Merge branch 'nlconvpa' into FEBioMFEMiFSI-gpu

Fix in the SPDE documentation - a factor of 2

.github/workflows/README.md

@@ -29,16 +29,12 @@ Runs a number of static repository-level sanity checks.
 
 - `branch-history` guards against accidental commits of large files using the `--history` option of the `config/githooks/pre-push` script.
 
-## `mfem-analysis.yml` (`build-analysis`)
-
-Checks if the code builds and satisfies minimal requirements.
-
-- `gitignore` builds hypre, METIS, and MFEM using `mfem/github-actions/build-hypre`, `mfem/github-actions/build-metis`, and `mfem/github-actions/build-mfem` and checks for correct `.gitignore` settings by running the `tests/scripts/gitignore` script.
-
 ## `builds-and-tests.yml`
 
 Runs a matrix of builds and tests runs with different compilers, OS, mfem/hypre settings, etc. Also processes and upload Codecov reports.
 
+One matrix job runs `tests/scripts/gitignore` after `make test-noclean` to check generated artifacts against `.gitignore`.
+
 Uses the following GitHub Actions from <https://github.com/mfem/github-actions>:
 
 - `mfem/github-actions/build-hypre`

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

@@ -111,6 +111,7 @@ jobs:
             build-system: make
             hypre-target: int64
             precision: fp64
+            gitignore-check: YES
           - os: ubuntu-latest
             target: opt
             codecov: NO
@@ -317,7 +318,13 @@ jobs:
       - name: tests
         if: matrix.build-system == 'make' && (matrix.target == 'opt' || matrix.os == 'ubuntu-latest')
         run: |
-          cd ${{ env.MFEM_TOP_DIR }} && make test
+          cd ${{ env.MFEM_TOP_DIR }}
+          if [[ "${{ matrix.gitignore-check }}" == "YES" ]]; then
+            make test-noclean
+          else
+            make test
+          fi
+        shell: bash
 
       - name: cmake checks
         if: matrix.build-system == 'cmake' && matrix.target == 'dbg'
@@ -369,3 +376,9 @@ jobs:
           directories: "fem general linalg mesh"
         env:
           CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
+
+      - name: gitignore
+        if: matrix.gitignore-check == 'YES'
+        run: |
+          cd ${{ env.MFEM_TOP_DIR }}/tests/scripts
+          ./runtest gitignore

.github/workflows/codeql-analysis.yml

@@ -14,9 +14,19 @@ name: "Static Analysis"
 on:
   push:
     branches: ["master", "next"]
+    paths-ignore: &docs-only-paths
+      - "**/*.md"
+      - "doc/**"
+      - ".binder/**"
+      - "CITATION.cff"
+      - "LICENSE"
+      - "NOTICE"
+      - "CHANGELOG"
+      - "INSTALL"
   pull_request:
     # The branches below must be a subset of the branches above
     branches: ["master"]
+    paths-ignore: *docs-only-paths
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}

.github/workflows/mfem-analysis.yml

@@ -1,102 +0,0 @@
-# Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
-# at the Lawrence Livermore National Laboratory. All Rights reserved. See files
-# LICENSE and NOTICE for details. LLNL-CODE-806117.
-#
-# This file is part of the MFEM library. For more information and source code
-# availability visit https://mfem.org.
-#
-# MFEM is free software; you can redistribute it and/or modify it under the
-# terms of the BSD-3 license. We welcome feedback and contributions, see file
-# CONTRIBUTING.md for details.
-
-name: "Build Analysis"
-
-permissions:
-  actions: write
-
-on:
-  push:
-    branches:
-      - master
-      - next
-  pull_request:
-  workflow_dispatch:
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
-
-env:
-  HYPRE_ARCHIVE: v2.19.0.tar.gz
-  HYPRE_TOP_DIR: hypre-2.19.0
-  METIS_ARCHIVE: metis-4.0.3.tar.gz
-  METIS_TOP_DIR: metis-4.0.3
-  COVERAGE_ENV: mfem-coverage
-  MFEM_ACTIONS_VERSION: v2.7
-
-jobs:
-  gitignore:
-    runs-on: ubuntu-latest
-
-    steps:
-      - name: checkout MFEM
-        uses: actions/checkout@v6
-        with:
-          path: mfem
-
-      - name: Get MPI (Linux)
-        run: |
-          sudo apt-get install openmpi-bin libopenmpi-dev
-          export OMPI_MCA_rmaps_base_oversubscribe=1
-
-      - name: Cache Hypre Install
-        id: hypre-cache
-        uses: actions/cache@v5
-        with:
-          path: ${{ env.HYPRE_TOP_DIR }}
-          key: ${{ runner.os }}-ompi-build-${{ env.HYPRE_TOP_DIR }}-int32-fp64-${{ env.MFEM_ACTIONS_VERSION }}
-
-      - name: Get Hypre
-        if: steps.hypre-cache.outputs.cache-hit != 'true'
-        uses: mfem/github-actions/build-hypre@v2.7
-        with:
-          archive: ${{ env.HYPRE_ARCHIVE }}
-          dir: ${{ env.HYPRE_TOP_DIR }}
-          target: int32
-          precision: fp64
-
-      - name: Cache Metis Install
-        id: metis-cache
-        uses: actions/cache@v5
-        with:
-          path: ${{ env.METIS_TOP_DIR }}
-          key: ${{ runner.os }}-build-${{ env.METIS_TOP_DIR }}-${{ env.MFEM_ACTIONS_VERSION }}
-
-      - name: Install Metis
-        if: steps.metis-cache.outputs.cache-hit != 'true'
-        uses: mfem/github-actions/build-metis@v2.7
-        with:
-          archive: ${{ env.METIS_ARCHIVE }}
-          dir: ${{ env.METIS_TOP_DIR }}
-
-      # MFEM build and test
-      - name: build-mfem
-        uses: mfem/github-actions/build-mfem@v2.7
-        with:
-          os: ${{ runner.os }}
-          target: opt
-          codecov: NO
-          mpi: par
-          build-system: make
-          hypre-dir: ${{ env.HYPRE_TOP_DIR }}
-          metis-dir: ${{ env.METIS_TOP_DIR }}
-          mfem-dir: mfem
-
-      - name: test (no clean)
-        run: |
-          cd mfem && make test-noclean
-
-      - name: gitignore
-        run: |
-          cd mfem/tests/scripts
-          ./runtest gitignore

.github/workflows/sanitizers.yml

@@ -17,7 +17,17 @@ permissions:
 on:
   push:
     branches: ["master", "next"]
+    paths-ignore: &docs-only-paths
+      - "**/*.md"
+      - "doc/**"
+      - ".binder/**"
+      - "CITATION.cff"
+      - "LICENSE"
+      - "NOTICE"
+      - "CHANGELOG"
+      - "INSTALL"
   pull_request:
+    paths-ignore: *docs-only-paths
   workflow_dispatch:
 
 concurrency:

CHANGELOG

@@ -15,6 +15,10 @@ Version 4.9.1 (development)
 
 Discretization improvements
 ---------------------------
+- Improved FindPointsGSLIB surface mesh capability with support for simplices
+  and an option to specify axis-aligned bounding box padding for near-surface
+  point queries.
+
 - Added GPU-enabled partial assembly for simplicial Bernstein H1 basis based on
   ragged tensor algorithms (see DOI: 10.1137/11082539X) for mass and diffusion
   integrators.
@@ -66,6 +70,17 @@ Linear and nonlinear solvers
 
 GPU computing
 -------------
+- Added PA gradient and diagonal support for VectorConvectionNLFIntegrator
+ (AssembleGradPA, AddMultGradPA, AssembleGradDiagonalPA).
+
+- Improved partial assembly for VectorDivergenceIntegrator with shared-memory
+  kernels, kernel registration, and transpose support.
+
+- Improved partial-assembly diagonal kernels for VectorMassIntegrator (shared-
+  memory specializations) and ElasticityIntegrator (no scratch Q-vector).
+
+- Added device assembly support for 3D H(curl) VectorFEDomainLFIntegrator.
+
 - Added NVIDIA cuDSS library interface. Implementation examples have been
   added to ex1 and ex1p. See https://developer.nvidia.com/cudss for more
   details. Supported versions >= 0.6.0.

CONTRIBUTING.md

@@ -6,7 +6,6 @@
 <a href="https://github.com/mfem/mfem/blob/master/LICENSE"><img alt="License" src="https://img.shields.io/badge/License-BSD-blue.svg"></a>
 <a href="https://github.com/mfem/mfem/releases/latest"><img alt="GitHub release" src="https://img.shields.io/github/v/release/mfem/mfem"></a>
 <a href="https://github.com/mfem/mfem/actions/workflows/repo-check.yml?query=branch%3Amaster"><img alt="Repo check" src="https://github.com/mfem/mfem/actions/workflows/repo-check.yml/badge.svg?branch=master"></a>
-<a href="https://github.com/mfem/mfem/actions/workflows/mfem-analysis.yml?query=branch%3Amaster"><img alt="Build Analysis" src="https://github.com/mfem/mfem/actions/workflows/mfem-analysis.yml/badge.svg?branch=master"></a>
 <a href="https://github.com/mfem/mfem/actions/workflows/builds-and-tests.yml?query=branch%3Amaster"><img alt="Builds and Tests" src="https://github.com/mfem/mfem/actions/workflows/builds-and-tests.yml/badge.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="https://docs.mfem.org/html/index.html"><img alt="Documentation" src="https://img.shields.io/badge/code-documented-brightgreen.svg"></a>

config/cmake/modules/FindSuiteSparse.cmake

@@ -22,15 +22,15 @@ include(MfemCmakeUtilities)
 mfem_find_package(SuiteSparse SuiteSparse SuiteSparse_DIR "" "" "" ""
   "Paths to headers required by SuiteSparse."
   "Libraries required by SuiteSparse."
-  ADD_COMPONENT "UMFPACK" "include;suitesparse" umfpack.h "lib" umfpack
-  ADD_COMPONENT "KLU" "include;suitesparse" klu.h "lib" klu
-  ADD_COMPONENT "AMD" "include;suitesparse" amd.h "lib" amd
-  ADD_COMPONENT "BTF" "include;suitesparse" btf.h "lib" btf
-  ADD_COMPONENT "CHOLMOD" "include;suitesparse" cholmod.h "lib" cholmod
-  ADD_COMPONENT "COLAMD" "include;suitesparse" colamd.h "lib" colamd
-  ADD_COMPONENT "CAMD" "include;suitesparse" camd.h "lib" camd
-  ADD_COMPONENT "CCOLAMD" "include;suitesparse" ccolamd.h "lib" ccolamd
-  ADD_COMPONENT "config" "include;suitesparse" SuiteSparse_config.h "lib"
+  ADD_COMPONENT "UMFPACK" "include;include/suitesparse;suitesparse" umfpack.h "lib" umfpack
+  ADD_COMPONENT "KLU" "include;include/suitesparse;suitesparse" klu.h "lib" klu
+  ADD_COMPONENT "AMD" "include;include/suitesparse;suitesparse" amd.h "lib" amd
+  ADD_COMPONENT "BTF" "include;include/suitesparse;suitesparse" btf.h "lib" btf
+  ADD_COMPONENT "CHOLMOD" "include;include/suitesparse;suitesparse" cholmod.h "lib" cholmod
+  ADD_COMPONENT "COLAMD" "include;include/suitesparse;suitesparse" colamd.h "lib" colamd
+  ADD_COMPONENT "CAMD" "include;include/suitesparse;suitesparse" camd.h "lib" camd
+  ADD_COMPONENT "CCOLAMD" "include;include/suitesparse;suitesparse" ccolamd.h "lib" ccolamd
+  ADD_COMPONENT "config" "include;include/suitesparse;suitesparse" SuiteSparse_config.h "lib"
     suitesparseconfig)
 
 if (SuiteSparse_FOUND AND METIS_VERSION_5)

fem/CMakeLists.txt

@@ -57,6 +57,8 @@ set(SRCS
   integ/lininteg_domain_grad.cpp
   integ/lininteg_domain_vectorfe.cpp
   integ/nonlininteg_vecconvection_pa.cpp
+  integ/nonlininteg_vecconvection_pa_diag.cpp
+  integ/nonlininteg_vecconvection_pa_grad.cpp
   integ/nonlininteg_vecconvection_mf.cpp
   coefficient.cpp
   complex_fem.cpp
@@ -133,7 +135,7 @@ set(SRCS
   tmop/assemble/diag2.cpp
   tmop/assemble/grad2_limit.cpp
   tmop/assemble/grad2.cpp
-  tmop/assemble/diag3_limit.cpp  
+  tmop/assemble/diag3_limit.cpp
   tmop/assemble/diag3.cpp
   tmop/assemble/grad3_limit.cpp
   tmop/assemble/grad3.cpp
@@ -204,7 +206,11 @@ set(HDRS
   integ/bilininteg_mass_kernels.hpp
   integ/bilininteg_mass_pa_simplices.hpp
   integ/bilininteg_vecdiffusion_pa.hpp
+  integ/bilininteg_vecdiv_pa.hpp
   integ/bilininteg_vecmass_pa.hpp
+  integ/nonlininteg_vecconvection_pa.hpp
+  integ/nonlininteg_vecconvection_pa_diag.hpp
+  integ/nonlininteg_vecconvection_pa_grad.hpp
   coefficient.hpp
   complex_fem.hpp
   convergence.hpp
@@ -311,6 +317,7 @@ set(HDRS
   tmop_tools.hpp
   tmop_amr.hpp
   gslib.hpp
+  gslib/gslib_kernel_helpers.hpp
   transfer.hpp
   hyperbolic.hpp
   integrator.hpp

fem/bilininteg.hpp

@@ -2689,14 +2689,22 @@ public:
    void AddMultMF(const Vector &x, Vector &y) const override;
    bool SupportsCeed() const override { return DeviceCanUseCeed(); }
 
+   // PA AddMultPA kernels
    using VectorMassAddMultPAType =
       void(*)(const int,  const int,
               const Array<real_t>&, const Vector&,
               const Vector&, Vector&, const int, const int);
-
    MFEM_REGISTER_KERNELS(VectorMassAddMultPA,
                          VectorMassAddMultPAType,
                          (int, int, int));
+
+   // PA DiagonalPA kernels
+   using VectorMassAssembleDiagonalPAType =
+      void(*)(const int, const int, const int,
+              const real_t*, const real_t*, real_t*);
+   MFEM_REGISTER_KERNELS(VectorMassAssembleDiagonalPA,
+                         VectorMassAssembleDiagonalPAType,
+                         (int /*dim*/, int /*q1d*/));
 };
 
 
@@ -3098,6 +3106,24 @@ public:
    void AddMultPA(const Vector &x, Vector &y) const override;
    void AddMultTransposePA(const Vector &x, Vector &y) const override;
 
+   using VectorDivergenceAddMultPAType =
+      void (*)(const int ne,
+               const Array<real_t> &b, const Array<real_t> &g, const Array<real_t> &bt,
+               const Vector &op, const Vector &x, Vector &y,
+               const int tr_d1d, const int te_d1d, const int q1d);
+   MFEM_REGISTER_KERNELS(VectorDivergenceAddMultPA,
+                         VectorDivergenceAddMultPAType,
+                         (int, int, int, int));
+
+   using VectorDivergenceAddMultTransposePAType =
+      void (*)(const int ne,
+               const Array<real_t> &bt, const Array<real_t> &gt, const Array<real_t> &b,
+               const Vector &q, const Vector &x, Vector &y,
+               const int tr_d1d, const int te_d1d, const int q1d);
+   MFEM_REGISTER_KERNELS(VectorDivergenceAddMultTransposePA,
+                         VectorDivergenceAddMultTransposePAType,
+                         (int, int, int, int));
+
    static const IntegrationRule &GetRule(const FiniteElement &trial_fe,
                                          const FiniteElement &test_fe,
                                          const ElementTransformation &Trans);

fem/coefficient.cpp

@@ -54,6 +54,8 @@ void Coefficient::Project(QuadratureFunction &qf)
    QuadratureSpaceBase &qspace = *qf.GetSpace();
    const int ne = qspace.GetNE();
    Vector values;
+   // GetValues makes a reference, but we need it to be valid on Host
+   qf.HostWrite();
    for (int iel = 0; iel < ne; ++iel)
    {
       qf.GetValues(iel, values);
@@ -327,6 +329,8 @@ void VectorCoefficient::Project(QuadratureFunction &qf)
    const int ne = qspace.GetNE();
    DenseMatrix values;
    Vector col;
+   // GetValues makes a reference, but we need it to be valid on Host
+   qf.HostWrite();
    for (int iel = 0; iel < ne; ++iel)
    {
       qf.GetValues(iel, values);
@@ -695,6 +699,8 @@ void MatrixCoefficient::Project(QuadratureFunction &qf, bool transpose)
    QuadratureSpaceBase &qspace = *qf.GetSpace();
    const int ne = qspace.GetNE();
    DenseMatrix values, matrix;
+   // GetValues makes a reference, but we need it to be valid on Host
+   qf.HostWrite();
    for (int iel = 0; iel < ne; ++iel)
    {
       qf.GetValues(iel, values);

fem/gslib.hpp

@@ -12,6 +12,9 @@
 #ifndef MFEM_GSLIB
 #define MFEM_GSLIB
 
+#include <map>
+#include <vector>
+
 #include "../config/config.hpp"
 #ifdef MFEM_USE_MPI
 #include "pgridfunc.hpp"
@@ -119,6 +122,11 @@ protected:
    // IntegrationRules for simplex->Quad/Hex and to project to p_max in-case of
    // p-refinement.
    Array<IntegrationRule *> ir_split;
+   /// Integration rules built at the field polynomial order (only for surface
+   /// meshes when mesh order is not the same as gridfunction order).
+   Array<IntegrationRule *> ir_split_sol;
+   /// Order at which #ir_split_sol was built; -1 means not built.
+   int ir_split_sol_order = -1;
    Array<FiniteElementSpace *> fes_rst_map; //FESpaces to map Quad/Hex->Simplex
    Array<GridFunction *> gf_rst_map; // GridFunctions to map Quad/Hex->Simplex
    FiniteElementCollection *fec_map_lin;
@@ -134,6 +142,8 @@ protected:
    AvgType avgtype;             // average type used for L2 functions
    Array<int> split_element_map;
    Array<int> split_element_index;
+   // Geometry::Type (as int) of the original element for each split quad.
+   Array<int> split_element_geom;
    int        NE_split_total;   // total number of elements after mesh splitting
    int        mesh_points_cnt;  // number of mesh nodes
    // Tolerance to ignore points found beyond the mesh boundary.
@@ -141,6 +151,12 @@ protected:
    double     bdr_tol;
    // Use CPU functions for Mesh/GridFunction on device for gslib1.0.7
    bool       gpu_to_cpu_fallback = false;
+   // Check if a point is inside the oriented bounding box of an
+   // element before the Newton iteration.
+   // Note: only used in MFEM implementation (not in gslib) which currently
+   // supports GPU kernels for area meshes in 2D, volume meshes in 3D,
+   // and surface meshes in 1D/2D/3D.
+   bool       obb_check = true;
 
    // Device specific data used for FindPoints
    struct DEV_STRUCT
@@ -162,11 +178,16 @@ protected:
       mutable double surf_dist_tol;
    } DEV;
 
-   /// Use GSLIB for communication and interpolation
+   // Helper function to setup and free gslib's crystal router.
+   void SetupCrystal(); // Called inside Setup and SetupSurf_base
+   void FreeCrystal();  // Called inside FreeData
+
+   /// Use GSLIB for communication and interpolation. Updates field_out on
+   /// host.
    virtual void InterpolateH1(const GridFunction &field_in, Vector &field_out,
                               const int field_out_ordering);
    /// Uses GSLIB Crystal Router for communication followed by MFEM's
-   /// interpolation functions
+   /// interpolation functions. Updates field_out on host.
    virtual void InterpolateGeneral(const GridFunction &field_in,
                                    Vector &field_out,
                                    const int field_out_ordering);
@@ -181,12 +202,26 @@ protected:
                                                  IntegrationRule *irule,
                                                  int order);
 
+   /** @brief Build integration rules at the given @a order for each split mesh
+    *  and store them in @a ir_out. Requires that \ref SetupSplitMeshes has
+    *  already been called. */
+   virtual void SetupIntegrationRules(const int order,
+                                      Array<IntegrationRule *> &ir_out);
+
    /** @brief Helper function that calls \ref SetupSplitMeshes and
-    * \ref SetupIntegrationRuleForSplitMesh. */
+    * \ref SetupIntegrationRules. */
    virtual void SetupSplitMeshesAndIntegrationRules(const int order);
 
-   /// Get GridFunction value at the points expected by GSLIB.
-   virtual void GetNodalValues(const GridFunction *gf_in, Vector &node_vals) const;
+   /** @brief Get GridFunction value at the points expected by GSLIB.
+    *  @param[in]  gf_in       Grid function to evaluate.
+    *  @param[out] node_vals   Output values.
+    *  @param[in]  ir_in       If non-null, use these rules instead of #ir_split.
+    *  @param[in]  by_element  If true, output has element-major layout
+    *                          [nel][vdim][ndofs]; otherwise component-major
+    *                          layout [vdim][total_pts]. */
+   virtual void GetNodalValues(const GridFunction *gf_in, Vector &node_vals,
+                               const Array<IntegrationRule *> *ir_in = nullptr,
+                               bool by_element = false) const;
 
    /** @brief Map {r,s,t} coordinates from [-1,1] to [0,1] for MFEM. For
     *  simplices, find the original element number (that was split into
@@ -293,9 +328,10 @@ protected:
                             const unsigned n,
                             const uint nel,
                             const unsigned m,
-                            const double bbox_tol,
+                            const double bbox_rel_size_inc,
                             const uint local_hash_size,
-                            const uint global_hash_size);
+                            const uint global_hash_size,
+                            const Vector *aabb_sz_inc);
 
    /// Preprocess 3D surface mesh needed for FindPoints.
    void findptssurf_setup_3(DEV_STRUCT &devs,
@@ -303,17 +339,47 @@ protected:
                             const unsigned n,
                             const uint nel,
                             const unsigned m,
-                            const double bbox_tol,
+                            const double bbox_rel_size_inc,
                             const uint local_hash_size,
                             const uint global_hash_size,
-                            const int rD);
+                            const int rD,
+                            const Vector *aabb_sz_inc);
 
+   /** @brief Shared implementation for the public surface-setup methods.
+    *
+    *  @details Initializes the surface-search data structures, builds the
+    *  split-element representation expected by gslib, and constructs the
+    *  element bounding boxes used by the MFEM surface kernels.
+    *
+    *  If @a aabb_sz_inc is null, the setup stores the default oriented
+    *  bounding boxes and uses @a bbox_rel_size_inc as their relative size
+    *  increase factor.
+    *
+    *  If @a aabb_sz_inc is non-null, the setup stores axis-aligned bounding
+    *  boxes only, applies the requested absolute AABB expansion in each
+    *  physical direction, and adjusts the tolerance @a bdr_tol so points
+    *  found in the expanded region are classified as border points.
+    *
+    *  @param[in] m                  Input surface mesh.
+    *  @param[in] bbox_rel_size_inc  Relative size increase applied when
+    *                                expanding each element bounding box during
+    *                                setup.
+    *  @param[in] aabb_sz_inc        Optional total absolute AABB expansion
+    *                                applied to the stored axis-aligned
+    *                                bounding boxes after construction.
+    *  @param[in] newt_tol           Newton tolerance for the point-search
+    *                                kernels.
+    */
+   void SetupSurf_Base(Mesh &m,
+                       const double bbox_rel_size_inc,
+                       const Vector *aabb_sz_inc,
+                       const double newt_tol);
 public:
    /// Serial constructor
    FindPointsGSLIB();
 
    /// Serial constructor + setup with given Mesh (see \ref Setup)
-   FindPointsGSLIB(Mesh &mesh_in, const double bb_t = 0.1,
+   FindPointsGSLIB(Mesh &mesh_in, const double bbox_rel_size_inc = 0.1,
                    const double newt_tol = 1.0e-12,
                    const int npt_max = 256);
 
@@ -322,7 +388,7 @@ public:
    FindPointsGSLIB(MPI_Comm comm_);
 
    /// Constructor + setup with given ParMesh (see \ref Setup)
-   FindPointsGSLIB(ParMesh &mesh_in, const double bb_t = 0.1,
+   FindPointsGSLIB(ParMesh &mesh_in, const double bbox_rel_size_inc = 0.1,
                    const double newt_tol = 1.0e-12,
                    const int npt_max = 256);
 #endif
@@ -338,23 +404,59 @@ public:
        Note: not tested with periodic (L2).
        Note: the input mesh \p m must have Nodes set.
 
-       @param[in] m         Input mesh.
-       @param[in] bb_t      (Optional) Relative size of bounding box around
-                            each element.
-       @param[in] newt_tol  (Optional) Newton tolerance for the gslib
-                            search methods.
-       @param[in] npt_max   (Optional) Number of points for simultaneous
-                            iteration. This alters performance and
-                            memory footprint.
+       @param[in] m                  Input mesh.
+       @param[in] bbox_rel_size_inc  (Optional) Relative size increase applied
+                                     when expanding each element bounding box.
+       @param[in] newt_tol           (Optional) Newton tolerance for the gslib
+                                     search methods.
+       @param[in] npt_max            (Optional) Number of points for
+                                     simultaneous iteration. This alters
+                                     performance and memory footprint.
    */
-   void Setup(Mesh &m, const double bb_t = 0.1, const double newt_tol = 1.0e-12,
+   void Setup(Mesh &m, const double bbox_rel_size_inc = 0.1,
+              const double newt_tol = 1.0e-12,
               const int npt_max = 256);
 
    /// Preprocess the surface mesh to compute data for FindPoints.
    void SetupSurf(Mesh &m,
-                  const double bb_t = 0.1,
-                  const double newt_tol = 1.0e-12,
-                  const int npt_max = 256);
+                  const double bbox_rel_size_inc = 0.1,
+                  const double newt_tol = 1.0e-12);
+
+   /** @brief Preprocess the surface mesh to compute data for FindPoints using
+    *  absolute AABB expansion.
+    *
+    *  @details This method computes only axis-aligned bounding boxes and
+    *  increases their total length by a user-specified amount in each
+    *  physical direction. The absolute AABB expansion is applied
+    *  symmetrically to the lower and upper bounds.
+    *
+    *  The size of @a aabb_sz_inc determines how the expansion values are
+    *  interpreted:
+    *  - `1`: one expansion value used in every direction for every element
+    *  - `NElements`: one expansion value per element, reused in x/y/z
+    *    directions
+    *  - `SpaceDim`: one expansion value per physical direction, reused for
+    *    every element
+    *  - `NElements*SpaceDim`: one expansion value per element and direction,
+    *    ordered as `(dx1,dy1,dz1, ... dxN,dyN,dzN)`
+    *
+    *  This method disables the oriented bounding-box precheck because the
+    *  stored boxes are modified only in their axis-aligned representation.
+    *
+    *  @param[in] m                  Input surface mesh.
+    *  @param[in] aabb_sz_inc        Total absolute AABB expansion applied in
+    *                                each physical direction to the stored
+    *                                axis-aligned bounding boxes.
+    *  @param[in] newt_tol           Newton tolerance for the point-search
+    *                                kernels.
+    *
+    *  @note We disable the oriented bounding box check with this setup.
+    *        @a bdr_tol is also adjusted so that all points in the AABBs can
+    *        be found.
+    */
+   void SetupSurfWithAABBExpansion(Mesh &m, const Vector &aabb_sz_inc,
+                                   const double newt_tol = 1.0e-12);
+
 
    /** @brief Searches positions given in physical space by \p point_pos.
 
@@ -401,7 +503,8 @@ public:
    /// Setup FindPoints and search positions
    void FindPoints(Mesh &m, const Vector &point_pos,
                    const int point_pos_ordering = Ordering::byNODES,
-                   const double bb_t = 0.1, const double newt_tol = 1.0e-12,
+                   const double bbox_rel_size_inc = 0.1,
+                   const double newt_tol = 1.0e-12,
                    const int npt_max = 256);
 
    /** @brief Interpolation of field values at prescribed reference space
@@ -413,7 +516,11 @@ public:
                               mesh that was given to Setup().
        @param[out] field_out  Interpolated values. For points that are not found
                               the value is set to #default_interp_value.
-                              The output ordering is determined from field_in.*/
+                              The output ordering is determined from field_in.
+
+       @note: field_out is moved to device if field_in is on device. Otherwise,
+              field_out memory allocation is not changed.
+   */
    virtual void Interpolate(const GridFunction &field_in, Vector &field_out);
 
    /// Interpolation of field values, with output ordering specification.
@@ -468,7 +575,12 @@ public:
     *  @details When using FindPoints, gslib may return points as found on the
     *  boundary even when they are slightly outside the domain. This tolerance
     *  is used to filter such points based on the distance^2 value and mark them
-    *  as not found.*/
+    *  as not found.
+    *
+    *  @note When the SetupSurfWithAABBExpansion method is used for surface
+    *  meshes, this tolerance is automatically computed based on the size of
+    *  expanded AABBs. Using this method will override that computed tolerance.
+    *  */
    virtual void SetDistanceToleranceForPointsFoundOnBoundary(double bdr_tol_)
    {
       bdr_tol = bdr_tol_;
@@ -603,25 +715,28 @@ public:
        Note: not tested with periodic meshes (L2).
        Note: the input mesh \p m must have Nodes set.
 
-       @param[in] m         Input mesh.
-       @param[in] meshid    A unique # for each overlapping mesh. This id is
-                            used to make sure that points being searched are not
-                            looked for in the mesh that they belong to.
-       @param[in] gfmax     (Optional) GridFunction in H1 that is used as a
-                            discriminator when one point is located in multiple
-                            meshes. The mesh that maximizes gfmax is chosen.
-                            For example, using the distance field based on the
-                            overlapping boundaries is helpful for convergence
-                            during Schwarz iterations.
-       @param[in] bb_t      (Optional) Relative size of bounding box around
-                            each element.
-       @param[in] newt_tol  (Optional) Newton tolerance for the gslib
-                            search methods.
-       @param[in] npt_max   (Optional) Number of points for simultaneous
-                            iteration. This alters performance and
-                            memory footprint.*/
-   void Setup(Mesh &m, const int meshid, GridFunction *gfmax = NULL,
-              const double bb_t = 0.1, const double newt_tol = 1.0e-12,
+       @param[in] m                  Input mesh.
+       @param[in] meshid             A unique # for each overlapping mesh.
+                                     This id is used to make sure that points
+                                     being searched are not looked for in the
+                                     mesh that they belong to.
+       @param[in] gfmax              (Optional) GridFunction in H1 that is used
+                                     as a discriminator when one point is
+                                     located in multiple meshes. The mesh that
+                                     maximizes gfmax is chosen. For example,
+                                     using the distance field based on the
+                                     overlapping boundaries is helpful for
+                                     convergence during Schwarz iterations.
+       @param[in] bbox_rel_size_inc  (Optional) Relative size increase applied
+                                     when expanding each element bounding box.
+       @param[in] newt_tol           (Optional) Newton tolerance for the gslib
+                                     search methods.
+       @param[in] npt_max            (Optional) Number of points for
+                                     simultaneous iteration. This alters
+                                     performance and memory footprint.*/
+   void Setup(Mesh &m, const int meshid, GridFunction *gfmax = nullptr,
+              const double bbox_rel_size_inc = 0.1,
+              const double newt_tol = 1.0e-12,
               const int npt_max = 256);
 
    /** Searches positions given in physical space by \p point_pos. All output
@@ -677,7 +792,7 @@ class GSOPGSLIB
 protected:
    struct gslib::crystal *cr;               // gslib's internal data
    struct gslib::comm *gsl_comm;            // gslib's internal data
-   struct gslib::gs_data *gsl_data = NULL;
+   struct gslib::gs_data *gsl_data = nullptr;
    int num_ids;
 
 public:

fem/gslib/findpts_local_2.cpp

@@ -11,7 +11,7 @@
 
 #include "../gslib.hpp"
 #include "../../general/forall.hpp"
-#include "../../linalg/kernels.hpp"
+#include "gslib_kernel_helpers.hpp"
 
 #ifdef MFEM_USE_GSLIB
 
@@ -27,8 +27,6 @@
 #pragma GCC diagnostic pop
 #endif
 
-#include <climits>
-
 namespace mfem
 {
 #if GSLIB_RELEASE_VERSION >= 10009
@@ -54,127 +52,14 @@ struct findptsElementGPT_t
    double x[DIM], jac[DIM * DIM], hes[4];
 };
 
-struct dbl_range_t
-{
-   double min, max;
-};
-struct obbox_t
-{
-   double c0[DIM], A[DIM * DIM];
-   dbl_range_t x[DIM];
-};
-
-struct findptsLocalHashData_t
-{
-   int hash_n;
-   dbl_range_t bnd[DIM];
-   double fac[DIM];
-   unsigned int *offset;
-   int max;
-};
-
-// Eval the ith Lagrange interpolant and its first derivative at x.
-// Note: lCoeff stores pre-computed coefficients for fast evaluation.
-static MFEM_HOST_DEVICE inline void lag_eval_first_der(double *p0, double x,
-                                                       int i, const double *z,
-                                                       const double *lCoeff,
-                                                       int pN)
-{
-   double u0 = 1, u1 = 0;
-   for (int j = 0; j < pN; ++j)
-   {
-      if (i != j)
-      {
-         double d_j = 2 * (x - z[j]);
-         u1 = d_j * u1 + u0;
-         u0 = d_j * u0;
-      }
-   }
-   p0[i] = lCoeff[i] * u0;
-   p0[pN+i] = 2.0 * lCoeff[i] * u1;
-}
-
-// Eval the ith Lagrange interpolant and its first and second derivative at x.
-// Note: lCoeff stores pre-computed coefficients for fast evaluation.
-static MFEM_HOST_DEVICE inline void lag_eval_second_der(double *p0, double x,
-                                                        int i, const double *z,
-                                                        const double *lCoeff,
-                                                        int pN)
-{
-   double u0 = 1, u1 = 0, u2 = 0;
-   for (int j = 0; j < pN; ++j)
-   {
-      if (i != j)
-      {
-         double d_j = 2 * (x - z[j]);
-         u2 = d_j * u2 + u1;
-         u1 = d_j * u1 + u0;
-         u0 = d_j * u0;
-      }
-   }
-   p0[i] = lCoeff[i] * u0;
-   p0[pN+i] = 2.0 * lCoeff[i] * u1;
-   p0[2*pN+i] = 8.0 * lCoeff[i] * u2;
-}
-
-// Axis-aligned bounding box test.
-static MFEM_HOST_DEVICE inline double AABB_test(const obbox_t *const b,
-                                                const double x[2])
-{
-   double test = 1;
-   for (int d = 0; d < 2; ++d)
-   {
-      double b_d = (x[d] - b->x[d].min) * (b->x[d].max - x[d]);
-      test = test < 0 ? test : b_d;
-   }
-   return test;
-}
-
-// Axis-aligned bounding box test followed by oriented bounding-box test.
-static MFEM_HOST_DEVICE inline double bbox_test(const obbox_t *const b,
-                                                const double x[2])
-{
-   const double bxyz = AABB_test(b, x);
-   if (bxyz < 0)
-   {
-      return bxyz;
-   }
-   else
-   {
-      double dxyz[2];
-      for (int d = 0; d < 2; ++d)
-      {
-         dxyz[d] = x[d] - b->c0[d];
-      }
-      double test = 1;
-      for (int d = 0; d < 2; ++d)
-      {
-         double rst = 0;
-         for (int e = 0; e < 2; ++e)
-         {
-            rst += b->A[d * 2 + e] * dxyz[e];
-         }
-         double brst = (rst + 1) * (1 - rst);
-         test = test < 0 ? test : brst;
-      }
-      return test;
-   }
-}
-
-// Element index corresponding to hash mesh that the point is located in.
-static MFEM_HOST_DEVICE inline int hash_index(const findptsLocalHashData_t *p,
-                                              const double x[2])
-{
-   const int n = p->hash_n;
-   int sum = 0;
-   for (int d = 2 - 1; d >= 0; --d)
-   {
-      sum *= n;
-      int i = (int)floor((x[d] - p->bnd[d].min) * p->fac[d]);
-      sum += i < 0 ? 0 : (n - 1 < i ? n - 1 : i);
-   }
-   return sum;
-}
+using dbl_range_t = gslib::dbl_range_t;
+using obbox_t = gslib::obbox_t<DIM>;
+using findptsLocalHashData_t = gslib::findptsLocalHashData_t<DIM>;
+using gslib::bbox_test;
+using gslib::hash_index;
+using gslib::l2norm2;
+using gslib::lag_eval_first_der;
+using gslib::lag_eval_second_der;
 
 /*Solve Ax=y. A is row-major */
 static MFEM_HOST_DEVICE inline void lin_solve_2(double x[2], const double A[4],
@@ -185,12 +70,6 @@ static MFEM_HOST_DEVICE inline void lin_solve_2(double x[2], const double A[4],
    x[1] = idet*(A[0]*y[1] - A[2]*y[0]);
 }
 
-/* L2 norm squared. */
-static MFEM_HOST_DEVICE inline double l2norm2(const double x[2])
-{
-   return x[0] * x[0] + x[1] * x[1];
-}
-
 /* the bit structure of flags is CSSRR
    the C bit --- 1<<4 --- is set when the point is converged
    RR is 0 = 00b if r is unconstrained,
@@ -352,7 +231,7 @@ static MFEM_HOST_DEVICE bool reject_prior_step_q(findptsElementPoint_t *res,
                                                  const findptsElementPoint_t *p,
                                                  const double tol)
 {
-   const double dist2 = l2norm2(resid);
+   const double dist2 = l2norm2<2>(resid);
    const double decr = p->dist2 - dist2;
    const double pred = p->dist2p;
    for (int d = 0; d < 2; ++d)
@@ -695,25 +574,25 @@ static MFEM_HOST_DEVICE double tensor_ig2_j(double *g_partials,
 }
 
 template<int T_D1D = 0>
-static void FindPointsLocal2D_Kernel(const int npt,
-                                     const double tol,
-                                     const double *x,
-                                     const int point_pos_ordering,
-                                     const double *xElemCoord,
-                                     const int nel,
-                                     const double *wtend,
-                                     const double *boxinfo,
-                                     const int hash_n,
-                                     const double *hashMin,
-                                     const double *hashFac,
-                                     unsigned int *hashOffset,
-                                     unsigned int *const code_base,
-                                     unsigned int *const el_base,
-                                     double *const r_base,
-                                     double *const dist2_base,
-                                     const double *gll1D,
-                                     const double *lagcoeff,
-                                     const int pN = 0)
+static void FindPointsLocal2DKernel(const int npt,
+                                    const double tol,
+                                    const double *x,
+                                    const int point_pos_ordering,
+                                    const double *xElemCoord,
+                                    const int nel,
+                                    const double *wtend,
+                                    const double *boxinfo,
+                                    const int hash_n,
+                                    const double *hashMin,
+                                    const double *hashFac,
+                                    unsigned int *hashOffset,
+                                    unsigned int *const code_base,
+                                    unsigned int *const el_base,
+                                    double *const r_base,
+                                    double *const dist2_base,
+                                    const double *gll1D,
+                                    const double *lagcoeff,
+                                    const int pN = 0)
 {
    const int MD1 = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
    const int D1D = T_D1D ? T_D1D : pN;
@@ -1175,30 +1054,45 @@ void FindPointsGSLIB::FindPointsLocal2(const Vector &point_pos,
    switch (DEV.dof1d)
    {
       case 2:
-         return FindPointsLocal2D_Kernel<2>(
-                   npt, DEV.newt_tol, pp, point_pos_ordering, pgslm, NE_split_total, pwt,
-                   pbb, DEV.lh_nx, plhm, plhf, plho, pcode, pelem, pref, pdist,
-                   pgll1d, plc);
+         FindPointsLocal2DKernel<2>(npt, DEV.newt_tol, pp,
+                                    point_pos_ordering, pgslm,
+                                    NE_split_total, pwt, pbb,
+                                    DEV.lh_nx, plhm, plhf, plho,
+                                    pcode, pelem, pref, pdist,
+                                    pgll1d, plc);
+         break;
       case 3:
-         return FindPointsLocal2D_Kernel<3>(
-                   npt, DEV.newt_tol, pp, point_pos_ordering, pgslm, NE_split_total, pwt,
-                   pbb, DEV.lh_nx, plhm, plhf, plho, pcode, pelem, pref, pdist,
-                   pgll1d, plc);
+         FindPointsLocal2DKernel<3>(npt, DEV.newt_tol, pp,
+                                    point_pos_ordering, pgslm,
+                                    NE_split_total, pwt, pbb,
+                                    DEV.lh_nx, plhm, plhf, plho,
+                                    pcode, pelem, pref, pdist,
+                                    pgll1d, plc);
+         break;
       case 4:
-         return FindPointsLocal2D_Kernel<4>(
-                   npt, DEV.newt_tol, pp, point_pos_ordering, pgslm, NE_split_total, pwt,
-                   pbb, DEV.lh_nx, plhm, plhf, plho, pcode, pelem, pref, pdist,
-                   pgll1d, plc);
+         FindPointsLocal2DKernel<4>(npt, DEV.newt_tol, pp,
+                                    point_pos_ordering, pgslm,
+                                    NE_split_total, pwt, pbb,
+                                    DEV.lh_nx, plhm, plhf, plho,
+                                    pcode, pelem, pref, pdist,
+                                    pgll1d, plc);
+         break;
       case 5:
-         return FindPointsLocal2D_Kernel<5>(
-                   npt, DEV.newt_tol, pp, point_pos_ordering, pgslm, NE_split_total, pwt,
-                   pbb, DEV.lh_nx, plhm, plhf, plho, pcode, pelem, pref, pdist,
-                   pgll1d, plc);
+         FindPointsLocal2DKernel<5>(npt, DEV.newt_tol, pp,
+                                    point_pos_ordering, pgslm,
+                                    NE_split_total, pwt, pbb,
+                                    DEV.lh_nx, plhm, plhf, plho,
+                                    pcode, pelem, pref, pdist,
+                                    pgll1d, plc);
+         break;
       default:
-         return FindPointsLocal2D_Kernel(npt, DEV.newt_tol, pp, point_pos_ordering,
-                                         pgslm, NE_split_total, pwt, pbb, DEV.lh_nx,
-                                         plhm, plhf, plho, pcode, pelem,
-                                         pref, pdist, pgll1d, plc, DEV.dof1d);
+         FindPointsLocal2DKernel(npt, DEV.newt_tol, pp,
+                                 point_pos_ordering, pgslm,
+                                 NE_split_total, pwt, pbb,
+                                 DEV.lh_nx, plhm, plhf, plho,
+                                 pcode, pelem, pref, pdist,
+                                 pgll1d, plc, DEV.dof1d);
+         break;
    }
 }
 #undef DIM2

fem/gslib/findpts_local_3.cpp

@@ -11,9 +11,7 @@
 
 #include "../gslib.hpp"
 #include "../../general/forall.hpp"
-#include "../../linalg/kernels.hpp"
-
-#include <climits>
+#include "gslib_kernel_helpers.hpp"
 
 #ifdef MFEM_USE_GSLIB
 
@@ -59,128 +57,15 @@ struct findptsElemPt
    double x[DIM], jac[DIM * DIM], hes[18];
 };
 
-struct dbl_range_t
-{
-   double min, max;
-};
-
-struct obbox_t
-{
-   double c0[DIM], A[DIM * DIM];
-   dbl_range_t x[DIM];
-};
-
-struct findptsLocalHashData_t
-{
-   int hash_n;
-   dbl_range_t bnd[DIM];
-   double fac[DIM];
-   unsigned int *offset;
-   // int max;
-};
-
-// Eval the ith Lagrange interpolant and its first derivative at x.
-// Note: lCoeff stores pre-computed coefficients for fast evaluation.
-static MFEM_HOST_DEVICE inline void lag_eval_first_der(double *p0, double x,
-                                                       int i, const double *z,
-                                                       const double *lCoeff,
-                                                       int pN)
-{
-   double u0 = 1, u1 = 0;
-   for (int j = 0; j < pN; ++j)
-   {
-      if (i != j)
-      {
-         double d_j = 2*(x-z[j]);
-         u1 = d_j*u1+u0;
-         u0 = d_j*u0;
-      }
-   }
-   p0[i] = lCoeff[i]*u0;
-   p0[pN+i] = 2.0*lCoeff[i]*u1;
-}
-
-// Eval the ith Lagrange interpolant and its first and second derivative at x.
-// Note: lCoeff stores pre-computed coefficients for fast evaluation.
-static MFEM_HOST_DEVICE inline void lag_eval_second_der(double *p0, double x,
-                                                        int i, const double *z,
-                                                        const double *lCoeff,
-                                                        int pN)
-{
-   double u0 = 1, u1 = 0, u2 = 0;
-   for (int j = 0; j < pN; ++j)
-   {
-      if (i != j)
-      {
-         double d_j = 2*(x-z[j]);
-         u2 = d_j*u2+u1;
-         u1 = d_j*u1+u0;
-         u0 = d_j*u0;
-      }
-   }
-   p0[i] = lCoeff[i]*u0;
-   p0[pN+i] = 2.0*lCoeff[i]*u1;
-   p0[2*pN+i] = 8.0*lCoeff[i]*u2;
-}
-
-// Axis-aligned bounding box test.
-static MFEM_HOST_DEVICE inline double AABB_test(const obbox_t *const b,
-                                                const double x[3])
-{
-   double b_d;
-   for (int d = 0; d < 3; ++d)
-   {
-      b_d = (x[d]-b->x[d].min)*(b->x[d].max-x[d]);
-      if (b_d < 0) { return b_d; }
-   }
-   return b_d;
-}
-
-// Axis-aligned bounding box test followed by oriented bounding-box test.
-static MFEM_HOST_DEVICE inline double bbox_test(const obbox_t *const b,
-                                                const double x[3])
-{
-   const double bxyz = AABB_test(b, x);
-   if (bxyz < 0)
-   {
-      return bxyz;
-   }
-   else
-   {
-      double dxyz[3];
-      for (int d = 0; d < 3; ++d)
-      {
-         dxyz[d] = x[d]-b->c0[d];
-      }
-      double test = 1;
-      for (int d = 0; d < 3; ++d)
-      {
-         double rst = 0;
-         for (int e = 0; e < 3; ++e)
-         {
-            rst += b->A[d*3+e]*dxyz[e];
-         }
-         double brst = (rst+1)*(1-rst);
-         test = test < 0 ? test : brst;
-      }
-      return test;
-   }
-}
-
-// Element index corresponding to hash mesh that the point is located in.
-static MFEM_HOST_DEVICE inline int hash_index(const findptsLocalHashData_t *p,
-                                              const double x[3])
-{
-   const int n = p->hash_n;
-   int sum = 0;
-   for (int d = 3-1; d >= 0; --d)
-   {
-      sum *= n;
-      int i = (int)floor((x[d]-p->bnd[d].min)*p->fac[d]);
-      sum += i < 0 ? 0 : (n-1 < i ? n-1 : i);
-   }
-   return sum;
-}
+using dbl_range_t = gslib::dbl_range_t;
+using obbox_t = gslib::obbox_t<DIM>;
+using findptsLocalHashData_t = gslib::findptsLocalHashData_t<DIM>;
+using gslib::bbox_test;
+using gslib::hash_index;
+using gslib::l2norm2;
+using gslib::lag_eval_first_der;
+using gslib::lag_eval_second_der;
+using gslib::lin_solve_sym_2;
 
 // Solve Ax=y. A is row-major.
 static MFEM_HOST_DEVICE inline void lin_solve_3(double x[3], const double A[9],
@@ -199,22 +84,6 @@ static MFEM_HOST_DEVICE inline void lin_solve_3(double x[3], const double A[9],
    x[2] = idet*(inv6*y[0]+inv7*y[1]+inv8*y[2]);
 }
 
-// Solve Ax=y. A is a symmetric 2x2 matrix.
-static MFEM_HOST_DEVICE inline void lin_solve_sym_2(double x[2],
-                                                    const double A[3],
-                                                    const double y[2])
-{
-   const double idet = 1 / (A[0]*A[2]-A[1]*A[1]);
-   x[0] = idet*(A[2]*y[0]-A[1]*y[1]);
-   x[1] = idet*(A[0]*y[1]-A[1]*y[0]);
-}
-
-// L2 norm.
-static MFEM_HOST_DEVICE inline double l2norm2(const double x[3])
-{
-   return x[0]*x[0]+x[1]*x[1]+x[2]*x[2];
-}
-
 /* the bit structure of flags is CTTSSRR
    the C bit --- 1<<6 --- is set when the point is converged
    RR is 0 = 00b if r is unconstrained,
@@ -459,7 +328,7 @@ static MFEM_HOST_DEVICE bool reject_prior_step_q(findptsPt *res,
                                                  const findptsPt *p,
                                                  const double tol)
 {
-   const double dist2 = l2norm2(resid);
+   const double dist2 = l2norm2<3>(resid);
    const double decr = p->dist2-dist2;
    const double pred = p->dist2p;
    for (int d = 0; d < 3; ++d)
@@ -1809,33 +1678,36 @@ void FindPointsGSLIB::FindPointsLocal3(const Vector &point_pos,
    {
       case 2:
          FindPointsLocal3DKernel<2>(npt, DEV.newt_tol, pp, point_pos_ordering,
-                                    pgslm, NE_split_total, pwt, pbb, DEV.lh_nx, plhm,
-                                    plhf, plho, pcode, pelem, pref, pdist, pgll1d,
-                                    plc);
+                                    pgslm, NE_split_total, pwt, pbb,
+                                    DEV.lh_nx, plhm, plhf, plho,
+                                    pcode, pelem, pref, pdist, pgll1d, plc);
          break;
       case 3:
          FindPointsLocal3DKernel<3>(npt, DEV.newt_tol, pp, point_pos_ordering,
-                                    pgslm, NE_split_total, pwt, pbb, DEV.lh_nx, plhm,
-                                    plhf, plho, pcode, pelem, pref, pdist, pgll1d,
-                                    plc);
+                                    pgslm, NE_split_total, pwt, pbb,
+                                    DEV.lh_nx, plhm, plhf, plho,
+                                    pcode, pelem, pref, pdist, pgll1d, plc);
          break;
       case 4:
          FindPointsLocal3DKernel<4>(npt, DEV.newt_tol, pp, point_pos_ordering,
-                                    pgslm, NE_split_total, pwt, pbb, DEV.lh_nx, plhm,
-                                    plhf, plho, pcode, pelem, pref, pdist, pgll1d,
-                                    plc);
+                                    pgslm, NE_split_total, pwt, pbb,
+                                    DEV.lh_nx, plhm, plhf, plho,
+                                    pcode, pelem, pref, pdist, pgll1d, plc);
          break;
       case 5:
          FindPointsLocal3DKernel<5>(npt, DEV.newt_tol, pp, point_pos_ordering,
-                                    pgslm, NE_split_total, pwt, pbb, DEV.lh_nx, plhm,
-                                    plhf, plho, pcode, pelem, pref, pdist, pgll1d,
-                                    plc);
+                                    pgslm, NE_split_total, pwt, pbb,
+                                    DEV.lh_nx, plhm, plhf, plho,
+                                    pcode, pelem, pref, pdist, pgll1d, plc);
          break;
       default:
-         FindPointsLocal3DKernel(npt, DEV.newt_tol, pp, point_pos_ordering, pgslm,
-                                 NE_split_total, pwt, pbb, DEV.lh_nx, plhm, plhf,
-                                 plho, pcode, pelem, pref, pdist, pgll1d, plc,
+         FindPointsLocal3DKernel(npt, DEV.newt_tol, pp,
+                                 point_pos_ordering, pgslm,
+                                 NE_split_total, pwt, pbb,
+                                 DEV.lh_nx, plhm, plhf, plho,
+                                 pcode, pelem, pref, pdist, pgll1d, plc,
                                  DEV.dof1d);
+         break;
    }
 }
 #undef pMax

fem/gslib/findptsedge_local_2.cpp

@@ -11,6 +11,7 @@
 
 #include "../gslib.hpp"
 #include "../../general/forall.hpp"
+#include "gslib_kernel_helpers.hpp"
 
 #ifdef MFEM_USE_GSLIB
 
@@ -52,113 +53,14 @@ struct findptsElementGPT_t
    double x[sDIM], jac[sDIM*rDIM], hes[sDIM*rDIM];
 };
 
-struct dbl_range_t
-{
-   double min, max;
-};
-
-struct obbox_t
-{
-   double c0[sDIM], A[sDIM*sDIM];
-   dbl_range_t x[sDIM];
-};
-
-struct findptsLocalHashData_t
-{
-   int hash_n;
-   dbl_range_t bnd[sDIM];
-   double fac[sDIM];
-   unsigned int *offset;
-};
-
-static MFEM_HOST_DEVICE inline void lag_eval_second_der(double *p0, double x,
-                                                        int i, const double *z,
-                                                        const double *lCoeff,
-                                                        int pN)
-{
-   double u0 = 1, u1 = 0, u2 = 0;
-   for (int j = 0; j < pN; ++j)
-   {
-      if (i != j)
-      {
-         double d_j = 2 * (x-z[j]);
-         u2 = d_j * u2 + u1;
-         u1 = d_j * u1 + u0;
-         u0 = d_j * u0;
-      }
-   }
-   double *p1 = p0 + pN, *p2 = p0 + 2 * pN;
-   p0[i] = lCoeff[i] * u0;
-   p1[i] = 2.0 * lCoeff[i] * u1;
-   p2[i] = 8.0 * lCoeff[i] * u2;
-}
-
-/* positive when possibly inside */
-static MFEM_HOST_DEVICE inline double obbox_axis_test(const obbox_t *const b,
-                                                      const double x[sDIM])
-{
-   double b_d;
-   for (int d=0; d<sDIM; ++d)
-   {
-      b_d = (x[d] - b->x[d].min) * (b->x[d].max - x[d]);
-      if (b_d < 0)   // if outside in any dimension
-      {
-         return b_d;
-      }
-   }
-   return b_d;       // only positive if inside
-}
-
-/* positive when given point is possibly inside given obbox b */
-static MFEM_HOST_DEVICE inline double obbox_test(const obbox_t *const b,
-                                                 const double x[sDIM])
-{
-   const double bxyz = obbox_axis_test(b,x);
-   if (bxyz<0)   // test if point is in AABB
-   {
-      return bxyz;
-   }
-   else   // test OBB only if inside AABB
-   {
-      double dxyz[sDIM];
-      for (int d=0; d<sDIM; ++d)
-      {
-         dxyz[d] = x[d] - b->c0[d];
-      }
-      double test = 1;
-      for (int d=0; d<sDIM; ++d)
-      {
-         double rst = 0;
-         for (int e=0; e<sDIM; ++e)
-         {
-            rst += b->A[d*2 + e] * dxyz[e];
-         }
-         double brst = (rst+1)*(1-rst);
-         test = test<0 ? test : brst;
-      }
-      return test;
-   }
-}
-
-/* Hash index in the hash table to the elements that possibly contain the point x */
-static MFEM_HOST_DEVICE inline int hash_index(const findptsLocalHashData_t *p,
-                                              const double x[2])
-{
-   const int n = p->hash_n;
-   int sum = 0;
-   for (int d=sDIM-1; d>=0; --d)
-   {
-      sum *= n;
-      int i = (int)floor((x[d] - p->bnd[d].min) * p->fac[d]);
-      sum += i<0 ? 0 : (n-1 < i ? n-1 : i);
-   }
-   return sum;
-}
-
-static MFEM_HOST_DEVICE inline double l2norm2(const double x[2])
-{
-   return x[0] * x[0] + x[1] * x[1];
-}
+using dbl_range_t = gslib::dbl_range_t;
+using obbox_t = gslib::obbox_t<sDIM>;
+using findptsLocalHashData_t = gslib::findptsLocalHashData_t<sDIM>;
+using gslib::AABB_test;
+using gslib::bbox_test;
+using gslib::hash_index;
+using gslib::l2norm2;
+using gslib::lag_eval_second_der;
 
 /* the bit structure of flags is CRR
    the C bit --- 1<<2 --- is set when the point is converged
@@ -187,29 +89,29 @@ static MFEM_HOST_DEVICE inline int point_index(const int x)
 /* check reduction in objective against prediction, and adjust
    trust region radius (p->tr) accordingly;
    may reject the prior step, returning 1; otherwise returns 0
-   sets out->dist2, out->index, out->x, out->oldr in any event,
-   leaving out->r, out->dr, out->flags to be set when returning 0 */
-static MFEM_HOST_DEVICE bool reject_prior_step_q(findptsElementPoint_t *out,
+   sets out_pt->dist2, out_pt->index, out_pt->x, out_pt->oldr in any event,
+   leaving out_pt->r, out_pt->dr, out_pt->flags to be set when returning 0 */
+static MFEM_HOST_DEVICE bool reject_prior_step_q(findptsElementPoint_t *out_pt,
                                                  const double resid[2],
                                                  const findptsElementPoint_t *p,
                                                  const double tol)
 {
-   const double dist2 = l2norm2(resid);
+   const double dist2 = l2norm2<2>(resid);
    const double decr = p->dist2 - dist2;
    const double pred = p->dist2p;
-   out->x[0] = p->x[0];
-   out->x[1] = p->x[1];
-   out->oldr = p->r;
-   out->dist2 = dist2;
+   out_pt->x[0] = p->x[0];
+   out_pt->x[1] = p->x[1];
+   out_pt->oldr = p->r;
+   out_pt->dist2 = dist2;
    if (decr >= 0.01*pred)
    {
       if (decr >= 0.9*pred)   // very good iteration
       {
-         out->tr = p->tr*2;
+         out_pt->tr = p->tr*2;
       }
       else                    // somewhat good iteration
       {
-         out->tr = p->tr;
+         out_pt->tr = p->tr;
       }
       return false;
    }
@@ -220,21 +122,21 @@ static MFEM_HOST_DEVICE bool reject_prior_step_q(findptsElementPoint_t *out,
          "very good iteration" --- this doubles the trust radius,
          which is why we divide by 4 below */
       double v0 = fabs(p->r - p->oldr);
-      out->tr = v0/4.0;
-      out->dist2 = p->dist2;
-      out->r = p->oldr;
-      out->flags = p->flags>>3;
-      out->dist2p = -HUGE_VAL;
+      out_pt->tr = v0/4.0;
+      out_pt->dist2 = p->dist2;
+      out_pt->r = p->oldr;
+      out_pt->flags = p->flags>>3;
+      out_pt->dist2p = -HUGE_VAL;
       if (pred < dist2*tol)
       {
-         out->flags |= CONVERGED_FLAG;
+         out_pt->flags |= CONVERGED_FLAG;
       }
       return true;
    }
 }
 
 static MFEM_HOST_DEVICE inline void newton_edge( findptsElementPoint_t *const
-                                                 out,
+                                                 out_pt,
                                                  const double jac[2],
                                                  const double rhess,
                                                  const double resid[2],
@@ -304,9 +206,9 @@ newton_edge_fin:
    {
       new_flags |= CONVERGED_FLAG;
    }
-   out->r = newr;
-   out->dist2p = -v;
-   out->flags = flags | new_flags | ((p->flags & FLAG_MASK)<<3);
+   out_pt->r = newr;
+   out_pt->dist2p = -v;
+   out_pt->flags = flags | new_flags | ((p->flags & FLAG_MASK)<<3);
 }
 
 static MFEM_HOST_DEVICE void seed_j( const double *elx[sDIM],
@@ -332,26 +234,27 @@ static MFEM_HOST_DEVICE void seed_j( const double *elx[sDIM],
 }
 
 template<int T_D1D = 0>
-static void FindPointsEdgeLocal2D_Kernel( const int     npt,
-                                          const double  tol,
-                                          const double  dist2tol,
-                                          const double  *x,
-                                          const int     point_pos_ordering,
-                                          const double  *xElemCoord,
-                                          const int     nel,
-                                          const double  *wtend,
-                                          const double  *boxinfo,
-                                          const int     hash_n,
-                                          const double  *hashMin,
-                                          const double  *hashFac,
-                                          unsigned int  *hashOffset,
-                                          unsigned int *const code_base,
-                                          unsigned int *const el_base,
-                                          double *const r_base,
-                                          double *const dist2_base,
-                                          const double  *gll1D,
-                                          const double  *lagcoeff,
-                                          const int     pN = 0 )
+static void FindPointsEdgeLocal2DKernel( const int     npt,
+                                         const double  tol,
+                                         const double  dist2tol,
+                                         const double  *x,
+                                         const int     point_pos_ordering,
+                                         const double  *xElemCoord,
+                                         const int     nel,
+                                         const double  *wtend,
+                                         const double  *boxinfo,
+                                         const bool    obb_check,
+                                         const int     hash_n,
+                                         const double  *hashMin,
+                                         const double  *hashFac,
+                                         unsigned int  *hashOffset,
+                                         unsigned int *const code_base,
+                                         unsigned int *const el_base,
+                                         double *const r_base,
+                                         double *const dist2_base,
+                                         const double  *gll1D,
+                                         const double  *lagcoeff,
+                                         const int     pN = 0 )
 {
    const int MD1   = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
    const int D1D   = T_D1D ? T_D1D : pN;
@@ -412,22 +315,34 @@ static void FindPointsEdgeLocal2D_Kernel( const int     npt,
       {
          const unsigned int el = *elp;
 
+         const int n_box_ents = obb_check ? (3*sDIM + sDIM2) : (2*sDIM);
+         bool pass_bb = true;
          obbox_t box;
-         int n_box_ents = 3*sDIM + sDIM2;
-
-         for (int idx = 0; idx < sDIM; ++idx)
+         if (obb_check)
          {
-            box.c0[idx] = boxinfo[n_box_ents*el + idx];
-            box.x[idx].min = boxinfo[n_box_ents*el + sDIM + idx];
-            box.x[idx].max = boxinfo[n_box_ents*el + 2*sDIM + idx];
+            for (int idx = 0; idx < sDIM; ++idx)
+            {
+               box.c0[idx] = boxinfo[n_box_ents*el + idx];
+               box.x[idx].min = boxinfo[n_box_ents*el + sDIM + idx];
+               box.x[idx].max = boxinfo[n_box_ents*el + 2*sDIM + idx];
+            }
+            for (int idx = 0; idx < sDIM2; ++idx)
+            {
+               box.A[idx] = boxinfo[n_box_ents*el + 3*sDIM + idx];
+            }
+            pass_bb = (bbox_test(&box, x_i) >= 0);
+         }
+         else
+         {
+            for (int d = 0; d < sDIM; ++d)
+            {
+               box.x[d].min = boxinfo[n_box_ents*el + d];
+               box.x[d].max = boxinfo[n_box_ents*el + sDIM + d];
+            }
+            pass_bb = (AABB_test(&box, x_i) >= 0);
          }
 
-         for (int idx = 0; idx < sDIM2; ++idx)
-         {
-            box.A[idx] = boxinfo[n_box_ents*el + 3*sDIM + idx];
-         }
-
-         if (obbox_test(&box,x_i)>=0)
+         if (pass_bb)
          {
             //------------ findpts_local ------------------
             {
@@ -516,11 +431,14 @@ static void FindPointsEdgeLocal2D_Kernel( const int     npt,
                            double *hess = jac + sDIM*rDIM;
 
                            findptsElementGEdge_t edge;
+                           for (int d=0; d<sDIM; ++d)
+                           {
+                              edge.x[d] = constraint_workspace + d*D1D;
+                           }
                            MFEM_FOREACH_THREAD(j,x,D1D)
                            {
                               for (int d=0; d<sDIM; ++d)
                               {
-                                 edge.x[d] = constraint_workspace + d*D1D;
                                  edge.x[d][j] = elx[d][j];
                               }
                            }
@@ -681,28 +599,41 @@ void FindPointsGSLIB::FindPointsEdgeLocal2( const Vector &point_pos,
    auto pgll1d = DEV.gll1d.ReadWrite(use_dev);
    auto plc = DEV.lagcoeff.Read(use_dev);
    double dist2tol = DEV.surf_dist_tol;
+   const bool obb_chk = obb_check;
    switch (DEV.dof1d)
    {
       case 2:
-         return FindPointsEdgeLocal2D_Kernel<2>(
-                   npt, DEV.newt_tol, dist2tol, pp, point_pos_ordering, pgslm,
-                   NE_split_total, pwt, pbb, DEV.lh_nx, plhm, plhf,
-                   plho, pcode, pelem, pref, pdist, pgll1d, plc);
+         FindPointsEdgeLocal2DKernel<2>(npt, DEV.newt_tol, dist2tol,
+                                        pp, point_pos_ordering, pgslm,
+                                        NE_split_total, pwt, pbb, obb_chk,
+                                        DEV.lh_nx, plhm, plhf, plho,
+                                        pcode, pelem, pref, pdist,
+                                        pgll1d, plc);
+         break;
       case 3:
-         return FindPointsEdgeLocal2D_Kernel<3>(
-                   npt, DEV.newt_tol, dist2tol, pp, point_pos_ordering, pgslm,
-                   NE_split_total, pwt, pbb, DEV.lh_nx, plhm, plhf,
-                   plho, pcode, pelem, pref, pdist, pgll1d, plc);
+         FindPointsEdgeLocal2DKernel<3>(npt, DEV.newt_tol, dist2tol,
+                                        pp, point_pos_ordering, pgslm,
+                                        NE_split_total, pwt, pbb, obb_chk,
+                                        DEV.lh_nx, plhm, plhf, plho,
+                                        pcode, pelem, pref, pdist,
+                                        pgll1d, plc);
+         break;
       case 4:
-         return FindPointsEdgeLocal2D_Kernel<4>(
-                   npt, DEV.newt_tol, dist2tol, pp, point_pos_ordering, pgslm,
-                   NE_split_total, pwt, pbb, DEV.lh_nx, plhm, plhf,
-                   plho, pcode, pelem, pref, pdist, pgll1d, plc);
+         FindPointsEdgeLocal2DKernel<4>(npt, DEV.newt_tol, dist2tol,
+                                        pp, point_pos_ordering, pgslm,
+                                        NE_split_total, pwt, pbb, obb_chk,
+                                        DEV.lh_nx, plhm, plhf, plho,
+                                        pcode, pelem, pref, pdist,
+                                        pgll1d, plc);
+         break;
       default:
-         return FindPointsEdgeLocal2D_Kernel(
-                   npt, DEV.newt_tol, dist2tol, pp, point_pos_ordering, pgslm,
-                   NE_split_total, pwt, pbb, DEV.lh_nx, plhm, plhf,
-                   plho, pcode, pelem, pref, pdist, pgll1d, plc, DEV.dof1d);
+         FindPointsEdgeLocal2DKernel(npt, DEV.newt_tol, dist2tol, pp,
+                                     point_pos_ordering, pgslm,
+                                     NE_split_total, pwt, pbb, obb_chk,
+                                     DEV.lh_nx, plhm, plhf, plho,
+                                     pcode, pelem, pref, pdist,
+                                     pgll1d, plc, DEV.dof1d);
+         break;
    }
 }
 #undef sDIM

fem/gslib/findptsedge_local_3.cpp

@@ -11,6 +11,7 @@
 
 #include "../gslib.hpp"
 #include "../../general/forall.hpp"
+#include "gslib_kernel_helpers.hpp"
 
 #ifdef MFEM_USE_GSLIB
 
@@ -54,117 +55,14 @@ struct findptsElementGPT_t
    double x[sDIM], jac[sDIM], hes[sDIM*(1+1)];
 };
 
-struct dbl_range_t
-{
-   double min, max;
-};
-
-struct obbox_t
-{
-   double c0[sDIM], A[sDIM*sDIM];
-   dbl_range_t x[sDIM];
-};
-
-struct findptsLocalHashData_t
-{
-   int hash_n;
-   dbl_range_t bnd[sDIM];
-   double fac[sDIM];
-   unsigned int *offset;
-};
-
-static MFEM_HOST_DEVICE inline void lag_eval_second_der(double *p0, double x,
-                                                        int i, const double *z,
-                                                        const double *lCoeff,
-                                                        int pN)
-{
-   double u0 = 1, u1 = 0, u2 = 0;
-   for (int j=0; j<pN; ++j)
-   {
-      if (i!=j)
-      {
-         double d_j = 2 * (x-z[j]);
-         u2 = d_j * u2 + u1;
-         u1 = d_j * u1 + u0;
-         u0 = d_j * u0;
-      }
-   }
-   double *p1 = p0 + pN, *p2 = p0 + 2 * pN;
-   p0[i] = lCoeff[i] * u0;
-   p1[i] = 2.0 * lCoeff[i] * u1;
-   p2[i] = 8.0 * lCoeff[i] * u2;
-}
-
-/* positive when possibly inside */
-static MFEM_HOST_DEVICE inline double obbox_axis_test(const obbox_t *const b,
-                                                      const double x[sDIM])
-{
-   double b_d;
-   for (int d=0; d<sDIM; ++d)
-   {
-      b_d = (x[d] - b->x[d].min) * (b->x[d].max - x[d]);
-      if (b_d < 0)   // if outside in any dimension
-      {
-         return b_d;
-      }
-   }
-   return b_d;       // only positive if inside in all dimensions
-}
-
-/* positive when possibly inside */
-static MFEM_HOST_DEVICE inline double obbox_test(const obbox_t *const b,
-                                                 const double x[sDIM])
-{
-   const double bxyz = obbox_axis_test(b, x);
-   if (bxyz<0)
-   {
-      return bxyz;
-   }
-   else
-   {
-      double dxyz[3];
-      // dxyz: distance of the point from the center of the OBB
-      for (int d=0; d<sDIM; ++d)
-      {
-         dxyz[d] = x[d] - b->c0[d];
-      }
-      // transform dxyz to the local coordinate system of the OBB,
-      // and check if the point is inside the OBB [-1,1]^sDIM
-      double test = 1;
-      for (int d=0; d<sDIM; ++d)
-      {
-         double rst = 0;
-         for (int e=0; e<sDIM; ++e)
-         {
-            rst += b->A[d*sDIM + e] * dxyz[e];
-         }
-         double brst = (rst+1)*(1-rst);
-         test = test<0 ? test : brst;
-      }
-      return test;
-   }
-}
-
-/* Hash index in the hash table to the elements that possibly contain the point x */
-static MFEM_HOST_DEVICE inline int hash_index(const findptsLocalHashData_t *p,
-                                              const double x[sDIM])
-{
-   const int n = p->hash_n;
-   int sum = 0;
-   for (int d=sDIM-1; d>=0; --d)
-   {
-      sum *= n;
-      int i = (int)floor((x[d] - p->bnd[d].min) * p->fac[d]);
-      sum += i<0 ? 0 : (n-1 < i ? n-1 : i);
-   }
-   return sum;
-}
-
-
-static MFEM_HOST_DEVICE inline double norm2(const double x[sDIM])
-{
-   return ( x[0]*x[0] + x[1]*x[1] + x[2]*x[2] );
-}
+using dbl_range_t = gslib::dbl_range_t;
+using obbox_t = gslib::obbox_t<sDIM>;
+using findptsLocalHashData_t = gslib::findptsLocalHashData_t<sDIM>;
+using gslib::AABB_test;
+using gslib::bbox_test;
+using gslib::hash_index;
+using gslib::l2norm2;
+using gslib::lag_eval_second_der;
 
 /* the bit structure of flags is CRR
    the C bit --- 1<<2 --- is set when the point is converged
@@ -175,47 +73,46 @@ static MFEM_HOST_DEVICE inline double norm2(const double x[sDIM])
 #define CONVERGED_FLAG (1u<<2)
 #define FLAG_MASK 0x07u
 
-/* returns the number of constrained reference coordinates, max 2
+/* returns the number of constrained reference coordinates, max 1
 */
 static MFEM_HOST_DEVICE inline int num_constrained(const int flags)
 {
-   const int y = (flags | flags>>1);
-   return (y & 1u) + (y>>2 & 1u);
+   return ((flags | flags>>1) & 1u);
 }
 
 static MFEM_HOST_DEVICE inline int point_index(const int x)
 {
-   return ((x>>1)&1u) | ((x>>2)&2u);
+   return ((x>>1)&1u);
 }
 
 /* check reduction in objective against prediction, and adjust
    trust region radius (p->tr) accordingly;
    may reject the prior step, returning 1; otherwise returns 0
-   sets out->dist2, out->index, out->x, out->oldr in any event,
-   leaving out->r, out->dr, out->flags to be set when returning 0 */
-static MFEM_HOST_DEVICE bool reject_prior_step_q(findptsElementPoint_t *out,
+   sets out_pt->dist2, out_pt->index, out_pt->x, out_pt->oldr in any event,
+   leaving out_pt->r, out_pt->dr, out_pt->flags to be set when returning 0 */
+static MFEM_HOST_DEVICE bool reject_prior_step_q(findptsElementPoint_t *out_pt,
                                                  const double resid[3],
                                                  const findptsElementPoint_t *p,
                                                  const double tol)
 {
-   const double dist2 = norm2(resid);
+   const double dist2 = l2norm2<sDIM>(resid);
    const double decr  = p->dist2 - dist2;
    const double pred  = p->dist2p;
    for (int d=0; d<sDIM; ++d)
    {
-      out->x[d] = p->x[d];
+      out_pt->x[d] = p->x[d];
    }
-   out->oldr = p->r;
-   out->dist2 = dist2;
+   out_pt->oldr = p->r;
+   out_pt->dist2 = dist2;
    if (decr>=0.01*pred)
    {
       if (decr>=0.9*pred)   // very good iteration
       {
-         out->tr = 2*p->tr;
+         out_pt->tr = 2*p->tr;
       }
       else   // good iteration
       {
-         out->tr = p->tr;
+         out_pt->tr = p->tr;
       }
       return false;
    }
@@ -226,21 +123,21 @@ static MFEM_HOST_DEVICE bool reject_prior_step_q(findptsElementPoint_t *out,
          "very good iteration" --- this doubles the trust radius,
          which is why we divide by 4 below */
       double v0 = fabs(p->r - p->oldr);
-      out->tr = v0/4.0;
-      out->dist2 = p->dist2;
-      out->r = p->oldr;
-      out->flags = p->flags>>3;
-      out->dist2p = -HUGE_VAL;
+      out_pt->tr = v0/4.0;
+      out_pt->dist2 = p->dist2;
+      out_pt->r = p->oldr;
+      out_pt->flags = p->flags>>3;
+      out_pt->dist2p = -HUGE_VAL;
       if (pred<dist2*tol)
       {
-         out->flags |= CONVERGED_FLAG;
+         out_pt->flags |= CONVERGED_FLAG;
       }
       return true;
    }
 }
 
 static MFEM_HOST_DEVICE inline void newton_edge(findptsElementPoint_t *const
-                                                out,
+                                                out_pt,
                                                 const double jac[sDIM*rDIM],
                                                 const double rhes,
                                                 const double resid[sDIM],
@@ -314,9 +211,9 @@ newton_edge_fin:
    {
       new_flags |= CONVERGED_FLAG;
    }
-   out->r = nr;
-   out->dist2p = -v;
-   out->flags = flags | new_flags | ((p->flags & FLAG_MASK)<<3);
+   out_pt->r = nr;
+   out_pt->dist2p = -v;
+   out_pt->flags = flags | new_flags | ((p->flags & FLAG_MASK)<<3);
 #undef EVAL
 }
 
@@ -338,31 +235,32 @@ static MFEM_HOST_DEVICE void seed_j(const double *elx[sDIM],
    {
       dx[d] = x[d] - elx[d][ir];
    }
-   dist2[ir] = norm2(dx);;
+   dist2[ir] = l2norm2(dx);
    r[ir] = z[ir];
 }
 
 template<int T_D1D = 0>
-static void FindPointsEdgeLocal3D_Kernel(const int npt,
-                                         const double tol,
-                                         const double dist2tol,
-                                         const double *x,
-                                         const int point_pos_ordering,
-                                         const double *xElemCoord,
-                                         const int nel,
-                                         const double *wtend,
-                                         const double *boxinfo,
-                                         const int hash_n,
-                                         const double *hashMin,
-                                         const double *hashFac,
-                                         unsigned int *hashOffset,
-                                         unsigned int *const code_base,
-                                         unsigned int *const el_base,
-                                         double *const r_base,
-                                         double *const dist2_base,
-                                         const double *gll1D,
-                                         const double *lagcoeff,
-                                         const int pN = 0)
+static void FindPointsEdgeLocal3DKernel(const int npt,
+                                        const double tol,
+                                        const double dist2tol,
+                                        const double *x,
+                                        const int point_pos_ordering,
+                                        const double *xElemCoord,
+                                        const int nel,
+                                        const double *wtend,
+                                        const double *boxinfo,
+                                        const bool obb_check,
+                                        const int hash_n,
+                                        const double *hashMin,
+                                        const double *hashFac,
+                                        unsigned int *hashOffset,
+                                        unsigned int *const code_base,
+                                        unsigned int *const el_base,
+                                        double *const r_base,
+                                        double *const dist2_base,
+                                        const double *gll1D,
+                                        const double *lagcoeff,
+                                        const int pN = 0)
 {
    const int MD1   = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
    const int D1D   = T_D1D ? T_D1D : pN;
@@ -419,21 +317,35 @@ static void FindPointsEdgeLocal3D_Kernel(const int npt,
       for (; elp!=ele; ++elp)
       {
          const unsigned int el = *elp;
+
+         const int n_box_ents = obb_check ? (3*sDIM + sDIM2) : (2*sDIM);
+         bool pass_bb = true;
          obbox_t box;
-         int n_box_ents = 3*sDIM + sDIM2;
-
-         for (int idx = 0; idx < sDIM; ++idx)
+         if (obb_check)
          {
-            box.c0[idx] = boxinfo[n_box_ents*el + idx];
-            box.x[idx].min = boxinfo[n_box_ents*el + sDIM + idx];
-            box.x[idx].max = boxinfo[n_box_ents*el + 2*sDIM + idx];
+            for (int idx = 0; idx < sDIM; ++idx)
+            {
+               box.c0[idx] = boxinfo[n_box_ents*el + idx];
+               box.x[idx].min = boxinfo[n_box_ents*el + sDIM + idx];
+               box.x[idx].max = boxinfo[n_box_ents*el + 2*sDIM + idx];
+            }
+            for (int idx = 0; idx < sDIM2; ++idx)
+            {
+               box.A[idx] = boxinfo[n_box_ents*el + 3*sDIM + idx];
+            }
+            pass_bb = (bbox_test(&box, x_i) >= 0);
          }
-         for (int idx = 0; idx < sDIM2; ++idx)
+         else
          {
-            box.A[idx] = boxinfo[n_box_ents*el + 3*sDIM + idx];
+            for (int d = 0; d < sDIM; ++d)
+            {
+               box.x[d].min = boxinfo[n_box_ents*el + d];
+               box.x[d].max = boxinfo[n_box_ents*el + sDIM + d];
+            }
+            pass_bb = (AABB_test(&box, x_i) >= 0);
          }
 
-         if (obbox_test(&box, x_i)>=0)
+         if (pass_bb)
          {
             //// findpts_local ////
             {
@@ -521,11 +433,14 @@ static void FindPointsEdgeLocal3D_Kernel(const int npt,
                            double *hess = jac + sDIM*rDIM;
 
                            findptsElementGEdge_t edge;
+                           for (int d=0; d<sDIM; ++d)
+                           {
+                              edge.x[d] = constraint_workspace + d*D1D;
+                           }
                            MFEM_FOREACH_THREAD(j,x,D1D)
                            {
                               for (int d=0; d<sDIM; ++d)
                               {
-                                 edge.x[d] = constraint_workspace + d*D1D;
                                  edge.x[d][j] = elx[d][j];
                               }
                            }
@@ -688,28 +603,41 @@ void FindPointsGSLIB::FindPointsEdgeLocal3(const Vector &point_pos,
    auto pgll1d = DEV.gll1d.ReadWrite(use_dev);
    auto plc = DEV.lagcoeff.Read(use_dev);
    double dist2tol = DEV.surf_dist_tol;
+   const bool obb_chk = obb_check;
    switch (DEV.dof1d)
    {
       case 2:
-         return FindPointsEdgeLocal3D_Kernel<2>(
-                   npt, DEV.newt_tol, dist2tol, pp, point_pos_ordering, pgslm,
-                   NE_split_total, pwt, pbb, DEV.lh_nx, plhm, plhf,
-                   plho, pcode, pelem, pref, pdist, pgll1d, plc);
+         FindPointsEdgeLocal3DKernel<2>(npt, DEV.newt_tol, dist2tol,
+                                        pp, point_pos_ordering, pgslm,
+                                        NE_split_total, pwt, pbb, obb_chk,
+                                        DEV.lh_nx, plhm, plhf, plho,
+                                        pcode, pelem, pref, pdist,
+                                        pgll1d, plc);
+         break;
       case 3:
-         return FindPointsEdgeLocal3D_Kernel<3>(
-                   npt, DEV.newt_tol, dist2tol, pp, point_pos_ordering, pgslm,
-                   NE_split_total, pwt, pbb, DEV.lh_nx, plhm, plhf,
-                   plho, pcode, pelem, pref, pdist, pgll1d, plc);
+         FindPointsEdgeLocal3DKernel<3>(npt, DEV.newt_tol, dist2tol,
+                                        pp, point_pos_ordering, pgslm,
+                                        NE_split_total, pwt, pbb, obb_chk,
+                                        DEV.lh_nx, plhm, plhf, plho,
+                                        pcode, pelem, pref, pdist,
+                                        pgll1d, plc);
+         break;
       case 4:
-         return FindPointsEdgeLocal3D_Kernel<4>(
-                   npt, DEV.newt_tol, dist2tol, pp, point_pos_ordering, pgslm,
-                   NE_split_total, pwt, pbb, DEV.lh_nx, plhm, plhf,
-                   plho, pcode, pelem, pref, pdist, pgll1d, plc);
+         FindPointsEdgeLocal3DKernel<4>(npt, DEV.newt_tol, dist2tol,
+                                        pp, point_pos_ordering, pgslm,
+                                        NE_split_total, pwt, pbb, obb_chk,
+                                        DEV.lh_nx, plhm, plhf, plho,
+                                        pcode, pelem, pref, pdist,
+                                        pgll1d, plc);
+         break;
       default:
-         return FindPointsEdgeLocal3D_Kernel(
-                   npt, DEV.newt_tol, dist2tol, pp, point_pos_ordering, pgslm,
-                   NE_split_total, pwt, pbb, DEV.lh_nx, plhm, plhf,
-                   plho, pcode, pelem, pref, pdist, pgll1d, plc, DEV.dof1d);
+         FindPointsEdgeLocal3DKernel(npt, DEV.newt_tol, dist2tol, pp,
+                                     point_pos_ordering, pgslm,
+                                     NE_split_total, pwt, pbb, obb_chk,
+                                     DEV.lh_nx, plhm, plhf, plho,
+                                     pcode, pelem, pref, pdist,
+                                     pgll1d, plc, DEV.dof1d);
+         break;
    }
 }
 #undef rDIM2

fem/gslib/findptssurf_local_3.cpp

@@ -11,6 +11,7 @@
 
 #include "../gslib.hpp"
 #include "../../general/forall.hpp"
+#include "gslib_kernel_helpers.hpp"
 #ifdef MFEM_USE_GSLIB
 
 #ifdef MFEM_HAVE_GCC_PRAGMA_DIAGNOSTIC
@@ -51,124 +52,15 @@ struct findptsElementGPT_t
    double x[sDIM], jac[sDIM*rDIM], hes[sDIM*(rDIM+1)];
 };
 
-struct dbl_range_t
-{
-   double min, max;
-};
-
-struct obbox_t
-{
-   double c0[sDIM], A[sDIM*sDIM];
-   dbl_range_t x[sDIM];
-};
-
-struct findptsLocalHashData_t
-{
-   int hash_n;
-   dbl_range_t bnd[sDIM];
-   double fac[sDIM];
-   unsigned int *offset;
-};
-
-static MFEM_HOST_DEVICE inline void lag_eval_second_der(double *p0, double x,
-                                                        int i, const double *z,
-                                                        const double *lCoeff,
-                                                        int pN)
-{
-   double u0 = 1, u1 = 0, u2 = 0;
-   for (int j = 0; j < pN; ++j)
-   {
-      if (i != j)
-      {
-         double d_j = 2 * (x - z[j]);
-         u2 = d_j * u2 + u1;
-         u1 = d_j * u1 + u0;
-         u0 = d_j * u0;
-      }
-   }
-   p0[i] = lCoeff[i] * u0;
-   p0[pN+i] = 2.0 * lCoeff[i] * u1;
-   p0[2*pN+i] = 8.0 * lCoeff[i] * u2;
-}
-
-/* positive when possibly inside */
-static MFEM_HOST_DEVICE inline double AABB_test(const obbox_t *const b,
-                                                const double x[sDIM])
-{
-   double b_d;
-   for (int d=0; d<sDIM; ++d)
-   {
-      b_d = (x[d] - b->x[d].min) * (b->x[d].max - x[d]);
-      if (b_d < 0)   // if outside in any dimension
-      {
-         return b_d;
-      }
-   }
-   return b_d;       // only positive if inside in all dimensions
-}
-
-/* positive when possibly inside */
-static MFEM_HOST_DEVICE inline double bbox_test(const obbox_t *const b,
-                                                const double x[sDIM])
-{
-   const double bxyz = AABB_test(b, x);
-   if (bxyz<0)
-   {
-      return bxyz;
-   }
-   else
-   {
-      double dxyz[3];
-      // dxyz: distance of the point from the center of the OBB
-      for (int d=0; d<sDIM; ++d)
-      {
-         dxyz[d] = x[d] - b->c0[d];
-      }
-      // tranform dxyz to the local coordinate system of the OBB,
-      // and check if the point is inside the OBB [-1,1]^sDIM
-      double test = 1;
-      for (int d=0; d<sDIM; ++d)
-      {
-         double rst = 0;
-         for (int e=0; e<sDIM; ++e)
-         {
-            rst += b->A[d*sDIM + e] * dxyz[e];
-         }
-         double brst = (rst+1)*(1-rst);
-         test = test<0 ? test : brst;
-      }
-      return test;
-   }
-}
-
-/* Hash index in the hash table to the elements that possibly contain the point x */
-static MFEM_HOST_DEVICE inline int hash_index(const findptsLocalHashData_t *p,
-                                              const double x[sDIM])
-{
-   const int n = p->hash_n;
-   int sum = 0;
-   for (int d=sDIM-1; d>=0; --d)
-   {
-      sum *= n;
-      int i = (int)floor((x[d] - p->bnd[d].min) * p->fac[d]);
-      sum += i<0 ? 0 : (n-1 < i ? n-1 : i);
-   }
-   return sum;
-}
-
-static MFEM_HOST_DEVICE inline void lin_solve_sym_2(double x[2],
-                                                    const double A[3],
-                                                    const double y[2])
-{
-   const double idet = 1 / (A[0] * A[2] - A[1] * A[1]);
-   x[0] = idet * (A[2] * y[0] - A[1] * y[1]);
-   x[1] = idet * (A[0] * y[1] - A[1] * y[0]);
-}
-
-static MFEM_HOST_DEVICE inline double l2norm2(const double x[sDIM])
-{
-   return ( x[0]*x[0] + x[1]*x[1] + x[2]*x[2]);
-}
+using dbl_range_t = gslib::dbl_range_t;
+using obbox_t = gslib::obbox_t<sDIM>;
+using findptsLocalHashData_t = gslib::findptsLocalHashData_t<sDIM>;
+using gslib::AABB_test;
+using gslib::bbox_test;
+using gslib::hash_index;
+using gslib::l2norm2;
+using gslib::lag_eval_second_der;
+using gslib::lin_solve_sym_2;
 
 /* the bit structure of flags is CSSRR
    the C bit --- 1<<4 --- is set when the point is converged
@@ -219,18 +111,10 @@ static MFEM_HOST_DEVICE inline int point_index(const int x)
    return ((x>>1)&1u) | ((x>>2)&2u);
 }
 
-static MFEM_HOST_DEVICE inline findptsElementGEdge_t
+static MFEM_HOST_DEVICE inline void
 get_edge(const double *elx[3], const double *wtend, int ei,
-         double *workspace, int &side_init, int jidx, int pN)
+         int &side_init, int jidx, int pN, findptsElementGEdge_t &edge)
 {
-   findptsElementGEdge_t edge;
-   for (int d=0; d<sDIM; ++d)
-   {
-      edge.x[d]     = workspace             + d*pN;
-      edge.dxdn[d]  = workspace +   sDIM*pN + d*pN;
-      edge.d2xdn[d] = workspace + 2*sDIM*pN + d*pN;
-   }
-
    // given edge index, compute normal and tangential directions
    const int dn = ei>>1, //0 for rmin/rmax, 1 for smin/smax
              de = plus_1_mod_2(dn); // 1 for rmin/rmax, 0 for smin/smax
@@ -256,7 +140,6 @@ get_edge(const double *elx[3], const double *wtend, int ei,
       edge.d2xdn[dd][jj] = sums_k[1];
 #undef ELX
    }
-   return edge;
 }
 
 static MFEM_HOST_DEVICE inline findptsElementGPT_t get_pt(const double *elx[3],
@@ -312,34 +195,34 @@ static MFEM_HOST_DEVICE inline findptsElementGPT_t get_pt(const double *elx[3],
 /* check reduction in objective against prediction, and adjust
    trust region radius (p->tr) accordingly;
    may reject the prior step, returning 1; otherwise returns 0
-   sets out->dist2, out->index, out->x, out->oldr in any event,
-   leaving out->r, out->dr, out->flags to be set when returning 0 */
-static MFEM_HOST_DEVICE bool reject_prior_step_q(findptsElementPoint_t *out,
+   sets out_pt->dist2, out_pt->index, out_pt->x, out_pt->oldr in any event,
+   leaving out_pt->r, out_pt->dr, out_pt->flags to be set when returning 0 */
+static MFEM_HOST_DEVICE bool reject_prior_step_q(findptsElementPoint_t *out_pt,
                                                  const double resid[3],
                                                  const findptsElementPoint_t *p,
                                                  const double tol)
 {
-   const double dist2 = l2norm2(resid);
+   const double dist2 = l2norm2<sDIM>(resid);
    const double decr  = p->dist2 - dist2;
    const double pred  = p->dist2p;
    for (int d=0; d<sDIM; ++d)
    {
-      out->x[d] = p->x[d];
+      out_pt->x[d] = p->x[d];
    }
    for (int d=0; d<rDIM; ++d)
    {
-      out->oldr[d] = p->r[d];
+      out_pt->oldr[d] = p->r[d];
    }
-   out->dist2 = dist2;
+   out_pt->dist2 = dist2;
    if (decr>=0.01*pred)
    {
       if (decr>=0.9*pred)   // very good iteration
       {
-         out->tr = 2*p->tr;
+         out_pt->tr = 2*p->tr;
       }
       else   // good iteration
       {
-         out->tr = p->tr;
+         out_pt->tr = p->tr;
       }
       return false;
    }
@@ -351,17 +234,17 @@ static MFEM_HOST_DEVICE bool reject_prior_step_q(findptsElementPoint_t *out,
          which is why we divide by 4 below */
       double v0 = fabs(p->r[0] - p->oldr[0]),
              v1 = fabs(p->r[1] - p->oldr[1]);
-      out->tr     = ( v0>v1 ? v0 : v1 )/4;
-      out->dist2  = p->dist2;
-      out->flags   = p->flags >> 5;
-      out->dist2p = -HUGE_VAL;
+      out_pt->tr     = ( v0>v1 ? v0 : v1 )/4;
+      out_pt->dist2  = p->dist2;
+      out_pt->flags   = p->flags >> 5;
+      out_pt->dist2p = -HUGE_VAL;
       for (int d=0; d<rDIM; ++d)
       {
-         out->r[d] = p->oldr[d];
+         out_pt->r[d] = p->oldr[d];
       }
       if (pred<dist2*tol)
       {
-         out->flags |= CONVERGED_FLAG;
+         out_pt->flags |= CONVERGED_FLAG;
       }
       return true;
    }
@@ -369,7 +252,7 @@ static MFEM_HOST_DEVICE bool reject_prior_step_q(findptsElementPoint_t *out,
 
 /* minimize ||resid - jac * dr||_2, with |dr| <= tr, |r0+dr|<=1
    (exact solution of trust region problem) */
-static MFEM_HOST_DEVICE void newton_face( findptsElementPoint_t *const out,
+static MFEM_HOST_DEVICE void newton_face( findptsElementPoint_t *const out_pt,
                                           const double jac[sDIM*rDIM],
                                           const double rhes[3],
                                           const double resid[sDIM],
@@ -540,19 +423,19 @@ newton_face_constrained:
    }
 
 newton_face_fin:
-   out->dist2p = -2*v;
+   out_pt->dist2p = -2*v;
    dr[0] = r[0] - p->r[0];
    dr[1] = r[1] - p->r[1];
    if ( fabs(dr[0])+fabs(dr[1]) < tol)
    {
       new_flags |= CONVERGED_FLAG;
    }
-   out->r[0] = r[0], out->r[1] = r[1];
-   out->flags = new_flags | ((p->flags & FLAG_MASK)<<5);
+   out_pt->r[0] = r[0], out_pt->r[1] = r[1];
+   out_pt->flags = new_flags | ((p->flags & FLAG_MASK)<<5);
 }
 
 static MFEM_HOST_DEVICE inline void newton_edge(findptsElementPoint_t *const
-                                                out,
+                                                out_pt,
                                                 const double jac[sDIM*rDIM],
                                                 const double rhes,
                                                 const double resid[sDIM],
@@ -637,10 +520,10 @@ newton_edge_fin:
    {
       new_flags |= CONVERGED_FLAG;
    }
-   out->r[de] = nr;
-   out->r[dn] = p->r[dn];
-   out->dist2p = -v;
-   out->flags = flags | new_flags | ((p->flags & FLAG_MASK)<<5);
+   out_pt->r[de] = nr;
+   out_pt->r[dn] = p->r[dn];
+   out_pt->dist2p = -v;
+   out_pt->flags = flags | new_flags | ((p->flags & FLAG_MASK)<<5);
 #undef EVAL
 }
 
@@ -676,26 +559,27 @@ static MFEM_HOST_DEVICE void seed_j(const double *elx[sDIM],
 // global memory access of element coordinates.
 // Are the structs being stored in "local memory" or registers?
 template<int T_D1D = 0>
-static void FindPointsSurfLocal3D_Kernel(const int npt,
-                                         const double tol,
-                                         const double dist2tol,
-                                         const double *x,
-                                         const int point_pos_ordering,
-                                         const double *xElemCoord,
-                                         const int nel,
-                                         const double *wtend,
-                                         const double *boxinfo,
-                                         const int hash_n,
-                                         const double *hashMin,
-                                         const double *hashFac,
-                                         unsigned int *hashOffset,
-                                         unsigned int *const code_base,
-                                         unsigned int *const el_base,
-                                         double *const r_base,
-                                         double *const dist2_base,
-                                         const double *gll1D,
-                                         const double *lagcoeff,
-                                         const int pN = 0)
+static void FindPointsSurfLocal3DKernel(const int npt,
+                                        const double tol,
+                                        const double dist2tol,
+                                        const double *x,
+                                        const int point_pos_ordering,
+                                        const double *xElemCoord,
+                                        const int nel,
+                                        const double *wtend,
+                                        const double *boxinfo,
+                                        const bool obb_check,
+                                        const int hash_n,
+                                        const double *hashMin,
+                                        const double *hashFac,
+                                        unsigned int *hashOffset,
+                                        unsigned int *const code_base,
+                                        unsigned int *const el_base,
+                                        double *const r_base,
+                                        double *const dist2_base,
+                                        const double *gll1D,
+                                        const double *lagcoeff,
+                                        const int pN = 0)
 {
    const int MD1   = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
    const int D1D   = T_D1D ? T_D1D : pN;
@@ -753,22 +637,36 @@ static void FindPointsSurfLocal3D_Kernel(const int npt,
       {
          const unsigned int el = *elp;
 
-         // construct obbox on the fly
+         const int n_box_ents = obb_check ? (3*sDIM + sDIM2) : (2*sDIM);
+         bool pass_bb = true;
          obbox_t box;
-         int n_box_ents = 3*sDIM + sDIM2;
-         for (int idx = 0; idx < sDIM; ++idx)
+         if (obb_check)
          {
-            box.c0[idx] = boxinfo[n_box_ents*el + idx];
-            box.x[idx].min = boxinfo[n_box_ents*el + sDIM + idx];
-            box.x[idx].max = boxinfo[n_box_ents*el + 2*sDIM + idx];
+            // construct obbox on the fly
+            for (int idx = 0; idx < sDIM; ++idx)
+            {
+               box.c0[idx] = boxinfo[n_box_ents*el + idx];
+               box.x[idx].min = boxinfo[n_box_ents*el + sDIM + idx];
+               box.x[idx].max = boxinfo[n_box_ents*el + 2*sDIM + idx];
+            }
+
+            for (int idx = 0; idx < sDIM2; ++idx)
+            {
+               box.A[idx] = boxinfo[n_box_ents*el + 3*sDIM + idx];
+            }
+            pass_bb = (bbox_test(&box, x_i) >= 0);
+         }
+         else
+         {
+            for (int d = 0; d < sDIM; ++d)
+            {
+               box.x[d].min = boxinfo[n_box_ents*el + d];
+               box.x[d].max = boxinfo[n_box_ents*el + sDIM + d];
+            }
+            pass_bb = (AABB_test(&box, x_i) >= 0);
          }
 
-         for (int idx = 0; idx < sDIM2; ++idx)
-         {
-            box.A[idx] = boxinfo[n_box_ents*el + 3*sDIM + idx];
-         }
-
-         if (bbox_test(&box, x_i) < 0) { continue; }
+         if (!pass_bb) { continue; }
 
          //// findpts_local ////
          {
@@ -968,13 +866,19 @@ static void FindPointsSurfLocal3D_Kernel(const int npt,
                         double *hes_T = jac + sDIM*rDIM;
                         double *hes   = hes_T + hes_count*sDIM;
                         findptsElementGEdge_t edge;
+                        for (int d=0; d<sDIM; ++d)
+                        {
+                           edge.x[d]     = constraint_workspace + d*D1D;
+                           edge.dxdn[d]  = constraint_workspace + d*D1D
+                                           + sDIM*D1D;
+                           edge.d2xdn[d] = constraint_workspace + d*D1D
+                                           + 2*sDIM*D1D;
+                        }
 
                         MFEM_FOREACH_THREAD(j,x,D1D*sDIM)
                         {
-                           // utilized first D1D threads
-                           edge = get_edge(elx, wtend, ei,
-                                           constraint_workspace, edge_init, j,
-                                           D1D);
+                           // One thread per physical component and edge DOF.
+                           get_edge(elx, wtend, ei, edge_init, j, D1D, edge);
                         }
                         MFEM_SYNC_THREAD;
 
@@ -1045,7 +949,15 @@ static void FindPointsSurfLocal3D_Kernel(const int npt,
                               steep *= tmp->r[dn];
                               if (steep<0)
                               {
-                                 newton_face( fpt,jac,hes,resid,tmp->flags&CONVERGED_FLAG,tmp,tol);
+                                 double face_hes[3] =
+                                 {
+                                    dn == 0 ? hes[2] : hes[0],
+                                    hes[1],
+                                    dn == 0 ? hes[0] : hes[2]
+                                 };
+                                 newton_face(fpt, jac, face_hes, resid,
+                                             tmp->flags & CONVERGED_FLAG,
+                                             tmp, tol);
                               }
                               else
                               {
@@ -1211,29 +1123,42 @@ void FindPointsGSLIB::FindPointsSurfLocal3(const Vector &point_pos,
    auto pgll1d = DEV.gll1d.ReadWrite(use_dev);
    auto plc = DEV.lagcoeff.Read(use_dev);
    double dist2tol = DEV.surf_dist_tol;
+   const bool obb_chk = obb_check;
 
    switch (DEV.dof1d)
    {
       case 2:
-         return FindPointsSurfLocal3D_Kernel<2>(
-                   npt, DEV.newt_tol, dist2tol, pp, point_pos_ordering, pgslm,
-                   NE_split_total, pwt, pbb, DEV.lh_nx, plhm, plhf,
-                   plho, pcode, pelem, pref, pdist, pgll1d, plc);
+         FindPointsSurfLocal3DKernel<2>(npt, DEV.newt_tol, dist2tol,
+                                        pp, point_pos_ordering, pgslm,
+                                        NE_split_total, pwt, pbb, obb_chk,
+                                        DEV.lh_nx, plhm, plhf, plho,
+                                        pcode, pelem, pref, pdist,
+                                        pgll1d, plc);
+         break;
       case 3:
-         return FindPointsSurfLocal3D_Kernel<3>(
-                   npt, DEV.newt_tol, dist2tol, pp, point_pos_ordering, pgslm,
-                   NE_split_total, pwt, pbb, DEV.lh_nx, plhm, plhf,
-                   plho, pcode, pelem, pref, pdist, pgll1d, plc);
+         FindPointsSurfLocal3DKernel<3>(npt, DEV.newt_tol, dist2tol,
+                                        pp, point_pos_ordering, pgslm,
+                                        NE_split_total, pwt, pbb, obb_chk,
+                                        DEV.lh_nx, plhm, plhf, plho,
+                                        pcode, pelem, pref, pdist,
+                                        pgll1d, plc);
+         break;
       case 4:
-         return FindPointsSurfLocal3D_Kernel<4>(
-                   npt, DEV.newt_tol, dist2tol, pp, point_pos_ordering, pgslm,
-                   NE_split_total, pwt, pbb, DEV.lh_nx, plhm, plhf,
-                   plho, pcode, pelem, pref, pdist, pgll1d, plc);
+         FindPointsSurfLocal3DKernel<4>(npt, DEV.newt_tol, dist2tol,
+                                        pp, point_pos_ordering, pgslm,
+                                        NE_split_total, pwt, pbb, obb_chk,
+                                        DEV.lh_nx, plhm, plhf, plho,
+                                        pcode, pelem, pref, pdist,
+                                        pgll1d, plc);
+         break;
       default:
-         return FindPointsSurfLocal3D_Kernel(
-                   npt, DEV.newt_tol, dist2tol, pp, point_pos_ordering, pgslm,
-                   NE_split_total, pwt, pbb, DEV.lh_nx, plhm, plhf,
-                   plho, pcode, pelem, pref, pdist, pgll1d, plc, DEV.dof1d);
+         FindPointsSurfLocal3DKernel(npt, DEV.newt_tol, dist2tol, pp,
+                                     point_pos_ordering, pgslm,
+                                     NE_split_total, pwt, pbb, obb_chk,
+                                     DEV.lh_nx, plhm, plhf, plho,
+                                     pcode, pelem, pref, pdist,
+                                     pgll1d, plc, DEV.dof1d);
+         break;
    }
 }
 

fem/gslib/gslib_kernel_helpers.hpp

@@ -0,0 +1,190 @@
+#ifndef MFEM_GSLIB_KERNEL_HELPERS_HPP
+#define MFEM_GSLIB_KERNEL_HELPERS_HPP
+
+#include "../../config/config.hpp"
+
+#include <cmath>
+
+namespace mfem
+{
+
+namespace gslib
+{
+
+struct dbl_range_t
+{
+   double min, max;
+};
+
+template <int SDIM>
+struct obbox_t
+{
+   double c0[SDIM], A[SDIM * SDIM];
+   dbl_range_t x[SDIM];
+};
+
+template <int SDIM>
+struct findptsLocalHashData_t
+{
+   int hash_n;
+   dbl_range_t bnd[SDIM];
+   double fac[SDIM];
+   unsigned int *offset;
+};
+
+// Eval the ith Lagrange interpolant at x.
+MFEM_HOST_DEVICE inline void lagrange_eval(double *p0, double x,
+                                           int i, int p_Nq,
+                                           double *z, double *lagrangeCoeff)
+{
+   double p_i = (1 << (p_Nq - 1));
+   for (int j = 0; j < p_Nq; ++j)
+   {
+      const double d_j = x - z[j];
+      p_i *= j == i ? 1 : d_j;
+   }
+   p0[i] = lagrangeCoeff[i] * p_i;
+}
+
+// Eval the ith Lagrange interpolant and its first derivative at x.
+MFEM_HOST_DEVICE inline void lag_eval_first_der(double *p0, double x,
+                                                int i, const double *z,
+                                                const double *lCoeff,
+                                                int pN)
+{
+   double u0 = 1, u1 = 0;
+   for (int j = 0; j < pN; ++j)
+   {
+      if (i != j)
+      {
+         const double d_j = 2 * (x - z[j]);
+         u1 = d_j * u1 + u0;
+         u0 = d_j * u0;
+      }
+   }
+   p0[i] = lCoeff[i] * u0;
+   p0[pN + i] = 2.0 * lCoeff[i] * u1;
+}
+
+// Eval the ith Lagrange interpolant and its first and second derivative at x.
+MFEM_HOST_DEVICE inline void lag_eval_second_der(double *p0, double x,
+                                                 int i, const double *z,
+                                                 const double *lCoeff,
+                                                 int pN)
+{
+   double u0 = 1, u1 = 0, u2 = 0;
+   for (int j = 0; j < pN; ++j)
+   {
+      if (i != j)
+      {
+         const double d_j = 2 * (x - z[j]);
+         u2 = d_j * u2 + u1;
+         u1 = d_j * u1 + u0;
+         u0 = d_j * u0;
+      }
+   }
+   p0[i] = lCoeff[i] * u0;
+   p0[pN + i] = 2.0 * lCoeff[i] * u1;
+   p0[2 * pN + i] = 8.0 * lCoeff[i] * u2;
+}
+
+// Solve Ax=y where A is a symmetric 2x2 matrix packed as {a00, a01, a11}.
+MFEM_HOST_DEVICE inline void lin_solve_sym_2(double x[2],
+                                             const double A[3],
+                                             const double y[2])
+{
+   const double idet = 1 / (A[0] * A[2] - A[1] * A[1]);
+   x[0] = idet * (A[2] * y[0] - A[1] * y[1]);
+   x[1] = idet * (A[0] * y[1] - A[1] * y[0]);
+}
+
+// Positive when the point is inside the axis-aligned bounding box.
+template <int SDIM>
+MFEM_HOST_DEVICE inline double AABB_test(const obbox_t<SDIM> *const b,
+                                         const double (&x)[SDIM])
+{
+   double test = 1.0;
+   for (int d = 0; d < SDIM; ++d)
+   {
+      const double b_d = (x[d] - b->x[d].min) * (b->x[d].max - x[d]);
+      test = test < 0.0 ? test : b_d;
+   }
+   return test;
+}
+
+// Positive when the point is inside the oriented bounding box.
+template <int SDIM>
+MFEM_HOST_DEVICE inline double bbox_test(const obbox_t<SDIM> *const b,
+                                         const double (&x)[SDIM])
+{
+   const double bxyz = AABB_test(b, x);
+   if (bxyz < 0.0)
+   {
+      return bxyz;
+   }
+
+   double dxyz[SDIM];
+   for (int d = 0; d < SDIM; ++d)
+   {
+      dxyz[d] = x[d] - b->c0[d];
+   }
+
+   double test = 1.0;
+   for (int d = 0; d < SDIM; ++d)
+   {
+      double rst = 0.0;
+      for (int e = 0; e < SDIM; ++e)
+      {
+         rst += b->A[d * SDIM + e] * dxyz[e];
+      }
+      const double brst = (rst + 1.0) * (1.0 - rst);
+      test = test < 0.0 ? test : brst;
+   }
+   return test;
+}
+
+// Hash index in the hash table for the point x.
+template <int SDIM>
+MFEM_HOST_DEVICE inline int hash_index(
+   const findptsLocalHashData_t<SDIM> *const p,
+   const double (&x)[SDIM])
+{
+   const int n = p->hash_n;
+   int sum = 0;
+   for (int d = SDIM - 1; d >= 0; --d)
+   {
+      sum *= n;
+      const int i = (int)floor((x[d] - p->bnd[d].min) * p->fac[d]);
+      sum += i < 0 ? 0 : (n - 1 < i ? n - 1 : i);
+   }
+   return sum;
+}
+
+// Squared Euclidean norm.
+template <int SDIM>
+MFEM_HOST_DEVICE inline double l2norm2(const double (&x)[SDIM])
+{
+   double sum = 0.0;
+   for (int d = 0; d < SDIM; ++d)
+   {
+      sum += x[d] * x[d];
+   }
+   return sum;
+}
+
+template <int SDIM>
+MFEM_HOST_DEVICE inline double l2norm2(const double *x)
+{
+   double sum = 0.0;
+   for (int d = 0; d < SDIM; ++d)
+   {
+      sum += x[d] * x[d];
+   }
+   return sum;
+}
+
+} // namespace gslib
+
+} // namespace mfem
+
+#endif

fem/gslib/interpolate_local_1.cpp

@@ -11,7 +11,7 @@
 
 #include "../gslib.hpp"
 #include "../../general/forall.hpp"
-#include "../../linalg/kernels.hpp"
+#include "gslib_kernel_helpers.hpp"
 
 #ifdef MFEM_USE_GSLIB
 
@@ -33,17 +33,7 @@ namespace mfem
 #define CODE_BORDER 1
 #define CODE_NOT_FOUND 2
 
-static MFEM_HOST_DEVICE void lagrange_eval(double *p0, double x,
-                                           int i, int p_Nq,
-                                           double *z, double *lagrangeCoeff)
-{
-   double p_i = (1 << (p_Nq - 1));
-   for (int j=0; j<p_Nq; ++j)
-   {
-      p_i *= j==i ? 1 : x-z[j];
-   }
-   p0[i] = lagrangeCoeff[i] * p_i;
-}
+using gslib::lagrange_eval;
 
 template<int T_D1D = 0>
 static void InterpolateLocal1DKernel(const double *const gf_in,
@@ -123,21 +113,26 @@ void FindPointsGSLIB::InterpolateLocal1( const Vector &field_in,
    auto plcf = DEV.lagcoeff_sol.ReadWrite(use_dev);
    switch (dof1Dsol)
    {
-      case 2: return InterpolateLocal1DKernel<2>(pfin, pgsl, pgslr, pfout,
-                                                    npt, ncomp,
-                                                    pgll, plcf);
-      case 3: return InterpolateLocal1DKernel<3>(pfin, pgsl, pgslr, pfout,
-                                                    npt, ncomp,
-                                                    pgll, plcf);
-      case 4: return InterpolateLocal1DKernel<4>(pfin, pgsl, pgslr, pfout,
-                                                    npt, ncomp,
-                                                    pgll, plcf);
-      case 5: return InterpolateLocal1DKernel<5>(pfin, pgsl, pgslr, pfout,
-                                                    npt, ncomp,
-                                                    pgll, plcf);
-      default: return InterpolateLocal1DKernel(pfin, pgsl, pgslr, pfout,
-                                                  npt, ncomp,
-                                                  pgll, plcf, dof1Dsol);
+      case 2:
+         InterpolateLocal1DKernel<2>(pfin, pgsl, pgslr, pfout,
+                                     npt, ncomp, pgll, plcf);
+         break;
+      case 3:
+         InterpolateLocal1DKernel<3>(pfin, pgsl, pgslr, pfout,
+                                     npt, ncomp, pgll, plcf);
+         break;
+      case 4:
+         InterpolateLocal1DKernel<4>(pfin, pgsl, pgslr, pfout,
+                                     npt, ncomp, pgll, plcf);
+         break;
+      case 5:
+         InterpolateLocal1DKernel<5>(pfin, pgsl, pgslr, pfout,
+                                     npt, ncomp, pgll, plcf);
+         break;
+      default:
+         InterpolateLocal1DKernel(pfin, pgsl, pgslr, pfout,
+                                  npt, ncomp, pgll, plcf, dof1Dsol);
+         break;
    }
 }
 #undef CODE_INTERNAL

fem/gslib/interpolate_local_2.cpp

@@ -11,6 +11,7 @@
 
 #include "../gslib.hpp"
 #include "../../general/forall.hpp"
+#include "gslib_kernel_helpers.hpp"
 
 #ifdef MFEM_USE_GSLIB
 
@@ -32,18 +33,7 @@ namespace mfem
 #define CODE_BORDER 1
 #define CODE_NOT_FOUND 2
 
-static MFEM_HOST_DEVICE void lagrange_eval(double *p0, double x,
-                                           int i, int p_Nq,
-                                           double *z, double *lagrangeCoeff)
-{
-   double p_i = (1 << (p_Nq - 1));
-   for (int j = 0; j < p_Nq; ++j)
-   {
-      double d_j = x - z[j];
-      p_i *= j == i ? 1 : d_j;
-   }
-   p0[i] = lagrangeCoeff[i] * p_i;
-}
+using gslib::lagrange_eval;
 
 template<int T_D1D = 0>
 static void InterpolateLocal2DKernel(const double *const gf_in,
@@ -132,21 +122,26 @@ void FindPointsGSLIB::InterpolateLocal2(const Vector &field_in,
    auto plcf = DEV.lagcoeff_sol.ReadWrite(use_dev);
    switch (dof1Dsol)
    {
-      case 2: return InterpolateLocal2DKernel<2>(pfin, pgsl, pgslr, pfout,
-                                                    npt, ncomp,
-                                                    pgll, plcf);
-      case 3: return InterpolateLocal2DKernel<3>(pfin, pgsl, pgslr, pfout,
-                                                    npt, ncomp,
-                                                    pgll, plcf);
-      case 4: return InterpolateLocal2DKernel<4>(pfin, pgsl, pgslr, pfout,
-                                                    npt, ncomp,
-                                                    pgll, plcf);
-      case 5: return InterpolateLocal2DKernel<5>(pfin, pgsl, pgslr, pfout,
-                                                    npt, ncomp,
-                                                    pgll, plcf);
-      default: return InterpolateLocal2DKernel(pfin, pgsl, pgslr, pfout,
-                                                  npt, ncomp,
-                                                  pgll, plcf, dof1Dsol);
+      case 2:
+         InterpolateLocal2DKernel<2>(pfin, pgsl, pgslr, pfout,
+                                     npt, ncomp, pgll, plcf);
+         break;
+      case 3:
+         InterpolateLocal2DKernel<3>(pfin, pgsl, pgslr, pfout,
+                                     npt, ncomp, pgll, plcf);
+         break;
+      case 4:
+         InterpolateLocal2DKernel<4>(pfin, pgsl, pgslr, pfout,
+                                     npt, ncomp, pgll, plcf);
+         break;
+      case 5:
+         InterpolateLocal2DKernel<5>(pfin, pgsl, pgslr, pfout,
+                                     npt, ncomp, pgll, plcf);
+         break;
+      default:
+         InterpolateLocal2DKernel(pfin, pgsl, pgslr, pfout,
+                                  npt, ncomp, pgll, plcf, dof1Dsol);
+         break;
    }
 }
 

fem/gslib/interpolate_local_3.cpp

@@ -11,6 +11,7 @@
 
 #include "../gslib.hpp"
 #include "../../general/forall.hpp"
+#include "gslib_kernel_helpers.hpp"
 
 #ifdef MFEM_USE_GSLIB
 
@@ -32,18 +33,7 @@ namespace mfem
 #define CODE_BORDER 1
 #define CODE_NOT_FOUND 2
 
-static MFEM_HOST_DEVICE void lagrange_eval(double *p0, double x,
-                                           int i, int p_Nq,
-                                           double *z, double *lagrangeCoeff)
-{
-   double p_i = (1 << (p_Nq - 1));
-   for (int j = 0; j < p_Nq; ++j)
-   {
-      double d_j = x - z[j];
-      p_i *= j == i ? 1 : d_j;
-   }
-   p0[i] = lagrangeCoeff[i] * p_i;
-}
+using gslib::lagrange_eval;
 
 template<int T_D1D = 0>
 static void InterpolateLocal3DKernel(const double *const gf_in,
@@ -135,21 +125,26 @@ void FindPointsGSLIB::InterpolateLocal3(const Vector &field_in,
    auto plcf = DEV.lagcoeff_sol.ReadWrite(use_dev);
    switch (dof1Dsol)
    {
-      case 2: return InterpolateLocal3DKernel<2>(pfin, pgsle, pgslr, pfout,
-                                                    npt, ncomp,
-                                                    pgll, plcf);
-      case 3: return InterpolateLocal3DKernel<3>(pfin, pgsle, pgslr, pfout,
-                                                    npt, ncomp,
-                                                    pgll, plcf);
-      case 4: return InterpolateLocal3DKernel<4>(pfin, pgsle, pgslr, pfout,
-                                                    npt, ncomp,
-                                                    pgll, plcf);
-      case 5: return InterpolateLocal3DKernel<5>(pfin, pgsle, pgslr, pfout,
-                                                    npt, ncomp,
-                                                    pgll, plcf);
-      default: return InterpolateLocal3DKernel(pfin, pgsle, pgslr, pfout,
-                                                  npt, ncomp,
-                                                  pgll, plcf, dof1Dsol);
+      case 2:
+         InterpolateLocal3DKernel<2>(pfin, pgsle, pgslr, pfout,
+                                     npt, ncomp, pgll, plcf);
+         break;
+      case 3:
+         InterpolateLocal3DKernel<3>(pfin, pgsle, pgslr, pfout,
+                                     npt, ncomp, pgll, plcf);
+         break;
+      case 4:
+         InterpolateLocal3DKernel<4>(pfin, pgsle, pgslr, pfout,
+                                     npt, ncomp, pgll, plcf);
+         break;
+      case 5:
+         InterpolateLocal3DKernel<5>(pfin, pgsle, pgslr, pfout,
+                                     npt, ncomp, pgll, plcf);
+         break;
+      default:
+         InterpolateLocal3DKernel(pfin, pgsle, pgslr, pfout,
+                                  npt, ncomp, pgll, plcf, dof1Dsol);
+         break;
    }
 }
 

fem/integ/bilininteg_elasticity_kernels.cpp

@@ -91,15 +91,15 @@ void ElasticityAddMultPA(const int dim, const int nDofs,
 void ElasticityAssembleDiagonalPA(const int dim, const int nDofs,
                                   const CoefficientVector &lambda,
                                   const CoefficientVector &mu, const GeometricFactors &geom,
-                                  const DofToQuad &maps, QuadratureFunction &QVec, Vector &diag)
+                                  const DofToQuad &maps, const IntegrationRule &ir, Vector &diag)
 {
    switch (dim)
    {
       case 2:
-         ElasticityAssembleDiagonalPA_<2>(nDofs, lambda, mu, geom, maps, QVec, diag);
+         ElasticityAssembleDiagonalPA_<2>(nDofs, lambda, mu, geom, maps, ir, diag);
          break;
       case 3:
-         ElasticityAssembleDiagonalPA_<3>(nDofs, lambda, mu, geom, maps, QVec, diag);
+         ElasticityAssembleDiagonalPA_<3>(nDofs, lambda, mu, geom, maps, ir, diag);
          break;
       default:
          MFEM_ABORT("Only dimensions 2 and 3 supported.");

fem/integ/bilininteg_elasticity_kernels.hpp

@@ -38,7 +38,6 @@
 #include "../../linalg/vector.hpp"
 #include "../../linalg/tensor.hpp"
 #include "../quadinterpolator.hpp"
-#include "../bilininteg.hpp"
 #include "../coefficient.hpp"
 #include "../qfunction.hpp"
 
@@ -133,12 +132,12 @@ void ElasticityAssembleEA(const int dim, const int i_block, const int j_block,
 /// @param[in] mu Quadrature function for second Lame param.
 /// @param[in] geom Geometric factors corresponding to fespace.
 /// @param[in] maps DofToQuad maps for one element (assume elements all same).
-/// @param QVec Scratch Q-Vector. nQuad x dim x dim x dim x dim x numEls.
+/// @param[in] ir Integration rule.
 /// @param[out] diag diagonal of A. nDofs x dim x numEls.
 void ElasticityAssembleDiagonalPA(const int dim, const int nDofs,
                                   const CoefficientVector &lambda,
                                   const CoefficientVector &mu, const GeometricFactors &geom,
-                                  const DofToQuad &maps, QuadratureFunction &QVec, Vector &diag);
+                                  const DofToQuad &maps, const IntegrationRule &ir, Vector &diag);
 
 /// Templated implementation of ElasticityAddMultPA.
 template<int dim, int i_block = -1, int j_block = -1>
@@ -280,77 +279,67 @@ void ElasticityAddMultPA_(const int nDofs, const FiniteElementSpace &fespace,
 template<int dim>
 void ElasticityAssembleDiagonalPA_(const int nDofs,
                                    const CoefficientVector &lambda,
-                                   const CoefficientVector &mu, const GeometricFactors &geom,
-                                   const DofToQuad &maps, QuadratureFunction &QVec, Vector &diag)
+                                   const CoefficientVector &mu,
+                                   const GeometricFactors &geom,
+                                   const DofToQuad &maps,
+                                   const IntegrationRule &ir,
+                                   Vector &diag)
 {
-   using future::tensor;
-   using future::make_tensor;
    using future::det;
    using future::inv;
+   using future::make_tensor;
+   using future::tensor;
 
    // Assuming all elements are the same
-   const auto &ir = QVec.GetIntRule(0);
    static constexpr int d = dim;
    const int numPoints = ir.GetNPoints();
-   const int numEls = lambda.Size()/numPoints;
+   const int numEls = lambda.Size() / numPoints;
+
    const auto lamDev = Reshape(lambda.Read(), numPoints, numEls);
    const auto muDev = Reshape(mu.Read(), numPoints, numEls);
    const auto J = Reshape(geom.J.Read(), numPoints, d, d, numEls);
-   auto Q = Reshape(QVec.ReadWrite(), numPoints, d,d, d, numEls);
    const real_t *ipWeights = ir.GetWeights().Read();
-   mfem::forall_2D(numEls, numPoints,1, [=] MFEM_HOST_DEVICE (int e)
-   {
-      MFEM_FOREACH_THREAD(p, x,numPoints)
-      {
-         auto invJ = inv(make_tensor<d, d>(
-         [&](int i, int j) { return J(p, i, j, e); }));
-         const real_t w = ipWeights[p] /det(invJ);
-         for (int n = 0; n < d; n++)
-         {
-            for (int m = 0; m < d; m++)
-            {
-               for (int q = 0; q < d; q++)
-               {
-                  // compute contraction of 4*sym(grad(u))sym(grad(v)) term.
-                  // this contraction could be made slightly cheaper using Voigt
-                  // notation, but repeated entries are summed for simplicity.
-                  real_t contraction = 0.;
-                  for (int a = 0; a < d; a++)
-                  {
-                     for (int b = 0; b < d; b++)
-                     {
-                        contraction += ((a == q)*invJ(m,b) + (b==q)*invJ(m,a))*((a == q)
-                                                                                *invJ(n, b) + (b==q)*invJ(n,a));
-                     }
-                  }
-                  // lambda*div(u)*div(v) + 2*mu*sym(grad(u))*sym(grad(v))
-                  // contraction = 4*sym(grad(u))sym(grad(v))
-                  Q(p,m,n,q,e) = w*(lamDev(p, e)*invJ(m,q)*invJ(n,q)
-                                    + 0.5*muDev(p, e)*contraction);
-               }
-            }
-         }
-      }
-   });
-
-   // Reduce quadrature function to an E-Vector
-   const auto QRead = Reshape(QVec.Read(), numPoints, d, d, d, numEls);
-   auto diagDev = Reshape(diag.Write(), nDofs, d, numEls);
    const auto G = Reshape(maps.G.Read(), numPoints, d, nDofs);
+   auto diagDev = Reshape(diag.Write(), nDofs, d, numEls);
+
    mfem::forall_2D(numEls, d, nDofs, [=] MFEM_HOST_DEVICE (int e)
    {
-      MFEM_FOREACH_THREAD(i, y, nDofs)
+      MFEM_FOREACH_THREAD_DIRECT(i, y, nDofs)
       {
-         MFEM_FOREACH_THREAD(q, x, d)
+         MFEM_FOREACH_THREAD_DIRECT(q, x, d)
          {
-            real_t sum = 0.;
-            for (int n = 0; n < d; n++)
+            real_t sum = 0.0;
+            for (int p = 0; p < numPoints; p++)
             {
-               for (int m = 0; m < d; m++)
+               const auto invJ = inv(make_tensor<d, d>([&](int r, int c)
                {
-                  for (int p = 0; p < numPoints; p++ )
+                  return J(p, r, c, e);
+               }));
+               const real_t w = ipWeights[p] / det(invJ);
+
+               for (int n = 0; n < d; n++)
+               {
+                  for (int m = 0; m < d; m++)
                   {
-                     sum += QRead(p,m,n,q,e)*G(p,m,i)*G(p,n,i);
+                     // compute contraction of 4*sym(grad(u))sym(grad(v)) term.
+                     // this contraction could be made slightly cheaper using Voigt
+                     // notation, but repeated entries are summed for simplicity.
+                     real_t contraction = 0.0;
+                     for (int a = 0; a < d; a++)
+                     {
+                        for (int b = 0; b < d; b++)
+                        {
+                           contraction +=
+                              ((a == q) * invJ(m, b) + (b == q) * invJ(m, a)) *
+                              ((a == q) * invJ(n, b) + (b == q) * invJ(n, a));
+                        }
+                     }
+                     // lambda*div(u)*div(v) + 2*mu*sym(grad(u))*sym(grad(v))
+                     // contraction = 4*sym(grad(u))sym(grad(v))
+                     const real_t Q =
+                        w * (lamDev(p, e) * invJ(m, q) * invJ(n, q)
+                             + 0.5 * muDev(p, e) * contraction);
+                     sum += Q * G(p, m, i) * G(p, n, i);
                   }
                }
             }

fem/integ/bilininteg_elasticity_pa.cpp

@@ -10,7 +10,6 @@
 // CONTRIBUTING.md for details.
 
 #include "../bilininteg.hpp"
-#include "../gridfunc.hpp"
 #include "../qfunction.hpp"
 #include "bilininteg_elasticity_kernels.hpp"
 
@@ -59,9 +58,8 @@ void ElasticityIntegrator::AssemblePA(const FiniteElementSpace &fes)
 
 void ElasticityIntegrator::AssembleDiagonalPA(Vector &diag)
 {
-   q_vec->SetVDim(vdim*vdim*vdim*vdim);
    internal::ElasticityAssembleDiagonalPA(vdim, ndofs, *lambda_quad, *mu_quad,
-                                          *geom, *maps, *q_vec, diag);
+                                          *geom, *maps, *IntRule, diag);
 }
 
 void ElasticityIntegrator::AddMultPA(const Vector &x, Vector &y) const

fem/integ/bilininteg_vecdiv_pa.hpp

@@ -0,0 +1,365 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+#pragma once
+
+#include "../../config/config.hpp"
+#include "../../general/array.hpp"
+#include "../../general/forall.hpp"
+#include "../../linalg/dtensor.hpp"
+#include "../../linalg/vector.hpp"
+#include "../bilininteg.hpp"
+#include "../kernels.hpp"
+
+namespace mfem
+{
+
+/// \cond DO_NOT_DOCUMENT
+
+namespace internal
+{
+
+// Shared memory PA Divergence Apply 2D kernel
+template<int T_TR_D1D = 0, int T_TE_D1D = 0, int T_Q1D = 0>
+inline void SmemPADivergenceApply2D(const int NE,
+                                    const Array<real_t> &b_,
+                                    const Array<real_t> &g_,
+                                    const Array<real_t> &bt_,
+                                    const Vector &q_,
+                                    const Vector &x_,
+                                    Vector &y_,
+                                    const int tr_d1d = 0,
+                                    const int te_d1d = 0,
+                                    const int q1d = 0)
+{
+   const int TR_D1D = T_TR_D1D ? T_TR_D1D : tr_d1d;
+   const int TE_D1D = T_TE_D1D ? T_TE_D1D : te_d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+
+   MFEM_VERIFY(TR_D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(TE_D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+
+   const auto B = b_.Read(), G = g_.Read(), Bt = bt_.Read();
+   const auto Q = Reshape(q_.Read(), Q1D, Q1D, 2, 2, NE);
+   const auto X = Reshape(x_.Read(), TR_D1D, TR_D1D, 2, NE);
+   auto Y = Reshape(y_.ReadWrite(), TE_D1D, TE_D1D, 1, NE);
+
+   mfem::forall_2D<T_Q1D * T_Q1D>(NE, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
+   {
+      constexpr int MQ1 = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+
+      MFEM_SHARED real_t smem[MQ1][MQ1];
+      MFEM_SHARED real_t sB[MQ1][MQ1], sG[MQ1][MQ1];
+
+      kernels::internal::vd_regs2d_t<2, 2, MQ1> g0, g1;
+      kernels::internal::v_regs2d_t<1, MQ1> r0, r1;
+
+      kernels::internal::LoadMatrix(TR_D1D, Q1D, B, sB);
+      kernels::internal::LoadMatrix(TR_D1D, Q1D, G, sG);
+
+      kernels::internal::LoadDofs2d(e, TR_D1D, X, g0);
+      kernels::internal::Grad2d(TR_D1D, Q1D, smem, sB, sG, g0, g1);
+
+      MFEM_FOREACH_THREAD_DIRECT(qy, y, Q1D)
+      {
+         MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+         {
+            r0[0][qy][qx] =
+               g1[0][0][qy][qx] * Q(qx, qy, 0, 0, e) +
+               g1[0][1][qy][qx] * Q(qx, qy, 1, 0, e) +
+               g1[1][0][qy][qx] * Q(qx, qy, 0, 1, e) +
+               g1[1][1][qy][qx] * Q(qx, qy, 1, 1, e);
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      kernels::internal::LoadMatrix<MQ1,true>(TE_D1D, Q1D, Bt, sB);
+      kernels::internal::EvalTranspose2d(TE_D1D, Q1D, smem, sB, r0, r1);
+      kernels::internal::WriteDofs2d(e, TE_D1D, r1, Y);
+   });
+}
+
+// Shared memory PA Divergence Apply 2D kernel transpose
+template<int T_TR_D1D = 0, int T_TE_D1D = 0, int T_Q1D = 0>
+inline void SmemPADivergenceApplyTranspose2D(const int NE,
+                                             const Array<real_t> &bt,
+                                             const Array<real_t> &gt,
+                                             const Array<real_t> &b,
+                                             const Vector &q_,
+                                             const Vector &x_,
+                                             Vector &y_,
+                                             const int tr_d1d = 0,
+                                             const int te_d1d = 0,
+                                             const int q1d = 0)
+{
+   const int TR_D1D = T_TR_D1D ? T_TR_D1D : tr_d1d;
+   const int TE_D1D = T_TE_D1D ? T_TE_D1D : te_d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+
+   MFEM_VERIFY(TR_D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(TE_D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+
+   const auto Bt = bt.Read(), Gt = gt.Read(), B = b.Read();
+   const auto Q = Reshape(q_.Read(), Q1D, Q1D, 2, 2, NE);
+   const auto X  = Reshape(x_.Read(), TE_D1D, TE_D1D, 1, NE);
+   auto Y  = Reshape(y_.ReadWrite(), TR_D1D, TR_D1D, 2, NE);
+
+   mfem::forall_2D<T_Q1D * T_Q1D>(NE, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
+   {
+      constexpr int MQ1 = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+
+      MFEM_SHARED real_t smem[MQ1][MQ1];
+      MFEM_SHARED real_t sB[MQ1][MQ1], sG[MQ1][MQ1];
+
+      kernels::internal::v_regs2d_t<1, MQ1> r0, r1;
+      kernels::internal::vd_regs2d_t<2, 2, MQ1> g0, g1;
+
+      kernels::internal::LoadMatrix(TE_D1D, Q1D, B, sB);
+      kernels::internal::LoadDofs2d(e, TE_D1D, X, r0);
+      kernels::internal::Eval2d(TE_D1D, Q1D, smem, sB, r0, r1);
+
+      MFEM_FOREACH_THREAD_DIRECT(qy, y, Q1D)
+      {
+         MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+         {
+            g0[0][0][qy][qx] = r1[0][qy][qx] * Q(qx, qy, 0, 0, e);
+            g0[0][1][qy][qx] = r1[0][qy][qx] * Q(qx, qy, 1, 0, e);
+            g0[1][0][qy][qx] = r1[0][qy][qx] * Q(qx, qy, 0, 1, e);
+            g0[1][1][qy][qx] = r1[0][qy][qx] * Q(qx, qy, 1, 1, e);
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      kernels::internal::LoadMatrix<MQ1,true>(TR_D1D, Q1D, Bt, sB);
+      kernels::internal::LoadMatrix<MQ1,true>(TR_D1D, Q1D, Gt, sG);
+      kernels::internal::GradTranspose2d(TR_D1D, Q1D, smem, sB, sG, g0, g1);
+      kernels::internal::WriteDofs2d(e, TR_D1D, g1, Y);
+   });
+}
+
+// Shared memory PA Divergence Apply 3D kernel transpose
+template<int T_TR_D1D = 0, int T_TE_D1D = 0, int T_Q1D = 0>
+inline void SmemPADivergenceApplyTranspose3D(const int NE,
+                                             const Array<real_t> &bt,
+                                             const Array<real_t> &gt,
+                                             const Array<real_t> &b,
+                                             const Vector &q_,
+                                             const Vector &x_,
+                                             Vector &y_,
+                                             int tr_d1d = 0,
+                                             int te_d1d = 0,
+                                             int q1d = 0)
+{
+   const int TR_D1D = T_TR_D1D ? T_TR_D1D : tr_d1d;
+   const int TE_D1D = T_TE_D1D ? T_TE_D1D : te_d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+
+   MFEM_VERIFY(TR_D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(TE_D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+
+   const auto Bt = bt.Read(), Gt = gt.Read(), B = b.Read();
+   const auto Q = Reshape(q_.Read(), Q1D, Q1D, Q1D, 3, 3, NE);
+   const auto X = Reshape(x_.Read(), TE_D1D, TE_D1D, TE_D1D, 1, NE);
+   auto Y = Reshape(y_.ReadWrite(), TR_D1D, TR_D1D, TR_D1D, 3, NE);
+
+   mfem::forall_2D<T_Q1D * T_Q1D>(NE, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
+   {
+      constexpr int MQ1 = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+
+      MFEM_SHARED real_t smem[MQ1][MQ1];
+      MFEM_SHARED real_t sB[MQ1][MQ1], sG[MQ1][MQ1];
+
+      kernels::internal::v_regs3d_t<1, MQ1> r0, r1;
+      kernels::internal::vd_regs3d_t<3, 3, MQ1> g0, g1;
+
+      kernels::internal::LoadMatrix(TE_D1D, Q1D, B, sB);
+      kernels::internal::LoadDofs3d(e, TE_D1D, X, r0);
+      kernels::internal::Eval3d(TE_D1D, Q1D, smem, sB, r0, r1);
+
+      for (int qz = 0; qz < Q1D; qz++)
+      {
+         MFEM_FOREACH_THREAD_DIRECT(qy, y, Q1D)
+         {
+            MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+            {
+               const auto r = r1[0][qz][qy][qx];
+               g0[0][0][qz][qy][qx] = r * Q(qx, qy, qz, 0, 0, e);
+               g0[0][1][qz][qy][qx] = r * Q(qx, qy, qz, 1, 0, e);
+               g0[0][2][qz][qy][qx] = r * Q(qx, qy, qz, 2, 0, e);
+
+               g0[1][0][qz][qy][qx] = r * Q(qx, qy, qz, 0, 1, e);
+               g0[1][1][qz][qy][qx] = r * Q(qx, qy, qz, 1, 1, e);
+               g0[1][2][qz][qy][qx] = r * Q(qx, qy, qz, 2, 1, e);
+
+               g0[2][0][qz][qy][qx] = r * Q(qx, qy, qz, 0, 2, e);
+               g0[2][1][qz][qy][qx] = r * Q(qx, qy, qz, 1, 2, e);
+               g0[2][2][qz][qy][qx] = r * Q(qx, qy, qz, 2, 2, e);
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      kernels::internal::LoadMatrix<MQ1,true>(TR_D1D, Q1D, Bt, sB);
+      kernels::internal::LoadMatrix<MQ1,true>(TR_D1D, Q1D, Gt, sG);
+      kernels::internal::GradTranspose3d(TR_D1D, Q1D, smem, sB, sG, g0, g1);
+      kernels::internal::WriteDofs3d(e, TR_D1D, g1, Y);
+   });
+}
+
+// Shared memory PA Divergence Apply 3D kernel
+template<int T_TR_D1D = 0, int T_TE_D1D = 0, int T_Q1D = 0>
+inline void SmemPADivergenceApply3D(const int NE,
+                                    const Array<real_t> &b_,
+                                    const Array<real_t> &g_,
+                                    const Array<real_t> &bt_,
+                                    const Vector &q_,
+                                    const Vector &x_,
+                                    Vector &y_,
+                                    const int tr_d1d = 0,
+                                    const int te_d1d = 0,
+                                    const int q1d = 0)
+{
+   const int TR_D1D = T_TR_D1D ? T_TR_D1D : tr_d1d;
+   const int TE_D1D = T_TE_D1D ? T_TE_D1D : te_d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+
+   MFEM_VERIFY(TR_D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(TE_D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+
+   const auto B = b_.Read(), G = g_.Read(), Bt = bt_.Read();
+   const auto Q = Reshape(q_.Read(), Q1D, Q1D, Q1D, 3,3, NE);
+   const auto X = Reshape(x_.Read(), TR_D1D, TR_D1D, TR_D1D, 3, NE);
+   auto Y = Reshape(y_.ReadWrite(), TE_D1D, TE_D1D, TE_D1D, 1, NE);
+
+   mfem::forall_2D<T_Q1D*T_Q1D>(NE, Q1D, Q1D, [=] MFEM_HOST_DEVICE(int e)
+   {
+      constexpr int MQ1 = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+
+      MFEM_SHARED real_t smem[MQ1][MQ1];
+      MFEM_SHARED real_t sB[MQ1][MQ1], sG[MQ1][MQ1];
+
+      kernels::internal::vd_regs3d_t<3, 3, MQ1> g0, g1;
+      kernels::internal::v_regs3d_t<1, MQ1> r0, r1;
+
+      kernels::internal::LoadMatrix(TR_D1D, Q1D, B, sB);
+      kernels::internal::LoadMatrix(TR_D1D, Q1D, G, sG);
+
+      kernels::internal::LoadDofs3d(e, TR_D1D, X, g0);
+      kernels::internal::Grad3d(TR_D1D, Q1D, smem, sB, sG, g0, g1);
+
+      for (int qz = 0; qz < Q1D; qz++)
+      {
+         MFEM_FOREACH_THREAD_DIRECT(qy, y, Q1D)
+         {
+            MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+            {
+               r0[0][qz][qy][qx] =
+                  // c = 0
+                  g1[0][0][qz][qy][qx] * Q(qx, qy, qz, 0, 0, e) +
+                  g1[0][1][qz][qy][qx] * Q(qx, qy, qz, 1, 0, e) +
+                  g1[0][2][qz][qy][qx] * Q(qx, qy, qz, 2, 0, e) +
+                  // c = 1
+                  g1[1][0][qz][qy][qx] * Q(qx, qy, qz, 0, 1, e) +
+                  g1[1][1][qz][qy][qx] * Q(qx, qy, qz, 1, 1, e) +
+                  g1[1][2][qz][qy][qx] * Q(qx, qy, qz, 2, 1, e) +
+                  // c = 2
+                  g1[2][0][qz][qy][qx] * Q(qx, qy, qz, 0, 2, e) +
+                  g1[2][1][qz][qy][qx] * Q(qx, qy, qz, 1, 2, e) +
+                  g1[2][2][qz][qy][qx] * Q(qx, qy, qz, 2, 2, e);
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      kernels::internal::LoadMatrix<MQ1, true>(TE_D1D, Q1D, Bt, sB);
+      kernels::internal::EvalTranspose3d(TE_D1D, Q1D, smem, sB, r0, r1);
+      kernels::internal::WriteDofs3d(e, TE_D1D, r1, Y);
+   });
+}
+
+} // namespace internal
+
+template<int DIM, int T_TR_D1D, int T_TE_D1D, int T_Q1D>
+VectorDivergenceIntegrator::VectorDivergenceAddMultPAType
+VectorDivergenceIntegrator::VectorDivergenceAddMultPA::Kernel()
+{
+   static_assert(T_TR_D1D <= T_Q1D && T_TE_D1D <= T_Q1D);
+   if constexpr (DIM == 2)
+   {
+      return internal::SmemPADivergenceApply2D<T_TR_D1D, T_TE_D1D, T_Q1D>;
+   }
+   else if constexpr (DIM == 3)
+   {
+      return internal::SmemPADivergenceApply3D<T_TR_D1D, T_TE_D1D, T_Q1D>;
+   }
+   else { MFEM_ABORT("Unsupported kernel"); }
+}
+
+inline VectorDivergenceIntegrator::VectorDivergenceAddMultPAType
+VectorDivergenceIntegrator::VectorDivergenceAddMultPA::Fallback
+(int dim, int tr_d1d, int te_d1d, int q1d)
+{
+   MFEM_VERIFY(tr_d1d <= q1d && te_d1d <= q1d, "");
+   MFEM_VERIFY(tr_d1d <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(te_d1d <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(q1d <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+   if (dim == 2)
+   {
+      return internal::SmemPADivergenceApply2D;
+   }
+   else if (dim == 3)
+   {
+      return internal::SmemPADivergenceApply3D;
+   }
+   else { MFEM_ABORT("Unsupported kernel"); }
+}
+
+template<int DIM, int T_TR_D1D, int T_TE_D1D, int T_Q1D>
+VectorDivergenceIntegrator::VectorDivergenceAddMultTransposePAType
+VectorDivergenceIntegrator::VectorDivergenceAddMultTransposePA::Kernel()
+{
+   static_assert(T_TR_D1D <= T_Q1D && T_TE_D1D <= T_Q1D);
+   if constexpr (DIM == 2)
+   {
+      return internal::SmemPADivergenceApplyTranspose2D<T_TR_D1D, T_TE_D1D, T_Q1D>;
+   }
+   else if constexpr (DIM == 3)
+   {
+      return internal::SmemPADivergenceApplyTranspose3D<T_TR_D1D, T_TE_D1D, T_Q1D>;
+   }
+   else { MFEM_ABORT("Unsupported kernel"); }
+}
+
+inline VectorDivergenceIntegrator::VectorDivergenceAddMultTransposePAType
+VectorDivergenceIntegrator::VectorDivergenceAddMultTransposePA::Fallback
+(int dim, int tr_d1d, int te_d1d, int q1d)
+{
+   MFEM_VERIFY(tr_d1d <= q1d && te_d1d <= q1d, "");
+   MFEM_VERIFY(tr_d1d <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(te_d1d <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(q1d <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+   if (dim == 2)
+   {
+      return internal::SmemPADivergenceApplyTranspose2D;
+   }
+   else if (dim == 3)
+   {
+      return internal::SmemPADivergenceApplyTranspose3D;
+   }
+   else { MFEM_ABORT("Unsupported kernel"); }
+}
+
+/// \endcond DO_NOT_DOCUMENT
+
+} // namespace mfem

fem/integ/bilininteg_vecmass_pa.cpp

@@ -205,157 +205,40 @@ void VectorMassIntegrator::AddMultPA(const Vector &x, Vector &y) const
 
 }
 
-template <const int T_D1D = 0, const int T_Q1D = 0>
-static void PAVectorMassAssembleDiagonal2D(const int NE,
-                                           const Array<real_t> &b,
-                                           const Vector &pa_data, Vector &diag,
-                                           const int d1d = 0, const int q1d = 0)
-{
-   constexpr int VDIM = 2;
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-   const auto B = Reshape(b.Read(), Q1D, D1D);
-   const auto D = Reshape(pa_data.Read(), Q1D, Q1D, NE);
-   auto Y = Reshape(diag.ReadWrite(), D1D, D1D, VDIM, NE);
-
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE(int e)
-   {
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-
-      real_t temp[max_Q1D][max_D1D];
-      for (int qx = 0; qx < Q1D; ++qx)
-      {
-         for (int dy = 0; dy < D1D; ++dy)
-         {
-            temp[qx][dy] = 0.0;
-            for (int qy = 0; qy < Q1D; ++qy)
-            {
-               temp[qx][dy] += B(qy, dy) * B(qy, dy) * D(qx, qy, e);
-            }
-         }
-      }
-      for (int dy = 0; dy < D1D; ++dy)
-      {
-         for (int dx = 0; dx < D1D; ++dx)
-         {
-            real_t temp1 = 0.0;
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               temp1 += B(qx, dx) * B(qx, dx) * temp[qx][dy];
-            }
-            Y(dx, dy, 0, e) = temp1;
-            Y(dx, dy, 1, e) = temp1;
-         }
-      }
-   });
-}
-
-template <const int T_D1D = 0, const int T_Q1D = 0>
-static void PAVectorMassAssembleDiagonal3D(const int NE,
-                                           const Array<real_t> &B_,
-                                           const Vector &pa_data, Vector &diag,
-                                           const int d1d = 0, const int q1d = 0)
-{
-   constexpr int VDIM = 3;
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-   const auto B = Reshape(B_.Read(), Q1D, D1D);
-   MFEM_VERIFY(pa_data.Size() == Q1D * Q1D * Q1D * NE, "pa_data size error");
-   const auto D = Reshape(pa_data.Read(), Q1D, Q1D, Q1D, NE);
-   auto Y = Reshape(diag.ReadWrite(), D1D, D1D, D1D, VDIM, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE(int e)
-   {
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      // the following variables are evaluated at compile time
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-
-      real_t temp[max_Q1D][max_Q1D][max_D1D];
-      for (int qx = 0; qx < Q1D; ++qx)
-      {
-         for (int qy = 0; qy < Q1D; ++qy)
-         {
-            for (int dz = 0; dz < D1D; ++dz)
-            {
-               temp[qx][qy][dz] = 0.0;
-               for (int qz = 0; qz < Q1D; ++qz)
-               {
-                  temp[qx][qy][dz] +=
-                     B(qz, dz) * B(qz, dz) * D(qx, qy, qz, e);
-               }
-            }
-         }
-      }
-      real_t temp2[max_Q1D][max_D1D][max_D1D];
-      for (int qx = 0; qx < Q1D; ++qx)
-      {
-         for (int dz = 0; dz < D1D; ++dz)
-         {
-            for (int dy = 0; dy < D1D; ++dy)
-            {
-               temp2[qx][dy][dz] = 0.0;
-               for (int qy = 0; qy < Q1D; ++qy)
-               {
-                  temp2[qx][dy][dz] +=
-                     B(qy, dy) * B(qy, dy) * temp[qx][qy][dz];
-               }
-            }
-         }
-      }
-      for (int dz = 0; dz < D1D; ++dz)
-      {
-         for (int dy = 0; dy < D1D; ++dy)
-         {
-            for (int dx = 0; dx < D1D; ++dx)
-            {
-               real_t temp3 = 0.0;
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  temp3 += B(qx, dx) * B(qx, dx) * temp2[qx][dy][dz];
-               }
-               Y(dx, dy, dz, 0, e) = temp3;
-               Y(dx, dy, dz, 1, e) = temp3;
-               Y(dx, dy, dz, 2, e) = temp3;
-            }
-         }
-      }
-   });
-}
-
-static void PAVectorMassAssembleDiagonal(const int dim, const int D1D,
-                                         const int Q1D, const int NE,
-                                         const Array<real_t> &B,
-                                         const Vector &pa_data,
-                                         Vector &diag)
-{
-   if (dim == 2)
-   {
-      return PAVectorMassAssembleDiagonal2D(NE, B, pa_data, diag, D1D, Q1D);
-   }
-   else if (dim == 3)
-   {
-      return PAVectorMassAssembleDiagonal3D(NE, B, pa_data, diag, D1D, Q1D);
-   }
-   MFEM_ABORT("Dimension not implemented.");
-}
-
 void VectorMassIntegrator::AssembleDiagonalPA(Vector &diag)
 {
-   if (DeviceCanUseCeed()) { ceedOp->GetDiagonal(diag); }
-   else
-   {
-      MFEM_VERIFY(coeff_vdim == 1, "coeff_vdim != 1");
-      MFEM_VERIFY(!VQ && !MQ, "VQ and MQ not supported");
-      PAVectorMassAssembleDiagonal(dim, dofs1D, quad1D, ne, maps->B, pa_data, diag);
-   }
+   if (DeviceCanUseCeed()) { return ceedOp->GetDiagonal(diag); }
+
+   MFEM_VERIFY(coeff_vdim == 1, "coeff_vdim != 1");
+   MFEM_VERIFY(!VQ && !MQ, "VQ and MQ not supported");
+
+   // Add the VectorMassAssembleDiagonalPA specializations
+   static const auto vector_mass_assemble_diagonal_kernel_specializations =
+      ( // 2D
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<2, 2>::Add(),
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<2, 3>::Add(),
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<2, 4>::Add(),
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<2, 5>::Add(),
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<2, 6>::Add(),
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<2, 7>::Add(),
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<2, 8>::Add(),
+         // 3D
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<3, 2>::Add(),
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<3, 3>::Add(),
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<3, 4>::Add(),
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<3, 5>::Add(),
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<3, 6>::Add(),
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<3, 7>::Add(),
+         VectorMassIntegrator::VectorMassAssembleDiagonalPA::Specialization<3, 8>::Add(),
+         true);
+   MFEM_CONTRACT_VAR(vector_mass_assemble_diagonal_kernel_specializations);
+
+   VectorMassAssembleDiagonalPA::Run(dim, quad1D, // templated arguments
+                                     ne, dofs1D, quad1D,
+                                     maps->B.Read(),
+                                     pa_data.Read(),
+                                     diag.ReadWrite());
+
 }
 
 } // namespace mfem

fem/integ/bilininteg_vecmass_pa.hpp

@@ -176,8 +176,146 @@ void SmemPAVectorMassApply3D(const int NE,
    });
 }
 
+template <int T_Q1D = 0, int T_MDQ = 16>
+static void SmemPAVectorMassAssembleDiagonal2D(const int ne,
+                                               const int d1d,
+                                               const int q1d,
+                                               const real_t *b_r,
+                                               const real_t *d_r,
+                                               real_t *y_rw)
+{
+   constexpr int VDIM = 2;
+
+   const int D1D = d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+
+   MFEM_VERIFY(Q1D <= T_MDQ && D1D <= Q1D, "");
+
+   const auto B = Reshape(b_r, Q1D, D1D);
+   const auto D = Reshape(d_r, Q1D, Q1D, ne);
+   auto Y = Reshape(y_rw, D1D, D1D, VDIM, ne);
+
+   mfem::forall_2D<T_Q1D*T_Q1D>(
+      ne, Q1D, Q1D, [=] MFEM_HOST_DEVICE(int e)
+   {
+      constexpr int MQ1 = T_Q1D ? T_Q1D : T_MDQ;
+
+      MFEM_SHARED real_t sm[MQ1][MQ1];
+
+      MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+      {
+         MFEM_FOREACH_THREAD_DIRECT(dy, y, D1D)
+         {
+            real_t u = 0.0;
+            for (int qy = 0; qy < Q1D; ++qy)
+            {
+               u += B(qy, dy) * B(qy, dy) * D(qx, qy, e);
+            }
+            sm[qx][dy] = u;
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      MFEM_FOREACH_THREAD_DIRECT(dy, y, D1D)
+      {
+         MFEM_FOREACH_THREAD_DIRECT(dx, x, D1D)
+         {
+            real_t u = 0.0;
+            for (int qx = 0; qx < Q1D; ++qx)
+            {
+               u += B(qx, dx) * B(qx, dx) * sm[qx][dy];
+            }
+            Y(dx, dy, 0, e) += u;
+            Y(dx, dy, 1, e) += u;
+         }
+      }
+   });
+}
+
+// T_MDQ <= 10 so the Q1D^3 thread block stays within the 1024/block GPU limit
+template <int T_Q1D = 0, int T_MDQ = 10>
+static void SmemPAVectorMassAssembleDiagonal3D(const int ne,
+                                               const int d1d,
+                                               const int q1d,
+                                               const real_t *b_r,
+                                               const real_t *d_r,
+                                               real_t *y_rw)
+{
+   constexpr int VDIM = 3;
+
+   const int D1D = d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+
+   MFEM_VERIFY(Q1D <= T_MDQ && D1D <= Q1D, "");
+
+   const auto B = Reshape(b_r, Q1D, D1D);
+   const auto D = Reshape(d_r, Q1D, Q1D, Q1D, ne);
+   auto Y = Reshape(y_rw, D1D, D1D, D1D, VDIM, ne);
+
+   mfem::forall_3D<T_Q1D*T_Q1D*T_Q1D>(
+      ne, Q1D, Q1D, Q1D, [=] MFEM_HOST_DEVICE(int e)
+   {
+      constexpr int MQ1 = T_Q1D ? T_Q1D : T_MDQ;
+
+      MFEM_SHARED real_t sm[2][MQ1][MQ1][MQ1];
+
+      MFEM_FOREACH_THREAD_DIRECT(dz, z, D1D)
+      {
+         MFEM_FOREACH_THREAD_DIRECT(qy, y, Q1D)
+         {
+            MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+            {
+               real_t u = 0.0;
+               for (int qz = 0; qz < Q1D; ++qz)
+               {
+                  u += B(qz, dz) * B(qz, dz) * D(qx, qy, qz, e);
+               }
+               sm[0][dz][qy][qx] = u;
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      MFEM_FOREACH_THREAD_DIRECT(dz, z, D1D)
+      {
+         MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+         {
+            MFEM_FOREACH_THREAD_DIRECT(dy, y, D1D)
+            {
+               real_t u = 0.0;
+               for (int qy = 0; qy < Q1D; ++qy)
+               {
+                  u += B(qy, dy) * B(qy, dy) * sm[0][dz][qy][qx];
+               }
+               sm[1][dz][dy][qx] = u;
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+
+      MFEM_FOREACH_THREAD_DIRECT(dz, z, D1D)
+      {
+         MFEM_FOREACH_THREAD_DIRECT(dy, y, D1D)
+         {
+            MFEM_FOREACH_THREAD_DIRECT(dx, x, D1D)
+            {
+               real_t u = 0.0;
+               for (int qx = 0; qx < Q1D; ++qx)
+               {
+                  u += B(qx, dx) * B(qx, dx) * sm[1][dz][dy][qx];
+               }
+               Y(dx, dy, dz, 0, e) += u;
+               Y(dx, dy, dz, 1, e) += u;
+               Y(dx, dy, dz, 2, e) += u;
+            }
+         }
+      }
+   });
+}
+
 } // namespace internal
 
+// AddMultPA kernels
 template<int DIM, int T_D1D, int T_Q1D>
 VectorMassIntegrator::VectorMassAddMultPAType
 VectorMassIntegrator::VectorMassAddMultPA::Kernel()
@@ -190,11 +328,11 @@ VectorMassIntegrator::VectorMassAddMultPA::Kernel()
    {
       return internal::SmemPAVectorMassApply3D<T_D1D, T_Q1D>;
    }
-   MFEM_ABORT("Unsupported kernel");
+   else { MFEM_ABORT("Unsupported kernel"); }
 }
 
 inline VectorMassIntegrator::VectorMassAddMultPAType
-VectorMassIntegrator::VectorMassAddMultPA::Fallback(int dim, int d1d, int q1d)
+VectorMassIntegrator::VectorMassAddMultPA::Fallback(int dim, int, int)
 {
    if (dim == 2)
    {
@@ -207,6 +345,36 @@ VectorMassIntegrator::VectorMassAddMultPA::Fallback(int dim, int d1d, int q1d)
    else { MFEM_ABORT("Unsupported kernel"); }
 }
 
+// DiagonalPA kernels
+template<int DIM, int T_Q1D>
+VectorMassIntegrator::VectorMassAssembleDiagonalPAType
+VectorMassIntegrator::VectorMassAssembleDiagonalPA::Kernel()
+{
+   if constexpr (DIM == 2)
+   {
+      return internal::SmemPAVectorMassAssembleDiagonal2D<T_Q1D>;
+   }
+   else if constexpr (DIM == 3)
+   {
+      return internal::SmemPAVectorMassAssembleDiagonal3D<T_Q1D>;
+   }
+   else { MFEM_ABORT("Unsupported kernel"); }
+}
+
+inline VectorMassIntegrator::VectorMassAssembleDiagonalPAType
+VectorMassIntegrator::VectorMassAssembleDiagonalPA::Fallback(int dim, int)
+{
+   if (dim == 2)
+   {
+      return internal::SmemPAVectorMassAssembleDiagonal2D;
+   }
+   else if (dim == 3)
+   {
+      return internal::SmemPAVectorMassAssembleDiagonal3D;
+   }
+   else { MFEM_ABORT("Unsupported kernel"); }
+}
+
 /// \endcond DO_NOT_DOCUMENT
 
 } // namespace mfem

fem/integ/lininteg_domain_kernels.hpp

@@ -307,6 +307,506 @@ DomainLFIntegrator::AssembleKernels::Kernel()
    MFEM_ABORT("");
 }
 
+template <int T_D1D = 0, int T_Q1D = 0>
+static void HdivDLFAssemble2D(const int ne, const Array<int> &markers,
+                              const Vector &jac, const Array<real_t> &weights,
+                              const Array<real_t> &testBO,
+                              const Array<real_t> &testBC, const Vector &coeff,
+                              Vector &y, const int d, const int q)
+{
+   MFEM_VERIFY(T_D1D || d <= DeviceDofQuadLimits::Get().HDIV_MAX_D1D,
+               "Problem size too large.");
+   MFEM_VERIFY(T_Q1D || q <= DeviceDofQuadLimits::Get().HDIV_MAX_Q1D,
+               "Problem size too large.");
+   MFEM_VERIFY(y.Size() == 2 * (d - 1) * d * ne, "");
+
+   constexpr int vdim = 2;
+   const auto F = coeff.Read();
+   const auto M = markers.Read();
+   const auto BO = Reshape(testBO.Read(), q, d-1);
+   const auto BC = Reshape(testBC.Read(), q, d);
+   const auto J = Reshape(jac.Read(), q, q, vdim, vdim, ne);
+   const auto W = Reshape(weights.Read(), q, q);
+   const bool cst = coeff.Size() == vdim;
+   const auto C = cst ? Reshape(F,vdim,1,1,1) : Reshape(F,vdim,q,q,ne);
+   auto Y = y.ReadWrite();
+
+   mfem::forall_3D(ne, q, q, vdim, [=] MFEM_HOST_DEVICE (int e)
+   {
+      constexpr int vdim = 2;
+      if (M[e] == 0) { return; } // ignore
+
+      constexpr int Q = T_Q1D ? T_Q1D : DofQuadLimits::HDIV_MAX_Q1D;
+      constexpr int D = T_D1D ? T_D1D : DofQuadLimits::HDIV_MAX_D1D;
+
+      MFEM_SHARED real_t sBot[Q*D];
+      MFEM_SHARED real_t sBct[Q*D];
+      MFEM_SHARED real_t sQQ[vdim*Q*Q];
+      MFEM_SHARED real_t sQD[vdim*Q*D];
+
+      // Bo and Bc into shared memory
+      const DeviceMatrix Bot(sBot, d-1, q);
+      kernels::internal::LoadB<D,Q>(d-1, q, BO, sBot);
+      const DeviceMatrix Bct(sBct, d, q);
+      kernels::internal::LoadB<D,Q>(d, q, BC, sBct);
+
+      const DeviceCube QQ(sQQ, q, q, vdim);
+      const DeviceCube QD(sQD, q, d, vdim);
+
+      MFEM_FOREACH_THREAD(vd,z,vdim)
+      {
+         const real_t cst_val_0 = C(0,0,0,0);
+         const real_t cst_val_1 = C(1,0,0,0);
+         MFEM_FOREACH_THREAD(y,y,q)
+         {
+            MFEM_FOREACH_THREAD(x,x,q)
+            {
+               const real_t J0 = J(x,y,0,vd,e);
+               const real_t J1 = J(x,y,1,vd,e);
+               const real_t C0 = cst ? cst_val_0 : C(0,x,y,e);
+               const real_t C1 = cst ? cst_val_1 : C(1,x,y,e);
+               QQ(x,y,vd) = W(x,y)*(J0*C0 + J1*C1);
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(vd,z,vdim)
+      {
+         const int nx = (vd == 0) ? d : d-1;
+         DeviceMatrix Btx = (vd == 0) ? Bct : Bot;
+         MFEM_FOREACH_THREAD(qy,y,q)
+         {
+            MFEM_FOREACH_THREAD(dx,x,nx)
+            {
+               real_t qd = 0.0;
+               for (int qx = 0; qx < q; ++qx)
+               {
+                  qd += QQ(qx,qy,vd) * Btx(dx,qx);
+               }
+               QD(dx,qy,vd) = qd;
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(vd,z,vdim)
+      {
+         const int nx = (vd == 0) ? d : d-1;
+         const int ny = (vd == 1) ? d : d-1;
+         DeviceMatrix Bty = (vd == 1) ? Bct : Bot;
+         DeviceTensor<4> Yxy(Y, nx, ny, vdim, ne);
+         MFEM_FOREACH_THREAD(dy,y,ny)
+         {
+            MFEM_FOREACH_THREAD(dx,x,nx)
+            {
+               real_t dd = 0.0;
+               for (int qy = 0; qy < q; ++qy)
+               {
+                  dd += QD(dx,qy,vd) * Bty(dy,qy);
+               }
+               Yxy(dx,dy,vd,e) += dd;
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+   });
+}
+
+template <int T_D1D = 0, int T_Q1D = 0>
+static void HdivDLFAssemble3D(const int ne, const Array<int> &markers,
+                              const Vector &jac, const Array<real_t> &weights,
+                              const Array<real_t> &testBO,
+                              const Array<real_t> &testBC, const Vector &coeff,
+                              Vector &y, const int d, const int q)
+{
+   MFEM_VERIFY(T_D1D || d <= DeviceDofQuadLimits::Get().HDIV_MAX_D1D,
+               "Problem size too large.");
+   MFEM_VERIFY(T_Q1D || q <= DeviceDofQuadLimits::Get().HDIV_MAX_Q1D,
+               "Problem size too large.");
+   MFEM_VERIFY(y.Size() == 3 * (d - 1) * (d - 1) * d * ne, "y wrong length");
+
+   constexpr int vdim = 3;
+   const auto F = coeff.Read();
+   const auto M = markers.Read();
+   const auto BO = Reshape(testBO.Read(), q, d-1);
+   const auto BC = Reshape(testBC.Read(), q, d);
+   const auto J = Reshape(jac.Read(), q, q, q, vdim, vdim, ne);
+   const auto W = Reshape(weights.Read(), q, q, q);
+   const bool cst = coeff.Size() == vdim;
+   const auto C = cst ? Reshape(F,vdim,1,1,1,1) : Reshape(F,vdim,q,q,q,ne);
+   auto Y = y.ReadWrite();
+
+   mfem::forall_3D(ne, q, q, vdim, [=] MFEM_HOST_DEVICE (int e)
+   {
+      constexpr int vdim = 3;
+      if (M[e] == 0) { return; } // ignore
+
+      constexpr int Q = T_Q1D ? T_Q1D : DofQuadLimits::HDIV_MAX_Q1D;
+      constexpr int D = T_D1D ? T_D1D : DofQuadLimits::HDIV_MAX_D1D;
+
+      MFEM_SHARED real_t sBot[Q*D];
+      MFEM_SHARED real_t sBct[Q*D];
+
+      // Bo and Bc into shared memory
+      const DeviceMatrix Bot(sBot, d-1, q);
+      kernels::internal::LoadB<D,Q>(d-1, q, BO, sBot);
+      const DeviceMatrix Bct(sBct, d, q);
+      kernels::internal::LoadB<D,Q>(d, q, BC, sBct);
+
+      MFEM_SHARED real_t sm0[vdim*Q*Q*Q];
+      MFEM_SHARED real_t sm1[vdim*Q*Q*Q];
+      DeviceTensor<4> QQQ(sm1, q, q, q, vdim);
+      DeviceTensor<4> DQQ(sm0, d, q, q, vdim);
+      DeviceTensor<4> DDQ(sm1, d, d, q, vdim);
+
+      MFEM_FOREACH_THREAD(vd,z,vdim)
+      {
+         const real_t cst_val_0 = C(0,0,0,0,0);
+         const real_t cst_val_1 = C(1,0,0,0,0);
+         const real_t cst_val_2 = C(2,0,0,0,0);
+         MFEM_FOREACH_THREAD(y,y,q)
+         {
+            MFEM_FOREACH_THREAD(x,x,q)
+            {
+               for (int z = 0; z < q; ++z)
+               {
+                  const real_t J0 = J(x,y,z,0,vd,e);
+                  const real_t J1 = J(x,y,z,1,vd,e);
+                  const real_t J2 = J(x,y,z,2,vd,e);
+                  const real_t C0 = cst ? cst_val_0 : C(0,x,y,z,e);
+                  const real_t C1 = cst ? cst_val_1 : C(1,x,y,z,e);
+                  const real_t C2 = cst ? cst_val_2 : C(2,x,y,z,e);
+                  QQQ(x,y,z,vd) = W(x,y,z)*(J0*C0 + J1*C1 + J2*C2);
+               }
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      // Apply Bt operator
+      MFEM_FOREACH_THREAD(vd,z,vdim)
+      {
+         const int nx = (vd == 0) ? d : d-1;
+         DeviceMatrix Btx = (vd == 0) ? Bct : Bot;
+         MFEM_FOREACH_THREAD(qy,y,q)
+         {
+            MFEM_FOREACH_THREAD(dx,x,nx)
+            {
+               real_t u[Q];
+               MFEM_UNROLL(Q)
+               for (int qz = 0; qz < q; ++qz) { u[qz] = 0.0; }
+               MFEM_UNROLL(Q)
+               for (int qx = 0; qx < q; ++qx)
+               {
+                  MFEM_UNROLL(Q)
+                  for (int qz = 0; qz < q; ++qz)
+                  {
+                     u[qz] += QQQ(qx,qy,qz,vd) * Btx(dx,qx);
+                  }
+               }
+               MFEM_UNROLL(Q)
+               for (int qz = 0; qz < q; ++qz) { DQQ(dx,qy,qz,vd) = u[qz]; }
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(vd,z,vdim)
+      {
+         const int nx = (vd == 0) ? d : d-1;
+         const int ny = (vd == 1) ? d : d-1;
+         DeviceMatrix Bty = (vd == 1) ? Bct : Bot;
+         MFEM_FOREACH_THREAD(dy,y,ny)
+         {
+            MFEM_FOREACH_THREAD(dx,x,nx)
+            {
+               real_t u[Q];
+               MFEM_UNROLL(Q)
+               for (int qz = 0; qz < q; ++qz) { u[qz] = 0.0; }
+               MFEM_UNROLL(Q)
+               for (int qy = 0; qy < q; ++qy)
+               {
+                  MFEM_UNROLL(Q)
+                  for (int qz = 0; qz < q; ++qz)
+                  {
+                     u[qz] += DQQ(dx,qy,qz,vd) * Bty(dy,qy);
+                  }
+               }
+               MFEM_UNROLL(Q)
+               for (int qz = 0; qz < q; ++qz) { DDQ(dx,dy,qz,vd) = u[qz]; }
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(vd,z,vdim)
+      {
+         const int nx = (vd == 0) ? d : d-1;
+         const int ny = (vd == 1) ? d : d-1;
+         const int nz = (vd == 2) ? d : d-1;
+         DeviceTensor<5> Yxyz(Y, nx, ny, nz, vdim, ne);
+         DeviceMatrix Btz = (vd == 2) ? Bct : Bot;
+         MFEM_FOREACH_THREAD(dy,y,ny)
+         {
+            MFEM_FOREACH_THREAD(dx,x,nx)
+            {
+               real_t u[D];
+               MFEM_UNROLL(D)
+               for (int dz = 0; dz < nz; ++dz) { u[dz] = 0.0; }
+               MFEM_UNROLL(Q)
+               for (int qz = 0; qz < q; ++qz)
+               {
+                  MFEM_UNROLL(D)
+                  for (int dz = 0; dz < nz; ++dz)
+                  {
+                     u[dz] += DDQ(dx,dy,qz,vd) * Btz(dz,qz);
+                  }
+               }
+               MFEM_UNROLL(D)
+               for (int dz = 0; dz < nz; ++dz) { Yxyz(dx,dy,dz,vd,e) += u[dz]; }
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+   });
+}
+
+/// @param ne number of elements
+/// @param markers array where entry markers[e] == 0 to skip assembly over
+/// element e element
+/// @param jac Spatial Jacobians evaluated at all quadrature points
+/// @param weights 1D quadrature weights
+/// @param testBO 1D open basis test functions
+/// @param testBC 1D closed basis test functions
+/// @param coeff coefficient values evaluated at quadrature points, possibly
+/// compressed.
+/// @param d number of 1D closed dofs
+/// @param q number of 1D quadrature points
+/// @tparam T_D1D maximum number of dofs along any direction, or 0
+/// @tparam T_Q1D maximum number of quadrature points along any direction, or 0
+template <int T_D1D = 0, int T_Q1D = 0>
+static void HcurlDLFAssemble3D(const int ne, const Array<int> &markers,
+                               const Vector &jac, const Array<real_t> &weights,
+                               const Array<real_t> &testBO,
+                               const Array<real_t> &testBC, const Vector &coeff,
+                               Vector &y, const int d, const int q)
+{
+   MFEM_VERIFY(T_D1D || d <= DeviceDofQuadLimits::Get().HCURL_MAX_D1D,
+               "Problem size too large.");
+   MFEM_VERIFY(T_Q1D || q <= DeviceDofQuadLimits::Get().HCURL_MAX_Q1D,
+               "Problem size too large.");
+   MFEM_VERIFY(y.Size() == 3 * (d - 1) * d * d * ne, "y wrong length");
+
+   constexpr int vdim = 3;
+   const auto F = coeff.Read();
+   const auto M = markers.Read();
+   const auto BO = Reshape(testBO.Read(), q, d-1);
+   const auto BC = Reshape(testBC.Read(), q, d);
+   const auto J = Reshape(jac.Read(), q, q, q, vdim, vdim, ne);
+   const auto W = Reshape(weights.Read(), q, q, q);
+   const bool cst = coeff.Size() == vdim;
+   const auto C = cst ? Reshape(F,vdim,1,1,1,1) : Reshape(F,vdim,q,q,q,ne);
+   auto Y = y.ReadWrite();
+
+   mfem::forall_3D(ne, q, q, vdim, [=] MFEM_HOST_DEVICE(int e)
+   {
+      if (M[e] == 0)
+      {
+         // ignore
+         return;
+      }
+
+      constexpr int vdim = 3;
+      constexpr int Q = T_Q1D ? T_Q1D : DofQuadLimits::HCURL_MAX_Q1D;
+      constexpr int D = T_D1D ? T_D1D : DofQuadLimits::HCURL_MAX_D1D;
+
+      MFEM_SHARED real_t sBot[Q * D];
+      MFEM_SHARED real_t sBct[Q * D];
+
+      // Bo and Bc into shared memory
+      const DeviceMatrix Bot(sBot, d - 1, q);
+      kernels::internal::LoadB<D, Q>(d - 1, q, BO, sBot);
+      const DeviceMatrix Bct(sBct, d, q);
+      kernels::internal::LoadB<D, Q>(d, q, BC, sBct);
+
+      MFEM_SHARED real_t sm0[vdim * Q * Q * Q];
+      MFEM_SHARED real_t sm1[vdim * Q * Q * Q];
+      DeviceTensor<4> QQQ(sm1, q, q, q, vdim);
+      DeviceTensor<4> DQQ(sm0, d, q, q, vdim);
+      DeviceTensor<4> DDQ(sm1, d, d, q, vdim);
+
+      const real_t cst_val_0 = C(0, 0, 0, 0, 0);
+      const real_t cst_val_1 = C(1, 0, 0, 0, 0);
+      const real_t cst_val_2 = C(2, 0, 0, 0, 0);
+
+      MFEM_FOREACH_THREAD(vd, z, vdim)
+      {
+         MFEM_FOREACH_THREAD(y, y, q)
+         {
+            MFEM_FOREACH_THREAD(x, x, q)
+            {
+               for (int z = 0; z < q; ++z)
+               {
+                  real_t curr[3];
+                  curr[0] = cst ? cst_val_0 : C(0, x, y, z, e);
+                  curr[1] = cst ? cst_val_1 : C(1, x, y, z, e);
+                  curr[2] = cst ? cst_val_2 : C(2, x, y, z, e);
+
+                  const real_t J11 = J(x, y, z, 0, 0, e);
+                  const real_t J21 = J(x, y, z, 1, 0, e);
+                  const real_t J31 = J(x, y, z, 2, 0, e);
+                  const real_t J12 = J(x, y, z, 0, 1, e);
+                  const real_t J22 = J(x, y, z, 1, 1, e);
+                  const real_t J32 = J(x, y, z, 2, 1, e);
+                  const real_t J13 = J(x, y, z, 0, 2, e);
+                  const real_t J23 = J(x, y, z, 1, 2, e);
+                  const real_t J33 = J(x, y, z, 2, 2, e);
+                  // adj(J)
+                  const real_t A11 = (J22 * J33) - (J23 * J32);
+                  const real_t A12 = (J32 * J13) - (J12 * J33);
+                  const real_t A13 = (J12 * J23) - (J22 * J13);
+                  const real_t A21 = (J31 * J23) - (J21 * J33);
+                  const real_t A22 = (J11 * J33) - (J13 * J31);
+                  const real_t A23 = (J21 * J13) - (J11 * J23);
+                  const real_t A31 = (J21 * J32) - (J31 * J22);
+                  const real_t A32 = (J31 * J12) - (J11 * J32);
+                  const real_t A33 = (J11 * J22) - (J12 * J21);
+                  const real_t A[9] = {A11, A12, A13, A21, A22,
+                                       A23, A31, A32, A33
+                                      };
+                  QQQ(x, y, z, vd) = W(x, y, z) * (A[vd * vdim] * curr[0] +
+                                                   A[vd * vdim + 1] * curr[1] +
+                                                   A[vd * vdim + 2] * curr[2]);
+               }
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      // Apply Bt operator
+      MFEM_FOREACH_THREAD(vd, z, vdim)
+      {
+         const int nx = (vd == 0) ? d - 1 : d;
+         DeviceMatrix Btx = (vd == 0) ? Bot : Bct;
+         MFEM_FOREACH_THREAD(qy, y, q)
+         {
+            MFEM_FOREACH_THREAD(dx, x, nx)
+            {
+               real_t u[Q];
+               MFEM_UNROLL(Q)
+               for (int qz = 0; qz < q; ++qz)
+               {
+                  u[qz] = 0.0;
+               }
+               MFEM_UNROLL(Q)
+               for (int qx = 0; qx < q; ++qx)
+               {
+                  MFEM_UNROLL(Q)
+                  for (int qz = 0; qz < q; ++qz)
+                  {
+                     u[qz] += QQQ(qx, qy, qz, vd) * Btx(dx, qx);
+                  }
+               }
+               MFEM_UNROLL(Q)
+               for (int qz = 0; qz < q; ++qz)
+               {
+                  DQQ(dx, qy, qz, vd) = u[qz];
+               }
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(vd, z, vdim)
+      {
+         const int nx = (vd == 0) ? d - 1 : d;
+         const int ny = (vd == 1) ? d - 1 : d;
+         DeviceMatrix Bty = (vd == 1) ? Bot : Bct;
+         MFEM_FOREACH_THREAD(dy, y, ny)
+         {
+            MFEM_FOREACH_THREAD(dx, x, nx)
+            {
+               real_t u[Q];
+               MFEM_UNROLL(Q)
+               for (int qz = 0; qz < q; ++qz)
+               {
+                  u[qz] = 0.0;
+               }
+               MFEM_UNROLL(Q)
+               for (int qy = 0; qy < q; ++qy)
+               {
+                  MFEM_UNROLL(Q)
+                  for (int qz = 0; qz < q; ++qz)
+                  {
+                     u[qz] += DQQ(dx, qy, qz, vd) * Bty(dy, qy);
+                  }
+               }
+               MFEM_UNROLL(Q)
+               for (int qz = 0; qz < q; ++qz)
+               {
+                  DDQ(dx, dy, qz, vd) = u[qz];
+               }
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      MFEM_FOREACH_THREAD(vd, z, vdim)
+      {
+         const int nx = (vd == 0) ? d - 1 : d;
+         const int ny = (vd == 1) ? d - 1 : d;
+         const int nz = (vd == 2) ? d - 1 : d;
+         DeviceTensor<5> Yxyz(Y, nx, ny, nz, vdim, ne);
+         DeviceMatrix Btz = (vd == 2) ? Bot : Bct;
+         MFEM_FOREACH_THREAD(dy, y, ny)
+         {
+            MFEM_FOREACH_THREAD(dx, x, nx)
+            {
+               real_t u[D];
+               MFEM_UNROLL(D)
+               for (int dz = 0; dz < nz; ++dz)
+               {
+                  u[dz] = 0.0;
+               }
+               MFEM_UNROLL(Q)
+               for (int qz = 0; qz < q; ++qz)
+               {
+                  MFEM_UNROLL(D)
+                  for (int dz = 0; dz < nz; ++dz)
+                  {
+                     u[dz] += DDQ(dx, dy, qz, vd) * Btz(dz, qz);
+                  }
+               }
+               MFEM_UNROLL(D)
+               for (int dz = 0; dz < nz; ++dz)
+               {
+                  Yxyz(dx, dy, dz, vd, e) += u[dz];
+               }
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+   });
+}
+
+template <FiniteElement::DerivType TestType, int DIM, int TEST_D1D, int Q1D>
+VectorFEDomainLFIntegrator::AssembleKernelType
+VectorFEDomainLFIntegrator::AssembleKernels::Kernel()
+{
+   if constexpr (TestType == FiniteElement::DIV)
+   {
+      if constexpr (DIM == 2)
+      {
+         return HdivDLFAssemble2D<TEST_D1D, Q1D>;
+      }
+      if constexpr (DIM == 3)
+      {
+         return HdivDLFAssemble3D<TEST_D1D, Q1D>;
+      }
+   }
+   if constexpr (TestType == FiniteElement::CURL)
+   {
+      if constexpr (DIM == 3)
+      {
+         return HcurlDLFAssemble3D<TEST_D1D, Q1D>;
+      }
+   }
+   MFEM_ABORT("");
+}
+
 /// \endcond DO_NOT_DOCUMENT
 
 } // namespace mfem

fem/integ/lininteg_domain_vectorfe.cpp

@@ -13,317 +13,76 @@
 #include "../../fem/kernels.hpp"
 #include "../fem.hpp"
 
+#include "lininteg_domain_kernels.hpp"
+
 namespace mfem
 {
 
-template<int T_D1D = 0, int T_Q1D = 0>
-static void HdivDLFAssemble2D(
-   const int ne, const int d, const int q, const int *markers, const real_t *bo,
-   const real_t *bc, const real_t *j, const real_t *weights,
-   const Vector &coeff, real_t *y)
+VectorFEDomainLFIntegrator::Kernels::Kernels()
 {
-   MFEM_VERIFY(T_D1D || d <= DeviceDofQuadLimits::Get().HDIV_MAX_D1D,
-               "Problem size too large.");
-   MFEM_VERIFY(T_Q1D || q <= DeviceDofQuadLimits::Get().HDIV_MAX_Q1D,
-               "Problem size too large.");
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 2, 1, 1>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 2, 2, 2>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 2, 3, 3>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 2, 4, 4>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 2, 5, 5>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 2, 6, 6>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 2, 7, 7>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 2, 8, 8>();
 
-   static constexpr int vdim = 2;
-   const auto F = coeff.Read();
-   const auto M = Reshape(markers, ne);
-   const auto BO = Reshape(bo, q, d-1);
-   const auto BC = Reshape(bc, q, d);
-   const auto J = Reshape(j, q, q, vdim, vdim, ne);
-   const auto W = Reshape(weights, q, q);
-   const bool cst = coeff.Size() == vdim;
-   const auto C = cst ? Reshape(F,vdim,1,1,1) : Reshape(F,vdim,q,q,ne);
-   auto Y = Reshape(y, 2*(d-1)*d, ne);
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 3, 1, 1>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 3, 2, 2>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 3, 3, 3>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 3, 4, 4>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 3, 5, 5>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 3, 6, 6>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 3, 7, 7>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::DIV, 3, 8, 8>();
 
-   mfem::forall_3D(ne, q, q, vdim, [=] MFEM_HOST_DEVICE (int e)
-   {
-      if (M(e) == 0) { return; } // ignore
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 1, 1>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 2, 2>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 3, 3>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 4, 4>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 5, 5>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 6, 6>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 7, 7>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 8, 8>();
 
-      constexpr int Q = T_Q1D ? T_Q1D : DofQuadLimits::HDIV_MAX_Q1D;
-      constexpr int D = T_D1D ? T_D1D : DofQuadLimits::HDIV_MAX_D1D;
-
-      MFEM_SHARED real_t sBot[Q*D];
-      MFEM_SHARED real_t sBct[Q*D];
-      MFEM_SHARED real_t sQQ[vdim*Q*Q];
-      MFEM_SHARED real_t sQD[vdim*Q*D];
-
-      // Bo and Bc into shared memory
-      const DeviceMatrix Bot(sBot, d-1, q);
-      kernels::internal::LoadB<D,Q>(d-1, q, BO, sBot);
-      const DeviceMatrix Bct(sBct, d, q);
-      kernels::internal::LoadB<D,Q>(d, q, BC, sBct);
-
-      const DeviceCube QQ(sQQ, q, q, vdim);
-      const DeviceCube QD(sQD, q, d, vdim);
-
-      MFEM_FOREACH_THREAD(vd,z,vdim)
-      {
-         const real_t cst_val_0 = C(0,0,0,0);
-         const real_t cst_val_1 = C(1,0,0,0);
-         MFEM_FOREACH_THREAD(y,y,q)
-         {
-            MFEM_FOREACH_THREAD(x,x,q)
-            {
-               const real_t J0 = J(x,y,0,vd,e);
-               const real_t J1 = J(x,y,1,vd,e);
-               const real_t C0 = cst ? cst_val_0 : C(0,x,y,e);
-               const real_t C1 = cst ? cst_val_1 : C(1,x,y,e);
-               QQ(x,y,vd) = W(x,y)*(J0*C0 + J1*C1);
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(vd,z,vdim)
-      {
-         const int nx = (vd == 0) ? d : d-1;
-         DeviceMatrix Btx = (vd == 0) ? Bct : Bot;
-         MFEM_FOREACH_THREAD(qy,y,q)
-         {
-            MFEM_FOREACH_THREAD(dx,x,nx)
-            {
-               real_t qd = 0.0;
-               for (int qx = 0; qx < q; ++qx)
-               {
-                  qd += QQ(qx,qy,vd) * Btx(dx,qx);
-               }
-               QD(dx,qy,vd) = qd;
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(vd,z,vdim)
-      {
-         const int nx = (vd == 0) ? d : d-1;
-         const int ny = (vd == 1) ? d : d-1;
-         DeviceMatrix Bty = (vd == 1) ? Bct : Bot;
-         DeviceTensor<4> Yxy(Y, nx, ny, vdim, ne);
-         MFEM_FOREACH_THREAD(dy,y,ny)
-         {
-            MFEM_FOREACH_THREAD(dx,x,nx)
-            {
-               real_t dd = 0.0;
-               for (int qy = 0; qy < q; ++qy)
-               {
-                  dd += QD(dx,qy,vd) * Bty(dy,qy);
-               }
-               Yxy(dx,dy,vd,e) += dd;
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-   });
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 1, 2>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 2, 3>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 3, 4>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 4, 5>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 5, 6>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 6, 7>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 7, 8>();
+   VectorFEDomainLFIntegrator::AddSpecialization<FiniteElement::CURL, 3, 8, 9>();
 }
 
-template<int T_D1D = 0, int T_Q1D = 0>
-static void HdivDLFAssemble3D(
-   const int ne, const int d, const int q, const int *markers, const real_t *bo,
-   const real_t *bc, const real_t *j, const real_t *weights,
-   const Vector &coeff, real_t *y)
+/// \cond DO_NOT_DOCUMENT
+VectorFEDomainLFIntegrator::AssembleKernelType
+VectorFEDomainLFIntegrator::AssembleKernels::Fallback(
+   FiniteElement::DerivType TestType, int DIM, int, int)
 {
-   MFEM_VERIFY(T_D1D || d <= DeviceDofQuadLimits::Get().HDIV_MAX_D1D,
-               "Problem size too large.");
-   MFEM_VERIFY(T_Q1D || q <= DeviceDofQuadLimits::Get().HDIV_MAX_Q1D,
-               "Problem size too large.");
-
-   static constexpr int vdim = 3;
-   const auto F = coeff.Read();
-   const auto M = Reshape(markers, ne);
-   const auto BO = Reshape(bo, q, d-1);
-   const auto BC = Reshape(bc, q, d);
-   const auto J = Reshape(j, q, q, q, vdim, vdim, ne);
-   const auto W = Reshape(weights, q, q, q);
-   const bool cst = coeff.Size() == vdim;
-   const auto C = cst ? Reshape(F,vdim,1,1,1,1) : Reshape(F,vdim,q,q,q,ne);
-   auto Y = Reshape(y, 2*(d-1)*(d-1)*d, ne);
-
-   mfem::forall_3D(ne, q, q, vdim, [=] MFEM_HOST_DEVICE (int e)
+   if (TestType == FiniteElement::DIV)
    {
-      if (M(e) == 0) { return; } // ignore
-
-      constexpr int Q = T_Q1D ? T_Q1D : DofQuadLimits::HDIV_MAX_Q1D;
-      constexpr int D = T_D1D ? T_D1D : DofQuadLimits::HDIV_MAX_D1D;
-
-      MFEM_SHARED real_t sBot[Q*D];
-      MFEM_SHARED real_t sBct[Q*D];
-
-      // Bo and Bc into shared memory
-      const DeviceMatrix Bot(sBot, d-1, q);
-      kernels::internal::LoadB<D,Q>(d-1, q, BO, sBot);
-      const DeviceMatrix Bct(sBct, d, q);
-      kernels::internal::LoadB<D,Q>(d, q, BC, sBct);
-
-      MFEM_SHARED real_t sm0[vdim*Q*Q*Q];
-      MFEM_SHARED real_t sm1[vdim*Q*Q*Q];
-      DeviceTensor<4> QQQ(sm1, q, q, q, vdim);
-      DeviceTensor<4> DQQ(sm0, d, q, q, vdim);
-      DeviceTensor<4> DDQ(sm1, d, d, q, vdim);
-
-      MFEM_FOREACH_THREAD(vd,z,vdim)
+      if (DIM == 2)
       {
-         const real_t cst_val_0 = C(0,0,0,0,0);
-         const real_t cst_val_1 = C(1,0,0,0,0);
-         const real_t cst_val_2 = C(2,0,0,0,0);
-         MFEM_FOREACH_THREAD(y,y,q)
-         {
-            MFEM_FOREACH_THREAD(x,x,q)
-            {
-               for (int z = 0; z < q; ++z)
-               {
-                  const real_t J0 = J(x,y,z,0,vd,e);
-                  const real_t J1 = J(x,y,z,1,vd,e);
-                  const real_t J2 = J(x,y,z,2,vd,e);
-                  const real_t C0 = cst ? cst_val_0 : C(0,x,y,z,e);
-                  const real_t C1 = cst ? cst_val_1 : C(1,x,y,z,e);
-                  const real_t C2 = cst ? cst_val_2 : C(2,x,y,z,e);
-                  QQQ(x,y,z,vd) = W(x,y,z)*(J0*C0 + J1*C1 + J2*C2);
-               }
-            }
-         }
+         return HdivDLFAssemble2D<0, 0>;
       }
-      MFEM_SYNC_THREAD;
-      // Apply Bt operator
-      MFEM_FOREACH_THREAD(vd,z,vdim)
+      if (DIM == 3)
       {
-         const int nx = (vd == 0) ? d : d-1;
-         DeviceMatrix Btx = (vd == 0) ? Bct : Bot;
-         MFEM_FOREACH_THREAD(qy,y,q)
-         {
-            MFEM_FOREACH_THREAD(dx,x,nx)
-            {
-               real_t u[Q];
-               MFEM_UNROLL(Q)
-               for (int qz = 0; qz < q; ++qz) { u[qz] = 0.0; }
-               MFEM_UNROLL(Q)
-               for (int qx = 0; qx < q; ++qx)
-               {
-                  MFEM_UNROLL(Q)
-                  for (int qz = 0; qz < q; ++qz)
-                  {
-                     u[qz] += QQQ(qx,qy,qz,vd) * Btx(dx,qx);
-                  }
-               }
-               MFEM_UNROLL(Q)
-               for (int qz = 0; qz < q; ++qz) { DQQ(dx,qy,qz,vd) = u[qz]; }
-            }
-         }
+         return HdivDLFAssemble3D<0, 0>;
       }
-      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(vd,z,vdim)
-      {
-         const int nx = (vd == 0) ? d : d-1;
-         const int ny = (vd == 1) ? d : d-1;
-         DeviceMatrix Bty = (vd == 1) ? Bct : Bot;
-         MFEM_FOREACH_THREAD(dy,y,ny)
-         {
-            MFEM_FOREACH_THREAD(dx,x,nx)
-            {
-               real_t u[Q];
-               MFEM_UNROLL(Q)
-               for (int qz = 0; qz < q; ++qz) { u[qz] = 0.0; }
-               MFEM_UNROLL(Q)
-               for (int qy = 0; qy < q; ++qy)
-               {
-                  MFEM_UNROLL(Q)
-                  for (int qz = 0; qz < q; ++qz)
-                  {
-                     u[qz] += DQQ(dx,qy,qz,vd) * Bty(dy,qy);
-                  }
-               }
-               MFEM_UNROLL(Q)
-               for (int qz = 0; qz < q; ++qz) { DDQ(dx,dy,qz,vd) = u[qz]; }
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-      MFEM_FOREACH_THREAD(vd,z,vdim)
-      {
-         const int nx = (vd == 0) ? d : d-1;
-         const int ny = (vd == 1) ? d : d-1;
-         const int nz = (vd == 2) ? d : d-1;
-         DeviceTensor<5> Yxyz(Y, nx, ny, nz, vdim, ne);
-         DeviceMatrix Btz = (vd == 2) ? Bct : Bot;
-         MFEM_FOREACH_THREAD(dy,y,ny)
-         {
-            MFEM_FOREACH_THREAD(dx,x,nx)
-            {
-               real_t u[D];
-               MFEM_UNROLL(D)
-               for (int dz = 0; dz < nz; ++dz) { u[dz] = 0.0; }
-               MFEM_UNROLL(Q)
-               for (int qz = 0; qz < q; ++qz)
-               {
-                  MFEM_UNROLL(D)
-                  for (int dz = 0; dz < nz; ++dz)
-                  {
-                     u[dz] += DDQ(dx,dy,qz,vd) * Btz(dz,qz);
-                  }
-               }
-               MFEM_UNROLL(D)
-               for (int dz = 0; dz < nz; ++dz) { Yxyz(dx,dy,dz,vd,e) += u[dz]; }
-            }
-         }
-      }
-      MFEM_SYNC_THREAD;
-   });
-}
-
-static void HdivDLFAssemble(const FiniteElementSpace &fes,
-                            const IntegrationRule *ir,
-                            const Array<int> &markers,
-                            const Vector &coeff,
-                            Vector &y)
-{
-   Mesh &mesh = *fes.GetMesh();
-   const int dim = mesh.Dimension();
-   const FiniteElement *el = fes.GetTypicalFE();
-   const auto *vel = dynamic_cast<const VectorTensorFiniteElement *>(el);
-   MFEM_VERIFY(vel != nullptr, "Must be VectorTensorFiniteElement");
-   const MemoryType mt = Device::GetDeviceMemoryType();
-   const DofToQuad &maps_o = vel->GetDofToQuadOpen(*ir, DofToQuad::TENSOR);
-   const DofToQuad &maps_c = vel->GetDofToQuad(*ir, DofToQuad::TENSOR);
-   const int d = maps_c.ndof, q = maps_c.nqpt;
-   constexpr int flags = GeometricFactors::JACOBIANS;
-   const GeometricFactors *geom = mesh.GetGeometricFactors(*ir, flags, mt);
-   decltype(&HdivDLFAssemble2D<>) ker =
-      dim == 2 ? HdivDLFAssemble2D<> : HdivDLFAssemble3D<>;
-
-   if (dim==2)
-   {
-      if (d==1 && q==1) { ker=HdivDLFAssemble2D<1,1>; }
-      if (d==2 && q==2) { ker=HdivDLFAssemble2D<2,2>; }
-      if (d==3 && q==3) { ker=HdivDLFAssemble2D<3,3>; }
-      if (d==4 && q==4) { ker=HdivDLFAssemble2D<4,4>; }
-      if (d==5 && q==5) { ker=HdivDLFAssemble2D<5,5>; }
-      if (d==6 && q==6) { ker=HdivDLFAssemble2D<6,6>; }
-      if (d==7 && q==7) { ker=HdivDLFAssemble2D<7,7>; }
-      if (d==8 && q==8) { ker=HdivDLFAssemble2D<8,8>; }
    }
-
-   if (dim==3)
+   else if (TestType == FiniteElement::CURL)
    {
-      if (d==2 && q==2) { ker=HdivDLFAssemble3D<2,2>; }
-      if (d==3 && q==3) { ker=HdivDLFAssemble3D<3,3>; }
-      if (d==4 && q==4) { ker=HdivDLFAssemble3D<4,4>; }
-      if (d==5 && q==5) { ker=HdivDLFAssemble3D<5,5>; }
-      if (d==6 && q==6) { ker=HdivDLFAssemble3D<6,6>; }
-      if (d==7 && q==7) { ker=HdivDLFAssemble3D<7,7>; }
-      if (d==8 && q==8) { ker=HdivDLFAssemble3D<8,8>; }
+      if (DIM == 3)
+      {
+         return HcurlDLFAssemble3D<0, 0>;
+      }
    }
-
-   MFEM_VERIFY(ker, "No kernel ndof " << d << " nqpt " << q);
-
-   const int ne = mesh.GetNE();
-   const int *M = markers.Read();
-   const real_t *Bo = maps_o.B.Read();
-   const real_t *Bc = maps_c.B.Read();
-   const real_t *J = geom->J.Read();
-   const real_t *W = ir->GetWeights().Read();
-   real_t *Y = y.ReadWrite();
-   ker(ne, d, q, M, Bo, Bc, J, W, coeff, Y);
+   MFEM_ABORT("");
 }
+/// \endcond DO_NOT_DOCUMENT
 
 void VectorFEDomainLFIntegrator::AssembleDevice(const FiniteElementSpace &fes,
                                                 const Array<int> &markers,
@@ -337,15 +96,23 @@ void VectorFEDomainLFIntegrator::AssembleDevice(const FiniteElementSpace &fes,
    QuadratureSpace qs(*fes.GetMesh(), *ir);
    CoefficientVector coeff(QF, qs, CoefficientStorage::COMPRESSED);
 
-   const int fe_type = fe.GetDerivType();
-   if (fe_type == FiniteElement::DIV)
-   {
-      HdivDLFAssemble(fes, ir, markers, coeff, b);
-   }
-   else
-   {
-      MFEM_ABORT("Not implemented.");
-   }
+   const FiniteElement::DerivType fe_type =
+      static_cast<FiniteElement::DerivType>(fe.GetDerivType());
+
+   Mesh &mesh = *fes.GetMesh();
+   const int dim = mesh.Dimension();
+   const FiniteElement *el = fes.GetTypicalFE();
+   const auto *vel = dynamic_cast<const VectorTensorFiniteElement *>(el);
+   MFEM_VERIFY(vel != nullptr, "Must be VectorTensorFiniteElement");
+   const MemoryType mt = Device::GetDeviceMemoryType();
+   const DofToQuad &maps_o = vel->GetDofToQuadOpen(*ir, DofToQuad::TENSOR);
+   const DofToQuad &maps_c = vel->GetDofToQuad(*ir, DofToQuad::TENSOR);
+   const int d = maps_c.ndof, q = maps_c.nqpt;
+   constexpr int flags = GeometricFactors::JACOBIANS;
+   const GeometricFactors *geom = mesh.GetGeometricFactors(*ir, flags, mt);
+
+   AssembleKernels::Run(fe_type, dim, d, q, mesh.GetNE(), markers, geom->J,
+                        ir->GetWeights(), maps_o.B, maps_c.B, coeff, b, d, q);
 }
 
 } // namespace mfem

fem/integ/nonlininteg_vecconvection_pa.cpp

@@ -9,21 +9,51 @@
 // terms of the BSD-3 license. We welcome feedback and contributions, see file
 // CONTRIBUTING.md for details.
 
-#include "../../general/forall.hpp"
-#include "../nonlininteg.hpp"
 #include "../ceed/integrators/nlconvection/nlconvection.hpp"
+#include "./nonlininteg_vecconvection_pa.hpp" // IWYU pragma: keep
+#include "./nonlininteg_vecconvection_pa_grad.hpp" // IWYU pragma: keep
+#include "./nonlininteg_vecconvection_pa_diag.hpp" // IWYU pragma: keep
 
 namespace mfem
 {
 
+VectorConvectionNLFIntegrator::Kernels::Kernels()
+{
+   // 2D
+   VectorConvectionNLFIntegrator::AddSpecialization<2, 2, 2>();
+   VectorConvectionNLFIntegrator::AddSpecialization<2, 2, 3>();
+   VectorConvectionNLFIntegrator::AddSpecialization<2, 3, 4>();
+   VectorConvectionNLFIntegrator::AddSpecialization<2, 3, 5>();
+   VectorConvectionNLFIntegrator::AddSpecialization<2, 4, 5>();
+   VectorConvectionNLFIntegrator::AddSpecialization<2, 4, 6>();
+   VectorConvectionNLFIntegrator::AddSpecialization<2, 5, 7>();
+   VectorConvectionNLFIntegrator::AddSpecialization<2, 5, 8>();
+   VectorConvectionNLFIntegrator::AddSpecialization<2, 6, 8>();
+   // 3D
+   VectorConvectionNLFIntegrator::AddSpecialization<3, 2, 3>();
+   VectorConvectionNLFIntegrator::AddSpecialization<3, 2, 4>();
+   VectorConvectionNLFIntegrator::AddSpecialization<3, 2, 5>();
+   VectorConvectionNLFIntegrator::AddSpecialization<3, 3, 4>();
+   VectorConvectionNLFIntegrator::AddSpecialization<3, 3, 5>();
+   VectorConvectionNLFIntegrator::AddSpecialization<3, 3, 6>();
+   VectorConvectionNLFIntegrator::AddSpecialization<3, 4, 5>();
+   VectorConvectionNLFIntegrator::AddSpecialization<3, 4, 6>();
+   VectorConvectionNLFIntegrator::AddSpecialization<3, 4, 7>();
+   VectorConvectionNLFIntegrator::AddSpecialization<3, 4, 8>();
+   VectorConvectionNLFIntegrator::AddSpecialization<3, 5, 6>();
+   VectorConvectionNLFIntegrator::AddSpecialization<3, 5, 7>();
+   VectorConvectionNLFIntegrator::AddSpecialization<3, 5, 8>();
+}
+
 void VectorConvectionNLFIntegrator::AssemblePA(const FiniteElementSpace &fes)
 {
    MFEM_ASSERT(fes.GetOrdering() == Ordering::byNODES,
                "PA Only supports Ordering::byNODES!");
    Mesh *mesh = fes.GetMesh();
    const FiniteElement &el = *fes.GetTypicalFE();
-   ElementTransformation &T = *mesh->GetTypicalElementTransformation();
-   const IntegrationRule *ir = IntRule ? IntRule : &GetRule(el, T);
+   ElementTransformation &Tr = *mesh->GetTypicalElementTransformation();
+   const IntegrationRule *ir = IntRule ? IntRule : &GetRule(el, Tr);
+
    if (DeviceCanUseCeed())
    {
       delete ceedOp;
@@ -39,769 +69,124 @@ void VectorConvectionNLFIntegrator::AssemblePA(const FiniteElementSpace &fes)
       }
       return;
    }
-   dim = mesh->Dimension();
-   ne = fes.GetMesh()->GetNE();
+
+   ne = mesh->GetNE();
    nq = ir->GetNPoints();
-   geom = mesh->GetGeometricFactors(*ir, GeometricFactors::JACOBIANS);
+   dim = mesh->Dimension();
+   MFEM_VERIFY(dim == 2 || dim == 3, "Dimension not supported");
+
+   const MemoryType mt = pa_mt == MemoryType::DEFAULT
+                         ? Device::GetDeviceMemoryType()
+                         : pa_mt;
+   pa_adj.SetSize(ne * nq * dim * dim, mt);
+   geom = mesh->GetGeometricFactors(*ir, GeometricFactors::JACOBIANS, mt);
    maps = &el.GetDofToQuad(*ir, DofToQuad::TENSOR);
-   pa_data.SetSize(ne * nq * dim * dim, Device::GetMemoryType());
-   real_t COEFF = 1.0;
-   if (Q)
-   {
-      ConstantCoefficient *cQ = dynamic_cast<ConstantCoefficient *>(Q);
-      MFEM_VERIFY(cQ != NULL, "only ConstantCoefficient is supported!");
-      COEFF = cQ->constant;
-   }
-   const int NE = ne;
-   const int NQ = nq;
-   auto W = ir->GetWeights().Read();
-   if (dim == 1)
-   {
-      MFEM_ABORT("dim==1 not supported!");
-   }
+   d1d = maps->ndof;
+   q1d = maps->nqpt;
+
+   QuadratureSpace qs(*mesh, *ir);
+   CoefficientVector coeff(Q, qs, CoefficientStorage::COMPRESSED);
+
+   const int nq1d = q1d * q1d * (dim==3 ? q1d : 1);
+   MFEM_VERIFY(coeff.Size() == 1 || coeff.Size() == nq1d*ne, "Invalid coeff");
+   MFEM_VERIFY(ir->GetWeights().Size() == nq1d, "Invalid weights size");
+
+   const auto w_r = ir->GetWeights().Read();
+   const bool const_coeff = coeff.Size() == 1;
+
    if (dim == 2)
    {
-      auto J = Reshape(geom->J.Read(), NQ, 2, 2, NE);
-      auto G = Reshape(pa_data.Write(), NQ, 2, 2, NE);
-      mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+      const int Q1D = q1d;
+      constexpr int VDIM = 2, DIM = 2;
+      const auto W = Reshape(w_r, Q1D, Q1D);
+      const auto C = const_coeff ?
+                     Reshape(coeff.Read(), 1, 1, 1) :
+                     Reshape(coeff.Read(), Q1D, Q1D, ne);
+      const auto J = Reshape(geom->J.Read(), Q1D, Q1D, VDIM, DIM, ne);
+      auto A = Reshape(pa_adj.Write(), VDIM, DIM, Q1D, Q1D, ne);
+
+      mfem::forall_2D(ne, Q1D, Q1D, [=] MFEM_HOST_DEVICE(int e)
       {
-         for (int q = 0; q < NQ; ++q)
+         MFEM_FOREACH_THREAD_DIRECT(qy, y, Q1D)
          {
-            const real_t J11 = J(q, 0, 0, e);
-            const real_t J12 = J(q, 0, 1, e);
-            const real_t J21 = J(q, 1, 0, e);
-            const real_t J22 = J(q, 1, 1, e);
-            // Store wq * Q * adj(J)
-            G(q, 0, 0, e) = W[q] * COEFF * J22;  // 1,1
-            G(q, 0, 1, e) = W[q] * COEFF * -J12; // 1,2
-            G(q, 1, 0, e) = W[q] * COEFF * -J21; // 2,1
-            G(q, 1, 1, e) = W[q] * COEFF * J11;  // 2,2
+            MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+            {
+               const real_t J11 = J(qx, qy, 0, 0, e), J12 = J(qx, qy, 0, 1, e);
+               const real_t J21 = J(qx, qy, 1, 0, e), J22 = J(qx, qy, 1, 1, e);
+               // adj(J)
+               const real_t A11 = +J22, A12 = -J12;
+               const real_t A21 = -J21, A22 = +J11;
+               // Store w * coeff * adj(J)
+               const real_t w = W(qx, qy);
+               const real_t c = const_coeff ? C(0, 0, 0) : C(qx, qy, e);
+               A(0, 0, qx, qy, e) = w * c * A11;
+               A(1, 0, qx, qy, e) = w * c * A12;
+               A(0, 1, qx, qy, e) = w * c * A21;
+               A(1, 1, qx, qy, e) = w * c * A22;
+            }
          }
       });
    }
-   if (dim == 3)
+   else if (dim == 3)
    {
-      auto J = Reshape(geom->J.Read(), NQ, 3, 3, NE);
-      auto G = Reshape(pa_data.Write(), NQ, 3, 3, NE);
-      mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+      const int Q1D = q1d;
+      constexpr int VDIM = 3, DIM = 3;
+      const auto W = Reshape(w_r, Q1D, Q1D, Q1D);
+      const auto C = const_coeff ?
+                     Reshape(coeff.Read(), 1, 1, 1, 1) :
+                     Reshape(coeff.Read(), Q1D, Q1D, Q1D, ne);
+      const auto J = Reshape(geom->J.Read(), Q1D, Q1D, Q1D, VDIM, DIM, ne);
+      auto A = Reshape(pa_adj.Write(), VDIM, DIM, Q1D, Q1D, Q1D, ne);
+
+      mfem::forall_3D(ne, Q1D, Q1D, Q1D, [=] MFEM_HOST_DEVICE(int e)
       {
-         for (int q = 0; q < NQ; ++q)
+         MFEM_FOREACH_THREAD_DIRECT(qz, z, Q1D)
          {
-            const real_t J11 = J(q, 0, 0, e);
-            const real_t J21 = J(q, 1, 0, e);
-            const real_t J31 = J(q, 2, 0, e);
-            const real_t J12 = J(q, 0, 1, e);
-            const real_t J22 = J(q, 1, 1, e);
-            const real_t J32 = J(q, 2, 1, e);
-            const real_t J13 = J(q, 0, 2, e);
-            const real_t J23 = J(q, 1, 2, e);
-            const real_t J33 = J(q, 2, 2, e);
-            const real_t cw = W[q] * COEFF;
-            // adj(J)
-            const real_t A11 = (J22 * J33) - (J23 * J32);
-            const real_t A12 = (J32 * J13) - (J12 * J33);
-            const real_t A13 = (J12 * J23) - (J22 * J13);
-            const real_t A21 = (J31 * J23) - (J21 * J33);
-            const real_t A22 = (J11 * J33) - (J13 * J31);
-            const real_t A23 = (J21 * J13) - (J11 * J23);
-            const real_t A31 = (J21 * J32) - (J31 * J22);
-            const real_t A32 = (J31 * J12) - (J11 * J32);
-            const real_t A33 = (J11 * J22) - (J12 * J21);
-            // Store wq * Q * adj(J)
-            G(q, 0, 0, e) = cw * A11; // 1,1
-            G(q, 0, 1, e) = cw * A12; // 1,2
-            G(q, 0, 2, e) = cw * A13; // 1,3
-            G(q, 1, 0, e) = cw * A21; // 2,1
-            G(q, 1, 1, e) = cw * A22; // 2,2
-            G(q, 1, 2, e) = cw * A23; // 2,3
-            G(q, 2, 0, e) = cw * A31; // 3,1
-            G(q, 2, 1, e) = cw * A32; // 3,2
-            G(q, 2, 2, e) = cw * A33; // 3,3
+            MFEM_FOREACH_THREAD_DIRECT(qy, y, Q1D)
+            {
+               MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+               {
+                  const real_t J11 = J(qx, qy, qz, 0, 0, e),
+                               J12 = J(qx, qy, qz, 0, 1, e),
+                               J13 = J(qx, qy, qz, 0, 2, e);
+                  const real_t J21 = J(qx, qy, qz, 1, 0, e),
+                               J22 = J(qx, qy, qz, 1, 1, e),
+                               J23 = J(qx, qy, qz, 1, 2, e);
+                  const real_t J31 = J(qx, qy, qz, 2, 0, e),
+                               J32 = J(qx, qy, qz, 2, 1, e),
+                               J33 = J(qx, qy, qz, 2, 2, e);
+                  const real_t c =
+                     const_coeff ? C(0, 0, 0, 0) : C(qx, qy, qz, e);
+                  const real_t cw = W(qx, qy, qz) * c;
+                  // adj(J)
+                  const real_t A11 = (J22 * J33) - (J23 * J32);
+                  const real_t A12 = (J32 * J13) - (J12 * J33);
+                  const real_t A13 = (J12 * J23) - (J22 * J13);
+                  const real_t A21 = (J31 * J23) - (J21 * J33);
+                  const real_t A22 = (J11 * J33) - (J13 * J31);
+                  const real_t A23 = (J21 * J13) - (J11 * J23);
+                  const real_t A31 = (J21 * J32) - (J31 * J22);
+                  const real_t A32 = (J31 * J12) - (J11 * J32);
+                  const real_t A33 = (J11 * J22) - (J12 * J21);
+                  // Store wq * coeff * adj(J)
+                  A(0, 0, qx, qy, qz, e) = cw * A11;
+                  A(1, 0, qx, qy, qz, e) = cw * A12;
+                  A(2, 0, qx, qy, qz, e) = cw * A13;
+                  A(0, 1, qx, qy, qz, e) = cw * A21;
+                  A(1, 1, qx, qy, qz, e) = cw * A22;
+                  A(2, 1, qx, qy, qz, e) = cw * A23;
+                  A(0, 2, qx, qy, qz, e) = cw * A31;
+                  A(1, 2, qx, qy, qz, e) = cw * A32;
+                  A(2, 2, qx, qy, qz, e) = cw * A33;
+               }
+            }
          }
       });
    }
-}
-
-// PA Convection NL 2D kernel
-template<int T_D1D = 0, int T_Q1D = 0>
-static void PAConvectionNLApply2D(const int NE,
-                                  const Array<real_t> &b,
-                                  const Array<real_t> &g,
-                                  const Array<real_t> &bt,
-                                  const Vector &q_,
-                                  const Vector &x_,
-                                  Vector &y_,
-                                  const int d1d = 0,
-                                  const int q1d = 0)
-{
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto G = Reshape(g.Read(), Q1D, D1D);
-   auto Bt = Reshape(bt.Read(), D1D, Q1D);
-   auto Q = Reshape(q_.Read(), Q1D * Q1D, 2, 2, NE);
-   auto x = Reshape(x_.Read(), D1D, D1D, 2, NE);
-   auto y = Reshape(y_.ReadWrite(), D1D, D1D, 2, NE);
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
+   else
    {
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-
-      real_t data[max_Q1D][max_Q1D][2];
-      real_t grad0[max_Q1D][max_Q1D][2];
-      real_t grad1[max_Q1D][max_Q1D][2];
-      real_t Z[max_Q1D][max_Q1D][2];
-      for (int qy = 0; qy < Q1D; ++qy)
-      {
-         for (int qx = 0; qx < Q1D; ++qx)
-         {
-            data[qy][qx][0] = 0.0;
-            data[qy][qx][1] = 0.0;
-            grad0[qy][qx][0] = 0.0;
-            grad0[qy][qx][1] = 0.0;
-            grad1[qy][qx][0] = 0.0;
-            grad1[qy][qx][1] = 0.0;
-         }
-      }
-      for (int dy = 0; dy < D1D; ++dy)
-      {
-         real_t dataX[max_Q1D][2];
-         real_t gradX0[max_Q1D][2];
-         real_t gradX1[max_Q1D][2];
-         for (int qx = 0; qx < Q1D; ++qx)
-         {
-            dataX[qx][0] = 0.0;
-            dataX[qx][1] = 0.0;
-            gradX0[qx][0] = 0.0;
-            gradX0[qx][1] = 0.0;
-            gradX1[qx][0] = 0.0;
-            gradX1[qx][1] = 0.0;
-         }
-         for (int dx = 0; dx < D1D; ++dx)
-         {
-            const real_t s0 = x(dx, dy, 0, e);
-            const real_t s1 = x(dx, dy, 1, e);
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               const real_t Bx = B(qx, dx);
-               const real_t Gx = G(qx, dx);
-               dataX[qx][0] += s0 * Bx;
-               dataX[qx][1] += s1 * Bx;
-               gradX0[qx][0] += s0 * Gx;
-               gradX0[qx][1] += s0 * Bx;
-               gradX1[qx][0] += s1 * Gx;
-               gradX1[qx][1] += s1 * Bx;
-            }
-         }
-         for (int qy = 0; qy < Q1D; ++qy)
-         {
-            const real_t By = B(qy, dy);
-            const real_t Gy = G(qy, dy);
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               data[qy][qx][0] += dataX[qx][0] * By;
-               data[qy][qx][1] += dataX[qx][1] * By;
-               grad0[qy][qx][0] += gradX0[qx][0] * By;
-               grad0[qy][qx][1] += gradX0[qx][1] * Gy;
-               grad1[qy][qx][0] += gradX1[qx][0] * By;
-               grad1[qy][qx][1] += gradX1[qx][1] * Gy;
-            }
-         }
-      }
-      for (int qy = 0; qy < Q1D; ++qy)
-      {
-         for (int qx = 0; qx < Q1D; ++qx)
-         {
-            const int q = qx + qy * Q1D;
-            const real_t u1 = data[qy][qx][0];
-            const real_t u2 = data[qy][qx][1];
-            const real_t grad00 = grad0[qy][qx][0];
-            const real_t grad01 = grad0[qy][qx][1];
-            const real_t grad10 = grad1[qy][qx][0];
-            const real_t grad11 = grad1[qy][qx][1];
-            const real_t Dxu1 = grad00 * Q(q, 0, 0, e) + grad01 * Q(q, 1, 0, e);
-            const real_t Dyu1 = grad00 * Q(q, 0, 1, e) + grad01 * Q(q, 1, 1, e);
-            const real_t Dxu2 = grad10 * Q(q, 0, 0, e) + grad11 * Q(q, 1, 0, e);
-            const real_t Dyu2 = grad10 * Q(q, 0, 1, e) + grad11 * Q(q, 1, 1, e);
-            Z[qy][qx][0] = u1 * Dxu1 + u2 * Dyu1;
-            Z[qy][qx][1] = u1 * Dxu2 + u2 * Dyu2;
-         }
-      }
-      for (int qy = 0; qy < Q1D; ++qy)
-      {
-         real_t Y[max_D1D][2];
-         for (int dx = 0; dx < D1D; ++dx)
-         {
-            Y[dx][0] = 0.0;
-            Y[dx][1] = 0.0;
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               const real_t Btx = Bt(dx, qx);
-               Y[dx][0] += Btx * Z[qy][qx][0];
-               Y[dx][1] += Btx * Z[qy][qx][1];
-            }
-         }
-         for (int dy = 0; dy < D1D; ++dy)
-         {
-            for (int dx = 0; dx < D1D; ++dx)
-            {
-               const real_t Bty = Bt(dy, qy);
-               y(dx, dy, 0, e) += Bty * Y[dx][0];
-               y(dx, dy, 1, e) += Bty * Y[dx][1];
-            }
-         }
-      }
-   });
-}
-
-// PA Convection NL 3D kernel
-template<int T_D1D = 0, int T_Q1D = 0>
-static void PAConvectionNLApply3D(const int NE,
-                                  const Array<real_t> &b,
-                                  const Array<real_t> &g,
-                                  const Array<real_t> &bt,
-                                  const Vector &q_,
-                                  const Vector &x_,
-                                  Vector &y_,
-                                  const int d1d = 0,
-                                  const int q1d = 0)
-{
-   constexpr int VDIM = 3;
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   MFEM_VERIFY(D1D <= DeviceDofQuadLimits::Get().MAX_D1D, "");
-   MFEM_VERIFY(Q1D <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
-
-   auto B = Reshape(b.Read(), Q1D, D1D);
-   auto G = Reshape(g.Read(), Q1D, D1D);
-   auto Bt = Reshape(bt.Read(), D1D, Q1D);
-   auto Q = Reshape(q_.Read(), Q1D * Q1D * Q1D, VDIM, VDIM, NE);
-   auto x = Reshape(x_.Read(), D1D, D1D, D1D, VDIM, NE);
-   auto y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, VDIM, NE);
-
-   mfem::forall(NE, [=] MFEM_HOST_DEVICE (int e)
-   {
-      constexpr int VDIM = 3;
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      constexpr int max_D1D = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
-      constexpr int max_Q1D = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
-
-      real_t data[max_Q1D][max_Q1D][max_Q1D][VDIM];
-      real_t grad0[max_Q1D][max_Q1D][max_Q1D][VDIM];
-      real_t grad1[max_Q1D][max_Q1D][max_Q1D][VDIM];
-      real_t grad2[max_Q1D][max_Q1D][max_Q1D][VDIM];
-      real_t Z[max_Q1D][max_Q1D][max_Q1D][VDIM];
-      for (int qz = 0; qz < Q1D; ++qz)
-      {
-         for (int qy = 0; qy < Q1D; ++qy)
-         {
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               data[qz][qy][qx][0] = 0.0;
-               data[qz][qy][qx][1] = 0.0;
-               data[qz][qy][qx][2] = 0.0;
-
-               grad0[qz][qy][qx][0] = 0.0;
-               grad0[qz][qy][qx][1] = 0.0;
-               grad0[qz][qy][qx][2] = 0.0;
-
-               grad1[qz][qy][qx][0] = 0.0;
-               grad1[qz][qy][qx][1] = 0.0;
-               grad1[qz][qy][qx][2] = 0.0;
-
-               grad2[qz][qy][qx][0] = 0.0;
-               grad2[qz][qy][qx][1] = 0.0;
-               grad2[qz][qy][qx][2] = 0.0;
-            }
-         }
-      }
-      for (int dz = 0; dz < D1D; ++dz)
-      {
-         real_t dataXY[max_Q1D][max_Q1D][VDIM];
-         real_t gradXY0[max_Q1D][max_Q1D][VDIM];
-         real_t gradXY1[max_Q1D][max_Q1D][VDIM];
-         real_t gradXY2[max_Q1D][max_Q1D][VDIM];
-         for (int qy = 0; qy < Q1D; ++qy)
-         {
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               dataXY[qy][qx][0] = 0.0;
-               dataXY[qy][qx][1] = 0.0;
-               dataXY[qy][qx][2] = 0.0;
-
-               gradXY0[qy][qx][0] = 0.0;
-               gradXY0[qy][qx][1] = 0.0;
-               gradXY0[qy][qx][2] = 0.0;
-
-               gradXY1[qy][qx][0] = 0.0;
-               gradXY1[qy][qx][1] = 0.0;
-               gradXY1[qy][qx][2] = 0.0;
-
-               gradXY2[qy][qx][0] = 0.0;
-               gradXY2[qy][qx][1] = 0.0;
-               gradXY2[qy][qx][2] = 0.0;
-            }
-         }
-         for (int dy = 0; dy < D1D; ++dy)
-         {
-            real_t dataX[max_Q1D][VDIM];
-            real_t gradX0[max_Q1D][VDIM];
-            real_t gradX1[max_Q1D][VDIM];
-            real_t gradX2[max_Q1D][VDIM];
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               dataX[qx][0] = 0.0;
-               dataX[qx][1] = 0.0;
-               dataX[qx][2] = 0.0;
-
-               gradX0[qx][0] = 0.0;
-               gradX0[qx][1] = 0.0;
-               gradX0[qx][2] = 0.0;
-
-               gradX1[qx][0] = 0.0;
-               gradX1[qx][1] = 0.0;
-               gradX1[qx][2] = 0.0;
-
-               gradX2[qx][0] = 0.0;
-               gradX2[qx][1] = 0.0;
-               gradX2[qx][2] = 0.0;
-            }
-            for (int dx = 0; dx < D1D; ++dx)
-            {
-               const real_t s0 = x(dx, dy, dz, 0, e);
-               const real_t s1 = x(dx, dy, dz, 1, e);
-               const real_t s2 = x(dx, dy, dz, 2, e);
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  const real_t Bx = B(qx, dx);
-                  const real_t Gx = G(qx, dx);
-
-                  dataX[qx][0] += s0 * Bx;
-                  dataX[qx][1] += s1 * Bx;
-                  dataX[qx][2] += s2 * Bx;
-
-                  gradX0[qx][0] += s0 * Gx;
-                  gradX0[qx][1] += s0 * Bx;
-                  gradX0[qx][2] += s0 * Bx;
-
-                  gradX1[qx][0] += s1 * Gx;
-                  gradX1[qx][1] += s1 * Bx;
-                  gradX1[qx][2] += s1 * Bx;
-
-                  gradX2[qx][0] += s2 * Gx;
-                  gradX2[qx][1] += s2 * Bx;
-                  gradX2[qx][2] += s2 * Bx;
-               }
-            }
-            for (int qy = 0; qy < Q1D; ++qy)
-            {
-               const real_t By = B(qy, dy);
-               const real_t Gy = G(qy, dy);
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  dataXY[qy][qx][0] += dataX[qx][0] * By;
-                  dataXY[qy][qx][1] += dataX[qx][1] * By;
-                  dataXY[qy][qx][2] += dataX[qx][2] * By;
-
-                  gradXY0[qy][qx][0] += gradX0[qx][0] * By;
-                  gradXY0[qy][qx][1] += gradX0[qx][1] * Gy;
-                  gradXY0[qy][qx][2] += gradX0[qx][2] * By;
-
-                  gradXY1[qy][qx][0] += gradX1[qx][0] * By;
-                  gradXY1[qy][qx][1] += gradX1[qx][1] * Gy;
-                  gradXY1[qy][qx][2] += gradX1[qx][2] * By;
-
-                  gradXY2[qy][qx][0] += gradX2[qx][0] * By;
-                  gradXY2[qy][qx][1] += gradX2[qx][1] * Gy;
-                  gradXY2[qy][qx][2] += gradX2[qx][2] * By;
-               }
-            }
-         }
-         for (int qz = 0; qz < Q1D; ++qz)
-         {
-            const real_t Bz = B(qz, dz);
-            const real_t Gz = G(qz, dz);
-            for (int qy = 0; qy < Q1D; ++qy)
-            {
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  data[qz][qy][qx][0] += dataXY[qy][qx][0] * Bz;
-                  data[qz][qy][qx][1] += dataXY[qy][qx][1] * Bz;
-                  data[qz][qy][qx][2] += dataXY[qy][qx][2] * Bz;
-
-                  grad0[qz][qy][qx][0] += gradXY0[qy][qx][0] * Bz;
-                  grad0[qz][qy][qx][1] += gradXY0[qy][qx][1] * Bz;
-                  grad0[qz][qy][qx][2] += gradXY0[qy][qx][2] * Gz;
-
-                  grad1[qz][qy][qx][0] += gradXY1[qy][qx][0] * Bz;
-                  grad1[qz][qy][qx][1] += gradXY1[qy][qx][1] * Bz;
-                  grad1[qz][qy][qx][2] += gradXY1[qy][qx][2] * Gz;
-
-                  grad2[qz][qy][qx][0] += gradXY2[qy][qx][0] * Bz;
-                  grad2[qz][qy][qx][1] += gradXY2[qy][qx][1] * Bz;
-                  grad2[qz][qy][qx][2] += gradXY2[qy][qx][2] * Gz;
-               }
-            }
-         }
-      }
-      for (int qz = 0; qz < Q1D; ++qz)
-      {
-         for (int qy = 0; qy < Q1D; ++qy)
-         {
-            for (int qx = 0; qx < Q1D; ++qx)
-            {
-               const int q = qx + Q1D * (qy + qz * Q1D);
-
-               const real_t u1 = data[qz][qy][qx][0];
-               const real_t u2 = data[qz][qy][qx][1];
-               const real_t u3 = data[qz][qy][qx][2];
-
-               const real_t grad00 = grad0[qz][qy][qx][0];
-               const real_t grad01 = grad0[qz][qy][qx][1];
-               const real_t grad02 = grad0[qz][qy][qx][2];
-
-               const real_t grad10 = grad1[qz][qy][qx][0];
-               const real_t grad11 = grad1[qz][qy][qx][1];
-               const real_t grad12 = grad1[qz][qy][qx][2];
-
-               const real_t grad20 = grad2[qz][qy][qx][0];
-               const real_t grad21 = grad2[qz][qy][qx][1];
-               const real_t grad22 = grad2[qz][qy][qx][2];
-
-               const real_t Dxu1 = grad00 * Q(q, 0, 0, e)
-                                   + grad01 * Q(q, 1, 0, e)
-                                   + grad02 * Q(q, 2, 0, e);
-               const real_t Dyu1 = grad00 * Q(q, 0, 1, e)
-                                   + grad01 * Q(q, 1, 1, e)
-                                   + grad02 * Q(q, 2, 1, e);
-               const real_t Dzu1 = grad00 * Q(q, 0, 2, e)
-                                   + grad01 * Q(q, 1, 2, e)
-                                   + grad02 * Q(q, 2, 2, e);
-
-               const real_t Dxu2 = grad10 * Q(q, 0, 0, e)
-                                   + grad11 * Q(q, 1, 0, e)
-                                   + grad12 * Q(q, 2, 0, e);
-               const real_t Dyu2 = grad10 * Q(q, 0, 1, e)
-                                   + grad11 * Q(q, 1, 1, e)
-                                   + grad12 * Q(q, 2, 1, e);
-               const real_t Dzu2 = grad10 * Q(q, 0, 2, e)
-                                   + grad11 * Q(q, 1, 2, e)
-                                   + grad12 * Q(q, 2, 2, e);
-
-               const real_t Dxu3 = grad20 * Q(q, 0, 0, e)
-                                   + grad21 * Q(q, 1, 0, e)
-                                   + grad22 * Q(q, 2, 0, e);
-               const real_t Dyu3 = grad20 * Q(q, 0, 1, e)
-                                   + grad21 * Q(q, 1, 1, e)
-                                   + grad22 * Q(q, 2, 1, e);
-               const real_t Dzu3 = grad20 * Q(q, 0, 2, e)
-                                   + grad21 * Q(q, 1, 2, e)
-                                   + grad22 * Q(q, 2, 2, e);
-
-               Z[qz][qy][qx][0] = u1 * Dxu1 + u2 * Dyu1 + u3 * Dzu1;
-               Z[qz][qy][qx][1] = u1 * Dxu2 + u2 * Dyu2 + u3 * Dzu2;
-               Z[qz][qy][qx][2] = u1 * Dxu3 + u2 * Dyu3 + u3 * Dzu3;
-            }
-         }
-      }
-      for (int qz = 0; qz < Q1D; ++qz)
-      {
-         real_t opXY[max_D1D][max_D1D][VDIM];
-         for (int dy = 0; dy < D1D; ++dy)
-         {
-            for (int dx = 0; dx < D1D; ++dx)
-            {
-               opXY[dy][dx][0] = 0.0;
-               opXY[dy][dx][1] = 0.0;
-               opXY[dy][dx][2] = 0.0;
-            }
-         }
-         for (int qy = 0; qy < Q1D; ++qy)
-         {
-            real_t opX[max_D1D][VDIM];
-            for (int dx = 0; dx < D1D; ++dx)
-            {
-               opX[dx][0] = 0.0;
-               opX[dx][1] = 0.0;
-               opX[dx][2] = 0.0;
-               for (int qx = 0; qx < Q1D; ++qx)
-               {
-                  const real_t Btx = Bt(dx, qx);
-                  opX[dx][0] += Btx * Z[qz][qy][qx][0];
-                  opX[dx][1] += Btx * Z[qz][qy][qx][1];
-                  opX[dx][2] += Btx * Z[qz][qy][qx][2];
-               }
-            }
-            for (int dy = 0; dy < D1D; ++dy)
-            {
-               for (int dx = 0; dx < D1D; ++dx)
-               {
-                  const real_t Bty = Bt(dy, qy);
-                  opXY[dy][dx][0] += Bty * opX[dx][0];
-                  opXY[dy][dx][1] += Bty * opX[dx][1];
-                  opXY[dy][dx][2] += Bty * opX[dx][2];
-               }
-            }
-         }
-         for (int dz = 0; dz < D1D; ++dz)
-         {
-            for (int dy = 0; dy < D1D; ++dy)
-            {
-               for (int dx = 0; dx < D1D; ++dx)
-               {
-                  const real_t Btz = Bt(dz, qz);
-                  y(dx, dy, dz, 0, e) += Btz * opXY[dy][dx][0];
-                  y(dx, dy, dz, 1, e) += Btz * opXY[dy][dx][1];
-                  y(dx, dy, dz, 2, e) += Btz * opXY[dy][dx][2];
-               }
-            }
-         }
-      }
-   });
-}
-
-template<int T_D1D = 0, int T_Q1D = 0, int T_MAX_D1D = 0, int T_MAX_Q1D = 0>
-static void SmemPAConvectionNLApply3D(const int NE,
-                                      const Array<real_t> &b_,
-                                      const Array<real_t> &g_,
-                                      const Vector &d_,
-                                      const Vector &x_,
-                                      Vector &y_,
-                                      const int d1d = 0,
-                                      const int q1d = 0)
-{
-   constexpr int VDIM = 3;
-   const int D1D = T_D1D ? T_D1D : d1d;
-   const int Q1D = T_Q1D ? T_Q1D : q1d;
-   constexpr int MD1 = T_D1D ? T_D1D : T_MAX_D1D;
-   constexpr int MQ1 = T_Q1D ? T_Q1D : T_MAX_Q1D;
-   MFEM_VERIFY(D1D <= MD1, "");
-   MFEM_VERIFY(Q1D <= MQ1, "");
-
-   auto b = Reshape(b_.Read(), Q1D, D1D);
-   auto g = Reshape(g_.Read(), Q1D, D1D);
-   auto D = Reshape(d_.Read(), Q1D * Q1D * Q1D, VDIM, VDIM, NE);
-   auto x = Reshape(x_.Read(), D1D, D1D, D1D, VDIM, NE);
-   auto Y = Reshape(y_.ReadWrite(), D1D, D1D, D1D, VDIM, NE);
-
-   mfem::forall_3D(NE, Q1D, Q1D, Q1D, [=] MFEM_HOST_DEVICE (int e)
-   {
-      const int tidz = MFEM_THREAD_ID(z);
-      const int D1D = T_D1D ? T_D1D : d1d;
-      const int Q1D = T_Q1D ? T_Q1D : q1d;
-      constexpr int MD1 = T_D1D ? T_D1D : T_MAX_D1D;
-      constexpr int MQ1 = T_Q1D ? T_Q1D : T_MAX_Q1D;
-      MFEM_SHARED real_t BG[2][MQ1 * MD1];
-      real_t(*B)[MD1] = (real_t(*)[MD1])(BG + 0);
-      real_t(*G)[MD1] = (real_t(*)[MD1])(BG + 1);
-      real_t(*Bt)[MQ1] = (real_t(*)[MQ1])(BG + 0);
-      MFEM_SHARED real_t U[2][MQ1][MQ1][MQ1];
-      MFEM_SHARED real_t sm0[3][MQ1 * MQ1 * MQ1];
-      MFEM_SHARED real_t sm1[3][MQ1 * MQ1 * MQ1];
-      real_t(*DDQ0)[MD1][MQ1] = (real_t(*)[MD1][MQ1])(sm0 + 0);
-      real_t(*DDQ1)[MD1][MQ1] = (real_t(*)[MD1][MQ1])(sm0 + 1);
-      real_t(*X)[MD1][MD1] = (real_t(*)[MD1][MD1])(sm0 + 2);
-      real_t(*DQQ0)[MQ1][MQ1] = (real_t(*)[MQ1][MQ1])(sm1 + 0);
-      real_t(*DQQ1)[MQ1][MQ1] = (real_t(*)[MQ1][MQ1])(sm1 + 1);
-      real_t(*DQQ2)[MQ1][MQ1] = (real_t(*)[MQ1][MQ1])(sm1 + 2);
-      real_t(*QQQ0)[MQ1][MQ1] = (real_t(*)[MQ1][MQ1])(sm0 + 0);
-      real_t(*QQQ1)[MQ1][MQ1] = (real_t(*)[MQ1][MQ1])(sm0 + 1);
-      real_t(*QQQ2)[MQ1][MQ1] = (real_t(*)[MQ1][MQ1])(sm0 + 2);
-      real_t(*QQD0)[MQ1][MD1] = (real_t(*)[MQ1][MD1])(sm1 + 0);
-      real_t(*QDD0)[MD1][MD1] = (real_t(*)[MD1][MD1])(sm0 + 0);
-      MFEM_SHARED real_t Z[MQ1][MQ1][MQ1];
-
-      for (int cy = 0; cy < VDIM; ++cy)
-      {
-         if (tidz == 0)
-         {
-            MFEM_FOREACH_THREAD(q, x, Q1D)
-            {
-               MFEM_FOREACH_THREAD(d, y, D1D)
-               {
-                  B[q][d] = b(q, d);
-                  G[q][d] = g(q, d);
-               }
-            }
-         }
-         MFEM_FOREACH_THREAD(qz, z, Q1D)
-         {
-            MFEM_FOREACH_THREAD(qy, y, Q1D)
-            {
-               MFEM_FOREACH_THREAD(qx, x, Q1D) { Z[qz][qy][qx] = 0.0; }
-            }
-         }
-         MFEM_SYNC_THREAD;
-         for (int c = 0; c < VDIM; ++c)
-         {
-            MFEM_FOREACH_THREAD(dz, z, D1D)
-            {
-               MFEM_FOREACH_THREAD(dy, y, D1D)
-               {
-                  MFEM_FOREACH_THREAD(dx, x, D1D)
-                  {
-                     X[dz][dy][dx] = x(dx, dy, dz, cy, e);
-                     U[0][dz][dy][dx] = x(dx, dy, dz, c, e);
-                  }
-               }
-            }
-            MFEM_SYNC_THREAD;
-            MFEM_FOREACH_THREAD(dz, z, D1D)
-            {
-               MFEM_FOREACH_THREAD(dy, y, D1D)
-               {
-                  MFEM_FOREACH_THREAD(qx, x, Q1D)
-                  {
-                     real_t u = 0.0;
-                     real_t v = 0.0;
-                     real_t z = 0.0;
-                     for (int dx = 0; dx < D1D; ++dx)
-                     {
-                        const real_t coord = X[dz][dy][dx];
-                        const real_t value = U[0][dz][dy][dx];
-                        u += coord * B[qx][dx];
-                        v += coord * G[qx][dx];
-                        z += value * B[qx][dx];
-                     }
-                     DDQ0[dz][dy][qx] = u;
-                     DDQ1[dz][dy][qx] = v;
-                     U[1][dz][dy][qx] = z;
-                  }
-               }
-            }
-            MFEM_SYNC_THREAD;
-            MFEM_FOREACH_THREAD(dz, z, D1D)
-            {
-               MFEM_FOREACH_THREAD(qy, y, Q1D)
-               {
-                  MFEM_FOREACH_THREAD(qx, x, Q1D)
-                  {
-                     real_t u = 0.0;
-                     real_t v = 0.0;
-                     real_t w = 0.0;
-                     real_t z = 0.0;
-                     for (int dy = 0; dy < D1D; ++dy)
-                     {
-                        u += DDQ1[dz][dy][qx] * B[qy][dy];
-                        v += DDQ0[dz][dy][qx] * G[qy][dy];
-                        w += DDQ0[dz][dy][qx] * B[qy][dy];
-                        z += U[1][dz][dy][qx] * B[qy][dy];
-                     }
-                     DQQ0[dz][qy][qx] = u;
-                     DQQ1[dz][qy][qx] = v;
-                     DQQ2[dz][qy][qx] = w;
-                     U[0][dz][qy][qx] = z;
-                  }
-               }
-            }
-            MFEM_SYNC_THREAD;
-            MFEM_FOREACH_THREAD(qz, z, Q1D)
-            {
-               MFEM_FOREACH_THREAD(qy, y, Q1D)
-               {
-                  MFEM_FOREACH_THREAD(qx, x, Q1D)
-                  {
-                     real_t u = 0.0;
-                     real_t v = 0.0;
-                     real_t w = 0.0;
-                     real_t z = 0.0;
-                     for (int dz = 0; dz < D1D; ++dz)
-                     {
-                        u += DQQ0[dz][qy][qx] * B[qz][dz];
-                        v += DQQ1[dz][qy][qx] * B[qz][dz];
-                        w += DQQ2[dz][qy][qx] * G[qz][dz];
-                        z += U[0][dz][qy][qx] * B[qz][dz];
-                     }
-                     QQQ0[qz][qy][qx] = u;
-                     QQQ1[qz][qy][qx] = v;
-                     QQQ2[qz][qy][qx] = w;
-                     U[1][qz][qy][qx] = z;
-                  }
-               }
-            }
-            MFEM_SYNC_THREAD;
-            MFEM_FOREACH_THREAD(qz, z, Q1D)
-            {
-               MFEM_FOREACH_THREAD(qy, y, Q1D)
-               {
-                  MFEM_FOREACH_THREAD(qx, x, Q1D)
-                  {
-                     const int q = qx + (qy + qz * Q1D) * Q1D;
-                     const real_t z = U[1][qz][qy][qx];
-                     const real_t gX = QQQ0[qz][qy][qx];
-                     const real_t gY = QQQ1[qz][qy][qx];
-                     const real_t gZ = QQQ2[qz][qy][qx];
-                     const real_t d = gX * D(q, 0, c, e) + gY * D(q, 1, c, e)
-                                      + gZ * D(q, 2, c, e);
-                     Z[qz][qy][qx] += z * d;
-                  }
-               }
-            }
-            MFEM_SYNC_THREAD;
-         } // for each conv component
-         if (tidz == 0)
-         {
-            MFEM_FOREACH_THREAD(d, y, D1D)
-            {
-               MFEM_FOREACH_THREAD(q, x, Q1D) { Bt[d][q] = b(q, d); }
-            }
-         }
-         MFEM_SYNC_THREAD;
-         MFEM_FOREACH_THREAD(qz, z, Q1D)
-         {
-            MFEM_FOREACH_THREAD(qy, y, Q1D)
-            {
-               MFEM_FOREACH_THREAD(dx, x, D1D)
-               {
-                  real_t u = 0.0;
-                  for (int qx = 0; qx < Q1D; ++qx)
-                  {
-                     u += Z[qz][qy][qx] * Bt[dx][qx];
-                  }
-                  QQD0[qz][qy][dx] = u;
-               }
-            }
-         }
-         MFEM_SYNC_THREAD;
-         MFEM_FOREACH_THREAD(qz, z, Q1D)
-         {
-            MFEM_FOREACH_THREAD(dy, y, D1D)
-            {
-               MFEM_FOREACH_THREAD(dx, x, D1D)
-               {
-                  real_t u = 0.0;
-                  for (int qy = 0; qy < Q1D; ++qy)
-                  {
-                     u += QQD0[qz][qy][dx] * Bt[dy][qy];
-                  }
-                  QDD0[qz][dy][dx] = u;
-               }
-            }
-         }
-         MFEM_SYNC_THREAD;
-         MFEM_FOREACH_THREAD(dz, z, D1D)
-         {
-            MFEM_FOREACH_THREAD(dy, y, D1D)
-            {
-               MFEM_FOREACH_THREAD(dx, x, D1D)
-               {
-                  real_t u = 0.0;
-                  for (int qz = 0; qz < Q1D; ++qz)
-                  {
-                     u += QDD0[qz][dy][dx] * Bt[dz][qz];
-                  }
-                  Y(dx, dy, dz, cy, e) += u;
-               }
-            }
-         }
-         MFEM_SYNC_THREAD;
-      }
-   });
+      MFEM_ABORT("dim " << dim << " not supported!");
+   }
 }
 
 void VectorConvectionNLFIntegrator::AddMultPA(const Vector &x, Vector &y) const
@@ -812,26 +197,13 @@ void VectorConvectionNLFIntegrator::AddMultPA(const Vector &x, Vector &y) const
    }
    else
    {
-      const int NE = ne;
-      const int D1D = maps->ndof;
-      const int Q1D = maps->nqpt;
-      const Vector &QV = pa_data;
-      const Array<real_t> &B = maps->B;
-      const Array<real_t> &G = maps->G;
-      const Array<real_t> &Bt = maps->Bt;
-      if (dim == 2)
-      {
-         return PAConvectionNLApply2D(NE, B, G, Bt, QV, x, y, D1D, Q1D);
-      }
-      if (dim == 3)
-      {
-         constexpr int T_MAX_D1D = 8;
-         constexpr int T_MAX_Q1D = 8;
-         MFEM_VERIFY(D1D <= T_MAX_D1D && Q1D <= T_MAX_Q1D, "Not yet implemented!");
-         return SmemPAConvectionNLApply3D<0, 0, T_MAX_D1D, T_MAX_Q1D>
-                (NE, B, G, QV, x, y, D1D, Q1D);
-      }
-      MFEM_ABORT("Not yet implemented!");
+      AddMultPAKernels::Run(dim, d1d, q1d, ne,
+                            maps->B.Read(),
+                            maps->G.Read(),
+                            pa_adj.Read(),
+                            x.Read(),
+                            y.ReadWrite(),
+                            d1d, q1d);
    }
 }
 

fem/integ/nonlininteg_vecconvection_pa.hpp

@@ -0,0 +1,209 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+#pragma once
+
+#include "../../config/config.hpp"
+#include "../../general/forall.hpp"
+#include "../../linalg/dtensor.hpp"
+#include "../kernels.hpp"
+#include "../nonlininteg.hpp"
+
+namespace mfem
+{
+
+/// \cond DO_NOT_DOCUMENT
+
+namespace internal
+{
+
+// PA Convection NL 2D kernel
+template<int T_D1D = 0, int T_Q1D = 0>
+inline void SmemPAConvectionNLApply2D(const int NE,
+                                      const real_t *b,
+                                      const real_t *g,
+                                      const real_t *a,
+                                      const real_t *x,
+                                      real_t *y,
+                                      const int d1d = 0,
+                                      const int q1d = 0)
+{
+   static constexpr int VDIM = 2, DIM = 2;
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+
+   const auto B = Reshape(b, Q1D, D1D);
+   const auto G = Reshape(g, Q1D, D1D);
+   const auto A = Reshape(a, VDIM, DIM, Q1D, Q1D, NE);
+   const auto X = Reshape(x, D1D, D1D, VDIM, NE);
+   auto Y = Reshape(y, D1D, D1D, VDIM, NE);
+
+   mfem::forall_2D<T_Q1D * T_Q1D>(NE, Q1D, Q1D, [=] MFEM_HOST_DEVICE(int e)
+   {
+      constexpr int MD1 = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
+      constexpr int MQ1 = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+
+      MFEM_SHARED real_t smem[MQ1][MQ1], sB[MD1][MQ1], sG[MD1][MQ1];
+
+      kernels::internal::vd_regs2d_t<VDIM, DIM, MQ1> g0, g1;
+      kernels::internal::v_regs2d_t<VDIM, MQ1> r0, r1;
+      kernels::internal::v_regs2d_t<VDIM, MQ1> s0, s1;
+
+      kernels::internal::LoadMatrix(D1D, Q1D, B, sB);
+      kernels::internal::LoadMatrix(D1D, Q1D, G, sG);
+
+      kernels::internal::LoadDofs2d(e, D1D, X, r0);
+      kernels::internal::Eval2d(D1D, Q1D, smem, sB, r0, r1); // u vector-value
+      kernels::internal::LoadDofs2d(e, D1D, X, g0);
+      kernels::internal::Grad2d(D1D, Q1D, smem, sB, sG, g0, g1); // u vector-gradient
+
+      MFEM_FOREACH_THREAD_DIRECT(qy, y, Q1D)
+      {
+         MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+         {
+            const future::tensor<real_t, 2> U =
+            {
+               r1[0][qy][qx], r1[1][qy][qx]
+            };
+            const future::tensor<real_t, 2,2> gradU = {{
+                  {g1[0][0][qy][qx], g1[1][0][qy][qx]},
+                  {g1[0][1][qy][qx], g1[1][1][qy][qx]},
+               }
+            };
+            const future::tensor<real_t, 2,2> Q = {{
+                  {A(0,0,qx,qy,e), A(1,0,qx,qy,e)},
+                  {A(0,1,qx,qy,e), A(1,1,qx,qy,e)},
+               }
+            };
+            const future::tensor<real_t, 2> conv = transpose(gradU) * (Q * U);
+            s0[0][qy][qx] = conv[0];
+            s0[1][qy][qx] = conv[1];
+         }
+      }
+      MFEM_SYNC_THREAD;
+      kernels::internal::EvalTranspose2d(D1D, Q1D, smem, sB, s0, s1);
+      kernels::internal::WriteDofs2d(e, D1D, s1, Y);
+   });
+}
+
+// PA Convection NL 3D kernel
+template<int T_D1D = 0, int T_Q1D = 0>
+inline void SmemPAConvectionNLApply3D(const int NE,
+                                      const real_t *b,
+                                      const real_t *g,
+                                      const real_t *a,
+                                      const real_t *x,
+                                      real_t *y,
+                                      const int d1d = 0,
+                                      const int q1d = 0)
+{
+   static constexpr int VDIM = 3, DIM = 3;
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+
+   const auto B = Reshape(b, Q1D, D1D);
+   const auto G = Reshape(g, Q1D, D1D);
+   const auto A = Reshape(a, VDIM, DIM, Q1D, Q1D, Q1D, NE);
+   const auto X = Reshape(x, D1D, D1D, D1D, VDIM, NE);
+   auto Y = Reshape(y, D1D, D1D, D1D, VDIM, NE);
+
+   mfem::forall_2D<T_Q1D*T_Q1D>(NE, Q1D, Q1D, [=] MFEM_HOST_DEVICE(int e)
+   {
+      constexpr int MD1 = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
+      constexpr int MQ1 = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+
+      MFEM_SHARED real_t smem[MQ1][MQ1], sB[MD1][MQ1], sG[MD1][MQ1];
+
+      kernels::internal::vd_regs3d_t<VDIM, DIM, MQ1> g0, g1;
+      kernels::internal::v_regs3d_t<VDIM, MQ1> r0, r1;
+      kernels::internal::v_regs3d_t<VDIM, MQ1> s0, s1;
+
+      kernels::internal::LoadMatrix(D1D, Q1D, B, sB);
+      kernels::internal::LoadMatrix(D1D, Q1D, G, sG);
+
+      kernels::internal::LoadDofs3d(e, D1D, X, r0);
+      kernels::internal::Eval3d(D1D, Q1D, smem, sB, r0, r1); // u vector-value
+      kernels::internal::LoadDofs3d(e, D1D, X, g0);
+      kernels::internal::Grad3d(D1D, Q1D, smem, sB, sG, g0, g1); // u vector-gradient
+
+      for (int qz = 0; qz < Q1D; qz++)
+      {
+         MFEM_FOREACH_THREAD_DIRECT(qy, y, Q1D)
+         {
+            MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+            {
+               const future::tensor<real_t, 3> U =
+               {
+                  r1[0][qz][qy][qx], r1[1][qz][qy][qx], r1[2][qz][qy][qx]
+               };
+               const future::tensor<real_t, 3,3> gradU = {{
+                     {g1[0][0][qz][qy][qx], g1[1][0][qz][qy][qx], g1[2][0][qz][qy][qx]},
+                     {g1[0][1][qz][qy][qx], g1[1][1][qz][qy][qx], g1[2][1][qz][qy][qx]},
+                     {g1[0][2][qz][qy][qx], g1[1][2][qz][qy][qx], g1[2][2][qz][qy][qx]}
+                  }
+               };
+               const future::tensor<real_t, 3,3> Q = {{
+                     {A(0,0,qx,qy,qz,e), A(1,0,qx,qy,qz,e), A(2,0,qx,qy,qz,e)},
+                     {A(0,1,qx,qy,qz,e), A(1,1,qx,qy,qz,e), A(2,1,qx,qy,qz,e)},
+                     {A(0,2,qx,qy,qz,e), A(1,2,qx,qy,qz,e), A(2,2,qx,qy,qz,e)}
+                  }
+               };
+               const future::tensor<real_t, 3> conv = transpose(gradU) * (Q * U);
+               s0[0][qz][qy][qx] = conv[0];
+               s0[1][qz][qy][qx] = conv[1];
+               s0[2][qz][qy][qx] = conv[2];
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      kernels::internal::EvalTranspose3d(D1D, Q1D, smem, sB, s0, s1);
+      kernels::internal::WriteDofs3d(e, D1D, s1, Y);
+   });
+}
+
+} // namespace internal
+
+template<int DIM, int T_D1D, int T_Q1D>
+VectorConvectionNLFIntegrator::AddMultPAType
+VectorConvectionNLFIntegrator::AddMultPAKernels::Kernel()
+{
+   static_assert(T_D1D <= T_Q1D, "d1d > q1d is not supported");
+   if constexpr (DIM == 2)
+   {
+      return internal::SmemPAConvectionNLApply2D<T_D1D, T_Q1D>;
+   }
+   else if constexpr (DIM == 3)
+   {
+      return internal::SmemPAConvectionNLApply3D<T_D1D, T_Q1D>;
+   }
+   else { MFEM_ABORT("Unsupported kernel"); }
+}
+
+inline VectorConvectionNLFIntegrator::AddMultPAType
+VectorConvectionNLFIntegrator::AddMultPAKernels::Fallback
+(int dim, int d1d, int q1d)
+{
+   MFEM_VERIFY(d1d <= q1d, "d1d > q1d is not supported");
+   MFEM_VERIFY(d1d <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(q1d <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+   if (dim == 2)
+   {
+      return internal::SmemPAConvectionNLApply2D<>;
+   }
+   else if (dim == 3)
+   {
+      return internal::SmemPAConvectionNLApply3D<>;
+   }
+   else { MFEM_ABORT("Unsupported kernel"); }
+}
+
+/// \endcond DO_NOT_DOCUMENT
+
+} // namespace mfem

fem/integ/nonlininteg_vecconvection_pa_diag.cpp

@@ -0,0 +1,50 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#include "../ceed/interface/util.hpp"
+#include "./nonlininteg_vecconvection_pa_diag.hpp" // IWYU pragma: keep
+
+namespace mfem
+{
+
+void VectorConvectionNLFIntegrator::AssembleGradDiagonalPA(Vector &de) const
+{
+   MFEM_VERIFY(!DeviceCanUseCeed(),
+               "VectorConvectionNLFIntegrator PA gradients are not supported "
+               "with the libCEED backend");
+
+   if (dim == 2)
+   {
+      GradDiagPA2D::Run(d1d, q1d, ne,
+                        maps->B.Read(),
+                        maps->G.Read(),
+                        pa_adj.Read(),
+                        pa_u.Read(),
+                        de.ReadWrite(),
+                        d1d, q1d);
+   }
+   else if (dim == 3)
+   {
+      GradDiagPA3D::Run(d1d, q1d, ne,
+                        maps->B.Read(),
+                        maps->G.Read(),
+                        pa_adj.Read(),
+                        pa_u.Read(),
+                        de.ReadWrite(),
+                        d1d, q1d);
+   }
+   else
+   {
+      MFEM_ABORT("Unsupported dimension");
+   }
+}
+
+} // namespace mfem

fem/integ/nonlininteg_vecconvection_pa_diag.hpp

@@ -0,0 +1,302 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+#pragma once
+
+#include "../../config/config.hpp"
+#include "../../general/forall.hpp"
+#include "../../linalg/dtensor.hpp"
+#include "../kernels.hpp"
+#include "../nonlininteg.hpp"
+
+namespace mfem
+{
+
+/// \cond DO_NOT_DOCUMENT
+
+namespace internal
+{
+
+template<int T_D1D = 0, int T_Q1D = 0>
+inline void SmemPAConvectionNLGradDiagonal2D(const int NE,
+                                             const real_t *b,
+                                             const real_t *g,
+                                             const real_t *a,
+                                             const real_t *u,
+                                             real_t *de,
+                                             const int d1d,
+                                             const int q1d)
+{
+   static constexpr int VDIM = 2, DIM = 2;
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+
+   const auto A = Reshape(a, VDIM, DIM, Q1D, Q1D, NE);
+   const auto U = Reshape(u, D1D, D1D, VDIM, NE);
+   auto D = Reshape(de, D1D, D1D, VDIM, NE);
+
+   mfem::forall_2D<T_Q1D * T_Q1D>(NE, Q1D, Q1D, [=] MFEM_HOST_DEVICE(int e)
+   {
+      constexpr int MD1 = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
+      constexpr int MQ1 = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+
+      MFEM_SHARED real_t sM[3][MQ1][MQ1], sQ[3][MQ1][MQ1];
+      MFEM_SHARED real_t sB[MD1][MQ1], sG[MD1][MQ1];
+
+      kernels::internal::v_regs2d_t<VDIM, MQ1> r0, r1;
+      kernels::internal::vd_regs2d_t<VDIM, DIM, MQ1> g0, g1;
+
+      kernels::internal::LoadMatrix(D1D, Q1D, b, sB);
+      kernels::internal::LoadMatrix(D1D, Q1D, g, sG);
+
+      kernels::internal::LoadDofs2d(e, D1D, U, r0);
+      kernels::internal::Eval2d(D1D, Q1D, sM[0], sB, r0, r1);
+
+      kernels::internal::LoadDofs2d(e, D1D, U, g0);
+      kernels::internal::Grad2d(D1D, Q1D, sM[0], sB, sG, g0, g1);
+
+      for (int v = 0; v < VDIM; ++v)
+      {
+         future::tensor<real_t, VDIM> e_v = {};
+         e_v[v] = real_t(1);
+         MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+         {
+            MFEM_FOREACH_THREAD_DIRECT(qy, y, Q1D)
+            {
+               const future::tensor<real_t, VDIM> u_val =
+               {
+                  r1[0][qy][qx], r1[1][qy][qx]
+               };
+               const future::tensor<real_t, VDIM, DIM> Q_adj =
+               {
+                  {  { A(0, 0, qx, qy, e), A(1, 0, qx, qy, e) },
+                     { A(0, 1, qx, qy, e), A(1, 1, qx, qy, e) }
+                  }
+               };
+               const future::tensor<real_t, VDIM, DIM> grad_U =
+               {
+                  {  { g1[0][0][qy][qx], g1[1][0][qy][qx] },
+                     { g1[0][1][qy][qx], g1[1][1][qy][qx] }
+                  }
+               };
+               const auto one = Q_adj * u_val;
+               const auto two = transpose(grad_U) * (Q_adj * e_v);
+               sQ[0][qx][qy] = one[0];
+               sQ[1][qx][qy] = one[1];
+               sQ[2][qx][qy] = two[v];
+            }
+         }
+         MFEM_SYNC_THREAD;
+
+         MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+         {
+            MFEM_FOREACH_THREAD_DIRECT(dy, y, D1D)
+            {
+               real_t s[3] = {};
+               for (int qy = 0; qy < Q1D; ++qy)
+               {
+                  const real_t By = sB[dy][qy], Gy = sG[dy][qy];
+                  s[0] += By * By * sQ[0][qx][qy];
+                  s[1] += Gy * By * sQ[1][qx][qy];
+                  s[2] += By * By * sQ[2][qx][qy];
+               }
+               sM[0][qx][dy] = s[0];
+               sM[1][qx][dy] = s[1];
+               sM[2][qx][dy] = s[2];
+            }
+         }
+         MFEM_SYNC_THREAD;
+
+         MFEM_FOREACH_THREAD_DIRECT(dy, y, D1D)
+         {
+            MFEM_FOREACH_THREAD_DIRECT(dx, x, D1D)
+            {
+               real_t d = 0.0;
+               for (int qx = 0; qx < Q1D; ++qx)
+               {
+                  const real_t Bx = sB[dx][qx], Gx = sG[dx][qx];
+                  d += Gx * Bx * sM[0][qx][dy] +
+                       Bx * Bx * sM[1][qx][dy] +
+                       Bx * Bx * sM[2][qx][dy];
+               }
+               D(dx, dy, v, e) += d;
+            }
+         }
+         MFEM_SYNC_THREAD;
+      }
+   });
+}
+
+template<int T_D1D = 0, int T_Q1D = 0>
+inline void SmemPAConvectionNLGradDiagonal3D(const int NE,
+                                             const real_t *b,
+                                             const real_t *g,
+                                             const real_t *a,
+                                             const real_t *u,
+                                             real_t *de,
+                                             const int d1d,
+                                             const int q1d)
+{
+   static constexpr int VDIM = 3, DIM = 3;
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+
+   const auto A = Reshape(a, VDIM, DIM, Q1D, Q1D, Q1D, NE);
+   const auto U = Reshape(u, D1D, D1D, D1D, VDIM, NE);
+   auto D = Reshape(de, D1D, D1D, D1D, VDIM, NE);
+
+   mfem::forall_2D<T_Q1D * T_Q1D>(NE, Q1D, Q1D, [=] MFEM_HOST_DEVICE(int e)
+   {
+      constexpr int MD1 = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
+      constexpr int MQ1 = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+
+      MFEM_SHARED real_t sM[4][MQ1][MQ1], sQ[4][MQ1][MQ1];
+      MFEM_SHARED real_t sB[MD1][MQ1], sG[MD1][MQ1];
+
+      kernels::internal::v_regs3d_t<VDIM, MQ1> r0, r1;
+      kernels::internal::vd_regs3d_t<VDIM, DIM, MQ1> g0, g1;
+
+      kernels::internal::LoadMatrix(D1D, Q1D, b, sB);
+      kernels::internal::LoadMatrix(D1D, Q1D, g, sG);
+
+      kernels::internal::LoadDofs3d(e, D1D, U, r0);
+      kernels::internal::Eval3d(D1D, Q1D, sM[0], sB, r0, r1);
+
+      kernels::internal::LoadDofs3d(e, D1D, U, g0);
+      kernels::internal::Grad3d(D1D, Q1D, sM[0], sB, sG, g0, g1);
+
+      for (int v = 0; v < VDIM; ++v)
+      {
+         future::tensor<real_t, VDIM> e_v = {};
+         e_v[v] = real_t(1);
+         for (int dz = 0; dz < D1D; ++dz)
+         {
+            MFEM_FOREACH_THREAD_DIRECT(qy, y, Q1D)
+            {
+               MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+               {
+                  real_t s[4] = {};
+                  for (int qz = 0; qz < Q1D; ++qz)
+                  {
+                     const future::tensor<real_t, VDIM> u_val =
+                     {
+                        r1[0][qz][qy][qx], r1[1][qz][qy][qx], r1[2][qz][qy][qx]
+                     };
+                     const future::tensor<real_t, VDIM, DIM> Q_adj = {{
+                           {A(0,0,qx,qy,qz,e), A(1,0,qx,qy,qz,e), A(2,0,qx,qy,qz,e)},
+                           {A(0,1,qx,qy,qz,e), A(1,1,qx,qy,qz,e), A(2,1,qx,qy,qz,e)},
+                           {A(0,2,qx,qy,qz,e), A(1,2,qx,qy,qz,e), A(2,2,qx,qy,qz,e)}
+                        }
+                     };
+                     const future::tensor<real_t, VDIM, DIM> grad_U = {{
+                           {g1[0][0][qz][qy][qx], g1[1][0][qz][qy][qx], g1[2][0][qz][qy][qx]},
+                           {g1[0][1][qz][qy][qx], g1[1][1][qz][qy][qx], g1[2][1][qz][qy][qx]},
+                           {g1[0][2][qz][qy][qx], g1[1][2][qz][qy][qx], g1[2][2][qz][qy][qx]}
+                        }
+                     };
+                     const auto one = Q_adj * u_val;
+                     const auto two = transpose(grad_U) * (Q_adj * e_v);
+
+                     const real_t Bz = sB[dz][qz], Gz = sG[dz][qz];
+                     s[0] += one[0] * Bz * Bz;
+                     s[1] += one[1] * Bz * Bz;
+                     s[2] += one[2] * Bz * Gz;
+                     s[3] += two[v] * Bz * Bz;
+                  }
+                  sQ[0][qx][qy] = s[0];
+                  sQ[1][qx][qy] = s[1];
+                  sQ[2][qx][qy] = s[2];
+                  sQ[3][qx][qy] = s[3];
+               }
+            }
+            MFEM_SYNC_THREAD;
+
+            MFEM_FOREACH_THREAD_DIRECT(dy, y, D1D)
+            {
+               MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+               {
+                  real_t s[4] = {};
+                  for (int qy = 0; qy < Q1D; ++qy)
+                  {
+                     const real_t By = sB[dy][qy], Gy = sG[dy][qy];
+                     s[0] += By * By * sQ[0][qx][qy];
+                     s[1] += Gy * By * sQ[1][qx][qy];
+                     s[2] += By * By * sQ[2][qx][qy];
+                     s[3] += By * By * sQ[3][qx][qy];
+                  }
+                  sM[0][dy][qx] = s[0];
+                  sM[1][dy][qx] = s[1];
+                  sM[2][dy][qx] = s[2];
+                  sM[3][dy][qx] = s[3];
+               }
+            }
+            MFEM_SYNC_THREAD;
+
+            MFEM_FOREACH_THREAD_DIRECT(dy, y, D1D)
+            {
+               MFEM_FOREACH_THREAD_DIRECT(dx, x, D1D)
+               {
+                  real_t d = 0.0;
+                  for (int qx = 0; qx < Q1D; ++qx)
+                  {
+                     const real_t Bx = sB[dx][qx], Gx = sG[dx][qx];
+                     d += Gx * Bx * sM[0][dy][qx];
+                     d += Bx * Bx * sM[1][dy][qx];
+                     d += Bx * Bx * sM[2][dy][qx];
+                     d += Bx * Bx * sM[3][dy][qx];
+                  }
+                  D(dx, dy, dz, v, e) += d;
+               }
+            }
+            MFEM_SYNC_THREAD;
+         }
+      }
+   });
+}
+
+} // namespace internal
+
+template<int T_D1D, int T_Q1D>
+VectorConvectionNLFIntegrator::GradDiagPAType
+VectorConvectionNLFIntegrator::GradDiagPA2D::Kernel()
+{
+   static_assert(T_D1D <= T_Q1D, "d1d > q1d is not supported");
+   return internal::SmemPAConvectionNLGradDiagonal2D<T_D1D, T_Q1D>;
+}
+
+inline VectorConvectionNLFIntegrator::GradDiagPAType
+VectorConvectionNLFIntegrator::GradDiagPA2D::Fallback(int d1d, int q1d)
+{
+   MFEM_VERIFY(d1d <= q1d, "d1d > q1d is not supported");
+   MFEM_VERIFY(d1d <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(q1d <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+   return internal::SmemPAConvectionNLGradDiagonal2D<>;
+}
+
+template<int T_D1D, int T_Q1D>
+VectorConvectionNLFIntegrator::GradDiagPAType
+VectorConvectionNLFIntegrator::GradDiagPA3D::Kernel()
+{
+   static_assert(T_D1D <= T_Q1D, "d1d > q1d is not supported");
+   return internal::SmemPAConvectionNLGradDiagonal3D<T_D1D, T_Q1D>;
+}
+
+inline VectorConvectionNLFIntegrator::GradDiagPAType
+VectorConvectionNLFIntegrator::GradDiagPA3D::Fallback(int d1d, int q1d)
+{
+   MFEM_VERIFY(d1d <= q1d, "d1d > q1d is not supported");
+   MFEM_VERIFY(d1d <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(q1d <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+   return internal::SmemPAConvectionNLGradDiagonal3D<>;
+}
+
+/// \endcond DO_NOT_DOCUMENT
+
+} // namespace mfem

fem/integ/nonlininteg_vecconvection_pa_grad.cpp

@@ -0,0 +1,64 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#include "../ceed/interface/util.hpp"
+#include "./nonlininteg_vecconvection_pa_grad.hpp" // IWYU pragma: keep
+
+namespace mfem
+{
+
+void VectorConvectionNLFIntegrator::AssembleGradPA(
+   const Vector &u, const FiniteElementSpace &fes)
+{
+   MFEM_VERIFY(!DeviceCanUseCeed(),
+               "VectorConvectionNLFIntegrator PA gradients are not supported "
+               "with the libCEED backend");
+
+   this->pa_u = u;
+   AssemblePA(fes);
+}
+
+void VectorConvectionNLFIntegrator::AddMultGradPA(const Vector &x,
+                                                  Vector &y) const
+{
+   MFEM_VERIFY(!DeviceCanUseCeed(),
+               "VectorConvectionNLFIntegrator PA gradients are not supported "
+               "with the libCEED backend");
+
+   if (dim == 2)
+   {
+      AddMultGradPA2D::Run(d1d, q1d, ne,
+                           maps->B.Read(),
+                           maps->G.Read(),
+                           pa_adj.Read(),
+                           pa_u.Read(),
+                           x.Read(),
+                           y.ReadWrite(),
+                           d1d, q1d);
+   }
+   else if (dim == 3)
+   {
+      AddMultGradPA3D::Run(d1d, q1d, ne,
+                           maps->B.Read(),
+                           maps->G.Read(),
+                           pa_adj.Read(),
+                           pa_u.Read(),
+                           x.Read(),
+                           y.ReadWrite(),
+                           d1d, q1d);
+   }
+   else
+   {
+      MFEM_ABORT("Unsupported dimension");
+   }
+}
+
+} // namespace mfem

fem/integ/nonlininteg_vecconvection_pa_grad.hpp

@@ -0,0 +1,257 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+#pragma once
+
+#include "../../config/config.hpp"
+#include "../../general/forall.hpp"
+#include "../../linalg/dtensor.hpp"
+#include "../kernels.hpp"
+#include "../nonlininteg.hpp"
+
+namespace mfem
+{
+
+/// \cond DO_NOT_DOCUMENT
+
+namespace internal
+{
+
+template<int T_D1D = 0, int T_Q1D = 0>
+inline void SmemPAConvectionNLGradApply2D(const int ne,
+                                          const real_t *b,
+                                          const real_t *g,
+                                          const real_t *a,
+                                          const real_t *u,
+                                          const real_t *du,
+                                          real_t *y,
+                                          const int d1d,
+                                          const int q1d)
+{
+   static constexpr int VDIM = 2, DIM = 2;
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+
+   const auto A = Reshape(a, VDIM, DIM, Q1D, Q1D, ne);
+   const auto U = Reshape(u, D1D, D1D, VDIM, ne);
+   const auto dU = Reshape(du, D1D, D1D, VDIM, ne);
+   auto Y = Reshape(y, D1D, D1D, VDIM, ne);
+
+   mfem::forall_2D<T_Q1D * T_Q1D>(ne, Q1D, Q1D, [=] MFEM_HOST_DEVICE(int e)
+   {
+      constexpr int MD1 = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
+      constexpr int MQ1 = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+
+      MFEM_SHARED real_t smem[MQ1][MQ1];
+      MFEM_SHARED real_t sB[MD1][MQ1], sG[MD1][MQ1];
+
+      kernels::internal::vd_regs2d_t<VDIM, DIM, MQ1> g0, g1, g2;
+      kernels::internal::v_regs2d_t<DIM, MQ1> r0, r1, r2;
+
+      kernels::internal::LoadMatrix(D1D, Q1D, b, sB);
+      kernels::internal::LoadMatrix(D1D, Q1D, g, sG);
+
+      kernels::internal::LoadDofs2d(e, D1D, dU, g0);
+      kernels::internal::Grad2d(D1D, Q1D, smem, sB, sG, g0, g1); // δu gradient
+
+      kernels::internal::LoadDofs2d(e, D1D, U, r0);
+      kernels::internal::Eval2d(D1D, Q1D, smem, sB, r0, r2); // u value
+
+      kernels::internal::LoadDofs2d(e, D1D, dU, r0);
+      kernels::internal::Eval2d(D1D, Q1D, smem, sB, r0, r1); // δu value
+
+      kernels::internal::LoadDofs2d(e, D1D, U, g0);
+      kernels::internal::Grad2d(D1D, Q1D, smem, sB, sG, g0, g2); // u gradient
+
+      MFEM_FOREACH_THREAD_DIRECT(qy, y, Q1D)
+      {
+         MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+         {
+            // First part of the Jacobian: u·∇δu
+            const future::tensor<real_t, DIM> u_val =
+            {
+               r2[0][qy][qx], r2[1][qy][qx]
+            };
+            const future::tensor<real_t, VDIM, DIM> Q_adj =
+            {
+               {  { A(0, 0, qx, qy, e), A(1, 0, qx, qy, e) },
+                  { A(0, 1, qx, qy, e), A(1, 1, qx, qy, e) }
+               }
+            };
+            const future::tensor<real_t, VDIM, DIM> grad_dU =
+            {
+               {  { g1[0][0][qy][qx], g1[1][0][qy][qx] },
+                  { g1[0][1][qy][qx], g1[1][1][qy][qx] }
+               }
+            };
+            const auto one = transpose(grad_dU) * (Q_adj * u_val);
+
+            // Second part of the Jacobian: δu·∇u
+            const future::tensor<real_t, DIM> du_val =
+            {
+               r1[0][qy][qx], r1[1][qy][qx]
+            };
+            const future::tensor<real_t, VDIM, DIM> grad_U =
+            {
+               {  { g2[0][0][qy][qx], g2[1][0][qy][qx] },
+                  { g2[0][1][qy][qx], g2[1][1][qy][qx] }
+               }
+            };
+            const auto two = transpose(grad_U) * (Q_adj * du_val);
+
+            // u⋅∇δu + δu⋅∇u
+            r0[0][qy][qx] = one[0] + two[0];
+            r0[1][qy][qx] = one[1] + two[1];
+         }
+      }
+      MFEM_SYNC_THREAD;
+      kernels::internal::EvalTranspose2d(D1D, Q1D, smem, sB, r0, r1);
+      kernels::internal::WriteDofs2d(e, D1D, r1, Y);
+   });
+}
+
+template<int T_D1D = 0, int T_Q1D = 0>
+inline void SmemPAConvectionNLGradApply3D(const int ne,
+                                          const real_t *b,
+                                          const real_t *g,
+                                          const real_t *a,
+                                          const real_t *u,
+                                          const real_t *du,
+                                          real_t *y,
+                                          const int d1d,
+                                          const int q1d)
+{
+   static constexpr int VDIM = 3, DIM = 3;
+   const int D1D = T_D1D ? T_D1D : d1d;
+   const int Q1D = T_Q1D ? T_Q1D : q1d;
+
+   const auto A = Reshape(a, VDIM, DIM, Q1D, Q1D, Q1D, ne);
+   const auto U = Reshape(u, D1D, D1D, D1D, VDIM, ne);
+   const auto dU = Reshape(du, D1D, D1D, D1D, VDIM, ne);
+   auto Y = Reshape(y, D1D, D1D, D1D, VDIM, ne);
+
+   mfem::forall_2D<T_Q1D * T_Q1D>(ne, Q1D, Q1D, [=] MFEM_HOST_DEVICE(int e)
+   {
+      constexpr int MD1 = T_D1D ? T_D1D : DofQuadLimits::MAX_D1D;
+      constexpr int MQ1 = T_Q1D ? T_Q1D : DofQuadLimits::MAX_Q1D;
+
+      MFEM_SHARED real_t smem[MQ1][MQ1];
+      MFEM_SHARED real_t sB[MD1][MQ1], sG[MD1][MQ1];
+
+      kernels::internal::v_regs3d_t<VDIM, MQ1> r0, r1, r2;
+      kernels::internal::vd_regs3d_t<VDIM, DIM, MQ1> g0, g1, g2;
+
+      kernels::internal::LoadMatrix(D1D, Q1D, b, sB);
+      kernels::internal::LoadMatrix(D1D, Q1D, g, sG);
+
+      kernels::internal::LoadDofs3d(e, D1D, dU, g0);
+      kernels::internal::Grad3d(D1D, Q1D, smem, sB, sG, g0, g1); // δu gradient
+
+      kernels::internal::LoadDofs3d(e, D1D, U, r0);
+      kernels::internal::Eval3d(D1D, Q1D, smem, sB, r0, r2); // u value
+
+      kernels::internal::LoadDofs3d(e, D1D, dU, r0);
+      kernels::internal::Eval3d(D1D, Q1D, smem, sB, r0, r1); // δu value
+
+      kernels::internal::LoadDofs3d(e, D1D, U, g0);
+      kernels::internal::Grad3d(D1D, Q1D, smem, sB, sG, g0, g2); // u gradient
+
+      for (int qz = 0; qz < Q1D; qz++)
+      {
+         MFEM_FOREACH_THREAD_DIRECT(qy, y, Q1D)
+         {
+            MFEM_FOREACH_THREAD_DIRECT(qx, x, Q1D)
+            {
+               // First part of the Jacobian: u·∇δu
+               const future::tensor<real_t, DIM> u_val =
+               {
+                  r2[0][qz][qy][qx],
+                  r2[1][qz][qy][qx],
+                  r2[2][qz][qy][qx]
+               };
+               const future::tensor<real_t, VDIM, DIM> Q_adj = {{
+                     {A(0,0,qx,qy,qz,e), A(1,0,qx,qy,qz,e), A(2,0,qx,qy,qz,e)},
+                     {A(0,1,qx,qy,qz,e), A(1,1,qx,qy,qz,e), A(2,1,qx,qy,qz,e)},
+                     {A(0,2,qx,qy,qz,e), A(1,2,qx,qy,qz,e), A(2,2,qx,qy,qz,e)}
+                  }
+               };
+               const future::tensor<real_t, DIM, DIM> grad_dU = {{
+                     {g1[0][0][qz][qy][qx], g1[1][0][qz][qy][qx], g1[2][0][qz][qy][qx]},
+                     {g1[0][1][qz][qy][qx], g1[1][1][qz][qy][qx], g1[2][1][qz][qy][qx]},
+                     {g1[0][2][qz][qy][qx], g1[1][2][qz][qy][qx], g1[2][2][qz][qy][qx]}
+                  }
+               };
+               const auto one = transpose(grad_dU) * (Q_adj * u_val);
+
+               // Second part of the Jacobian: δu·∇u
+               const future::tensor<real_t, DIM> du_val =
+               {
+                  r1[0][qz][qy][qx], r1[1][qz][qy][qx], r1[2][qz][qy][qx]
+               };
+               const future::tensor<real_t, VDIM, DIM> grad_U = {{
+                     {g2[0][0][qz][qy][qx], g2[1][0][qz][qy][qx], g2[2][0][qz][qy][qx]},
+                     {g2[0][1][qz][qy][qx], g2[1][1][qz][qy][qx], g2[2][1][qz][qy][qx]},
+                     {g2[0][2][qz][qy][qx], g2[1][2][qz][qy][qx], g2[2][2][qz][qy][qx]}
+                  }
+               };
+               const auto two = transpose(grad_U) * (Q_adj * du_val);
+
+               // u⋅∇δu + δu⋅∇u
+               r0[0][qz][qy][qx] = one[0] + two[0];
+               r0[1][qz][qy][qx] = one[1] + two[1];
+               r0[2][qz][qy][qx] = one[2] + two[2];
+            }
+         }
+      }
+      MFEM_SYNC_THREAD;
+      kernels::internal::EvalTranspose3d(D1D, Q1D, smem, sB, r0, r1);
+      kernels::internal::WriteDofs3d(e, D1D, r1, Y);
+   });
+}
+
+} // namespace internal
+
+template<int T_D1D, int T_Q1D>
+VectorConvectionNLFIntegrator::AddMultGradPAType
+VectorConvectionNLFIntegrator::AddMultGradPA2D::Kernel()
+{
+   static_assert(T_D1D <= T_Q1D, "d1d > q1d is not supported");
+   return internal::SmemPAConvectionNLGradApply2D<T_D1D, T_Q1D>;
+}
+
+inline VectorConvectionNLFIntegrator::AddMultGradPAType
+VectorConvectionNLFIntegrator::AddMultGradPA2D::Fallback(int d1d, int q1d)
+{
+   MFEM_VERIFY(d1d <= q1d, "d1d > q1d is not supported");
+   MFEM_VERIFY(d1d <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(q1d <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+   return internal::SmemPAConvectionNLGradApply2D<>;
+}
+
+template<int T_D1D, int T_Q1D>
+VectorConvectionNLFIntegrator::AddMultGradPAType
+VectorConvectionNLFIntegrator::AddMultGradPA3D::Kernel()
+{
+   static_assert(T_D1D <= T_Q1D, "d1d > q1d is not supported");
+   return internal::SmemPAConvectionNLGradApply3D<T_D1D, T_Q1D>;
+}
+
+inline VectorConvectionNLFIntegrator::AddMultGradPAType
+VectorConvectionNLFIntegrator::AddMultGradPA3D::Fallback(int d1d, int q1d)
+{
+   MFEM_VERIFY(d1d <= q1d, "d1d > q1d is not supported");
+   MFEM_VERIFY(d1d <= DeviceDofQuadLimits::Get().MAX_D1D, "");
+   MFEM_VERIFY(q1d <= DeviceDofQuadLimits::Get().MAX_Q1D, "");
+   return internal::SmemPAConvectionNLGradApply3D<>;
+}
+
+/// \endcond DO_NOT_DOCUMENT
+
+} // namespace mfem

fem/kernels.hpp

@@ -83,7 +83,7 @@ constexpr int SetMaxOf(int n) { return NextMultipleOf<4>(n); }
 #endif // CUDA/HIP && DEVICE_COMPILE
 
 /// Load 2D matrix into shared memory
-template <int MQ1>
+template <int MQ1, bool TRANSPOSE = false>
 inline MFEM_HOST_DEVICE void LoadMatrix(const int d1d, const int q1d,
                                         const real_t *M, real_t (*N)[MQ1])
 {
@@ -91,7 +91,14 @@ inline MFEM_HOST_DEVICE void LoadMatrix(const int d1d, const int q1d,
    {
       MFEM_FOREACH_THREAD_DIRECT(qx, x, q1d)
       {
-         N[dy][qx] = M[dy * q1d + qx];
+         if constexpr (TRANSPOSE)
+         {
+            N[dy][qx] = M[qx * d1d + dy];
+         }
+         else
+         {
+            N[dy][qx] = M[dy * q1d + qx];
+         }
       }
    }
    MFEM_SYNC_THREAD;

fem/lininteg.cpp

@@ -471,6 +471,13 @@ void VectorBoundaryLFIntegrator::AssembleRHSElementVect(
    }
 }
 
+VectorFEDomainLFIntegrator::VectorFEDomainLFIntegrator(
+   VectorCoefficient &F, const IntegrationRule *ir)
+   : DeltaLFIntegrator(F, ir), QF(F)
+{
+   static Kernels kernels{};
+}
+
 void VectorFEDomainLFIntegrator::AssembleRHSElementVect(
    const FiniteElement &el, ElementTransformation &Tr, Vector &elvect)
 {

fem/lininteg.hpp

@@ -369,8 +369,8 @@ private:
    Vector vec;
 
 public:
-   VectorFEDomainLFIntegrator(VectorCoefficient &F)
-      : DeltaLFIntegrator(F), QF(F) { }
+   VectorFEDomainLFIntegrator(VectorCoefficient &F,
+                              const IntegrationRule *ir = nullptr);
 
    void AssembleRHSElementVect(const FiniteElement &el,
                                ElementTransformation &Tr,
@@ -387,6 +387,40 @@ public:
                        Vector &b) override;
 
    using LinearFormIntegrator::AssembleRHSElementVect;
+
+   /// @param ne number of elements
+   /// @param markers array where entry markers[e] == 0 to skip assembly over
+   /// element e element
+   /// @param jac Spatial Jacobians evaluated at all quadrature points
+   /// @param weights 1D quadrature weights
+   /// @param testBO 1D open basis test functions
+   /// @param testBC 1D closed basis test functions
+   /// @param coeff coefficient values evaluated at quadrature points, possibly
+   /// compressed.
+   /// @param d number of 1D closed dofs
+   /// @param q number of 1D quadrature points
+   using AssembleKernelType = void (*)(const int NE, const Array<int> &markers,
+                                       const Vector &jac,
+                                       const Array<real_t> &weights,
+                                       const Array<real_t> &testBO,
+                                       const Array<real_t> &testBC,
+                                       const Vector &coeff, Vector &y,
+                                       const int testd1d, const int q1d);
+
+   /// parameters: test_fetype, ndims, test_d1d, q1d
+   MFEM_REGISTER_KERNELS(AssembleKernels, AssembleKernelType,
+                         (FiniteElement::DerivType, int, int, int));
+
+   struct Kernels
+   {
+      Kernels();
+   };
+
+   template <FiniteElement::DerivType TestType, int DIM, int TEST_D1D, int Q1D>
+   static void AddSpecialization()
+   {
+      AssembleKernels::Specialization<TestType, DIM, TEST_D1D, Q1D>::Add();
+   }
 };
 
 /// $ (Q, \mathrm{curl}(v))_{\Omega} $ for Nedelec Elements

fem/nonlinearform_ext.cpp

@@ -100,6 +100,17 @@ PANonlinearFormExtension::Gradient::Gradient(const PANonlinearFormExtension &e):
 
 void PANonlinearFormExtension::Gradient::AssembleGrad(const Vector &g)
 {
+   if (DeviceCanUseCeed())
+   {
+      for (int i = 0; i < ext.dnfi.Size(); ++i)
+      {
+         MFEM_VERIFY(dynamic_cast<VectorConvectionNLFIntegrator *>
+                     (ext.dnfi[i]) == nullptr,
+                     "VectorConvectionNLFIntegrator PA gradients are not supported "
+                     "with the libCEED backend");
+      }
+   }
+
    ext.elemR->Mult(g, ext.xe);
    for (int i = 0; i < ext.dnfi.Size(); ++i)
    {

fem/nonlininteg.cpp

@@ -954,4 +954,74 @@ void SkewSymmetricVectorConvectionNLFIntegrator::AssembleElementGrad(
    }
 }
 
+void ConvectiveVectorConvectionNLFIntegrator::AssemblePA(
+   const FiniteElementSpace &)
+{
+   MFEM_ABORT("ConvectiveVectorConvectionNLFIntegrator does not support "
+              "partial assembly; use VectorConvectionNLFIntegrator");
+}
+
+void ConvectiveVectorConvectionNLFIntegrator::AssembleGradPA(
+   const Vector &, const FiniteElementSpace &)
+{
+   MFEM_ABORT("ConvectiveVectorConvectionNLFIntegrator does not support "
+              "partial assembly; use VectorConvectionNLFIntegrator");
+}
+
+void ConvectiveVectorConvectionNLFIntegrator::AddMultPA(
+   const Vector &, Vector &) const
+{
+   MFEM_ABORT("ConvectiveVectorConvectionNLFIntegrator does not support "
+              "partial assembly; use VectorConvectionNLFIntegrator");
+}
+
+void ConvectiveVectorConvectionNLFIntegrator::AddMultGradPA(
+   const Vector &, Vector &) const
+{
+   MFEM_ABORT("ConvectiveVectorConvectionNLFIntegrator does not support "
+              "partial assembly; use VectorConvectionNLFIntegrator");
+}
+
+void ConvectiveVectorConvectionNLFIntegrator::AssembleGradDiagonalPA(
+   Vector &) const
+{
+   MFEM_ABORT("ConvectiveVectorConvectionNLFIntegrator does not support "
+              "partial assembly; use VectorConvectionNLFIntegrator");
+}
+
+void SkewSymmetricVectorConvectionNLFIntegrator::AssemblePA(
+   const FiniteElementSpace &)
+{
+   MFEM_ABORT("SkewSymmetricVectorConvectionNLFIntegrator does not support "
+              "partial assembly; use VectorConvectionNLFIntegrator");
+}
+
+void SkewSymmetricVectorConvectionNLFIntegrator::AssembleGradPA(
+   const Vector &, const FiniteElementSpace &)
+{
+   MFEM_ABORT("SkewSymmetricVectorConvectionNLFIntegrator does not support "
+              "partial assembly; use VectorConvectionNLFIntegrator");
+}
+
+void SkewSymmetricVectorConvectionNLFIntegrator::AddMultPA(
+   const Vector &, Vector &) const
+{
+   MFEM_ABORT("SkewSymmetricVectorConvectionNLFIntegrator does not support "
+              "partial assembly; use VectorConvectionNLFIntegrator");
+}
+
+void SkewSymmetricVectorConvectionNLFIntegrator::AddMultGradPA(
+   const Vector &, Vector &) const
+{
+   MFEM_ABORT("SkewSymmetricVectorConvectionNLFIntegrator does not support "
+              "partial assembly; use VectorConvectionNLFIntegrator");
+}
+
+void SkewSymmetricVectorConvectionNLFIntegrator::AssembleGradDiagonalPA(
+   Vector &) const
+{
+   MFEM_ABORT("SkewSymmetricVectorConvectionNLFIntegrator does not support "
+              "partial assembly; use VectorConvectionNLFIntegrator");
+}
+
 }

fem/nonlininteg.hpp

@@ -18,6 +18,7 @@
 #include "fespace.hpp"
 #include "ceed/interface/operator.hpp"
 #include "integrator.hpp"
+#include "kernel_dispatch.hpp"
 
 namespace mfem
 {
@@ -384,15 +385,17 @@ private:
    DenseMatrix dshape, dshapex, EF, gradEF, ELV, elmat_comp;
    Vector shape;
    // PA extension
-   Vector pa_data;
+   int dim, ne, nq, d1d, q1d;
+   Vector pa_adj, pa_u;
    const DofToQuad *maps;         ///< Not owned
    const GeometricFactors *geom;  ///< Not owned
-   int dim, ne, nq;
 
 public:
-   VectorConvectionNLFIntegrator(Coefficient &q): Q(&q) { }
+   struct Kernels { Kernels(); };
 
-   VectorConvectionNLFIntegrator() = default;
+   VectorConvectionNLFIntegrator(Coefficient &q): Q(&q) { static Kernels kernels; }
+
+   VectorConvectionNLFIntegrator() { static Kernels kernels; }
 
    static const IntegrationRule &GetRule(const FiniteElement &fe,
                                          const ElementTransformation &T);
@@ -411,12 +414,55 @@ public:
 
    void AssemblePA(const FiniteElementSpace &fes) override;
 
-   void AssembleMF(const FiniteElementSpace &fes) override;
+   void AssembleGradPA(const Vector &x, const FiniteElementSpace &fes) override;
 
    void AddMultPA(const Vector &x, Vector &y) const override;
 
-   void AddMultMF(const Vector &x, Vector &y) const override;
+   using AddMultPAType =
+      void(*)(const int ne, const real_t *B, const real_t *G, const real_t *A,
+              const real_t *x, real_t *y,
+              const int d1d, const int q1d);
+   MFEM_REGISTER_KERNELS(AddMultPAKernels, AddMultPAType, (int, int, int));
 
+   void AddMultGradPA(const Vector &x, Vector &y) const override;
+
+   using AddMultGradPAType =
+      void(*)(const int ne, const real_t *B, const real_t *G, const real_t *A,
+              const real_t *u, const real_t *x, real_t *y,
+              const int d1d, const int q1d);
+
+   MFEM_REGISTER_KERNELS(AddMultGradPA2D, AddMultGradPAType, (int, int));
+   MFEM_REGISTER_KERNELS(AddMultGradPA3D, AddMultGradPAType, (int, int));
+
+   void AssembleGradDiagonalPA(Vector &) const override;
+
+   using GradDiagPAType =
+      void (*)(const int ne, const real_t *B, const real_t *G, const real_t *A,
+               const real_t *u, real_t *y,
+               const int d1d, const int q1d);
+
+   MFEM_REGISTER_KERNELS(GradDiagPA2D, GradDiagPAType, (int, int));
+   MFEM_REGISTER_KERNELS(GradDiagPA3D, GradDiagPAType, (int, int));
+
+   template <int DIM, int D1D, int Q1D>
+   static void AddSpecialization()
+   {
+      AddMultPAKernels::Specialization<DIM, D1D, Q1D>::Add();
+      if constexpr (DIM == 2)
+      {
+         AddMultGradPA2D::Specialization<D1D, Q1D>::Add();
+         GradDiagPA2D::Specialization<D1D, Q1D>::Add();
+      }
+      else if constexpr (DIM == 3)
+      {
+         AddMultGradPA3D::Specialization<D1D, Q1D>::Add();
+         GradDiagPA3D::Specialization<D1D, Q1D>::Add();
+      }
+   }
+
+   void AssembleMF(const FiniteElementSpace &fes) override;
+
+   void AddMultMF(const Vector &x, Vector &y) const override;
 
 protected:
    const IntegrationRule* GetDefaultIntegrationRule(
@@ -430,7 +476,8 @@ protected:
 
 
 /** This class is used to assemble the convective form of the nonlinear term
-    arising in the Navier-Stokes equations $(u \cdot \nabla v, w )$ */
+    arising in the Navier-Stokes equations $(u \cdot \nabla v, w )$.
+    Partial assembly is not supported; use VectorConvectionNLFIntegrator. */
 class ConvectiveVectorConvectionNLFIntegrator :
    public VectorConvectionNLFIntegrator
 {
@@ -448,12 +495,20 @@ public:
                             ElementTransformation &trans,
                             const Vector &elfun,
                             DenseMatrix &elmat) override;
+
+   using NonlinearFormIntegrator::AssemblePA;
+   void AssemblePA(const FiniteElementSpace &fes) override;
+   void AssembleGradPA(const Vector &x, const FiniteElementSpace &fes) override;
+   void AddMultPA(const Vector &x, Vector &y) const override;
+   void AddMultGradPA(const Vector &x, Vector &y) const override;
+   void AssembleGradDiagonalPA(Vector &diag) const override;
 };
 
 
 /** This class is used to assemble the skew-symmetric form of the nonlinear term
     arising in the Navier-Stokes equations
-    $.5*(u \cdot \nabla v, w ) - .5*(u \cdot \nabla w, v )$ */
+    $.5*(u \cdot \nabla v, w ) - .5*(u \cdot \nabla w, v )$.
+    Partial assembly is not supported; use VectorConvectionNLFIntegrator. */
 class SkewSymmetricVectorConvectionNLFIntegrator :
    public VectorConvectionNLFIntegrator
 {
@@ -471,6 +526,13 @@ public:
                             ElementTransformation &trans,
                             const Vector &elfun,
                             DenseMatrix &elmat) override;
+
+   using NonlinearFormIntegrator::AssemblePA;
+   void AssemblePA(const FiniteElementSpace &fes) override;
+   void AssembleGradPA(const Vector &x, const FiniteElementSpace &fes) override;
+   void AddMultPA(const Vector &x, Vector &y) const override;
+   void AddMultGradPA(const Vector &x, Vector &y) const override;
+   void AssembleGradDiagonalPA(Vector &diag) const override;
 };
 
 }

fem/particleset.cpp

@@ -10,6 +10,7 @@
 // CONTRIBUTING.md for details.
 
 #include "particleset.hpp"
+#include "../general/forall.hpp"
 
 #if defined(MFEM_USE_MPI) && defined(MFEM_USE_GSLIB)
 
@@ -225,6 +226,7 @@ void ParticleSet::AddParticles(const Array<IDType> &new_ids,
       }
    }
    // Add new ids
+   ids.HostReadWrite();
    ids.Append(new_ids);
 
    // Update data
@@ -244,6 +246,102 @@ void ParticleSet::AddParticles(const Array<IDType> &new_ids,
 #if defined(MFEM_USE_MPI) && defined(MFEM_USE_GSLIB)
 
 /// \cond DO_NOT_DOCUMENT
+// Static helper: gather selected particle-vector entries into a compact buffer.
+// nvcc does not allow extended host/device lambdas in non-public members.
+static void GatherParticleVectorDevice(const ParticleVector &pv,
+                                       const Array<int> &send_idxs,
+                                       Vector &send_data,
+                                       int nsend)
+{
+   const int vdim = pv.GetVDim();
+   const int ordering = pv.GetOrdering();
+   const int num_particles = pv.GetNumParticles();
+   const MemoryClass device_mc = Device::GetDeviceMemoryClass();
+   send_data.SetSize(nsend*vdim);
+   real_t *d_send_data =
+      send_data.GetMemory().Write(device_mc, send_data.Size());
+   const real_t *d_src = pv.GetMemory().Read(device_mc, pv.Size());
+   const int *d_send_idxs = send_idxs.GetMemory().Read(device_mc, nsend);
+
+   mfem::forall(nsend, [=] MFEM_HOST_DEVICE (int i)
+   {
+      const int p = d_send_idxs[i];
+      const int offset = (ordering == Ordering::byVDIM) ? p * vdim : p;
+      const int stride = (ordering == Ordering::byVDIM) ? 1 : num_particles;
+
+      for (int c = 0; c < vdim; c++)
+      {
+         d_send_data[i*vdim + c] = d_src[offset + c*stride];
+      }
+   });
+}
+
+// Static helper: gather selected tag values into a compact buffer.
+// nvcc does not allow extended host/device lambdas in non-public members.
+static void GatherParticleTagsDevice(const Array<int> &tag,
+                                     const Array<int> &send_idxs,
+                                     Array<int> &send_tag,
+                                     int nsend)
+{
+   const MemoryClass device_mc = Device::GetDeviceMemoryClass();
+   send_tag.SetSize(nsend);
+   int *d_send_tag = send_tag.GetMemory().Write(device_mc, nsend);
+   const int *d_tag = tag.GetMemory().Read(device_mc, tag.Size());
+   const int *d_send_idxs = send_idxs.GetMemory().Read(device_mc, nsend);
+
+   mfem::forall(nsend, [=] MFEM_HOST_DEVICE (int i)
+   {
+      d_send_tag[i] = d_tag[d_send_idxs[i]];
+   });
+}
+
+// Static helper: scatter compact particle-vector entries to particle storage.
+// nvcc does not allow extended host/device lambdas in non-public members.
+static void ScatterParticleVectorDevice(ParticleVector &pv,
+                                        const Vector &recv_data,
+                                        const Array<int> &recv_locs,
+                                        int nrecv)
+{
+   const int vdim = pv.GetVDim();
+   const int ordering = pv.GetOrdering();
+   const int num_particles = pv.GetNumParticles();
+   const MemoryClass device_mc = Device::GetDeviceMemoryClass();
+   const real_t *d_recv_data =
+      recv_data.GetMemory().Read(device_mc, recv_data.Size());
+   const int *d_recv_locs = recv_locs.GetMemory().Read(device_mc, nrecv);
+   real_t *d_dst = pv.GetMemory().ReadWrite(device_mc, pv.Size());
+
+   mfem::forall(nrecv, [=] MFEM_HOST_DEVICE (int i)
+   {
+      const int p = d_recv_locs[i];
+      const int offset = (ordering == Ordering::byVDIM) ? p * vdim : p;
+      const int stride = (ordering == Ordering::byVDIM) ? 1 : num_particles;
+
+      for (int c = 0; c < vdim; c++)
+      {
+         d_dst[offset + c*stride] = d_recv_data[i*vdim + c];
+      }
+   });
+}
+
+// Static helper: scatter compact tag values to particle storage.
+// nvcc does not allow extended host/device lambdas in non-public members.
+static void ScatterParticleTagsDevice(Array<int> &tag,
+                                      const Array<int> &recv_tag,
+                                      const Array<int> &recv_locs,
+                                      int nrecv)
+{
+   const MemoryClass device_mc = Device::GetDeviceMemoryClass();
+   const int *d_recv_tag = recv_tag.GetMemory().Read(device_mc, nrecv);
+   const int *d_recv_locs = recv_locs.GetMemory().Read(device_mc, nrecv);
+   int *d_tag = tag.GetMemory().ReadWrite(device_mc, tag.Size());
+
+   mfem::forall(nrecv, [=] MFEM_HOST_DEVICE (int i)
+   {
+      d_tag[d_recv_locs[i]] = d_recv_tag[i];
+   });
+}
+
 template<size_t NBytes>
 void ParticleSet::TransferParticlesImpl(ParticleSet &pset,
                                         const Array<int> &send_idxs,
@@ -266,37 +364,108 @@ void ParticleSet::TransferParticlesImpl(ParticleSet &pset,
    array_init(parr_t, &gsl_arr, send_idxs.Size());
    pdata_arr = (parr_t*) gsl_arr.ptr;
 
+   int nparticles = pset.GetNParticles();
+   int nsend      = send_idxs.Size();
    gsl_arr.n = send_idxs.Size();
+
+   const int *h_send_idxs_initial = send_idxs.HostRead();
+   const IDType *h_ids = pset.GetIDs().HostRead();
    for (int i = 0; i < send_idxs.Size(); i++)
    {
       parr_t &pdata = pdata_arr[i];
-      pdata.id = pset.GetIDs()[send_idxs[i]];
+      pdata.id = h_ids[h_send_idxs_initial[i]];
+   }
 
-      // Copy particle data directly into pdata
-      size_t counter = 0;
-      for (int f = -1; f < pset.GetNFields(); f++)
+   // Pack coords and fields into the GSLIB send buffer. Device-resident data
+   // is first gathered into a compact device buffer so that only selected
+   // particles are copied back to host. Host-resident data is packed directly.
+   int max_vdim = pset.Coords().GetVDim();
+   for (int f = 0; f < pset.GetNFields(); f++)
+   {
+      int f_vdim = pset.Field(f).GetVDim();
+      if (f_vdim > max_vdim) { max_vdim = f_vdim; }
+   }
+   Vector send_data;
+   Array<int> send_tag;
+   if (Device::IsEnabled())
+   {
+      send_data.SetSize(nsend * max_vdim); // allocate max size over all fields
+      send_tag.SetSize(nsend);
+   }
+
+   size_t counter = 0;
+   for (int f = -1; f < pset.GetNFields(); f++)
+   {
+      const ParticleVector &pv = f == -1 ? pset.Coords() : pset.Field(f);
+      const int vdim = pv.GetVDim();
+      const int ordering = pv.GetOrdering();
+      const int num_particles = pv.GetNumParticles();
+      const bool use_dev = Device::IsEnabled() && pv.UseDevice();
+
+      if (use_dev)
       {
-         ParticleVector &pv = (f == -1 ? pset.Coords() : pset.Field(f));
-         for (int c = 0; c < pv.GetVDim(); c++)
+         GatherParticleVectorDevice(pv, send_idxs, send_data, nsend);
+
+         const real_t *h_send_data = send_data.HostRead();
+         for (int i = 0; i < nsend; i++)
          {
-            std::memcpy(pdata.data.data() + counter, &pv(send_idxs[i], c),
-                        sizeof(real_t));
-            counter += sizeof(real_t);
+            std::memcpy(pdata_arr[i].data.data() + counter,
+                        h_send_data + i*vdim, vdim * sizeof(real_t));
+         }
+      }
+      else
+      {
+         const real_t *h_src = pv.HostRead();
+         const int *h_send_idxs = send_idxs.HostRead();
+         for (int i = 0; i < nsend; i++)
+         {
+            parr_t &pdata = pdata_arr[i];
+            const int p = h_send_idxs[i];
+            const int offset = (ordering == Ordering::byVDIM) ? p * vdim : p;
+            const int stride = (ordering == Ordering::byVDIM) ? 1 :
+                               num_particles;
+
+            for (int c = 0; c < vdim; c++)
+            {
+               std::memcpy(pdata.data.data() + counter + c*sizeof(real_t),
+                           h_src + offset + c*stride, sizeof(real_t));
+            }
          }
       }
 
-      // Copy tags
-      for (int t = 0; t < pset.GetNTags(); t++)
-      {
-         Array<int> &tag_arr = pset.Tag(t);
-         std::memcpy(pdata.data.data() + counter, &tag_arr[send_idxs[i]],
-                     sizeof(int));
-         counter += sizeof(int);
-      }
+      counter += vdim*sizeof(real_t);
    }
 
-   int nparticles = pset.GetNParticles();
-   int nsend      = send_idxs.Size();
+   // Pack tags after all real_t data. Each tag uses the same selective
+   // device gather path when its Array is device-resident.
+   for (int t = 0; t < pset.GetNTags(); t++)
+   {
+      const Array<int> &tag = pset.Tag(t);
+      const size_t tag_counter = counter + t*sizeof(int);
+      const bool use_dev = Device::IsEnabled() && tag.UseDevice();
+
+      if (use_dev)
+      {
+         GatherParticleTagsDevice(tag, send_idxs, send_tag, nsend);
+
+         const int *h_send_tag = send_tag.HostRead();
+         for (int i = 0; i < nsend; i++)
+         {
+            std::memcpy(pdata_arr[i].data.data() + tag_counter,
+                        h_send_tag + i, sizeof(int));
+         }
+      }
+      else
+      {
+         const int *h_tag = tag.HostRead();
+         const int *h_send_idxs = send_idxs.HostRead();
+         for (int i = 0; i < nsend; i++)
+         {
+            std::memcpy(pdata_arr[i].data.data() + tag_counter,
+                        h_tag + h_send_idxs[i], sizeof(int));
+         }
+      }
+   }
 
    // Transfer particles
    sarray_transfer_ext(parr_t, &gsl_arr, send_ranks.GetData(),
@@ -304,11 +473,20 @@ void ParticleSet::TransferParticlesImpl(ParticleSet &pset,
 
    // Make sure we have enough space for received particles
    int nrecv = (int) gsl_arr.n;
+
+   Vector recv_data;
+   Array<int> recv_tag;
+   if (Device::IsEnabled())
+   {
+      recv_data.SetSize(nrecv * max_vdim);
+      recv_tag.SetSize(nrecv);
+   }
+
    int ndelete = nsend - nrecv;
    if (ndelete > 0)
    {
       // Remove unneeded particles
-      auto datap = const_cast<int*>(send_idxs.GetData());
+      auto datap = const_cast<int*>(send_idxs.HostRead());
       Array<int> delete_idxs(datap + nrecv, ndelete);
       pset.RemoveParticles(delete_idxs);
    }
@@ -319,47 +497,133 @@ void ParticleSet::TransferParticlesImpl(ParticleSet &pset,
 
    pdata_arr = (parr_t*) gsl_arr.ptr;
 
-   // Add newly-recvd data directly to active state
+   // Make a list of new IDs to add
+   int num_new = nrecv > nsend ? nrecv - nsend : 0;
+   Array<IDType> new_ids(num_new);
+   for (int i = 0; i < num_new; i++)
+   {
+      new_ids[i] = pdata_arr[nsend + i].id;
+   }
+
+   // Add particles in batch
+   Array<int> new_indices;
+   if (num_new > 0)
+   {
+      pset.AddParticles(new_ids, &new_indices);
+   }
+
+   // Map each received packet to the local particle slot it updates.
+   Array<int> recv_locs(nrecv);
+   int *h_recv_locs = recv_locs.HostWrite();
+   const int *h_send_idxs_recv = send_idxs.HostRead();
    for (int i = 0; i < nrecv; i++)
    {
       parr_t &pdata = pdata_arr[i];
-      IDType id = pdata.id;
-
-      int new_loc_idx;
       if (i < nsend) // update existing particle
       {
-         new_loc_idx = send_idxs[i];
-         pset.UpdateID(new_loc_idx, id);
+         h_recv_locs[i] = h_send_idxs_recv[i];
+         pset.UpdateID(h_recv_locs[i], pdata.id);
       }
       else
       {
-         // add new particle
-         Array<int> idx_temp;
-         pset.AddParticles(Array<IDType>({id}), &idx_temp);
-         new_loc_idx = idx_temp[0]; // Get index of newly-added particle
+         h_recv_locs[i] = new_indices[i - nsend];
       }
+   }
 
-      size_t counter = 0;
-      for (int f = -1; f < pset.GetNFields(); f++)
+   // Unpack coords and fields from GSLIB host packets. Device-resident
+   // destinations use a compact host buffer followed by a device scatter.
+   size_t recv_counter = 0;
+   for (int f = -1; f < pset.GetNFields(); f++)
+   {
+      ParticleVector &pv = (f == -1 ? pset.Coords() : pset.Field(f));
+      const int vdim = pv.GetVDim();
+      const int ordering = pv.GetOrdering();
+      const int num_particles = pv.GetNumParticles();
+      const bool use_dev = Device::IsEnabled() && pv.UseDevice();
+
+      if (use_dev)
       {
-         ParticleVector &pv = (f == -1 ? pset.Coords() : pset.Field(f));
-         for (int c = 0; c < pv.GetVDim(); c++)
+         recv_data.SetSize(nrecv*vdim);
+         real_t *h_recv_data = recv_data.HostWrite();
+
+         for (int i = 0; i < nrecv; i++)
          {
-            real_t& val = pv(new_loc_idx, c);
-            std::memcpy(&val, pdata.data.data() + counter, sizeof(real_t));
-            counter += sizeof(real_t);
+            std::memcpy(h_recv_data + i*vdim,
+                        pdata_arr[i].data.data() + recv_counter,
+                        vdim*sizeof(real_t));
+         }
+
+         ScatterParticleVectorDevice(pv, recv_data, recv_locs, nrecv);
+      }
+      else
+      {
+         real_t *h_dst = pv.HostReadWrite();
+         const int *h_recv_locs_read = recv_locs.HostRead();
+         for (int i = 0; i < nrecv; i++)
+         {
+            parr_t &pdata = pdata_arr[i];
+            const int p = h_recv_locs_read[i];
+            const int offset = (ordering == Ordering::byVDIM) ? p * vdim : p;
+            const int stride = (ordering == Ordering::byVDIM) ? 1 :
+                               num_particles;
+
+            for (int c = 0; c < vdim; c++)
+            {
+               std::memcpy(h_dst + offset + c*stride,
+                           pdata.data.data() + recv_counter + c*sizeof(real_t),
+                           sizeof(real_t));
+            }
          }
       }
 
-      for (int t = 0; t < pset.GetNTags(); t++)
+      recv_counter += vdim*sizeof(real_t);
+   }
+
+   // Unpack tags after all real_t data, using the same compact scatter path
+   // for device-resident tag arrays.
+   for (int t = 0; t < pset.GetNTags(); t++)
+   {
+      Array<int> &tag = pset.Tag(t);
+      const size_t tag_counter = recv_counter + t*sizeof(int);
+      const bool use_dev = Device::IsEnabled() && tag.UseDevice();
+
+      if (use_dev)
       {
-         Array<int> &tag_arr = pset.Tag(t);
-         std::memcpy(&tag_arr[new_loc_idx],
-                     pdata.data.data() + counter, sizeof(int));
-         counter += sizeof(int);
+         recv_tag.SetSize(nrecv);
+         int *h_recv_tag = recv_tag.HostWrite();
+
+         for (int i = 0; i < nrecv; i++)
+         {
+            std::memcpy(h_recv_tag + i,
+                        pdata_arr[i].data.data() + tag_counter, sizeof(int));
+         }
+
+         ScatterParticleTagsDevice(tag, recv_tag, recv_locs, nrecv);
+      }
+      else
+      {
+         int *h_tag = tag.HostReadWrite();
+         const int *h_recv_locs_read = recv_locs.HostRead();
+         for (int i = 0; i < nrecv; i++)
+         {
+            std::memcpy(h_tag + h_recv_locs_read[i],
+                        pdata_arr[i].data.data() + tag_counter, sizeof(int));
+         }
       }
    }
    array_free(&gsl_arr);
+
+   // Restore Device validity if needed
+   for (int f = -1; f < pset.GetNFields(); f++)
+   {
+      ParticleVector &pv = (f == -1 ? pset.Coords() : pset.Field(f));
+      pv.ReadWrite(pv.UseDevice());
+   }
+   for (int t = 0; t < pset.GetNTags(); t++)
+   {
+      Array<int> &tag_arr = pset.Tag(t);
+      if (tag_arr.UseDevice()) { tag_arr.ReadWrite(true); }
+   }
 }
 
 template<size_t NBytes>
@@ -526,11 +790,14 @@ ParticleSet::ParticleSet(int id_stride_, IDType id_counter_, int num_particles,
                          int dim, Ordering::Type coords_ordering, const Array<int> &field_vdims,
                          const Array<Ordering::Type> &field_orderings,
                          const Array<const char*> &field_names_, int num_tags,
-                         const Array<const char*> &tag_names_)
+                         const Array<const char*> &tag_names_,
+                         bool use_device)
    : id_stride(id_stride_),
      id_counter(id_counter_),
      coords(dim, coords_ordering)
 {
+   if (use_device) { coords.UseDevice(true); }
+
    // Initialize fields
    for (int f = 0; f < field_vdims.Size(); f++)
    {
@@ -580,21 +847,22 @@ bool ParticleSet::IsValidParticle(const Particle &p) const
 }
 
 ParticleSet::ParticleSet(int num_particles, int dim,
-                         Ordering::Type coords_ordering)
+                         Ordering::Type coords_ordering,
+                         bool use_device)
    : ParticleSet(1, 0, num_particles, dim, coords_ordering, Array<int>(),
                  Array<Ordering::Type>(), Array<const char*>(), 0,
-                 Array<const char*>())
+                 Array<const char*>(), use_device)
 {
 
 }
 
 ParticleSet::ParticleSet(int num_particles, int dim,
                          const Array<int> &field_vdims, int num_tags,
-                         Ordering::Type all_ordering)
+                         Ordering::Type all_ordering, bool use_device)
    : ParticleSet(1, 0, num_particles, dim, all_ordering, field_vdims,
                  GetOrderingArray(all_ordering, field_vdims.Size()),
                  GetEmptyNameArray(field_vdims.Size()), num_tags,
-                 GetEmptyNameArray(num_tags))
+                 GetEmptyNameArray(num_tags), use_device)
 {
 }
 
@@ -602,11 +870,11 @@ ParticleSet::ParticleSet(int num_particles, int dim,
                          const Array<int> &field_vdims, const Array<const
                          char*> &field_names_, int num_tags,
                          const Array<const char*> &tag_names_,
-                         Ordering::Type all_ordering)
+                         Ordering::Type all_ordering, bool use_device)
    : ParticleSet(1, 0, num_particles, dim, all_ordering, field_vdims,
                  GetOrderingArray(all_ordering, field_vdims.Size()),
                  field_names_, num_tags,
-                 tag_names_)
+                 tag_names_, use_device)
 {
 
 }
@@ -616,9 +884,9 @@ ParticleSet::ParticleSet(int num_particles, int dim,
                          const Array<int> &field_vdims,
                          const Array<Ordering::Type> &field_orderings,
                          const Array<const char*> &field_names_, int num_tags,
-                         const Array<const char*> &tag_names_)
+                         const Array<const char*> &tag_names_, bool use_device)
    : ParticleSet(1, 0, num_particles, dim, coords_ordering, field_vdims,
-                 field_orderings, field_names_, num_tags, tag_names_)
+                 field_orderings, field_names_, num_tags, tag_names_, use_device)
 {
 
 }
@@ -627,21 +895,21 @@ ParticleSet::ParticleSet(int num_particles, int dim,
 
 #ifdef MFEM_USE_MPI
 ParticleSet::ParticleSet(MPI_Comm comm_, int rank_num_particles, int dim,
-                         Ordering::Type coords_ordering)
+                         Ordering::Type coords_ordering, bool use_device)
    : ParticleSet(comm_, rank_num_particles, dim, coords_ordering, Array<int>(),
                  Array<Ordering::Type>(), Array<const char*>(), 0,
-                 Array<const char*>())
+                 Array<const char*>(), use_device)
 {
 
 };
 
 ParticleSet::ParticleSet(MPI_Comm comm_, int rank_num_particles, int dim,
                          const Array<int> &field_vdims, int num_tags,
-                         Ordering::Type all_ordering)
+                         Ordering::Type all_ordering, bool use_device)
    : ParticleSet(comm_, rank_num_particles, dim, all_ordering, field_vdims,
                  GetOrderingArray(all_ordering, field_vdims.Size()),
                  GetEmptyNameArray(field_vdims.Size()), num_tags,
-                 GetEmptyNameArray(num_tags))
+                 GetEmptyNameArray(num_tags), use_device)
 {
 
 }
@@ -650,11 +918,11 @@ ParticleSet::ParticleSet(MPI_Comm comm_, int rank_num_particles, int dim,
                          const Array<int> &field_vdims, const Array<const
                          char*> &field_names_,
                          int num_tags, const Array<const char*> &tag_names_,
-                         Ordering::Type all_ordering)
+                         Ordering::Type all_ordering, bool use_device)
    : ParticleSet(comm_, rank_num_particles, dim, all_ordering, field_vdims,
                  GetOrderingArray(all_ordering, field_vdims.Size()),
                  field_names_, num_tags,
-                 tag_names_)
+                 tag_names_, use_device)
 {
 
 }
@@ -664,7 +932,7 @@ ParticleSet::ParticleSet(MPI_Comm comm_, int rank_num_particles, int dim,
                          const Array<int> &field_vdims,
                          const Array<Ordering::Type> &field_orderings,
                          const Array<const char*> &field_names_, int num_tags,
-                         const Array<const char*> &tag_names_)
+                         const Array<const char*> &tag_names_, bool use_device)
    : ParticleSet(GetSize(comm_), (IDType)GetRank(comm_),
                  rank_num_particles,
                  dim,
@@ -673,7 +941,7 @@ ParticleSet::ParticleSet(MPI_Comm comm_, int rank_num_particles, int dim,
                  field_orderings,
                  field_names_,
                  num_tags,
-                 tag_names_)
+                 tag_names_, use_device)
 {
    comm = comm_;
 #ifdef MFEM_USE_GSLIB
@@ -705,6 +973,7 @@ int ParticleSet::AddField(int vdim, Ordering::Type field_ordering,
    }
    fields.emplace_back(std::make_unique<ParticleVector>(vdim, field_ordering,
                                                         GetNParticles()));
+   if (coords.UseDevice()) { fields.back()->UseDevice(true); }
    field_names.emplace_back(field_name_str);
 
    return GetNFields() - 1;
@@ -718,6 +987,7 @@ int ParticleSet::AddTag(const char* tag_name)
       tag_name_str = GetDefaultTagName(tag_names.size());
    }
    tags.emplace_back(std::make_unique<Array<int>>(GetNParticles()));
+   if (coords.UseDevice()) { tags.back()->GetMemory().UseDevice(true); }
    tag_names.emplace_back(tag_name_str);
 
    return GetNTags() - 1;
@@ -782,7 +1052,7 @@ Particle ParticleSet::GetParticle(int i) const
 
    for (int t = 0; t < GetNTags(); t++)
    {
-      p.Tag(t) = Tag(t)[i];
+      p.Tag(t) = Tag(t).HostRead()[i];
    }
 
    return p;
@@ -790,13 +1060,21 @@ Particle ParticleSet::GetParticle(int i) const
 
 bool ParticleSet::IsParticleRefValid() const
 {
-   if (coords.GetOrdering() == Ordering::byNODES)
+   if (coords.GetOrdering() == Ordering::byNODES || coords.UseDevice())
    {
       return false;
    }
    for (int f = 0; f < GetNFields(); f++)
    {
-      if (fields[f]->GetOrdering() == Ordering::byNODES)
+      if (fields[f]->GetOrdering() == Ordering::byNODES ||
+          fields[f]->UseDevice())
+      {
+         return false;
+      }
+   }
+   for (int t = 0; t < GetNTags(); t++)
+   {
+      if (tags[t]->UseDevice())
       {
          return false;
       }
@@ -806,6 +1084,10 @@ bool ParticleSet::IsParticleRefValid() const
 
 Particle ParticleSet::GetParticleRef(int i)
 {
+   MFEM_ASSERT(IsParticleRefValid(),
+               "GetParticleRef is only valid when coordinates and fields are "
+               "ordered byVDIM and particle data is host-resident.");
+
    Particle p = CreateParticle();
 
    Coords().GetValuesRef(i, p.Coords());
@@ -839,7 +1121,7 @@ void ParticleSet::SetParticle(int i, const Particle &p)
 
    for (int t = 0; t < GetNTags(); t++)
    {
-      Tag(t)[i] = p.Tag(t);
+      Tag(t).HostReadWrite()[i] = p.Tag(t);
    }
 }
 
@@ -900,6 +1182,15 @@ void ParticleSet::PrintCSV(const char *fname, const Array<int> &field_idxs,
 #ifdef MFEM_USE_MPI
    int rank = GetRank(comm);
 #endif // MFEM_USE_MPI
+   // make sure we can read tag data on host. fields and coords will be read as
+   // needed in the loop below, so we don't need to pre-read them here.
+   for (int i = 0; i < GetNTags(); i++)
+   {
+      tags[i]->HostRead();
+   }
+   ids.HostRead();
+
+   // Write particle data
    for (int i = 0; i < GetNParticles(); i++)
    {
       ss_data << ids[i];

fem/particleset.hpp

@@ -211,6 +211,12 @@ public:
  *  byVDIM). The unique_ptrs to all the ParticleVectors are stored in the
  *  std::vector \ref fields.
  *
+ *  @par Device Behavior:
+ *  When a ParticleSet is constructed with \p use_device=true, \ref coords and
+ *  all ParticleVector fields are marked to use device memory. Fields added
+ *  later through \ref AddField inherit the current device mode (through
+ *  \ref coords).
+ *
  *  @par Tags:
  *  Tags represent integers associated with each particle. For a given tag,
  *  all particle data are stored in a single Array<int>. The unique_ptrs to all
@@ -369,7 +375,10 @@ protected:
     *  ID of a particle.
     */
    void UpdateID(int local_idx, IDType new_global_id)
-   { ids[local_idx] = new_global_id; }
+   {
+      ids.HostReadWrite();
+      ids[local_idx] = new_global_id;
+   }
 
    /** @brief Create a Particle object with the same spatial dimension,
     *  number of fields and field vdims, and number of tags as this ParticleSet.
@@ -399,12 +408,14 @@ protected:
     *  @param[in] field_names_        Array of field names.
     *  @param[in] num_tags            Number of tags to register.
     *  @param[in] tag_names_          Array of tag names.
+    *  @param[in] use_device          Use device memory for particle fields.
     */
    ParticleSet(int id_stride_, IDType id_counter_, int num_particles, int dim,
                Ordering::Type coords_ordering, const Array<int> &field_vdims,
                const Array<Ordering::Type> &field_orderings,
                const Array<const char*> &field_names_, int num_tags,
-               const Array<const char*> &tag_names_);
+               const Array<const char*> &tag_names_,
+               bool use_device);
 
 public:
 
@@ -413,9 +424,12 @@ public:
     *  @param[in] num_particles       Number of particles to initialize.
     *  @param[in] dim                 Particle spatial dimension.
     *  @param[in] coords_ordering     Ordering of coordinates.
+    *  @param[in] use_device          (Optional) Use device memory for particle
+    *                                 fields.
     */
    ParticleSet(int num_particles, int dim,
-               Ordering::Type coords_ordering=Ordering::byVDIM);
+               Ordering::Type coords_ordering=Ordering::byVDIM,
+               bool use_device=false);
 
    /** @brief Construct a serial ParticleSet with specified fields and tags at
     *  construction.
@@ -426,9 +440,12 @@ public:
     *  @param[in] num_tags            Number of tags to register.
     *  @param[in] all_ordering        (Optional) Ordering of coordinates and
     *                                 field ParticleVector.
+    *  @param[in] use_device          (Optional) Use device memory for particle
+    *                                 fields.
     */
    ParticleSet(int num_particles, int dim, const Array<int> &field_vdims,
-               int num_tags, Ordering::Type all_ordering=Ordering::byVDIM);
+               int num_tags, Ordering::Type all_ordering=Ordering::byVDIM,
+               bool use_device=false);
 
    /** @brief Construct a serial ParticleSet with specified fields and tags at
     *  construction, with names.
@@ -441,11 +458,14 @@ public:
     *  @param[in] tag_names_          Array of tag names.
     *  @param[in] all_ordering        (Optional) Ordering of coordinates and
     *                                 field ParticleVector.
+    *  @param[in] use_device          (Optional) Use device memory for particle
+    *                                 fields.
     */
    ParticleSet(int num_particles, int dim, const Array<int> &field_vdims,
                const Array<const char*> &field_names_, int num_tags,
                const Array<const char*> &tag_names_,
-               Ordering::Type all_ordering=Ordering::byVDIM);
+               Ordering::Type all_ordering=Ordering::byVDIM,
+               bool use_device=false);
 
    /** @brief Comprehensive serial constructor of ParticleSet.
     *
@@ -457,12 +477,15 @@ public:
     *  @param[in] field_names_        Array of field names.
     *  @param[in] num_tags            Number of tags to register.
     *  @param[in] tag_names_          Array of tag names.
+    *  @param[in] use_device          (Optional) Use device memory for particle
+    *                                 fields.
     */
    ParticleSet(int num_particles, int dim, Ordering::Type coords_ordering,
                const Array<int> &field_vdims,
                const Array<Ordering::Type> &field_orderings,
                const Array<const char*> &field_names_, int num_tags,
-               const Array<const char*> &tag_names_);
+               const Array<const char*> &tag_names_,
+               bool use_device=false);
 
 #ifdef MFEM_USE_MPI
    /** @brief Construct a parallel ParticleSet.
@@ -471,9 +494,12 @@ public:
     *  @param[in] rank_num_particles  Number of particles to initialize.
     *  @param[in] dim                 Particle spatial dimension.
     *  @param[in] coords_ordering     (Optional) Ordering of coordinates.
+    *  @param[in] use_device          (Optional) Use device memory for particle
+    *                                 fields.
     */
    ParticleSet(MPI_Comm comm_, int rank_num_particles, int dim,
-               Ordering::Type coords_ordering=Ordering::byVDIM);
+               Ordering::Type coords_ordering=Ordering::byVDIM,
+               bool use_device=false);
 
    /** @brief Construct a parallel ParticleSet with specified fields and tags
     *  at construction.
@@ -485,10 +511,13 @@ public:
     *  @param[in] num_tags            Number of tags to register.
     *  @param[in] all_ordering        (Optional) Ordering of coordinates and
     *                                 field ParticleVector.
+    *  @param[in] use_device          (Optional) Use device memory for particle
+    *                                 fields.
     */
    ParticleSet(MPI_Comm comm_, int rank_num_particles, int dim,
                const Array<int> &field_vdims, int num_tags,
-               Ordering::Type all_ordering=Ordering::byVDIM);
+               Ordering::Type all_ordering=Ordering::byVDIM,
+               bool use_device=false);
 
    /** @brief Construct a parallel ParticleSet with specified fields and tags
     *  at construction, with names (for PrintCSV()).
@@ -502,12 +531,15 @@ public:
     *  @param[in] tag_names_          Array of tag names.
     *  @param[in] all_ordering        (Optional) Ordering of coordinates and
     *                                 field ParticleVector.
+    *  @param[in] use_device          (Optional) Use device memory for particle
+    *                                 fields.
     */
    ParticleSet(MPI_Comm comm_, int rank_num_particles, int dim,
                const Array<int> &field_vdims,
                const Array<const char*> &field_names_,
                int num_tags, const Array<const char*> &tag_names_,
-               Ordering::Type all_ordering=Ordering::byVDIM);
+               Ordering::Type all_ordering=Ordering::byVDIM,
+               bool use_device=false);
 
    /** @brief Comprehensive parallel constructor of ParticleSet.
     *
@@ -520,12 +552,15 @@ public:
     *  @param[in] field_names_        Array of field names.
     *  @param[in] num_tags            Number of tags to register.
     *  @param[in] tag_names_          Array of tag names.
+    *  @param[in] use_device          (Optional) Use device memory for particle
+    *                                 fields.
     */
    ParticleSet(MPI_Comm comm_, int rank_num_particles, int dim,
                Ordering::Type coords_ordering, const Array<int> &field_vdims,
                const Array<Ordering::Type> &field_orderings,
                const Array<const char*> &field_names_, int num_tags,
-               const Array<const char*> &tag_names_);
+               const Array<const char*> &tag_names_,
+               bool use_device=false);
 
    /// Get the MPI communicator for this ParticleSet.
    MPI_Comm GetComm() const { return comm; };
@@ -545,6 +580,8 @@ public:
     *  @param[in] field_ordering   (Optional) Ordering::Type of the field.
     *  @param[in] field_name       (Optional) Name of the field.
     *
+    *  @note New fields inherit the current device mode of \ref coords.
+    *
     *  @return Index of the newly-added field.
     */
    int AddField(int vdim, Ordering::Type field_ordering=Ordering::byVDIM,
@@ -637,8 +674,8 @@ public:
 
    /** @brief Determine if GetParticleRef is valid.
     *
-    * If coordinates and all fields are ordered byVDIM, then returns true.
-    * Otherwise, false.
+    * Returns true when coordinates and all fields are ordered byVDIM and
+    * particle data is host-resident. Otherwise, false.
     */
    bool IsParticleRefValid() const;
 

general/backends.hpp

@@ -38,15 +38,6 @@
 #define CUB_IGNORE_DEPRECATED_CPP_DIALECT
 #define THRUST_IGNORE_DEPRECATED_CPP_DIALECT
 
-// MFEM only supports using RAJA/CAMP backends in default stream mode because
-// memory calls are performed outside of the RAJA ecosystem
-#ifndef CAMP_USE_PLATFORM_DEFAULT_STREAM
-#define CAMP_USE_PLATFORM_DEFAULT_STREAM 1
-#else
-#if !CAMP_USE_PLATFORM_DEFAULT_STREAM
-#error "MFEM only supports RAJA/CAMP with the default platform stream."
-#endif
-#endif
 #include "RAJA/RAJA.hpp"
 #if defined(RAJA_ENABLE_CUDA) && !defined(MFEM_USE_CUDA)
 #error When RAJA is built with CUDA, MFEM_USE_CUDA=YES is required

general/device.cpp

@@ -581,7 +581,9 @@ void Device::Setup(const std::string &device_option, const int device_id)
    if (Allows(Backend::CUDA)) { CudaDeviceSetup(dev, ngpu); }
    if (Allows(Backend::HIP)) { HipDeviceSetup(dev, ngpu); }
    if (Allows(Backend::RAJA_CUDA) || Allows(Backend::RAJA_HIP))
-   { RajaDeviceSetup(dev, ngpu); }
+   {
+      RajaDeviceSetup(dev, ngpu);
+   }
    // The check for MFEM_USE_OCCA is in the function OccaDeviceSetup().
    if (Allows(Backend::OCCA_MASK)) { OccaDeviceSetup(dev); }
    if (Allows(Backend::CEED_MASK))

general/device.hpp

@@ -16,6 +16,11 @@
 #include "globals.hpp"
 #include "mem_manager.hpp"
 
+#ifdef MFEM_USE_RAJA
+#include "RAJA/RAJA.hpp"
+#endif
+
+#include <memory>
 #include <string>
 
 namespace mfem
@@ -266,6 +271,18 @@ public:
    static inline bool Allows(unsigned long b_mask)
    { return Get().backends & b_mask; }
 
+#if defined(MFEM_USE_RAJA) &&                                                  \
+   (defined(RAJA_ENABLE_CUDA) || defined(RAJA_ENABLE_HIP))
+   static inline auto GetRajaResource()
+   {
+#if defined(RAJA_ENABLE_CUDA)
+      return RAJA::resources::Cuda::CudaFromStream(0, Get().GetId());
+#elif defined(RAJA_ENABLE_HIP)
+      return RAJA::resources::Hip::HipFromStream(0, Get().GetId());
+#endif
+   }
+#endif
+
    /** @brief Get the current Host MemoryType. This is the MemoryType used by
        most MFEM classes when allocating memory used on the host.
    */

general/forall.hpp

@@ -317,8 +317,8 @@ template <typename DBODY>
 void RajaCuWrap1D(const int N, DBODY &&d_body)
 {
    //true denotes asynchronous kernel
-   RAJA::forall<RAJA::cuda_exec<MFEM_CUDA_BLOCKS,true>>(RAJA::RangeSegment(0,N),
-                                                        d_body);
+   RAJA::forall<RAJA::cuda_exec<MFEM_CUDA_BLOCKS, true> >(
+      Device::GetRajaResource(), RAJA::RangeSegment(0, N), d_body);
 }
 
 template <typename DBODY>
@@ -331,9 +331,9 @@ void RajaCuWrap2D(const int N, DBODY &&d_body,
    using namespace RAJA;
    using RAJA::RangeSegment;
 
-   launch<cuda_launch_policy>
-   (LaunchParams(Teams(G), Threads(X, Y, BZ)),
-    [=] RAJA_DEVICE (LaunchContext ctx)
+   launch<cuda_launch_policy>(Device::GetRajaResource(),
+                              LaunchParams(Teams(G), Threads(X, Y, BZ)),
+                              [=] RAJA_DEVICE(LaunchContext ctx)
    {
 
       loop<cuda_teams_x>(ctx, RangeSegment(0, G), [&] (const int n)
@@ -349,7 +349,6 @@ void RajaCuWrap2D(const int N, DBODY &&d_body,
          });
 
       });
-
    });
 
    MFEM_GPU_CHECK(cudaGetLastError());
@@ -365,9 +364,9 @@ void RajaCuWrap2DLaunchBounds(const int N, DBODY &&d_body, const int X,
    using namespace RAJA;
    using RAJA::RangeSegment;
 
-   launch<cuda_launch_bounds_policy<LB> >
-   (LaunchParams(Teams(G), Threads(X, Y, BZ)),
-    [=] RAJA_DEVICE(LaunchContext ctx)
+   launch<cuda_launch_bounds_policy<LB> >(
+      Device::GetRajaResource(), LaunchParams(Teams(G), Threads(X, Y, BZ)),
+      [=] RAJA_DEVICE(LaunchContext ctx)
    {
       loop<cuda_teams_x>(ctx, RangeSegment(0, G), [&] (const int n)
       {
@@ -390,13 +389,12 @@ void RajaCuWrap3D(const int N, DBODY &&d_body,
    using namespace RAJA;
    using RAJA::RangeSegment;
 
-   launch<cuda_launch_policy>
-   (LaunchParams(Teams(GRID), Threads(X, Y, Z)),
-    [=] RAJA_DEVICE (LaunchContext ctx)
+   launch<cuda_launch_policy>(Device::GetRajaResource(),
+                              LaunchParams(Teams(GRID), Threads(X, Y, Z)),
+                              [=] RAJA_DEVICE(LaunchContext ctx)
    {
 
       loop<cuda_teams_x>(ctx, RangeSegment(0, N), d_body);
-
    });
 
    MFEM_GPU_CHECK(cudaGetLastError());
@@ -410,12 +408,10 @@ void RajaCuWrap3DLaunchBounds(const int N, DBODY &&d_body,
    using namespace RAJA;
    using RAJA::RangeSegment;
 
-   launch<cuda_launch_bounds_policy<LB> >
-   (LaunchParams(Teams(GRID), Threads(X, Y, Z)),
-    [=] RAJA_DEVICE(LaunchContext ctx)
-   {
-      loop<cuda_teams_x>(ctx, RangeSegment(0, N), d_body);
-   });
+   launch<cuda_launch_bounds_policy<LB> >(
+      Device::GetRajaResource(), LaunchParams(Teams(GRID), Threads(X, Y, Z)),
+      [=] RAJA_DEVICE(LaunchContext ctx)
+   { loop<cuda_teams_x>(ctx, RangeSegment(0, N), d_body); });
    MFEM_GPU_CHECK(cudaGetLastError());
 }
 
@@ -484,8 +480,8 @@ template <typename DBODY>
 void RajaHipWrap1D(const int N, DBODY &&d_body)
 {
    //true denotes asynchronous kernel
-   RAJA::forall<RAJA::hip_exec<MFEM_HIP_BLOCKS,true>>(RAJA::RangeSegment(0,N),
-                                                      d_body);
+   RAJA::forall<RAJA::hip_exec<MFEM_HIP_BLOCKS,true> >(RAJA::RangeSegment(0,N),
+                                                       d_body);
 }
 
 template <typename DBODY>
@@ -498,9 +494,9 @@ void RajaHipWrap2D(const int N, DBODY &&d_body,
    using namespace RAJA;
    using RAJA::RangeSegment;
 
-   launch<hip_launch_policy>
-   (LaunchParams(Teams(G), Threads(X, Y, BZ)),
-    [=] RAJA_DEVICE (LaunchContext ctx)
+   launch<hip_launch_policy>(Device::GetRajaResource(),
+                             LaunchParams(Teams(G), Threads(X, Y, BZ)),
+                             [=] RAJA_DEVICE(LaunchContext ctx)
    {
 
       loop<hip_teams_x>(ctx, RangeSegment(0, G), [&] (const int n)
@@ -516,7 +512,6 @@ void RajaHipWrap2D(const int N, DBODY &&d_body,
          });
 
       });
-
    });
 
    MFEM_GPU_CHECK(hipGetLastError());
@@ -532,9 +527,9 @@ void RajaHipWrap2DLaunchBounds(const int N, DBODY &&d_body, const int X,
    using namespace RAJA;
    using RAJA::RangeSegment;
 
-   launch<hip_launch_bounds_policy<LB> >
-   (LaunchParams(Teams(G), Threads(X, Y, BZ)),
-    [=] RAJA_DEVICE(LaunchContext ctx)
+   launch<hip_launch_bounds_policy<LB> >(
+      Device::GetRajaResource(), LaunchParams(Teams(G), Threads(X, Y, BZ)),
+      [=] RAJA_DEVICE(LaunchContext ctx)
    {
       loop<hip_teams_x>(ctx, RangeSegment(0, G), [&] (const int n)
       {
@@ -557,13 +552,12 @@ void RajaHipWrap3D(const int N, DBODY &&d_body,
    using namespace RAJA;
    using RAJA::RangeSegment;
 
-   launch<hip_launch_policy>
-   (LaunchParams(Teams(GRID), Threads(X, Y, Z)),
-    [=] RAJA_DEVICE (LaunchContext ctx)
+   launch<hip_launch_policy>(Device::GetRajaResource(),
+                             LaunchParams(Teams(GRID), Threads(X, Y, Z)),
+                             [=] RAJA_DEVICE(LaunchContext ctx)
    {
 
       loop<hip_teams_x>(ctx, RangeSegment(0, N), d_body);
-
    });
 
    MFEM_GPU_CHECK(hipGetLastError());
@@ -577,12 +571,10 @@ void RajaHipWrap3DLaunchBounds(const int N, DBODY &&d_body, const int X,
    using namespace RAJA;
    using RAJA::RangeSegment;
 
-   launch<hip_launch_bounds_policy<LB> >
-   (LaunchParams(Teams(GRID), Threads(X, Y, Z)),
-    [=] RAJA_DEVICE(LaunchContext ctx)
-   {
-      loop<hip_teams_x>(ctx, RangeSegment(0, N), d_body);
-   });
+   launch<hip_launch_bounds_policy<LB> >(
+      Device::GetRajaResource(), LaunchParams(Teams(GRID), Threads(X, Y, Z)),
+      [=] RAJA_DEVICE(LaunchContext ctx)
+   { loop<hip_teams_x>(ctx, RangeSegment(0, N), d_body); });
    MFEM_GPU_CHECK(hipGetLastError());
 }
 

linalg/ginkgo.hpp

@@ -39,6 +39,7 @@ namespace Ginkgo
 {
 
 template <typename T> using gko_array = gko::array<T>;
+#if defined(MFEM_USE_MPI) && GINKGO_BUILD_MPI
 // for inter-operability with hypre integer types
 using gko_hypre_int =
    std::conditional_t<sizeof(HYPRE_Int) == sizeof(std::int32_t), std::int32_t,
@@ -50,6 +51,7 @@ static_assert(!std::is_void_v<gko_hypre_int>,
               "HYPRE_Int type is incompatible with Ginkgo");
 static_assert(!std::is_void_v<gko_hypre_bigint>,
               "HYPRE_BigInt type is incompatible with Ginkgo");
+#endif
 
 /**
 * Helper class for a case where a wrapped MFEM Vector

linalg/hypre.cpp

@@ -2872,8 +2872,8 @@ void HypreParMatrix::Destroy()
    if (HypreUsingGPU() && ParCSROwner && (diagOwner < 0 || offdOwner < 0))
    {
       // Put the "host" or "hypre" pointers in {i,j,data} of A->{diag,offd}, so
-      // that they can be destroyed by hypre when hypre_ParCSRMatrixDestroy(A)
-      // is called below.
+      // that they can be destroyed by mfem_hypre_TFree_host() or hypre when
+      // hypre_ParCSRMatrixDestroy(A) is called below, respectively.
 
       // Check that if both diagOwner and offdOwner are negative then they have
       // the same value.
@@ -2882,7 +2882,33 @@ void HypreParMatrix::Destroy()
 
       MemoryClass mc = (diagOwner == -1 || offdOwner == -1) ?
                        Device::GetHostMemoryClass() : GetHypreMemoryClass();
-      Write(mc, diagOwner < 0, offdOwner <0);
+      Write(mc, diagOwner < 0, offdOwner < 0);
+      if (diagOwner == -1)
+      {
+         // Note: mfem_hypre_TFree_host() sets the pointer to NULL.
+         mfem_hypre_TFree_host(hypre_CSRMatrixI(A->diag));
+         if (hypre_CSRMatrixOwnsData(A->diag))
+         {
+            mfem_hypre_TFree_host(hypre_CSRMatrixJ(A->diag));
+            mfem_hypre_TFree_host(hypre_CSRMatrixData(A->diag));
+         }
+#if MFEM_HYPRE_VERSION >= 21800
+         hypre_CSRMatrixMemoryLocation(A->diag) = GetHypreMemoryLocation();
+#endif
+      }
+      if (offdOwner == -1)
+      {
+         // Note: mfem_hypre_TFree_host() sets the pointer to NULL.
+         mfem_hypre_TFree_host(hypre_CSRMatrixI(A->offd));
+         if (hypre_CSRMatrixOwnsData(A->offd))
+         {
+            mfem_hypre_TFree_host(hypre_CSRMatrixJ(A->offd));
+            mfem_hypre_TFree_host(hypre_CSRMatrixData(A->offd));
+         }
+#if MFEM_HYPRE_VERSION >= 21800
+         hypre_CSRMatrixMemoryLocation(A->offd) = GetHypreMemoryLocation();
+#endif
+      }
    }
 #endif
 

linalg/hypre.hpp

@@ -432,10 +432,11 @@ private:
    // and A->col_map_offd.
    // The possible values for diagOwner are:
    //  -1: no special treatment of A->diag (default)
-   //      when hypre is built with CUDA support, A->diag owns the "host"
-   //      pointers (according to A->diag->owns_data)
-   //  -2: used when hypre is built with CUDA support, A->diag owns the "hypre"
-   //      pointers (according to A->diag->owns_data)
+   //      when hypre is using GPU, A->diag owns the "host" pointers (according
+   //      to A->diag->owns_data); these host pointers are freed by MFEM using
+   //      hypre's host deallocation macros
+   //  -2: used when hypre is using GPU, A->diag owns the "hypre" pointers
+   //      (according to A->diag->owns_data)
    //   0: prevent hypre from destroying A->diag->{i,j,data}
    //   1: same as 0, plus own the "host" A->diag->{i,j}
    //   2: same as 0, plus own the "host" A->diag->data

linalg/particlevector.cpp

@@ -10,6 +10,7 @@
 // CONTRIBUTING.md for details.
 
 #include "particlevector.hpp"
+#include "../general/forall.hpp"
 
 namespace mfem
 {
@@ -46,20 +47,38 @@ void ParticleVector::GetValues(int i, Vector &nvals) const
 {
    nvals.SetSize(vdim);
 
-   if (ordering == Ordering::byNODES)
+   const bool nvals_use_dev = nvals.UseDevice();
+   // Use ParticleVector's device flag to minimize movement from large source
+   const bool use_dev = UseDevice();
+   const auto d_src = Read(use_dev);
+   auto d_dest = nvals.Write(use_dev);
+
+   const int vdim_ = vdim;
+   const int ordering_ = (int)ordering;
+   const int nv = (ordering == Ordering::byNODES) ? size / vdim : 0;
+
+   mfem::forall_switch(use_dev, vdim_, [=] MFEM_HOST_DEVICE (int c)
    {
-      int nv = GetNumParticles();
-      for (int c = 0; c < vdim; c++)
+      if (ordering_ == Ordering::byNODES)
       {
-         nvals[c] = Vector::operator[](i+nv*c);
+         d_dest[c] = d_src[i + nv*c];
       }
+      else
+      {
+         d_dest[c] = d_src[c + vdim_*i];
+      }
+   });
+
+   // If nvals was not using device but ParticleVector is, copy back to host
+   if (!nvals_use_dev && use_dev)
+   {
+      nvals.HostRead();
+      nvals.UseDevice(false);
    }
-   else
+   // If nvals was using device but ParticleVector is not, copy back to device
+   if (!use_dev && nvals_use_dev)
    {
-      for (int c = 0; c < vdim; c++)
-      {
-         nvals[c] = Vector::operator[](c+vdim*i);
-      }
+      nvals.Read();
    }
 }
 
@@ -99,21 +118,27 @@ void ParticleVector::GetComponentsRef(int vd, Vector &nref)
 
 void ParticleVector::SetValues(int i, const Vector &nvals)
 {
-   if (ordering == Ordering::byNODES)
+   const bool use_dev = UseDevice(); // use ParticleVector's device flag
+   const auto mc = use_dev ? Device::GetDeviceMemoryClass()
+                   : Device::GetHostMemoryClass();
+   auto d_dest = ReadWrite(use_dev);
+   const auto d_src = nvals.GetMemory().Read(mc, nvals.Size());
+
+   const int vdim_ = vdim;
+   const int ordering_ = (int)ordering;
+   const int nv = (ordering == Ordering::byNODES) ? size / vdim : 0;
+
+   mfem::forall_switch(use_dev, vdim_, [=] MFEM_HOST_DEVICE (int c)
    {
-      int nv = GetNumParticles();
-      for (int c = 0; c < vdim; c++)
+      if (ordering_ == Ordering::byNODES)
       {
-         Vector::operator[](i + c*nv) = nvals[c];
+         d_dest[i + c*nv] = d_src[c];
       }
-   }
-   else
-   {
-      for (int c = 0; c < vdim; c++)
+      else
       {
-         Vector::operator[](c + i*vdim) = nvals[c];
+         d_dest[c + i*vdim_] = d_src[c];
       }
-   }
+   });
 }
 
 void ParticleVector::SetComponents(int vd, const Vector &comp)
@@ -144,6 +169,9 @@ real_t& ParticleVector::operator()(int i, int comp)
                "Component index " << comp <<
                " is invalid for vector dimension " << vdim);
 
+   // non-const so we make host flag valid in case user modifies data
+   HostReadWrite();
+
    if (ordering == Ordering::byNODES)
    {
       return Vector::operator[](i + comp*GetNumParticles());
@@ -163,6 +191,8 @@ const real_t& ParticleVector::operator()(int i, int comp) const
                "Component index " << comp <<
                " is invalid for vector dimension " << vdim);
 
+   HostRead();
+
    if (ordering == Ordering::byNODES)
    {
       return Vector::operator[](i + comp*GetNumParticles());
@@ -240,9 +270,37 @@ void ParticleVector::SetVDim(int vdim_, bool keep_data)
 
 void ParticleVector::SetOrdering(Ordering::Type ordering_, bool keep_data)
 {
-   if (keep_data)
+   if (keep_data && ordering != ordering_)
    {
-      Ordering::Reorder(*this, vdim, ordering, ordering_);
+      int num_particles = GetNumParticles();
+      // create deep copy of old data that will be copied
+      Vector old_data(*this);
+
+      const bool use_dev = UseDevice();
+      const auto d_src = old_data.Read(use_dev);
+      auto d_dest = Write(use_dev);
+
+      const int vdim_ = vdim;
+      const int size_ = size;
+
+      if (ordering_ == Ordering::byNODES) // byVDIM -> byNODES
+      {
+         mfem::forall_switch(use_dev, size_, [=] MFEM_HOST_DEVICE (int k)
+         {
+            int i = k / vdim_; // src particle index
+            int d = k % vdim_; // src component index
+            d_dest[i + d * num_particles] = d_src[k];
+         });
+      }
+      else // byNODES -> byVDIM
+      {
+         mfem::forall_switch(use_dev, size_, [=] MFEM_HOST_DEVICE (int k)
+         {
+            int d = k / num_particles; // src component index
+            int i = k % num_particles; // src particle index
+            d_dest[d + i * vdim_] = d_src[k];
+         });
+      }
    }
    ordering = ordering_;
 }
@@ -270,32 +328,47 @@ void ParticleVector::SetNumParticles(int num_vectors, bool keep_data)
 
       if (!keep_data) { return; }
 
+      const bool use_dev = UseDevice();
+      auto d_dest = this->ReadWrite(use_dev);
+
       if (ordering == Ordering::byNODES)
       {
-         // Shift entries for byNODES
-         for (int c = vdim-1; c > 0; c--)
-         {
-            for (int i = old_nv-1; i >= 0; i--)
-            {
-               Vector::operator[](i+c*num_vectors) = Vector::operator[](i+c*old_nv);
-            }
-         }
+         // create deep copy of old data that will be copied
+         Vector old_slice;
+         old_slice.MakeRef(*this, 0, old_nv * vdim);
+         Vector old_copy(old_slice);
 
-         // Zero-out data now associated with new Vectors
-         for (int c = 0; c < vdim; c++)
+         const auto d_src = old_copy.Read(use_dev);
+         const int vdim_ = vdim;
+
+         // Shift entries for byNODES
+         mfem::forall_switch(use_dev, old_nv * vdim_,
+                             [=] MFEM_HOST_DEVICE (int k)
          {
-            for (int i = old_nv; i < num_vectors; i++)
-            {
-               Vector::operator[](i+c*num_vectors) = 0.0;
-            }
-         }
+            const int d = k / old_nv;
+            const int i = k % old_nv;
+            d_dest[i + d*num_vectors] = d_src[k];
+         });
+
+         // Zero-out new data slots
+         const int diff = num_vectors - old_nv;
+         mfem::forall_switch(use_dev, diff * vdim,
+                             [=] MFEM_HOST_DEVICE (int k)
+         {
+            const int d = k / diff;
+            const int i = k % diff;
+            d_dest[d * num_vectors + old_nv + i] = 0.0;
+         });
       }
       else // byVDIM
       {
-         for (int i = old_nv*vdim; i < num_vectors*vdim; i++)
+         const int start_idx = old_nv * vdim;
+         const int end_idx = num_vectors * vdim;
+         const int diff = end_idx - start_idx;
+         mfem::forall_switch(use_dev, diff, [=] MFEM_HOST_DEVICE (int i)
          {
-            data[i] = 0.0;
-         }
+            d_dest[start_idx + i] = 0.0;
+         });
       }
    }
    else // Else just remove the trailing vector data

miniapps/common/particles_extras.cpp

@@ -126,6 +126,7 @@ void VisualizeParticles(socketstream &sock, const char* vishost, int visport,
    {
       Vector pcoords;
       pset.Coords().GetValues(i, pcoords);
+      pcoords.HostRead();
       if (dim == 2)
       {
          Add2DPoint(pcoords, particles_mesh, psize);
@@ -139,6 +140,7 @@ void VisualizeParticles(socketstream &sock, const char* vishost, int visport,
 
    FiniteElementSpace fes(&particles_mesh, &l2fec, 1);
    GridFunction gf(&fes);
+   gf.HostWrite();
 
    for (int i = 0; i < pset.GetNParticles(); i++)
    {
@@ -193,6 +195,7 @@ void ParticleTrajectories::AddSegmentStart()
    {
       Vector pcoords;
       pset.Coords().GetValues(i, pcoords);
+      pcoords.HostRead();
       segment_meshes.front().AddVertex(pcoords);
    }
 }
@@ -213,6 +216,7 @@ void ParticleTrajectories::SetSegmentEnd()
       {
          Vector pcoords;
          pset.Coords().GetValues(pidx, pcoords);
+         pcoords.HostRead();
          segment_meshes.front().AddVertex(pcoords);
       }
       else // Otherwise set its end vertex == start vertex

miniapps/electromagnetics/lorentz.cpp

@@ -91,6 +91,7 @@ struct LorentzContext
    int nt = 1000;                  // number of timesteps
    int redist_interval = 5;        // redistribution interval
    int redist_mesh = 0;            // redistribution mesh: 0: E mesh, 1: B mesh
+   std::string device_config = "cpu";
 } ctx;
 
 /// This class implements the Boris algorithm as described in the article
@@ -130,7 +131,7 @@ protected:
 public:
 
    Boris(MPI_Comm comm, GridFunction *E_gf_, GridFunction *B_gf_,
-         int nparticles, Ordering::Type pdata_ordering);
+         int nparticles, Ordering::Type pdata_ordering, bool use_device);
 
    /// Find Particles in mesh corresponding to E and B fields
    void FindParticles();
@@ -139,9 +140,12 @@ public:
    /// right after FindParticles has been called.
    void EvaluateFieldsAtParticles();
 
-   /// Advance particles one time step using Boris algorithm
+   /// Advance particles one time step using Boris algorithm. Host version.
    void Step(real_t &t, real_t &dt);
 
+   /// Advance particles one time step using Boris algorithm. Device version.
+   void StepDevice(real_t &t, real_t &dt);
+
    /// Remove lost particles and return their indices
    Array<int> RemoveLostParticles();
 
@@ -235,6 +239,8 @@ int main(int argc, char *argv[])
    args.AddOption(&vis_interval, "-vf", "--vis-interval",
                   "GLVis visualization update after this many timesteps. "
                   "0 means no visualization.");
+   args.AddOption(&ctx.device_config, "-d", "--device",
+                  "Device configuration definition string.");
 
    args.Parse();
    if (!args.Good())
@@ -251,6 +257,10 @@ int main(int argc, char *argv[])
       args.PrintOptions(cout);
    }
 
+   Device device(ctx.device_config);
+   if (Mpi::Root()) { device.Print(); }
+   bool use_device = (ctx.device_config != "cpu") && Device::IsEnabled();
+
    std::unique_ptr<VisItDataCollection> E_dc, B_dc;
    ParGridFunction *E_gf = nullptr, *B_gf = nullptr;
    Vector bb_xmin, bb_xmax;
@@ -266,6 +276,7 @@ int main(int argc, char *argv[])
          return 1;
       }
       E_gf->ParFESpace()->GetParMesh()->GetBoundingBox(bb_xmin, bb_xmax, 2);
+      E_gf->UseDevice(use_device);
    }
 
    // Read B field if provided
@@ -280,6 +291,7 @@ int main(int argc, char *argv[])
       }
       Vector bb_xmint, bb_xmaxt;
       B_gf->ParFESpace()->GetParMesh()->GetBoundingBox(bb_xmint, bb_xmaxt, 2);
+      B_gf->UseDevice(use_device);
       if (ctx.E.coll_name != "")
       {
          // compute intersection of bounding boxes
@@ -302,10 +314,14 @@ int main(int argc, char *argv[])
    // Initialize particles
    int num_particles = ctx.npt/num_ranks +
                        (rank < (ctx.npt % num_ranks) ? 1 : 0);
-   Boris boris(MPI_COMM_WORLD, E_gf, B_gf, num_particles, ordering_type);
+   Boris boris(MPI_COMM_WORLD, E_gf, B_gf, num_particles, ordering_type,
+               use_device);
    InitializeChargedParticles(boris.GetParticles(), ctx.x_min, ctx.x_max,
                               ctx.p_min, ctx.p_max, ctx.m, ctx.q);
+
+   Array<int> removed_idxs_dummy;
    boris.FindParticles();
+   boris.Redistribute(ctx.redist_mesh, removed_idxs_dummy);
    boris.EvaluateFieldsAtParticles();
 
    real_t t = 0.0;
@@ -329,7 +345,14 @@ int main(int argc, char *argv[])
    for (int step = 1; step <= ctx.nt; step++)
    {
       // Step the Boris algorithm
-      boris.Step(t, dt);
+      if (use_device)
+      {
+         boris.StepDevice(t, dt);
+      }
+      else
+      {
+         boris.Step(t, dt);
+      }
       if (Mpi::Root())
       {
          mfem::out << "Step: " << step << " | Time: " << t << endl;
@@ -397,7 +420,7 @@ void Boris::ParticleStep(Particle &part, real_t &dt)
 }
 
 Boris::Boris(MPI_Comm comm, GridFunction *E_gf_, GridFunction *B_gf_,
-             int nparticles, Ordering::Type pdata_ordering)
+             int nparticles, Ordering::Type pdata_ordering, bool use_device)
    : E_gf(E_gf_),
      B_gf(B_gf_),
      E_finder(comm),
@@ -426,6 +449,7 @@ Boris::Boris(MPI_Comm comm, GridFunction *E_gf_, GridFunction *B_gf_,
    }
 
    int dim = E_mesh ? E_mesh->SpaceDimension() : B_mesh->SpaceDimension();
+   MFEM_VERIFY(dim == 3, "Only 3D meshes are currently supported.");
 
    pxB_.SetSize(dim); pm_.SetSize(dim); pp_.SetSize(dim);
 
@@ -435,7 +459,8 @@ Boris::Boris(MPI_Comm comm, GridFunction *E_gf_, GridFunction *B_gf_,
    Array<int> field_vdims({1, 1, dim, dim, dim});
 
    charged_particles = std::make_unique<ParticleSet>
-                       (comm, nparticles, dim, field_vdims, 0, pdata_ordering);
+                       (comm, nparticles, dim, field_vdims, 0, pdata_ordering,
+                        use_device);
 }
 
 void Boris::FindParticles()
@@ -481,7 +506,6 @@ void Boris::Step(real_t &t, real_t &dt)
 {
    // Interpolate E and B fields onto particles
    EvaluateFieldsAtParticles();
-
    // Individually step each particle. If all ParticleSet fields are ordered
    // byVDIM, we can use GetParticleRef for better performance.
    if (charged_particles->IsParticleRefValid())
@@ -509,6 +533,112 @@ void Boris::Step(real_t &t, real_t &dt)
    t += dt;
 }
 
+void Boris::StepDevice(real_t &t, real_t &dt)
+{
+   // Interpolate E and B fields onto particles
+   EvaluateFieldsAtParticles();
+   const int N = charged_particles->GetNParticles();
+   auto &X = charged_particles->Coords();
+   auto &M = charged_particles->Field(MASS);
+   auto &Q = charged_particles->Field(CHARGE);
+   auto &P = charged_particles->Field(MOM);
+   auto &E = charged_particles->Field(EFIELD);
+   auto &B = charged_particles->Field(BFIELD);
+
+   const int dim = X.GetVDim();
+
+   // Capture orderings for each field to ensure correct access
+   const bool byVDIM_X = (X.GetOrdering() == Ordering::byVDIM);
+   const bool byVDIM_P = (P.GetOrdering() == Ordering::byVDIM);
+   const bool byVDIM_E = (E.GetOrdering() == Ordering::byVDIM);
+   const bool byVDIM_B = (B.GetOrdering() == Ordering::byVDIM);
+
+   auto d_x = X.ReadWrite();
+   auto d_m = M.Read();
+   auto d_q = Q.Read();
+   auto d_p = P.ReadWrite();
+   auto d_e = E.Read();
+   auto d_b = B.Read();
+
+   mfem::forall(N, [=] MFEM_HOST_DEVICE (int i)
+   {
+      const real_t m = d_m[i];
+      const real_t q = d_q[i];
+
+      real_t x[3], p[3], e[3], b[3];
+      // Load data
+      for (int d = 0; d < dim; d++)
+      {
+         x[d] = d_x[byVDIM_X ? i * dim + d : i + d * N];
+         p[d] = d_p[byVDIM_P ? i * dim + d : i + d * N];
+         e[d] = d_e[byVDIM_E ? i * dim + d : i + d * N];
+         b[d] = d_b[byVDIM_B ? i * dim + d : i + d * N];
+      }
+
+      // Boris algorithm implementation
+      real_t pm[3], pxB[3], pp[3];
+
+      // Compute half of the contribution from q E
+      // pm = p + 0.5 * dt * q * e
+      for (int d = 0; d < dim; d++)
+      {
+         pm[d] = p[d] + (0.5 * dt * q) * e[d];
+      }
+
+      // Compute the contribution from q p x B
+      real_t B2 = 0.0;
+      for (int d = 0; d < dim; d++) { B2 += b[d] * b[d]; }
+
+      // ... along pm x B
+      // pxB = pm x b
+      pxB[0] = pm[1] * b[2] - pm[2] * b[1];
+      pxB[1] = pm[2] * b[0] - pm[0] * b[2];
+      pxB[2] = pm[0] * b[1] - pm[1] * b[0];
+
+      // pp = a1 * pxB
+      const real_t a1 = 4.0 * dt * q * m;
+      for (int d = 0; d < dim; d++) { pp[d] = a1 * pxB[d]; }
+
+      // ... along pm
+      // pp += a2 * pm
+      const real_t a2 = 4.0 * m * m - dt * dt * q * q * B2;
+      for (int d = 0; d < dim; d++) { pp[d] += a2 * pm[d]; }
+
+      // ... along B
+      real_t b_dot_pm = 0.0;
+      for (int d = 0; d < dim; d++) { b_dot_pm += b[d] * pm[d]; }
+      const real_t a3 = 2.0 * dt * dt * q * q * b_dot_pm;
+      // pp += a3 * b
+      for (int d = 0; d < dim; d++) { pp[d] += a3 * b[d]; }
+
+      // scale by common denominator
+      const real_t a4 = 4.0 * m * m + dt * dt * q * q * B2;
+      for (int d = 0; d < dim; d++) { pp[d] /= a4; }
+
+      // Update the momentum
+      // p = pp + 0.5 * dt * q * e
+      for (int d = 0; d < dim; d++)
+      {
+         p[d] = pp[d] + (0.5 * dt * q) * e[d];
+      }
+
+      // Update the position
+      // x += (dt / m) * p
+      // Store back to global arrays
+      for (int d = 0; d < dim; d++)
+      {
+         d_p[byVDIM_P ? i * dim + d : i + d * N] = p[d];
+         d_x[byVDIM_X ? i * dim + d : i + d * N] = x[d] + (dt / m) * p[d];
+      }
+   });
+
+   // Find updated particle locations in E and B field meshes
+   FindParticles();
+
+   // Update time
+   t += dt;
+}
+
 Array<int> Boris::RemoveLostParticles()
 {
    Array<int> lost_idxs;
@@ -617,6 +747,11 @@ void InitializeChargedParticles(ParticleSet &charged_particles,
    ParticleVector &M = charged_particles.Field(Boris::MASS);
    ParticleVector &Q = charged_particles.Field(Boris::CHARGE);
 
+   X.HostWrite();
+   P.HostWrite();
+   M.HostWrite();
+   Q.HostWrite();
+
    for (int i = 0; i < charged_particles.GetNParticles(); i++)
    {
       for (int d = 0; d < dim; d++)
@@ -643,4 +778,9 @@ void InitializeChargedParticles(ParticleSet &charged_particles,
       M(i) = m;
       Q(i) = q;
    }
+
+   X.Read();
+   P.Read();
+   M.Read();
+   Q.Read();
 }

miniapps/gslib/findpts.cpp

@@ -49,6 +49,9 @@
 //    findpts -m ../../data/ref-square.mesh -o 2 -mo 1 -random 1 -surf
 //    findpts -m ../../data/ref-cube.mesh -o 2 -mo 1 -random 1 -surf
 //    findpts -m ../../data/square-disc-p2.mesh -o 4 -mo 2 -random 1 -surf
+// Surface meshes + bounding box size increase:
+//    findpts -m ../../data/square-disc-p2.mesh -o 4 -mo 2 -random 1 -surf -sabs 0.1
+//    findpts -m ../../data/tinyzoo-3d.mesh -o 4 -mo 2 -random 1 -surf -sabs 0.1
 
 #include "mfem.hpp"
 #include "../common/mfem-common.hpp"
@@ -109,6 +112,7 @@ int main (int argc, char *argv[])
    int randomization     = 0;
    int npt               = 100;
    bool surface          = false;
+   double surf_aabb_sz_inc = 0.0;
 
    // Parse command-line options.
    OptionsParser args(argc, argv);
@@ -150,6 +154,9 @@ int main (int argc, char *argv[])
    args.AddOption(&surface, "-surf", "--surface", "-no-surf",
                   "--no-surface",
                   "Extract surface mesh from volume mesh.");
+   args.AddOption(&surf_aabb_sz_inc, "-sabs", "--surface-aabb-size-inc",
+                  "Absolute AABB expansion applied to surface-search "
+                  "axis-aligned bounding boxes in FindPointsGSLIB surface meshes.");
 
    args.Parse();
    if (!args.Good())
@@ -384,8 +391,17 @@ int main (int argc, char *argv[])
 
    // Find and Interpolate FE function values on the desired points.
    Vector interp_vals(pts_cnt*vec_dim);
-   FindPointsGSLIB finder(*mesh);
-   finder.SetDistanceToleranceForPointsFoundOnBoundary(10);
+   FindPointsGSLIB finder;
+   if (surface && surf_aabb_sz_inc > 0.0)
+   {
+      Vector bb_size({surf_aabb_sz_inc});
+      finder.SetupSurfWithAABBExpansion(*mesh, bb_size);
+   }
+   else
+   {
+      finder.Setup(*mesh);
+      // finder.SetDistanceToleranceForPointsFoundOnBoundary(10);
+   }
    finder.SetL2AvgType(FindPointsGSLIB::NONE);
    finder.Interpolate(vxyz, field_vals, interp_vals, point_ordering);
    Array<unsigned int> code_out    = finder.GetCode();
@@ -424,7 +440,7 @@ int main (int argc, char *argv[])
         << "Searched points:     "   << pts_cnt
         << "\nFound points:        " << found
         << "\nMax interp error:    " << max_err
-        << "\nMax dist (of found): " << max_dist
+        << "\nMax dist^2 (of found): " << max_dist
         << "\nPoints not found:    " << not_found;
    if (randomization == 1)
    {

miniapps/gslib/pfindpts.cpp

@@ -48,11 +48,14 @@
 // Device runs:
 //    mpirun -np 2 pfindpts -m ../../data/inline-quad.mesh -o 3 -mo 2 -random 1 -d debug
 //    mpirun -np 2 pfindpts -m ../../data/amr-quad.mesh -rs 1 -o 4 -mo 2 -random 1 -npt 100 -d debug
-//    mpirun -np 2 pfindpts -m ../../data/inline-hex.mesh -o 3 -mo 2 -random 1 -d debug
+//    mpirun -np 2 pfindpts -m ../../data/inline-hex.mesh -o 3 -mo 2 -random 1 -d debug -ft 1
 // Surface meshes:
 //    mpirun -np 4 pfindpts -m ../../data/square-disc-p2.mesh -o 4 -mo 2 -vis -random 1 -surf
 //    mpirun -np 4 pfindpts -m ../../data/star-q3.mesh -o 6 -mo 3 -vis -random 1 -surf
 //    mpirun -np 4 pfindpts -m ../../data/fichera-q2.mesh -o 6 -mo 3 -vis -random 1 -surf
+// Surface meshes + bounding box size increase:
+//    mpirun -np 4 pfindpts -m ../../data/square-disc-p2.mesh -o 4 -mo 2 -vis -random 1 -surf -sabs 0.1
+//    mpirun -np 4 pfindpts -m ../../data/tinyzoo-3d.mesh -o 4 -mo 2 -vis -random 1 -surf -sabs 0.1
 
 #include "mfem.hpp"
 #include "../common/mfem-common.hpp"
@@ -102,6 +105,7 @@ int main (int argc, char *argv[])
    int randomization     = 0;
    int npt               = 100; //points per proc
    bool surface          = false;
+   double surf_aabb_sz_inc = 0.0;
 
    // Parse command-line options.
    OptionsParser args(argc, argv);
@@ -145,7 +149,9 @@ int main (int argc, char *argv[])
    args.AddOption(&surface, "-surf", "--surface", "-no-surf",
                   "--no-surface",
                   "Extract surface mesh from volume mesh.");
-
+   args.AddOption(&surf_aabb_sz_inc, "-sabs", "--surface-aabb-size-inc",
+                  "Absolute AABB expansion applied to surface-search "
+                  "axis-aligned bounding boxes in FindPointsGSLIB surface meshes.");
    args.Parse();
    if (!args.Good())
    {
@@ -343,7 +349,7 @@ int main (int argc, char *argv[])
             Geometry::GetRandomPoint(geom, ip);
             if (j < npt_face_per_elem)
             {
-               ip.x = 0.0; // force point to be on the face
+               ip.x = 0.0; // force point to be on a face
                npt_total_face++;
             }
             Vector pos_i(sdim);
@@ -373,8 +379,17 @@ int main (int argc, char *argv[])
 
    // Find and Interpolate FE function values on the desired points.
    Vector interp_vals(pts_cnt*vec_dim);
-   FindPointsGSLIB finder(pmesh);
-   finder.SetDistanceToleranceForPointsFoundOnBoundary(10);
+   FindPointsGSLIB finder;
+   if (surface && surf_aabb_sz_inc > 0.0)
+   {
+      Vector bb_size({surf_aabb_sz_inc});
+      finder.SetupSurfWithAABBExpansion(pmesh, bb_size);
+   }
+   else
+   {
+      finder.Setup(pmesh);
+   }
+   // finder.SetDistanceToleranceForPointsFoundOnBoundary(1e-10);
    // Enable GPU to CPU fallback for GPUData only if you are using an older
    // version of GSLIB.
    // finder.SetGPUtoCPUFallback(true);
@@ -456,10 +471,11 @@ int main (int argc, char *argv[])
            << "\nPoints on faces:      " << face_pts << " out of "
            << npt_total_face
            << "\nMax interp error:     " << max_error
-           << "\nMax dist (of found):  " << max_dist
+           << "\nMax dist^2 (of found): " << max_dist
            << endl;
    }
 
+
    delete fec;
 
    if (randomization != 0)

miniapps/spde/spde_solver.hpp

@@ -83,7 +83,7 @@ real_t IntegrateBC(const ParGridFunction &x, const Array<int> &bdr,
 /// where A is
 /// A = div ( Theta(x) grad + Id ) u(x)
 /// and alpha is given as
-/// alpha = (2 nu + dim) / 2.
+/// alpha = (2 nu + dim) / 4.
 /// Theta (anisotropy tensor) and nu (smoothness) can be specified in the
 /// constructor. Traditionally, the SPDE method requires the specification of
 /// a white noise right hands side. SPDESolver accepts arbitrary right hand

tests/benchmarks/CMakeLists.txt

@@ -56,3 +56,4 @@ add_benchmark(elasticity)
 add_benchmark(tmop)
 add_benchmark(vector)
 add_benchmark(virtuals)
+add_benchmark(nlvc)

tests/benchmarks/bench_nlvc.cpp

@@ -0,0 +1,244 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#include "bench.hpp" // IWYU pragma: keep
+
+#ifdef MFEM_USE_BENCHMARK
+
+#include <cassert>
+#include <cstdlib>
+#include <functional>
+
+#include "fem/qinterp/grad.hpp"
+
+using namespace mfem;
+
+// Custom benchmark arguments generator ///////////////////////////////////////
+static void CustomArguments(bm::Benchmark *b) noexcept
+{
+   constexpr int MAX_NDOFS = 8 * 1024 * (mfem_use_gpu ? 1024 : 8);
+
+   const auto orders = { 6, 5, 4, 3, 2, 1 };
+
+   constexpr auto ndofs = [](int n) constexpr noexcept -> int
+   {
+      return (n + 1) * (n + 1) * (n + 1);
+   };
+
+   constexpr auto inc = [](int n) constexpr noexcept -> int
+   {
+      return n < 160 ?  4 : n < 240 ?  8 : n < 320 ? 16 : 32;
+   };
+
+   for (auto p : orders)
+   {
+      for (int n = (mfem_use_gpu ? 16 : 8); ndofs(n) <= MAX_NDOFS; n += inc(n))
+      {
+         b->Args({p, n});
+      }
+   }
+}
+
+/// Basic Kernels Specializations /////////////////////////////////////////////
+static void AddBasicKernelSpecializations()
+{
+   using Grad = QuadratureInterpolator::GradKernels;
+   // 2D
+   Grad::Specialization<2, QVectorLayout::byNODES, false, 2,2,7>::Add();
+   Grad::Specialization<2, QVectorLayout::byNODES, false, 2,2,8>::Add();
+   Grad::Specialization<2, QVectorLayout::byNODES, false, 2,2,10>::Add();
+   // 3D
+   Grad::Specialization<3, QVectorLayout::byNODES, false, 3,2,7>::Add();
+   Grad::Specialization<3, QVectorLayout::byNODES, false, 3,2,9>::Add();
+   Grad::Specialization<3, QVectorLayout::byNODES, false, 3,2,10>::Add();
+}
+
+/// VectorConvectionNLFBenchmark //////////////////////////////////////////////
+template <int DIM>
+struct VectorConvectionNLFBenchmark
+{
+   const int p, c, q, n, nx, ny, nz;
+   const std::function<Mesh()> MakeCartesianMesh = [&]()
+   {
+      if constexpr (DIM == 2)
+      {
+         return Mesh::MakeCartesian2D(nx, ny, Element::QUADRILATERAL);
+      }
+      else
+      {
+         return Mesh::MakeCartesian3D(nx, ny, nz, Element::HEXAHEDRON);
+      }
+   };
+   Mesh mesh;
+   H1_FECollection fec;
+   FiniteElementSpace fes;
+   const Geometry::Type geom_type;
+   IntegrationRules irs;
+   const IntegrationRule *ir;
+   ConstantCoefficient const_coeff { M_2_SQRTPI };
+   NonlinearFormIntegrator *nlfi;
+   NonlinearForm nlf;
+   Operator *grad;
+   GridFunction x, dx, y_pa;
+   Vector xe, dxe, ye;
+   const int dofs;
+   const int q1d;
+   double mdofs{};
+
+   VectorConvectionNLFBenchmark(int p, int side):
+      p(p), c(side), q(2 * p + 3), n((assert(c >= p), c / p)),
+      nx(n + (p * (n + 1) * p * n * p * n < c * c * c ? 1 : 0)),
+      ny(n + (p * (n + 1) * p * (n + 1) * p * n < c * c * c ? 1 : 0)), nz(n),
+      mesh(MakeCartesianMesh()),
+      fec(p, DIM),
+      fes(&mesh, &fec, DIM),
+      geom_type(mesh.GetTypicalElementGeometry()),
+      irs(0, Quadrature1D::GaussLegendre),
+      ir(&irs.Get(geom_type, q)),
+      nlfi(new VectorConvectionNLFIntegrator(const_coeff)),
+      nlf(&fes),
+      x(&fes),
+      dx(&fes),
+      y_pa(&fes),
+      dofs(fes.GetTrueVSize()),
+      q1d(IntRules.Get(Geometry::SEGMENT, ir->GetOrder()).GetNPoints())
+   {
+      MFEM_VERIFY(q1d*q1d*(DIM == 3 ? q1d : 1) == ir->GetNPoints(), "");
+
+      nlf.SetAssemblyLevel(AssemblyLevel::PARTIAL);
+      nlf.AddDomainIntegrator(nlfi);
+      nlf.Setup();
+
+      dx.Randomize(0x9e3779b9), x.Randomize(0x100001b3);
+
+      grad = &nlf.GetGradient(x);
+
+      const Table &el2dof = fes.GetElementToDofTable();
+      const int e_size = el2dof.Size_of_connections()*fes.GetVDim();
+      const auto R = fes.GetElementRestriction(ElementDofOrdering::LEXICOGRAPHIC);
+      MFEM_VERIFY(e_size == R->Height(), "Input/Output E-vector size mismatch!");
+      xe.SetSize(R->Height()), dxe.SetSize(R->Height()), ye.SetSize(R->Height());
+      xe.UseDevice(true), dxe.UseDevice(true), ye.UseDevice(true);
+      xe.Randomize(0x100001b3), dxe.Randomize(0x9e3779b9), ye = 0.0;
+
+      mdofs = 0.0;
+   }
+
+   void Setup()
+   {
+      nlfi->AssembleGradPA(xe, fes);
+      MFEM_DEVICE_SYNC;
+      mdofs += this->MDofs();
+   }
+
+   void AddMult()
+   {
+      nlf.AddMult(x, y_pa);
+      MFEM_DEVICE_SYNC;
+      mdofs += this->MDofs();
+   }
+
+   void AddMultPA()
+   {
+      nlfi->AddMultPA(xe, ye);
+      MFEM_DEVICE_SYNC;
+      mdofs += this->MDofs();
+   }
+
+   void AddMultGrad()
+   {
+      grad->Mult(dx, y_pa);
+      MFEM_DEVICE_SYNC;
+      mdofs += this->MDofs();
+   }
+
+   void AddMultGradPA()
+   {
+      nlfi->AddMultGradPA(dxe, ye);
+      MFEM_DEVICE_SYNC;
+      mdofs += this->MDofs();
+   }
+
+   void AssembleGradDiagonal()
+   {
+      grad->AssembleDiagonal(ye);
+      MFEM_DEVICE_SYNC;
+      mdofs += this->MDofs();
+   }
+
+   [[nodiscard]] double SumMdofs() const noexcept { return mdofs; }
+
+   [[nodiscard]] double MDofs() const noexcept { return 1e-6 * dofs; }
+};
+
+///////////////////////////////////////////////////////////////////////////////
+#define RegisterVectorConvectionNLFBenchmark(Benchmark, DIM)         \
+   static void Benchmark##DIM##d(bm::State &state)                   \
+   {                                                                 \
+      const auto order = static_cast<int>(state.range(0));           \
+      const auto side = static_cast<int>(state.range(1));            \
+      VectorConvectionNLFBenchmark<DIM> ker(order, side);            \
+      while (state.KeepRunning()) { ker.Benchmark(); }               \
+      bm::Counter::Flags flags = bm::Counter::kIsRate;               \
+      state.counters["MDof/s"] = bm::Counter(ker.SumMdofs(), flags); \
+      state.counters["Dofs"] = bm::Counter(ker.dofs);                \
+      state.counters["p"] = bm::Counter(order);                      \
+   }                                                                 \
+   BENCHMARK(Benchmark##DIM##d)                                      \
+      ->Apply(CustomArguments)                                       \
+      ->Unit(bm::kMillisecond)
+
+RegisterVectorConvectionNLFBenchmark(Setup,3);
+RegisterVectorConvectionNLFBenchmark(AddMult,3);
+RegisterVectorConvectionNLFBenchmark(AddMultPA,3);
+RegisterVectorConvectionNLFBenchmark(AddMultGrad,3);
+RegisterVectorConvectionNLFBenchmark(AddMultGradPA,3);
+RegisterVectorConvectionNLFBenchmark(AssembleGradDiagonal,3);
+
+RegisterVectorConvectionNLFBenchmark(Setup,2);
+RegisterVectorConvectionNLFBenchmark(AddMult,2);
+RegisterVectorConvectionNLFBenchmark(AddMultPA,2);
+RegisterVectorConvectionNLFBenchmark(AddMultGrad,2);
+RegisterVectorConvectionNLFBenchmark(AddMultGradPA,2);
+RegisterVectorConvectionNLFBenchmark(AssembleGradDiagonal,2);
+
+/// main //////////////////////////////////////////////////////////////////////
+int main(int argc, char *argv[])
+{
+   AddBasicKernelSpecializations();
+
+   bm::ConsoleReporter CR;
+   bm::Initialize(&argc, argv);
+
+   // Device setup, cpu by default
+   std::string device_context = "cpu";
+   const auto global_context = bmi::GetGlobalContext();
+   if (global_context != nullptr)
+   {
+      const auto device = global_context->find("device");
+      if (device != global_context->end())
+      {
+         mfem::out << device->first << " : "
+                   << device->second << std::endl;
+         device_context = device->second;
+      }
+   }
+   Device device(device_context.c_str());
+   device.Print();
+
+   if (bm::ReportUnrecognizedArguments(argc, argv)) { return EXIT_FAILURE; }
+
+   bm::RunSpecifiedBenchmarks(&CR);
+
+   return EXIT_SUCCESS;
+}
+
+#endif // MFEM_USE_BENCHMARK

tests/benchmarks/makefile

@@ -21,7 +21,7 @@ MFEM_LIB_FILE = mfem_is_not_built
 -include $(CONFIG_MK)
 
 SEQ_TESTS = bench_assembly_levels bench_ceed bench_dg_amr bench_elasticity \
-            bench_tmop bench_vector bench_virtuals
+            bench_nlvc bench_tmop bench_vector bench_virtuals
 PAR_TESTS = 
 ifeq ($(MFEM_USE_MPI),NO)
    TESTS = $(SEQ_TESTS)

tests/unit/CMakeLists.txt

@@ -147,6 +147,8 @@ set(UNIT_TESTS_SRCS
   fem/test_pa_grad.cpp
   fem/test_pa_idinterp.cpp
   fem/test_pa_kernels.cpp
+  fem/test_pa_vecdiv.cpp
+  fem/test_pa_nlvc.cpp
   fem/test_pa_simplices.cpp
   fem/test_particleset.cpp
   fem/test_pgridfunc_save_serial.cpp

tests/unit/fem/test_gslib.cpp

@@ -12,6 +12,8 @@
 #include "unit_tests.hpp"
 #include "mfem.hpp"
 
+#include <random>
+
 using namespace mfem;
 #ifdef MFEM_USE_GSLIB
 namespace gslib_test
@@ -36,6 +38,165 @@ void F_exact(const Vector &p, Vector &F)
 
 enum class Space { H1, L2 };
 
+enum class SurfaceMeshType { Segment2D, Segment3D, Quad3D, Tri3D };
+
+const char *SurfaceMeshName(const SurfaceMeshType type)
+{
+   switch (type)
+   {
+      case SurfaceMeshType::Segment2D: return "segment-2d";
+      case SurfaceMeshType::Segment3D: return "segment-3d";
+      case SurfaceMeshType::Quad3D:    return "quad-3d";
+      case SurfaceMeshType::Tri3D:     return "tri-3d";
+   }
+   return "unknown";
+}
+
+int SurfaceSpaceDim(const SurfaceMeshType type)
+{
+   switch (type)
+   {
+      case SurfaceMeshType::Segment2D: return 2;
+      case SurfaceMeshType::Segment3D: return 3;
+      case SurfaceMeshType::Quad3D:    return 3;
+      case SurfaceMeshType::Tri3D:     return 3;
+   }
+   return -1;
+}
+
+Mesh MakeSurfaceMesh(const SurfaceMeshType type, const int ne)
+{
+   switch (type)
+   {
+      case SurfaceMeshType::Segment2D:
+         return Mesh::MakeCartesian1D(ne);
+      case SurfaceMeshType::Segment3D:
+         return Mesh::MakeCartesian1D(ne);
+      case SurfaceMeshType::Quad3D:
+         return Mesh::MakeCartesian2D(ne, ne, Element::QUADRILATERAL);
+      case SurfaceMeshType::Tri3D:
+         return Mesh::MakeCartesian2D(ne, ne, Element::TRIANGLE);
+   }
+   MFEM_ABORT("Unknown surface mesh type.");
+   return Mesh();
+}
+
+void GetSurfaceInteriorPoints(Mesh &mesh, const int npt_per_el,
+                              const int ordering, Vector &xyz,
+                              const int p0 = 0)
+{
+   MFEM_VERIFY(mesh.GetNodes() != nullptr, "Mesh nodes are required.");
+
+   const int sdim = mesh.SpaceDimension();
+   const int npt = xyz.Size()/sdim;
+   MFEM_VERIFY((p0 + mesh.GetNE()*npt_per_el)*sdim <= xyz.Size(),
+               "Output vector is too small.");
+   Vector point(sdim);
+   std::mt19937 gen(123);
+   std::uniform_real_distribution<double> uni(0.01, 0.99);
+   int p = p0;
+
+   for (int e = 0; e < mesh.GetNE(); e++)
+   {
+      ElementTransformation *T = mesh.GetElementTransformation(e);
+      const Geometry::Type geom = mesh.GetElementBaseGeometry(e);
+      for (int j = 0; j < npt_per_el; j++)
+      {
+         IntegrationPoint ip;
+         real_t xv = uni(gen);
+         if (geom == Geometry::SEGMENT)
+         {
+            ip.x = xv;
+         }
+         else if (geom == Geometry::SQUARE)
+         {
+            ip.Set2(xv, uni(gen));
+         }
+         else
+         {
+            MFEM_VERIFY(geom == Geometry::TRIANGLE,
+                        "Unsupported surface element geometry.");
+            ip.Set2(xv, uni(gen)*(1.0 - xv));
+         }
+         T->Transform(ip, point);
+         for (int d = 0; d < sdim; d++)
+         {
+            const int idx = (ordering == Ordering::byNODES) ?
+                            d*npt + p :
+                            p*sdim + d;
+            xyz(idx) = point(d);
+         }
+         p++;
+      }
+   }
+}
+
+void GetSurfaceBoundaryPoints(Mesh &mesh, const int npt_per_el,
+                              const int ordering, Vector &xyz,
+                              const int p0 = 0)
+{
+   MFEM_VERIFY(mesh.GetNodes() != nullptr, "Mesh nodes are required.");
+
+   const int sdim = mesh.SpaceDimension();
+   const int npt = xyz.Size()/sdim;
+   MFEM_VERIFY((p0 + mesh.GetNE()*npt_per_el)*sdim <= xyz.Size(),
+               "Output vector is too small.");
+   Vector point(sdim);
+   std::mt19937 gen(246);
+   std::uniform_real_distribution<double> uni(0.01, 0.99);
+   int p = p0;
+
+   for (int e = 0; e < mesh.GetNE(); e++)
+   {
+      ElementTransformation *T = mesh.GetElementTransformation(e);
+      const Geometry::Type geom = mesh.GetElementBaseGeometry(e);
+      for (int j = 0; j < npt_per_el; j++)
+      {
+         IntegrationPoint ip;
+         if (geom == Geometry::SEGMENT)
+         {
+            MFEM_VERIFY(npt_per_el == 2,
+                        "Segment boundary sampling requires npt_per_el = 2.");
+            ip.x = (j == 0) ? 0.0 : 1.0;
+         }
+         else
+         {
+            const double t = uni(gen);
+            if (geom == Geometry::SQUARE)
+            {
+               switch (j % 4)
+               {
+                  case 0: ip.Set2(t, 0.0);     break;
+                  case 1: ip.Set2(1.0, t);     break;
+                  case 2: ip.Set2(t, 1.0);     break;
+                  case 3: ip.Set2(0.0, t);     break;
+               }
+            }
+            else
+            {
+               MFEM_VERIFY(geom == Geometry::TRIANGLE,
+                           "Unsupported surface element geometry.");
+               switch (j % 3)
+               {
+                  case 0: ip.Set2(t, 0.0);     break;
+                  case 1: ip.Set2(t, 1.0 - t); break;
+                  case 2: ip.Set2(0.0, t);     break;
+               }
+            }
+         }
+         T->Transform(ip, point);
+         for (int d = 0; d < sdim; d++)
+         {
+            const int idx = (ordering == Ordering::byNODES) ?
+                            d*npt + p :
+                            p*sdim + d;
+            xyz(idx) = point(d);
+         }
+         p++;
+      }
+   }
+}
+
 TEST_CASE("GSLIBInterpolate", "[GSLIBInterpolate][GSLIB]")
 {
    auto space               = GENERATE(Space::H1, Space::L2);
@@ -190,6 +351,92 @@ TEST_CASE("GSLIBInterpolate", "[GSLIBInterpolate][GSLIB]")
    delete c_fec;
 }
 
+TEST_CASE("GSLIBSurfInterpolate", "[GSLIBSurfInterpolate][GSLIB]")
+{
+   auto surface_mesh_type   = GENERATE(SurfaceMeshType::Segment2D,
+                                       SurfaceMeshType::Segment3D,
+                                       SurfaceMeshType::Quad3D,
+                                       SurfaceMeshType::Tri3D);
+   func_order               = GENERATE(1, 2);
+   int mesh_order           = GENERATE(1, 2);
+   int mesh_node_ordering   = GENERATE(0, 1);
+   int point_ordering       = GENERATE(0, 1);
+   int ncomp                = GENERATE(1, 2);
+   int gf_ordering          = GENERATE(0, 1);
+   int func_out_ordering    = GENERATE(0, 1);
+
+   const char *mesh_name = SurfaceMeshName(surface_mesh_type);
+   CAPTURE(mesh_name, func_order, mesh_order, mesh_node_ordering,
+           point_ordering, ncomp, gf_ordering, func_out_ordering);
+
+   if (ncomp == 1 && gf_ordering == 1)
+   {
+      return;
+   }
+
+   Mesh mesh = MakeSurfaceMesh(surface_mesh_type, 4);
+   const int sdim = SurfaceSpaceDim(surface_mesh_type);
+   mesh.SetCurvature(mesh_order, false, sdim, mesh_node_ordering);
+
+   H1_FECollection c_fec(func_order, mesh.Dimension());
+   FiniteElementSpace c_fespace(&mesh, &c_fec, ncomp, gf_ordering);
+   GridFunction field_vals(&c_fespace);
+
+   VectorFunctionCoefficient F(ncomp, F_exact);
+   field_vals.ProjectCoefficient(F);
+
+   const int npt_per_el = 8;
+   const int pts_cnt = mesh.GetNE()*npt_per_el;
+   Vector vxyz(pts_cnt*sdim);
+   GetSurfaceInteriorPoints(mesh, npt_per_el, point_ordering, vxyz);
+
+   Vector interp_vals(pts_cnt*ncomp);
+   FindPointsGSLIB finder;
+   finder.SetupSurf(mesh);
+   finder.SetL2AvgType(FindPointsGSLIB::NONE);
+   finder.Interpolate(vxyz, field_vals, interp_vals, point_ordering,
+                      func_out_ordering);
+
+   Array<unsigned int> code_out = finder.GetCode();
+   Vector dist_p_out = finder.GetDist();
+
+   int not_found = 0;
+   double err = 0.0, max_err = 0.0, max_dist = 0.0;
+   Vector pos(sdim);
+   Vector exact_val(ncomp);
+
+   for (int i = 0; i < pts_cnt; i++)
+   {
+      max_dist = std::max(max_dist, dist_p_out(i));
+      for (int d = 0; d < sdim; d++)
+      {
+         const int idx = (point_ordering == Ordering::byNODES) ?
+                         d*pts_cnt + i :
+                         i*sdim + d;
+         pos(d) = vxyz(idx);
+      }
+      F_exact(pos, exact_val);
+      for (int j = 0; j < ncomp; j++)
+      {
+         if (code_out[i] < 2)
+         {
+            err = func_out_ordering == Ordering::byNODES ?
+                  fabs(exact_val(j) - interp_vals[i + j*pts_cnt]) :
+                  fabs(exact_val(j) - interp_vals[i*ncomp + j]);
+            max_err = std::max(max_err, err);
+         }
+         else if (j == 0)
+         {
+            not_found++;
+         }
+      }
+   }
+
+   REQUIRE(max_err < 1e-12);
+   REQUIRE(max_dist < 1e-10);
+   REQUIRE(not_found == 0);
+}
+
 // Generates meshes with different element types, followed by points at
 // element faces and interior, and finally checks to see if these points are
 // correctly detected at element boundary or not.
@@ -257,9 +504,8 @@ TEST_CASE("GSLIBFindAtElementBoundary",
       int nptface = xyz.Size()/dim;
 
       // Generate points inside each element
-      FiniteElementCollection *l2_fec = new L2_FECollection(l2_order, dim);
-      FiniteElementSpace l2_fespace =
-         FiniteElementSpace(&mesh, l2_fec, 1);
+      L2_FECollection l2_fec(l2_order, dim);
+      FiniteElementSpace l2_fespace(&mesh, &l2_fec, 1);
       DenseMatrix vals;
       DenseMatrix tr;
       for (int e = 0; e < mesh.GetNE(); e++)
@@ -295,7 +541,92 @@ TEST_CASE("GSLIBFindAtElementBoundary",
          cmax = std::max(code_out[i], cmax);
       }
       REQUIRE((cmin == 0 && cmax == 0)); // should be found inside element
-      delete l2_fec;
+   }
+}
+
+TEST_CASE("GSLIBSurfFindAtElementBoundary",
+          "[GSLIBSurfFindAtElementBoundary][GSLIB]")
+{
+   auto surface_mesh_type = GENERATE(SurfaceMeshType::Segment2D,
+                                     SurfaceMeshType::Segment3D,
+                                     SurfaceMeshType::Quad3D,
+                                     SurfaceMeshType::Tri3D);
+   const char *mesh_name = SurfaceMeshName(surface_mesh_type);
+   CAPTURE(mesh_name);
+
+   Mesh mesh = MakeSurfaceMesh(surface_mesh_type, 4);
+   const int sdim = SurfaceSpaceDim(surface_mesh_type);
+   mesh.SetCurvature(2, false, sdim);
+
+   const int nptface_per_el = (mesh.Dimension() == 1) ? 2 : 8;
+   const int nptint_per_el = 8;
+   const int nptface = mesh.GetNE()*nptface_per_el;
+   const int nptint = mesh.GetNE()*nptint_per_el;
+   Vector xyz((nptface + nptint)*sdim);
+   GetSurfaceBoundaryPoints(mesh, nptface_per_el, Ordering::byVDIM, xyz, 0);
+   GetSurfaceInteriorPoints(mesh, nptint_per_el, Ordering::byVDIM, xyz,
+                            nptface);
+
+   FindPointsGSLIB finder;
+   finder.SetupSurf(mesh);
+   finder.FindPoints(xyz, Ordering::byVDIM);
+   Array<unsigned int> code_out = finder.GetCode();
+
+   for (int i = 0; i < nptface; i++)
+   {
+      REQUIRE(code_out[i] == 1);
+   }
+   for (int i = nptface; i < nptface + nptint; i++)
+   {
+      REQUIRE(code_out[i] == 0);
+   }
+}
+
+TEST_CASE("GSLIBSurfAABBExpansion", "[GSLIBSurfAABBExpansion][GSLIB]")
+{
+   auto surface_mesh_type = GENERATE(SurfaceMeshType::Segment2D,
+                                     SurfaceMeshType::Segment3D,
+                                     SurfaceMeshType::Quad3D,
+                                     SurfaceMeshType::Tri3D);
+   const char *mesh_name = SurfaceMeshName(surface_mesh_type);
+   CAPTURE(mesh_name);
+
+   constexpr double offset = 1.0e-3;
+   const int npt_per_el = 8;
+
+   Mesh mesh = MakeSurfaceMesh(surface_mesh_type, 4);
+   const int sdim = SurfaceSpaceDim(surface_mesh_type);
+   mesh.SetCurvature(2, false, sdim);
+
+   const int npt = mesh.GetNE()*npt_per_el;
+   Vector xyz(npt*sdim);
+   GetSurfaceInteriorPoints(mesh, npt_per_el, Ordering::byVDIM, xyz);
+   // offset them to move away from the surface
+   const int off_d = (surface_mesh_type == SurfaceMeshType::Segment2D) ? 1 : 2;
+   for (int i = 0; i < npt; i++)
+   {
+      xyz(i*sdim + off_d) += offset;
+   }
+
+   FindPointsGSLIB finder;
+   finder.SetupSurf(mesh, 0.0);
+   finder.FindPoints(xyz, Ordering::byVDIM);
+   Array<unsigned int> code_no_pad = finder.GetCode();
+
+   for (int i = 0; i < code_no_pad.Size(); i++)
+   {
+      REQUIRE(code_no_pad[i] == 2);
+   }
+
+   // make aabb at least big enough to include the offset points
+   Vector aabb_sz_inc({2.1*offset});
+   finder.SetupSurfWithAABBExpansion(mesh, aabb_sz_inc);
+   finder.FindPoints(xyz, Ordering::byVDIM);
+   Array<unsigned int> code_with_pad = finder.GetCode();
+
+   for (int i = 0; i < npt; i++)
+   {
+      REQUIRE(code_with_pad[i] == 1);
    }
 }
 
@@ -322,9 +653,8 @@ TEST_CASE("GSLIBInterpolateL2ElementBoundary",
    mesh.SetCurvature(mesh_order);
 
    // Set GridFunction to be interpolated
-   FiniteElementCollection *c_fec = new L2_FECollection(3, dim);
-   FiniteElementSpace c_fespace =
-      FiniteElementSpace(&mesh, c_fec, 1);
+   L2_FECollection c_fec(3, dim);
+   FiniteElementSpace c_fespace(&mesh, &c_fec, 1);
    GridFunction field_vals(&c_fespace);
    Array<int> dofs;
    double leftval = 1.0;
@@ -366,7 +696,6 @@ TEST_CASE("GSLIBInterpolateL2ElementBoundary",
    REQUIRE(interp_vals(0) == MFEM_Approx(0.5*(leftval+rightval)));
 
    finder.FreeData();
-   delete c_fec;
 }
 
 #ifdef MFEM_USE_MPI

tests/unit/fem/test_linearform_ext.cpp

@@ -328,7 +328,7 @@ TEST_CASE("Linear Form Extension", "[LinearFormExtension], [GPU]")
    }
 }
 
-TEST_CASE("H(div) Linear Form Extension", "[LinearFormExtension], [GPU]")
+TEST_CASE("Vector FE Linear Form Extension", "[LinearFormExtension], [GPU]")
 {
    const bool all = launch_all_non_regression_tests;
 
@@ -341,26 +341,44 @@ TEST_CASE("H(div) Linear Form Extension", "[LinearFormExtension], [GPU]")
    Mesh mesh(mesh_file);
    const int dim = mesh.Dimension();
 
-   CAPTURE(mesh_file, dim, p);
+   {
+      const auto space_type =
+         dim == 3 ? GENERATE(FiniteElement::DIV, FiniteElement::CURL)
+         : FiniteElement::DIV;
 
-   RT_FECollection fec(p, dim);
-   FiniteElementSpace fes(&mesh, &fec);
+      CAPTURE(mesh_file, dim, p, space_type);
 
-   VectorFunctionCoefficient coeff(dim, fvec_dim);
+      std::unique_ptr<FiniteElementCollection> fec;
 
-   LinearForm d1(&fes);
-   d1.AddDomainIntegrator(new VectorFEDomainLFIntegrator(coeff));
-   d1.UseFastAssembly(true);
-   d1.Assemble();
+      switch (space_type)
+      {
+         case FiniteElement::DIV:
+            fec.reset(new RT_FECollection(p, dim));
+            break;
+         case FiniteElement::CURL:
+            fec.reset(new ND_FECollection(p, dim));
+            break;
+         default:
+            MFEM_ABORT("unsupported space type");
+      }
+      FiniteElementSpace fes(&mesh, fec.get());
 
-   LinearForm d2(&fes);
-   d2.AddDomainIntegrator(new VectorFEDomainLFIntegrator(coeff));
-   d2.UseFastAssembly(false);
-   d2.Assemble();
+      VectorFunctionCoefficient coeff(dim, fvec_dim);
 
-   CAPTURE(d1.Norml2(), d2.Norml2());
-   d1 -= d2;
-   REQUIRE(d1.Norml2() == MFEM_Approx(0.0));
+      LinearForm d1(&fes);
+      d1.AddDomainIntegrator(new VectorFEDomainLFIntegrator(coeff));
+      d1.UseFastAssembly(true);
+      d1.Assemble();
+
+      LinearForm d2(&fes);
+      d2.AddDomainIntegrator(new VectorFEDomainLFIntegrator(coeff));
+      d2.UseFastAssembly(false);
+      d2.Assemble();
+
+      CAPTURE(d1.Norml2(), d2.Norml2());
+      d1 -= d2;
+      REQUIRE(d1.Norml2() == MFEM_Approx(0.0));
+   }
 }
 
 #ifdef MFEM_USE_MPI

tests/unit/fem/test_pa_diagonal.cpp

@@ -320,7 +320,8 @@ double test_vdiag_pa(int dim, int order)
 }
 
 TEST_CASE("Vector Mass Diagonal PA",
-          "[AssembleDiagonal][PartialAssembly][VectorPA][VectorDiagonalPA][VectorMassPA][CUDA]")
+          "[AssembleDiagonal][PartialAssembly]"
+          "[VectorPA][VectorDiagonalPA][VectorMassPA][GPU]")
 {
    const auto DIM = GENERATE(2, 3);
    const auto P = GENERATE(1, 2, 3);
@@ -328,8 +329,48 @@ TEST_CASE("Vector Mass Diagonal PA",
    REQUIRE(test_vdiag_pa<VectorMassIntegrator>(DIM,P) == MFEM_Approx(0.0));
 }
 
+TEST_CASE("Vector Mass Diagonal PA accumulate",
+          "[AssembleDiagonal][PartialAssembly]"
+          "[VectorPA][VectorDiagonalPA][VectorMassPA][GPU]")
+{
+   const auto dim = GENERATE(2, 3);
+   const auto order = GENERATE(1, 2, 3);
+   CAPTURE(dim, order);
+
+   Mesh mesh;
+   if (dim == 2)
+   {
+      mesh = Mesh::MakeCartesian2D(1, 1, Element::QUADRILATERAL);
+   }
+   else
+   {
+      mesh = Mesh::MakeCartesian3D(1, 1, 1, Element::HEXAHEDRON);
+   }
+
+   H1_FECollection fec(order, dim);
+   FiniteElementSpace fes(&mesh, &fec, dim);
+
+   VectorMassIntegrator integ;
+   integ.AssemblePA(fes);
+
+   const int n = fes.GetVSize();
+   Vector base(n), from_zero(n), from_nonzero(n);
+   base.Randomize(1);
+
+   from_zero = 0.0;
+   integ.AssembleDiagonalPA(from_zero);
+
+   from_nonzero = base;
+   integ.AssembleDiagonalPA(from_nonzero);
+
+   from_nonzero -= base;
+   from_nonzero -= from_zero;
+   REQUIRE(from_nonzero.Normlinf() == MFEM_Approx(0.0));
+}
+
 TEST_CASE("Vector Diffusion Diagonal PA",
-          "[AssembleDiagonal][PartialAssembly][VectorPA][VectorDiagonalPA][VectorDiffusionPA][CUDA]")
+          "[AssembleDiagonal][PartialAssembly]"
+          "[VectorPA][VectorDiagonalPA][VectorDiffusionPA][GPU]")
 {
    const auto DIM = GENERATE(2, 3);
    const auto P = GENERATE(1, 2, 3);
@@ -337,6 +378,46 @@ TEST_CASE("Vector Diffusion Diagonal PA",
    REQUIRE(test_vdiag_pa<VectorDiffusionIntegrator>(DIM,P) == MFEM_Approx(0.0));
 }
 
+TEST_CASE("Elasticity Diagonal PA",
+          "[AssembleDiagonal][PartialAssembly][ElasticityPA][GPU]")
+{
+   const auto dim = GENERATE(2, 3);
+   const auto order = GENERATE(1, 2);
+   CAPTURE(dim, order);
+
+   Mesh mesh;
+   if (dim == 2)
+   {
+      mesh = Mesh::MakeCartesian2D(2, 2, Element::QUADRILATERAL, 0, 1.0, 1.0);
+   }
+   else
+   {
+      mesh = Mesh::MakeCartesian3D(2, 2, 2, Element::HEXAHEDRON, 1.0, 1.0, 1.0);
+   }
+
+   H1_FECollection fec(order, dim);
+   FiniteElementSpace fes(&mesh, &fec, dim, Ordering::byNODES);
+
+   ConstantCoefficient lambda(1.0), mu(1.0);
+
+   BilinearForm form_pa(&fes);
+   form_pa.SetAssemblyLevel(AssemblyLevel::PARTIAL);
+   form_pa.AddDomainIntegrator(new ElasticityIntegrator(lambda, mu));
+   form_pa.Assemble();
+
+   BilinearForm form_fa(&fes);
+   form_fa.AddDomainIntegrator(new ElasticityIntegrator(lambda, mu));
+   form_fa.Assemble();
+   form_fa.Finalize();
+
+   Vector diag_pa(fes.GetVSize()), diag_fa(fes.GetVSize());
+   form_pa.AssembleDiagonal(diag_pa);
+   form_fa.SpMat().GetDiag(diag_fa);
+
+   diag_fa -= diag_pa;
+   REQUIRE(diag_fa.Normlinf() == MFEM_Approx(0.0));
+}
+
 TEST_CASE("Hcurl/Hdiv diagonal PA",
           "[GPU][PartialAssembly][AssembleDiagonal]")
 {

tests/unit/fem/test_pa_kernels.cpp

@@ -17,6 +17,8 @@
 #include "unit_tests.hpp"
 #include "mfem.hpp"
 
+#include "fem/integ/bilininteg_vecmass_pa.hpp" // IWYU pragma: keep
+
 using namespace mfem;
 
 namespace pa_kernels
@@ -443,13 +445,76 @@ real_t test_pa_vector_integrator(int dim, int sdim)
    return y_fa.Norml2();
 }
 
+void test_pa_vector_mass_kernels(int dim, int p, int q_order)
+{
+   CAPTURE(dim, p, q_order);
+
+   Mesh mesh = MakeCartesianNonaligned(dim, 2);
+   mesh.SetCurvature(p, false, dim);
+   const IntegrationRule &ir =
+      IntRules.Get(mesh.GetTypicalElementGeometry(), q_order);
+
+   H1_FECollection fec(p, dim);
+   FiniteElementSpace fes(&mesh, &fec, dim);
+
+   GridFunction x(&fes), y_fa(&fes), y_pa(&fes);
+   x.Randomize(1);
+
+   BilinearForm blf_fa(&fes);
+   blf_fa.SetAssemblyLevel(AssemblyLevel::LEGACY);
+   // blf_fa will take ownership of integ_fa
+   auto *integ_fa = new VectorMassIntegrator;
+   integ_fa->SetIntRule(&ir);
+   blf_fa.AddDomainIntegrator(integ_fa);
+   blf_fa.Assemble();
+   blf_fa.Finalize();
+   blf_fa.Mult(x, y_fa);
+
+   Vector diag_fa(fes.GetVSize());
+   blf_fa.SpMat().GetDiag(diag_fa);
+
+   BilinearForm blf_pa(&fes);
+   blf_pa.SetAssemblyLevel(AssemblyLevel::PARTIAL);
+   // blf_pa will take ownership of integ_pa
+   auto *integ_pa = new VectorMassIntegrator;
+   integ_pa->SetIntRule(&ir);
+   blf_pa.AddDomainIntegrator(integ_pa);
+   blf_pa.Assemble();
+   blf_pa.Mult(x, y_pa);
+
+   Vector diag_pa(fes.GetVSize());
+   blf_pa.AssembleDiagonal(diag_pa);
+
+   y_fa -= y_pa;
+   diag_fa -= diag_pa;
+
+   REQUIRE(y_fa.Norml2() == MFEM_Approx(0.0));
+   REQUIRE(diag_fa.Normlinf() == MFEM_Approx(0.0));
+}
+
 TEST_CASE("PA Vector Mass",
           "[PartialAssembly][VectorPA][VectorMassPA][GPU]")
 {
-   const auto DIM = GENERATE(2, 3);
-   CAPTURE(DIM);
-   REQUIRE(test_pa_vector_integrator<VectorMassIntegrator>(DIM, DIM)
-           == MFEM_Approx(0.0));
+   using Mass = VectorMassIntegrator;
+
+   SECTION("built-in specializations")
+   {
+      const auto DIM = GENERATE(2, 3);
+      CAPTURE(DIM);
+      REQUIRE(test_pa_vector_integrator<Mass>(DIM, DIM) == MFEM_Approx(0.0));
+   }
+
+   SECTION("user specializations")
+   {
+      constexpr int user_dim = 2, user_d1d = 2, user_q1d = 9;
+      using AddMult = VectorMassIntegrator::VectorMassAddMultPA;
+      using Diag = VectorMassIntegrator::VectorMassAssembleDiagonalPA;
+      AddMult::Specialization<user_dim, user_d1d, user_q1d>::Add();
+      Diag::Specialization<user_dim, user_q1d>::Add();
+
+      constexpr int p = 1, q_order = 2*user_q1d - 1;
+      test_pa_vector_mass_kernels(user_dim, p, q_order);
+   }
 }
 
 TEST_CASE("PA Vector Diffusion",
@@ -462,7 +527,7 @@ TEST_CASE("PA Vector Diffusion",
 }
 
 TEST_CASE("PA Vector Diffusion 2D/3D",
-          "[PartialAssembly][VectorPA][VectorDiffusionPA][CUDA]")
+          "[PartialAssembly][VectorPA][VectorDiffusionPA][GPU]")
 {
    const int DIM = 2, SDIM = 3;
    CAPTURE(DIM, SDIM);

tests/unit/fem/test_pa_nlvc.cpp

@@ -0,0 +1,135 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#ifdef _WIN32
+#define _USE_MATH_DEFINES
+#include <cmath>
+#endif
+
+#include "unit_tests.hpp"
+#include "mfem.hpp"
+
+#include "fem/integ/nonlininteg_vecconvection_pa.hpp" // IWYU pragma: keep
+#include "fem/integ/nonlininteg_vecconvection_pa_diag.hpp" // IWYU pragma: keep
+#include "fem/integ/nonlininteg_vecconvection_pa_grad.hpp" // IWYU pragma: keep
+#include "fem/qinterp/grad.hpp" // IWYU pragma: keep
+
+using namespace mfem;
+
+namespace pa_kernels
+{
+
+template<int DIM>
+void test_nl_convection_pa_grad(const char *filename, int p)
+{
+   CAPTURE(filename, DIM, p);
+
+   Mesh mesh(filename);
+   MFEM_VERIFY(mesh.Dimension() == DIM, "Mesh dimension mismatch");
+
+   H1_FECollection fec(p, DIM);
+   FiniteElementSpace fes(&mesh, &fec, DIM);
+
+   GridFunction x(&fes), dx(&fes), y_fa(&fes), y_pa(&fes);
+   x.Randomize(0x100001b3);
+   dx.Randomize(0x9e3779b9);
+
+   ConstantCoefficient const_coeff(M_2_SQRTPI);
+   FunctionCoefficient funct_coeff([](const Vector &x)
+   { return M_1_PI + x[0] * x[0]; });
+
+   NonlinearForm nlf_fa(&fes);
+   nlf_fa.AddDomainIntegrator(new VectorConvectionNLFIntegrator);
+   nlf_fa.AddDomainIntegrator(new VectorConvectionNLFIntegrator(const_coeff));
+   nlf_fa.AddDomainIntegrator(new VectorConvectionNLFIntegrator(funct_coeff));
+
+   NonlinearForm nlf_pa(&fes);
+   nlf_pa.SetAssemblyLevel(AssemblyLevel::PARTIAL);
+   nlf_pa.AddDomainIntegrator(new VectorConvectionNLFIntegrator);
+   nlf_pa.AddDomainIntegrator(new VectorConvectionNLFIntegrator(const_coeff));
+   nlf_pa.AddDomainIntegrator(new VectorConvectionNLFIntegrator(funct_coeff));
+   nlf_pa.Setup();
+
+   SECTION("Action")
+   {
+      nlf_fa.Mult(x, y_fa), nlf_pa.Mult(x, y_pa);
+      y_fa -= y_pa;
+      REQUIRE(y_fa.Norml2() == MFEM_Approx(0.0));
+   }
+
+   SECTION("Gradient")
+   {
+      Operator &nlf_fa_grad = nlf_fa.GetGradient(x);
+      Operator &nlf_pa_grad = nlf_pa.GetGradient(x);
+      nlf_pa_grad.Mult(dx, y_pa);
+      nlf_fa_grad.Mult(dx, y_fa);
+      y_fa -= y_pa;
+      REQUIRE(y_fa.Norml2() == MFEM_Approx(0.0));
+   }
+
+   SECTION("Diagonal")
+   {
+      Vector diag_fa(fes.GetVSize()), diag_pa(fes.GetVSize());
+      dynamic_cast<SparseMatrix &>(nlf_fa.GetGradient(x)).GetDiag(diag_fa);
+      nlf_pa.GetGradient(x).AssembleDiagonal(diag_pa);
+      diag_fa -= diag_pa;
+      REQUIRE(diag_fa.Norml2() == MFEM_Approx(0.0));
+   }
+}
+
+TEST_CASE("NL Convection PA Gradient",
+          "[PartialAssembly][NonlinearPA][GPU][NLConv]")
+{
+   const auto p_base = {1, 2}, p_extra = {3, 4};
+   const auto p = MFEM_GENERATE_RANGES(p_base, p_extra);
+
+   using Grad = QuadratureInterpolator::GradKernels;
+   using NLVC = VectorConvectionNLFIntegrator;
+   static const auto specializations =
+      (Grad::Specialization<2, QVectorLayout::byNODES, false, 2, 2, 7>::Add(),
+       Grad::Specialization<2, QVectorLayout::byNODES, false, 2, 3, 7>::Add(),
+       Grad::Specialization<2, QVectorLayout::byNODES, false, 2, 4, 8>::Add(),
+       Grad::Specialization<3, QVectorLayout::byNODES, false, 3, 2, 7>::Add(),
+       Grad::Specialization<3, QVectorLayout::byNODES, false, 3, 3, 7>::Add(),
+       Grad::Specialization<3, QVectorLayout::byNODES, false, 3, 3, 8>::Add(),
+       Grad::Specialization<3, QVectorLayout::byNODES, false, 3, 4, 5>::Add(),
+       Grad::Specialization<3, QVectorLayout::byNODES, false, 3, 4, 9>::Add(),
+       // User specialization example for p=4 (D1D=5, Q1D=9) in 3D
+       NLVC::AddSpecialization<3, 5, 9>(),
+       true);
+   MFEM_CONTRACT_VAR(specializations);
+
+   SECTION("2D")
+   {
+      const auto meshs = { "../../data/inline-quad.mesh" };
+      const auto extra = { "../../data/star-q2.mesh",
+                           "../../data/star-q3.mesh",
+                           "../../data/rt-2d-q3.mesh",
+                           "../../data/periodic-square.mesh"
+                         };
+      test_nl_convection_pa_grad<2>(MFEM_GENERATE_RANGES(meshs, extra), p);
+   }
+
+   SECTION("3D")
+   {
+      const auto meshs = { "../../data/inline-hex.mesh" };
+      const auto extra = { "../../data/fichera.mesh",
+                           "../../data/beam-hex.mesh",
+                           "../../data/toroid-hex.mesh",
+                           "../../data/fichera-q2.mesh",
+                           "../../data/fichera-q3.mesh",
+                           "../../data/periodic-cube.mesh"
+                         };
+      test_nl_convection_pa_grad<3>(MFEM_GENERATE_RANGES(meshs, extra), p);
+   }
+}
+
+} // namespace pa_kernels

tests/unit/fem/test_pa_vecdiv.cpp

@@ -0,0 +1,153 @@
+// Copyright (c) 2010-2025, Lawrence Livermore National Security, LLC. Produced
+// at the Lawrence Livermore National Laboratory. All Rights reserved. See files
+// LICENSE and NOTICE for details. LLNL-CODE-806117.
+//
+// This file is part of the MFEM library. For more information and source code
+// availability visit https://mfem.org.
+//
+// MFEM is free software; you can redistribute it and/or modify it under the
+// terms of the BSD-3 license. We welcome feedback and contributions, see file
+// CONTRIBUTING.md for details.
+
+#ifdef _WIN32
+#define _USE_MATH_DEFINES
+#include <cmath>
+#endif
+
+#include "unit_tests.hpp"
+#include "mfem.hpp"
+
+#include "fem/integ/bilininteg_vecdiv_pa.hpp" // IWYU pragma: keep
+#include "fem/qinterp/grad.hpp" // IWYU pragma: keep
+
+#include <algorithm>
+#include <utility>
+
+using namespace mfem;
+
+namespace pa_kernels
+{
+
+template <typename INTEGRATOR, bool TRANSPOSE>
+void pa_mixed_test(FiniteElementSpace &fes1,
+                   FiniteElementSpace &fes2,
+                   const IntegrationRule &ir)
+{
+   MixedBilinearForm bform_pa(&fes1, &fes2);
+   if constexpr (TRANSPOSE)
+   {
+      auto *integ = new TransposeIntegrator(new INTEGRATOR);
+      integ->SetIntRule(&ir);
+      bform_pa.AddDomainIntegrator(integ);
+   }
+   else
+   {
+      auto *integ = new INTEGRATOR;
+      integ->SetIntRule(&ir);
+      bform_pa.AddDomainIntegrator(integ);
+   }
+   bform_pa.SetAssemblyLevel(AssemblyLevel::PARTIAL);
+   bform_pa.Assemble();
+
+   MixedBilinearForm bform_fa(&fes1, &fes2);
+   if constexpr (TRANSPOSE)
+   {
+      auto *integ = new TransposeIntegrator(new INTEGRATOR);
+      integ->SetIntRule(&ir);
+      bform_fa.AddDomainIntegrator(integ);
+   }
+   else
+   {
+      auto *integ = new INTEGRATOR;
+      integ->SetIntRule(&ir);
+      bform_fa.AddDomainIntegrator(integ);
+   }
+   bform_fa.Assemble();
+   bform_fa.Finalize();
+
+   GridFunction x(&fes1), y_pa(&fes2), y_fa(&fes2);
+   x.Randomize(0x100001b3);
+
+   bform_pa.Mult(x, y_pa);
+   bform_fa.Mult(x, y_fa);
+
+   y_pa -= y_fa;
+   REQUIRE(y_pa.Normlinf() == MFEM_Approx(0.0));
+}
+
+template<int DIM>
+void test_pa_divergence(const char *filename, int vp, int sp)
+{
+   CAPTURE(filename, DIM, vp, sp);
+
+   Mesh mesh(filename);
+   MFEM_VERIFY(mesh.Dimension() == DIM, "Mesh dimension mismatch");
+
+   // Vector
+   H1_FECollection vfec(vp, DIM);
+   FiniteElementSpace vfes(&mesh, &vfec, DIM);
+
+   // Scalar
+   H1_FECollection sfec(sp, DIM);
+   FiniteElementSpace sfes(&mesh, &sfec);
+
+   // Shared-memory PA kernels require q1d >= max(trial_d1d, test_d1d)
+   const auto &trial_fe = *vfes.GetTypicalFE();
+   const auto &test_fe = *sfes.GetTypicalFE();
+   const auto &Trans = *mesh.GetTypicalElementTransformation();
+   int order = Trans.OrderGrad(&trial_fe) + test_fe.GetOrder() + Trans.OrderJ();
+   const int min_q1d = std::max(vp, sp) + 1;
+   order = std::max(order, 2 * min_q1d - 1);
+   const IntegrationRule &ir = IntRules.Get(trial_fe.GetGeomType(), order);
+
+   pa_mixed_test<VectorDivergenceIntegrator, false>(vfes, sfes, ir);
+   pa_mixed_test<VectorDivergenceIntegrator, true>(sfes, vfes, ir);
+}
+
+TEST_CASE("VecDivPA", "[PartialAssembly][VecDivPA][GPU]")
+{
+   if (static auto done = false; !std::exchange(done, true))
+   {
+      using Grad = QuadratureInterpolator::GradKernels;
+      Grad::Specialization<2, QVectorLayout::byNODES, false, 2, 3, 5>::Add();
+      Grad::Specialization<3, QVectorLayout::byNODES, false, 3, 3, 7>::Add();
+      Grad::Specialization<3, QVectorLayout::byNODES, false, 3, 4, 9>::Add();
+
+      using VDiv = VectorDivergenceIntegrator::VectorDivergenceAddMultPA;
+      using VDivT = VectorDivergenceIntegrator::VectorDivergenceAddMultTransposePA;
+      VDiv::Specialization<2, 2, 3, 3>::Add();
+      VDivT::Specialization<2, 2, 3, 3>::Add();
+   }
+
+   // Vector (vp) and scalar (sp) space orders
+   const auto vp_base = {1, 2}, vp_extra = {3, 4};
+   const auto sp_base = {1, 2}, sp_extra = {3, 4};
+   const auto vp = MFEM_GENERATE_RANGES(vp_base, vp_extra);
+   const auto sp = MFEM_GENERATE_RANGES(sp_base, sp_extra);
+
+   SECTION("2D")
+   {
+      const auto meshs = { "../../data/inline-quad.mesh" };
+      const auto extra = { "../../data/star-q2.mesh",
+                           "../../data/star-q3.mesh",
+                           "../../data/rt-2d-q3.mesh",
+                           "../../data/periodic-square.mesh"
+                         };
+      test_pa_divergence<2>(MFEM_GENERATE_RANGES(meshs, extra), vp, sp);
+   }
+
+   SECTION("3D")
+   {
+      const auto meshs = { "../../data/inline-hex.mesh" };
+      const auto extra = { "../../data/fichera.mesh",
+                           "../../data/beam-hex.mesh",
+                           "../../data/toroid-hex.mesh",
+                           "../../data/fichera-q2.mesh",
+                           "../../data/fichera-q3.mesh",
+                           "../../data/periodic-cube.mesh"
+                         };
+      test_pa_divergence<3>(MFEM_GENERATE_RANGES(meshs, extra), vp, sp);
+   }
+}
+
+} // namespace pa_kernels

tests/unit/fem/test_particleset.cpp

@@ -200,10 +200,71 @@ int CheckArrayEquality(const Array<T> &arr1, const Array<T> &arr2)
    return wrong_ct;
 }
 
+// Apply a deterministic perturbation to particle data on host.
+void PerturbParticleDataOnHost(std::vector<Particle> &particles)
+{
+   for (auto &p : particles)
+   {
+      for (int f = -1; f < p.GetNFields(); f++)
+      {
+         Vector &field = f == -1 ? p.Coords() : p.Field(f);
+         field.HostReadWrite();
+         const real_t scale = (f == -1) ? 0.001 : 1.0;
+         for (int c = 0; c < field.Size(); c++)
+         {
+            field(c) += scale * (f + c + 2);
+         }
+      }
+
+      for (int t = 0; t < p.GetNTags(); t++)
+      {
+         p.Tag(t) += t + 1;
+      }
+   }
+}
+
+// Apply a deterministic perturbation to particle data on device.
+void PerturbParticleDataOnDevice(ParticleSet &pset)
+{
+   const int np = pset.GetNParticles();
+
+   // Shift coordinates and fields using the same per-component formula while
+   // honoring the ParticleVector ordering selected by the test.
+   for (int f = -1; f < pset.GetNFields(); f++)
+   {
+      ParticleVector &field = f == -1 ? pset.Coords() : pset.Field(f);
+      const int vdim = field.GetVDim();
+      const bool by_vdim = (field.GetOrdering() == Ordering::byVDIM);
+      const real_t scale = (f == -1) ? 0.001 : 1.0;
+      auto d_field = field.ReadWrite();
+
+      mfem::forall(np, [=] MFEM_HOST_DEVICE (int i)
+      {
+         for (int c = 0; c < vdim; c++)
+         {
+            const int idx = by_vdim ? i * vdim + c : i + c * np;
+            d_field[idx] += scale * (f + c + 2);
+         }
+      });
+   }
+
+   for (int t = 0; t < pset.GetNTags(); t++)
+   {
+      Array<int> &tag = pset.Tag(t);
+      auto d_tag = tag.ReadWrite();
+
+      mfem::forall(np, [=] MFEM_HOST_DEVICE (int i)
+      {
+         d_tag[i] += t + 1;
+      });
+   }
+}
+
 void TestRedistribute(Ordering::Type ordering)
 {
    int size = Mpi::WorldSize();
    int rank = Mpi::WorldRank();
+   const bool use_device = Device::IsEnabled();
 
    // Create a 3D hex mesh
    Mesh m = Mesh::MakeCartesian3D(N_e, N_e, N_e, Element::Type::HEXAHEDRON);
@@ -252,15 +313,22 @@ void TestRedistribute(Ordering::Type ordering)
    SECTION(std::string("Ordering: ") +
            (ordering == Ordering::byNODES ? "byNODES" : "byVDIM"))
    {
-      // Add the particles uniquely to each rank particleset
       ParticleSet pset(MPI_COMM_WORLD, 0, SpaceDim, FieldVDims,
-                       NumTags, ordering);
+                       NumTags, ordering, use_device);
+      CHECK(pset.IsParticleRefValid() ==
+            (!use_device && ordering == Ordering::byVDIM));
 
       for (int i = 0; i < N_rank; i++)
       {
          pset.AddParticle(all_particles[i*size+rank]);
       }
 
+      if (use_device)
+      {
+         PerturbParticleDataOnDevice(pset);
+         PerturbParticleDataOnHost(all_particles);
+      }
+
       // Find points
       FindPointsGSLIB finder(MPI_COMM_WORLD);
       finder.Setup(pmesh);
@@ -270,6 +338,7 @@ void TestRedistribute(Ordering::Type ordering)
       int code_1_count = 0;
       int code_2_count = 0;
       const Array<unsigned int> &code = finder.GetCode();
+      code.HostRead();
       for (int i = 0; i < code.Size(); i++)
       {
          if (code[i] == 1)
@@ -292,6 +361,7 @@ void TestRedistribute(Ordering::Type ordering)
       finder.FindPoints(pset.Coords(), ordering);
 
       const Array<unsigned int> &procs = finder.GetProc();
+      procs.HostRead();
 
       int wrong_proc_count = 0;
       for (int i = 0; i < procs.Size(); i++)
@@ -307,6 +377,11 @@ void TestRedistribute(Ordering::Type ordering)
 
       // Check that coordinates + fields + tags are all still correct
       int wrong_particle_count = 0;
+      pset.GetIDs().HostRead();
+      for (int t = 0; t < pset.GetNTags(); t++)
+      {
+         pset.Tag(t).HostRead();
+      }
       for (int i = 0; i < pset.GetNParticles(); i++)
       {
          Particle &actual_p = all_particles[pset.GetIDs()[i]];
@@ -317,13 +392,13 @@ void TestRedistribute(Ordering::Type ordering)
             wrong_particle_count++;
          }
       }
-      MPI_Allreduce(MPI_IN_PLACE, &wrong_proc_count, 1, MPI_INT, MPI_SUM,
+      MPI_Allreduce(MPI_IN_PLACE, &wrong_particle_count, 1, MPI_INT, MPI_SUM,
                     MPI_COMM_WORLD);
       CHECK(wrong_particle_count == 0);
    }
 }
 
-TEST_CASE("Particle Redistribution", "[ParticleSet][Parallel]")
+TEST_CASE("Particle Redistribution", "[ParticleSet][Parallel][GPU]")
 {
    TestRedistribute(Ordering::byNODES);
    TestRedistribute(Ordering::byVDIM);

tests/unit/linalg/test_complex_operator.cpp

@@ -113,3 +113,65 @@ TEST_CASE("ComplexOperator Quaternion Tests", "[ComplexOperator]")
       REQUIRE(qikx.Normlinf() < tol);
    }
 }
+
+#ifdef MFEM_USE_MPI
+
+TEST_CASE("ComplexHypreParMatrix GetSystemMatrix",
+          "[ComplexOperator][Parallel][GPU]")
+{
+   // This test reproduces the issue described in PR #5200 on GitHub. See also
+   // the follow up PR #5346.
+
+   // 1. Construct ComplexHypreParMatrix similar to ex25p.
+   const char mesh_file[] = "../../data/inline-quad.mesh";
+   Mesh *mesh = new Mesh(mesh_file, 1, 1);
+   int dim = mesh->Dimension();
+   int ref_levels = 1;
+   for (int l = 0; l < ref_levels; l++)
+   {
+      mesh->UniformRefinement();
+   }
+   ParMesh pmesh(MPI_COMM_WORLD, *mesh);
+   delete mesh;
+   int par_ref_levels = 1;
+   for (int l = 0; l < par_ref_levels; l++)
+   {
+      pmesh.UniformRefinement();
+   }
+   int order = 1;
+   ND_FECollection fec(order, dim);
+   ParFiniteElementSpace fespace(&pmesh, &fec);
+   Array<int> ess_tdof_list;
+   Array<int> ess_bdr;
+   if (pmesh.bdr_attributes.Size())
+   {
+      ess_bdr.SetSize(pmesh.bdr_attributes.Max());
+      ess_bdr = 1;
+   }
+   fespace.GetEssentialTrueDofs(ess_bdr, ess_tdof_list);
+   ComplexOperator::Convention conv = ComplexOperator::HERMITIAN;
+   VectorConstantCoefficient f(Vector{1_r, 2_r});
+   ParComplexLinearForm b(&fespace, conv);
+   b.AddDomainIntegrator(NULL, new VectorFEDomainLFIntegrator(f));
+   b = 0.0;
+   b.Assemble();
+   ParComplexGridFunction x(&fespace);
+   x = 0.0;
+   ConstantCoefficient one(1_r);
+   ParSesquilinearForm a(&fespace, conv);
+   a.AddDomainIntegrator(new CurlCurlIntegrator(one),
+                         new CurlCurlIntegrator(one));
+   a.AddDomainIntegrator(new VectorFEMassIntegrator(one),
+                         new VectorFEMassIntegrator(one));
+   a.Assemble();
+   OperatorPtr Ah;
+   Vector B, X;
+   a.FormLinearSystem(ess_tdof_list, x, b, Ah, X, B);
+
+   // 2. Test the call to ComplexHypreParMatrix::GetSystemMatrix and destroying
+   //    the returned matrix.
+   HypreParMatrix *A = Ah.As<ComplexHypreParMatrix>()->GetSystemMatrix();
+   delete A;
+}
+
+#endif // MFEM_USE_MPI

tests/unit/unit_tests.hpp

@@ -36,4 +36,11 @@ inline Approx MFEM_Approx(double val,
    return Approx(val).margin(abs_tol).epsilon(rel_tol);
 }
 
+/** @brief Generate values from @a base, and also from @a extra if the
+    command line '--all' option is provided. */
+#define MFEM_GENERATE_RANGES(base, extra) \
+   (!launch_all_non_regression_tests \
+    ? GENERATE_COPY(from_range(base)) \
+    : GENERATE_COPY(from_range(base), from_range(extra)))
+
 #endif