More attempts to make HYPRE sparse matrix work

Add AMG solver option
Fix HYPRE build dependencies
2018-07-13 16:45:23 -07:00 · 2018-07-11 10:51:39 -07:00 · 2018-07-05 15:34:07 -07:00 · 2018-07-05 14:57:07 -07:00 · 2018-07-04 14:52:10 -07:00 · 2018-07-03 14:04:06 -07:00
361 changed files with 19602 additions and 55651 deletions
@@ -43,9 +43,7 @@ before_build:
 build_script:
 - cmake --build build_parallel
 - cmake --build build_serial
- cmake --build build_serial --target exec

 after_build:
 # - cmake --build build_parallel --target check
- cmake --build build_serial --target RUN_TESTS
-
+- cmake --build build_serial --target check
@@ -45,8 +45,6 @@ examples/ex[1-9]
 examples/ex[1-9]p
 examples/ex1[04-9]
 examples/ex1[0-9]p
-examples/ex2[0-9]
-examples/ex2[0-9]p

 examples/refined.mesh
 examples/displaced.mesh
@@ -78,13 +76,6 @@ examples/vortex-?-init.*
 examples/vortex-?-final.*
 examples/deformation.*
 examples/pressure.*
-examples/ex20.dat
-examples/ex20p_?????.dat
-examples/gnuplot_ex20.inp
-examples/gnuplot_ex20p.inp
-examples/ex22*.mesh
-examples/ex22*.sol
-examples/ex22p_*.*

 examples/sundials/ex9
 examples/sundials/ex1[06]
@@ -121,15 +112,6 @@ examples/petsc/deformed.*
 examples/petsc/velocity.*
 examples/petsc/elastic_energy.*

-examples/pumi/ex1
-examples/pumi/ex[126]p
-
-examples/pumi/refined.mesh
-examples/pumi/sol.gf
-examples/pumi/mesh.*
-examples/pumi/sol.*
-examples/pumi/displaced.mesh
-
 miniapps/electromagnetics/volta
 miniapps/electromagnetics/tesla
 miniapps/electromagnetics/maxwell
@@ -142,20 +124,16 @@ miniapps/electromagnetics/Joule_*

 miniapps/meshing/mobius-strip
 miniapps/meshing/klein-bottle
-miniapps/meshing/toroid
 miniapps/meshing/mesh-explorer
 miniapps/meshing/shaper
-miniapps/meshing/extruder
 miniapps/meshing/mesh-optimizer
 miniapps/meshing/pmesh-optimizer

 miniapps/meshing/mobius-strip.mesh
 miniapps/meshing/klein-bottle.mesh
-miniapps/meshing/toroid-*.mesh
 miniapps/meshing/mesh-explorer.mesh
 miniapps/meshing/partitioning.txt
 miniapps/meshing/shaper.mesh
-miniapps/meshing/extruder.mesh
 miniapps/meshing/optimized*
 miniapps/meshing/perturbed*

@@ -169,7 +147,6 @@ miniapps/performance/sol.*
 miniapps/tools/display-basis
 miniapps/tools/load-dc
 miniapps/tools/convert-dc
-miniapps/tools/lor-transfer

 miniapps/nurbs/ex1
 miniapps/nurbs/ex1p
@@ -179,7 +156,3 @@ miniapps/nurbs/mesh.*
 miniapps/nurbs/sol.*
 miniapps/nurbs/mode_*
 miniapps/nurbs/Example1*
-
-# Unit test binary and outputs
-tests/unit/output_meshes
-tests/unit/unit_tests
@@ -205,7 +205,6 @@ install:
        else
            echo "Reusing cached hypre-2.10.0b/";
        fi;
-        ln -s hypre-2.10.0b hypre;
     else
        echo "Serial build, not using hypre";
     fi
@@ -8,192 +8,11 @@
                               http://mfem.org


-Version 4.0-RC2, Apr 24, 2019
-=============================
-
-Requirements and Limitations
----------------------------
- This is a release candidate for mfem-4.0.
- Use at your own risk -- not everything will work and the API may change.
- We are looking for feedback from friendly users.
- Unlike previous MFEM releases, this version requires a C++11 compiler.
-
- GPU-related limitations:
-  * Hypre preconditioners are not yet available in GPU mode.
-  * Only constant coefficients are currently supported on GPUs.
-  * Full-assembly (on device), element assembly, and matrix-free bilinear forms
-    are not supported yet. Element batching is currently ignored.
-  * Partial assembly kernels are not implemented yet for simplices.
-
-GPU support
-----------
- Added initial support for hardware devices, such as GPUs, and programming
-  models, such as CUDA, OCCA, RAJA and OpenMP.
-
- The GPU/device support is based on MFEM's new backends and kernels working
-  seamlessly with a new lightweight device/host memory manager. The kernels can
-  be implemented either in OCCA, or as a simple wrapper around for-loops, which
-  can then be dispatched to RAJA and native backends. See the files forall.hpp
-  and mem_manager.hpp in the general/ directory.
-
- Several of the MFEM example codes (ex1, ex1p, ex6, and ex6p) can now take
-  advantage of GPU acceleration with the backend selectable at runtime. Many of
-  the linear algebra and finite element operations (e.g. partially assembled
-  bilinear forms) have been extended to take advantage of kernel acceleration by
-  simply replacing loops with the MFEM_FORALL() macro.
-
- In addition to pure CUDA, the library currently supports OCCA, RAJA and OpenMP
-  kernels, which could be mixed and matched in different parts of the same
-  application. We plan on adding support for more programming models and devices
-  in the future, without the need for significant modifications in user code.
-  The list of current backends is: "occa-cuda", "raja-cuda", "cuda", "occa-omp",
-  "raja-omp", "omp", "occa-cpu", "raja-cpu", and "cpu".
-
-Discretization improvements
---------------------------
- Added support for a general "low-order refined"-to-"high-order" transfer of
-  GridFunction data from a "low-order refined" (LOR) space defined on a refined
-  mesh to a "high-order" (HO) finite element space defined on a coarse mesh. See
-  the new classes InterpolationGridTransfer and L2ProjectionGridTransfer and the
-  new LOR Transfer miniapp: miniapps/tools/lor-transfer.cpp.
-
- Added support for derefinement of vector (RT + ND) spaces.
-
- Added element flux, and flux energy computation in class ElasticityIntegrator,
-  allowing for the use of Zienkiewicz-Zhu type error estimators with the
-  integrator. For an illustration of this addition, see the new Example 22.
-
- Added a variety of coefficients which are sums or products of existing
-  coefficients as well as grid function coefficients which return the
-  divergence, gradient, or curl of their GridFunctions.
-
-Support for wedge elements and meshes with mixed element types
--------------------------------------------------------------
- Added support for wedge-shaped mesh elements of arbitrary order (with Geometry
-  type PRISM) which have two triangular faces and three quadrilateral faces.
-  Several examples of such meshes can be found in the data/ directory.
-
- Added H1 and L2 finite elements of arbitrary order for Wedge elements.
-
- Added support for mixed meshes containing triangles and quadrilaterals in 2D
-  or tetrahedra, wedges, and hexahedra in 3D. This includes support for uniform
-  refinement of such meshes. Several examples of such meshes can be found in the
-  data/ directory.
-
- Added support for reading and writing linear and quadratic meshes containing
-  wedge elements in VTK mesh format. Several examples of such meshes can be
-  found in the data/ directory.
-
-Other meshing improvements
--------------------------
- Improved the uniform refinement of tetrahedral meshes (also part of the
-  uniform refinement of mixed 3D meshes). The previous refinement algorithm is
-  still available as an option in Mesh::UniformRefinement. Both can be used in
-  the updated Mesh Explorer miniapp.
-
- The local tetrahedral mesh refinement algorithm in serial and in parallel now
-  follows precisely the paper:
-
-     D. Arnold, A. Mukherjee, and L. Pouly, "Locally Adapted Tetrahedral Meshes
-     Using Bisection", SIAM J. Sci. Comput. 22 (2000), 431–448.
-
-  This guarantees that the shape regularity of the elements will be preserved
-  under refinement.
-
- Added support for parallel communication groups on non-conforming meshes.
-
- Improved parallel partitioning of non-conforming meshes. If the coarse mesh
-  elements are ordered as a sequence of face-neighbors, the parallel partitions
-  are now guaranteed to be continuous. To that end, inline quadrilateral and
-  hexahedral meshes are now by default ordered along a space-filling curve.
-
- A boundary in a NURBS mesh can now be connected with another boundary. Such a
-  periodic NURBS mesh is a simple way to impose periodic boundary conditions.
-
- Added support for reading linear and quadratic 2D quadrilateral and triangular
-  Cubit meshes.
-
- The TMOP mesh optimization algorithms were extended to support user-defined
-  space-dependent limiting terms. Improved the TMOP objective functions by more
-  accurate normalization of the different terms.
-
-New and updated examples and miniapps
-------------------------------------
- Added a new meshing miniapp, Toroid, which can produce a variety of torus
-  shaped meshes by twisting a stack of wedges or hexahedra.
-
- Added a new meshing miniapp, Extruder, that demonstrates the capability to
-  produce 3D meshes by extruding 2D meshes.
-
- Added a simple miniapp, LOR Transfer, for visualizing the actions of the
-  transfer operators between a high-order and a low-order refined spaces.
-
- Added a new example, Example 20/20p, that solves a system of 1D ODEs derived
-  from a Hamiltonian. The example demonstrates the use of the variable order,
-  symplectic integration algorithm implemented in class SIAVSolver.
-
- Added a new example, Example 22/22p, that illustrates the use of AMR to solve
-  a linear elasticity problem. This is an extension of Example 2/2p.
-
-New and improved solvers and preconditioners
--------------------------------------------
- Added support for parallel ILU preconditioning via hypre's Euclid solver.
-
- Added support for STRUMPACK v3 with a small API change in the class
-  STRUMPACKSolver, see "API changes" below.
-
-Miscellaneous
-------------
- In SparseMatrix added the option to perform MultTranspose() by matvec with
-  computed and stored transpose matrix. This is required for deterministic
-  results when using devices such as CUDA and OpenMP.
-
- Added unit tests based on the Catch++ library.
-
- Renamed the option MFEM_USE_OPENMP to MFEM_USE_LEGACY_OPENMP. This legacy
-  option is deprecated and planned for removal in a future release. The original
-  option name, MFEM_USE_OPENMP, is now used to enable the new OpenMP backends in
-  the new kernels.
-
- Altered the way FGMRES counts its iterations so that it matches GMRES.
-
- Various other simplifications, extensions, and bugfixes in the code.
-
-API changes
-----------
- In multiple places, use Geometry::Type instead of int, where appropriate.
- In multiple places, use Element::Type instead of int, where appropriate.
- The Mesh methods GetElementBaseGeometry and GetBdrElementBaseGeometry no
-  longer have a default value for their parameter, they only work with an
-  explicitly given index.
- In class Mesh, added methods useful for queries regarding the types of
-  elements present in the mesh: HasGeometry, GetNumGeometries, GetGeometries,
-  and class Mesh::GeometryList.
- The struct CoarseFineTransformations (returned by the Mesh method
-  GetRefinementTransforms) now stores the embedding matrices separately for each
-  Geometry::Type.
- In class ParMesh, replaced the method GroupNFaces with two new methods:
-  GroupNTriangles and GroupNQuadrilaterals. Also, replaced GroupFace with two
-  methods: GroupTriangle and GroupQuadrilateral.
- In class ParMesh, made the two RefineGroups methods protected.
- Removed the virtual method Element::GetRefinementFlag, it is only used by the
-  derived class Tetrahedron.
- Added new methods: Array::CopyTo, Tetrahedron::Init.
- In class STRUMPACKSolver, the method SetMC64Job() was replaced by the new
-  methods: DisableMatching(), EnableMatching(), and EnableParallelMatching().
-
-
-Version 3.4, released on May 29, 2018
-=====================================
-
-More general and efficient mesh adaptivity
------------------------------------------
- Added support for PUMI, the Parallel Unstructured Mesh Infrastructure from
-  https://scorec.rpi.edu/pumi. PUMI is an unstructured, distributed mesh data
-  management system that is capable of handling general non-manifold models and
-  effectively supports automated adaptive analysis. PUMI enables for the first
-  time support for parallel unstructured modifications of MFEM meshes.
+Version 3.3.3 (development)
+===========================

+More efficient non-conforming adaptive mesh refinement
+------------------------------------------------------
 - Significantly reduced MPI communication in the construction of the parallel
  prolongation matrix in ParFiniteElementSpace, for much improved parallel
  scaling of non-conforming AMR on hundreds of thousands of MPI tasks. The
@@ -262,11 +81,6 @@ New and updated examples and miniapps
  NURBS meshes in the miniapps/nurbs directory. Currently the directory contains
  variable order NURBS versions of examples 1, 1p and 11p.

- Added PUMI versions of examples ex1, ex1p, ex2 and ex6p in a new examples/pumi
-  directory. The new examples demonstrate the PUMI APIs for parallel and serial
-  mesh loading (ex1 and ex1p), applying BCs using classification (ex2), and
-  performing parallel mesh adaptation (ex6p).
-
 - Added two new miniapps related to DataCollection I/O in miniapps/tools:
  load-dc.cpp can be used to visualize fields saved via DataCollection classes;
  convert-dc.cpp demonstrates how to convert between MFEM's different concrete
@@ -13,11 +13,6 @@ cmake_minimum_required(VERSION 2.8.11)
 set(USER_CONFIG "${CMAKE_CURRENT_SOURCE_DIR}/config/user.cmake" CACHE PATH
  "Path to optional user configuration file.")

-# Require C++11 and disable compiler-specific extensions
-set(CMAKE_CXX_STANDARD 11)
-set(CMAKE_CXX_STANDARD_REQUIRED ON)
-set(CMAKE_CXX_EXTENSIONS OFF)
-
 # Load user settings before the defaults - this way the defaults will not
 # overwrite the user set options. If the user has not set all options, we still
 # have the defaults.
@@ -50,7 +45,7 @@ project(mfem NONE)
 # Current version of MFEM, see also `makefile`.
 #   mfem_VERSION = (string)
 #   MFEM_VERSION = (int)   [automatically derived from mfem_VERSION]
-set(${PROJECT_NAME}_VERSION 3.4.1)
+set(${PROJECT_NAME}_VERSION 3.3.3)

 # Prohibit in-source build
 if (${PROJECT_SOURCE_DIR} STREQUAL ${PROJECT_BINARY_DIR})
@@ -86,13 +81,6 @@ include("${CMAKE_CURRENT_SOURCE_DIR}/config/XSDKDefaults.cmake")

 # Enable languages.
 enable_language(CXX)
-if (MFEM_USE_CUDA)
-   # MFEM_USE_CUDA requires CMake 3.8 or newer (for direct CUDA support)
-   cmake_minimum_required(VERSION 3.8 FATAL_ERROR)
-   enable_language(CUDA)
-   message(STATUS "Using CUDA architecture: ${CUDA_ARCH}")
-endif()
-
 if (XSDK_ENABLE_C)
   enable_language(C)
 endif()
@@ -151,7 +139,7 @@ if (MFEM_USE_MPI)
    set(PETSC_INCLUDE_DIRS ${PETSC_INCLUDES})
  endif()
 else()
-  set(PKGS_NEED_MPI SUPERLU PETSC STRUMPACK PUMI)
+  set(PKGS_NEED_MPI SUPERLU PETSC STRUMPACK)
  foreach(PKG IN LISTS PKGS_NEED_MPI)
    if (MFEM_USE_${PKG})
      message(STATUS "Disabling package ${PKG} - requires MPI")
@@ -182,11 +170,12 @@ if (MFEM_USE_LAPACK)
 endif()

 # OpenMP
-if (MFEM_USE_OPENMP OR MFEM_USE_LEGACY_OPENMP)
-  if (NOT MFEM_THREAD_SAFE AND MFEM_USE_LEGACY_OPENMP)
-    message(FATAL_ERROR " *** MFEM_USE_LEGACY_OPENMP requires MFEM_THREAD_SAFE=ON.")
+if (MFEM_USE_OPENMP)
+  if (MFEM_THREAD_SAFE)
+    find_package(OpenMP REQUIRED)
+  else()
+    message(FATAL_ERROR " *** MFEM_USE_OPENMP requires MFEM_THREAD_SAFE=ON.")
  endif()
-  find_package(OpenMP REQUIRED)
 endif()

 # SuiteSparse (before SUNDIALS which may depend on KLU)
@@ -257,48 +246,6 @@ if (MFEM_USE_SIDRE)
  find_package(Axom REQUIRED Sidre SLIC axom_utils)
 endif()

-# PUMI
-if (MFEM_USE_PUMI)
-  # If PUMI_DIR was specified, only link to that directory,
-  # i.e. don't link to another installation in /usr/lib by mistake
-  find_package(SCOREC 2.1.0 REQUIRED OPTIONAL_COMPONENTS gmi_sim
-    CONFIG PATHS ${PUMI_DIR} NO_DEFAULT_PATH)
-  if (SCOREC_FOUND)
-    # Define a header file with the MFEM_USE_SIMMETRIX preprocessor variable
-    set(MFEM_USE_SIMMETRIX ${SCOREC_gmi_sim_FOUND})
-    set(PUMI_FOUND ${SCOREC_FOUND})
-    get_target_property(PUMI_INCLUDE_DIRS
-      SCOREC::apf INTERFACE_INCLUDE_DIRECTORIES)
-    set(PUMI_LIBRARIES SCOREC::core)
-  endif()
-endif()
-
-# CUDA
-if (MFEM_USE_CUDA)
-   set(CMAKE_CUDA_STANDARD 11)
-   set(CMAKE_CUDA_STANDARD_REQUIRED ON)
-   set(CMAKE_CUDA_EXTENSIONS OFF)
-   set(CMAKE_CUDA_FLAGS "-arch=${CUDA_ARCH} --expt-extended-lambda"
-       CACHE STRING "CUDA flags set for MFEM" FORCE)
-   if (MFEM_USE_MPI)
-      set(CUDA_CCBIN_COMPILER ${MPI_CXX_COMPILER})
-   else()
-      set(CUDA_CCBIN_COMPILER ${CMAKE_CXX_COMPILER})
-   endif()
-   string(APPEND CMAKE_CUDA_FLAGS " -ccbin ${CUDA_CCBIN_COMPILER}")
-   set(MFEM_USE_MM YES CACHE BOOL "Enable MFEM's memory manager" FORCE)
-endif()
-
-# OCCA
-if (MFEM_USE_OCCA)
-   find_package(OCCA REQUIRED)
-endif()
-
-# RAJA
-if (MFEM_USE_RAJA)
-   find_package(RAJA REQUIRED)
-endif()
-
 # MFEM_TIMER_TYPE
 if (NOT DEFINED MFEM_TIMER_TYPE)
  if (APPLE)
@@ -323,8 +270,8 @@ endif()
 #    integers, the METIS header (with 32-bit indices, as used by mfem) needs to
 #    be before SuiteSparse.
 set(MFEM_TPLS MPI_CXX OPENMP BLAS LAPACK METIS HYPRE SuiteSparse SUNDIALS PETSC
-    MESQUITE SuperLUDist STRUMPACK AXOM CONDUIT GECKO GNUTLS NETCDF MPFR PUMI
-    POSIXCLOCKS MFEMBacktrace ZLIB OCCA RAJA)
+    MESQUITE SuperLUDist STRUMPACK AXOM CONDUIT GECKO GNUTLS NETCDF MPFR POSIXCLOCKS
+    MFEMBacktrace ZLIB)
 # Add all *_FOUND libraries in the variable TPL_LIBRARIES.
 set(TPL_LIBRARIES "")
 set(TPL_INCLUDE_DIRS "")
@@ -349,6 +296,9 @@ message(STATUS "MFEM build type: CMAKE_BUILD_TYPE = ${CMAKE_BUILD_TYPE}")
 message(STATUS "MFEM version: v${MFEM_VERSION_STRING}")
 message(STATUS "MFEM git string: ${MFEM_GIT_STRING}")

+# Windows specific
+set(_USE_MATH_DEFINES ${WIN32})
+
 #-------------------------------------------------------------------------------
 # Define and configure the MFEM library
 #-------------------------------------------------------------------------------
@@ -360,13 +310,6 @@ set(MFEM_SOURCE_DIRS general linalg mesh fem)
 foreach(DIR IN LISTS MFEM_SOURCE_DIRS)
  add_subdirectory(${DIR})
 endforeach()
-
-if (MFEM_USE_CUDA)
-   foreach(file IN LISTS SOURCES)
-      set_property(SOURCE ${file} PROPERTY LANGUAGE CUDA)
-   endforeach()
-endif()
-
 add_subdirectory(config)
 set(MASTER_HEADERS
  ${PROJECT_SOURCE_DIR}/mfem.hpp
@@ -377,11 +320,6 @@ set(CMAKE_INSTALL_RPATH_USE_LINK_PATH ON CACHE BOOL "")
 set(CMAKE_INSTALL_RPATH "${_lib_path}" CACHE PATH "")
 set(CMAKE_INSTALL_NAME_DIR "${_lib_path}" CACHE PATH "")

-set(MFEM_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR} CACHE PATH
-    "The MFEM source directory" FORCE)
-set(MFEM_INSTALL_DIR ${CMAKE_INSTALL_PREFIX} CACHE PATH
-    "The MFEM install directory" FORCE)
-
 # Declaring the library
 add_library(mfem ${SOURCES} ${HEADERS} ${MASTER_HEADERS})
 # message(STATUS "TPL_LIBRARIES = ${TPL_LIBRARIES}")
@@ -434,9 +372,6 @@ endif()
 # Enable testing if required
 if (MFEM_ENABLE_TESTING)
  enable_testing()
-  set(MFEM_ALL_TESTS_TARGET_NAME tests)
-  add_mfem_target(${MFEM_ALL_TESTS_TARGET_NAME} OFF)
-  add_subdirectory(tests EXCLUDE_FROM_ALL)
 endif()

 # Define a target that all examples and miniapps will depend on.
@@ -456,9 +391,7 @@ add_subdirectory(miniapps EXCLUDE_FROM_ALL)
 # Target to build all executables, i.e. everything.
 add_custom_target(exec)
 add_dependencies(exec
-  ${MFEM_ALL_EXAMPLES_TARGET_NAME}
-  ${MFEM_ALL_MINIAPPS_TARGET_NAME}
-  ${MFEM_ALL_TESTS_TARGET_NAME})
+  ${MFEM_ALL_EXAMPLES_TARGET_NAME} ${MFEM_ALL_MINIAPPS_TARGET_NAME})
 # Here, we want to "add_dependencies(test exec)". However, dependencies for
 # 'test' (and other built-in targets) can not be added with add_dependencies():
 #  - https://gitlab.kitware.com/cmake/cmake/issues/8438
@@ -479,12 +412,12 @@ endif()
 # Add 'check' target - quick test
 if (NOT MFEM_USE_MPI)
  add_custom_target(check
-    ${CMAKE_CTEST_COMMAND} -R \"^ex1_ser\" -C ${CMAKE_CFG_INTDIR}
+    ${CMAKE_CTEST_COMMAND} -R '^ex1_ser' -C ${CMAKE_CFG_INTDIR}
    USES_TERMINAL)
  add_dependencies(check ex1)
 else()
  add_custom_target(check
-    ${CMAKE_CTEST_COMMAND} -R \"^ex1p\" -C ${CMAKE_CFG_INTDIR}
+    ${CMAKE_CTEST_COMMAND} -R '^ex1p' -C ${CMAKE_CFG_INTDIR}
    USES_TERMINAL)
  add_dependencies(check ex1p)
 endif()
@@ -529,13 +462,6 @@ install(DIRECTORY ${MFEM_SOURCE_DIRS}
  DESTINATION ${INSTALL_INCLUDE_DIR}/mfem
  FILES_MATCHING PATTERN "*.hpp")

-# Install the okl files
-if (MFEM_USE_OCCA)
-  install(DIRECTORY ${MFEM_SOURCE_DIRS}
-    DESTINATION ${INSTALL_INCLUDE_DIR}/mfem
-    FILES_MATCHING PATTERN "*.okl")
-endif()
-
 # Install ${HEADERS}
 # ---
 # foreach (HDR ${HEADERS})
@@ -601,10 +527,3 @@ install(FILES
 # Install the export set for use with the install-tree
 install(EXPORT ${PROJECT_NAME_UC}Targets
    DESTINATION ${INSTALL_CMAKE_DIR})
-
-#-------------------------------------------------------------------------------
-# Create 'config.mk' from 'config.mk.in' for the build and install locations and
-# define install rules for 'config.mk' and 'test.mk'
-#-------------------------------------------------------------------------------
-
-mfem_export_mk_files()
@@ -1,15 +1,3 @@
-<p align="center">
-<a href="http://mfem.org/"><img alt="mfem" src="http://mfem.org/img/logo-300.png"></a>
-</p>
-
-<p align="center">
-<a href="https://github.com/mfem/mfem/blob/master/COPYRIGHT"><img alt="License" src="https://img.shields.io/badge/License-LGPL--2.1-brightgreen.svg"></a>
-<a href="https://travis-ci.org/mfem/mfem"><img alt="Build Status" src="https://travis-ci.org/mfem/mfem.svg?branch=master"></a>
-<a href="https://ci.appveyor.com/project/mfem/mfem"><img alt="Build Status" src="https://ci.appveyor.com/api/projects/status/19non9sqm6msi2wy?svg=true"></a>
-<a href="http://mfem.github.io/doxygen/html/index.html"><img alt="Doxygen" src="https://img.shields.io/badge/code-documented-brightgreen.svg"></a>
-</p>
-
-
 # How to Contribute

 The MFEM team welcomes contributions at all levels: bugfixes; code
@@ -28,7 +16,6 @@ See the [Quick Summary](#quick-summary) section for the main highlights of our
 GitHub workflow. For more details, consult the following sections and refer
 back to them before issuing pull requests:

- [Code Overview](#code-overview)
 - [GitHub Workflow](#github-workflow)
  - [MFEM Organization](#mfem-organization)
  - [New Feature Development](#new-feature-development)
@@ -66,112 +53,6 @@ Origin](#developers-certificate-of-origin-11) at the end of this file.*
 - Don't hesitate to [contact us](#contact-information) if you have any questions.


-### Code Overview
-
- The MFEM library uses object-orient design principles which reflect, in code,
-  the independent mathematical concepts of meshing, linear algebra and finite
-  element spaces and operators.
-
- The MFEM source code has the following structure:
-  ```
-  .
-  ├── config
-  │   └── cmake
-  │       └── modules
-  ├── data
-  ├── doc
-  │   └── web
-  │       └── examples
-  ├── examples
-  │   ├── petsc
-  │   ├── pumi
-  │   └── sundials
-  ├── fem
-  ├── general
-  ├── linalg
-  ├── mesh
-  ├── miniapps
-  │   ├── common
-  │   ├── electromagnetics
-  │   ├── meshing
-  │   ├── nurbs
-  │   ├── performance
-  │   └── tools
-  └── tests
-      ├── unit
-      │   ├── ...
-      └── ...
-
-  ```
-
- The main directories are `fem/`, `mesh/` and `linalg/` containing the C++
-  classes implementing the finite element, mesh and linear algebra concepts
-  respectively.
-
- The main mesh classes are:
-  + [`Mesh`](http://mfem.github.io/doxygen/html/classmfem_1_1Mesh.html)
-  + [`NCMesh`](http://mfem.github.io/doxygen/html/classmfem_1_1NCMesh.html)
-  + [`Element`](http://mfem.github.io/doxygen/html/classmfem_1_1Element.html)
-  + [`ElementTransformation`](http://mfem.github.io/doxygen/html/classmfem_1_1ElementTransformation.html)
-
- The main finite element classes are:
-  + [`FiniteElement`](http://mfem.github.io/doxygen/html/classmfem_1_1FiniteElement.html)
-  + [`FiniteElementCollection`](http://mfem.github.io/doxygen/html/classmfem_1_1FiniteElement.html)
-  + [`FiniteElementSpace`](http://mfem.github.io/doxygen/html/classmfem_1_1FiniteElementSpace.html)
-  + [`GridFunction`](http://mfem.github.io/doxygen/html/classmfem_1_1GridFunction.html)
-  + [`BilinearFormIntegrator`](http://mfem.github.io/doxygen/html/classmfem_1_1BilinearFormIntegrator.html) and [`LinearFormIntegrator`](http://mfem.github.io/doxygen/html/classmfem_1_1LinearFormIntegrator.html)
-  + [`LinearForm`](http://mfem.github.io/doxygen/html/classmfem_1_1LinearFormIntegrator.html), [`BilinearForm`](http://mfem.github.io/doxygen/html/classmfem_1_1BilinearForm.html) and [`MixedBilinearForm`](http://mfem.github.io/doxygen/html/classmfem_1_1MixedBilinearForm.html)
-
- The main linear algebra classes and sources are
-  + [`Operator`](http://mfem.github.io/doxygen/html/classmfem_1_1Operator.html) and [`BilinearForm`](http://mfem.github.io/doxygen/html/classmfem_1_1BilinearForm.html)
-  + [`Vector`](http://mfem.github.io/doxygen/html/classmfem_1_1BilinearForm.html) and [`LinearForm`](http://mfem.github.io/doxygen/html/classmfem_1_1LinearForm.html)
-  + [`DenseMatrix`](http://mfem.github.io/doxygen/html/classmfem_1_1DenseMatrix.html) and [`SparseMatrix`](http://mfem.github.io/doxygen/html/classmfem_1_1SparseMatrix.html)
-  + Sparse [smoothers](http://mfem.github.io/doxygen/html/sparsesmoothers_8hpp.html) and linear [solvers](http://mfem.github.io/doxygen/html/solvers_8hpp.html)
-
- Parallel MPI objects in MFEM inherit their serial counterparts, so a parallel
-  mesh for example is just a serial mesh on each task plus the information on
-  shared geometric entities between different tasks. The parallel source files
-  have a `p` prefix, e.g. `pmesh.cpp` vs. the serial `mesh.cpp`.
-
- The main parallel classes are
-  + [`ParMesh`](http://mfem.github.io/doxygen/html/solvers_8hpp.html)
-  + [`ParNCMesh`](http://mfem.github.io/doxygen/html/classmfem_1_1ParMesh.html)
-  + [`ParFiniteElementSpace`](http://mfem.github.io/doxygen/html/classmfem_1_1ParFiniteElementSpace.html)
-  + [`ParGridFunction`](http://mfem.github.io/doxygen/html/classmfem_1_1ParGridFunction.html)
-  + [`ParBilinearForm`](http://mfem.github.io/doxygen/html/classmfem_1_1ParBilinearForm.html) and [`ParLinearForm`](http://mfem.github.io/doxygen/html/classmfem_1_1ParLinearForm.html)
-  + [`HypreParMatrix`](http://mfem.github.io/doxygen/html/classmfem_1_1HypreParMatrix.html) and [`HypreParVector`](http://mfem.github.io/doxygen/html/classmfem_1_1HypreParVector.html)
-  + [`HypreSolver`](http://mfem.github.io/doxygen/html/classmfem_1_1HypreSolver.html) and other [hypre classes](http://mfem.github.io/doxygen/html/hypre_8hpp.html)
-
- GPU and multi-core CPU support is based on device kernels supporting different
-  backends (CUDA, OCCA, RAJA, OpenMP, etc.) and an internal lightweight
-  device/host memory manager.
-
- The main device-relevant classes and sources are:
-  + [`Device`](http://mfem.github.io/doxygen/html/device_8hpp.html)
-  + [`MemoryManager`](http://mfem.github.io/doxygen/html/mem_manager_8hpp.html)
-  + the [`MFEM_FORALL`](http://mfem.github.io/doxygen/html/forall_8hpp.html) macro
-  + the [`cuda.hpp`](http://mfem.github.io/doxygen/html/cuda_8hpp.html) and [`occa.hpp`](http://mfem.github.io/doxygen/html/occa_8hpp.html) files
-
- The `general/` directory contains C++ classes that serve as utilities for
-  communication, error handling, arrays, (Boolean) tables, timing, etc.
-
- The `config/` directory contains build-related files, both for the plain
-  Makefile and the CMake build options.
-
- The `doc/` directory contains configuration for the Doxygen code documentation
-  that can either be build locally, or browsed online at
-  http://mfem.github.io/doxygen/html/index.html.
-
- The `data/` directory contains a collection of small mesh files, that are used
-  in the simple example codes and more fully-featured mini applications in the
-  `examples/` and `miniapps/` directories.
-
- The `tests/` directory contains a unit test suite and will later contain more
-  tests that run example codes.
-
- See also the [code overview](http://mfem.org/code-overview/) section on the
-  MFEM website.
-
 ## GitHub Workflow

 The GitHub organization, https://github.com/mfem, is the main developer hub for
@@ -241,7 +122,7 @@ will allow us to reach you directly with project announcements.
  # Work on "feature-dev", add local commits
  # ...

-  # (One time only) push the branch to github and setup your local
+  # One time only) push the branch to github and setup your local
  # branch to track the github branch (for "git pull"):
  git push -u origin feature-dev

@@ -334,9 +215,9 @@ Before a PR can be merged, it should satisfy the following:
    - [ ] Is this a new feature users need to be aware of? New or updated example or miniapp?
    - [ ] Does it make sense to create a new section in the `CHANGELOG` to group with other related features?
 - [ ] Update `INSTALL`:
-    - [ ] Had a new optional library been added? (*Make sure the external library is licensed under LGPL, not GPL!*)
+    - [ ] Has a new optional library been added? (*Make sure the external library is licensed under LGPL, not GPL!*)
    - [ ] Does `make` or `cmake` have a new target?
-    - [ ] Did the requirements or the installation process change? *(rare)*
+    - [ ] Did the requirements or the installation process change? *(rare)*.
 - [ ] Update `.gitignore`:
    - [ ] Check if `make distclean; git status` shows any files that are generated from the source but we don't want to track in the repository.
    - [ ] Add new patterns (just for the new files above) and re-run the above test.
@@ -376,10 +257,10 @@ Before a PR can be merged, it should satisfy the following:
   - [ ] If this is a major new feature, consider mentioning in the short summary inside `README` *(rare)*.
   - [ ] List major new classes in `doc/CodeDocumentation.dox` *(rare)*.
 - [ ] Update this checklist, if the new pull request affects it.
- [ ] Run the unit tests and make sure they all pass `make unittest`.
 - [ ] (LLNL only) Clone the `tests` repository and run the following tests, see `mfem/tests/README.md`:
   - [ ] `compilers`
   - [ ] `memcheck`
+   - [ ] `unit-test`
   - [ ] `documentation`
 - [ ] (LLNL only) After merging:
   - [ ] Regenerate `README.html` files from companion documentation pull requests.
@@ -451,7 +332,7 @@ MFEM uses a `master`/`next`-branch workflow as described below:
    - [ ] `CHANGELOG`
    - [ ] `makefile`
    - [ ] `CMakeLists.txt`
-    - [ ] `doc/CodeDocumentation.conf.in`
+    - [ ] `doc/CodeDocumentation.conf`
 - [ ] (LLNL only) Make sure all `README.html` files in the source repo are up to date.
 - [ ] Tag the repository:

@@ -490,6 +371,7 @@ MFEM uses a `master`/`next`-branch workflow as described below:
  - `mfem:gh-next` -- Bleeding-edge development version, may be broken, use at
     your own risk.

+
 ## Automated Testing

 MFEM has several levels of automated testing running on GitHub, as well as on
@@ -13,38 +13,14 @@ of MFEM is a (modern) C++ compiler, such as g++. The parallel version of MFEM
 requires an MPI C++ compiler, as well as the following external libraries:

 - hypre (a library of high-performance preconditioners)
-  https://github.com/hypre-space/hypre
+  http://www.llnl.gov/CASC/hypre

 - METIS (a family of multilevel partitioning algorithms)
  http://glaros.dtc.umn.edu/gkhome/metis/metis/overview

-The hypre dependency can be downloaded as a tarball from GitHub or from the
-project webpage https://www.llnl.gov/casc/hypre. For example, the 2.16.0 release
-of hypre is available at
-
-   https://github.com/hypre-space/hypre/archive/v2.16.0.tar.gz
-
 The METIS dependency can be disabled but that is not generally recommended, see
 the option MFEM_USE_METIS.

-MFEM also includes support for devices such as GPUs, and programming models such
-as CUDA, OCCA, OpenMP and RAJA.
-
- Starting with version 4.0, MFEM requires a C++11 compiler
-
- CUDA support requires an NVIDIA GPU and an installation of the CUDA Toolkit
-  https://developer.nvidia.com/cuda-toolkit
-
- OCCA support requires the OCCA library
-  https://libocca.org
-
- OpenMP support requires a compiler implementing the OpenMP API
-  https://www.openmp.org
-
- RAJA support requires installation of the RAJA performance portability layer
-  with (optionally) support for CUDA and OpenMP
-  https://github.com/LLNL/RAJA
-
 The library supports two build systems: one based on GNU make, and a second one
 based on CMake. Both build systems are described below. Some hints for building
 without GNU make or CMake can be found at the end of this file.
@@ -54,11 +30,8 @@ following package managers:

 - Spack, https://github.com/spack/spack
 - OpenHPC, http://openhpc.community
- Homebrew/Science, https://github.com/Homebrew/homebrew-science (deprecated)
+- Homebrew/Science, https://github.com/Homebrew/homebrew-science

-We also recommend downloading and building the MFEM-based GLVis visualization
-tool which can be used to visualize the meshes and solution in MFEM's examples
-and miniapps. See http://glvis.org and http://mfem.org/building.

 Quick start with GNU make
 =========================
@@ -66,15 +39,11 @@ Serial build:
   make serial -j 4

 Parallel build:
-   (download hypre and METIS 4 from above URLs)
+   (download hypre 2.10.0b and METIS 4 from above URLs)
   (build METIS 4 in ../metis-4.0 relative to mfem/)
-   (build hypre in ../hypre relative to mfem/)
+   (build hypre 2.10.0b in ../hypre-2.10.0b relative to mfem/)
   make parallel -j 4

-CUDA build:
-   make cuda -j 4
-   (build for a specific compute capability: 'make cuda -j 4 CUDA_ARCH=sm_30')
-
 Example codes (serial/parallel, depending on the build):
   cd examples
   make -j 4
@@ -85,6 +54,7 @@ Build everything (library, examples and miniapps) with current configuration:
 Quick-check the build by running Example 1/1p (optional):
   make check

+
 Quick start with CMake
 ======================
 Serial build:
@@ -93,19 +63,13 @@ Serial build:
   make -j 4   (assuming "UNIX Makefiles" generator)

 Parallel build:
-   (download hypre and METIS 4 from above URLs)
+   (download hypre 2.10.0b and METIS 4 from above URLs)
   (build METIS 4 in ../metis-4.0 relative to mfem/)
-   (build hypre in ../hypre relative to mfem/)
+   (build hypre 2.10.0b in ../hypre-2.10.0b relative to mfem/)
   mkdir <mfem-build-dir> ; cd <mfem-build-dir>
   cmake <mfem-source-dir> -DMFEM_USE_MPI=YES
   make -j 4

-CUDA build:
-   (this build requires CMake 3.8 or newer)
-   mkdir <mfem-build-dir> ; cd <mfem-build-dir>
-   cmake <mfem-source-dir> -DMFEM_USE_CUDA=YES
-   make -j 4
-
 Example codes (serial/parallel, depending on the build):
   make examples -j 4

@@ -165,10 +129,6 @@ are also defined:
   make parallel -> Builds parallel optimized version of the library
   make debug    -> Builds serial debug version of the library
   make pdebug   -> Builds parallel debug version of the library
-   make cuda     -> Builds serial cuda optimized version of the library
-   make pcuda    -> Builds parallel cuda optimized version of the library
-   make cudebug  -> Builds serial cuda debug version of the library
-   make pcudebug -> Builds parallel cuda debug version of the library

 Note that any of the above shortcuts accept configuration options, either at the
 command line or through a user configuration file.
@@ -230,9 +190,8 @@ Configuration options (GNU make)
 See the configuration file config/defaults.mk for the default settings.

 Compilers:
-   CXX      - C++ compiler, serial build
-   MPICXX   - MPI C++ compiler, parallel build
-   CUDA_CXX - The CUDA compiler, 'nvcc'
+   CXX    - C++ compiler, serial build
+   MPICXX - MPI C++ compiler, parallel build

 Compiler options:
   OPTIM_FLAGS - Options for optimized build
@@ -268,7 +227,7 @@ MFEM_DEBUG = YES/NO
   and consistency checks that may simplify bug-hunting.

 MFEM_USE_EXCEPTIONS = YES/NO
-   Enable the use of exceptions. In particular, modifies the default behavior
+   Enable the use of exceptions. In particular, modifies the default bahavior
   when errors are encountered: throw an exception, instead of aborting.

 MFEM_USE_LIBUNWIND = YES/NO
@@ -288,12 +247,8 @@ MFEM_THREAD_SAFE = YES/NO
   Use thread-safe implementation for some classes/methods. This comes at the
   cost of extra memory allocation and de-allocation.

-MFEM_USE_LEGACY_OPENMP = YES/NO
-   Enable (basic) experimental OpenMP support. Requires MFEM_THREAD_SAFE.
-   This option is deprecated.
-
 MFEM_USE_OPENMP = YES/NO
-   Enable the OpenMP backend.
+   Enable (basic) experimental OpenMP support. Requires MFEM_THREAD_SAFE.

 MFEM_USE_MEMALLOC = YES/NO
   Internal MFEM option: enable batch allocation for some small objects.
@@ -397,39 +352,6 @@ MFEM_USE_GZSTREAM = YES/NO
   before attempting to use it with MFEM.
   When enabled, this option uses the ZLIB_* library options, see below.

-MFEM_USE_PUMI = YES/NO
-   Enable the usage of PUMI (https://scorec.rpi.edu/pumi/) in MFEM. The Parallel
-   Unstructured Mesh Infrastructure (PUMI) is an unstructured, distributed mesh
-   data management system that is capable of handling general non-manifold
-   models and effectively supports automated adaptive analysis. PUMI enables
-   support for parallel unstructured mesh modifications in MFEM.
-
-MFEM_USE_MM = YES/NO
-   Enables support for the MFEM's memory manager (MM), which is required to
-   support devices with different memory spaces. This option is required when
-   CUDA support is enabled, i.e. when MFEM_USE_CUDA=YES.
-
-MFEM_USE_CUDA = YES/NO
-   Enables support for CUDA devices in MFEM. CUDA is a parallel computing
-   platform and programming model for general computing on graphical processing
-   units (GPUs). This option requires MFEM_USE_MM. The variable CUDA_ARCH is
-   used to specify the CUDA compute capability used during compilation (by
-   default, CUDA_ARCH=sm_60). When enabled, this option uses the CUDA_* build
-   options, see below.
-
-MFEM_USE_RAJA = YES/NO
-   Enable support for the RAJA performance portability layer in MFEM. RAJA
-   provides a portable abstraction for loops, supporting different programming
-   model backends. When using RAJA built with CUDA support, CUDA support must be
-   also enabled in MFEM, i.e. MFEM_USE_CUDA=YES must be set.
-
-MFEM_USE_OCCA = YES/NO
-   Enables support for the OCCA library in MFEM. OCCA is an open-source library
-   which aims to make it easy to program different types of devices (e.g. CPU,
-   GPU, FPGA) by providing an unified API for interacting with JIT-compiled
-   backends. In order to use the OCCA CUDA backend, CUDA support must be enabled
-   in MFEM as well, i.e. MFEM_USE_CUDA=YES must be set.
-
 MFEM_BUILD_TAG = (any value)
   An optional tag to characterize the build. Exported to config/config.mk.
   Can be used to identify the MFEM build from other makefiles.
@@ -451,7 +373,7 @@ directory and use the string @MFEM_DIR@, e.g. HYPRE_OPT = -I@MFEM_DIR@/../hypre.
 The specific libraries and their options are:

 - HYPRE, required for the parallel build, i.e. when MFEM_USE_MPI = YES.
-  URL: https://github.com/hypre-space/hypre and https://www.llnl.gov/casc/hypre
+  URL: http://www.llnl.gov/CASC/hypre
  Options: HYPRE_OPT, HYPRE_LIB.

 - METIS, used when MFEM_USE_METIS = YES. If using METIS 5, set
@@ -465,8 +387,7 @@ The specific libraries and their options are:
       http://math-atlas.sourceforge.net (ATLAS)
  Options: LAPACK_OPT (currently not used/needed), LAPACK_LIB.

- OpenMP (optional), usually part of compiler, used when either MFEM_USE_OPENMP
-  or MFEM_USE_LEGACY_OPENMP is set to YES.
+- OpenMP (optional), usually part of compiler, used when MFEM_USE_OPENMP = YES.
  Options: OPENMP_OPT, OPENMP_LIB.

 - High-resolution POSIX clocks: when using MFEM_TIMER_TYPE = 2, it may be
@@ -498,8 +419,7 @@ The specific libraries and their options are:

 - STRUMPACK (optional), used when MFEM_USE_STRUMPACK = YES. Note that STRUMPACK
  requires the PT-Scotch and Scalapack libraries as well as ParMETIS, which
-  includes METIS 5 in its distribution. Starting with STRUMPACK v2.2.0, ParMETIS
-  and PT-Scotch are optional dependencies.
+  includes METIS 5 in its distribution.
  The support for STRUMPACK was added in MFEM v3.3.2 and it requires STRUMPACK
  2.0.0 or later.
  URL: http://portal.nersc.gov/project/sparse/strumpack
@@ -541,22 +461,6 @@ The specific libraries and their options are:
       https://support.hdfgroup.org/HDF5 (HDF5)
  Options: CONDUIT_OPT, CONDUIT_LIB.

- PUMI, used when MFEM_USE_PUMI = YES.
-  URL: https://scorec.rpi.edu/pumi
-  Options: PUMI_OPT, PUMI_LIB.
-
- CUDA, used when MFEM_USE_CUDA = YES.
-  URL: https://developer.nvidia.com/cuda-toolkit
-  Options: CUDA_CXX, CUDA_ARCH, CUDA_OPT, CUDA_LIB.
-
- OCCA, used when MFEM_USE_OCCA = YES.
-  URL: https://libocca.org
-  Options: OCCA_DIR, OCCA_OPT, OCCA_LIB.
-
- RAJA, used when MFEM_USE_RAJA = YES.
-  URL: https://github.com/LLNL/RAJA
-  Options: RAJA_DIR, RAJA_OPT, RAJA_LIB.
-
 - MPFR (optional), used when MFEM_USE_MPFR = YES.
  URL: http://mpfr.org, it depends on the GMP library: https://gmplib.org
  Options: MPFR_OPT, MPFR_LIB.
@@ -661,8 +565,6 @@ Configuration variables (CMake)
 ===============================
 See the configuration file config/defaults.cmake for the default settings.

-Note: the option MFEM_USE_CUDA requires CMake version 3.8 or newer!
-
 Non-standard CMake variables for compilers:
   CXX    - If set, overwrite the auto-detected C++ compiler, serial build
   MPICXX - If set, overwrite the auto-detected MPI C++ compiler, parallel build
@@ -680,7 +582,6 @@ MFEM_USE_METIS - Set to ${MFEM_USE_MPI}, can be overwritten.
 MFEM_USE_LIBUNWIND
 MFEM_USE_LAPACK
 MFEM_THREAD_SAFE
-MFEM_USE_LEGACY_OPENMP
 MFEM_USE_OPENMP
 MFEM_USE_MEMALLOC
 MFEM_TIMER_TYPE - Set automatically, can be overwritten.
@@ -692,11 +593,6 @@ MFEM_USE_GNUTLS
 MFEM_USE_NETCDF
 MFEM_USE_MPFR
 MFEM_USE_GZSTREAM
-MFEM_USE_PUMI
-MFEM_USE_CUDA
-MFEM_USE_OCCA
-MFEM_USE_RAJA
-MFEM_USE_MM

 The following options are CMake specific:

@@ -742,9 +638,6 @@ The CMake build system adds auto-detection for the following packages/libraries:
 - MPFR
 - LIBUNWIND
 - POSIXCLOCKS
- - PUMI
- - OCCA
- - RAJA

 The following built-in CMake packages are also used:

@@ -1,5 +1,5 @@
-                  GNU LESSER GENERAL PUBLIC LICENSE
-                       Version 2.1, February 1999
+		  GNU LESSER GENERAL PUBLIC LICENSE
+		       Version 2.1, February 1999

 Copyright (C) 1991, 1999 Free Software Foundation, Inc.
 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
@@ -10,7 +10,7 @@
 as the successor of the GNU Library Public License, version 2, hence
 the version number 2.1.]

-                            Preamble
+			    Preamble

  The licenses for most software are designed to take away your
 freedom to share and change it.  By contrast, the GNU General Public
@@ -112,7 +112,7 @@ modification follow.  Pay close attention to the difference between a
 former contains code derived from the library, whereas the latter must
 be combined with the library in order to run.

-                  GNU LESSER GENERAL PUBLIC LICENSE
+		  GNU LESSER GENERAL PUBLIC LICENSE
   TERMS AND CONDITIONS FOR COPYING, DISTRIBUTION AND MODIFICATION

  0. This License Agreement applies to any software library or other
@@ -146,7 +146,7 @@ such a program is covered only if its contents constitute a work based
 on the Library (independent of the use of the Library in a tool for
 writing it).  Whether that is true depends on what the Library does
 and what the program that uses the Library does.
-
+  
  1. You may copy and distribute verbatim copies of the Library's
 complete source code as you receive it, in any medium, provided that
 you conspicuously and appropriately publish on each copy an
@@ -432,7 +432,7 @@ decision will be guided by the two goals of preserving the free status
 of all derivatives of our free software and of promoting the sharing
 and reuse of software generally.

-                            NO WARRANTY
+			    NO WARRANTY

  15. BECAUSE THE LIBRARY IS LICENSED FREE OF CHARGE, THERE IS NO
 WARRANTY FOR THE LIBRARY, TO THE EXTENT PERMITTED BY APPLICABLE LAW.
@@ -455,7 +455,7 @@ FAILURE OF THE LIBRARY TO OPERATE WITH ANY OTHER SOFTWARE), EVEN IF
 SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH
 DAMAGES.

-                     END OF TERMS AND CONDITIONS
+		     END OF TERMS AND CONDITIONS

           How to Apply These Terms to Your New Libraries

@@ -485,8 +485,7 @@ convey the exclusion of warranty; and each file should have at least the

    You should have received a copy of the GNU Lesser General Public
    License along with this library; if not, write to the Free Software
-    Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301
-    USA
+    Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA

 Also add information on how to contact you by electronic and paper mail.

@@ -495,10 +494,11 @@ school, if any, to sign a "copyright disclaimer" for the library, if
 necessary.  Here is a sample; alter the names:

  Yoyodyne, Inc., hereby disclaims all copyright interest in the
-  library `Frob' (a library for tweaking knobs) written by James Random
-  Hacker.
+  library `Frob' (a library for tweaking knobs) written by James Random Hacker.

  <signature of Ty Coon>, 1 April 1990
  Ty Coon, President of Vice

 That's all there is to it!
+
+
@@ -8,19 +8,15 @@
                               http://mfem.org

 MFEM is a modular parallel C++ library for finite element methods. Its goal is
-to enable high-performance scalable finite element discretization research and
-application development on a wide variety of platforms, ranging from laptops to
-supercomputers.
+to enable the research and development of scalable finite element discretization
+and solver algorithms through general finite element abstractions, accurate and
+flexible visualization, and tight integration with the hypre library.

-* For building instructions, see the file INSTALL, or type "make help".
+For building instructions, see the file INSTALL, or type "make help". Copyright
+information and licensing restrictions can be found in the file COPYRIGHT.

-* Copyright and licensing information can be found in the file COPYRIGHT.
-
-* The best starting point for new users interested in MFEM's features is the
-  interactive documentation in examples/README.html.
-
-* Developers interested in contributing to the library, should read the
-  instructions and documentation in the CONTRIBUTING.md file.
+The best starting point for new users interested in MFEM's features is the
+interactive documentation in examples/README.html.

 Conceptually, MFEM can be viewed as a finite element toolbox that provides the
 building blocks for developing finite element algorithms in a manner similar to
@@ -39,30 +35,29 @@ conforming and non-conforming (AMR) adaptive refinement. Arbitrary element
 transformations, allowing for high-order mesh elements with curved boundaries,
 are also supported.

-When used as a "finite element to linear algebra translator", MFEM can take a
-problem described in terms of finite element-type objects, and produce the
-corresponding linear algebra vectors and fully or partially assembled operators,
-e.g. in the form of global sparse matrices or matrix-free operators. The library
-includes simple smoothers and Krylov solvers, such as PCG, MINRES and GMRES, as
-well as support for sequential sparse direct solvers from the SuiteSparse
+MFEM is commonly used as a "finite element to linear algebra translator", since
+it can take a problem described in terms of finite element-type objects, and
+produce the corresponding linear algebra vectors and sparse matrices. In order
+to facilitate this, MFEM uses compressed sparse row (CSR) sparse matrix storage
+and includes simple smoothers and Krylov solvers, such as PCG, MINRES and GMRES,
+as well as support for sequential sparse direct solvers from the SuiteSparse
 library. Nonlinear solvers (the Newton method), eigensolvers (LOBPCG), and
 several explicit and implicit Runge-Kutta time integrators are also available.

 MFEM supports MPI-based parallelism throughout the library, and can readily be
-used as a scalable unstructured finite element problem generator. As of version
-4.0, MFEM offers initial support for GPU acceleration, and programming models,
-such as CUDA, OCCA, RAJA and OpenMP. MFEM-based applications require minimal
-changes to switch from a serial to a high-performing MPI-parallel version of the
-code, where they can take advantage of the integrated linear solvers from the
-hypre library.  Comprehensive support for other external packages, e.g. PETSc
-and SUNDIALS is also included, giving access to many additional linear and
-nonlinear solvers, preconditioners, time integrators, etc.
+used as a scalable unstructured finite element problem generator. MFEM-based
+applications require minimal changes to transition from a serial to a
+high-performing parallel version of the code, where they can take advantage of
+the integrated scalable linear solvers from the hypre library. Comprehensive
+support for other external packages, e.g. PETSc and SUNDIALS is also included,
+giving access to many additional linear and nonlinear solvers, preconditioners,
+time integrators, etc.

 For examples of using MFEM, see the examples/ and miniapps/ directories, as well
 as the OpenGL visualization tool GLVis which is available at http://glvis.org.

-This project is released under the LGPL v2.1 license with static linking
-exception. See files COPYRIGHT and LICENSE file for full details.
+This project is released under the LGPL v2.1 license. See LICENSE file for full
+details.

 LLNL Release Number: LLNL-CODE-443211
 DOI: 10.11578/dc.20171025.1248
@@ -9,5 +9,23 @@
 // terms of the GNU Lesser General Public License (as published by the Free
 // Software Foundation) version 2.1 dated February 1999.

-#define CATCH_CONFIG_MAIN     // This tells Catch to provide a main() - only do this in one cpp file
-#include "catch.hpp"
+#ifndef MFEM_BACKENDS_ALL_HPP
+#define MFEM_BACKENDS_ALL_HPP
+
+#include "../config/config.hpp"
+
+#ifdef MFEM_USE_BACKENDS
+
+#include "base/backend.hpp"
+
+#ifdef MFEM_USE_OCCA
+#include "occa/backend.hpp"
+#endif
+
+#ifdef MFEM_USE_OMP
+#include "omp/backend.hpp"
+#endif
+
+#endif // MFEM_USE_BACKENDS
+
+#endif // MFEM_BACKENDS_ALL_HPP
@@ -0,0 +1,215 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_BASE_ARRAY_HPP
+#define MFEM_BACKENDS_BASE_ARRAY_HPP
+
+#include "../../config/config.hpp"
+#ifdef MFEM_USE_BACKENDS
+
+#include "layout.hpp"
+#include "utils.hpp"
+
+namespace mfem
+{
+
+/// Extension to the template class Array<T>
+class PArray : public RefCounted
+{
+protected:
+   /// Layout with shared ownership (smart pointer)
+   DLayout layout;
+
+
+   /**
+       @name Virtual interface
+    */
+   ///@{
+
+   virtual void *DoGetData() const = 0;
+
+   /** @brief Create and return a new array (in @a *clone) of the same dynamic
+       type as this array using the same layout and ItemSize().
+
+       Set @a *clone to NULL if allocation fails.
+
+       If @a copy_data is true, the contents of this array is copied to the new
+       array; otherwise, the new array remains uninitialized.
+
+       If @a buffer is not NULL, return the array data of the newly created
+       object (in @a *buffer) , if it is stored as a contiguous array on the
+       host; otherwise, set @a *buffer to NULL. */
+   virtual PArray *DoClone(bool copy_data, void **buffer,
+                           std::size_t item_size) const = 0;
+
+   /// Resize the array, reallocating its data if necessary.
+   /** If @a buffer is not NULL, return the array data (in @a *buffer), if it
+       is stored as a contiguous array on the host; otherwise, set @a *buffer to
+       NULL. Returns 0 on success and non-zero otherwise, e.g. if memory
+       allocation fails.
+
+       If the @a new_layout is not supported, a non-zero error code will be
+       returned.
+
+       The @a new_layout has to be valid, i.e. new_layout != NULL and
+       new_layout->HasEngine() == true.
+
+       @note If reallocation is performed, the previous content of the array is
+       NOT copied to the new location. */
+   virtual int DoResize(PLayout &new_layout, void **buffer,
+                        std::size_t item_size) = 0;
+
+   /** @brief Get access to the contents of the array in host memory, as a
+       contiguous array. */
+   /** If the array data is stored as a contiguous array in host memory, return
+       a pointer to it. Otherwise, copy the data to @a buffer (if @a buffer is
+       not NULL) and return @a buffer.
+       @note If not NULL, @a buffer is assumed to be of size greater than or
+       equal to Size(). */
+   virtual void *DoPullData(void *buffer, std::size_t item_size) = 0;
+
+   /** @brief Set all entries of the array to the (single) value pointed to by
+       @a value_ptr. */
+   virtual void DoFill(const void *value_ptr, std::size_t item_size) = 0;
+
+   /** @brief Set all Size() entries of the array from the given contiguous
+       array, @a src_buffer, on the host. */
+   virtual void DoPushData(const void *src_buffer, std::size_t item_size) = 0;
+
+   /// Copy the data from @a src to @a *this.
+   /** Both arrays must have the same dynamic type, layout, and item_size. */
+   virtual void DoAssign(const PArray &src, std::size_t item_size) = 0;
+
+   ///@}
+   // End: Virtual interface
+
+public:
+   /** @brief The @a layout parameter will be reference counted and therefore it
+       should be dynamically allocated. */
+   /** The @a layout must be valid in the sense that layout != NULL and
+       layout->HasEngine() == true. */
+   PArray(PLayout &p_layout)
+      : layout(&p_layout)
+   {
+      MFEM_ASSERT(layout && layout->HasEngine(), "invalid layout");
+   }
+
+   virtual ~PArray() { }
+
+   /// Get the current size of the array.
+   std::size_t Size() const { return layout->Size(); }
+
+   /// Get the current layout of the array.
+   PLayout &GetLayout() const { return *layout; }
+
+   /// TODO
+   template <typename derived_t>
+   derived_t &As() { return dynamic_cast<derived_t&>(*this); }
+
+   /// TODO
+   template <typename derived_t>
+   const derived_t &As() const { return dynamic_cast<const derived_t&>(*this); }
+
+   // TODO: Error handling ... handle errors at the Engine level, at the class
+   //       level, or at the method level?
+
+   // TODO: Asynchronous execution interface ...
+
+
+   /**
+       @name Public virtual interface
+    */
+   ///@{
+
+   template <typename T=void>
+   T* GetData() const { return (T*) DoGetData(); }
+
+   /** @brief Create and return a new array (in @a *clone) of the same dynamic
+       type as this array using the same layout and ItemSize().
+
+       Set @a *clone to NULL if allocation fails.
+
+       If @a copy_data is true, the contents of this array is copied to the new
+       array; otherwise, the new array remains uninitialized.
+
+       If @a buffer is not NULL, return the array data of the newly created
+       object (in @a *buffer) , if it is stored as a contiguous array on the
+       host; otherwise, set @a *buffer to NULL. */
+   template <typename T>
+   DArray Clone(bool copy_data, T **buffer) const
+   { return DArray(DoClone(copy_data, (void**)buffer, sizeof(T))); }
+
+   /// Resize the array, reallocating its data if necessary.
+   /** If @a buffer is not NULL, return the array data (in @a *buffer), if it
+       is stored as a contiguous array on the host; otherwise, set @a *buffer to
+       NULL. Returns 0 on success and non-zero otherwise, e.g. if memory
+       allocation fails.
+
+       If the @a new_layout is not supported, a non-zero error code will be
+       returned.
+
+       The @a new_layout has to be valid, i.e. new_layout != NULL and
+       new_layout->HasEngine() == true.
+
+       @note If reallocation is performed, the previous content of the array is
+       NOT copied to the new location. */
+   template <typename T>
+   int Resize(PLayout &new_layout, T **buffer)
+   { return DoResize(new_layout, (void**)buffer, sizeof(T)); }
+
+   /// Shortcut for Resize(*layout, buffer).
+   /** This method is useful for updating the array after its layout is changed
+       externally. */
+   template <typename T>
+   int Update(T **buffer)
+   { return DoResize(*layout, (void**)buffer, sizeof(T)); }
+
+   /// Shortcut for layout->Resize(new_size) followed by Update()
+   template <typename T>
+   int Resize(std::size_t new_size, T **buffer)
+   { layout->Resize(new_size); return Update(buffer); }
+
+   /** @brief Get access to the contents of the array in host memory, as a
+       contiguous array. */
+   /** If the array data is stored as a contiguous array in host memory, return
+       a pointer to it. Otherwise, copy the data to @a buffer (if @a buffer is
+       not NULL) and return @a buffer.
+       @note If not NULL, @a buffer is assumed to be of size greater than or
+       equal to Size(). */
+   template <typename T>
+   T *PullData(T *buffer)
+   { return Size() ? (T*)DoPullData((void*)buffer, sizeof(T)) : NULL; }
+
+   /** @brief Set all entries of the array to the (single) value pointed to by
+       @a value_ptr. */
+   template <typename T>
+   void Fill(const T &value) { if (Size()) { DoFill(&value, sizeof(T)); } }
+
+   /** @brief Set all Size() entries of the array from the given contiguous
+       array, @a src_buffer, on the host. */
+   template <typename T>
+   void PushData(const T *src_buffer)
+   { if (Size()) { DoPushData(src_buffer, sizeof(T)); } }
+
+   /// Copy the data from @a src to @a *this.
+   /** Both arrays must have the same dynamic type, layout, and entry type. */
+   template <typename T>
+   void Assign(const PArray &src) { if (Size()) { DoAssign(src, sizeof(T)); } }
+
+   ///@}
+   // End: Virtual interface
+};
+
+} // namespace mfem
+
+#endif // MFEM_USE_BACKENDS
+
+#endif // MFEM_BACKENDS_BASE_ARRAY_HPP
@@ -0,0 +1,57 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_BASE_BACKEND_HPP
+#define MFEM_BACKENDS_BASE_BACKEND_HPP
+
+#include "../../config/config.hpp"
+#ifdef MFEM_USE_BACKENDS
+
+#include "memory_resource.hpp"
+#include "engine.hpp"
+#include "array.hpp"
+#include "vector.hpp"
+#include "fespace.hpp"
+#include "bilinearform.hpp"
+
+#include <string>
+
+#ifdef MFEM_USE_MPI
+#include <mpi.h>
+#endif
+
+namespace mfem
+{
+
+/// TODO
+class Backend
+{
+public:
+   /// TODO
+   virtual ~Backend() { }
+
+   /// TODO
+   virtual bool Supports(const std::string &engine_spec) const = 0;
+
+   /// TODO
+   virtual Engine *Create(const std::string &engine_spec) = 0;
+
+#ifdef MFEM_USE_MPI
+   /// TODO
+   virtual Engine *Create(MPI_Comm comm, const std::string &engine_spec) = 0;
+#endif
+};
+
+} // namespace mfem
+
+#endif // MFEM_USE_BACKENDS
+
+#endif // MFEM_BACKENDS_BASE_BACKEND_HPP
@@ -0,0 +1,72 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_BASE_BILINEARFORM_HPP
+#define MFEM_BACKENDS_BASE_BILINEARFORM_HPP
+
+#include "../../config/config.hpp"
+#ifdef MFEM_USE_BACKENDS
+
+#include "engine.hpp"
+
+namespace mfem
+{
+
+class Vector;
+class OperatorHandle;
+class BilinearForm;
+
+/// TODO: doxygen
+class PBilinearForm : public RefCounted
+{
+protected:
+   /// Engine with shared ownership
+   SharedPtr<const Engine> engine;
+   /// Not owned.
+   BilinearForm *bform;
+
+public:
+   /// TODO: doxygen
+   PBilinearForm(const Engine &e, BilinearForm &bf)
+      : engine(&e), bform(&bf) { }
+
+   /// Virtual destructor
+   virtual ~PBilinearForm() { }
+
+   /// Get the associated Engine
+   const Engine &GetEngine() const { return *engine; }
+
+   /// Assemble the PBilinearForm.
+   /** This method is called from the method BilinearForm::Assemble() of the
+       associated BilinearForm #bform.
+       @returns True, if the host assembly should be skipped. */
+   virtual bool Assemble() = 0;
+
+   /// TODO: doxygen
+   virtual void FormSystemMatrix(const Array<int> &ess_tdof_list,
+                                 OperatorHandle &A) = 0;
+
+   /// TODO: doxygen
+   virtual void FormLinearSystem(const Array<int> &ess_tdof_list,
+                                 Vector &x, Vector &b,
+                                 OperatorHandle &A, Vector &X, Vector &B,
+                                 int copy_interior) = 0;
+
+   /// TODO: doxygen
+   virtual void RecoverFEMSolution(const Vector &X, const Vector &b,
+                                   Vector &x) = 0;
+};
+
+} // namespace mfem
+
+#endif // MFEM_USE_BACKENDS
+
+#endif // MFEM_BACKENDS_BASE_BILINEARFORM_HPP
@@ -0,0 +1,29 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#ifdef MFEM_USE_BACKENDS
+
+#include "engine.hpp"
+#include "fespace.hpp"
+#include "bilinearform.hpp"
+
+namespace mfem
+{
+
+DFiniteElementSpace Engine::MakeFESpace(FiniteElementSpace &fes) const
+{
+   return DFiniteElementSpace(new PFiniteElementSpace(*this, fes));
+}
+
+} // namespace mfem
+
+#endif // MFEM_USE_BACKENDS
@@ -0,0 +1,210 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_BASE_ENGINE_HPP
+#define MFEM_BACKENDS_BASE_ENGINE_HPP
+
+#include "../../config/config.hpp"
+#ifdef MFEM_USE_BACKENDS
+
+#include "../../general/scalars.hpp"
+#include "memory_resource.hpp"
+#include "smart_pointers.hpp"
+#include "utils.hpp"
+
+#ifdef MFEM_USE_MPI
+#include <mpi.h>
+#endif
+
+namespace mfem
+{
+
+// Forward declarations.
+class Backend;
+template <typename T> class Array;
+class Vector;
+class Operator;
+class FiniteElementSpace;
+class LinearForm;
+class BilinearForm;
+class MixedBilinearForm;
+class NonlinearForm;
+
+
+/// In parallel, each MPI rank will usually create a single engine.
+class Engine : public RefCounted
+{
+protected:
+   Backend *backend; ///< Backend that created the engine. Not owned.
+
+#ifdef MFEM_USE_MPI
+   MPI_Comm comm; ///< Associated MPI communicator (may be MPI_COMM_NULL).
+#endif
+
+   /// Number of memory resources used by the Engine.
+   int num_mem_res;
+   /// Number of workers used by the Engine.
+   int num_workers;
+
+   /// Memory resources used by the engine - array of pointers.
+   /** Both the array and the entries are owned. */
+   MemoryResource **memory_resources;
+
+   /// Relative computational speed of the workers. Owned.
+   double *workers_weights;
+
+   /// For each worker, which memory resource it uses.
+   int *workers_mem_res;
+
+public:
+   /// TODO: doxygen
+   Engine(Backend *b, int n_mem, int n_workers)
+      : backend(b),
+#ifdef MFEM_USE_MPI
+        comm(MPI_COMM_NULL),
+#endif
+        num_mem_res(n_mem),
+        num_workers(n_workers),
+        memory_resources(new MemoryResource*[num_mem_res]()),
+        workers_weights(new double[num_workers]()),
+        workers_mem_res(new int[num_workers]())
+   { /* Note: all arrays are value-initialized with zeros. */ }
+
+   /// TODO: doxygen
+   virtual ~Engine()
+   {
+      delete [] workers_mem_res;
+      delete [] workers_weights;
+      for (int i = 0; i < num_mem_res; i++)
+      {
+         delete memory_resources[i];
+      }
+      delete [] memory_resources;
+   }
+
+
+   /**
+       @name Machine resources interface
+    */
+   ///@{
+
+#ifdef MFEM_USE_MPI
+   /// Get the associated MPI_Comm
+   MPI_Comm GetComm() const { return comm; }
+#endif
+
+   /// TODO
+   int GetNumMemRes() const { return num_mem_res; }
+
+   /// TODO
+   MemoryResource &GetMemRes(int idx) const { return *memory_resources[idx]; }
+
+   /// TODO
+   int GetNumWorkers() const { return num_workers; }
+
+   /// TODO
+   const double *GetWorkersWeights() const { return workers_weights; }
+
+   /// TODO
+   const int *GetWorkersMemRes() const { return workers_mem_res; }
+
+   ///@}
+   // End: Machine resources interface
+
+   /// TODO
+   template <typename derived_t>
+   derived_t &As() { *util::As<derived_t>(this); }
+
+   /// TODO
+   template <typename derived_t>
+   const derived_t &As() const { *util::As<const derived_t>(this); }
+
+
+   // TODO: Error handling ... handle errors at the Engine level, at the class
+   //       level, or at the method level?
+
+
+   /**
+       @name Virtual interface: finite element data structures and algorithms
+    */
+   ///@{
+
+   // TODO: Asynchronous execution in this class ...
+
+   /// Allocate and return a new layout for the given @a size.
+   /** The layout decomposition is determined automatically by the Engine using
+       a deterministic algorithm: calls to this method with the same @a size
+       will produce the same result, as long as the Engine remains unmodified
+       between the calls.
+
+       The returned object is allocated with operator new and must be
+       deallocated by the caller.
+
+       TODO: Returns NULL if memory allocation fails?
+   */
+   virtual DLayout MakeLayout(std::size_t size) const = 0;
+
+   /// Allocate and return a new layout for the given worker decomposition.
+   /** The returned object is allocated with operator new and must be
+       deallocated by the caller.
+
+       TODO: Returns NULL if memory allocation fails?
+
+       The @a offsets should satisfy: offsets.Size() == number of workers + 1,
+       offsets[0] == 0, and offsets[i] <= offsets[i+1], for i: 0 <= i < number
+       of workers. */
+   virtual DLayout MakeLayout(const Array<std::size_t> &offsets) const = 0;
+
+   // Note: There may be other ways to construct layouts in the future, e.g.
+   //       block-vector layouts, or multi-vector layouts.
+
+   /// TODO
+   virtual DArray MakeArray(PLayout &layout, std::size_t item_size) const = 0;
+
+   /// Allocate and return a new vector using the given @a layout.
+   /** The returned object is a smart pointer that will automatically deallocate
+       the vector.
+
+       TODO: Produce an error if memory allocation fails?
+
+       Only layouts returned by this Engine are guaranteed to be supported.
+       Using a type that is not supported will produce an error. */
+   virtual DVector MakeVector(PLayout &layout,
+                              int type_id = ScalarId<double>::value) const = 0;
+
+   /// TODO: doxygen
+   virtual DFiniteElementSpace MakeFESpace(FiniteElementSpace &fes) const;
+
+   /// TODO: doxygen
+   virtual DBilinearForm MakeBilinearForm(BilinearForm &bf) const = 0;
+
+
+   // Question: How do we construct coefficients?
+
+
+   /// FIXME - What will the actual parameters be?
+   virtual void AssembleLinearForm(LinearForm &l_form) const = 0;
+
+   /// FIXME - What will the actual parameters be?
+   virtual Operator *MakeOperator(const MixedBilinearForm &mbl_form) const = 0;
+
+   /// FIXME - What will the actual parameters be?
+   virtual Operator *MakeOperator(const NonlinearForm &nl_form) const = 0;
+
+   ///@}
+   // End: Virtual interface
+};
+
+} // namespace mfem
+
+#endif // MFEM_USE_BACKENDS
+
+#endif // MFEM_BACKENDS_BASE_ENGINE_HPP
@@ -0,0 +1,61 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_BASE_FE_SPACE_HPP
+#define MFEM_BACKENDS_BASE_FE_SPACE_HPP
+
+#include "../../config/config.hpp"
+#ifdef MFEM_USE_BACKENDS
+
+#include "engine.hpp"
+#include "utils.hpp"
+
+namespace mfem
+{
+
+class FiniteElementSpace;
+
+/// TODO: doxygen
+class PFiniteElementSpace : public RefCounted
+{
+protected:
+   /// Engine with shared ownership
+   SharedPtr<const Engine> engine;
+   /// Not owned.
+   FiniteElementSpace *fes;
+
+public:
+   /// TODO: doxygen
+   PFiniteElementSpace(const Engine &e, FiniteElementSpace &fespace)
+      : engine(&e), fes(&fespace) { }
+
+   /// Virtual destructor
+   virtual ~PFiniteElementSpace() { }
+
+   /// Get the associated engine
+   const Engine &GetEngine() const { return *engine; }
+
+   mfem::FiniteElementSpace* GetFESpace() const { return fes; }
+
+   /// TODO
+   template <typename derived_t>
+   derived_t &As() { return *util::As<derived_t>(this); }
+
+   /// TODO
+   template <typename derived_t>
+   const derived_t &As() const { return *util::As<const derived_t>(this); }
+};
+
+} // namespace mfem
+
+#endif // MFEM_USE_BACKENDS
+
+#endif // MFEM_BACKENDS_BASE_FE_SPACE_HPP
@@ -0,0 +1,104 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_BASE_LAYOUT_HPP
+#define MFEM_BACKENDS_BASE_LAYOUT_HPP
+
+#include "../../config/config.hpp"
+#ifdef MFEM_USE_BACKENDS
+
+#include "smart_pointers.hpp"
+#include "engine.hpp"
+
+namespace mfem
+{
+
+/// Polymorphic layout (array/vector layout descriptor)
+class PLayout : public RefCounted
+{
+protected:
+   /// Engine with shared ownership
+   SharedPtr<const Engine> engine;
+   std::size_t size;
+
+   template <typename DObject>
+   struct Maker
+   {
+      template <typename entry_t>
+      static DObject MakeNew(PLayout &layout);
+   };
+
+public:
+   explicit PLayout(std::size_t s = 0) : engine(NULL), size(s) { }
+
+   explicit PLayout(const Engine &e, std::size_t s = 0)
+      : engine(&e), size(s) { }
+
+   virtual ~PLayout() { }
+
+   /**
+       @name Virtual interface
+    */
+   ///@{
+
+   /// Resize the layout
+   virtual void Resize(std::size_t new_size) { size = new_size; }
+
+   /// Resize the layout based on the given worker offsets
+   virtual void Resize(const Array<std::size_t> &offsets)
+   { MFEM_ABORT("method not supported"); }
+
+   ///@}
+   // End: Virtual interface
+
+   /// Layouts without engine cannot create DArray, DVector, etc.
+   bool HasEngine() const { return engine != NULL; }
+
+   /// TODO: doxygen
+   const Engine &GetEngine() const { return *engine; }
+
+   /// TODO: doxygen
+   std::size_t Size() const { return size; }
+
+   /// TODO
+   template <typename derived_t>
+   derived_t &As() { return *util::As<derived_t>(this); }
+
+   /// TODO
+   template <typename derived_t>
+   const derived_t &As() const { return *util::As<const derived_t>(this); }
+
+   /// TODO: doxygen
+   template <typename DObject, typename entry_t>
+   DObject Make()
+   {
+      MFEM_ASSERT(HasEngine(), "this method requires an Engine");
+      return Maker<DObject>::template MakeNew<entry_t>(*this);
+   }
+};
+
+template <> struct PLayout::Maker<DArray>
+{
+   template <typename entry_t> static DArray MakeNew(PLayout &layout)
+   { return layout.GetEngine().MakeArray(layout, sizeof(entry_t)); }
+};
+
+template <> struct PLayout::Maker<DVector>
+{
+   template <typename entry_t> static DVector MakeNew(PLayout &layout)
+   { return layout.GetEngine().MakeVector(layout, ScalarId<entry_t>::value); }
+};
+
+} // namespace mfem
+
+#endif // MFEM_USE_BACKENDS
+
+#endif // MFEM_BACKENDS_BASE_LAYOUT_HPP
@@ -0,0 +1,59 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#ifdef MFEM_USE_BACKENDS
+
+#include "memory_resource.hpp"
+#include "../../general/error.hpp"
+
+#include <cstdlib>
+#include <cstring>
+#include <cerrno>
+
+namespace mfem
+{
+
+void *NewDeleteMemoryResource::DoAllocate(std::size_t bytes,
+                                          std::size_t alignment)
+{
+   void *p = ::operator new[](bytes);
+   MFEM_VERIFY(!alignment || (std::size_t)(p) % alignment == 0,
+               "invalid alignment");
+   return p;
+}
+
+void NewDeleteMemoryResource::DoDeallocate(void *p, std::size_t bytes,
+                                           std::size_t alignment)
+{
+   ::operator delete[](p);
+}
+
+
+void *AlignedMemoryResource::DoAllocate(std::size_t bytes,
+                                        std::size_t alignment)
+{
+   void *p;
+   if (!alignment) { alignment = sizeof(long double); }
+   MFEM_VERIFY(posix_memalign(&p, alignment, bytes) == 0,
+               "error in posix_memalign(): " << strerror(errno));
+   return p;
+}
+
+void AlignedMemoryResource::DoDeallocate(void *p, std::size_t bytes,
+                                         std::size_t alignment)
+{
+   free(p);
+}
+
+} // namespace mfem
+
+#endif // MFEM_USE_BACKENDS
@@ -0,0 +1,70 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_BASE_MEMORY_RESOURCE_HPP
+#define MFEM_BACKENDS_BASE_MEMORY_RESOURCE_HPP
+
+#include "../../config/config.hpp"
+#ifdef MFEM_USE_BACKENDS
+
+#include <cstddef>
+
+namespace mfem
+{
+
+/// Polymorphic memory resource. Similar to C++17's std::pmr::memory_resource.
+class MemoryResource
+{
+protected:
+   virtual void *DoAllocate(std::size_t bytes, std::size_t alignment) = 0;
+   virtual void DoDeallocate(void* p, std::size_t bytes,
+                             std::size_t alignment) = 0;
+
+public:
+   // Implicitly defined default & copy constructors
+
+   /// Virtual destructor.
+   virtual ~MemoryResource() { }
+
+   /// If alignment == 0, use default alignment.
+   void *Allocate(std::size_t bytes, std::size_t alignment = 0)
+   { return DoAllocate(bytes, alignment); }
+
+   /// If alignment == 0, use default alignment.
+   void Deallocate(void *p, std::size_t bytes, std::size_t alignment = 0)
+   { DoDeallocate(p, bytes, alignment); }
+};
+
+
+/** @brief Dynamic host memory resource using operator new[](std::size_t) for
+    allocation and operator delete[](void*) for deallocation. */
+class NewDeleteMemoryResource : public MemoryResource
+{
+protected:
+   virtual void *DoAllocate(std::size_t bytes, std::size_t alignment);
+   virtual void DoDeallocate(void *p, std::size_t bytes, std::size_t alignment);
+};
+
+
+/** @brief Dynamic host memory resource using posix_memalign() for aligned
+    allocation and free() for deallocation. */
+class AlignedMemoryResource : public MemoryResource
+{
+protected:
+   virtual void *DoAllocate(std::size_t bytes, std::size_t alignment);
+   virtual void DoDeallocate(void *p, std::size_t bytes, std::size_t alignment);
+};
+
+} // namespace mfem
+
+#endif // MFEM_USE_BACKENDS
+
+#endif // MFEM_BACKENDS_BASE_MEMORY_RESOURCE_HPP
@@ -0,0 +1,233 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_BASE_SMART_POINTERS_HPP
+#define MFEM_BACKENDS_BASE_SMART_POINTERS_HPP
+
+#include "../../config/config.hpp"
+#ifdef MFEM_USE_BACKENDS
+
+#include "utils.hpp"
+#include "../../general/error.hpp"
+#include <cstddef>
+
+// #define MFEM_TRACE_SHARED_PTR
+#ifdef MFEM_TRACE_SHARED_PTR
+#include "../../general/globals.hpp"
+#endif
+
+namespace mfem
+{
+
+/// Base class for classes with simple reference counting.
+/** Reference counting is performed by the class SharedPtr. */
+class RefCounted
+{
+private:
+   mutable unsigned ref_count;
+
+   /// Only class SharedPtr can access ref_count.
+   template <typename T> friend class SharedPtr;
+
+public:
+   RefCounted() : ref_count(0) { }
+
+   /** @brief Prevent SharedPtr objects from deleting this object by
+       incrementing the reference counter by one. */
+   void DontDelete() const { ++ref_count; }
+};
+
+
+/** @brief Smart pointer class that manages objects of type T derived from class
+    RefCounted. */
+/** This class is generally meant to work with dynamically allocated object,
+    specifically objects allocated with operator new(). It will invoke operator
+    delete() to destroy the managed object when its reference counter reaches
+    zero. This behavior can be overriden by calling RefCounted::DontDelete() to
+    ensure that an object will not be deleted by a SharedPtr that holds a
+    pointer to it.
+    @note This class is NOT thread-safe and does not support circular ownership.
+ */
+template <typename T>
+class SharedPtr
+{
+public:
+   typedef T stored_type;
+
+private:
+   T *ptr;
+
+   void Init(T *new_ptr)
+   {
+      ptr = new_ptr;
+      if (ptr) { ++ptr->RefCounted::ref_count; }
+#ifdef MFEM_TRACE_SHARED_PTR
+#elif 0
+      mfem::out << "  [" << _MFEM_FUNC_NAME << "]: ptr = " << ptr;
+      if (ptr)
+      {
+         mfem::out << ", new ref_count = " << ptr->RefCounted::ref_count;
+      }
+      mfem::out << '\n';
+#endif
+   }
+   void Destroy()
+   {
+      MFEM_ASSERT(!ptr || ptr->RefCounted::ref_count >= 1, "invalid use");
+      if (ptr && --ptr->RefCounted::ref_count == 0) { delete ptr; }
+#ifdef MFEM_TRACE_SHARED_PTR
+#elif 0
+      mfem::out << "  [" << _MFEM_FUNC_NAME << "]: ptr = " << ptr;
+      if (ptr)
+      {
+         mfem::out << ", new ref_count = " << ptr->RefCounted::ref_count;
+      }
+      mfem::out << '\n';
+#endif
+   }
+
+public:
+   SharedPtr() : ptr(NULL)
+   {
+#ifdef MFEM_TRACE_SHARED_PTR
+      mfem::out << '[' << _MFEM_FUNC_NAME << "]: ptr = " << ptr << '\n';
+#endif
+   }
+
+   SharedPtr(const SharedPtr &other)
+   {
+#ifdef MFEM_TRACE_SHARED_PTR
+      mfem::out << '[' << _MFEM_FUNC_NAME << "]\n";
+#endif
+      Init(other.ptr);
+   }
+
+   template <typename U>
+   SharedPtr(const SharedPtr<U> &other)
+   {
+#ifdef MFEM_TRACE_SHARED_PTR
+      mfem::out << '[' << _MFEM_FUNC_NAME << "]\n";
+#endif
+      Init(other.Get());
+   }
+
+   explicit SharedPtr(T *p)
+   {
+#ifdef MFEM_TRACE_SHARED_PTR
+      mfem::out << '[' << _MFEM_FUNC_NAME << "]\n";
+#endif
+      Init(p);
+   }
+
+   ~SharedPtr()
+   {
+#ifdef MFEM_TRACE_SHARED_PTR
+      mfem::out << '[' << _MFEM_FUNC_NAME << "]\n";
+#endif
+      Destroy();
+   }
+
+   SharedPtr &operator=(const SharedPtr &other)
+   {
+#ifdef MFEM_TRACE_SHARED_PTR
+      mfem::out << '[' << _MFEM_FUNC_NAME << "]\n";
+#endif
+      Reset(other.ptr); return *this;
+   }
+
+   template <typename U>
+   SharedPtr &operator=(const SharedPtr<U> &other)
+   {
+#ifdef MFEM_TRACE_SHARED_PTR
+      mfem::out << '[' << _MFEM_FUNC_NAME << "]\n";
+#endif
+      Reset(other.Get()); return *this;
+   }
+
+   T &operator*() const { return *ptr; }
+   T *operator->() const { return ptr; }
+
+   operator bool() const { return ptr; }
+   bool operator!() const { return !ptr; }
+
+   template <typename U>
+   bool operator==(const SharedPtr<U> &other) const
+   { return ptr == other.Ptr(); }
+   template <typename U>
+   bool operator!=(const SharedPtr<U> &other) const
+   { return ptr != other.Ptr(); }
+
+   template <typename U>
+   bool operator==(const U &p) const { return ptr == (void*) p; }
+   template <typename U>
+   bool operator!=(const U &p) const { return ptr != (void*) p; }
+
+   T *Get() const { return ptr; }
+
+   /// TODO
+   template <typename derived_t>
+   derived_t *As() const { return util::As<derived_t>(ptr); }
+
+   unsigned UseCount() const { return ptr ? ptr->RefCounted::ref_count : 0; }
+
+   void Reset()
+   {
+#ifdef MFEM_TRACE_SHARED_PTR
+      mfem::out << '[' << _MFEM_FUNC_NAME << "]\n";
+#endif
+      Destroy();
+      ptr = NULL;
+   }
+
+   /// The type U* needs to be implicitly convertible to T*
+   template <typename U>
+   void Reset(U *new_ptr)
+   {
+#ifdef MFEM_TRACE_SHARED_PTR
+      mfem::out << '[' << _MFEM_FUNC_NAME << "]\n";
+#endif
+      if (ptr != new_ptr) { Destroy(); Init(new_ptr); }
+   }
+
+   void Swap(SharedPtr &other)
+   {
+#ifdef MFEM_TRACE_SHARED_PTR
+      mfem::out << '[' << _MFEM_FUNC_NAME << "]\n";
+#endif
+      std::swap(ptr, other.ptr);
+   }
+};
+
+
+template <class T>
+inline void Swap(SharedPtr<T> &a, SharedPtr<T> &b) { a.Swap(b); }
+
+
+class PLayout;
+typedef SharedPtr<PLayout> DLayout;
+
+class PArray;
+typedef SharedPtr<PArray> DArray;
+
+class PVector;
+typedef SharedPtr<PVector> DVector;
+
+class PFiniteElementSpace;
+typedef SharedPtr<PFiniteElementSpace> DFiniteElementSpace;
+
+class PBilinearForm;
+typedef SharedPtr<PBilinearForm> DBilinearForm;
+
+} // namespace mfem
+
+#endif // MFEM_USE_BACKENDS
+
+#endif // MFEM_BACKENDS_BASE_SMART_POINTERS_HPP
@@ -0,0 +1,52 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_BASE_UTILS_HPP
+#define MFEM_BACKENDS_BASE_UTILS_HPP
+
+#include "../../config/config.hpp"
+#ifdef MFEM_USE_BACKENDS
+
+#include "../../general/error.hpp"
+
+namespace mfem
+{
+
+namespace util
+{
+
+//
+// Inline methods
+//
+
+/// TODO: doxygen
+template <typename derived_t, typename base_t>
+inline derived_t *As(base_t *base_obj)
+{
+   MFEM_ASSERT(dynamic_cast<derived_t*>(base_obj) != NULL,
+               "invalid object type");
+   return static_cast<derived_t*>(base_obj);
+}
+
+/// TODO: doxygen
+template <typename derived_t, typename base_t>
+inline derived_t *Is(base_t *base_obj)
+{
+   return dynamic_cast<derived_t*>(base_obj);
+}
+
+} // namespace mfem::util
+
+} // namespace mfem
+
+#endif // MFEM_USE_BACKENDS
+
+#endif // MFEM_BACKENDS_BASE_UTILS_HPP
@@ -0,0 +1,153 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_BASE_VECTOR_HPP
+#define MFEM_BACKENDS_BASE_VECTOR_HPP
+
+#include "../../config/config.hpp"
+#ifdef MFEM_USE_BACKENDS
+
+#include "../../general/scalars.hpp"
+#include "array.hpp"
+
+namespace mfem
+{
+
+/// Polymorphic vector - array of scalars.
+class PVector : virtual public PArray
+{
+protected:
+   /**
+       @name Virtual interface
+    */
+   ///@{
+
+   /** @brief Create and return a new vector of the same dynamic type as this
+       vector using the same layout with entries specified by @a buffer_type_id
+       which should be a constant defined by the `value` field in a
+       specialization of the template class mfem::ScalarId.
+
+       Returns NULL if allocation fails.
+
+       If @a copy_data is true, the contents of this vector is copied to the new
+       vector; otherwise, the new vector remains uninitialized.
+
+       If @a buffer is not NULL, return the vector data of the newly created
+       object (in @a *buffer), if it is stored as a contiguous array on the
+       host; otherwise, set @a *buffer to NULL. */
+   virtual PVector *DoVectorClone(bool copy_data, void **buffer,
+                                  int buffer_type_id) const = 0;
+
+   /** @brief Compute and return the dot product of @a *this and @a x. In the
+       case of an MPI-parallel vector, the result must be the MPI-global dot
+       product. */
+   /** Both vectors must have the same dynamic type and layout. */
+   virtual void DoDotProduct(const PVector &x, void *result,
+                             int result_type_id) const = 0;
+
+   // TODO: add reduction operations: min, max, sum
+
+   /// Perform the operation @a *this = @a a @a x + @a b @a y.
+   /** Rules:
+       - the dynamic type of both @a x and @a y is the same as that of @a *this
+       - if @a a == 0, neither @a x nor its data are accessed
+       - if @a b == 0, neither @a y nor its data are accessed
+       - @a x's data is never the same as @a y's data, unless @a a == 0, or
+         @a b == 0
+       - @a x's data or @a y's data may be the same as the data of @a *this
+       - all accessed vectors, @a x, @a y, and @a *this have the same layout. */
+   virtual void DoAxpby(const void *a, const PVector &x,
+                        const void *b, const PVector &y,
+                        int ab_type_id) = 0;
+
+   ///@}
+   // End: Virtual interface
+
+public:
+   /** @brief Create a PVector. */
+   /** The @a layout must be valid in the sense that layout != NULL and
+       layout->HasEngine() == true. */
+   PVector(PLayout &p_layout)
+      : PArray(p_layout) { }
+
+   template <typename derived_t>
+   derived_t &As() { return *util::As<derived_t>(this); }
+
+   template <typename derived_t>
+   const derived_t &As() const { return *util::As<const derived_t>(this); }
+
+
+   // TODO: Error handling ... handle errors at the Engine level, at the class
+   //       level, or at the method level?
+
+   // TODO: Asynchronous execution interface ...
+
+   // TODO: Multi-vector interface ...
+
+
+   /**
+       @name Public virtual interface
+    */
+   ///@{
+
+   /** @brief Create and return a new vector of the same dynamic type as this
+       vector using the same layout with entries of type @a scalar_t.
+
+       If @a copy_data is true, the contents of this vector is copied to the new
+       vector; otherwise, the new vector remains uninitialized.
+
+       If @a buffer is not NULL, return the vector data of the newly created
+       object (in @a *buffer) , if it is stored as a contiguous array on the
+       host; otherwise, set @a *buffer to NULL. */
+   template <typename scalar_t>
+   DVector Clone(bool copy_data, scalar_t **buffer) const
+   {
+      return DVector(DoVectorClone(copy_data, (void**)buffer,
+                                   ScalarId<scalar_t>::value));
+   }
+
+   /** @brief Compute and return the dot product of @a *this and @a x. In the
+       case of an MPI-parallel vector, the result must be the MPI-global dot
+       product. */
+   /** Both vectors must have the same dynamic type and layout. */
+   template <typename scalar_t>
+   scalar_t DotProduct(const PVector &x) const
+   {
+      scalar_t result;
+      DoDotProduct(x, &result, ScalarId<scalar_t>::value);
+      return result;
+   }
+
+   // TODO: add reduction operations: min, max, sum
+
+   /// Perform the operation @a *this = @a a @a x + @a b @a y.
+   /** Rules:
+       - the dynamic type of both @a x and @a y is the same as that of @a *this
+       - if @a a == 0, neither @a x nor its data are accessed
+       - if @a b == 0, neither @a y nor its data are accessed
+       - @a x's data is never the same as @a y's data, unless @a a == 0, or
+         @a b == 0
+       - @a x's data or @a y's data may be the same as the data of @a *this
+       - all accessed vectors, @a x, @a y, and @a *this have the same layout. */
+   template <typename scalar_t>
+   void Axpby(const scalar_t &a, const PVector &x,
+              const scalar_t &b, const PVector &y)
+   { if (Size()) { DoAxpby(&a, x, &b, y, ScalarId<scalar_t>::value); } }
+
+   ///@}
+   // End: Virtual interface
+};
+
+} // namespace mfem
+
+#endif // MFEM_USE_BACKENDS
+
+#endif // MFEM_BACKENDS_BASE_VECTOR_HPP
@@ -0,0 +1,66 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+/*
+  ---[ Defines Known At Compile-Time ]------------
+  ELEMENT_BATCH   : How many elements are in each
+  .                   computation batch
+  NUM_DOFS_1D     : Dofs in the 1D segments
+  NUM_DOFS_2D     : Dofs in the 2D faces
+  NUM_DOFS_3D     : Dofs in the 3D domain
+  NUM_QUAD_1D     : Dofs in the 1D segments
+  NUM_QUAD_2D     : Dofs in the 2D faces
+  NUM_QUAD_3D     : Dofs in the 3D domain
+  NUM_MAX_1D      : max(NUM_QUAD_1D, NUM_DOFS_1D)
+  NUM_QUAD_DOFS_1D: NUM_QUAD_1D * NUM_DOFS_1D
+  COEFF_ARGS      : Code that passes required arguments to the kernel
+  COEFF           : Code that computes the coefficient
+  ================================================
+
+  [MISSING]
+  - Add support to auto-pick @dim and use @idxOrder on stack arrays
+  |   double a[2][2];
+  |   a[0][1]; <-- regular index
+  |   a(0,1);  <-- uses @idxOrder a[0][1] or a[1][0]
+  - Add support for @idxOrder to change indexing order after allocation
+  |   double a[2][2] @idxOrder(0,1);
+  |   a(0,1) -> a[1][0]
+  |   @set(a, idxOrder(1,0));
+  |   a(0,1) -> a[0][1]
+  - Add support to iterate over loop depending on mode
+  |   for(i; @inner) {
+  |     for(0 < j < N) {} <-- ++j or j += block?
+  |   }
+*/
+
+#include "mfem-occa://defines.okl"
+
+#if USING_TENSOR_OPS
+#  ifdef OCCA_USING_GPU
+#    if USING_LOW_ORDER
+#      include "mfem-occa://diffusion/tensor/gpuHighOrder.okl"
+#    else
+#      include "mfem-occa://diffusion/tensor/gpuHighOrder.okl"
+#    endif
+#  else
+#    include "mfem-occa://diffusion/tensor/cpu.okl"
+#  endif
+#else
+#  ifdef OCCA_USING_GPU
+#    if USING_LOW_ORDER
+#      include "mfem-occa://diffusion/simplex/gpuHighOrder.okl"
+#    else
+#      include "mfem-occa://diffusion/simplex/gpuHighOrder.okl"
+#    endif
+#  else
+#    include "mfem-occa://diffusion/simplex/cpu.okl"
+#  endif
+#endif
@@ -0,0 +1,123 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "array.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+PArray *Array::DoClone(bool copy_data, void **buffer,
+                       std::size_t item_size) const
+{
+   Array *new_array = new Array(OccaLayout(), item_size);
+   if (copy_data)
+   {
+      new_array->slice.copyFrom(slice);
+   }
+   if (buffer)
+   {
+      *buffer = new_array->GetBuffer();
+   }
+   return new_array;
+}
+
+int Array::DoResize(PLayout &new_layout, void **buffer,
+                    std::size_t item_size)
+{
+   MFEM_ASSERT(dynamic_cast<Layout *>(&new_layout) != NULL,
+               "new_layout is not an OCCA Layout");
+   Layout *lt = static_cast<Layout *>(&new_layout);
+   layout.Reset(lt); // Reset() checks if the pointer is the same
+   int err = ResizeData(lt, item_size);
+   if (!err && buffer)
+   {
+      *buffer = GetBuffer();
+   }
+   return err;
+}
+
+void *Array::DoPullData(void *buffer, std::size_t item_size)
+{
+   // called only when Size() != 0
+
+   if (!slice.getDevice().hasSeparateMemorySpace())
+   {
+      return slice.ptr();
+   }
+   if (buffer)
+   {
+      slice.copyTo(buffer);
+   }
+   return buffer;
+}
+
+void Array::DoFill(const void *value_ptr, std::size_t item_size)
+{
+   // called only when Size() != 0
+
+   switch (item_size)
+   {
+      case sizeof(int8_t):
+         OccaFill((const int8_t *)value_ptr);
+         break;
+      case sizeof(int16_t):
+         OccaFill((const int16_t *)value_ptr);
+         break;
+      case sizeof(int32_t):
+         OccaFill((const int32_t *)value_ptr);
+         break;
+      // case sizeof(int64_t):
+      //    OccaFill((const int64_t *)value_ptr);
+      //    break;
+      case sizeof(double):
+         OccaFill((const double *)value_ptr);
+         break;
+      // case sizeof(::occa::double2):
+      //    OccaFill((const ::occa::double2 *)value_ptr);
+      //    break;
+      default:
+         MFEM_ABORT("item_size = " << item_size << " is not supported");
+   }
+}
+
+void Array::DoPushData(const void *src_buffer, std::size_t item_size)
+{
+   // called only when Size() != 0
+
+   if (slice.getDevice().hasSeparateMemorySpace() || slice.ptr() != src_buffer)
+   {
+      slice.copyFrom(src_buffer);
+   }
+}
+
+void Array::DoAssign(const PArray &src, std::size_t item_size)
+{
+   // called only when Size() != 0
+
+   // Note: static_cast can not be used here since PArray is a virtual base
+   //       class.
+   const Array *source = dynamic_cast<const Array *>(&src);
+   MFEM_ASSERT(source != NULL, "invalid source Array type");
+   MFEM_ASSERT(Size() == source->Size(), "");
+   slice.copyFrom(source->slice);
+}
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,133 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_ARRAY_HPP
+#define MFEM_BACKENDS_OCCA_ARRAY_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include <occa.hpp>
+#include "layout.hpp"
+#include "../base/array.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+class Array : public virtual PArray
+{
+protected:
+   //
+   // Inherited fields
+   //
+   // DLayout layout;
+
+   // Always true: Size()*item_size == slice.size() <= data.size()
+   mutable ::occa::memory data, slice;
+
+   //
+   // Virtual interface
+   //
+
+   virtual void *DoGetData() const { return GetBuffer(); }
+
+   virtual PArray *DoClone(bool copy_data, void **buffer,
+                           std::size_t item_size) const;
+
+   virtual int DoResize(PLayout &new_layout, void **buffer,
+                        std::size_t item_size);
+
+   virtual void *DoPullData(void *buffer, std::size_t item_size);
+
+   virtual void DoFill(const void *value_ptr, std::size_t item_size);
+
+   virtual void DoPushData(const void *src_buffer, std::size_t item_size);
+
+   virtual void DoAssign(const PArray &src, std::size_t item_size);
+
+   //
+   // Auxiliary methods
+   //
+
+   inline void *GetBuffer() const;
+
+   inline int ResizeData(const Layout *lt, std::size_t item_size);
+
+   template <typename T>
+   inline void OccaFill(const T *val_ptr)
+   { ::occa::linalg::operator_eq<T>(slice, *val_ptr); }
+
+public:
+   Array(Layout &lt, std::size_t item_size)
+      : PArray(lt),
+        data(lt.Alloc(lt.Size()*item_size)),
+        slice(data)
+   { }
+
+   virtual ~Array() { }
+
+   inline void MakeRef(Array &master);
+
+   Layout &OccaLayout() const
+   { return *static_cast<Layout *>(layout.Get()); }
+
+   ::occa::memory &OccaMem() { return slice; }
+   const ::occa::memory &OccaMem() const { return slice; }
+};
+
+
+//
+// Inline methods
+//
+
+inline void *Array::GetBuffer() const
+{
+   if (!slice.getDevice().hasSeparateMemorySpace())
+   {
+      return slice.ptr();
+   }
+   return NULL;
+}
+
+inline int Array::ResizeData(const Layout *lt, std::size_t item_size)
+{
+   const std::size_t new_bytes = lt->Size()*item_size;
+   if (data.size() < new_bytes ||
+       data.getDHandle() != lt->OccaEngine().GetDevice().getDHandle())
+   {
+      data = lt->Alloc(new_bytes);
+      slice = data;
+      // If memory allocation fails - an exception is thrown.
+   }
+   else if (slice.size() != new_bytes)
+   {
+      slice = data.slice(0, new_bytes);
+   }
+   return 0;
+}
+
+inline void Array::MakeRef(Array &master)
+{
+   layout = master.layout;
+   data = master.data;
+   slice = master.slice;
+}
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_ARRAY_HPP
@@ -0,0 +1,47 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "backend.hpp"
+#include "engine.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+bool Backend::Supports(const std::string &engine_spec) const
+{
+   // TODO: check if 'engine_spec' is valid OCCA string.
+   return true;
+}
+
+mfem::Engine *Create(const std::string &engine_spec)
+{
+   return new Engine(engine_spec);
+}
+
+#ifdef MFEM_USE_MPI
+mfem::Engine *Create(MPI_Comm comm, const std::string &engine_spec)
+{
+   return new Engine(comm, engine_spec);
+}
+#endif
+
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,49 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_BACKEND_HPP
+#define MFEM_BACKENDS_OCCA_BACKEND_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+// Only the Backend and Engine classes should be exposed through "backend.hpp"
+#include "../base/backend.hpp"
+#include "engine.hpp"
+#include <occa.hpp>
+
+namespace mfem
+{
+
+namespace occa
+{
+
+class Backend : public mfem::Backend
+{
+public:
+   virtual ~Backend();
+
+   virtual bool Supports(const std::string &engine_spec) const;
+
+   virtual mfem::Engine *Create(const std::string &engine_spec);
+
+#ifdef MFEM_USE_MPI
+   virtual mfem::Engine *Create(MPI_Comm comm, const std::string &engine_spec);
+#endif
+};
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_BACKEND_HPP
@@ -0,0 +1,514 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "backend.hpp"
+#include "bilininteg.hpp"
+#include "../../fem/bilinearform.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+OccaBilinearForm::OccaBilinearForm(FiniteElementSpace *ofespace_) :
+   Operator(ofespace_->OccaVLayout()),
+   localX((ofespace_->OccaEVLayout().DontDelete(), ofespace_->OccaEVLayout())),
+   localY((ofespace_->OccaEVLayout().DontDelete(), ofespace_->OccaEVLayout()))
+{
+   Init(ofespace_->OccaEngine(), ofespace_, ofespace_);
+}
+
+OccaBilinearForm::OccaBilinearForm(FiniteElementSpace *otrialFESpace_,
+                                   FiniteElementSpace *otestFESpace_) :
+   Operator(otrialFESpace_->OccaVLayout(),
+            otestFESpace_->OccaVLayout()),
+   localX((otrialFESpace_->OccaEVLayout().DontDelete(), otrialFESpace_->OccaEVLayout())),
+   localY((otestFESpace_->OccaEVLayout().DontDelete(), otestFESpace_->OccaEVLayout()))
+{
+   Init(otrialFESpace_->OccaEngine(), otrialFESpace_, otestFESpace_);
+}
+
+void OccaBilinearForm::Init(const Engine &e,
+                            FiniteElementSpace *otrialFESpace_,
+                            FiniteElementSpace *otestFESpace_)
+{
+   engine.Reset(&e);
+
+   otrialFESpace = otrialFESpace_;
+   trialFESpace  = otrialFESpace_->GetFESpace();
+
+   otestFESpace = otestFESpace_;
+   testFESpace  = otestFESpace_->GetFESpace();
+
+   mesh = trialFESpace->GetMesh();
+
+   const int elements = GetNE();
+
+   const int trialVDim = trialFESpace->GetVDim();
+
+   const int trialLocalDofs = otrialFESpace->GetLocalDofs();
+   const int testLocalDofs  = otestFESpace->GetLocalDofs();
+
+   // First-touch policy when running with OpenMP
+   if (GetDevice().mode() == "OpenMP")
+   {
+      const std::string &okl_path = OccaEngine().GetOklPath();
+      const std::string &okl_defines = OccaEngine().GetOklDefines();
+      ::occa::kernel initLocalKernel =
+         GetDevice().buildKernel(okl_path + "utils.okl",
+                                 "InitLocalVector",
+                                 okl_defines);
+
+      const std::size_t sd = sizeof(double);
+      const uint64_t trialEntries = sd * (elements * trialLocalDofs);
+      const uint64_t testEntries  = sd * (elements * testLocalDofs);
+      for (int v = 0; v < trialVDim; ++v)
+      {
+         const uint64_t trialOffset = v * trialEntries;
+         const uint64_t testOffset  = v * testEntries;
+
+         initLocalKernel(elements, trialLocalDofs,
+                         localX.OccaMem().slice(trialOffset, trialEntries));
+         initLocalKernel(elements, testLocalDofs,
+                         localY.OccaMem().slice(testOffset, testEntries));
+      }
+   }
+}
+
+int OccaBilinearForm::BaseGeom() const
+{
+   return mesh->GetElementBaseGeometry();
+}
+
+int OccaBilinearForm::GetDim() const
+{
+   return mesh->Dimension();
+}
+
+int64_t OccaBilinearForm::GetNE() const
+{
+   return mesh->GetNE();
+}
+
+Mesh& OccaBilinearForm::GetMesh() const
+{
+   return *mesh;
+}
+
+FiniteElementSpace& OccaBilinearForm::GetTrialOccaFESpace() const
+{
+   return *otrialFESpace;
+}
+
+FiniteElementSpace& OccaBilinearForm::GetTestOccaFESpace() const
+{
+   return *otestFESpace;
+}
+
+mfem::FiniteElementSpace& OccaBilinearForm::GetTrialFESpace() const
+{
+   return *trialFESpace;
+}
+
+mfem::FiniteElementSpace& OccaBilinearForm::GetTestFESpace() const
+{
+   return *testFESpace;
+}
+
+int64_t OccaBilinearForm::GetTrialNDofs() const
+{
+   return trialFESpace->GetNDofs();
+}
+
+int64_t OccaBilinearForm::GetTestNDofs() const
+{
+   return testFESpace->GetNDofs();
+}
+
+int64_t OccaBilinearForm::GetTrialVDim() const
+{
+   return trialFESpace->GetVDim();
+}
+
+int64_t OccaBilinearForm::GetTestVDim() const
+{
+   return testFESpace->GetVDim();
+}
+
+const FiniteElement& OccaBilinearForm::GetTrialFE(const int i) const
+{
+   return *(trialFESpace->GetFE(i));
+}
+
+const FiniteElement& OccaBilinearForm::GetTestFE(const int i) const
+{
+   return *(testFESpace->GetFE(i));
+}
+
+// Adds new Domain Integrator.
+void OccaBilinearForm::AddDomainIntegrator(OccaIntegrator *integrator,
+                                           const ::occa::properties &props)
+{
+   AddIntegrator(integrator, props, DomainIntegrator);
+}
+
+// Adds new Boundary Integrator.
+void OccaBilinearForm::AddBoundaryIntegrator(OccaIntegrator *integrator,
+                                             const ::occa::properties &props)
+{
+   AddIntegrator(integrator, props, BoundaryIntegrator);
+}
+
+// Adds new interior Face Integrator.
+void OccaBilinearForm::AddInteriorFaceIntegrator(OccaIntegrator *integrator,
+                                                 const ::occa::properties &props)
+{
+   AddIntegrator(integrator, props, InteriorFaceIntegrator);
+}
+
+// Adds new boundary Face Integrator.
+void OccaBilinearForm::AddBoundaryFaceIntegrator(OccaIntegrator *integrator,
+                                                 const ::occa::properties &props)
+{
+   AddIntegrator(integrator, props, BoundaryFaceIntegrator);
+}
+
+// Adds Integrator based on OccaIntegratorType
+void OccaBilinearForm::AddIntegrator(OccaIntegrator *integrator,
+                                     const ::occa::properties &props,
+                                     const OccaIntegratorType itype)
+{
+   if (integrator == NULL)
+   {
+      std::stringstream error_ss;
+      error_ss << "OccaBilinearForm::";
+      switch (itype)
+      {
+         case DomainIntegrator      : error_ss << "AddDomainIntegrator";       break;
+         case BoundaryIntegrator    : error_ss << "AddBoundaryIntegrator";     break;
+         case InteriorFaceIntegrator: error_ss << "AddInteriorFaceIntegrator"; break;
+         case BoundaryFaceIntegrator: error_ss << "AddBoundaryFaceIntegrator"; break;
+      }
+      error_ss << " (...):\n"
+               << "  Integrator is NULL";
+      const std::string error = error_ss.str();
+      mfem_error(error.c_str());
+   }
+   integrator->SetupIntegrator(*this, baseKernelProps + props, itype);
+   integrators.push_back(integrator);
+}
+
+const mfem::Operator* OccaBilinearForm::GetTrialProlongation() const
+{
+   return otrialFESpace->GetProlongationOperator();
+}
+
+const mfem::Operator* OccaBilinearForm::GetTestProlongation() const
+{
+   return otestFESpace->GetProlongationOperator();
+}
+
+const mfem::Operator* OccaBilinearForm::GetTrialRestriction() const
+{
+   return otrialFESpace->GetRestrictionOperator();
+}
+
+const mfem::Operator* OccaBilinearForm::GetTestRestriction() const
+{
+   return otestFESpace->GetRestrictionOperator();
+}
+
+void OccaBilinearForm::Assemble()
+{
+   // [MISSING] Find geometric information that is needed by intergrators
+   //             to share between integrators.
+   const int integratorCount = (int) integrators.size();
+   for (int i = 0; i < integratorCount; ++i)
+   {
+      integrators[i]->Assemble();
+   }
+}
+
+void OccaBilinearForm::FormLinearSystem(const mfem::Array<int> &constraintList,
+                                        mfem::Vector &x, mfem::Vector &b,
+                                        mfem::Operator *&Aout,
+                                        mfem::Vector &X, mfem::Vector &B,
+                                        int copy_interior)
+{
+   FormOperator(constraintList, Aout);
+   InitRHS(constraintList, x, b, Aout, X, B, copy_interior);
+}
+
+void OccaBilinearForm::FormOperator(const mfem::Array<int> &constraintList,
+                                    mfem::Operator *&Aout)
+{
+   const mfem::Operator *trialP = GetTrialProlongation();
+   const mfem::Operator *testP  = GetTestProlongation();
+   mfem::Operator *rap = this;
+
+   if (trialP)
+   {
+      rap = new RAPOperator(*testP, *this, *trialP);
+   }
+
+   Aout = new OccaConstrainedOperator(rap, constraintList,
+                                      rap != this);
+}
+
+void OccaBilinearForm::InitRHS(const mfem::Array<int> &constraintList,
+                               mfem::Vector &x, mfem::Vector &b,
+                               mfem::Operator *A,
+                               mfem::Vector &X, mfem::Vector &B,
+                               int copy_interior)
+{
+   const std::string okl_defines = OccaEngine().GetOklDefines();
+
+   // FIXME: move these kernels to the Backend?
+   static ::occa::kernelBuilder get_subvector_builder =
+      ::occa::linalg::customLinearMethod(
+         "vector_get_subvector",
+
+         "const int dof_i = v2[i];"
+         "v0[i] = dof_i >= 0 ? v1[dof_i] : -v1[-dof_i - 1];",
+
+         "defines: {"
+         "  VTYPE0: 'double',"
+         "  VTYPE1: 'double',"
+         "  VTYPE2: 'int',"
+         "  TILESIZE: 128,"
+         "}" + okl_defines);
+
+   static ::occa::kernelBuilder set_subvector_builder =
+      ::occa::linalg::customLinearMethod(
+         "vector_set_subvector",
+         "const int dof_i = v2[i];"
+         "if (dof_i >= 0) { v0[dof_i]      = v1[i]; }"
+         "else            { v0[-dof_i - 1] = -v1[i]; }",
+
+         "defines: {"
+         "  VTYPE0: 'double',"
+         "  VTYPE1: 'double',"
+         "  VTYPE2: 'int',"
+         "  TILESIZE: 128,"
+         "}" + okl_defines);
+
+   const mfem::Operator *P = GetTrialProlongation();
+   const mfem::Operator *R = GetTrialRestriction();
+
+   if (P)
+   {
+      // Variational restriction with P
+      B.Resize(P->InLayout());
+      P->MultTranspose(b, B);
+      X.Resize(R->OutLayout());
+      R->Mult(x, X);
+   }
+   else
+   {
+      // rap, X and B point to the same data as this, x and b
+      X.MakeRef(x);
+      B.MakeRef(b);
+   }
+
+   if (!copy_interior && constraintList.Size() > 0)
+   {
+      ::occa::kernel get_subvector_kernel =
+         get_subvector_builder.build(GetDevice());
+      ::occa::kernel set_subvector_kernel =
+         set_subvector_builder.build(GetDevice());
+
+      const Array &constrList = constraintList.Get_PArray()->As<Array>();
+      Vector subvec(constrList.OccaLayout());
+
+      get_subvector_kernel(constraintList.Size(),
+                           subvec.OccaMem(),
+                           X.Get_PVector()->As<Vector>().OccaMem(),
+                           constrList.OccaMem());
+
+      X.Fill(0.0);
+
+      set_subvector_kernel(constraintList.Size(),
+                           X.Get_PVector()->As<Vector>().OccaMem(),
+                           subvec.OccaMem(),
+                           constrList.OccaMem());
+   }
+
+   OccaConstrainedOperator *cA = dynamic_cast<OccaConstrainedOperator*>(A);
+   if (cA)
+   {
+      cA->EliminateRHS(X.Get_PVector()->As<Vector>(),
+                       B.Get_PVector()->As<Vector>());
+   }
+   else
+   {
+      mfem_error("OccaBilinearForm::InitRHS expects an OccaConstrainedOperator");
+   }
+}
+
+// Matrix vector multiplication.
+void OccaBilinearForm::Mult_(const Vector &x, Vector &y) const
+{
+   otrialFESpace->GlobalToLocal(x, localX);
+   localY.Fill<double>(0.0);
+
+   const int integratorCount = (int) integrators.size();
+   for (int i = 0; i < integratorCount; ++i)
+   {
+      integrators[i]->MultAdd(localX, localY);
+   }
+
+   otestFESpace->LocalToGlobal(localY, y);
+}
+
+// Matrix transpose vector multiplication.
+void OccaBilinearForm::MultTranspose_(const Vector &x, Vector &y) const
+{
+   otestFESpace->GlobalToLocal(x, localX);
+   localY.Fill<double>(0.0);
+
+   const int integratorCount = (int) integrators.size();
+   for (int i = 0; i < integratorCount; ++i)
+   {
+      integrators[i]->MultTransposeAdd(localX, localY);
+   }
+
+   otrialFESpace->LocalToGlobal(localY, y);
+}
+
+void OccaBilinearForm::OccaRecoverFEMSolution(const mfem::Vector &X,
+                                              const mfem::Vector &b,
+                                              mfem::Vector &x)
+{
+   const mfem::Operator *P = this->GetTrialProlongation();
+   if (P)
+   {
+      // Apply conforming prolongation
+      x.Resize(P->OutLayout());
+      P->Mult(X, x);
+   }
+   // Otherwise X and x point to the same data
+}
+
+// Frees memory bilinear form.
+OccaBilinearForm::~OccaBilinearForm()
+{
+   // Make sure all integrators free their data
+   IntegratorVector::iterator it = integrators.begin();
+   while (it != integrators.end())
+   {
+      delete *it;
+      ++it;
+   }
+}
+
+
+void BilinearForm::InitOccaBilinearForm()
+{
+   // Init 'obform' using 'bform'
+   MFEM_ASSERT(bform != NULL, "");
+   MFEM_ASSERT(obform == NULL, "");
+
+   FiniteElementSpace &ofes =
+      bform->FESpace()->Get_PFESpace()->As<FiniteElementSpace>();
+   obform = new OccaBilinearForm(&ofes);
+
+   // Transfer domain integrators
+   mfem::Array<mfem::BilinearFormIntegrator*> &dbfi = *bform->GetDBFI();
+   for (int i = 0; i < dbfi.Size(); i++)
+   {
+      std::string integ_name(dbfi[i]->Name());
+      Coefficient *scal_coeff = dbfi[i]->GetScalarCoefficient();
+      ConstantCoefficient *const_coeff =
+         dynamic_cast<ConstantCoefficient*>(scal_coeff);
+      // TODO: other types of coefficients ...
+      double val = const_coeff ? const_coeff->constant : 1.0;
+      OccaCoefficient ocoeff(obform->OccaEngine(), val);
+
+      OccaIntegrator *ointeg = NULL;
+
+      if (integ_name == "(undefined)")
+      {
+         MFEM_ABORT("BilinearFormIntegrator does not define Name()");
+      }
+      else if (integ_name == "diffusion")
+      {
+         ointeg = new OccaDiffusionIntegrator(ocoeff);
+      }
+      else
+      {
+         MFEM_ABORT("BilinearFormIntegrator [Name() = " << integ_name
+                    << "] is not supported");
+      }
+
+      const mfem::IntegrationRule *ir = dbfi[i]->GetIntRule();
+      if (ir) { ointeg->SetIntegrationRule(*ir); }
+
+      obform->AddDomainIntegrator(ointeg);
+   }
+
+   // TODO: other types of integrators ...
+}
+
+bool BilinearForm::Assemble()
+{
+   if (obform == NULL) { InitOccaBilinearForm(); }
+
+   obform->Assemble();
+
+   return true; // --> host assembly is not needed
+}
+
+void BilinearForm::FormSystemMatrix(const mfem::Array<int> &ess_tdof_list,
+                                    mfem::OperatorHandle &A)
+{
+   if (A.Type() == mfem::Operator::ANY_TYPE)
+   {
+      mfem::Operator *Aout = NULL;
+      obform->FormOperator(ess_tdof_list, Aout);
+      A.Reset(Aout);
+   }
+   else
+   {
+      MFEM_ABORT("Operator::Type is not supported, type = " << A.Type());
+   }
+}
+
+void BilinearForm::FormLinearSystem(const mfem::Array<int> &ess_tdof_list,
+                                    mfem::Vector &x, mfem::Vector &b,
+                                    mfem::OperatorHandle &A,
+                                    mfem::Vector &X, mfem::Vector &B,
+                                    int copy_interior)
+{
+   FormSystemMatrix(ess_tdof_list, A);
+   obform->InitRHS(ess_tdof_list, x, b, A.Ptr(), X, B, copy_interior);
+}
+
+void BilinearForm::RecoverFEMSolution(const mfem::Vector &X,
+                                      const mfem::Vector &b,
+                                      mfem::Vector &x)
+{
+   obform->OccaRecoverFEMSolution(X, b, x);
+}
+
+BilinearForm::~BilinearForm()
+{
+   delete obform;
+}
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,213 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_BILINEAR_FORM_HPP
+#define MFEM_BACKENDS_OCCA_BILINEAR_FORM_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "fespace.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+enum OccaIntegratorType
+{
+   DomainIntegrator       = 0,
+   BoundaryIntegrator     = 1,
+   InteriorFaceIntegrator = 2,
+   BoundaryFaceIntegrator = 3
+};
+
+class OccaIntegrator;
+
+
+/** Class for bilinear form - "Matrix" with associated FE space and
+    BLFIntegrators. */
+class OccaBilinearForm : public Operator
+{
+   friend class OccaIntegrator;
+
+protected:
+   typedef std::vector<OccaIntegrator*> IntegratorVector;
+
+   SharedPtr<const Engine> engine;
+
+   // State information
+   mutable mfem::Mesh *mesh;
+
+   mutable FiniteElementSpace *otrialFESpace;
+   mutable mfem::FiniteElementSpace *trialFESpace;
+
+   mutable FiniteElementSpace *otestFESpace;
+   mutable mfem::FiniteElementSpace *testFESpace;
+
+   IntegratorVector integrators;
+
+   // Device data
+   ::occa::properties baseKernelProps;
+
+   // The input and output vectors are mapped to local nodes for efficient
+   // operations. In other words, they are E-vectors.
+   // The size is: (number of elements) * (nodes in element) * (vector dim)
+   mutable Vector localX, localY;
+
+public:
+   OccaBilinearForm(FiniteElementSpace *ofespace_);
+
+   OccaBilinearForm(FiniteElementSpace *otrialFESpace_,
+                    FiniteElementSpace *otestFESpace_);
+
+   void Init(const Engine &e,
+             FiniteElementSpace *otrialFESpace_,
+             FiniteElementSpace *otestFESpace_);
+
+   const Engine &OccaEngine() const { return *engine; }
+
+   ::occa::device GetDevice(int idx = 0) const
+   { return engine->GetDevice(idx); }
+
+   // Useful mesh Information
+   int BaseGeom() const;
+   int GetDim() const;
+   int64_t GetNE() const;
+
+   mfem::Mesh& GetMesh() const;
+
+   FiniteElementSpace& GetTrialOccaFESpace() const;
+   FiniteElementSpace& GetTestOccaFESpace() const;
+
+   mfem::FiniteElementSpace& GetTrialFESpace() const;
+   mfem::FiniteElementSpace& GetTestFESpace() const;
+
+   // Useful FE information
+   int64_t GetTrialNDofs() const;
+   int64_t GetTestNDofs() const;
+
+   int64_t GetTrialVDim() const;
+   int64_t GetTestVDim() const;
+
+   const mfem::FiniteElement& GetTrialFE(const int i) const;
+   const mfem::FiniteElement& GetTestFE(const int i) const;
+
+   // Adds new Domain Integrator.
+   void AddDomainIntegrator(OccaIntegrator *integrator,
+                            const ::occa::properties &props =
+                               ::occa::properties());
+
+   // Adds new Boundary Integrator.
+   void AddBoundaryIntegrator(OccaIntegrator *integrator,
+                              const ::occa::properties &props =
+                                 ::occa::properties());
+
+   // Adds new interior Face Integrator.
+   void AddInteriorFaceIntegrator(OccaIntegrator *integrator,
+                                  const ::occa::properties &props =
+                                     ::occa::properties());
+
+   // Adds new boundary Face Integrator.
+   void AddBoundaryFaceIntegrator(OccaIntegrator *integrator,
+                                  const ::occa::properties &props =
+                                     ::occa::properties());
+
+   // Adds Integrator based on OccaIntegratorType
+   void AddIntegrator(OccaIntegrator *integrator,
+                      const ::occa::properties &props,
+                      const OccaIntegratorType itype);
+
+   virtual const mfem::Operator *GetTrialProlongation() const;
+   virtual const mfem::Operator *GetTestProlongation() const;
+
+   virtual const mfem::Operator *GetTrialRestriction() const;
+   virtual const mfem::Operator *GetTestRestriction() const;
+
+   // Assembles the form i.e. sums over all domain/bdr integrators.
+   virtual void Assemble();
+
+   void FormLinearSystem(const mfem::Array<int> &constraintList,
+                         mfem::Vector &x, mfem::Vector &b,
+                         mfem::Operator *&Aout,
+                         mfem::Vector &X, mfem::Vector &B,
+                         int copy_interior = 0);
+
+   void FormOperator(const mfem::Array<int> &constraintList,
+                     mfem::Operator *&Aout);
+
+   void InitRHS(const mfem::Array<int> &constraintList,
+                mfem::Vector &x, mfem::Vector &b,
+                mfem::Operator *Aout,
+                mfem::Vector &X, mfem::Vector &B,
+                int copy_interior = 0);
+
+   // overrides
+   virtual void Mult_(const Vector &x, Vector &y) const;
+   virtual void MultTranspose_(const Vector &x, Vector &y) const;
+
+   void OccaRecoverFEMSolution(const mfem::Vector &X, const mfem::Vector &b,
+                               mfem::Vector &x);
+
+   // Destroys bilinear form.
+   ~OccaBilinearForm();
+};
+
+
+/// TODO: doxygen
+class BilinearForm : public mfem::PBilinearForm
+{
+protected:
+   //
+   // Inherited fields
+   //
+   // SharedPtr<const mfem::Engine> engine;
+   // mfem::BilinearForm *bform;
+   OccaBilinearForm *obform;
+
+   // Called from Assemble() if obform is NULL to initialize obform.
+   void InitOccaBilinearForm();
+
+public:
+   /// TODO: doxygen
+   BilinearForm(const Engine &e, mfem::BilinearForm &bf)
+      : mfem::PBilinearForm(e, bf), obform(NULL) { }
+
+   /// Virtual destructor
+   virtual ~BilinearForm();
+
+   /// Assemble the PBilinearForm.
+   /** This method is called from the method mfem::BilinearForm::Assemble() of
+       the associated mfem::BilinearForm, #bform.
+       @returns True, if the host assembly should NOT be performed. */
+   virtual bool Assemble();
+
+   virtual void FormSystemMatrix(const mfem::Array<int> &ess_tdof_list,
+                                 mfem::OperatorHandle &A);
+
+   virtual void FormLinearSystem(const mfem::Array<int> &ess_tdof_list,
+                                 mfem::Vector &x, mfem::Vector &b,
+                                 mfem::OperatorHandle &A,
+                                 mfem::Vector &X, mfem::Vector &B,
+                                 int copy_interior);
+
+   virtual void RecoverFEMSolution(const mfem::Vector &X, const mfem::Vector &b,
+                                   mfem::Vector &x);
+};
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_BILINEAR_FORM_HPP
@@ -0,0 +1,956 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "bilininteg.hpp"
+#include "../../fem/fem.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+std::map<std::string, OccaDofQuadMaps> OccaDofQuadMaps::AllDofQuadMaps;
+
+OccaGeometry OccaGeometry::Get(::occa::device device,
+                               FiniteElementSpace &ofespace,
+                               const mfem::IntegrationRule &ir,
+                               const int flags)
+{
+   OccaGeometry geom;
+
+   mfem::Mesh &mesh = *(ofespace.GetMesh());
+   if (!mesh.GetNodes())
+   {
+      mesh.SetCurvature(1, false, -1, mfem::Ordering::byVDIM);
+   }
+   mfem::GridFunction &nodes = *(mesh.GetNodes());
+   const mfem::FiniteElementSpace &fespace = *(nodes.FESpace());
+   const mfem::FiniteElement &fe = *(fespace.GetFE(0));
+
+   const int dims     = fe.GetDim();
+   const int elements = fespace.GetNE();
+   const int numDofs  = fe.GetDof();
+   const int numQuad  = ir.GetNPoints();
+
+   MFEM_ASSERT(dims == mesh.SpaceDimension(), "");
+
+   geom.meshNodes.allocate(device,
+                           dims, numDofs, elements);
+
+   const mfem::Table &e2dTable = fespace.GetElementToDofTable();
+   const int *elementMap = e2dTable.GetJ();
+   nodes.Pull();
+   for (int e = 0; e < elements; ++e)
+   {
+      for (int dof = 0; dof < numDofs; ++dof)
+      {
+         const int gid = elementMap[dof + numDofs*e];
+         for (int dim = 0; dim < dims; ++dim)
+         {
+            geom.meshNodes(dim, dof, e) = nodes[fespace.DofToVDof(gid,dim)];
+         }
+      }
+   }
+   geom.meshNodes.keepInDevice();
+
+   if (flags & Jacobian)
+   {
+      geom.J.allocate(device,
+                      dims, dims, numQuad, elements);
+   }
+   else
+   {
+      geom.J.allocate(device, 1);
+   }
+   if (flags & JacobianInv)
+   {
+      geom.invJ.allocate(device,
+                         dims, dims, numQuad, elements);
+   }
+   else
+   {
+      geom.invJ.allocate(device, 1);
+   }
+   if (flags & JacobianDet)
+   {
+      geom.detJ.allocate(device,
+                         numQuad, elements);
+   }
+   else
+   {
+      geom.detJ.allocate(device, 1);
+   }
+
+   geom.J.stopManaging();
+   geom.invJ.stopManaging();
+   geom.detJ.stopManaging();
+
+   OccaDofQuadMaps &maps = OccaDofQuadMaps::GetSimplexMaps(device, fe, ir);
+
+   ::occa::properties props;
+   props["defines/NUM_DOFS"] = numDofs;
+   props["defines/NUM_QUAD"] = numQuad;
+   props["defines/STORE_JACOBIAN"]     = (flags & Jacobian);
+   props["defines/STORE_JACOBIAN_INV"] = (flags & JacobianInv);
+   props["defines/STORE_JACOBIAN_DET"] = (flags & JacobianDet);
+
+   const std::string &okl_path = ofespace.OccaEngine().GetOklPath();
+   const std::string &okl_defines = ofespace.OccaEngine().GetOklDefines();
+   ::occa::kernel init = device.buildKernel(okl_path + "geometry.okl",
+                                            stringWithDim("InitGeometryInfo",
+                                                          fe.GetDim()),
+                                            props + okl_defines);
+   init(elements,
+        maps.dofToQuadD,
+        geom.meshNodes,
+        geom.J, geom.invJ, geom.detJ);
+
+   return geom;
+}
+
+OccaDofQuadMaps::OccaDofQuadMaps() :
+   hash() {}
+
+OccaDofQuadMaps::OccaDofQuadMaps(const OccaDofQuadMaps &maps)
+{
+   *this = maps;
+}
+
+OccaDofQuadMaps& OccaDofQuadMaps::operator = (const OccaDofQuadMaps &maps)
+{
+   hash = maps.hash;
+   dofToQuad   = maps.dofToQuad;
+   dofToQuadD  = maps.dofToQuadD;
+   quadToDof   = maps.quadToDof;
+   quadToDofD  = maps.quadToDofD;
+   quadWeights = maps.quadWeights;
+   return *this;
+}
+
+OccaDofQuadMaps& OccaDofQuadMaps::Get(::occa::device device,
+                                      const FiniteElementSpace &fespace,
+                                      const mfem::IntegrationRule &ir,
+                                      const bool transpose)
+{
+   return Get(device,
+              *fespace.GetFE(0),
+              *fespace.GetFE(0),
+              ir,
+              transpose);
+}
+
+OccaDofQuadMaps& OccaDofQuadMaps::Get(::occa::device device,
+                                      const mfem::FiniteElement &fe,
+                                      const mfem::IntegrationRule &ir,
+                                      const bool transpose)
+{
+   return Get(device, fe, fe, ir, transpose);
+}
+
+OccaDofQuadMaps& OccaDofQuadMaps::Get(::occa::device device,
+                                      const FiniteElementSpace &trialFESpace,
+                                      const FiniteElementSpace &testFESpace,
+                                      const mfem::IntegrationRule &ir,
+                                      const bool transpose)
+{
+   return Get(device,
+              *trialFESpace.GetFE(0),
+              *testFESpace.GetFE(0),
+              ir,
+              transpose);
+}
+
+OccaDofQuadMaps& OccaDofQuadMaps::Get(::occa::device device,
+                                      const mfem::FiniteElement &trialFE,
+                                      const mfem::FiniteElement &testFE,
+                                      const mfem::IntegrationRule &ir,
+                                      const bool transpose)
+{
+   return (dynamic_cast<const mfem::TensorBasisElement*>(&trialFE)
+           ? GetTensorMaps(device, trialFE, testFE, ir, transpose)
+           : GetSimplexMaps(device, trialFE, testFE, ir, transpose));
+}
+
+OccaDofQuadMaps& OccaDofQuadMaps::GetTensorMaps(::occa::device device,
+                                                const mfem::FiniteElement &fe,
+                                                const mfem::IntegrationRule &ir,
+                                                const bool transpose)
+{
+   return GetTensorMaps(device,
+                        fe, fe,
+                        ir, transpose);
+}
+
+OccaDofQuadMaps& OccaDofQuadMaps::GetTensorMaps(::occa::device device,
+                                                const mfem::FiniteElement &trialFE,
+                                                const mfem::FiniteElement &testFE,
+                                                const mfem::IntegrationRule &ir,
+                                                const bool transpose)
+{
+   const mfem::TensorBasisElement &trialTFE =
+      dynamic_cast<const mfem::TensorBasisElement&>(trialFE);
+   const mfem::TensorBasisElement &testTFE =
+      dynamic_cast<const mfem::TensorBasisElement&>(testFE);
+
+   std::stringstream ss;
+   ss << ::occa::hash(device)
+      << "Tensor"
+      << "O1:"  << trialFE.GetOrder()
+      << "O2:"  << testFE.GetOrder()
+      << "BT1:" << trialTFE.GetBasisType()
+      << "BT2:" << testTFE.GetBasisType()
+      << "Q:"   << ir.GetNPoints();
+   std::string hash = ss.str();
+
+   // If we've already made the dof-quad maps, reuse them
+   OccaDofQuadMaps &maps = AllDofQuadMaps[hash];
+   if (!maps.hash.size())
+   {
+      // Create the dof-quad maps
+      maps.hash = hash;
+
+      OccaDofQuadMaps trialMaps = GetD2QTensorMaps(device, trialFE, ir);
+      OccaDofQuadMaps testMaps  = GetD2QTensorMaps(device, testFE , ir, true);
+
+      maps.dofToQuad   = trialMaps.dofToQuad;
+      maps.dofToQuadD  = trialMaps.dofToQuadD;
+      maps.quadToDof   = testMaps.dofToQuad;
+      maps.quadToDofD  = testMaps.dofToQuadD;
+      maps.quadWeights = testMaps.quadWeights;
+   }
+   return maps;
+}
+
+OccaDofQuadMaps OccaDofQuadMaps::GetD2QTensorMaps(::occa::device device,
+                                                  const mfem::FiniteElement &fe,
+                                                  const mfem::IntegrationRule &ir,
+                                                  const bool transpose)
+{
+   const mfem::TensorBasisElement &tfe =
+      dynamic_cast<const mfem::TensorBasisElement&>(fe);
+
+   const mfem::Poly_1D::Basis &basis = tfe.GetBasis1D();
+   const int order = fe.GetOrder();
+   // [MISSING] Get 1D dofs
+   const int dofs = order + 1;
+   const int dims = fe.GetDim();
+
+   // Create the dof -> quadrature point map
+   const mfem::IntegrationRule &ir1D =
+      mfem::IntRules.Get(mfem::Geometry::SEGMENT, ir.GetOrder());
+   const int quadPoints = ir1D.GetNPoints();
+   const int quadPoints2D = quadPoints*quadPoints;
+   const int quadPoints3D = quadPoints2D*quadPoints;
+   const int quadPointsND = ((dims == 1) ? quadPoints :
+                             ((dims == 2) ? quadPoints2D : quadPoints3D));
+
+   OccaDofQuadMaps maps;
+   // Initialize the dof -> quad mapping
+   maps.dofToQuad.allocate(device,
+                           quadPoints, dofs);
+   maps.dofToQuadD.allocate(device,
+                            quadPoints, dofs);
+
+   double *quadWeights1DData = NULL;
+
+   if (transpose)
+   {
+      maps.dofToQuad.reindex(1,0);
+      maps.dofToQuadD.reindex(1,0);
+      // Initialize quad weights only for transpose
+      maps.quadWeights.allocate(device,
+                                quadPointsND);
+      quadWeights1DData = new double[quadPoints];
+   }
+
+   mfem::Vector d2q(dofs);
+   mfem::Vector d2qD(dofs);
+   for (int q = 0; q < quadPoints; ++q)
+   {
+      const mfem::IntegrationPoint &ip = ir1D.IntPoint(q);
+      basis.Eval(ip.x, d2q, d2qD);
+      if (transpose)
+      {
+         quadWeights1DData[q] = ip.weight;
+      }
+      for (int d = 0; d < dofs; ++d)
+      {
+         maps.dofToQuad(q, d)  = d2q[d];
+         maps.dofToQuadD(q, d) = d2qD[d];
+      }
+   }
+
+   maps.dofToQuad.keepInDevice();
+   maps.dofToQuadD.keepInDevice();
+
+   if (transpose)
+   {
+      for (int q = 0; q < quadPointsND; ++q)
+      {
+         const int qx = q % quadPoints;
+         const int qz = q / quadPoints2D;
+         const int qy = (q - qz*quadPoints2D) / quadPoints;
+         double w = quadWeights1DData[qx];
+         if (dims > 1)
+         {
+            w *= quadWeights1DData[qy];
+         }
+         if (dims > 2)
+         {
+            w *= quadWeights1DData[qz];
+         }
+         maps.quadWeights[q] = w;
+      }
+      maps.quadWeights.keepInDevice();
+      delete [] quadWeights1DData;
+   }
+
+   return maps;
+}
+
+OccaDofQuadMaps& OccaDofQuadMaps::GetSimplexMaps(::occa::device device,
+                                                 const mfem::FiniteElement &fe,
+                                                 const mfem::IntegrationRule &ir,
+                                                 const bool transpose)
+{
+   return GetSimplexMaps(device,
+                         fe, fe,
+                         ir, transpose);
+}
+
+OccaDofQuadMaps& OccaDofQuadMaps::GetSimplexMaps(::occa::device device,
+                                                 const mfem::FiniteElement &trialFE,
+                                                 const mfem::FiniteElement &testFE,
+                                                 const mfem::IntegrationRule &ir,
+                                                 const bool transpose)
+{
+   std::stringstream ss;
+   ss << ::occa::hash(device)
+      << "Simplex"
+      << "O1:" << trialFE.GetOrder()
+      << "O2:" << testFE.GetOrder()
+      << "Q:"  << ir.GetNPoints();
+   std::string hash = ss.str();
+
+   // If we've already made the dof-quad maps, reuse them
+   OccaDofQuadMaps &maps = AllDofQuadMaps[hash];
+   if (!maps.hash.size())
+   {
+      // Create the dof-quad maps
+      maps.hash = hash;
+
+      OccaDofQuadMaps trialMaps = GetD2QSimplexMaps(device, trialFE, ir);
+      OccaDofQuadMaps testMaps  = GetD2QSimplexMaps(device, testFE , ir, true);
+
+      maps.dofToQuad   = trialMaps.dofToQuad;
+      maps.dofToQuadD  = trialMaps.dofToQuadD;
+      maps.quadToDof   = testMaps.dofToQuad;
+      maps.quadToDofD  = testMaps.dofToQuadD;
+      maps.quadWeights = testMaps.quadWeights;
+   }
+   return maps;
+}
+
+OccaDofQuadMaps OccaDofQuadMaps::GetD2QSimplexMaps(::occa::device device,
+                                                   const mfem::FiniteElement &fe,
+                                                   const mfem::IntegrationRule &ir,
+                                                   const bool transpose)
+{
+   const int dims = fe.GetDim();
+   const int numDofs = fe.GetDof();
+   const int numQuad = ir.GetNPoints();
+
+   OccaDofQuadMaps maps;
+   // Initialize the dof -> quad mapping
+   maps.dofToQuad.allocate(device,
+                           numQuad, numDofs);
+   maps.dofToQuadD.allocate(device,
+                            dims, numQuad, numDofs);
+
+   if (transpose)
+   {
+      maps.dofToQuad.reindex(1,0);
+      maps.dofToQuadD.reindex(1,0);
+      // Initialize quad weights only for transpose
+      maps.quadWeights.allocate(device,
+                                numQuad);
+   }
+
+   mfem::Vector d2q(numDofs);
+   mfem::DenseMatrix d2qD(numDofs, dims);
+   for (int q = 0; q < numQuad; ++q)
+   {
+      const mfem::IntegrationPoint &ip = ir.IntPoint(q);
+      if (transpose)
+      {
+         maps.quadWeights[q] = ip.weight;
+      }
+      fe.CalcShape(ip, d2q);
+      fe.CalcDShape(ip, d2qD);
+      for (int d = 0; d < numDofs; ++d)
+      {
+         const double w = d2q[d];
+         maps.dofToQuad(q, d) = w;
+         for (int dim = 0; dim < dims; ++dim)
+         {
+            const double wD = d2qD(d, dim);
+            maps.dofToQuadD(dim, q, d) = wD;
+         }
+      }
+   }
+
+   maps.dofToQuad.keepInDevice();
+   maps.dofToQuadD.keepInDevice();
+   if (transpose)
+   {
+      maps.quadWeights.keepInDevice();
+   }
+
+   return maps;
+}
+
+//---[ Integrator Defines ]-----------
+std::string stringWithDim(const std::string &s, const int dim)
+{
+   std::string ret = s;
+   ret += ('0' + (char) dim);
+   ret += 'D';
+   return ret;
+}
+
+int closestWarpBatchTo(const int value)
+{
+   return ((value + 31) / 32) * 32;
+}
+
+int closestMultipleWarpBatch(const int multiple, const int maxSize)
+{
+   if (multiple > maxSize)
+   {
+      return maxSize;
+   }
+   int batch = (32 / multiple);
+   int minDiff = 32 - (multiple * batch);
+   for (int i = 64; i <= maxSize; i += 32)
+   {
+      const int newDiff = i - (multiple * (i / multiple));
+      if (newDiff < minDiff)
+      {
+         batch = (i / multiple);
+         minDiff = newDiff;
+      }
+   }
+   return batch;
+}
+
+void SetProperties(FiniteElementSpace &fespace,
+                   const mfem::IntegrationRule &ir,
+                   ::occa::properties &props)
+{
+   SetProperties(fespace, fespace, ir, props);
+}
+
+void SetProperties(FiniteElementSpace &trialFESpace,
+                   FiniteElementSpace &testFESpace,
+                   const mfem::IntegrationRule &ir,
+                   ::occa::properties &props)
+{
+   props["defines/TRIAL_VDIM"] = trialFESpace.GetVDim();
+   props["defines/TEST_VDIM"]  = testFESpace.GetVDim();
+   props["defines/NUM_DIM"]    = trialFESpace.GetDim();
+
+   if (trialFESpace.hasTensorBasis())
+   {
+      SetTensorProperties(trialFESpace, testFESpace, ir, props);
+   }
+   else
+   {
+      SetSimplexProperties(trialFESpace, testFESpace, ir, props);
+   }
+}
+
+void SetTensorProperties(FiniteElementSpace &fespace,
+                         const mfem::IntegrationRule &ir,
+                         ::occa::properties &props)
+{
+   SetTensorProperties(fespace, fespace, ir, props);
+}
+
+void SetTensorProperties(FiniteElementSpace &trialFESpace,
+                         FiniteElementSpace &testFESpace,
+                         const mfem::IntegrationRule &ir,
+                         ::occa::properties &props)
+{
+   const mfem::FiniteElement &trialFE = *(trialFESpace.GetFE(0));
+   const mfem::FiniteElement &testFE = *(testFESpace.GetFE(0));
+
+   const mfem::IntegrationRule &ir1D =
+      mfem::IntRules.Get(mfem::Geometry::SEGMENT, ir.GetOrder());
+
+   const int trialDofs = trialFE.GetDof();
+   const int testDofs  = testFE.GetDof();
+   const int numQuad  = ir.GetNPoints();
+
+   const int trialDofs1D = trialFE.GetOrder() + 1;
+   const int testDofs1D  = testFE.GetOrder() + 1;
+   const int quad1D  = ir1D.GetNPoints();
+   int trialDofsND = trialDofs1D;
+   int testDofsND  = testDofs1D;
+   int quadND  = quad1D;
+
+   const bool trialByVDIM = (trialFESpace.GetOrdering() == mfem::Ordering::byVDIM);
+   const bool testByVDIM  = (testFESpace.GetOrdering()  == mfem::Ordering::byVDIM);
+
+   props["defines/ORDERING_BY_NODES"] = 0;
+   props["defines/ORDERING_BY_VDIM"]  = 1;
+   props["defines/VDIM_ORDERING"]  = (int) trialByVDIM;
+   props["defines/TRIAL_ORDERING"] = (int) trialByVDIM;
+   props["defines/TEST_ORDERING"]  = (int) testByVDIM;
+
+   props["defines/USING_TENSOR_OPS"] = 1;
+   props["defines/NUM_DOFS"]  = trialDofs;
+   props["defines/NUM_QUAD"]  = numQuad;
+
+   props["defines/TRIAL_DOFS"] = trialDofs;
+   props["defines/TEST_DOFS"]  = testDofs;
+
+   for (int d = 1; d <= 3; ++d)
+   {
+      if (d > 1)
+      {
+         trialDofsND *= trialDofs1D;
+         testDofsND  *= testDofs1D;
+         quadND *= quad1D;
+      }
+      props["defines"][stringWithDim("NUM_DOFS_", d)] = trialDofsND;
+      props["defines"][stringWithDim("NUM_QUAD_", d)] = quadND;
+
+      props["defines"][stringWithDim("TRIAL_DOFS_", d)] = trialDofsND;
+      props["defines"][stringWithDim("TEST_DOFS_" , d)] = testDofsND;
+   }
+
+   // 1D Defines
+   const int m1InnerBatch = 32 * ((quad1D + 31) / 32);
+   props["defines/A1_ELEMENT_BATCH"]       = closestMultipleWarpBatch(quad1D, 512);
+   props["defines/M1_OUTER_ELEMENT_BATCH"] = closestMultipleWarpBatch(m1InnerBatch,
+                                                                      512);
+   props["defines/M1_INNER_ELEMENT_BATCH"] = m1InnerBatch;
+
+   // 2D Defines
+   props["defines/A2_ELEMENT_BATCH"] = 1;
+   props["defines/A2_QUAD_BATCH"]    = 1;
+   props["defines/M2_ELEMENT_BATCH"] = 32;
+
+   // 3D Defines
+   const int a3QuadBatch = closestMultipleWarpBatch(quadND, 512);
+   props["defines/A3_ELEMENT_BATCH"] = closestMultipleWarpBatch(a3QuadBatch, 512);
+   props["defines/A3_QUAD_BATCH"]    = a3QuadBatch;
+}
+
+void SetSimplexProperties(FiniteElementSpace &fespace,
+                          const mfem::IntegrationRule &ir,
+                          ::occa::properties &props)
+{
+   SetSimplexProperties(fespace, fespace, ir, props);
+}
+
+void SetSimplexProperties(FiniteElementSpace &trialFESpace,
+                          FiniteElementSpace &testFESpace,
+                          const mfem::IntegrationRule &ir,
+                          ::occa::properties &props)
+{
+   const mfem::FiniteElement &trialFE = *(trialFESpace.GetFE(0));
+   const mfem::FiniteElement &testFE = *(testFESpace.GetFE(0));
+
+   const int trialDofs = trialFE.GetDof();
+   const int testDofs  = testFE.GetDof();
+   const int numQuad = ir.GetNPoints();
+   const int maxDQ   = std::max(std::max(trialDofs, testDofs), numQuad);
+
+   const bool trialByVDIM = (trialFESpace.GetOrdering() == mfem::Ordering::byVDIM);
+   const bool testByVDIM  = (testFESpace.GetOrdering()  == mfem::Ordering::byVDIM);
+
+   props["defines/ORDERING_BY_NODES"] = 0;
+   props["defines/ORDERING_BY_VDIM"]  = 1;
+   props["defines/VDIM_ORDERING"]  = (int) trialByVDIM;
+   props["defines/TRIAL_ORDERING"] = (int) trialByVDIM;
+   props["defines/TEST_ORDERING"]  = (int) testByVDIM;
+
+   props["defines/USING_TENSOR_OPS"] = 0;
+   props["defines/NUM_DOFS"]  = trialDofs;
+   props["defines/NUM_QUAD"]  = numQuad;
+
+   props["defines/TRIAL_DOFS"] = trialDofs;
+   props["defines/TEST_DOFS"]  = testDofs;
+
+   // 2D Defines
+   const int quadBatch = closestWarpBatchTo(numQuad);
+   props["defines/A2_ELEMENT_BATCH"] = closestMultipleWarpBatch(quadBatch, 2048);
+   props["defines/A2_QUAD_BATCH"]    = quadBatch;
+   props["defines/M2_INNER_BATCH"]   = closestWarpBatchTo(maxDQ);
+
+   // 3D Defines
+   props["defines/A3_ELEMENT_BATCH"] = closestMultipleWarpBatch(quadBatch, 2048);
+   props["defines/A3_QUAD_BATCH"]    = quadBatch;
+   props["defines/M3_INNER_BATCH"]   = closestWarpBatchTo(maxDQ);
+}
+
+
+//---[ Base Integrator ]--------------
+OccaIntegrator::OccaIntegrator(const Engine &e)
+   : engine(&e),
+     bform(),
+     mesh(),
+     otrialFESpace(),
+     otestFESpace(),
+     trialFESpace(),
+     testFESpace(),
+     itype(DomainIntegrator),
+     ir(NULL),
+     hasTensorBasis(false) { }
+
+OccaIntegrator::~OccaIntegrator() {}
+
+void OccaIntegrator::SetupMaps()
+{
+   maps = OccaDofQuadMaps::Get(GetDevice(),
+                               *otrialFESpace,
+                               *otestFESpace,
+                               *ir);
+
+   mapsTranspose = OccaDofQuadMaps::Get(GetDevice(),
+                                        *otestFESpace,
+                                        *otrialFESpace,
+                                        *ir);
+}
+
+FiniteElementSpace& OccaIntegrator::GetTrialOccaFESpace() const
+{
+   return *otrialFESpace;
+}
+
+FiniteElementSpace& OccaIntegrator::GetTestOccaFESpace() const
+{
+   return *otestFESpace;
+}
+
+mfem::FiniteElementSpace& OccaIntegrator::GetTrialFESpace() const
+{
+   return *trialFESpace;
+}
+
+mfem::FiniteElementSpace& OccaIntegrator::GetTestFESpace() const
+{
+   return *testFESpace;
+}
+
+void OccaIntegrator::SetIntegrationRule(const mfem::IntegrationRule &ir_)
+{
+   ir = &ir_;
+}
+
+const mfem::IntegrationRule& OccaIntegrator::GetIntegrationRule() const
+{
+   return *ir;
+}
+
+OccaDofQuadMaps& OccaIntegrator::GetDofQuadMaps()
+{
+   return maps;
+}
+
+void OccaIntegrator::SetupIntegrator(OccaBilinearForm &bform_,
+                                     const ::occa::properties &props_,
+                                     const OccaIntegratorType itype_)
+{
+   MFEM_ASSERT(engine == &bform_.OccaEngine(), "");
+   bform     = &bform_;
+   mesh      = &(bform_.GetMesh());
+
+   otrialFESpace = &(bform_.GetTrialOccaFESpace());
+   otestFESpace  = &(bform_.GetTestOccaFESpace());
+
+   trialFESpace = &(bform_.GetTrialFESpace());
+   testFESpace  = &(bform_.GetTestFESpace());
+
+   hasTensorBasis = otrialFESpace->hasTensorBasis();
+
+   props = props_;
+   itype = itype_;
+
+   if (ir == NULL)
+   {
+      SetupIntegrationRule();
+   }
+
+   SetupMaps();
+
+   SetProperties(*otrialFESpace,
+                 *otestFESpace,
+                 *ir,
+                 props);
+
+   Setup();
+}
+
+OccaGeometry OccaIntegrator::GetGeometry(const int flags)
+{
+   return OccaGeometry::Get(GetDevice(), *otrialFESpace, *ir, flags);
+}
+
+::occa::kernel OccaIntegrator::GetAssembleKernel(const ::occa::properties
+                                                 &props)
+{
+   const mfem::FiniteElement &fe = *(trialFESpace->GetFE(0));
+   return GetKernel(stringWithDim("Assemble", fe.GetDim()),
+                    props);
+}
+
+::occa::kernel OccaIntegrator::GetMultAddKernel(const ::occa::properties &props)
+{
+   const mfem::FiniteElement &fe = *(trialFESpace->GetFE(0));
+   return GetKernel(stringWithDim("MultAdd", fe.GetDim()),
+                    props);
+}
+
+::occa::kernel OccaIntegrator::GetKernel(const std::string &kernelName,
+                                         const ::occa::properties &props)
+{
+   const std::string filename = GetName() + ".okl";
+   const std::string &okl_path = OccaEngine().GetOklPath();
+   const std::string &okl_defines = OccaEngine().GetOklDefines();
+   return GetDevice().buildKernel(okl_path + filename,
+                                  kernelName,
+                                  props + okl_defines);
+}
+//====================================
+
+
+//---[ Diffusion Integrator ]---------
+OccaDiffusionIntegrator::OccaDiffusionIntegrator(const OccaCoefficient &coeff_)
+   :
+   OccaIntegrator(coeff_.OccaEngine()),
+   coeff(coeff_),
+   assembledOperator(*(new Layout(coeff_.OccaEngine(), 0)))
+{
+   coeff.SetName("COEFF");
+}
+
+OccaDiffusionIntegrator::~OccaDiffusionIntegrator() {}
+
+
+std::string OccaDiffusionIntegrator::GetName()
+{
+   return "DiffusionIntegrator";
+}
+
+void OccaDiffusionIntegrator::SetupIntegrationRule()
+{
+   const FiniteElement &trialFE = *(trialFESpace->GetFE(0));
+   const FiniteElement &testFE  = *(testFESpace->GetFE(0));
+   ir = &mfem::DiffusionIntegrator::GetRule(trialFE, testFE);
+}
+
+void OccaDiffusionIntegrator::Setup()
+{
+   ::occa::properties kernelProps = props;
+
+   coeff.Setup(*this, kernelProps);
+
+   // Setup assemble and mult kernels
+   assembleKernel = GetAssembleKernel(kernelProps);
+   multKernel     = GetMultAddKernel(kernelProps);
+}
+
+void OccaDiffusionIntegrator::Assemble()
+{
+   const mfem::FiniteElement &fe = *(trialFESpace->GetFE(0));
+
+   const int dims = fe.GetDim();
+   const int symmDims = (dims * (dims + 1)) / 2; // 1x1: 1, 2x2: 3, 3x3: 6
+
+   const int elements = trialFESpace->GetNE();
+   const int quadraturePoints = ir->GetNPoints();
+
+   OccaGeometry geom = GetGeometry(OccaGeometry::Jacobian);
+
+   assembledOperator.Resize<double>(symmDims * quadraturePoints * elements,
+                                    NULL);
+
+   assembleKernel((int) mesh->GetNE(),
+                  maps.quadWeights,
+                  geom.J,
+                  coeff,
+                  assembledOperator.OccaMem());
+}
+
+void OccaDiffusionIntegrator::MultAdd(Vector &x, Vector &y)
+{
+   // Note: x and y are E-vectors
+
+   multKernel((int) mesh->GetNE(),
+              maps.dofToQuad,
+              maps.dofToQuadD,
+              maps.quadToDof,
+              maps.quadToDofD,
+              assembledOperator.OccaMem(),
+              x.OccaMem(), y.OccaMem());
+}
+//====================================
+
+
+//---[ Mass Integrator ]--------------
+OccaMassIntegrator::OccaMassIntegrator(const OccaCoefficient &coeff_) :
+   OccaIntegrator(coeff_.OccaEngine()),
+   coeff(coeff_),
+   assembledOperator(*(new Layout(coeff_.OccaEngine(), 0)))
+{
+   coeff.SetName("COEFF");
+}
+
+OccaMassIntegrator::~OccaMassIntegrator() {}
+
+std::string OccaMassIntegrator::GetName()
+{
+   return "MassIntegrator";
+}
+
+void OccaMassIntegrator::SetupIntegrationRule()
+{
+   const mfem::FiniteElement &trialFE = *(trialFESpace->GetFE(0));
+   const mfem::FiniteElement &testFE  = *(testFESpace->GetFE(0));
+   mfem::ElementTransformation &T = *trialFESpace->GetElementTransformation(0);
+   ir = &mfem::MassIntegrator::GetRule(trialFE, testFE, T);
+}
+
+void OccaMassIntegrator::Setup()
+{
+   ::occa::properties kernelProps = props;
+
+   coeff.Setup(*this, kernelProps);
+
+   // Setup assemble and mult kernels
+   assembleKernel = GetAssembleKernel(kernelProps);
+   multKernel     = GetMultAddKernel(kernelProps);
+}
+
+void OccaMassIntegrator::Assemble()
+{
+   if (assembledOperator.Size())
+   {
+      return;
+   }
+
+   const int elements = trialFESpace->GetNE();
+   const int quadraturePoints = ir->GetNPoints();
+
+   OccaGeometry geom = GetGeometry(OccaGeometry::Jacobian);
+
+   assembledOperator.Resize<double>(quadraturePoints * elements, NULL);
+
+   assembleKernel((int) mesh->GetNE(),
+                  maps.quadWeights,
+                  geom.J,
+                  coeff,
+                  assembledOperator.OccaMem());
+}
+
+void OccaMassIntegrator::SetOperator(Vector &v)
+{
+   assembledOperator = v;
+}
+
+void OccaMassIntegrator::MultAdd(Vector &x, Vector &y)
+{
+   multKernel((int) mesh->GetNE(),
+              maps.dofToQuad,
+              maps.dofToQuadD,
+              maps.quadToDof,
+              maps.quadToDofD,
+              assembledOperator.OccaMem(),
+              x.OccaMem(), y.OccaMem());
+}
+//====================================
+
+
+//---[ Vector Mass Integrator ]--------------
+OccaVectorMassIntegrator::OccaVectorMassIntegrator(const OccaCoefficient &
+                                                   coeff_)
+   :
+   OccaIntegrator(coeff_.OccaEngine()),
+   coeff(coeff_),
+   assembledOperator(*(new Layout(coeff_.OccaEngine(), 0)))
+{
+   coeff.SetName("COEFF");
+}
+
+OccaVectorMassIntegrator::~OccaVectorMassIntegrator() {}
+
+std::string OccaVectorMassIntegrator::GetName()
+{
+   return "VectorMassIntegrator";
+}
+
+void OccaVectorMassIntegrator::SetupIntegrationRule()
+{
+   const mfem::FiniteElement &trialFE = *(trialFESpace->GetFE(0));
+   const mfem::FiniteElement &testFE  = *(testFESpace->GetFE(0));
+   mfem::ElementTransformation &T = *trialFESpace->GetElementTransformation(0);
+   ir = &mfem::MassIntegrator::GetRule(trialFE, testFE, T);
+}
+
+void OccaVectorMassIntegrator::Setup()
+{
+   ::occa::properties kernelProps = props;
+
+   coeff.Setup(*this, kernelProps);
+
+   // Setup assemble and mult kernels
+   assembleKernel = GetAssembleKernel(kernelProps);
+   multKernel     = GetMultAddKernel(kernelProps);
+}
+
+void OccaVectorMassIntegrator::Assemble()
+{
+   const int elements = trialFESpace->GetNE();
+   const int quadraturePoints = ir->GetNPoints();
+
+   OccaGeometry geom = GetGeometry(OccaGeometry::Jacobian);
+
+   assembledOperator.Resize<double>(quadraturePoints * elements, NULL);
+
+   assembleKernel((int) mesh->GetNE(),
+                  maps.quadWeights,
+                  geom.J,
+                  coeff,
+                  assembledOperator.OccaMem());
+}
+
+void OccaVectorMassIntegrator::MultAdd(Vector &x, Vector &y)
+{
+   multKernel((int) mesh->GetNE(),
+              maps.dofToQuad,
+              maps.dofToQuadD,
+              maps.quadToDof,
+              maps.quadToDofD,
+              assembledOperator.OccaMem(),
+              x.OccaMem(), y.OccaMem());
+}
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,323 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_BILIN_INTEG_HPP
+#define MFEM_BACKENDS_OCCA_BILIN_INTEG_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "fespace.hpp"
+#include "bilinearform.hpp"
+#include "coefficient.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+class OccaGeometry
+{
+public:
+   ::occa::array<double> meshNodes;
+   ::occa::array<double> J, invJ, detJ;
+
+   // byVDIM  -> [x y z x y z x y z]
+   // byNodes -> [x x x y y y z z z]
+   static const int Jacobian    = (1 << 0);
+   static const int JacobianInv = (1 << 1);
+   static const int JacobianDet = (1 << 2);
+
+   static OccaGeometry Get(::occa::device device,
+                           FiniteElementSpace &ofespace,
+                           const IntegrationRule &ir,
+                           const int flags = (Jacobian    |
+                                              JacobianInv |
+                                              JacobianDet));
+};
+
+class OccaDofQuadMaps
+{
+private:
+   // Reuse dof-quad maps
+   static std::map<std::string, OccaDofQuadMaps> AllDofQuadMaps;
+   std::string hash;
+
+public:
+   // Local stiffness matrices (B and B^T operators)
+   ::occa::array<double, ::occa::dynamic> dofToQuad, dofToQuadD; // B
+   ::occa::array<double, ::occa::dynamic> quadToDof, quadToDofD; // B^T
+   ::occa::array<double> quadWeights;
+
+   OccaDofQuadMaps();
+   OccaDofQuadMaps(const OccaDofQuadMaps &maps);
+   OccaDofQuadMaps& operator = (const OccaDofQuadMaps &maps);
+
+   // [[x y] [x y] [x y]]
+   // [[x y z] [x y z] [x y z]]
+   // mfem::GridFunction* mfem::Mesh::GetNodes() { return Nodes; }
+
+   // mfem::FiniteElementSpace *Nodes->FESpace()
+   // 25
+   // 1D [x x x x x x]
+   // 2D [x y x y x y]
+   // GetVdim()
+   // 3D ordering == byVDIM  -> [x y z x y z x y z x y z x y z x y z]
+   //    ordering == byNODES -> [x x x x x x y y y y y y z z z z z z]
+   static OccaDofQuadMaps& Get(::occa::device device,
+                               const FiniteElementSpace &fespace,
+                               const mfem::IntegrationRule &ir,
+                               const bool transpose = false);
+
+   static OccaDofQuadMaps& Get(::occa::device device,
+                               const mfem::FiniteElement &fe,
+                               const mfem::IntegrationRule &ir,
+                               const bool transpose = false);
+
+   static OccaDofQuadMaps& Get(::occa::device device,
+                               const FiniteElementSpace &trialFESpace,
+                               const FiniteElementSpace &testFESpace,
+                               const mfem::IntegrationRule &ir,
+                               const bool transpose = false);
+
+   static OccaDofQuadMaps& Get(::occa::device device,
+                               const mfem::FiniteElement &trialFE,
+                               const mfem::FiniteElement &testFE,
+                               const mfem::IntegrationRule &ir,
+                               const bool transpose = false);
+
+   static OccaDofQuadMaps& GetTensorMaps(::occa::device device,
+                                         const mfem::FiniteElement &fe,
+                                         const mfem::IntegrationRule &ir,
+                                         const bool transpose = false);
+
+   static OccaDofQuadMaps& GetTensorMaps(::occa::device device,
+                                         const mfem::FiniteElement &trialFE,
+                                         const mfem::FiniteElement &testFE,
+                                         const mfem::IntegrationRule &ir,
+                                         const bool transpose = false);
+
+   static OccaDofQuadMaps GetD2QTensorMaps(::occa::device device,
+                                           const mfem::FiniteElement &fe,
+                                           const mfem::IntegrationRule &ir,
+                                           const bool transpose = false);
+
+   static OccaDofQuadMaps& GetSimplexMaps(::occa::device device,
+                                          const mfem::FiniteElement &fe,
+                                          const mfem::IntegrationRule &ir,
+                                          const bool transpose = false);
+
+   static OccaDofQuadMaps& GetSimplexMaps(::occa::device device,
+                                          const mfem::FiniteElement &trialFE,
+                                          const mfem::FiniteElement &testFE,
+                                          const mfem::IntegrationRule &ir,
+                                          const bool transpose = false);
+
+   static OccaDofQuadMaps GetD2QSimplexMaps(::occa::device device,
+                                            const mfem::FiniteElement &fe,
+                                            const mfem::IntegrationRule &ir,
+                                            const bool transpose = false);
+};
+
+//---[ Define Methods ]---------------
+std::string stringWithDim(const std::string &s, const int dim);
+int closestWarpBatch(const int multiple, const int maxSize);
+
+void SetProperties(FiniteElementSpace &fespace,
+                   const mfem::IntegrationRule &ir,
+                   ::occa::properties &props);
+
+void SetProperties(FiniteElementSpace &trialFESpace,
+                   FiniteElementSpace &testFESpace,
+                   const mfem::IntegrationRule &ir,
+                   ::occa::properties &props);
+
+void SetTensorProperties(FiniteElementSpace &fespace,
+                         const mfem::IntegrationRule &ir,
+                         ::occa::properties &props);
+
+void SetTensorProperties(FiniteElementSpace &trialFESpace,
+                         FiniteElementSpace &testFESpace,
+                         const IntegrationRule &ir,
+                         ::occa::properties &props);
+
+void SetSimplexProperties(FiniteElementSpace &fespace,
+                          const IntegrationRule &ir,
+                          ::occa::properties &props);
+
+void SetSimplexProperties(FiniteElementSpace &trialFESpace,
+                          FiniteElementSpace &testFESpace,
+                          const IntegrationRule &ir,
+                          ::occa::properties &props);
+
+//---[ Base Integrator ]--------------
+class OccaIntegrator
+{
+protected:
+   SharedPtr<const Engine> engine;
+
+   OccaBilinearForm *bform;
+   mfem::Mesh *mesh;
+
+   FiniteElementSpace *otrialFESpace;
+   FiniteElementSpace *otestFESpace;
+
+   mfem::FiniteElementSpace *trialFESpace;
+   mfem::FiniteElementSpace *testFESpace;
+
+   ::occa::properties props;
+   OccaIntegratorType itype;
+
+   const IntegrationRule *ir;
+   bool hasTensorBasis;
+   OccaDofQuadMaps maps;
+   OccaDofQuadMaps mapsTranspose;
+
+public:
+   OccaIntegrator(const Engine &e);
+   virtual ~OccaIntegrator();
+
+   const Engine &OccaEngine() const { return *engine; }
+
+   ::occa::device GetDevice(int idx = 0) const
+   { return engine->GetDevice(idx); }
+
+   virtual std::string GetName() = 0;
+
+   FiniteElementSpace& GetTrialOccaFESpace() const;
+   FiniteElementSpace& GetTestOccaFESpace() const;
+
+   mfem::FiniteElementSpace& GetTrialFESpace() const;
+   mfem::FiniteElementSpace& GetTestFESpace() const;
+
+   void SetIntegrationRule(const mfem::IntegrationRule &ir_);
+   const mfem::IntegrationRule& GetIntegrationRule() const;
+
+   OccaDofQuadMaps& GetDofQuadMaps();
+
+   void SetupMaps();
+
+   virtual void SetupIntegrationRule() = 0;
+
+   virtual void SetupIntegrator(OccaBilinearForm &bform_,
+                                const ::occa::properties &props_,
+                                const OccaIntegratorType itype_);
+
+   virtual void Setup() = 0;
+
+   virtual void Assemble() = 0;
+   /// This method works on E-vectors!
+   virtual void MultAdd(Vector &x, Vector &y) = 0;
+
+   virtual void MultTransposeAdd(Vector &x, Vector &y)
+   {
+      mfem_error("OccaIntegrator::MultTransposeAdd() is not overloaded!");
+   }
+
+   OccaGeometry GetGeometry(const int flags = (OccaGeometry::Jacobian    |
+                                               OccaGeometry::JacobianInv |
+                                               OccaGeometry::JacobianDet));
+
+   ::occa::kernel GetAssembleKernel(const ::occa::properties &props);
+   ::occa::kernel GetMultAddKernel(const ::occa::properties &props);
+
+   ::occa::kernel GetKernel(const std::string &kernelName,
+                            const ::occa::properties &props);
+};
+//====================================
+
+
+//---[ Diffusion Integrator ]---------
+class OccaDiffusionIntegrator : public OccaIntegrator
+{
+private:
+   OccaCoefficient coeff;
+
+   ::occa::kernel assembleKernel, multKernel;
+
+   Vector assembledOperator;
+
+public:
+   OccaDiffusionIntegrator(const OccaCoefficient &coeff_);
+   virtual ~OccaDiffusionIntegrator();
+
+   virtual std::string GetName();
+
+   virtual void SetupIntegrationRule();
+
+   virtual void Setup();
+
+   virtual void Assemble();
+   virtual void MultAdd(Vector &x, Vector &y);
+};
+//====================================
+
+
+//---[ Mass Integrator ]--------------
+class OccaMassIntegrator : public OccaIntegrator
+{
+private:
+   OccaCoefficient coeff;
+
+   ::occa::kernel assembleKernel, multKernel;
+
+   Vector assembledOperator;
+
+public:
+   OccaMassIntegrator(const OccaCoefficient &coeff_);
+   virtual ~OccaMassIntegrator();
+
+   virtual std::string GetName();
+
+   virtual void SetupIntegrationRule();
+
+   virtual void Setup();
+
+   virtual void Assemble();
+   void SetOperator(Vector &v);
+
+   virtual void MultAdd(Vector &x, Vector &y);
+};
+//====================================
+
+//---[ Vector Mass Integrator ]--------------
+class OccaVectorMassIntegrator : public OccaIntegrator
+{
+private:
+   OccaCoefficient coeff;
+
+   ::occa::kernel assembleKernel, multKernel;
+
+   Vector assembledOperator;
+
+public:
+   OccaVectorMassIntegrator(const OccaCoefficient &coeff_);
+   virtual ~OccaVectorMassIntegrator();
+
+   virtual std::string GetName();
+
+   virtual void SetupIntegrationRule();
+
+   virtual void Setup();
+
+   virtual void Assemble();
+
+   virtual void MultAdd(Vector &x, Vector &y);
+};
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_BILIN_INTEG_HPP
@@ -0,0 +1,344 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "coefficient.hpp"
+#include "bilininteg.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+//---[ Parameter ]------------
+OccaParameter::~OccaParameter() {}
+
+void OccaParameter::Setup(OccaIntegrator &integ,
+                          ::occa::properties &props) {}
+
+::occa::kernelArg OccaParameter::KernelArgs()
+{
+   return ::occa::kernelArg();
+}
+//====================================
+
+
+//---[ Include Parameter ]------------
+OccaIncludeParameter::OccaIncludeParameter(const std::string &filename_) :
+   filename(filename_) {}
+
+OccaParameter* OccaIncludeParameter::Clone()
+{
+   return new OccaIncludeParameter(filename);
+}
+
+void OccaIncludeParameter::Setup(OccaIntegrator &integ,
+                                 ::occa::properties &props)
+{
+   props["headers"].asArray() += "#include " + filename;
+}
+//====================================
+
+
+//---[ Source Parameter ]------------
+OccaSourceParameter::OccaSourceParameter(const std::string &source_) :
+   source(source_) {}
+
+OccaParameter* OccaSourceParameter::Clone()
+{
+   return new OccaSourceParameter(source);
+}
+
+void OccaSourceParameter::Setup(OccaIntegrator &integ,
+                                ::occa::properties &props)
+{
+   props["headers"].asArray() += source;
+}
+//====================================
+
+//---[ Vector Parameter ]-------
+OccaVectorParameter::OccaVectorParameter(const std::string &name_,
+                                         Vector &v_,
+                                         const bool useRestrict_) :
+   name(name_),
+   v(v_),
+   useRestrict(useRestrict_),
+   attr("") {}
+
+OccaVectorParameter::OccaVectorParameter(const std::string &name_,
+                                         Vector &v_,
+                                         const std::string &attr_,
+                                         const bool useRestrict_) :
+   name(name_),
+   v(v_),
+   useRestrict(useRestrict_),
+   attr(attr_) {}
+
+OccaParameter* OccaVectorParameter::Clone()
+{
+   return new OccaVectorParameter(name, v, attr, useRestrict);
+}
+
+void OccaVectorParameter::Setup(OccaIntegrator &integ,
+                                ::occa::properties &props)
+{
+   std::string &args = (props["defines/COEFF_ARGS"]
+                        .asString()
+                        .string());
+   args += "const double *";
+   if (useRestrict)
+   {
+      args += " restrict ";
+   }
+   args += name;
+   if (attr.size())
+   {
+      args += ' ';
+      args += attr;
+   }
+   args += ",\n";
+}
+
+::occa::kernelArg OccaVectorParameter::KernelArgs()
+{
+   return ::occa::kernelArg(v.OccaMem());
+}
+//====================================
+
+
+//---[ GridFunction Parameter ]-------
+OccaGridFunctionParameter::OccaGridFunctionParameter(const std::string &name_,
+                                                     OccaGridFunction &gf_,
+                                                     const bool useRestrict_)
+   : name(name_),
+     gf(gf_),
+     gfQuad(*(new Layout(gf_.OccaLayout().OccaEngine(), 0))),
+     useRestrict(useRestrict_) {}
+
+OccaParameter* OccaGridFunctionParameter::Clone()
+{
+   OccaGridFunctionParameter *param =
+      new OccaGridFunctionParameter(name, gf, useRestrict);
+   param->gfQuad.MakeRef(gfQuad);
+   return param;
+}
+
+void OccaGridFunctionParameter::Setup(OccaIntegrator &integ,
+                                      ::occa::properties &props)
+{
+
+   std::string &args = (props["defines/COEFF_ARGS"]
+                        .asString()
+                        .string());
+   args += "const double *";
+   if (useRestrict)
+   {
+      args += " restrict ";
+   }
+   args += name;
+   args += " @dim(NUM_QUAD, numElements),\n";
+
+   gf.ToQuad(integ.GetIntegrationRule(), gfQuad);
+}
+
+::occa::kernelArg OccaGridFunctionParameter::KernelArgs()
+{
+   return gfQuad.OccaMem();
+}
+//====================================
+
+
+//---[ Coefficient ]------------------
+OccaCoefficient::OccaCoefficient(const Engine &e, const double value) :
+   engine(&e),
+   integ(NULL),
+   name("COEFF")
+{
+   coeffValue = value;
+}
+
+OccaCoefficient::OccaCoefficient(const Engine &e, const std::string &source) :
+   engine(&e),
+   integ(NULL),
+   name("COEFF")
+{
+   coeffValue = source;
+}
+
+OccaCoefficient::OccaCoefficient(const Engine &e, const char *source) :
+   engine(&e),
+   integ(NULL),
+   name("COEFF")
+{
+   coeffValue = source;
+}
+
+OccaCoefficient::OccaCoefficient(const OccaCoefficient &coeff) :
+   engine(coeff.engine),
+   integ(NULL),
+   name(coeff.name),
+   coeffValue(coeff.coeffValue)
+{
+
+   const int paramCount = (int) coeff.params.size();
+   for (int i = 0; i < paramCount; ++i)
+   {
+      params.push_back(coeff.params[i]->Clone());
+   }
+}
+
+OccaCoefficient::~OccaCoefficient()
+{
+   const int paramCount = (int) params.size();
+   for (int i = 0; i < paramCount; ++i)
+   {
+      delete params[i];
+   }
+}
+
+OccaCoefficient& OccaCoefficient::SetName(const std::string &name_)
+{
+   name = name_;
+   return *this;
+}
+
+void OccaCoefficient::Setup(OccaIntegrator &integ_,
+                            ::occa::properties &props_)
+{
+   integ = &integ_;
+
+   const int paramCount = (int) params.size();
+   props_["defines"][name + "_ARGS"] = "";
+   for (int i = 0; i < paramCount; ++i)
+   {
+      params[i]->Setup(integ_, props_);
+   }
+   props_["defines"][name] = coeffValue;
+
+   props = props_;
+}
+
+OccaCoefficient& OccaCoefficient::Add(OccaParameter *param)
+{
+   params.push_back(param);
+   return *this;
+}
+
+OccaCoefficient& OccaCoefficient::IncludeHeader(const std::string &filename)
+{
+   return Add(new OccaIncludeParameter(filename));
+}
+
+OccaCoefficient& OccaCoefficient::IncludeSource(const std::string &source)
+{
+   return Add(new OccaSourceParameter(source));
+}
+
+OccaCoefficient& OccaCoefficient::AddVector(const std::string &name_,
+                                            Vector &v,
+                                            const bool useRestrict)
+{
+   return Add(new OccaVectorParameter(name_, v, useRestrict));
+}
+
+OccaCoefficient& OccaCoefficient::AddVector(const std::string &name_,
+                                            Vector &v,
+                                            const std::string &attr,
+                                            const bool useRestrict)
+{
+   return Add(new OccaVectorParameter(name_, v, attr, useRestrict));
+}
+
+OccaCoefficient& OccaCoefficient::AddGridFunction(const std::string &name_,
+                                                  OccaGridFunction &gf,
+                                                  const bool useRestrict)
+{
+   return Add(new OccaGridFunctionParameter(name_, gf, useRestrict));
+}
+
+bool OccaCoefficient::IsConstant()
+{
+   return coeffValue.isNumber();
+}
+
+double OccaCoefficient::GetConstantValue()
+{
+   if (!IsConstant())
+   {
+      mfem_error("OccaCoefficient is not constant");
+   }
+   return coeffValue.number();
+}
+
+Vector OccaCoefficient::Eval()
+{
+   if (integ == NULL)
+   {
+      mfem_error("OccaCoefficient requires a Setup() call before Eval()");
+   }
+
+   mfem::FiniteElementSpace &fespace = integ->GetTrialFESpace();
+   const mfem::IntegrationRule &ir   = integ->GetIntegrationRule();
+
+   const int elements = fespace.GetNE();
+   const int numQuad  = ir.GetNPoints();
+
+   Vector quadCoeff(*(new Layout(OccaEngine(), numQuad * elements)));
+   Eval(quadCoeff);
+   return quadCoeff;
+}
+
+void OccaCoefficient::Eval(Vector &quadCoeff)
+{
+   const std::string &okl_path = OccaEngine().GetOklPath();
+   const std::string &okl_defines = OccaEngine().GetOklDefines();
+   static ::occa::kernelBuilder builder =
+      ::occa::kernelBuilder::fromFile(okl_path + "coefficient.okl",
+                                      "CoefficientEval", okl_defines);
+
+   if (integ == NULL)
+   {
+      mfem_error("OccaCoefficient requires a Setup() call before Eval()");
+   }
+
+   const int elements = integ->GetTrialFESpace().GetNE();
+
+   ::occa::properties kernelProps = props;
+   if (name != "COEFF")
+   {
+      kernelProps["defines/COEFF"]      = name;
+      kernelProps["defines/COEFF_ARGS"] = name + "_ARGS";
+   }
+   kernelProps += okl_defines;
+
+   ::occa::kernel evalKernel = builder.build(GetDevice(), kernelProps);
+   evalKernel(elements, *this, quadCoeff.OccaMem());
+}
+
+OccaCoefficient::operator ::occa::kernelArg ()
+{
+   ::occa::kernelArg kArg;
+   const int paramCount = (int) params.size();
+   for (int i = 0; i < paramCount; ++i)
+   {
+      kArg.add(params[i]->KernelArgs());
+   }
+   return kArg;
+}
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,284 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_COEFFICIENT_HPP
+#define MFEM_BACKENDS_OCCA_COEFFICIENT_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "vector.hpp"
+#include "gridfunc.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+class OccaIntegrator;
+
+
+class OccaParameter
+{
+public:
+   virtual ~OccaParameter();
+
+   virtual OccaParameter* Clone() = 0;
+
+   virtual void Setup(OccaIntegrator &integ,
+                      ::occa::properties &props);
+
+   virtual ::occa::kernelArg KernelArgs();
+};
+
+
+//---[ Include Parameter ]------------
+class OccaIncludeParameter : public OccaParameter
+{
+private:
+   std::string filename;
+
+public:
+   OccaIncludeParameter(const std::string &filename_);
+
+   virtual OccaParameter* Clone();
+
+   virtual void Setup(OccaIntegrator &integ,
+                      ::occa::properties &props);
+};
+//====================================
+
+
+//---[ Source Parameter ]------------
+class OccaSourceParameter : public OccaParameter
+{
+private:
+   std::string source;
+
+public:
+   OccaSourceParameter(const std::string &filename_);
+
+   virtual OccaParameter* Clone();
+
+   virtual void Setup(OccaIntegrator &integ,
+                      ::occa::properties &props);
+};
+//====================================
+
+
+//---[ Define Parameter ]------------
+template <class TM>
+class OccaDefineParameter : public OccaParameter
+{
+private:
+   const std::string name;
+   TM value;
+
+public:
+   OccaDefineParameter(const std::string &name_,
+                       const TM &value_) :
+      name(name_),
+      value(value_) {}
+
+   virtual OccaParameter* Clone()
+   {
+      return new OccaDefineParameter(name, value);
+   }
+
+   virtual void Setup(OccaIntegrator &integ,
+                      ::occa::properties &props)
+   {
+      props["defines"][name] = value;
+   }
+};
+//====================================
+
+
+//---[ Variable Parameter ]-----------
+template <class TM>
+class OccaVariableParameter : public OccaParameter
+{
+private:
+   const std::string name;
+   const TM &value;
+
+public:
+   OccaVariableParameter(const std::string &name_,
+                         const TM &value_) :
+      name(name_),
+      value(value_) {}
+
+   virtual OccaParameter* Clone()
+   {
+      return new OccaVariableParameter(name, value);
+   }
+
+   virtual void Setup(OccaIntegrator &integ,
+                      ::occa::properties &props)
+   {
+      std::string &args = (props["defines/COEFF_ARGS"]
+                           .asString()
+                           .string());
+      // const TM name,\n"
+      args += "const ";
+      args += ::occa::primitiveinfo<TM>::name;
+      args += ' ';
+      args += name;
+      args += ",\n";
+   }
+
+   virtual ::occa::kernelArg KernelArgs()
+   {
+      return ::occa::kernelArg(value);
+   }
+};
+//====================================
+
+
+//---[ Vector Parameter ]-------
+class OccaVectorParameter : public OccaParameter
+{
+private:
+   const std::string name;
+   Vector v;
+   bool useRestrict;
+   std::string attr;
+
+public:
+   OccaVectorParameter(const std::string &name_,
+                       Vector &v_,
+                       const bool useRestrict_ = false);
+
+   OccaVectorParameter(const std::string &name_,
+                       Vector &v_,
+                       const std::string &attr_,
+                       const bool useRestrict_ = false);
+
+   virtual OccaParameter* Clone();
+
+   virtual void Setup(OccaIntegrator &integ,
+                      ::occa::properties &props);
+
+   virtual ::occa::kernelArg KernelArgs();
+};
+//====================================
+
+
+//---[ GridFunction Parameter ]-------
+class OccaGridFunctionParameter : public OccaParameter
+{
+private:
+   const std::string name;
+   OccaGridFunction &gf;
+   Vector gfQuad;
+   bool useRestrict;
+
+public:
+   OccaGridFunctionParameter(const std::string &name_,
+                             OccaGridFunction &gf_,
+                             const bool useRestrict_ = false);
+
+   virtual OccaParameter* Clone();
+
+   virtual void Setup(OccaIntegrator &integ,
+                      ::occa::properties &props);
+
+   virtual ::occa::kernelArg KernelArgs();
+};
+//====================================
+
+
+//---[ Coefficient ]------------------
+// [MISSING]
+// Needs to know about the integrator's
+//   - fespace
+//   - ir
+// Step where parameters that need the ir get called for setup
+// For example, GridFunction (d, e) -> (q, e)
+class OccaCoefficient
+{
+private:
+   SharedPtr<const Engine> engine;
+
+   OccaIntegrator *integ;
+
+   std::string name;
+   ::occa::json coeffValue;
+
+   ::occa::properties props;
+   std::vector<OccaParameter*> params;
+
+public:
+   OccaCoefficient(const Engine &e, const double value = 1.0);
+   OccaCoefficient(const Engine &e, const std::string &source);
+   OccaCoefficient(const Engine &e, const char *source);
+   ~OccaCoefficient();
+
+   OccaCoefficient(const OccaCoefficient &coeff);
+
+   const Engine &OccaEngine() const { return *engine; }
+
+   ::occa::device GetDevice(int idx = 0) const
+   { return engine->GetDevice(idx); }
+
+   OccaCoefficient& SetName(const std::string &name_);
+
+   void Setup(OccaIntegrator &integ_,
+              ::occa::properties &props_);
+
+   OccaCoefficient& Add(OccaParameter *param);
+
+   OccaCoefficient& IncludeHeader(const std::string &filename);
+   OccaCoefficient& IncludeSource(const std::string &source);
+
+   template <class TM>
+   OccaCoefficient& AddDefine(const std::string &name_, const TM &value)
+   {
+      return Add(new OccaDefineParameter<TM>(name_, value));
+   }
+
+   template <class TM>
+   OccaCoefficient& AddVariable(const std::string &name_, const TM &value)
+   {
+      return Add(new OccaVariableParameter<TM>(name_, value));
+   }
+
+   OccaCoefficient& AddVector(const std::string &name_,
+                              Vector &v,
+                              const bool useRestrict = false);
+
+
+   OccaCoefficient& AddVector(const std::string &name_,
+                              Vector &v,
+                              const std::string &attr,
+                              const bool useRestrict = false);
+
+   OccaCoefficient& AddGridFunction(const std::string &name_,
+                                    OccaGridFunction &gf,
+                                    const bool useRestrict = false);
+
+   bool IsConstant();
+   double GetConstantValue();
+
+   Vector Eval();
+   void Eval(Vector &quadCoeff);
+
+   operator ::occa::kernelArg ();
+};
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_COEFFICIENT_HPP
@@ -0,0 +1,40 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_OCCA_DEFINES
+#define MFEM_OCCA_DEFINES
+
+#ifndef USING_TENSOR_OPS
+#  define USING_TENSOR_OPS 0
+#endif
+
+#ifdef OCCA_USING_GPU
+#  define GPU_ORDER_2(I0, I1)         @dimOrder(I0, I1)
+#  define GPU_ORDER_3(I0, I1, I2)     @dimOrder(I0, I1, I2)
+#  define GPU_ORDER_4(I0, I1, I2, I3) @dimOrder(I0, I1, I2, I3)
+#else
+#  define GPU_ORDER_2(I0, I1)         @dimOrder(0, 1)
+#  define GPU_ORDER_3(I0, I1, I2)     @dimOrder(0, 1, 2)
+#  define GPU_ORDER_4(I0, I1, I2, I3) @dimOrder(0, 1, 2, 3)
+#endif
+
+#ifndef COEFF
+#  define COEFF 1.0
+#  define COEFF_ARGS
+#endif
+
+#if USING_TENSOR_OPS
+#  include "mfem-occa://defines/tensor.okl"
+#else
+#  include "mfem-occa://defines/simplex.okl"
+#endif
+
+#endif
@@ -0,0 +1,40 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#define USING_LOW_ORDER 1
+#define USING_HI_ORDER  0
+
+typedef double* DofToQuad_t @dim(NUM_QUAD, NUM_DOFS);
+typedef double* DofToQuadD2D_t @dim(2, NUM_QUAD, NUM_DOFS);
+typedef double* DofToQuadD3D_t @dim(3, NUM_QUAD, NUM_DOFS);
+
+typedef double* QuadToDof_t @dim(NUM_DOFS, NUM_QUAD);
+typedef double* QuadToDofD2D_t @dim(2, NUM_DOFS, NUM_QUAD);
+typedef double* QuadToDofD3D_t @dim(3, NUM_DOFS, NUM_QUAD);
+
+typedef double* Jacobian2D_t @dim(2, 2, NUM_QUAD, numElements);
+typedef double* Jacobian3D_t @dim(3, 3, NUM_QUAD, numElements);
+
+typedef double* SymmOperator2D_t @dim(3, NUM_QUAD, numElements);
+typedef double* SymmOperator3D_t @dim(6, NUM_QUAD, numElements);
+
+typedef double* DLocal_t @dim(NUM_DOFS, numElements);
+typedef double* QLocal_t @dim(NUM_QUAD, numElements);
+
+#if VDIM_ORDERING == ORDERING_BY_VDIM
+typedef double* DVLocal_t @dim(NUM_VDIM, NUM_DOFS, numElements);
+typedef double* QVLocal_t @dim(NUM_VDIM, NUM_QUAD, numElements);
+#else
+typedef double* DVLocal_t @dim(NUM_VDIM, NUM_DOFS, numElements) @dimOrder(2,0,1);
+typedef double* QVLocal_t @dim(NUM_VDIM, NUM_QUAD, numElements) @dimOrder(2,0,1);
+#endif
+
+typedef int* DLocalMap_t @dim(NUM_DOFS, numElements);
@@ -0,0 +1,85 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#if NUM_QUAD_1D < NUM_DOFS_1D
+#  define NUM_MAX_1D NUM_DOFS_1D
+#else
+#  define NUM_MAX_1D NUM_QUAD_1D
+#endif
+
+#define NUM_MAX_2D (NUM_MAX_1D * NUM_MAX_1D)
+
+#define NUM_QUAD_DOFS_1D (NUM_QUAD_1D * NUM_DOFS_1D)
+
+#define QUAD_2D_ID(X, Y) (X + ((Y) * NUM_QUAD_1D))
+#define DOFS_2D_ID(X, Y) (X + ((Y) * NUM_DOFS_1D))
+
+#define QUAD_3D_ID(X, Y, Z) (X + ((Y) * NUM_QUAD_1D) + ((Z) * NUM_QUAD_2D))
+#define DOFS_3D_ID(X, Y, Z) (X + ((Y) * NUM_DOFS_1D) + ((Z) * NUM_DOFS_2D))
+
+#if NUM_MAX_1D < 8
+#  define USING_LOW_ORDER 1
+#  define USING_HI_ORDER  0
+#else
+#  define USING_LOW_ORDER 0
+#  define USING_HI_ORDER  1
+#endif
+
+#define M1_ELEMENT_BATCHES (M1_OUTER_ELEMENT_BATCH * M1_INNER_ELEMENT_BATCH)
+
+typedef double* DofToQuad_t @dim(NUM_QUAD_1D, NUM_DOFS_1D);
+typedef double* QuadToDof_t @dim(NUM_DOFS_1D, NUM_QUAD_1D);
+
+typedef double* Jacobian_t   @dim(NUM_DIM, NUM_DIM, numElements);
+typedef double* Jacobian1D_t @dim(NUM_QUAD_1D, numElements);
+typedef double* Jacobian2D_t @dim(2, 2, NUM_QUAD_2D, numElements);
+typedef double* Jacobian3D_t @dim(3, 3, NUM_QUAD_3D, numElements);
+
+typedef double* SymmOperator1D_t @dim(NUM_QUAD_1D, numElements);
+typedef double* SymmOperator2D_t @dim(3, NUM_QUAD_2D, numElements);
+typedef double* SymmOperator3D_t @dim(6, NUM_QUAD_3D, numElements);
+
+typedef double* DLocal_t   @dim(NUM_DOFS, numElements);
+typedef double* DLocal1D_t @dim(NUM_DOFS_1D, numElements);
+typedef double* DLocal2D_t @dim(NUM_DOFS_1D, NUM_DOFS_1D, numElements);
+typedef double* DLocal3D_t @dim(NUM_DOFS_1D, NUM_DOFS_1D, NUM_DOFS_1D, numElements);
+
+typedef double* QLocal_t   @dim(NUM_QUAD, numElements);
+typedef double* QLocal1D_t @dim(NUM_QUAD_1D, numElements);
+typedef double* QLocal2D_t @dim(NUM_QUAD_1D, NUM_QUAD_1D, numElements);
+typedef double* QLocal3D_t @dim(NUM_QUAD_1D, NUM_QUAD_1D, NUM_QUAD_1D, numElements);
+
+#if VDIM_ORDERING == ORDERING_BY_VDIM
+typedef double* DVLocal_t   @dim(NUM_VDIM, NUM_DOFS, numElements);
+typedef double* DVLocal1D_t @dim(NUM_VDIM, NUM_DOFS_1D, numElements);
+typedef double* DVLocal2D_t @dim(NUM_VDIM, NUM_DOFS_1D, NUM_DOFS_1D, numElements);
+typedef double* DVLocal3D_t @dim(NUM_VDIM, NUM_DOFS_1D, NUM_DOFS_1D, NUM_DOFS_1D, numElements);
+
+typedef double* QVLocal_t   @dim(NUM_VDIM, NUM_QUAD, numElements);
+typedef double* QVLocal1D_t @dim(NUM_VDIM, NUM_QUAD_1D, numElements);
+typedef double* QVLocal2D_t @dim(NUM_VDIM, NUM_QUAD_1D, NUM_QUAD_1D, numElements);
+typedef double* QVLocal3D_t @dim(NUM_VDIM, NUM_QUAD_1D, NUM_QUAD_1D, NUM_QUAD_1D, numElements);
+#else
+typedef double* DVLocal_t   @dim(NUM_VDIM, NUM_DOFS, numElements) @dimOrder(2,0,1);
+typedef double* DVLocal1D_t @dim(NUM_VDIM, NUM_DOFS_1D, numElements) @dimOrder(2,0,1);
+typedef double* DVLocal2D_t @dim(NUM_VDIM, NUM_DOFS_1D, NUM_DOFS_1D, numElements) @dimOrder(3,0,1,2);
+typedef double* DVLocal3D_t @dim(NUM_VDIM, NUM_DOFS_1D, NUM_DOFS_1D, NUM_DOFS_1D, numElements) @dimOrder(4,0,1,2,3);
+
+typedef double* QVLocal_t   @dim(NUM_VDIM, NUM_QUAD, numElements) @dimOrder(2,0,1);
+typedef double* QVLocal1D_t @dim(NUM_VDIM, NUM_QUAD_1D, numElements) @dimOrder(2,0,1);
+typedef double* QVLocal2D_t @dim(NUM_VDIM, NUM_QUAD_1D, NUM_QUAD_1D, numElements) @dimOrder(3,0,1,2);
+typedef double* QVLocal3D_t @dim(NUM_VDIM, NUM_QUAD_1D, NUM_QUAD_1D, NUM_QUAD_1D, numElements) @dimOrder(4,0,1,2,3);
+#endif
+
+typedef int* DLocalMap_t   @dim(NUM_DOFS, numElements);
+typedef int* DLocalMap1D_t @dim(NUM_DOFS_1D, numElements);
+typedef int* DLocalMap2D_t @dim(NUM_DOFS_1D, NUM_DOFS_1D, numElements);
+typedef int* DLocalMap3D_t @dim(NUM_DOFS_1D, NUM_DOFS_1D, NUM_DOFS_1D, numElements);
@@ -0,0 +1,168 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "mfem-occa://defines.okl"
+
+//---[ 2D ]-----------------------------
+@kernel void Assemble2D(const int numElements,
+                        const double * restrict quadWeights,
+                        const Jacobian2D_t restrict J,
+                        COEFF_ARGS
+                        SymmOperator2D_t restrict oper) {
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int q = 0; q < NUM_QUAD; ++q; @inner) {
+      const double J11 = J(0, 0, q, e), J12 = J(1, 0, q, e);
+      const double J21 = J(0, 1, q, e), J22 = J(1, 1, q, e);
+
+      const double c_detJ = quadWeights[q] * COEFF / ((J11 * J22) - (J21 * J12));
+
+      oper(0, q, e) =  c_detJ * (J12*J12 + J22*J22); // (1,1)
+      oper(1, q, e) = -c_detJ * (J12*J11 + J22*J21); // (1,2) + (2,1)
+      oper(2, q, e) =  c_detJ * (J11*J11 + J21*J21); // (2,2)
+    }
+  }
+}
+
+@kernel void MultAdd2D(const int numElements,
+                       const DofToQuad_t restrict dofToQuad,
+                       const DofToQuadD2D_t restrict dofToQuadD,
+                       const QuadToDof_t restrict quadToDof,
+                       const QuadToDofD2D_t restrict quadToDofD,
+                       const SymmOperator2D_t restrict oper,
+                       const DLocal_t restrict solIn,
+                       DLocal_t restrict solOut) {
+  // Iterate over elements
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int dummy = 0; dummy < 1; ++dummy; @inner) {
+      double r_sol[NUM_DOFS];
+      for (int d = 0; d < NUM_DOFS; ++d) {
+        r_sol[d] = 0;
+      }
+
+      for (int q = 0; q < NUM_QUAD; ++q) {
+        double gradX = 0, gradY = 0;
+        for (int d = 0; d < NUM_DOFS; ++d) {
+          const double s = solIn(d, e);
+          gradX += s * quadToDofD(0, d, q);
+          gradY += s * quadToDofD(1, d, q);
+        }
+        const double O11 = oper(0, q, e);
+        const double O12 = oper(1, q, e);
+        const double O22 = oper(2, q, e);
+
+        const double gradX2 = (O11 * gradX) + (O12 * gradY);
+        const double gradY2 = (O12 * gradX) + (O22 * gradY);
+
+        for (int d = 0; d < NUM_DOFS; ++d) {
+          r_sol[d] += ((gradX2 * quadToDofD(0, d, q)) +
+                       (gradY2 * quadToDofD(1, d, q)));
+        }
+      }
+
+      for (int d = 0; d < NUM_DOFS; ++d) {
+        solOut(d, e) += r_sol[d];
+      }
+    }
+  }
+}
+//======================================
+
+
+//---[ 3D ]-----------------------------
+@kernel void Assemble3D(const int numElements,
+                        const double * restrict quadWeights,
+                        const Jacobian3D_t restrict J,
+                        COEFF_ARGS
+                        SymmOperator3D_t restrict oper) {
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int q = 0; q < NUM_QUAD; ++q; @inner) {
+      const double J11 = J(0, 0, q, e), J12 = J(1, 0, q, e), J13 = J(2, 0, q, e);
+      const double J21 = J(0, 1, q, e), J22 = J(1, 1, q, e), J23 = J(2, 1, q, e);
+      const double J31 = J(0, 2, q, e), J32 = J(1, 2, q, e), J33 = J(2, 2, q, e);
+
+      const double detJ = ((J11 * J22 * J33) + (J12 * J23 * J31) + (J13 * J21 * J32) -
+                           (J13 * J22 * J31) - (J12 * J21 * J33) - (J11 * J23 * J32));
+
+      const double c_detJ = quadWeights[q] * COEFF / detJ;
+
+      // adj(J)
+      const double A11 = (J22 * J33) - (J23 * J32);
+      const double A12 = (J23 * J31) - (J21 * J33);
+      const double A13 = (J21 * J32) - (J22 * J31);
+
+      const double A21 = (J13 * J32) - (J12 * J33);
+      const double A22 = (J11 * J33) - (J13 * J31);
+      const double A23 = (J12 * J31) - (J11 * J32);
+
+      const double A31 = (J12 * J23) - (J13 * J22);
+      const double A32 = (J13 * J21) - (J11 * J23);
+      const double A33 = (J11 * J22) - (J12 * J21);
+
+      // adj(J)^Tadj(J)
+      oper(0, q, e) = c_detJ * (A11*A11 + A21*A21 + A31*A31); // (1,1)
+      oper(1, q, e) = c_detJ * (A11*A12 + A21*A22 + A31*A32); // (1,2) + (2,1)
+      oper(2, q, e) = c_detJ * (A11*A13 + A21*A23 + A31*A33); // (1,3) + (3,1)
+      oper(3, q, e) = c_detJ * (A12*A12 + A22*A22 + A32*A32); // (2,2)
+      oper(4, q, e) = c_detJ * (A12*A13 + A22*A23 + A32*A33); // (2,3) + (3,2)
+      oper(5, q, e) = c_detJ * (A13*A13 + A23*A23 + A33*A33); // (3,3)
+    }
+  }
+}
+
+@kernel void MultAdd3D(const int numElements,
+                       const DofToQuad_t restrict dofToQuad,
+                       const DofToQuadD3D_t restrict dofToQuadD,
+                       const QuadToDof_t restrict quadToDof,
+                       const QuadToDofD3D_t restrict quadToDofD,
+                       const SymmOperator3D_t restrict oper,
+                       const DLocal_t restrict solIn,
+                       DLocal_t restrict solOut) {
+  // Iterate over elements
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int dummy = 0; dummy < 1; ++dummy; @inner) {
+      double r_sol[NUM_DOFS];
+      for (int d = 0; d < NUM_DOFS; ++d) {
+        r_sol[d] = 0;
+      }
+
+      for (int q = 0; q < NUM_QUAD; ++q) {
+        double gradX = 0, gradY = 0, gradZ = 0;
+        for (int d = 0; d < NUM_DOFS; ++d) {
+          const double s = solIn(d, e);
+          gradX += s * quadToDofD(0, d, q);
+          gradY += s * quadToDofD(1, d, q);
+          gradZ += s * quadToDofD(2, d, q);
+        }
+        const double O11 = oper(0, q, e);
+        const double O12 = oper(1, q, e);
+        const double O13 = oper(2, q, e);
+        const double O22 = oper(3, q, e);
+        const double O23 = oper(4, q, e);
+        const double O33 = oper(5, q, e);
+
+        const double gradX2 = (O11 * gradX) + (O12 * gradY) + (O13 * gradZ);
+        const double gradY2 = (O12 * gradX) + (O22 * gradY) + (O23 * gradZ);
+        const double gradZ2 = (O13 * gradX) + (O23 * gradY) + (O33 * gradZ);
+
+        for (int d = 0; d < NUM_DOFS; ++d) {
+          r_sol[d] += ((gradX2 * quadToDofD(0, d, q)) +
+                       (gradY2 * quadToDofD(1, d, q)) +
+                       (gradZ2 * quadToDofD(2, d, q)));
+        }
+      }
+
+      for (int d = 0; d < NUM_DOFS; ++d) {
+        solOut(d, e) += r_sol[d];
+      }
+    }
+  }
+}
+//======================================
@@ -0,0 +1,182 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "mfem-occa://defines.okl"
+
+//---[ 2D ]-----------------------------
+@kernel void Assemble2D(const int numElements,
+                        const double *quadWeights,
+                        const Jacobian2D_t J,
+                        COEFF_ARGS
+                        SymmOperator2D_t oper) {
+  for (int eOff = 0; eOff < numElements; eOff += A2_ELEMENT_BATCH; @outer) {
+    for (int e = eOff; e < (eOff + A2_ELEMENT_BATCH); ++e; @inner) {
+      if (e < numElements) {
+        for (int qOff = 0; qOff < A2_QUAD_BATCH; ++qOff; @inner) {
+          for (int q = qOff; q < NUM_QUAD; q += A2_QUAD_BATCH) {
+            const double J11 = J(0, 0, q, e), J12 = J(1, 0, q, e);
+            const double J21 = J(0, 1, q, e), J22 = J(1, 1, q, e);
+
+            const double c_detJ = quadWeights[q] * COEFF / ((J11 * J22) - (J21 * J12));
+
+            oper(0, q, e) =  c_detJ * (J12*J12 + J22*J22); // (1,1)
+            oper(1, q, e) = -c_detJ * (J12*J11 + J22*J21); // (1,2) + (2,1)
+            oper(2, q, e) =  c_detJ * (J11*J11 + J21*J21); // (2,2)
+          }
+        }
+      }
+    }
+  }
+}
+
+@kernel void MultAdd2D(const int numElements,
+                       const DofToQuad_t restrict dofToQuad,
+                       const DofToQuadD2D_t restrict dofToQuadD,
+                       const QuadToDof_t restrict quadToDof,
+                       const QuadToDofD2D_t restrict quadToDofD,
+                       const SymmOperator2D_t restrict oper,
+                       const DLocal_t restrict solIn,
+                       DLocal_t restrict solOut) {
+  // Iterate over elements
+  for (int e = 0; e < numElements; ++e; @outer) {
+    @shared double s_gradX[NUM_QUAD];
+    @shared double s_gradY[NUM_QUAD];
+
+    for (int qOff = 0; qOff < M2_INNER_BATCH; ++qOff; @inner) {
+      for (int q = qOff; q < NUM_QUAD; q += M2_INNER_BATCH) {
+        double gradX = 0, gradY = 0;
+        for (int d = 0; d < NUM_DOFS; ++d) {
+          const double s = solIn(d, e);
+          gradX += s * quadToDofD(0, d, q);
+          gradY += s * quadToDofD(1, d, q);
+        }
+        const double O11 = oper(0, q, e);
+        const double O12 = oper(1, q, e);
+        const double O22 = oper(2, q, e);
+
+        s_gradX[q] = (O11 * gradX) + (O12 * gradY);
+        s_gradY[q] = (O12 * gradX) + (O22 * gradY);
+      }
+    }
+
+    for (int dOff = 0; dOff < M2_INNER_BATCH; ++dOff) {
+      for (int d = dOff; d < NUM_DOFS; d += M2_INNER_BATCH) {
+        double r_sol = 0;
+        for (int q = 0; q < NUM_QUAD; ++q) {
+          // FIXME: s_gradX and s_gradY are @shared used outside of @inner
+          r_sol += ((s_gradX[q] * quadToDofD(0, d, q)) +
+                    (s_gradY[q] * quadToDofD(1, d, q)));
+        }
+        solOut(d, e) += r_sol;
+      }
+    }
+  }
+}
+//======================================
+
+
+//---[ 3D ]-----------------------------
+@kernel void Assemble3D(const int numElements,
+                        const double *quadWeights,
+                        const Jacobian3D_t J,
+                        COEFF_ARGS
+                        SymmOperator3D_t oper) {
+  for (int eOff = 0; eOff < numElements; eOff += A3_ELEMENT_BATCH; @outer) {
+    for (int e = eOff; e < (eOff + A3_ELEMENT_BATCH); ++e; @inner) {
+      if (e < numElements) {
+        for (int qOff = 0; qOff < A3_QUAD_BATCH; ++qOff; @inner) {
+          for (int q = qOff; q < NUM_QUAD; q += A3_QUAD_BATCH) {
+            const double J11 = J(0, 0, q, e), J12 = J(1, 0, q, e), J13 = J(2, 0, q, e);
+            const double J21 = J(0, 1, q, e), J22 = J(1, 1, q, e), J23 = J(2, 1, q, e);
+            const double J31 = J(0, 2, q, e), J32 = J(1, 2, q, e), J33 = J(2, 2, q, e);
+
+            const double detJ = ((J11 * J22 * J33) + (J12 * J23 * J31) + (J13 * J21 * J32) -
+                                 (J13 * J22 * J31) - (J12 * J21 * J33) - (J11 * J23 * J32));
+
+            const double c_detJ = quadWeights[q] * COEFF / detJ;
+
+            // adj(J)
+            const double A11 = (J22 * J33) - (J23 * J32);
+            const double A12 = (J23 * J31) - (J21 * J33);
+            const double A13 = (J21 * J32) - (J22 * J31);
+
+            const double A21 = (J13 * J32) - (J12 * J33);
+            const double A22 = (J11 * J33) - (J13 * J31);
+            const double A23 = (J12 * J31) - (J11 * J32);
+
+            const double A31 = (J12 * J23) - (J13 * J22);
+            const double A32 = (J13 * J21) - (J11 * J23);
+            const double A33 = (J11 * J22) - (J12 * J21);
+
+            // adj(J)^Tadj(J)
+            oper(0, q, e) = c_detJ * (A11*A11 + A21*A21 + A31*A31); // (1,1)
+            oper(1, q, e) = c_detJ * (A11*A12 + A21*A22 + A31*A32); // (1,2) + (2,1)
+            oper(2, q, e) = c_detJ * (A11*A13 + A21*A23 + A31*A33); // (1,3) + (3,1)
+            oper(3, q, e) = c_detJ * (A12*A12 + A22*A22 + A32*A32); // (2,2)
+            oper(4, q, e) = c_detJ * (A12*A13 + A22*A23 + A32*A33); // (2,3) + (3,2)
+            oper(5, q, e) = c_detJ * (A13*A13 + A23*A23 + A33*A33); // (3,3)
+          }
+        }
+      }
+    }
+  }
+}
+
+@kernel void MultAdd3D(const int numElements,
+                       const DofToQuad_t restrict dofToQuad,
+                       const DofToQuadD3D_t restrict dofToQuadD,
+                       const QuadToDof_t restrict quadToDof,
+                       const QuadToDofD3D_t restrict quadToDofD,
+                       const SymmOperator3D_t restrict oper,
+                       const DLocal_t restrict solIn,
+                       DLocal_t restrict solOut) {
+  // Iterate over elements
+  for (int e = 0; e < numElements; ++e; @outer) {
+    @shared double s_gradX[NUM_QUAD];
+    @shared double s_gradY[NUM_QUAD];
+    @shared double s_gradZ[NUM_QUAD];
+
+    for (int qOff = 0; qOff < M3_INNER_BATCH; ++qOff; @inner) {
+      for (int q = qOff; q < NUM_QUAD; q += M3_INNER_BATCH) {
+        double gradX = 0, gradY = 0, gradZ = 0;
+        for (int d = 0; d < NUM_DOFS; ++d) {
+          const double s = solIn(d, e);
+          gradX += s * quadToDofD(0, d, q);
+          gradY += s * quadToDofD(1, d, q);
+          gradZ += s * quadToDofD(2, d, q);
+        }
+        const double O11 = oper(0, q, e);
+        const double O12 = oper(1, q, e);
+        const double O13 = oper(2, q, e);
+        const double O22 = oper(3, q, e);
+        const double O23 = oper(4, q, e);
+        const double O33 = oper(5, q, e);
+
+        s_gradX[q] = (O11 * gradX) + (O12 * gradY) + (O13 * gradZ);
+        s_gradY[q] = (O12 * gradX) + (O22 * gradY) + (O23 * gradZ);
+        s_gradZ[q] = (O13 * gradX) + (O23 * gradY) + (O33 * gradZ);
+      }
+    }
+
+    for (int dOff = 0; dOff < M3_INNER_BATCH; ++dOff) {
+      for (int d = dOff; d < NUM_DOFS; d += M3_INNER_BATCH) {
+        double r_sol = 0;
+        for (int q = 0; q < NUM_QUAD; ++q) {
+          r_sol += ((s_gradX[q] * quadToDofD(0, d, q)) +
+                    (s_gradY[q] * quadToDofD(1, d, q)) +
+                    (s_gradZ[q] * quadToDofD(2, d, q)));
+        }
+        solOut(d, e) += r_sol;
+      }
+    }
+  }
+}
+//======================================
@@ -0,0 +1,370 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "mfem-occa://defines.okl"
+
+//---[ 1D ]-----------------------------
+@kernel void Assemble1D(const int numElements,
+                        const double * restrict quadWeights,
+                        const Jacobian1D_t restrict J,
+                        COEFF_ARGS
+                        SymmOperator1D_t restrict oper) {
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int q = 0; q < NUM_QUAD_1D; ++q; @inner) {
+      oper(q, e) = quadWeights[q] * COEFF / J(q, e);
+    }
+  }
+}
+
+@kernel void MultAdd1D(const int numElements,
+                       const DofToQuad_t restrict dofToQuad,
+                       const DofToQuad_t restrict dofToQuadD,
+                       const QuadToDof_t restrict quadToDof,
+                       const QuadToDof_t restrict quadToDofD,
+                       const SymmOperator1D_t restrict oper,
+                       const DLocal1D_t restrict solIn,
+                       DLocal1D_t restrict solOut) {
+  // Iterate over elements
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int dummy = 0; dummy < 1; ++dummy; @inner) {
+      double grad[NUM_QUAD_1D];
+      for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+        grad[qx] = 0;
+      }
+      for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+        const double s = solIn(dx, e);
+        for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+          grad[qx] += s * dofToQuadD(qx, dx);
+        }
+      }
+      for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+        grad[qx] *= oper(qx, e);
+      }
+      for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+        const double gradX = grad[qx];
+        for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+          solOut(dx, e) += gradX * quadToDofD(dx, qx);
+        }
+      }
+    }
+  }
+}
+//======================================
+
+
+//---[ 2D ]-----------------------------
+@kernel void Assemble2D(const int numElements,
+                        const double * restrict quadWeights,
+                        const Jacobian2D_t restrict J,
+                        COEFF_ARGS
+                        SymmOperator2D_t restrict oper) {
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int q = 0; q < NUM_QUAD_2D; ++q; @inner) {
+      const double J11 = J(0, 0, q, e), J12 = J(1, 0, q, e);
+      const double J21 = J(0, 1, q, e), J22 = J(1, 1, q, e);
+
+      const double c_detJ = quadWeights[q] * COEFF / ((J11 * J22) - (J21 * J12));
+
+      oper(0, q, e) =  c_detJ * (J21*J21 + J22*J22); // (1,1)
+      oper(1, q, e) = -c_detJ * (J21*J11 + J22*J12); // (1,2), (2,1)
+      oper(2, q, e) =  c_detJ * (J11*J11 + J12*J12); // (2,2)
+    }
+  }
+}
+
+@kernel void MultAdd2D(const int numElements,
+                       const DofToQuad_t restrict dofToQuad,
+                       const DofToQuad_t restrict dofToQuadD,
+                       const QuadToDof_t restrict quadToDof,
+                       const QuadToDof_t restrict quadToDofD,
+                       const SymmOperator2D_t restrict oper,
+                       const DLocal2D_t restrict solIn,
+                       DLocal2D_t restrict solOut) {
+  // Iterate over elements
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int dummy = 0; dummy < 1; ++dummy; @inner) {
+      double grad[NUM_QUAD_1D][NUM_QUAD_1D][2];
+      for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+        for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+          grad[qy][qx][0] = 0;
+          grad[qy][qx][1] = 0;
+        }
+      }
+
+      for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+        double gradX[NUM_QUAD_1D][2];
+        for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+          gradX[qx][0] = 0;
+          gradX[qx][1] = 0;
+        }
+
+        for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+          const double s = solIn(dx, dy, e);
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            gradX[qx][0] += s * dofToQuad(qx, dx);
+            gradX[qx][1] += s * dofToQuadD(qx, dx);
+          }
+        }
+
+        for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+          const double wy  = dofToQuad(qy, dy);
+          const double wDy = dofToQuadD(qy, dy);
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            grad[qy][qx][0] += gradX[qx][1] * wy;
+            grad[qy][qx][1] += gradX[qx][0] * wDy;
+          }
+        }
+      }
+
+      // Calculate Dxy, xDy in plane
+      for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+        for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+          const int q = QUAD_2D_ID(qx, qy);
+          const double O11 = oper(0, q, e);
+          const double O12 = oper(1, q, e);
+          const double O22 = oper(2, q, e);
+
+          const double gradX = grad[qy][qx][0];
+          const double gradY = grad[qy][qx][1];
+
+          grad[qy][qx][0] = (O11 * gradX) + (O12 * gradY);
+          grad[qy][qx][1] = (O12 * gradX) + (O22 * gradY);
+        }
+      }
+
+      for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+        double gradX[NUM_DOFS_1D][2];
+        for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+          gradX[dx][0] = 0;
+          gradX[dx][1] = 0;
+        }
+
+        for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+          const double gX = grad[qy][qx][0];
+          const double gY = grad[qy][qx][1];
+          for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+            const double wx  = quadToDof(dx, qx);
+            const double wDx = quadToDofD(dx, qx);
+            gradX[dx][0] += gX * wDx;
+            gradX[dx][1] += gY * wx;
+          }
+        }
+
+        for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+          const double wy  = quadToDof(dy, qy);
+          const double wDy = quadToDofD(dy, qy);
+          for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+            solOut(dx, dy, e) += ((gradX[dx][0] * wy) +
+                                  (gradX[dx][1] * wDy));
+          }
+        }
+      }
+    }
+  }
+}
+//======================================
+
+
+//---[ 3D ]-----------------------------
+@kernel void Assemble3D(const int numElements,
+                        const double * restrict quadWeights,
+                        const Jacobian3D_t restrict J,
+                        COEFF_ARGS
+                        SymmOperator3D_t restrict oper) {
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int q = 0; q < NUM_QUAD_3D; ++q; @inner) {
+      const double J11 = J(0, 0, q, e), J12 = J(1, 0, q, e), J13 = J(2, 0, q, e);
+      const double J21 = J(0, 1, q, e), J22 = J(1, 1, q, e), J23 = J(2, 1, q, e);
+      const double J31 = J(0, 2, q, e), J32 = J(1, 2, q, e), J33 = J(2, 2, q, e);
+
+      const double detJ = ((J11 * J22 * J33) + (J12 * J23 * J31) + (J13 * J21 * J32) -
+                           (J13 * J22 * J31) - (J12 * J21 * J33) - (J11 * J23 * J32));
+
+      const double c_detJ = quadWeights[q] * COEFF / detJ;
+
+      // adj(J)
+      const double A11 = (J22 * J33) - (J23 * J32);
+      const double A12 = (J23 * J31) - (J21 * J33);
+      const double A13 = (J21 * J32) - (J22 * J31);
+
+      const double A21 = (J13 * J32) - (J12 * J33);
+      const double A22 = (J11 * J33) - (J13 * J31);
+      const double A23 = (J12 * J31) - (J11 * J32);
+
+      const double A31 = (J12 * J23) - (J13 * J22);
+      const double A32 = (J13 * J21) - (J11 * J23);
+      const double A33 = (J11 * J22) - (J12 * J21);
+
+      // adj(J)^Tadj(J)
+      oper(0, q, e) = c_detJ * (A11*A11 + A21*A21 + A31*A31); // (1,1)
+      oper(1, q, e) = c_detJ * (A11*A12 + A21*A22 + A31*A32); // (1,2), (2,1)
+      oper(2, q, e) = c_detJ * (A11*A13 + A21*A23 + A31*A33); // (1,3), (3,1)
+      oper(3, q, e) = c_detJ * (A12*A12 + A22*A22 + A32*A32); // (2,2)
+      oper(4, q, e) = c_detJ * (A12*A13 + A22*A23 + A32*A33); // (2,3), (3,2)
+      oper(5, q, e) = c_detJ * (A13*A13 + A23*A23 + A33*A33); // (3,3)
+    }
+  }
+}
+
+@kernel void MultAdd3D(const int numElements,
+                       const DofToQuad_t restrict dofToQuad,
+                       const DofToQuad_t restrict dofToQuadD,
+                       const QuadToDof_t restrict quadToDof,
+                       const QuadToDof_t restrict quadToDofD,
+                       const SymmOperator3D_t restrict oper,
+                       const DLocal3D_t restrict solIn,
+                       DLocal3D_t restrict solOut) {
+  // Iterate over elements
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int dummy = 0; dummy < 1; ++dummy; @inner) {
+      double grad[NUM_QUAD_1D][NUM_QUAD_1D][NUM_QUAD_1D][4];
+      for (int qz = 0; qz < NUM_QUAD_1D; ++qz) {
+        for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            grad[qz][qy][qx][0] = 0;
+            grad[qz][qy][qx][1] = 0;
+            grad[qz][qy][qx][2] = 0;
+          }
+        }
+      }
+
+      for (int dz = 0; dz < NUM_DOFS_1D; ++dz) {
+        double gradXY[NUM_QUAD_1D][NUM_QUAD_1D][4];
+        for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            gradXY[qy][qx][0] = 0;
+            gradXY[qy][qx][1] = 0;
+            gradXY[qy][qx][2] = 0;
+          }
+        }
+
+        for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+          double gradX[NUM_QUAD_1D][2];
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            gradX[qx][0] = 0;
+            gradX[qx][1] = 0;
+          }
+
+          for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+            const double s = solIn(dx, dy, dz, e);
+            for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+              gradX[qx][0] += s * dofToQuad(qx, dx);
+              gradX[qx][1] += s * dofToQuadD(qx, dx);
+            }
+          }
+
+          for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+            const double wy  = dofToQuad(qy, dy);
+            const double wDy = dofToQuadD(qy, dy);
+            for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+              const double wx  = gradX[qx][0];
+              const double wDx = gradX[qx][1];
+              gradXY[qy][qx][0] += wDx * wy;
+              gradXY[qy][qx][1] += wx  * wDy;
+              gradXY[qy][qx][2] += wx  * wy;
+            }
+          }
+        }
+
+        for (int qz = 0; qz < NUM_QUAD_1D; ++qz) {
+          const double wz  = dofToQuad(qz, dz);
+          const double wDz = dofToQuadD(qz, dz);
+          for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+            for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+              grad[qz][qy][qx][0] += gradXY[qy][qx][0] * wz;
+              grad[qz][qy][qx][1] += gradXY[qy][qx][1] * wz;
+              grad[qz][qy][qx][2] += gradXY[qy][qx][2] * wDz;
+            }
+          }
+        }
+      }
+
+      // Calculate Dxyz, xDyz, xyDz in plane
+      for (int qz = 0; qz < NUM_QUAD_1D; ++qz) {
+        for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            const int q = QUAD_3D_ID(qx, qy, qz);
+            const double O11 = oper(0, q, e);
+            const double O12 = oper(1, q, e);
+            const double O13 = oper(2, q, e);
+            const double O22 = oper(3, q, e);
+            const double O23 = oper(4, q, e);
+            const double O33 = oper(5, q, e);
+
+            const double gradX = grad[qz][qy][qx][0];
+            const double gradY = grad[qz][qy][qx][1];
+            const double gradZ = grad[qz][qy][qx][2];
+
+            grad[qz][qy][qx][0] = (O11 * gradX) + (O12 * gradY) + (O13 * gradZ);
+            grad[qz][qy][qx][1] = (O12 * gradX) + (O22 * gradY) + (O23 * gradZ);
+            grad[qz][qy][qx][2] = (O13 * gradX) + (O23 * gradY) + (O33 * gradZ);
+          }
+        }
+      }
+
+      for (int qz = 0; qz < NUM_QUAD_1D; ++qz) {
+        double gradXY[NUM_DOFS_1D][NUM_DOFS_1D][4];
+        for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+          for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+            gradXY[dy][dx][0] = 0;
+            gradXY[dy][dx][1] = 0;
+            gradXY[dy][dx][2] = 0;
+          }
+        }
+
+        for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+          double gradX[NUM_DOFS_1D][4];
+          for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+            gradX[dx][0] = 0;
+            gradX[dx][1] = 0;
+            gradX[dx][2] = 0;
+          }
+
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            const double gX = grad[qz][qy][qx][0];
+            const double gY = grad[qz][qy][qx][1];
+            const double gZ = grad[qz][qy][qx][2];
+            for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+              const double wx  = quadToDof(dx, qx);
+              const double wDx = quadToDofD(dx, qx);
+              gradX[dx][0] += gX * wDx;
+              gradX[dx][1] += gY * wx;
+              gradX[dx][2] += gZ * wx;
+            }
+          }
+
+          for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+            const double wy  = quadToDof(dy, qy);
+            const double wDy = quadToDofD(dy, qy);
+            for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+              gradXY[dy][dx][0] += gradX[dx][0] * wy;
+              gradXY[dy][dx][1] += gradX[dx][1] * wDy;
+              gradXY[dy][dx][2] += gradX[dx][2] * wy;
+            }
+          }
+        }
+
+        for (int dz = 0; dz < NUM_DOFS_1D; ++dz) {
+          const double wz  = quadToDof(dz, qz);
+          const double wDz = quadToDofD(dz, qz);
+          for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+            for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+              solOut(dx, dy, dz, e) += ((gradXY[dy][dx][0] * wz) +
+                                        (gradXY[dy][dx][1] * wz) +
+                                        (gradXY[dy][dx][2] * wDz));
+            }
+          }
+        }
+      }
+    }
+  }
+}
+//======================================
@@ -0,0 +1,433 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "mfem-occa://defines.okl"
+
+//---[ 1D ]-----------------------------
+@kernel void Assemble1D(const int numElements,
+                        const double *quadWeights,
+                        const Jacobian1D_t J,
+                        COEFF_ARGS
+                        SymmOperator1D_t oper) {
+  for (int eOff = 0; eOff < numElements; eOff += A1_ELEMENT_BATCH; @outer) {
+    for (int e = eOff; e < (eOff + A1_ELEMENT_BATCH); ++e; @inner) {
+      if (e < numElements) {
+        for (int q = 0; q < NUM_QUAD_1D; ++q; @inner) {
+          oper(q, e) = quadWeights[q] * COEFF / J(q, e);
+        }
+      }
+    }
+  }
+}
+
+@kernel void MultAdd1D(const int numElements,
+                       const DofToQuad_t restrict dofToQuad,
+                       const DofToQuad_t restrict dofToQuadD,
+                       const QuadToDof_t restrict quadToDof,
+                       const QuadToDof_t restrict quadToDofD,
+                       const SymmOperator1D_t restrict oper,
+                       const DLocal1D_t restrict solIn,
+                       DLocal1D_t restrict solOut) {
+  // Iterate over elements
+  for (int eOff = 0; eOff < numElements; eOff += M1_ELEMENT_BATCHES; @outer) {
+    @shared double s_dofToQuadD[NUM_QUAD_DOFS_1D] @dim(NUM_QUAD_1D, NUM_DOFS_1D);
+    @shared double s_quadToDofD[NUM_QUAD_DOFS_1D] @dim(NUM_DOFS_1D, NUM_QUAD_1D);
+
+    @exclusive double grad[NUM_QUAD_1D];
+
+    for (int el = 0; el < M1_INNER_ELEMENT_BATCH; ++el; @inner) {
+      for (int i = el; i < NUM_QUAD_DOFS_1D; i += M1_INNER_ELEMENT_BATCH) {
+        s_dofToQuadD[i] = dofToQuadD[i];
+        s_quadToDofD[i] = quadToDofD[i];
+      }
+    }
+
+    for (int b = 0; b < M1_OUTER_ELEMENT_BATCH; ++b) {
+      for (int el = 0; el < M1_INNER_ELEMENT_BATCH; ++el; @inner) {
+        const int e = eOff + b*M1_INNER_ELEMENT_BATCH + el;
+        if (e < numElements) {
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            grad[qx] = 0;
+          }
+          for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+            const double s = solIn(dx, e);
+            for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+              grad[qx] += s * s_dofToQuadD(qx, dx);
+            }
+          }
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            grad[qx] *= oper(qx, e);
+          }
+          for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+            double s = 0;
+            for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+              s += grad[qx] * s_quadToDofD(dx, qx);
+            }
+            solOut(dx, e) += s;
+          }
+        }
+      }
+    }
+  }
+}
+//======================================
+
+
+//---[ 2D ]-----------------------------
+@kernel void Assemble2D(const int numElements,
+                        const double *quadWeights,
+                        const Jacobian2D_t J,
+                        COEFF_ARGS
+                        SymmOperator2D_t oper) {
+  for (int eOff = 0; eOff < numElements; eOff += A2_ELEMENT_BATCH; @outer) {
+    for (int e = eOff; e < (eOff + A2_ELEMENT_BATCH); ++e; @inner) {
+      if (e < numElements) {
+        for (int qOff = 0; qOff < A2_QUAD_BATCH; ++qOff; @inner) {
+          for (int q = qOff; q < NUM_QUAD_2D; q += A2_QUAD_BATCH) {
+            const double J11 = J(0, 0, q, e), J12 = J(1, 0, q, e);
+            const double J21 = J(0, 1, q, e), J22 = J(1, 1, q, e);
+
+            const double c_detJ = quadWeights[q] * COEFF / ((J11 * J22) - (J21 * J12));
+
+            oper(0, q, e) =  c_detJ * (J21*J21 + J22*J22); // (1,1)
+            oper(1, q, e) = -c_detJ * (J21*J11 + J22*J12); // (1,2), (2,1)
+            oper(2, q, e) =  c_detJ * (J11*J11 + J12*J12); // (2,2)
+          }
+        }
+      }
+    }
+  }
+}
+
+@kernel void MultAdd2D(const int numElements,
+                       const DofToQuad_t restrict dofToQuad,
+                       const DofToQuad_t restrict dofToQuadD,
+                       const QuadToDof_t restrict quadToDof,
+                       const QuadToDof_t restrict quadToDofD,
+                       const SymmOperator2D_t restrict oper,
+                       const DLocal2D_t restrict solIn,
+                       DLocal2D_t restrict solOut) {
+  // Iterate over elements
+  for (int eOff = 0; eOff < numElements; eOff += M2_ELEMENT_BATCH; @outer) {
+    // Store dof <--> quad mappings
+    @shared double s_dofToQuad[NUM_QUAD_DOFS_1D] @dim(NUM_QUAD_1D, NUM_DOFS_1D);
+    @shared double s_dofToQuadD[NUM_QUAD_DOFS_1D] @dim(NUM_QUAD_1D, NUM_DOFS_1D);
+    @shared double s_quadToDof[NUM_QUAD_DOFS_1D] @dim(NUM_DOFS_1D, NUM_QUAD_1D);
+    @shared double s_quadToDofD[NUM_QUAD_DOFS_1D] @dim(NUM_DOFS_1D, NUM_QUAD_1D);
+
+    // Store xy planes in shared memory
+    @shared double s_xy[NUM_QUAD_DOFS_1D] @dim(NUM_DOFS_1D, NUM_QUAD_1D);
+    @shared double s_xDy[NUM_QUAD_DOFS_1D] @dim(NUM_DOFS_1D, NUM_QUAD_1D);
+    @shared double s_grad[2 * NUM_QUAD_2D] @dim(2, NUM_QUAD_1D, NUM_QUAD_1D);
+
+    @exclusive double r_x[NUM_MAX_1D];
+    @exclusive double r_y[NUM_QUAD_1D];
+
+    for (int x = 0; x < NUM_MAX_1D; ++x; @inner) {
+      for (int id = x; id < NUM_QUAD_DOFS_1D; id += NUM_MAX_1D) {
+        s_dofToQuad[id]  = dofToQuad[id];
+        s_dofToQuadD[id] = dofToQuadD[id];
+        s_quadToDof[id]  = quadToDof[id];
+        s_quadToDofD[id] = quadToDofD[id];
+      }
+    }
+
+    for (int e = eOff; e < (eOff + M2_ELEMENT_BATCH); ++e) {
+      if (e < numElements) {
+        for (int dx = 0; dx < NUM_DOFS_1D; ++dx; @inner) {
+          if (dx < NUM_DOFS_1D) {
+            for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+              s_xy(dx, qy) = 0;
+              s_xDy(dx, qy) = 0;
+            }
+            for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+              r_x[dy] = solIn(dx, dy, e);
+            }
+            for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+              double xy = 0;
+              double xDy = 0;
+              for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+                xy  += r_x[dy] * s_dofToQuad(qy, dy);
+                xDy += r_x[dy] * s_dofToQuadD(qy, dy);
+              }
+              s_xy(dx, qy)  = xy;
+              s_xDy(dx, qy) = xDy;
+            }
+          }
+        }
+        for (int qy = 0; qy < NUM_MAX_1D; ++qy; @inner) {
+          if (qy < NUM_QUAD_1D) {
+            for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+              double gradX = 0, gradY = 0;
+              for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+                gradX += s_xy(dx, qy)  * s_dofToQuadD(qx, dx);
+                gradY += s_xDy(dx, qy) * s_dofToQuad(qx, dx);
+              }
+
+              const int q = QUAD_2D_ID(qx, qy);
+              const double O11 = oper(0, q, e);
+              const double O12 = oper(1, q, e);
+              const double O22 = oper(2, q, e);
+
+              s_grad(0, qx, qy) = (O11 * gradX) + (O12 * gradY);
+              s_grad(1, qx, qy) = (O12 * gradX) + (O22 * gradY);
+            }
+          }
+        }
+
+        for (int qx = 0; qx < NUM_QUAD_1D; ++qx; @inner) {
+          if (qx < NUM_QUAD_1D) {
+            for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+              s_xy(dy, qx) = 0;
+              s_xDy(dy, qx) = 0;
+            }
+            for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+              r_x[qy] = s_grad(0, qx, qy);
+              r_y[qy] = s_grad(1, qx, qy);
+            }
+            for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+              double xy  = 0;
+              double xDy = 0;
+              for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+                xy  += r_x[qy] * s_quadToDof(dy, qy);
+                xDy += r_y[qy] * s_quadToDofD(dy, qy);
+              }
+              s_xy(dy, qx) = xy;
+              s_xDy(dy, qx) = xDy;
+            }
+          }
+        }
+        for (int dx = 0; dx < NUM_DOFS_1D; ++dx; @inner) {
+          if (dx < NUM_DOFS_1D) {
+            for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+              double s = 0;
+              for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+                s += ((s_xy(dy, qx) * s_quadToDofD(dx, qx)) +
+                      (s_xDy(dy, qx) * s_quadToDof(dx, qx)));
+              }
+              solOut(dx, dy, e) += s;
+            }
+          }
+        }
+      }
+    }
+  }
+}
+//======================================
+
+
+//---[ 3D ]-----------------------------
+@kernel void Assemble3D(const int numElements,
+                        const double *quadWeights,
+                        const Jacobian3D_t J,
+                        COEFF_ARGS
+                        SymmOperator3D_t oper) {
+  for (int eOff = 0; eOff < numElements; eOff += A3_ELEMENT_BATCH; @outer) {
+    for (int e = eOff; e < (eOff + A3_ELEMENT_BATCH); ++e; @inner) {
+      if (e < numElements) {
+        for (int qOff = 0; qOff < A3_QUAD_BATCH; ++qOff; @inner) {
+          for (int q = qOff; q < NUM_QUAD_3D; q += A3_QUAD_BATCH) {
+            const double J11 = J(0, 0, q, e), J12 = J(1, 0, q, e), J13 = J(2, 0, q, e);
+            const double J21 = J(0, 1, q, e), J22 = J(1, 1, q, e), J23 = J(2, 1, q, e);
+            const double J31 = J(0, 2, q, e), J32 = J(1, 2, q, e), J33 = J(2, 2, q, e);
+
+            const double detJ = ((J11 * J22 * J33) + (J12 * J23 * J31) + (J13 * J21 * J32) -
+                                 (J13 * J22 * J31) - (J12 * J21 * J33) - (J11 * J23 * J32));
+
+            const double c_detJ = quadWeights[q] * COEFF / detJ;
+
+            // adj(J)
+            const double A11 = (J22 * J33) - (J23 * J32);
+            const double A12 = (J23 * J31) - (J21 * J33);
+            const double A13 = (J21 * J32) - (J22 * J31);
+
+            const double A21 = (J13 * J32) - (J12 * J33);
+            const double A22 = (J11 * J33) - (J13 * J31);
+            const double A23 = (J12 * J31) - (J11 * J32);
+
+            const double A31 = (J12 * J23) - (J13 * J22);
+            const double A32 = (J13 * J21) - (J11 * J23);
+            const double A33 = (J11 * J22) - (J12 * J21);
+
+            // adj(J)^Tadj(J)
+            oper(0, q, e) = c_detJ * (A11*A11 + A21*A21 + A31*A31); // (1,1)
+            oper(1, q, e) = c_detJ * (A11*A12 + A21*A22 + A31*A32); // (1,2), (2,1)
+            oper(2, q, e) = c_detJ * (A11*A13 + A21*A23 + A31*A33); // (1,3), (3,1)
+            oper(3, q, e) = c_detJ * (A12*A12 + A22*A22 + A32*A32); // (2,2)
+            oper(4, q, e) = c_detJ * (A12*A13 + A22*A23 + A32*A33); // (2,3), (3,2)
+            oper(5, q, e) = c_detJ * (A13*A13 + A23*A23 + A33*A33); // (3,3)
+          }
+        }
+      }
+    }
+  }
+}
+
+@kernel void MultAdd3D(const int numElements,
+                       const DofToQuad_t restrict dofToQuad,
+                       const DofToQuad_t restrict dofToQuadD,
+                       const QuadToDof_t restrict quadToDof,
+                       const QuadToDof_t restrict quadToDofD,
+                       const SymmOperator3D_t restrict oper,
+                       const DLocal3D_t restrict solIn,
+                       DLocal3D_t restrict solOut) {
+  // Iterate over elements
+  for (int e = 0; e < numElements; ++e; @outer) {
+    // Store dof <--> quad mappings
+    @shared double s_dofToQuad[NUM_QUAD_DOFS_1D] @dim(NUM_QUAD_1D, NUM_DOFS_1D);
+    @shared double s_dofToQuadD[NUM_QUAD_DOFS_1D] @dim(NUM_QUAD_1D, NUM_DOFS_1D);
+    @shared double s_quadToDof[NUM_QUAD_DOFS_1D] @dim(NUM_DOFS_1D, NUM_QUAD_1D);
+    @shared double s_quadToDofD[NUM_QUAD_DOFS_1D] @dim(NUM_DOFS_1D, NUM_QUAD_1D);
+
+    // Store xy planes in shared memory
+    @shared double s_z[NUM_MAX_2D] @dim(NUM_MAX_1D, NUM_MAX_1D);
+    @shared double s_Dz[NUM_MAX_2D] @dim(NUM_MAX_1D, NUM_MAX_1D);
+    @shared double s_xyDz[NUM_QUAD_2D] @dim(NUM_QUAD_1D, NUM_QUAD_1D);
+
+    // Store z axis as registers
+    @exclusive double r_qz[NUM_QUAD_1D];
+    @exclusive double r_qDz[NUM_QUAD_1D];
+    @exclusive double r_dDxyz[NUM_DOFS_1D];
+    @exclusive double r_dxDyz[NUM_DOFS_1D];
+    @exclusive double r_dxyDz[NUM_DOFS_1D];
+
+    for (int y = 0; y < NUM_MAX_1D; ++y; @inner) {
+      for (int x = 0; x < NUM_MAX_1D; ++x; @inner) {
+        const int id = (y * NUM_MAX_1D) + x;
+        // Fetch Q <--> D maps
+        if (id < NUM_QUAD_DOFS_1D) {
+          s_dofToQuad[id]  = dofToQuad[id];
+          s_dofToQuadD[id] = dofToQuadD[id];
+          s_quadToDof[id]  = quadToDof[id];
+          s_quadToDofD[id] = quadToDofD[id];
+        }
+        // Initialize our Z axis
+        for (int qz = 0; qz < NUM_QUAD_1D; ++qz) {
+          r_qz[qz] = 0;
+          r_qDz[qz] = 0;
+        }
+        // Initialize our solution updates in the Z axis
+        for (int dz = 0; dz < NUM_DOFS_1D; ++dz) {
+          r_dDxyz[dz] = 0;
+          r_dxDyz[dz] = 0;
+          r_dxyDz[dz] = 0;
+        }
+      }
+    }
+
+    for (int dy = 0; dy < NUM_MAX_1D; ++dy; @inner) {
+      for (int dx = 0; dx < NUM_MAX_1D; ++dx; @inner) {
+        if ((dx < NUM_DOFS_1D) && (dy < NUM_DOFS_1D)) {
+          for (int dz = 0; dz < NUM_DOFS_1D; ++dz) {
+            const double s = solIn(dx, dy, dz, e);
+            // Calculate D -> Q in the Z axis
+            for (int qz = 0; qz < NUM_QUAD_1D; ++qz) {
+              r_qz[qz]  += s * s_dofToQuad(qz, dz);
+              r_qDz[qz] += s * s_dofToQuadD(qz, dz);
+            }
+          }
+        }
+      }
+    }
+    // For each xy plane
+    for (int qz = 0; qz < NUM_QUAD_1D; ++qz) {
+      // Fill xy plane at given z position
+      for (int dy = 0; dy < NUM_MAX_1D; ++dy; @inner) {
+        for (int dx = 0; dx < NUM_MAX_1D; ++dx; @inner) {
+          if ((dx < NUM_DOFS_1D) && (dy < NUM_DOFS_1D)) {
+            s_z(dx, dy)  = r_qz[qz];
+            s_Dz(dx, dy) = r_qDz[qz];
+          }
+        }
+      }
+      // Calculate Dxyz, xDyz, xyDz in plane
+      for (int qy = 0; qy < NUM_MAX_1D; ++qy; @inner) {
+        for (int qx = 0; qx < NUM_MAX_1D; ++qx; @inner) {
+          if ((qx < NUM_QUAD_1D) && (qy < NUM_QUAD_1D)) {
+            double Dxyz = 0;
+            double xDyz = 0;
+            double xyDz = 0;
+            for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+              const double wy  = s_dofToQuad(qy, dy);
+              const double wDy = s_dofToQuadD(qy, dy);
+              for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+                const double wx  = s_dofToQuad(qx, dx);
+                const double wDx = s_dofToQuadD(qx, dx);
+                const double z  = s_z(dx, dy);
+                const double Dz = s_Dz(dx, dy);
+                Dxyz += wDx * wy  * z;
+                xDyz += wx  * wDy * z;
+                xyDz += wx  * wy  * Dz;
+              }
+            }
+
+            const int q = QUAD_3D_ID(qx, qy, qz);
+            const double O11 = oper(0, q, e);
+            const double O12 = oper(1, q, e);
+            const double O13 = oper(2, q, e);
+            const double O22 = oper(3, q, e);
+            const double O23 = oper(4, q, e);
+            const double O33 = oper(5, q, e);
+
+            const double qDxyz = (O11 * Dxyz) + (O12 * xDyz) + (O13 * xyDz);
+            const double qxDyz = (O12 * Dxyz) + (O22 * xDyz) + (O23 * xyDz);
+            const double qxyDz = (O13 * Dxyz) + (O23 * xDyz) + (O33 * xyDz);
+
+            for (int dz = 0; dz < NUM_DOFS_1D; ++dz) {
+              const double wz  = s_quadToDof(dz, qz);
+              const double wDz = s_quadToDofD(dz, qz);
+              r_dDxyz[dz] += wz  * qDxyz;
+              r_dxDyz[dz] += wz  * qxDyz;
+              r_dxyDz[dz] += wDz * qxyDz;
+            }
+          }
+        }
+      }
+    }
+    // Iterate over xy planes to compute solution
+    for (int dz = 0; dz < NUM_DOFS_1D; ++dz) {
+      // Place xy plane in shared memory
+      for (int qy = 0; qy < NUM_MAX_1D; ++qy; @inner) {
+        for (int qx = 0; qx < NUM_MAX_1D; ++qx; @inner) {
+          if ((qx < NUM_QUAD_1D) && (qy < NUM_QUAD_1D)) {
+            s_z(qx, qy)    = r_dDxyz[dz];
+            s_Dz(qx, qy)   = r_dxDyz[dz];
+            s_xyDz(qx, qy) = r_dxyDz[dz];
+          }
+        }
+      }
+      // Finalize solution in xy plane
+      for (int dy = 0; dy < NUM_MAX_1D; ++dy; @inner) {
+        for (int dx = 0; dx < NUM_MAX_1D; ++dx; @inner) {
+          if ((dx < NUM_DOFS_1D) && (dy < NUM_DOFS_1D)) {
+            double solZ = 0;
+            for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+              const double wy  = s_quadToDof(dy, qy);
+              const double wDy = s_quadToDofD(dy, qy);
+              for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+                const double wx  = s_quadToDof(dx, qx);
+                const double wDx = s_quadToDofD(dx, qx);
+                const double Dxyz = s_z(qx, qy);
+                const double xDyz = s_Dz(qx, qy);
+                const double xyDz = s_xyDz(qx, qy);
+                solZ += ((wDx * wy  * Dxyz) +
+                         (wx  * wDy * xDyz) +
+                         (wx  * wy  * xyDz));
+              }
+            }
+            solOut(dx, dy, dz, e) += solZ;
+          }
+        }
+      }
+    }
+  }
+}
+//======================================
@@ -0,0 +1,140 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "backend.hpp"
+#include "url_handler.hpp"
+#include "bilinearform.hpp"
+#include "../../general/array.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+bool Engine::fileOpenerRegistered = false;
+
+void Engine::Init(const std::string &engine_spec)
+{
+   //
+   // Initialize inherited fields
+   //
+   memory_resources[0] = NULL;
+   workers_weights[0]= 1.0;
+   workers_mem_res[0] = 0;
+
+   //
+   // Initialize the OCCA engine
+   //
+   ::occa::properties props(engine_spec);
+   device = new ::occa::device[1];
+   device[0].setup(props);
+
+   okl_path = "mfem-occa://";
+   // okl_defines = "...";
+   if (!fileOpenerRegistered)
+   {
+      // The directories from "MFEM_OCCA_OKL_PATH", if any, have the highest
+      // priority.
+      FileOpener *fo = new FileOpener("mfem-occa://", "MFEM_OCCA_OKL_PATH");
+      // Next in priority is the source path, if it exists.
+      std::string mfem_src_prefix = mfem::GetSourcePath();
+      fo->AddDir(mfem_src_prefix + "/backends/occa");
+      // And last in priority is the install path, if it exists.
+      std::string mfem_install_prefix = mfem::GetInstallPath();
+      fo->AddDir(mfem_install_prefix + "/lib/mfem/occa");
+      ::occa::io::fileOpener::add(fo);
+      fileOpenerRegistered = true;
+   }
+}
+
+Engine::Engine(const std::string &engine_spec)
+   : mfem::Engine(NULL, 1, 1)
+{
+   Init(engine_spec);
+}
+
+#ifdef MFEM_USE_MPI
+Engine::Engine(MPI_Comm _comm, const std::string &engine_spec)
+   : mfem::Engine(NULL, 1, 1)
+{
+   comm = _comm;
+   Init(engine_spec);
+}
+#endif
+
+DLayout Engine::MakeLayout(std::size_t size) const
+{
+   return DLayout(new Layout(*this, size));
+}
+
+DLayout Engine::MakeLayout(const mfem::Array<std::size_t> &offsets) const
+{
+   MFEM_ASSERT(offsets.Size() == 2,
+               "multiple workers are not supported yet");
+   return DLayout(new Layout(*this, offsets.Last()));
+}
+
+DArray Engine::MakeArray(PLayout &layout, std::size_t item_size) const
+{
+   MFEM_ASSERT(dynamic_cast<Layout *>(&layout) != NULL,
+               "invalid input layout");
+   Layout *lt = static_cast<Layout *>(&layout);
+   return DArray(new Array(*lt, item_size));
+}
+
+DVector Engine::MakeVector(PLayout &layout, int type_id) const
+{
+   MFEM_ASSERT(type_id == ScalarId<double>::value, "invalid type_id");
+   MFEM_ASSERT(dynamic_cast<Layout *>(&layout) != NULL,
+               "invalid input layout");
+   Layout *lt = static_cast<Layout *>(&layout);
+   return DVector(new Vector(*lt));
+}
+
+DFiniteElementSpace Engine::MakeFESpace(mfem::FiniteElementSpace &fespace) const
+{
+   return DFiniteElementSpace(new FiniteElementSpace(*this, fespace));
+}
+
+DBilinearForm Engine::MakeBilinearForm(mfem::BilinearForm &bf) const
+{
+   return DBilinearForm(new BilinearForm(*this, bf));
+}
+
+void Engine::AssembleLinearForm(LinearForm &l_form) const
+{
+   /// FIXME - What will the actual parameters be?
+   MFEM_ABORT("FIXME");
+}
+
+mfem::Operator *Engine::MakeOperator(const MixedBilinearForm &mbl_form) const
+{
+   /// FIXME - What will the actual parameters be?
+   MFEM_ABORT("FIXME");
+   return NULL;
+}
+
+mfem::Operator *Engine::MakeOperator(const NonlinearForm &nl_form) const
+{
+   /// FIXME - What will the actual parameters be?
+   MFEM_ABORT("FIXME");
+   return NULL;
+}
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,111 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_ENGINE_HPP
+#define MFEM_BACKENDS_OCCA_ENGINE_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "../base/backend.hpp"
+#include <occa.hpp>
+
+namespace mfem
+{
+
+namespace occa
+{
+
+class Engine : public mfem::Engine
+{
+protected:
+   //
+   // Inherited fields
+   //
+   // mfem::Backend *backend;
+#ifdef MFEM_USE_MPI
+   // MPI_Comm comm;
+#endif
+   // int num_mem_res;
+   // int num_workers;
+   // MemoryResource **memory_resources;
+   // double *workers_weights;
+   // int *workers_mem_res;
+
+   static bool fileOpenerRegistered;
+   ::occa::device *device; // An array of OCCA devices
+   std::string okl_path, okl_defines;
+
+   void Init(const std::string &engine_spec);
+
+public:
+   Engine(const std::string &engine_spec);
+
+#ifdef MFEM_USE_MPI
+   Engine(MPI_Comm comm, const std::string &engine_spec);
+#endif
+
+   virtual ~Engine() { delete [] device; }
+
+   /**
+       @name OCCA specific interface, used by other objects in the OCCA backend
+    */
+   ///@{
+
+   ::occa::device GetDevice(int idx = 0) const { return device[idx]; }
+
+   /// TODO: doxygen
+   const std::string &GetOklPath() const { return okl_path; }
+
+   /// TODO: doxygen
+   const std::string &GetOklDefines() const { return okl_defines; }
+
+   ///@}
+   // End: OCCA specific interface
+
+   /**
+       @name Virtual interface: finite element data structures and algorithms
+    */
+   ///@{
+
+   virtual DLayout MakeLayout(std::size_t size) const;
+   virtual DLayout MakeLayout(const mfem::Array<std::size_t> &offsets) const;
+
+   virtual DArray MakeArray(PLayout &layout, std::size_t item_size) const;
+
+   virtual DVector MakeVector(PLayout &layout,
+                              int type_id = ScalarId<double>::value) const;
+
+   virtual DFiniteElementSpace MakeFESpace(mfem::FiniteElementSpace &
+                                           fespace) const;
+
+   virtual DBilinearForm MakeBilinearForm(mfem::BilinearForm &bf) const;
+
+   /// FIXME - What will the actual parameters be?
+   virtual void AssembleLinearForm(LinearForm &l_form) const;
+
+   /// FIXME - What will the actual parameters be?
+   virtual mfem::Operator *MakeOperator(const MixedBilinearForm &mbl_form) const;
+
+   /// FIXME - What will the actual parameters be?
+   virtual mfem::Operator *MakeOperator(const NonlinearForm &nl_form) const;
+
+   ///@}
+   // End: Virtual interface
+};
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_ENGINE_HPP
@@ -0,0 +1,174 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "backend.hpp"
+#include "fespace.hpp"
+#include "interpolation.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+FiniteElementSpace::FiniteElementSpace(const Engine &e,
+                                       mfem::FiniteElementSpace &fespace)
+   : PFiniteElementSpace(e, fespace),
+     e_layout(e, 0) // resized in SetupLocalGlobalMaps()
+{
+   vdim     = fespace.GetVDim();
+   ordering = fespace.GetOrdering();
+
+   SetupLocalGlobalMaps();
+   SetupOperators();
+   SetupKernels();
+}
+
+FiniteElementSpace::~FiniteElementSpace()
+{
+   delete [] elementDofMap;
+   delete [] elementDofMapInverse;
+   delete restrictionOp;
+   delete prolongationOp;
+}
+
+void FiniteElementSpace::SetupLocalGlobalMaps()
+{
+   const mfem::FiniteElement &fe = *(fes->GetFE(0));
+   const mfem::TensorBasisElement *el =
+      dynamic_cast<const mfem::TensorBasisElement*>(&fe);
+
+   const mfem::Table &e2dTable = fes->GetElementToDofTable();
+   const int *elementMap = e2dTable.GetJ();
+   const int elements = fes->GetNE();
+
+   globalDofs = fes->GetNDofs();
+   localDofs  = fe.GetDof();
+
+   e_layout.Resize(localDofs * elements * fes->GetVDim());
+
+   elementDofMap        = new int[localDofs];
+   elementDofMapInverse = new int[localDofs];
+   if (el)
+   {
+      ::memcpy(elementDofMap,
+               el->GetDofMap().GetData(),
+               localDofs * sizeof(int));
+   }
+   else
+   {
+      for (int i = 0; i < localDofs; ++i)
+      {
+         elementDofMap[i] = i;
+      }
+   }
+   for (int i = 0; i < localDofs; ++i)
+   {
+      elementDofMapInverse[elementDofMap[i]] = i;
+   }
+
+   // Allocate device offsets and indices
+   globalToLocalOffsets.allocate(GetDevice(),
+                                 globalDofs + 1);
+   globalToLocalIndices.allocate(GetDevice(),
+                                 localDofs, elements);
+   localToGlobalMap.allocate(GetDevice(),
+                             localDofs, elements);
+
+   int *offsets = globalToLocalOffsets.ptr();
+   int *indices = globalToLocalIndices.ptr();
+   int *l2gMap  = localToGlobalMap.ptr();
+
+   // We'll be keeping a count of how many local nodes point
+   //   to its global dof
+   for (int i = 0; i <= globalDofs; ++i)
+   {
+      offsets[i] = 0;
+   }
+
+   for (int e = 0; e < elements; ++e)
+   {
+      for (int d = 0; d < localDofs; ++d)
+      {
+         const int gid = elementMap[localDofs*e + d];
+         ++offsets[gid + 1];
+      }
+   }
+   // Aggregate to find offsets for each global dof
+   for (int i = 1; i <= globalDofs; ++i)
+   {
+      offsets[i] += offsets[i - 1];
+   }
+   // For each global dof, fill in all local nodes that point
+   //   to it
+   for (int e = 0; e < elements; ++e)
+   {
+      for (int d = 0; d < localDofs; ++d)
+      {
+         const int gid = elementMap[localDofs*e + elementDofMap[d]];
+         const int lid = localDofs*e + d;
+         indices[offsets[gid]++] = lid;
+         l2gMap[lid] = gid;
+      }
+   }
+   // We shifted the offsets vector by 1 by using it
+   //   as a counter. Now we shift it back.
+   for (int i = globalDofs; i > 0; --i)
+   {
+      offsets[i] = offsets[i - 1];
+   }
+   offsets[0] = 0;
+
+   globalToLocalOffsets.keepInDevice();
+   globalToLocalIndices.keepInDevice();
+   localToGlobalMap.keepInDevice();
+}
+
+void FiniteElementSpace::SetupOperators()
+{
+   const mfem::SparseMatrix *R = fes->GetRestrictionMatrix();
+   const mfem::Operator *P = fes->GetProlongationMatrix();
+   CreateRPOperators(OccaVLayout(), OccaTrueVLayout(),
+                     R, P,
+                     restrictionOp,
+                     prolongationOp);
+}
+
+void FiniteElementSpace::SetupKernels()
+{
+   ::occa::properties props("defines: {"
+                            "  TILESIZE: 256,"
+                            "}");
+   props["defines/NUM_VDIM"] = vdim;
+
+   props["defines/ORDERING_BY_NODES"] = 0;
+   props["defines/ORDERING_BY_VDIM"]  = 1;
+   props["defines/VDIM_ORDERING"] = (int) (ordering == Ordering::byVDIM);
+
+   ::occa::device device = GetDevice();
+   const std::string &okl_path = OccaEngine().GetOklPath();
+   const std::string &okl_defines = OccaEngine().GetOklDefines();
+   globalToLocalKernel = device.buildKernel(okl_path + "fespace.okl",
+                                            "GlobalToLocal",
+                                            props + okl_defines);
+   localToGlobalKernel = device.buildKernel(okl_path + "fespace.okl",
+                                            "LocalToGlobal",
+                                            props + okl_defines);
+}
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,146 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_FE_SPACE_HPP
+#define MFEM_BACKENDS_OCCA_FE_SPACE_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "engine.hpp"
+#include "operator.hpp"
+#include "../../fem/fem.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+/// TODO: doxygen
+class FiniteElementSpace : public mfem::PFiniteElementSpace
+{
+protected:
+   //
+   // Inherited fields
+   //
+   // SharedPtr<const mfem::Engine> engine;
+   // mfem::FiniteElementSpace *fes;
+
+   Layout e_layout;
+
+   int *elementDofMap;
+   int *elementDofMapInverse;
+
+   ::occa::array<int> globalToLocalOffsets;
+   ::occa::array<int> globalToLocalIndices;
+   ::occa::array<int> localToGlobalMap;
+   ::occa::kernel globalToLocalKernel, localToGlobalKernel;
+
+   mfem::Ordering::Type ordering;
+
+   int globalDofs, localDofs;
+   int vdim;
+
+   mfem::Operator *restrictionOp, *prolongationOp;
+
+   void SetupLocalGlobalMaps();
+   void SetupOperators();
+   void SetupKernels();
+
+public:
+   /// TODO: doxygen
+   FiniteElementSpace(const Engine &e, mfem::FiniteElementSpace &fespace);
+
+   /// Virtual destructor
+   virtual ~FiniteElementSpace();
+
+   /// TODO: doxygen
+   const Engine &OccaEngine() const
+   { return *static_cast<const Engine *>(engine.Get()); }
+
+   /// TODO: doxygen
+   ::occa::device GetDevice(int idx = 0) const
+   { return OccaEngine().GetDevice(idx); }
+
+   mfem::Mesh* GetMesh() const { return fes->GetMesh(); }
+
+   Layout &OccaVLayout() const
+   { return *fes->GetVLayout().As<Layout>(); }
+
+   Layout &OccaTrueVLayout() const
+   { return *fes->GetTrueVLayout().As<Layout>(); }
+
+   Layout &OccaEVLayout() { return e_layout; }
+
+#ifdef MFEM_USE_MPI
+   bool isDistributed() const { return (OccaEngine().GetComm() != MPI_COMM_NULL); }
+#else
+   bool isDistributed() const { return false; }
+#endif
+
+   bool hasTensorBasis() const
+   { return dynamic_cast<const mfem::TensorBasisElement*>(fes->GetFE(0)); }
+
+   mfem::Ordering::Type GetOrdering() const { return ordering; }
+
+   int GetGlobalDofs() const { return globalDofs; }
+   int GetLocalDofs() const { return localDofs; }
+
+   int GetDim() const { return fes->GetMesh()->Dimension(); }
+   int GetVDim() const { return vdim; }
+
+   int GetVSize() const { return globalDofs * vdim; }
+   int GetTrueVSize() const { return fes->GetTrueVSize(); }
+   int GetGlobalVSize() const { return globalDofs*vdim; /* FIXME: MPI */ }
+   int GetGlobalTrueVSize() const { return fes->GetTrueVSize(); }
+
+   int GetNE() const { return fes->GetNE(); }
+
+   const mfem::FiniteElementCollection* FEColl() const
+   { return fes->FEColl(); }
+   const mfem::FiniteElement* GetFE(const int idx) const
+   { return fes->GetFE(idx); }
+
+   const int* GetElementDofMap() const { return elementDofMap; }
+   const int* GetElementDofMapInverse() const { return elementDofMapInverse; }
+
+   const mfem::Operator* GetRestrictionOperator() { return restrictionOp; }
+   const mfem::Operator* GetProlongationOperator() { return prolongationOp; }
+
+   const ::occa::array<int> GetLocalToGlobalMap() const
+   { return localToGlobalMap; }
+
+   void GlobalToLocal(const Vector &globalVec, Vector &localVec) const
+   {
+      globalToLocalKernel(globalDofs,
+                          localDofs * fes->GetNE(),
+                          globalToLocalOffsets,
+                          globalToLocalIndices,
+                          globalVec.OccaMem(), localVec.OccaMem());
+   }
+   void LocalToGlobal(const Vector &localVec, Vector &globalVec) const
+   {
+      localToGlobalKernel(globalDofs,
+                          localDofs * fes->GetNE(),
+                          globalToLocalOffsets,
+                          globalToLocalIndices,
+                          localVec.OccaMem(), globalVec.OccaMem());
+   }
+};
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_FE_SPACE_HPP
@@ -0,0 +1,67 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+/*
+  ---[ Defines Known At Compile-Time ]------------
+  TILESIZE : Tilesize for iterating over entries
+  ================================================
+*/
+
+#if VDIM_ORDERING == ORDERING_BY_VDIM
+typedef double *Global_t @dim(NUM_VDIM, globalEntries);
+typedef double *Local_t  @dim(NUM_VDIM, localEntries);
+#else
+typedef double *Global_t @dim(NUM_VDIM, globalEntries) @dimOrder(1, 0);
+typedef double *Local_t  @dim(NUM_VDIM, localEntries)  @dimOrder(1, 0);
+#endif
+
+@kernel void GlobalToLocal(const int globalEntries,
+                           const int localEntries,
+                           const int * restrict offsets,
+                           const int * restrict indices,
+                           const Global_t restrict globalX,
+                           Local_t restrict localX) {
+
+  for (int i = 0; i < globalEntries; ++i; @tile(TILESIZE, @outer, @inner)) {
+    if (i < globalEntries) {
+      const int offset = offsets[i];
+      const int nextOffset = offsets[i + 1];
+      for (int v = 0; v < NUM_VDIM; ++v) {
+        const double dofValue = globalX(v, i);
+        for (int j = offset; j < nextOffset; ++j) {
+          localX(v, indices[j]) = dofValue;
+        }
+      }
+    }
+  }
+}
+
+@kernel void LocalToGlobal(const int globalEntries,
+                           const int localEntries,
+                           const int * restrict offsets,
+                           const int * restrict indices,
+                           const Local_t restrict localX,
+                           Global_t restrict globalX) {
+
+  for (int i = 0; i < globalEntries; ++i; @tile(TILESIZE, @outer, @inner)) {
+    if (i < globalEntries) {
+      const int offset = offsets[i];
+      const int nextOffset = offsets[i + 1];
+      for (int v = 0; v < NUM_VDIM; ++v) {
+        double dofValue = 0;
+        for (int j = offset; j < nextOffset; ++j) {
+          dofValue += localX(v, indices[j]);
+        }
+        globalX(v, i) = dofValue;
+      }
+    }
+  }
+}
@@ -0,0 +1,181 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef STORE_JACOBIAN
+#  define STORE_JACOBIAN 1
+#endif
+
+#ifndef STORE_JACOBIAN_INV
+#  define STORE_JACOBIAN_INV 1
+#endif
+
+#ifndef STORE_JACOBIAN_DET
+#  define STORE_JACOBIAN_DET 1
+#endif
+
+typedef double* Local1D_t @dim(1, NUM_DOFS, numElements);
+typedef double* Local2D_t @dim(2, NUM_DOFS, numElements);
+typedef double* Local3D_t @dim(3, NUM_DOFS, numElements);
+
+typedef double* QLocal_t @dim(NUM_QUAD, numElements);
+
+typedef double* DofToQuadD1D_t @dim(NUM_QUAD, NUM_DOFS);
+typedef double* DofToQuadD2D_t @dim(2, NUM_QUAD, NUM_DOFS);
+typedef double* DofToQuadD3D_t @dim(3, NUM_QUAD, NUM_DOFS);
+
+typedef double* Jacobian1D_t @dim(NUM_QUAD, numElements);
+typedef double* Jacobian2D_t @dim(2, 2, NUM_QUAD, numElements);
+typedef double* Jacobian3D_t @dim(3, 3, NUM_QUAD, numElements);
+
+@kernel void InitGeometryInfo1D(const int numElements,
+                                const DofToQuadD1D_t restrict dofToQuadD,
+                                const Local1D_t restrict nodes,
+                                Jacobian1D_t restrict J,
+                                Jacobian1D_t restrict invJ,
+                                QLocal_t restrict detJ) {
+
+  for (int e = 0; e < numElements; ++e; @outer) {
+    @shared double s_nodes[NUM_DOFS];
+
+    for (int q = 0; q < NUM_QUAD; ++q; @inner) {
+      for (int d = q; d < NUM_DOFS; d += NUM_QUAD) {
+        s_nodes[d] = nodes(0, d, e);
+      }
+    }
+    for (int q = 0; q < NUM_QUAD; ++q; @inner) {
+      double J11 = 0;
+      for (int d = 0; d < NUM_DOFS; ++d) {
+        const double wx = dofToQuadD(q, d);
+        J11 += wx * s_nodes[d];
+      }
+#if STORE_JACOBIAN
+      J(q, e) = J11;
+#endif
+#if STORE_JACOBIAN_INV
+      invJ(q, e) = 1.0 / J11;
+#endif
+#if STORE_JACOBIAN_DET
+      detJ(q, e) = J11;
+#endif
+    }
+  }
+}
+
+@kernel void InitGeometryInfo2D(const int numElements,
+                                const DofToQuadD2D_t restrict dofToQuadD,
+                                const Local2D_t restrict nodes,
+                                Jacobian2D_t restrict J,
+                                Jacobian2D_t restrict invJ,
+                                QLocal_t restrict detJ) {
+
+  for (int e = 0; e < numElements; ++e; @outer) {
+    @shared double s_nodes[2 * NUM_DOFS] @dim(2, NUM_DOFS);
+
+    for (int q = 0; q < NUM_QUAD; ++q; @inner) {
+      for (int d = q; d < NUM_DOFS; d += NUM_QUAD) {
+        s_nodes(0, d) = nodes(0, d, e);
+        s_nodes(1, d) = nodes(1, d, e);
+      }
+    }
+    for (int q = 0; q < NUM_QUAD; ++q; @inner) {
+      double J11 = 0, J12 = 0;
+      double J21 = 0, J22 = 0;
+      for (int d = 0; d < NUM_DOFS; ++d) {
+        const double wx = dofToQuadD(0, q, d);
+        const double wy = dofToQuadD(1, q, d);
+        const double x = s_nodes(0, d);
+        const double y = s_nodes(1, d);
+        J11 += (wx * x); J12 += (wx * y);
+        J21 += (wy * x); J22 += (wy * y);
+      }
+#if STORE_JACOBIAN_INV || STORE_JACOBIAN_DET
+      const double r_detJ = (J11 * J22) - (J12 * J21);
+#endif
+#if STORE_JACOBIAN
+      J(0, 0, q, e) = J11; J(1, 0, q, e) = J12;
+      J(0, 1, q, e) = J21; J(1, 1, q, e) = J22;
+#endif
+#if STORE_JACOBIAN_INV
+      const double r_idetJ = 1.0 / r_detJ;
+      invJ(0, 0, q, e) =  J22 * r_idetJ;
+      invJ(1, 0, q, e) = -J12 * r_idetJ;
+
+      invJ(0, 1, q, e) = -J21 * r_idetJ;
+      invJ(1, 1, q, e) =  J11 * r_idetJ;
+#endif
+#if STORE_JACOBIAN_DET
+      detJ(q, e) = r_detJ;
+#endif
+    }
+  }
+}
+
+@kernel void InitGeometryInfo3D(const int numElements,
+                                const DofToQuadD3D_t restrict dofToQuadD,
+                                const Local3D_t restrict nodes,
+                                Jacobian3D_t restrict J,
+                                Jacobian3D_t restrict invJ,
+                                QLocal_t restrict detJ) {
+
+  for (int e = 0; e < numElements; ++e; @outer) {
+    @shared double s_nodes[3 * NUM_DOFS] @dim(3, NUM_DOFS);
+
+    for (int q = 0; q < NUM_QUAD; ++q; @inner) {
+      for (int d = q; d < NUM_DOFS; d += NUM_QUAD) {
+        s_nodes(0, d) = nodes(0, d, e);
+        s_nodes(1, d) = nodes(1, d, e);
+        s_nodes(2, d) = nodes(2, d, e);
+      }
+    }
+    for (int q = 0; q < NUM_QUAD; ++q; @inner) {
+      double J11 = 0, J12 = 0, J13 = 0;
+      double J21 = 0, J22 = 0, J23 = 0;
+      double J31 = 0, J32 = 0, J33 = 0;
+      for (int d = 0; d < NUM_DOFS; ++d) {
+        const double wx = dofToQuadD(0, q, d);
+        const double wy = dofToQuadD(1, q, d);
+        const double wz = dofToQuadD(2, q, d);
+        const double x = s_nodes(0, d);
+        const double y = s_nodes(1, d);
+        const double z = s_nodes(2, d);
+        J11 += (wx * x); J12 += (wx * y); J13 += (wx * z);
+        J21 += (wy * x); J22 += (wy * y); J23 += (wy * z);
+        J31 += (wz * x); J32 += (wz * y); J33 += (wz * z);
+      }
+#if STORE_JACOBIAN_INV || STORE_JACOBIAN_DET
+      const double r_detJ = ((J11 * J22 * J33) + (J12 * J23 * J31) + (J13 * J21 * J32) -
+                             (J13 * J22 * J31) - (J12 * J21 * J33) - (J11 * J23 * J32));
+#endif
+#if STORE_JACOBIAN
+      J(0, 0, q, e) = J11; J(1, 0, q, e) = J12; J(2, 0, q, e) = J13;
+      J(0, 1, q, e) = J21; J(1, 1, q, e) = J22; J(2, 1, q, e) = J23;
+      J(0, 2, q, e) = J31; J(1, 2, q, e) = J32; J(2, 2, q, e) = J33;
+#endif
+#if STORE_JACOBIAN_INV
+      const double r_idetJ = 1.0 / r_detJ;
+      invJ(0, 0, q, e) = r_idetJ * ((J22 * J33) - (J23 * J32));
+      invJ(1, 0, q, e) = r_idetJ * ((J32 * J13) - (J33 * J12));
+      invJ(2, 0, q, e) = r_idetJ * ((J12 * J23) - (J13 * J22));
+
+      invJ(0, 1, q, e) = r_idetJ * ((J23 * J31) - (J21 * J33));
+      invJ(1, 1, q, e) = r_idetJ * ((J33 * J11) - (J31 * J13));
+      invJ(2, 1, q, e) = r_idetJ * ((J13 * J21) - (J11 * J23));
+
+      invJ(0, 2, q, e) = r_idetJ * ((J21 * J32) - (J22 * J31));
+      invJ(1, 2, q, e) = r_idetJ * ((J31 * J12) - (J32 * J11));
+      invJ(2, 2, q, e) = r_idetJ * ((J11 * J22) - (J12 * J21));
+#endif
+#if STORE_JACOBIAN_DET
+      detJ(q, e) = r_detJ;
+#endif
+    }
+  }
+}
@@ -0,0 +1,195 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "gridfunc.hpp"
+#include "bilininteg.hpp"
+#include "../../fem/gridfunc.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+std::map<std::string, ::occa::kernel> gridFunctionKernels;
+
+::occa::kernel GetGridFunctionKernel(::occa::device device,
+                                     FiniteElementSpace &fespace,
+                                     const mfem::IntegrationRule &ir)
+{
+   const int numQuad = ir.GetNPoints();
+
+   const FiniteElement &fe = *(fespace.GetFE(0));
+   const int dim  = fe.GetDim();
+   const int vdim = fespace.GetVDim();
+
+   std::stringstream ss;
+   ss << ::occa::hash(device)
+      << "FEColl : " << fespace.FEColl()->Name()
+      << "Quad: "    << numQuad
+      << "Dim: "     << dim
+      << "VDim: "    << vdim;
+   std::string hash = ss.str();
+
+   // Kernel defines
+   ::occa::properties props;
+   props["defines/NUM_VDIM"] = vdim;
+
+   SetProperties(fespace, ir, props);
+
+   ::occa::kernel kernel = gridFunctionKernels[hash];
+   if (!kernel.isInitialized())
+   {
+      const std::string &okl_path = fespace.OccaEngine().GetOklPath();
+      kernel = device.buildKernel(okl_path + "gridfunc.okl",
+                                  stringWithDim("GridFuncToQuad", dim),
+                                  props);
+   }
+   return kernel;
+}
+
+// OccaGridFunction::OccaGridFunction() :
+//    Vector(),
+//    ofespace(NULL),
+//    sequence(0) {}
+
+OccaGridFunction::OccaGridFunction(FiniteElementSpace *ofespace_)
+   : PArray(ofespace_->OccaVLayout()),
+     Array(ofespace_->OccaVLayout(), sizeof(double)),
+     Vector(ofespace_->OccaVLayout()),
+     ofespace(ofespace_),
+     sequence(0) {}
+
+// OccaGridFunction::OccaGridFunction(OccaFiniteElementSpace *ofespace_,
+//                                    OccaVectorRef ref) :
+//    OccaVector(ref),
+//    ofespace(ofespace_),
+//    sequence(0) {}
+
+OccaGridFunction::OccaGridFunction(const OccaGridFunction &v)
+   : PArray(v),
+     Array(v),
+     Vector(v),
+     ofespace(v.ofespace),
+     sequence(v.sequence) {}
+
+OccaGridFunction& OccaGridFunction::operator = (double value)
+{
+   Fill(value);
+   return *this;
+}
+
+OccaGridFunction& OccaGridFunction::operator = (const Vector &v)
+{
+   Assign<double>(v);
+   return *this;
+}
+
+// OccaGridFunction& OccaGridFunction::operator = (const OccaVectorRef &v)
+// {
+//    OccaVector::operator = (v);
+//    return *this;
+// }
+
+OccaGridFunction& OccaGridFunction::operator = (const OccaGridFunction &v)
+{
+   Assign<double>(v);
+   return *this;
+}
+
+// void OccaGridFunction::SetGridFunction(mfem::GridFunction &gf)
+// {
+//    Vector v = *this;
+//    gf.MakeRef(ofespace->GetFESpace(), v, 0);
+//    // Make gf the owner of the data
+//    v.Swap(gf);
+// }
+
+void OccaGridFunction::GetTrueDofs(Vector &v)
+{
+   const mfem::Operator *R = ofespace->GetRestrictionOperator();
+   if (!R)
+   {
+      v.MakeRef(*this);
+   }
+   else
+   {
+      v.Resize<double>(R->OutLayout(), NULL);
+      mfem::Vector mfem_v(v);
+      R->Mult(this->Wrap(), mfem_v);
+   }
+}
+
+void OccaGridFunction::SetFromTrueDofs(Vector &v)
+{
+   const mfem::Operator *P = ofespace->GetProlongationOperator();
+   if (!P)
+   {
+      MakeRef(v);
+   }
+   else
+   {
+      Resize<double>(P->OutLayout(), NULL);
+      mfem::Vector mfem_this(*this);
+      P->Mult(v.Wrap(), mfem_this);
+   }
+}
+
+mfem::FiniteElementSpace* OccaGridFunction::GetFESpace()
+{
+   return ofespace->GetFESpace();
+}
+
+const mfem::FiniteElementSpace* OccaGridFunction::GetFESpace() const
+{
+   return ofespace->GetFESpace();
+}
+
+void OccaGridFunction::ToQuad(const IntegrationRule &ir, Vector &quadValues)
+{
+   const Engine &engine = OccaLayout().OccaEngine();
+   ::occa::device device = engine.GetDevice();
+
+   OccaDofQuadMaps &maps = OccaDofQuadMaps::Get(device, *ofespace, ir);
+
+   const int elements = ofespace->GetNE();
+   const int numQuad  = ir.GetNPoints();
+   quadValues.Resize<double>(*(new Layout(engine, numQuad * elements)), NULL);
+
+   ::occa::kernel g2qKernel = GetGridFunctionKernel(device, *ofespace, ir);
+   g2qKernel(elements,
+             maps.dofToQuad,
+             ofespace->GetLocalToGlobalMap(),
+             this->OccaMem(),
+             quadValues.OccaMem());
+}
+
+void OccaGridFunction::Distribute(const Vector &v)
+{
+   if (ofespace->isDistributed())
+   {
+      mfem::Vector mfem_this(*this);
+      ofespace->GetProlongationOperator()->Mult(v.Wrap(), mfem_this);
+   }
+   else
+   {
+      *this = v;
+   }
+}
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,83 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_GRID_FUNC_HPP
+#define MFEM_BACKENDS_OCCA_GRID_FUNC_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "vector.hpp"
+#include "fespace.hpp"
+
+namespace mfem
+{
+
+class IntegrationRule;
+class GridFunction;
+
+namespace occa
+{
+
+class OccaIntegrator;
+class OccaDofQuadMaps;
+
+// TODO: make this object part of the backend or the engine.
+extern std::map<std::string, ::occa::kernel> gridFunctionKernels;
+
+// TODO: make this a method of the backend or the engine.
+::occa::kernel GetGridFunctionKernel(::occa::device device,
+                                     FiniteElementSpace &fespace,
+                                     const mfem::IntegrationRule &ir);
+
+class OccaGridFunction : public Vector
+{
+protected:
+   FiniteElementSpace *ofespace;
+   long sequence;
+
+   ::occa::kernel gridFuncToQuad[3];
+
+public:
+   // OccaGridFunction();
+
+   OccaGridFunction(FiniteElementSpace *ofespace_);
+
+   // OccaGridFunction(FiniteElementSpace *ofespace_,
+   //                  OccaVectorRef ref);
+
+   OccaGridFunction(const OccaGridFunction &gf);
+
+   OccaGridFunction& operator = (double value);
+   OccaGridFunction& operator = (const Vector &v);
+   // OccaGridFunction& operator = (const OccaVectorRef &v);
+   OccaGridFunction& operator = (const OccaGridFunction &gf);
+
+   // void SetGridFunction(mfem::GridFunction &gf);
+
+   void GetTrueDofs(Vector &v);
+   void SetFromTrueDofs(Vector &v);
+
+   mfem::FiniteElementSpace* GetFESpace();
+   const mfem::FiniteElementSpace* GetFESpace() const;
+
+   void ToQuad(const mfem::IntegrationRule &ir, Vector &quadValues);
+
+   void Distribute(const Vector &v);
+};
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_GRID_FUNC_HPP
@@ -0,0 +1,26 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "mfem-occa://defines.okl"
+
+#if USING_TENSOR_OPS
+#  if OCCA_USING_CPU
+#    include "mfem-occa://gridfunc/tensor/cpu.okl"
+#  else
+#    include "mfem-occa://gridfunc/tensor/gpuHighOrder.okl"
+#  endif
+#else
+#  if OCCA_USING_CPU
+#    include "mfem-occa://gridfunc/simplex/cpu.okl"
+#  else
+#    include "mfem-occa://gridfunc/simplex/gpuHighOrder.okl"
+#  endif
+#endif
@@ -0,0 +1,63 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "mfem-occa://defines.okl"
+
+//---[ 2D ]-----------------------------
+@kernel void GridFuncToQuad2D(const int numElements,
+                              const DofToQuad_t restrict dofToQuad,
+                              const DLocalMap_t restrict l2gMap,
+                              const double * restrict gf,
+                              QVLocal_t restrict out) {
+  // Iterate over elements
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int dummy = 0; dummy < 1; ++dummy; @inner) {
+      for (int d = 0; d < NUM_DOFS; ++d) {
+        const int gid = l2gMap(d, e);
+        for (int v = 0; v < NUM_VDIM; ++v) {
+          const double r_gf = gf[v + gid*NUM_VDIM];
+          double r_out = 0;
+          for (int q = 0; q < NUM_QUAD; ++q) {
+            r_out += r_gf * dofToQuad(d, q);
+          }
+          out(v, d, e) = r_out;
+        }
+      }
+    }
+  }
+}
+//======================================
+
+
+//---[ 3D ]-----------------------------
+@kernel void GridFuncToQuad3D(const int numElements,
+                              const DofToQuad_t restrict dofToQuad,
+                              const DLocalMap_t restrict l2gMap,
+                              const double * restrict gf,
+                              QVLocal_t restrict out) {
+  // Iterate over elements
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int dummy = 0; dummy < 1; ++dummy; @inner) {
+      for (int d = 0; d < NUM_DOFS; ++d) {
+        const int gid = l2gMap(d, e);
+        for (int v = 0; v < NUM_VDIM; ++v) {
+          const double r_gf = gf[v + gid*NUM_VDIM];
+          double r_out = 0;
+          for (int q = 0; q < NUM_QUAD; ++q) {
+            r_out += r_gf * dofToQuad(d, q);
+          }
+          out(v, d, e) = r_out;
+        }
+      }
+    }
+  }
+}
+//======================================
@@ -0,0 +1,79 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "mfem-occa://defines.okl"
+
+//---[ 2D ]-----------------------------
+@kernel void GridFuncToQuad2D(const int numElements,
+                              const DofToQuad_t restrict dofToQuad,
+                              const DLocalMap_t restrict l2gMap,
+                              const double * restrict gf,
+                              QVLocal_t restrict out) {
+  for (int e = 0; e < numElements; ++e; @outer) {
+    @shared double s_gf[NUM_VDIM][NUM_DOFS];
+
+    for (int dOff = 0; dOff < M2_INNER_BATCH; ++dOff; @inner) {
+      for (int d = dOff; d < NUM_DOFS; d += M2_INNER_BATCH) {
+        const int gid = l2gMap(d, e);
+        for (int v = 0; v < NUM_VDIM; ++v) {
+          s_gf[v][d] = gf[v + gid*NUM_VDIM]];
+        }
+      }
+    }
+
+    for (int qOff = 0; qOff < M2_INNER_BATCH; ++qOff; @inner) {
+      for (int q = qOff; q < NUM_QUAD; q += M2_INNER_BATCH) {
+        for (int v = 0; v < NUM_VDIM; ++v) {
+          double r_out = 0;
+          for (int d = 0; d < NUM_DOFS; ++d) {
+            r_out += s_gf[v][d] * dofToQuad(d, q);
+          }
+          out(v, q, e) = r_out;
+        }
+      }
+    }
+  }
+}
+//======================================
+
+
+//---[ 3D ]-----------------------------
+@kernel void GridFuncToQuad3D(const int numElements,
+                              const DofToQuad_t restrict dofToQuad,
+                              const DLocalMap_t restrict l2gMap,
+                              const double * restrict gf,
+                              QVLocal_t restrict out) {
+  for (int e = 0; e < numElements; ++e; @outer) {
+    @shared double s_gf[NUM_VDIM][NUM_DOFS];
+
+    for (int dOff = 0; dOff < M3_INNER_BATCH; ++dOff; @inner) {
+      for (int d = dOff; d < NUM_DOFS; d += M3_INNER_BATCH) {
+        const int gid = l2gMap(d, e);
+        for (int v = 0; v < NUM_VDIM; ++v) {
+          s_gf[v][d] = gf[v + gid*NUM_VDIM]];
+        }
+      }
+    }
+
+    for (int qOff = 0; qOff < M3_INNER_BATCH; ++qOff; @inner) {
+      for (int q = qOff; q < NUM_QUAD; q += M3_INNER_BATCH) {
+        for (int v = 0; v < NUM_VDIM; ++v) {
+          double r_out = 0;
+          for (int d = 0; d < NUM_DOFS; ++d) {
+            r_out += s_gf[v][d] * dofToQuad(d, q);
+          }
+          out(v, q, e) = r_out;
+        }
+      }
+    }
+  }
+}
+//======================================
@@ -0,0 +1,188 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "mfem-occa://defines.okl"
+
+//---[ 1D ]-----------------------------
+@kernel void GridFuncToQuad1D(const int numElements,
+                              const DofToQuad_t restrict dofToQuad,
+                              const DLocalMap1D_t restrict l2gMap,
+                              const double * restrict gf,
+                              QVLocal1D_t restrict out) {
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int dummy = 0; dummy < 1; ++dummy; @inner) {
+      double r_out[NUM_VDIM][NUM_QUAD_1D];
+      for (int v = 0; v < NUM_VDIM; ++v) {
+        for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+          r_out[v][qx] = 0;
+        }
+      }
+      for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+        const int gid = l2gMap(dx, e);
+        for (int v = 0; v < NUM_VDIM; ++v) {
+          const double r_gf = gf[v + gid*NUM_VDIM];
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            r_out[v][qx] += r_gf * dofToQuad(qx, dx);
+          }
+        }
+      }
+      for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+        for (int v = 0; v < NUM_VDIM; ++v) {
+          out(v, qx, e) = r_out[v][qx];
+        }
+      }
+    }
+  }
+}
+//======================================
+
+
+//---[ 2D ]-----------------------------
+@kernel void GridFuncToQuad2D(const int numElements,
+                              const DofToQuad_t restrict dofToQuad,
+                              const DLocalMap2D_t restrict l2gMap,
+                              const double * restrict gf,
+                              QVLocal2D_t restrict out) {
+  // Iterate over elements
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int dummy = 0; dummy < 1; ++dummy; @inner) {
+      double out_xy[NUM_VDIM][NUM_QUAD_1D][NUM_QUAD_1D];
+      for (int v = 0; v < NUM_VDIM; ++v) {
+        for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            out_xy[v][qy][qx] = 0;
+          }
+        }
+      }
+
+      for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+        double out_x[NUM_VDIM][NUM_QUAD_1D];
+        for (int v = 0; v < NUM_VDIM; ++v) {
+          for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+            out_x[v][qy] = 0;
+          }
+        }
+
+        for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+          const int gid = l2gMap(dx, dy, e);
+          for (int v = 0; v < NUM_VDIM; ++v) {
+            const double r_gf = gf[v + gid*NUM_VDIM];
+            for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+              out_x[v][qy] += r_gf * dofToQuad(qy, dx);
+            }
+          }
+        }
+
+        for (int v = 0; v < NUM_VDIM; ++v) {
+          for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+            const double d2q = dofToQuad(qy, dy);
+            for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+              out_xy[v][qy][qx] += d2q * out_x[v][qx];
+            }
+          }
+        }
+      }
+
+      for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+        for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+          for (int v = 0; v < NUM_VDIM; ++v) {
+            out(v, qx, qy, e) = out_xy[v][qy][qx];
+          }
+        }
+      }
+    }
+  }
+}
+//======================================
+
+
+//---[ 3D ]-----------------------------
+@kernel void GridFuncToQuad3D(const int numElements,
+                              const DofToQuad_t restrict dofToQuad,
+                              const DLocalMap3D_t restrict l2gMap,
+                              const double * restrict gf,
+                              QVLocal3D_t restrict out) {
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int dummy = 0; dummy < 1; ++dummy; @inner) {
+      double out_xyz[NUM_VDIM][NUM_QUAD_1D][NUM_QUAD_1D][NUM_QUAD_1D];
+      for (int v = 0; v < NUM_VDIM; ++v) {
+        for (int qz = 0; qz < NUM_QUAD_1D; ++qz) {
+          for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+            for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+              out_xyz[v][qz][qy][qx] = 0;
+            }
+          }
+        }
+      }
+
+      for (int dz = 0; dz < NUM_DOFS_1D; ++dz) {
+        double out_xy[NUM_VDIM][NUM_QUAD_1D][NUM_QUAD_1D];
+        for (int v = 0; v < NUM_VDIM; ++v) {
+          for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+            for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+              out_xy[v][qy][qx] = 0;
+            }
+          }
+        }
+
+        for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+          double out_x[NUM_VDIM][NUM_QUAD_1D];
+          for (int v = 0; v < NUM_VDIM; ++v) {
+            for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+              out_x[v][qx] = 0;
+            }
+          }
+
+          for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+            const int gid = l2gMap(dx, dy, dz, e);
+            for (int v = 0; v < NUM_VDIM; ++v) {
+              const double r_gf = gf[v + gid*NUM_VDIM];
+              for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+                out_x[v][qx] += r_gf * dofToQuad(qx, dx);
+              }
+            }
+          }
+
+          for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+            const double wy = dofToQuad(qy, dy);
+            for (int v = 0; v < NUM_VDIM; ++v) {
+              for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+                out_xy[v][qy][qx] += wy * out_x[v][qx];
+              }
+            }
+          }
+        }
+
+        for (int qz = 0; qz < NUM_QUAD_1D; ++qz) {
+          const double wz = dofToQuad(qz, dz);
+          for (int v = 0; v < NUM_VDIM; ++v) {
+            for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+              for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+                out_xyz[v][qz][qy][qx] += wz * out_xy[v][qy][qx];
+              }
+            }
+          }
+        }
+      }
+
+      for (int qz = 0; qz < NUM_QUAD_1D; ++qz) {
+        for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            for (int v = 0; v < NUM_VDIM; ++v) {
+              out(v, qx, qy, qz, e) = out_xyz[v][qz][qy][qx];
+            }
+          }
+        }
+      }
+    }
+  }
+}
+//======================================
@@ -0,0 +1,183 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "mfem-occa://defines.okl"
+
+//---[ 1D ]-----------------------------
+@kernel void GridFuncToQuad1D(const int numElements,
+                              const DofToQuad_t restrict dofToQuad,
+                              const DLocalMap1D_t restrict l2gMap,
+                              const double * restrict gf,
+                              QLocal1D_t restrict out) {
+  // Iterate over elements
+  for (int eOff = 0; eOff < numElements; eOff += M1_ELEMENT_BATCHES; @outer) {
+    @shared double s_dofToQuad[NUM_QUAD_DOFS_1D] @dim(NUM_QUAD_1D, NUM_DOFS_1D);
+    @exclusive double r_out[NUM_QUAD_1D];
+
+    for (int el = 0; el < M1_INNER_ELEMENT_BATCH; ++el; @inner) {
+      for (int i = el; i < NUM_QUAD_DOFS_1D; i += M1_INNER_ELEMENT_BATCH) {
+        s_dofToQuad[i] = dofToQuad[i];
+      }
+    }
+
+    for (int b = 0; b < M1_OUTER_ELEMENT_BATCH; ++b) {
+      for (int el = 0; el < M1_INNER_ELEMENT_BATCH; ++el; @inner) {
+        const int e = eOff + b*M1_INNER_ELEMENT_BATCH + el;
+        if (e < numElements) {
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            r_out[qx] = 0;
+          }
+          for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+            const double r_gf = gf[l2gMap(dx, e)];
+            for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+              r_out[qx] += r_gf * s_dofToQuad(qx, dx);
+            }
+          }
+          for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+            out(qx, e) = r_out[qx];
+          }
+        }
+      }
+    }
+  }
+}
+//======================================
+
+
+//---[ 2D ]-----------------------------
+@kernel void GridFuncToQuad2D(const int numElements,
+                              const DofToQuad_t restrict dofToQuad,
+                              const DLocalMap2D_t restrict l2gMap,
+                              const double * restrict gf,
+                              QLocal2D_t restrict out) {
+  // Iterate over elements
+  for (int eOff = 0; eOff < numElements; eOff += M2_ELEMENT_BATCH; @outer) {
+    // Store dof <--> quad mappings
+    @shared double s_dofToQuad[NUM_QUAD_DOFS_1D] @dim(NUM_QUAD_1D, NUM_DOFS_1D);
+
+    // Store xy planes in shared memory
+    @shared double s_xy[NUM_QUAD_DOFS_1D] @dim(NUM_DOFS_1D, NUM_QUAD_1D);
+
+    for (int x = 0; x < NUM_MAX_1D; ++x; @inner) {
+      for (int id = x; id < NUM_QUAD_DOFS_1D; id += NUM_MAX_1D) {
+        s_dofToQuad[id] = dofToQuad[id];
+      }
+    }
+
+    for (int e = eOff; e < (eOff + M2_ELEMENT_BATCH); ++e) {
+      if (e < numElements) {
+        for (int dx = 0; dx < NUM_MAX_1D; ++dx; @inner) {
+          if (dx < NUM_DOFS_1D) {
+            double r_x[NUM_DOFS_1D];
+            for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+              r_x[dy] = gf[l2gMap(dx, dy, e)];
+            }
+            for (int qy = 0; qy < NUM_QUAD_1D; ++qy) {
+              double xy = 0;
+              for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+                xy += r_x[dy] * s_dofToQuad(qy, dy);
+              }
+              s_xy(dx, qy) = xy;
+            }
+          }
+        }
+        for (int qy = 0; qy < NUM_MAX_1D; ++qy; @inner) {
+          if (qy < NUM_QUAD_1D) {
+            for (int qx = 0; qx < NUM_QUAD_1D; ++qx) {
+              double val = 0;
+              for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+                val += s_xy(dx, qy) * s_dofToQuad(qx, dx);
+              }
+              out(qx, qy, e) = val;
+            }
+          }
+        }
+      }
+    }
+  }
+}
+//======================================
+
+
+//---[ 3D ]-----------------------------
+@kernel void GridFuncToQuad3D(const int numElements,
+                              const DofToQuad_t restrict dofToQuad,
+                              const DLocalMap3D_t restrict l2gMap,
+                              const double * restrict gf,
+                              QLocal3D_t restrict out) {
+  // Iterate over elements
+  for (int e = 0; e < numElements; ++e; @outer) {
+    // Store dof <--> quad mappings
+    @shared double s_dofToQuad[NUM_QUAD_DOFS_1D] @dim(NUM_QUAD_1D, NUM_DOFS_1D);
+
+    // Store xy planes in shared memory
+    @shared double s_z[NUM_MAX_2D] @dim(NUM_MAX_1D, NUM_MAX_1D);
+
+    // Store z axis as registers
+    @exclusive double r_qz[NUM_QUAD_1D];
+
+    for (int y = 0; y < NUM_MAX_1D; ++y; @inner) {
+      for (int x = 0; x < NUM_MAX_1D; ++x; @inner) {
+        const int id = (y * NUM_MAX_1D) + x;
+        // Fetch Q <--> D maps
+        if (id < NUM_QUAD_DOFS_1D) {
+          s_dofToQuad[id] = dofToQuad[id];
+        }
+        // Initialize our Z axis
+        for (int qz = 0; qz < NUM_QUAD_1D; ++qz) {
+          r_qz[qz] = 0;
+        }
+      }
+    }
+
+    for (int dy = 0; dy < NUM_MAX_1D; ++dy; @inner) {
+      for (int dx = 0; dx < NUM_MAX_1D; ++dx; @inner) {
+        if ((dx < NUM_DOFS_1D) && (dy < NUM_DOFS_1D)) {
+          for (int dz = 0; dz < NUM_DOFS_1D; ++dz) {
+            const double val = gf[l2gMap(dx, dy, dz, e)];
+            // Calculate D -> Q in the Z axis
+            for (int qz = 0; qz < NUM_QUAD_1D; ++qz) {
+              r_qz[qz] += val * s_dofToQuad(qz, dz);
+            }
+          }
+        }
+      }
+    }
+    // For each xy plane
+    for (int qz = 0; qz < NUM_QUAD_1D; ++qz) {
+      // Fill xy plane at given z position
+      for (int dy = 0; dy < NUM_MAX_1D; ++dy; @inner) {
+        for (int dx = 0; dx < NUM_MAX_1D; ++dx; @inner) {
+          if ((dx < NUM_DOFS_1D) && (dy < NUM_DOFS_1D)) {
+            s_z(dx, dy) = r_qz[qz];
+          }
+        }
+      }
+      // Calculate Dxyz, xDyz, xyDz in plane
+      for (int qy = 0; qy < NUM_MAX_1D; ++qy; @inner) {
+        for (int qx = 0; qx < NUM_MAX_1D; ++qx; @inner) {
+          if ((qx < NUM_QUAD_1D) && (qy < NUM_QUAD_1D)) {
+            double val = 0;
+            for (int dy = 0; dy < NUM_DOFS_1D; ++dy) {
+              const double wy = s_dofToQuad(qy, dy);
+              for (int dx = 0; dx < NUM_DOFS_1D; ++dx) {
+                const double wx = s_dofToQuad(qx, dx);
+                val += wx * wy * s_z(dx, dy);
+              }
+            }
+            out(qx, qy, qz, e) = val;
+          }
+        }
+      }
+    }
+  }
+}
+//======================================
@@ -0,0 +1,162 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "interpolation.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+void CreateRPOperators(Layout &v_layout, Layout &t_layout,
+                       const mfem::SparseMatrix *R, const mfem::Operator *P,
+                       mfem::Operator *&OccaR, mfem::Operator *&OccaP)
+{
+   if (!P)
+   {
+      OccaR = new IdentityOperator(t_layout);
+      OccaP = new IdentityOperator(t_layout);
+      return;
+   }
+
+   const mfem::SparseMatrix *pmat = dynamic_cast<const mfem::SparseMatrix*>(P);
+   ::occa::device device = v_layout.OccaEngine().GetDevice();
+
+   if (R)
+   {
+      OccaSparseMatrix *occaR =
+         CreateMappedSparseMatrix(v_layout, t_layout, *R);
+      ::occa::array<int> reorderIndices = occaR->reorderIndices;
+      delete occaR;
+
+      OccaR = new RestrictionOperator(v_layout, t_layout, reorderIndices);
+   }
+
+   if (pmat)
+   {
+      const mfem::SparseMatrix *pmatT = Transpose(*pmat);
+
+      OccaSparseMatrix *occaP  =
+         CreateMappedSparseMatrix(t_layout, v_layout, *pmat);
+      OccaSparseMatrix *occaPT =
+         CreateMappedSparseMatrix(v_layout, t_layout, *pmatT);
+
+      OccaP = new ProlongationOperator(*occaP, *occaPT);
+   }
+   else
+   {
+      OccaP = new ProlongationOperator(t_layout, v_layout, P);
+   }
+}
+
+RestrictionOperator::RestrictionOperator(Layout &in_layout, Layout &out_layout,
+                                         ::occa::array<int> indices) :
+   Operator(in_layout, out_layout)
+{
+
+   entries     = indices.size() / 2;
+   trueIndices = indices;
+
+   // FIXME: paths ...
+   ::occa::device device = in_layout.OccaEngine().GetDevice();
+   const std::string &okl_path = in_layout.OccaEngine().GetOklPath();
+   const std::string &okl_defines = in_layout.OccaEngine().GetOklDefines();
+   multOp = device.buildKernel(okl_path + "mappings.okl",
+                               "ExtractSubVector",
+                               "defines: { TILESIZE: 256 }" + okl_defines);
+
+   multTransposeOp = device.buildKernel(okl_path + "mappings.okl",
+                                        "SetSubVector",
+                                        "defines: { TILESIZE: 256 }" +
+                                        okl_defines);
+}
+
+void RestrictionOperator::Mult_(const Vector &x, Vector &y) const
+{
+   multOp(entries, trueIndices, x.OccaMem(), y.OccaMem());
+}
+
+void RestrictionOperator::MultTranspose_(const Vector &x, Vector &y) const
+{
+   y.Fill<double>(0.0);
+   multTransposeOp(entries, trueIndices, x.OccaMem(), y.OccaMem());
+}
+
+ProlongationOperator::ProlongationOperator(OccaSparseMatrix &multOp_,
+                                           OccaSparseMatrix &multTransposeOp_) :
+   Operator(multOp_),
+   pmat(NULL),
+   multOp(multOp_),
+   multTransposeOp(multTransposeOp_) {}
+
+ProlongationOperator::ProlongationOperator(Layout &in_layout,
+                                           Layout &out_layout,
+                                           const mfem::Operator *pmat_) :
+   Operator(in_layout, out_layout),
+   pmat(pmat_),
+   multOp(*this),
+   multTransposeOp(*this)
+{ }
+
+void ProlongationOperator::Mult_(const Vector &x, Vector &y) const
+{
+   MFEM_VERIFY(pmat == NULL, "");
+   multOp.Mult_(x, y);
+}
+
+void ProlongationOperator::MultTranspose_(const Vector &x, Vector &y) const
+{
+   MFEM_VERIFY(pmat == NULL, "");
+   multTransposeOp.Mult_(x, y);
+}
+
+void ProlongationOperator::Mult(const mfem::Vector &x, mfem::Vector &y) const
+{
+   if (pmat)
+   {
+      // FIXME: create an OCCA version of 'pmat'
+      x.Pull();
+      y.Pull(false);
+      pmat->Mult(x, y);
+      y.Push();
+   }
+   else
+   {
+      multOp.Mult(x, y);
+   }
+}
+
+void ProlongationOperator::MultTranspose(const mfem::Vector &x,
+                                         mfem::Vector &y) const
+{
+   if (pmat)
+   {
+      // FIXME: create an OCCA version of 'pmat'
+      x.Pull();
+      y.Pull(false);
+      pmat->MultTranspose(x, y);
+      y.Push();
+   }
+   else
+   {
+      multTransposeOp.Mult(x, y);
+   }
+}
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,79 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_INTERPOLATION_HPP
+#define MFEM_BACKENDS_OCCA_INTERPOLATION_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include <occa.hpp>
+#include "vector.hpp"
+#include "engine.hpp"
+#include "sparsemat.hpp"
+#include "../../fem/fem.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+// [MISSING] Proper destructors
+void CreateRPOperators(Layout &v_layout, Layout &t_layout,
+                       const mfem::SparseMatrix *R, const mfem::Operator *P,
+                       mfem::Operator *&OccaR, mfem::Operator *&OccaP);
+
+class RestrictionOperator : public Operator
+{
+protected:
+   int entries;
+   ::occa::array<int> trueIndices;
+   ::occa::kernel multOp, multTransposeOp;
+
+public:
+   RestrictionOperator(Layout &in_layout, Layout &out_layout,
+                       ::occa::array<int> indices);
+
+   // overrides
+   virtual void Mult_(const Vector &x, Vector &y) const;
+   virtual void MultTranspose_(const Vector &x, Vector &y) const;
+};
+
+class ProlongationOperator : public Operator
+{
+protected:
+   const mfem::Operator *pmat;
+   OccaSparseMatrix multOp, multTransposeOp;
+
+public:
+   ProlongationOperator(OccaSparseMatrix &multOp_,
+                        OccaSparseMatrix &multTransposeOp_);
+
+   ProlongationOperator(Layout &in_layout, Layout &out_layout,
+                        const mfem::Operator *pmat_);
+
+   // overrides
+   virtual void Mult_(const Vector &x, Vector &y) const;
+   virtual void MultTranspose_(const Vector &x, Vector &y) const;
+
+   // overrides
+   virtual void Mult(const mfem::Vector &x, mfem::Vector &y) const;
+   virtual void MultTranspose(const mfem::Vector &x, mfem::Vector &y) const;
+};
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_INTERPOLATION_HPP
@@ -9,35 +9,32 @@
 // terms of the GNU Lesser General Public License (as published by the Free
 // Software Foundation) version 2.1 dated February 1999.

-#ifndef MFEM_OCCA_HPP
-#define MFEM_OCCA_HPP
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)

-#include "../config/config.hpp"
-
-#ifdef MFEM_USE_OCCA
-#include <occa.hpp>
-
-#if defined(MFEM_USE_CUDA) && OCCA_CUDA_ENABLED
-#include <occa/modes/cuda/utils.hpp>
-#endif
-
-typedef occa::device OccaDevice;
-typedef occa::memory OccaMemory;
-
-#else // MFEM_USE_OCCA
-
-typedef void* OccaDevice;
-typedef void* OccaMemory;
-
-#endif // MFEM_USE_OCCA
+#include "layout.hpp"
+#include "../../general/array.hpp"

 namespace mfem
 {

-// Function called when the pointer 'a' needs to be passed to an OCCA kernel.
-OccaMemory OccaPtr(const void *a);
-OccaDevice OccaDev();
+namespace occa
+{
+
+void Layout::Resize(std::size_t new_size)
+{
+   size = new_size;
+}
+
+void Layout::Resize(const Array<std::size_t> &offsets)
+{
+   MFEM_ASSERT(offsets.Size() == 2,
+               "multiple workers are not supported yet");
+   size = offsets.Last();
+}
+
+} // namespace mfem::occa

 } // namespace mfem

-#endif // MFEM_OCCA_HPP
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,68 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_LAYOUT_HPP
+#define MFEM_BACKENDS_OCCA_LAYOUT_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "../base/layout.hpp"
+#include "engine.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+class Layout : public PLayout
+{
+protected:
+   //
+   // Inherited fields
+   //
+   // SharedPtr<const mfem::Engine> engine;
+   // std::size_t size;
+
+public:
+   Layout(const Engine &e, std::size_t s = 0) : PLayout(e, s) { }
+
+   const Engine &OccaEngine() const
+   { return *static_cast<const Engine *>(engine.Get()); }
+
+   ::occa::memory Alloc(std::size_t bytes) const
+   { return OccaEngine().GetDevice().malloc(bytes); }
+
+   virtual ~Layout() { }
+
+   /**
+       @name Virtual interface
+    */
+   ///@{
+
+   /// Resize the layout
+   virtual void Resize(std::size_t new_size);
+
+   /// Resize the layout based on the given worker offsets
+   virtual void Resize(const Array<std::size_t> &offsets);
+
+   ///@}
+   // End: Virtual interface
+};
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_LAYOUT_HPP
@@ -0,0 +1,54 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+/*
+  ---[ Defines Known At Compile-Time ]------------
+  TILESIZE : Tilesize for iterating over entries
+  ================================================
+*/
+
+@kernel void ExtractSubVector(const int entries,
+                              const int * restrict indices,
+                              const double * restrict in,
+                              double * restrict out) {
+
+  for (int i = 0; i < entries; ++i; @tile(TILESIZE, @outer, @inner)) {
+    if (i < entries) {
+      out[i] = in[indices[i]];
+    }
+  }
+}
+
+@kernel void SetSubVector(const int entries,
+                          const int * restrict indices,
+                          const double * restrict in,
+                          double * restrict out) {
+
+  for (int i = 0; i < entries; ++i; @tile(TILESIZE, @outer, @inner)) {
+    if (i < entries) {
+      out[indices[i]] = in[i];
+    }
+  }
+}
+
+@kernel void MapSubVector(const int entries,
+                          const int * restrict indices,
+                          const double * restrict in,
+                          double * restrict out) {
+
+  for (int i = 0; i < entries; ++i; @tile(TILESIZE, @outer, @inner)) {
+    if (i < entries) {
+      const int fromIdx = indices[2*i + 0];
+      const int toIdx   = indices[2*i + 1];
+      out[toIdx] = in[fromIdx];
+    }
+  }
+}
@@ -0,0 +1,135 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "operator.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+// FIXME: move this object to the Backend?
+::occa::kernelBuilder OccaConstrainedOperator::mapDofBuilder =
+   ::occa::linalg::customLinearMethod(
+      "vector_map_dofs",
+
+      "const int idx = v2[i];"
+      "v0[idx] = v1[idx];",
+
+      "defines: {"
+      "  VTYPE0: 'double',"
+      "  VTYPE1: 'double',"
+      "  VTYPE2: 'int',"
+      "  TILESIZE: 128,"
+      "}");
+
+// FIXME: move this object to the Backend?
+::occa::kernelBuilder OccaConstrainedOperator::clearDofBuilder =
+   ::occa::linalg::customLinearMethod(
+      "vector_clear_dofs",
+
+      "v0[v1[i]] = 0.0;",
+
+      "defines: {"
+      "  VTYPE0: 'double',"
+      "  VTYPE1: 'int',"
+      "  TILESIZE: 128,"
+      "}");
+
+OccaConstrainedOperator::OccaConstrainedOperator(
+   mfem::Operator *A_,
+   const mfem::Array<int> &constraintList_,
+   bool own_A_)
+
+   : Operator(A_->InLayout()->As<Layout>()),
+     z(OutLayout_()),
+     w(OutLayout_()),
+     mfem_z((z.DontDelete(), z)),
+     mfem_w((w.DontDelete(), w))
+{
+   Setup(OutLayout_().OccaEngine().GetDevice(), A_, constraintList_, own_A_);
+}
+
+void OccaConstrainedOperator::Setup(::occa::device device_,
+                                    mfem::Operator *A_,
+                                    const mfem::Array<int> &constraintList_,
+                                    bool own_A_)
+{
+   device = device_;
+
+   A = A_;
+   own_A = own_A_;
+
+   constraintIndices = constraintList_.Size();
+   constraintList = constraintList_.Get_PArray()->As<Array>().OccaMem();
+}
+
+void OccaConstrainedOperator::EliminateRHS(const Vector &x, Vector &b) const
+{
+   const std::string &okl_defines = InLayout_().OccaEngine().GetOklDefines();
+   ::occa::kernel mapDofs = mapDofBuilder.build(device, okl_defines);
+
+   w.Fill<double>(0.0);
+
+   if (constraintIndices)
+   {
+      mapDofs(constraintIndices, w.OccaMem(), x.OccaMem(), constraintList);
+   }
+
+   A->Mult(mfem_w, mfem_z);
+
+   b.Axpby<double>(1.0, b, -1.0, z);
+
+   if (constraintIndices)
+   {
+      mapDofs(constraintIndices, b.OccaMem(), x.OccaMem(), constraintList);
+   }
+}
+
+void OccaConstrainedOperator::Mult_(const Vector &x, Vector &y) const
+{
+   mfem::Vector mfem_y(y);
+   if (constraintIndices == 0)
+   {
+      A->Mult(x.Wrap(), mfem_y);
+      return;
+   }
+
+   const std::string &okl_defines = InLayout_().OccaEngine().GetOklDefines();
+   ::occa::kernel mapDofs   = mapDofBuilder.build(device, okl_defines);
+   ::occa::kernel clearDofs = clearDofBuilder.build(device, okl_defines);
+
+   z.Assign<double>(x); // z = x
+
+   clearDofs(constraintIndices, z.OccaMem(), constraintList);
+
+   A->Mult(mfem_z, mfem_y);
+
+   mapDofs(constraintIndices, y.OccaMem(), x.OccaMem(), constraintList);
+}
+
+OccaConstrainedOperator::~OccaConstrainedOperator()
+{
+   if (own_A)
+   {
+      delete A;
+   }
+}
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,129 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_OPERATOR_HPP
+#define MFEM_BACKENDS_OCCA_OPERATOR_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "vector.hpp"
+#include "../../linalg/operator.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+class Operator : public mfem::Operator
+{
+public:
+   /// Creare an operator with the same dimensions as @a orig.
+   Operator(const Operator &orig)
+      : mfem::Operator(orig) { }
+
+   Operator(Layout &layout)
+      : mfem::Operator(layout) { }
+
+   Operator(Layout &in_layout, Layout &out_layout)
+      : mfem::Operator(in_layout, out_layout) { }
+
+   Layout &InLayout_() const
+   { return *static_cast<Layout*>(in_layout.Get()); }
+
+   Layout &OutLayout_() const
+   { return *static_cast<Layout*>(out_layout.Get()); }
+
+   virtual void Mult_(const Vector &x, Vector &y) const = 0;
+
+   virtual void MultTranspose_(const Vector &x, Vector &y) const
+   { MFEM_ABORT("method is not supported"); }
+
+   // override
+   virtual void Mult(const mfem::Vector &x, mfem::Vector &y) const
+   {
+      Mult_(x.Get_PVector()->As<Vector>(),
+            y.Get_PVector()->As<Vector>());
+   }
+
+   // override
+   virtual void MultTranspose(const mfem::Vector &x, mfem::Vector &y) const
+   {
+      MultTranspose_(x.Get_PVector()->As<Vector>(),
+                     y.Get_PVector()->As<Vector>());
+   }
+};
+
+
+class OccaConstrainedOperator : public Operator
+{
+protected:
+   ::occa::device device;
+
+   mfem::Operator *A;              //< The unconstrained Operator.
+   bool own_A;                     //< Ownership flag for A.
+   ::occa::memory constraintList;  //< List of constrained indices/dofs.
+   int constraintIndices;
+   mutable Vector z, w;       //< Auxiliary vectors.
+   mutable mfem::Vector mfem_z, mfem_w; // Wrap z, w
+
+   static ::occa::kernelBuilder mapDofBuilder, clearDofBuilder;
+
+public:
+   /** @brief Constructor from a general Operator and a list of essential
+       indices/dofs.
+
+       Specify the unconstrained operator @a *A and a @a list of indices to
+       constrain, i.e. each entry @a list[i] represents an essential-dof. If the
+       ownership flag @a own_A is true, the operator @a *A will be destroyed
+       when this object is destroyed. */
+   OccaConstrainedOperator(mfem::Operator *A_,
+                           const mfem::Array<int> &constraintList_,
+                           bool own_A_ = false);
+
+   void Setup(::occa::device device_,
+              mfem::Operator *A_,
+              const mfem::Array<int> &constraintList_,
+              bool own_A_ = false);
+
+   /** @brief Eliminate "essential boundary condition" values specified in @a x
+       from the given right-hand side @a b.
+
+       Performs the following steps:
+
+       z = A((0,x_b));  b_i -= z_i;  b_b = x_b;
+
+       where the "_b" subscripts denote the essential (boundary) indices/dofs of
+       the vectors, and "_i" -- the rest of the entries. */
+   void EliminateRHS(const Vector &x, Vector &b) const;
+
+   /** @brief Constrained operator action.
+
+       Performs the following steps:
+
+       z = A((x_i,0));  y_i = z_i;  y_b = x_b;
+
+       where the "_b" subscripts denote the essential (boundary) indices/dofs of
+       the vectors, and "_i" -- the rest of the entries. */
+   virtual void Mult_(const Vector &x, Vector &y) const;
+
+   // Destructor: destroys the unconstrained Operator @a A if @a own_A is true.
+   virtual ~OccaConstrainedOperator();
+};
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_OPERATOR_HPP
@@ -0,0 +1,57 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+/*
+  ---[ Defines Known At Compile-Time ]------------
+  TILESIZE : Tilesize for iterating over dofs
+  ================================================
+*/
+
+@kernel void Mult(const int entries,
+                  const int * restrict offsets,
+                  const int * restrict indices,
+                  const double * restrict weights,
+                  const double * restrict in,
+                  double * restrict out) {
+
+  for (int i = 0; i < entries; ++i; @tile(TILESIZE, @outer, @inner)) {
+    if (i < entries) {
+      const int offset = offsets[i];
+      const int nextOffset = offsets[i + 1];
+      double value = 0;
+      for (int j = offset; j < nextOffset; ++j) {
+        value += weights[j] * in[indices[j]];
+      }
+      out[i] = value;
+    }
+  }
+}
+
+@kernel void MappedMult(const int entries,
+                        const int * restrict offsets,
+                        const int * restrict indices,
+                        const double * restrict weights,
+                        const int * restrict outIndices,
+                        const double * restrict in,
+                        double * restrict out) {
+
+  for (int i = 0; i < entries; ++i; @tile(TILESIZE, @outer, @inner)) {
+    if (i < entries) {
+      const int offset = offsets[i];
+      const int nextOffset = offsets[i + 1];
+      double value = 0;
+      for (int j = offset; j < nextOffset; ++j) {
+        value += weights[j] * in[indices[j]];
+      }
+      out[outIndices[i]] = value;
+    }
+  }
+}
@@ -0,0 +1,248 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "sparsemat.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+OccaSparseMatrix::OccaSparseMatrix(Layout &in_layout, Layout &out_layout,
+                                   const mfem::SparseMatrix &m,
+                                   const ::occa::properties &props) :
+   Operator(in_layout, out_layout)
+{
+
+   Setup(in_layout.OccaEngine().GetDevice(), m, props);
+}
+
+OccaSparseMatrix::OccaSparseMatrix(Layout &in_layout, Layout &out_layout,
+                                   const mfem::SparseMatrix &m,
+                                   ::occa::array<int> reorderIndices_,
+                                   ::occa::array<int> mappedIndices_,
+                                   const ::occa::properties &props) :
+   Operator(in_layout, out_layout)
+{
+
+   Setup(in_layout.OccaEngine().GetDevice(), m,
+         reorderIndices, mappedIndices_, props);
+}
+
+OccaSparseMatrix::OccaSparseMatrix(Layout &in_layout, Layout &out_layout,
+                                   ::occa::array<int> offsets_,
+                                   ::occa::array<int> indices_,
+                                   ::occa::array<double> weights_,
+                                   const ::occa::properties &props) :
+   Operator(in_layout, out_layout),
+   offsets(offsets_),
+   indices(indices_),
+   weights(weights_)
+{
+
+   SetupKernel(in_layout.OccaEngine().GetDevice(), props);
+}
+
+OccaSparseMatrix::OccaSparseMatrix(Layout &in_layout, Layout &out_layout,
+                                   ::occa::array<int> offsets_,
+                                   ::occa::array<int> indices_,
+                                   ::occa::array<double> weights_,
+                                   ::occa::array<int> reorderIndices_,
+                                   ::occa::array<int> mappedIndices_,
+                                   const ::occa::properties &props) :
+   Operator(in_layout, out_layout),
+   offsets(offsets_),
+   indices(indices_),
+   weights(weights_),
+   reorderIndices(reorderIndices_),
+   mappedIndices(mappedIndices_)
+{
+
+   SetupKernel(in_layout.OccaEngine().GetDevice(), props);
+}
+
+void OccaSparseMatrix::Setup(::occa::device device, const mfem::SparseMatrix &m,
+                             const ::occa::properties &props)
+{
+   Setup(device, m, ::occa::array<int>(), ::occa::array<int>(), props);
+}
+
+void OccaSparseMatrix::Setup(::occa::device device, const SparseMatrix &m,
+                             ::occa::array<int> reorderIndices_,
+                             ::occa::array<int> mappedIndices_,
+                             const ::occa::properties &props)
+{
+
+   const int nnz = m.GetI()[height];
+   offsets.allocate(device,
+                    height + 1, m.GetI());
+   indices.allocate(device,
+                    nnz, m.GetJ());
+   weights.allocate(device,
+                    nnz, m.GetData());
+
+   offsets.keepInDevice();
+   indices.keepInDevice();
+   weights.keepInDevice();
+
+   reorderIndices = reorderIndices_;
+   mappedIndices  = mappedIndices_;
+
+   SetupKernel(device, props);
+}
+
+void OccaSparseMatrix::SetupKernel(::occa::device device,
+                                   const ::occa::properties &props)
+{
+
+   const bool hasOutIndices = mappedIndices.isInitialized();
+
+   const ::occa::properties defaultProps("defines: {"
+                                         "  TILESIZE: 256,"
+                                         "}");
+
+   const std::string &okl_path = InLayout_().OccaEngine().GetOklPath();
+   const std::string &okl_defines = InLayout_().OccaEngine().GetOklDefines();
+   mapKernel = device.buildKernel(okl_path + "mappings.okl",
+                                  "MapSubVector",
+                                  defaultProps + props + okl_defines);
+
+   multKernel = device.buildKernel(okl_path + "sparse.okl",
+                                   hasOutIndices ? "MappedMult" : "Mult",
+                                   defaultProps + props + okl_defines);
+}
+
+void OccaSparseMatrix::Mult_(const Vector &x, Vector &y) const
+{
+   if (reorderIndices.isInitialized() ||
+       mappedIndices.isInitialized())
+   {
+      if (reorderIndices.isInitialized())
+      {
+         mapKernel((int) (reorderIndices.size() / 2),
+                   reorderIndices,
+                   x.OccaMem(), y.OccaMem());
+      }
+      if (mappedIndices.isInitialized())
+      {
+         multKernel((int) (mappedIndices.size()),
+                    offsets, indices, weights,
+                    mappedIndices,
+                    x.OccaMem(), y.OccaMem());
+      }
+   }
+   else
+   {
+      multKernel((int) height,
+                 offsets, indices, weights,
+                 x.OccaMem(), y.OccaMem());
+   }
+}
+
+
+OccaSparseMatrix* CreateMappedSparseMatrix(Layout &in_layout,
+                                           Layout &out_layout,
+                                           const mfem::SparseMatrix &m,
+                                           const ::occa::properties &props)
+{
+   const int mHeight = m.Height();
+   // const int mWidth  = m.Width();
+
+   // Count indices that are only reordered (true dofs)
+   const int *I = m.GetI();
+   const int *J = m.GetJ();
+   const double *D = m.GetData();
+
+   int trueCount = 0;
+   for (int i = 0; i < mHeight; ++i)
+   {
+      trueCount += ((I[i + 1] - I[i]) == 1);
+   }
+   const int dupCount = (mHeight - trueCount);
+
+   // Create the reordering map for entries that aren't modified (true dofs)
+   ::occa::device device(in_layout.OccaEngine().GetDevice());
+   ::occa::array<int> reorderIndices(device,
+                                     2 * trueCount);
+   ::occa::array<int> mappedIndices, offsets, indices;
+   ::occa::array<double> weights;
+
+   if (dupCount)
+   {
+      mappedIndices.allocate(device,
+                             dupCount);
+   }
+   int trueIdx = 0, dupIdx = 0;
+   for (int i = 0; i < mHeight; ++i)
+   {
+      const int i1 = I[i];
+      if ((I[i + 1] - i1) == 1)
+      {
+         reorderIndices[trueIdx++] = J[i1];
+         reorderIndices[trueIdx++] = i;
+      }
+      else
+      {
+         mappedIndices[dupIdx++] = i;
+      }
+   }
+   reorderIndices.keepInDevice();
+
+   if (dupCount)
+   {
+      mappedIndices.keepInDevice();
+
+      // Extract sparse matrix without reordered identity
+      const int dupNnz = I[mHeight] - trueCount;
+
+      offsets.allocate(device,
+                       dupCount + 1);
+      indices.allocate(device,
+                       dupNnz);
+      weights.allocate(device,
+                       dupNnz);
+
+      int nnz = 0;
+      offsets[0] = 0;
+      for (int i = 0; i < dupCount; ++i)
+      {
+         const int idx = mappedIndices[i];
+         const int offStart = I[idx];
+         const int offEnd   = I[idx + 1];
+         offsets[i + 1] = offsets[i] + (offEnd - offStart);
+         for (int j = offStart; j < offEnd; ++j)
+         {
+            indices[nnz] = J[j];
+            weights[nnz] = D[j];
+            ++nnz;
+         }
+      }
+
+      offsets.keepInDevice();
+      indices.keepInDevice();
+      weights.keepInDevice();
+   }
+
+   return new OccaSparseMatrix(in_layout, out_layout,
+                               offsets, indices, weights,
+                               reorderIndices, mappedIndices,
+                               props);
+}
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,95 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_SPARSE_MAT_HPP
+#define MFEM_BACKENDS_OCCA_SPARSE_MAT_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include <occa.hpp>
+#include "vector.hpp"
+#include "engine.hpp"
+#include "operator.hpp"
+#include "../../linalg/sparsemat.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+/// TODO: doxygen
+class OccaSparseMatrix : public Operator
+{
+public:
+   ::occa::array<int> offsets, indices;
+   ::occa::array<double> weights;
+   ::occa::array<int> reorderIndices, mappedIndices;
+   ::occa::kernel mapKernel, multKernel;
+
+   /// Construct an empty OccaSparseMatrix.
+   OccaSparseMatrix(const Operator &orig)
+      : Operator(orig) { }
+
+   OccaSparseMatrix(Layout &in_layout, Layout &out_layout,
+                    const mfem::SparseMatrix &m,
+                    const ::occa::properties &props = ::occa::properties());
+
+   OccaSparseMatrix(Layout &in_layout, Layout &out_layout,
+                    const mfem::SparseMatrix &m,
+                    ::occa::array<int> reorderIndices_,
+                    ::occa::array<int> mappedIndices_,
+                    const ::occa::properties &props = ::occa::properties());
+
+   OccaSparseMatrix(Layout &in_layout, Layout &out_layout,
+                    ::occa::array<int> offsets_,
+                    ::occa::array<int> indices_,
+                    ::occa::array<double> weights_,
+                    const ::occa::properties &props = ::occa::properties());
+
+   OccaSparseMatrix(Layout &in_layout, Layout &out_layout,
+                    ::occa::array<int> offsets_,
+                    ::occa::array<int> indices_,
+                    ::occa::array<double> weights_,
+                    ::occa::array<int> reorderIndices_,
+                    ::occa::array<int> mappedIndices_,
+                    const ::occa::properties &props = ::occa::properties());
+
+   void Setup(::occa::device device, const mfem::SparseMatrix &m,
+              const ::occa::properties &props);
+
+   void Setup(::occa::device device, const mfem::SparseMatrix &m,
+              ::occa::array<int> reorderIndices_,
+              ::occa::array<int> mappedIndices_,
+              const ::occa::properties &props);
+
+   void SetupKernel(::occa::device device,
+                    const ::occa::properties &props);
+
+   // override
+   virtual void Mult_(const Vector &x, Vector &y) const;
+};
+
+
+/// TODO: doxygen
+OccaSparseMatrix* CreateMappedSparseMatrix(
+   Layout &in_layout, Layout &out_layout,
+   const mfem::SparseMatrix &m,
+   const ::occa::properties &props = ::occa::properties());
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_SPARSE_MAT_HPP
@@ -0,0 +1,81 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "url_handler.hpp"
+#include "../../general/error.hpp"
+#include <cstdlib>
+#include <sys/stat.h>
+
+namespace mfem
+{
+
+namespace occa
+{
+
+FileOpener::FileOpener(const std::string &prefix,
+                       const std::string &env_variable)
+   : pfx(prefix)
+{
+   const char *env_path = getenv(env_variable.c_str());
+   if (!env_path) { return; }
+   std::string path(env_path);
+   for (std::size_t start = 0, end; start < path.size(); start = end + 1)
+   {
+      end = path.find(':', start);
+      if (end == std::string::npos)
+      {
+         AddDir(path.substr(start, end));
+         break;
+      }
+      AddDir(path.substr(start, end - start));
+   }
+}
+
+bool FileOpener::AddDir(const std::string &dir)
+{
+   if (dir.size() == 0 || dir[0] != '/') { return false; }
+   struct stat dir_stat;
+   if (stat(dir.c_str(), &dir_stat)) { return false; }
+   if (!S_ISDIR(dir_stat.st_mode)) { return false; }
+   paths.push_back(dir + (*dir.rbegin() == '/' ? "" : "/"));
+   return true;
+}
+
+bool FileOpener::handles(const std::string &filename)
+{
+   return filename.size() >= pfx.size() &&
+          filename.compare(0, pfx.size(), pfx) == 0;
+}
+
+std::string FileOpener::expand(const std::string &filename)
+{
+   std::string sfx(filename.substr(pfx.size()));
+   for (std::size_t i = 0; i < paths.size(); i++)
+   {
+      std::string file = paths[i] + sfx;
+      struct stat file_stat;
+      if (stat(file.c_str(), &file_stat) == 0 && S_ISREG(file_stat.st_mode))
+      {
+         return file;
+      }
+   }
+   MFEM_ABORT("invalid url: " << filename);
+   return sfx;
+}
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,47 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_URL_HANDLER_HPP
+#define MFEM_BACKENDS_OCCA_URL_HANDLER_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include <occa.hpp>
+
+namespace mfem
+{
+
+namespace occa
+{
+
+class FileOpener : public ::occa::io::fileOpener
+{
+protected:
+   std::string pfx; // prefix, e.g. "mfem://"
+   std::vector<std::string> paths; // paths to search for prefix replacement
+
+public:
+   FileOpener(const std::string &prefix, const std::string &env_variable);
+
+   bool AddDir(const std::string &dir);
+
+   virtual bool handles(const std::string &filename);
+   virtual std::string expand(const std::string &filename);
+};
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_URL_HANDLER_HPP
@@ -9,20 +9,14 @@
 // terms of the GNU Lesser General Public License (as published by the Free
 // Software Foundation) version 2.1 dated February 1999.

-#include "mfem.hpp"
-using namespace mfem;
+typedef double* Local_t @dim(numDofs, numElements);

-#include "catch.hpp"
-#include "general/text.hpp"
-
-TEST_CASE("String Manipulation", "[General]")
-{
-   SECTION("String Conversion")
-   {
-      SECTION("Integer")
-      {
-         REQUIRE(to_int(to_string(12)) == 12);
-         REQUIRE(to_int(to_string(-1234)) == -1234);
-      }
-   }
+@kernel void InitLocalVector(const int numElements,
+                             const int numDofs,
+                             Local_t restrict sol) {
+  for (int e = 0; e < numElements; ++e; @outer) {
+    for (int d = 0; d < numDofs; ++d; @inner) {
+      sol(d, e) = 0;
+    }
+  }
 }
@@ -0,0 +1,204 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include "vector.hpp"
+#include "../../linalg/vector.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+PVector *Vector::DoVectorClone(bool copy_data, void **buffer,
+                               int buffer_type_id) const
+{
+   MFEM_ASSERT(buffer_type_id == ScalarId<double>::value, "");
+   Vector *new_vector = new Vector(OccaLayout());
+   if (copy_data)
+   {
+      new_vector->slice.copyFrom(slice);
+   }
+   if (buffer)
+   {
+      *buffer = new_vector->GetBuffer();
+   }
+   return new_vector;
+}
+
+void Vector::DoDotProduct(const PVector &x, void *result,
+                          int result_type_id) const
+{
+   // Can be called when Size() == 0, e.g. when an MPI-parallel vector has a
+   // local size of 0.
+
+   MFEM_ASSERT(result_type_id == ScalarId<double>::value, "");
+   double *res = (double *)result;
+   MFEM_ASSERT(dynamic_cast<const Vector *>(&x) != NULL, "invalid Vector type");
+   const Vector *xp = static_cast<const Vector *>(&x);
+   MFEM_ASSERT(this->Size() == xp->Size(), "");
+   *res = ::occa::linalg::dot<double, double, double>(this->slice, xp->slice);
+
+#ifdef MFEM_USE_MPI
+   double local_dot = *res;
+   if (IsParallel())
+   {
+      MPI_Allreduce(&local_dot, res, 1, MPI_DOUBLE, MPI_SUM,
+                    OccaLayout().OccaEngine().GetComm());
+   }
+#endif
+}
+
+void Vector::DoAxpby(const void *a, const PVector &x,
+                     const void *b, const PVector &y,
+                     int ab_type_id)
+{
+   const std::string &okl_defines = OccaLayout().OccaEngine().GetOklDefines();
+
+   //
+   // TODO: move all kernel builders to class mfem::occa::Backend
+   //
+   static ::occa::kernelBuilder axpby1_builder =
+      ::occa::linalg::customLinearMethod(
+         "mfem_occa_axpby1",
+         "v0[i] = c0 * v1[i];",
+         "defines: {"
+         "  CTYPE0: 'double',"
+         "  VTYPE0: 'double',"
+         "  VTYPE1: 'double',"
+         "  TILESIZE: '128',"
+         "}");
+
+   static ::occa::kernelBuilder axpby2_builder =
+      ::occa::linalg::customLinearMethod(
+         "mfem_occa_axpby2",
+         "v0[i] = c0 * v0[i] + c1 * v1[i];",
+         "defines: {"
+         "  CTYPE0: 'double',"
+         "  CTYPE1: 'double',"
+         "  VTYPE0: 'double',"
+         "  VTYPE1: 'double',"
+         "  TILESIZE: '128',"
+         "}");
+
+   static ::occa::kernelBuilder axpby3_builder =
+      ::occa::linalg::customLinearMethod(
+         "mfem_occa_axpby3",
+         "v0[i] = c0 * v1[i] + c1 * v2[i];",
+         "defines: {"
+         "  CTYPE0: 'double',"
+         "  CTYPE1: 'double',"
+         "  VTYPE0: 'double',"
+         "  VTYPE1: 'double',"
+         "  VTYPE2: 'double',"
+         "  TILESIZE: '128',"
+         "}");
+
+   // called only when Size() != 0
+
+   MFEM_ASSERT(ab_type_id == ScalarId<double>::value, "");
+   const double da = *static_cast<const double *>(a);
+   const double db = *static_cast<const double *>(b);
+   MFEM_ASSERT(da == 0.0 || dynamic_cast<const Vector *>(&x) != NULL,
+               "invalid Vector x");
+   MFEM_ASSERT(db == 0.0 || dynamic_cast<const Vector *>(&y) != NULL,
+               "invalid Vector y");
+   const Vector *xp = static_cast<const Vector *>(&x);
+   const Vector *yp = static_cast<const Vector *>(&y);
+
+   MFEM_ASSERT(da == 0.0 || this->Size() == xp->Size(), "");
+   MFEM_ASSERT(db == 0.0 || this->Size() == yp->Size(), "");
+
+   if (da == 0.0)
+   {
+      if (db == 0.0)
+      {
+         OccaFill(&da);
+      }
+      else
+      {
+         if (this->slice == yp->slice)
+         {
+            // *this *= db
+            ::occa::linalg::operator_mult_eq(slice, db);
+         }
+         else
+         {
+            // *this = db * y
+            ::occa::kernel kernel = axpby1_builder.build(slice.getDevice(),
+                                                         okl_defines);
+            kernel((int)Size(), db, slice, yp->slice);
+         }
+      }
+   }
+   else
+   {
+      if (db == 0.0)
+      {
+         if (this->slice == xp->slice)
+         {
+            // *this *= da
+            ::occa::linalg::operator_mult_eq(slice, da);
+         }
+         else
+         {
+            // *this = da * x
+            ::occa::kernel kernel = axpby1_builder.build(slice.getDevice(),
+                                                         okl_defines);
+            kernel((int)Size(), da, slice, xp->slice);
+         }
+      }
+      else
+      {
+         MFEM_ASSERT(xp->slice != yp->slice, "invalid input");
+         if (this->slice == xp->slice)
+         {
+            // *this = da * (*this) + db * y
+            ::occa::kernel kernel = axpby2_builder.build(slice.getDevice(),
+                                                         okl_defines);
+            kernel((int)Size(), da, db, slice, yp->slice);
+         }
+         else if (this->slice == yp->slice)
+         {
+            // *this = da * x + db * (*this)
+            ::occa::kernel kernel = axpby2_builder.build(slice.getDevice(),
+                                                         okl_defines);
+            kernel((int)Size(), db, da, slice, xp->slice);
+         }
+         else
+         {
+            // *this = da * x + db * y
+            ::occa::kernel kernel = axpby3_builder.build(slice.getDevice(),
+                                                         okl_defines);
+            kernel((int)Size(), da, db, slice, xp->slice, yp->slice);
+         }
+      }
+   }
+}
+
+mfem::Vector Vector::Wrap()
+{
+   return mfem::Vector(*this);
+}
+
+const mfem::Vector Vector::Wrap() const
+{
+   return mfem::Vector(*const_cast<Vector*>(this));
+}
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
@@ -0,0 +1,74 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OCCA_VECTOR_HPP
+#define MFEM_BACKENDS_OCCA_VECTOR_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#include <occa.hpp>
+#include "../base/vector.hpp"
+#include "array.hpp"
+
+namespace mfem
+{
+
+namespace occa
+{
+
+class Vector : virtual public Array, public PVector
+{
+protected:
+   //
+   // Inherited fields
+   //
+   // DLayout layout;
+
+   /**
+       @name Virtual interface
+    */
+   ///@{
+
+   virtual PVector *DoVectorClone(bool copy_data, void **buffer,
+                                  int buffer_type_id) const;
+
+   virtual void DoDotProduct(const PVector &x, void *result,
+                             int result_type_id) const;
+
+   virtual void DoAxpby(const void *a, const PVector &x,
+                        const void *b, const PVector &y,
+                        int ab_type_id);
+
+   ///@}
+   // End: Virtual interface
+
+public:
+   Vector(Layout &lt)
+      : PArray(lt), Array(lt, sizeof(double)), PVector(lt)
+   { }
+
+   mfem::Vector Wrap();
+
+   const mfem::Vector Wrap() const;
+
+#if defined(MFEM_USE_MPI)
+   bool IsParallel() const { return (OccaLayout().OccaEngine().GetComm() != MPI_COMM_NULL); }
+#endif
+};
+
+} // namespace mfem::occa
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OCCA)
+
+#endif // MFEM_BACKENDS_OCCA_VECTOR_HPP
@@ -0,0 +1,675 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) &&               \
+   defined(MFEM_USE_OMP) &&                     \
+   defined(MFEM_USE_ACROTENSOR)
+
+#include "adiffusioninteg.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+PAIntegrator::PAIntegrator(Coefficient &q, FiniteElementSpace &f)
+{
+   Q = &q;
+   ofes = &f;
+   fes = ofes->GetFESpace();
+   onGPU = (ofes->OmpEngine().ExecTarget() == Device);
+   fe = fes->GetFE(0);
+   tfe = dynamic_cast<const TensorBasisElement*>(fe);
+   if (tfe)
+   {
+      tDofMap = tfe->GetDofMap();
+   }
+   else
+   {
+      tDofMap.SetSize(nDof);
+      for (int i = 0; i < nDof; ++i)
+      {
+         tDofMap[i] = i;
+      }
+   }
+
+   nElem  = fes->GetNE();
+   GeomType = fe->GetGeomType();
+   FEOrder = fe->GetOrder();
+   nDim    = fe->GetDim();
+   nDof   = fe->GetDof();
+
+   ElementTransformation *Trans = fes->GetElementTransformation(0);
+   int irorder = 2*fe->GetOrder() + Trans->OrderW();
+   ir = &IntRules.Get(GeomType, irorder);
+   nQuad = ir->GetNPoints();
+   hasTensorBasis = tfe ? true : false;
+
+   if (nDim > 3)
+   {
+      mfem_error("AcroIntegrator tensor computations don't support dim > 3.");
+   }
+}
+
+PAIntegrator::~PAIntegrator()
+{
+
+}
+
+AcroDiffusionIntegrator::AcroDiffusionIntegrator(Coefficient &q, FiniteElementSpace &f) :
+   PAIntegrator(q,f)
+{
+   if (onGPU)
+   {
+      //TE.SetExecutorType("OneOutPerThread");
+      TE.SetExecutorType("Cuda");
+      //TODO:  Set to an existing cuda context if one exists
+   }
+   else
+   {
+      TE.SetExecutorType("CPUInterpreted");
+   }
+
+   const IntegrationRule *ir1D = &IntRules.Get(Geometry::SEGMENT, ir->GetOrder());
+   nDof1D = FEOrder + 1;
+   nQuad1D = ir1D->GetNPoints();
+
+   if (hasTensorBasis)
+   {
+      H1_FECollection fec(FEOrder,1);
+      const FiniteElement *fe1D = fec.FiniteElementForGeometry(Geometry::SEGMENT);
+      mfem::Vector eval(nDof1D);
+      DenseMatrix deval(nDof1D,1);
+      B.Init(nQuad1D, nDof1D);
+      G.Init(nQuad1D, nDof1D);
+      std::vector<int> wdims(nDim, nQuad1D);
+      W.Init(wdims);
+
+      mfem::Vector w(nQuad1D);
+      for (int k = 0; k < nQuad1D; ++k)
+      {
+         const IntegrationPoint &ip = ir1D->IntPoint(k);
+         fe1D->CalcShape(ip, eval);
+         fe1D->CalcDShape(ip, deval);
+
+         B(k,0) = eval(0);
+         B(k,nDof1D-1) = eval(1);
+         G(k,0) = deval(0,0);
+         G(k,nDof1D-1) = deval(1,0);
+         for (int i = 1; i < nDof1D-1; ++i)
+         {
+            B(k,i) = eval(i+1);
+            G(k,i) = deval(i+1,0);
+         }
+         w(k) = ip.weight;
+      }
+
+      if (nDim == 1)
+      {
+         for (int k1 = 0; k1 < nQuad1D; ++k1)
+         {
+            W(k1) = w(k1);
+         }
+      }
+      else if (nDim == 2)
+      {
+         for (int k1 = 0; k1 < nQuad1D; ++k1)
+         {
+            for (int k2 = 0; k2 < nQuad1D; ++k2)
+            {
+               W(k1,k2) = w(k1)*w(k2);
+            }
+         }
+      }
+      else if (nDim == 3)
+      {
+         for (int k1 = 0; k1 < nQuad1D; ++k1)
+         {
+            for (int k2 = 0; k2 < nQuad1D; ++k2)
+            {
+               for (int k3 = 0; k3 < nQuad1D; ++k3)
+               {
+                  W(k1,k2,k3) = w(k1)*w(k2)*w(k3);
+               }
+            }
+         }
+      }
+   }
+   else
+   {
+      mfem::Vector eval(nDof);
+      DenseMatrix deval(nDof,nDim);
+      G.Init(nQuad, nDof,nDim);
+      W.Init(nQuad);
+      for (int k = 0; k < nQuad; ++k)
+      {
+         const IntegrationPoint &ip = ir->IntPoint(k);
+         fe->CalcDShape(ip, deval);
+         for (int i = 0; i < nDof; ++i)
+         {
+            for (int d = 0; d < nDim; ++d)
+            {
+               G(k,i,d) = deval(i,d);
+            }
+         }
+         W(k) = ip.weight;
+      }
+   }
+
+   if (onGPU)
+   {
+      B.MapToGPU();
+      G.MapToGPU();
+      W.MapToGPU();
+   }
+
+   // Assemble in the constructor!
+   BatchedPartialAssemble();
+}
+
+
+AcroDiffusionIntegrator::~AcroDiffusionIntegrator()
+{
+   for (int i = 0; i < Btil.Size(); i++) delete Btil[i];
+}
+
+
+void AcroDiffusionIntegrator::ComputeBTilde()
+{
+   Btil.SetSize(nDim);
+   for (int d = 0; d < nDim; ++d)
+   {
+      Btil[d] = new acro::Tensor(nDim, nDim, nQuad1D, nDof1D, nDof1D);
+      for (int m = 0; m < nDim; ++m)
+      {
+         for (int n = 0; n < nDim; ++n)
+         {
+            acro::Tensor &BGM = (m == d) ? G : B;
+            acro::Tensor &BGN = (n == d) ? G : B;
+            for (int k = 0; k < nQuad1D; ++k)
+            {
+               for (int i = 0; i < nDof1D; ++i)
+               {
+                  for (int j = 0; j < nDof1D; ++j)
+                  {
+                     (*Btil[d])(m, n, k, i, j) = BGM(k,i)*BGN(k,j);
+                  }
+               }
+            }
+         }
+      }
+   }
+}
+
+
+void AcroDiffusionIntegrator::BatchedPartialAssemble()
+{
+   //Initilze the tensors
+   acro::Tensor J,Jinv,Jdet,C;
+   if (hasTensorBasis)
+   {
+      const IntegrationRule *ir1D = &IntRules.Get(Geometry::SEGMENT, ir->GetOrder());
+      IntegrationPoint ip;
+      if (nDim == 1)
+      {
+         D.Init(nElem, nDim, nDim, nQuad1D);
+         J.Init(nElem, nQuad1D, nDim, nDim);
+         Jinv.Init(nElem, nQuad1D, nDim, nDim);
+         Jdet.Init(nElem, nQuad1D);
+         C.Init(nElem, nQuad1D);
+
+         for (int e = 0; e < nElem; ++e)
+         {
+            ElementTransformation *Trans = fes->GetElementTransformation(e);
+            for (int k1 = 0; k1 < nQuad1D; ++k1)
+            {
+               ip.x = ir1D->IntPoint(k1).x;
+               ip.y = 0.0;
+               ip.z = 0.0;
+               Trans->SetIntPoint(&ip);
+               C(e,k1) = Q->Eval(*Trans, ip);
+               const DenseMatrix &JMat = Trans->Jacobian();
+               for (int m = 0; m < nDim; ++m)
+               {
+                  for (int n = 0; n < nDim; ++n)
+                  {
+                     J(e,k1,m,n) = JMat.Elem(m,n);
+                  }
+               }
+            }
+         }
+      }
+      else if (nDim == 2)
+      {
+         D.Init(nElem, nDim, nDim, nQuad1D, nQuad1D);
+         J.Init(nElem, nQuad1D, nQuad1D, nDim, nDim);
+         Jinv.Init(nElem, nQuad1D, nQuad1D, nDim, nDim);
+         Jdet.Init(nElem, nQuad1D, nQuad1D);
+         C.Init(nElem, nQuad1D, nQuad1D);
+
+         for (int e = 0; e < nElem; ++e)
+         {
+            ElementTransformation *Trans = fes->GetElementTransformation(e);
+            for (int k1 = 0; k1 < nQuad1D; ++k1)
+            {
+               for (int k2 = 0; k2 < nQuad1D; ++k2)
+               {
+                  ip.x = ir1D->IntPoint(k1).x;
+                  ip.y = ir1D->IntPoint(k2).y;
+                  ip.z = 0.0;
+                  Trans->SetIntPoint(&ip);
+                  C(e,k1,k2) = Q->Eval(*Trans, ip);
+                  const DenseMatrix &JMat = Trans->Jacobian();
+                  for (int m = 0; m < nDim; ++m)
+                  {
+                     for (int n = 0; n < nDim; ++n)
+                     {
+                        J(e,k1,k2,m,n) = JMat.Elem(m,n);
+                     }
+                  }
+               }
+            }
+         }
+      }
+      else if (nDim == 3)
+      {
+         D.Init(nElem, nDim, nDim, nQuad1D, nQuad1D, nQuad1D);
+         J.Init(nElem, nQuad1D, nQuad1D, nQuad1D, nDim, nDim);
+         Jinv.Init(nElem, nQuad1D, nQuad1D, nQuad1D, nDim, nDim);
+         Jdet.Init(nElem, nQuad1D, nQuad1D, nQuad1D);
+         C.Init(nElem, nQuad1D, nQuad1D, nQuad1D);
+
+         for (int e = 0; e < nElem; ++e)
+         {
+            ElementTransformation *Trans = fes->GetElementTransformation(e);
+            for (int k1 = 0; k1 < nQuad1D; ++k1)
+            {
+               for (int k2 = 0; k2 < nQuad1D; ++k2)
+               {
+                  for (int k3 = 0; k3 < nQuad1D; ++k3)
+                  {
+                     ip.x = ir1D->IntPoint(k1).x;
+                     ip.y = ir1D->IntPoint(k2).y;
+                     ip.z = ir1D->IntPoint(k3).z;
+                     Trans->SetIntPoint(&ip);
+                     C(e,k1,k2,k3) = Q->Eval(*Trans, ip);
+                     const DenseMatrix &JMat = Trans->Jacobian();
+                     for (int m = 0; m < nDim; ++m)
+                     {
+                        for (int n = 0; n < nDim; ++n)
+                        {
+                           J(e,k1,k2,k3,m,n) = JMat.Elem(m,n);
+                        }
+                     }
+                  }
+               }
+            }
+         }
+      }
+   }
+   else
+   {
+      D.Init(nElem, nDim, nDim, nQuad);
+      J.Init(nElem, nQuad, nDim, nDim);
+      Jinv.Init(nElem, nQuad, nDim, nDim);
+      Jdet.Init(nElem, nQuad);
+      C.Init(nElem, nQuad);
+
+      for (int e = 0; e < nElem; ++e)
+      {
+         ElementTransformation *Trans = fes->GetElementTransformation(e);
+         for (int k = 0; k < nQuad; ++k)
+         {
+            const IntegrationPoint &ip = ir->IntPoint(k);
+            Trans->SetIntPoint(&ip);
+            C(e,k) = Q->Eval(*Trans, ip);
+            const DenseMatrix &JMat = Trans->Jacobian();
+            for (int m = 0; m < nDim; ++m)
+            {
+               for (int n = 0; n < nDim; ++n)
+               {
+                  J(e,k,m,n) = JMat.Elem(m,n);
+               }
+            }
+         }
+      }
+   }
+
+   TE.BatchMatrixInvDet(Jinv, Jdet, J);
+
+   if (hasTensorBasis)
+   {
+      if (nDim == 1)
+      {
+         TE("D_e_m_n_k = W_k C_e_k Jdet_e_k Jinv_e_k_m_j Jinv_e_k_n_j",
+            D, W, C, Jdet, Jinv, Jinv);
+      }
+      else if (nDim == 2)
+      {
+         TE("D_e_m_n_k1_k2 = W_k1_k2 C_e_k1_k2 Jdet_e_k1_k2 Jinv_e_k1_k2_m_j Jinv_e_k1_k2_n_j",
+            D, W, C, Jdet, Jinv, Jinv);
+      }
+      else if (nDim == 3)
+      {
+         TE("D_e_m_n_k1_k2_k3 = W_k1_k2_k3 C_e_k1_k2_k3 Jdet_e_k1_k2_k3 Jinv_e_k1_k2_k3_n_j Jinv_e_k1_k2_k3_m_j",
+            D, W, C, Jdet, Jinv, Jinv);
+      }
+   }
+   else
+   {
+      TE("D_e_m_n_k = W_k C_e_k Jdet_e_k Jinv_e_k_m_j Jinv_e_k_n_j",
+         D, W, C, Jdet, Jinv, Jinv);
+   }
+}
+
+
+void AcroDiffusionIntegrator::BatchedAssembleElementMatrices(DenseTensor &elmats)
+{
+   if (hasTensorBasis && Btil.Size() == 0)
+   {
+      ComputeBTilde();
+   }
+
+   if (!D.IsInitialized())
+   {
+      BatchedPartialAssemble();
+   }
+
+   if (!S.IsInitialized())
+   {
+      if (hasTensorBasis)
+      {
+         if (nDim == 1)
+         {
+            S.Init(nElem, nDof1D, nDof1D);
+         }
+         else if (nDim == 2)
+         {
+            S.Init(nElem, nDof1D, nDof1D, nDof1D, nDof1D);
+         }
+         else if (nDim == 3)
+         {
+            S.Init(nElem, nDof1D, nDof1D, nDof1D, nDof1D, nDof1D, nDof1D);
+         }
+      }
+      else
+      {
+         S.Init(nElem, nDof, nDof);
+      }
+      if (onGPU) {S.SwitchToGPU();}
+   }
+
+
+   if (hasTensorBasis) {
+      if (nDim == 1) {
+         TE("S_e_i1_j1 = Btil_m_n_k1_i1_j1 D_e_m_n_k1",
+            S, *Btil[0], D);
+      }
+      else if (nDim == 2)
+      {
+         TE("S_e_i1_i2_j1_j2 = Btil1_m_n_k1_i1_j1 Btil2_m_n_k2_i2_j2 D_e_m_n_k1_k2",
+            S, *Btil[0], *Btil[1], D);
+      }
+      else if (nDim == 3)
+      {
+         TE("S_e_i1_i2_i3_j1_j2_j3 = Btil1_m_n_k1_i1_j1 Btil2_m_n_k2_i2_j2 Btil3_m_n_k3_i3_j3 D_e_m_n_k1_k2_k3",
+            S, *Btil[0], *Btil[1], *Btil[2], D);
+      }
+   }
+   else
+   {
+      TE("S_e_i_j = G_k_i_m G_k_i_n D_e_m_n_k",
+         S, G, G, D);
+   }
+
+   S.MoveFromGPU();
+   for (int e = 0; e < nElem; ++e)
+   {
+      for (int ei = 0; ei < nDof; ++ei)
+      {
+         for (int ej = 0; ej < nDof; ++ej)
+         {
+            elmats(tDofMap[ei], tDofMap[ej], e) = S[e*nDof*nDof + ei*nDof + ej];
+         }
+      }
+   }
+}
+
+
+void AcroDiffusionIntegrator::ComputeElementMatrices(Vector &elmats)
+{
+   if (hasTensorBasis && Btil.Size() == 0)
+   {
+      ComputeBTilde();
+   }
+
+   if (!D.IsInitialized())
+   {
+      BatchedPartialAssemble();
+   }
+
+   if (!S.IsInitialized())
+   {
+      if (hasTensorBasis)
+      {
+         if (nDim == 1)
+         {
+            S.Init(nElem, nDof1D, nDof1D);
+         }
+         else if (nDim == 2)
+         {
+            S.Init(nElem, nDof1D, nDof1D, nDof1D, nDof1D);
+         }
+         else if (nDim == 3)
+         {
+            S.Init(nElem, nDof1D, nDof1D, nDof1D, nDof1D, nDof1D, nDof1D);
+         }
+      }
+      else
+      {
+         S.Init(nElem, nDof, nDof);
+      }
+      if (onGPU) {S.SwitchToGPU();}
+   }
+
+
+   if (hasTensorBasis) {
+      if (nDim == 1) {
+         TE("S_e_i1_j1 += Btil_m_n_k1_i1_j1 D_e_m_n_k1",
+            S, *Btil[0], D);
+      }
+      else if (nDim == 2)
+      {
+         TE("S_e_i1_i2_j1_j2 += Btil1_m_n_k1_i1_j1 Btil2_m_n_k2_i2_j2 D_e_m_n_k1_k2",
+            S, *Btil[0], *Btil[1], D);
+      }
+      else if (nDim == 3)
+      {
+         TE("S_e_i1_i2_i3_j1_j2_j3 += Btil1_m_n_k1_i1_j1 Btil2_m_n_k2_i2_j2 Btil3_m_n_k3_i3_j3 D_e_m_n_k1_k2_k3",
+            S, *Btil[0], *Btil[1], *Btil[2], D);
+      }
+   }
+   else
+   {
+      TE("S_e_i_j += G_k_i_m G_k_i_n D_e_m_n_k",
+         S, G, G, D);
+   }
+
+   S.MoveFromGPU();
+
+   double *edata = elmats.GetData<double>();
+   for (int e = 0; e < nElem; ++e)
+   {
+      const int e_offset = e * nDof * nDof;
+      for (int ei = 0; ei < nDof; ++ei)
+      {
+         const int offset = e_offset + ei * tDofMap[ei] * nDof;
+         for (int ej = 0; ej < nDof; ++ej)
+         {
+            const int index = offset + tDofMap[ej];
+            edata[index] = S[e*nDof*nDof + ei*nDof + ej];
+         }
+      }
+   }
+}
+
+void AcroDiffusionIntegrator::ReassembleOperator()
+{
+   BatchedPartialAssemble();
+}
+
+void AcroDiffusionIntegrator::PAMult(const Vector &x, Vector &y)
+{
+   MFEM_ASSERT(hasTensorBasis,"AcroDiffusionIntegrator PAMult on simplices not supported");
+
+   if (!U.IsInitialized())
+   {
+      // NOTE: x and y are already sized for the fespace in the constructor
+      double *Xptr = const_cast<double*>(x.GetData<double>());
+      double *Yptr = y.GetData<double>();
+      if (nDim == 1) {
+         X.Init(nElem,nDof1D,Xptr,Xptr,onGPU);
+         Y.Init(nElem,nDof1D,Yptr,Yptr,onGPU);
+         U.Init(nDim, nElem, nQuad1D);
+         Z.Init(nDim, nElem, nQuad1D);
+         if (onGPU)
+         {
+            U.SwitchToGPU();
+            Z.SwitchToGPU();
+         }
+      }
+      else if (nDim == 2)
+      {
+         X.Init(nElem,nDof1D,nDof1D,Xptr,Xptr,onGPU);
+         Y.Init(nElem,nDof1D,nDof1D,Yptr,Yptr,onGPU);
+         U.Init(nDim, nElem, nQuad1D, nQuad1D);
+         Z.Init(nDim, nElem, nQuad1D, nQuad1D);
+         T1.Init(nElem,nDof1D,nQuad1D);
+         if (onGPU)
+         {
+            U.SwitchToGPU();
+            Z.SwitchToGPU();
+            T1.SwitchToGPU();
+         }
+      }
+      else if (nDim == 3)
+      {
+         X.Init(nElem,nDof1D,nDof1D,nDof1D,Xptr,Xptr,onGPU);
+         Y.Init(nElem,nDof1D,nDof1D,nDof1D,Yptr,Yptr,onGPU);
+         U.Init(nDim, nElem, nQuad1D, nQuad1D, nQuad1D);
+         Z.Init(nDim, nElem, nQuad1D, nQuad1D, nQuad1D);
+         T1.Init(nElem, nDof1D, nQuad1D, nQuad1D);
+         T2.Init(nElem, nDof1D, nDof1D, nQuad1D);
+         if (onGPU)
+         {
+            U.SwitchToGPU();
+            Z.SwitchToGPU();
+            T1.SwitchToGPU();
+            T2.SwitchToGPU();
+         }
+      }
+   }
+   else
+   {
+      // NOTE: x and y are already sized for the fespace in the constructor
+      double *Xptr = const_cast<double*>(x.GetData<double>());
+      double *Yptr = y.GetData<double>();
+      X.Retarget(Xptr,Xptr);
+      Y.Retarget(Yptr,Yptr);
+   }
+
+   acro::SliceTensor U1,U2,U3,Z1,Z2,Z3;
+   if (nDim == 1)
+   {
+      TE("U_n_e_k1 = G_k1_i1 X_e_i1", U, G, X);
+      TE("Z_m_e_k1 = D_e_m_n_k1 U_n_e_k1", Z, D, U);
+      TE("Y_e_i1 = G_k1_i1 Z_m_e_k1", Y, G, Z);
+   }
+   else if (nDim == 2)
+   {
+      U1.SliceInit(U, 0); U2.SliceInit(U, 1);
+      Z1.SliceInit(Z, 0); Z2.SliceInit(Z, 1);
+
+      //U1_e_k1_k2 = G_k1_i1 B_k2_i2 X_e_i1_i2
+      TE("BX_e_i1_k2 = B_k2_i2 X_e_i2_i1", T1, B, X);
+      TE("U1_e_k1_k2 = G_k1_i1 BX_e_i1_k2", U1, G, T1);
+
+      //U2_e_k1_k2 = B_k1_i1 G_k2_i2 X_e_i1_i2
+      TE("GX_e_i1_k2 = G_k2_i2 X_e_i2_i1", T1, G, X);
+      TE("U2_e_k1_k2 = B_k1_i1 GX_e_i1_k2", U2, B, T1);
+
+      TE("Z_m_e_k1_k2 = D_e_m_n_k1_k2 U_n_e_k1_k2", Z, D, U);
+
+      //Y_e_i1_i2 = G_k1_i1 B_k2_i2 Z1_e_k1_k2
+      TE("BZ1_e_i2_k1 = B_k2_i2 Z1_e_k1_k2", T1, B, Z1);
+      TE("Y_e_i2_i1 = G_k1_i1 BZ1_e_i2_k1", Y, G, T1);
+
+      //Y_e_i1_i2 += B_k1_i1 G_k2_i2 Z2_e_k1_k2
+      TE("GZ2_e_i2_k1 = G_k2_i2 Z2_e_k1_k2", T1, G, Z2);
+      TE("Y_e_i2_i1 += B_k1_i1 GZ2_e_i2_k1", Y, B, T1);
+   }
+   else if (nDim == 3)
+   {
+      U1.SliceInit(U, 0); U2.SliceInit(U, 1); U3.SliceInit(U, 2);
+      Z1.SliceInit(Z, 0); Z2.SliceInit(Z, 1); Z3.SliceInit(Z, 2);
+
+      TE.BeginMultiKernelLaunch();
+      //U1_e_k1_k2_k3 = G_k1_i1 B_k2_i2 B_k3_i3 X_e_i1_i2_i3
+      TE("T2_e_i1_i2_k3 = B_k3_i3 X_e_i1_i2_i3", T2, B, X);
+      TE("T1_e_i1_k2_k3 = B_k2_i2 T2_e_i1_i2_k3", T1, B, T2);
+      TE("U1_e_k1_k2_k3 = G_k1_i1 T1_e_i1_k2_k3", U1, G, T1);
+
+      //U2_e_k1_k2_k3 = B_k1_i1 G_k2_i2 B_k3_i3 X_e_i1_i2_i3
+      TE("T1_e_i1_k2_k3 = G_k2_i2 T2_e_i1_i2_k3", T1, G, T2);
+      TE("U2_e_k1_k2_k3 = B_k1_i1 T1_e_i1_k2_k3", U2, B, T1);
+
+      //U3_e_k1_k2_k3 = B_k1_i1 B_k2_i2 G_k3_i3 X_e_i1_i2_i3
+      TE("T2_e_i1_i2_k3 = G_k3_i3 X_e_i1_i2_i3", T2, G, X);
+      TE("T1_e_i1_k2_k3 = B_k2_i2 T2_e_i1_i2_k3", T1, B, T2);
+      TE("U3_e_k1_k2_k3 = B_k1_i1 T1_e_i1_k2_k3", U3, B, T1);
+
+      TE("Z_m_e_k1_k2_k3 = D_e_m_n_k1_k2_k3 U_n_e_k1_k2_k3", Z, D, U);
+
+      //Y_e_i1_i2_i3 =  G_k1_i1 B_k2_i2 B_k3_i3 Z1_e_k1_k2_k3
+      TE("T1_e_i3_k1_k2 = B_k3_i3 Z1_e_k1_k2_k3", T1, B, Z1);
+      TE("T2_e_i2_i3_k1 = B_k2_i2 T1_e_i3_k1_k2", T2, B, T1);
+      TE("Y_e_i1_i2_i3 = G_k1_i1 T2_e_i2_i3_k1", Y, G, T2);
+
+      //Y_e_i1_i2_i3 +=  B_k1_i1 G_k2_i2 B_k3_i3 Z2_e_k1_k2_k3
+      TE("T1_e_i3_k1_k2 = B_k3_i3 Z2_e_k1_k2_k3", T1, B, Z2);
+      TE("T2_e_i2_i3_k1 = G_k2_i2 T1_e_i3_k1_k2", T2, G, T1);
+      TE("Y_e_i1_i2_i3 += B_k1_i1 T2_e_i2_i3_k1", Y, B, T2);
+
+      //Y_e_i1_i2_i3 +=  B_k1_i1 B_k2_i2 G_k3_i3 Z3_e_k1_k2_k3
+      TE("T1_e_i3_k1_k2 = G_k3_i3 Z3_e_k1_k2_k3", T1, G, Z3);
+      TE("T2_e_i2_i3_k1 = B_k2_i2 T1_e_i3_k1_k2", T2, B, T1);
+      TE("Y_e_i1_i2_i3 += B_k1_i1 T2_e_i2_i3_k1", Y, B, T2);
+      TE.EndMultiKernelLaunch();
+   }
+}
+
+void AcroDiffusionIntegrator::MultAdd(const Vector &x, Vector &y) const
+{
+   const_cast<AcroDiffusionIntegrator*>(this)->PAMult(x, y);
+}
+
+void AcroDiffusionIntegrator::MultTransposeAdd(const Vector &x, Vector &y) const
+{
+   mfem_error("Not supported");
+}
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif
@@ -0,0 +1,95 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OMP_ADIFFUSIONINTEG_HPP
+#define MFEM_BACKENDS_OMP_ADIFFUSIONINTEG_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) &&               \
+   defined(MFEM_USE_OMP) &&                     \
+   defined(MFEM_USE_ACROTENSOR)
+
+#include "../../fem/bilininteg.hpp"
+#include "../../fem/fem.hpp"
+#include "vector.hpp"
+#include "fespace.hpp"
+#include "bilinearform.hpp"
+#include "AcroTensor.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+class PAIntegrator : public TensorBilinearFormIntegrator
+{
+protected:
+   Coefficient *Q;
+   FiniteElementSpace *ofes;
+   mfem::FiniteElementSpace *fes;
+   const FiniteElement *fe;
+   const TensorBasisElement *tfe;
+   const IntegrationRule *ir;
+   mfem::Array<int> tDofMap;
+   int GeomType;
+   int FEOrder;
+   bool onGPU;
+   bool hasTensorBasis;
+   int nDim;
+   int nElem;
+   int nDof;
+   int nQuad;
+
+public:
+   PAIntegrator(Coefficient &q, FiniteElementSpace &f);
+   virtual ~PAIntegrator();
+};
+
+class AcroDiffusionIntegrator : public PAIntegrator
+{
+private:
+   acro::TensorEngine TE;
+   int nDof1D;
+   int nQuad1D;
+
+   acro::Tensor B, G;         //Basis and dbasis evaluated on the quad points
+   acro::Tensor W;            //Integration weights
+   mfem::Array<acro::Tensor*> Btil; //Btilde used to compute stiffness matrix
+   acro::Tensor D;            //Product of integration weight, physical consts, and element shape info
+   acro::Tensor S;            //The assembled local stiffness matrices
+   acro::Tensor U, Z, T1, T2; //Intermediate computations for tensor product partial assembly
+   acro::Tensor X, Y;
+
+   void ComputeBTilde();
+
+public:
+   AcroDiffusionIntegrator(BilinearFormIntegrator *integ);
+   AcroDiffusionIntegrator(Coefficient &q, FiniteElementSpace &f);
+   virtual ~AcroDiffusionIntegrator();
+
+   void BatchedPartialAssemble();
+   void BatchedAssembleElementMatrices(DenseTensor &elmats);
+   void ComputeElementMatrices(Vector &elmats);
+   void PAMult(const Vector &x, Vector &y);
+   virtual void MultTransposeAdd(const Vector &x, Vector &y) const;
+   virtual void MultAdd(const Vector &x, Vector &y) const;
+   virtual void ReassembleOperator();
+};
+
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif
+
+#endif
@@ -0,0 +1,128 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include <cstring>
+#include "array.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+PArray *Array::DoClone(bool copy_data, void **buffer,
+                       std::size_t item_size) const
+{
+   Array *new_array = new Array(OmpLayout(), item_size);
+   if (copy_data)
+   {
+      if (!ComputeOnDevice())
+         std::memcpy(new_array->GetData<void>(), data, bytes);
+      else
+      {
+         char *new_data = new_array->GetData<char>();
+         const bool use_target = ComputeOnDevice();
+         const bool use_parallel = Size() > 1000;
+#pragma omp target teams distribute parallel for       \
+   if (target: use_target) if (parallel: use_parallel) \
+   is_device_ptr(new_data)
+         for (std::size_t i = 0; i < bytes; i++) new_data[i] = data[i];
+      }
+   }
+   if (buffer)
+   {
+      *buffer = new_array->GetData<void>();
+   }
+   return new_array;
+}
+
+int Array::DoResize(PLayout &new_layout, void **buffer,
+                    std::size_t item_size)
+{
+   MFEM_ASSERT(dynamic_cast<Layout *>(&new_layout) != NULL,
+               "new_layout is not an OMP Layout");
+   Layout *lt = static_cast<Layout *>(&new_layout);
+   layout.Reset(lt); // Reset() checks if the pointer is the same
+   int err = ResizeData(lt, item_size);
+   if (!err && buffer)
+   {
+      *buffer = GetData<void>();
+   }
+   return err;
+}
+
+void *Array::DoPullData(void *buffer, std::size_t item_size)
+{
+   // called only when Size() != 0
+
+   if (!IsUnifiedMemory() && ComputeOnDevice() && (buffer != NULL))
+   {
+#pragma omp target update from(data)
+      std::memcpy(buffer, data, bytes);
+   }
+   else
+   {
+      buffer = data;
+   }
+
+   return buffer;
+}
+
+void Array::DoFill(const void *value_ptr, std::size_t item_size)
+{
+   // called only when Size() != 0
+
+   switch (item_size)
+   {
+   case sizeof(int):
+      OmpFill((const int *)value_ptr);
+      break;
+   case sizeof(double):
+      OmpFill((const double *)value_ptr);
+      break;
+   default:
+      MFEM_ABORT("item_size = " << item_size << " is not supported");
+   }
+}
+
+void Array::DoPushData(const void *src_buffer, std::size_t item_size)
+{
+   // called only when Size() != 0
+
+   std::memcpy(data, (char *) src_buffer, bytes);
+
+   if ((!IsUnifiedMemory() && ComputeOnDevice()) && (data != src_buffer))
+   {
+#pragma omp target update to(data)
+   }
+}
+
+void Array::DoAssign(const PArray &src, std::size_t item_size)
+{
+   // called only when Size() != 0
+
+   // Note: static_cast can not be used here since PArray is a virtual base
+   //       class.
+   const Array *source = dynamic_cast<const Array *>(&src);
+   MFEM_ASSERT(source != NULL, "invalid source Array type");
+   MFEM_ASSERT(Size() == source->Size(), "");
+   // All arrays from this engine are of the same type, so we can simply check *this and assume the same is used in src.
+   DoPushData(source->GetData<void>(), item_size);
+}
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
@@ -0,0 +1,143 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OMP_ARRAY_HPP
+#define MFEM_BACKENDS_OMP_ARRAY_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "layout.hpp"
+#include "../base/array.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+class Array : public virtual mfem::PArray
+{
+protected:
+   //
+   // Inherited fields
+   //
+   // DLayout layout;
+
+   bool own_data;
+   std::size_t bytes;
+   char *data;
+
+   //
+   // Virtual interface
+   //
+
+   virtual void *DoGetData() const { return (void *) data; }
+
+   virtual PArray *DoClone(bool copy_data, void **buffer,
+                           std::size_t item_size) const;
+
+   virtual int DoResize(PLayout &new_layout, void **buffer,
+                        std::size_t item_size);
+
+   virtual void *DoPullData(void *buffer, std::size_t item_size);
+
+   virtual void DoFill(const void *value_ptr, std::size_t item_size);
+
+   virtual void DoPushData(const void *src_buffer, std::size_t item_size);
+
+   virtual void DoAssign(const PArray &src, std::size_t item_size);
+
+   //
+   // Auxiliary methods
+   //
+
+   inline int ResizeData(const Layout *lt, std::size_t item_size);
+
+   inline bool IsUnifiedMemory() const { return OmpLayout().OmpEngine().UnifiedMemory(); }
+
+   template <typename T>
+   void OmpFill(const T *pval)
+   {
+      T *ptr = (T*) data;
+      T val = *pval;
+      const bool use_target = ComputeOnDevice();
+      const bool use_parallel = (use_target || layout->Size() > 1000);
+      const std::size_t size = layout->Size();
+#pragma omp target teams distribute parallel for        \
+   if (target: use_target)                              \
+   if (parallel: use_parallel) map (to: ptr, val)
+      for (int i = 0; i < size; i++) ptr[i] = val;
+   }
+
+public:
+   Array(Layout &lt, std::size_t item_size)
+      : PArray(lt),
+        own_data(true),
+        bytes(lt.Size() * item_size),
+        data(static_cast<char *>(lt.Alloc(bytes)))
+   {
+#pragma omp target enter data map(alloc:data[:bytes]) if (!IsUnifiedMemory() && ComputeOnDevice())
+   }
+
+   Array(const Array &array)
+      : PArray(array.GetLayout()),
+        own_data(false),
+        bytes(array.bytes),
+        data(array.data) { }
+
+   inline bool ComputeOnDevice() const { return (OmpLayout().OmpEngine().ExecTarget() == Device); }
+
+   virtual ~Array()
+   {
+#pragma omp target exit data map(delete:data[:bytes]) if (!IsUnifiedMemory() && ComputeOnDevice())
+      if (own_data) layout->As<Layout>().Dealloc(data);
+   }
+
+   inline void MakeRef(Array &master);
+
+   Layout &OmpLayout() const
+   { return *static_cast<Layout *>(layout.Get()); }
+};
+
+
+//
+// Inline methods
+//
+
+inline int Array::ResizeData(const Layout *lt, std::size_t item_size)
+{
+   const std::size_t new_bytes = lt->Size() * item_size;
+   if (bytes < new_bytes)
+   {
+#pragma omp target exit data map(delete:data)
+      OmpLayout().Dealloc(data);
+      data = static_cast<char *>(OmpLayout().Alloc(new_bytes));
+      MFEM_VERIFY(data != NULL, "");
+      // If memory allocation fails - an exception is thrown.
+#pragma omp target enter data map(alloc:data[:new_bytes])
+   }
+   return 0;
+}
+
+inline void Array::MakeRef(Array &master)
+{
+   layout = master.layout;
+   data = master.data;
+}
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#endif // MFEM_BACKENDS_OMP_ARRAY_HPP
@@ -0,0 +1,46 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "backend.hpp"
+#include "engine.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+bool Backend::Supports(const std::string &engine_spec) const
+{
+   return true;
+}
+
+mfem::Engine *Create(const std::string &engine_spec)
+{
+   return new Engine(engine_spec);
+}
+
+#ifdef MFEM_USE_MPI
+mfem::Engine *Create(MPI_Comm comm, const std::string &engine_spec)
+{
+   return new Engine(comm, engine_spec);
+}
+#endif
+
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
@@ -0,0 +1,48 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OMP_BACKEND_HPP
+#define MFEM_BACKENDS_OMP_BACKEND_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+// Only the Backend and Engine classes should be exposed through "backend.hpp"
+#include "../base/backend.hpp"
+#include "engine.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+class Backend : public mfem::Backend
+{
+public:
+   virtual ~Backend();
+
+   virtual bool Supports(const std::string &engine_spec) const;
+
+   virtual mfem::Engine *Create(const std::string &engine_spec);
+
+#ifdef MFEM_USE_MPI
+   virtual mfem::Engine *Create(MPI_Comm comm, const std::string &engine_spec);
+#endif
+};
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#endif // MFEM_BACKENDS_OMP_BACKEND_HPP
@@ -0,0 +1,399 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "backend.hpp"
+#include "bilinearform.hpp"
+#include "adiffusioninteg.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+BilinearForm::~BilinearForm()
+{
+   // Make sure all integrators free their data
+   for (int i = 0; i < tbfi.Size(); i++) delete tbfi[i];
+
+   delete element_matrices;
+}
+
+void BilinearForm::TransferIntegrators()
+{
+   mfem::Array<mfem::BilinearFormIntegrator*> &dbfi = *bform->GetDBFI();
+   for (int i = 0; i < dbfi.Size(); i++)
+   {
+      std::string integ_name(dbfi[i]->Name());
+      Coefficient *scal_coeff = dbfi[i]->GetScalarCoefficient();
+      // ConstantCoefficient *const_coeff =
+      //    dynamic_cast<ConstantCoefficient*>(scal_coeff);
+      // // TODO: other types of coefficients ...
+      // double val = const_coeff ? const_coeff->constant : 1.0;
+
+      if (integ_name == "(undefined)")
+      {
+         MFEM_ABORT("BilinearFormIntegrator does not define Name()");
+      }
+      else if (integ_name == "diffusion")
+      {
+         switch (OmpEngine().IntegType())
+         {
+         case Acrotensor:
+            tbfi.Append(new AcroDiffusionIntegrator(*scal_coeff, bform->FESpace()->Get_PFESpace()->As<FiniteElementSpace>()));
+            break;
+         default:
+            mfem_error("integrator is not supported for any MultType");
+            break;
+         }
+      }
+      else
+      {
+         MFEM_ABORT("BilinearFormIntegrator [Name() = " << integ_name
+                    << "] is not supported");
+      }
+   }
+}
+
+void BilinearForm::InitRHS(const mfem::Array<int> &ess_tdof_list,
+                           mfem::Vector &mfem_x, mfem::Vector &mfem_b,
+                           mfem::OperatorHandle &A,
+                           mfem::Vector &mfem_X, mfem::Vector &mfem_B,
+                           int copy_interior) const
+{
+   const mfem::Operator *P = GetProlongation();
+   const mfem::Operator *R = GetRestriction();
+
+   if (P)
+   {
+      // Variational restriction with P
+      mfem_B.Resize(P->InLayout());
+      P->MultTranspose(mfem_b, mfem_B);
+      mfem_X.Resize(R->OutLayout());
+      R->Mult(mfem_x, mfem_X);
+   }
+   else
+   {
+      // rap, X and B point to the same data as this, x and b
+      mfem_X.MakeRef(mfem_x);
+      mfem_B.MakeRef(mfem_b);
+   }
+
+   if (A.Type() != mfem::Operator::ANY_TYPE)
+   {
+      A.EliminateBC(mat_e, ess_tdof_list, mfem_X, mfem_B);
+   }
+
+   if (!copy_interior && ess_tdof_list.Size() > 0)
+   {
+      Vector &X = mfem_X.Get_PVector()->As<Vector>();
+      const Array &constraint_list = ess_tdof_list.Get_PArray()->As<Array>();
+
+      double *X_data = X.GetData<double>();
+      const int* constraint_data = constraint_list.GetData<int>();
+
+      Vector subvec(constraint_list.OmpLayout());
+      double *subvec_data = subvec.GetData<double>();
+
+      const std::size_t num_constraint = constraint_list.Size();
+      const bool use_target = constraint_list.ComputeOnDevice();
+      const bool use_parallel = (use_target || num_constraint > 1000);
+
+      // This operation is a general version of mfem::Vector::SetSubVectorComplement()
+      // {
+#pragma omp target teams distribute parallel for        \
+   map(to: subvec_data, constraint_data, X_data)        \
+   if (target: use_target)                              \
+   if (parallel: use_parallel)
+      for (std::size_t i = 0; i < num_constraint; i++) subvec_data[i] = X_data[constraint_data[i]];
+
+      X.Fill(0.0);
+
+#pragma omp target teams distribute parallel for        \
+   map(to: X_data, constraint_data, subvec_data)        \
+   if (target: use_target)                              \
+   if (parallel: use_parallel)
+      for (std::size_t i = 0; i < num_constraint; i++) X_data[constraint_data[i]] = subvec_data[i];
+      // }
+   }
+
+   if (A.Type() == mfem::Operator::ANY_TYPE)
+   {
+      ConstrainedOperator *A_constrained = static_cast<ConstrainedOperator*>(A.Ptr());
+      A_constrained->EliminateRHS(mfem_X, mfem_B);
+   }
+}
+
+
+bool BilinearForm::Assemble()
+{
+   if (!has_assembled)
+   {
+      TransferIntegrators();
+      has_assembled = true;
+   }
+
+   return true;
+}
+
+void BilinearForm::ComputeElementMatrices()
+{
+   // Only called if performing full assembly
+   const int nelements = trial_fes->GetFESpace()->GetNE();
+   const int trial_ndofs = trial_fes->GetFESpace()->GetFE(0)->GetDof() * trial_fes->GetFESpace()->GetVDim();
+   const int test_ndofs = test_fes->GetFESpace()->GetFE(0)->GetDof() * test_fes->GetFESpace()->GetVDim();
+   const std::size_t length = nelements * trial_ndofs * test_ndofs;
+
+   if (!element_matrices) element_matrices = new mfem::Vector(*(new Layout(OmpEngine(), length)));
+   else element_matrices->Push();
+
+   element_matrices->Fill(0.0);
+   Vector &elmats = element_matrices->Get_PVector()->As<Vector>();
+
+   tbfi[0]->ComputeElementMatrices(elmats);
+
+   if (tbfi.Size() > 1)
+   {
+      for (int k = 1; k < tbfi.Size(); k++)
+      {
+         tbfi[k]->ComputeElementMatrices(elmats);
+      }
+   }
+}
+
+void BilinearForm::FormSystemMatrix(const mfem::Array<int> &ess_tdof_list,
+                                    mfem::OperatorHandle &A)
+{
+   if (A.Type() == mfem::Operator::ANY_TYPE)
+   {
+      // FIXME: Support different test and trial spaces (MixedBilinearForm)
+      const mfem::Operator *P = GetProlongation();
+
+      mfem::Operator *rap = this;
+      if (P != NULL) rap = new mfem::RAPOperator(*P, *this, *P);
+
+      A.Reset(new ConstrainedOperator(rap, ess_tdof_list, (rap != this)));
+
+      return;
+   }
+   else
+   {
+      // ASSUMPTION: some sort of sparse matrix
+      // Compute the local matrices (stored in bform->element_matrices
+      ComputeElementMatrices();
+      bform->AllocateMatrix();
+      mfem::SparseMatrix &mat = bform->SpMat();
+
+      element_matrices->Pull();
+      double *data = element_matrices->GetData();
+
+      const bool skip_zeros = true;
+      mfem::Array<int> tr_vdofs, te_vdofs;
+      for (int i = 0; i < trial_fes->GetFESpace()->GetNE(); i++)
+      {
+         trial_fes->GetFESpace()->GetElementVDofs(i, tr_vdofs);
+         test_fes->GetFESpace()->GetElementVDofs(i, te_vdofs);
+         const mfem::DenseMatrix elmat(data, te_vdofs.Size(), tr_vdofs.Size());
+         mat.AddSubMatrix(te_vdofs, tr_vdofs, elmat, skip_zeros);
+         data += tr_vdofs.Size() * te_vdofs.Size();
+      }
+   }
+
+   if (A.Type() == mfem::Operator::MFEM_SPARSEMAT)
+   {
+      // This works because the FormSystemMatrix call with an explicit
+      // SparseMatrix doesnt call the backend version... This might
+      // change in the future.
+      bform->FormSystemMatrix(ess_tdof_list, static_cast<mfem::SparseMatrix&>(*A.Ptr()));
+   }
+#ifdef MFEM_USE_MPI
+   else if (A.Type() == mfem::Operator::Hypre_ParCSR)
+   {
+      mfem::SparseMatrix &mat = bform->SpMat();
+      mfem::ParBilinearForm *pbform = dynamic_cast<mfem::ParBilinearForm*>(bform);
+
+      const bool skip_zeros = false;
+      mat.Finalize(skip_zeros);
+
+      // -------- FOR SOME VERY AGGREVATING REASON THIS DOESN'T WORK ---------
+      // mfem::ParFiniteElementSpace *pfes = pbform->ParFESpace();
+      // OperatorHandle dA(Operator::Hypre_ParCSR);
+      // // construct a parallel block-diagonal matrix 'A' based on 'a'
+      // dA.MakeSquareBlockDiag(pfes->GetComm(), *engine->MakeLayout(pfes->GlobalTrueVSize()),
+      //                        pfes->GetDofOffsets(), &mat);
+      // OperatorHandle Ph(pfes->Dof_TrueDof_Matrix());
+      // A.MakePtAP(dA, Ph);
+      // A.SetOperatorOwner(false);
+      // -------- BUT THIS DOES ---------
+      pbform->ParallelAssemble(A, &mat);
+      A.SetOperatorOwner(false);
+      // ---------------------
+      mat.Clear();
+      mat_e.Clear();
+      std::cout << "operator size (FormSystemMatrix): " << A.Ptr()->InLayout()->Size() << " " << A.Ptr()->OutLayout()->Size() << std::endl;
+
+      mat_e.EliminateRowsCols(A, ess_tdof_list);
+   }
+#endif
+   else
+   {
+      MFEM_ABORT("Operator::Type is not supported, type = " << A.Type());
+   }
+}
+
+void BilinearForm::FormLinearSystem(const mfem::Array<int> &ess_tdof_list,
+                                    mfem::Vector &x, mfem::Vector &b,
+                                    mfem::OperatorHandle &A, mfem::Vector &X, mfem::Vector &B,
+                                    int copy_interior)
+{
+   FormSystemMatrix(ess_tdof_list, A);
+   std::cout << "operator size (FormLinearSystem 1): " << A.Ptr()->InLayout()->Size() << " " << A.Ptr()->OutLayout()->Size() << std::endl;
+   InitRHS(ess_tdof_list, x, b, A, X, B, copy_interior);
+}
+
+void BilinearForm::RecoverFEMSolution(const mfem::Vector &X, const mfem::Vector &b,
+                                      mfem::Vector &x)
+{
+   const mfem::Operator *P = GetProlongation();
+   if (P)
+   {
+      // Apply conforming prolongation
+      x.Resize(P->OutLayout());
+      P->Mult(X, x);
+   }
+   // Otherwise X and x point to the same data
+}
+
+void BilinearForm::Mult(const mfem::Vector &x, mfem::Vector &y) const
+{
+   trial_fes->ToEVector(x.Get_PVector()->As<Vector>(), x_local);
+
+   y_local.Fill<double>(0.0);
+   for (int i = 0; i < tbfi.Size(); i++) tbfi[i]->MultAdd(x_local, y_local);
+
+   test_fes->ToLVector(y_local, y.Get_PVector()->As<Vector>());
+}
+
+void BilinearForm::MultTranspose(const mfem::Vector &x, mfem::Vector &y) const
+{ mfem_error("mfem::omp::BilinearForm::MultTranspose() is not supported!"); }
+
+
+ConstrainedOperator::ConstrainedOperator(mfem::Operator *A_,
+                                         const mfem::Array<int> &constraint_list_,
+                                         bool own_A_)
+   : Operator(A_->InLayout()->As<Layout>()),
+     A(A_),
+     own_A(own_A_),
+     // FIXME: @dudouit1 has a general fix for this
+     constraint_list(constraint_list_.Get_PArray()->As<Array>()),
+     z(OutLayout()->As<Layout>()),
+     w(OutLayout()->As<Layout>()),
+     mfem_z((z.DontDelete(), z)),
+     mfem_w((w.DontDelete(), w)) { }
+
+void ConstrainedOperator::EliminateRHS(const mfem::Vector &mfem_x, mfem::Vector &mfem_b) const
+{
+   w.Fill<double>(0.0);
+
+   const Vector &x = mfem_x.Get_PVector()->As<Vector>();
+   Vector &b = mfem_b.Get_PVector()->As<Vector>();
+
+   const double *x_data = x.GetData<double>();
+   double *b_data = b.GetData<double>();
+   double *w_data = w.GetData<double>();
+   const int* constraint_data = constraint_list.GetData<int>();
+
+   const std::size_t num_constraint = constraint_list.Size();
+   const bool use_target = constraint_list.ComputeOnDevice();
+   const bool use_parallel = (use_target || num_constraint > 1000);
+
+   if (num_constraint > 0)
+   {
+#pragma omp target teams distribute parallel for             \
+   map(to: w_data, constraint_data, x_data)                  \
+   if (target: use_target)                                   \
+   if (parallel: use_parallel)
+      for (std::size_t i = 0; i < num_constraint; i++)
+         w_data[constraint_data[i]] = x_data[constraint_data[i]];
+   }
+
+   A->Mult(mfem_w, mfem_z);
+
+   b.Axpby<double>(1.0, b, -1.0, z);
+
+   if (num_constraint > 0)
+   {
+#pragma omp target teams distribute parallel for        \
+   map(to: b_data, constraint_data, x_data)             \
+   if (target: use_target)                              \
+   if (parallel: use_parallel)
+      for (std::size_t i = 0; i < num_constraint; i++)
+         b_data[constraint_data[i]] = x_data[constraint_data[i]];
+   }
+}
+
+void ConstrainedOperator::Mult(const mfem::Vector &mfem_x, mfem::Vector &mfem_y) const
+{
+   if (constraint_list.Size() == 0)
+   {
+      A->Mult(mfem_x, mfem_y);
+      return;
+   }
+
+   const Vector &x = mfem_x.Get_PVector()->As<Vector>();
+   Vector &y = mfem_y.Get_PVector()->As<Vector>();
+
+   const double *x_data = x.GetData<double>();
+   double *y_data = y.GetData<double>();
+   double *z_data = z.GetData<double>();
+   const int* constraint_data = constraint_list.GetData<int>();
+
+   const std::size_t num_constraint = constraint_list.Size();
+   const bool use_target = constraint_list.ComputeOnDevice();
+   const bool use_parallel = (use_target || num_constraint > 1000);
+
+   z.Assign<double>(x); // z = x
+
+   // z[constraint_list] = 0.0
+#pragma omp target teams distribute parallel for             \
+   map(to: z_data, constraint_data)                          \
+   if (target: use_target)                                   \
+   if (parallel: use_parallel)
+   for (std::size_t i = 0; i < num_constraint; i++)
+      z_data[constraint_data[i]] = 0.0;
+
+   // y = A * z
+   A->Mult(mfem_z, mfem_y);
+
+   // y[constraint_list] = x[constraint_list]
+#pragma omp target teams distribute parallel for             \
+   map(to: y_data, constraint_data, x_data)                  \
+   if (target: use_target)                                   \
+   if (parallel: use_parallel)
+   for (std::size_t i = 0; i < num_constraint; i++)
+      y_data[constraint_data[i]] = x_data[constraint_data[i]];
+}
+
+// Destructor: destroys the unconstrained Operator @a A if @a own_A is true.
+ConstrainedOperator::~ConstrainedOperator()
+{
+   if (own_A) delete A;
+}
+
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
@@ -0,0 +1,176 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OMP_BILINEARFORM_HPP
+#define MFEM_BACKENDS_OMP_BILINEARFORM_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "fespace.hpp"
+#include "array.hpp"
+#include "vector.hpp"
+#include "../../fem/bilininteg.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+class TensorBilinearFormIntegrator
+{
+public:
+   virtual ~TensorBilinearFormIntegrator() { }
+
+   virtual void ReassembleOperator() = 0;
+
+   virtual void ComputeElementMatrices(Vector &element_matrices)
+   { mfem_error("TensorBilinaerFormIntegrator::ComputeElementMatrices is not overloaded"); }
+
+   virtual void MultAdd(const Vector &x, Vector &y) const = 0;
+
+   virtual void Mult(const Vector &x, Vector &y) const
+   { y.Fill<double>(0.0); MultAdd(x, y); }
+};
+
+/// TODO: doxygen
+class BilinearForm : public mfem::PBilinearForm, public mfem::Operator
+{
+protected:
+   //
+   // Inherited fields
+   //
+   // SharedPtr<const mfem::Engine> engine;
+   // mfem::BilinearForm *bform;
+
+   mfem::Array<TensorBilinearFormIntegrator*> tbfi;
+   bool has_assembled;
+
+   mutable FiniteElementSpace *trial_fes, *test_fes;
+
+   mutable Vector x_local, y_local;
+
+   mfem::Vector *element_matrices;
+   OperatorHandle mat_e;
+
+   void TransferIntegrators();
+
+   void ComputeElementMatrices();
+
+   void InitRHS(const mfem::Array<int> &constraint_list,
+                mfem::Vector &mfem_x, mfem::Vector &mfem_b,
+                mfem::OperatorHandle &A,
+                mfem::Vector &mfem_X, mfem::Vector &mfem_B,
+                int copy_interior = 0) const;
+
+public:
+   /// TODO: doxygen
+   BilinearForm(const Engine &e, mfem::BilinearForm &bf)
+      : mfem::PBilinearForm(e, bf),
+        // FIXME: for mixed bilinear forms
+        mfem::Operator(*bf.FESpace()->GetVLayout().As<Layout>()),
+        tbfi(),
+        has_assembled(false),
+        trial_fes(&bf.FESpace()->Get_PFESpace()->As<FiniteElementSpace>()),
+        test_fes(&bf.FESpace()->Get_PFESpace()->As<FiniteElementSpace>()),
+        x_local(trial_fes->GetELayout()),
+        y_local(test_fes->GetELayout()),
+        element_matrices(NULL),
+        mat_e() { }
+
+   /// Virtual destructor
+   virtual ~BilinearForm();
+
+   /// Return the engine as an OpenMP engine
+   const Engine &OmpEngine() { return static_cast<const Engine&>(*engine); }
+
+   /** @brief Prolongation operator from linear algebra (linear system) vectors,
+       to input vectors for the operator. `NULL` means identity. */
+   virtual const Operator *GetProlongation() const { return trial_fes->GetProlongation(); }
+
+   /** @brief Restriction operator from input vectors for the operator to linear
+       algebra (linear system) vectors. `NULL` means identity. */
+   virtual const Operator *GetRestriction() const  { return test_fes->GetRestriction(); }
+
+   /// Assemble the PBilinearForm.
+   /** This method is called from the method BilinearForm::Assemble() of the
+       associated BilinearForm #bform.
+       @returns True, if the host assembly should be skipped. */
+   virtual bool Assemble();
+
+   /// TODO: doxygen
+   virtual void FormSystemMatrix(const mfem::Array<int> &ess_tdof_list,
+                                 mfem::OperatorHandle &A);
+
+   /// TODO: doxygen
+   virtual void FormLinearSystem(const mfem::Array<int> &ess_tdof_list,
+                                 mfem::Vector &x, mfem::Vector &b,
+                                 mfem::OperatorHandle &A, mfem::Vector &mfem_X, mfem::Vector &mfem_B,
+                                 int copy_interior);
+
+   /// TODO: doxygen
+   virtual void RecoverFEMSolution(const mfem::Vector &mfem_X, const mfem::Vector &mfem_b,
+                                   mfem::Vector &mfem_x);
+
+   /// Operator application: `y=A(x)`.
+   virtual void Mult(const mfem::Vector &mfem_x, mfem::Vector &mfem_y) const;
+
+   /** @brief Action of the transpose operator: `y=A^t(x)`. The default behavior
+       in class Operator is to generate an error. */
+   virtual void MultTranspose(const mfem::Vector &mfem_x, mfem::Vector &mfem_y) const;
+};
+
+class ConstrainedOperator : public mfem::Operator
+{
+   const mfem::Operator *A;
+   const bool own_A;
+   const Array constraint_list;
+   mutable Vector z, w;
+   mutable mfem::Vector mfem_z, mfem_w;
+
+public:
+   ConstrainedOperator(mfem::Operator *A_,
+                       const mfem::Array<int> &constraint_list_,
+                       bool own_A_ = false);
+
+   // Destructor: destroys the unconstrained Operator @a A if @a own_A is true.
+   virtual ~ConstrainedOperator();
+
+   /** @brief Eliminate "essential boundary condition" values specified in @a x
+       from the given right-hand side @a b.
+
+       Performs the following steps:
+
+       z = A((0,x_b));  b_i -= z_i;  b_b = x_b;
+
+       where the "_b" subscripts denote the essential (boundary) indices/dofs of
+       the vectors, and "_i" -- the rest of the entries. */
+   void EliminateRHS(const mfem::Vector &mfem_x, mfem::Vector &mfem_b) const;
+
+   /** @brief Constrained operator action.
+
+       Performs the following steps:
+
+       z = A((x_i,0));  y_i = z_i;  y_b = x_b;
+
+       where the "_b" subscripts denote the essential (boundary) indices/dofs of
+       the vectors, and "_i" -- the rest of the entries. */
+   virtual void Mult(const mfem::Vector &mfem_x, mfem::Vector &mfem_y) const;
+};
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#endif // MFEM_BACKENDS_OMP_BILINEAR_FORM_HPP
@@ -0,0 +1,253 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "engine.hpp"
+#include "array.hpp"
+#include "layout.hpp"
+#include "vector.hpp"
+#include "fespace.hpp"
+#include "bilinearform.hpp"
+#include "memory_resource.hpp"
+
+#include <map>
+
+namespace mfem
+{
+
+namespace omp
+{
+
+typedef std::map<std::string, std::string> keyval_pair_t;
+
+template<typename T, typename P>
+static T remove_if(T beg, T end, P pred)
+{
+   T dest = beg;
+   for (T itr = beg;itr != end; ++itr)
+      if (!pred(*itr))
+         *(dest++) = *itr;
+   return dest;
+}
+
+void parse_token(const std::string &token, std::string &key, std::string &val)
+{
+   std::size_t sep = token.find_first_of(':');
+   if (sep > token.size()) mfem_error("Parse error");
+
+   key = token.substr(0, sep);
+   key.erase(mfem::omp::remove_if(key.begin(), key.end(), isspace), key.end());
+   key.erase(std::remove(key.begin(), key.end(), '\''), key.end());
+
+   val = token.substr(sep+1);
+   val.erase(mfem::omp::remove_if(val.begin(), val.end(), isspace), val.end());
+   val.erase(std::remove(val.begin(), val.end(), '\''), val.end());
+}
+
+keyval_pair_t parse_engine_spec(const std::string &engine_spec)
+{
+   keyval_pair_t map;
+   std::size_t token_extent = 0;
+   std::string key, val;
+   while (token_extent < engine_spec.size())
+   {
+      const std::string remaining(engine_spec, token_extent);
+
+      std::size_t next_comma = remaining.find_first_of(',');
+      if (next_comma == std::string::npos) next_comma = engine_spec.size() - 1;
+
+      const std::string token(remaining, 0, next_comma);
+      parse_token(token, key, val);
+
+      map[key] = val;
+      token_extent += next_comma+1;
+   }
+   return map;
+}
+
+void Engine::Init(const std::string &engine_spec)
+{
+   keyval_pair_t tokens(parse_engine_spec(engine_spec));
+   keyval_pair_t::iterator it;
+
+   it = tokens.find("exec_target");
+   if (it != tokens.end())
+   {
+      if (!std::strncmp(it->second.data(), "device", 6))
+      {
+         exec_target = Device;
+         device_number = 0;
+      }
+      else if (!std::strncmp(it->second.data(), "host", 4))
+      {
+         exec_target = Host;
+         device_number = -1;
+      }
+      else
+      {
+         mfem_error("Parse error. Possible values for exec_target are: ['host', 'device']");
+      }
+   }
+   else
+   {
+      // Default to host if not specified
+      mfem::out << "Did not specify exec_target. Defaulting to host..." << std::endl;
+      exec_target = Host;
+      device_number = -1;
+   }
+
+   it = tokens.find("mem_type");
+   if (it != tokens.end())
+   {
+      if (!std::strncmp(it->second.data(), "unified", 7))
+      {
+#if defined(MFEM_USE_CUDAUM)
+         memory_resources[0] = new UnifiedMemoryResource();
+         unified_memory = true;
+#else
+         mfem_error("Have not compiled support for CUDA unified memory.");
+#endif
+      }
+      else if (!std::strncmp(it->second.data(), "separate", 4))
+      {
+         memory_resources[0] = new NewDeleteMemoryResource();
+         unified_memory = false;
+      }
+      else
+      {
+         mfem_error("Parse error. Possible values for mem_type are: ['separate', 'unified']");
+      }
+   }
+   else {
+      if (exec_target == Device)
+      {
+#if defined(MFEM_USE_CUDAUM)
+         mfem::out << "Did not specify mem_type in engine spec. Defaulting to unified memory..." << std::endl;
+         // Default to unified memory
+         memory_resources[0] = new UnifiedMemoryResource();
+         unified_memory = true;
+#else
+         mfem::out << "Did not specify mem_type in engine spec. Defaulting to standard host memory..." << std::endl;
+         memory_resources[0] = new NewDeleteMemoryResource();
+         unified_memory = false;
+#endif
+      }
+      else
+      {
+         mfem::out << "Did not specify mem_type in engine spec. Defaulting to standard host memory..." << std::endl;
+         memory_resources[0] = new NewDeleteMemoryResource();
+         unified_memory = false;
+      }
+   }
+
+   it = tokens.find("mult_engine");
+   if (it != tokens.end())
+   {
+      if (!std::strncmp(it->second.data(), "acrotensor", 10))
+      {
+         mult_type = Acrotensor;
+      }
+      else
+      {
+         mfem_error("Parse error. Possible values for mem_type are: ['acrotensor'].");
+      }
+   }
+   else
+   {
+      mfem::out << "Did not specify mult_engine in engine spec. Defaulting to Acrotensor..." << std::endl;
+#ifndef MFEM_USE_ACROTENSOR
+      mfem_error("Must compile with Acrotensor support");
+#endif
+      mult_type = Acrotensor;
+   }
+}
+
+Engine::Engine(const std::string &engine_spec)
+   : mfem::Engine(NULL, 1, 1)
+{
+   Init(engine_spec);
+}
+
+#ifdef MFEM_USE_MPI
+Engine::Engine(MPI_Comm _comm, const std::string &engine_spec)
+   : mfem::Engine(NULL, 1, 1)
+{
+   comm = _comm;
+   Init(engine_spec);
+}
+#endif
+
+DLayout Engine::MakeLayout(std::size_t size) const
+{
+   return DLayout(new Layout(*this, size));
+}
+
+DLayout Engine::MakeLayout(const mfem::Array<std::size_t> &offsets) const
+{
+   MFEM_ASSERT(offsets.Size() == 2,
+               "multiple workers are not supported yet");
+   return DLayout(new Layout(*this, offsets.Last()));
+}
+
+DArray Engine::MakeArray(PLayout &layout, std::size_t item_size) const
+{
+   MFEM_ASSERT(dynamic_cast<Layout *>(&layout) != NULL,
+               "invalid input layout");
+   Layout *lt = static_cast<Layout *>(&layout);
+   return DArray(new Array(*lt, item_size));
+}
+
+DVector Engine::MakeVector(PLayout &layout, int type_id) const
+{
+   MFEM_ASSERT(type_id == ScalarId<double>::value, "invalid type_id");
+   MFEM_ASSERT(dynamic_cast<Layout *>(&layout) != NULL,
+               "invalid input layout");
+   Layout *lt = static_cast<Layout *>(&layout);
+   return DVector(new Vector(*lt));
+}
+
+DFiniteElementSpace Engine::MakeFESpace(mfem::FiniteElementSpace &fespace) const
+{
+   return DFiniteElementSpace(new FiniteElementSpace(*this, fespace));
+}
+
+DBilinearForm Engine::MakeBilinearForm(mfem::BilinearForm &bf) const
+{
+   return DBilinearForm(new BilinearForm(*this, bf));
+}
+
+void Engine::AssembleLinearForm(LinearForm &l_form) const
+{
+   /// FIXME - What will the actual parameters be?
+   MFEM_ABORT("FIXME");
+}
+
+mfem::Operator *Engine::MakeOperator(const MixedBilinearForm &mbl_form) const
+{
+   /// FIXME - What will the actual parameters be?
+   MFEM_ABORT("FIXME");
+   return NULL;
+}
+
+mfem::Operator *Engine::MakeOperator(const NonlinearForm &nl_form) const
+{
+   /// FIXME - What will the actual parameters be?
+   MFEM_ABORT("FIXME");
+   return NULL;
+}
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
@@ -0,0 +1,123 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OMP_ENGINE_HPP
+#define MFEM_BACKENDS_OMP_ENGINE_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "../base/engine.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+enum ExecutionTarget { Host, Device };
+
+enum IntegratorType { Acrotensor };
+
+class Engine : public mfem::Engine
+{
+protected:
+   //
+   // Inherited fields
+   //
+   // mfem::Backend *backend;
+#ifdef MFEM_USE_MPI
+   // MPI_Comm comm;
+#endif
+   // int num_mem_res;
+   // int num_workers;
+   // MemoryResource **memory_resources;
+   // double *workers_weights;
+   // int *workers_mem_res;
+
+   enum ExecutionTarget exec_target;
+   bool unified_memory;
+   int device_number;
+   IntegratorType mult_type;
+
+   void Init(const std::string &engine_spec);
+
+public:
+   Engine(const std::string &engine_spec);
+
+#ifdef MFEM_USE_MPI
+   Engine(MPI_Comm comm, const std::string &engine_spec);
+#endif
+
+   virtual ~Engine() { }
+
+   /**
+      @name OMP specific interface, used by other objects in the OMP backend
+   */
+   ///@{
+
+   IntegratorType IntegType() const { return mult_type; }
+
+   ExecutionTarget ExecTarget() const { return exec_target; }
+
+   inline bool UnifiedMemory() const { return unified_memory; }
+
+   void* Malloc(std::size_t bytes) const
+   {
+      return memory_resources[0]->Allocate(bytes, 16);
+   }
+
+   void Dealloc(void *ptr, std::size_t bytes = 0) const
+   {
+      memory_resources[0]->Deallocate(ptr, bytes);
+   }
+
+   ///@}
+   // End: OMP specific interface
+
+   /**
+      @name Virtual interface: finite element data structures and algorithms
+   */
+   ///@{
+
+   virtual DLayout MakeLayout(std::size_t size) const;
+   virtual DLayout MakeLayout(const mfem::Array<std::size_t> &offsets) const;
+
+   virtual DArray MakeArray(PLayout &layout, std::size_t item_size) const;
+
+   virtual DVector MakeVector(PLayout &layout,
+                              int type_id = ScalarId<double>::value) const;
+
+   virtual DFiniteElementSpace MakeFESpace(mfem::FiniteElementSpace &
+                                           fespace) const;
+
+   virtual DBilinearForm MakeBilinearForm(mfem::BilinearForm &bf) const;
+
+   /// FIXME - What will the actual parameters be?
+   virtual void AssembleLinearForm(LinearForm &l_form) const;
+
+   /// FIXME - What will the actual parameters be?
+   virtual mfem::Operator *MakeOperator(const MixedBilinearForm &mbl_form) const;
+
+   /// FIXME - What will the actual parameters be?
+   virtual mfem::Operator *MakeOperator(const NonlinearForm &nl_form) const;
+
+   ///@}
+   // End: Virtual interface
+};
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#endif // MFEM_BACKENDS_OMP_ENGINE_HPP
@@ -0,0 +1,237 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "fespace.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+FiniteElementSpace::FiniteElementSpace(const Engine &e,
+                                       mfem::FiniteElementSpace &fespace)
+   : PFiniteElementSpace(e, fespace),
+     e_layout(e, 0),
+     tensor_offsets(NULL),
+     tensor_indices(NULL),
+     prolongation(NULL),
+     restriction(NULL)
+{
+   std::size_t lsize = 0;
+   for (int e = 0; e < fespace.GetNE(); e++) { lsize += fespace.GetFE(e)->GetDof(); }
+   e_layout.Resize(lsize);
+   // The e_layout will be stored inside multiple shared DLayout objects
+   e_layout.DontDelete();
+}
+
+void FiniteElementSpace::BuildDofMaps()
+{
+   mfem::FiniteElementSpace *mfem_fes = GetFESpace();
+
+   const int local_size = GetELayout().Size();
+   const int global_size = mfem_fes->GetVLayout()->Size();
+   const int vdim = mfem_fes->GetVDim();
+
+   // Now we can allocate and fill the global map
+   tensor_offsets = new mfem::Array<int>(*(new Layout(OmpEngine(), global_size + 1)));
+   tensor_indices = new mfem::Array<int>(*(new Layout(OmpEngine(), local_size)));
+
+   mfem::Array<int> &offsets = *tensor_offsets;
+   mfem::Array<int> &indices = *tensor_indices;
+
+   mfem::Array<int> global_map(local_size);
+   mfem::Array<int> elem_vdof;
+
+   int offset = 0;
+   for (int e = 0; e < mfem_fes->GetNE(); e++)
+   {
+      const FiniteElement *fe = mfem_fes->GetFE(e);
+      const int dofs = fe->GetDof();
+      const TensorBasisElement *tfe = dynamic_cast<const TensorBasisElement *>(fe);
+      const mfem::Array<int> &dof_map = tfe->GetDofMap();
+
+      mfem_fes->GetElementVDofs(e, elem_vdof);
+
+      for (int vd = 0; vd < vdim; vd++)
+         for (int i = 0; i < dofs; i++)
+         {
+            global_map[offset + dofs*vd + i] = elem_vdof[dofs*vd + dof_map[i]];
+         }
+      offset += dofs * vdim;
+   }
+
+   // global_map[i] = index in global vector for local dof i
+   // NOTE: multiple i values will yield same global_map[i] for shared DOF.
+
+   // We want to now invert this map so we have indices[j] = (local dof for global dof j).
+
+   // Zero the offset vector
+   offsets = 0;
+
+   // Keep track of how many local dof point to its global dof
+   // Count how many times each dof gets hit
+   for (int i = 0; i < local_size; i++)
+   {
+      const int g = global_map[i];
+      ++offsets[g + 1];
+   }
+   // Aggregate the offsets
+   for (int i = 1; i <= global_size; i++)
+   {
+      offsets[i] += offsets[i - 1];
+   }
+
+   for (int i = 0; i < local_size; i++)
+   {
+      const int g = global_map[i];
+      indices[offsets[g]++] = i;
+   }
+
+   // Shift the offset vector back by one, since it was used as a
+   // counter above.
+   for (int i = global_size; i > 0; i--)
+   {
+      offsets[i] = offsets[i - 1];
+   }
+   offsets[0] = 0;
+
+   offsets.Push();
+   indices.Push();
+}
+
+/// Convert an E vector to L vector
+void FiniteElementSpace::ToLVector(const Vector &e_vector, Vector &l_vector)
+{
+   if (tensor_indices == NULL) BuildDofMaps();
+
+   if (l_vector.Size() != (std::size_t) GetFESpace()->GetVSize())
+   {
+      l_vector.Resize<double>(GetFESpace()->GetVLayout(), NULL);
+   }
+
+   const int lsize = l_vector.Size();
+   const int *offsets = tensor_offsets->Get_PArray()->As<Array>().GetData<int>();
+   const int *indices = tensor_indices->Get_PArray()->As<Array>().GetData<int>();
+
+   const double *e_data = e_vector.GetData<double>();
+   double *l_data = l_vector.GetData<double>();
+
+   const bool use_target = l_vector.ComputeOnDevice();
+   const bool use_parallel = (use_target || lsize > 1000);
+
+#pragma omp target teams distribute parallel for        \
+   map (to: offsets, indices, l_data, e_data)           \
+   if (target: use_target)                              \
+   if (parallel: use_parallel)
+   for (int i = 0; i < lsize; i++)
+   {
+      const int offset = offsets[i];
+      const int next_offset = offsets[i + 1];
+      double dof_value = 0;
+      for (int j = offset; j < next_offset; j++)
+      {
+         dof_value += e_data[indices[j]];
+      }
+      l_data[i] = dof_value;
+   }
+}
+
+/// Covert an L vector to E vector
+void FiniteElementSpace::ToEVector(const Vector &l_vector, Vector &e_vector)
+{
+   if (tensor_indices == NULL) BuildDofMaps();
+
+   if (e_vector.Size() != (std::size_t) e_layout.Size())
+   {
+      e_vector.Resize<double>(GetELayout(), NULL);
+   }
+
+   const int lsize = l_vector.Size();
+   const int *offsets = tensor_offsets->Get_PArray()->As<Array>().GetData<int>();
+   const int *indices = tensor_indices->Get_PArray()->As<Array>().GetData<int>();
+
+   const double *l_data = l_vector.GetData<double>();
+   double *e_data = e_vector.GetData<double>();
+
+   const bool use_target = l_vector.ComputeOnDevice();
+   const bool use_parallel = (use_target || lsize > 1000);
+
+#pragma omp target teams distribute parallel for         \
+   map (to: offsets, indices, l_data, e_data)            \
+   if (target: use_target)                               \
+   if (parallel: use_parallel)
+   for (int i = 0; i < lsize; i++)
+   {
+      const int offset = offsets[i];
+      const int next_offset = offsets[i + 1];
+      const double dof_value = l_data[i];
+      for (int j = offset; j < next_offset; j++)
+      {
+         e_data[indices[j]] = dof_value;
+      }
+   }
+}
+
+/// Get the finite element space prolongation matrix
+const Operator *FiniteElementSpace::GetProlongation() const
+{
+   // FIXME: This relies on unified memory if using a device other than the CPU
+   if (!prolongation)
+   {
+      Layout &v_layout = GetVLayout();
+      Layout &t_layout = GetTrueVLayout();
+
+      const mfem::Operator *op = GetFESpace()->GetProlongationMatrix();
+      if (!op)
+      {
+         prolongation = new mfem::IdentityOperator(t_layout);
+      }
+      else
+      {
+         prolongation = new BackendOperator(t_layout, v_layout, op);
+      }
+   }
+   return prolongation;
+}
+
+/// Get the finite element space restriction matrix
+const Operator *FiniteElementSpace::GetRestriction() const
+{
+   // FIXME: This relies on unified memory if using a device other than the CPU
+   if (!restriction)
+   {
+      Layout &v_layout = GetVLayout();
+      Layout &t_layout = GetTrueVLayout();
+
+      const mfem::Operator *op = GetFESpace()->GetRestrictionMatrix();
+      if (!op)
+      {
+         restriction = new mfem::IdentityOperator(t_layout);
+      }
+      else
+      {
+         restriction = new BackendOperator(v_layout, t_layout, op);
+      }
+   }
+   return restriction;
+}
+
+
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
@@ -0,0 +1,106 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OMP_FESPACE_HPP
+#define MFEM_BACKENDS_OMP_FESPACE_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "engine.hpp"
+#include "array.hpp"
+#include "vector.hpp"
+#include "../../fem/fem.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+/*
+  Wraps an mfem::Operator that does not contain layout information.
+ */
+class BackendOperator : public mfem::Operator
+{
+   const mfem::Operator *op;
+
+public:
+   BackendOperator(Layout &in_layout, Layout &out_layout,
+                   const mfem::Operator *op_) : Operator(in_layout, out_layout), op(op_) { }
+
+   virtual void Mult(const mfem::Vector &x, mfem::Vector &y) const { op->Mult(x, y); }
+   virtual void MultTranspose(const mfem::Vector &x, mfem::Vector &y) const { op->MultTranspose(x, y); }
+};
+
+/// TODO: doxygen
+class FiniteElementSpace : public mfem::PFiniteElementSpace
+{
+protected:
+   //
+   // Inherited fields
+   //
+   // SharedPtr<const mfem::Engine> engine;
+   // mfem::FiniteElementSpace *fes;
+
+   Layout e_layout;
+
+   mfem::Array<int> *tensor_offsets, *tensor_indices;
+
+   mutable mfem::Operator *prolongation, *restriction;
+
+   void BuildDofMaps();
+
+public:
+   /// Nearly-empty class that stores a pointer to a mfem::FiniteElementSpace instance and the engine
+   FiniteElementSpace(const Engine &e, mfem::FiniteElementSpace &fespace);
+
+   /// Virtual destructor
+   virtual ~FiniteElementSpace()
+   {
+      delete tensor_offsets;
+      delete tensor_indices;
+      delete prolongation;
+      delete restriction;
+   }
+
+   Layout &GetELayout() { return e_layout; }
+
+   Layout &GetVLayout() const
+   { return *fes->GetVLayout().As<Layout>(); }
+
+   Layout &GetTrueVLayout() const
+   { return *fes->GetTrueVLayout().As<Layout>(); }
+
+   /// Return the engine as an OpenMP engine
+   const Engine &OmpEngine() { return static_cast<const Engine&>(*engine); }
+
+   /// Convert an E vector to L vector
+   void ToLVector(const Vector &e_vector, Vector &l_vector);
+
+   /// Covert an L vector to E vector
+   void ToEVector(const Vector &l_vector, Vector &e_vector);
+
+   /// Get the finite element space prolongation matrix
+   const Operator *GetProlongation() const;
+
+   /// Get the finite element space restriction matrix
+   const Operator *GetRestriction() const;
+
+};
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#endif // MFEM_BACKENDS_OMP_FESPACE_HPP
@@ -0,0 +1,40 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "layout.hpp"
+#include "../../general/array.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+void Layout::Resize(std::size_t new_size)
+{
+   size = new_size;
+}
+
+void Layout::Resize(const Array<std::size_t> &offsets)
+{
+   MFEM_ASSERT(offsets.Size() == 2,
+               "multiple workers are not supported yet");
+   size = offsets.Last();
+}
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
@@ -0,0 +1,71 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OMP_LAYOUT_HPP
+#define MFEM_BACKENDS_OMP_LAYOUT_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "../base/layout.hpp"
+#include "engine.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+class Layout : public mfem::PLayout
+{
+protected:
+   //
+   // Inherited fields
+   //
+   // SharedPtr<const mfem::Engine> engine;
+   // std::size_t size;
+
+public:
+   Layout(const Engine &e, std::size_t s = 0) : PLayout(e, s) { }
+
+   const Engine &OmpEngine() const
+   { return *static_cast<const Engine *>(engine.Get()); }
+
+   void *Alloc(std::size_t bytes) const
+   { return OmpEngine().Malloc(bytes); }
+
+   void Dealloc(void *ptr) const
+   { return OmpEngine().Dealloc(ptr); }
+
+   virtual ~Layout() { }
+
+   /**
+      @name Virtual interface
+   */
+   ///@{
+
+   /// Resize the layout
+   virtual void Resize(std::size_t new_size);
+
+   /// Resize the layout based on the given worker offsets
+   virtual void Resize(const Array<std::size_t> &offsets);
+
+   ///@}
+   // End: Virtual interface
+};
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#endif // MFEM_BACKENDS_OMP_LAYOUT_HPP
@@ -0,0 +1,57 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "memory_resource.hpp"
+#include "../../general/error.hpp"
+
+#ifdef MFEM_USE_CUDAUM
+#include "cuda_runtime.h"
+#include "cuda.h"
+#endif
+
+namespace mfem
+{
+
+namespace omp
+{
+
+#ifdef MFEM_USE_CUDAUM
+void *UnifiedMemoryResource::DoAllocate(std::size_t bytes,
+                                        std::size_t alignment)
+{
+   void *p = NULL;
+   if (bytes > 0)
+   {
+      cudaError_t ret = cudaMallocManaged(&p, bytes);
+      MFEM_VERIFY(ret == cudaSuccess, "");
+   }
+   return p;
+}
+
+void UnifiedMemoryResource::DoDeallocate(void *p, std::size_t bytes,
+                                         std::size_t alignment)
+{
+   if (p != NULL)
+   {
+      cudaFree(p);
+   }
+}
+#endif
+
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
@@ -0,0 +1,44 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OMP_MEMORY_RESOURCE_HPP
+#define MFEM_BACKENDS_OMP_MEMORY_RESOURCE_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "../../backends/base/memory_resource.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+/// Polymorphic memory resource. Similar to C++17's std::pmr::memory_resource.
+
+#ifdef MFEM_USE_CUDAUM
+/** @brief Memory resource using unified memory. */
+class UnifiedMemoryResource : public MemoryResource
+{
+protected:
+   virtual void *DoAllocate(std::size_t bytes, std::size_t alignment);
+   virtual void DoDeallocate(void *p, std::size_t bytes, std::size_t alignment);
+};
+#endif
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#endif // MFEM_BACKENDS_OMP_MEMORY_RESOURCE_HPP
@@ -0,0 +1,205 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "vector.hpp"
+#include "../../linalg/vector.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+PVector *Vector::DoVectorClone(bool copy_data, void **buffer,
+                               int buffer_type_id) const
+{
+   MFEM_ASSERT(buffer_type_id == ScalarId<double>::value, "");
+   Vector *new_vector = new Vector(OmpLayout());
+   if (copy_data)
+   {
+      const std::size_t total_size = sizeof(double) * OmpLayout().Size();
+      if (!ComputeOnDevice())
+         std::memcpy(new_vector->GetData<void>(), data, total_size);
+      else
+      {
+         char *new_data = new_vector->GetData<char>();
+#pragma omp target teams distribute parallel for is_device_ptr(new_data)
+         for (std::size_t i = 0; i < total_size; i++) new_data[i] = data[i];
+      }
+   }
+   if (buffer)
+   {
+      *buffer = new_vector->GetData<void>();
+   }
+   return new_vector;
+}
+
+void Vector::DoDotProduct(const PVector &x, void *result,
+                          int result_type_id) const
+{
+   // Can be called when Size() == 0, e.g. when an MPI-parallel vector has a
+   // local size of 0.
+
+   MFEM_ASSERT(result_type_id == ScalarId<double>::value, "");
+   double *res = (double *)result;
+   double local_dot = 0.;
+   MFEM_ASSERT(dynamic_cast<const Vector *>(&x) != NULL, "invalid Vector type");
+   const Vector *xp = static_cast<const Vector *>(&x);
+   MFEM_ASSERT(this->Size() == xp->Size(), "");
+
+   const double *ptr = GetData<double>();
+   const double *xptr = xp->GetData<double>();
+   const std::size_t size = Size();
+
+   if (!ComputeOnDevice())
+   {
+      for (std::size_t i = 0; i < size; i++) local_dot += ptr[i] * xptr[i];
+   }
+   else
+   {
+#pragma omp target teams distribute parallel for map(to: ptr, xptr) reduction(+:local_dot)
+      for (std::size_t i = 0; i < size; i++) local_dot += ptr[i] * xptr[i];
+   }
+
+   *res = local_dot;
+#ifdef MFEM_USE_MPI
+   MPI_Comm comm = OmpLayout().OmpEngine().GetComm();
+   if (comm != MPI_COMM_NULL)
+   {
+      MPI_Allreduce(&local_dot, res, 1, MPI_DOUBLE, MPI_SUM, comm);
+   }
+#endif
+}
+
+void Vector::DoAxpby(const void *a, const PVector &x,
+                     const void *b, const PVector &y,
+                     int ab_type_id)
+{
+   // called only when Size() != 0
+   MFEM_ASSERT(ab_type_id == ScalarId<double>::value, "");
+
+   const double da = *static_cast<const double *>(a);
+   const double db = *static_cast<const double *>(b);
+   MFEM_ASSERT(da == 0.0 || dynamic_cast<const Vector *>(&x) != NULL,
+               "invalid Vector x");
+   MFEM_ASSERT(db == 0.0 || dynamic_cast<const Vector *>(&y) != NULL,
+               "invalid Vector y");
+   const Vector *xp = static_cast<const Vector *>(&x);
+   const Vector *yp = static_cast<const Vector *>(&y);
+
+   MFEM_ASSERT(da == 0.0 || this->Size() == xp->Size(), "");
+   MFEM_ASSERT(db == 0.0 || this->Size() == yp->Size(), "");
+
+   const std::size_t size = Size();
+   const std::size_t critical_size = 1000;
+   const double *xd = xp->GetData<double>();
+   const double *yd = yp->GetData<double>();
+   double *td = GetData<double>();
+
+   const bool use_target = ComputeOnDevice();
+   const bool use_parallel = (use_target || size > critical_size);
+
+   if (da == 0.0)
+   {
+      if (db == 0.0)
+      {
+         OmpFill(&da);
+      }
+      else
+      {
+         if (td == yd)
+         {
+            // *this *= db
+#pragma omp target teams distribute parallel for        \
+   if (target: use_target)                              \
+   if (parallel: use_parallel) map (to: db)
+            for (std::size_t i = 0; i < size; i++) td[i] *= db;
+         }
+         else
+         {
+            // *this = db * y
+#pragma omp target teams distribute parallel for        \
+   if (target: use_target)                              \
+   if (parallel: use_parallel) map (to: yd, db)
+            for (std::size_t i = 0; i < size; i++) td[i] = yd[i] * db;
+         }
+      }
+   }
+   else
+   {
+      if (db == 0.0)
+      {
+         if (td == xd)
+         {
+            // *this *= da
+#pragma omp target teams distribute parallel for        \
+   if (target: use_target)                              \
+   if (parallel: use_parallel) map (to: da)
+            for (std::size_t i = 0; i < size; i++) td[i] *= da;
+         }
+         else
+         {
+            // *this = da * x
+#pragma omp target teams distribute parallel for        \
+   if (target: use_target)                              \
+   if (parallel: use_parallel) map (to: xd, da)
+            for (std::size_t i = 0; i < size; i++) td[i] = xd[i] * da;
+         }
+      }
+      else
+      {
+         MFEM_ASSERT(xd != yd, "invalid input");
+         if (td == xd)
+         {
+            // *this = da * (*this) + db * y
+#pragma omp target teams distribute parallel for        \
+   if (target: use_target)                              \
+   if (parallel: use_parallel) map (to: da, td, db, yd)
+            for (std::size_t i = 0; i < size; i++) td[i] = da * td[i] + db * yd[i];
+         }
+         else if (td == yd)
+         {
+            // *this = da * x + db * (*this)
+#pragma omp target teams distribute parallel for        \
+   if (target: use_target)                              \
+   if (parallel: use_parallel) map (to: da, xd, db, td)
+            for (std::size_t i = 0; i < size; i++) td[i] = da * xd[i] + db * td[i];
+         }
+         else
+         {
+            // *this = da * x + db * y
+#pragma omp target teams distribute parallel for        \
+   if (target: use_target)                              \
+   if (parallel: use_parallel) map (to: da, xd, db, yd)
+            for (std::size_t i = 0; i < size; i++) td[i] = da * xd[i] + db * yd[i];
+         }
+      }
+   }
+}
+
+mfem::Vector Vector::Wrap()
+{
+   return mfem::Vector(*this);
+}
+
+const mfem::Vector Vector::Wrap() const
+{
+   return mfem::Vector(*const_cast<Vector*>(this));
+}
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
@@ -0,0 +1,71 @@
+// Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at
+// the Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights
+// reserved. See file COPYRIGHT for details.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability see http://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the GNU Lesser General Public License (as published by the Free
+// Software Foundation) version 2.1 dated February 1999.
+
+#ifndef MFEM_BACKENDS_OMP_VECTOR_HPP
+#define MFEM_BACKENDS_OMP_VECTOR_HPP
+
+#include "../../config/config.hpp"
+#if defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#include "../base/vector.hpp"
+#include "array.hpp"
+
+namespace mfem
+{
+
+namespace omp
+{
+
+class Vector : virtual public Array, public mfem::PVector
+{
+protected:
+   //
+   // Inherited fields
+   //
+   // DLayout layout;
+   // char *data;
+   // std::size_t size;
+
+   /**
+      @name Virtual interface
+   */
+   ///@{
+
+   virtual PVector *DoVectorClone(bool copy_data, void **buffer,
+                                  int buffer_type_id) const;
+
+   virtual void DoDotProduct(const PVector &x, void *result,
+                             int result_type_id) const;
+
+   virtual void DoAxpby(const void *a, const PVector &x,
+                        const void *b, const PVector &y,
+                        int ab_type_id);
+
+   ///@}
+   // End: Virtual interface
+
+public:
+   Vector(Layout &lt)
+      : PArray(lt), Array(lt, sizeof(double)), PVector(lt)
+   { }
+
+   mfem::Vector Wrap();
+
+   const mfem::Vector Wrap() const;
+};
+
+} // namespace mfem::omp
+
+} // namespace mfem
+
+#endif // defined(MFEM_USE_BACKENDS) && defined(MFEM_USE_OMP)
+
+#endif // MFEM_BACKENDS_OMP_VECTOR_HPP
@@ -74,7 +74,7 @@

 IF (NOT COMMAND PRINT_VAR)
  FUNCTION(PRINT_VAR  VAR_NAME)
-    MESSAGE(STATUS "${VAR_NAME} = '${${VAR_NAME}}'")
+    MESSAGE("-- " "${VAR_NAME} = '${${VAR_NAME}}'")
  ENDFUNCTION()
 ENDIF()

@@ -166,108 +166,17 @@ IF (USE_XSDK_DEFAULTS)
    ENDIF()
    XSDK_HANDLE_LANG_DEFAULTS(Fortran  FC  "FFLAGS;FCFLAGS")
  ENDIF()
-
+  
  # Set XSDK defaults for other CMake variables
-
+  
  IF ("${BUILD_SHARED_LIBS}"  STREQUAL  "")
    MESSAGE("-- " "XSDK: Setting default BUILD_SHARED_LIBS=TRUE")
    SET(BUILD_SHARED_LIBS  TRUE  CACHE  BOOL  "Set by default in XSDK mode")
  ENDIF()
-
+  
  IF ("${CMAKE_BUILD_TYPE}"  STREQUAL  "")
    MESSAGE("-- " "XSDK: Setting default CMAKE_BUILD_TYPE=DEBUG")
    SET(CMAKE_BUILD_TYPE  DEBUG  CACHE  STRING  "Set by default in XSDK mode")
  ENDIF()

 ENDIF()
-
-
-##################################################################################
-#
-#               MFEM-specific additions: set TPL MFEM_USE_* defaults
-#
-##################################################################################
-
-IF (DEFINED TPL_ENABLE_MPI)
-  SET(MFEM_USE_MPI ${TPL_ENABLE_MPI} CACHE BOOL "Enable MPI parallel build" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_METIS)
-  SET(MFEM_USE_METIS ${TPL_ENABLE_METIS} CACHE BOOL "Enable METIS usage" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_GZSTREAM)
-  SET(MFEM_USE_GZSTREAM ${TPL_ENABLE_GZSTREAM} CACHE BOOL "Enable gzstream for compressed data streams." FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_LIBUNWIND)
-  SET(MFEM_USE_LIBUNWIND ${TPL_ENABLE_LIBUNWIND} CACHE BOOL "Enable backtrace for errors." FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_LAPACK)
-  SET(MFEM_USE_LAPACK ${TPL_ENABLE_LAPACK} CACHE BOOL "Enable LAPACK usage" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_SUNDIALS)
-  SET(MFEM_USE_SUNDIALS ${TPL_ENABLE_SUNDIALS} CACHE BOOL "Enable SUNDIALS usage" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_MESQUITE)
-  SET(MFEM_USE_MESQUITE ${TPL_ENABLE_MESQUITE} CACHE BOOL "Enable MESQUITE usage" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_SUITESPARSE)
-  SET(MFEM_USE_SUITESPARSE ${TPL_ENABLE_SUITESPARSE} CACHE BOOL "Enable SuiteSparse usage" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_SUPERLU)
-  SET(MFEM_USE_SUPERLU ${TPL_ENABLE_SUPERLU} CACHE BOOL "Enable SuperLU_DIST usage" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_STRUMPACK)
-  SET(MFEM_USE_STRUMPACK ${TPL_ENABLE_STRUMPACK} CACHE BOOL "Enable STRUMPACK usage" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_GECKO)
-  SET(MFEM_USE_GECKO ${TPL_ENABLE_GECKO} CACHE BOOL "Enable GECKO usage" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_GNUTLS)
-  SET(MFEM_USE_GNUTLS ${TPL_ENABLE_GNUTLS} CACHE BOOL "Enable GNUTLS usage" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_NETCDF)
-  SET(MFEM_USE_NETCDF ${TPL_ENABLE_NETCDF} CACHE BOOL "Enable NETCDF usage" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_PETSC)
-  SET(MFEM_USE_PETSC ${TPL_ENABLE_PETSC} CACHE BOOL "Enable PETSc support." FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_MPFR)
-  SET(MFEM_USE_MPFR ${TPL_ENABLE_MPFR} CACHE BOOL "Enable MPFR usage." FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_SIDRE)
-  SET(MFEM_USE_SIDRE ${TPL_ENABLE_SIDRE} CACHE BOOL "Enable Axom/Sidre usage" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_CONDUIT)
-  SET(MFEM_USE_CONDUIT ${TPL_ENABLE_CONDUIT} CACHE BOOL "Enable Conduit usage" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_PUMI)
-  SET(MFEM_USE_PUMI ${TPL_ENABLE_PUMI} CACHE BOOL "Enable PUMI" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_CUDA)
-  SET(MFEM_USE_CUDA ${TPL_ENABLE_CUDA} CACHE BOOL "Enable CUDA" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_OCCA)
-  SET(MFEM_USE_OCCA ${TPL_ENABLE_OCCA} CACHE BOOL "Enable OCCA" FORCE)
-ENDIF()
-
-IF (DEFINED TPL_ENABLE_RAJA)
-  SET(MFEM_USE_RAJA ${TPL_ENABLE_RAJA} CACHE BOOL "Enable RAJA" FORCE)
-ENDIF()
@@ -25,7 +25,6 @@ set(MFEM_USE_LIBUNWIND @MFEM_USE_LIBUNWIND@)
 set(MFEM_USE_LAPACK @MFEM_USE_LAPACK@)
 set(MFEM_THREAD_SAFE @MFEM_THREAD_SAFE@)
 set(MFEM_USE_OPENMP @MFEM_USE_OPENMP@)
-set(MFEM_USE_LEGACY_OPENMP @MFEM_USE_LEGACY_OPENMP@)
 set(MFEM_USE_MEMALLOC @MFEM_USE_MEMALLOC@)
 set(MFEM_TIMER_TYPE @MFEM_TIMER_TYPE@)
 set(MFEM_USE_SUNDIALS @MFEM_USE_SUNDIALS@)
@@ -40,11 +39,6 @@ set(MFEM_USE_PETSC @MFEM_USE_PETSC@)
 set(MFEM_USE_MPFR @MFEM_USE_MPFR@)
 set(MFEM_USE_SIDRE @MFEM_USE_SIDRE@)
 set(MFEM_USE_CONDUIT @MFEM_USE_CONDUIT@)
-set(MFEM_USE_PUMI @MFEM_USE_PUMI@)
-set(MFEM_USE_MM @MFEM_USE_MM@)
-set(MFEM_USE_CUDA @MFEM_USE_CUDA@)
-set(MFEM_USE_OCCA @MFEM_USE_OCCA@)
-set(MFEM_USE_RAJA @MFEM_USE_RAJA@)

 set(MFEM_CXX_COMPILER "@CMAKE_CXX_COMPILER@")
 set(MFEM_CXX_FLAGS "@CMAKE_CXX_FLAGS@")
@@ -30,12 +30,6 @@
 #define MFEM_VERSION_MINOR (((MFEM_VERSION)/100)%100)
 #define MFEM_VERSION_PATCH ((MFEM_VERSION)%100)

-// MFEM source directory.
-#define MFEM_SOURCE_DIR "@MFEM_SOURCE_DIR@"
-
-// MFEM install directory.
-#define MFEM_INSTALL_DIR "@MFEM_INSTALL_DIR@"
-
 // Description of the git commit used to build MFEM.
 #cmakedefine MFEM_GIT_STRING "@MFEM_GIT_STRING@"

@@ -68,12 +62,9 @@
 // allocation and de-allocation.
 #cmakedefine MFEM_THREAD_SAFE

-// Enable the OpenMP backend.
+// Enable experimental OpenMP support. Requires MFEM_THREAD_SAFE.
 #cmakedefine MFEM_USE_OPENMP

-// [Deprecated] Enable experimental OpenMP support. Requires MFEM_THREAD_SAFE.
-#cmakedefine MFEM_USE_LEGACY_OPENMP
-
 // Enable MFEM functionality based on the Mesquite library.
 #cmakedefine MFEM_USE_MESQUITE

@@ -107,22 +98,6 @@
 // Enable MFEM functionality based on Conduit
 #cmakedefine MFEM_USE_CONDUIT

-// Enable MFEM functionality based on the PUMI library
-#cmakedefine MFEM_USE_PUMI
-
-// Build the GPU/CUDA-enabled version of the MFEM library.
-// Requires a CUDA compiler (nvcc).
-#cmakedefine MFEM_USE_CUDA
-
-// Enable MFEM functionality based on the RAJA library
-#cmakedefine MFEM_USE_RAJA
-
-// Enable MFEM functionality based on the OCCA library
-#cmakedefine MFEM_USE_OCCA
-
-// Enable MFEM's internal Memory Manager (needed e.g. for MFEM_USE_CUDA)
-#cmakedefine MFEM_USE_MM
-
 // Which library functions to use in class StopWatch for measuring time.
 // For a list of the available options, see INSTALL.
 // If not defined, an option is selected automatically.
@@ -131,11 +106,11 @@
 // Enable MFEM functionality based on the SUNDIALS libraries.
 #cmakedefine MFEM_USE_SUNDIALS

+// Windows specific options
+// Macro needed to get defines like M_PI from <cmath>. (Visual Studio C++ only?)
+#cmakedefine _USE_MATH_DEFINES
+
 // Version of HYPRE used for building MFEM.
 #cmakedefine MFEM_HYPRE_VERSION @MFEM_HYPRE_VERSION@

-// Macro defined when PUMI is built with support for the Simmetrix SimModSuite
-// library.
-#cmakedefine MFEM_USE_SIMMETRIX
-
 #endif // MFEM_CONFIG_HEADER
@@ -1,19 +0,0 @@
-# Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at the
-# Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights reserved.
-# See file COPYRIGHT for details.
-#
-# This file is part of the MFEM library. For more information and source code
-# availability see http://mfem.org.
-#
-# MFEM is free software; you can redistribute it and/or modify it under the
-# terms of the GNU Lesser General Public License (as published by the Free
-# Software Foundation) version 2.1 dated February 1999.
-
-# Defines the following variables:
-#   - OCCA_FOUND
-#   - OCCA_LIBRARIES
-#   - OCCA_INCLUDE_DIRS
-
-include(MfemCmakeUtilities)
-mfem_find_package(OCCA OCCA OCCA_DIR "include" "occa.hpp" "lib" "occa"
-  "Paths to headers required by OCCA." "Libraries required by OCCA.")
@@ -1,30 +0,0 @@
-# Copyright (c) 2010, Lawrence Livermore National Security, LLC. Produced at the
-# Lawrence Livermore National Laboratory. LLNL-CODE-443211. All Rights reserved.
-# See file COPYRIGHT for details.
-#
-# This file is part of the MFEM library. For more information and source code
-# availability see http://mfem.org.
-#
-# MFEM is free software; you can redistribute it and/or modify it under the
-# terms of the GNU Lesser General Public License (as published by the Free
-# Software Foundation) version 2.1 dated February 1999.
-
-# Defines the following variables:
-#   - RAJA_FOUND
-#   - RAJA_LIBRARIES
-#   - RAJA_INCLUDE_DIRS
-
-include(MfemCmakeUtilities)
-mfem_find_package(RAJA RAJA RAJA_DIR "include" "RAJA/RAJA.hpp" "lib" "RAJA"
-  "Paths to headers required by RAJA." "Libraries required by RAJA.")
-
-if (NOT RAJA_CONFIG_CMAKE)
-   set(RAJA_CONFIG_CMAKE "${RAJA_DIR}/share/raja/cmake/raja-config.cmake")
-endif()
-if (EXISTS "${RAJA_CONFIG_CMAKE}")
-   include("${RAJA_CONFIG_CMAKE}")
-   if (ENABLE_CUDA AND NOT MFEM_USE_CUDA)
-      message(FATAL_ERROR
-         "RAJA is built with CUDA: MFEM_USE_CUDA=YES is required")
-   endif()
-endif()
@@ -229,17 +229,6 @@ endfunction(mfem_find_component)
 function(mfem_find_package Name Prefix DirVar IncSuffixes Header LibSuffixes
         Lib IncDoc LibDoc)

-  # If we have the TPL_ versions of _INCLUDE_DIRS and _LIBRARIES then set the
-  # standard ${Prefix} versions
-  if (TPL_${Prefix}_INCLUDE_DIRS)
-    set(${Prefix}_INCLUDE_DIRS ${TPL_${Prefix}_INCLUDE_DIRS} CACHE STRING
-      "TPL_${Prefix}_INCLUDE_DIRS was found." FORCE)
-  endif()
-  if (TPL_${Prefix}_LIBRARIES)
-    set(${Prefix}_LIBRARIES ${TPL_${Prefix}_LIBRARIES} CACHE STRING
-      "TPL_${Prefix}_LIBRARIES was found." FORCE)
-  endif()
-
  # Quick return
  if (${Prefix}_FOUND)
    return()
@@ -696,164 +685,3 @@ function(mfem_find_library Name Prefix Lib LibDoc CheckVar CheckSrc)
  endif()

 endfunction(mfem_find_library)
-
-
-#
-#   Function that creates 'config.mk' from 'config.mk.in' for the both the
-#   build- and the install-locations and define install rules for 'config.mk'
-#   and 'test.mk'.
-#
-function(mfem_export_mk_files)
-
-  # Define a few auxiliary variables (not written to 'config.mk')
-  string(TOUPPER "${CMAKE_BUILD_TYPE}" BUILD_TYPE)
-  # CMAKE_SHARED_LIBRARY_RUNTIME_C_FLAG -> '-Wl,-rpath,'
-  set(shared_link_flag ${CMAKE_SHARED_LIBRARY_RUNTIME_C_FLAG})
-  if (NOT shared_link_flag)
-    set(shared_link_flag "-Wl,-rpath,")
-  endif()
-
-  # Convert Boolean vars to YES/NO without writting the values to cache
-  set(CONFIG_MK_BOOL_VARS MFEM_USE_MPI MFEM_USE_METIS MFEM_USE_METIS_5
-      MFEM_DEBUG MFEM_USE_EXCEPTIONS MFEM_USE_GZSTREAM MFEM_USE_LIBUNWIND
-      MFEM_USE_LAPACK MFEM_THREAD_SAFE MFEM_USE_OPENMP MFEM_USE_LEGACY_OPENMP
-      MFEM_USE_MEMALLOC MFEM_USE_SUNDIALS MFEM_USE_MESQUITE MFEM_USE_SUITESPARSE
-      MFEM_USE_SUPERLU MFEM_USE_STRUMPACK MFEM_USE_GECKO MFEM_USE_GNUTLS
-      MFEM_USE_NETCDF MFEM_USE_PETSC MFEM_USE_MPFR MFEM_USE_SIDRE
-      MFEM_USE_CONDUIT MFEM_USE_PUMI MFEM_USE_MM MFEM_USE_CUDA MFEM_USE_OCCA
-      MFEM_USE_RAJA)
-  foreach(var ${CONFIG_MK_BOOL_VARS})
-    if (${var})
-      set(${var} YES)
-    else()
-      set(${var} NO)
-    endif()
-  endforeach()
-  # TODO: Add support for MFEM_USE_CUDA=YES
-  set(MFEM_CXX ${CMAKE_CXX_COMPILER})
-  set(MFEM_CPPFLAGS "")
-  string(STRIP "${CMAKE_CXX_FLAGS_${BUILD_TYPE}} ${CMAKE_CXX_FLAGS}"
-         MFEM_CXXFLAGS)
-  set(MFEM_TPLFLAGS "")
-  foreach(dir ${MFEM_TPL_INCLUDE_DIRS})
-    set(MFEM_TPLFLAGS "${MFEM_TPLFLAGS} -I${dir}")
-  endforeach()
-  # TODO: MFEM_TPLFLAGS: add other TPL flags, in addition to the -I flags.
-  set(MFEM_INCFLAGS "-I\$(MFEM_INC_DIR) \$(MFEM_TPLFLAGS)")
-  set(MFEM_PICFLAG "")
-  if (BUILD_SHARED_LIBS)
-    set(MFEM_PICFLAG "${CMAKE_SHARED_LIBRARY_CXX_FLAGS}")
-  endif()
-  set(MFEM_FLAGS "\$(MFEM_CPPFLAGS) \$(MFEM_CXXFLAGS) \$(MFEM_INCFLAGS)")
-  # TPL link flags: set below
-  set(MFEM_EXT_LIBS "")
-  if (BUILD_SHARED_LIBS)
-    set(MFEM_LIBS "${shared_link_flag}\$(MFEM_LIB_DIR) -L\$(MFEM_LIB_DIR)")
-    set(MFEM_LIBS "${MFEM_LIBS} -lmfem \$(MFEM_EXT_LIBS)")
-    if (APPLE)
-      set(SO_VER ".${mfem_VERSION}${CMAKE_SHARED_LIBRARY_SUFFIX}")
-    else()
-      set(SO_VER "${CMAKE_SHARED_LIBRARY_SUFFIX}.${mfem_VERSION}")
-    endif()
-    set(MFEM_LIB_FILE "\$(MFEM_LIB_DIR)/libmfem${SO_VER}")
-    set(MFEM_SHARED YES)
-    set(MFEM_STATIC NO)
-  else()
-    set(MFEM_LIBS "-L\$(MFEM_LIB_DIR) -lmfem \$(MFEM_EXT_LIBS)")
-    set(MFEM_LIB_FILE "\$(MFEM_LIB_DIR)/libmfem.a")
-    set(MFEM_SHARED NO)
-    set(MFEM_STATIC YES)
-  endif()
-  set(MFEM_BUILD_TAG "${CMAKE_SYSTEM}")
-  set(MFEM_PREFIX "${CMAKE_INSTALL_PREFIX}")
-  # For the next 4 variable, these are the values for the build-tree version of
-  # 'config.mk'
-  set(MFEM_INC_DIR "${PROJECT_BINARY_DIR}")
-  set(MFEM_LIB_DIR "${PROJECT_BINARY_DIR}")
-  set(MFEM_TEST_MK "${PROJECT_SOURCE_DIR}/config/test.mk")
-  set(MFEM_CONFIG_EXTRA "MFEM_BUILD_DIR ?= ${PROJECT_BINARY_DIR}")
-  set(MFEM_MPIEXEC ${MPIEXEC})
-  if (NOT MFEM_MPIEXEC)
-    set(MFEM_MPIEXEC "mpirun")
-  endif()
-  set(MFEM_MPIEXEC_NP ${MPIEXEC_NUMPROC_FLAG})
-  if (NOT MFEM_MPIEXEC_NP)
-    set(MFEM_MPIEXEC_NP "-np")
-  endif()
-  # MFEM_MPI_NP is already set
-  # Define the variable 'MFEM_EXT_LIBS': handle PUMI libs
-  if ("${MFEM_USE_PUMI}" STREQUAL "YES")
-    message(STATUS "simmodsuite_dir = '${SIMMODSUITE_DIR}'")
-    get_target_property(liblist ${PUMI_LIBRARIES} INTERFACE_LINK_LIBRARIES)
-    set(pumi_dep_libs "${liblist}")
-    foreach(pumilib ${liblist})
-      get_target_property(libdeps ${pumilib} INTERFACE_LINK_LIBRARIES)
-      if (NOT "${libdeps}" MATCHES "libdeps-NOTFOUND")
-        list(APPEND pumi_dep_libs ${libdeps})
-      endif()
-    endforeach()
-    list(REMOVE_DUPLICATES pumi_dep_libs)
-    foreach(pumilib ${pumi_dep_libs})
-      unset(lib CACHE)
-      string(REGEX REPLACE "^SCOREC::" "" libname ${pumilib})
-      string(FIND "${pumilib}" ".a" staticlib)
-      string(FIND "${pumilib}" ".so" sharedlib)
-      find_library(lib ${libname} PATHS ${PUMI_DIR}/lib NO_DEFUALT_PATH)
-      if (NOT "${sharedlib}" MATCHES "-1" OR
-          NOT "${staticlib}" MATCHES "-1"   )
-        set(MFEM_EXT_LIBS "${pumilib} ${MFEM_EXT_LIBS}")
-      elseif (NOT "${lib}" MATCHES "lib-NOTFOUND")
-        set(MFEM_EXT_LIBS "${lib} ${MFEM_EXT_LIBS}")
-      elseif ("${lib}" MATCHES "lib-NOTFOUND" AND
-              NOT "${libname}" MATCHES "can" AND
-              NOT "${libname}" MATCHES "pthread")
-        message(FATAL_ERROR "SCOREC lib ${libname} not found")
-      endif()
-    endforeach()
-  endif()
-  # Define the variable 'MFEM_EXT_LIBS': handle other (not PUMI) libs
-  foreach(lib ${TPL_LIBRARIES})
-    get_filename_component(suffix ${lib} EXT)
-    # handle interfaces (e.g., SCOREC::apf)
-    if ("${lib}" MATCHES "SCOREC::.*")
-    elseif (NOT "${lib}" MATCHES "SCOREC::.*" AND "${lib}" MATCHES ".*::.*")
-      message(FATAL_ERROR "***** interface lib found ... exiting *****")
-      # handle static and shared libs
-    elseif ("${suffix}" STREQUAL "${CMAKE_SHARED_LIBRARY_SUFFIX}")
-      get_filename_component(dir ${lib} DIRECTORY)
-      get_filename_component(fullLibName ${lib} NAME_WE)
-      string(REGEX REPLACE "^lib" "" libname ${fullLibName})
-      set(MFEM_EXT_LIBS
-          "${MFEM_EXT_LIBS} ${shared_link_flag}${dir} -L${dir} -l${libname}")
-    else()
-      set(MFEM_EXT_LIBS "${MFEM_EXT_LIBS} ${lib}")
-    endif()
-  endforeach()
-
-  # Create the build-tree version of 'config.mk'
-  configure_file(
-    "${PROJECT_SOURCE_DIR}/config/config.mk.in"
-    "${PROJECT_BINARY_DIR}/config/config.mk")
-  # Copy 'test.mk' from the source-tree to the build-tree
-  configure_file(
-    "${PROJECT_SOURCE_DIR}/config/test.mk"
-    "${PROJECT_BINARY_DIR}/config/test.mk" COPYONLY)
-
-  # Update variables for the install-tree version of 'config.mk'
-  set(MFEM_INC_DIR "${CMAKE_INSTALL_PREFIX}/include")
-  set(MFEM_LIB_DIR "${CMAKE_INSTALL_PREFIX}/lib")
-  set(MFEM_TEST_MK "${CMAKE_INSTALL_PREFIX}/share/mfem/test.mk")
-  set(MFEM_CONFIG_EXTRA "")
-
-  # Create the install-tree version of 'config.mk'
-  configure_file(
-    "${PROJECT_SOURCE_DIR}/config/config.mk.in"
-    "${PROJECT_BINARY_DIR}/config/config-install.mk")
-
-  # Install rules for 'config.mk' and 'test.mk'
-  install(FILES ${PROJECT_SOURCE_DIR}/config/test.mk
-    DESTINATION ${CMAKE_INSTALL_PREFIX}/share/mfem/)
-  install(FILES ${PROJECT_BINARY_DIR}/config/config-install.mk
-    DESTINATION ${CMAKE_INSTALL_PREFIX}/share/mfem/ RENAME config.mk)
-
-endfunction()
@@ -15,9 +15,6 @@
 //
 // Otherwise, use the local file: _config.hpp.

-#ifndef MFEM_CONFIG_HPP
-#define MFEM_CONFIG_HPP
-
 #ifdef MFEM_BUILD_DIR
 #define MFEM_QUOTE(a) #a
 #define MFEM_MAKE_PATH(x,y) MFEM_QUOTE(x/y)
@@ -26,18 +23,6 @@
 #include "_config.hpp"
 #endif

-// Common configuration macros
-
-#if (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 7)) || defined(__clang__)
-#define MFEM_HAVE_GCC_PRAGMA_DIAGNOSTIC
-#endif
-
-// Windows specific options
-#ifdef _WIN32
-// Macro needed to get defines like M_PI from <cmath>. (Visual Studio C++ only?)
-#define _USE_MATH_DEFINES
-#endif
-
 // Check dependencies:

 // Options that require MPI
@@ -51,14 +36,12 @@
 #ifdef MFEM_USE_PETSC
 #error Building with PETSc (MFEM_USE_PETSC=YES) requires MPI (MFEM_USE_MPI=YES)
 #endif
-#ifdef MFEM_USE_PUMI
-#error Building with PUMI (MFEM_USE_PUMI=YES) requires MPI (MFEM_USE_MPI=YES)
-#endif
 #endif // MFEM_USE_MPI not defined

-// CUDA requires the memory manager
-#if defined(MFEM_USE_CUDA) && !defined(MFEM_USE_MM)
-#error Building with CUDA (MFEM_USE_CUDA=YES) requires MFEM_USE_MM=YES
+// Macro that returns its first arg when MFEM_USE_BACKENDS is defined, and its
+// second arg if it is not defined.
+#ifdef MFEM_USE_BACKENDS
+#define MFEM_IF_BACKENDS(x,y) (x)
+#else
+#define MFEM_IF_BACKENDS(x,y) (y)
 #endif
-
-#endif // MFEM_CONFIG_HPP
@@ -30,15 +30,15 @@
 #define MFEM_VERSION_MINOR (((MFEM_VERSION)/100)%100)
 #define MFEM_VERSION_PATCH ((MFEM_VERSION)%100)

+// Description of the git commit used to build MFEM.
+// #define MFEM_GIT_STRING "@MFEM_GIT_STRING@"
+
 // The absolute path of the MFEM source prefix
 // #define MFEM_SOURCE_DIR "@MFEM_SOURCE_DIR@"

 // The absolute path of the MFEM installation prefix
 // #define MFEM_INSTALL_DIR "@MFEM_INSTALL_DIR@"

-// Description of the git commit used to build MFEM.
-// #define MFEM_GIT_STRING "@MFEM_GIT_STRING@"
-
 // Build the parallel MFEM library.
 // Requires an MPI compiler, and the libraries HYPRE and METIS.
 // #define MFEM_USE_MPI
@@ -68,12 +68,9 @@
 // allocation and de-allocation.
 // #define MFEM_THREAD_SAFE

-// Enable the OpenMP backend.
+// Enable experimental OpenMP support. Requires MFEM_THREAD_SAFE.
 // #define MFEM_USE_OPENMP

-// [Deprecated] Enable experimental OpenMP support. Requires MFEM_THREAD_SAFE.
-// #define MFEM_USE_LEGACY_OPENMP
-
 // Internal MFEM option: enable group/batch allocation for some small objects.
 // #define MFEM_USE_MEMALLOC

@@ -118,27 +115,28 @@
 // Enable functionality based on the MPFR library.
 // #define MFEM_USE_MPFR

-// Enable MFEM functionality based on the PUMI library
-// #define MFEM_USE_PUMI
+// Enable the use of MFEM backends.
+// #define MFEM_USE_BACKENDS

-// Build the GPU/CUDA-enabled version of the MFEM library.
-// Requires a CUDA compiler (nvcc).
-// #define MFEM_USE_CUDA
-
-// Enable functionality based on the RAJA library.
-// #define MFEM_USE_RAJA
-
-// Enable functionality based on the OCCA library.
+// Enable the OCCA backend.
 // #define MFEM_USE_OCCA

-// Enable MFEM's internal Memory Manager (needed e.g. for MFEM_USE_CUDA)
-// #define MFEM_USE_MM
+// Enable the OMP backend.
+// #define MFEM_USE_OMP
+
+// Enable use of acrotensor in backends.
+// #define MFEM_USE_ACROTENSOR
+
+// Enable use of unified memory.
+// #define MFEM_USE_CUDAUM
+
+// Windows specific options
+#ifdef _WIN32
+// Macro needed to get defines like M_PI from <cmath>. (Visual Studio C++ only?)
+#define _USE_MATH_DEFINES
+#endif

 // Version of HYPRE used for building MFEM.
 // #define MFEM_HYPRE_VERSION @MFEM_HYPRE_VERSION@

-// Macro defined when PUMI is built with support for the Simmetrix SimModSuite
-// library.
-// #define MFEM_USE_SIMMETRIX
-
 #endif // MFEM_CONFIG_HEADER
@@ -10,41 +10,40 @@
 # Software Foundation) version 2.1 dated February 1999.

 # Variables corresponding to defines in config.hpp (YES, NO, or value)
-MFEM_VERSION           = @MFEM_VERSION@
-MFEM_VERSION_STRING    = @MFEM_VERSION_STRING@
-MFEM_SOURCE_DIR        = @MFEM_SOURCE_DIR@
-MFEM_INSTALL_DIR       = @MFEM_INSTALL_DIR@
-MFEM_GIT_STRING        = @MFEM_GIT_STRING@
-MFEM_USE_MPI           = @MFEM_USE_MPI@
-MFEM_USE_METIS         = @MFEM_USE_METIS@
-MFEM_USE_METIS_5       = @MFEM_USE_METIS_5@
-MFEM_DEBUG             = @MFEM_DEBUG@
-MFEM_USE_EXCEPTIONS    = @MFEM_USE_EXCEPTIONS@
-MFEM_USE_GZSTREAM      = @MFEM_USE_GZSTREAM@
-MFEM_USE_LIBUNWIND     = @MFEM_USE_LIBUNWIND@
-MFEM_USE_LAPACK        = @MFEM_USE_LAPACK@
-MFEM_THREAD_SAFE       = @MFEM_THREAD_SAFE@
-MFEM_USE_LEGACY_OPENMP = @MFEM_USE_LEGACY_OPENMP@
-MFEM_USE_OPENMP        = @MFEM_USE_OPENMP@
-MFEM_USE_MEMALLOC      = @MFEM_USE_MEMALLOC@
-MFEM_TIMER_TYPE        = @MFEM_TIMER_TYPE@
-MFEM_USE_SUNDIALS      = @MFEM_USE_SUNDIALS@
-MFEM_USE_MESQUITE      = @MFEM_USE_MESQUITE@
-MFEM_USE_SUITESPARSE   = @MFEM_USE_SUITESPARSE@
-MFEM_USE_SUPERLU       = @MFEM_USE_SUPERLU@
-MFEM_USE_STRUMPACK     = @MFEM_USE_STRUMPACK@
-MFEM_USE_GECKO         = @MFEM_USE_GECKO@
-MFEM_USE_GNUTLS        = @MFEM_USE_GNUTLS@
-MFEM_USE_NETCDF        = @MFEM_USE_NETCDF@
-MFEM_USE_PETSC         = @MFEM_USE_PETSC@
-MFEM_USE_MPFR          = @MFEM_USE_MPFR@
-MFEM_USE_SIDRE         = @MFEM_USE_SIDRE@
-MFEM_USE_CONDUIT       = @MFEM_USE_CONDUIT@
-MFEM_USE_PUMI          = @MFEM_USE_PUMI@
-MFEM_USE_CUDA          = @MFEM_USE_CUDA@
-MFEM_USE_RAJA          = @MFEM_USE_RAJA@
-MFEM_USE_OCCA          = @MFEM_USE_OCCA@
-MFEM_USE_MM            = @MFEM_USE_MM@
+MFEM_VERSION         = @MFEM_VERSION@
+MFEM_VERSION_STRING  = @MFEM_VERSION_STRING@
+MFEM_GIT_STRING      = @MFEM_GIT_STRING@
+MFEM_SOURCE_DIR      = @MFEM_SOURCE_DIR@
+MFEM_INSTALL_DIR     = @MFEM_INSTALL_DIR@
+MFEM_USE_MPI         = @MFEM_USE_MPI@
+MFEM_USE_METIS       = @MFEM_USE_METIS@
+MFEM_USE_METIS_5     = @MFEM_USE_METIS_5@
+MFEM_DEBUG           = @MFEM_DEBUG@
+MFEM_USE_EXCEPTIONS  = @MFEM_USE_EXCEPTIONS@
+MFEM_USE_GZSTREAM    = @MFEM_USE_GZSTREAM@
+MFEM_USE_LIBUNWIND   = @MFEM_USE_LIBUNWIND@
+MFEM_USE_LAPACK      = @MFEM_USE_LAPACK@
+MFEM_THREAD_SAFE     = @MFEM_THREAD_SAFE@
+MFEM_USE_OPENMP      = @MFEM_USE_OPENMP@
+MFEM_USE_MEMALLOC    = @MFEM_USE_MEMALLOC@
+MFEM_TIMER_TYPE      = @MFEM_TIMER_TYPE@
+MFEM_USE_SUNDIALS    = @MFEM_USE_SUNDIALS@
+MFEM_USE_MESQUITE    = @MFEM_USE_MESQUITE@
+MFEM_USE_SUITESPARSE = @MFEM_USE_SUITESPARSE@
+MFEM_USE_SUPERLU     = @MFEM_USE_SUPERLU@
+MFEM_USE_STRUMPACK   = @MFEM_USE_STRUMPACK@
+MFEM_USE_GECKO       = @MFEM_USE_GECKO@
+MFEM_USE_GNUTLS      = @MFEM_USE_GNUTLS@
+MFEM_USE_NETCDF      = @MFEM_USE_NETCDF@
+MFEM_USE_PETSC       = @MFEM_USE_PETSC@
+MFEM_USE_MPFR        = @MFEM_USE_MPFR@
+MFEM_USE_SIDRE       = @MFEM_USE_SIDRE@
+MFEM_USE_CONDUIT     = @MFEM_USE_CONDUIT@
+MFEM_USE_BACKENDS    = @MFEM_USE_BACKENDS@
+MFEM_USE_OCCA        = @MFEM_USE_OCCA@
+MFEM_USE_OMP         = @MFEM_USE_OMP@
+MFEM_USE_ACROTENSOR  = @MFEM_USE_ACROTENSOR@
+MFEM_USE_CUDAUM      = @MFEM_USE_CUDAUM@

 # Compiler, compile options, and link options
 MFEM_CXX       = @MFEM_CXX@
@@ -72,8 +71,5 @@ MFEM_MPIEXEC    = @MFEM_MPIEXEC@
 MFEM_MPIEXEC_NP = @MFEM_MPIEXEC_NP@
 MFEM_MPI_NP     = @MFEM_MPI_NP@

-# The NVCC compiler cannot link with -x=cu
-MFEM_LINK_FLAGS := $(filter-out -x=cu, $(MFEM_FLAGS))
-
 # Optional extra configuration
@MFEM_CONFIG_EXTRA@
@@ -26,8 +26,7 @@ option(MFEM_USE_GZSTREAM "Enable gzstream for compressed data streams." OFF)
 option(MFEM_USE_LIBUNWIND "Enable backtrace for errors." OFF)
 option(MFEM_USE_LAPACK "Enable LAPACK usage" OFF)
 option(MFEM_THREAD_SAFE "Enable thread safety" OFF)
-option(MFEM_USE_OPENMP "Enable the OpenMP backend" OFF)
-option(MFEM_USE_LEGACY_OPENMP "Enable legacy OpenMP usage" OFF)
+option(MFEM_USE_OPENMP "Enable OpenMP usage" OFF)
 option(MFEM_USE_MEMALLOC "Enable the internal MEMALLOC option." ON)
 option(MFEM_USE_SUNDIALS "Enable SUNDIALS usage" OFF)
 option(MFEM_USE_MESQUITE "Enable MESQUITE usage" OFF)
@@ -41,13 +40,6 @@ option(MFEM_USE_PETSC "Enable PETSc support." OFF)
 option(MFEM_USE_MPFR "Enable MPFR usage." OFF)
 option(MFEM_USE_SIDRE "Enable Axom/Sidre usage" OFF)
 option(MFEM_USE_CONDUIT "Enable Conduit usage" OFF)
-option(MFEM_USE_PUMI "Enable PUMI" OFF)
-option(MFEM_USE_MM "Enable MFEM's memory manager" OFF)
-option(MFEM_USE_CUDA "Enable CUDA" OFF)
-option(MFEM_USE_OCCA "Enable OCCA" OFF)
-option(MFEM_USE_RAJA "Enable RAJA" OFF)
-
-set(MFEM_MPI_NP 4 CACHE STRING "Number of processes used for MPI tests")

 # Allow a user to disable testing, examples, and/or miniapps at CONFIGURE TIME
 # if they don't want/need them (e.g. if MFEM is "just a dependency" and all they
@@ -63,16 +55,13 @@ option(MFEM_ENABLE_MINIAPPS "Build all of the miniapps" OFF)
 # set(CXX g++)
 # set(MPICXX mpicxx)

-# Set the target CUDA architecture
-set(CUDA_ARCH "sm_60" CACHE STRING "Target CUDA architecture.")
-
 set(MFEM_DIR ${CMAKE_CURRENT_SOURCE_DIR})

 # The *_DIR paths below will be the first place searched for the corresponding
 # headers and library. If these fail, then standard cmake search is performed.
 # Note: if the variables are already in the cache, they are not overwritten.

-set(HYPRE_DIR "${MFEM_DIR}/../hypre/src/hypre" CACHE PATH
+set(HYPRE_DIR "${MFEM_DIR}/../hypre-2.10.0b/src/hypre" CACHE PATH
    "Path to the hypre library.")
 # If hypre was compiled to depend on BLAS and LAPACK:
 # set(HYPRE_REQUIRED_PACKAGES "BLAS" "LAPACK" CACHE STRING
@@ -110,7 +99,6 @@ set(SuperLUDist_REQUIRED_PACKAGES "MPI" "BLAS" "ParMETIS" CACHE STRING
 set(STRUMPACK_DIR "${MFEM_DIR}/../STRUMPACK-build" CACHE PATH
    "Path to the STRUMPACK library.")
 # STRUMPACK may also depend on "OpenMP", depending on how it was compiled.
-# Starting with v2.2.0 of STRUMPACK, ParMETIS and Scotch are optional.
 set(STRUMPACK_REQUIRED_PACKAGES "MPI" "MPI_Fortran" "ParMETIS" "METIS"
    "ScaLAPACK" "Scotch/ptscotch/ptscotcherr/scotch/scotcherr" CACHE STRING
    "Additional packages required by STRUMPACK.")
@@ -118,8 +106,7 @@ set(STRUMPACK_REQUIRED_PACKAGES "MPI" "MPI_Fortran" "ParMETIS" "METIS"
 # set(STRUMPACK_REQUIRED_LIBRARIES "gfortran" "mpi_mpifh" CACHE STRING
 #     "Additional libraries required by STRUMPACK.")

-# The Scotch library, required by STRUMPACK <= v2.1.0, optional in STRUMPACK >=
-# v2.2.0.
+# The Scotch library, required by STRUMPACK
 set(Scotch_DIR "${MFEM_DIR}/../scotch_6.0.4" CACHE PATH
    "Path to the Scotch and PT-Scotch libraries.")
 set(Scotch_REQUIRED_PACKAGES "Threads" CACHE STRING
@@ -158,12 +145,6 @@ set(AXOM_DIR "${MFEM_DIR}/../axom" CACHE PATH "Path to the Axom library.")
 set(Axom_REQUIRED_PACKAGES "Conduit/relay" CACHE STRING
    "Additional packages required by Axom.")

-set(PUMI_DIR "${MFEM_DIR}/../pumi-2.1.0" CACHE STRING
-    "Directory where PUMI is installed")
-
-set(OCCA_DIR "${MFEM_DIR}/../occa" CACHE PATH "Path to OCCA")
-set(RAJA_DIR "${MFEM_DIR}/../raja" CACHE PATH "Path to RAJA")
-
 set(BLAS_INCLUDE_DIRS "" CACHE STRING "Path to BLAS headers.")
 set(BLAS_LIBRARIES "" CACHE STRING "The BLAS library.")
 set(LAPACK_INCLUDE_DIRS "" CACHE STRING "Path to LAPACK headers.")
@@ -21,13 +21,8 @@ NOTMAC := $(subst Darwin,,$(shell uname -s))
 CXX = g++
 MPICXX = mpicxx

-BASE_FLAGS  = -std=c++11
-OPTIM_FLAGS = -O3 $(BASE_FLAGS)
-DEBUG_FLAGS = -g $(XCOMPILER)-Wall $(BASE_FLAGS)
-
-# Prefixes for passing flags to the compiler and linker when using CXX or MPICXX
-CXX_XCOMPILER =
-CXX_XLINKER   = -Wl,
+OPTIM_FLAGS = -O3
+DEBUG_FLAGS = -g -Wall

 # Destination location of make install
 # PREFIX = $(HOME)/mfem
@@ -38,41 +33,33 @@ INSTALL = /usr/bin/install
 STATIC = YES
 SHARED = NO

-# CUDA configuration options
-CUDA_CXX = nvcc
-CUDA_ARCH = sm_60
-CUDA_FLAGS = -x=cu --expt-extended-lambda -arch=$(CUDA_ARCH)
-# Prefixes for passing flags to the host compiler and linker when using CUDA_CXX
-CUDA_XCOMPILER = -Xcompiler=
-CUDA_XLINKER   = -Xlinker=
-
 ifneq ($(NOTMAC),)
   AR      = ar
   ARFLAGS = cruv
   RANLIB  = ranlib
-   PICFLAG = $(XCOMPILER)-fPIC
+   PICFLAG = -fPIC
   SO_EXT  = so
   SO_VER  = so.$(MFEM_VERSION_STRING)
-   BUILD_SOFLAGS = -shared $(XLINKER)-soname,libmfem.$(SO_VER)
-   BUILD_RPATH = $(XLINKER)-rpath,$(BUILD_REAL_DIR)
+   BUILD_SOFLAGS = -shared -Wl,-soname,libmfem.$(SO_VER)
+   BUILD_RPATH = -Wl,-rpath,$(BUILD_REAL_DIR)
   INSTALL_SOFLAGS = $(BUILD_SOFLAGS)
-   INSTALL_RPATH = $(XLINKER)-rpath,@MFEM_LIB_DIR@
+   INSTALL_RPATH = -Wl,-rpath,@MFEM_LIB_DIR@
 else
   # Silence "has no symbols" warnings on Mac OS X
   AR      = ar
   ARFLAGS = Scruv
   RANLIB  = ranlib -no_warning_for_no_symbols
-   PICFLAG = $(XCOMPILER)-fPIC
+   PICFLAG = -fPIC
   SO_EXT  = dylib
   SO_VER  = $(MFEM_VERSION_STRING).dylib
-   MAKE_SOFLAGS = $(XLINKER)-dylib,-install_name,$(1)/libmfem.$(SO_VER),\
+   MAKE_SOFLAGS = -Wl,-dylib,-install_name,$(1)/libmfem.$(SO_VER),\
      -compatibility_version,$(MFEM_VERSION_STRING),\
      -current_version,$(MFEM_VERSION_STRING),\
      -undefined,dynamic_lookup
   BUILD_SOFLAGS = $(subst $1 ,,$(call MAKE_SOFLAGS,$(BUILD_REAL_DIR)))
-   BUILD_RPATH = $(XLINKER)-undefined,dynamic_lookup
+   BUILD_RPATH = -Wl,-undefined,dynamic_lookup
   INSTALL_SOFLAGS = $(subst $1 ,,$(call MAKE_SOFLAGS,$(MFEM_LIB_DIR)))
-   INSTALL_RPATH = $(XLINKER)-undefined,dynamic_lookup
+   INSTALL_RPATH = -Wl,-undefined,dynamic_lookup
 endif

 # Set CXXFLAGS to overwrite the default selection of DEBUG_FLAGS/OPTIM_FLAGS
@@ -95,36 +82,37 @@ MFEM_MPI_NP = 4
 # config.hpp. The values below are the defaults for generating the actual values
 # in config.mk and config.hpp.

-MFEM_USE_MPI           = NO
-MFEM_USE_METIS         = $(MFEM_USE_MPI)
-MFEM_USE_METIS_5       = NO
-MFEM_DEBUG             = NO
-MFEM_USE_EXCEPTIONS    = NO
-MFEM_USE_GZSTREAM      = NO
-MFEM_USE_LIBUNWIND     = NO
-MFEM_USE_LAPACK        = NO
-MFEM_THREAD_SAFE       = NO
-MFEM_USE_OPENMP        = NO
-MFEM_USE_LEGACY_OPENMP = NO
-MFEM_USE_MEMALLOC      = YES
-MFEM_TIMER_TYPE        = $(if $(NOTMAC),2,4)
-MFEM_USE_SUNDIALS      = NO
-MFEM_USE_MESQUITE      = NO
-MFEM_USE_SUITESPARSE   = NO
-MFEM_USE_SUPERLU       = NO
-MFEM_USE_STRUMPACK     = NO
-MFEM_USE_GECKO         = NO
-MFEM_USE_GNUTLS        = NO
-MFEM_USE_NETCDF        = NO
-MFEM_USE_PETSC         = NO
-MFEM_USE_MPFR          = NO
-MFEM_USE_SIDRE         = NO
-MFEM_USE_CONDUIT       = NO
-MFEM_USE_PUMI          = NO
-MFEM_USE_CUDA          = NO
-MFEM_USE_RAJA          = NO
-MFEM_USE_OCCA          = NO
-MFEM_USE_MM            = NO
+MFEM_USE_MPI         = NO
+# FIXME: add MFEM_USE_BACKENDS, MFEM_USE_OCCA to the CMake build system
+MFEM_USE_BACKENDS    = YES
+MFEM_USE_OCCA        = YES
+MFEM_USE_METIS       = $(MFEM_USE_MPI)
+MFEM_USE_METIS_5     = NO
+MFEM_DEBUG           = NO
+MFEM_USE_EXCEPTIONS  = NO
+MFEM_USE_GZSTREAM    = NO
+MFEM_USE_LIBUNWIND   = NO
+MFEM_USE_LAPACK      = NO
+MFEM_THREAD_SAFE     = NO
+MFEM_USE_OPENMP      = NO
+MFEM_USE_MEMALLOC    = YES
+MFEM_TIMER_TYPE      = $(if $(NOTMAC),2,4)
+MFEM_USE_SUNDIALS    = NO
+MFEM_USE_MESQUITE    = NO
+MFEM_USE_SUITESPARSE = NO
+MFEM_USE_SUPERLU     = NO
+MFEM_USE_STRUMPACK   = NO
+MFEM_USE_GECKO       = NO
+MFEM_USE_GNUTLS      = NO
+MFEM_USE_NETCDF      = NO
+MFEM_USE_PETSC       = NO
+MFEM_USE_MPFR        = NO
+MFEM_USE_SIDRE       = NO
+MFEM_USE_CONDUIT     = NO
+# FIXME: add MFEM_USE_OMP and MFEM_USE_ACROTENSOR to the CMake build system
+MFEM_USE_OMP         = NO
+MFEM_USE_ACROTENSOR  = NO
+MFEM_USE_CUDAUM      = NO

 # Compile and link options for zlib.
 ZLIB_DIR =
@@ -136,7 +124,7 @@ LIBUNWIND_OPT = -g
 LIBUNWIND_LIB = $(if $(NOTMAC),-lunwind -ldl,)

 # HYPRE library configuration (needed to build the parallel version)
-HYPRE_DIR = @MFEM_DIR@/../hypre/src/hypre
+HYPRE_DIR = @MFEM_DIR@/../hypre-2.10.0b/src/hypre
 HYPRE_OPT = -I$(HYPRE_DIR)/include
 HYPRE_LIB = -L$(HYPRE_DIR)/lib -lHYPRE

@@ -154,8 +142,6 @@ ifeq ($(MFEM_USE_SUPERLU)$(MFEM_USE_STRUMPACK),NONO)
 else
   # ParMETIS: currently needed by SuperLU or STRUMPACK. We assume that METIS 5
   # (included with ParMETIS) is installed in the same location.
-   # Starting with STRUMPACK v2.2.0, ParMETIS is an optional dependency while
-   # METIS is still required.
   METIS_DIR = @MFEM_DIR@/../parmetis-4.0.3
   METIS_OPT = -I$(METIS_DIR)/include
   METIS_LIB = -L$(METIS_DIR)/lib -lparmetis -lmetis
@@ -167,7 +153,7 @@ LAPACK_OPT =
 LAPACK_LIB = $(if $(NOTMAC),-llapack -lblas,-framework Accelerate)

 # OpenMP configuration
-OPENMP_OPT = $(XCOMPILER)-fopenmp
+OPENMP_OPT = -fopenmp
 OPENMP_LIB =

 # Used when MFEM_TIMER_TYPE = 2
@@ -203,8 +189,7 @@ SUPERLU_DIR = @MFEM_DIR@/../SuperLU_DIST_5.1.0
 SUPERLU_OPT = -I$(SUPERLU_DIR)/SRC
 SUPERLU_LIB = -Wl,-rpath,$(SUPERLU_DIR)/SRC -L$(SUPERLU_DIR)/SRC -lsuperlu_dist

-# SCOTCH library configuration (required by STRUMPACK <= v2.1.0, optional in
-# STRUMPACK >= v2.2.0)
+# SCOTCH library configuration (required by STRUMPACK)
 SCOTCH_DIR = @MFEM_DIR@/../scotch_6.0.4
 SCOTCH_OPT = -I$(SCOTCH_DIR)/include
 SCOTCH_LIB = -L$(SCOTCH_DIR)/lib -lptscotch -lptscotcherr -lscotch -lscotcherr\
@@ -293,29 +278,30 @@ SIDRE_LIB = \
   -Wl,-rpath,$(HDF5_DIR)/lib -L$(HDF5_DIR)/lib \
   -lsidre -lslic -laxom_utils -lconduit -lconduit_relay -lhdf5 $(ZLIB_LIB) -ldl

-# PUMI
-# Note that PUMI_DIR is needed -- it is used to check for gmi_sim.h
-PUMI_DIR = @MFEM_DIR@/../pumi-2.1.0
-PUMI_OPT = -I$(PUMI_DIR)/include
-PUMI_LIB = -L$(PUMI_DIR)/lib -lpumi -lcrv -lma -lmds -lapf -lpcu -lgmi -lparma\
-   -llion -lmth -lapf_zoltan -lspr
-
-# CUDA library configuration (currently not needed)
-CUDA_OPT =
-CUDA_LIB =
-
-# OCCA library configuration
 OCCA_DIR = @MFEM_DIR@/../occa
 OCCA_OPT = -I$(OCCA_DIR)/include
-OCCA_LIB = $(XLINKER)-rpath,$(OCCA_DIR)/lib -L$(OCCA_DIR)/lib -locca
+OCCA_LIB = -Wl,-rpath,$(OCCA_DIR)/lib -L$(OCCA_DIR)/lib -locca

-# RAJA library configuration
-RAJA_DIR = @MFEM_DIR@/../raja
-RAJA_OPT = -I$(RAJA_DIR)/include
-ifdef CUB_DIR
-   RAJA_OPT += -I$(CUB_DIR)
+CUDA_DIR = /usr/local/cuda
+CUDAUM_LIB = -L$(CUDA_DIR)/lib64 -lcudart
+CUDAUM_OPT = -I$(CUDA_DIR)/include
+
+OMP_OPT = -qsmp=omp -qoffload
+
+ACROTENSOR_DIR = @MFEM_DIR@/../acrotensor
+ACROTENSOR_OPT = -std=c++11 -I$(ACROTENSOR_DIR)/inc
+ACROTENSOR_LIB = -Wl,-rpath,$(ACROTENSOR_DIR)/lib/shared -L$(ACROTENSOR_DIR)/lib/shared -lacrotensor
+# If Acrotensor was compile with CUDA support, but MFEM_USE_CUDAUM==NO, then uncomment the lines below
+# ACROTENSOR_OPT += -I$(CUDA_DIR)/include
+# ACROTENSOR_LIB += -L$(CUDA_DIR)/lib64 -lcuda -lcudart -lnvrtc
+ifeq ($(MFEM_USE_CUDAUM),YES)
+ifeq ($(MFEM_USE_MPI),YES)
+# HYPRE needs some extra libraries in parallel on the GPU
+# FIXME: We need another solution for compilers other than XL for the
+#        dlink CUDA step, but fixes need to happen elsewhere as well.
+HYPRE_LIB += -qcuda -lcublas -lcusparse -lnvToolsExt
+endif
 endif
-RAJA_LIB = $(XLINKER)-rpath,$(RAJA_DIR)/lib -L$(RAJA_DIR)/lib -lRAJA

 # If YES, enable some informational messages
 VERBOSE = NO
@@ -38,9 +38,6 @@ all: header config-mk
 MPI = $(MFEM_USE_MPI:NO=)
 GHV = get_hypre_version
 GHV_FLAGS = $(subst @MFEM_DIR@,$(if $(MFEM_DIR),$(MFEM_DIR),..),$(HYPRE_OPT))
-SMX = $(if $(MFEM_USE_PUMI:NO=),MFEM_USE_SIMMETRIX)
-SMX_PATH = $(PUMI_DIR)/include/gmi_sim.h
-SMX_FILE = $(subst @MFEM_DIR@,$(if $(MFEM_DIR),$(MFEM_DIR),..),$(SMX_PATH))

 $(GHV): $(SRC)$(GHV).cpp
 	$(call mfem-info, Determining HYPRE version ...)
@@ -55,16 +52,10 @@ get-hypre-version: $(GHV).out
 	$(info HYPRE version: $(MFEM_HYPRE_VERSION)),\
 	$(error Unable to determine HYPRE version))

-check-smx:
-	$(call mfem-info, Checking for Simmetrix header [$(SMX_FILE)] ...)
-	$(eval MFEM_USE_SIMMETRIX:=$(if $(wildcard $(SMX_FILE)),YES,NO))
-	$(call mfem-info, MFEM_USE_SIMMETRIX = $(MFEM_USE_SIMMETRIX))
-	$(eval export MFEM_USE_SIMMETRIX)
-
-header: $(if $(MPI),get-hypre-version,) $(if $(SMX),check-smx)
+header: $(if $(MPI),get-hypre-version,)
 	$(call mfem-info, Writing $(CONFIG_HPP) ...)
 	@set -- && \
-	for def in $${MFEM_DEFINES} $(if $(MPI),MFEM_HYPRE_VERSION) $(SMX); do \
+	for def in $${MFEM_DEFINES} $(if $(MPI),MFEM_HYPRE_VERSION,); do \
 	   eval var=\$$$$def && \
 	   if [ "NO" != "$${var}" ]; then \
 	      set -- "$$@" -e "s|// \(#define $${def} \)|\1|" && \
@@ -18,8 +18,6 @@ run_prefix=""
 run_vg="valgrind --leak-check=full --show-reachable=yes --track-origins=yes"
 run_suffix="-no-vis"
 skip_gen_meshes="yes"
-# filter-out device runs ("no") or non-device runs ("yes"):
-device_runs="no"
 cur_dir="${PWD}"
 mfem_dir="$(cd "$(dirname "$0")"/.. && pwd)"
 mfem_build_dir=""
@@ -32,7 +30,7 @@ groups_serial=(
 '"examples"
   "Examples:"
   "examples"
-   "ex{,1,2}[0-9].cpp"'
+   "ex{,1}[0-9].cpp"'
 #   "ex1.cpp"'
 '"sundials"
   "SUNDIALS examples:"
@@ -46,15 +44,14 @@ groups_serial=(
 '"meshing"
   "Meshing miniapps:"
   "miniapps/meshing"
-   "mobius-strip.cpp klein-bottle.cpp extruder.cpp toroid.cpp
-    mesh-optimizer.cpp"'
+   "mobius-strip.cpp klein-bottle.cpp mesh-optimizer.cpp"'
 )
 # Parallel groups
 groups_parallel=(
 '"examples"
   "Examples:"
   "examples"
-   "ex{,1,2}[0-9]p.cpp"'
+   "ex{,1}[0-9]p.cpp"'
 #   "ex1p.cpp"'
 '"sundials"
   "SUNDIALS examples:"
@@ -84,7 +81,7 @@ groups_all=(
 '"examples"
   "Examples:"
   "examples"
-   "ex\"{,1,2}[0-9]\"{,p}.cpp"'
+   "ex\"{,1}[0-9]\"{,p}.cpp"'
 '"sundials"
   "SUNDIALS examples:"
   "examples/sundials"
@@ -100,8 +97,7 @@ groups_all=(
 '"meshing"
   "Meshing miniapps:"
   "miniapps/meshing"
-   "mobius-strip.cpp klein-bottle.cpp extruder.cpp toroid.cpp
-    {,p}mesh-optimizer.cpp"'
+   "mobius-strip.cpp klein-bottle.cpp {,p}mesh-optimizer.cpp"'
 '"electromagnetics"
   "Electromagnetics miniapps:"
   "miniapps/electromagnetics"
@@ -150,11 +146,6 @@ function extract_sample_runs()
   if [ "$skip_gen_meshes" == "yes" ]; then
      runs=`printf "%s" "$runs" | grep -v ".* -m .*\.gen"`
   fi
-   if [ "$device_runs" == "yes" ]; then
-      runs=`printf "%s" "$runs" | grep ".* -d .*"`
-   else
-      runs=`printf "%s" "$runs" | grep -v ".* -d .*"`
-   fi
   IFS=$'\n'
   runs=(${runs})
   IFS="${old_IFS}"
@@ -176,15 +167,9 @@ function help_message()
      -g <dir> <pattern>
                  Specify explicitly a group (dir + file pattern) to run; This
                  option can be used multiple times to define multiple groups
-      -dev        configure only sample runs using devices.
-                  To test with a parallel build, the parallel (-p|-par) option
-                  should be set first on the command line.
      -v          Enable valgrind
      -o <dir>    [${output_dir:-"<empty>: output goes to stdout"}]
                  If not empty, save output to files inside <dir>
-      -d <dir>    [${mfem_build_dir}]
-                  If <dir> is different from <mfem_dir> then use an
-                  out-of-source build in <dir>
      -j <np>     [${make_j}] Specify the number of jobs to use for building
      -c|-color   Always use colors for the status messages: OK, FAILED, etc
      -b|-built   Do NOT rebuild the library and the executables
@@ -211,8 +196,8 @@ function help_message()
         Their values can also 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>
-         to use an out-of-source build
+         Set this variable to something different from <mfem_dir> to use an
+         out-of-source build

   For other valid variables, see the script source.

@@ -263,7 +248,7 @@ case "$1" in
   -h|-help)
      opt_help="yes"
      ;;
-   -p|-par)
+   -p|-parallel)
      mfem_config="MFEM_USE_MPI=YES MFEM_DEBUG=NO"
      ;;
   -g)
@@ -274,11 +259,6 @@ case "$1" in
      groups=("${groups[@]}" "${test_group}")
      shift 2
      ;;
-   -dev)
-       device_runs="yes"
-       mfem_config+=" MFEM_USE_CUDA=YES MFEM_USE_MM=YES \
-MFEM_USE_OCCA=YES MFEM_USE_RAJA=YES MFEM_USE_OPENMP=YES"
-      ;;
   -v)
      valgrind="yes"
      ;;
@@ -286,10 +266,6 @@ MFEM_USE_OCCA=YES MFEM_USE_RAJA=YES MFEM_USE_OPENMP=YES"
      shift
      output_dir="$1"
      ;;
-   -d)
-      shift
-      mfem_build_dir="$1"
-      ;;
   -j)
      shift
      make_j="-j $1"
@@ -309,10 +285,6 @@ MFEM_USE_OCCA=YES MFEM_USE_RAJA=YES MFEM_USE_OPENMP=YES"
   -n)
      run_prefix="echo"
      ;;
-   -*)
-      echo "unknown option: '$1'"
-      exit 1
-      ;;
   *=*)
      eval $1
      ;;
@@ -14,13 +14,11 @@
 # Colors used below:
 # green    '\033[0;32m'
 # red      '\033[0;31m'
-# yellow   '\033[0;33m'
 # no color '\033[0m'
 COLOR_PRINT = if [ -t 1 ]; then \
   printf $(1)$(2)'\033[0m'$(3); else printf $(2)$(3); fi
 PRINT_OK = $(call COLOR_PRINT,'\033[0;32m',OK,"  ($$1 $$2)\n")
 PRINT_FAILED = $(call COLOR_PRINT,'\033[0;31m',FAILED,"  ($$1 $$2)\n")
-PRINT_SKIP = $(call COLOR_PRINT,'\033[0;33m',SKIP,"\n")

 # Timing support
 define TIMECMD_detect
@@ -38,7 +36,7 @@ export TIME='%es %MkB %x'; \
 set -- $$($(1) $(SHELL) -c "$(2)" 2>&1); while [ "$$#" -gt 3 ]; do shift; done
 endef
 define TIMECMD.NOTGNU
-set -- $$($(1) -l $(SHELL) -c "{ $(2); } > /dev/null 2>&1" 2>&1; echo $$?); \
+set -- $$($(1) -l $(SHELL) -c "$(2)" 2>&1; echo $$?); \
 set -- "$$1"s "$$(($$7/1024))"kB "$${60}"
 endef
 define TIMECMD.BASH
@@ -60,8 +58,7 @@ endif
 # Test runs of the examples/miniapps with parameters - check exit code
 mfem-test = \
   printf "   $(3) [$(2) $(1) ... ]: "; \
-   $(call $(TIMEFUN),$(TIMECMD),$(2) ./$(1) $(if $(5),,-no-vis )$(4) \
-     > $(1).stderr 2>&1); \
+   $(call $(TIMEFUN),$(TIMECMD),$(2) ./$(1) -no-vis $(4) > $(1).stderr 2>&1); \
   if [ "$$3" = 0 ]; \
   then $(PRINT_OK); else $(PRINT_FAILED); cat $(1).stderr; fi; \
   rm -f $(1).stderr; exit $$3
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Johann Dahm	17bfaa3a33	More attempts to make HYPRE sparse matrix work	2018-07-13 16:45:23 -07:00
Johann Dahm	4c93356add	Add AMG solver option	2018-07-11 10:51:39 -07:00
Johann Dahm	114033d21d	Fix HYPRE build dependencies	2018-07-05 15:34:07 -07:00
Johann Dahm	c09f2f5252	Fix hack and generalize engine string parser	2018-07-05 14:57:07 -07:00
Johann Dahm	e81bc38599	Fix memory leak	2018-07-04 14:52:10 -07:00
Johann Dahm	208346aa76	Add DontDelete() on e_layout in FESpace in OMP backend	2018-07-03 14:04:06 -07:00
Johann Dahm	d186f5e4a8	Merge branch 'engines-dev' into acro-engine-dev	2018-07-02 22:39:42 -07:00
Johann Dahm	bdc8e0c16a	Move GetFESpace to mfem::PFiniteElementSpace	2018-07-02 22:38:22 -07:00
Johann Dahm	f021470c43	Merge branch 'engines-dev' into acro-engine-dev	2018-07-02 22:35:06 -07:00
Johann Dahm	3a6ef2cd85	Delete underlying OccaBilinearForm but keep ref to e_layouts	2018-07-02 22:33:47 -07:00
Johann Dahm	46ccd04462	Merge branch 'engines-dev' into acro-engine-dev	2018-07-02 22:11:45 -07:00
Johann Dahm	073380beb4	Prolongation and Restriction - and some temporary hacks	2018-07-02 14:35:42 -07:00
Johann Dahm	1ff9c6fec1	Fix backend Vector::DoDotProduct() and misc	2018-06-28 15:46:22 -07:00
Johann Dahm	0fe5682334	Works on a single GPU	2018-06-28 11:59:19 -07:00
Johann Dahm	824b48b2a6	Fixes for CPU mode	2018-06-22 09:54:42 -07:00
Johann Dahm	52e32a6eb5	Fix compiler errors for host-only compilation	2018-06-20 15:06:01 -07:00
Johann Dahm	f849ee7aa5	Initial work toward an OpenMP/Acrotensor backend	2018-06-19 10:28:21 -07:00
Veselin Dobrev	d49258aaaa	Update okl files for OCCA v1.0.0.	2018-06-14 16:48:43 -07:00
Veselin Dobrev	abfa3bc631	Make DevExtension::Pull() a const method -- this makes it possible to call Pull() for the input mfem::Vector inside Operator::Mult() implementation. Simplify the implementation of mfem::occa::ProlongationOperator which still needs OCCA operator(s) for the MPI case.	2018-06-13 20:11:06 -07:00
Veselin Dobrev	4f11a7194d	Run 'make style'. A few small tweaks in ex1d and ex1pd.	2018-06-12 18:10:38 -07:00
Johann Dahm	bfaf7a8da9	Fix ex1d and correct tolerance on CG solver in parallel	2018-06-07 11:45:58 -07:00
Johann Dahm	65e1de0f2b	OCCA backend works in parallel on ex1pd	2018-06-04 10:37:38 -07:00
Johann Dahm	8980df563e	Toward parallel engines support * Add engine in parfespace and parbilinearform * Add parallel ex1 with engines	2018-05-22 10:05:15 -07:00
Veselin Dobrev	b79c9fc31c	When updating libmfem.a remove the old archive first. This prevents an issue when updating the archive in cases when we have object files with the same name from different directories.	2018-05-15 20:09:37 -07:00
Veselin Dobrev	65e1fe7365	Add new files forgotten in the last commit	2018-05-15 18:37:37 -07:00
Veselin Dobrev	60834fc386	Add a custom "mfem-occa://" file-opener With this file-opener there is no need for the makefile okl-hacks. The path specification "mfem-occa://a/b.ext" will be resolved by searching for the file 'a/b.ext' in a list of directories, currently defined from the list (only existing directories are added to the list): * All directories specified by the environment variable MFEM_OCCA_OKL_PATH - list of directories separated by ':'. * The MFEM source directory + "/backends/occa". * The MFEM install prefix + "/lib/mfem/occa". The order of the above list also defines the directory search order.	2018-05-15 18:22:34 -07:00
Veselin Dobrev	7a1b66b203	Update some doxygen documentation and comments	2018-05-15 18:04:12 -07:00
Johann Dahm	7d0d3bb6e1	Fix typo in simplex gpuHighOrder file	2018-05-11 14:17:07 -07:00
Veselin Dobrev	788bcea676	Adding an acknowledgment: The first MFEM backend (backends/occa) is based on the OCCA library with code taken from the previous work in the branch occa-dev, developed by David Medina (@dmed256) and others.	2018-05-09 19:06:41 -07:00
Veselin Dobrev	2c3ce1b4db	Initial commit of new draft MFEM extensions to support GPUs and other accelerators.	2018-05-09 17:11:38 -07:00