdr = Reshape(ea_data_bdr.ReadWrite(), faceDofs, faceDofs, nf);
-      mfem::forall(nf, [=] MFEM_HOST_DEVICE (int f)
+      MFEM_FORALL(f, nf,
       {
          for (int i = 0; i < faceDofs; i++)
          {

fem/bilininteg_vecdiffusion.cpp

@@ -9,15 +9,19 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../../general/forall.hpp"
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
-#include "../qfunction.hpp"
-#include "../ceed/integrators/diffusion/diffusion.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
+#include "qfunction.hpp"
+#include "ceed/integrators/diffusion/diffusion.hpp"
+
+using namespace std;
 
 namespace mfem
 {
 
+// PA Vector Diffusion Integrator
+
 // PA Diffusion Assemble 2D kernel
 static void PAVectorDiffusionSetup2D(const int Q1D,
                                      const int NE,
@@ -37,7 +41,7 @@ static void PAVectorDiffusionSetup2D(const int Q1D,
                   Reshape(c.Read(), NQ, NE);
 
 
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       for (int q = 0; q < NQ; ++q)
       {
@@ -73,7 +77,7 @@ static void PAVectorDiffusionSetup3D(const int Q1D,
                   Reshape(c.Read(), NQ,NE);
 
 
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       for (int q = 0; q < NQ; ++q)
       {
@@ -87,8 +91,8 @@ static void PAVectorDiffusionSetup3D(const int Q1D,
          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);
+         /* */               J21 * (J12 * J33 - J32 * J13) +
+         /* */               J31 * (J12 * J23 - J22 * J13);
 
          const double C1 = const_c ? C(0,0) : C(q,e);
 
@@ -193,7 +197,7 @@ void VectorDiffusionIntegrator::AssemblePA(const FiniteElementSpace &fes)
       const auto C = const_c ? Reshape(coeff.Read(), 1,1) :
                      Reshape(coeff.Read(), NQ,ne);
 
-      mfem::forall(ne, [=] MFEM_HOST_DEVICE (int e)
+      MFEM_FORALL(e, ne,
       {
          for (int q = 0; q < NQ; ++q)
          {
@@ -222,209 +226,6 @@ void VectorDiffusionIntegrator::AssemblePA(const FiniteElementSpace &fes)
    }
 }
 
-template<int T_D1D = 0, int T_Q1D = 0>
-static void PAVectorDiffusionDiagonal2D(const int NE,
-                                        const Array<double> &b,
-                                        const Array<double> &g,
-                                        const Vector &d,
-                                        Vector &y,
-                                        const int d1d = 0,
-                                        const int q1d = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   MFEM_VERIFY(D1D <= MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto G = Reshape(g.Read(), Q1D, D1D);
-   // note the different shape for D, this is a (symmetric) matrix so we only
-   // store necessary entries
-   auto D = Reshape(d.Read(), Q1D*Q1D, 3, NE);
-   auto Y = Reshape(y.ReadWrite(), D1D, D1D, 2, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
-   {
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      constexpr int MD1 = T_D1D ? T_D1D : MAX_D1D;
-      constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
-      // gradphi \cdot Q \gradphi has four terms
-      double QD0[MQ1][MD1];
-      double QD1[MQ1][MD1];
-      double QD2[MQ1][MD1];
-      for (int qx = 0; qx < Q1D; ++qx)
-      {
-         for (int dy = 0; dy < D1D; ++dy)
-         {
-            QD0[qx][dy] = 0.0;
-            QD1[qx][dy] = 0.0;
-            QD2[qx][dy] = 0.0;
-            for (int qy = 0; qy < Q1D; ++qy)
-            {
-               const int q = qx + qy * Q1D;
-               const double D0 = D(q,0,e);
-               const double D1 = D(q,1,e);
-               const double D2 = D(q,2,e);
-               QD0[qx][dy] += B(qy, dy) * B(qy, dy) * D0;
-               QD1[qx][dy] += B(qy, dy) * G(qy, dy) * D1;
-               QD2[qx][dy] += G(qy, dy) * G(qy, dy) * D2;
-            }
-         }
-      }
-      for (int dy = 0; dy < D1D; ++dy)
-      {
-         for (int dx = 0; dx < D1D; ++dx)
-         {
-            double temp = 0.0;
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               temp += G(qx, dx) * G(qx, dx) * QD0[qx][dy];
-               temp += G(qx, dx) * B(qx, dx) * QD1[qx][dy];
-               temp += B(qx, dx) * G(qx, dx) * QD1[qx][dy];
-               temp += B(qx, dx) * B(qx, dx) * QD2[qx][dy];
-            }
-            Y(dx,dy,0,e) += temp;
-            Y(dx,dy,1,e) += temp;
-         }
-      }
-   });
-}
-
-template<int T_D1D = 0, int T_Q1D = 0>
-static void PAVectorDiffusionDiagonal3D(const int NE,
-                                        const Array<double> &b,
-                                        const Array<double> &g,
-                                        const Vector &d,
-                                        Vector &y,
-                                        const int d1d = 0,
-                                        const int q1d = 0)
-{
-   constexpr int DIM = 3;
-   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;
-   MFEM_VERIFY(D1D <= MD1, "");
-   MFEM_VERIFY(Q1D <= MQ1, "");
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto G = Reshape(g.Read(), Q1D, D1D);
-   auto Q = Reshape(d.Read(), Q1D*Q1D*Q1D, 6, NE);
-   auto Y = Reshape(y.ReadWrite(), D1D, D1D, D1D, 3, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
-   {
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      constexpr int MD1 = T_D1D ? T_D1D : MAX_D1D;
-      constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
-      double QQD[MQ1][MQ1][MD1];
-      double QDD[MQ1][MD1][MD1];
-      for (int i = 0; i < DIM; ++i)
-      {
-         for (int j = 0; j < DIM; ++j)
-         {
-            // first tensor contraction, along z direction
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               for (int qy = 0; qy < Q1D; ++qy)
-               {
-                  for (int dz = 0; dz < D1D; ++dz)
-                  {
-                     QQD[qx][qy][dz] = 0.0;
-                     for (int qz = 0; qz < Q1D; ++qz)
-                     {
-                        const int q = qx + (qy + qz * Q1D) * Q1D;
-                        const int k = j >= i ?
-                                      3 - (3-i)*(2-i)/2 + j:
-                                      3 - (3-j)*(2-j)/2 + i;
-                        const double O = Q(q,k,e);
-                        const double Bz = B(qz,dz);
-                        const double Gz = G(qz,dz);
-                        const double L = i==2 ? Gz : Bz;
-                        const double R = j==2 ? Gz : Bz;
-                        QQD[qx][qy][dz] += L * O * R;
-                     }
-                  }
-               }
-            }
-            // second tensor contraction, along y direction
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               for (int dz = 0; dz < D1D; ++dz)
-               {
-                  for (int dy = 0; dy < D1D; ++dy)
-                  {
-                     QDD[qx][dy][dz] = 0.0;
-                     for (int qy = 0; qy < Q1D; ++qy)
-                     {
-                        const double By = B(qy,dy);
-                        const double Gy = G(qy,dy);
-                        const double L = i==1 ? Gy : By;
-                        const double R = j==1 ? Gy : By;
-                        QDD[qx][dy][dz] += L * QQD[qx][qy][dz] * R;
-                     }
-                  }
-               }
-            }
-            // third tensor contraction, along x direction
-            for (int dz = 0; dz < D1D; ++dz)
-            {
-               for (int dy = 0; dy < D1D; ++dy)
-               {
-                  for (int dx = 0; dx < D1D; ++dx)
-                  {
-                     double temp = 0.0;
-                     for (int qx = 0; qx < Q1D; ++qx)
-                     {
-                        const double Bx = B(qx,dx);
-                        const double Gx = G(qx,dx);
-                        const double L = i==0 ? Gx : Bx;
-                        const double R = j==0 ? Gx : Bx;
-                        temp += L * QDD[qx][dy][dz] * R;
-                     }
-                     Y(dx, dy, dz, 0, e) += temp;
-                     Y(dx, dy, dz, 1, e) += temp;
-                     Y(dx, dy, dz, 2, e) += temp;
-                  }
-               }
-            }
-         }
-      }
-   });
-}
-
-static void PAVectorDiffusionAssembleDiagonal(const int dim,
-                                              const int D1D,
-                                              const int Q1D,
-                                              const int NE,
-                                              const Array<double> &B,
-                                              const Array<double> &G,
-                                              const Vector &op,
-                                              Vector &y)
-{
-   if (dim == 2)
-   {
-      return PAVectorDiffusionDiagonal2D(NE, B, G, op, y, D1D, Q1D);
-   }
-   else if (dim == 3)
-   {
-      return PAVectorDiffusionDiagonal3D(NE, B, G, op, y, D1D, Q1D);
-   }
-   MFEM_ABORT("Dimension not implemented.");
-}
-
-void VectorDiffusionIntegrator::AssembleDiagonalPA(Vector &diag)
-{
-   if (DeviceCanUseCeed())
-   {
-      ceedOp->GetDiagonal(diag);
-   }
-   else
-   {
-      PAVectorDiffusionAssembleDiagonal(dim, dofs1D, quad1D, ne,
-                                        maps->B, maps->G,
-                                        pa_data, diag);
-   }
-}
-
 // PA Diffusion Apply 2D kernel
 template<int T_D1D = 0, int T_Q1D = 0, int T_VDIM = 0> static
 void PAVectorDiffusionApply2D(const int NE,
@@ -451,7 +252,7 @@ void PAVectorDiffusionApply2D(const int NE,
    auto D = Reshape(d_.Read(), Q1D*Q1D, 3, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, VDIM, NE);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, VDIM, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -572,7 +373,7 @@ void PAVectorDiffusionApply3D(const int NE,
    auto op = Reshape(op_.Read(), Q1D*Q1D*Q1D, 6, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, D1D, VDIM, NE);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, VDIM, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -780,4 +581,212 @@ void VectorDiffusionIntegrator::AddMultPA(const Vector &x, Vector &y) const
    }
 }
 
+template<int T_D1D = 0, int T_Q1D = 0>
+static void PAVectorDiffusionDiagonal2D(const int NE,
+                                        const Array<double> &b,
+                                        const Array<double> &g,
+                                        const Vector &d,
+                                        Vector &y,
+                                        const int d1d = 0,
+                                        const int q1d = 0)
+{
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+   MFEM_VERIFY(D1D <= MAX_D1D, "");
+   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
+   auto B = Reshape(b.Read(), Q1D, D1D);
+   auto G = Reshape(g.Read(), Q1D, D1D);
+   // note the different shape for D, this is a (symmetric) matrix so we only
+   // store necessary entries
+   auto D = Reshape(d.Read(), Q1D*Q1D, 3, NE);
+   auto Y = Reshape(y.ReadWrite(), D1D, D1D, 2, NE);
+   MFEM_FORALL(e, NE,
+   {
+      const int D1D = T_D1D ? T_D1D : d1d;
+      const int Q1D = T_Q1D ? T_Q1D : q1d;
+      constexpr int MD1 = T_D1D ? T_D1D : MAX_D1D;
+      constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
+      // gradphi \cdot Q \gradphi has four terms
+      double QD0[MQ1][MD1];
+      double QD1[MQ1][MD1];
+      double QD2[MQ1][MD1];
+      for (int qx = 0; qx < Q1D; ++qx)
+      {
+         for (int dy = 0; dy < D1D; ++dy)
+         {
+            QD0[qx][dy] = 0.0;
+            QD1[qx][dy] = 0.0;
+            QD2[qx][dy] = 0.0;
+            for (int qy = 0; qy < Q1D; ++qy)
+            {
+               const int q = qx + qy * Q1D;
+               const double D0 = D(q,0,e);
+               const double D1 = D(q,1,e);
+               const double D2 = D(q,2,e);
+               QD0[qx][dy] += B(qy, dy) * B(qy, dy) * D0;
+               QD1[qx][dy] += B(qy, dy) * G(qy, dy) * D1;
+               QD2[qx][dy] += G(qy, dy) * G(qy, dy) * D2;
+            }
+         }
+      }
+      for (int dy = 0; dy < D1D; ++dy)
+      {
+         for (int dx = 0; dx < D1D; ++dx)
+         {
+            double temp = 0.0;
+            for (int qx = 0; qx < Q1D; ++qx)
+            {
+               temp += G(qx, dx) * G(qx, dx) * QD0[qx][dy];
+               temp += G(qx, dx) * B(qx, dx) * QD1[qx][dy];
+               temp += B(qx, dx) * G(qx, dx) * QD1[qx][dy];
+               temp += B(qx, dx) * B(qx, dx) * QD2[qx][dy];
+            }
+            Y(dx,dy,0,e) += temp;
+            Y(dx,dy,1,e) += temp;
+         }
+      }
+   });
+}
+
+template<int T_D1D = 0, int T_Q1D = 0>
+static void PAVectorDiffusionDiagonal3D(const int NE,
+                                        const Array<double> &b,
+                                        const Array<double> &g,
+                                        const Vector &d,
+                                        Vector &y,
+                                        const int d1d = 0,
+                                        const int q1d = 0)
+{
+   constexpr int DIM = 3;
+   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;
+   MFEM_VERIFY(D1D <= MD1, "");
+   MFEM_VERIFY(Q1D <= MQ1, "");
+   auto B = Reshape(b.Read(), Q1D, D1D);
+   auto G = Reshape(g.Read(), Q1D, D1D);
+   auto Q = Reshape(d.Read(), Q1D*Q1D*Q1D, 6, NE);
+   auto Y = Reshape(y.ReadWrite(), D1D, D1D, D1D, 3, NE);
+   MFEM_FORALL(e, NE,
+   {
+      const int D1D = T_D1D ? T_D1D : d1d;
+      const int Q1D = T_Q1D ? T_Q1D : q1d;
+      constexpr int MD1 = T_D1D ? T_D1D : MAX_D1D;
+      constexpr int MQ1 = T_Q1D ? T_Q1D : MAX_Q1D;
+      double QQD[MQ1][MQ1][MD1];
+      double QDD[MQ1][MD1][MD1];
+      for (int i = 0; i < DIM; ++i)
+      {
+         for (int j = 0; j < DIM; ++j)
+         {
+            // first tensor contraction, along z direction
+            for (int qx = 0; qx < Q1D; ++qx)
+            {
+               for (int qy = 0; qy < Q1D; ++qy)
+               {
+                  for (int dz = 0; dz < D1D; ++dz)
+                  {
+                     QQD[qx][qy][dz] = 0.0;
+                     for (int qz = 0; qz < Q1D; ++qz)
+                     {
+                        const int q = qx + (qy + qz * Q1D) * Q1D;
+                        const int k = j >= i ?
+                        3 - (3-i)*(2-i)/2 + j:
+                        3 - (3-j)*(2-j)/2 + i;
+                        const double O = Q(q,k,e);
+                        const double Bz = B(qz,dz);
+                        const double Gz = G(qz,dz);
+                        const double L = i==2 ? Gz : Bz;
+                        const double R = j==2 ? Gz : Bz;
+                        QQD[qx][qy][dz] += L * O * R;
+                     }
+                  }
+               }
+            }
+            // second tensor contraction, along y direction
+            for (int qx = 0; qx < Q1D; ++qx)
+            {
+               for (int dz = 0; dz < D1D; ++dz)
+               {
+                  for (int dy = 0; dy < D1D; ++dy)
+                  {
+                     QDD[qx][dy][dz] = 0.0;
+                     for (int qy = 0; qy < Q1D; ++qy)
+                     {
+                        const double By = B(qy,dy);
+                        const double Gy = G(qy,dy);
+                        const double L = i==1 ? Gy : By;
+                        const double R = j==1 ? Gy : By;
+                        QDD[qx][dy][dz] += L * QQD[qx][qy][dz] * R;
+                     }
+                  }
+               }
+            }
+            // third tensor contraction, along x direction
+            for (int dz = 0; dz < D1D; ++dz)
+            {
+               for (int dy = 0; dy < D1D; ++dy)
+               {
+                  for (int dx = 0; dx < D1D; ++dx)
+                  {
+                     double temp = 0.0;
+                     for (int qx = 0; qx < Q1D; ++qx)
+                     {
+                        const double Bx = B(qx,dx);
+                        const double Gx = G(qx,dx);
+                        const double L = i==0 ? Gx : Bx;
+                        const double R = j==0 ? Gx : Bx;
+                        temp += L * QDD[qx][dy][dz] * R;
+                     }
+                     Y(dx, dy, dz, 0, e) += temp;
+                     Y(dx, dy, dz, 1, e) += temp;
+                     Y(dx, dy, dz, 2, e) += temp;
+                  }
+               }
+            }
+         }
+      }
+   });
+}
+
+static void PAVectorDiffusionAssembleDiagonal(const int dim,
+                                              const int D1D,
+                                              const int Q1D,
+                                              const int NE,
+                                              const Array<double> &B,
+                                              const Array<double> &G,
+                                              const Vector &op,
+                                              Vector &y)
+{
+   if (dim == 2)
+   {
+      return PAVectorDiffusionDiagonal2D(NE, B, G, op, y, D1D, Q1D);
+   }
+   else if (dim == 3)
+   {
+      return PAVectorDiffusionDiagonal3D(NE, B, G, op, y, D1D, Q1D);
+   }
+   MFEM_ABORT("Dimension not implemented.");
+}
+
+void VectorDiffusionIntegrator::AssembleDiagonalPA(Vector &diag)
+{
+   if (DeviceCanUseCeed())
+   {
+      ceedOp->GetDiagonal(diag);
+   }
+   else
+   {
+      PAVectorDiffusionAssembleDiagonal(dim,
+                                        dofs1D,
+                                        quad1D,
+                                        ne,
+                                        maps->B,
+                                        maps->G,
+                                        pa_data,
+                                        diag);
+   }
+}
+
 } // namespace mfem

fem/bilininteg_vecdiffusion_mf.cpp

@@ -9,9 +9,12 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
-#include "../ceed/integrators/diffusion/diffusion.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
+#include "ceed/integrators/diffusion/diffusion.hpp"
+
+using namespace std;
 
 namespace mfem
 {

fem/bilininteg_vecmass.cpp

@@ -9,14 +9,19 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../../general/forall.hpp"
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
-#include "../ceed/integrators/mass/mass.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
+#include "ceed/integrators/mass/mass.hpp"
+
+using namespace std;
 
 namespace mfem
 {
 
+// PA Mass Integrator
+
+// PA Mass Assemble kernel
 void VectorMassIntegrator::AssemblePA(const FiniteElementSpace &fes)
 {
    // Assuming the same element type
@@ -69,7 +74,7 @@ void VectorMassIntegrator::AssemblePA(const FiniteElementSpace &fes)
       auto w = ir->GetWeights().Read();
       auto J = Reshape(geom->J.Read(), NQ,2,2,NE);
       auto v = Reshape(pa_data.Write(), NQ, NE);
-      mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+      MFEM_FORALL(e, NE,
       {
          for (int q = 0; q < NQ; ++q)
          {
@@ -90,7 +95,7 @@ void VectorMassIntegrator::AssemblePA(const FiniteElementSpace &fes)
       auto W = ir->GetWeights().Read();
       auto J = Reshape(geom->J.Read(), NQ,3,3,NE);
       auto v = Reshape(pa_data.Write(), NQ,NE);
-      mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+      MFEM_FORALL(e, NE,
       {
          for (int q = 0; q < NQ; ++q)
          {
@@ -98,177 +103,16 @@ void VectorMassIntegrator::AssemblePA(const FiniteElementSpace &fes)
             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);
+            /* */               J21 * (J12 * J33 - J32 * J13) +
+            /* */               J31 * (J12 * J23 - J22 * J13);
             v(q,e) = W[q] * constant * detJ;
          }
       });
    }
 }
 
-template<const int T_D1D = 0, const int T_Q1D = 0>
-static void PAVectorMassAssembleDiagonal2D(const int NE,
-                                           const Array<double> &B_,
-                                           const Array<double> &Bt_,
-                                           const Vector &op_,
-                                           Vector &diag_,
-                                           const int d1d = 0,
-                                           const int q1d = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   constexpr int VDIM = 2;
-   MFEM_VERIFY(D1D <= MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
-   auto B = Reshape(B_.Read(), Q1D, D1D);
-   auto op = Reshape(op_.Read(), Q1D, Q1D, NE);
-   auto y = Reshape(diag_.ReadWrite(), D1D, D1D, VDIM, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
-   {
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      constexpr int max_D1D = T_D1D ? T_D1D : MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : MAX_Q1D;
-
-      double temp[max_Q1D][max_D1D];
-      for (int qx = 0; qx < Q1D; ++qx)
-      {
-         for (int dy = 0; dy < D1D; ++dy)
-         {
-            temp[qx][dy] = 0.0;
-            for (int qy = 0; qy < Q1D; ++qy)
-            {
-               temp[qx][dy] += B(qy, dy) * B(qy, dy) * op(qx, qy, e);
-            }
-         }
-      }
-      for (int dy = 0; dy < D1D; ++dy)
-      {
-         for (int dx = 0; dx < D1D; ++dx)
-         {
-            double temp1 = 0.0;
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               temp1 += B(qx, dx) * B(qx, dx) * temp[qx][dy];
-            }
-            y(dx, dy, 0, e) = temp1;
-            y(dx, dy, 1, e) = temp1;
-         }
-      }
-   });
-}
-
-template<const int T_D1D = 0, const int T_Q1D = 0>
-static void PAVectorMassAssembleDiagonal3D(const int NE,
-                                           const Array<double> &B_,
-                                           const Array<double> &Bt_,
-                                           const Vector &op_,
-                                           Vector &diag_,
-                                           const int d1d = 0,
-                                           const int q1d = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   constexpr int VDIM = 3;
-   MFEM_VERIFY(D1D <= MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
-   auto B = Reshape(B_.Read(), Q1D, D1D);
-   auto op = Reshape(op_.Read(), Q1D, Q1D, Q1D, NE);
-   auto y = Reshape(diag_.ReadWrite(), D1D, D1D, D1D, VDIM, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
-   {
-      const int D1D = T_D1D ? T_D1D : d1d; // nvcc workaround
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      // the following variables are evaluated at compile time
-      constexpr int max_D1D = T_D1D ? T_D1D : MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : MAX_Q1D;
-
-      double temp[max_Q1D][max_Q1D][max_D1D];
-      for (int qx = 0; qx < Q1D; ++qx)
-      {
-         for (int qy = 0; qy < Q1D; ++qy)
-         {
-            for (int dz = 0; dz < D1D; ++dz)
-            {
-               temp[qx][qy][dz] = 0.0;
-               for (int qz = 0; qz < Q1D; ++qz)
-               {
-                  temp[qx][qy][dz] += B(qz, dz) * B(qz, dz) * op(qx, qy, qz, e);
-               }
-            }
-         }
-      }
-      double temp2[max_Q1D][max_D1D][max_D1D];
-      for (int qx = 0; qx < Q1D; ++qx)
-      {
-         for (int dz = 0; dz < D1D; ++dz)
-         {
-            for (int dy = 0; dy < D1D; ++dy)
-            {
-               temp2[qx][dy][dz] = 0.0;
-               for (int qy = 0; qy < Q1D; ++qy)
-               {
-                  temp2[qx][dy][dz] += B(qy, dy) * B(qy, dy) * temp[qx][qy][dz];
-               }
-            }
-         }
-      }
-      for (int dz = 0; dz < D1D; ++dz)
-      {
-         for (int dy = 0; dy < D1D; ++dy)
-         {
-            for (int dx = 0; dx < D1D; ++dx)
-            {
-               double temp3 = 0.0;
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  temp3 += B(qx, dx) * B(qx, dx)
-                           * temp2[qx][dy][dz];
-               }
-               y(dx, dy, dz, 0, e) = temp3;
-               y(dx, dy, dz, 1, e) = temp3;
-               y(dx, dy, dz, 2, e) = temp3;
-            }
-         }
-      }
-   });
-}
-
-static void PAVectorMassAssembleDiagonal(const int dim,
-                                         const int D1D,
-                                         const int Q1D,
-                                         const int NE,
-                                         const Array<double> &B,
-                                         const Array<double> &Bt,
-                                         const Vector &op,
-                                         Vector &y)
-{
-   if (dim == 2)
-   {
-      return PAVectorMassAssembleDiagonal2D(NE, B, Bt, op, y, D1D, Q1D);
-   }
-   else if (dim == 3)
-   {
-      return PAVectorMassAssembleDiagonal3D(NE, B, Bt, op, y, D1D, Q1D);
-   }
-   MFEM_ABORT("Dimension not implemented.");
-}
-
-void VectorMassIntegrator::AssembleDiagonalPA(Vector &diag)
-{
-   if (DeviceCanUseCeed())
-   {
-      ceedOp->GetDiagonal(diag);
-   }
-   else
-   {
-      PAVectorMassAssembleDiagonal(dim, dofs1D, quad1D, ne,
-                                   maps->B, maps->Bt,
-                                   pa_data, diag);
-   }
-}
-
-template<const int T_D1D = 0, const int T_Q1D = 0>
+template<const int T_D1D = 0,
+         const int T_Q1D = 0>
 static void PAVectorMassApply2D(const int NE,
                                 const Array<double> &B_,
                                 const Array<double> &Bt_,
@@ -288,7 +132,7 @@ static void PAVectorMassApply2D(const int NE,
    auto op = Reshape(op_.Read(), Q1D, Q1D, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, VDIM, NE);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, VDIM, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int D1D = T_D1D ? T_D1D : d1d; // nvcc workaround
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -364,7 +208,8 @@ static void PAVectorMassApply2D(const int NE,
    });
 }
 
-template<const int T_D1D = 0, const int T_Q1D = 0>
+template<const int T_D1D = 0,
+         const int T_Q1D = 0>
 static void PAVectorMassApply3D(const int NE,
                                 const Array<double> &B_,
                                 const Array<double> &Bt_,
@@ -384,7 +229,7 @@ static void PAVectorMassApply3D(const int NE,
    auto op = Reshape(op_.Read(), Q1D, Q1D, Q1D, NE);
    auto x = Reshape(x_.Read(), D1D, D1D, D1D, VDIM, NE);
    auto y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, VDIM, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL(e, NE,
    {
       const int D1D = T_D1D ? T_D1D : d1d;
       const int Q1D = T_Q1D ? T_Q1D : q1d;
@@ -542,4 +387,171 @@ void VectorMassIntegrator::AddMultPA(const Vector &x, Vector &y) const
    }
 }
 
+template<const int T_D1D = 0, const int T_Q1D = 0>
+static void PAVectorMassAssembleDiagonal2D(const int NE,
+                                           const Array<double> &B_,
+                                           const Array<double> &Bt_,
+                                           const Vector &op_,
+                                           Vector &diag_,
+                                           const int d1d = 0,
+                                           const int q1d = 0)
+{
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+   constexpr int VDIM = 2;
+   MFEM_VERIFY(D1D <= MAX_D1D, "");
+   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
+   auto B = Reshape(B_.Read(), Q1D, D1D);
+   auto op = Reshape(op_.Read(), Q1D, Q1D, NE);
+   auto y = Reshape(diag_.ReadWrite(), D1D, D1D, VDIM, NE);
+   MFEM_FORALL(e, NE,
+   {
+      const int D1D = T_D1D ? T_D1D : d1d;
+      const int Q1D = T_Q1D ? T_Q1D : q1d;
+      constexpr int max_D1D = T_D1D ? T_D1D : MAX_D1D;
+      constexpr int max_Q1D = T_Q1D ? T_Q1D : MAX_Q1D;
+
+      double temp[max_Q1D][max_D1D];
+      for (int qx = 0; qx < Q1D; ++qx)
+      {
+         for (int dy = 0; dy < D1D; ++dy)
+         {
+            temp[qx][dy] = 0.0;
+            for (int qy = 0; qy < Q1D; ++qy)
+            {
+               temp[qx][dy] += B(qy, dy) * B(qy, dy) * op(qx, qy, e);
+            }
+         }
+      }
+      for (int dy = 0; dy < D1D; ++dy)
+      {
+         for (int dx = 0; dx < D1D; ++dx)
+         {
+            double temp1 = 0.0;
+            for (int qx = 0; qx < Q1D; ++qx)
+            {
+               temp1 += B(qx, dx) * B(qx, dx) * temp[qx][dy];
+            }
+            y(dx, dy, 0, e) = temp1;
+            y(dx, dy, 1, e) = temp1;
+         }
+      }
+   });
+}
+
+template<const int T_D1D = 0, const int T_Q1D = 0>
+static void PAVectorMassAssembleDiagonal3D(const int NE,
+                                           const Array<double> &B_,
+                                           const Array<double> &Bt_,
+                                           const Vector &op_,
+                                           Vector &diag_,
+                                           const int d1d = 0,
+                                           const int q1d = 0)
+{
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+   constexpr int VDIM = 3;
+   MFEM_VERIFY(D1D <= MAX_D1D, "");
+   MFEM_VERIFY(Q1D <= MAX_Q1D, "");
+   auto B = Reshape(B_.Read(), Q1D, D1D);
+   auto op = Reshape(op_.Read(), Q1D, Q1D, Q1D, NE);
+   auto y = Reshape(diag_.ReadWrite(), D1D, D1D, D1D, VDIM, NE);
+   MFEM_FORALL(e, NE,
+   {
+      const int D1D = T_D1D ? T_D1D : d1d; // nvcc workaround
+      const int Q1D = T_Q1D ? T_Q1D : q1d;
+      // the following variables are evaluated at compile time
+      constexpr int max_D1D = T_D1D ? T_D1D : MAX_D1D;
+      constexpr int max_Q1D = T_Q1D ? T_Q1D : MAX_Q1D;
+
+      double temp[max_Q1D][max_Q1D][max_D1D];
+      for (int qx = 0; qx < Q1D; ++qx)
+      {
+         for (int qy = 0; qy < Q1D; ++qy)
+         {
+            for (int dz = 0; dz < D1D; ++dz)
+            {
+               temp[qx][qy][dz] = 0.0;
+               for (int qz = 0; qz < Q1D; ++qz)
+               {
+                  temp[qx][qy][dz] += B(qz, dz) * B(qz, dz) * op(qx, qy, qz, e);
+               }
+            }
+         }
+      }
+      double temp2[max_Q1D][max_D1D][max_D1D];
+      for (int qx = 0; qx < Q1D; ++qx)
+      {
+         for (int dz = 0; dz < D1D; ++dz)
+         {
+            for (int dy = 0; dy < D1D; ++dy)
+            {
+               temp2[qx][dy][dz] = 0.0;
+               for (int qy = 0; qy < Q1D; ++qy)
+               {
+                  temp2[qx][dy][dz] += B(qy, dy) * B(qy, dy) * temp[qx][qy][dz];
+               }
+            }
+         }
+      }
+      for (int dz = 0; dz < D1D; ++dz)
+      {
+         for (int dy = 0; dy < D1D; ++dy)
+         {
+            for (int dx = 0; dx < D1D; ++dx)
+            {
+               double temp3 = 0.0;
+               for (int qx = 0; qx < Q1D; ++qx)
+               {
+                  temp3 += B(qx, dx) * B(qx, dx)
+                           * temp2[qx][dy][dz];
+               }
+               y(dx, dy, dz, 0, e) = temp3;
+               y(dx, dy, dz, 1, e) = temp3;
+               y(dx, dy, dz, 2, e) = temp3;
+            }
+         }
+      }
+   });
+}
+
+static void PAVectorMassAssembleDiagonal(const int dim,
+                                         const int D1D,
+                                         const int Q1D,
+                                         const int NE,
+                                         const Array<double> &B,
+                                         const Array<double> &Bt,
+                                         const Vector &op,
+                                         Vector &y)
+{
+   if (dim == 2)
+   {
+      return PAVectorMassAssembleDiagonal2D(NE, B, Bt, op, y, D1D, Q1D);
+   }
+   else if (dim == 3)
+   {
+      return PAVectorMassAssembleDiagonal3D(NE, B, Bt, op, y, D1D, Q1D);
+   }
+   MFEM_ABORT("Dimension not implemented.");
+}
+
+void VectorMassIntegrator::AssembleDiagonalPA(Vector &diag)
+{
+   if (DeviceCanUseCeed())
+   {
+      ceedOp->GetDiagonal(diag);
+   }
+   else
+   {
+      PAVectorMassAssembleDiagonal(dim,
+                                   dofs1D,
+                                   quad1D,
+                                   ne,
+                                   maps->B,
+                                   maps->Bt,
+                                   pa_data,
+                                   diag);
+   }
+}
+
 } // namespace mfem

fem/bilininteg_vecmass_mf.cpp

@@ -9,13 +9,19 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
-#include "../ceed/integrators/mass/mass.hpp"
+#include "../general/forall.hpp"
+#include "bilininteg.hpp"
+#include "gridfunc.hpp"
+#include "ceed/integrators/mass/mass.hpp"
+
+using namespace std;
 
 namespace mfem
 {
 
+// MF Mass Integrator
+
+// MF Mass Assemble kernel
 void VectorMassIntegrator::AssembleMF(const FiniteElementSpace &fes)
 {
    // Assuming the same element type

fem/ceed/interface/coefficient.hpp

@@ -288,7 +288,7 @@ void InitCoefficientWithIndices(mfem::Coefficient *Q, mfem::Mesh &mesh,
       auto in = Reshape(qFun.Read(), nq, ne);
       auto d_indices = Read(m_indices, nelem);
       auto out = Reshape(ceedCoeff->coeff.Write(), nq, nelem);
-      mfem::forall(nelem * nq, [=] MFEM_HOST_DEVICE (int i)
+      MFEM_FORALL(i, nelem * nq,
       {
          const int q = i%nq;
          const int sub_e = i/nq;
@@ -378,7 +378,7 @@ void InitCoefficientWithIndices(mfem::VectorCoefficient *VQ, mfem::Mesh &mesh,
       auto in = Reshape(qFun.Read(), dim, nq, ne);
       auto d_indices = Read(m_indices, nelem);
       auto out = Reshape(ceedCoeff->coeff.Write(), dim, nq, nelem);
-      mfem::forall(nelem * nq, [=] MFEM_HOST_DEVICE (int i)
+      MFEM_FORALL(i, nelem * nq,
       {
          const int q = i%nq;
          const int sub_e = i/nq;

fem/ceed/solvers/algebraic.cpp

@@ -519,7 +519,7 @@ int CeedVectorPointwiseMult(CeedVector a, const CeedVector b)
    ierr = CeedVectorGetArray(a, mem, &a_data); CeedChk(ierr);
    ierr = CeedVectorGetArrayRead(b, mem, &b_data); CeedChk(ierr);
    MFEM_VERIFY(int(length) == length, "length overflow");
-   mfem::forall(length, [=] MFEM_HOST_DEVICE (int i)
+   MFEM_FORALL(i, length,
    {a_data[i] *= b_data[i];});
 
    ierr = CeedVectorRestoreArray(a, &a_data); CeedChk(ierr);
@@ -593,7 +593,7 @@ void AlgebraicInterpolation::MultTranspose(const mfem::Vector& x,
                                  &multiplicitydata); PCeedChk(ierr);
    ierr = CeedVectorGetArrayWrite(fine_work, mem, &workdata); PCeedChk(ierr);
    MFEM_VERIFY((int)length == length, "length overflow");
-   mfem::forall(length, [=] MFEM_HOST_DEVICE (int i)
+   MFEM_FORALL(i, length,
    {workdata[i] = in_ptr[i] * multiplicitydata[i];});
    ierr = CeedVectorRestoreArrayRead(fine_multiplicity_r,
                                      &multiplicitydata);

fem/coefficient.cpp

@@ -144,54 +144,11 @@ double FunctionCoefficient::Eval(ElementTransformation & T,
    }
 }
 
-double CartesianCoefficient::Eval(ElementTransformation & T,
-                                  const IntegrationPoint & ip)
-{
-   T.Transform(ip, transip);
-   return transip[comp];
-}
-
-double CylindricalRadialCoefficient::Eval(ElementTransformation & T,
-                                          const IntegrationPoint & ip)
-{
-   T.Transform(ip, transip);
-   return sqrt(transip[0] * transip[0] + transip[1] * transip[1]);
-}
-
-double CylindricalAzimuthalCoefficient::Eval(ElementTransformation & T,
-                                             const IntegrationPoint & ip)
-{
-   T.Transform(ip, transip);
-   return atan2(transip[1], transip[0]);
-}
-
-double SphericalRadialCoefficient::Eval(ElementTransformation & T,
-                                        const IntegrationPoint & ip)
-{
-   T.Transform(ip, transip);
-   return sqrt(transip * transip);
-}
-
-double SphericalAzimuthalCoefficient::Eval(ElementTransformation & T,
-                                           const IntegrationPoint & ip)
-{
-   T.Transform(ip, transip);
-   return atan2(transip[1], transip[0]);
-}
-
-double SphericalPolarCoefficient::Eval(ElementTransformation & T,
-                                       const IntegrationPoint & ip)
-{
-   T.Transform(ip, transip);
-   return atan2(sqrt(transip[0] * transip[0] + transip[1] * transip[1]),
-                transip[2]);
-}
-
 double GridFunctionCoefficient::Eval (ElementTransformation &T,
                                       const IntegrationPoint &ip)
 {
    Mesh *gf_mesh = GridF->FESpace()->GetMesh();
-   if (T.mesh->GetNE() == gf_mesh->GetNE())
+   if (T.mesh == gf_mesh)
    {
       return GridF->GetValue(T, ip, Component);
    }
@@ -356,13 +313,6 @@ void PWVectorCoefficient::Eval(Vector &V, ElementTransformation &T,
    V = 0.0;
 }
 
-void PositionVectorCoefficient::Eval(Vector &V, ElementTransformation &T,
-                                     const IntegrationPoint &ip)
-{
-   V.SetSize(vdim);
-   T.Transform(ip, V);
-}
-
 void VectorFunctionCoefficient::Eval(Vector &V, ElementTransformation &T,
                                      const IntegrationPoint &ip)
 {
@@ -446,7 +396,7 @@ void VectorGridFunctionCoefficient::Eval(Vector &V, ElementTransformation &T,
                                          const IntegrationPoint &ip)
 {
    Mesh *gf_mesh = GridFunc->FESpace()->GetMesh();
-   if (T.mesh->GetNE() == gf_mesh->GetNE())
+   if (T.mesh == gf_mesh)
    {
       GridFunc->GetVectorValue(T, ip, V);
    }
@@ -494,7 +444,7 @@ void GradientGridFunctionCoefficient::Eval(Vector &V, ElementTransformation &T,
                                            const IntegrationPoint &ip)
 {
    Mesh *gf_mesh = GridFunc->FESpace()->GetMesh();
-   if (T.mesh->GetNE() == gf_mesh->GetNE())
+   if (T.mesh == gf_mesh)
    {
       GridFunc->GetGradient(T, V);
    }
@@ -535,7 +485,7 @@ void CurlGridFunctionCoefficient::Eval(Vector &V, ElementTransformation &T,
                                        const IntegrationPoint &ip)
 {
    Mesh *gf_mesh = GridFunc->FESpace()->GetMesh();
-   if (T.mesh->GetNE() == gf_mesh->GetNE())
+   if (T.mesh == gf_mesh)
    {
       GridFunc->GetCurl(T, V);
    }
@@ -557,7 +507,7 @@ double DivergenceGridFunctionCoefficient::Eval(ElementTransformation &T,
                                                const IntegrationPoint &ip)
 {
    Mesh *gf_mesh = GridFunc->FESpace()->GetMesh();
-   if (T.mesh->GetNE() == gf_mesh->GetNE())
+   if (T.mesh == gf_mesh)
    {
       return GridFunc->GetDivergence(T);
    }

fem/coefficient.hpp

@@ -258,124 +258,6 @@ public:
                        const IntegrationPoint &ip);
 };
 
-/// A common base class for returning individual components of the domain's
-/// Cartesian coordinates.
-class CartesianCoefficient : public Coefficient
-{
-protected:
-   int comp;
-   mutable Vector transip;
-
-   /// @a comp_ index of the desired component (0 -> x, 1 -> y, 2 -> z)
-   CartesianCoefficient(int comp_) : comp(comp_), transip(3) {}
-
-public:
-   /// Evaluate the coefficient at @a ip.
-   virtual double Eval(ElementTransformation &T,
-                       const IntegrationPoint &ip);
-};
-
-/// Scalar coefficient which returns the x-component of the evaluation point
-class CartesianXCoefficient : public CartesianCoefficient
-{
-public:
-   CartesianXCoefficient() : CartesianCoefficient(0) {}
-};
-
-/// Scalar coefficient which returns the y-component of the evaluation point
-class CartesianYCoefficient : public CartesianCoefficient
-{
-public:
-   CartesianYCoefficient() : CartesianCoefficient(1) {}
-};
-
-/// Scalar coefficient which returns the z-component of the evaluation point
-class CartesianZCoefficient : public CartesianCoefficient
-{
-public:
-   CartesianZCoefficient() : CartesianCoefficient(2) {}
-};
-
-/// Scalar coefficient which returns the radial distance from the axis of
-/// the evaluation point in the cylindrical coordinate system
-class CylindricalRadialCoefficient : public Coefficient
-{
-private:
-   mutable Vector transip;
-
-public:
-   CylindricalRadialCoefficient() : transip(3) {}
-
-   /// Evaluate the coefficient at @a ip.
-   virtual double Eval(ElementTransformation &T,
-                       const IntegrationPoint &ip);
-};
-
-/// Scalar coefficient which returns the angular position or azimuth (often
-/// denoted by theta) of the evaluation point in the cylindrical coordinate
-/// system
-class CylindricalAzimuthalCoefficient : public Coefficient
-{
-private:
-   mutable Vector transip;
-
-public:
-   CylindricalAzimuthalCoefficient() : transip(3) {}
-
-   /// Evaluate the coefficient at @a ip.
-   virtual double Eval(ElementTransformation &T,
-                       const IntegrationPoint &ip);
-};
-
-/// Scalar coefficient which returns the height or altitude of
-/// the evaluation point in the cylindrical coordinate system
-typedef CartesianZCoefficient CylindricalZCoefficient;
-
-/// Scalar coefficient which returns the radial distance from the origin of
-/// the evaluation point in the spherical coordinate system
-class SphericalRadialCoefficient : public Coefficient
-{
-private:
-   mutable Vector transip;
-
-public:
-   SphericalRadialCoefficient() : transip(3) {}
-
-   /// Evaluate the coefficient at @a ip.
-   virtual double Eval(ElementTransformation &T,
-                       const IntegrationPoint &ip);
-};
-
-/// Scalar coefficient which returns the azimuthal angle (often denoted by phi)
-/// of the evaluation point in the spherical coordinate system
-class SphericalAzimuthalCoefficient : public Coefficient
-{
-private:
-   mutable Vector transip;
-
-public:
-   SphericalAzimuthalCoefficient() : transip(3) {}
-
-   /// Evaluate the coefficient at @a ip.
-   virtual double Eval(ElementTransformation &T,
-                       const IntegrationPoint &ip);
-};
-
-/// Scalar coefficient which returns the polar angle (often denoted by theta)
-/// of the evaluation point in the spherical coordinate system
-class SphericalPolarCoefficient : public Coefficient
-{
-private:
-   mutable Vector transip;
-
-public:
-   SphericalPolarCoefficient() : transip(3) {}
-
-   /// Evaluate the coefficient at @a ip.
-   virtual double Eval(ElementTransformation &T,
-                       const IntegrationPoint &ip);
-};
-
 class GridFunction;
 
 /// Coefficient defined by a GridFunction. This coefficient is mesh dependent.
@@ -718,22 +600,6 @@ public:
    using VectorCoefficient::Eval;
 };
 
-/// A vector coefficient which returns the physical location of the
-/// evaluation point in the Cartesian coordinate system.
-class PositionVectorCoefficient : public VectorCoefficient
-{
-public:
-
-   PositionVectorCoefficient(int dim) : VectorCoefficient(dim) {}
-
-   using VectorCoefficient::Eval;
-   /// Evaluate the vector coefficient at @a ip.
-   virtual void Eval(Vector &V, ElementTransformation &T,
-                     const IntegrationPoint &ip);
-
-   virtual ~PositionVectorCoefficient() { }
-};
-
 /// A general vector function coefficient
 class VectorFunctionCoefficient : public VectorCoefficient
 {

fem/complex_fem.cpp

@@ -497,7 +497,7 @@ SesquilinearForm::FormLinearSystem(const Array<int> &ess_tdof_list,
          auto d_X_r = X_r.Read();
          auto d_X_i = X_i.Read();
          auto d_idx = ess_tdof_list.Read();
-         mfem::forall(n, [=] MFEM_HOST_DEVICE (int i)
+         MFEM_FORALL(i, n,
          {
             const int j = d_idx[i];
             d_B_r[j] = d_X_r[j];
@@ -1230,7 +1230,7 @@ ParSesquilinearForm::FormLinearSystem(const Array<int> &ess_tdof_list,
       auto d_X_r = X_r.Read();
       auto d_X_i = X_i.Read();
       auto d_idx = ess_tdof_list.Read();
-      mfem::forall(n, [=] MFEM_HOST_DEVICE (int i)
+      MFEM_FORALL(i, n,
       {
          const int j = d_idx[i];
          d_B_r[j] = d_X_r[j];

fem/dgmassinv.cpp

@@ -107,7 +107,7 @@ void DGMassInverse::Update()
 {
    M->Assemble();
    M->AssembleDiagonal(diag_inv);
-   diag_inv.Reciprocal();
+   internal::MakeReciprocal(diag_inv.Size(), diag_inv.ReadWrite());
 }
 
 DGMassInverse::~DGMassInverse()
@@ -168,7 +168,7 @@ void DGMassInverse::DGMassCGIteration(const Vector &b_, Vector &u_) const
 
    constexpr int NB = Q1D ? Q1D : 1; // block size
 
-   mfem::forall_2D(NE, NB, NB, [=] MFEM_HOST_DEVICE (int e)
+   MFEM_FORALL_2D(e, NE, NB, NB, 1,
    {
       constexpr int NB = Q1D ? Q1D : 1; // redefine here for some compilers
 

fem/dgmassinv.hpp

@@ -87,8 +87,6 @@ public:
    ///
    /// If @ref iterative_mode is @a true, @a u is used as an initial guess.
    void Mult(const Vector &b, Vector &u) const;
-   /// Same as Mult() since the mass matrix is symmetric.
-   void MultTranspose(const Vector &b, Vector &u) const { Mult(b, u); }
    /// Not implemented. Aborts.
    void SetOperator(const Operator &op);
    /// Set the relative tolerance.
@@ -103,8 +101,8 @@ public:
    ~DGMassInverse();
 
    /// @brief Solve the system M b = u. <b>Not part of the public interface.</b>
-   /// @note This member function must be public because it defines an
-   /// extended lambda used in an mfem::forall kernel (nvcc limitation)
+   /// @note This member function must be public because it contains an
+   /// MFEM_FORALL kernel (nvcc limitation)
    template<int DIM, int D1D = 0, int Q1D = 0>
    void DGMassCGIteration(const Vector &b_, Vector &u_) const;
 };

fem/dgmassinv_kernels.hpp

@@ -12,9 +12,9 @@
 #ifndef MFEM_DGMASSINV_KERNELS_HPP
 #define MFEM_DGMASSINV_KERNELS_HPP
 
+#include "bilininteg_mass_pa.hpp"
 #include "../linalg/kernels.hpp"
 #include "kernels.hpp"
-#include "integ/bilininteg_mass_kernels.hpp"
 
 namespace mfem
 {
@@ -22,6 +22,11 @@ namespace mfem
 namespace internal
 {
 
+void MakeReciprocal(int n, double *x)
+{
+   MFEM_FORALL(i, n, x[i] = 1.0/x[i]; );
+}
+
 template <int DIM, int D1D, int Q1D>
 MFEM_HOST_DEVICE inline
 void DGMassApply(const int e,

fem/fe/fe_base.cpp

@@ -401,7 +401,7 @@ const DofToQuad &FiniteElement::GetDofToQuad(const IntegrationRule &ir,
          }
       }
    }
-   else if (range_type == VECTOR)
+   else
    {
       d2q->B.SetSize(nqpt*dim*dof);
       d2q->Bt.SetSize(dof*nqpt*dim);
@@ -419,10 +419,6 @@ const DofToQuad &FiniteElement::GetDofToQuad(const IntegrationRule &ir,
          }
       }
    }
-   else
-   {
-      // Skip B and Bt for unknown range type
-   }
    switch (deriv_type)
    {
       case GRAD:
@@ -476,7 +472,7 @@ const DofToQuad &FiniteElement::GetDofToQuad(const IntegrationRule &ir,
             {
                for (int j = 0; j < dof; j++)
                {
-                  d2q->G[i+nqpt*(d+cdim*j)] = d2q->Gt[j+dof*(i+nqpt*d)] = curlshape(j, d);
+                  d2q->G[i+nqpt*(d+dim*j)] = d2q->Gt[j+dof*(i+nqpt*d)] = curlshape(j, d);
                }
             }
          }
@@ -484,8 +480,7 @@ const DofToQuad &FiniteElement::GetDofToQuad(const IntegrationRule &ir,
       }
       case NONE:
       default:
-         // Skip G and Gt for unknown derivative type
-         break;
+         MFEM_ABORT("invalid finite element derivative type");
    }
    dof2quad_array.Append(d2q);
    return *d2q;

fem/fe/fe_base.hpp

@@ -1288,9 +1288,9 @@ public:
    const DofToQuad &GetDofToQuad(const IntegrationRule &ir,
                                  DofToQuad::Mode mode) const override
    {
-      return (mode == DofToQuad::FULL) ?
-             FiniteElement::GetDofToQuad(ir, mode) :
-             GetTensorDofToQuad(*this, ir, mode, basis1d, true, dof2quad_array);
+      MFEM_VERIFY(mode != DofToQuad::FULL, "invalid mode requested");
+      return GetTensorDofToQuad(*this, ir, mode, basis1d, true,
+                                dof2quad_array);
    }
 
    const DofToQuad &GetDofToQuadOpen(const IntegrationRule &ir,

fem/fespace.cpp

@@ -385,38 +385,21 @@ void FiniteElementSpace::BuildBdrElementToDofTable() const
    if (bdr_elem_dof) { return; }
 
    Table *bel_dof = new Table;
-   Table *bel_fos = (mesh->Dimension() == 3) ? (new Table) : NULL;
    Array<int> dofs;
-   int F, Fo;
    bel_dof->MakeI(mesh->GetNBE());
-   if (bel_fos) { bel_fos->MakeI(mesh->GetNBE()); }
    for (int i = 0; i < mesh->GetNBE(); i++)
    {
       GetBdrElementDofs(i, dofs);
       bel_dof->AddColumnsInRow(i, dofs.Size());
-
-      if (bel_fos)
-      {
-         bel_fos->AddAColumnInRow(i);
-      }
    }
    bel_dof->MakeJ();
-   if (bel_fos) { bel_fos->MakeJ(); }
    for (int i = 0; i < mesh->GetNBE(); i++)
    {
       GetBdrElementDofs(i, dofs);
       bel_dof->AddConnections(i, (int *)dofs, dofs.Size());
-
-      if (bel_fos)
-      {
-         mesh->GetBdrElementFace(i, &F, &Fo);
-         bel_fos->AddConnection(i, Fo);
-      }
    }
    bel_dof->ShiftUpI();
-   if (bel_fos) { bel_fos->ShiftUpI(); }
    bdr_elem_dof = bel_dof;
-   bdr_elem_fos = bel_fos;
 }
 
 void FiniteElementSpace::BuildFaceToDofTable() const
@@ -1559,10 +1542,6 @@ FiniteElementSpace::RefinementOperator::~RefinementOperator()
 {
    delete old_elem_dof;
    delete old_elem_fos;
-   for (int i=0; i<old_DoFTrans.Size(); i++)
-   {
-      delete old_DoFTrans[i];
-   }
 }
 
 void FiniteElementSpace::RefinementOperator

fem/fespace.hpp

@@ -994,9 +994,7 @@ public:
 
    /// @brief Returns indices of degrees of freedom for the @a i'th element.
    /// The returned indices are offsets into an @ref ldof vector with @b vdim
-   /// not necessarily equal to 1. The returned indexes are always ordered
-   /// byNODES, irrespective of whether the space is byNODES or byVDIM.
-   /// See also GetElementDofs().
+   /// not necessarily equal to 1. See also GetElementDofs().
    ///
    /// @note In many cases the returned DofTransformation object will be NULL.
    /// In other cases see the documentation of the DofTransformation class for

fem/gridfunc.cpp

@@ -397,6 +397,8 @@ void GridFunction::GetNodalValues(int i, Array<double> &nval, int vdim) const
 {
    Array<int> vdofs;
 
+   int k;
+
    DofTransformation * doftrans = fes->GetElementVDofs(i, vdofs);
    const FiniteElement *FElem = fes->GetFE(i);
    const IntegrationRule *ElemVert =
@@ -417,7 +419,7 @@ void GridFunction::GetNodalValues(int i, Array<double> &nval, int vdim) const
       Vector shape(dof);
       if (FElem->GetMapType() == FiniteElement::VALUE)
       {
-         for (int k = 0; k < n; k++)
+         for (k = 0; k < n; k++)
          {
             FElem->CalcShape(ElemVert->IntPoint(k), shape);
             nval[k] = shape * (&loc_data[dof * vdim]);
@@ -426,7 +428,7 @@ void GridFunction::GetNodalValues(int i, Array<double> &nval, int vdim) const
       else
       {
          ElementTransformation *Tr = fes->GetElementTransformation(i);
-         for (int k = 0; k < n; k++)
+         for (k = 0; k < n; k++)
          {
             Tr->SetIntPoint(&ElemVert->IntPoint(k));
             FElem->CalcPhysShape(*Tr, shape);
@@ -438,7 +440,7 @@ void GridFunction::GetNodalValues(int i, Array<double> &nval, int vdim) const
    {
       ElementTransformation *Tr = fes->GetElementTransformation(i);
       DenseMatrix vshape(dof, FElem->GetDim());
-      for (int k = 0; k < n; k++)
+      for (k = 0; k < n; k++)
       {
          Tr->SetIntPoint(&ElemVert->IntPoint(k));
          FElem->CalcVShape(*Tr, vshape);
@@ -2399,11 +2401,7 @@ void GridFunction::ProjectDeltaCoefficient(DeltaCoefficient &delta_coeff,
                                      loc_mass);
             vals.SetSize(fe->GetDof());
             fe->ProjectDelta(j, vals);
-            const DofTransformation* const doftrans = fes->GetElementVDofs(i, vdofs);
-            if (doftrans)
-            {
-               doftrans->TransformPrimal(vals);
-            }
+            fes->GetElementVDofs(i, vdofs);
             SetSubVector(vdofs, vals);
             loc_mass_vals.SetSize(vals.Size());
             loc_mass.Mult(vals, loc_mass_vals);
@@ -2757,11 +2755,7 @@ void GridFunction::ProjectBdrCoefficientNormal(
          CalcOrtho(T->Jacobian(), nor);
          lvec(j) = (vc * nor);
       }
-      const DofTransformation* const doftrans = fes->GetBdrElementDofs(i, dofs);
-      if (doftrans)
-      {
-         doftrans->TransformPrimal(lvec);
-      }
+      fes->GetBdrElementDofs(i, dofs);
       SetSubVector(dofs, lvec);
    }
 #endif
@@ -4037,19 +4031,11 @@ double ZZErrorEstimator(BilinearFormIntegrator &blfi,
       {
          if (with_subdomains && ufes->GetAttribute(i) != s) { continue; }
 
-         const DofTransformation* const utrans = ufes->GetElementVDofs(i, udofs);
-         const DofTransformation* const ftrans = ffes->GetElementVDofs(i, fdofs);
+         ufes->GetElementVDofs(i, udofs);
+         ffes->GetElementVDofs(i, fdofs);
 
          u.GetSubVector(udofs, ul);
          flux.GetSubVector(fdofs, fla);
-         if (utrans)
-         {
-            utrans->InvTransformPrimal(ul);
-         }
-         if (ftrans)
-         {
-            ftrans->InvTransformPrimal(fla);
-         }
 
          Transf = ufes->GetElementTransformation(i);
          blfi.ComputeElementFlux(*ufes->GetFE(i), *Transf, ul,
@@ -4344,12 +4330,8 @@ double LSZZErrorEstimator(BilinearFormIntegrator &blfi,  // input
                                                     flux_order));
          int num_integration_pts = ir->GetNPoints();
 
-         const DofTransformation* const utrans = ufes->GetElementVDofs(ielem, udofs);
+         ufes->GetElementVDofs(ielem, udofs);
          u.GetSubVector(udofs, ul);
-         if (utrans)
-         {
-            utrans->InvTransformPrimal(ul);
-         }
          Transf = ufes->GetElementTransformation(ielem);
          FiniteElement *dummy = nullptr;
          blfi.ComputeElementFlux(*ufes->GetFE(ielem), *Transf, ul,

fem/gridfunc.hpp

@@ -684,10 +684,6 @@ public:
    /// Transform by the Space UpdateMatrix (e.g., on Mesh change).
    virtual void Update();
 
-   /** Return update counter, similar to Mesh::GetSequence(). Used to
-       check if it is up to date with the space. */
-   long GetSequence() const { return fes_sequence; }
-
    FiniteElementSpace *FESpace() { return fes; }
    const FiniteElementSpace *FESpace() const { return fes; }
 

fem/gslib.cpp

@@ -108,7 +108,8 @@ void FindPointsGSLIB::Setup(Mesh &m, const double bb_t, const double newt_tol,
                             const int npt_max)
 {
    MFEM_VERIFY(m.GetNodes() != NULL, "Mesh nodes are required.");
-   const int meshOrder = m.GetNodes()->FESpace()->GetMaxElementOrder();
+   MFEM_VERIFY(!(m.GetNodes()->FESpace()->IsVariableOrder()),
+               "Variable order mesh is not currently supported.");
 
    // call FreeData if FindPointsGSLIB::Setup has been called already
    if (setupflag) { FreeData(); }
@@ -116,36 +117,30 @@ void FindPointsGSLIB::Setup(Mesh &m, const double bb_t, const double newt_tol,
    crystal_init(cr, gsl_comm);
    mesh = &m;
    dim  = mesh->Dimension();
-   unsigned dof1D = meshOrder + 1;
+   const FiniteElement *fe = mesh->GetNodalFESpace()->GetFE(0);
+   unsigned dof1D = fe->GetOrder() + 1;
 
    SetupSplitMeshes();
    if (dim == 2)
    {
       if (ir_split[0]) { delete ir_split[0]; ir_split[0] = NULL; }
       ir_split[0] = new IntegrationRule(3*pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[0], ir_split[0], meshOrder);
-
-      if (ir_split[1]) { delete ir_split[1]; ir_split[1] = NULL; }
-      ir_split[1] = new IntegrationRule(pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[1], ir_split[1], meshOrder);
+      SetupIntegrationRuleForSplitMesh(mesh_split[0], ir_split[0], fe->GetOrder());
    }
    else if (dim == 3)
    {
-      if (ir_split[0]) { delete ir_split[0]; ir_split[0] = NULL; }
-      ir_split[0] = new IntegrationRule(pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[0], ir_split[0], meshOrder);
 
       if (ir_split[1]) { delete ir_split[1]; ir_split[1] = NULL; }
       ir_split[1] = new IntegrationRule(4*pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[1], ir_split[1], meshOrder);
+      SetupIntegrationRuleForSplitMesh(mesh_split[1], ir_split[1], fe->GetOrder());
 
       if (ir_split[2]) { delete ir_split[2]; ir_split[2] = NULL; }
       ir_split[2] = new IntegrationRule(3*pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[2], ir_split[2], meshOrder);
+      SetupIntegrationRuleForSplitMesh(mesh_split[2], ir_split[2], fe->GetOrder());
 
       if (ir_split[3]) { delete ir_split[3]; ir_split[3] = NULL; }
       ir_split[3] = new IntegrationRule(8*pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[3], ir_split[3], meshOrder);
+      SetupIntegrationRuleForSplitMesh(mesh_split[3], ir_split[3], fe->GetOrder());
    }
 
    GetNodalValues(mesh->GetNodes(), gsl_mesh);
@@ -184,7 +179,7 @@ void FindPointsGSLIB::FindPoints(const Vector &point_pos,
    gsl_ref.SetSize(points_cnt * dim);
    gsl_dist.SetSize(points_cnt);
 
-   auto xvFill = [&](const double *xv_base[], unsigned xv_stride[])
+   auto xvFill = [&](const double *xv_base[], unsigned xv_stride[], int dim)
    {
       for (int d = 0; d < dim; d++)
       {
@@ -204,7 +199,7 @@ void FindPointsGSLIB::FindPoints(const Vector &point_pos,
    {
       const double *xv_base[2];
       unsigned xv_stride[2];
-      xvFill(xv_base, xv_stride);
+      xvFill(xv_base, xv_stride, dim);
       findpts_2(gsl_code.GetData(), sizeof(unsigned int),
                 gsl_proc.GetData(), sizeof(unsigned int),
                 gsl_elem.GetData(), sizeof(unsigned int),
@@ -216,7 +211,7 @@ void FindPointsGSLIB::FindPoints(const Vector &point_pos,
    {
       const double *xv_base[3];
       unsigned xv_stride[3];
-      xvFill(xv_base, xv_stride);
+      xvFill(xv_base, xv_stride, dim);
       findpts_3(gsl_code.GetData(), sizeof(unsigned int),
                 gsl_proc.GetData(), sizeof(unsigned int),
                 gsl_elem.GetData(), sizeof(unsigned int),
@@ -338,14 +333,9 @@ void FindPointsGSLIB::SetupSplitMeshes()
             (*gf_rst_map[0])(j+k*npt) = quad_v[j][k];
          }
       }
-
-      mesh_split[1] = new Mesh(Mesh::MakeCartesian2D(1, 1,
-                                                     Element::QUADRILATERAL));
    }
    else if (mesh->Dimension() == 3)
    {
-      mesh_split[0] = new Mesh(Mesh::MakeCartesian3D(1, 1, 1,
-                                                     Element::HEXAHEDRON));
       // Tetrahedron
       {
          int Nvert = 15;
@@ -575,12 +565,11 @@ void FindPointsGSLIB::GetNodalValues(const GridFunction *gf_in,
    const GridFunction *nodes     = gf_in;
    const FiniteElementSpace *fes = nodes->FESpace();
    const int NE                  = mesh->GetNE();
-   const int vdim                = fes->GetVDim();
+   const int vdim                = gf_in->FESpace()->GetVDim();
 
    IntegrationRule *ir_split_temp = NULL;
 
-   const int maxOrder = fes->GetMaxElementOrder();
-   const int dof_1D =  maxOrder+1;
+   const int dof_1D =  nodes->FESpace()->GetFE(0)->GetOrder()+1;
    const int pts_el = std::pow(dof_1D, dim);
    const int pts_cnt = NE_split_total * pts_el;
    node_vals.SetSize(vdim * pts_cnt);
@@ -590,7 +579,7 @@ void FindPointsGSLIB::GetNodalValues(const GridFunction *gf_in,
 
    for (int e = 0; e < NE; e++)
    {
-      const FiniteElement *fe   = fes->GetFE(e);
+      const FiniteElement *fe   = nodes->FESpace()->GetFE(e);
       const Geometry::Type gt   = fe->GetGeomType();
       bool el_to_split = true;
       if (gt == Geometry::TRIANGLE)
@@ -609,22 +598,16 @@ void FindPointsGSLIB::GetNodalValues(const GridFunction *gf_in,
       {
          ir_split_temp = ir_split[3];
       }
-      else if (gt == Geometry::SQUARE)
+      else if (gt == Geometry::SQUARE || gt == Geometry::CUBE)
       {
-         ir_split_temp = ir_split[1];
-         el_to_split = gf_in->FESpace()->IsVariableOrder();
-      }
-      else if (gt == Geometry::CUBE)
-      {
-         ir_split_temp = ir_split[0];
-         el_to_split = gf_in->FESpace()->IsVariableOrder();
+         el_to_split = false;
       }
       else
       {
          MFEM_ABORT("Unsupported geometry type.");
       }
 
-      if (el_to_split) // Triangle/Tet/Prism or Quads/Hex but variable order
+      if (el_to_split) // Triangle/Tet/Prism
       {
          // Fill gsl_mesh with location of split points.
          Vector locval(vdim);
@@ -639,7 +622,7 @@ void FindPointsGSLIB::GetNodalValues(const GridFunction *gf_in,
             gsl_mesh_pt_index++;
          }
       }
-      else // Quad/Hex and constant polynomial order
+      else // Quad/Hex
       {
          const int dof_cnt_split = fe->GetDof();
 
@@ -820,8 +803,8 @@ void FindPointsGSLIB::MapRefPosAndElemIndices()
 void FindPointsGSLIB::Interpolate(const GridFunction &field_in,
                                   Vector &field_out)
 {
-   const int  gf_order   = field_in.FESpace()->GetMaxElementOrder(),
-              mesh_order = mesh->GetNodalFESpace()->GetMaxElementOrder();
+   const int  gf_order   = field_in.FESpace()->GetFE(0)->GetOrder(),
+              mesh_order = mesh->GetNodalFESpace()->GetFE(0)->GetOrder();
 
    const FiniteElementCollection *fec_in =  field_in.FESpace()->FEColl();
    const H1_FECollection *fec_h1 = dynamic_cast<const H1_FECollection *>(fec_in);
@@ -829,8 +812,7 @@ void FindPointsGSLIB::Interpolate(const GridFunction &field_in,
 
    if (fec_h1 && gf_order == mesh_order &&
        fec_h1->GetBasisType() == BasisType::GaussLobatto &&
-       field_in.FESpace()->IsVariableOrder() ==
-       mesh->GetNodalFESpace()->IsVariableOrder())
+       !field_in.FESpace()->IsVariableOrder())
    {
       InterpolateH1(field_in, field_out);
       return;
@@ -904,21 +886,12 @@ void FindPointsGSLIB::InterpolateH1(const GridFunction &field_in,
                                     Vector &field_out)
 {
    FiniteElementSpace ind_fes(mesh, field_in.FESpace()->FEColl());
-   if (field_in.FESpace()->IsVariableOrder())
-   {
-      for (int e = 0; e < ind_fes.GetMesh()->GetNE(); e++)
-      {
-         ind_fes.SetElementOrder(e, field_in.FESpace()->GetElementOrder(e));
-      }
-      ind_fes.Update(false);
-   }
    GridFunction field_in_scalar(&ind_fes);
    Vector node_vals;
 
    const int ncomp      = field_in.FESpace()->GetVDim(),
-             points_fld = field_in.Size() / ncomp;
-   MFEM_VERIFY(points_cnt == gsl_code.Size(),
-               "FindPointsGSLIB::InterpolateH1: Inconsistent size of gsl_code");
+             points_fld = field_in.Size() / ncomp,
+             points_cnt = gsl_code.Size();
 
    field_out.SetSize(points_cnt*ncomp);
    field_out = default_interp_value;
@@ -1138,7 +1111,8 @@ void OversetFindPointsGSLIB::Setup(Mesh &m, const int meshid,
                                    const int npt_max)
 {
    MFEM_VERIFY(m.GetNodes() != NULL, "Mesh nodes are required.");
-   const int meshOrder = m.GetNodes()->FESpace()->GetMaxElementOrder();
+   MFEM_VERIFY(!(m.GetNodes()->FESpace()->IsVariableOrder()),
+               "Variable order mesh is not currently supported.");
 
    // FreeData if OversetFindPointsGSLIB::Setup has been called already
    if (setupflag) { FreeData(); }
@@ -1154,29 +1128,21 @@ void OversetFindPointsGSLIB::Setup(Mesh &m, const int meshid,
    {
       if (ir_split[0]) { delete ir_split[0]; ir_split[0] = NULL; }
       ir_split[0] = new IntegrationRule(3*pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[0], ir_split[0], meshOrder);
-
-      if (ir_split[1]) { delete ir_split[1]; ir_split[1] = NULL; }
-      ir_split[1] = new IntegrationRule(pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[1], ir_split[1], meshOrder);
+      SetupIntegrationRuleForSplitMesh(mesh_split[0], ir_split[0], fe->GetOrder());
    }
    else if (dim == 3)
    {
-      if (ir_split[0]) { delete ir_split[0]; ir_split[0] = NULL; }
-      ir_split[0] = new IntegrationRule(pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[0], ir_split[0], meshOrder);
-
       if (ir_split[1]) { delete ir_split[1]; ir_split[1] = NULL; }
       ir_split[1] = new IntegrationRule(4*pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[1], ir_split[1], meshOrder);
+      SetupIntegrationRuleForSplitMesh(mesh_split[1], ir_split[1], fe->GetOrder());
 
       if (ir_split[2]) { delete ir_split[2]; ir_split[2] = NULL; }
       ir_split[2] = new IntegrationRule(3*pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[2], ir_split[2], meshOrder);
+      SetupIntegrationRuleForSplitMesh(mesh_split[2], ir_split[2], fe->GetOrder());
 
       if (ir_split[3]) { delete ir_split[3]; ir_split[3] = NULL; }
       ir_split[3] = new IntegrationRule(8*pow(dof1D, dim));
-      SetupIntegrationRuleForSplitMesh(mesh_split[3], ir_split[3], meshOrder);
+      SetupIntegrationRuleForSplitMesh(mesh_split[3], ir_split[3], fe->GetOrder());
    }
 
    GetNodalValues(mesh->GetNodes(), gsl_mesh);

fem/gslib.hpp

@@ -57,9 +57,7 @@ public:
 protected:
    Mesh *mesh;
    Array<Mesh *> mesh_split;  // Meshes used to split simplices.
-   // IntegrationRules for simplex->Quad/Hex and to project to highest polynomial
-   // order in-case of p-refinement.
-   Array<IntegrationRule *> ir_split;
+   Array<IntegrationRule *> ir_split; // IntegrationRules for simplex->Quad/Hex
    Array<FiniteElementSpace *>
    fes_rst_map; // FESpaces to map info Quad/Hex->Simplex
    Array<GridFunction *> gf_rst_map; // GridFunctions to map info Quad/Hex->Simplex

fem/integ/bilininteg_curlcurl_pa.cpp

@@ -1,205 +0,0 @@
-// Copyright (c) 2010-2023, Lawrence Livermore National Security, LLC. Produced
-// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
-// LICENSE and NOTICE for details. LLNL-CODE-806117.
-//
-// This file is part of the MFEM library. For more information and source code
-// availability visit https://mfem.org.
-//
-// MFEM is free software; you can redistribute it and/or modify it under the
-// terms of the BSD-3 license. We welcome feedback and contributions, see file
-// CONTRIBUTING.md for details.
-
-#include "../qfunction.hpp"
-#include "bilininteg_hcurl_kernels.hpp"
-
-namespace mfem
-{
-
-void CurlCurlIntegrator::AssemblePA(const FiniteElementSpace &fes)
-{
-   // Assumes tensor-product elements
-   Mesh *mesh = fes.GetMesh();
-   const FiniteElement *fel = fes.GetFE(0);
-
-   const VectorTensorFiniteElement *el =
-      dynamic_cast<const VectorTensorFiniteElement*>(fel);
-   MFEM_VERIFY(el != NULL, "Only VectorTensorFiniteElement is supported!");
-
-   const IntegrationRule *ir
-      = IntRule ? IntRule : &MassIntegrator::GetRule(*el, *el,
-                                                     *mesh->GetElementTransformation(0));
-
-   const int dims = el->GetDim();
-   MFEM_VERIFY(dims == 2 || dims == 3, "");
-
-   nq = ir->GetNPoints();
-   dim = mesh->Dimension();
-   MFEM_VERIFY(dim == 2 || dim == 3, "");
-
-   ne = fes.GetNE();
-   geom = mesh->GetGeometricFactors(*ir, GeometricFactors::JACOBIANS);
-   mapsC = &el->GetDofToQuad(*ir, DofToQuad::TENSOR);
-   mapsO = &el->GetDofToQuadOpen(*ir, DofToQuad::TENSOR);
-   dofs1D = mapsC->ndof;
-   quad1D = mapsC->nqpt;
-
-   MFEM_VERIFY(dofs1D == mapsO->ndof + 1 && quad1D == mapsO->nqpt, "");
-
-   QuadratureSpace qs(*mesh, *ir);
-   CoefficientVector coeff(qs, CoefficientStorage::SYMMETRIC);
-   if (Q) { coeff.Project(*Q); }
-   else if (MQ) { coeff.ProjectTranspose(*MQ); }
-   else if (DQ) { coeff.Project(*DQ); }
-   else { coeff.SetConstant(1.0); }
-
-   const int coeff_dim = coeff.GetVDim();
-   symmetric = (coeff_dim != dim*dim);
-   const int sym_dims = (dims * (dims + 1)) / 2; // 1x1: 1, 2x2: 3, 3x3: 6
-   const int ndata = (dim == 2) ? 1 : (symmetric ? sym_dims : dim*dim);
-   pa_data.SetSize(ndata * nq * ne, Device::GetMemoryType());
-
-   if (el->GetDerivType() != mfem::FiniteElement::CURL)
-   {
-      MFEM_ABORT("Unknown kernel.");
-   }
-
-   if (dim == 3)
-   {
-      internal::PACurlCurlSetup3D(quad1D, coeff_dim, ne, ir->GetWeights(), geom->J,
-                                  coeff, pa_data);
-   }
-   else
-   {
-      internal::PACurlCurlSetup2D(quad1D, ne, ir->GetWeights(), geom->J, coeff,
-                                  pa_data);
-   }
-}
-
-void CurlCurlIntegrator::AssembleDiagonalPA(Vector& diag)
-{
-   if (dim == 3)
-   {
-      if (Device::Allows(Backend::DEVICE_MASK))
-      {
-         const int ID = (dofs1D << 4) | quad1D;
-         switch (ID)
-         {
-            case 0x23:
-               return internal::SmemPACurlCurlAssembleDiagonal3D<2,3>(
-                         dofs1D,
-                         quad1D,
-                         symmetric, ne,
-                         mapsO->B, mapsC->B,
-                         mapsO->G, mapsC->G,
-                         pa_data, diag);
-            case 0x34:
-               return internal::SmemPACurlCurlAssembleDiagonal3D<3,4>(
-                         dofs1D,
-                         quad1D,
-                         symmetric, ne,
-                         mapsO->B, mapsC->B,
-                         mapsO->G, mapsC->G,
-                         pa_data, diag);
-            case 0x45:
-               return internal::SmemPACurlCurlAssembleDiagonal3D<4,5>(
-                         dofs1D,
-                         quad1D,
-                         symmetric, ne,
-                         mapsO->B, mapsC->B,
-                         mapsO->G, mapsC->G,
-                         pa_data, diag);
-            case 0x56:
-               return internal::SmemPACurlCurlAssembleDiagonal3D<5,6>(
-                         dofs1D,
-                         quad1D,
-                         symmetric, ne,
-                         mapsO->B, mapsC->B,
-                         mapsO->G, mapsC->G,
-                         pa_data, diag);
-            default:
-               return internal::SmemPACurlCurlAssembleDiagonal3D(
-                         dofs1D, quad1D,
-                         symmetric, ne,
-                         mapsO->B, mapsC->B,
-                         mapsO->G, mapsC->G,
-                         pa_data, diag);
-         }
-      }
-      else
-      {
-         internal::PACurlCurlAssembleDiagonal3D(dofs1D, quad1D, symmetric, ne,
-                                                mapsO->B, mapsC->B,
-                                                mapsO->G, mapsC->G,
-                                                pa_data, diag);
-      }
-   }
-   else if (dim == 2)
-   {
-      internal::PACurlCurlAssembleDiagonal2D(dofs1D, quad1D, ne,
-                                             mapsO->B, mapsC->G, pa_data, diag);
-   }
-   else
-   {
-      MFEM_ABORT("Unsupported dimension!");
-   }
-}
-
-void CurlCurlIntegrator::AddMultPA(const Vector &x, Vector &y) const
-{
-   if (dim == 3)
-   {
-      if (Device::Allows(Backend::DEVICE_MASK))
-      {
-         const int ID = (dofs1D << 4) | quad1D;
-         switch (ID)
-         {
-            case 0x23:
-               return internal::SmemPACurlCurlApply3D<2,3>(
-                         dofs1D, quad1D,
-                         symmetric, ne,
-                         mapsO->B, mapsC->B, mapsO->Bt, mapsC->Bt,
-                         mapsC->G, mapsC->Gt, pa_data, x, y);
-            case 0x34:
-               return internal::SmemPACurlCurlApply3D<3,4>(
-                         dofs1D, quad1D,
-                         symmetric, ne,
-                         mapsO->B, mapsC->B, mapsO->Bt, mapsC->Bt,
-                         mapsC->G, mapsC->Gt, pa_data, x, y);
-            case 0x45:
-               return internal::SmemPACurlCurlApply3D<4,5>(
-                         dofs1D, quad1D,
-                         symmetric, ne,
-                         mapsO->B, mapsC->B, mapsO->Bt, mapsC->Bt,
-                         mapsC->G, mapsC->Gt, pa_data, x, y);
-            case 0x56:
-               return internal::SmemPACurlCurlApply3D<5,6>(
-                         dofs1D, quad1D,
-                         symmetric, ne,
-                         mapsO->B, mapsC->B, mapsO->Bt, mapsC->Bt,
-                         mapsC->G, mapsC->Gt, pa_data, x, y);
-            default:
-               return internal::SmemPACurlCurlApply3D(
-                         dofs1D, quad1D, symmetric, ne,
-                         mapsO->B, mapsC->B, mapsO->Bt, mapsC->Bt,
-                         mapsC->G, mapsC->Gt, pa_data, x, y);
-         }
-      }
-      else
-      {
-         internal::PACurlCurlApply3D(dofs1D, quad1D, symmetric, ne, mapsO->B, mapsC->B,
-                                     mapsO->Bt, mapsC->Bt, mapsC->G, mapsC->Gt,
-                                     pa_data, x, y);
-      }
-   }
-   else if (dim == 2)
-   {
-      internal::PACurlCurlApply2D(dofs1D, quad1D, ne, mapsO->B, mapsO->Bt,
-                                  mapsC->G, mapsC->Gt, pa_data, x, y);
-   }
-   else
-   {
-      MFEM_ABORT("Unsupported dimension!");
-   }
-}
-
-} // namespace mfem

fem/integ/bilininteg_diffusion_kernels.cpp

@@ -1,578 +0,0 @@
-// Copyright (c) 2010-2023, Lawrence Livermore National Security, LLC. Produced
-// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
-// LICENSE and NOTICE for details. LLNL-CODE-806117.
-//
-// This file is part of the MFEM library. For more information and source code
-// availability visit https://mfem.org.
-//
-// MFEM is free software; you can redistribute it and/or modify it under the
-// terms of the BSD-3 license. We welcome feedback and contributions, see file
-// CONTRIBUTING.md for details.
-
-#include "bilininteg_diffusion_kernels.hpp"
-
-namespace mfem
-{
-
-namespace internal
-{
-
-template<>
-void PADiffusionSetup2D<2>(const int Q1D,
-                           const int coeffDim,
-                           const int NE,
-                           const Array<double> &w,
-                           const Vector &j,
-                           const Vector &c,
-                           Vector &d);
-
-template<>
-void PADiffusionSetup2D<3>(const int Q1D,
-                           const int coeffDim,
-                           const int NE,
-                           const Array<double> &w,
-                           const Vector &j,
-                           const Vector &c,
-                           Vector &d);
-
-void PADiffusionSetup(const int dim,
-                      const int sdim,
-                      const int D1D,
-                      const int Q1D,
-                      const int coeffDim,
-                      const int NE,
-                      const Array<double> &W,
-                      const Vector &J,
-                      const Vector &C,
-                      Vector &D)
-{
-   if (dim == 1) { MFEM_ABORT("dim==1 not supported in PADiffusionSetup"); }
-   if (dim == 2)
-   {
-#ifdef MFEM_USE_OCCA
-      if (DeviceCanUseOcca())
-      {
-         OccaPADiffusionSetup2D(D1D, Q1D, NE, W, J, C, D);
-         return;
-      }
-#else
-      MFEM_CONTRACT_VAR(D1D);
-#endif // MFEM_USE_OCCA
-      if (sdim == 2) { PADiffusionSetup2D<2>(Q1D, coeffDim, NE, W, J, C, D); }
-      if (sdim == 3) { PADiffusionSetup2D<3>(Q1D, coeffDim, NE, W, J, C, D); }
-   }
-   if (dim == 3)
-   {
-#ifdef MFEM_USE_OCCA
-      if (DeviceCanUseOcca())
-      {
-         OccaPADiffusionSetup3D(D1D, Q1D, NE, W, J, C, D);
-         return;
-      }
-#endif // MFEM_USE_OCCA
-      PADiffusionSetup3D(Q1D, coeffDim, NE, W, J, C, D);
-   }
-}
-
-template<>
-void PADiffusionSetup2D<2>(const int Q1D,
-                           const int coeffDim,
-                           const int NE,
-                           const Array<double> &w,
-                           const Vector &j,
-                           const Vector &c,
-                           Vector &d)
-{
-   const bool symmetric = (coeffDim != 4);
-   const bool const_c = c.Size() == 1;
-   MFEM_VERIFY(coeffDim < 3 ||
-               !const_c, "Constant matrix coefficient not supported");
-   const auto W = Reshape(w.Read(), Q1D,Q1D);
-   const auto J = Reshape(j.Read(), Q1D,Q1D,2,2,NE);
-   const auto C = const_c ? Reshape(c.Read(), 1,1,1,1) :
-                  Reshape(c.Read(), coeffDim,Q1D,Q1D,NE);
-   auto D = Reshape(d.Write(), Q1D,Q1D, symmetric ? 3 : 4, NE);
-   mfem::forall_2D(NE, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
-   {
-      MFEM_FOREACH_THREAD(qx,x,Q1D)
-      {
-         MFEM_FOREACH_THREAD(qy,y,Q1D)
-         {
-            const double J11 = J(qx,qy,0,0,e);
-            const double J21 = J(qx,qy,1,0,e);
-            const double J12 = J(qx,qy,0,1,e);
-            const double J22 = J(qx,qy,1,1,e);
-            const double w_detJ = W(qx,qy) / ((J11*J22)-(J21*J12));
-            if (coeffDim == 3 || coeffDim == 4) // Matrix coefficient
-            {
-               // First compute entries of R = MJ^{-T}, without det J factor.
-               const double M11 = C(0,qx,qy,e);
-               const double M12 = C(1,qx,qy,e);
-               const double M21 = symmetric ? M12 : C(2,qx,qy,e);
-               const double M22 = symmetric ? C(2,qx,qy,e) : C(3,qx,qy,e);
-               const double R11 = M11*J22 - M12*J12;
-               const double R21 = M21*J22 - M22*J12;
-               const double R12 = -M11*J21 + M12*J11;
-               const double R22 = -M21*J21 + M22*J11;
-
-               // Now set y to J^{-1}R.
-               D(qx,qy,0,e) = w_detJ * ( J22*R11 - J12*R21); // 1,1
-               D(qx,qy,1,e) = w_detJ * (-J21*R11 + J11*R21); // 2,1
-               D(qx,qy,2,e) = w_detJ * (symmetric ? (-J21*R12 + J11*R22) :
-                                        (J22*R12 - J12*R22)); // 2,2 or 1,2
-               if (!symmetric)
-               {
-                  D(qx,qy,3,e) = w_detJ * (-J21*R12 + J11*R22); // 2,2
-               }
-            }
-            else // Vector or scalar coefficient
-            {
-               const double C1 = const_c ? C(0,0,0,0) : C(0,qx,qy,e);
-               const double C2 = const_c ? C(0,0,0,0) :
-                                 (coeffDim == 2 ? C(1,qx,qy,e) : C(0,qx,qy,e));
-
-               D(qx,qy,0,e) =  w_detJ * (C2*J12*J12 + C1*J22*J22); // 1,1
-               D(qx,qy,1,e) = -w_detJ * (C2*J12*J11 + C1*J22*J21); // 1,2
-               D(qx,qy,2,e) =  w_detJ * (C2*J11*J11 + C1*J21*J21); // 2,2
-            }
-         }
-      }
-   });
-}
-
-template<>
-void PADiffusionSetup2D<3>(const int Q1D,
-                           const int coeffDim,
-                           const int NE,
-                           const Array<double> &w,
-                           const Vector &j,
-                           const Vector &c,
-                           Vector &d)
-{
-   MFEM_VERIFY(coeffDim == 1, "Matrix and vector coefficients not supported");
-   constexpr int DIM = 2;
-   constexpr int SDIM = 3;
-   const bool const_c = c.Size() == 1;
-   const auto W = Reshape(w.Read(), Q1D,Q1D);
-   const auto J = Reshape(j.Read(), Q1D,Q1D,SDIM,DIM,NE);
-   const auto C = const_c ? Reshape(c.Read(), 1,1,1) :
-                  Reshape(c.Read(), Q1D,Q1D,NE);
-   auto D = Reshape(d.Write(), Q1D,Q1D, 3, NE);
-   mfem::forall_2D(NE, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
-   {
-      MFEM_FOREACH_THREAD(qx,x,Q1D)
-      {
-         MFEM_FOREACH_THREAD(qy,y,Q1D)
-         {
-            const double wq = W(qx,qy);
-            const double J11 = J(qx,qy,0,0,e);
-            const double J21 = J(qx,qy,1,0,e);
-            const double J31 = J(qx,qy,2,0,e);
-            const double J12 = J(qx,qy,0,1,e);
-            const double J22 = J(qx,qy,1,1,e);
-            const double J32 = J(qx,qy,2,1,e);
-            const double E = J11*J11 + J21*J21 + J31*J31;
-            const double G = J12*J12 + J22*J22 + J32*J32;
-            const double F = J11*J12 + J21*J22 + J31*J32;
-            const double iw = 1.0 / sqrt(E*G - F*F);
-            const double coeff = const_c ? C(0,0,0) : C(qx,qy,e);
-            const double alpha = wq * coeff * iw;
-            D(qx,qy,0,e) =  alpha * G; // 1,1
-            D(qx,qy,1,e) = -alpha * F; // 1,2
-            D(qx,qy,2,e) =  alpha * E; // 2,2
-         }
-      }
-   });
-}
-
-void PADiffusionSetup3D(const int Q1D,
-                        const int coeffDim,
-                        const int NE,
-                        const Array<double> &w,
-                        const Vector &j,
-                        const Vector &c,
-                        Vector &d)
-{
-   const bool symmetric = (coeffDim != 9);
-   const bool const_c = c.Size() == 1;
-   MFEM_VERIFY(coeffDim < 6 ||
-               !const_c, "Constant matrix coefficient not supported");
-   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(), 1,1,1,1,1) :
-                  Reshape(c.Read(), coeffDim,Q1D,Q1D,Q1D,NE);
-   auto D = Reshape(d.Write(), Q1D,Q1D,Q1D, symmetric ? 6 : 9, NE);
-   mfem::forall_3D(NE, Q1D, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
-   {
-      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 w_detJ = W(qx,qy,qz) / 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);
-
-               if (coeffDim == 6 || coeffDim == 9) // Matrix coefficient version
-               {
-                  // Compute entries of R = MJ^{-T} = M adj(J)^T, without det J.
-                  const double M11 = C(0, qx,qy,qz, e);
-                  const double M12 = C(1, qx,qy,qz, e);
-                  const double M13 = C(2, qx,qy,qz, e);
-                  const double M21 = (!symmetric) ? C(3, qx,qy,qz, e) : M12;
-                  const double M22 = (!symmetric) ? C(4, qx,qy,qz, e) : C(3, qx,qy,qz, e);
-                  const double M23 = (!symmetric) ? C(5, qx,qy,qz, e) : C(4, qx,qy,qz, e);
-                  const double M31 = (!symmetric) ? C(6, qx,qy,qz, e) : M13;
-                  const double M32 = (!symmetric) ? C(7, qx,qy,qz, e) : M23;
-                  const double M33 = (!symmetric) ? C(8, qx,qy,qz, e) : C(5, qx,qy,qz, e);
-
-                  const double R11 = M11*A11 + M12*A12 + M13*A13;
-                  const double R12 = M11*A21 + M12*A22 + M13*A23;
-                  const double R13 = M11*A31 + M12*A32 + M13*A33;
-                  const double R21 = M21*A11 + M22*A12 + M23*A13;
-                  const double R22 = M21*A21 + M22*A22 + M23*A23;
-                  const double R23 = M21*A31 + M22*A32 + M23*A33;
-                  const double R31 = M31*A11 + M32*A12 + M33*A13;
-                  const double R32 = M31*A21 + M32*A22 + M33*A23;
-                  const double R33 = M31*A31 + M32*A32 + M33*A33;
-
-                  // Now set D to J^{-1} R = adj(J) R
-                  D(qx,qy,qz,0,e) = w_detJ * (A11*R11 + A12*R21 + A13*R31); // 1,1
-                  const double D12 = w_detJ * (A11*R12 + A12*R22 + A13*R32);
-                  D(qx,qy,qz,1,e) = D12; // 1,2
-                  D(qx,qy,qz,2,e) = w_detJ * (A11*R13 + A12*R23 + A13*R33); // 1,3
-
-                  const double D22 = w_detJ * (A21*R12 + A22*R22 + A23*R32);
-                  const double D23 = w_detJ * (A21*R13 + A22*R23 + A23*R33);
-
-                  const double D33 = w_detJ * (A31*R13 + A32*R23 + A33*R33);
-
-                  D(qx,qy,qz,4,e) = symmetric ? D23 : D22; // 2,3 or 2,2
-                  D(qx,qy,qz,5,e) = symmetric ? D33 : D23; // 3,3 or 2,3
-
-                  if (symmetric)
-                  {
-                     D(qx,qy,qz,3,e) = D22; // 2,2
-                  }
-                  else
-                  {
-                     D(qx,qy,qz,3,e) = w_detJ * (A21*R11 + A22*R21 + A23*R31); // 2,1
-                     D(qx,qy,qz,6,e) = w_detJ * (A31*R11 + A32*R21 + A33*R31); // 3,1
-                     D(qx,qy,qz,7,e) = w_detJ * (A31*R12 + A32*R22 + A33*R32); // 3,2
-                     D(qx,qy,qz,8,e) = D33; // 3,3
-                  }
-               }
-               else  // Vector or scalar coefficient version
-               {
-                  const double C1 = const_c ? C(0,0,0,0,0) : C(0,qx,qy,qz,e);
-                  const double C2 = const_c ? C(0,0,0,0,0) :
-                                    (coeffDim == 3 ? C(1,qx,qy,qz,e) : C(0,qx,qy,qz,e));
-                  const double C3 = const_c ? C(0,0,0,0,0) :
-                                    (coeffDim == 3 ? C(2,qx,qy,qz,e) : C(0,qx,qy,qz,e));
-
-                  // detJ J^{-1} J^{-T} = (1/detJ) adj(J) adj(J)^T
-                  D(qx,qy,qz,0,e) = w_detJ * (C1*A11*A11 + C2*A12*A12 + C3*A13*A13); // 1,1
-                  D(qx,qy,qz,1,e) = w_detJ * (C1*A11*A21 + C2*A12*A22 + C3*A13*A23); // 2,1
-                  D(qx,qy,qz,2,e) = w_detJ * (C1*A11*A31 + C2*A12*A32 + C3*A13*A33); // 3,1
-                  D(qx,qy,qz,3,e) = w_detJ * (C1*A21*A21 + C2*A22*A22 + C3*A23*A23); // 2,2
-                  D(qx,qy,qz,4,e) = w_detJ * (C1*A21*A31 + C2*A22*A32 + C3*A23*A33); // 3,2
-                  D(qx,qy,qz,5,e) = w_detJ * (C1*A31*A31 + C2*A32*A32 + C3*A33*A33); // 3,3
-               }
-            }
-         }
-      }
-   });
-}
-
-#ifdef MFEM_USE_OCCA
-void OccaPADiffusionSetup2D(const int D1D,
-                            const int Q1D,
-                            const int NE,
-                            const Array<double> &W,
-                            const Vector &J,
-                            const Vector &C,
-                            Vector &op)
-{
-   occa::properties props;
-   props["defines/D1D"] = D1D;
-   props["defines/Q1D"] = Q1D;
-   const occa::memory o_W = OccaMemoryRead(W.GetMemory(), W.Size());
-   const occa::memory o_J = OccaMemoryRead(J.GetMemory(), J.Size());
-   const occa::memory o_C = OccaMemoryRead(C.GetMemory(), C.Size());
-   occa::memory o_op = OccaMemoryWrite(op.GetMemory(), op.Size());
-   const bool const_c = C.Size() == 1;
-   const occa_id_t id = std::make_pair(D1D,Q1D);
-   static occa_kernel_t OccaDiffSetup2D_ker;
-   if (OccaDiffSetup2D_ker.find(id) == OccaDiffSetup2D_ker.end())
-   {
-      const occa::kernel DiffusionSetup2D =
-         mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
-                                     "DiffusionSetup2D", props);
-      OccaDiffSetup2D_ker.emplace(id, DiffusionSetup2D);
-   }
-   OccaDiffSetup2D_ker.at(id)(NE, o_W, o_J, o_C, o_op, const_c);
-}
-
-void OccaPADiffusionSetup3D(const int D1D,
-                            const int Q1D,
-                            const int NE,
-                            const Array<double> &W,
-                            const Vector &J,
-                            const Vector &C,
-                            Vector &op)
-{
-   occa::properties props;
-   props["defines/D1D"] = D1D;
-   props["defines/Q1D"] = Q1D;
-   const occa::memory o_W = OccaMemoryRead(W.GetMemory(), W.Size());
-   const occa::memory o_J = OccaMemoryRead(J.GetMemory(), J.Size());
-   const occa::memory o_C = OccaMemoryRead(C.GetMemory(), C.Size());
-   occa::memory o_op = OccaMemoryWrite(op.GetMemory(), op.Size());
-   const bool const_c = C.Size() == 1;
-   const occa_id_t id = std::make_pair(D1D,Q1D);
-   static occa_kernel_t OccaDiffSetup3D_ker;
-   if (OccaDiffSetup3D_ker.find(id) == OccaDiffSetup3D_ker.end())
-   {
-      const occa::kernel DiffusionSetup3D =
-         mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
-                                     "DiffusionSetup3D", props);
-      OccaDiffSetup3D_ker.emplace(id, DiffusionSetup3D);
-   }
-   OccaDiffSetup3D_ker.at(id)(NE, o_W, o_J, o_C, o_op, const_c);
-}
-#endif // MFEM_USE_OCCA
-
-void PADiffusionAssembleDiagonal(const int dim,
-                                 const int D1D,
-                                 const int Q1D,
-                                 const int NE,
-                                 const bool symm,
-                                 const Array<double> &B,
-                                 const Array<double> &G,
-                                 const Vector &D,
-                                 Vector &Y)
-{
-   if (dim == 2)
-   {
-      switch ((D1D << 4 ) | Q1D)
-      {
-         case 0x22: return SmemPADiffusionDiagonal2D<2,2,8>(NE,symm,B,G,D,Y);
-         case 0x33: return SmemPADiffusionDiagonal2D<3,3,8>(NE,symm,B,G,D,Y);
-         case 0x44: return SmemPADiffusionDiagonal2D<4,4,4>(NE,symm,B,G,D,Y);
-         case 0x55: return SmemPADiffusionDiagonal2D<5,5,4>(NE,symm,B,G,D,Y);
-         case 0x66: return SmemPADiffusionDiagonal2D<6,6,2>(NE,symm,B,G,D,Y);
-         case 0x77: return SmemPADiffusionDiagonal2D<7,7,2>(NE,symm,B,G,D,Y);
-         case 0x88: return SmemPADiffusionDiagonal2D<8,8,1>(NE,symm,B,G,D,Y);
-         case 0x99: return SmemPADiffusionDiagonal2D<9,9,1>(NE,symm,B,G,D,Y);
-         default: return PADiffusionDiagonal2D(NE,symm,B,G,D,Y,D1D,Q1D);
-      }
-   }
-   else if (dim == 3)
-   {
-      switch ((D1D << 4 ) | Q1D)
-      {
-         case 0x22: return SmemPADiffusionDiagonal3D<2,2>(NE,symm,B,G,D,Y);
-         case 0x23: return SmemPADiffusionDiagonal3D<2,3>(NE,symm,B,G,D,Y);
-         case 0x34: return SmemPADiffusionDiagonal3D<3,4>(NE,symm,B,G,D,Y);
-         case 0x45: return SmemPADiffusionDiagonal3D<4,5>(NE,symm,B,G,D,Y);
-         case 0x46: return SmemPADiffusionDiagonal3D<4,6>(NE,symm,B,G,D,Y);
-         case 0x56: return SmemPADiffusionDiagonal3D<5,6>(NE,symm,B,G,D,Y);
-         case 0x67: return SmemPADiffusionDiagonal3D<6,7>(NE,symm,B,G,D,Y);
-         case 0x78: return SmemPADiffusionDiagonal3D<7,8>(NE,symm,B,G,D,Y);
-         case 0x89: return SmemPADiffusionDiagonal3D<8,9>(NE,symm,B,G,D,Y);
-         case 0x9A: return SmemPADiffusionDiagonal3D<9,10>(NE,symm,B,G,D,Y);
-         default: return PADiffusionDiagonal3D(NE,symm,B,G,D,Y,D1D,Q1D);
-      }
-   }
-   MFEM_ABORT("Unknown kernel.");
-}
-
-void PADiffusionApply(const int dim,
-                      const int D1D,
-                      const int Q1D,
-                      const int NE,
-                      const bool symm,
-                      const Array<double> &B,
-                      const Array<double> &G,
-                      const Array<double> &Bt,
-                      const Array<double> &Gt,
-                      const Vector &D,
-                      const Vector &X,
-                      Vector &Y)
-{
-#ifdef MFEM_USE_OCCA
-   if (DeviceCanUseOcca())
-   {
-      if (dim == 2)
-      {
-         OccaPADiffusionApply2D(D1D,Q1D,NE,B,G,Bt,Gt,D,X,Y);
-         return;
-      }
-      if (dim == 3)
-      {
-         OccaPADiffusionApply3D(D1D,Q1D,NE,B,G,Bt,Gt,D,X,Y);
-         return;
-      }
-      MFEM_ABORT("OCCA PADiffusionApply unknown kernel!");
-   }
-#endif // MFEM_USE_OCCA
-   const int id = (D1D << 4) | Q1D;
-
-   if (dim == 2)
-   {
-      switch (id)
-      {
-         case 0x22: return SmemPADiffusionApply2D<2,2,16>(NE,symm,B,G,D,X,Y);
-         case 0x33: return SmemPADiffusionApply2D<3,3,16>(NE,symm,B,G,D,X,Y);
-         case 0x44: return SmemPADiffusionApply2D<4,4,8>(NE,symm,B,G,D,X,Y);
-         case 0x55: return SmemPADiffusionApply2D<5,5,8>(NE,symm,B,G,D,X,Y);
-         case 0x66: return SmemPADiffusionApply2D<6,6,4>(NE,symm,B,G,D,X,Y);
-         case 0x77: return SmemPADiffusionApply2D<7,7,4>(NE,symm,B,G,D,X,Y);
-         case 0x88: return SmemPADiffusionApply2D<8,8,2>(NE,symm,B,G,D,X,Y);
-         case 0x99: return SmemPADiffusionApply2D<9,9,2>(NE,symm,B,G,D,X,Y);
-         default:   return PADiffusionApply2D(NE,symm,B,G,Bt,Gt,D,X,Y,D1D,Q1D);
-      }
-   }
-
-   if (dim == 3)
-   {
-      switch (id)
-      {
-         case 0x22: return SmemPADiffusionApply3D<2,2>(NE,symm,B,G,D,X,Y);
-         case 0x23: return SmemPADiffusionApply3D<2,3>(NE,symm,B,G,D,X,Y);
-         case 0x34: return SmemPADiffusionApply3D<3,4>(NE,symm,B,G,D,X,Y);
-         case 0x45: return SmemPADiffusionApply3D<4,5>(NE,symm,B,G,D,X,Y);
-         case 0x46: return SmemPADiffusionApply3D<4,6>(NE,symm,B,G,D,X,Y);
-         case 0x56: return SmemPADiffusionApply3D<5,6>(NE,symm,B,G,D,X,Y);
-         case 0x58: return SmemPADiffusionApply3D<5,8>(NE,symm,B,G,D,X,Y);
-         case 0x67: return SmemPADiffusionApply3D<6,7>(NE,symm,B,G,D,X,Y);
-         case 0x78: return SmemPADiffusionApply3D<7,8>(NE,symm,B,G,D,X,Y);
-         case 0x89: return SmemPADiffusionApply3D<8,9>(NE,symm,B,G,D,X,Y);
-         default:   return PADiffusionApply3D(NE,symm,B,G,Bt,Gt,D,X,Y,D1D,Q1D);
-      }
-   }
-   MFEM_ABORT("Unknown kernel: 0x"<<std::hex << id << std::dec);
-}
-
-#ifdef MFEM_USE_OCCA
-void OccaPADiffusionApply2D(const int D1D,
-                            const int Q1D,
-                            const int NE,
-                            const Array<double> &B,
-                            const Array<double> &G,
-                            const Array<double> &Bt,
-                            const Array<double> &Gt,
-                            const Vector &D,
-                            const Vector &X,
-                            Vector &Y)
-{
-   occa::properties props;
-   props["defines/D1D"] = D1D;
-   props["defines/Q1D"] = Q1D;
-   const occa::memory o_B = OccaMemoryRead(B.GetMemory(), B.Size());
-   const occa::memory o_G = OccaMemoryRead(G.GetMemory(), G.Size());
-   const occa::memory o_Bt = OccaMemoryRead(Bt.GetMemory(), Bt.Size());
-   const occa::memory o_Gt = OccaMemoryRead(Gt.GetMemory(), Gt.Size());
-   const occa::memory o_D = OccaMemoryRead(D.GetMemory(), D.Size());
-   const occa::memory o_X = OccaMemoryRead(X.GetMemory(), X.Size());
-   occa::memory o_Y = OccaMemoryReadWrite(Y.GetMemory(), Y.Size());
-   const occa_id_t id = std::make_pair(D1D,Q1D);
-   if (!Device::Allows(Backend::OCCA_CUDA))
-   {
-      static occa_kernel_t OccaDiffApply2D_cpu;
-      if (OccaDiffApply2D_cpu.find(id) == OccaDiffApply2D_cpu.end())
-      {
-         const occa::kernel DiffusionApply2D_CPU =
-            mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
-                                        "DiffusionApply2D_CPU", props);
-         OccaDiffApply2D_cpu.emplace(id, DiffusionApply2D_CPU);
-      }
-      OccaDiffApply2D_cpu.at(id)(NE, o_B, o_G, o_Bt, o_Gt, o_D, o_X, o_Y);
-   }
-   else
-   {
-      static occa_kernel_t OccaDiffApply2D_gpu;
-      if (OccaDiffApply2D_gpu.find(id) == OccaDiffApply2D_gpu.end())
-      {
-         const occa::kernel DiffusionApply2D_GPU =
-            mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
-                                        "DiffusionApply2D_GPU", props);
-         OccaDiffApply2D_gpu.emplace(id, DiffusionApply2D_GPU);
-      }
-      OccaDiffApply2D_gpu.at(id)(NE, o_B, o_G, o_Bt, o_Gt, o_D, o_X, o_Y);
-   }
-}
-
-void OccaPADiffusionApply3D(const int D1D,
-                            const int Q1D,
-                            const int NE,
-                            const Array<double> &B,
-                            const Array<double> &G,
-                            const Array<double> &Bt,
-                            const Array<double> &Gt,
-                            const Vector &D,
-                            const Vector &X,
-                            Vector &Y)
-{
-   occa::properties props;
-   props["defines/D1D"] = D1D;
-   props["defines/Q1D"] = Q1D;
-   const occa::memory o_B = OccaMemoryRead(B.GetMemory(), B.Size());
-   const occa::memory o_G = OccaMemoryRead(G.GetMemory(), G.Size());
-   const occa::memory o_Bt = OccaMemoryRead(Bt.GetMemory(), Bt.Size());
-   const occa::memory o_Gt = OccaMemoryRead(Gt.GetMemory(), Gt.Size());
-   const occa::memory o_D = OccaMemoryRead(D.GetMemory(), D.Size());
-   const occa::memory o_X = OccaMemoryRead(X.GetMemory(), X.Size());
-   occa::memory o_Y = OccaMemoryReadWrite(Y.GetMemory(), Y.Size());
-   const occa_id_t id = std::make_pair(D1D,Q1D);
-   if (!Device::Allows(Backend::OCCA_CUDA))
-   {
-      static occa_kernel_t OccaDiffApply3D_cpu;
-      if (OccaDiffApply3D_cpu.find(id) == OccaDiffApply3D_cpu.end())
-      {
-         const occa::kernel DiffusionApply3D_CPU =
-            mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
-                                        "DiffusionApply3D_CPU", props);
-         OccaDiffApply3D_cpu.emplace(id, DiffusionApply3D_CPU);
-      }
-      OccaDiffApply3D_cpu.at(id)(NE, o_B, o_G, o_Bt, o_Gt, o_D, o_X, o_Y);
-   }
-   else
-   {
-      static occa_kernel_t OccaDiffApply3D_gpu;
-      if (OccaDiffApply3D_gpu.find(id) == OccaDiffApply3D_gpu.end())
-      {
-         const occa::kernel DiffusionApply3D_GPU =
-            mfem::OccaDev().buildKernel("occa://mfem/fem/occa.okl",
-                                        "DiffusionApply3D_GPU", props);
-         OccaDiffApply3D_gpu.emplace(id, DiffusionApply3D_GPU);
-      }
-      OccaDiffApply3D_gpu.at(id)(NE, o_B, o_G, o_Bt, o_Gt, o_D, o_X, o_Y);
-   }
-}
-#endif // MFEM_USE_OCCA
-
-} // namespace internal
-
-} // namespace mfem

fem/integ/bilininteg_diffusion_pa.cpp

@@ -1,118 +0,0 @@
-// Copyright (c) 2010-2023, Lawrence Livermore National Security, LLC. Produced
-// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
-// LICENSE and NOTICE for details. LLNL-CODE-806117.
-//
-// This file is part of the MFEM library. For more information and source code
-// availability visit https://mfem.org.
-//
-// MFEM is free software; you can redistribute it and/or modify it under the
-// terms of the BSD-3 license. We welcome feedback and contributions, see file
-// CONTRIBUTING.md for details.
-
-#include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
-#include "../qfunction.hpp"
-#include "../ceed/integrators/diffusion/diffusion.hpp"
-#include "bilininteg_diffusion_kernels.hpp"
-
-namespace mfem
-{
-
-void DiffusionIntegrator::AssemblePA(const FiniteElementSpace &fes)
-{
-   const MemoryType mt = (pa_mt == MemoryType::DEFAULT) ?
-                         Device::GetDeviceMemoryType() : pa_mt;
-   // Assuming the same element type
-   fespace = &fes;
-   Mesh *mesh = fes.GetMesh();
-   if (mesh->GetNE() == 0) { return; }
-   const FiniteElement &el = *fes.GetFE(0);
-   const IntegrationRule *ir = IntRule ? IntRule : &GetRule(el, el);
-   if (DeviceCanUseCeed())
-   {
-      delete ceedOp;
-      MFEM_VERIFY(!VQ && !MQ,
-                  "Only scalar coefficient supported for DiffusionIntegrator"
-                  " with libCEED");
-      const bool mixed = mesh->GetNumGeometries(mesh->Dimension()) > 1 ||
-                         fes.IsVariableOrder();
-      if (mixed)
-      {
-         ceedOp = new ceed::MixedPADiffusionIntegrator(*this, fes, Q);
-      }
-      else
-      {
-         ceedOp = new ceed::PADiffusionIntegrator(fes, *ir, Q);
-      }
-      return;
-   }
-   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();
-   geom = mesh->GetGeometricFactors(*ir, GeometricFactors::JACOBIANS, mt);
-   const int sdim = mesh->SpaceDimension();
-   maps = &el.GetDofToQuad(*ir, DofToQuad::TENSOR);
-   dofs1D = maps->ndof;
-   quad1D = maps->nqpt;
-
-   QuadratureSpace qs(*mesh, *ir);
-   CoefficientVector coeff(qs, CoefficientStorage::COMPRESSED);
-
-   if (MQ) { coeff.ProjectTranspose(*MQ); }
-   else if (VQ) { coeff.Project(*VQ); }
-   else if (Q) { coeff.Project(*Q); }
-   else { coeff.SetConstant(1.0); }
-
-   const int coeff_dim = coeff.GetVDim();
-   symmetric = (coeff_dim != dims*dims);
-   const int pa_size = symmetric ? symmDims : dims*dims;
-
-   pa_data.SetSize(pa_size * nq * ne, mt);
-   internal::PADiffusionSetup(dim, sdim, dofs1D, quad1D, coeff_dim, ne,
-                              ir->GetWeights(), geom->J, coeff, pa_data);
-}
-
-void DiffusionIntegrator::AssembleDiagonalPA(Vector &diag)
-{
-   if (DeviceCanUseCeed())
-   {
-      ceedOp->GetDiagonal(diag);
-   }
-   else
-   {
-      if (pa_data.Size()==0) { AssemblePA(*fespace); }
-      internal::PADiffusionAssembleDiagonal(dim, dofs1D, quad1D, ne, symmetric,
-                                            maps->B, maps->G, pa_data, diag);
-   }
-}
-
-void DiffusionIntegrator::AddMultPA(const Vector &x, Vector &y) const
-{
-   if (DeviceCanUseCeed())
-   {
-      ceedOp->AddMult(x, y);
-   }
-   else
-   {
-      internal::PADiffusionApply(dim, dofs1D, quad1D, ne, symmetric,
-                                 maps->B, maps->G, maps->Bt, maps->Gt,
-                                 pa_data, x, y);
-   }
-}
-
-void DiffusionIntegrator::AddMultTransposePA(const Vector &x, Vector &y) const
-{
-   if (symmetric)
-   {
-      AddMultPA(x, y);
-   }
-   else
-   {
-      MFEM_ABORT("DiffusionIntegrator::AddMultTransposePA only implemented in "
-                 "the symmetric case.")
-   }
-}
-
-} // namespace mfem