Use MFEM_ASSERT_KERNEL in DeviceTensor

Batched linear algebra with GPU backends

.gitignore

@@ -272,16 +272,27 @@ miniapps/navier/*_output
 
 miniapps/nurbs/nurbs_ex1
 miniapps/nurbs/nurbs_ex1p
+miniapps/nurbs/nurbs_ex3
+miniapps/nurbs/nurbs_ex5
 miniapps/nurbs/nurbs_ex11p
+miniapps/nurbs/nurbs_ex24
+miniapps/nurbs/nurbs_solenoidal
 miniapps/nurbs/nurbs_printfunc
 miniapps/nurbs/nurbs_patch_ex1
 miniapps/nurbs/nurbs_curveint
 miniapps/nurbs/refined.mesh
 miniapps/nurbs/mesh.*
+miniapps/nurbs/sol_?.gf
 miniapps/nurbs/sol.*
 miniapps/nurbs/mode_*
 miniapps/nurbs/Example1*
+miniapps/nurbs/Example3*
+miniapps/nurbs/Example5*
+miniapps/nurbs/Solenoidal*
+miniapps/nurbs/ParaView
 miniapps/nurbs/sin-fit.mesh
+miniapps/nurbs/ex5.mesh
+miniapps/nurbs/exsol.mesh
 miniapps/nurbs/CurveInt
 miniapps/nurbs/nurbs_naca_cmesh
 miniapps/nurbs/naca-cmesh.mesh

.gitlab-ci.yml

@@ -22,7 +22,7 @@ include:
 # the "needs" keyword and express the DAG of jobs for more efficiency.
 # - We use setup and setup_baseline phases to download content outside of mfem
 #   directory.
-# - Allocate/Release is where quartz resource are allocated/released once for all.
+# - Allocate/Release is where ruby resource are allocated/released once for all.
 # - Build and Test is where we build and MFEM for multiple toolchains.
 # - Baseline_checks gathers baseline-type test suites execution
 # - Baseline_publish, only available on master, allows to update baseline
@@ -53,7 +53,7 @@ variables:
   AUTOTEST_COMMIT: "YES"
 
 # Trigger subpipelines:
-quartz-build-and-test:
+ruby-build-and-test:
   stage: sub-pipelines
   variables:
     # Explicitly pass down values that we want to be able to set when triggering
@@ -61,10 +61,10 @@ quartz-build-and-test:
     AUTOTEST: "${AUTOTEST}"
     AUTOTEST_COMMIT: "${AUTOTEST_COMMIT}"
   trigger:
-    include: .gitlab/quartz-build-and-test.yml
+    include: .gitlab/ruby-build-and-test.yml
     strategy: depend
 
-quartz-baseline:
+ruby-baseline:
   stage: sub-pipelines
   variables:
     # Explicitly pass down values that we want to be able to set when triggering
@@ -73,7 +73,7 @@ quartz-baseline:
     AUTOTEST: "${AUTOTEST}"
     AUTOTEST_COMMIT: "${AUTOTEST_COMMIT}"
   trigger:
-    include: .gitlab/quartz-baseline.yml
+    include: .gitlab/ruby-baseline.yml
     strategy: depend
 
 lassen-build-and-test:

.gitlab/README.md

@@ -24,7 +24,7 @@ and `test type`.
 
 Machines typically include:
 
-* Quartz: Intel bi-socket x86
+* Ruby: 2nd Gen Intel Xeon (Cascade Lake)
 * Lassen: Power9 + Nvidia GPU
 * Corona: AMD GPU
 
@@ -76,13 +76,13 @@ with a spack spec of MFEM, within the limits permitted by the MFEM spack
 package.
 
 In any build-and-test sub-pipeline a job basically consists in defining the
-spack spec to use. Adding a job on quartz for example resumes to:
+spack spec to use. Adding a job on ruby for example resumes to:
 
 ```yaml
 <job_name>:
   variables:
     SPEC: "<spack_spec>"
-  extends: .build_and_test_on_quartz
+  extends: .build_and_test_on_ruby
 ```
 
 The remaining and non trivial work is to make sure this spec is working. To

.gitlab/configs/common.yml

@@ -24,7 +24,7 @@ variables:
 # TODO: add a clean-up mechanism
   BUILD_ROOT: ${USER_CI_TOP_DIR}/${CI_PROJECT_NAME}-${MACHINE_NAME}-pipeline-${CI_PIPELINE_ID}
 
-# On LLNL's quartz, there is only one allocation shared among jobs in order to
+# On LLNL's ruby, there is only one allocation shared among jobs in order to
 # save time and resource. This allocation has to be uniquely named so that we
 # are sure to retrieve it.
   ALLOC_NAME: ${CI_PROJECT_NAME}_ci_${CI_PIPELINE_ID}

.gitlab/configs/ruby-config.yml

@@ -9,17 +9,17 @@
 # terms of the BSD-3 license. We welcome feedback and contributions, see file
 # CONTRIBUTING.md for details.
 
-# GitLab pipelines configurations for the Quartz machine at LLNL
+# GitLab pipelines configurations for the Ruby machine at LLNL
 variables:
-  MACHINE_NAME: quartz
+  MACHINE_NAME: ruby
 
-.on_quartz:
+.on_ruby:
   tags:
     - shell
-    - quartz
+    - ruby
   rules:
-    # Don't run quartz jobs if...
-    - if: '$CI_COMMIT_BRANCH =~ /_qnone/ || $ON_QUARTZ == "OFF"'
+    # Don't run ruby jobs if...
+    - if: '$CI_COMMIT_BRANCH =~ /_qnone/ || $ON_RUBY == "OFF"'
       when: never
     # Don't run autotest update if...
     - if: '$CI_JOB_NAME =~ /report/ && $AUTOTEST != "YES"'
@@ -40,13 +40,13 @@ variables:
     - when: on_success
 
 # Spack helped builds
-# Generic quartz build job, extending build script
-.build_and_test_on_quartz:
-  extends: [.on_quartz]
+# Generic ruby build job, extending build script
+.build_and_test_on_ruby:
+  extends: [.on_ruby]
   stage: build_and_test
   script:
     # THREADS is used by 'tests/gitlab/build_and_test', run below
-    - export THREADS=12
+    - export THREADS=16
     - echo ${ALLOC_NAME}
     - export JOBID=$(squeue -h --name=${ALLOC_NAME} --format=%A)
     - echo ${JOBID}

.gitlab/configs/setup-baseline.yml

@@ -18,7 +18,7 @@
 setup_baseline:
   tags:
     - shell
-    - quartz
+    - ruby
   stage: setup
   variables:
     GIT_STRATEGY: none

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

@@ -16,7 +16,7 @@
 setup:
   tags:
     - shell
-    - quartz
+    - ruby
   stage: setup
   variables:
     GIT_STRATEGY: none

.gitlab/ruby-baseline.yml

@@ -19,8 +19,8 @@ stages:
   - cleanup
   - baseline_publish
 
-baselinecheck_mfem_intel_quartz:
-  extends: [.on_quartz]
+baselinecheck_mfem_intel_ruby:
+  extends: [.on_ruby]
   stage: baseline_check
   variables:
     # TPLS_DIR is used in .gitlab/scripts/baseline to provide the tpls location
@@ -32,7 +32,7 @@ baselinecheck_mfem_intel_quartz:
     - echo ${BUILD_ROOT}
     - echo ${TPLS_DIR}
     # Used by the tests in MFEM/tests:
-    - export MFEM_TEST_NP=32
+    - export MFEM_TEST_NP=48
     # The next script uses the following environment variables:
     # * BASELINE_TEST, SYS_TYPE, CI_PROJECT_DIR, ARTIFACTS_DIR,
     # * BUILD_ROOT, TPLS_DIR, MACHINE_NAME
@@ -44,18 +44,16 @@ baselinecheck_mfem_intel_quartz:
   allow_failure: true
 
 cleanup:
-  extends: .on_quartz
+  extends: .on_ruby
   stage: cleanup
   variables:
     GIT_STRATEGY: none
   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]
+  extends: [.on_ruby]
   stage: baseline_report
   script:
     - echo ${MACHINE_NAME}
@@ -115,8 +113,8 @@ report_baseline:
         exit $err
       ) 9> autotest.lock
 
-baselinepublish_mfem_quartz:
-  extends: [.on_quartz]
+baselinepublish_mfem_ruby:
+  extends: [.on_ruby]
   stage: baseline_publish
   rules:
     # - if: '$CI_COMMIT_BRANCH == "master" || $REBASELINE == "YES"'
@@ -131,5 +129,5 @@ baselinepublish_mfem_quartz:
 
 include:
   - local: .gitlab/configs/common.yml
-  - local: .gitlab/configs/quartz-config.yml
+  - local: .gitlab/configs/ruby-config.yml
   - local: .gitlab/configs/setup-baseline.yml

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

@@ -19,54 +19,54 @@ stages:
 allocate_resource:
   variables:
     GIT_STRATEGY: none
-  extends: .on_quartz
+  extends: .on_ruby
   stage: allocate_resource
   script:
     - echo ${ALLOC_NAME}
     - salloc --exclusive --nodes=1 --reservation=ci --time=60 --no-shell --job-name=${ALLOC_NAME}
   timeout: 6h
 
-# GitLab jobs for the Quartz machine at LLNL
+# GitLab jobs for the Ruby machine at LLNL
 debug_ser_gcc_10:
   variables:
     SPEC: "%gcc@10.3.1 +debug~mpi"
-  extends: .build_and_test_on_quartz
+  extends: .build_and_test_on_ruby
 
 debug_par_gcc_10:
   variables:
     SPEC: "%gcc@10.3.1 +debug+mpi"
-  extends: .build_and_test_on_quartz
+  extends: .build_and_test_on_ruby
 
 opt_ser_gcc_10:
   variables:
     SPEC: "%gcc@10.3.1 ~mpi"
-  extends: .build_and_test_on_quartz
+  extends: .build_and_test_on_ruby
 
 opt_par_gcc_10:
   variables:
     SPEC: "%gcc@10.3.1"
-  extends: .build_and_test_on_quartz
+  extends: .build_and_test_on_ruby
 
 opt_par_gcc_10_sundials:
   variables:
     SPEC: "%gcc@10.3.1 +sundials"
-  extends: .build_and_test_on_quartz
+  extends: .build_and_test_on_ruby
 
 opt_par_gcc_10_petsc:
   variables:
     SPEC: "%gcc@10.3.1 +petsc ^petsc+mumps~superlu-dist"
-  extends: .build_and_test_on_quartz
+  extends: .build_and_test_on_ruby
 
 opt_par_gcc_10_pumi:
   variables:
     SPEC: "%gcc@10.3.1 +pumi"
-  extends: .build_and_test_on_quartz
+  extends: .build_and_test_on_ruby
 
 # Release
 release_resource:
   variables:
     GIT_STRATEGY: none
-  extends: .on_quartz
+  extends: .on_ruby
   stage: release_resource_and_report
   script:
     - echo ${ALLOC_NAME}
@@ -78,17 +78,17 @@ release_resource:
 report_job_success:
   stage: release_resource_and_report
   extends:
-    - .on_quartz
+    - .on_ruby
     - .report_job_success
 
 report_job_failure:
   stage: release_resource_and_report
   extends:
-    - .on_quartz
+    - .on_ruby
     - .report_job_failure
 
 include:
   - local: .gitlab/configs/common.yml
-  - local: .gitlab/configs/quartz-config.yml
+  - local: .gitlab/configs/ruby-config.yml
   - local: .gitlab/configs/setup-build-and-test.yml
   - local: .gitlab/configs/report-build-and-test.yml

.gitlab/scripts/baseline

@@ -14,7 +14,7 @@
 # locals
 glob_err=${BASELINE_TEST}.err
 base=${BASELINE_TEST}-${SYS_TYPE}
-if [[ "${MACHINE_NAME}" == "quartz" ]]; then
+if [[ "${MACHINE_NAME}" == "ruby" ]]; then
   base="${BASELINE_TEST}-${MACHINE_NAME}"
 fi
 base_diff=${base}.diff
@@ -31,8 +31,8 @@ cd tests
 mkdir _${BASELINE_TEST} && cd _${BASELINE_TEST}
 
 # run
-if [[ "${MACHINE_NAME}" == "quartz" || "${MACHINE_NAME}" == "ruby" ]]; then
-  salloc --nodes=1 --reservation=ci ../runtest ../../mfem "${BASELINE_TEST} ${TPLS_DIR}"
+if [[ "${MACHINE_NAME}" == "ruby" ]]; then
+  salloc --nodes=1 --exclusive --reservation=ci ../runtest ../../mfem "${BASELINE_TEST} ${TPLS_DIR}"
 elif [[ ${MACHINE_NAME} == "corona" ]]; then
   salloc --nodes=1 -t 60 -p pbatch ../runtest ../../mfem "${BASELINE_TEST} ${TPLS_DIR}"
 elif [[ ${MACHINE_NAME} == "lassen" ]]; then
@@ -41,11 +41,11 @@ else
   echo "Unknown machine: MACHINE_NAME=$MACHINE_NAME"
   exit 1
 fi
+status="$?"
 
 # post
 mkdir ${artifacts_path}
 
-status=0
 if [[ -f ${BASELINE_TEST}.out ]]; then
   cp ${BASELINE_TEST}.out ${artifacts_path}
 fi

.gitlab/scripts/rebaseline

@@ -11,7 +11,7 @@
 # terms of the BSD-3 license. We welcome feedback and contributions, see file
 # CONTRIBUTING.md for details.
 
-# There will be collision between corona and quartz baselines.
+# There will be collision between corona and ruby baselines.
 # Once the corresponding files have been generated, we can switch to machine
 # specific ref.
 ARTIFACT_PATH=${CI_PROJECT_DIR}/${ARTIFACTS_DIR}/baseline-${SYS_TYPE}
@@ -21,7 +21,7 @@ PATCH_FILE=${ARTIFACT_PATH}.patch
 FULL_FILE=${ARTIFACT_PATH}.out
 DIFF_FILE=${ARTIFACT_PATH}.diff
 
-# There will be collision between corona and quartz baselines.
+# There will be collision between corona and ruby baselines.
 # Once the corresponding files have been generated, we can switch to machine
 # specific ref.
 SAVED_NAME=baseline-${SYS_TYPE}.saved

CHANGELOG

@@ -11,9 +11,48 @@
 Version 4.7.1 (development)
 ===========================
 
+Discretization improvements
+---------------------------
+- Added NURBS-based H(div) and H(curl) elements in 2D and 3D. Only on single
+  patch meshes. Only implemented for serial computations.
+
+- Added support for boundary constraints to the hybridization class.
+
+Meshing improvements
+--------------------
+- The ExodusII reader now handles pyramid and wedge element types. Mixed meshes
+  are also supported.
+
+New and updated examples and miniapps
+-------------------------------------
+- Added miniapps to demonstrate the H(div) and H(curl) NURBS elements.
+
 - Added an MFEM example for the eikonal equation. This new solver is based on
   the proximal Galerkin method introduced by Keith and Surowiec.
 
+GPU computing
+-------------
+- Added support for GPU-accelerated batched linear algebra (using cuBLAS,
+  hipBLAS, MAGMA, or native MFEM functionality) through the BatchedLinAlg class.
+
+Miscellaneous
+-------------
+- Refactored the `ARKStepSolver` class (ARKODE interface) to use
+  `TimeDependentOperator::Mult` only when the associated ODE operator is
+  expressed in explicit form (i.e., `TimeDependentOperator::isExplicit()`),
+  otherwise `TimeDependentOperator::ExplicitMult` is used. A check has been
+  added to `ARKStepSolver` to verify that the associated ODE operator is not in
+  explicit form when a mass matrix solver is enabled via a call to either the
+  `UseMFEMMassLinearSolver` or `UseSundialsMassLinearSolver` methods. This is
+  because enabling a mass matrix solver assumes that F(u,k,t) = M k in the
+  associated ODE operator.
+
+- Added support for custom interpolation procedure in FindPointsGSLIB.
+
+API changes
+-----------
+- API change: in class GridFunction, 'fec' was renamed to 'fec_owned'.
+
 
 Version 4.7, released on May 7, 2024
 ====================================
@@ -38,6 +77,9 @@ Meshing improvements
 
 - Added support for internal boundary elements in nonconforming meshes.
 
+- Added ExodusII output capability. The writer can handle first-order (Pyramid5,
+  Wedge6, Hex8, Tet4) and second-order FE types (Pyramid14, Wedge18, Hex27, Tet10).
+
 - The ReadCubit Genesis mesh importer has been rewritten to improve readability.
 
 Discretization improvements

CMakeLists.txt

@@ -146,7 +146,9 @@ if (MFEM_USE_CUDA)
    set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} ${CUDA_FLAGS}")
    find_package(CUDAToolkit REQUIRED)
    set(CUSPARSE_FOUND TRUE)
+   set(CUBLAS_FOUND TRUE)
    get_target_property(CUSPARSE_LIBRARIES CUDA::cusparse LOCATION)
+   get_target_property(CUBLAS_LIBRARIES CUDA::cublas LOCATION)
 endif()
 
 if (XSDK_ENABLE_C)
@@ -231,6 +233,7 @@ if (MFEM_USE_HIP)
     list(INSERT CMAKE_PREFIX_PATH 0 ${ROCM_PATH})
   endif()
   find_package(HIP REQUIRED)
+  find_package(HIPBLAS REQUIRED)
   find_package(HIPSPARSE REQUIRED)
 endif()
 
@@ -396,6 +399,10 @@ if (MFEM_USE_AMGX)
   find_package(AMGX REQUIRED)
 endif()
 
+if (MFEM_USE_MAGMA)
+  find_package(MAGMA REQUIRED)
+endif()
+
 if (MFEM_USE_CONDUIT)
     find_package(Conduit REQUIRED conduit relay blueprint)
 endif()
@@ -557,8 +564,9 @@ 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 CUSPARSE MKL_CPARDISO MKL_PARDISO AMGX CALIPER CODIPACK
-    BENCHMARK PARELAG TRIBOL MPI_CXX HIP HIPSPARSE MOONOLITH BLITZ ALGOIM ENZYME)
+    ADIOS2 MKL_CPARDISO MKL_PARDISO AMGX MAGMA CUSPARSE CUBLAS CALIPER CODIPACK
+    BENCHMARK PARELAG TRIBOL MPI_CXX HIP HIPBLAS HIPSPARSE MOONOLITH BLITZ
+    ALGOIM ENZYME)
 
 # Add all *_FOUND libraries in the variable TPL_LIBRARIES.
 set(TPL_LIBRARIES "")
@@ -673,7 +681,7 @@ if (NOT ("${PROJECT_SOURCE_DIR}" STREQUAL "${PROJECT_BINARY_DIR}"))
 #include \"${PROJECT_SOURCE_DIR}/${Header}\"
 ")
 
-    execute_process(COMMAND ${CMAKE_COMMAND} -E copy_if_different 
+    execute_process(COMMAND ${CMAKE_COMMAND} -E copy_if_different
       "${PROJECT_BINARY_DIR}/${Header}.tmp"
       "${PROJECT_BINARY_DIR}/${Header}"
     )
@@ -687,7 +695,7 @@ if (NOT ("${PROJECT_SOURCE_DIR}" STREQUAL "${PROJECT_BINARY_DIR}"))
 #include \"mfem/${Header}\"
 ")
 
-    execute_process(COMMAND ${CMAKE_COMMAND} -E copy_if_different 
+    execute_process(COMMAND ${CMAKE_COMMAND} -E copy_if_different
       "${PROJECT_BINARY_DIR}/InstallHeaders/${Header}.tmp"
       "${PROJECT_BINARY_DIR}/InstallHeaders/${Header}"
     )

INSTALL

@@ -273,7 +273,13 @@ Installation options:
    PREFIX  - Specify the installation directory. The library (libmfem.a) will be
              installed in $(PREFIX)/lib, the headers in $(PREFIX)/include, and
              the configuration makefile (config.mk) in $(PREFIX)/share/mfem.
-   INSTALL - Specify the install program, e.g /usr/bin/install
+   INSTALL - Specify the install program, default = /usr/bin/install
+   INSTALL_DEF_PERM - Specify the default install permissions. This affects
+                      headers and configuration makefiles, default = 644
+   INSTALL_BIN_PERM - Specify the install permissions for binaries. This only
+                      affects the shared version of the library, default = 755
+   INSTALL_DIR_PERM - Specify the install permissions for directories and,
+                      on macOS/BSD, for symlinks as well, default = 755
 
 MFEM library features/options (GNU make)
 ----------------------------------------
@@ -388,6 +394,11 @@ MFEM_USE_AMGX = YES/NO
    Allows the user to use SparseMatrices and HypreParMatrices to solve linear
    systems with the routines from the AmgX library.
 
+MFEM_USE_MAGMA = YES/NO
+   Enable MFEM functionality based on the MAGMA high-performance linear algebra
+   library. The MAGMA library provides a BLAS/LAPACK interface, with
+   implementations that have been optimized for Nvidia and AMD GPUs.
+
 MFEM_USE_GNUTLS = YES/NO
    Enable secure socket support in class socketstream, using the auxiliary
    GnuTLS_* classes, based on the GnuTLS library. This option may be useful in
@@ -699,6 +710,11 @@ The specific libraries and their options are:
   Options: AMGX_OPT, AMGX_LIB.
   Versions: AmgX >= 2.1, older versions may work too.
 
+- MAGMA (optional), used with MFEM_USE_MAGMA = YES.
+  URL: https://icl.utk.edu/magma/
+  Options: MAGMA_OPT, MAGMA_LIB
+  Versions: MAGMA >= 2.8.0
+
 - GnuTLS (optional), used when MFEM_USE_GNUTLS = YES. On most Linux systems,
   GnuTLS is available as a development package, e.g. gnutls-devel. On Mac OS X,
   one can get the library through the Homebrew package manager (http://brew.sh).

config/cmake/MFEMConfig.cmake.in

@@ -37,6 +37,7 @@ set(MFEM_USE_MUMPS @MFEM_USE_MUMPS@)
 set(MFEM_USE_STRUMPACK @MFEM_USE_STRUMPACK@)
 set(MFEM_USE_GINKGO @MFEM_USE_GINKGO@)
 set(MFEM_USE_AMGX @MFEM_USE_AMGX@)
+set(MFEM_USE_MAGMA @MFEM_USE_MAGMA@)
 set(MFEM_USE_HIOP @MFEM_USE_HIOP@)
 set(MFEM_USE_GNUTLS @MFEM_USE_GNUTLS@)
 set(MFEM_USE_GSLIB @MFEM_USE_GSLIB@)

config/cmake/config.hpp.in

@@ -114,6 +114,9 @@
 // Enable MFEM functionality based on the AmgX library.
 #cmakedefine MFEM_USE_AMGX
 
+// Enable MFEM functionality based on the MAGMA library.
+#cmakedefine MFEM_USE_MAGMA
+
 // Enable secure socket streams based on the GNUTLS library.
 #cmakedefine MFEM_USE_GNUTLS
 

config/cmake/modules/FindMAGMA.cmake

@@ -0,0 +1,37 @@
+# Copyright (c) 2010-2024, 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:
+#   - MAGMA_FOUND
+#   - MAGMA_LIBRARIES
+#   - MAGMA_INCLUDE_DIRS
+
+include(MfemCmakeUtilities)
+mfem_find_package(MAGMA MAGMA MAGMA_DIR "include" "magma.h" "lib" "magma"
+  "Paths to headers required by MAGMA." "Libraries required by MAGMA.")
+
+if (MAGMA_FOUND AND MFEM_USE_CUDA)
+  get_target_property(CUSPARSE_LIBRARIES CUDA::cusparse LOCATION)
+  get_target_property(CUBLAS_LIBRARIES CUDA::cublas LOCATION)
+  list(APPEND MAGMA_LIBRARIES ${CUSPARSE_LIBRARIES} ${CUBLAS_LIBRARIES})
+  set(MAGMA_LIBRARIES ${MAGMA_LIBRARIES} CACHE STRING
+      "MAGMA libraries + dependencies." FORCE)
+  message(STATUS "Updated MAGMA_LIBRARIES: ${MAGMA_LIBRARIES}")
+endif()
+
+if (MAGMA_FOUND AND MFEM_USE_HIP)
+  find_package(HIPBLAS REQUIRED)
+  find_package(HIPSPARSE REQUIRED)
+  list(APPEND MAGMA_LIBRARIES ${HIPBLAS_LIBRARIES} ${HIPSPARSE_LIBRARIES})
+  set(MAGMA_LIBRARIES ${MAGMA_LIBRARIES} CACHE STRING
+      "MAGMA libraries + dependencies." FORCE)
+  message(STATUS "Updated MAGMA_LIBRARIES: ${MAGMA_LIBRARIES}")
+endif()

config/cmake/modules/MfemCmakeUtilities.cmake

@@ -846,14 +846,14 @@ function(mfem_export_mk_files)
       MFEM_USE_ZLIB MFEM_USE_LIBUNWIND MFEM_USE_LAPACK MFEM_THREAD_SAFE
       MFEM_USE_LEGACY_OPENMP MFEM_USE_OPENMP MFEM_USE_MEMALLOC MFEM_USE_SUNDIALS
       MFEM_USE_SUITESPARSE MFEM_USE_SUPERLU MFEM_USE_SUPERLU5 MFEM_USE_MUMPS
-      MFEM_USE_STRUMPACK MFEM_USE_GINKGO MFEM_USE_AMGX MFEM_USE_GNUTLS
-      MFEM_USE_NETCDF MFEM_USE_PETSC MFEM_USE_SLEPC MFEM_USE_MPFR MFEM_USE_SIDRE
-      MFEM_USE_FMS MFEM_USE_CONDUIT MFEM_USE_PUMI MFEM_USE_HIOP MFEM_USE_GSLIB
-      MFEM_USE_CUDA MFEM_USE_HIP MFEM_USE_RAJA MFEM_USE_OCCA MFEM_USE_CEED
-      MFEM_USE_CALIPER MFEM_USE_UMPIRE MFEM_USE_SIMD MFEM_USE_ADIOS2
-      MFEM_USE_MKL_CPARDISO MFEM_USE_MKL_PARDISO MFEM_USE_ADFORWARD
-      MFEM_USE_CODIPACK MFEM_USE_BENCHMARK MFEM_USE_PARELAG MFEM_USE_TRIBOL
-      MFEM_USE_MOONOLITH MFEM_USE_ALGOIM MFEM_USE_ENZYME)
+      MFEM_USE_STRUMPACK MFEM_USE_GINKGO MFEM_USE_AMGX MFEM_USE_MAGMA
+      MFEM_USE_GNUTLS MFEM_USE_NETCDF MFEM_USE_PETSC MFEM_USE_SLEPC
+      MFEM_USE_MPFR MFEM_USE_SIDRE MFEM_USE_FMS MFEM_USE_CONDUIT MFEM_USE_PUMI
+      MFEM_USE_HIOP MFEM_USE_GSLIB MFEM_USE_CUDA MFEM_USE_HIP MFEM_USE_RAJA
+      MFEM_USE_OCCA MFEM_USE_CEED MFEM_USE_CALIPER MFEM_USE_UMPIRE MFEM_USE_SIMD
+      MFEM_USE_ADIOS2 MFEM_USE_MKL_CPARDISO MFEM_USE_MKL_PARDISO
+      MFEM_USE_ADFORWARD MFEM_USE_CODIPACK MFEM_USE_BENCHMARK MFEM_USE_PARELAG
+      MFEM_USE_TRIBOL MFEM_USE_MOONOLITH MFEM_USE_ALGOIM MFEM_USE_ENZYME)
   foreach(var ${CONFIG_MK_BOOL_VARS})
     if (${var})
       set(${var} YES)

config/config.hpp.in

@@ -114,6 +114,9 @@
 // Enable MFEM functionality based on the AmgX library.
 // #define MFEM_USE_AMGX
 
+// Enable MFEM functionality based on the MAGMA library.
+// #define MFEM_USE_MAGMA
+
 // Enable secure socket streams based on the GNUTLS library.
 // #define MFEM_USE_GNUTLS
 

config/config.mk.in

@@ -38,6 +38,7 @@ MFEM_USE_MUMPS         = @MFEM_USE_MUMPS@
 MFEM_USE_STRUMPACK     = @MFEM_USE_STRUMPACK@
 MFEM_USE_GINKGO        = @MFEM_USE_GINKGO@
 MFEM_USE_AMGX          = @MFEM_USE_AMGX@
+MFEM_USE_MAGMA         = @MFEM_USE_MAGMA@
 MFEM_USE_GNUTLS        = @MFEM_USE_GNUTLS@
 MFEM_USE_NETCDF        = @MFEM_USE_NETCDF@
 MFEM_USE_PETSC         = @MFEM_USE_PETSC@

config/defaults.cmake

@@ -40,6 +40,7 @@ option(MFEM_USE_MUMPS "Enable MUMPS usage" OFF)
 option(MFEM_USE_STRUMPACK "Enable STRUMPACK usage" OFF)
 option(MFEM_USE_GINKGO "Enable Ginkgo usage" OFF)
 option(MFEM_USE_AMGX "Enable AmgX usage" OFF)
+option(MFEM_USE_MAGMA "Enable MAGMA usage" OFF)
 option(MFEM_USE_GNUTLS "Enable GNUTLS usage" OFF)
 option(MFEM_USE_GSLIB "Enable GSLIB usage" OFF)
 option(MFEM_USE_NETCDF "Enable NETCDF usage" OFF)
@@ -183,6 +184,10 @@ set(Ginkgo_DIR "${MFEM_DIR}/../ginkgo" CACHE PATH "Path to the Ginkgo library.")
 
 set(AMGX_DIR "${MFEM_DIR}/../amgx" CACHE PATH "Path to AmgX")
 
+set(MAGMA_DIR "${MFEM_DIR}/../magma" CACHE PATH "Path to MAGMA")
+set(MAGMA_REQUIRED_PACKAGES "BLAS" "LAPACK" CACHE STRING
+    "Additional packages required by MAGMA.")
+
 set(GNUTLS_DIR "" CACHE PATH "Path to the GnuTLS library.")
 
 set(GSLIB_DIR "" CACHE PATH "Path to the GSLIB library.")
@@ -259,7 +264,7 @@ set(PARELAG_LIBRARIES "${PARELAG_DIR}/build/src/libParELAG.a" CACHE STRING
     "The ParELAG library.")
 
 set(TRIBOL_DIR "${MFEM_DIR}/../tribol" CACHE PATH "Path to Tribol")
-set(Tribol_REQUIRED_PACKAGES "Axom/core/mint/slam/slic" CACHE STRING 
+set(Tribol_REQUIRED_PACKAGES "Axom/core/mint/slam/slic" CACHE STRING
     "Additional packages required by Tribol")
 
 set(BLAS_INCLUDE_DIRS "" CACHE STRING "Path to BLAS headers.")

config/defaults.mk

@@ -95,6 +95,10 @@ else
    # Silence unused command line argument warnings when generating dependencies
    # with mpicxx and clang
    DEP_FLAGS := -Wno-unused-command-line-argument $(DEP_FLAGS)
+   # Silence "ignoring duplicate libraries" warnings on new (Xcode 15) linker
+   ifneq (,$(findstring PROJECT:dyld,$(shell ld -v 2>&1)))
+      LDFLAGS_INTERNAL = -Xlinker -no_warn_duplicate_libraries
+   endif
 endif
 
 # Set CXXFLAGS to overwrite the default selection of DEBUG_FLAGS/OPTIM_FLAGS
@@ -139,6 +143,7 @@ MFEM_USE_MUMPS         = NO
 MFEM_USE_STRUMPACK     = NO
 MFEM_USE_GINKGO        = NO
 MFEM_USE_AMGX          = NO
+MFEM_USE_MAGMA         = NO
 MFEM_USE_GNUTLS        = NO
 MFEM_USE_NETCDF        = NO
 MFEM_USE_PETSC         = NO
@@ -390,6 +395,11 @@ AMGX_DIR = @MFEM_DIR@/../amgx
 AMGX_OPT = -I$(AMGX_DIR)/include
 AMGX_LIB = -L$(AMGX_DIR)/lib -lamgx -lcusparse -lcusolver -lcublas -lnvToolsExt
 
+# MAGMA library configuration
+MAGMA_DIR = @MFEM_DIR@/../magma
+MAGMA_OPT = -I$(MAGMA_DIR)/include
+MAGMA_LIB = -L$(MAGMA_DIR)/lib -l:libmagma.a -lcublas -lcusparse $(LAPACK_LIB)
+
 # GnuTLS library configuration
 GNUTLS_OPT =
 GNUTLS_LIB = -lgnutls
@@ -497,11 +507,11 @@ GSLIB_LIB = -L$(GSLIB_DIR)/lib -lgs
 
 # CUDA library configuration
 CUDA_OPT =
-CUDA_LIB = -lcusparse
+CUDA_LIB = -lcusparse -lcublas
 
 # HIP library configuration
 HIP_OPT =
-HIP_LIB = -L$(HIP_DIR)/lib $(XLINKER)-rpath,$(HIP_DIR)/lib -lhipsparse
+HIP_LIB = -L$(HIP_DIR)/lib $(XLINKER)-rpath,$(HIP_DIR)/lib -lhipsparse -lhipblas
 
 # OCCA library configuration
 OCCA_DIR = @MFEM_DIR@/../occa

config/sample-runs.sh

@@ -32,7 +32,7 @@ groups_serial=(
 '"examples"
    "Examples:"
    "examples"
-   "ex{,1,2,3}[0-9].cpp"'
+   "ex{,[1-9]}[0-9].cpp"'
 #   "ex1.cpp"'
 '"sundials"
    "SUNDIALS examples:"
@@ -58,6 +58,10 @@ groups_serial=(
    "HiOp examples:"
    "examples/hiop"
    "ex9.cpp"'
+'"moonolith"
+   "Moonolith examples:"
+   "examples/moonolith"
+   "ex1.cpp"'
 '"pumi"
    "PUMI examples:"
    "examples/pumi"
@@ -66,25 +70,38 @@ groups_serial=(
 '"meshing"
    "Meshing miniapps:"
    "miniapps/meshing"
-   "mobius-strip.cpp klein-bottle.cpp extruder.cpp toroid.cpp
+   "mobius-strip.cpp klein-bottle.cpp extruder.cpp toroid.cpp mesh-quality.cpp
+    polar-nc.cpp reflector.cpp shaper.cpp trimmer.cpp twist.cpp
     mesh-optimizer.cpp minimal-surface.cpp"'
 '"adjoint"
    "Adjoint miniapps:"
    "miniapps/adjoint"
    "cvsRoberts_ASAi_dns.cpp"'
+'"autodiff"
+   "Autodiff miniapps:"
+   "miniapps/autodiff"
+   "seq_example.cpp seq_test.cpp"' # 'seq_test.cpp' has no sample runs
+'"dpg"
+   "DPG miniapps:"
+   "miniapps/dpg"
+   "{acoustics,convection-diffusion,diffusion,maxwell}.cpp"'
 '"gslib"
    "GSLIB miniapps:"
    "miniapps/gslib"
    "field-diff.cpp field-interp.cpp findpts.cpp schwarz_ex1.cpp "'
+# todo: miniapps/mtop
 '"nurbs"
    "NURBS miniapps:"
    "miniapps/nurbs"
    "nurbs_ex1.cpp"'
+# todo: add other nurbs miniapps
+# todo: miniapps/solvers (serial)
 '"tools"
    "Tools miniapps:"
    "miniapps/tools"
    "convert-dc.cpp display-basis.cpp get-values.cpp load-dc.cpp
     lor-transfer.cpp"'
+# todo: add other tools miniapps
 '"toys"
    "Toys miniapps:"
    "miniapps/toys"
@@ -100,7 +117,7 @@ groups_parallel=(
 '"examples"
    "Examples:"
    "examples"
-   "ex{,1,2,3}[0-9]p.cpp"'
+   "ex{,[1-9]}[0-9]p.cpp"'
 #   "ex1p.cpp"'
 '"sundials"
    "SUNDIALS examples:"
@@ -126,6 +143,10 @@ groups_parallel=(
    "HiOp examples:"
    "examples/hiop"
    "ex9p.cpp"'
+'"moonolith"
+   "Moonolith examples:"
+   "examples/moonolith"
+   "ex{1,2}p.cpp"'
 '"pumi"
    "PUMI examples:"
    "examples/pumi"
@@ -138,24 +159,41 @@ groups_parallel=(
 '"meshing"
    "Meshing miniapps:"
    "miniapps/meshing"
-   "pmesh-optimizer.cpp pmesh-fitting.cpp pminimal-surface.cpp"'
+   "pmesh-optimizer.cpp pmesh-fitting.cpp pminimal-surface.cpp
+    fit-node-position.cpp"'
 '"electromagnetics"
    "Electromagnetics miniapps:"
    "miniapps/electromagnetics"
    "joule.cpp"'
-#   "{volta,tesla,joule}.cpp"' # todo: multiline sample runs
+#   "{joule,maxwell,tesla,volta}.cpp"' # todo: multiline sample runs
 '"adjoint"
    "Adjoint miniapps:"
    "miniapps/adjoint"
    "adjoint_advection_diffusion.cpp"'
+'"autodiff"
+   "Autodiff miniapps:"
+   "miniapps/autodiff"
+   "par_example.cpp"'
+'"dpg"
+   "DPG miniapps:"
+   "miniapps/dpg"
+   "p{acoustics,convection-diffusion,diffusion,maxwell}.cpp"'
 '"gslib"
    "GSLIB miniapps:"
    "miniapps/gslib"
    "pfindpts.cpp schwarz_ex1p.cpp"'
+'"hdiv-linear-solver"
+   "H(div) linear solver miniapps:"
+   "miniapps/hdiv-linear-solver"
+   "grad_div.cpp darcy.cpp"'
+# 'miniapps/hooke/hooke.cpp' has no sample runs
+# todo: miniapps/mtop
+# todo: miniapps/multidomain
 '"navier"
    "Navier miniapps:"
    "miniapps/navier"
    "navier_cht.cpp"'
+# todo: add other navier miniapps
 '"nurbs"
    "NURBS miniapps:"
    "miniapps/nurbs"
@@ -164,14 +202,18 @@ groups_parallel=(
    "Shifted miniapps:"
    "miniapps/shifted"
    "distance.cpp"'
+# todo: add other shifted miniapps
 '"solvers"
    "Solvers miniapps:"
    "miniapps/solvers"
    "block-solvers.cpp"'
+# todo: add other solvers miniapps
+# todo: miniapps/spde
 '"tools"
    "Tools miniapps:"
    "miniapps/tools"
-   "convert-cd.cpp get-values.cpp load-dc.cpp"'
+   "convert-dc.cpp get-values.cpp load-dc.cpp"'
+# todo: add other tools miniapps
 '"convergence"
    "Convergence tests:"
    "tests/convergence"
@@ -186,7 +228,7 @@ groups_all=(
 '"examples"
    "Examples:"
    "examples"
-   "ex\"{,1,2,3}[0-9]\"{,p}.cpp"'
+   "ex\"{,[1-9]}[0-9]\"{,p}.cpp"'
 '"sundials"
    "SUNDIALS examples:"
    "examples/sundials"
@@ -215,10 +257,14 @@ groups_all=(
    "HiOp examples:"
    "examples/hiop"
    "ex9.cpp ex9p.cpp"'
+'"moonolith"
+   "Moonolith examples:"
+   "examples/moonolith"
+   "ex1.cpp ex{1,2}p.cpp"'
 '"pumi"
    "PUMI examples:"
    "examples/pumi"
-   "ex1.cpp ex1p.cpp ex2.cpp ex6p.cpp"'
+   "ex1.cpp ex2.cpp ex1p.cpp ex6p.cpp"'
 '"superlu"
    "Superlu examples:"
    "examples/superlu"
@@ -226,43 +272,67 @@ groups_all=(
 '"meshing"
    "Meshing miniapps:"
    "miniapps/meshing"
-   "mobius-strip.cpp klein-bottle.cpp extruder.cpp toroid.cpp
-    {,p}mesh-optimizer.cpp pmesh-fitting.cpp {,p}minimal-surface.cpp"'
+   "mobius-strip.cpp klein-bottle.cpp extruder.cpp toroid.cpp mesh-quality.cpp
+    polar-nc.cpp reflector.cpp shaper.cpp trimmer.cpp twist.cpp
+    {,p}mesh-optimizer.cpp pmesh-fitting.cpp {,p}minimal-surface.cpp
+    fit-node-position.cpp"'
 '"electromagnetics"
    "Electromagnetics miniapps:"
    "miniapps/electromagnetics"
    "joule.cpp"'
-#   "{volta,tesla,joule}.cpp"' # todo: multiline sample runs
+#   "{joule,maxwell,tesla,volta}.cpp"' # todo: multiline sample runs
 '"adjoint"
    "Adjoint miniapps:"
    "miniapps/adjoint"
-   "adjoint_advection_diffusion.cpp cvsRoberts_ASAi_dns.cpp"'
+   "cvsRoberts_ASAi_dns.cpp adjoint_advection_diffusion.cpp"'
+'"autodiff"
+   "Autodiff miniapps:"
+   "miniapps/autodiff"
+   "seq_example.cpp seq_test.cpp par_example.cpp"'
+# 'seq_test.cpp' has no sample runs
+'"dpg"
+   "DPG miniapps:"
+   "miniapps/dpg"
+   "{,p}{acoustics,convection-diffusion,diffusion,maxwell}.cpp"'
 '"gslib"
    "GSLIB miniapps:"
    "miniapps/gslib"
    "field-diff.cpp field-interp.cpp findpts.cpp schwarz_ex1.cpp pfindpts.cpp
     schwarz_ex1p.cpp"'
+'"hdiv-linear-solver"
+   "H(div) linear solver miniapps:"
+   "miniapps/hdiv-linear-solver"
+   "grad_div.cpp darcy.cpp"'
+# 'miniapps/hooke/hooke.cpp' has no sample runs
+# todo: miniapps/mtop
+# todo: miniapps/multidomain
 '"navier"
    "Navier miniapps:"
    "miniapps/navier"
    "navier_cht.cpp"'
+# todo: add other navier miniapps
 '"nurbs"
    "NURBS miniapps:"
    "miniapps/nurbs"
    "nurbs_ex1.cpp nurbs_ex1p.cpp nurbs_ex11p.cpp"'
+# todo: add other nurbs miniapps
 '"shifted"
    "Shifted miniapps:"
    "miniapps/shifted"
    "distance.cpp"'
+# todo: add other shifted miniapps
 '"solvers"
    "Solvers miniapps:"
    "miniapps/solvers"
    "block-solvers.cpp"'
+# todo: add other solvers miniapps
+# todo: miniapps/spde
 '"tools"
    "Tools miniapps:"
    "miniapps/tools"
    "convert-dc.cpp display-basis.cpp get-values.cpp load-dc.cpp
     lor-transfer.cpp"'
+# todo: add other tools miniapps
 '"toys"
    "Toys miniapps:"
    "miniapps/toys"
@@ -386,7 +456,7 @@ function help_message()
       mfem_config [${mfem_config}]
          Set MFEM configuration options
       make [${make}], mpiexec [${mpiexec}], mpiexec_np [${mpiexec_np}]
-         Their values can also set using the respective uppercase environment
+         Their values can also be set using the respective uppercase environment
          variable
       mfem_build_dir [${mfem_build_dir}]
          Same as '-d': set this variable to something different from <mfem_dir>

data/pipe-nurbs-2d.mesh

@@ -18,9 +18,9 @@ elements
 boundary
 4
 1 1 0 1
-1 1 2 3
-1 1 3 0
-1 1 1 2
+2 1 2 3
+3 1 3 0
+4 1 1 2
 
 edges
 4

doc/CodeDocumentation.conf.in

@@ -938,6 +938,7 @@ INPUT                  = @MFEM_SOURCE_DIR@/doc/CodeDocumentation.dox \
                          @MFEM_SOURCE_DIR@/config \
                          @MFEM_SOURCE_DIR@/general \
                          @MFEM_SOURCE_DIR@/linalg \
+                         @MFEM_SOURCE_DIR@/linalg/batched \
                          @MFEM_SOURCE_DIR@/linalg/simd \
                          @MFEM_SOURCE_DIR@/mesh \
                          @MFEM_SOURCE_DIR@/mesh/submesh \
@@ -1049,7 +1050,8 @@ RECURSIVE              = NO
 EXCLUDE                = @MFEM_SOURCE_DIR@/config/_config.hpp \
                          @MFEM_SOURCE_DIR@/config/get_hypre_version.cpp \
                          @MFEM_SOURCE_DIR@/general/tinyxml2.h \
-                         @MFEM_SOURCE_DIR@/general/tinyxml2.cpp
+                         @MFEM_SOURCE_DIR@/general/tinyxml2.cpp \
+                         @MFEM_SOURCE_DIR@/linalg/lapack.hpp
 
 # The EXCLUDE_SYMLINKS tag can be used to select whether or not files or
 # directories that are symbolic links (a Unix file system feature) are excluded

doc/CodeDocumentation.dox

@@ -182,6 +182,21 @@ namespace mfem {
  *   <a class="el" href="examples_2superlu_2ex1p_8cpp_source.html">1p</a>,
  *   demonstrating the use of MFEM's \link superlu.hpp SuperLU integration\endlink.
  *
+ * <H4>NURBS Examples</H4>
+ * - Variants of Examples
+ *  <a class="el" href="nurbs__ex1_8cpp_source.html">1</a>,
+ *  <a class="el" href="nurbs__ex1p_8cpp_source.html">1p</a>,
+ *  <a class="el" href="nurbs__ex3_8cpp_source.html">3</a>,
+ *  <a class="el" href="nurbs__ex5_8cpp_source.html">5</a>,
+ *  <a class="el" href="nurbs__ex11p_8cpp_source.html">11p</a>, and
+ *  <a class="el" href="nurbs__ex24_8cpp_source.html">24</a>,
+ *   demonstrating howto perform NURBS-based Isogeometric Analysis.
+ * - Variant of Example <a class="el" href="nurbs__patch__ex1_8cpp_source.html">1</a>: demonstrates the use of patch integration
+ * - <a class="el" href="nurbs__solenoidal_8cpp_source.html">NURBS Divergence-free</a>: solve a solenoidal vector projection with NURBS-based H(div) elements
+ * - <a class="el" href="nurbs__curveint_8cpp_source.html">NURBS Interpolation</a>: NURBS interpolation of given geometry
+ * - <a class="el" href="nurbs__naca__cmesh_8cpp_source.html">NURBS NACA Mesher</a>: generate NURBS based mesh around a NACA foil
+ * - <a class="el" href="nurbs__printfunc_8cpp_source.html">NURBS Printer</a>: print the NURBS-basis
+ *
  * <H3>Miniapps</H3>
  * - <a class="el" href="volta_8cpp_source.html">Volta</a>: simple electrostatics simulation code
  * - <a class="el" href="tesla_8cpp_source.html">Tesla</a>: simple magnetostatics simulation code

examples/amgx/makefile

@@ -12,11 +12,10 @@
 # Use the MFEM build directory
 MFEM_DIR ?= ../..
 MFEM_BUILD_DIR ?= ../..
+MFEM_INSTALL_DIR ?= ../../mfem
 SRC = $(if $(MFEM_DIR:../..=),$(MFEM_DIR)/examples/amgx/,)
-CONFIG_MK = $(MFEM_BUILD_DIR)/config/config.mk
-# Use the MFEM install directory
-# MFEM_INSTALL_DIR = ../../mfem
-# CONFIG_MK = $(MFEM_INSTALL_DIR)/share/mfem/config.mk
+CONFIG_MK = $(or $(wildcard $(MFEM_BUILD_DIR)/config/config.mk),\
+   $(wildcard $(MFEM_INSTALL_DIR)/share/mfem/config.mk))
 
 MFEM_LIB_FILE = mfem_is_not_built
 -include $(CONFIG_MK)

examples/caliper/makefile

@@ -12,11 +12,10 @@
 # Use the MFEM build directory
 MFEM_DIR ?= ../..
 MFEM_BUILD_DIR ?= ../..
+MFEM_INSTALL_DIR ?= ../../mfem
 SRC = $(if $(MFEM_DIR:../..=),$(MFEM_DIR)/examples/caliper,)
-CONFIG_MK = $(MFEM_BUILD_DIR)/config/config.mk
-# Use the MFEM install directory
-# MFEM_INSTALL_DIR = ../../mfem
-# CONFIG_MK = $(MFEM_INSTALL_DIR)/share/mfem/config.mk
+CONFIG_MK = $(or $(wildcard $(MFEM_BUILD_DIR)/config/config.mk),\
+   $(wildcard $(MFEM_INSTALL_DIR)/share/mfem/config.mk))
 
 MFEM_LIB_FILE = mfem_is_not_built
 -include $(CONFIG_MK)

examples/ex23.cpp

@@ -44,7 +44,7 @@ protected:
    BilinearForm *M;
    BilinearForm *K;
 
-   SparseMatrix Mmat, Kmat, Kmat0;
+   SparseMatrix Mmat, Kmat;
    SparseMatrix *T; // T = M + dt K
    real_t current_dt;
 
@@ -83,25 +83,24 @@ WaveOperator::WaveOperator(FiniteElementSpace &f,
    : SecondOrderTimeDependentOperator(f.GetTrueVSize(), (real_t) 0.0),
      fespace(f), M(NULL), K(NULL), T(NULL), current_dt(0.0), z(height)
 {
-   const real_t rel_tol = 1e-8;
-
-   fespace.GetEssentialTrueDofs(ess_bdr, ess_tdof_list);
-
+   // Assemble Laplace matrix
    c2 = new ConstantCoefficient(speed*speed);
-
    K = new BilinearForm(&fespace);
    K->AddDomainIntegrator(new DiffusionIntegrator(*c2));
    K->Assemble();
 
-   Array<int> dummy;
-   K->FormSystemMatrix(dummy, Kmat0);
-   K->FormSystemMatrix(ess_tdof_list, Kmat);
-
+   // Assemble Mass matrix
    M = new BilinearForm(&fespace);
    M->AddDomainIntegrator(new MassIntegrator());
    M->Assemble();
+
+   // Apply Bcs
+   fespace.GetEssentialTrueDofs(ess_bdr, ess_tdof_list);
+   K->FormSystemMatrix(ess_tdof_list, Kmat);
    M->FormSystemMatrix(ess_tdof_list, Mmat);
 
+   // Configure preconditioner
+   const real_t rel_tol = 1e-8;
    M_solver.iterative_mode = false;
    M_solver.SetRelTol(rel_tol);
    M_solver.SetAbsTol(0.0);
@@ -110,14 +109,13 @@ WaveOperator::WaveOperator(FiniteElementSpace &f,
    M_solver.SetPreconditioner(M_prec);
    M_solver.SetOperator(Mmat);
 
+   // Configure solver
    T_solver.iterative_mode = false;
    T_solver.SetRelTol(rel_tol);
    T_solver.SetAbsTol(0.0);
    T_solver.SetMaxIter(100);
    T_solver.SetPrintLevel(0);
    T_solver.SetPreconditioner(T_prec);
-
-   T = NULL;
 }
 
 void WaveOperator::Mult(const Vector &u, const Vector &du_dt,
@@ -126,9 +124,11 @@ void WaveOperator::Mult(const Vector &u, const Vector &du_dt,
    // Compute:
    //    d2udt2 = M^{-1}*-K(u)
    // for d2udt2
-   Kmat.Mult(u, z);
+   K->FullMult(u, z);
    z.Neg(); // z = -z
+   z.SetSubVector(ess_tdof_list, 0.0);
    M_solver.Mult(z, d2udt2);
+   d2udt2.SetSubVector(ess_tdof_list, 0.0);
 }
 
 void WaveOperator::ImplicitSolve(const real_t fac0, const real_t fac1,
@@ -142,14 +142,11 @@ void WaveOperator::ImplicitSolve(const real_t fac0, const real_t fac1,
       T = Add(1.0, Mmat, fac0, Kmat);
       T_solver.SetOperator(*T);
    }
-   Kmat0.Mult(u, z);
+   K->FullMult(u, z);
    z.Neg();
-
-   for (int i = 0; i < ess_tdof_list.Size(); i++)
-   {
-      z[ess_tdof_list[i]] = 0.0;
-   }
+   z.SetSubVector(ess_tdof_list, 0.0);
    T_solver.Mult(z, d2udt2);
+   d2udt2.SetSubVector(ess_tdof_list, 0.0);
 }
 
 void WaveOperator::SetParameters(const Vector &u)
@@ -314,7 +311,6 @@ int main(int argc, char *argv[])
          ess_bdr = 0;
       }
    }
-
    WaveOperator oper(fespace, ess_bdr, speed);
 
    u_gf.SetFromTrueDofs(u);

examples/ex40.cpp

@@ -67,6 +67,8 @@ public:
    ZCoefficient(int vdim, GridFunction &psi_, real_t alpha_ = 1.0)
       : VectorCoefficient(vdim), psi(&psi_), alpha(alpha_) { }
 
+   using VectorCoefficient::Eval;
+
    virtual void Eval(Vector &V, ElementTransformation &T,
                      const IntegrationPoint &ip);
    void SetAlpha(real_t alpha_) { alpha = alpha_; }

examples/ex40p.cpp

@@ -67,6 +67,8 @@ public:
    ZCoefficient(int vdim, ParGridFunction &psi_, real_t alpha_ = 1.0)
       : VectorCoefficient(vdim), psi(&psi_), alpha(alpha_) { }
 
+   using VectorCoefficient::Eval;
+
    virtual void Eval(Vector &V, ElementTransformation &T,
                      const IntegrationPoint &ip);
    void SetAlpha(real_t alpha_) { alpha = alpha_; }

examples/ginkgo/makefile

@@ -12,11 +12,10 @@
 # Use the MFEM build directory
 MFEM_DIR ?= ../..
 MFEM_BUILD_DIR ?= ../..
+MFEM_INSTALL_DIR ?= ../../mfem
 SRC = $(if $(MFEM_DIR:../..=),$(MFEM_DIR)/examples/ginkgo/,)
-CONFIG_MK = $(MFEM_BUILD_DIR)/config/config.mk
-# Use the MFEM install directory
-# MFEM_INSTALL_DIR = ../../mfem
-# CONFIG_MK = $(MFEM_INSTALL_DIR)/share/mfem/config.mk
+CONFIG_MK = $(or $(wildcard $(MFEM_BUILD_DIR)/config/config.mk),\
+   $(wildcard $(MFEM_INSTALL_DIR)/share/mfem/config.mk))
 
 MFEM_LIB_FILE = mfem_is_not_built
 -include $(CONFIG_MK)

examples/hiop/ex9p.cpp

@@ -96,6 +96,7 @@ public:
    {
       Vector w_glob(width);
       pfes.Dof_TrueDof_Matrix()->MultTranspose(w, w_glob);
+      w_glob.HostReadWrite(); // read+write -> can use w_glob(i) (non-const)
       for (int i = 0; i < width; i++) { grad(0, i) = w_glob(i); }
    }
 

examples/hiop/makefile

@@ -12,11 +12,10 @@
 # Use the MFEM build directory
 MFEM_DIR ?= ../..
 MFEM_BUILD_DIR ?= ../..
+MFEM_INSTALL_DIR ?= ../../mfem
 SRC = $(if $(MFEM_DIR:../..=),$(MFEM_DIR)/examples/hiop/,)
-CONFIG_MK = $(MFEM_BUILD_DIR)/config/config.mk
-# Use the MFEM install directory
-# MFEM_INSTALL_DIR = ../../mfem
-# CONFIG_MK = $(MFEM_INSTALL_DIR)/share/mfem/config.mk
+CONFIG_MK = $(or $(wildcard $(MFEM_BUILD_DIR)/config/config.mk),\
+   $(wildcard $(MFEM_INSTALL_DIR)/share/mfem/config.mk))
 
 MFEM_LIB_FILE = mfem_is_not_built
 -include $(CONFIG_MK)

examples/makefile

@@ -12,11 +12,10 @@
 # Use the MFEM build directory
 MFEM_DIR ?= ..
 MFEM_BUILD_DIR ?= ..
+MFEM_INSTALL_DIR ?= ../mfem
 SRC = $(if $(MFEM_DIR:..=),$(MFEM_DIR)/examples/,)
-CONFIG_MK = $(MFEM_BUILD_DIR)/config/config.mk
-# Use the MFEM install directory
-# MFEM_INSTALL_DIR = ../mfem
-# CONFIG_MK = $(MFEM_INSTALL_DIR)/share/mfem/config.mk
+CONFIG_MK = $(or $(wildcard $(MFEM_BUILD_DIR)/config/config.mk),\
+   $(wildcard $(MFEM_INSTALL_DIR)/share/mfem/config.mk))
 
 MFEM_LIB_FILE = mfem_is_not_built
 -include $(CONFIG_MK)

examples/moonolith/makefile

@@ -12,11 +12,10 @@
 # Use the MFEM build directory
 MFEM_DIR ?= ../..
 MFEM_BUILD_DIR ?= ../..
+MFEM_INSTALL_DIR ?= ../../mfem
 SRC = $(if $(MFEM_DIR:../..=),$(MFEM_DIR)/examples/moonolith/,)
-CONFIG_MK = $(MFEM_BUILD_DIR)/config/config.mk
-# Use the MFEM install directory
-# MFEM_INSTALL_DIR = ../../mfem
-# CONFIG_MK = $(MFEM_INSTALL_DIR)/share/mfem/config.mk
+CONFIG_MK = $(or $(wildcard $(MFEM_BUILD_DIR)/config/config.mk),\
+   $(wildcard $(MFEM_INSTALL_DIR)/share/mfem/config.mk))
 
 MFEM_LIB_FILE = mfem_is_not_built
 -include $(CONFIG_MK)

examples/petsc/makefile

@@ -12,11 +12,10 @@
 # Use the MFEM build directory
 MFEM_DIR ?= ../..
 MFEM_BUILD_DIR ?= ../..
+MFEM_INSTALL_DIR ?= ../../mfem
 SRC = $(if $(MFEM_DIR:../..=),$(MFEM_DIR)/examples/petsc/,)
-CONFIG_MK = $(MFEM_BUILD_DIR)/config/config.mk
-# Use the MFEM install directory
-# MFEM_INSTALL_DIR = ../../mfem
-# CONFIG_MK = $(MFEM_INSTALL_DIR)/share/mfem/config.mk
+CONFIG_MK = $(or $(wildcard $(MFEM_BUILD_DIR)/config/config.mk),\
+   $(wildcard $(MFEM_INSTALL_DIR)/share/mfem/config.mk))
 
 MFEM_LIB_FILE = mfem_is_not_built
 -include $(CONFIG_MK)

examples/pumi/ex1.cpp

@@ -66,7 +66,6 @@ int main(int argc, char *argv[])
 {
    // 1. Initialize MPI (required by PUMI) and HYPRE.
    Mpi::Init(argc, argv);
-   int num_procs = Mpi::WorldSize();
    int myid = Mpi::WorldRank();
    Hypre::Init();
 

examples/pumi/ex2.cpp

@@ -80,8 +80,6 @@ int main(int argc, char *argv[])
 {
    // 1. Initialize MPI (required by PUMI) and HYPRE.
    Mpi::Init(argc, argv);
-   int num_proc = Mpi::WorldSize();
-   int myId = Mpi::WorldRank();
    Hypre::Init();
 
    // 2. Parse command-line options.

examples/pumi/makefile

@@ -12,11 +12,10 @@
 # Use the MFEM build directory
 MFEM_DIR ?= ../..
 MFEM_BUILD_DIR ?= ../..
+MFEM_INSTALL_DIR ?= ../../mfem
 SRC = $(if $(MFEM_DIR:../..=),$(MFEM_DIR)/examples/pumi/,)
-CONFIG_MK = $(MFEM_BUILD_DIR)/config/config.mk
-# Use the MFEM install directory
-# MFEM_INSTALL_DIR = ../../mfem
-# CONFIG_MK = $(MFEM_INSTALL_DIR)/share/mfem/config.mk
+CONFIG_MK = $(or $(wildcard $(MFEM_BUILD_DIR)/config/config.mk),\
+   $(wildcard $(MFEM_INSTALL_DIR)/share/mfem/config.mk))
 
 MFEM_LIB_FILE = mfem_is_not_built
 -include $(CONFIG_MK)

examples/sundials/CMakeLists.txt

@@ -31,11 +31,21 @@ include_directories(BEFORE ${PROJECT_BINARY_DIR})
 add_custom_target(test_sundials
   ${CMAKE_CTEST_COMMAND} -R sundials USES_TERMINAL)
 
-# Add one executable per cpp file, adding "sundials_" as prefix. Sets
-# "test_sundials" as a target that depends on the given examples.
+# Add one executable per cpp file, adding "sundials_" as prefix so the CMake
+# target is unique from those in the non-SUNDIALS examples. Also sets
+# "test_sundials" as a target that depends on the given SUNDIALS examples.
 set(PFX sundials_)
 add_mfem_examples(SUNDIALS_EXAMPLES_SRCS ${PFX} "" test_sundials)
 
+# Remove "sundials_" prefix from exectuable name for consistency with GNU build
+# system.
+foreach(SRC_FILE ${SUNDIALS_EXAMPLES_SRCS})
+  get_filename_component(SRC_FILENAME ${SRC_FILE} NAME)
+  string(REPLACE ".cpp" "" TARGET_NAME "${PFX}${SRC_FILENAME}")
+  string(REPLACE ${PFX} "" EXE_NAME ${TARGET_NAME})
+  set_target_properties(${TARGET_NAME} PROPERTIES OUTPUT_NAME ${EXE_NAME})
+endforeach()
+
 # Testing.
 # The SUNDIALS tests can be run separately using the target "test_sundials"
 # which builds the examples and runs:
@@ -51,7 +61,10 @@ if (MFEM_ENABLE_TESTING)
   set(EX10_COMMON_OPTS -m ../../data/beam-quad.mesh -o 2 -s 5 -dt 0.15 -tf 6 -vs 10)
   set(EX10_TEST_OPTS ${EX10_COMMON_OPTS}  -r  2)
   set(EX10P_TEST_OPTS ${EX10_COMMON_OPTS} -rp 1)
-  # Example 16: use the default options
+  # Example 16: test ARKODE with implicit time stepping using mass form
+  set(EX16_COMMON_OPTS -s 15)
+  set(EX16_TEST_OPTS ${EX16_COMMON_OPTS})
+  set(EX16P_TEST_OPTS ${EX16_COMMON_OPTS})
 
   # Add the tests: one test per source file.
   foreach(SRC_FILE ${SUNDIALS_EXAMPLES_SRCS})

examples/sundials/ex10.cpp

@@ -1,7 +1,9 @@
 //                                MFEM Example 10
 //                             SUNDIALS Modification
 //
-// Compile with: make ex10
+// Compile with:
+//    make ex10             (GNU make)
+//    make sundials_ex10    (CMake)
 //
 // Sample runs:
 //    ex10 -m ../../data/beam-quad.mesh -r 2 -o 2 -s 12 -dt 0.15 -vs 10

examples/sundials/ex10p.cpp

@@ -1,7 +1,9 @@
 //                       MFEM Example 10 - Parallel Version
 //                             SUNDIALS Modification
 //
-// Compile with: make ex10p
+// Compile with:
+//    make ex10p            (GNU make)
+//    make sundials_ex10p   (CMake)
 //
 // Sample runs:
 //    mpirun -np 4 ex10p -m ../../data/beam-quad.mesh -rp 1 -o 2 -s 12 -dt 0.15 -vs 10

examples/sundials/ex16.cpp

@@ -1,15 +1,21 @@
 //                                MFEM Example 16
 //                             SUNDIALS Modification
 //
-// Compile with: make ex16
+// Compile with:
+//    make ex16             (GNU make)
+//    make sundials_ex16    (CMake)
 //
 // Sample runs:  ex16
 //               ex16 -m ../../data/inline-tri.mesh
 //               ex16 -m ../../data/disc-nurbs.mesh -tf 2
 //               ex16 -s 12 -a 0.0 -k 1.0
+//               ex16 -s 15 -a 0.0 -k 1.0
 //               ex16 -s 8 -a 1.0 -k 0.0 -dt 1e-4 -tf 5e-2 -vs 25
+//               ex16 -s 11 -a 1.0 -k 0.0 -dt 1e-4 -tf 5e-2 -vs 25
 //               ex16 -s 9 -a 0.5 -k 0.5 -o 4 -dt 1e-4 -tf 2e-2 -vs 25
+//               ex16 -s 12 -a 0.5 -k 0.5 -o 4 -dt 1e-4 -tf 2e-2 -vs 25
 //               ex16 -s 10 -dt 1.0e-4 -tf 4.0e-2 -vs 40
+//               ex16 -s 13 -dt 1.0e-4 -tf 4.0e-2 -vs 40
 //               ex16 -m ../../data/fichera-q2.mesh
 //               ex16 -m ../../data/escher.mesh
 //               ex16 -m ../../data/beam-tet.mesh -tf 10 -dt 0.1
@@ -37,75 +43,102 @@
 using namespace std;
 using namespace mfem;
 
-/** After spatial discretization, the conduction model can be written as:
+/** After spatial discretization, the conduction model is expressed as
  *
- *     du/dt = M^{-1}(-Ku)
+ *   M du/dt = - K(u) u
  *
  *  where u is the vector representing the temperature, M is the mass matrix,
- *  and K is the diffusion operator with diffusivity depending on u:
+ *  and K(u) is the diffusion operator with diffusivity depending on u:
  *  (\kappa + \alpha u).
  *
- *  Class ConductionOperator represents the right-hand side of the above ODE.
+ *  Class ConductionOperatorOperator represents the above ODE operator in the
+ *  general form F(u, k, t) = G(u, t) where
+ *
+ *    1. F(u, du/dt, t) = du/dt            (ODE is expressed in EXPLICIT form)
+ *       G(u, t) = - inv(M) K(u) u
+ *    2. F(u, du/dt, t) = M du/dt          (ODE is expressed in IMPLICIT form)
+ *       G(u, t) = - K(u) u
  */
 class ConductionOperator : public TimeDependentOperator
 {
-protected:
    FiniteElementSpace &fespace;
    Array<int> ess_tdof_list; // this list remains empty for pure Neumann b.c.
 
-   BilinearForm *M;
-   BilinearForm *K;
+   BilinearForm M;
+   SparseMatrix Mmat;
 
-   SparseMatrix Mmat, Kmat;
-   SparseMatrix *T; // T = M + dt K
+   const real_t alpha, kappa;
+   std::unique_ptr<BilinearForm> K;
+   SparseMatrix Kmat;
+
+   std::unique_ptr<SparseMatrix> T; // T = M + gam K(u)
 
    CGSolver M_solver; // Krylov solver for inverting the mass matrix M
    DSmoother M_prec;  // Preconditioner for the mass matrix M
 
-   CGSolver T_solver; // Implicit solver for T = M + dt K
+   CGSolver T_solver; // Implicit solver for T = M + gam K(u)
    DSmoother T_prec;  // Preconditioner for the implicit solver
 
-   double alpha, kappa;
-
    mutable Vector z; // auxiliary vector
 
 public:
-   ConductionOperator(FiniteElementSpace &f, double alpha, double kappa,
-                      const Vector &u);
 
-   virtual void Mult(const Vector &u, Vector &du_dt) const;
+   ConductionOperator(FiniteElementSpace &f, const real_t alpha,
+                      const real_t kappa, const Vector &u,
+                      const Type &ode_expression_type);
 
-   /** Solve the Backward-Euler equation: k = f(u + dt*k, t), for the unknown k.
-       This is the only requirement for high-order SDIRK implicit integration.*/
-   virtual void ImplicitSolve(const double dt, const Vector &u, Vector &k);
+   // Compute K(u_n) for use as an approximation in - K(u) u
+   void SetConductionTensor(const Vector &u);
 
-   /// Custom Jacobian system solver for the SUNDIALS time integrators.
-   /** For the ODE system represented by ConductionOperator
+   /** Compute G(u, t) as defined in the IMPLICIT expression form of the ODE
+    operator, i.e., @a v = - K(u_n) @a u. Note that K(u_n) is an
+      approximation to K(u). */
+   void ExplicitMult(const Vector &u, Vector &v) const override;
 
-           M du/dt = -K(u),
+   /** Solve for k in F(u, k, t) = G(u, t) for either EXPLICIT or IMPLICIT
+       expression forms of the ODE operator, i.e., @a k = - inv(M) K(u_n) @a u.
+       Note that K(u_n) is an approximation to K(u). */
+   void Mult(const Vector &u, Vector &k) const override;
 
-       this class facilitates the solution of linear systems of the form
+   /** Solve for k in F(u + gam*k, k, t) = G(u + gam*k, t) for either EXPLICIT
+       or IMPLICIT expression forms of the ODE operator, i.e.,
+       [ M + @a gam K(u_n) ] @a k = - K(u_n) @a u . Note that K(u_n) is an
+       approximation to K(u). */
+   void ImplicitSolve(const real_t gam, const Vector &u, Vector &k) override;
 
-           (M + γK) y = M b,
+   /** Setup to solve for dk in [dF/dk + gam*dF/du - gam*dG/du] dk = G - F for
+       either EXPLICIT or IMPLICIT expression forms of the ODE operator, i.e.,
+       [M - @a gam Jf(u)] dk = G - F, where Jf(u) is an approximation of the
+       Jacobian of -K(u) u. The approximation chosen here is Jf(u) = -K(u_n). */
+   int SUNImplicitSetup(const Vector &u, const Vector &fu, int jok, int *jcur,
+                        real_t gam) override;
 
-       for given b, u (not used), and γ = GetTimeStep(). */
+   /** Solve for @a dk in the system in SUNImplicitSetup to the given tolerance,
+       with the residual @a r providing either
+        1. @a r = G - F = inv(M) f(u) - k       (EXPLICIT expression form)
+        1. @a r = G - F = f(u) - M k            (IMPLICIT expression form)
+       */
+   int SUNImplicitSolve(const Vector &r, Vector &dk, real_t tol) override;
 
-   /** Setup the system (M + dt K) x = M b. This method is used by the implicit
-       SUNDIALS solvers. */
-   virtual int SUNImplicitSetup(const Vector &x, const Vector &fx,
-                                int jok, int *jcur, double gamma);
+   int SUNMassSetup() override;
 
-   /** Solve the system (M + dt K) x = M b. This method is used by the implicit
-       SUNDIALS solvers. */
-   virtual int SUNImplicitSolve(const Vector &b, Vector &x, double tol);
+   int SUNMassSolve(const Vector &b, Vector &x, real_t tol) override;
 
-   /// Update the diffusion BilinearForm K using the given true-dof vector `u`.
-   void SetParameters(const Vector &u);
-
-   virtual ~ConductionOperator();
+   int SUNMassMult(const Vector &x, Vector &v) override;
 };
 
-double InitialTemperature(const Vector &x);
+real_t InitialTemperature(const Vector &x)
+{
+   if (x.Norml2() < 0.5)
+   {
+      return 2.0;
+   }
+   else
+   {
+      return 1.0;
+   }
+}
+
 
 int main(int argc, char *argv[])
 {
@@ -117,16 +150,16 @@ int main(int argc, char *argv[])
    int ref_levels = 2;
    int order = 2;
    int ode_solver_type = 9; // CVODE implicit BDF
-   double t_final = 0.5;
-   double dt = 1.0e-2;
-   double alpha = 1.0e-2;
-   double kappa = 0.5;
+   real_t t_final = 0.5;
+   real_t dt = 1.0e-2;
+   real_t alpha = 1.0e-2;
+   real_t kappa = 0.5;
    bool visualization = true;
    bool visit = false;
    int vis_steps = 5;
 
    // Relative and absolute tolerances for CVODE and ARKODE.
-   const double reltol = 1e-4, abstol = 1e-4;
+   const real_t reltol = 1e-4, abstol = 1e-4;
 
    int precision = 8;
    cout.precision(precision);
@@ -151,7 +184,10 @@ int main(int argc, char *argv[])
                   "9  - CVODE (implicit BDF),\n\t"
                   "10 - ARKODE (default explicit),\n\t"
                   "11 - ARKODE (explicit Fehlberg-6-4-5),\n\t"
-                  "12 - ARKODE (default impicit).");
+                  "12 - ARKODE (default implicit),\n\t"
+                  "13 - ARKODE (default explicit with MFEM mass solve),\n\t"
+                  "14 - ARKODE (explicit Fehlberg-6-4-5 with MFEM mass solve),\n\t"
+                  "15 - ARKODE (default implicit with MFEM mass solve).");
    args.AddOption(&t_final, "-tf", "--t-final",
                   "Final time; start time is 0.");
    args.AddOption(&dt, "-dt", "--time-step",
@@ -174,16 +210,13 @@ int main(int argc, char *argv[])
       args.PrintUsage(cout);
       return 1;
    }
-   if (ode_solver_type < 1 || ode_solver_type > 12)
-   {
-      cout << "Unknown ODE solver type: " << ode_solver_type << '\n';
-      return 3;
-   }
    args.PrintOptions(cout);
 
+   bool use_mass_solver = ode_solver_type >= 13;
+
    // 2. Read the mesh from the given mesh file. We can handle triangular,
    //    quadrilateral, tetrahedral and hexahedral meshes with the same code.
-   Mesh *mesh = new Mesh(mesh_file, 1, 1);
+   std::unique_ptr<Mesh> mesh(new Mesh(mesh_file, 1, 1));
    int dim = mesh->Dimension();
 
    // 3. Refine the mesh to increase the resolution. In this example we do
@@ -197,7 +230,7 @@ int main(int argc, char *argv[])
    // 4. Define the vector finite element space representing the current and the
    //    initial temperature, u_ref.
    H1_FECollection fe_coll(order, dim);
-   FiniteElementSpace fespace(mesh, &fe_coll);
+   FiniteElementSpace fespace(mesh.get(), &fe_coll);
 
    int fe_size = fespace.GetTrueVSize();
    cout << "Number of temperature unknowns: " << fe_size << endl;
@@ -211,8 +244,17 @@ int main(int argc, char *argv[])
    Vector u;
    u_gf.GetTrueDofs(u);
 
-   // 6. Initialize the conduction operator and the visualization.
-   ConductionOperator oper(fespace, alpha, kappa, u);
+   // 6. Initialize the conduction ODE operator and the visualization.
+   ConductionOperator::Type ode_expression_type;
+   if (use_mass_solver)
+   {
+      ode_expression_type = ConductionOperator::Type::IMPLICIT;
+   }
+   else
+   {
+      ode_expression_type = ConductionOperator::Type::EXPLICIT;
+   }
+   ConductionOperator oper(fespace, alpha, kappa, u, ode_expression_type);
 
    u_gf.SetFromTrueDofs(u);
    {
@@ -224,7 +266,7 @@ int main(int argc, char *argv[])
       u_gf.Save(osol);
    }
 
-   VisItDataCollection visit_dc("Example16", mesh);
+   VisItDataCollection visit_dc("Example16", mesh.get());
    visit_dc.RegisterField("temperature", &u_gf);
    if (visit)
    {
@@ -258,52 +300,75 @@ int main(int argc, char *argv[])
    }
 
    // 7. Define the ODE solver used for time integration.
-   double t = 0.0;
-   ODESolver *ode_solver = NULL;
-   CVODESolver *cvode = NULL;
-   ARKStepSolver *arkode = NULL;
+   real_t t = 0.0;
+   std::unique_ptr<ODESolver> ode_solver;
    switch (ode_solver_type)
    {
       // MFEM explicit methods
-      case 1: ode_solver = new ForwardEulerSolver; break;
-      case 2: ode_solver = new RK2Solver(0.5); break; // midpoint method
-      case 3: ode_solver = new RK3SSPSolver; break;
-      case 4: ode_solver = new RK4Solver; break;
+      case 1: ode_solver = std::make_unique<ForwardEulerSolver>(); break;
+      case 2: ode_solver = std::make_unique<RK2Solver>(0.5); break; // midpoint method
+      case 3: ode_solver = std::make_unique<RK3SSPSolver>(); break;
+      case 4: ode_solver = std::make_unique<RK4Solver>(); break;
       // MFEM implicit L-stable methods
-      case 5: ode_solver = new BackwardEulerSolver; break;
-      case 6: ode_solver = new SDIRK23Solver(2); break;
-      case 7: ode_solver = new SDIRK33Solver; break;
+      case 5: ode_solver = std::make_unique<BackwardEulerSolver>(); break;
+      case 6: ode_solver = std::make_unique<SDIRK23Solver>(2); break;
+      case 7: ode_solver = std::make_unique<SDIRK33Solver>(); break;
       // CVODE
       case 8:
-         cvode = new CVODESolver(CV_ADAMS);
-         cvode->Init(oper);
-         cvode->SetSStolerances(reltol, abstol);
-         cvode->SetMaxStep(dt);
-         ode_solver = cvode; break;
       case 9:
-         cvode = new CVODESolver(CV_BDF);
+      {
+         int cvode_solver_type;
+         if (ode_solver_type == 8)
+         {
+            cvode_solver_type = CV_ADAMS;
+         }
+         else
+         {
+            cvode_solver_type = CV_BDF;
+         }
+         std::unique_ptr<CVODESolver> cvode(new CVODESolver(cvode_solver_type));
          cvode->Init(oper);
          cvode->SetSStolerances(reltol, abstol);
          cvode->SetMaxStep(dt);
-         ode_solver = cvode; break;
+         ode_solver = std::move(cvode);
+         break;
+      }
       // ARKODE
       case 10:
       case 11:
-         arkode = new ARKStepSolver(ARKStepSolver::EXPLICIT);
+      case 12:
+      case 13:
+      case 14:
+      case 15:
+      {
+         ARKStepSolver::Type arkode_solver_type;
+         if (ode_solver_type == 12 || ode_solver_type == 15)
+         {
+            arkode_solver_type = ARKStepSolver::IMPLICIT;
+         }
+         else
+         {
+            arkode_solver_type = ARKStepSolver::EXPLICIT;
+         }
+         std::unique_ptr<ARKStepSolver> arkode(
+            new ARKStepSolver(arkode_solver_type));
          arkode->Init(oper);
          arkode->SetSStolerances(reltol, abstol);
          arkode->SetMaxStep(dt);
-         if (ode_solver_type == 11)
+         if (ode_solver_type == 11 || ode_solver_type == 14)
          {
             arkode->SetERKTableNum(ARKODE_FEHLBERG_13_7_8);
          }
-         ode_solver = arkode; break;
-      case 12:
-         arkode = new ARKStepSolver(ARKStepSolver::IMPLICIT);
-         arkode->Init(oper);
-         arkode->SetSStolerances(reltol, abstol);
-         arkode->SetMaxStep(dt);
-         ode_solver = arkode; break;
+         if (use_mass_solver)
+         {
+            arkode->UseMFEMMassLinearSolver(SUNFALSE);
+         }
+         ode_solver = std::move(arkode);
+         break;
+      }
+      default:
+         cout << "Unknown ODE solver type: " << ode_solver_type << '\n';
+         return 3;
    }
 
    // Initialize MFEM integrators, SUNDIALS integrators are initialized above
@@ -311,8 +376,14 @@ int main(int argc, char *argv[])
 
    // Since we want to update the diffusion coefficient after every time step,
    // we need to use the "one-step" mode of the SUNDIALS solvers.
-   if (cvode) { cvode->SetStepMode(CV_ONE_STEP); }
-   if (arkode) { arkode->SetStepMode(ARK_ONE_STEP); }
+   if (CVODESolver* cvode = dynamic_cast<CVODESolver*>(ode_solver.get()))
+   {
+      cvode->SetStepMode(CV_ONE_STEP);
+   }
+   else if (ARKStepSolver* arkode = dynamic_cast<ARKStepSolver*>(ode_solver.get()))
+   {
+      arkode->SetStepMode(ARK_ONE_STEP);
+   }
 
    // 8. Perform time-integration (looping over the time iterations, ti, with a
    //    time-step dt).
@@ -323,7 +394,7 @@ int main(int argc, char *argv[])
    bool last_step = false;
    for (int ti = 1; !last_step; ti++)
    {
-      double dt_real = min(dt, t_final - t);
+      real_t dt_real = min(dt, t_final - t);
 
       // Note that since we are using the "one-step" mode of the SUNDIALS
       // solvers, they will, generally, step over the final time and will not
@@ -337,8 +408,14 @@ int main(int argc, char *argv[])
       if (last_step || (ti % vis_steps) == 0)
       {
          cout << "step " << ti << ", t = " << t << endl;
-         if (cvode) { cvode->PrintInfo(); }
-         if (arkode) { arkode->PrintInfo(); }
+         if (CVODESolver* cvode = dynamic_cast<CVODESolver*>(ode_solver.get()))
+         {
+            cvode->PrintInfo();
+         }
+         else if (ARKStepSolver* arkode = dynamic_cast<ARKStepSolver*>(ode_solver.get()))
+         {
+            arkode->PrintInfo();
+         }
 
          u_gf.SetFromTrueDofs(u);
          if (visualization)
@@ -353,137 +430,153 @@ int main(int argc, char *argv[])
             visit_dc.Save();
          }
       }
-      oper.SetParameters(u);
+      oper.SetConductionTensor(u);
    }
    tic_toc.Stop();
    cout << "Done, " << tic_toc.RealTime() << "s." << endl;
 
    // 9. Save the final solution. This output can be viewed later using GLVis:
    //    "glvis -m ex16.mesh -g ex16-final.gf".
-   {
-      ofstream osol("ex16-final.gf");
-      osol.precision(precision);
-      u_gf.Save(osol);
-   }
-
-   // 10. Free the used memory.
-   delete ode_solver;
-   delete mesh;
+   u_gf.Save("ex16-final.gf", precision);
 
    return 0;
 }
 
-ConductionOperator::ConductionOperator(FiniteElementSpace &f, double al,
-                                       double kap, const Vector &u)
-   : TimeDependentOperator(f.GetTrueVSize(), 0.0), fespace(f), M(NULL), K(NULL),
-     T(NULL), z(height)
+ConductionOperator::ConductionOperator(FiniteElementSpace &fes,
+                                       const real_t alpha, const real_t kappa,
+                                       const Vector &u,
+                                       const Type &ode_expression_type)
+   : TimeDependentOperator(fes.GetTrueVSize(), 0.0, ode_expression_type),
+     fespace(fes), alpha(alpha), kappa(kappa), M(&fespace), z(height)
 {
-   const double rel_tol = 1e-8;
+   // specify a relative tolerance for all solves with MFEM integrators
+   const real_t rel_tol = 1e-8;
 
-   M = new BilinearForm(&fespace);
-   M->AddDomainIntegrator(new MassIntegrator());
-   M->Assemble();
-   M->FormSystemMatrix(ess_tdof_list, Mmat);
+   M.AddDomainIntegrator(new MassIntegrator());
+   M.Assemble();
+   M.FormSystemMatrix(ess_tdof_list, Mmat);
 
    M_solver.iterative_mode = false;
-   M_solver.SetRelTol(rel_tol);
+   M_solver.SetRelTol(rel_tol); // will be overwritten with SUNDIALS integrators
    M_solver.SetAbsTol(0.0);
    M_solver.SetMaxIter(50);
    M_solver.SetPrintLevel(0);
    M_solver.SetPreconditioner(M_prec);
    M_solver.SetOperator(Mmat);
 
-   alpha = al;
-   kappa = kap;
-
    T_solver.iterative_mode = false;
-   T_solver.SetRelTol(rel_tol);
+   T_solver.SetRelTol(rel_tol); // will be overwritten with SUNDIALS integrators
    T_solver.SetAbsTol(0.0);
    T_solver.SetMaxIter(100);
    T_solver.SetPrintLevel(0);
    T_solver.SetPreconditioner(T_prec);
 
-   SetParameters(u);
+   SetConductionTensor(u);
 }
 
-void ConductionOperator::Mult(const Vector &u, Vector &du_dt) const
-{
-   // Compute:
-   //    du_dt = M^{-1}*-K(u)
-   // for du_dt
-   Kmat.Mult(u, z);
-   z.Neg(); // z = -z
-   M_solver.Mult(z, du_dt);
-}
-
-void ConductionOperator::ImplicitSolve(const double dt,
-                                       const Vector &u, Vector &du_dt)
-{
-   // Solve the equation:
-   //    du_dt = M^{-1}*[-K(u + dt*du_dt)]
-   // for du_dt
-   if (T) { delete T; }
-   T = Add(1.0, Mmat, dt, Kmat);
-   T_solver.SetOperator(*T);
-   Kmat.Mult(u, z);
-   z.Neg();
-   T_solver.Mult(z, du_dt);
-}
-
-void ConductionOperator::SetParameters(const Vector &u)
+void ConductionOperator::SetConductionTensor(const Vector &u)
 {
+   // Compute K(u_n).
    GridFunction u_alpha_gf(&fespace);
    u_alpha_gf.SetFromTrueDofs(u);
    for (int i = 0; i < u_alpha_gf.Size(); i++)
    {
       u_alpha_gf(i) = kappa + alpha*u_alpha_gf(i);
    }
-
-   delete K;
-   K = new BilinearForm(&fespace);
-
    GridFunctionCoefficient u_coeff(&u_alpha_gf);
 
+   K = std::make_unique<BilinearForm>(&fespace);
    K->AddDomainIntegrator(new DiffusionIntegrator(u_coeff));
    K->Assemble();
    K->FormSystemMatrix(ess_tdof_list, Kmat);
 }
 
-int ConductionOperator::SUNImplicitSetup(const Vector &x,
-                                         const Vector &fx, int jok, int *jcur,
-                                         double gamma)
+void ConductionOperator::ExplicitMult(const Vector &u, Vector &v) const
 {
-   // Setup the ODE Jacobian T = M + gamma K.
-   if (T) { delete T; }
-   T = Add(1.0, Mmat, gamma, Kmat);
+   // Compute - K(u_n) u.
+   Kmat.Mult(u, v);
+   v.Neg();
+}
+
+void ConductionOperator::Mult(const Vector &u, Vector &k) const
+{
+   // Compute - inv(M) K(u_n) u.
+   ExplicitMult(u, z);
+   M_solver.Mult(z, k);
+}
+
+void ConductionOperator::ImplicitSolve(const real_t gam, const Vector &u,
+                                       Vector &k)
+{
+   // Solve for k in M k = - K(u_n) [u + gam*k].
+   ExplicitMult(u, z);
+   T = std::unique_ptr<SparseMatrix>(Add(1.0, Mmat, gam, Kmat));
    T_solver.SetOperator(*T);
-   *jcur = 1;
-   return (0);
+   T_solver.Mult(z, k);
 }
 
-int ConductionOperator::SUNImplicitSolve(const Vector &b, Vector &x, double tol)
+int ConductionOperator::SUNImplicitSetup(const Vector &u, const Vector &fu,
+                                         int jok, int *jcur, real_t gam)
 {
-   // Solve the system A x = z => (M - gamma K) x = M b.
-   Mmat.Mult(b, z);
-   T_solver.Mult(z, x);
-   return (0);
+   // Compute T = M + gamma K(u_n).
+   T = std::unique_ptr<SparseMatrix>(Add(1.0, Mmat, gam, Kmat));
+   T_solver.SetOperator(*T);
+   *jcur = SUNTRUE; // this should eventually only be set true if K(u) is used
+   return SUNLS_SUCCESS;
 }
 
-ConductionOperator::~ConductionOperator()
+int ConductionOperator::SUNImplicitSolve(const Vector &r, Vector &dk,
+                                         real_t tol)
 {
-   delete T;
-   delete M;
-   delete K;
-}
-
-double InitialTemperature(const Vector &x)
-{
-   if (x.Norml2() < 0.5)
+   // Solve the system [M + gamma K(u_n)] dk = - K(u_n) u - M k.
+   // What value r is providing depends on the ODE expression form:
+   //   EXPLICIT form: r = -inv(M) K(u_n) u - k
+   //   IMPLICIT form: r = -K(u_n) u - M k
+   T_solver.SetRelTol(tol);
+   if (isExplicit())
    {
-      return 2.0;
+      Mmat.Mult(r, z);
+      T_solver.Mult(z, dk);
    }
    else
    {
-      return 1.0;
+      T_solver.Mult(r, dk);
+   }
+   if (T_solver.GetConverged())
+   {
+      return SUNLS_SUCCESS;
+   }
+   else
+   {
+      return SUNLS_CONV_FAIL;
    }
 }
+
+int ConductionOperator::SUNMassSetup()
+{
+   // Do nothing b/c mass solver was setup in constructor.
+   return SUNLS_SUCCESS;
+}
+
+int ConductionOperator::SUNMassSolve(const Vector &b, Vector &x, real_t tol)
+{
+   // Solve the system M x = b.
+   M_solver.SetRelTol(tol);
+   M_solver.Mult(b, x);
+   if (M_solver.GetConverged())
+   {
+      return SUNLS_SUCCESS;
+   }
+   else
+   {
+      return SUNLS_CONV_FAIL;
+   }
+}
+
+int ConductionOperator::SUNMassMult(const Vector &x, Vector &v)
+{
+   // Compute M x.
+   Mmat.Mult(x, v);
+   return SUNLS_SUCCESS;
+}
+

examples/sundials/ex16p.cpp

@@ -1,16 +1,22 @@
 //                         MFEM Example 16 - Parallel Version
 //                              SUNDIALS Modification
 //
-// Compile with: make ex16p
+// Compile with:
+//    make ex16p            (GNU make)
+//    make sundials_ex16p   (CMake)
 //
 // Sample runs:
 //     mpirun -np 4 ex16p
 //     mpirun -np 4 ex16p -m ../../data/inline-tri.mesh
 //     mpirun -np 4 ex16p -m ../../data/disc-nurbs.mesh -tf 2
 //     mpirun -np 4 ex16p -s 12 -a 0.0 -k 1.0
+//     mpirun -np 4 ex16p -s 15 -a 0.0 -k 1.0
 //     mpirun -np 4 ex16p -s 8 -a 1.0 -k 0.0 -dt 4e-6 -tf 2e-2 -vs 50
+//     mpirun -np 4 ex16p -s 11 -a 1.0 -k 0.0 -dt 4e-6 -tf 2e-2 -vs 50
 //     mpirun -np 8 ex16p -s 9 -a 0.5 -k 0.5 -o 4 -dt 8e-6 -tf 2e-2 -vs 50
+//     mpirun -np 8 ex16p -s 12 -a 0.5 -k 0.5 -o 4 -dt 8e-6 -tf 2e-2 -vs 50
 //     mpirun -np 4 ex16p -s 10 -dt 2.0e-4 -tf 4.0e-2
+//     mpirun -np 4 ex16p -s 13 -dt 2.0e-4 -tf 4.0e-2
 //     mpirun -np 16 ex16p -m ../../data/fichera-q2.mesh
 //     mpirun -np 16 ex16p -m ../../data/escher-p2.mesh
 //     mpirun -np 8 ex16p -m ../../data/beam-tet.mesh -tf 10 -dt 0.1
@@ -38,66 +44,102 @@
 using namespace std;
 using namespace mfem;
 
-/** After spatial discretization, the conduction model can be written as:
+/** After spatial discretization, the conduction model is expressed as
  *
- *     du/dt = M^{-1}(-Ku)
+ *   M du/dt = - K(u) u
  *
  *  where u is the vector representing the temperature, M is the mass matrix,
- *  and K is the diffusion operator with diffusivity depending on u:
+ *  and K(u) is the diffusion operator with diffusivity depending on u:
  *  (\kappa + \alpha u).
  *
- *  Class ConductionOperator represents the right-hand side of the above ODE.
+ *  Class ConductionOperatorOperator represents the above ODE operator in the
+ *  general form F(u, k, t) = G(u, t) where either
+ *
+ *    1. F(u, du/dt, t) = du/dt            (ODE is expressed in EXPLICIT form)
+ *       G(u, t) = - inv(M) K(u) u
+ *    2. F(u, du/dt, t) = M du/dt          (ODE is expressed in IMPLICIT form)
+ *       G(u, t) = - K(u) u
  */
 class ConductionOperator : public TimeDependentOperator
 {
-protected:
    ParFiniteElementSpace &fespace;
    Array<int> ess_tdof_list; // this list remains empty for pure Neumann b.c.
 
-   ParBilinearForm *M;
-   ParBilinearForm *K;
-
+   ParBilinearForm M;
    HypreParMatrix Mmat;
+
+   const real_t alpha, kappa;
+   std::unique_ptr<BilinearForm> K;
    HypreParMatrix Kmat;
-   HypreParMatrix *T; // T = M + dt K
-   double current_dt;
 
-   CGSolver M_solver;    // Krylov solver for inverting the mass matrix M
-   HypreSmoother M_prec; // Preconditioner for the mass matrix M
+   std::unique_ptr<HypreParMatrix> T; // T = M + gam K(u)
 
-   CGSolver T_solver;    // Implicit solver for T = M + dt K
-   HypreSmoother T_prec; // Preconditioner for the implicit solver
+   CGSolver M_solver; // Krylov solver for inverting the mass matrix M
+   HypreSmoother M_prec;  // Preconditioner for the mass matrix M
 
-   double alpha, kappa;
+   CGSolver T_solver; // Implicit solver for T = M + gam K(u)
+   HypreSmoother T_prec;  // Preconditioner for the implicit solver
 
    mutable Vector z; // auxiliary vector
 
 public:
-   ConductionOperator(ParFiniteElementSpace &f, double alpha, double kappa,
-                      const Vector &u);
 
-   virtual void Mult(const Vector &u, Vector &du_dt) const;
+   ConductionOperator(ParFiniteElementSpace &f, const real_t alpha,
+                      const real_t kappa, const Vector &u,
+                      const Type &ode_expression_type);
 
-   /** Solve the Backward-Euler equation: k = f(u + dt*k, t), for the unknown k.
-       This is the only requirement for high-order SDIRK implicit integration.*/
-   virtual void ImplicitSolve(const double dt, const Vector &u, Vector &k);
+   // Compute K(u_n) for use as an approximation in - K(u) u
+   void SetConductionTensor(const Vector &u);
 
-   /** Setup the system (M + dt K) x = M b. This method is used by the implicit
-       SUNDIALS solvers. */
-   virtual int SUNImplicitSetup(const Vector &x, const Vector &fx,
-                                int jok, int *jcur, double gamma);
+   /** Compute G(u, t) as defined in the IMPLICIT expression form of the ODE
+       operator, i.e., @a v = - K(u_n) @a u. Note that K(u_n) is an
+       approximation to K(u). */
+   void ExplicitMult(const Vector &u, Vector &v) const override;
 
-   /** Solve the system (M + dt K) x = M b. This method is used by the implicit
-       SUNDIALS solvers. */
-   virtual int SUNImplicitSolve(const Vector &b, Vector &x, double tol);
+   /** Solve for k in F(u, k, t) = G(u, t) for either EXPLICIT or IMPLICIT
+       expression forms of the ODE operator, i.e., @a k = - inv(M) K(u_n) @a u.
+       Note that K(u_n) is an approximation to K(u). */
+   void Mult(const Vector &u, Vector &k) const override;
 
-   /// Update the diffusion BilinearForm K using the given true-dof vector `u`.
-   void SetParameters(const Vector &u);
+   /** Solve for k in F(u + gam*k, k, t) = G(u + gam*k, t) for either EXPLICIT
+       or IMPLICIT expression forms of the ODE operator, i.e.,
+       [ M + @a gam K(u_n) ] @a k = - K(u_n) @a u . Note that K(u_n) is an
+       approximation to K(u). */
+   void ImplicitSolve(const real_t gam, const Vector &u, Vector &k) override;
 
-   virtual ~ConductionOperator();
+   /** Setup to solve for dk in [dF/dk + gam*dF/du - gam*dG/du] dk = G - F for
+       either EXPLICIT or IMPLICIT expression forms of the ODE operator, i.e.,
+       [M - @a gam Jf(u)] dk = G - F, where Jf(u) is an approximation of the
+       Jacobian of -K(u) u. The approximation chosen here is Jf(u) = -K(u_n). */
+   int SUNImplicitSetup(const Vector &u, const Vector &fu, int jok, int *jcur,
+                        real_t gam) override;
+
+   /** Solve for @a dk in the system in SUNImplicitSetup to the given tolerance,
+       with the residual @a r providing either
+        1. @a r = G - F = inv(M) f(u) - k       (EXPLICIT expression form)
+        1. @a r = G - F = f(u) - M k            (IMPLICIT expression form)
+       */
+   int SUNImplicitSolve(const Vector &r, Vector &dk, real_t tol) override;
+
+   int SUNMassSetup() override;
+
+   int SUNMassSolve(const Vector &b, Vector &x, real_t tol) override;
+
+   int SUNMassMult(const Vector &x, Vector &v) override;
 };
 
-double InitialTemperature(const Vector &x);
+real_t InitialTemperature(const Vector &x)
+{
+   if (x.Norml2() < 0.5)
+   {
+      return 2.0;
+   }
+   else
+   {
+      return 1.0;
+   }
+}
+
 
 int main(int argc, char *argv[])
 {
@@ -114,16 +156,16 @@ int main(int argc, char *argv[])
    int par_ref_levels = 1;
    int order = 2;
    int ode_solver_type = 9; // CVODE implicit BDF
-   double t_final = 0.5;
-   double dt = 1.0e-2;
-   double alpha = 1.0e-2;
-   double kappa = 0.5;
+   real_t t_final = 0.5;
+   real_t dt = 1.0e-2;
+   real_t alpha = 1.0e-2;
+   real_t kappa = 0.5;
    bool visualization = true;
    bool visit = false;
    int vis_steps = 5;
 
    // Relative and absolute tolerances for CVODE and ARKODE.
-   const double reltol = 1e-4, abstol = 1e-4;
+   const real_t reltol = 1e-4, abstol = 1e-4;
 
    int precision = 8;
    cout.precision(precision);
@@ -150,7 +192,10 @@ int main(int argc, char *argv[])
                   "9  - CVODE (implicit BDF),\n\t"
                   "10 - ARKODE (default explicit),\n\t"
                   "11 - ARKODE (explicit Fehlberg-6-4-5),\n\t"
-                  "12 - ARKODE (default impicit).");
+                  "12 - ARKODE (default implicit),\n\t"
+                  "13 - ARKODE (default explicit with MFEM mass solve),\n\t"
+                  "14 - ARKODE (explicit Fehlberg-6-4-5 with MFEM mass solve),\n\t"
+                  "15 - ARKODE (default implicit with MFEM mass solve).");
    args.AddOption(&t_final, "-tf", "--t-final",
                   "Final time; start time is 0.");
    args.AddOption(&dt, "-dt", "--time-step",
@@ -174,40 +219,33 @@ int main(int argc, char *argv[])
       return 1;
    }
 
-   if (myid == 0)
+   if (Mpi::Root())
    {
       args.PrintOptions(cout);
    }
 
-   // check for valid ODE solver option
-   if (ode_solver_type < 1 || ode_solver_type > 12)
-   {
-      if (myid == 0)
-      {
-         cout << "Unknown ODE solver type: " << ode_solver_type << '\n';
-      }
-      return 1;
-   }
+   bool use_mass_solver = ode_solver_type >= 13;
 
-   // 3. Read the serial mesh from the given mesh file on all processors. We can
+   // 3. Define a parallel mesh by a partitioning of a serial mesh. Read the
+   //    serial mesh from the given mesh file on all processors. We can
    //    handle triangular, quadrilateral, tetrahedral and hexahedral meshes
    //    with the same code.
-   Mesh *mesh = new Mesh(mesh_file, 1, 1);
-   int dim = mesh->Dimension();
-
-   // 4. Refine the mesh in serial to increase the resolution. In this example
-   //    we do 'ser_ref_levels' of uniform refinement, where 'ser_ref_levels' is
-   //    a command-line parameter.
-   for (int lev = 0; lev < ser_ref_levels; lev++)
+   std::unique_ptr<ParMesh> pmesh;
    {
-      mesh->UniformRefinement();
-   }
+      std::unique_ptr<Mesh> mesh(new Mesh(mesh_file, 1, 1));
 
-   // 5. Define a parallel mesh by a partitioning of the serial mesh. Refine
-   //    this mesh further in parallel to increase the resolution. Once the
-   //    parallel mesh is defined, the serial mesh can be deleted.
-   ParMesh *pmesh = new ParMesh(MPI_COMM_WORLD, *mesh);
-   delete mesh;
+      // 4. Refine the mesh in serial to increase the resolution. In this example
+      //    we do 'ser_ref_levels' of uniform refinement, where 'ser_ref_levels' is
+      //    a command-line parameter.
+      for (int lev = 0; lev < ser_ref_levels; lev++)
+      {
+         mesh->UniformRefinement();
+      }
+
+      // 5. Refine this mesh further in parallel to increase the resolution.
+      //    Once the parallel mesh is defined, the serial mesh can be deleted.
+      pmesh = std::make_unique<ParMesh>(MPI_COMM_WORLD, *mesh);
+   }
    for (int lev = 0; lev < par_ref_levels; lev++)
    {
       pmesh->UniformRefinement();
@@ -215,8 +253,9 @@ int main(int argc, char *argv[])
 
    // 6. Define the vector finite element space representing the current and the
    //    initial temperature, u_ref.
+   int dim = pmesh->Dimension();
    H1_FECollection fe_coll(order, dim);
-   ParFiniteElementSpace fespace(pmesh, &fe_coll);
+   ParFiniteElementSpace fespace(pmesh.get(), &fe_coll);
 
    int fe_size = fespace.GlobalTrueVSize();
    if (myid == 0)
@@ -233,8 +272,17 @@ int main(int argc, char *argv[])
    Vector u;
    u_gf.GetTrueDofs(u);
 
-   // 8. Initialize the conduction operator and the VisIt visualization.
-   ConductionOperator oper(fespace, alpha, kappa, u);
+   // 8. Initialize the conduction ODE operator and the visualization.
+   ConductionOperator::Type ode_expression_type;
+   if (use_mass_solver)
+   {
+      ode_expression_type = ConductionOperator::Type::IMPLICIT;
+   }
+   else
+   {
+      ode_expression_type = ConductionOperator::Type::EXPLICIT;
+   }
+   ConductionOperator oper(fespace, alpha, kappa, u, ode_expression_type);
 
    u_gf.SetFromTrueDofs(u);
    {
@@ -249,7 +297,7 @@ int main(int argc, char *argv[])
       u_gf.Save(osol);
    }
 
-   VisItDataCollection visit_dc("Example16-Parallel", pmesh);
+   VisItDataCollection visit_dc("Example16-Parallel", pmesh.get());
    visit_dc.RegisterField("temperature", &u_gf);
    if (visit)
    {
@@ -293,52 +341,76 @@ int main(int argc, char *argv[])
    }
 
    // 9. Define the ODE solver used for time integration.
-   double t = 0.0;
-   ODESolver *ode_solver = NULL;
-   CVODESolver *cvode = NULL;
-   ARKStepSolver *arkode = NULL;
+   real_t t = 0.0;
+   std::unique_ptr<ODESolver> ode_solver;
    switch (ode_solver_type)
    {
       // MFEM explicit methods
-      case 1: ode_solver = new ForwardEulerSolver; break;
-      case 2: ode_solver = new RK2Solver(0.5); break; // midpoint method
-      case 3: ode_solver = new RK3SSPSolver; break;
-      case 4: ode_solver = new RK4Solver; break;
+      case 1: ode_solver = std::make_unique<ForwardEulerSolver>(); break;
+      case 2: ode_solver = std::make_unique<RK2Solver>(0.5); break; // midpoint method
+      case 3: ode_solver = std::make_unique<RK3SSPSolver>(); break;
+      case 4: ode_solver = std::make_unique<RK4Solver>(); break;
       // MFEM implicit L-stable methods
-      case 5: ode_solver = new BackwardEulerSolver; break;
-      case 6: ode_solver = new SDIRK23Solver(2); break;
-      case 7: ode_solver = new SDIRK33Solver; break;
+      case 5: ode_solver = std::make_unique<BackwardEulerSolver>(); break;
+      case 6: ode_solver = std::make_unique<SDIRK23Solver>(2); break;
+      case 7: ode_solver = std::make_unique<SDIRK33Solver>(); break;
       // CVODE
       case 8:
-         cvode = new CVODESolver(MPI_COMM_WORLD, CV_ADAMS);
-         cvode->Init(oper);
-         cvode->SetSStolerances(reltol, abstol);
-         cvode->SetMaxStep(dt);
-         ode_solver = cvode; break;
       case 9:
-         cvode = new CVODESolver(MPI_COMM_WORLD, CV_BDF);
+      {
+         int cvode_solver_type;
+         if (ode_solver_type == 8)
+         {
+            cvode_solver_type = CV_ADAMS;
+         }
+         else
+         {
+            cvode_solver_type = CV_BDF;
+         }
+         std::unique_ptr<CVODESolver> cvode(
+            new CVODESolver(MPI_COMM_WORLD, cvode_solver_type));
          cvode->Init(oper);
          cvode->SetSStolerances(reltol, abstol);
          cvode->SetMaxStep(dt);
-         ode_solver = cvode; break;
+         ode_solver = std::move(cvode);
+         break;
+      }
       // ARKODE
       case 10:
       case 11:
-         arkode = new ARKStepSolver(MPI_COMM_WORLD, ARKStepSolver::EXPLICIT);
+      case 12:
+      case 13:
+      case 14:
+      case 15:
+      {
+         ARKStepSolver::Type arkode_solver_type;
+         if (ode_solver_type == 12 || ode_solver_type == 15)
+         {
+            arkode_solver_type = ARKStepSolver::IMPLICIT;
+         }
+         else
+         {
+            arkode_solver_type = ARKStepSolver::EXPLICIT;
+         }
+         std::unique_ptr<ARKStepSolver> arkode(
+            new ARKStepSolver(MPI_COMM_WORLD, arkode_solver_type));
          arkode->Init(oper);
          arkode->SetSStolerances(reltol, abstol);
          arkode->SetMaxStep(dt);
-         if (ode_solver_type == 11)
+         if (ode_solver_type == 11 || ode_solver_type == 14)
          {
             arkode->SetERKTableNum(ARKODE_FEHLBERG_13_7_8);
          }
-         ode_solver = arkode; break;
-      case 12:
-         arkode = new ARKStepSolver(MPI_COMM_WORLD, ARKStepSolver::IMPLICIT);
-         arkode->Init(oper);
-         arkode->SetSStolerances(reltol, abstol);
-         arkode->SetMaxStep(dt);
-         ode_solver = arkode; break;
+         if (use_mass_solver)
+         {
+            arkode->UseMFEMMassLinearSolver(SUNFALSE);
+         }
+         ode_solver = std::move(arkode);
+         break;
+      }
+      default:
+         cout << "Unknown ODE solver type: " << ode_solver_type << '\n';
+         return 3;
    }
 
    // Initialize MFEM integrators, SUNDIALS integrators are initialized above
@@ -346,12 +418,18 @@ int main(int argc, char *argv[])
 
    // Since we want to update the diffusion coefficient after every time step,
    // we need to use the "one-step" mode of the SUNDIALS solvers.
-   if (cvode) { cvode->SetStepMode(CV_ONE_STEP); }
-   if (arkode) { arkode->SetStepMode(ARK_ONE_STEP); }
+   if (CVODESolver* cvode = dynamic_cast<CVODESolver*>(ode_solver.get()))
+   {
+      cvode->SetStepMode(CV_ONE_STEP);
+   }
+   else if (ARKStepSolver* arkode = dynamic_cast<ARKStepSolver*>(ode_solver.get()))
+   {
+      arkode->SetStepMode(ARK_ONE_STEP);
+   }
 
    // 10. Perform time-integration (looping over the time iterations, ti, with a
    //     time-step dt).
-   if (myid == 0)
+   if (Mpi::Root())
    {
       cout << "Integrating the ODE ..." << endl;
    }
@@ -361,7 +439,7 @@ int main(int argc, char *argv[])
    bool last_step = false;
    for (int ti = 1; !last_step; ti++)
    {
-      double dt_real = min(dt, t_final - t);
+      real_t dt_real = min(dt, t_final - t);
 
       // Note that since we are using the "one-step" mode of the SUNDIALS
       // solvers, they will, generally, step over the final time and will not
@@ -377,8 +455,14 @@ int main(int argc, char *argv[])
          if (myid == 0)
          {
             cout << "step " << ti << ", t = " << t << endl;
-            if (cvode) { cvode->PrintInfo(); }
-            if (arkode) { arkode->PrintInfo(); }
+            if (CVODESolver* cvode = dynamic_cast<CVODESolver*>(ode_solver.get()))
+            {
+               cvode->PrintInfo();
+            }
+            else if (ARKStepSolver* arkode = dynamic_cast<ARKStepSolver*>(ode_solver.get()))
+            {
+               arkode->PrintInfo();
+            }
          }
 
          u_gf.SetFromTrueDofs(u);
@@ -395,46 +479,38 @@ int main(int argc, char *argv[])
             visit_dc.Save();
          }
       }
-      oper.SetParameters(u);
+      oper.SetConductionTensor(u);
    }
    tic_toc.Stop();
-   if (myid == 0)
+   if (Mpi::Root())
    {
       cout << "Done, " << tic_toc.RealTime() << "s." << endl;
    }
 
    // 11. Save the final solution in parallel. This output can be viewed later
    //     using GLVis: "glvis -np <np> -m ex16-mesh -g ex16-final".
-   {
-      ostringstream sol_name;
-      sol_name << "ex16-final." << setfill('0') << setw(6) << myid;
-      ofstream osol(sol_name.str().c_str());
-      osol.precision(precision);
-      u_gf.Save(osol);
-   }
-
-   // 12. Free the used memory.
-   delete ode_solver;
-   delete pmesh;
+   u_gf.Save("ex16-final", precision);
 
    return 0;
 }
 
-ConductionOperator::ConductionOperator(ParFiniteElementSpace &f, double al,
-                                       double kap, const Vector &u)
-   : TimeDependentOperator(f.GetTrueVSize(), 0.0), fespace(f), M(NULL), K(NULL),
-     T(NULL),
-     M_solver(f.GetComm()), T_solver(f.GetComm()), z(height)
+ConductionOperator::ConductionOperator(ParFiniteElementSpace &fes,
+                                       const real_t alpha, const real_t kappa,
+                                       const Vector &u,
+                                       const Type &ode_expression_type)
+   : TimeDependentOperator(fes.GetTrueVSize(), 0.0, ode_expression_type),
+     fespace(fes), alpha(alpha), kappa(kappa), M(&fespace),
+     M_solver(fes.GetComm()), T_solver(fes.GetComm()), z(height)
 {
-   const double rel_tol = 1e-8;
+   // specify a relative tolerance for all solves with MFEM integrators
+   const real_t rel_tol = 1e-8;
 
-   M = new ParBilinearForm(&fespace);
-   M->AddDomainIntegrator(new MassIntegrator());
-   M->Assemble(0); // keep sparsity pattern of M and K the same
-   M->FormSystemMatrix(ess_tdof_list, Mmat);
+   M.AddDomainIntegrator(new MassIntegrator());
+   M.Assemble(0); // keep zeros to keep sparsity pattern of M and K the same
+   M.FormSystemMatrix(ess_tdof_list, Mmat);
 
    M_solver.iterative_mode = false;
-   M_solver.SetRelTol(rel_tol);
+   M_solver.SetRelTol(rel_tol); // will be overwritten with SUNDIALS integrators
    M_solver.SetAbsTol(0.0);
    M_solver.SetMaxIter(100);
    M_solver.SetPrintLevel(0);
@@ -442,97 +518,118 @@ ConductionOperator::ConductionOperator(ParFiniteElementSpace &f, double al,
    M_solver.SetPreconditioner(M_prec);
    M_solver.SetOperator(Mmat);
 
-   alpha = al;
-   kappa = kap;
-
    T_solver.iterative_mode = false;
-   T_solver.SetRelTol(rel_tol);
+   T_solver.SetRelTol(rel_tol); // will be overwritten with SUNDIALS integrators
    T_solver.SetAbsTol(0.0);
    T_solver.SetMaxIter(100);
    T_solver.SetPrintLevel(0);
    T_solver.SetPreconditioner(T_prec);
 
-   SetParameters(u);
+   SetConductionTensor(u);
 }
 
-void ConductionOperator::Mult(const Vector &u, Vector &du_dt) const
-{
-   // Compute:
-   //    du_dt = M^{-1}*-K(u)
-   // for du_dt
-   Kmat.Mult(u, z);
-   z.Neg(); // z = -z
-   M_solver.Mult(z, du_dt);
-}
-
-void ConductionOperator::ImplicitSolve(const double dt,
-                                       const Vector &u, Vector &du_dt)
-{
-   // Solve the equation:
-   //    du_dt = M^{-1}*[-K(u + dt*du_dt)]
-   // for du_dt
-   if (T) { delete T; }
-   T = Add(1.0, Mmat, dt, Kmat);
-   T_solver.SetOperator(*T);
-   Kmat.Mult(u, z);
-   z.Neg();
-   T_solver.Mult(z, du_dt);
-}
-
-int ConductionOperator::SUNImplicitSetup(const Vector &x,
-                                         const Vector &fx, int jok, int *jcur,
-                                         double gamma)
-{
-   // Setup the ODE Jacobian T = M + gamma K.
-   if (T) { delete T; }
-   T = Add(1.0, Mmat, gamma, Kmat);
-   T_solver.SetOperator(*T);
-   *jcur = 1;
-   return (0);
-}
-
-int ConductionOperator::SUNImplicitSolve(const Vector &b, Vector &x, double tol)
-{
-   // Solve the system A x = z => (M - gamma K) x = M b.
-   Mmat.Mult(b, z);
-   T_solver.Mult(z, x);
-   return (0);
-}
-
-void ConductionOperator::SetParameters(const Vector &u)
+void ConductionOperator::SetConductionTensor(const Vector &u)
 {
+   // Compute K(u_n).
    ParGridFunction u_alpha_gf(&fespace);
    u_alpha_gf.SetFromTrueDofs(u);
    for (int i = 0; i < u_alpha_gf.Size(); i++)
    {
       u_alpha_gf(i) = kappa + alpha*u_alpha_gf(i);
    }
-
-   delete K;
-   K = new ParBilinearForm(&fespace);
-
    GridFunctionCoefficient u_coeff(&u_alpha_gf);
 
+   K = std::make_unique<ParBilinearForm>(&fespace);
    K->AddDomainIntegrator(new DiffusionIntegrator(u_coeff));
-   K->Assemble(0); // keep sparsity pattern of M and K the same
+   K->Assemble(0); // keep zeros to keep sparsity pattern of M and K the same
    K->FormSystemMatrix(ess_tdof_list, Kmat);
 }
 
-ConductionOperator::~ConductionOperator()
+void ConductionOperator::ExplicitMult(const Vector &u, Vector &v) const
 {
-   delete T;
-   delete M;
-   delete K;
+   // Compute - K(u_n) u.
+   Kmat.Mult(u, v);
+   v.Neg();
 }
 
-double InitialTemperature(const Vector &x)
+void ConductionOperator::Mult(const Vector &u, Vector &k) const
 {
-   if (x.Norml2() < 0.5)
+   // Compute - inv(M) K(u_n) u.
+   ExplicitMult(u, z);
+   M_solver.Mult(z, k);
+}
+
+void ConductionOperator::ImplicitSolve(const real_t gam, const Vector &u,
+                                       Vector &k)
+{
+   // Solve for k in M k = - K(u_n) [u + gam*k].
+   ExplicitMult(u, z);
+   T = std::unique_ptr<HypreParMatrix>(Add(1.0, Mmat, gam, Kmat));
+   T_solver.SetOperator(*T);
+   T_solver.Mult(z, k);
+}
+
+int ConductionOperator::SUNImplicitSetup(const Vector &u, const Vector &fu,
+                                         int jok, int *jcur, real_t gam)
+{
+   // Compute T = M + gamma K(u_n).
+   T = std::unique_ptr<HypreParMatrix>(Add(1.0, Mmat, gam, Kmat));
+   T_solver.SetOperator(*T);
+   *jcur = SUNTRUE; // this should eventually only be set true if K(u) is used
+   return SUNLS_SUCCESS;
+}
+
+int ConductionOperator::SUNImplicitSolve(const Vector &r, Vector &dk,
+                                         real_t tol)
+{
+   // Solve the system [M + gamma K(u_n)] dk = - K(u_n) u - M k.
+   // What value r is providing depends on the ODE expression form:
+   //   EXPLICIT form: r = -inv(M) K(u_n) u - k
+   //   IMPLICIT form: r = -K(u_n) u - M k
+   T_solver.SetRelTol(tol);
+   if (isExplicit())
    {
-      return 2.0;
+      Mmat.Mult(r, z);
+      T_solver.Mult(z, dk);
    }
    else
    {
-      return 1.0;
+      T_solver.Mult(r, dk);
+   }
+   if (T_solver.GetConverged())
+   {
+      return SUNLS_SUCCESS;
+   }
+   else
+   {
+      return SUNLS_CONV_FAIL;
    }
 }
+
+int ConductionOperator::SUNMassSetup()
+{
+   // Do nothing b/c mass solver was setup in constructor.
+   return SUNLS_SUCCESS;
+}
+
+int ConductionOperator::SUNMassSolve(const Vector &b, Vector &x, real_t tol)
+{
+   // Solve the system M x = b.
+   M_solver.SetRelTol(tol);
+   M_solver.Mult(b, x);
+   if (M_solver.GetConverged())
+   {
+      return SUNLS_SUCCESS;
+   }
+   else
+   {
+      return SUNLS_CONV_FAIL;
+   }
+}
+
+int ConductionOperator::SUNMassMult(const Vector &x, Vector &v)
+{
+   // Compute M x.
+   Mmat.Mult(x, v);
+   return SUNLS_SUCCESS;
+}

examples/sundials/ex9.cpp

@@ -1,7 +1,9 @@
 //                                MFEM Example 9
 //                             SUNDIALS Modification
 //
-// Compile with: make ex9
+// Compile with:
+//    make ex9              (GNU make)
+//    make sundials_ex9     (CMake)
 //
 // Sample runs:
 //    ex9 -m ../../data/periodic-segment.mesh -p 0 -r 2 -s 7 -dt 0.005

examples/sundials/ex9p.cpp

@@ -1,7 +1,9 @@
 //                       MFEM Example 9 - Parallel Version
 //                             SUNDIALS Modification
 //
-// Compile with: make ex9p
+// Compile with:
+//    make ex9p             (GNU make)
+//    make sundials_ex9p    (CMake)
 //
 // Sample runs:
 //    mpirun -np 4 ex9p -m ../../data/periodic-segment.mesh -p 1 -rp 1 -s 7 -dt 0.0025

examples/sundials/makefile

@@ -12,11 +12,10 @@
 # Use the MFEM build directory
 MFEM_DIR ?= ../..
 MFEM_BUILD_DIR ?= ../..
+MFEM_INSTALL_DIR ?= ../../mfem
 SRC = $(if $(MFEM_DIR:../..=),$(MFEM_DIR)/examples/sundials/,)
-CONFIG_MK = $(MFEM_BUILD_DIR)/config/config.mk
-# Use the MFEM install directory
-# MFEM_INSTALL_DIR = ../../mfem
-# CONFIG_MK = $(MFEM_INSTALL_DIR)/share/mfem/config.mk
+CONFIG_MK = $(or $(wildcard $(MFEM_BUILD_DIR)/config/config.mk),\
+   $(wildcard $(MFEM_INSTALL_DIR)/share/mfem/config.mk))
 
 MFEM_LIB_FILE = mfem_is_not_built
 -include $(CONFIG_MK)
@@ -100,6 +99,12 @@ ex10-test-seq: ex10
 	@$(call mfem-test,$<,, $(SERIAL_NAME),$(EX10_ARGS))
 ex10p-test-par: ex10p
 	@$(call mfem-test,$<, $(RUN_MPI), $(PARALLEL_NAME),$(EX10P_ARGS))
+# Example 16: test ARKODE with implicit time stepping using mass form
+EX16_COMMON_ARGS := -s 15
+ex16-test-seq: ex16
+	@$(call mfem-test,$<,, $(SERIAL_NAME),$(EX16_COMMON_ARGS))
+ex16p-test-par: ex16p
+	@$(call mfem-test,$<, $(RUN_MPI), $(PARALLEL_NAME),$(EX16_COMMON_ARGS))
 
 # Testing: "test" target and mfem-test* variables are defined in config/test.mk
 

examples/superlu/makefile

@@ -12,11 +12,10 @@
 # Use the MFEM build directory
 MFEM_DIR ?= ../..
 MFEM_BUILD_DIR ?= ../..
+MFEM_INSTALL_DIR ?= ../../mfem
 SRC = $(if $(MFEM_DIR:../..=),$(MFEM_DIR)/examples/superlu/,)
-CONFIG_MK = $(MFEM_BUILD_DIR)/config/config.mk
-# Use the MFEM install directory
-# MFEM_INSTALL_DIR = ../../mfem
-# CONFIG_MK = $(MFEM_INSTALL_DIR)/share/mfem/config.mk
+CONFIG_MK = $(or $(wildcard $(MFEM_BUILD_DIR)/config/config.mk),\
+   $(wildcard $(MFEM_INSTALL_DIR)/share/mfem/config.mk))
 
 MFEM_LIB_FILE = mfem_is_not_built
 -include $(CONFIG_MK)

fem/bilinearform.cpp

@@ -280,7 +280,7 @@ void BilinearForm::AddBdrFaceIntegrator(BilinearFormIntegrator *bfi,
    boundary_face_integs_marker.Append(&bdr_marker);
 }
 
-void BilinearForm::ComputeElementMatrix(int i, DenseMatrix &elmat)
+void BilinearForm::ComputeElementMatrix(int i, DenseMatrix &elmat) const
 {
    if (element_matrices)
    {
@@ -308,7 +308,7 @@ void BilinearForm::ComputeElementMatrix(int i, DenseMatrix &elmat)
    }
 }
 
-void BilinearForm::ComputeBdrElementMatrix(int i, DenseMatrix &elmat)
+void BilinearForm::ComputeBdrElementMatrix(int i, DenseMatrix &elmat) const
 {
    if (boundary_integs.Size())
    {
@@ -329,6 +329,79 @@ void BilinearForm::ComputeBdrElementMatrix(int i, DenseMatrix &elmat)
    }
 }
 
+void BilinearForm::ComputeFaceMatrix(int i, DenseMatrix &elmat) const
+{
+   FaceElementTransformations *tr;
+   Mesh *mesh = fes -> GetMesh();
+   tr = mesh -> GetFaceElementTransformations (i);
+
+   const FiniteElement *fe1, *fe2;
+   fe1 = fes->GetFE(tr->Elem1No);
+   if (tr->Elem2No >= 0)
+   {
+      fe2 = fes->GetFE(tr->Elem2No);
+   }
+   else
+   {
+      // The fe2 object is really a dummy and not used on the
+      // boundaries, but we can't dereference a NULL pointer, and we don't
+      // want to actually make a fake element.
+      fe2 = fe1;
+   }
+
+   if (interior_face_integs.Size())
+   {
+      interior_face_integs[0] -> AssembleFaceMatrix (*fe1, *fe2, *tr, elmat);
+      for (int k = 1; k < interior_face_integs.Size(); k++)
+      {
+         interior_face_integs[k] -> AssembleFaceMatrix (*fe1, *fe2, *tr, elemmat);
+         elmat += elemmat;
+      }
+   }
+   else
+   {
+      int ndof = fe1->GetDof() * fes->GetVDim();
+      if (tr->Elem2No >= 0)
+      {
+         ndof += fe2->GetDof() * fes->GetVDim();
+      }
+
+      elmat.SetSize(ndof);
+      elmat = 0.0;
+   }
+}
+
+void BilinearForm::ComputeBdrFaceMatrix(int i, DenseMatrix &elmat) const
+{
+   FaceElementTransformations *tr;
+   Mesh *mesh = fes -> GetMesh();
+   tr = mesh -> GetBdrFaceTransformations (i);
+
+   const FiniteElement *fe1, *fe2;
+
+   fe1 = fes -> GetFE (tr -> Elem1No);
+   // The fe2 object is really a dummy and not used on the boundaries,
+   // but we can't dereference a NULL pointer, and we don't want to
+   // actually make a fake element.
+   fe2 = fe1;
+
+   if (boundary_face_integs.Size())
+   {
+      boundary_face_integs[0] -> AssembleFaceMatrix (*fe1, *fe2, *tr, elmat);
+      for (int k = 1; k < boundary_face_integs.Size(); k++)
+      {
+         boundary_face_integs[k] -> AssembleFaceMatrix (*fe1, *fe2, *tr, elemmat);
+         elmat += elemmat;
+      }
+   }
+   else
+   {
+      int ndof = fe1->GetDof() * fes->GetVDim();
+      elmat.SetSize(ndof);
+      elmat = 0.0;
+   }
+}
+
 void BilinearForm::AssembleElementMatrix(
    int i, const DenseMatrix &elmat, int skip_zeros)
 {
@@ -1692,7 +1765,7 @@ void MixedBilinearForm::ConformingAssemble()
 }
 
 
-void MixedBilinearForm::ComputeElementMatrix(int i, DenseMatrix &elmat)
+void MixedBilinearForm::ComputeElementMatrix(int i, DenseMatrix &elmat) const
 {
    if (domain_integs.Size())
    {
@@ -1717,7 +1790,7 @@ void MixedBilinearForm::ComputeElementMatrix(int i, DenseMatrix &elmat)
    }
 }
 
-void MixedBilinearForm::ComputeBdrElementMatrix(int i, DenseMatrix &elmat)
+void MixedBilinearForm::ComputeBdrElementMatrix(int i, DenseMatrix &elmat) const
 {
    if (boundary_integs.Size())
    {
@@ -1742,6 +1815,94 @@ void MixedBilinearForm::ComputeBdrElementMatrix(int i, DenseMatrix &elmat)
    }
 }
 
+void MixedBilinearForm::ComputeTraceFaceMatrix(int i, DenseMatrix &elmat) const
+{
+   FaceElementTransformations *ftr;
+   Mesh *mesh = test_fes -> GetMesh();
+   ftr = mesh->GetFaceElementTransformations(i);
+   MFEM_ASSERT(ftr, "No associated face transformation.");
+
+   const FiniteElement *trial_face_fe, *test_fe1, *test_fe2;
+
+   trial_face_fe = trial_fes->GetFaceElement(i);
+   test_fe1 = test_fes->GetFE(ftr->Elem1No);
+   if (ftr->Elem2No >= 0)
+   {
+      test_fe2 = test_fes->GetFE(ftr->Elem2No);
+   }
+   else
+   {
+      // The test_fe2 object is really a dummy and not used on the
+      // boundaries, but we can't dereference a NULL pointer, and we don't
+      // want to actually make a fake element.
+      test_fe2 = test_fe1;
+   }
+
+   if (trace_face_integs.Size())
+   {
+      trace_face_integs[0]->AssembleFaceMatrix(*trial_face_fe, *test_fe1, *test_fe2,
+                                               *ftr, elmat);
+      for (int k = 1; k < trace_face_integs.Size(); k++)
+      {
+         trace_face_integs[k]->AssembleFaceMatrix(*trial_face_fe, *test_fe1, *test_fe2,
+                                                  *ftr, elemmat);
+         elmat += elemmat;
+      }
+   }
+   else
+   {
+      const int tr_face_dofs = trial_face_fe->GetDof() * trial_fes->GetVDim();
+      int te_dofs = test_fe1->GetDof() * test_fes->GetVDim();
+      if (ftr->Elem2No >= 0)
+      {
+         te_dofs += test_fe2->GetDof() * test_fes->GetVDim();
+      }
+
+      elmat.SetSize(te_dofs, tr_face_dofs);
+      elmat = 0.0;
+   }
+}
+
+void MixedBilinearForm::ComputeBdrTraceFaceMatrix(int i,
+                                                  DenseMatrix &elmat) const
+{
+   FaceElementTransformations *ftr;
+   Mesh *mesh = test_fes -> GetMesh();
+   ftr = mesh->GetBdrFaceTransformations(i);
+   MFEM_ASSERT(ftr, "No associated boundary face.");
+
+   const FiniteElement *trial_face_fe, *test_fe1, *test_fe2;
+   int iface = mesh->GetBdrElementFaceIndex(i);
+   trial_face_fe = trial_fes->GetFaceElement(iface);
+   test_fe1 = test_fes->GetFE(ftr->Elem1No);
+   // The test_fe2 object is really a dummy and not used on the
+   // boundaries, but we can't dereference a NULL pointer, and we don't
+   // want to actually make a fake element.
+   test_fe2 = test_fe1;
+
+   if (boundary_trace_face_integs.Size())
+   {
+      boundary_trace_face_integs[0]->AssembleFaceMatrix(*trial_face_fe, *test_fe1,
+                                                        *test_fe2,
+                                                        *ftr, elmat);
+      for (int k = 1; k < boundary_trace_face_integs.Size(); k++)
+      {
+         boundary_trace_face_integs[k]->AssembleFaceMatrix(*trial_face_fe, *test_fe1,
+                                                           *test_fe2,
+                                                           *ftr, elemmat);
+         elmat += elemmat;
+      }
+   }
+   else
+   {
+      const int tr_face_dofs = trial_face_fe->GetDof() * trial_fes->GetVDim();
+      int te_dofs = test_fe1->GetDof() * test_fes->GetVDim();
+
+      elmat.SetSize(te_dofs, tr_face_dofs);
+      elmat = 0.0;
+   }
+}
+
 void MixedBilinearForm::AssembleElementMatrix(
    int i, const DenseMatrix &elmat, int skip_zeros)
 {

fem/bilinearform.hpp

@@ -119,8 +119,8 @@ protected:
    Array<BilinearFormIntegrator*> boundary_face_integs;
    Array<Array<int>*> boundary_face_integs_marker; ///< Entries are not owned.
 
-   DenseMatrix elemmat;
-   Array<int>  vdofs;
+   mutable DenseMatrix elemmat;
+   mutable Array<int>  vdofs;
 
    DenseTensor *element_matrices; ///< Owned.
 
@@ -580,10 +580,18 @@ public:
        or the one stored internally by a prior call of ComputeElementMatrices()
        is returned when available.
    */
-   void ComputeElementMatrix(int i, DenseMatrix &elmat);
+   void ComputeElementMatrix(int i, DenseMatrix &elmat) const;
 
    /// Compute the boundary element matrix of the given boundary element
-   void ComputeBdrElementMatrix(int i, DenseMatrix &elmat);
+   /** @note The boundary attribute markers of the integrators are ignored. */
+   void ComputeBdrElementMatrix(int i, DenseMatrix &elmat) const;
+
+   /// Compute the face matrix of the given face element
+   void ComputeFaceMatrix(int i, DenseMatrix &elmat) const;
+
+   /// Compute the boundary face matrix of the given boundary element
+   /** @note The boundary attribute markers of the integrators are ignored. */
+   void ComputeBdrFaceMatrix(int i, DenseMatrix &elmat) const;
 
    /// Assemble the given element matrix
    /** The element matrix @a elmat is assembled for the element @a i, i.e.
@@ -771,8 +779,8 @@ protected:
    /// Entries are not owned.
    Array<Array<int>*> boundary_trace_face_integs_marker;
 
-   DenseMatrix elemmat;
-   Array<int>  trial_vdofs, test_vdofs;
+   mutable DenseMatrix elemmat;
+   mutable Array<int>  trial_vdofs, test_vdofs;
 
 private:
    /// Copy construction is not supported; body is undefined.
@@ -944,10 +952,18 @@ public:
    void ConformingAssemble();
 
    /// Compute the element matrix of the given element
-   void ComputeElementMatrix(int i, DenseMatrix &elmat);
+   void ComputeElementMatrix(int i, DenseMatrix &elmat) const;
 
    /// Compute the boundary element matrix of the given boundary element
-   void ComputeBdrElementMatrix(int i, DenseMatrix &elmat);
+   /** @note The boundary attribute markers of the integrators are ignored. */
+   void ComputeBdrElementMatrix(int i, DenseMatrix &elmat) const;
+
+   /// Compute the trace face matrix of the given face element
+   void ComputeTraceFaceMatrix(int i, DenseMatrix &elmat) const;
+
+   /// Compute the boundary trace face matrix of the given boundary element
+   /** @note The boundary attribute markers of the integrators are ignored. */
+   void ComputeBdrTraceFaceMatrix(int i, DenseMatrix &elmat) const;
 
    /// Assemble the given element matrix
    /** The element matrix @a elmat is assembled for the element @a i, i.e.

fem/bilininteg.cpp

@@ -1222,7 +1222,8 @@ real_t DiffusionIntegrator::ComputeFluxEnergy
    for (int i = 0; i < ir->GetNPoints(); i++)
    {
       const IntegrationPoint &ip = ir->IntPoint(i);
-      fluxelem.CalcShape(ip, shape);
+      Trans.SetIntPoint(&ip);
+      fluxelem.CalcPhysShape(Trans, shape);
 
       pointflux = 0.0;
       for (int k = 0; k < spaceDim; k++)
@@ -1233,7 +1234,6 @@ real_t DiffusionIntegrator::ComputeFluxEnergy
          }
       }
 
-      Trans.SetIntPoint(&ip);
       real_t w = Trans.Weight() * ip.weight;
 
       if (MQ)
@@ -1410,9 +1410,7 @@ void BoundaryMassIntegrator::AssembleFaceMatrix(
       // Set the integration point in the face and the neighboring element
       Trans.SetAllIntPoints(&ip);
 
-      // Access the neighboring element's integration point
-      const IntegrationPoint &eip = Trans.GetElement1IntPoint();
-      el1.CalcShape(eip, shape);
+      el1.CalcPhysShape(*Trans.Elem1, shape);
 
       w = Trans.Weight() * ip.weight;
       if (Q)
@@ -1582,9 +1580,9 @@ void VectorMassIntegrator::AssembleElementMatrix
    for (int s = 0; s < ir->GetNPoints(); s++)
    {
       const IntegrationPoint &ip = ir->IntPoint(s);
-      el.CalcShape(ip, shape);
-
       Trans.SetIntPoint (&ip);
+      el.CalcPhysShape(Trans, shape);
+
       norm = ip.weight * Trans.Weight();
 
       MultVVt(shape, partelmat);
@@ -1666,10 +1664,10 @@ void VectorMassIntegrator::AssembleElementMatrix2(
    for (int s = 0; s < ir->GetNPoints(); s++)
    {
       const IntegrationPoint &ip = ir->IntPoint(s);
-      trial_fe.CalcShape(ip, shape);
-      test_fe.CalcShape(ip, te_shape);
-
       Trans.SetIntPoint(&ip);
+      trial_fe.CalcPhysShape(Trans, shape);
+      test_fe.CalcPhysShape(Trans, te_shape);
+
       norm = ip.weight * Trans.Weight();
 
       MultVWt(te_shape, shape, partelmat);
@@ -1897,12 +1895,12 @@ void VectorFECurlIntegrator::AssembleElementMatrix2(
          if ( trial_fe.GetMapType() == mfem::FiniteElement::H_CURL )
          {
             trial_fe.CalcCurlShape(ip, curlshapeTrial_dFT);
-            test_fe.CalcShape(ip, shapeTest);
+            test_fe.CalcPhysShape(Trans, shapeTest);
          }
          else
          {
             test_fe.CalcCurlShape(ip, curlshapeTrial_dFT);
-            trial_fe.CalcShape(ip, shapeTest);
+            trial_fe.CalcPhysShape(Trans, shapeTest);
          }
       }
 
@@ -1925,6 +1923,89 @@ void VectorFECurlIntegrator::AssembleElementMatrix2(
    }
 }
 
+void VectorFEBoundaryFluxIntegrator::AssembleElementMatrix(
+   const FiniteElement &el, ElementTransformation &Tr,
+   DenseMatrix &elmat)
+{
+   int nd = el.GetDof();
+   real_t w;
+
+#ifdef MFEM_THREAD_SAFE
+   Vector shape;
+#endif
+   elmat.SetSize(nd);
+   shape.SetSize(nd);
+
+   const IntegrationRule *ir = IntRule;
+   if (ir == NULL)
+   {
+      int intorder = 2*el.GetOrder() + Tr.OrderW();  // <----------
+      ir = &IntRules.Get(el.GetGeomType(), intorder);
+   }
+
+   elmat = 0.0;
+   for (int i = 0; i < ir->GetNPoints(); i++)
+   {
+      const IntegrationPoint &ip = ir->IntPoint(i);
+      el.CalcShape(ip, shape);
+
+      Tr.SetIntPoint (&ip);
+      w = ip.weight / Tr.Weight();
+
+      if (Q)
+      {
+         w *= Q->Eval(Tr, ip);
+      }
+
+      AddMult_a_VVt(w, shape, elmat);
+   }
+}
+
+void VectorFEBoundaryFluxIntegrator::AssembleElementMatrix2(
+   const FiniteElement &trial_fe,
+   const FiniteElement &test_fe,
+   ElementTransformation &Tr,
+   DenseMatrix &elmat)
+{
+   int tr_nd = trial_fe.GetDof();
+   int te_nd = test_fe.GetDof();
+   real_t w;
+
+#ifdef MFEM_THREAD_SAFE
+   Vector shape, te_shape;
+#endif
+   elmat.SetSize(te_nd, tr_nd);
+   shape.SetSize(tr_nd);
+   te_shape.SetSize(te_nd);
+
+   const IntegrationRule *ir = IntRule;
+   if (ir == NULL)
+   {
+      int order = trial_fe.GetOrder() + test_fe.GetOrder() + Tr.OrderW();
+
+      ir = &IntRules.Get(trial_fe.GetGeomType(), order);
+   }
+
+   elmat = 0.0;
+   for (int i = 0; i < ir->GetNPoints(); i++)
+   {
+      const IntegrationPoint &ip = ir->IntPoint(i);
+      trial_fe.CalcShape(ip, shape);
+      test_fe.CalcShape(ip, te_shape);
+
+      Tr.SetIntPoint (&ip);
+      w = ip.weight / Tr.Weight();
+
+      if (Q)
+      {
+         w *= Q->Eval(Tr, ip);
+      }
+
+      te_shape *= w;
+      AddMultVWt(te_shape, shape, elmat);
+   }
+}
+
 void DerivativeIntegrator::AssembleElementMatrix2 (
    const FiniteElement &trial_fe,
    const FiniteElement &test_fe,
@@ -1981,7 +2062,7 @@ void DerivativeIntegrator::AssembleElementMatrix2 (
       det = Trans.Weight();
       Mult (dshape, invdfdx, dshapedxt);
 
-      test_fe.CalcShape(ip, shape);
+      test_fe.CalcPhysShape(Trans, shape);
 
       for (l = 0; l < trial_nd; l++)
       {
@@ -2566,7 +2647,7 @@ void VectorFEMassIntegrator::AssembleElementMatrix2(
          Trans.SetIntPoint (&ip);
 
          trial_fe.CalcVShape(Trans, trial_vshape);
-         test_fe.CalcShape(ip, shape);
+         test_fe.CalcPhysShape(Trans, shape);
 
          w = ip.weight * Trans.Weight();
          if (DQ)
@@ -2726,11 +2807,11 @@ void VectorDivergenceIntegrator::AssembleElementMatrix2(
    for (int i = 0; i < ir -> GetNPoints(); i++)
    {
       const IntegrationPoint &ip = ir->IntPoint(i);
+      Trans.SetIntPoint (&ip);
 
       trial_fe.CalcDShape (ip, dshape);
-      test_fe.CalcShape (ip, shape);
+      test_fe.CalcPhysShape (Trans, shape);
 
-      Trans.SetIntPoint (&ip);
       CalcAdjugate(Trans.Jacobian(), Jadj);
 
       Mult (dshape, Jadj, gshape);
@@ -3231,11 +3312,11 @@ real_t ElasticityIntegrator::ComputeFluxEnergy(const FiniteElement &fluxelem,
    for (int i = 0; i < ir->GetNPoints(); i++)
    {
       const IntegrationPoint &ip = ir->IntPoint(i);
-      fluxelem.CalcShape(ip, shape);
+      Trans.SetIntPoint(&ip);
+      fluxelem.CalcPhysShape(Trans, shape);
 
       flux_mat.MultTranspose(shape, pointstress);
 
-      Trans.SetIntPoint(&ip);
       real_t w = Trans.Weight() * ip.weight;
 
       M = mu->Eval(Trans, ip);
@@ -3342,7 +3423,7 @@ void DGTraceIntegrator::AssembleFaceMatrix(const FiniteElement &el1,
       const IntegrationPoint &eip1 = Trans.GetElement1IntPoint();
       const IntegrationPoint &eip2 = Trans.GetElement2IntPoint();
 
-      el1.CalcShape(eip1, shape1);
+      el1.CalcPhysShape(*Trans.Elem1, shape1);
 
       u->Eval(vu, *Trans.Elem1, eip1);
 
@@ -3389,7 +3470,7 @@ void DGTraceIntegrator::AssembleFaceMatrix(const FiniteElement &el1,
 
       if (ndof2)
       {
-         el2.CalcShape(eip2, shape2);
+         el2.CalcPhysShape(*Trans.Elem2, shape2);
 
          if (w != 0.0)
             for (int i = 0; i < ndof2; i++)
@@ -3939,19 +4020,14 @@ void TraceJumpIntegrator::AssembleFaceMatrix(
       // Set the integration point in the face and the neighboring elements
       Trans.SetAllIntPoints(&ip);
 
-      // Access the neighboring elements' integration points
-      // Note: eip2 will only contain valid data if Elem2 exists
-      const IntegrationPoint &eip1 = Trans.GetElement1IntPoint();
-      const IntegrationPoint &eip2 = Trans.GetElement2IntPoint();
-
       // Trace finite element shape function
       trial_face_fe.CalcShape(ip, face_shape);
       // Side 1 finite element shape function
-      test_fe1.CalcShape(eip1, shape1);
+      test_fe1.CalcPhysShape(*Trans.Elem1, shape1);
       if (ndof2)
       {
          // Side 2 finite element shape function
-         test_fe2.CalcShape(eip2, shape2);
+         test_fe2.CalcPhysShape(*Trans.Elem2, shape2);
       }
       w = ip.weight;
       if (trial_face_fe.GetMapType() == FiniteElement::VALUE)

fem/bilininteg.hpp

@@ -2612,6 +2612,25 @@ public:
                                        DenseMatrix &elmat);
 };
 
+/// Integrator for (Q u.n, v.n) for RT elements
+class VectorFEBoundaryFluxIntegrator : public BilinearFormIntegrator
+{
+   Coefficient *Q;
+#ifndef MFEM_THREAD_SAFE
+   Vector shape, te_shape;
+#endif
+public:
+   VectorFEBoundaryFluxIntegrator() { Q = NULL; }
+   VectorFEBoundaryFluxIntegrator(Coefficient &q) { Q = &q; }
+   void AssembleElementMatrix(const FiniteElement &el,
+                              ElementTransformation &Trans,
+                              DenseMatrix &elmat) override;
+   void AssembleElementMatrix2(const FiniteElement &trial_fe,
+                               const FiniteElement &test_fe,
+                               ElementTransformation &Trans,
+                               DenseMatrix &elmat) override;
+};
+
 /// Class for integrating $ (Q \partial_i(u), v) $ where $u$ and $v$ are scalars
 class DerivativeIntegrator : public BilinearFormIntegrator
 {

fem/datacollection.cpp

@@ -943,6 +943,7 @@ void ParaViewDataCollection::Save()
             pvtu_out << "<PDataArray type=\"" << GetDataTypeString()
                      << "\" Name=\"" << field_it.first
                      << "\" NumberOfComponents=\"" << vec_dim << "\" "
+                     << VTKComponentLabels(vec_dim) << " "
                      << "format=\"" << GetDataFormatString() << "\" />\n";
          }
          pvtu_out << "</PPointData>\n";
@@ -977,6 +978,7 @@ void ParaViewDataCollection::Save()
          pvtu_out << "<PDataArray type=\"" << GetDataTypeString()
                   << "\" Name=\"" << q_field_name
                   << "\" NumberOfComponents=\"" << vec_dim << "\" "
+                  << VTKComponentLabels(vec_dim) << " "
                   << "format=\"" << GetDataFormatString() << "\" />\n";
          pvtu_out << "</PPointData>\n";
          WritePVTUFooter(pvtu_out, q_field_name);
@@ -1069,8 +1071,9 @@ void ParaViewDataCollection::SaveGFieldVTU(std::ostream &os, int ref_,
    int vec_dim = it->second->VectorDim();
    os << "<DataArray type=\"" << GetDataTypeString()
       << "\" Name=\"" << it->first
-      << "\" NumberOfComponents=\"" << vec_dim << "\""
-      << " format=\"" << GetDataFormatString() << "\" >" << '\n';
+      << "\" NumberOfComponents=\"" << vec_dim << "\" "
+      << VTKComponentLabels(vec_dim) << " "
+      << "format=\"" << GetDataFormatString() << "\" >" << '\n';
    if (vec_dim == 1)
    {
       // scalar data

fem/eltrans.hpp

@@ -52,6 +52,15 @@ protected:
    const DenseMatrix &EvalTransAdjugateJ();
    const DenseMatrix &EvalInverseJ();
 
+   /// @name Tolerance used for point comparisons
+   ///@{
+#ifdef MFEM_USE_DOUBLE
+   static constexpr real_t tol_0 = 1e-15;
+#elif defined(MFEM_USE_SINGLE)
+   static constexpr real_t tol_0 = 1e-7;
+#endif
+   ///@}
+
 public:
 
    /** This enumeration declares the values stored in
@@ -176,7 +185,7 @@ public:
        returned. This method is not 100 percent reliable for non-linear
        transformations. */
    virtual int TransformBack(const Vector &pt, IntegrationPoint &ip,
-                             const real_t phys_tol = 1e-15) = 0;
+                             const real_t phys_tol = tol_0) = 0;
 
    virtual ~ElementTransformation() { }
 };
@@ -281,9 +290,15 @@ public:
         rel_qpts_order(-1),
         solver_type(NewtonElementProject),
         max_iter(16),
+#ifdef MFEM_USE_DOUBLE
         ref_tol(1e-15),
         phys_rtol(1e-15),
         ip_tol(1e-8),
+#elif defined(MFEM_USE_SINGLE)
+        ref_tol(1e-7),
+        phys_rtol(1e-7),
+        ip_tol(1e-4),
+#endif
         print_level(-1)
    { }
 
@@ -449,7 +464,7 @@ public:
        returned. This method is not 100 percent reliable for non-linear
        transformations. */
    virtual int TransformBack(const Vector & v, IntegrationPoint & ip,
-                             const real_t phys_rel_tol = 1e-15)
+                             const real_t phys_rel_tol = tol_0)
    {
       InverseElementTransformation inv_tr(this);
       inv_tr.SetPhysicalRelTol(phys_rel_tol);

fem/fe/fe_base.hpp

@@ -394,7 +394,32 @@ public:
    /// Get a const reference to the nodes of the element
    const IntegrationRule & GetNodes() const { return Nodes; }
 
-   // virtual functions for finite elements on vector spaces
+   /** @brief Evaluate the Hessians of all shape functions of a scalar finite
+       element in reference space at the given point @a ip. */
+   /** Each row of the result DenseMatrix @a Hessian contains upper triangular
+       part of the Hessian of one shape function.
+       The order in 2D is {u_xx, u_xy, u_yy}.
+       The size (#dof x (#dim (#dim+1)/2) of @a Hessian must be set in advance.*/
+   virtual void CalcHessian(const IntegrationPoint &ip,
+                            DenseMatrix &Hessian) const;
+
+   /** @brief Evaluate the Hessian of all shape functions of a scalar finite
+       element in physical space at the given point @a ip. */
+   /** The size (#dof, #dim*(#dim+1)/2) of @a Hessian must be set in advance. */
+   void CalcPhysHessian(ElementTransformation &Trans,
+                        DenseMatrix& Hessian) const;
+
+   /** @brief Evaluate the Laplacian of all shape functions of a scalar finite
+       element in physical space at the given point @a ip. */
+   /** The size (#dof) of @a Laplacian must be set in advance. */
+   void CalcPhysLaplacian(ElementTransformation &Trans,
+                          Vector& Laplacian) const;
+
+   /** @brief Evaluate the Laplacian of all shape functions of a scalar finite
+       element in physical space at the given point @a ip. */
+   /** The size (#dof) of @a Laplacian must be set in advance. */
+   void CalcPhysLinLaplacian(ElementTransformation &Trans,
+                             Vector& Laplacian) const;
 
    /** @brief Evaluate the values of all shape functions of a *vector* finite
        element in reference space at the given point @a ip. */
@@ -454,30 +479,6 @@ public:
    */
    virtual void GetFaceDofs(int face, int **dofs, int *ndofs) const;
 
-   /** @brief Evaluate the Hessians of all shape functions of a scalar finite
-       element in reference space at the given point @a ip. */
-   /** Each row of the result DenseMatrix @a Hessian contains upper triangular
-       part of the Hessian of one shape function.
-       The order in 2D is {u_xx, u_xy, u_yy}.
-       The size (#dof x (#dim (#dim+1)/2) of @a Hessian must be set in advance.*/
-   virtual void CalcHessian(const IntegrationPoint &ip,
-                            DenseMatrix &Hessian) const;
-
-   /** @brief Evaluate the Hessian of all shape functions of a scalar finite
-       element in reference space at the given point @a ip. */
-   /** The size (#dof, #dim*(#dim+1)/2) of @a Hessian must be set in advance. */
-   virtual void CalcPhysHessian(ElementTransformation &Trans,
-                                DenseMatrix& Hessian) const;
-
-   /** @brief Evaluate the Laplacian of all shape functions of a scalar finite
-       element in reference space at the given point @a ip. */
-   /** The size (#dof) of @a Laplacian must be set in advance. */
-   virtual void CalcPhysLaplacian(ElementTransformation &Trans,
-                                  Vector& Laplacian) const;
-
-   virtual void CalcPhysLinLaplacian(ElementTransformation &Trans,
-                                     Vector& Laplacian) const;
-
    /** @brief Return the local interpolation matrix @a I (Dof x Dof) where the
        fine element is the image of the base geometry under the given
        transformation. */

fem/fe/fe_nurbs.cpp

@@ -398,8 +398,621 @@ void NURBS3DFiniteElement::CalcHessian (const IntegrationPoint &ip,
       hessian(o,5) = hessian(o,5)*sum
                      - 2*du(o,1)*sum*dsum[1]
                      + u[o]*sum*(2*dsum[1]*dsum[1] - d2sum[5]);
-
    }
 }
 
+
+void NURBS_HDiv2DFiniteElement::SetOrder() const
+{
+   orders[0] = kv[0]->GetOrder();
+   orders[1] = kv[1]->GetOrder();
+
+   if (kv1[0]) { delete kv1[0]; }
+   if (kv1[1]) { delete kv1[1]; }
+
+   kv1[0] = kv[0]->DegreeElevate(1);
+   kv1[1] = kv[1]->DegreeElevate(1);
+
+   shape_x.SetSize(orders[0]+1);
+   shape_y.SetSize(orders[1]+1);
+
+   dshape_x.SetSize(orders[0]+1);
+   dshape_y.SetSize(orders[1]+1);
+
+   d2shape_x.SetSize(orders[0]+1);
+   d2shape_y.SetSize(orders[1]+1);
+
+   shape1_x.SetSize(orders[0]+2);
+   shape1_y.SetSize(orders[1]+2);
+
+   dshape1_x.SetSize(orders[0]+2);
+   dshape1_y.SetSize(orders[1]+2);
+
+   d2shape1_x.SetSize(orders[0]+2);
+   d2shape1_y.SetSize(orders[1]+2);
+
+   order = max(orders[0]+1, orders[1]+1);
+   dof = (orders[0] + 2)*(orders[1] + 1)
+         + (orders[1] + 1)*(orders[1] + 2);
+   u.SetSize(dof);
+   du.SetSize(dof);
+   weights.SetSize(dof);
+}
+
+void NURBS_HDiv2DFiniteElement::CalcVShape(const IntegrationPoint &ip,
+                                           DenseMatrix &shape) const
+{
+   kv[0]->CalcShape(shape_x, ijk[0], ip.x);
+   kv[1]->CalcShape(shape_y, ijk[1], ip.y);
+
+   kv1[0]->CalcShape(shape1_x, ijk[0], ip.x);
+   kv1[1]->CalcShape(shape1_y, ijk[1], ip.y);
+
+   int o = 0;
+   for (int j = 0; j <= orders[1]; j++)
+   {
+      const real_t sy = shape_y(j);
+      for (int i = 0; i <= orders[0]+1; i++, o++)
+      {
+         shape(o,0) = shape1_x(i)*sy;
+         shape(o,1) = 0.0;
+      }
+   }
+
+   for (int j = 0; j <= orders[1]+1; j++)
+   {
+      const real_t sy1 = shape1_y(j);
+      for (int i = 0; i <= orders[0]; i++, o++)
+      {
+         shape(o,0) = 0.0;
+         shape(o,1) = shape_x(i)*sy1;
+      }
+   }
+}
+
+void NURBS_HDiv2DFiniteElement::CalcVShape(ElementTransformation &Trans,
+                                           DenseMatrix &shape) const
+{
+   CalcVShape(Trans.GetIntPoint(), shape);
+   const DenseMatrix & J = Trans.Jacobian();
+   MFEM_ASSERT(J.Width() == 2 && J.Height() == 2,
+               "NURBS_HDiv2DFiniteElement cannot be embedded in "
+               "3 dimensional spaces");
+   for (int i=0; i<dof; i++)
+   {
+      real_t sx = shape(i, 0);
+      real_t sy = shape(i, 1);
+      shape(i, 0) = sx * J(0, 0) + sy * J(0, 1);
+      shape(i, 1) = sx * J(1, 0) + sy * J(1, 1);
+   }
+   shape *= (1.0 / Trans.Weight());
+}
+
+void NURBS_HDiv2DFiniteElement::CalcDivShape(const IntegrationPoint &ip,
+                                             Vector &divshape) const
+{
+   kv[0]->CalcShape ( shape_x, ijk[0], ip.x);
+   kv[1]->CalcShape ( shape_y, ijk[1], ip.y);
+
+   kv1[0]->CalcDShape(dshape1_x, ijk[0], ip.x);
+   kv1[1]->CalcDShape(dshape1_y, ijk[1], ip.y);
+
+   int o = 0;
+   for (int j = 0; j <= orders[1]; j++)
+   {
+      const real_t sy = shape_y(j);
+      for (int i = 0; i <= orders[0]+1; i++, o++)
+      {
+         divshape(o) = dshape1_x(i)*sy;
+      }
+   }
+
+   for (int j = 0; j <= orders[1]+1; j++)
+   {
+      const real_t dsy1 = dshape1_y(j);
+      for (int i = 0; i <= orders[0]; i++, o++)
+      {
+         divshape(o) = shape_x(i)*dsy1;
+      }
+   }
+}
+
+NURBS_HDiv2DFiniteElement::~NURBS_HDiv2DFiniteElement()
+{
+   if (kv1[0]) { delete kv1[0]; }
+   if (kv1[1]) { delete kv1[1]; }
+}
+
+
+void NURBS_HDiv3DFiniteElement::SetOrder() const
+{
+   orders[0] = kv[0]->GetOrder();
+   orders[1] = kv[1]->GetOrder();
+   orders[2] = kv[2]->GetOrder();
+
+   if (kv1[0]) { delete kv1[0]; }
+   if (kv1[1]) { delete kv1[1]; }
+   if (kv1[2]) { delete kv1[2]; }
+
+   kv1[0] = kv[0]->DegreeElevate(1);
+   kv1[1] = kv[1]->DegreeElevate(1);
+   kv1[2] = kv[2]->DegreeElevate(1);
+
+   shape_x.SetSize(orders[0]+1);
+   shape_y.SetSize(orders[1]+1);
+   shape_z.SetSize(orders[2]+1);
+
+   dshape_x.SetSize(orders[0]+1);
+   dshape_y.SetSize(orders[1]+1);
+   dshape_z.SetSize(orders[2]+1);
+
+   d2shape_x.SetSize(orders[0]+1);
+   d2shape_y.SetSize(orders[1]+1);
+   d2shape_z.SetSize(orders[2]+1);
+
+   shape1_x.SetSize(orders[0]+2);
+   shape1_y.SetSize(orders[1]+2);
+   shape1_z.SetSize(orders[2]+2);
+
+   dshape1_x.SetSize(orders[0]+2);
+   dshape1_y.SetSize(orders[1]+2);
+   dshape1_z.SetSize(orders[2]+2);
+
+   d2shape1_x.SetSize(orders[0]+2);
+   d2shape1_y.SetSize(orders[1]+2);
+   d2shape1_z.SetSize(orders[2]+2);
+
+   order = max(orders[0]+1, max( orders[1]+1, orders[2]+1));
+   dof = (orders[0] + 2)*(orders[1] + 1)*(orders[2] + 1) +
+         (orders[0] + 1)*(orders[1] + 2)*(orders[2] + 1) +
+         (orders[0] + 1)*(orders[1] + 1)*(orders[2] + 2);
+   u.SetSize(dof);
+   du.SetSize(dof);
+   weights.SetSize(dof);
+}
+
+void NURBS_HDiv3DFiniteElement::CalcVShape(const IntegrationPoint &ip,
+                                           DenseMatrix &shape) const
+{
+   kv[0]->CalcShape(shape_x, ijk[0], ip.x);
+   kv[1]->CalcShape(shape_y, ijk[1], ip.y);
+   kv[2]->CalcShape(shape_z, ijk[2], ip.z);
+
+   kv1[0]->CalcShape(shape1_x, ijk[0], ip.x);
+   kv1[1]->CalcShape(shape1_y, ijk[1], ip.y);
+   kv1[2]->CalcShape(shape1_z, ijk[2], ip.z);
+
+   shape = 0.0;
+   int o = 0;
+   for (int k = 0; k <= orders[2]; k++)
+   {
+      const real_t sz = shape_z(k);
+      for (int j = 0; j <= orders[1]; j++)
+      {
+         const real_t sy_sz = shape_y(j)*sz;
+         for (int i = 0; i <= orders[0]+1; i++, o++)
+         {
+            shape(o,0) = shape1_x(i)*sy_sz;
+         }
+      }
+   }
+
+   for (int  k = 0; k <= orders[2]; k++)
+   {
+      const real_t sz = shape_z(k);
+      for (int j = 0; j <= orders[1]+1; j++)
+      {
+         const real_t sy1_sz = shape1_y(j)*sz;
+         for (int i = 0; i <= orders[0]; i++, o++)
+         {
+            shape(o,1) = shape_x(i)*sy1_sz;
+         }
+      }
+   }
+
+   for (int  k = 0; k <= orders[2]+1; k++)
+   {
+      const real_t sz1 = shape1_z(k);
+      for (int j = 0; j <= orders[1]; j++)
+      {
+         const real_t sy_sz1 = shape_y(j)*sz1;
+         for (int i = 0; i <= orders[0]; i++, o++)
+         {
+            shape(o,2) = shape_x(i)*sy_sz1;
+         }
+      }
+   }
+}
+
+void NURBS_HDiv3DFiniteElement::CalcVShape(ElementTransformation &Trans,
+                                           DenseMatrix &shape) const
+{
+   CalcVShape(Trans.GetIntPoint(), shape);
+   const DenseMatrix & J = Trans.Jacobian();
+   MFEM_ASSERT(J.Width() == 3 && J.Height() == 3,
+               "RT_R2D_FiniteElement cannot be embedded in "
+               "3 dimensional spaces");
+   for (int i=0; i<dof; i++)
+   {
+      real_t sx = shape(i, 0);
+      real_t sy = shape(i, 1);
+      real_t sz = shape(i, 2);
+      shape(i, 0) = sx * J(0, 0) + sy * J(0, 1) + sz * J(0, 2);
+      shape(i, 1) = sx * J(1, 0) + sy * J(1, 1) + sz * J(1, 2);
+      shape(i, 2) = sx * J(2, 0) + sy * J(2, 1) + sz * J(2, 2);
+   }
+   shape *= (1.0 / Trans.Weight());
+}
+
+void NURBS_HDiv3DFiniteElement::CalcDivShape(const IntegrationPoint &ip,
+                                             Vector &divshape) const
+{
+   kv[0]->CalcShape ( shape_x, ijk[0], ip.x);
+   kv[1]->CalcShape ( shape_y, ijk[1], ip.y);
+   kv[2]->CalcShape ( shape_z, ijk[2], ip.z);
+
+   kv1[0]->CalcDShape(dshape1_x, ijk[0], ip.x);
+   kv1[1]->CalcDShape(dshape1_y, ijk[1], ip.y);
+   kv1[2]->CalcDShape(dshape1_z, ijk[2], ip.z);
+
+   int o = 0;
+   for (int  k = 0; k <= orders[2]; k++)
+   {
+      const real_t sz = shape_z(k);
+      for (int j = 0; j <= orders[1]; j++)
+      {
+         const real_t sy_sz = shape_y(j)*sz;
+         for (int i = 0; i <= orders[0]+1; i++, o++)
+         {
+            divshape(o) = dshape1_x(i)*sy_sz;
+         }
+      }
+   }
+
+   for (int  k = 0; k <= orders[2]; k++)
+   {
+      const real_t sz = shape_z(k);
+      for (int j = 0; j <= orders[1]+1; j++)
+      {
+         const real_t dy1_sz = dshape1_y(j)*sz;
+         for (int i = 0; i <= orders[0]; i++, o++)
+         {
+            divshape(o) = shape_x(i)*dy1_sz;
+         }
+      }
+   }
+
+   for (int  k = 0; k <= orders[2]+1; k++)
+   {
+      const real_t dz1 = dshape1_z(k);
+      for (int j = 0; j <= orders[1]; j++)
+      {
+         const real_t sy_dz1 = shape_y(j)*dz1;
+         for (int i = 0; i <= orders[0]; i++, o++)
+         {
+            divshape(o) = shape_x(i)*sy_dz1;
+         }
+      }
+   }
+}
+
+NURBS_HDiv3DFiniteElement::~NURBS_HDiv3DFiniteElement()
+{
+   if (kv1[0]) { delete kv1[0]; }
+   if (kv1[1]) { delete kv1[1]; }
+   if (kv1[2]) { delete kv1[2]; }
+}
+
+void NURBS_HCurl2DFiniteElement::SetOrder() const
+{
+   orders[0] = kv[0]->GetOrder();
+   orders[1] = kv[1]->GetOrder();
+
+   if (kv1[0]) { delete kv1[0]; }
+   if (kv1[1]) { delete kv1[1]; }
+
+   kv1[0] = kv[0]->DegreeElevate(1);
+   kv1[1] = kv[1]->DegreeElevate(1);
+
+   shape_x.SetSize(orders[0]+1);
+   shape_y.SetSize(orders[1]+1);
+
+   dshape_x.SetSize(orders[0]+1);
+   dshape_y.SetSize(orders[1]+1);
+
+   d2shape_x.SetSize(orders[0]+1);
+   d2shape_y.SetSize(orders[1]+1);
+
+   shape1_x.SetSize(orders[0]+2);
+   shape1_y.SetSize(orders[1]+2);
+
+   dshape1_x.SetSize(orders[0]+2);
+   dshape1_y.SetSize(orders[1]+2);
+
+   d2shape1_x.SetSize(orders[0]+2);
+   d2shape1_y.SetSize(orders[1]+2);
+
+   order = max(orders[0]+1, orders[1]+1);
+   dof = (orders[0] + 1)*(orders[1] + 2)
+         + (orders[1] + 2)*(orders[1] + 1);
+   u.SetSize(dof);
+   du.SetSize(dof);
+   weights.SetSize(dof);
+}
+
+void NURBS_HCurl2DFiniteElement::CalcVShape(const IntegrationPoint &ip,
+                                            DenseMatrix &shape) const
+{
+   kv[0]->CalcShape(shape_x, ijk[0], ip.x);
+   kv[1]->CalcShape(shape_y, ijk[1], ip.y);
+
+   kv1[0]->CalcShape(shape1_x, ijk[0], ip.x);
+   kv1[1]->CalcShape(shape1_y, ijk[1], ip.y);
+
+   int o = 0;
+   for (int j = 0; j <= orders[1]+1; j++)
+   {
+      const real_t sy1 = shape1_y(j);
+      for (int i = 0; i <= orders[0]; i++, o++)
+      {
+         shape(o,0) = shape_x(i)*sy1;
+         shape(o,1) = 0.0;
+      }
+   }
+
+   for (int j = 0; j <= orders[1]; j++)
+   {
+      const real_t sy = shape_y(j);
+      for (int i = 0; i <= orders[0]+1; i++, o++)
+      {
+         shape(o,0) = 0.0;
+         shape(o,1) = shape1_x(i)*sy;
+      }
+   }
+}
+
+void NURBS_HCurl2DFiniteElement::CalcVShape(ElementTransformation &Trans,
+                                            DenseMatrix &shape) const
+{
+   CalcVShape(Trans.GetIntPoint(), shape);
+   const DenseMatrix & JI = Trans.InverseJacobian();
+   MFEM_ASSERT(JI.Width() == 2 && JI.Height() == 2,
+               "NURBS_HCurl2DFiniteElement cannot be embedded in "
+               "3 dimensional spaces");
+   for (int i=0; i<dof; i++)
+   {
+      real_t sx = shape(i, 0);
+      real_t sy = shape(i, 1);
+      shape(i, 0) = sx * JI(0, 0) + sy * JI(1, 0);
+      shape(i, 1) = sx * JI(0, 1) + sy * JI(1, 1);
+   }
+}
+
+void NURBS_HCurl2DFiniteElement::CalcCurlShape(const IntegrationPoint &ip,
+                                               DenseMatrix &curl_shape) const
+{
+   kv[0]->CalcShape ( shape_x, ijk[0], ip.x);
+   kv[1]->CalcShape ( shape_y, ijk[1], ip.y);
+
+   kv1[0]->CalcDShape(dshape1_x, ijk[0], ip.x);
+   kv1[1]->CalcDShape(dshape1_y, ijk[1], ip.y);
+
+   int o = 0;
+   for (int j = 0; j <= orders[1]+1; j++)
+   {
+      const real_t dsy1 = dshape1_y(j);
+      for (int i = 0; i <= orders[0]; i++, o++)
+      {
+         curl_shape(o,0) = -shape_x(i)*dsy1;
+      }
+   }
+
+   for (int j = 0; j <= orders[1]; j++)
+   {
+      const real_t sy = shape_y(j);
+      for (int i = 0; i <= orders[0]+1; i++, o++)
+      {
+         curl_shape(o,0) = dshape1_x(i)*sy;
+      }
+   }
+}
+
+NURBS_HCurl2DFiniteElement::~NURBS_HCurl2DFiniteElement()
+{
+   if (kv1[0]) { delete kv1[0]; }
+   if (kv1[1]) { delete kv1[1]; }
+}
+
+
+void NURBS_HCurl3DFiniteElement::SetOrder() const
+{
+   orders[0] = kv[0]->GetOrder();
+   orders[1] = kv[1]->GetOrder();
+   orders[2] = kv[2]->GetOrder();
+
+   if (kv1[0]) { delete kv1[0]; }
+   if (kv1[1]) { delete kv1[1]; }
+   if (kv1[2]) { delete kv1[2]; }
+
+   kv1[0] = kv[0]->DegreeElevate(1);
+   kv1[1] = kv[1]->DegreeElevate(1);
+   kv1[2] = kv[2]->DegreeElevate(1);
+
+   shape_x.SetSize(orders[0]+1);
+   shape_y.SetSize(orders[1]+1);
+   shape_z.SetSize(orders[2]+1);
+
+   dshape_x.SetSize(orders[0]+1);
+   dshape_y.SetSize(orders[1]+1);
+   dshape_z.SetSize(orders[2]+1);
+
+   d2shape_x.SetSize(orders[0]+1);
+   d2shape_y.SetSize(orders[1]+1);
+   d2shape_z.SetSize(orders[2]+1);
+
+   shape1_x.SetSize(orders[0]+2);
+   shape1_y.SetSize(orders[1]+2);
+   shape1_z.SetSize(orders[2]+2);
+
+   dshape1_x.SetSize(orders[0]+2);
+   dshape1_y.SetSize(orders[1]+2);
+   dshape1_z.SetSize(orders[2]+2);
+
+   d2shape1_x.SetSize(orders[0]+2);
+   d2shape1_y.SetSize(orders[1]+2);
+   d2shape1_z.SetSize(orders[2]+2);
+
+   order = max(orders[0]+1, max( orders[1]+1, orders[2]+1));
+   dof = (orders[0] + 1)*(orders[1] + 2)*(orders[2] + 2) +
+         (orders[0] + 2)*(orders[1] + 1)*(orders[2] + 2) +
+         (orders[0] + 2)*(orders[1] + 2)*(orders[2] + 1);
+   u.SetSize(dof);
+   du.SetSize(dof);
+   weights.SetSize(dof);
+}
+
+void NURBS_HCurl3DFiniteElement::CalcVShape(const IntegrationPoint &ip,
+                                            DenseMatrix &shape) const
+{
+   kv[0]->CalcShape(shape_x, ijk[0], ip.x);
+   kv[1]->CalcShape(shape_y, ijk[1], ip.y);
+   kv[2]->CalcShape(shape_z, ijk[2], ip.z);
+
+   kv1[0]->CalcShape(shape1_x, ijk[0], ip.x);
+   kv1[1]->CalcShape(shape1_y, ijk[1], ip.y);
+   kv1[2]->CalcShape(shape1_z, ijk[2], ip.z);
+
+   shape = 0.0;
+   int o = 0;
+   for (int  k = 0; k <= orders[2]+1; k++)
+   {
+      const real_t sz1 = shape1_z(k);
+      for (int j = 0; j <= orders[1]+1; j++)
+      {
+         const real_t sy1_sz1 = shape1_y(j)*sz1;
+         for (int i = 0; i <= orders[0]; i++, o++)
+         {
+            shape(o,0) = shape_x(i)*sy1_sz1;
+         }
+      }
+   }
+
+   for (int  k = 0; k <= orders[2]+1; k++)
+   {
+      const real_t sz1 = shape1_z(k);
+      for (int j = 0; j <= orders[1]; j++)
+      {
+         const real_t sy_sz1 = shape_y(j)*sz1;
+         for (int i = 0; i <= orders[0]+1; i++, o++)
+         {
+            shape(o,1) = shape1_x(i)*sy_sz1;
+         }
+      }
+   }
+
+   for (int  k = 0; k <= orders[2]; k++)
+   {
+      const real_t sz = shape_z(k);
+      for (int j = 0; j <= orders[1]+1; j++)
+      {
+         const real_t sy1_sz = shape1_y(j)*sz;
+         for (int i = 0; i <= orders[0]+1; i++, o++)
+         {
+            shape(o,2) = shape1_x(i)*sy1_sz;
+         }
+      }
+   }
+}
+
+void NURBS_HCurl3DFiniteElement::CalcVShape(ElementTransformation &Trans,
+                                            DenseMatrix &shape) const
+{
+   CalcVShape(Trans.GetIntPoint(), shape);
+   const DenseMatrix & JI = Trans.InverseJacobian();
+   MFEM_ASSERT(JI.Width() == 3 && JI.Height() == 3,
+               "NURBS_HCurl3DFiniteElement must be in a"
+               "3 dimensional spaces");
+   for (int i=0; i<dof; i++)
+   {
+      real_t sx = shape(i, 0);
+      real_t sy = shape(i, 1);
+      real_t sz = shape(i, 2);
+      shape(i, 0) = sx * JI(0, 0) + sy * JI(1, 0) + sz * JI(2, 0);
+      shape(i, 1) = sx * JI(0, 1) + sy * JI(1, 1) + sz * JI(2, 1);
+      shape(i, 2) = sx * JI(0, 2) + sy * JI(1, 2) + sz * JI(2, 2);
+   }
+}
+
+void NURBS_HCurl3DFiniteElement::CalcCurlShape(const IntegrationPoint &ip,
+                                               DenseMatrix &curl_shape) const
+{
+   kv[0]->CalcShape ( shape_x, ijk[0], ip.x);
+   kv[1]->CalcShape ( shape_y, ijk[1], ip.y);
+   kv[2]->CalcShape ( shape_z, ijk[2], ip.z);
+
+   kv1[0]->CalcShape(shape1_x, ijk[0], ip.x);
+   kv1[1]->CalcShape(shape1_y, ijk[1], ip.y);
+   kv1[2]->CalcShape(shape1_z, ijk[2], ip.z);
+
+   kv1[0]->CalcDShape(dshape1_x, ijk[0], ip.x);
+   kv1[1]->CalcDShape(dshape1_y, ijk[1], ip.y);
+   kv1[2]->CalcDShape(dshape1_z, ijk[2], ip.z);
+
+   int o = 0;
+   for (int  k = 0; k <= orders[2]+1; k++)
+   {
+      const real_t sz1 = shape1_z(k), dsz1 = dshape1_z(k);
+      for (int j = 0; j <= orders[1]+1; j++)
+      {
+         const real_t sy1_dsz1 = shape1_y(j)*dsz1,
+                      dsy1_sz1 = dshape1_y(j)*sz1;
+         for (int i = 0; i <= orders[0]; i++, o++)
+         {
+            curl_shape(o,0) = 0.0;
+            curl_shape(o,1) =  shape_x(i)*sy1_dsz1;
+            curl_shape(o,2) = -shape_x(i)*dsy1_sz1;
+         }
+      }
+   }
+
+   for (int  k = 0; k <= orders[2]+1; k++)
+   {
+      const real_t sz1 = shape1_z(k), dsz1 = dshape1_z(k);
+      for (int j = 0; j <= orders[1]; j++)
+      {
+         const real_t sy_dsz1 = shape_y(j)*dsz1,
+                      sy_sz1 = shape_y(j)*sz1;
+         for (int i = 0; i <= orders[0]+1; i++, o++)
+         {
+            curl_shape(o,0) = -shape1_x(i)*sy_dsz1;
+            curl_shape(o,1) = 0.0;
+            curl_shape(o,2) = dshape1_x(i)*sy_sz1;
+         }
+      }
+   }
+
+   for (int  k = 0; k <= orders[2]; k++)
+   {
+      const real_t sz = shape_z(k);
+      for (int j = 0; j <= orders[1]+1; j++)
+      {
+         const real_t sy1_sz = shape1_y(j)*sz,
+                      dsy1_sz = dshape1_y(j)*sz;
+         for (int i = 0; i <= orders[0]+1; i++, o++)
+         {
+            curl_shape(o,0) = shape1_x(i)*dsy1_sz;
+            curl_shape(o,1) = -dshape1_x(i)*sy1_sz;
+            curl_shape(o,2) = 0.0;
+         }
+      }
+   }
+}
+
+NURBS_HCurl3DFiniteElement::~NURBS_HCurl3DFiniteElement()
+{
+   if (kv1[0]) { delete kv1[0]; }
+   if (kv1[1]) { delete kv1[1]; }
+   if (kv1[2]) { delete kv1[2]; }
+}
+
 }

fem/fe/fe_nurbs.hpp

@@ -20,7 +20,7 @@ namespace mfem
 class KnotVector;
 
 /// An arbitrary order and dimension NURBS element
-class NURBSFiniteElement : public ScalarFiniteElement
+class NURBSFiniteElement
 {
 protected:
    mutable Array <const KnotVector*> kv;
@@ -30,31 +30,34 @@ protected:
 
 public:
    /** @brief Construct NURBSFiniteElement with given
-       @param D    Reference space dimension
-       @param G    Geometry type (of type Geometry::Type)
-       @param Do   Number of degrees of freedom in the FiniteElement
-       @param O    Order/degree of the FiniteElement
-       @param F    FunctionSpace type of the FiniteElement
+       @param dim    Reference space dimension
     */
-   NURBSFiniteElement(int D, Geometry::Type G, int Do, int O, int F)
-      : ScalarFiniteElement(D, G, Do, O, F)
+   NURBSFiniteElement(int dim)
    {
       ijk = NULL;
       patch = elem = -1;
       kv.SetSize(dim);
-      weights.SetSize(dof);
-      weights = 1.0;
    }
 
+   /// Resets the patch and element data stored in the element
    void                 Reset      ()         const { patch = elem = -1; }
+   /// Set which IJK in patch should be evaluated
    void                 SetIJK     (const int *IJK) const { ijk = IJK; }
+   /// Get which patch is currently considered
    int                  GetPatch   ()         const { return patch; }
+   /// Set which patch should be evaluated
    void                 SetPatch   (int p)    const { patch = p; }
+   /// Set which elemenet should be evaluated
    int                  GetElement ()         const { return elem; }
+   /// Get which element is currently considered
    void                 SetElement (int e)    const { elem = e; }
+   /// Get the KnotVectors
    Array <const KnotVector*> &KnotVectors()   const { return kv; }
+   /// Get the Weights
    Vector              &Weights    ()         const { return weights; }
-   /// Update the NURBSFiniteElement according to the currently set knot vectors
+   /// Update the polynomial order according to the currently set knotvectors
+   /// Resizes all internal data members to have the correct size
+   /// related to the polynomial order
    virtual void         SetOrder   ()         const { }
 
    /// Returns the indices (i,j) in 2D or (i,j,k) in 3D of this element in the
@@ -64,7 +67,8 @@ public:
 
 
 /// An arbitrary order 1D NURBS element on a segment
-class NURBS1DFiniteElement : public NURBSFiniteElement
+class NURBS1DFiniteElement : public ScalarFiniteElement,
+   public NURBSFiniteElement
 {
 protected:
    mutable Vector shape_x;
@@ -72,7 +76,8 @@ protected:
 public:
    /// Construct the NURBS1DFiniteElement of order @a p
    NURBS1DFiniteElement(int p)
-      : NURBSFiniteElement(1, Geometry::SEGMENT, p + 1, p, FunctionSpace::Qk),
+      : ScalarFiniteElement(1, Geometry::SEGMENT, p + 1, p, FunctionSpace::Qk),
+        NURBSFiniteElement(1),
         shape_x(p + 1) { }
 
    virtual void SetOrder() const;
@@ -84,7 +89,8 @@ public:
 };
 
 /// An arbitrary order 2D NURBS element on a square
-class NURBS2DFiniteElement : public NURBSFiniteElement
+class NURBS2DFiniteElement : public ScalarFiniteElement,
+   public NURBSFiniteElement
 {
 protected:
    mutable Vector u, shape_x, shape_y, dshape_x, dshape_y, d2shape_x, d2shape_y;
@@ -93,16 +99,18 @@ protected:
 public:
    /// Construct the NURBS2DFiniteElement of order @a p
    NURBS2DFiniteElement(int p)
-      : NURBSFiniteElement(2, Geometry::SQUARE, (p + 1)*(p + 1), p,
-                           FunctionSpace::Qk),
+      : ScalarFiniteElement(2, Geometry::SQUARE, (p + 1)*(p + 1), p,
+                            FunctionSpace::Qk),
+        NURBSFiniteElement(2),
         u(dof), shape_x(p + 1), shape_y(p + 1), dshape_x(p + 1),
         dshape_y(p + 1), d2shape_x(p + 1), d2shape_y(p + 1), du(dof,2)
    { orders[0] = orders[1] = p; }
 
    /// Construct the NURBS2DFiniteElement with x-order @a px and y-order @a py
    NURBS2DFiniteElement(int px, int py)
-      : NURBSFiniteElement(2, Geometry::SQUARE, (px + 1)*(py + 1),
-                           std::max(px, py), FunctionSpace::Qk),
+      : ScalarFiniteElement(2, Geometry::SQUARE, (px + 1)*(py + 1),
+                            std::max(px, py), FunctionSpace::Qk),
+        NURBSFiniteElement(2),
         u(dof), shape_x(px + 1), shape_y(py + 1), dshape_x(px + 1),
         dshape_y(py + 1), d2shape_x(px + 1), d2shape_y(py + 1), du(dof,2)
    { orders[0] = px; orders[1] = py; }
@@ -116,7 +124,8 @@ public:
 };
 
 /// An arbitrary order 3D NURBS element on a cube
-class NURBS3DFiniteElement : public NURBSFiniteElement
+class NURBS3DFiniteElement : public ScalarFiniteElement,
+   public NURBSFiniteElement
 {
 protected:
    mutable Vector u, shape_x, shape_y, shape_z;
@@ -127,8 +136,9 @@ protected:
 public:
    /// Construct the NURBS3DFiniteElement of order @a p
    NURBS3DFiniteElement(int p)
-      : NURBSFiniteElement(3, Geometry::CUBE, (p + 1)*(p + 1)*(p + 1), p,
-                           FunctionSpace::Qk),
+      : ScalarFiniteElement(3, Geometry::CUBE, (p + 1)*(p + 1)*(p + 1), p,
+                            FunctionSpace::Qk),
+        NURBSFiniteElement(3),
         u(dof), shape_x(p + 1), shape_y(p + 1), shape_z(p + 1),
         dshape_x(p + 1), dshape_y(p + 1), dshape_z(p + 1),
         d2shape_x(p + 1), d2shape_y(p + 1), d2shape_z(p + 1), du(dof,3)
@@ -137,8 +147,9 @@ public:
    /// Construct the NURBS3DFiniteElement with x-order @a px and y-order @a py
    /// and z-order @a pz
    NURBS3DFiniteElement(int px, int py, int pz)
-      : NURBSFiniteElement(3, Geometry::CUBE, (px + 1)*(py + 1)*(pz + 1),
-                           std::max(std::max(px,py),pz), FunctionSpace::Qk),
+      : ScalarFiniteElement(3, Geometry::CUBE, (px + 1)*(py + 1)*(pz + 1),
+                            std::max(std::max(px,py),pz), FunctionSpace::Qk),
+        NURBSFiniteElement(2),
         u(dof), shape_x(px + 1), shape_y(py + 1), shape_z(pz + 1),
         dshape_x(px + 1), dshape_y(py + 1), dshape_z(pz + 1),
         d2shape_x(px + 1), d2shape_y(py + 1), d2shape_z(pz + 1), du(dof,3)
@@ -152,6 +163,352 @@ public:
                              DenseMatrix &hessian) const;
 };
 
+
+/** An arbitrary order H(div)-conforming 2D NURBS element on a square.
+    More details in the following papers:
+
+[1] Annalisa Buffa, Carlo De Falco, Giancarlo Sangalli
+"Isogeometric analysis: stable elements for the 2D Stokes equation."
+International Journal for Numerical Methods in Fluids 65 (11‐12) 1407-1422
+
+[2] John A Evans, Thomas JR Hughes
+"Isogeometric divergence-conforming B-splines for the unsteady Navier–Stokes equations."
+Journal of Computational Physics (241) 141-167
+*/
+class NURBS_HDiv2DFiniteElement : public VectorFiniteElement,
+   public NURBSFiniteElement
+{
+protected:
+   mutable Vector shape_x, shape_y, dshape_x, dshape_y, d2shape_x, d2shape_y;
+   mutable Vector shape1_x, shape1_y, dshape1_x, dshape1_y, d2shape1_x, d2shape1_y;
+   mutable Vector u;
+   mutable DenseMatrix du;
+   mutable Array <const KnotVector*> kv1;
+
+public:
+   /// Construct the NURBS_HDiv2DFiniteElement of order @a p
+   NURBS_HDiv2DFiniteElement(int p)
+      : VectorFiniteElement(2, Geometry::SQUARE, 2*(p + 1)*(p + 2), p,
+                            H_DIV,FunctionSpace::Qk),
+        NURBSFiniteElement(2),
+        shape_x(p + 1), shape_y(p + 1), dshape_x(p + 1),
+        dshape_y(p + 1), d2shape_x(p + 1), d2shape_y(p + 1),
+        shape1_x(p + 2), shape1_y(p + 2), dshape1_x(p + 2),
+        dshape1_y(p + 2), d2shape1_x(p + 2), d2shape1_y(p + 2),
+        u(dof), du(dof,2)
+   {
+      orders[0] = orders[1] = p;
+      kv1.SetSize(dim);
+      kv1[0] = nullptr;
+      kv1[1] = nullptr;
+   }
+
+   /// Construct the NURBS_HDiv2DFiniteElement with x-order @a px and y-order @a py
+   NURBS_HDiv2DFiniteElement(int px, int py)
+      : VectorFiniteElement(2, Geometry::SQUARE,
+                            (px + 2)*(py + 1)+(px + 1)*(py + 2),
+                            std::max(px, py), H_DIV, FunctionSpace::Qk),
+        NURBSFiniteElement(2),
+        shape_x(px + 1), shape_y(py + 1), dshape_x(px + 1),
+        dshape_y(py + 1), d2shape_x(px + 1), d2shape_y(py + 1),
+        shape1_x(px + 2), shape1_y(py + 2), dshape1_x(px + 2),
+        dshape1_y(py + 2), d2shape1_x(px + 2), d2shape1_y(py + 2),
+        u(dof), du(dof,2)
+   {
+      orders[0] = px; orders[1] = py;
+      kv1.SetSize(dim);
+      kv1[0] = nullptr;
+      kv1[1] = nullptr;
+   }
+
+   virtual void SetOrder() const;
+
+   virtual void CalcVShape(const IntegrationPoint &ip,
+                           DenseMatrix &shape) const;
+
+   /** @brief Evaluate the values of all shape functions of a *vector* finite
+       element in physical space at the point described by @a Trans. */
+   /** Each row of the result DenseMatrix @a shape contains the components of
+       one vector shape function. The size (#dof x SDim) of @a shape must be set
+       in advance, where SDim >= #dim is the physical space dimension as
+       described by @a Trans. */
+   virtual void CalcVShape(ElementTransformation &Trans,
+                           DenseMatrix &shape) const;
+
+   /** @brief Evaluate the divergence of all shape functions of a *vector*
+       finite element in reference space at the given point @a ip. */
+   /** The size (#dof) of the result Vector @a divshape must be set in advance.
+    */
+   virtual void CalcDivShape(const IntegrationPoint &ip,
+                             Vector &divshape) const;
+
+   ~NURBS_HDiv2DFiniteElement();
+};
+
+
+/** An arbitrary order H(div)-conforming 3D NURBS element on a cube
+    More details in the following papers:
+
+    [1] Annalisa Buffa, Carlo De Falco, Giancarlo Sangalli
+    "Isogeometric analysis: stable elements for the 2D Stokes equation."
+    International Journal for Numerical Methods in Fluids 65 (11‐12) 1407-1422
+
+    [2] John A Evans, Thomas JR Hughes
+    "Isogeometric divergence-conforming B-splines for the unsteady
+    Navier–Stokes equations."
+    Journal of Computational Physics (241) 141-167 */
+class NURBS_HDiv3DFiniteElement : public VectorFiniteElement,
+   public NURBSFiniteElement
+{
+protected:
+   mutable Vector shape_x, shape_y, shape_z;
+   mutable Vector dshape_x, dshape_y, dshape_z;
+   mutable Vector d2shape_x, d2shape_y, d2shape_z;
+   mutable Vector shape1_x, shape1_y, shape1_z;
+   mutable Vector dshape1_x, dshape1_y, dshape1_z;
+   mutable Vector d2shape1_x, d2shape1_y, d2shape1_z;
+   mutable Vector u;
+   mutable DenseMatrix du;
+   mutable Array <const KnotVector*> kv1;
+
+public:
+   /// Construct the NURBS_HDiv3DFiniteElement of order @a p
+   NURBS_HDiv3DFiniteElement(int p)
+      : VectorFiniteElement(3, Geometry::CUBE, 3*(p + 1)*(p + 1)*(p + 2),
+                            p, H_DIV,FunctionSpace::Qk),
+        NURBSFiniteElement(3),
+        shape_x(p + 1), shape_y(p + 1), shape_z(p + 1),
+        dshape_x(p + 1), dshape_y(p + 1), dshape_z(p + 1),
+        d2shape_x(p + 1), d2shape_y(p + 1), d2shape_z(p + 1),
+        shape1_x(p + 2), shape1_y(p + 2), shape1_z(p + 2),
+        dshape1_x(p + 2), dshape1_y(p + 2),dshape1_z(p + 2),
+        d2shape1_x(p + 2), d2shape1_y(p + 2), d2shape1_z(p + 2),
+        u(dof), du(dof,3)
+   {
+      orders[0] = orders[1] = orders[2] = p;
+      kv1.SetSize(dim);
+      kv1[0] = nullptr;
+      kv1[1] = nullptr;
+      kv1[2] = nullptr;
+   }
+
+   /// Construct the NURBS_HDiv3DFiniteElement with x-order @a px, y-order @a py and z-order @a pz
+   NURBS_HDiv3DFiniteElement(int px, int py, int pz)
+      : VectorFiniteElement(3, Geometry::CUBE,
+                            (px + 2)*(py + 1)*(pz + 1) +
+                            (px + 1)*(py + 2)*(pz + 1) +
+                            (px + 1)*(py + 1)*(pz + 2),
+                            std::max(px, py), H_DIV, FunctionSpace::Qk),
+        NURBSFiniteElement(3),
+        shape_x(px + 1), shape_y(py + 1), shape_z(pz + 1),
+        dshape_x(px + 1), dshape_y(py + 1), dshape_z(pz + 1),
+        d2shape_x(px + 1), d2shape_y(py + 1), d2shape_z(pz + 1),
+        shape1_x(px + 2), shape1_y(py + 2), shape1_z(pz + 2),
+        dshape1_x(px + 2), dshape1_y(py + 2),dshape1_z(pz + 2),
+        d2shape1_x(px + 2), d2shape1_y(py + 2), d2shape1_z(pz + 2),
+        u(dof), du(dof,3)
+   {
+      orders[0] = px; orders[1] = py; orders[2] = pz;
+      kv1.SetSize(dim);
+      kv1[0] = nullptr;
+      kv1[1] = nullptr;
+      kv1[2] = nullptr;
+   }
+
+   virtual void SetOrder() const;
+
+   virtual void CalcVShape(const IntegrationPoint &ip,
+                           DenseMatrix &shape) const;
+
+   /** @brief Evaluate the values of all shape functions of a *vector* finite
+       element in physical space at the point described by @a Trans. */
+   /** Each row of the result DenseMatrix @a shape contains the components of
+       one vector shape function. The size (#dof x SDim) of @a shape must be set
+       in advance, where SDim >= #dim is the physical space dimension as
+       described by @a Trans. */
+   virtual void CalcVShape(ElementTransformation &Trans,
+                           DenseMatrix &shape) const;
+
+   /** @brief Evaluate the divergence of all shape functions of a *vector*
+       finite element in reference space at the given point @a ip. */
+   /** The size (#dof) of the result Vector @a divshape must be set in advance.
+    */
+   virtual void CalcDivShape(const IntegrationPoint &ip,
+                             Vector &divshape) const;
+
+   ~NURBS_HDiv3DFiniteElement();
+};
+
+
+/** An arbitrary order H(curl)-conforming 2D NURBS element on a square
+    More details in the following paper:
+
+    [1] Annalisa Buffa, Giancarlo Sangalli, Rafael Vázquez
+    "Isogeometric analysis in electromagnetics: B-splines approximation."
+    Computer Methods in Applied Mechanics and Engineering (199) 1143-1152 */
+class NURBS_HCurl2DFiniteElement : public VectorFiniteElement,
+   public NURBSFiniteElement
+{
+protected:
+   mutable Vector shape_x, shape_y, dshape_x, dshape_y, d2shape_x, d2shape_y;
+   mutable Vector shape1_x, shape1_y, dshape1_x, dshape1_y, d2shape1_x, d2shape1_y;
+   mutable Vector u;
+   mutable DenseMatrix du;
+   mutable Array <const KnotVector*> kv1;
+
+public:
+   /// Construct the NURBS_HCurl2DFiniteElement of order @a p
+   NURBS_HCurl2DFiniteElement(int p)
+      : VectorFiniteElement(2, Geometry::SQUARE, 2*(p + 1)*(p + 2), p,
+                            H_CURL,FunctionSpace::Qk),
+        NURBSFiniteElement(2),
+        shape_x(p + 1), shape_y(p + 1), dshape_x(p + 1),
+        dshape_y(p + 1), d2shape_x(p + 1), d2shape_y(p + 1),
+        shape1_x(p + 2), shape1_y(p + 2), dshape1_x(p + 2),
+        dshape1_y(p + 2), d2shape1_x(p + 2), d2shape1_y(p + 2),
+        u(dof), du(dof,2)
+   {
+      orders[0] = orders[1] = p;
+      kv1.SetSize(dim);
+      kv1[0] = nullptr;
+      kv1[1] = nullptr;
+   }
+
+   /// Construct the NURBS_HCurl2DFiniteElement with x-order @a px and y-order @a py
+   NURBS_HCurl2DFiniteElement(int px, int py)
+      : VectorFiniteElement(2, Geometry::SQUARE,
+                            (px + 1)*(py + 2)+(px + 2)*(py + 1),
+                            std::max(px, py), H_CURL, FunctionSpace::Qk),
+        NURBSFiniteElement(2),
+        shape_x(px + 1), shape_y(py + 1), dshape_x(px + 1),
+        dshape_y(py + 1), d2shape_x(px + 1), d2shape_y(py + 1),
+        shape1_x(px + 2), shape1_y(py + 2), dshape1_x(px + 2),
+        dshape1_y(py + 2), d2shape1_x(px + 2), d2shape1_y(py + 2),
+        u(dof), du(dof,2)
+   {
+      orders[0] = px; orders[1] = py;
+      kv1.SetSize(dim);
+      kv1[0] = nullptr;
+      kv1[1] = nullptr;
+   }
+
+   virtual void SetOrder() const;
+
+   virtual void CalcVShape(const IntegrationPoint &ip,
+                           DenseMatrix &shape) const;
+
+   /** @brief Evaluate the values of all shape functions of a *vector* finite
+       element in physical space at the point described by @a Trans. */
+   /** Each row of the result DenseMatrix @a shape contains the components of
+       one vector shape function. The size (#dof x SDim) of @a shape must be set
+       in advance, where SDim >= #dim is the physical space dimension as
+       described by @a Trans. */
+   virtual void CalcVShape(ElementTransformation &Trans,
+                           DenseMatrix &shape) const;
+
+   /** @brief Evaluate the curl of all shape functions of a *vector* finite
+       element in reference space at the given point @a ip. */
+   /** Each row of the result DenseMatrix @a curl_shape contains the components
+       of the curl of one vector shape function. The size (#dof x CDim) of
+       @a curl_shape must be set in advance, where CDim = 3 for #dim = 3 and
+       CDim = 1 for #dim = 2. */
+   virtual void CalcCurlShape(const IntegrationPoint &ip,
+                              DenseMatrix &curl_shape) const;
+
+   ~NURBS_HCurl2DFiniteElement();
+};
+
+
+/** An arbitrary order H(curl)-conforming 3D NURBS element on a cube
+    More details in the following paper:
+
+    [1] Annalisa Buffa, Giancarlo Sangalli, Rafael Vázquez
+    "Isogeometric analysis in electromagnetics: B-splines approximation."
+    Computer Methods in Applied Mechanics and Engineering (199) 1143-1152 */
+class NURBS_HCurl3DFiniteElement : public VectorFiniteElement,
+   public NURBSFiniteElement
+{
+protected:
+   mutable Vector shape_x, shape_y, shape_z;
+   mutable Vector dshape_x, dshape_y, dshape_z;
+   mutable Vector d2shape_x, d2shape_y, d2shape_z;
+   mutable Vector shape1_x, shape1_y, shape1_z;
+   mutable Vector dshape1_x, dshape1_y, dshape1_z;
+   mutable Vector d2shape1_x, d2shape1_y, d2shape1_z;
+   mutable Vector u;
+   mutable DenseMatrix du;
+   mutable Array <const KnotVector*> kv1;
+
+public:
+   /// Construct the NURBS_HCurl3DFiniteElement of order @a p
+   NURBS_HCurl3DFiniteElement(int p)
+      : VectorFiniteElement(3, Geometry::CUBE, 3*(p + 1)*(p + 2)*(p + 2), p,
+                            H_CURL,FunctionSpace::Qk),
+        NURBSFiniteElement(3),
+        shape_x(p + 1), shape_y(p + 1), shape_z(p + 1),
+        dshape_x(p + 1), dshape_y(p + 1), dshape_z(p + 1),
+        d2shape_x(p + 1), d2shape_y(p + 1), d2shape_z(p + 1),
+        shape1_x(p + 2), shape1_y(p + 2), shape1_z(p + 2),
+        dshape1_x(p + 2), dshape1_y(p + 2),dshape1_z(p + 2),
+        d2shape1_x(p + 2), d2shape1_y(p + 2), d2shape1_z(p + 2),
+        u(dof), du(dof,3)
+   {
+      orders[0] = orders[1] = orders[2] = p;
+      kv1.SetSize(dim);
+      kv1[0] = nullptr;
+      kv1[1] = nullptr;
+      kv1[2] = nullptr;
+   }
+
+   /// Construct the NURBS_HCurl3DFiniteElement with x-order @a px, y-order @a py and z-order @a pz
+   NURBS_HCurl3DFiniteElement(int px, int py, int pz)
+      : VectorFiniteElement(3, Geometry::CUBE,
+                            (px + 1)*(py + 2)*(pz + 2) +
+                            (px + 2)*(py + 1)*(pz + 2) +
+                            (px + 2)*(py + 2)*(pz + 1),
+                            std::max(std::max(px, py), pz), H_CURL, FunctionSpace::Qk),
+        NURBSFiniteElement(3),
+        shape_x(px + 1), shape_y(py + 1), shape_z(pz + 1),
+        dshape_x(px + 1), dshape_y(py + 1), dshape_z(pz + 1),
+        d2shape_x(px + 1), d2shape_y(py + 1), d2shape_z(pz + 1),
+        shape1_x(px + 2), shape1_y(py + 2), shape1_z(pz + 2),
+        dshape1_x(px + 2), dshape1_y(py + 2),dshape1_z(pz + 2),
+        d2shape1_x(px + 2), d2shape1_y(py + 2), d2shape1_z(pz + 2),
+        u(dof), du(dof,3)
+   {
+      orders[0] = px; orders[1] = py; orders[2] = pz;
+      kv1.SetSize(dim);
+      kv1[0] = nullptr;
+      kv1[1] = nullptr;
+      kv1[2] = nullptr;
+   }
+
+   virtual void SetOrder() const;
+
+   virtual void CalcVShape(const IntegrationPoint &ip,
+                           DenseMatrix &shape) const;
+
+   /** @brief Evaluate the values of all shape functions of a *vector* finite
+       element in physical space at the point described by @a Trans. */
+   /** Each row of the result DenseMatrix @a shape contains the components of
+       one vector shape function. The size (#dof x SDim) of @a shape must be set
+       in advance, where SDim >= #dim is the physical space dimension as
+       described by @a Trans. */
+   virtual void CalcVShape(ElementTransformation &Trans,
+                           DenseMatrix &shape) const;
+
+   /** @brief Evaluate the curl of all shape functions of a *vector* finite
+       element in reference space at the given point @a ip. */
+   /** Each row of the result DenseMatrix @a curl_shape contains the components
+       of the curl of one vector shape function. The size (#dof x CDim) of
+       @a curl_shape must be set in advance, where CDim = 3 for #dim = 3 and
+       CDim = 1 for #dim = 2. */
+   virtual void CalcCurlShape(const IntegrationPoint &ip,
+                              DenseMatrix &curl_shape) const;
+
+   ~NURBS_HCurl3DFiniteElement();
+};
+
 } // namespace mfem
 
 #endif

fem/fe_coll.cpp

@@ -344,6 +344,32 @@ FiniteElementCollection *FiniteElementCollection::New(const char *name)
    {
       fec = new Local_FECollection(name + 6);
    }
+   else if (!strncmp(name, "NURBS_HDiv", 10))
+   {
+      if (name[10] != '\0')
+      {
+         // "NURBS" + "number" --> fixed order nurbs collection
+         fec = new NURBS_HDivFECollection(atoi(name + 10));
+      }
+      else
+      {
+         // "NURBS" --> variable order nurbs collection
+         fec = new NURBS_HDivFECollection();
+      }
+   }
+   else if (!strncmp(name, "NURBS_HCurl", 11))
+   {
+      if (name[11] != '\0')
+      {
+         // "NURBS" + "number" --> fixed order nurbs collection
+         fec = new NURBS_HCurlFECollection(atoi(name + 11));
+      }
+      else
+      {
+         // "NURBS" --> variable order nurbs collection
+         fec = new NURBS_HCurlFECollection();
+      }
+   }
    else if (!strncmp(name, "NURBS", 5))
    {
       if (name[5] != '\0')
@@ -3533,4 +3559,192 @@ FiniteElementCollection *NURBSFECollection::GetTraceCollection() const
    return NULL;
 }
 
+
+NURBS_HDivFECollection::NURBS_HDivFECollection(int Order, const int dim)
+   : NURBSFECollection((Order == VariableOrder) ? 1 : Order)
+{
+   const int order = (Order == VariableOrder) ? 1 : Order;
+
+   SegmentFE       = new NURBS1DFiniteElement(order);
+   QuadrilateralFE = new NURBS2DFiniteElement(order);
+
+   QuadrilateralVFE  = new NURBS_HDiv2DFiniteElement(order);
+   ParallelepipedVFE = new NURBS_HDiv3DFiniteElement(order);
+
+   if (dim != -1) { SetDim(dim); }
+   SetOrder(Order);
+}
+
+void NURBS_HDivFECollection::SetDim(int dim)
+{
+   if (dim == 2)
+   {
+      sFE = SegmentFE;
+      qFE = QuadrilateralVFE;
+      hFE = nullptr;
+   }
+   else if (dim == 3)
+   {
+      sFE = nullptr;
+      qFE = QuadrilateralFE;
+      hFE = ParallelepipedVFE;
+   }
+   else
+   {
+      mfem::err<<"Dimension = "<<dim<<endl;
+      mfem_error ("NURBS_HDivFECollection: wrong dimension!");
+   }
+}
+
+NURBS_HDivFECollection::~NURBS_HDivFECollection()
+{
+   delete SegmentFE;
+   delete QuadrilateralFE;
+   delete QuadrilateralVFE;
+   delete ParallelepipedVFE;
+}
+
+const FiniteElement *
+NURBS_HDivFECollection::FiniteElementForGeometry(Geometry::Type GeomType) const
+{
+   switch (GeomType)
+   {
+      case Geometry::SEGMENT:     return sFE;
+      case Geometry::SQUARE:      return qFE;
+      case Geometry::CUBE:        return hFE;
+      default:
+         if (error_mode == RETURN_NULL) { return nullptr; }
+         mfem_error ("NURBS_HDivFECollection: unknown geometry type.");
+   }
+   return QuadrilateralFE; // Make some compilers happy
+}
+
+void NURBS_HDivFECollection::SetOrder(int Order) const
+{
+   mOrder = Order;
+   if (Order != VariableOrder)
+   {
+      snprintf(name, 16, "NURBS_HDiv%i", Order);
+   }
+   else
+   {
+      snprintf(name, 16, "NURBS_HDiv");
+   }
+}
+
+int NURBS_HDivFECollection::DofForGeometry(Geometry::Type GeomType) const
+{
+   mfem_error("NURBS_HDivFECollection::DofForGeometry");
+   return 0; // Make some compilers happy
+}
+
+const int *NURBS_HDivFECollection::DofOrderForOrientation(
+   Geometry::Type GeomType,
+   int Or) const
+{
+   mfem_error("NURBS_HDivFECollection::DofOrderForOrientation");
+   return NULL;
+}
+
+FiniteElementCollection *NURBS_HDivFECollection::GetTraceCollection() const
+{
+   MFEM_ABORT("NURBS finite elements can not be statically condensed!");
+   return NULL;
+}
+
+NURBS_HCurlFECollection::NURBS_HCurlFECollection(int Order, const int dim)
+   : NURBSFECollection((Order == VariableOrder) ? 1 : Order)
+{
+   const int order = (Order == VariableOrder) ? 1 : Order;
+
+   SegmentFE       = new NURBS1DFiniteElement(order+1);
+   QuadrilateralFE = new NURBS2DFiniteElement(order+1);
+
+   QuadrilateralVFE  = new NURBS_HCurl2DFiniteElement(order);
+   ParallelepipedVFE = new NURBS_HCurl3DFiniteElement(order);
+   if (dim != -1) { SetDim(dim); }
+   SetOrder(Order);
+}
+
+void NURBS_HCurlFECollection::SetDim(int dim)
+{
+   if (dim == 2)
+   {
+      sFE = SegmentFE;
+      qFE = QuadrilateralVFE;
+      hFE = nullptr;
+   }
+   else if (dim == 3)
+   {
+      sFE = nullptr;
+      qFE = QuadrilateralFE;
+      hFE = ParallelepipedVFE;
+   }
+   else
+   {
+      mfem::err<<"Dimension = "<<dim<<endl;
+      mfem_error ("NURBS_HCurlFECollection: wrong dimension!");
+   }
+}
+
+
+
+NURBS_HCurlFECollection::~NURBS_HCurlFECollection()
+{
+   delete SegmentFE;
+   delete QuadrilateralFE;
+   delete QuadrilateralVFE;
+   delete ParallelepipedVFE;
+}
+
+const FiniteElement *
+NURBS_HCurlFECollection::FiniteElementForGeometry(Geometry::Type GeomType) const
+{
+   switch (GeomType)
+   {
+      case Geometry::SEGMENT:     return sFE;
+      case Geometry::SQUARE:      return qFE;
+      case Geometry::CUBE:        return hFE;
+      default:
+         if (error_mode == RETURN_NULL) { return nullptr; }
+         mfem_error ("NURBS_HCurlFECollection: unknown geometry type.");
+   }
+   return QuadrilateralFE; // Make some compilers happy
+}
+
+void NURBS_HCurlFECollection::SetOrder(int Order) const
+{
+   mOrder = Order;
+   if (Order != VariableOrder)
+   {
+      snprintf(name, 16, "NURBS_HCurl%i", Order);
+   }
+   else
+   {
+      snprintf(name, 16, "NURBS_HCurl");
+   }
+}
+
+int NURBS_HCurlFECollection::DofForGeometry(Geometry::Type GeomType) const
+{
+   mfem_error("NURBS_HCurlFECollection::DofForGeometry");
+   return 0; // Make some compilers happy
+}
+
+const int *NURBS_HCurlFECollection::DofOrderForOrientation(
+   Geometry::Type GeomType,
+   int Or) const
+{
+   mfem_error("NURBS_HCurlFECollection::DofOrderForOrientation");
+   return NULL;
+}
+
+FiniteElementCollection *NURBS_HCurlFECollection::GetTraceCollection() const
+{
+   MFEM_ABORT("NURBS finite elements can not be statically condensed!");
+   return NULL;
+}
+
+
+
 }

fem/fe_coll.hpp

@@ -680,8 +680,8 @@ public:
 /// Arbitrary order non-uniform rational B-splines (NURBS) finite elements.
 class NURBSFECollection : public FiniteElementCollection
 {
-private:
-   PointFiniteElement   *PointFE;
+protected:
+   PointFiniteElement *PointFE;
    NURBS1DFiniteElement *SegmentFE;
    NURBS2DFiniteElement *QuadrilateralFE;
    NURBS3DFiniteElement *ParallelepipedFE;
@@ -701,13 +701,15 @@ public:
       order, or VariableOrder (default). */
    explicit NURBSFECollection(int Order = VariableOrder);
 
-   void Reset() const
+   virtual void Reset() const
    {
       SegmentFE->Reset();
       QuadrilateralFE->Reset();
       ParallelepipedFE->Reset();
    }
 
+   virtual void SetDim(const int dim) {};
+
    /** @brief Get the order of the NURBS collection: either a positive number,
        when using fixed order, or VariableOrder. */
    /** @note Not to be confused with FiniteElementCollection::GetOrder(). */
@@ -715,7 +717,7 @@ public:
 
    /** @brief Set the order and the name, based on the given @a Order: either a
        positive number for fixed order, or VariableOrder. */
-   void SetOrder(int Order) const;
+   virtual void SetOrder(int Order) const;
 
    const FiniteElement *
    FiniteElementForGeometry(Geometry::Type GeomType) const override;
@@ -734,6 +736,109 @@ public:
    virtual ~NURBSFECollection();
 };
 
+/// Arbitrary order H(div) NURBS finite elements.
+class NURBS_HDivFECollection : public NURBSFECollection
+{
+private:
+
+   NURBS1DFiniteElement *SegmentFE;
+   NURBS2DFiniteElement *QuadrilateralFE;
+
+   NURBS_HDiv2DFiniteElement *QuadrilateralVFE;
+   NURBS_HDiv3DFiniteElement *ParallelepipedVFE;
+
+   FiniteElement *sFE;
+   FiniteElement *qFE;
+   FiniteElement *hFE;
+
+public:
+
+   /** @brief The parameter @a Order must be either a positive number, for fixed
+      order, or VariableOrder (default). */
+   explicit NURBS_HDivFECollection(int Order = VariableOrder, const int vdim = -1);
+
+   virtual void Reset() const override
+   {
+      SegmentFE->Reset();
+      QuadrilateralFE->Reset();
+      QuadrilateralVFE->Reset();
+      ParallelepipedVFE->Reset();
+   }
+
+   virtual void SetDim(const int dim) override;
+
+   /** @brief Set the order and the name, based on the given @a Order: either a
+       positive number for fixed order, or VariableOrder. */
+   virtual void SetOrder(int Order) const override;
+
+   const FiniteElement *
+   FiniteElementForGeometry(Geometry::Type GeomType) const override;
+
+   int DofForGeometry(Geometry::Type GeomType) const override;
+
+   const int *DofOrderForOrientation(Geometry::Type GeomType,
+                                     int Or) const override;
+
+   const char *Name() const override { return name; }
+
+   int GetContType() const override { return CONTINUOUS; }
+
+   FiniteElementCollection *GetTraceCollection() const override;
+
+   virtual ~NURBS_HDivFECollection();
+};
+
+/// Arbitrary order H(curl) NURBS finite elements.
+class NURBS_HCurlFECollection : public NURBSFECollection
+{
+private:
+   NURBS1DFiniteElement *SegmentFE;
+   NURBS2DFiniteElement *QuadrilateralFE;
+
+   NURBS_HCurl2DFiniteElement *QuadrilateralVFE;
+   NURBS_HCurl3DFiniteElement *ParallelepipedVFE;
+
+   FiniteElement *sFE;
+   FiniteElement *qFE;
+   FiniteElement *hFE;
+public:
+
+   /** @brief The parameter @a Order must be either a positive number, for fixed
+      order, or VariableOrder (default). */
+   explicit NURBS_HCurlFECollection(int Order = VariableOrder,
+                                    const int vdim = -1);
+
+   virtual void Reset() const override
+   {
+      SegmentFE->Reset();
+      QuadrilateralFE->Reset();
+      QuadrilateralVFE->Reset();
+      ParallelepipedVFE->Reset();
+   }
+
+   virtual void SetDim(const int dim) override;
+
+   /** @brief Set the order and the name, based on the given @a Order: either a
+       positive number for fixed order, or VariableOrder. */
+   virtual void SetOrder(int Order) const override;
+
+   const FiniteElement *
+   FiniteElementForGeometry(Geometry::Type GeomType) const override;
+
+   int DofForGeometry(Geometry::Type GeomType) const override;
+
+   const int *DofOrderForOrientation(Geometry::Type GeomType,
+                                     int Or) const override;
+
+   const char *Name() const override { return name; }
+
+   int GetContType() const override { return CONTINUOUS; }
+
+   FiniteElementCollection *GetTraceCollection() const override;
+
+   virtual ~NURBS_HCurlFECollection();
+};
+
 /// Piecewise-(bi/tri)linear continuous finite elements.
 class LinearFECollection : public FiniteElementCollection
 {

fem/fespace.cpp

@@ -1525,6 +1525,67 @@ SparseMatrix *FiniteElementSpace::RefinementMatrix_main(
    return P;
 }
 
+SparseMatrix *FiniteElementSpace::VariableOrderRefinementMatrix(
+   const int coarse_ndofs, const Table &coarse_elem_dof) const
+{
+   MFEM_VERIFY(mesh->GetLastOperation() == Mesh::REFINE, "");
+
+   Array<int> dofs, coarse_dofs, coarse_vdofs;
+   Vector row;
+
+   Mesh::GeometryList elem_geoms(*mesh);
+
+   SparseMatrix *P = new SparseMatrix(GetVSize(), coarse_ndofs*vdim);
+
+   Array<int> mark(P->Height());
+   mark = 0;
+
+   const CoarseFineTransformations &rtrans = mesh->GetRefinementTransforms();
+   DenseMatrix lP;
+   IsoparametricTransformation isotr;
+   for (int k = 0; k < mesh->GetNE(); k++)
+   {
+      const Embedding &emb = rtrans.embeddings[k];
+      const Geometry::Type geom = mesh->GetElementBaseGeometry(k);
+
+      const FiniteElement *fe = GetFE(k);
+      isotr.SetIdentityTransformation(geom);
+      const int ldof = fe->GetDof();
+      lP.SetSize(ldof, ldof);
+      const DenseTensor &pmats = rtrans.point_matrices[geom];
+      isotr.SetPointMat(pmats(emb.matrix));
+      fe->GetLocalInterpolation(isotr, lP);
+
+      const int fine_ldof = lP.Height();
+
+      elem_dof->GetRow(k, dofs);
+      coarse_elem_dof.GetRow(emb.parent, coarse_dofs);
+
+      for (int vd = 0; vd < vdim; vd++)
+      {
+         coarse_dofs.Copy(coarse_vdofs);
+         DofsToVDofs(vd, coarse_vdofs, coarse_ndofs);
+
+         for (int i = 0; i < fine_ldof; i++)
+         {
+            const int r = DofToVDof(dofs[i], vd);
+            int m = (r >= 0) ? r : (-1 - r);
+
+            if (!mark[m])
+            {
+               lP.GetRow(i, row);
+               P->SetRow(r, coarse_vdofs, row);
+               mark[m] = 1;
+            }
+         }
+      }
+   }
+
+   MFEM_VERIFY(mark.Sum() == P->Height(), "Not all rows of P set.");
+   P->Finalize();
+   return P;
+}
+
 void FiniteElementSpace::GetLocalRefinementMatrices(
    Geometry::Type geom, DenseTensor &localP) const
 {
@@ -1556,15 +1617,20 @@ SparseMatrix* FiniteElementSpace::RefinementMatrix(int old_ndofs,
                "Previous mesh is not coarser.");
 
    Mesh::GeometryList elem_geoms(*mesh);
-
-   DenseTensor localP[Geometry::NumGeom];
-   for (int i = 0; i < elem_geoms.Size(); i++)
+   if (!IsVariableOrder())
    {
-      GetLocalRefinementMatrices(elem_geoms[i], localP[elem_geoms[i]]);
+      DenseTensor localP[Geometry::NumGeom];
+      for (int i = 0; i < elem_geoms.Size(); i++)
+      {
+         GetLocalRefinementMatrices(elem_geoms[i], localP[elem_geoms[i]]);
+      }
+      return RefinementMatrix_main(old_ndofs, *old_elem_dof, old_elem_fos,
+                                   localP);
+   }
+   else
+   {
+      return VariableOrderRefinementMatrix(old_ndofs, *old_elem_dof);
    }
-
-   return RefinementMatrix_main(old_ndofs, *old_elem_dof, old_elem_fos,
-                                localP);
 }
 
 FiniteElementSpace::RefinementOperator::RefinementOperator(
@@ -1582,9 +1648,12 @@ FiniteElementSpace::RefinementOperator::RefinementOperator(
 
    Mesh::GeometryList elem_geoms(*fespace->GetMesh());
 
-   for (int i = 0; i < elem_geoms.Size(); i++)
+   if (!fespace->IsVariableOrder())
    {
-      fespace->GetLocalRefinementMatrices(elem_geoms[i], localP[elem_geoms[i]]);
+      for (int i = 0; i < elem_geoms.Size(); i++)
+      {
+         fespace->GetLocalRefinementMatrices(elem_geoms[i], localP[elem_geoms[i]]);
+      }
    }
 
    ConstructDoFTransArray();
@@ -1597,10 +1666,13 @@ FiniteElementSpace::RefinementOperator::RefinementOperator(
 {
    Mesh::GeometryList elem_geoms(*fespace->GetMesh());
 
-   for (int i = 0; i < elem_geoms.Size(); i++)
+   if (!fespace->IsVariableOrder())
    {
-      fespace->GetLocalRefinementMatrices(*coarse_fes, elem_geoms[i],
-                                          localP[elem_geoms[i]]);
+      for (int i = 0; i < elem_geoms.Size(); i++)
+      {
+         fespace->GetLocalRefinementMatrices(*coarse_fes, elem_geoms[i],
+                                             localP[elem_geoms[i]]);
+      }
    }
 
    // Make a copy of the coarse elem_dof Table.
@@ -1676,11 +1748,25 @@ void FiniteElementSpace::RefinementOperator::Mult(const Vector &x,
 
    Vector subY, subX;
 
+   DenseMatrix eP;
+   IsoparametricTransformation isotr;
+
    for (int k = 0; k < mesh_ref->GetNE(); k++)
    {
       const Embedding &emb = trans_ref.embeddings[k];
       const Geometry::Type geom = mesh_ref->GetElementBaseGeometry(k);
-      const DenseMatrix &lP = localP[geom](emb.matrix);
+      if (fespace->IsVariableOrder())
+      {
+         const FiniteElement *fe = fespace->GetFE(k);
+         isotr.SetIdentityTransformation(geom);
+         const int ldof = fe->GetDof();
+         eP.SetSize(ldof, ldof);
+         const DenseTensor &pmats = trans_ref.point_matrices[geom];
+         isotr.SetPointMat(pmats(emb.matrix));
+         fe->GetLocalInterpolation(isotr, eP);
+      }
+      const DenseMatrix &lP = (fespace->IsVariableOrder()) ? eP : localP[geom](
+                                 emb.matrix);
 
       subY.SetSize(lP.Height());
 
@@ -1745,11 +1831,28 @@ void FiniteElementSpace::RefinementOperator::MultTranspose(const Vector &x,
 
    Vector subY, subX, subYt;
 
+   DenseMatrix eP;
+   IsoparametricTransformation isotr;
+   const FiniteElement *fe = nullptr;
+
    for (int k = 0; k < mesh_ref->GetNE(); k++)
    {
       const Embedding &emb = trans_ref.embeddings[k];
       const Geometry::Type geom = mesh_ref->GetElementBaseGeometry(k);
-      const DenseMatrix &lP = localP[geom](emb.matrix);
+
+      if (fespace->IsVariableOrder())
+      {
+         fe = fespace->GetFE(k);
+         isotr.SetIdentityTransformation(geom);
+         const int ldof = fe->GetDof();
+         eP.SetSize(ldof);
+         const DenseTensor &pmats = trans_ref.point_matrices[geom];
+         isotr.SetPointMat(pmats(emb.matrix));
+         fe->GetLocalInterpolation(isotr, eP);
+      }
+
+      const DenseMatrix &lP = (fespace->IsVariableOrder()) ? eP : localP[geom](
+                                 emb.matrix);
 
       DofTransformation *doftrans = fespace->GetElementDofs(k, f_dofs);
       old_elem_dof->GetRow(emb.parent, c_dofs);
@@ -2108,9 +2211,12 @@ SparseMatrix* FiniteElementSpace::DerefinementMatrix(int old_ndofs,
    Mesh::GeometryList elem_geoms(*mesh);
 
    DenseTensor localR[Geometry::NumGeom];
-   for (int i = 0; i < elem_geoms.Size(); i++)
+   if (!IsVariableOrder())
    {
-      GetLocalDerefinementMatrices(elem_geoms[i], localR[elem_geoms[i]]);
+      for (int i = 0; i < elem_geoms.Size(); i++)
+      {
+         GetLocalDerefinementMatrices(elem_geoms[i], localR[elem_geoms[i]]);
+      }
    }
 
    SparseMatrix *R = new SparseMatrix(ndofs*vdim, old_ndofs*vdim);
@@ -2125,14 +2231,34 @@ SparseMatrix* FiniteElementSpace::DerefinementMatrix(int old_ndofs,
 
    bool is_dg = FEColl()->GetContType() == FiniteElementCollection::DISCONTINUOUS;
    int num_marked = 0;
+   const FiniteElement *fe = nullptr;
+   DenseMatrix localRVO; //for variable order only
    for (int k = 0; k < dtrans.embeddings.Size(); k++)
    {
       const Embedding &emb = dtrans.embeddings[k];
       Geometry::Type geom = mesh->GetElementBaseGeometry(emb.parent);
-      DenseMatrix &lR = localR[geom](emb.matrix);
+
+      if (IsVariableOrder())
+      {
+         fe = GetFE(emb.parent);
+         const DenseTensor &pmats = dtrans.point_matrices[geom];
+         const int ldof = fe->GetDof();
+
+         IsoparametricTransformation isotr;
+         isotr.SetIdentityTransformation(geom);
+
+         localRVO.SetSize(ldof, ldof);
+         isotr.SetPointMat(pmats(emb.matrix));
+         // Local restriction is size ldofxldof assuming that the parent and
+         // child are of same polynomial order.
+         fe->GetLocalRestriction(isotr, localRVO);
+      }
+      DenseMatrix &lR = IsVariableOrder() ? localRVO : localR[geom](emb.matrix);
 
       elem_dof->GetRow(emb.parent, dofs);
       old_elem_dof->GetRow(k, old_dofs);
+      MFEM_VERIFY(old_dofs.Size() == dofs.Size(),
+                  "Parent and child must have same #dofs.");
 
       for (int vd = 0; vd < vdim; vd++)
       {
@@ -2158,7 +2284,7 @@ SparseMatrix* FiniteElementSpace::DerefinementMatrix(int old_ndofs,
       }
    }
 
-   if (!is_dg)
+   if (!is_dg && !IsVariableOrder())
    {
       MFEM_VERIFY(num_marked == R->Height(),
                   "internal error: not all rows of R were set.");
@@ -2216,6 +2342,7 @@ void FiniteElementSpace::Constructor(Mesh *mesh_, NURBSExtension *NURBSext_,
 
    const NURBSFECollection *nurbs_fec =
       dynamic_cast<const NURBSFECollection *>(fec_);
+
    if (nurbs_fec)
    {
       MFEM_VERIFY(mesh_->NURBSext, "NURBS FE space requires a NURBS mesh.");
@@ -2312,12 +2439,63 @@ void FiniteElementSpace::UpdateNURBS()
    face_dof = NULL;
    face_to_be.DeleteAll();
 
+   // Depending on the element type create the appropriate extensions
+   // for the individual components.
    dynamic_cast<const NURBSFECollection *>(fec)->Reset();
 
-   ndofs = NURBSext->GetNDof();
-   elem_dof = NURBSext->GetElementDofTable();
-   bdr_elem_dof = NURBSext->GetBdrElementDofTable();
+   if (dynamic_cast<const NURBS_HDivFECollection *>(fec))
+   {
+      VNURBSext.SetSize(mesh->Dimension());
+      for (int d = 0; d < mesh->Dimension(); d++)
+      {
+         VNURBSext[d] = NURBSext->GetDivExtension(d);
+      }
+   }
 
+   if (dynamic_cast<const NURBS_HCurlFECollection *>(fec))
+   {
+      VNURBSext.SetSize(mesh->Dimension());
+      for (int d = 0; d < mesh->Dimension(); d++)
+      {
+         VNURBSext[d] = NURBSext->GetCurlExtension(d);
+      }
+   }
+
+   // If required: concatenate the dof tables of the individual components into
+   // one dof table for the vector fespace.
+   if (VNURBSext.Size() == 2)
+   {
+      int offset1 = VNURBSext[0]->GetNDof();
+      ndofs = VNURBSext[0]->GetNDof() + VNURBSext[1]->GetNDof();
+
+      // Merge Tables
+      elem_dof = new Table(*VNURBSext[0]->GetElementDofTable(),
+                           *VNURBSext[1]->GetElementDofTable(),offset1 );
+
+      bdr_elem_dof = new Table(*VNURBSext[0]->GetBdrElementDofTable(),
+                               *VNURBSext[1]->GetBdrElementDofTable(),offset1);
+   }
+   else if (VNURBSext.Size() == 3)
+   {
+      int offset1 = VNURBSext[0]->GetNDof();
+      int offset2 = offset1 + VNURBSext[1]->GetNDof();
+      ndofs = offset2 + VNURBSext[2]->GetNDof();
+
+      // Merge Tables
+      elem_dof = new Table(*VNURBSext[0]->GetElementDofTable(),
+                           *VNURBSext[1]->GetElementDofTable(),offset1,
+                           *VNURBSext[2]->GetElementDofTable(),offset2);
+
+      bdr_elem_dof = new Table(*VNURBSext[0]->GetBdrElementDofTable(),
+                               *VNURBSext[1]->GetBdrElementDofTable(),offset1,
+                               *VNURBSext[2]->GetBdrElementDofTable(),offset2);
+   }
+   else
+   {
+      ndofs = NURBSext->GetNDof();
+      elem_dof = NURBSext->GetElementDofTable();
+      bdr_elem_dof = NURBSext->GetBdrElementDofTable();
+   }
    mesh_sequence = mesh->GetSequence();
    sequence++;
 }
@@ -3319,11 +3497,21 @@ void FiniteElementSpace::Destroy()
    dof_elem_array.DeleteAll();
    dof_ldof_array.DeleteAll();
 
+   for (int i = 0; i < VNURBSext.Size(); i++)
+   {
+      delete VNURBSext[i];
+   }
+
    if (NURBSext)
    {
       if (own_ext) { delete NURBSext; }
       delete face_dof;
       face_to_be.DeleteAll();
+      if (VNURBSext.Size() > 0 )
+      {
+         delete elem_dof;
+         delete bdr_elem_dof;
+      }
    }
    else
    {
@@ -3335,6 +3523,8 @@ void FiniteElementSpace::Destroy()
       delete [] bdofs;
    }
    ceed::RemoveBasisAndRestriction(this);
+
+
 }
 
 void FiniteElementSpace::DestroyDoFTransArray()
@@ -3353,19 +3543,27 @@ void FiniteElementSpace::GetTransferOperator(
 
    if (T.Type() == Operator::MFEM_SPARSEMAT)
    {
-      Mesh::GeometryList elem_geoms(*mesh);
-
-      DenseTensor localP[Geometry::NumGeom];
-      for (int i = 0; i < elem_geoms.Size(); i++)
+      if (!IsVariableOrder())
       {
-         GetLocalRefinementMatrices(coarse_fes, elem_geoms[i],
-                                    localP[elem_geoms[i]]);
+         Mesh::GeometryList elem_geoms(*mesh);
+
+         DenseTensor localP[Geometry::NumGeom];
+         for (int i = 0; i < elem_geoms.Size(); i++)
+         {
+            GetLocalRefinementMatrices(coarse_fes, elem_geoms[i],
+                                       localP[elem_geoms[i]]);
+         }
+         T.Reset(RefinementMatrix_main(coarse_fes.GetNDofs(),
+                                       coarse_fes.GetElementToDofTable(),
+                                       coarse_fes.
+                                       GetElementToFaceOrientationTable(),
+                                       localP));
+      }
+      else
+      {
+         T.Reset(VariableOrderRefinementMatrix(coarse_fes.GetNDofs(),
+                                               coarse_fes.GetElementToDofTable()));
       }
-      T.Reset(RefinementMatrix_main(coarse_fes.GetNDofs(),
-                                    coarse_fes.GetElementToDofTable(),
-                                    coarse_fes.
-                                    GetElementToFaceOrientationTable(),
-                                    localP));
    }
    else
    {
@@ -3416,19 +3614,33 @@ void FiniteElementSpace::GetTrueTransferOperator(
 
 void FiniteElementSpace::UpdateElementOrders()
 {
-   const CoarseFineTransformations &cf_tr = mesh->GetRefinementTransforms();
-
    Array<char> new_order(mesh->GetNE());
    switch (mesh->GetLastOperation())
    {
       case Mesh::REFINE:
       {
+         const CoarseFineTransformations &cf_tr = mesh->GetRefinementTransforms();
          for (int i = 0; i < mesh->GetNE(); i++)
          {
             new_order[i] = elem_order[cf_tr.embeddings[i].parent];
          }
          break;
       }
+      case Mesh::DEREFINE:
+      {
+         const CoarseFineTransformations &cf_tr =
+            mesh->ncmesh->GetDerefinementTransforms();
+         Table coarse_to_fine;
+         cf_tr.MakeCoarseToFineTable(coarse_to_fine);
+         Array<int> tabrow;
+         for (int i = 0; i < coarse_to_fine.Size(); i++)
+         {
+            coarse_to_fine.GetRow(i, tabrow);
+            //For now we require that all children are of same polynomial order.
+            new_order[i] = elem_order[tabrow[0]];
+         }
+         break;
+      }
       default:
          MFEM_ABORT("not implemented yet");
    }
@@ -3523,11 +3735,23 @@ void FiniteElementSpace::Update(bool want_transform)
          {
             BuildConformingInterpolation();
             Th.Reset(DerefinementMatrix(old_ndofs, old_elem_dof, old_elem_fos));
-            if (cP && cR)
+            if (IsVariableOrder())
             {
-               Th.SetOperatorOwner(false);
-               Th.Reset(new TripleProductOperator(cP.get(), cR.get(), Th.Ptr(),
-                                                  false, false, true));
+               if (cP && cR_hp)
+               {
+                  Th.SetOperatorOwner(false);
+                  Th.Reset(new TripleProductOperator(cP.get(), cR_hp.get(), Th.Ptr(),
+                                                     false, false, true));
+               }
+            }
+            else
+            {
+               if (cP && cR)
+               {
+                  Th.SetOperatorOwner(false);
+                  Th.Reset(new TripleProductOperator(cP.get(), cR.get(), Th.Ptr(),
+                                                     false, false, true));
+               }
             }
             break;
          }
@@ -3640,6 +3864,7 @@ FiniteElementCollection *FiniteElementSpace::Load(Mesh *m, std::istream &input)
    input >> ord;
 
    NURBSFECollection *nurbs_fec = dynamic_cast<NURBSFECollection*>(r_fec);
+   if (nurbs_fec) { nurbs_fec->SetDim(m->Dimension()); }
    NURBSExtension *nurbs_ext = NULL;
    if (fes_format == 90) // original format, v0.9
    {

fem/fespace.hpp

@@ -268,6 +268,10 @@ protected:
    Array<int> dof_elem_array, dof_ldof_array;
 
    NURBSExtension *NURBSext;
+   /** array of NURBS extension for H(div) and H(curl) vector elements.
+       For each direction an extension is created from the base NURBSext,
+       with an increase in order in the appropriate direction. */
+   Array<NURBSExtension*> VNURBSext;
    int own_ext;
    mutable Array<int> face_to_be; // NURBS FE space only
 
@@ -469,6 +473,11 @@ protected:
                                        const Table *coarse_elem_fos,
                                        const DenseTensor localP[]) const;
 
+   /* This method returns the Refinement matrix (i.e., the embedding)
+      from a coarse variable-order fes to a fine fes (after a geometric refinement) */
+   SparseMatrix *VariableOrderRefinementMatrix(const int coarse_ndofs,
+                                               const Table &coarse_elem_dof) const;
+
    void GetLocalRefinementMatrices(Geometry::Type geom,
                                    DenseTensor &localP) const;
    void GetLocalDerefinementMatrices(Geometry::Type geom,
@@ -517,6 +526,8 @@ protected:
                                                const Array<int> *perm);
 
 public:
+
+
    /** @brief Default constructor: the object is invalid until initialized using
        the method Load(). */
    FiniteElementSpace();
@@ -644,7 +655,10 @@ public:
    const ElementRestrictionOperator *GetElementRestriction(
       ElementDofOrdering e_ordering) const;
 
-   /// Return an Operator that converts L-vectors to E-vectors on each face.
+   /** @brief Return an Operator that converts L-vectors to E-vectors on each
+       face. */
+   /** @warning only meshes with tensor-product elements are currently
+       supported. */
    virtual const FaceRestriction *GetFaceRestriction(
       ElementDofOrdering f_ordering, FaceType,
       L2FaceValues mul = L2FaceValues::DoubleValued) const;

fem/gridfunc.cpp

@@ -12,6 +12,8 @@
 // Implementation of GridFunction
 
 #include "gridfunc.hpp"
+#include "linearform.hpp"
+#include "bilinearform.hpp"
 #include "quadinterpolator.hpp"
 #include "../mesh/nurbs.hpp"
 #include "../general/text.hpp"
@@ -39,7 +41,7 @@ GridFunction::GridFunction(Mesh *m, std::istream &input)
    UseDevice(true);
 
    fes = new FiniteElementSpace;
-   fec = fes->Load(m, input);
+   fec_owned = fes->Load(m, input);
 
    skip_comment_lines(input, '#');
    istream::int_type next_char = input.peek();
@@ -81,10 +83,10 @@ GridFunction::GridFunction(Mesh *m, GridFunction *gf_array[], int num_pieces)
    int vdim, ordering;
 
    fes = gf_array[0]->FESpace();
-   fec = FiniteElementCollection::New(fes->FEColl()->Name());
+   fec_owned = FiniteElementCollection::New(fes->FEColl()->Name());
    vdim = fes->GetVDim();
    ordering = fes->GetOrdering();
-   fes = new FiniteElementSpace(m, fec, vdim, ordering);
+   fes = new FiniteElementSpace(m, fec_owned, vdim, ordering);
    SetSize(fes->GetVSize());
 
    if (m->NURBSext)
@@ -153,11 +155,11 @@ GridFunction::GridFunction(Mesh *m, GridFunction *gf_array[], int num_pieces)
 
 void GridFunction::Destroy()
 {
-   if (fec)
+   if (fec_owned)
    {
       delete fes;
-      delete fec;
-      fec = NULL;
+      delete fec_owned;
+      fec_owned = NULL;
    }
 }
 
@@ -325,10 +327,9 @@ int GridFunction::VectorDim() const
    const FiniteElement *fe;
    if (!fes->GetNE())
    {
-      const FiniteElementCollection *fe_coll = fes->FEColl();
       static const Geometry::Type geoms[3] =
       { Geometry::SEGMENT, Geometry::TRIANGLE, Geometry::TETRAHEDRON };
-      fe = fe_coll->
+      fe = fes->FEColl()->
            FiniteElementForGeometry(geoms[fes->GetMesh()->Dimension()-1]);
    }
    else
@@ -350,7 +351,8 @@ int GridFunction::CurlDim() const
    {
       static const Geometry::Type geoms[3] =
       { Geometry::SEGMENT, Geometry::TRIANGLE, Geometry::TETRAHEDRON };
-      fe = fec->FiniteElementForGeometry(geoms[fes->GetMesh()->Dimension()-1]);
+      fe = fes->FEColl()->
+           FiniteElementForGeometry(geoms[fes->GetMesh()->Dimension()-1]);
    }
    else
    {
@@ -2372,19 +2374,48 @@ void GridFunction::ProjectCoefficient(Coefficient &coeff)
 
    if (delta_c == NULL)
    {
-      Array<int> vdofs;
-      Vector vals;
-
-      for (int i = 0; i < fes->GetNE(); i++)
+      if (fes->GetNURBSext() == NULL)
       {
-         doftrans = fes->GetElementVDofs(i, vdofs);
-         vals.SetSize(vdofs.Size());
-         fes->GetFE(i)->Project(coeff, *fes->GetElementTransformation(i), vals);
-         if (doftrans)
+         Array<int> vdofs;
+         Vector vals;
+
+         for (int i = 0; i < fes->GetNE(); i++)
          {
-            doftrans->TransformPrimal(vals);
+            doftrans = fes->GetElementVDofs(i, vdofs);
+            vals.SetSize(vdofs.Size());
+            fes->GetFE(i)->Project(coeff, *fes->GetElementTransformation(i), vals);
+            if (doftrans)
+            {
+               doftrans->TransformPrimal(vals);
+            }
+            SetSubVector(vdofs, vals);
          }
-         SetSubVector(vdofs, vals);
+      }
+      else
+      {
+         // Define and assemble linear form
+         LinearForm b(fes);
+         b.AddDomainIntegrator(new DomainLFIntegrator(coeff));
+         b.Assemble();
+
+         // Define and assemble bilinear form
+         BilinearForm a(fes);
+         a.AddDomainIntegrator(new MassIntegrator());
+         a.Assemble();
+
+         // Set solver and preconditioner
+         SparseMatrix A(a.SpMat());
+         GSSmoother  prec(A);
+         CGSolver cg;
+         cg.SetOperator(A);
+         cg.SetPreconditioner(prec);
+         cg.SetRelTol(1e-12);
+         cg.SetMaxIter(1000);
+         cg.SetPrintLevel(0);
+
+         // Solve and get solution
+         *this = 0.0;
+         cg.Mult(b,*this);
       }
    }
    else
@@ -2425,22 +2456,54 @@ void GridFunction::ProjectCoefficient(
 
 void GridFunction::ProjectCoefficient(VectorCoefficient &vcoeff)
 {
-   int i;
-   Array<int> vdofs;
-   Vector vals;
-
-   DofTransformation * doftrans = NULL;
-
-   for (i = 0; i < fes->GetNE(); i++)
+   if (fes->GetNURBSext() == NULL)
    {
-      doftrans = fes->GetElementVDofs(i, vdofs);
-      vals.SetSize(vdofs.Size());
-      fes->GetFE(i)->Project(vcoeff, *fes->GetElementTransformation(i), vals);
-      if (doftrans)
+
+      int i;
+      Array<int> vdofs;
+      Vector vals;
+
+      DofTransformation * doftrans = NULL;
+
+      for (i = 0; i < fes->GetNE(); i++)
       {
-         doftrans->TransformPrimal(vals);
+         doftrans = fes->GetElementVDofs(i, vdofs);
+         vals.SetSize(vdofs.Size());
+         fes->GetFE(i)->Project(vcoeff, *fes->GetElementTransformation(i), vals);
+         if (doftrans)
+         {
+            doftrans->TransformPrimal(vals);
+         }
+         SetSubVector(vdofs, vals);
       }
-      SetSubVector(vdofs, vals);
+
+   }
+
+   else
+   {
+      // Define and assemble linear form
+      LinearForm b(fes);
+      b.AddDomainIntegrator(new VectorFEDomainLFIntegrator(vcoeff));
+      b.Assemble();
+
+      // Define and assemble bilinear form
+      BilinearForm a(fes);
+      a.AddDomainIntegrator(new VectorFEMassIntegrator());
+      a.Assemble();
+
+      // Set solver and preconditioner
+      SparseMatrix A(a.SpMat());
+      GSSmoother  prec(A);
+      CGSolver cg;
+      cg.SetOperator(A);
+      cg.SetPreconditioner(prec);
+      cg.SetRelTol(1e-12);
+      cg.SetMaxIter(1000);
+      cg.SetPrintLevel(0);
+
+      // Solve and get solution
+      *this = 0.0;
+      cg.Mult(b,*this);
    }
 }
 
@@ -3926,7 +3989,7 @@ void GridFunction::LegacyNCReorder()
       mesh->GetEdgeVertices(i, ev);
       if (old_vertex[ev[0]] > old_vertex[ev[1]])
       {
-         const int *ind = fec->DofOrderForOrientation(Geometry::SEGMENT, -1);
+         const int *ind = fes->FEColl()->DofOrderForOrientation(Geometry::SEGMENT, -1);
 
          fes->GetEdgeInteriorDofs(i, dofs);
          for (int k = 0; k < dofs.Size(); k++)
@@ -4518,6 +4581,11 @@ GridFunction *Extrude1DGridFunction(Mesh *mesh, Mesh *mesh2d,
    {
       solfec2d = new L2_FECollection(atoi(name + 7), 2);
    }
+   else if (!strncmp(name, "L2Int_", 6))
+   {
+      solfec2d = new L2_FECollection(atoi(name + 7), 2, BasisType::GaussLegendre,
+                                     FiniteElement::INTEGRAL);
+   }
    else
    {
       mfem::err << "Extrude1DGridFunction : unknown FE collection : "

fem/gridfunc.hpp

@@ -30,14 +30,14 @@ namespace mfem
 class GridFunction : public Vector
 {
 protected:
-   /// FE space on which the grid function lives. Owned if #fec is not NULL.
+   /// FE space on which the grid function lives. Owned if #fec_owned is not NULL.
    FiniteElementSpace *fes;
 
    /** @brief Used when the grid function is read from a file. It can also be
        set explicitly, see MakeOwner().
 
        If not NULL, this pointer is owned by the GridFunction. */
-   FiniteElementCollection *fec;
+   FiniteElementCollection *fec_owned;
 
    long fes_sequence; // see FiniteElementSpace::sequence, Mesh::sequence
 
@@ -72,16 +72,16 @@ protected:
 
 public:
 
-   GridFunction() { fes = NULL; fec = NULL; fes_sequence = 0; UseDevice(true); }
+   GridFunction() { fes = NULL; fec_owned = NULL; fes_sequence = 0; UseDevice(true); }
 
    /// Copy constructor. The internal true-dof vector #t_vec is not copied.
    GridFunction(const GridFunction &orig)
-      : Vector(orig), fes(orig.fes), fec(NULL), fes_sequence(orig.fes_sequence)
+      : Vector(orig), fes(orig.fes), fec_owned(NULL), fes_sequence(orig.fes_sequence)
    { UseDevice(true); }
 
    /// Construct a GridFunction associated with the FiniteElementSpace @a *f.
    GridFunction(FiniteElementSpace *f) : Vector(f->GetVSize())
-   { fes = f; fec = NULL; fes_sequence = f->GetSequence(); UseDevice(true); }
+   { fes = f; fec_owned = NULL; fes_sequence = f->GetSequence(); UseDevice(true); }
 
    /// Construct a GridFunction using previously allocated array @a data.
    /** The GridFunction does not assume ownership of @a data which is assumed to
@@ -91,13 +91,13 @@ public:
     */
    GridFunction(FiniteElementSpace *f, real_t *data)
       : Vector(data, f->GetVSize())
-   { fes = f; fec = NULL; fes_sequence = f->GetSequence(); UseDevice(true); }
+   { fes = f; fec_owned = NULL; fes_sequence = f->GetSequence(); UseDevice(true); }
 
    /** @brief Construct a GridFunction using previously allocated Vector @a base
        starting at the given offset, @a base_offset. */
    GridFunction(FiniteElementSpace *f, Vector &base, int base_offset = 0)
       : Vector(base, base_offset, f->GetVSize())
-   { fes = f; fec = NULL; fes_sequence = f->GetSequence(); UseDevice(true); }
+   { fes = f; fec_owned = NULL; fes_sequence = f->GetSequence(); UseDevice(true); }
 
    /// Construct a GridFunction on the given Mesh, using the data from @a input.
    /** The content of @a input should be in the format created by the method
@@ -116,12 +116,12 @@ public:
    GridFunction &operator=(const GridFunction &rhs)
    { return operator=((const Vector &)rhs); }
 
-   /// Make the GridFunction the owner of #fec and #fes.
-   /** If the new FiniteElementCollection, @a fec_, is NULL, ownership of #fec
+   /// Make the GridFunction the owner of #fec_owned and #fes.
+   /** If the new FiniteElementCollection, @a fec_, is NULL, ownership of #fec_owned
        and #fes is taken away. */
-   void MakeOwner(FiniteElementCollection *fec_) { fec = fec_; }
+   void MakeOwner(FiniteElementCollection *fec_) { fec_owned = fec_; }
 
-   FiniteElementCollection *OwnFEC() { return fec; }
+   FiniteElementCollection *OwnFEC() { return fec_owned; }
 
    int VectorDim() const;
    int CurlDim() const;
@@ -387,7 +387,8 @@ public:
    /** @brief Project @a coeff Coefficient to @a this GridFunction. The
        projection computation depends on the choice of the FiniteElementSpace
        #fes. Note that this is usually interpolation at the degrees of freedom
-       in each element (not L2 projection). */
+       in each element (not L2 projection). For NURBS spaces these degrees of
+       freedom are not available and L2 projection is resorted to as fallback. */
    virtual void ProjectCoefficient(Coefficient &coeff);
 
    /** @brief Project @a coeff Coefficient to @a this GridFunction, using one
@@ -398,7 +399,8 @@ public:
    /** @brief Project @a vcoeff VectorCoefficient to @a this GridFunction. The
        projection computation depends on the choice of the FiniteElementSpace
        #fes. Note that this is usually interpolation at the degrees of freedom
-       in each element (not L2 projection).*/
+       in each element (not L2 projection). For NURBS spaces these degrees of
+       freedom are not available and L2 projection is resorted to as fallback. */
    void ProjectCoefficient(VectorCoefficient &vcoeff);
 
    /** @brief Project @a vcoeff VectorCoefficient to @a this GridFunction, using

fem/gslib.cpp

@@ -1168,6 +1168,120 @@ void FindPointsGSLIB::InterpolateGeneral(const GridFunction &field_in,
    } // parallel
 }
 
+void FindPointsGSLIB::DistributePointInfoToOwningMPIRanks(
+   Array<unsigned int> &recv_elem, Vector &recv_ref,
+   Array<unsigned int> &recv_code)
+{
+   MFEM_VERIFY(points_cnt,
+               "Invalid size. Please make sure to call FindPoints method "
+               "before calling this function.");
+
+   // Pack data to send via crystal router
+   struct gslib::array *outpt = new gslib::array;
+
+   struct out_pt { double rst[3]; uint index, elem, proc, code; };
+   struct out_pt *pt;
+   array_init(struct out_pt, outpt, points_cnt);
+   outpt->n=points_cnt;
+   pt = (struct out_pt *)outpt->ptr;
+
+   for (int index = 0; index < points_cnt; index++)
+   {
+      pt->index = index;
+      pt->elem = gsl_mfem_elem[index];
+      pt->proc  = gsl_proc[index];
+      pt->code = gsl_code[index];
+      for (int d = 0; d < dim; ++d)
+      {
+         pt->rst[d]= gsl_mfem_ref(index*dim + d);
+      }
+      ++pt;
+   }
+
+   // Transfer data to target MPI ranks
+   sarray_transfer(struct out_pt, outpt, proc, 1, cr);
+
+   // Store received data
+   const int points_recv = outpt->n;
+   recv_proc.SetSize(points_recv);
+   recv_elem.SetSize(points_recv);
+   recv_index.SetSize(points_recv);
+   recv_code.SetSize(points_recv);
+   recv_ref.SetSize(points_recv*dim);
+
+   pt = (struct out_pt *)outpt->ptr;
+   for (int index = 0; index < points_recv; index++)
+   {
+      recv_index[index] = pt->index;
+      recv_elem[index] = pt->elem;
+      recv_proc[index] = pt->proc;
+      recv_code[index] = pt->code;
+      for (int d = 0; d < dim; ++d)
+      {
+         recv_ref(index*dim + d)= pt->rst[d];
+      }
+      ++pt;
+   }
+
+   array_free(outpt);
+   delete outpt;
+}
+
+void FindPointsGSLIB::DistributeInterpolatedValues(const Vector &int_vals,
+                                                   const int vdim,
+                                                   const int ordering,
+                                                   Vector &field_out) const
+{
+   const int points_recv = recv_index.Size();;
+   MFEM_VERIFY(points_recv == 0 ||
+               int_vals.Size() % points_recv == 0,
+               "Incompatible size. Please return interpolated values"
+               "corresponding to points received using"
+               "SendCoordinatesToOwningProcessors.");
+   field_out.SetSize(points_cnt*vdim);
+
+   for (int v = 0; v < vdim; v++)
+   {
+      // Pack data to send via crystal router
+      struct gslib::array *outpt = new gslib::array;
+      struct out_pt { double val; uint index, proc; };
+      struct out_pt *pt;
+      array_init(struct out_pt, outpt, points_recv);
+      outpt->n=points_recv;
+      pt = (struct out_pt *)outpt->ptr;
+      for (int index = 0; index < points_recv; index++)
+      {
+         pt->index = recv_index[index];
+         pt->proc  = recv_proc[index];
+         pt->val = ordering == Ordering::byNODES ?
+                   int_vals(index + v*points_recv) :
+                   int_vals(index*vdim + v);
+         ++pt;
+      }
+
+      // Transfer data to target MPI ranks
+      sarray_transfer(struct out_pt, outpt, proc, 1, cr);
+
+      // Store received data
+      MFEM_VERIFY(outpt->n == points_cnt, "Incompatible size. Number of points "
+                  "received does not match the number of points originally "
+                  "found using FindPoints.");
+
+      pt = (struct out_pt *)outpt->ptr;
+      for (int index = 0; index < points_cnt; index++)
+      {
+         int idx = ordering == Ordering::byNODES ?
+                   pt->index + v*points_cnt :
+                   pt->index*vdim + v;
+         field_out(idx) = pt->val;
+         ++pt;
+      }
+
+      array_free(outpt);
+      delete outpt;
+   }
+}
+
 void OversetFindPointsGSLIB::Setup(Mesh &m, const int meshid,
                                    GridFunction *gfmax,
                                    const double bb_t, const double newt_tol,

fem/gslib.hpp

@@ -34,7 +34,7 @@ namespace mfem
  *
  *  There are three key functions in FindPointsGSLIB:
  *
- *  1. Setup - constructs the internal data structures of gslib.
+ *  1. Setup - constructs the internal data structures of gslib. See \ref Setup.
  *
  *  2. FindPoints - for any given arbitrary set of points in physical space,
  *     gslib finds the element number, MPI rank, and the reference space
@@ -45,12 +45,23 @@ namespace mfem
  *     on an element edge/face or near the domain boundary, and gslib also
  *     returns a distance to the border. Points near (but outside) the domain
  *     boundary must then be marked as not found using the distance returned
- *     by gslib.
+ *     by gslib. See \ref FindPoints.
  *
  *  3. Interpolate - Interpolates any grid function at the points found using 2.
+ *     For functions in L2 finite element space, use \ref SetL2AvgType to
+ *     specify how to interpolate values at points located at element boundaries
+ *     where the function might be multi-valued. See \ref Interpolate.
  *
- *  FindPointsGSLIB provides interface to use these functions individually or
- *  using a single call.
+ *  FindPointsGSLIB also provides interface to use these functions through a
+ *  single call.
+ *
+ *  For custom interpolation (e.g., evaluating strain rate tensor), we provide
+ *  functions that use gslib to send element index and corresponding
+ *  reference-space coordinates for each point to the mpi rank that the element
+ *  is located on. Then, custom interpolation can be defined locally by the user
+ *  before sending the values back to mpi ranks where the query originated from.
+ *  See \ref DistributePointInfoToOwningMPIRanks and
+ *  \ref DistributeInterpolatedValues.
  */
 class FindPointsGSLIB
 {
@@ -74,7 +85,8 @@ protected:
    int dim, points_cnt;
    Array<unsigned int> gsl_code, gsl_proc, gsl_elem, gsl_mfem_elem;
    Vector gsl_mesh, gsl_ref, gsl_dist, gsl_mfem_ref;
-   bool setupflag;              // flag to indicate whether gslib data has been setup
+   Array<unsigned int> recv_proc, recv_index; // data for custom interpolation
+   bool setupflag;              // flag to indicate if gslib data has been setup
    double default_interp_value; // used for points that are not found in the mesh
    AvgType avgtype;             // average type used for L2 functions
    Array<int> split_element_map;
@@ -118,9 +130,9 @@ public:
    virtual ~FindPointsGSLIB();
 
    /** Initializes the internal mesh in gslib, by sending the positions of the
-       Gauss-Lobatto nodes of the input Mesh object @a m.
+       Gauss-Lobatto nodes of the input Mesh object \p m.
        Note: not tested with periodic (L2).
-       Note: the input mesh @a m must have Nodes set.
+       Note: the input mesh \p m must have Nodes set.
 
        @param[in] m         Input mesh.
        @param[in] bb_t      (Optional) Relative size of bounding box around
@@ -133,9 +145,9 @@ public:
    void Setup(Mesh &m, const double bb_t = 0.1,
               const double newt_tol = 1.0e-12,
               const int npt_max = 256);
-   /** Searches positions given in physical space by @a point_pos.
+   /** Searches positions given in physical space by \p point_pos.
        These positions can be ordered byNodes: (XXX...,YYY...,ZZZ) or
-       byVDim: (XYZ,XYZ,....XYZ) specified by @a point_pos_ordering.
+       byVDim: (XYZ,XYZ,....XYZ) specified by \p point_pos_ordering.
        This function populates the following member variables:
        #gsl_code        Return codes for each point: inside element (0),
                         element boundary (1), not found (2).
@@ -164,20 +176,20 @@ public:
    /** Interpolation of field values at prescribed reference space positions.
        @param[in] field_in    Function values that will be interpolated on the
                               reference positions. Note: it is assumed that
-                              @a field_in is in H1 and in the same space as the
+                              \p field_in is in H1 and in the same space as the
                               mesh that was given to Setup().
        @param[out] field_out  Interpolated values. For points that are not found
                               the value is set to #default_interp_value. */
    virtual void Interpolate(const GridFunction &field_in, Vector &field_out);
    /** Search positions and interpolate. The ordering (byNODES or byVDIM) of
-       the output values in @a field_out corresponds to the ordering used
-       in the input GridFunction @a field_in. */
+       the output values in \p field_out corresponds to the ordering used
+       in the input GridFunction \p field_in. */
    void Interpolate(const Vector &point_pos, const GridFunction &field_in,
                     Vector &field_out,
                     int point_pos_ordering = Ordering::byNODES);
    /** Setup FindPoints, search positions and interpolate. The ordering (byNODES
-       or byVDIM) of the output values in @a field_out corresponds to the
-       ordering used in the input GridFunction @a field_in. */
+       or byVDIM) of the output values in \p field_out corresponds to the
+       ordering used in the input GridFunction \p field_in. */
    void Interpolate(Mesh &m, const Vector &point_pos,
                     const GridFunction &field_in, Vector &field_out,
                     int point_pos_ordering = Ordering::byNODES);
@@ -225,6 +237,41 @@ public:
    /// Return reference coordinates in [-1,1] (internal range in GSLIB) for each
    /// point found by FindPoints.
    virtual const Vector &GetGSLIBReferencePosition() const { return gsl_ref; }
+
+   /** @name Methods to support a custom interpolation procedure.
+       \brief The physical-space point that the user seeks to interpolate at
+       could be located inside an element on another mpi rank.
+       To enable a custom interpolation procedure (e.g., strain tensor computation)
+       we need a mechanism to first send element indices and reference-space
+       coordinates to the mpi-ranks where each point is found. Then the custom
+       interpolation can be done locally by the user before sending the
+       interpolated values back to the mpi-ranks that the query originated from.
+       Example usage looks something like this:
+
+       FindPoints() -> DistributePointInfoToOwningMPIRanks() -> Computation by
+       user -> DistributeInterpolatedValues().
+   */
+   ///@{
+   /// Distribute element indices in #gsl_mfem_elem, the reference coordinates
+   /// #gsl_mfem_ref, and the code #gsl_code to the corresponding mpi-rank
+   /// #gsl_proc for each point. The received information is provided locally
+   /// in \p recv_elem, \p recv_ref (ordered by vdim), and \p recv_code.
+   /// Note: The user can send empty Array/Vectors to the method as they are
+   /// appropriately sized and filled internally.
+   virtual void DistributePointInfoToOwningMPIRanks(
+      Array<unsigned int> &recv_elem, Vector &recv_ref,
+      Array<unsigned int> &recv_code);
+   /// Return interpolated values back to the mpi-ranks #recv_proc that had
+   /// sent the element indices and corresponding reference-space coordinates.
+   /// Specify \p vdim and \p ordering (by nodes or by vdim) based on how the
+   /// \p int_vals are structured. The received values are filled in
+   /// \p field_out consistent with the original ordering of the points that
+   /// were used in \ref FindPoints.
+   virtual void DistributeInterpolatedValues(const Vector &int_vals,
+                                             const int vdim,
+                                             const int ordering,
+                                             Vector &field_out) const;
+   ///@}
 };
 
 /** \brief OversetFindPointsGSLIB enables use of findpts for arbitrary number of
@@ -249,9 +296,9 @@ public:
 #endif
 
    /** Initializes the internal mesh in gslib, by sending the positions of the
-       Gauss-Lobatto nodes of the input Mesh object @a m.
+       Gauss-Lobatto nodes of the input Mesh object \p m.
        Note: not tested with periodic meshes (L2).
-       Note: the input mesh @a m must have Nodes set.
+       Note: the input mesh \p m must have Nodes set.
 
        @param[in] m         Input mesh.
        @param[in] meshid    A unique # for each overlapping mesh. This id is
@@ -274,12 +321,12 @@ public:
               const double bb_t = 0.1, const double newt_tol = 1.0e-12,
               const int npt_max = 256);
 
-   /** Searches positions given in physical space by @a point_pos. All output
+   /** Searches positions given in physical space by \p point_pos. All output
        Arrays and Vectors are expected to have the correct size.
 
        @param[in]  point_pos           Positions to be found.
        @param[in]  point_id            Index of the mesh that the point belongs
-                                       to (corresponding to @a meshid in Setup).
+                                       to (corresponding to \p meshid in Setup).
        @param[in]  point_pos_ordering  Ordering of the points:
                                        byNodes: (XXX...,YYY...,ZZZ) or
                                        byVDim: (XYZ,XYZ,....XYZ) */
@@ -342,7 +389,7 @@ public:
    enum GSOp {ADD, MUL, MIN, MAX};
 
    /// Update the identifiers used for the gather-scatter operator.
-   /// Same @a ids get grouped together and id == 0 does not participate.
+   /// Same \p ids get grouped together and id == 0 does not participate.
    /// See class description.
    void UpdateIdentifiers(const Array<long long> &ids);
 

fem/hybridization.cpp

@@ -29,8 +29,8 @@ namespace mfem
 
 Hybridization::Hybridization(FiniteElementSpace *fespace,
                              FiniteElementSpace *c_fespace)
-   : fes(fespace), c_fes(c_fespace), c_bfi(NULL), Ct(NULL), H(NULL),
-     Af_data(NULL), Af_ipiv(NULL)
+   : fes(fespace), c_fes(c_fespace), c_bfi(NULL), extern_bdr_constr_integs(0),
+     Ct(NULL), H(NULL), Af_data(NULL), Af_ipiv(NULL)
 {
 #ifdef MFEM_USE_MPI
    pC = P_pc = NULL;
@@ -49,6 +49,11 @@ Hybridization::~Hybridization()
    delete H;
    delete Ct;
    delete c_bfi;
+   if (!extern_bdr_constr_integs)
+   {
+      for (int k=0; k < boundary_constraint_integs.Size(); k++)
+      { delete boundary_constraint_integs[k]; }
+   }
 }
 
 void Hybridization::ConstructC()
@@ -57,6 +62,15 @@ void Hybridization::ConstructC()
    int num_hat_dofs = hat_offsets[NE];
    Array<int> vdofs, c_vdofs;
 
+#if defined(MFEM_USE_DOUBLE)
+   constexpr real_t mtol = 1e-12;
+#elif defined(MFEM_USE_SINGLE)
+   constexpr real_t mtol = 4e-6;
+#else
+#error "Only single and double precision are supported!"
+   constexpr real_t mtol = 1.;
+#endif
+
    int c_num_face_nbr_dofs = 0;
 #ifdef MFEM_USE_MPI
    ParFiniteElementSpace *c_pfes = dynamic_cast<ParFiniteElementSpace*>(c_fes);
@@ -125,9 +139,73 @@ void Hybridization::ConstructC()
                                    *fes->GetFE(FTr->Elem2No),
                                    *FTr, elmat);
          // zero-out small elements in elmat
-         elmat.Threshold(1e-12 * elmat.MaxMaxNorm());
+         elmat.Threshold(mtol * elmat.MaxMaxNorm());
          Ct->AddSubMatrix(vdofs, c_vdofs, elmat, skip_zeros);
       }
+
+      if (boundary_constraint_integs.Size())
+      {
+         const FiniteElement *fe1, *fe2;
+         const FiniteElement *face_el;
+
+         // Which 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 < boundary_constraint_integs.Size(); k++)
+         {
+            if (boundary_constraint_integs_marker[k] == NULL)
+            {
+               bdr_attr_marker = 1;
+               break;
+            }
+            Array<int> &bdr_marker = *boundary_constraint_integs_marker[k];
+            MFEM_ASSERT(bdr_marker.Size() == bdr_attr_marker.Size(),
+                        "invalid boundary marker for boundary face integrator #"
+                        << k << ", counting from zero");
+            for (int i = 0; i < bdr_attr_marker.Size(); i++)
+            {
+               bdr_attr_marker[i] |= bdr_marker[i];
+            }
+         }
+
+         for (int i = 0; i < fes->GetNBE(); i++)
+         {
+            const int bdr_attr = mesh->GetBdrAttribute(i);
+            if (bdr_attr_marker[bdr_attr-1] == 0) { continue; }
+
+            FTr = mesh->GetBdrFaceTransformations(i);
+            if (!FTr) { continue; }
+
+            int o1 = hat_offsets[FTr->Elem1No];
+            int s1 = hat_offsets[FTr->Elem1No+1] - o1;
+
+            vdofs.SetSize(s1);
+            for (int j = 0; j < s1; j++)
+            {
+               vdofs[j] = o1 + j;
+            }
+            int iface = mesh->GetBdrElementFaceIndex(i);
+            c_fes->GetFaceVDofs(iface, c_vdofs);
+            face_el = c_fes->GetFaceElement(iface);
+            fe1 = fes -> GetFE (FTr -> Elem1No);
+            // The fe2 object is really a dummy and not used on the boundaries,
+            // but we can't dereference a NULL pointer, and we don't want to
+            // actually make a fake element.
+            fe2 = fe1;
+            for (int k = 0; k < boundary_constraint_integs.Size(); k++)
+            {
+               if (boundary_constraint_integs_marker[k] &&
+                   (*boundary_constraint_integs_marker[k])[bdr_attr-1] == 0) { continue; }
+
+               boundary_constraint_integs[k]->AssembleFaceMatrix(*face_el, *fe1, *fe2, *FTr,
+                                                                 elmat);
+               // zero-out small elements in elmat
+               elmat.Threshold(mtol * elmat.MaxMaxNorm());
+               Ct->AddSubMatrix(vdofs, c_vdofs, elmat, skip_zeros);
+            }
+         }
+      }
 #ifdef MFEM_USE_MPI
       if (pmesh)
       {
@@ -167,7 +245,7 @@ void Hybridization::ConstructC()
             fe = fes->GetFE(FTr->Elem1No);
             c_bfi->AssembleFaceMatrix(*face_fe, *fe, *fe, *FTr, elmat);
             // zero-out small elements in elmat
-            elmat.Threshold(1e-12 * elmat.MaxMaxNorm());
+            elmat.Threshold(mtol * elmat.MaxMaxNorm());
             Ct->AddSubMatrix(vdofs, c_vdofs, elmat, skip_zeros);
          }
          if (glob_num_shared_slave_faces)

fem/hybridization.hpp

@@ -64,6 +64,12 @@ protected:
    FiniteElementSpace *fes, *c_fes;
    BilinearFormIntegrator *c_bfi;
 
+   /// Set of constraint boundary face integrators to be applied.
+   Array<BilinearFormIntegrator*> boundary_constraint_integs;
+   Array<Array<int>*>             boundary_constraint_integs_marker;
+   /// Indicates if the boundary_constraint_integs integrators are owned externally
+   bool extern_bdr_constr_integs;
+
    SparseMatrix *Ct, *H;
 
    Array<int> hat_offsets, hat_dofs_marker;
@@ -106,6 +112,33 @@ public:
    void SetConstraintIntegrator(BilinearFormIntegrator *c_integ)
    { delete c_bfi; c_bfi = c_integ; }
 
+   /** Add the boundary face integrator that will be used to construct the
+       constraint matrix C. The Hybridization object assumes ownership of the
+       integrator, i.e. it will delete the integrator when destroyed. */
+   void AddBdrConstraintIntegrator(BilinearFormIntegrator *c_integ)
+   {
+      boundary_constraint_integs.Append(c_integ);
+      boundary_constraint_integs_marker.Append(
+         NULL); // NULL marker means apply everywhere
+   }
+   void AddBdrConstraintIntegrator(BilinearFormIntegrator *c_integ,
+                                   Array<int> &bdr_marker)
+   {
+      boundary_constraint_integs.Append(c_integ);
+      boundary_constraint_integs_marker.Append(&bdr_marker);
+   }
+
+   /// Access all integrators added with AddBdrConstraintIntegrator().
+   Array<BilinearFormIntegrator*> *GetBCBFI() { return &boundary_constraint_integs; }
+
+   /// Access all boundary markers added with AddBdrConstraintIntegrator().
+   /** If no marker was specified when the integrator was added, the
+       corresponding pointer (to Array<int>) will be NULL. */
+   Array<Array<int>*> *GetBCBFI_Marker() { return &boundary_constraint_integs_marker; }
+
+   /// Indicate that boundary constraint integrators are not owned
+   void UseExternalBdrConstraintIntegrators() { extern_bdr_constr_integs = true; }
+
    /// Prepare the Hybridization object for assembly.
    void Init(const Array<int> &ess_tdof_list);
 

fem/intrules_cut.cpp

@@ -220,8 +220,8 @@ void MomentFittingIntRules::ComputeSurfaceWeights1D(ElementTransformation& Tr)
    {
       IntegrationPoint ip2;
       ip2.x = .5;
-      while (LvlSet->Eval(Tr, ip2) > 1e-12
-             || LvlSet->Eval(Tr, ip2) < -1e-12)
+      while (LvlSet->Eval(Tr, ip2) > tol_1
+             || LvlSet->Eval(Tr, ip2) < -tol_1)
       {
          if (LvlSet->Eval(Tr, ip0) * LvlSet->Eval(Tr, ip2) < 0.)
          {
@@ -237,12 +237,12 @@ void MomentFittingIntRules::ComputeSurfaceWeights1D(ElementTransformation& Tr)
       intp.x = ip2.x;
       intp.weight = 1. / Tr.Weight();
    }
-   else if (LvlSet->Eval(Tr, ip0) > 0. && LvlSet->Eval(Tr, ip1) <= 1e-12)
+   else if (LvlSet->Eval(Tr, ip0) > 0. && LvlSet->Eval(Tr, ip1) <= tol_1)
    {
       intp.x = 1.;
       intp.weight = 1. / Tr.Weight();
    }
-   else if (LvlSet->Eval(Tr, ip1) > 0. && LvlSet->Eval(Tr, ip0) <= 1e-12)
+   else if (LvlSet->Eval(Tr, ip1) > 0. && LvlSet->Eval(Tr, ip0) <= tol_1)
    {
       intp.x = 0.;
       intp.weight = 1. / Tr.Weight();
@@ -290,8 +290,8 @@ void MomentFittingIntRules::ComputeVolumeWeights1D(ElementTransformation& Tr,
          }
       }
    }
-   else if (LvlSet->Eval(Tr, ip0) <= -1e-12
-            || LvlSet->Eval(Tr, ip1) <= -1e-12)
+   else if (LvlSet->Eval(Tr, ip0) <= -tol_1
+            || LvlSet->Eval(Tr, ip1) <= -tol_1)
    {
       for (int ip = 0; ip < ir.GetNPoints(); ip++)
       {
@@ -356,24 +356,24 @@ void MomentFittingIntRules::ComputeSurfaceWeights2D(ElementTransformation& Tr)
       IntegrationPoint ipB;
       Trafo.TransformBack(pointB, ipB);
 
-      if (LvlSet->Eval(Trafo, ipA) < -1e-12
-          || LvlSet->Eval(Trafo, ipB) < -1e-12)
+      if (LvlSet->Eval(Trafo, ipA) < -tol_1
+          || LvlSet->Eval(Trafo, ipB) < -tol_1)
       {
          interior = false;
       }
 
-      if (LvlSet->Eval(Trafo, ipA) > -1e-12
-          && LvlSet->Eval(Trafo, ipB) > -1e-12)
+      if (LvlSet->Eval(Trafo, ipA) > -tol_1
+          && LvlSet->Eval(Trafo, ipB) > -tol_1)
       {
          layout = Layout::inside;
       }
-      else if (LvlSet->Eval(Trafo, ipA) > 1e-15
+      else if (LvlSet->Eval(Trafo, ipA) > tol_2
                && LvlSet->Eval(Trafo, ipB) <= 0.)
       {
          layout = Layout::intersected;
       }
       else if (LvlSet->Eval(Trafo, ipA) <= 0.
-               && LvlSet->Eval(Trafo, ipB) > 1e-15)
+               && LvlSet->Eval(Trafo, ipB) > tol_2)
       {
          layout = Layout::intersected;
          Vector temp(pointA.Size());
@@ -399,10 +399,10 @@ void MomentFittingIntRules::ComputeSurfaceWeights2D(ElementTransformation& Tr)
          IntegrationPoint ip;
          Trafo.TransformBack(mid, ip);
 
-         while (LvlSet->Eval(Trafo, ip) > 1e-12
-                || LvlSet->Eval(Trafo, ip) < -1e-12)
+         while (LvlSet->Eval(Trafo, ip) > tol_1
+                || LvlSet->Eval(Trafo, ip) < -tol_1)
          {
-            if (LvlSet->Eval(Trafo, ip) > 1e-12)
+            if (LvlSet->Eval(Trafo, ip) > tol_1)
             {
                pointC = mid;
             }
@@ -539,7 +539,7 @@ void MomentFittingIntRules::ComputeSurfaceWeights2D(ElementTransformation& Tr)
       temp2 = 0.;
       for (int i = 0; i < nBasis; i++)
       {
-         if (SVD.Singularvalue(i) > 1e-12)
+         if (SVD.Singularvalue(i) > tol_1)
          {
             temp2(i) = temp(i) / SVD.Singularvalue(i);
          }
@@ -606,24 +606,24 @@ void MomentFittingIntRules::ComputeVolumeWeights2D(ElementTransformation& Tr,
       IntegrationPoint ipB;
       Trafo.TransformBack(pointB, ipB);
 
-      if (LvlSet->Eval(Trafo, ipA) < -1e-12
-          || LvlSet->Eval(Trafo, ipB) < -1e-12)
+      if (LvlSet->Eval(Trafo, ipA) < -tol_1
+          || LvlSet->Eval(Trafo, ipB) < -tol_1)
       {
          interior = false;
       }
 
-      if (LvlSet->Eval(Trafo, ipA) > -1e-12
-          && LvlSet->Eval(Trafo, ipB) > -1e-12)
+      if (LvlSet->Eval(Trafo, ipA) > -tol_1
+          && LvlSet->Eval(Trafo, ipB) > -tol_1)
       {
          layout = Layout::inside;
       }
-      else if (LvlSet->Eval(Trafo, ipA) > 1e-15
+      else if (LvlSet->Eval(Trafo, ipA) > tol_2
                && LvlSet->Eval(Trafo, ipB) <= 0.)
       {
          layout = Layout::intersected;
       }
       else if (LvlSet->Eval(Trafo, ipA) <= 0.
-               && LvlSet->Eval(Trafo, ipB) > 1e-15)
+               && LvlSet->Eval(Trafo, ipB) > tol_2)
       {
          layout = Layout::intersected;
          Vector temp(pointA.Size());
@@ -648,10 +648,10 @@ void MomentFittingIntRules::ComputeVolumeWeights2D(ElementTransformation& Tr,
          IntegrationPoint ip;
          Trafo.TransformBack(mid, ip);
 
-         while (LvlSet->Eval(Trafo, ip) > 1e-12
-                || LvlSet->Eval(Trafo, ip) < -1e-12)
+         while (LvlSet->Eval(Trafo, ip) > tol_1
+                || LvlSet->Eval(Trafo, ip) < -tol_1)
          {
-            if (LvlSet->Eval(Trafo, ip) > 1e-12)
+            if (LvlSet->Eval(Trafo, ip) > tol_1)
             {
                pointC = mid;
             }
@@ -786,7 +786,7 @@ void MomentFittingIntRules::ComputeVolumeWeights2D(ElementTransformation& Tr,
       VolumeSVD->LeftSingularvectors().MultTranspose(RHS, temp);
       for (int i = 0; i < nBasisVolume; i++)
       {
-         if (VolumeSVD->Singularvalue(i) > 1e-12)
+         if (VolumeSVD->Singularvalue(i) > tol_1)
          {
             temp2(i) = temp(i) / VolumeSVD->Singularvalue(i);
          }
@@ -865,18 +865,18 @@ void MomentFittingIntRules::ComputeSurfaceWeights3D(ElementTransformation& Tr)
       IntegrationPoint ipD;
       Trafo.TransformBack(pointD, ipD);
 
-      if (LvlSet->Eval(Trafo, ipA) < -1e-12
-          || LvlSet->Eval(Trafo, ipB) < -1e-12
-          || LvlSet->Eval(Trafo, ipC) < -1e-12
-          || LvlSet->Eval(Trafo, ipD) < -1e-12)
+      if (LvlSet->Eval(Trafo, ipA) < -tol_1
+          || LvlSet->Eval(Trafo, ipB) < -tol_1
+          || LvlSet->Eval(Trafo, ipC) < -tol_1
+          || LvlSet->Eval(Trafo, ipD) < -tol_1)
       {
          interior = false;
       }
 
-      if (LvlSet->Eval(Trafo, ipA) > -1e-12
-          || LvlSet->Eval(Trafo, ipB) > -1e-12
-          || LvlSet->Eval(Trafo, ipC) > -1e-12
-          || LvlSet->Eval(Trafo, ipD) > -1e-12)
+      if (LvlSet->Eval(Trafo, ipA) > -tol_1
+          || LvlSet->Eval(Trafo, ipB) > -tol_1
+          || LvlSet->Eval(Trafo, ipC) > -tol_1
+          || LvlSet->Eval(Trafo, ipD) > -tol_1)
       {
          element_int = true;
       }
@@ -978,7 +978,7 @@ void MomentFittingIntRules::ComputeSurfaceWeights3D(ElementTransformation& Tr)
       temp2 = 0.;
       for (int i = 0; i < nBasis; i++)
       {
-         if (SVD.Singularvalue(i) > 1e-12)
+         if (SVD.Singularvalue(i) > tol_1)
          {
             temp2(i) = temp(i) / SVD.Singularvalue(i);
          }
@@ -1047,18 +1047,18 @@ void MomentFittingIntRules::ComputeVolumeWeights3D(ElementTransformation& Tr,
       IntegrationPoint ipD;
       Trafo.TransformBack(pointD, ipD);
 
-      if (LvlSet->Eval(Trafo, ipA) < -1e-12
-          || LvlSet->Eval(Trafo, ipB) < -1e-12
-          || LvlSet->Eval(Trafo, ipC) < -1e-12
-          || LvlSet->Eval(Trafo, ipD) < -1e-12)
+      if (LvlSet->Eval(Trafo, ipA) < -tol_1
+          || LvlSet->Eval(Trafo, ipB) < -tol_1
+          || LvlSet->Eval(Trafo, ipC) < -tol_1
+          || LvlSet->Eval(Trafo, ipD) < -tol_1)
       {
          interior = false;
       }
 
-      if (LvlSet->Eval(Trafo, ipA) > -1e-12
-          || LvlSet->Eval(Trafo, ipB) > -1e-12
-          || LvlSet->Eval(Trafo, ipC) > -1e-12
-          || LvlSet->Eval(Trafo, ipD) > -1e-12)
+      if (LvlSet->Eval(Trafo, ipA) > -tol_1
+          || LvlSet->Eval(Trafo, ipB) > -tol_1
+          || LvlSet->Eval(Trafo, ipC) > -tol_1
+          || LvlSet->Eval(Trafo, ipD) > -tol_1)
       {
          element_int = true;
       }
@@ -1159,7 +1159,7 @@ void MomentFittingIntRules::ComputeVolumeWeights3D(ElementTransformation& Tr,
       VolumeSVD->LeftSingularvectors().MultTranspose(RHS, temp);
       temp2 = 0.;
       for (int i = 0; i < nBasisVolume; i++)
-         if (VolumeSVD->Singularvalue(i) > 1e-12)
+         if (VolumeSVD->Singularvalue(i) > tol_1)
          {
             temp2(i) = temp(i) / VolumeSVD->Singularvalue(i);
          }

fem/intrules_cut.hpp

@@ -36,6 +36,17 @@ protected:
    /// Space order for the LS projection.
    int lsOrder;
 
+   /// @name Tolerances used for point comparisons
+   ///@{
+#ifdef MFEM_USE_DOUBLE
+   static constexpr real_t tol_1 = 1e-12;
+   static constexpr real_t tol_2 = 1e-15;
+#elif defined(MFEM_USE_SINGLE)
+   static constexpr real_t tol_1 = 1e-5;
+   static constexpr real_t tol_2 = 1e-7;
+#endif
+   ///@}
+
    /** @brief Constructor to set up the generated cut IntegrationRules.
 
        @param [in] order  Order of the constructed IntegrationRule.

fem/lininteg.cpp

@@ -651,6 +651,39 @@ void VectorFEBoundaryFluxLFIntegrator::AssembleRHSElementVect(
    }
 }
 
+void VectorFEBoundaryNormalLFIntegrator::AssembleRHSElementVect(
+   const FiniteElement &el, ElementTransformation &Tr, Vector &elvect)
+{
+   int dim = el.GetDim()+1;
+   int dof = el.GetDof();
+   Vector nor(dim), Fvec(dim);
+
+   shape.SetSize(dof);
+   elvect.SetSize(dof);
+   elvect = 0.0;
+
+   const IntegrationRule *ir = IntRule;
+   if (ir == NULL)
+   {
+      int intorder = 2 * el.GetOrder() + Tr.OrderW();  // <----------
+      ir = &IntRules.Get(el.GetGeomType(), intorder);
+   }
+
+   for (int i = 0; i < ir->GetNPoints(); i++)
+   {
+      const IntegrationPoint &ip = ir->IntPoint(i);
+
+      Tr.SetIntPoint(&ip);
+      CalcOrtho(Tr.Jacobian(), nor);
+      F.Eval(Fvec, Tr, ip);
+      real_t val = ip.weight * (Fvec*nor) / Tr.Weight();
+
+      el.CalcShape(ip, shape);
+
+      elvect.Add(val, shape);
+   }
+}
+
 void VectorFEBoundaryTangentLFIntegrator::AssembleRHSElementVect(
    const FiniteElement &el, ElementTransformation &Tr, Vector &elvect)
 {

fem/lininteg.hpp

@@ -470,6 +470,25 @@ public:
                                Vector &b);
 };
 
+/** Class for boundary integration of (f.n, v.n) for vector coefficient f and
+    RT vector test function v. This integrator works with RT spaces defined
+    using the RT_FECollection class. */
+class VectorFEBoundaryNormalLFIntegrator : public LinearFormIntegrator
+{
+private:
+   VectorCoefficient &F;
+   Vector shape;
+
+public:
+   VectorFEBoundaryNormalLFIntegrator(VectorCoefficient &f) : F(f) { }
+
+   virtual void AssembleRHSElementVect(const FiniteElement &el,
+                                       ElementTransformation &Tr,
+                                       Vector &elvect) override;
+
+   using LinearFormIntegrator::AssembleRHSElementVect;
+};
+
 /// Class for boundary integration $ L(v) = (n \times f, v) $
 class VectorFEBoundaryTangentLFIntegrator : public LinearFormIntegrator
 {

fem/pgridfunc.cpp

@@ -39,9 +39,10 @@ ParGridFunction::ParGridFunction(ParMesh *pmesh, const GridFunction *gf,
 {
    const FiniteElementSpace *glob_fes = gf->FESpace();
    // duplicate the FiniteElementCollection from 'gf'
-   fec = FiniteElementCollection::New(glob_fes->FEColl()->Name());
+   fec_owned = FiniteElementCollection::New(glob_fes->FEColl()->Name());
    // create a local ParFiniteElementSpace from the global one:
-   fes = pfes = new ParFiniteElementSpace(pmesh, glob_fes, partitioning, fec);
+   fes = pfes = new ParFiniteElementSpace(pmesh, glob_fes, partitioning,
+                                          fec_owned);
    SetSize(pfes->GetVSize());
 
    if (partitioning)
@@ -81,7 +82,7 @@ ParGridFunction::ParGridFunction(ParMesh *pmesh, std::istream &input)
    : GridFunction(pmesh, input)
 {
    // Convert the FiniteElementSpace, fes, to a ParFiniteElementSpace:
-   pfes = new ParFiniteElementSpace(pmesh, fec, fes->GetVDim(),
+   pfes = new ParFiniteElementSpace(pmesh, fec_owned, fes->GetVDim(),
                                     fes->GetOrdering());
    delete fes;
    fes = pfes;

fem/qfunction.cpp

@@ -207,8 +207,9 @@ void QuadratureFunction::SaveVTU(std::ostream &os, VTKFormat format,
 
    os << "<PointData>\n";
    os << "<DataArray type=\"" << type_str << "\" Name=\"" << field_name
-      << "\" format=\"" << fmt_str << "\" NumberOfComponents=\"" << vdim
-      << "\">\n";
+      << "\" format=\"" << fmt_str << "\" NumberOfComponents=\"" << vdim << "\" "
+      << VTKComponentLabels(vdim) << " "
+      << ">\n";
    for (int i = 0; i < ne; i++)
    {
       DenseMatrix vals;

fem/restriction.cpp

@@ -74,6 +74,7 @@ ElementRestriction::ElementRestriction(const FiniteElementSpace &f,
          ++offsets[gid + 1];
       }
    }
+   max_connectivity = offsets.Max();
    // Aggregate to find offsets for each global dof
    for (int i = 1; i <= ndofs; ++i)
    {
@@ -320,7 +321,7 @@ static MFEM_HOST_DEVICE int GetAndIncrementNnzIndex(const int i_L, int* I)
 
 int ElementRestriction::FillI(SparseMatrix &mat) const
 {
-   static constexpr int Max = MaxNbNbr;
+   const int max_connect = max_connectivity;
    const int all_dofs = ndofs;
    const int vd = vdim;
    const int elt_dofs = dof;
@@ -332,21 +333,20 @@ int ElementRestriction::FillI(SparseMatrix &mat) const
    {
       I[i_L] = 0;
    });
-   mfem::forall(ne*elt_dofs, [=] MFEM_HOST_DEVICE (int l_dof)
+   mfem::forall_2D(ne*elt_dofs, 1, 1, [=] MFEM_HOST_DEVICE (int l_dof)
    {
+      int *shared = DynamicSharedMemory::Get<int>();
+
       const int e = l_dof/elt_dofs;
       const int i = l_dof%elt_dofs;
 
-      int i_elts[Max];
+      int *i_elts = shared;
+
       const int i_gm = e*elt_dofs + i;
       const int i_L = d_gather_map[i_gm];
       const int i_offset = d_offsets[i_L];
       const int i_next_offset = d_offsets[i_L+1];
       const int i_nbElts = i_next_offset - i_offset;
-      MFEM_ASSERT_KERNEL(
-         i_nbElts <= Max,
-         "The connectivity of this mesh is beyond the max, increase the "
-         "MaxNbNbr variable to comply with your mesh.");
       for (int e_i = 0; e_i < i_nbElts; ++e_i)
       {
          const int i_E = d_indices[i_offset+e_i];
@@ -359,17 +359,13 @@ int ElementRestriction::FillI(SparseMatrix &mat) const
          const int j_offset = d_offsets[j_L];
          const int j_next_offset = d_offsets[j_L+1];
          const int j_nbElts = j_next_offset - j_offset;
-         MFEM_ASSERT_KERNEL(
-            j_nbElts <= Max,
-            "The connectivity of this mesh is beyond the max, increase the "
-            "MaxNbNbr variable to comply with your mesh.");
          if (i_nbElts == 1 || j_nbElts == 1) // no assembly required
          {
             GetAndIncrementNnzIndex(i_L, I);
          }
          else // assembly required
          {
-            int j_elts[Max];
+            int *j_elts = shared + max_connect;
             for (int e_j = 0; e_j < j_nbElts; ++e_j)
             {
                const int j_E = d_indices[j_offset+e_j];
@@ -383,7 +379,7 @@ int ElementRestriction::FillI(SparseMatrix &mat) const
             }
          }
       }
-   });
+   }, 2*max_connectivity*sizeof(int));
    // We need to sum the entries of I, we do it on CPU as it is very sequential.
    auto h_I = mat.HostReadWriteI();
    const int nTdofs = vd*all_dofs;
@@ -402,10 +398,10 @@ int ElementRestriction::FillI(SparseMatrix &mat) const
 void ElementRestriction::FillJAndData(const Vector &ea_data,
                                       SparseMatrix &mat) const
 {
-   static constexpr int Max = MaxNbNbr;
    const int all_dofs = ndofs;
    const int vd = vdim;
    const int elt_dofs = dof;
+   const int max_connect = max_connectivity;
    auto I = mat.ReadWriteI();
    auto J = mat.WriteJ();
    auto Data = mat.WriteData();
@@ -413,22 +409,21 @@ void ElementRestriction::FillJAndData(const Vector &ea_data,
    auto d_indices = indices.Read();
    auto d_gather_map = gather_map.Read();
    auto mat_ea = Reshape(ea_data.Read(), elt_dofs, elt_dofs, ne);
-   mfem::forall(ne*elt_dofs, [=] MFEM_HOST_DEVICE (int l_dof)
+   mfem::forall_2D(ne*elt_dofs, 1, 1, [=] MFEM_HOST_DEVICE (int l_dof)
    {
+      int *shared = DynamicSharedMemory::Get<int>();
+
       const int e = l_dof/elt_dofs;
       const int i = l_dof%elt_dofs;
 
-      int i_elts[Max];
-      int i_B[Max];
+      int *i_elts = shared;
+      int *i_B = shared + max_connect;
+
       const int i_gm = e*elt_dofs + i;
       const int i_L = d_gather_map[i_gm];
       const int i_offset = d_offsets[i_L];
       const int i_next_offset = d_offsets[i_L+1];
       const int i_nbElts = i_next_offset - i_offset;
-      MFEM_ASSERT_KERNEL(
-         i_nbElts <= Max,
-         "The connectivity of this mesh is beyond the max, increase the "
-         "MaxNbNbr variable to comply with your mesh.");
       for (int e_i = 0; e_i < i_nbElts; ++e_i)
       {
          const int i_E = d_indices[i_offset+e_i];
@@ -450,8 +445,8 @@ void ElementRestriction::FillJAndData(const Vector &ea_data,
          }
          else // assembly required
          {
-            int j_elts[Max];
-            int j_B[Max];
+            int *j_elts = shared + 2*max_connect;
+            int *j_B = shared + 3*max_connect;
             for (int e_j = 0; e_j < j_nbElts; ++e_j)
             {
                const int j_E = d_indices[j_offset+e_j];
@@ -483,7 +478,7 @@ void ElementRestriction::FillJAndData(const Vector &ea_data,
             }
          }
       }
-   });
+   }, 4*max_connectivity*sizeof(int));
    // We need to shift again the entries of I, we do it on CPU as it is very
    // sequential.
    auto h_I = mat.HostReadWriteI();
@@ -751,11 +746,11 @@ void ConformingFaceRestriction::CheckFESpace(const ElementDofOrdering
 #ifdef MFEM_DEBUG
    const FiniteElement *fe0 = fes.GetFE(0);
    const TensorBasisElement *tfe = dynamic_cast<const TensorBasisElement*>(fe0);
-   MFEM_VERIFY(tfe != NULL &&
-               (tfe->GetBasisType()==BasisType::GaussLobatto ||
-                tfe->GetBasisType()==BasisType::Positive),
-               "Only Gauss-Lobatto and Bernstein basis are supported in "
-               "ConformingFaceRestriction.");
+   MFEM_VERIFY(tfe != NULL,
+               "ConformingFaceRestriction only supports TensorBasisElements");
+   MFEM_VERIFY(tfe->GetBasisType()==BasisType::GaussLobatto ||
+               tfe->GetBasisType()==BasisType::Positive,
+               "ConformingFaceRestriction only supports Gauss-Lobatto and Bernstein bases");
 
    // Assuming all finite elements are using Gauss-Lobatto.
    const bool dof_reorder = (f_ordering == ElementDofOrdering::LEXICOGRAPHIC);
@@ -855,7 +850,8 @@ void ConformingFaceRestriction::SetFaceDofsScatterIndices(
                "This method should not be used on nonconforming coarse faces.");
    MFEM_ASSERT(face.element[0].orientation==0,
                "FaceRestriction used on degenerated mesh.");
-   MFEM_CONTRACT_VAR(f_ordering); // not supported yet
+   MFEM_VERIFY(f_ordering == ElementDofOrdering::LEXICOGRAPHIC,
+               "NATIVE ordering is not supported yet");
 
    fes.GetFE(0)->GetFaceMap(face.element[0].local_face_id, face_map);
 
@@ -883,7 +879,8 @@ void ConformingFaceRestriction::SetFaceDofsGatherIndices(
 {
    MFEM_ASSERT(!(face.IsNonconformingCoarse()),
                "This method should not be used on nonconforming coarse faces.");
-   MFEM_CONTRACT_VAR(f_ordering); // not supported yet
+   MFEM_VERIFY(f_ordering == ElementDofOrdering::LEXICOGRAPHIC,
+               "NATIVE ordering is not supported yet");
 
    fes.GetFE(0)->GetFaceMap(face.element[0].local_face_id, face_map);
 

fem/restriction.hpp

@@ -39,11 +39,6 @@ public:
     objects, see FiniteElementSpace::GetElementRestriction(). */
 class ElementRestriction : public ElementRestrictionOperator
 {
-private:
-   /** This number defines the maximum number of elements any dof can belong to
-       for the FillSparseMatrix method. */
-   static const int MaxNbNbr = 16;
-
 protected:
    const FiniteElementSpace &fes;
    const int ne;
@@ -55,6 +50,7 @@ protected:
    Array<int> offsets;
    Array<int> indices;
    Array<int> gather_map;
+   int max_connectivity;
 
 public:
    ElementRestriction(const FiniteElementSpace&, ElementDofOrdering);

fem/tmop.cpp

@@ -3390,11 +3390,12 @@ real_t TMOP_Integrator::GetElementEnergy(const FiniteElement &el,
 
          const IntegrationPoint &ip_s = ir_s->IntPoint(s);
          Tpr->SetIntPoint(&ip_s);
+         double w = surf_fit_coeff->Eval(*Tpr, ip_s) * surf_fit_normal *
+                    1.0 / surf_fit_dof_count[scalar_dof_id];
 
          if (surf_fit_gf)
          {
-            energy += surf_fit_coeff->Eval(*Tpr, ip_s) * surf_fit_normal *
-                      sigma_e(s) * sigma_e(s);
+            energy += w * sigma_e(s) * sigma_e(s);
          }
          if (surf_fit_pos)
          {
@@ -3405,8 +3406,7 @@ real_t TMOP_Integrator::GetElementEnergy(const FiniteElement &el,
                pos(d) = PMatI(s, d);
                pos_target(d) = (*surf_fit_pos)(vdofs[d*dof + s]);
             }
-            energy += surf_fit_coeff->Eval(*Tpr, ip_s) * surf_fit_normal *
-                      surf_fit_limiter->Eval(pos, pos_target, 1.0);
+            energy += w * surf_fit_limiter->Eval(pos, pos_target, 1.0);
          }
       }
    }

fem/tmop.hpp

@@ -217,6 +217,21 @@ public:
    virtual WorstCaseType GetWorstCaseType() { return wctype; }
 };
 
+/// 0 metric
+class TMOP_Metric_000 : public TMOP_QualityMetric
+{
+public:
+   // W = 0.
+   virtual real_t EvalW(const DenseMatrix &Jpt) const {return 0.0;}
+
+   virtual void EvalP(const DenseMatrix &Jpt, DenseMatrix &P) const {P = 0.0;}
+
+   virtual void AssembleH(const DenseMatrix &Jpt, const DenseMatrix &DS,
+                          const real_t weight, DenseMatrix &A) const {A = 0.0;}
+
+   virtual int Id() const { return 0; }
+};
+
 /// 2D non-barrier metric without a type.
 class TMOP_Metric_001 : public TMOP_QualityMetric
 {

fem/transfer.cpp

@@ -257,11 +257,12 @@ void L2ProjectionGridTransfer::L2Projection::BuildHo2Lor(
 
 void L2ProjectionGridTransfer::L2Projection::ElemMixedMass(
    Geometry::Type geom, const FiniteElement& fe_ho,
-   const FiniteElement& fe_lor, ElementTransformation* el_tr,
+   const FiniteElement& fe_lor, ElementTransformation* tr_ho,
+   ElementTransformation* tr_lor,
    IntegrationPointTransformation& ip_tr,
    DenseMatrix& M_mixed_el) const
 {
-   int order = fe_lor.GetOrder() + fe_ho.GetOrder() + el_tr->OrderW();
+   int order = fe_lor.GetOrder() + fe_ho.GetOrder() + tr_lor->OrderW();
    const IntegrationRule* ir = &IntRules.Get(geom, order);
    M_mixed_el = 0.0;
    for (int i = 0; i < ir->GetNPoints(); i++)
@@ -272,11 +273,16 @@ void L2ProjectionGridTransfer::L2Projection::ElemMixedMass(
       Vector shape_lor(fe_lor.GetDof());
       fe_lor.CalcShape(ip_lor, shape_lor);
       Vector shape_ho(fe_ho.GetDof());
-      fe_ho.CalcShape(ip_ho, shape_ho);
-      el_tr->SetIntPoint(&ip_lor);
+      tr_ho->SetIntPoint(&ip_ho);
+      fe_ho.CalcPhysShape(*tr_ho, shape_ho);
+      tr_lor->SetIntPoint(&ip_lor);
       // For now we use the geometry information from the LOR space, which means
       // we won't be mass conservative if the mesh is curved
-      real_t w = el_tr->Weight() * ip_lor.weight;
+      real_t w = ip_lor.weight;
+      if (fe_lor.GetMapType() == FiniteElement::VALUE)
+      {
+         w *= tr_lor->Weight();
+      }
       shape_lor *= w;
       AddMultVWt(shape_lor, shape_ho, M_mixed_el);
    }
@@ -344,6 +350,8 @@ L2ProjectionGridTransfer::L2ProjectionL2Space::L2ProjectionL2Space(
       int ndof_ho = fe_ho.GetDof();
       int ndof_lor = fe_lor.GetDof();
 
+      ElementTransformation *tr_ho = fes_ho.GetElementTransformation(iho);
+
       emb_tr.SetIdentityTransformation(geom);
       const DenseTensor &pmats = cf_tr.point_matrices[geom];
 
@@ -369,8 +377,8 @@ L2ProjectionGridTransfer::L2ProjectionL2Space::L2ProjectionL2Space(
       {
          // Assemble the low-order refined mass matrix and invert locally
          int ilor = lor_els[iref];
-         ElementTransformation *el_tr = fes_lor.GetElementTransformation(ilor);
-         mi.AssembleElementMatrix(fe_lor, *el_tr, M_lor_el);
+         ElementTransformation *tr_lor = fes_lor.GetElementTransformation(ilor);
+         mi.AssembleElementMatrix(fe_lor, *tr_lor, M_lor_el);
          M_lor.CopyMN(M_lor_el, iref*ndof_lor, iref*ndof_lor);
          Minv_lor_el.Factor();
          Minv_lor_el.GetInverseMatrix(M_lor_el);
@@ -385,7 +393,7 @@ L2ProjectionGridTransfer::L2ProjectionL2Space::L2ProjectionL2Space(
          // within the coarse high-order element in reference space
          emb_tr.SetPointMat(pmats(cf_tr.embeddings[ilor].matrix));
 
-         ElemMixedMass(geom, fe_ho, fe_lor, el_tr, ip_tr, M_mixed_el);
+         ElemMixedMass(geom, fe_ho, fe_lor, tr_ho, tr_lor, ip_tr, M_mixed_el);
 
          M_mixed.CopyMN(M_mixed_el, iref*ndof_lor, 0);
       }
@@ -880,6 +888,8 @@ std::unique_ptr<SparseMatrix>>
       const FiniteElement& fe_ho = *fes_ho.GetFE(iho);
       const FiniteElement& fe_lor = *fes_lor.GetFE(lor_els[0]);
 
+      ElementTransformation *tr_ho = fes_ho.GetElementTransformation(iho);
+
       emb_tr.SetIdentityTransformation(geom);
       const DenseTensor& pmats = cf_tr.point_matrices[geom];
 
@@ -891,13 +901,13 @@ std::unique_ptr<SparseMatrix>>
       for (int iref = 0; iref < nref; ++iref)
       {
          int ilor = lor_els[iref];
-         ElementTransformation* el_tr = fes_lor.GetElementTransformation(ilor);
+         ElementTransformation* tr_lor = fes_lor.GetElementTransformation(ilor);
 
          // Create the transformation that embeds the fine low-order element
          // within the coarse high-order element in reference space
          emb_tr.SetPointMat(pmats(cf_tr.embeddings[ilor].matrix));
 
-         ElemMixedMass(geom, fe_ho, fe_lor, el_tr, ip_tr, M_LH_el);
+         ElemMixedMass(geom, fe_ho, fe_lor, tr_ho, tr_lor, ip_tr, M_LH_el);
 
          Array<int> dofs_lor(nedof_lor);
          fes_lor.GetElementDofs(ilor, dofs_lor);
@@ -1233,6 +1243,8 @@ void PRefinementTransferOperator::Mult(const Vector& x, Vector& y) const
 
    int vdim = lFESpace.GetVDim();
 
+   y = 0.0;
+
    for (int i = 0; i < mesh->GetNE(); i++)
    {
       DofTransformation * doftrans_h = hFESpace.GetElementDofs(i, h_dofs);

fem/transfer.hpp

@@ -203,7 +203,8 @@ protected:
                        const CoarseFineTransformations& cf_tr);
 
       void ElemMixedMass(Geometry::Type geom, const FiniteElement& fe_ho,
-                         const FiniteElement& fe_lor, ElementTransformation* el_tr,
+                         const FiniteElement& fe_lor, ElementTransformation* tr_ho,
+                         ElementTransformation* tr_lor,
                          IntegrationPointTransformation& ip_tr,
                          DenseMatrix& M_mixed_el) const;
    };

general/backends.hpp

@@ -63,9 +63,9 @@
 #define MFEM_FOREACH_THREAD(i,k,N) for(int i=0; i<N; i++)
 #endif
 
-// 'double' atomicAdd implementation for previous versions of CUDA
+// 'double' and 'float' atomicAdd implementation for previous versions of CUDA
 #if defined(MFEM_USE_CUDA) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 600
-MFEM_DEVICE inline real_t atomicAdd(real_t *add, real_t val)
+MFEM_DEVICE inline mfem::real_t atomicAdd(mfem::real_t *add, mfem::real_t val)
 {
    unsigned long long int *ptr = (unsigned long long int *) add;
    unsigned long long int old = *ptr, reg;

general/forall.hpp

@@ -154,6 +154,47 @@ private:
    }
 };
 
+class DynamicSharedMemory
+{
+   int host_capacity = 0;
+   void *host_data = nullptr;
+   DynamicSharedMemory() = default;
+   ~DynamicSharedMemory()
+   {
+      std::free(host_data); // no-op if host_data is nullptr
+   }
+   static DynamicSharedMemory &Instance()
+   {
+      static DynamicSharedMemory instance;
+      return instance;
+   }
+public:
+   template <typename T>
+   static MFEM_HOST_DEVICE inline T* Get()
+   {
+#if defined(__CUDA_ARCH__)
+      extern __shared__ T shared_mem[];
+      return shared_mem;
+#elif defined(__HIP_DEVICE_COMPILE__)
+      extern __shared__ T shared_mem[];
+      return shared_mem;
+#else
+      return (T*)Instance().host_data;
+#endif
+   }
+
+   static void EnsureHostCapacity(const int nbytes)
+   {
+      DynamicSharedMemory &instance = Instance();
+      if (instance.host_capacity < nbytes)
+      {
+         std::free(instance.host_data);
+         instance.host_data = std::malloc(nbytes);
+         instance.host_capacity = nbytes;
+      }
+   }
+};
+
 // MFEM pragma macros that can be used inside MFEM_FORALL macros.
 #define MFEM_PRAGMA(X) _Pragma(#X)
 
@@ -251,7 +292,8 @@ void RajaCuWrap1D(const int N, DBODY &&d_body)
 
 template <typename DBODY>
 void RajaCuWrap2D(const int N, DBODY &&d_body,
-                  const int X, const int Y, const int BZ)
+                  const int X, const int Y, const int BZ,
+                  const int shared_nbytes)
 {
    MFEM_VERIFY(N>0, "");
    MFEM_VERIFY(BZ>0, "");
@@ -261,7 +303,7 @@ void RajaCuWrap2D(const int N, DBODY &&d_body,
    using RAJA::RangeSegment;
 
    launch<cuda_launch_policy>
-   (LaunchParams(Teams(G), Threads(X, Y, BZ)),
+   (LaunchParams(Teams(G), Threads(X, Y, BZ), shared_nbytes),
     [=] RAJA_DEVICE (LaunchContext ctx)
    {
 
@@ -286,7 +328,8 @@ void RajaCuWrap2D(const int N, DBODY &&d_body,
 
 template <typename DBODY>
 void RajaCuWrap3D(const int N, DBODY &&d_body,
-                  const int X, const int Y, const int Z, const int G)
+                  const int X, const int Y, const int Z, const int G,
+                  const int shared_nbytes)
 {
    MFEM_VERIFY(N>0, "");
    const int GRID = G == 0 ? N : G;
@@ -294,7 +337,7 @@ void RajaCuWrap3D(const int N, DBODY &&d_body,
    using RAJA::RangeSegment;
 
    launch<cuda_launch_policy>
-   (LaunchParams(Teams(GRID), Threads(X, Y, Z)),
+   (LaunchParams(Teams(GRID), Threads(X, Y, Z), shared_nbytes),
     [=] RAJA_DEVICE (LaunchContext ctx)
    {
 
@@ -324,9 +367,10 @@ struct RajaCuWrap<2>
 {
    template <const int BLCK = MFEM_CUDA_BLOCKS, typename DBODY>
    static void run(const int N, DBODY &&d_body,
-                   const int X, const int Y, const int Z, const int G)
+                   const int X, const int Y, const int Z, const int G,
+                   const int shared_nbytes)
    {
-      RajaCuWrap2D(N, d_body, X, Y, Z);
+      RajaCuWrap2D(N, d_body, X, Y, Z, shared_nbytes);
    }
 };
 
@@ -335,9 +379,10 @@ struct RajaCuWrap<3>
 {
    template <const int BLCK = MFEM_CUDA_BLOCKS, typename DBODY>
    static void run(const int N, DBODY &&d_body,
-                   const int X, const int Y, const int Z, const int G)
+                   const int X, const int Y, const int Z, const int G,
+                   const int shared_nbytes)
    {
-      RajaCuWrap3D(N, d_body, X, Y, Z, G);
+      RajaCuWrap3D(N, d_body, X, Y, Z, G, shared_nbytes);
    }
 };
 
@@ -353,7 +398,8 @@ void RajaHipWrap1D(const int N, DBODY &&d_body)
 
 template <typename DBODY>
 void RajaHipWrap2D(const int N, DBODY &&d_body,
-                   const int X, const int Y, const int BZ)
+                   const int X, const int Y, const int BZ,
+                   const int shared_nbytes)
 {
    MFEM_VERIFY(N>0, "");
    MFEM_VERIFY(BZ>0, "");
@@ -363,7 +409,7 @@ void RajaHipWrap2D(const int N, DBODY &&d_body,
    using RAJA::RangeSegment;
 
    launch<hip_launch_policy>
-   (LaunchParams(Teams(G), Threads(X, Y, BZ)),
+   (LaunchParams(Teams(G), Threads(X, Y, BZ), shared_nbytes),
     [=] RAJA_DEVICE (LaunchContext ctx)
    {
 
@@ -388,7 +434,8 @@ void RajaHipWrap2D(const int N, DBODY &&d_body,
 
 template <typename DBODY>
 void RajaHipWrap3D(const int N, DBODY &&d_body,
-                   const int X, const int Y, const int Z, const int G)
+                   const int X, const int Y, const int Z, const int G,
+                   const int shared_nbytes)
 {
    MFEM_VERIFY(N>0, "");
    const int GRID = G == 0 ? N : G;
@@ -396,7 +443,7 @@ void RajaHipWrap3D(const int N, DBODY &&d_body,
    using RAJA::RangeSegment;
 
    launch<hip_launch_policy>
-   (LaunchParams(Teams(GRID), Threads(X, Y, Z)),
+   (LaunchParams(Teams(GRID), Threads(X, Y, Z), shared_nbytes),
     [=] RAJA_DEVICE (LaunchContext ctx)
    {
 
@@ -426,9 +473,10 @@ struct RajaHipWrap<2>
 {
    template <const int BLCK = MFEM_CUDA_BLOCKS, typename DBODY>
    static void run(const int N, DBODY &&d_body,
-                   const int X, const int Y, const int Z, const int G)
+                   const int X, const int Y, const int Z, const int G,
+                   const int shared_nbytes)
    {
-      RajaHipWrap2D(N, d_body, X, Y, Z);
+      RajaHipWrap2D(N, d_body, X, Y, Z, shared_nbytes);
    }
 };
 
@@ -437,9 +485,10 @@ struct RajaHipWrap<3>
 {
    template <const int BLCK = MFEM_CUDA_BLOCKS, typename DBODY>
    static void run(const int N, DBODY &&d_body,
-                   const int X, const int Y, const int Z, const int G)
+                   const int X, const int Y, const int Z, const int G,
+                   const int shared_nbytes)
    {
-      RajaHipWrap3D(N, d_body, X, Y, Z, G);
+      RajaHipWrap3D(N, d_body, X, Y, Z, G, shared_nbytes);
    }
 };
 
@@ -516,24 +565,25 @@ void CuWrap1D(const int N, DBODY &&d_body)
 
 template <typename DBODY>
 void CuWrap2D(const int N, DBODY &&d_body,
-              const int X, const int Y, const int BZ)
+              const int X, const int Y, const int BZ, const int shared_nbytes)
 {
    if (N==0) { return; }
    MFEM_VERIFY(BZ>0, "");
    const int GRID = (N+BZ-1)/BZ;
    const dim3 BLCK(X,Y,BZ);
-   CuKernel2D<<<GRID,BLCK>>>(N,d_body);
+   CuKernel2D<<<GRID,BLCK,shared_nbytes>>>(N,d_body);
    MFEM_GPU_CHECK(cudaGetLastError());
 }
 
 template <typename DBODY>
 void CuWrap3D(const int N, DBODY &&d_body,
-              const int X, const int Y, const int Z, const int G)
+              const int X, const int Y, const int Z, const int G,
+              const int shared_nbytes)
 {
    if (N==0) { return; }
    const int GRID = G == 0 ? N : G;
    const dim3 BLCK(X,Y,Z);
-   CuKernel3D<<<GRID,BLCK>>>(N,d_body);
+   CuKernel3D<<<GRID,BLCK,shared_nbytes>>>(N,d_body);
    MFEM_GPU_CHECK(cudaGetLastError());
 }
 
@@ -545,9 +595,10 @@ struct CuWrap<1>
 {
    template <const int BLCK = MFEM_CUDA_BLOCKS, typename DBODY>
    static void run(const int N, DBODY &&d_body,
-                   const int X, const int Y, const int Z, const int G)
+                   const int X, const int Y, const int Z, const int G,
+                   const int shared_nbytes)
    {
-      CuWrap1D<BLCK>(N, d_body);
+      CuWrap1D<BLCK>(N, d_body, shared_nbytes);
    }
 };
 
@@ -556,9 +607,10 @@ struct CuWrap<2>
 {
    template <const int BLCK = MFEM_CUDA_BLOCKS, typename DBODY>
    static void run(const int N, DBODY &&d_body,
-                   const int X, const int Y, const int Z, const int G)
+                   const int X, const int Y, const int Z, const int G,
+                   const int shared_nbytes)
    {
-      CuWrap2D(N, d_body, X, Y, Z);
+      CuWrap2D(N, d_body, X, Y, Z, shared_nbytes);
    }
 };
 
@@ -567,9 +619,10 @@ struct CuWrap<3>
 {
    template <const int BLCK = MFEM_CUDA_BLOCKS, typename DBODY>
    static void run(const int N, DBODY &&d_body,
-                   const int X, const int Y, const int Z, const int G)
+                   const int X, const int Y, const int Z, const int G,
+                   const int shared_nbytes)
    {
-      CuWrap3D(N, d_body, X, Y, Z, G);
+      CuWrap3D(N, d_body, X, Y, Z, G, shared_nbytes);
    }
 };
 
@@ -612,23 +665,24 @@ void HipWrap1D(const int N, DBODY &&d_body)
 
 template <typename DBODY>
 void HipWrap2D(const int N, DBODY &&d_body,
-               const int X, const int Y, const int BZ)
+               const int X, const int Y, const int BZ, const int shared)
 {
    if (N==0) { return; }
    const int GRID = (N+BZ-1)/BZ;
    const dim3 BLCK(X,Y,BZ);
-   hipLaunchKernelGGL(HipKernel2D,GRID,BLCK,0,0,N,d_body);
+   hipLaunchKernelGGL(HipKernel2D,GRID,BLCK,shared,0,N,d_body);
    MFEM_GPU_CHECK(hipGetLastError());
 }
 
 template <typename DBODY>
 void HipWrap3D(const int N, DBODY &&d_body,
-               const int X, const int Y, const int Z, const int G)
+               const int X, const int Y, const int Z, const int G,
+               const int shared)
 {
    if (N==0) { return; }
    const int GRID = G == 0 ? N : G;
    const dim3 BLCK(X,Y,Z);
-   hipLaunchKernelGGL(HipKernel3D,GRID,BLCK,0,0,N,d_body);
+   hipLaunchKernelGGL(HipKernel3D,GRID,BLCK,shared,0,N,d_body);
    MFEM_GPU_CHECK(hipGetLastError());
 }
 
@@ -651,9 +705,10 @@ struct HipWrap<2>
 {
    template <const int BLCK = MFEM_CUDA_BLOCKS, typename DBODY>
    static void run(const int N, DBODY &&d_body,
-                   const int X, const int Y, const int Z, const int G)
+                   const int X, const int Y, const int Z, const int G,
+                   int shared_nbytes)
    {
-      HipWrap2D(N, d_body, X, Y, Z);
+      HipWrap2D(N, d_body, X, Y, Z, shared_nbytes);
    }
 };
 
@@ -662,9 +717,10 @@ struct HipWrap<3>
 {
    template <const int BLCK = MFEM_CUDA_BLOCKS, typename DBODY>
    static void run(const int N, DBODY &&d_body,
-                   const int X, const int Y, const int Z, const int G)
+                   const int X, const int Y, const int Z, const int G,
+                   int shared_nbytes)
    {
-      HipWrap3D(N, d_body, X, Y, Z, G);
+      HipWrap3D(N, d_body, X, Y, Z, G, shared_nbytes);
    }
 };
 
@@ -676,7 +732,7 @@ template <const int DIM, typename d_lambda, typename h_lambda>
 inline void ForallWrap(const bool use_dev, const int N,
                        d_lambda &&d_body, h_lambda &&h_body,
                        const int X=0, const int Y=0, const int Z=0,
-                       const int G=0)
+                       const int G=0, const int shared_nbytes=0)
 {
    MFEM_CONTRACT_VAR(X);
    MFEM_CONTRACT_VAR(Y);
@@ -689,7 +745,7 @@ inline void ForallWrap(const bool use_dev, const int N,
    // If Backend::RAJA_CUDA is allowed, use it
    if (Device::Allows(Backend::RAJA_CUDA))
    {
-      return RajaCuWrap<DIM>::run(N, d_body, X, Y, Z, G);
+      return RajaCuWrap<DIM>::run(N, d_body, X, Y, Z, G, shared_nbytes);
    }
 #endif
 
@@ -697,7 +753,7 @@ inline void ForallWrap(const bool use_dev, const int N,
    // If Backend::RAJA_HIP is allowed, use it
    if (Device::Allows(Backend::RAJA_HIP))
    {
-      return RajaHipWrap<DIM>::run(N, d_body, X, Y, Z, G);
+      return RajaHipWrap<DIM>::run(N, d_body, X, Y, Z, G, shared_nbytes);
    }
 #endif
 
@@ -705,7 +761,7 @@ inline void ForallWrap(const bool use_dev, const int N,
    // If Backend::CUDA is allowed, use it
    if (Device::Allows(Backend::CUDA))
    {
-      return CuWrap<DIM>::run(N, d_body, X, Y, Z, G);
+      return CuWrap<DIM>::run(N, d_body, X, Y, Z, G, shared_nbytes);
    }
 #endif
 
@@ -713,10 +769,14 @@ inline void ForallWrap(const bool use_dev, const int N,
    // If Backend::HIP is allowed, use it
    if (Device::Allows(Backend::HIP))
    {
-      return HipWrap<DIM>::run(N, d_body, X, Y, Z, G);
+      return HipWrap<DIM>::run(N, d_body, X, Y, Z, G, shared_nbytes);
    }
 #endif
 
+   // The remaining backends run on host. Ensure there is enough host shared
+   // memory available.
+   DynamicSharedMemory::EnsureHostCapacity(shared_nbytes);
+
    // If Backend::DEBUG_DEVICE is allowed, use it
    if (Device::Allows(Backend::DEBUG_DEVICE)) { goto backend_cpu; }
 
@@ -745,42 +805,49 @@ backend_cpu:
 template <const int DIM, typename lambda>
 inline void ForallWrap(const bool use_dev, const int N, lambda &&body,
                        const int X=0, const int Y=0, const int Z=0,
-                       const int G=0)
+                       const int G=0, const int shared_nbytes=0)
 {
-   ForallWrap<DIM>(use_dev, N, body, body, X, Y, Z, G);
+   ForallWrap<DIM>(use_dev, N, body, body, X, Y, Z, G, shared_nbytes);
 }
 
 template<typename lambda>
-inline void forall(int N, lambda &&body) { ForallWrap<1>(true, N, body); }
-
-template<typename lambda>
-inline void forall_switch(bool use_dev, int N, lambda &&body)
+inline void forall(int N, lambda &&body)
 {
-   ForallWrap<1>(use_dev, N, body);
+   ForallWrap<1>(true, N, body);
 }
 
 template<typename lambda>
-inline void forall_2D(int N, int X, int Y, lambda &&body)
+inline void forall_switch(bool use_dev, int N, lambda &&body,
+                          int shared_nbytes=0)
 {
-   ForallWrap<2>(true, N, body, X, Y, 1);
+   ForallWrap<1>(use_dev, N, body, 0, 0, 0, 0, shared_nbytes);
 }
 
 template<typename lambda>
-inline void forall_2D_batch(int N, int X, int Y, int BZ, lambda &&body)
+inline void forall_2D(int N, int X, int Y, lambda &&body, int shared_nbytes=0)
 {
-   ForallWrap<2>(true, N, body, X, Y, BZ);
+   ForallWrap<2>(true, N, body, X, Y, 1, 0, shared_nbytes);
 }
 
 template<typename lambda>
-inline void forall_3D(int N, int X, int Y, int Z, lambda &&body)
+inline void forall_2D_batch(int N, int X, int Y, int BZ, lambda &&body,
+                            int shared_nbytes=0)
 {
-   ForallWrap<3>(true, N, body, X, Y, Z, 0);
+   ForallWrap<2>(true, N, body, X, Y, BZ, 0, shared_nbytes);
 }
 
 template<typename lambda>
-inline void forall_3D_grid(int N, int X, int Y, int Z, int G, lambda &&body)
+inline void forall_3D(int N, int X, int Y, int Z, lambda &&body,
+                      int shared_nbytes=0)
 {
-   ForallWrap<3>(true, N, body, X, Y, Z, G);
+   ForallWrap<3>(true, N, body, X, Y, Z, 0, shared_nbytes);
+}
+
+template<typename lambda>
+inline void forall_3D_grid(int N, int X, int Y, int Z, int G, lambda &&body,
+                           int shared_nbytes=0)
+{
+   ForallWrap<3>(true, N, body, X, Y, Z, G, shared_nbytes);
 }
 
 #ifdef MFEM_USE_MPI
@@ -801,12 +868,12 @@ inline void hypre_forall_cpu(int N, lambda &&body)
 // executes on the GPU device that hypre was configured with at build time.
 #if defined(HYPRE_USING_GPU)
 template<typename lambda>
-inline void hypre_forall_gpu(int N, lambda &&body)
+inline void hypre_forall_gpu(int N, lambda &&body, int shared=0)
 {
 #if defined(HYPRE_USING_CUDA)
-   CuWrap1D(N, body);
+   CuWrap1D(N, body, shared);
 #elif defined(HYPRE_USING_HIP)
-   HipWrap1D(N, body);
+   HipWrap1D(N, body, shared);
 #else
 #error Unknown HYPRE GPU backend!
 #endif

general/isockstream.cpp

@@ -51,7 +51,7 @@ int isockstream::establish()
 {
    // char myname[129];
    char   myname[] = "localhost";
-   int    sfd;
+   int    sfd = -1;
    struct addrinfo hints, *res, *rp;
 
    memset(&hints, 0, sizeof(hints));

general/socketstream.cpp

@@ -58,7 +58,7 @@ public:
       int err_flag = WSAStartup(MAKEWORD(2,2), &wsaData);
       if (err_flag != 0)
       {
-         mfem::out << "Error occurred during initialization of WinSock."
+         mfem::err << "Error occurred during initialization of WinSock."
                    << std::endl;
          return;
       }
@@ -108,11 +108,19 @@ int socketbuf::open(const char hostname[], int port)
    hints.ai_socktype = SOCK_STREAM;
    hints.ai_flags = 0;
    hints.ai_protocol = 0;
+   // On Windows, the following need to be set to 0; also required by POSIX.
+   hints.ai_addrlen = 0;
+   hints.ai_canonname = NULL;
+   hints.ai_addr = NULL;
+   hints.ai_next = NULL;
 
    std::string portStr = std::to_string(port);
    int s = getaddrinfo(hostname, portStr.c_str(), &hints, &res);
    if (s != 0)
    {
+#ifdef MFEM_DEBUG
+      mfem::err << "Error in getaddrinfo(): code = " << s << std::endl;
+#endif
       socket_descriptor = -3;
       return -1;
    }
@@ -177,7 +185,7 @@ int socketbuf::sync()
       if (bw < 0)
       {
 #ifdef MFEM_DEBUG
-         mfem::out << "Error in send(): " << strerror(errno) << std::endl;
+         mfem::err << "Error in send(): " << strerror(errno) << std::endl;
 #endif
          setp(pptr() - n, obuf + buflen);
          pbump(n);
@@ -200,7 +208,7 @@ socketbuf::int_type socketbuf::underflow()
 #ifdef MFEM_DEBUG
       if (br < 0)
       {
-         mfem::out << "Error in recv(): " << strerror(errno) << std::endl;
+         mfem::err << "Error in recv(): " << strerror(errno) << std::endl;
       }
 #endif
       setg(NULL, NULL, NULL);
@@ -249,7 +257,7 @@ std::streamsize socketbuf::xsgetn(char_type *s__, std::streamsize n__)
 #ifdef MFEM_DEBUG
          if (br < 0)
          {
-            mfem::out << "Error in recv(): " << strerror(errno) << std::endl;
+            mfem::err << "Error in recv(): " << strerror(errno) << std::endl;
          }
 #endif
          return (n__ - remain);
@@ -286,7 +294,7 @@ std::streamsize socketbuf::xsputn(const char_type *s__, std::streamsize n__)
       if (bw < 0)
       {
 #ifdef MFEM_DEBUG
-         mfem::out << "Error in send(): " << strerror(errno) << std::endl;
+         mfem::err << "Error in send(): " << strerror(errno) << std::endl;
 #endif
          return (n__ - remain);
       }
@@ -433,7 +441,7 @@ static int mfem_gnutls_verify_callback(gnutls_session_t session)
    int ret = gnutls_certificate_verify_peers3(session, hostname, &status);
    if (ret < 0)
    {
-      mfem::out << "Error in gnutls_certificate_verify_peers3:"
+      mfem::err << "Error in gnutls_certificate_verify_peers3:"
                 << gnutls_strerror(ret) << std::endl;
       return GNUTLS_E_CERTIFICATE_ERROR;
    }
@@ -445,7 +453,7 @@ static int mfem_gnutls_verify_callback(gnutls_session_t session)
             status, type, &status_str, 0);
    if (ret < 0)
    {
-      mfem::out << "Error in gnutls_certificate_verification_status_print:"
+      mfem::err << "Error in gnutls_certificate_verification_status_print:"
                 << gnutls_strerror(ret) << std::endl;
       return GNUTLS_E_CERTIFICATE_ERROR;
    }
@@ -456,7 +464,7 @@ static int mfem_gnutls_verify_callback(gnutls_session_t session)
    int ret = gnutls_certificate_verify_peers2(session, &status);
    if (ret < 0)
    {
-      mfem::out << "Error in gnutls_certificate_verify_peers2:"
+      mfem::err << "Error in gnutls_certificate_verify_peers2:"
                 << gnutls_strerror(ret) << std::endl;
       return GNUTLS_E_CERTIFICATE_ERROR;
    }
@@ -643,7 +651,7 @@ void GnuTLS_socketbuf::start_session()
       status.print_on_error("gnutls_priority_set_direct");
       if (!status.good())
       {
-         mfem::out << "Error ptr = \"" << err_ptr << '"' << std::endl;
+         mfem::err << "Error ptr = \"" << err_ptr << '"' << std::endl;
       }
    }
 
@@ -973,10 +981,10 @@ GnuTLS_session_params &socketstream::add_socket()
          GNUTLS_CLIENT);
       if (!params->status.good())
       {
-         mfem::out << "  public key   = " << pubkey << '\n'
+         mfem::err << "  public key   = " << pubkey << '\n'
                    << "  private key  = " << privkey << '\n'
                    << "  trusted keys = " << trustedkeys << std::endl;
-         mfem::out << "Error setting GLVis client parameters.\n"
+         mfem::err << "Error setting GLVis client parameters.\n"
                    "Use the following GLVis script to create your GLVis keys:\n"
                    "   bash glvis-keygen.sh [\"Your Name\"] [\"Your Email\"]"
                    << std::endl;

general/socketstream.hpp

@@ -108,7 +108,7 @@ public:
    void print_on_error(const char *msg) const
    {
       if (good()) { return; }
-      mfem::out << "Error in " << msg << ": " << gnutls_strerror(res)
+      mfem::err << "Error in " << msg << ": " << gnutls_strerror(res)
                 << std::endl;
    }
 };

general/table.cpp

@@ -37,6 +37,78 @@ Table::Table(const Table &table)
    }
 }
 
+Table::Table(const Table &table1,
+             const Table &table2, int offset)
+{
+   MFEM_ASSERT(table1.size == table2.size,
+               "Tables have different sizes can not merge.");
+   size = table1.size;
+
+   const int nnz = table1.I[size] + table2.I[size];
+   I.New(size+1, table1.I.GetMemoryType());
+   J.New(nnz, table1.J.GetMemoryType());
+
+   I[0] = 0;
+   Array<int> row;
+   for (int i = 0; i < size; i++)
+   {
+      I[i+1] = I[i];
+
+      table1.GetRow(i, row);
+      for (int r = 0; r < row.Size(); r++,  I[i+1] ++)
+      {
+         J[ I[i+1] ] = row[r];
+      }
+
+      table2.GetRow(i, row);
+      for (int r = 0; r < row.Size(); r++,  I[i+1] ++)
+      {
+         J[ I[i+1] ] = (row[r] < 0) ? row[r] - offset : row[r] + offset;
+      }
+
+   }
+}
+
+Table::Table(const Table &table1,
+             const Table &table2, int offset2,
+             const Table &table3, int offset3)
+{
+   MFEM_ASSERT(table1.size == table2.size,
+               "Tables have different sizes can not merge.");
+   MFEM_ASSERT(table1.size == table3.size,
+               "Tables have different sizes can not merge.");
+   size = table1.size;
+
+   const int nnz = table1.I[size] + table2.I[size] + table3.I[size];
+   I.New(size+1, table1.I.GetMemoryType());
+   J.New(nnz, table1.J.GetMemoryType());
+
+   I[0] = 0;
+   Array<int> row;
+   for (int i = 0; i < size; i++)
+   {
+      I[i+1] = I[i];
+
+      table1.GetRow(i, row);
+      for (int r = 0; r < row.Size(); r++,  I[i+1] ++)
+      {
+         J[ I[i+1] ] = row[r];
+      }
+
+      table2.GetRow(i, row);
+      for (int r = 0; r < row.Size(); r++,  I[i+1] ++)
+      {
+         J[ I[i+1] ] = (row[r] < 0) ? row[r] - offset2 : row[r] + offset2;
+      }
+
+      table3.GetRow(i, row);
+      for (int r = 0; r < row.Size(); r++,  I[i+1] ++)
+      {
+         J[ I[i+1] ] = (row[r] < 0) ? row[r] - offset3 : row[r] + offset3;
+      }
+   }
+}
+
 Table& Table::operator=(const Table &rhs)
 {
    Clear();

general/table.hpp

@@ -58,6 +58,14 @@ public:
    /// Copy constructor
    Table(const Table &);
 
+   /** Merge constructors
+       This is used to combine two or three tables into one table.*/
+   Table(const Table &table1,
+         const Table &table2, int offset2);
+   Table(const Table &table1,
+         const Table &table2, int offset2,
+         const Table &table3, int offset3);
+
    /// Assignment operator: deep copy
    Table& operator=(const Table &rhs);
 

general/version.cpp

@@ -73,6 +73,9 @@ const char *GetConfigStr()
 #ifdef MFEM_USE_AMGX
       "MFEM_USE_AMGX\n"
 #endif
+#ifdef MFEM_USE_MAGMA
+      "MFEM_USE_MAGMA\n"
+#endif
 #ifdef MFEM_USE_CEED
       "MFEM_USE_CEED\n"
 #endif

linalg/CMakeLists.txt

@@ -11,6 +11,11 @@
 
 list(APPEND SRCS
   auxiliary.cpp
+  batched/batched.cpp
+  batched/gpu_blas.cpp
+  batched/magma.cpp
+  batched/native.cpp
+  batched/solver.cpp
   blockmatrix.cpp
   blockoperator.cpp
   blockvector.cpp
@@ -31,6 +36,11 @@ list(APPEND SRCS
 
 list(APPEND HDRS
   auxiliary.hpp
+  batched/batched.hpp
+  batched/gpu_blas.hpp
+  batched/magma.hpp
+  batched/native.hpp
+  batched/solver.hpp
   blockmatrix.hpp
   blockoperator.hpp
   blockvector.hpp
@@ -44,6 +54,7 @@ list(APPEND HDRS
   handle.hpp
   invariants.hpp
   kernels.hpp
+  lapack.hpp
   linalg.hpp
   matrix.hpp
   ode.hpp

linalg/batched/batched.cpp

@@ -0,0 +1,110 @@
+// Copyright (c) 2010-2024, 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 "batched.hpp"
+#include "native.hpp"
+#include "gpu_blas.hpp"
+#include "magma.hpp"
+
+namespace mfem
+{
+
+BatchedLinAlg::BatchedLinAlg()
+{
+   backends[NATIVE].reset(new NativeBatchedLinAlg);
+
+   if (Device::Allows(mfem::Backend::CUDA_MASK | mfem::Backend::HIP_MASK))
+   {
+#ifdef MFEM_USE_CUDA_OR_HIP
+      backends[GPU_BLAS].reset(new GPUBlasBatchedLinAlg);
+#endif
+
+#ifdef MFEM_USE_MAGMA
+      backends[MAGMA].reset(new MagmaBatchedLinAlg);
+#endif
+
+#if defined(MFEM_USE_MAGMA)
+      active_backend = MAGMA;
+#elif defined(MFEM_USE_CUDA_OR_HIP)
+      active_backend = GPU_BLAS;
+#else
+      active_backend = NATIVE;
+#endif
+   }
+   else
+   {
+      active_backend = NATIVE;
+   }
+}
+
+BatchedLinAlg &BatchedLinAlg::Instance()
+{
+   static BatchedLinAlg instance;
+   return instance;
+}
+
+void BatchedLinAlg::AddMult(const DenseTensor &A, const Vector &x, Vector &y,
+                            real_t alpha, real_t beta)
+{
+   Get(Instance().active_backend).AddMult(A, x, y, alpha, beta);
+}
+
+void BatchedLinAlg::Mult(const DenseTensor &A, const Vector &x, Vector &y)
+{
+   Get(Instance().active_backend).Mult(A, x, y);
+}
+
+void BatchedLinAlg::Invert(DenseTensor &A)
+{
+   Get(Instance().active_backend).Invert(A);
+}
+
+void BatchedLinAlg::LUFactor(DenseTensor &A, Array<int> &P)
+{
+   Get(Instance().active_backend).LUFactor(A, P);
+}
+
+void BatchedLinAlg::LUSolve(const DenseTensor &A, const Array<int> &P,
+                            Vector &x)
+{
+   Get(Instance().active_backend).LUSolve(A, P, x);
+}
+
+bool BatchedLinAlg::IsAvailable(BatchedLinAlg::Backend backend)
+{
+   return Instance().backends[backend] != nullptr;
+}
+
+void BatchedLinAlg::SetActiveBackend(BatchedLinAlg::Backend backend)
+{
+   MFEM_VERIFY(IsAvailable(backend), "Requested backend not supported.");
+   Instance().active_backend = backend;
+}
+
+BatchedLinAlg::Backend BatchedLinAlg::GetActiveBackend()
+{
+   return Instance().active_backend;
+}
+
+const BatchedLinAlgBase &BatchedLinAlg::Get(BatchedLinAlg::Backend backend)
+{
+   auto &backend_ptr = Instance().backends[backend];
+   MFEM_VERIFY(backend_ptr, "Requested backend not supported.")
+   return *backend_ptr;
+}
+
+void BatchedLinAlgBase::Mult(const DenseTensor &A, const Vector &x,
+                             Vector &y) const
+{
+   AddMult(A, x, y, 1.0, 0.0);
+}
+
+}

linalg/batched/batched.hpp

@@ -0,0 +1,130 @@
+// Copyright (c) 2010-2024, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#ifndef MFEM_BATCHED_LINALG
+#define MFEM_BATCHED_LINALG
+
+#include "../../config/config.hpp"
+#include "../densemat.hpp"
+#include <array>
+#include <memory>
+
+namespace mfem
+{
+
+/// @brief Class for performing batched linear algebra operations, potentially
+/// using accelerated algorithms (GPU BLAS or MAGMA). Accessed using static
+/// member functions.
+///
+/// The static member functions will delegate to the active backend (which can
+/// be set using SetActiveBackend(), see BatchedLinAlg::Backend for all
+/// available backends and the order in which they will be chosen initially).
+/// Operations can be performed directly with a specific backend using Get().
+class BatchedLinAlg
+{
+public:
+   /// @brief Available backends for implementations of batched algorithms.
+   ///
+   /// The initially active backend will be the first available backend in this
+   /// order: MAGMA, GPU_BLAS, NATIVE.
+   enum Backend
+   {
+      /// @brief The standard MFEM backend, implemented using mfem::forall
+      /// kernels. Not as performant as the other kernels.
+      NATIVE,
+      /// @brief Either cuBLAS or hipBLAS, depending on whether MFEM is using
+      /// CUDA or HIP. Not available otherwise.
+      GPU_BLAS,
+      /// MAGMA backend, only available if MFEM is compiled with MAGMA support.
+      MAGMA,
+      /// Counter for the number of backends.
+      NUM_BACKENDS
+   };
+private:
+   /// All available backends. Unavailble backends will be nullptr.
+   std::array<std::unique_ptr<class BatchedLinAlgBase>,
+          Backend::NUM_BACKENDS> backends;
+   Backend active_backend;
+   /// Default constructor. Private.
+   BatchedLinAlg();
+   /// Return the singleton instance.
+   static BatchedLinAlg &Instance();
+public:
+   /// @brief Computes $y = \alpha A x + \beta y$.
+   ///
+   /// $A$ is a block diagonal matrix, represented by the DenseTensor @a A with
+   /// shape (m, n, n_mat). $x$ has shape (n, k, n_mat), and $y$ has shape
+   /// (m, k, n_mat).
+   static void AddMult(const DenseTensor &A, const Vector &x, Vector &y,
+                       real_t alpha = 1.0, real_t beta = 1.0);
+   /// Computes $y = A x$ (e.g. by calling @ref AddMult "AddMult(A,x,y,1,0)").
+   static void Mult(const DenseTensor &A, const Vector &x, Vector &y);
+   /// @brief Replaces the block diagonal matrix $A$ with its inverse $A^{-1}$.
+   ///
+   /// $A$ is represented by the DenseTensor @a A with shape (m, m, n_mat).
+   static void Invert(DenseTensor &A);
+   /// @brief Replaces the block diagonal matrix $A$ with its LU factors. The
+   /// pivots are stored in @a P.
+   ///
+   /// $A$ is represented by the DenseTensor @a A with shape (n, n, n_mat). On
+   /// output, $P$ has shape (n, n_mat).
+   static void LUFactor(DenseTensor &A, Array<int> &P);
+   /// @brief Replaces $x$ with $A^{-1} x$, given the LU factors @a A and pivots
+   /// @a P of the block-diagonal matrix $A$.
+   ///
+   /// The LU factors and pivots of $A$ should be obtained by first calling
+   /// LUFactor(). $A$ has shape (n, n, n_mat) and $x$ has shape (n, n_rhs,
+   /// n_mat).
+   ///
+   /// @warning LUSolve() and LUFactor() should be called using the same backend
+   /// because of potential incompatibilities (e.g. 0-based or 1-based
+   /// indexing).
+   static void LUSolve(const DenseTensor &A, const Array<int> &P, Vector &x);
+   /// @brief Returns true if the requested backend is available.
+   ///
+   /// The available backends depend on which third-party libraries MFEM is
+   /// compiled with, and whether the the CUDA/HIP device is enabled.
+   static bool IsAvailable(Backend backend);
+   /// Set the default backend for batched linear algebra operations.
+   static void SetActiveBackend(Backend backend);
+   /// Get the default backend for batched linear algebra operations.
+   static Backend GetActiveBackend();
+   /// @brief Get the BatchedLinAlgBase object associated with a specific
+   /// backend.
+   ///
+   /// This allows the user to perform specific operations with a backend
+   /// different from the active backend.
+   static const BatchedLinAlgBase &Get(Backend backend);
+};
+
+/// Abstract base clase for batched linear algebra operations.
+class BatchedLinAlgBase
+{
+public:
+   /// See BatchedLinAlg::AddMult.
+   virtual void AddMult(const DenseTensor &A, const Vector &x, Vector &y,
+                        real_t alpha = 1.0, real_t beta = 1.0) const = 0;
+   /// See BatchedLinAlg::Mult.
+   virtual void Mult(const DenseTensor &A, const Vector &x, Vector &y) const;
+   /// See BatchedLinAlg::Invert.
+   virtual void Invert(DenseTensor &A) const = 0;
+   /// See BatchedLinAlg::LUFactor.
+   virtual void LUFactor(DenseTensor &A, Array<int> &P) const = 0;
+   /// See BatchedLinAlg::LUSolve.
+   virtual void LUSolve(const DenseTensor &LU, const Array<int> &P,
+                        Vector &x) const = 0;
+   /// Virtual destructor.
+   virtual ~BatchedLinAlgBase() { }
+};
+
+} // namespace mfem
+
+#endif

linalg/batched/gpu_blas.cpp

@@ -0,0 +1,198 @@
+// Copyright (c) 2010-2024, 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 "gpu_blas.hpp"
+#include "../../general/forall.hpp"
+
+#if defined(MFEM_USE_CUDA)
+#define MFEM_cu_or_hip(stub) cu##stub
+#define MFEM_CU_or_HIP(stub) CU##stub
+#elif defined(MFEM_USE_HIP)
+#define MFEM_cu_or_hip(stub) hip##stub
+#define MFEM_CU_or_HIP(stub) HIP##stub
+#endif
+
+#define MFEM_CONCAT(x, y, z) MFEM_CONCAT_(x, y, z)
+#define MFEM_CONCAT_(x, y, z) x ## y ## z
+
+#ifdef MFEM_USE_SINGLE
+#define MFEM_GPUBLAS_PREFIX(stub) MFEM_CONCAT(MFEM_cu_or_hip(blas), S, stub)
+#elif defined(MFEM_USE_DOUBLE)
+#define MFEM_GPUBLAS_PREFIX(stub) MFEM_CONCAT(MFEM_cu_or_hip(blas), D, stub)
+#endif
+
+#define MFEM_BLAS_SUCCESS MFEM_CU_or_HIP(BLAS_STATUS_SUCCESS)
+
+namespace mfem
+{
+
+GPUBlas &GPUBlas::Instance()
+{
+   static GPUBlas instance;
+   return instance;
+}
+
+GPUBlas::HandleType GPUBlas::Handle()
+{
+   return Instance().handle;
+}
+
+#ifndef MFEM_USE_CUDA_OR_HIP
+
+GPUBlas::GPUBlas() { }
+GPUBlas::~GPUBlas() { }
+void GPUBlas::EnableAtomics() { }
+void GPUBlas::DisableAtomics() { }
+
+#else
+
+using blasStatus_t = MFEM_cu_or_hip(blasStatus_t);
+
+GPUBlas::GPUBlas()
+{
+   blasStatus_t status = MFEM_cu_or_hip(blasCreate)(&handle);
+   MFEM_VERIFY(status == MFEM_BLAS_SUCCESS, "Cannot initialize GPU BLAS.");
+}
+
+GPUBlas::~GPUBlas()
+{
+   MFEM_cu_or_hip(blasDestroy)(handle);
+}
+
+void GPUBlas::EnableAtomics()
+{
+   const blasStatus_t status = MFEM_cu_or_hip(blasSetAtomicsMode)(
+                                  Handle(), MFEM_CU_or_HIP(BLAS_ATOMICS_ALLOWED));
+   MFEM_VERIFY(status == MFEM_BLAS_SUCCESS, "GPU BLAS error.");
+}
+
+void GPUBlas::DisableAtomics()
+{
+   const blasStatus_t status = MFEM_cu_or_hip(blasSetAtomicsMode)(
+                                  Handle(), MFEM_CU_or_HIP(BLAS_ATOMICS_NOT_ALLOWED));
+   MFEM_VERIFY(status == MFEM_BLAS_SUCCESS, "GPU BLAS error.");
+}
+
+void GPUBlasBatchedLinAlg::AddMult(const DenseTensor &A, const Vector &x,
+                                   Vector &y, real_t alpha, real_t beta) const
+{
+   const int m = A.SizeI();
+   const int n = A.SizeJ();
+   const int n_mat = A.SizeK();
+   const int k = x.Size() / n / n_mat;
+
+   auto d_A = mfem::Reshape(A.Read(), m, n, n_mat);
+   auto d_x = mfem::Reshape(x.Read(), n, k, n_mat);
+   auto d_y = mfem::Reshape(beta == 0.0 ? y.Write() : y.ReadWrite(), m, k, n_mat);
+
+   const auto op = MFEM_CU_or_HIP(BLAS_OP_N);
+
+   const blasStatus_t status = MFEM_GPUBLAS_PREFIX(gemmStridedBatched)(
+                                  GPUBlas::Handle(), op, op, m, k, n, &alpha,
+                                  d_A, m, m*n, d_x, n, n*k, &beta, d_y, m, m*k,
+                                  n_mat);
+   MFEM_VERIFY(status == MFEM_BLAS_SUCCESS, "GPU BLAS error.");
+}
+
+void GPUBlasBatchedLinAlg::LUFactor(DenseTensor &A, Array<int> &P) const
+{
+   const int n = A.SizeI();
+   const int n_mat = A.SizeK();
+
+   P.SetSize(n*n_mat);
+
+   Array<int> info_array(n_mat);
+
+   real_t *A_base = A.ReadWrite();
+   Array<real_t*> A_ptrs(n_mat);
+   real_t **d_A_ptrs = A_ptrs.Write();
+   mfem::forall(n_mat, [=] MFEM_HOST_DEVICE (int i)
+   {
+      d_A_ptrs[i] = A_base + i*n*n;
+   });
+
+   const blasStatus_t status = MFEM_GPUBLAS_PREFIX(getrfBatched)(
+                                  GPUBlas::Handle(), n, d_A_ptrs, n, P.Write(),
+                                  info_array.Write(), n_mat);
+   MFEM_VERIFY(status == MFEM_BLAS_SUCCESS, "");
+}
+
+void GPUBlasBatchedLinAlg::LUSolve(
+   const DenseTensor &LU, const Array<int> &P, Vector &x) const
+{
+   const int n = LU.SizeI();
+   const int n_mat = LU.SizeK();
+   const int n_rhs = x.Size() / n / n_mat;
+
+   Array<real_t*> A_ptrs(n_mat);
+   real_t **d_A_ptrs = A_ptrs.Write();
+   Array<real_t*> B_ptrs(n_mat);
+   real_t **d_B_ptrs = B_ptrs.Write();
+
+   {
+      real_t *A_base = const_cast<real_t*>(LU.Read());
+      real_t *B_base = x.ReadWrite();
+      mfem::forall(n_mat, [=] MFEM_HOST_DEVICE (int i)
+      {
+         d_A_ptrs[i] = A_base + i*n*n;
+         d_B_ptrs[i] = B_base + i*n*n_rhs;
+      });
+   }
+
+   int info = 0;
+   const blasStatus_t status = MFEM_GPUBLAS_PREFIX(getrsBatched)(
+                                  GPUBlas::Handle(), MFEM_CU_or_HIP(BLAS_OP_N),
+                                  n, n_rhs, d_A_ptrs, n, P.Read(), d_B_ptrs, n,
+                                  &info, n_mat);
+   MFEM_VERIFY(status == MFEM_BLAS_SUCCESS, "");
+}
+
+void GPUBlasBatchedLinAlg::Invert(DenseTensor &A) const
+{
+   const int n = A.SizeI();
+   const int n_mat = A.SizeK();
+
+   DenseTensor LU(A.SizeI(), A.SizeJ(), A.SizeK());
+   LU.Write();
+   LU.GetMemory().CopyFrom(A.GetMemory(), A.TotalSize());
+
+   Array<real_t*> LU_ptrs(n_mat);
+   Array<real_t*> A_ptrs(n_mat);
+   real_t **d_A_ptrs = A_ptrs.Write();
+   real_t **d_LU_ptrs = LU_ptrs.Write();
+   {
+      real_t *A_base = A.ReadWrite();
+      real_t *LU_base = LU.Write();
+      mfem::forall(n_mat, [=] MFEM_HOST_DEVICE (int i)
+      {
+         d_A_ptrs[i] = A_base + i*n*n;
+         d_LU_ptrs[i] = LU_base + i*n*n;
+      });
+   }
+
+   Array<int> P(n*n_mat);
+   Array<int> info_array(n_mat);
+   blasStatus_t status;
+
+   status = MFEM_GPUBLAS_PREFIX(getrfBatched)(
+               GPUBlas::Handle(), n, d_LU_ptrs, n, P.Write(),
+               info_array.Write(), n_mat);
+   MFEM_VERIFY(status == MFEM_BLAS_SUCCESS, "");
+
+   status = MFEM_GPUBLAS_PREFIX(getriBatched)(
+               GPUBlas::Handle(), n, d_LU_ptrs, n, P.ReadWrite(), d_A_ptrs, n,
+               info_array.Write(), n_mat);
+   MFEM_VERIFY(status == MFEM_BLAS_SUCCESS, "");
+}
+
+#endif
+
+} // namespace mfem

linalg/batched/gpu_blas.hpp

@@ -0,0 +1,71 @@
+// Copyright (c) 2010-2024, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#ifndef MFEM_GPU_BLAS_LINALG
+#define MFEM_GPU_BLAS_LINALG
+
+#include "batched.hpp"
+#include "../../general/backends.hpp"
+#include <cstddef> // std::nullptr_t
+
+#if defined(MFEM_USE_CUDA)
+#include <cublas_v2.h>
+#elif defined(MFEM_USE_HIP)
+#include <hipblas/hipblas.h>
+#endif
+
+namespace mfem
+{
+
+/// @brief Singleton class represented a cuBLAS or hipBLAS handle.
+///
+/// If MFEM is compiled without CUDA or HIP, then this class has no effect.
+class GPUBlas
+{
+#if defined(MFEM_USE_CUDA)
+   using HandleType = cublasHandle_t;
+#elif defined(MFEM_USE_HIP)
+   using HandleType = hipblasHandle_t;
+#else
+   using HandleType = std::nullptr_t;
+#endif
+
+   HandleType handle = nullptr; ///< The internal handle.
+   GPUBlas(); ///< Create the handle.
+   ~GPUBlas(); ///< Destroy the handle.
+   static GPUBlas &Instance(); ///< Get the unique instnce.
+public:
+   /// Return the handle, creating it if needed.
+   static HandleType Handle();
+   /// Enable atomic operations.
+   static void EnableAtomics();
+   /// Disable atomic operations.
+   static void DisableAtomics();
+};
+
+#ifdef MFEM_USE_CUDA_OR_HIP
+
+class GPUBlasBatchedLinAlg : public BatchedLinAlgBase
+{
+public:
+   void AddMult(const DenseTensor &A, const Vector &x, Vector &y,
+                real_t alpha = 1.0, real_t beta = 1.0) const override;
+   void Invert(DenseTensor &A) const override;
+   void LUFactor(DenseTensor &A, Array<int> &P) const override;
+   void LUSolve(const DenseTensor &LU, const Array<int> &P,
+                Vector &x) const override;
+};
+
+#endif // MFEM_USE_CUDA_OR_HIP
+
+} // namespace mfem
+
+#endif // MFEM_GPU_BLAS_LINALG

linalg/batched/magma.cpp

@@ -0,0 +1,177 @@
+// Copyright (c) 2010-2024, 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 "magma.hpp"
+#include "../lapack.hpp"
+#include "../../general/forall.hpp"
+
+#ifdef MFEM_USE_MAGMA
+
+#ifdef MFEM_USE_SINGLE
+#define MFEM_MAGMA_PREFIX(stub) magma_s ## stub
+#define MFEM_MAGMABLAS_PREFIX(stub) magmablas_s ## stub
+#elif defined(MFEM_USE_DOUBLE)
+#define MFEM_MAGMA_PREFIX(stub) magma_d ## stub
+#define MFEM_MAGMABLAS_PREFIX(stub) magmablas_d ## stub
+#endif
+
+namespace mfem
+{
+
+Magma::Magma()
+{
+   const magma_int_t status = magma_init();
+   MFEM_VERIFY(status == MAGMA_SUCCESS, "Error initializing MAGMA.");
+   magma_device_t dev;
+   magma_getdevice(&dev);
+   magma_queue_create(dev, &queue);
+}
+
+Magma::~Magma()
+{
+   magma_queue_destroy(queue);
+   const magma_int_t status = magma_finalize();
+   MFEM_VERIFY(status == MAGMA_SUCCESS, "Error finalizing MAGMA.");
+}
+
+Magma &Magma::Instance()
+{
+   static Magma magma;
+   return magma;
+}
+
+magma_queue_t Magma::Queue()
+{
+   return Instance().queue;
+}
+
+void MagmaBatchedLinAlg::AddMult(const DenseTensor &A, const Vector &x,
+                                 Vector &y, real_t alpha, real_t beta) const
+{
+   const int m = A.SizeI();
+   const int n = A.SizeJ();
+   const int n_mat = A.SizeK();
+   const int k = x.Size() / n / n_mat;
+
+   auto d_A = mfem::Reshape(A.Read(), m, n, n_mat);
+   auto d_x = mfem::Reshape(x.Read(), n, k, n_mat);
+   auto d_y = mfem::Reshape(beta == 0.0 ? y.Write() : y.ReadWrite(), m, k, n_mat);
+
+   MFEM_MAGMABLAS_PREFIX(gemm_batched_strided)(
+      MagmaNoTrans, MagmaNoTrans, m, k, n, alpha, d_A, m, m*n, d_x, n, n*k,
+      beta, d_y, m, m*k, n_mat, Magma::Queue());
+}
+
+void MagmaBatchedLinAlg::LUFactor(DenseTensor &A, Array<int> &P) const
+{
+   const int n = A.SizeI();
+   const int n_mat = A.SizeK();
+
+   P.SetSize(n*n_mat);
+
+   real_t *A_base = A.ReadWrite();
+   int *P_base = P.ReadWrite();
+
+   Array<real_t*> A_ptrs(n_mat);
+   Array<int*> P_ptrs(n_mat);
+   real_t **d_A_ptrs = A_ptrs.Write();
+   int **d_P_ptrs = P_ptrs.Write();
+   mfem::forall(n_mat, [=] MFEM_HOST_DEVICE (int i)
+   {
+      d_A_ptrs[i] = A_base + i*n*n;
+      d_P_ptrs[i] = P_base + i*n;
+   });
+
+   Array<int> info_array(n_mat);
+   const magma_int_t status = MFEM_MAGMA_PREFIX(getrf_batched)(
+                                 n, n, d_A_ptrs, n, d_P_ptrs,
+                                 info_array.Write(), n_mat, Magma::Queue());
+   MFEM_VERIFY(status == MAGMA_SUCCESS, "");
+}
+
+void MagmaBatchedLinAlg::LUSolve(
+   const DenseTensor &LU, const Array<int> &P, Vector &x) const
+{
+   const int n = LU.SizeI();
+   const int n_mat = LU.SizeK();
+   const int n_rhs = x.Size() / n / n_mat;
+
+   Array<real_t*> A_ptrs(n_mat);
+   Array<real_t*> B_ptrs(n_mat);
+   Array<int*> P_ptrs(n_mat);
+   real_t **d_A_ptrs = A_ptrs.Write();
+   real_t **d_B_ptrs = B_ptrs.Write();
+   int **d_P_ptrs = P_ptrs.Write();
+
+   {
+      real_t *A_base = const_cast<real_t*>(LU.Read());
+      real_t *B_base = x.ReadWrite();
+      int *P_base = const_cast<int*>(P.Read());
+      mfem::forall(n_mat, [=] MFEM_HOST_DEVICE (int i)
+      {
+         d_A_ptrs[i] = A_base + i*n*n;
+         d_B_ptrs[i] = B_base + i*n*n_rhs;
+         d_P_ptrs[i] = P_base + i*n;
+      });
+   }
+
+   const magma_int_t status = MFEM_MAGMA_PREFIX(getrs_batched)(
+                                 MagmaNoTrans, n, n_rhs, d_A_ptrs, n, d_P_ptrs,
+                                 d_B_ptrs, n, n_mat, Magma::Queue());
+   MFEM_VERIFY(status == MAGMA_SUCCESS, "");
+}
+
+void MagmaBatchedLinAlg::Invert(DenseTensor &A) const
+{
+   const int n = A.SizeI();
+   const int n_mat = A.SizeK();
+
+   DenseTensor LU(A.SizeI(), A.SizeJ(), A.SizeK());
+   LU.Write();
+   LU.GetMemory().CopyFrom(A.GetMemory(), A.TotalSize());
+
+   Array<int> P(n*n_mat);
+
+   Array<real_t*> LU_ptrs(n_mat);
+   Array<real_t*> A_ptrs(n_mat);
+   Array<int*> P_ptrs(n_mat);
+   real_t **d_A_ptrs = A_ptrs.Write();
+   real_t **d_LU_ptrs = LU_ptrs.Write();
+   int **d_P_ptrs = P_ptrs.Write();
+   {
+      real_t *A_base = A.ReadWrite();
+      real_t *LU_base = LU.Write();
+      int *P_base = P.Write();
+      mfem::forall(n_mat, [=] MFEM_HOST_DEVICE (int i)
+      {
+         d_A_ptrs[i] = A_base + i*n*n;
+         d_LU_ptrs[i] = LU_base + i*n*n;
+         d_P_ptrs[i] = P_base + i*n;
+      });
+   }
+
+   Array<int> info_array(n_mat);
+   magma_int_t status;
+
+   status = MFEM_MAGMA_PREFIX(getrf_batched)(
+               n, n, d_A_ptrs, n, d_P_ptrs, info_array.Write(), n_mat,
+               Magma::Queue());
+   MFEM_VERIFY(status == MAGMA_SUCCESS, "");
+
+   status = MFEM_MAGMA_PREFIX(getri_outofplace_batched)(
+               n, d_LU_ptrs, n, d_P_ptrs, d_A_ptrs, n, info_array.Write(),
+               n_mat, Magma::Queue());
+   MFEM_VERIFY(status == MAGMA_SUCCESS, "");
+}
+
+} // namespace mfem
+
+#endif

linalg/batched/magma.hpp

@@ -0,0 +1,51 @@
+// Copyright (c) 2010-2024, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#ifndef MFEM_MAGMA_LINALG
+#define MFEM_MAGMA_LINALG
+
+#include "batched.hpp"
+
+#ifdef MFEM_USE_MAGMA
+
+#include <magma_v2.h>
+
+namespace mfem
+{
+
+class MagmaBatchedLinAlg : public BatchedLinAlgBase
+{
+public:
+   void AddMult(const DenseTensor &A, const Vector &x, Vector &y,
+                real_t alpha = 1.0, real_t beta = 1.0) const override;
+   void Invert(DenseTensor &A) const override;
+   void LUFactor(DenseTensor &A, Array<int> &P) const override;
+   void LUSolve(const DenseTensor &A, const Array<int> &P,
+                Vector &x) const override;
+};
+
+/// Singleton class for interfacing with the MAGMA library.
+class Magma
+{
+   magma_queue_t queue; ///< The default MAGMA queue.
+   Magma(); ///< Initialize the MAGMA library.
+   ~Magma(); ///< Finalize the MAGMA library.
+   static Magma &Instance(); ///< Get the unique instance of this class.
+public:
+   /// Return the queue, creating it if needed.
+   static magma_queue_t Queue();
+};
+
+} // namespace mfem
+
+#endif
+
+#endif

linalg/batched/native.cpp

@@ -0,0 +1,140 @@
+// Copyright (c) 2010-2024, 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 "../kernels.hpp"
+#include "native.hpp"
+#include "../dtensor.hpp"
+#include "../../general/forall.hpp"
+
+namespace mfem
+{
+
+void NativeBatchedLinAlg::AddMult(const DenseTensor &A, const Vector &x,
+                                  Vector &y, real_t alpha, real_t beta) const
+{
+   const int m = A.SizeI();
+   const int n = A.SizeJ();
+   const int n_mat = A.SizeK();
+   const int k = x.Size() / n / n_mat;
+
+   auto d_A = mfem::Reshape(A.Read(), m, n, n_mat);
+   auto d_x = mfem::Reshape(x.Read(), n, k, n_mat);
+   auto d_y = mfem::Reshape(beta == 0.0 ? y.Write() : y.ReadWrite(), m, k, n_mat);
+
+   mfem::forall(n_mat, [=] MFEM_HOST_DEVICE (int i)
+   {
+      kernels::AddMult(m, k, n, &d_A(0,0,i), &d_x(0,0,i), &d_y(0,0,i),
+                       alpha, beta);
+   });
+
+   // Alternative approach, threading also over the second index. Which one is
+   // better?
+
+   // mfem::forall(n_mat * k, [=] MFEM_HOST_DEVICE (int idx)
+   // {
+   //    const int i = idx % k;
+   //    const int j = idx / k;
+   //    kernels::Mult(m, n, &d_A(0,0,j), &d_x(0,i,j), &d_y(0,i,j));
+   // });
+}
+
+void NativeBatchedLinAlg::Invert(DenseTensor &A) const
+{
+   MFEM_ABORT("");
+}
+
+void NativeBatchedLinAlg::LUFactor(DenseTensor &A, Array<int> &P) const
+{
+   constexpr real_t tol = 0.0; // Make this user-adjustable?
+   const int m = A.SizeI();
+   const int NE = A.SizeK();
+   P.SetSize(m*NE);
+
+   auto data_all = mfem::Reshape(A.ReadWrite(), m, m, NE);
+   auto ipiv_all = mfem::Reshape(P.Write(), m, NE);
+   Array<bool> pivot_flag(1);
+   pivot_flag[0] = true;
+   bool *d_pivot_flag = pivot_flag.ReadWrite();
+
+   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   {
+      for (int i = 0; i < m; i++)
+      {
+         // pivoting
+         {
+            int piv = i;
+            real_t a = fabs(data_all(piv,i,e));
+            for (int j = i+1; j < m; j++)
+            {
+               const real_t b = fabs(data_all(j,i,e));
+               if (b > a)
+               {
+                  a = b;
+                  piv = j;
+               }
+            }
+            ipiv_all(i,e) = piv;
+            if (piv != i)
+            {
+               // swap rows i and piv in both L and U parts
+               for (int j = 0; j < m; j++)
+               {
+                  mfem::kernels::internal::Swap<real_t>(data_all(i,j,e), data_all(piv,j,e));
+               }
+            }
+         } // pivot end
+
+         if (abs(data_all(i,i,e)) <= tol)
+         {
+            d_pivot_flag[0] = false;
+         }
+
+         const real_t a_ii_inv = 1.0 / data_all(i,i,e);
+         for (int j = i+1; j < m; j++)
+         {
+            data_all(j,i,e) *= a_ii_inv;
+         }
+
+         for (int k = i+1; k < m; k++)
+         {
+            const real_t a_ik = data_all(i,k,e);
+            for (int j = i+1; j < m; j++)
+            {
+               data_all(j,k,e) -= a_ik * data_all(j,i,e);
+            }
+         }
+      }
+   });
+
+   MFEM_VERIFY(pivot_flag.HostRead()[0], "Batch LU factorization failed");
+}
+
+void NativeBatchedLinAlg::LUSolve(const DenseTensor &LU, const Array<int> &P,
+                                  Vector &x) const
+{
+   const int m = LU.SizeI();
+   const int n_mat = LU.SizeK();
+   const int n_rhs = x.Size() / m / n_mat;
+
+   auto d_LU = mfem::Reshape(LU.Read(), m, m, n_mat);
+   auto d_P = mfem::Reshape(P.Read(), m, n_mat);
+   auto d_x = mfem::Reshape(x.Write(), m, n_rhs, n_mat);
+
+   mfem::forall(n_mat * n_rhs, [=] MFEM_HOST_DEVICE (int idx)
+   {
+      const int i_rhs = idx % n_rhs;
+      const int i_mat = idx / n_rhs;
+
+      kernels::LUSolve(&d_LU(0,0,i_mat), m, &d_P(0,i_mat), &d_x(0,i_rhs,i_mat));
+   });
+}
+
+}

linalg/batched/native.hpp

@@ -0,0 +1,33 @@
+// Copyright (c) 2010-2024, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#ifndef MFEM_NATIVE_LINALG
+#define MFEM_NATIVE_LINALG
+
+#include "batched.hpp"
+
+namespace mfem
+{
+
+class NativeBatchedLinAlg : public BatchedLinAlgBase
+{
+public:
+   void AddMult(const DenseTensor &A, const Vector &x, Vector &y,
+                real_t alpha, real_t beta) const override;
+   void Invert(DenseTensor &A) const override;
+   void LUFactor(DenseTensor &A, Array<int> &P) const override;
+   void LUSolve(const DenseTensor &LU, const Array<int> &P,
+                Vector &x) const override;
+};
+
+} // namespace mfem
+
+#endif

linalg/batched/solver.cpp

@@ -0,0 +1,50 @@
+// Copyright (c) 2010-2024, 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 "solver.hpp"
+
+namespace mfem
+{
+
+BatchedDirectSolver::BatchedDirectSolver(const DenseTensor &A_, Mode mode_,
+                                         BatchedLinAlg::Backend backend_)
+   : A(A_), mode(mode_), backend(backend_)
+{
+   MFEM_VERIFY(A.SizeI() == A.SizeJ(), "Blocks must be square.");
+   if (mode == LU)
+   {
+      BatchedLinAlg::Get(backend).LUFactor(A, P);
+   }
+   else
+   {
+      BatchedLinAlg::Get(backend).Invert(A);
+   }
+}
+
+void BatchedDirectSolver::Mult(const Vector &x, Vector &y) const
+{
+   if (mode == LU)
+   {
+      y = x;
+      BatchedLinAlg::Get(backend).LUSolve(A, P, y);
+   }
+   else
+   {
+      BatchedLinAlg::Get(backend).Mult(A, x, y);
+   }
+}
+
+void BatchedDirectSolver::SetOperator(const Operator &op)
+{
+   MFEM_ABORT("Not supported.");
+}
+
+} // namespace mfem

linalg/batched/solver.hpp

@@ -0,0 +1,59 @@
+// Copyright (c) 2010-2024, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#ifndef MFEM_BATCHED_SOLVER
+#define MFEM_BATCHED_SOLVER
+
+#include "batched.hpp"
+#include "../operator.hpp"
+
+namespace mfem
+{
+
+/// @brief Solve block-diagonal systems using batched LU or inverses.
+///
+/// LU factorization is more numerically stable, but exposes less fine-grained
+/// parallelism. Inverse matrices have worse conditioning (and increased setup
+/// time), but solving the system is more efficient in parallel (e.g. on GPUs).
+class BatchedDirectSolver : public Solver
+{
+public:
+   /// %Solver mode: whether to use LU factorization or inverses.
+   enum Mode
+   {
+      LU, ///< LU factorization.
+      INVERSE ///< Inverse matrices.
+   };
+protected:
+   DenseTensor A; ///< The LU factors/inverses of the input matrices.
+   Array<int> P; ///< Pivots (needed only for LU factors).
+   Mode mode; ///< Solver mode.
+   BatchedLinAlg::Backend backend; ///< Requested batched linear algebra backend.
+public:
+   /// @brief Constructor.
+   ///
+   /// The DenseTensor @a A_ has dimensions $(m, m, n)$, and represents a block
+   /// diagonal matrix $A$ with $n$ blocks of size $m \times m$.
+   ///
+   /// A deep copy is made of the input @a A_, and so it does not need to be
+   /// retained by the caller.
+   BatchedDirectSolver(const DenseTensor &A_, Mode mode_,
+                       BatchedLinAlg::Backend backend_ =
+                          BatchedLinAlg::GetActiveBackend());
+   /// Sets $y = A^{-1} x$.
+   void Mult(const Vector &x, Vector &y) const;
+   /// Not supported (aborts).
+   void SetOperator(const Operator &op);
+};
+
+} // namespace mfem
+
+#endif

linalg/complex_densemat.cpp

@@ -10,60 +10,9 @@
 // CONTRIBUTING.md for details.
 
 #include "complex_densemat.hpp"
+#include "lapack.hpp"
 #include <complex>
 
-#ifdef MFEM_USE_LAPACK
-#ifdef MFEM_USE_SINGLE
-extern "C" void
-cgetrf_(int *, int *, std::complex<float> *, int *, int *, int *);
-extern "C" void
-cgetrs_(char *, int *, int *, std::complex<float> *, int *, int *,
-        std::complex<float> *, int *, int *);
-extern "C" void
-cgetri_(int *, std::complex<float> *, int *, int *,
-        std::complex<float> *, int *, int *);
-extern "C" void
-ctrsm_(char *, char *, char *, char *, int *, int *, std::complex<float> *,
-       std::complex<float> *, int *, std::complex<float> *, int *);
-extern "C" void
-cpotrf_(char *, int *, std::complex<float> *, int *, int *);
-
-extern "C" void
-ctrtrs_(char *, char*, char *, int *, int *, std::complex<float> *, int *,
-        std::complex<float> *, int *, int *);
-extern "C" void
-cpotri_(char *, int *, std::complex<float> *, int*, int *);
-
-extern "C" void
-cpotrs_(char *, int *, int *, std::complex<float> *, int *,
-        std::complex<float> *, int *, int *);
-#elif defined MFEM_USE_DOUBLE
-extern "C" void
-zgetrf_(int *, int *, std::complex<double> *, int *, int *, int *);
-extern "C" void
-zgetrs_(char *, int *, int *, std::complex<double> *, int *, int *,
-        std::complex<double> *, int *, int *);
-extern "C" void
-zgetri_(int *, std::complex<double> *, int *, int *,
-        std::complex<double> *, int *, int *);
-extern "C" void
-ztrsm_(char *, char *, char *, char *, int *, int *, std::complex<double> *,
-       std::complex<double> *, int *, std::complex<double> *, int *);
-extern "C" void
-zpotrf_(char *, int *, std::complex<double> *, int *, int *);
-
-extern "C" void
-ztrtrs_(char *, char*, char *, int *, int *, std::complex<double> *, int *,
-        std::complex<double> *, int *, int *);
-extern "C" void
-zpotri_(char *, int *, std::complex<double> *, int*, int *);
-
-extern "C" void
-zpotrs_(char *, int *, int *, std::complex<double> *, int *,
-        std::complex<double> *, int *, int *);
-#endif
-#endif
-
 namespace mfem
 {
 
@@ -175,35 +124,17 @@ ComplexDenseMatrix * ComplexDenseMatrix::ComputeInverse()
    std::complex<real_t> qwork, *work;
    int    info;
 
-#ifdef MFEM_USE_SINGLE
-   cgetrf_(&w, &w, data, &w, ipiv, &info);
-#elif defined MFEM_USE_DOUBLE
-   zgetrf_(&w, &w, data, &w, ipiv, &info);
-#else
-   MFEM_ABORT("Floating point type undefined");
-#endif
+   MFEM_LAPACK_COMPLEX(getrf_)(&w, &w, data, &w, ipiv, &info);
    if (info)
    {
       mfem_error("DenseMatrix::Invert() : Error in ZGETRF");
    }
 
-#ifdef MFEM_USE_SINGLE
-   cgetri_(&w, data, &w, ipiv, &qwork, &lwork, &info);
-#elif defined MFEM_USE_DOUBLE
-   zgetri_(&w, data, &w, ipiv, &qwork, &lwork, &info);
-#else
-   MFEM_ABORT("Floating point type undefined");
-#endif
+   MFEM_LAPACK_COMPLEX(getri_)(&w, data, &w, ipiv, &qwork, &lwork, &info);
    lwork = (int) qwork.real();
    work = new std::complex<real_t>[lwork];
 
-#ifdef MFEM_USE_SINGLE
-   cgetri_(&w, data, &w, ipiv, work, &lwork, &info);
-#elif defined MFEM_USE_DOUBLE
-   zgetri_(&w, data, &w, ipiv, work, &lwork, &info);
-#else
-   MFEM_ABORT("Floating point type undefined");
-#endif
+   MFEM_LAPACK_COMPLEX(getri_)(&w, data, &w, ipiv, work, &lwork, &info);
    if (info)
    {
       mfem_error("DenseMatrix::Invert() : Error in ZGETRI");
@@ -493,11 +424,7 @@ bool ComplexLUFactors::Factor(int m, real_t TOL)
 #ifdef MFEM_USE_LAPACK
    int info = 0;
    MFEM_VERIFY(data, "Matrix data not set");
-#ifdef MFEM_USE_SINGLE
-   if (m) { cgetrf_(&m, &m, data, &m, ipiv, &info); }
-#elif defined MFEM_USE_DOUBLE
-   if (m) { zgetrf_(&m, &m, data, &m, ipiv, &info); }
-#endif
+   if (m) { MFEM_LAPACK_COMPLEX(getrf_)(&m, &m, data, &m, ipiv, &info); }
    return info == 0;
 #else
    // compiling without LAPACK
@@ -659,13 +586,10 @@ void ComplexLUFactors::Solve(int m, int n, real_t *X_r, real_t * X_i) const
    std::complex<real_t> * x = ComplexFactors::RealToComplex(m*n,X_r,X_i);
    char trans = 'N';
    int  info = 0;
-#ifdef MFEM_USE_SINGLE
-   if (m > 0 && n > 0) { cgetrs_(&trans, &m, &n, data, &m, ipiv, x, &m, &info); }
-#elif defined MFEM_USE_DOUBLE
-   if (m > 0 && n > 0) { zgetrs_(&trans, &m, &n, data, &m, ipiv, x, &m, &info); }
-#else
-   MFEM_ABORT("Floating point type undefined");
-#endif
+   if (m > 0 && n > 0)
+   {
+      MFEM_LAPACK_COMPLEX(getrs_)(&trans, &m, &n, data, &m, ipiv, x, &m, &info);
+   }
    MFEM_VERIFY(!info, "LAPACK: error in ZGETRS");
    ComplexFactors::ComplexToReal(m*n,x,X_r,X_i);
    delete [] x;
@@ -685,15 +609,8 @@ void ComplexLUFactors::RightSolve(int m, int n, real_t *X_r, real_t * X_i) const
    if (m > 0 && n > 0)
    {
       std::complex<real_t> alpha(1.0,0.0);
-#ifdef MFEM_USE_SINGLE
-      ctrsm_(&side,&u_ch,&n_ch,&n_ch,&n,&m,&alpha,data,&m,X,&n);
-      ctrsm_(&side,&l_ch,&n_ch,&u_ch,&n,&m,&alpha,data,&m,X,&n);
-#elif defined MFEM_USE_DOUBLE
-      ztrsm_(&side,&u_ch,&n_ch,&n_ch,&n,&m,&alpha,data,&m,X,&n);
-      ztrsm_(&side,&l_ch,&n_ch,&u_ch,&n,&m,&alpha,data,&m,X,&n);
-#else
-      MFEM_ABORT("Floating point type undefined");
-#endif
+      MFEM_LAPACK_COMPLEX(trsm_)(&side,&u_ch,&n_ch,&n_ch,&n,&m,&alpha,data,&m,X,&n);
+      MFEM_LAPACK_COMPLEX(trsm_)(&side,&l_ch,&n_ch,&u_ch,&n,&m,&alpha,data,&m,X,&n);
    }
 #else
    // compiling without LAPACK
@@ -815,13 +732,7 @@ bool ComplexCholeskyFactors::Factor(int m, real_t TOL)
    int info = 0;
    char uplo = 'L';
    MFEM_VERIFY(data, "Matrix data not set");
-#ifdef MFEM_USE_SINGLE
-   if (m) {cpotrf_(&uplo, &m, data, &m, &info);}
-#elif defined MFEM_USE_DOUBLE
-   if (m) {zpotrf_(&uplo, &m, data, &m, &info);}
-#else
-   MFEM_ABORT("Floating point type undefined");
-#endif
+   if (m) { MFEM_LAPACK_COMPLEX(potrf_)(&uplo, &m, data, &m, &info); }
    return info == 0;
 #else
    // Cholesky–Crout algorithm
@@ -921,13 +832,8 @@ void ComplexCholeskyFactors::LSolve(int m, int n, real_t * X_r,
    char diag = 'N';
    int info = 0;
 
-#ifdef MFEM_USE_SINGLE
-   ctrtrs_(&uplo, &trans, &diag, &m, &n, data, &m, x, &m, &info);
-#elif defined MFEM_USE_DOUBLE
-   ztrtrs_(&uplo, &trans, &diag, &m, &n, data, &m, x, &m, &info);
-#else
-   MFEM_ABORT("Floating point type undefined");
-#endif
+   MFEM_LAPACK_COMPLEX(trtrs_)(&uplo, &trans, &diag, &m, &n, data, &m, x, &m,
+                               &info);
    MFEM_VERIFY(!info, "ComplexCholeskyFactors:LSolve:: info");
 #else
    for (int k = 0; k < n; k++)
@@ -960,13 +866,8 @@ void ComplexCholeskyFactors::USolve(int m, int n, real_t * X_r,
    char diag = 'N';
    int info = 0;
 
-#ifdef MFEM_USE_SINGLE
-   ctrtrs_(&uplo, &trans, &diag, &m, &n, data, &m, x, &m, &info);
-#elif defined MFEM_USE_DOUBLE
-   ztrtrs_(&uplo, &trans, &diag, &m, &n, data, &m, x, &m, &info);
-#else
-   MFEM_ABORT("Floating point type undefined");
-#endif
+   MFEM_LAPACK_COMPLEX(trtrs_)(&uplo, &trans, &diag, &m, &n, data, &m, x, &m,
+                               &info);
    MFEM_VERIFY(!info, "ComplexCholeskyFactors:USolve:: info");
 #else
    // X <- L^{-t} X
@@ -994,13 +895,7 @@ void ComplexCholeskyFactors::Solve(int m, int n, real_t * X_r,
    char uplo = 'L';
    int info = 0;
    std::complex<real_t> *x = ComplexFactors::RealToComplex(m*n,X_r,X_i);
-#ifdef MFEM_USE_SINGLE
-   cpotrs_(&uplo, &m, &n, data, &m, x, &m, &info);
-#elif defined MFEM_USE_DOUBLE
-   zpotrs_(&uplo, &m, &n, data, &m, x, &m, &info);
-#else
-   MFEM_ABORT("Floating point type undefined");
-#endif
+   MFEM_LAPACK_COMPLEX(potrs_)(&uplo, &m, &n, data, &m, x, &m, &info);
    MFEM_VERIFY(!info, "ComplexCholeskyFactors:Solve:: info");
    ComplexFactors::ComplexToReal(m*n,x,X_r,X_i);
    delete x;
@@ -1026,15 +921,8 @@ void ComplexCholeskyFactors::RightSolve(int m, int n, real_t * X_r,
    std::complex<real_t> alpha(1.0,0.0);
    if (m > 0 && n > 0)
    {
-#ifdef MFEM_USE_SINGLE
-      ctrsm_(&side,&uplo,&transt,&diag,&n,&m,&alpha,data,&m,x,&n);
-      ctrsm_(&side,&uplo,&trans,&diag,&n,&m,&alpha,data,&m,x,&n);
-#elif defined MFEM_USE_DOUBLE
-      ztrsm_(&side,&uplo,&transt,&diag,&n,&m,&alpha,data,&m,x,&n);
-      ztrsm_(&side,&uplo,&trans,&diag,&n,&m,&alpha,data,&m,x,&n);
-#else
-      MFEM_ABORT("Floating point type undefined");
-#endif
+      MFEM_LAPACK_COMPLEX(trsm_)(&side,&uplo,&transt,&diag,&n,&m,&alpha,data,&m,x,&n);
+      MFEM_LAPACK_COMPLEX(trsm_)(&side,&uplo,&trans,&diag,&n,&m,&alpha,data,&m,x,&n);
    }
 #else
    // X <- X L^{-H}
@@ -1085,13 +973,7 @@ void ComplexCholeskyFactors::GetInverseMatrix(int m, real_t * X_r,
    }
    char uplo = 'L';
    int info = 0;
-#ifdef MFEM_USE_SINGLE
-   cpotri_(&uplo, &m, X, &m, &info);
-#elif defined MFEM_USE_DOUBLE
-   zpotri_(&uplo, &m, X, &m, &info);
-#else
-   MFEM_ABORT("Floating point type undefined");
-#endif
+   MFEM_LAPACK_COMPLEX(potri_)(&uplo, &m, X, &m, &info);
    MFEM_VERIFY(!info, "ComplexCholeskyFactors:GetInverseMatrix:: info");
    // fill in the upper triangular part
    for (int i = 0; i<m; i++)